fs/ceph/mds_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/fs.h>
   5 #include <linux/wait.h>
   6 #include <linux/slab.h>
   7 #include <linux/gfp.h>
   8 #include <linux/sched.h>
   9 #include <linux/debugfs.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/bits.h>
  13 #include <linux/ktime.h>
  14 #include <linux/bitmap.h>
  15 #include <linux/mnt_idmapping.h>
  16
  17 #include "super.h"
  18 #include "mds_client.h"
  19 #include "crypto.h"
  20
  21 #include <linux/ceph/ceph_features.h>
  22 #include <linux/ceph/messenger.h>
  23 #include <linux/ceph/decode.h>
  24 #include <linux/ceph/pagelist.h>
  25 #include <linux/ceph/auth.h>
  26 #include <linux/ceph/debugfs.h>
  27
  28 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
  29
  30 /*
  31  * A cluster of MDS (metadata server) daemons is responsible for
  32  * managing the file system namespace (the directory hierarchy and
  33  * inodes) and for coordinating shared access to storage.  Metadata is
  34  * partitioning hierarchically across a number of servers, and that
  35  * partition varies over time as the cluster adjusts the distribution
  36  * in order to balance load.
  37  *
  38  * The MDS client is primarily responsible to managing synchronous
  39  * metadata requests for operations like open, unlink, and so forth.
  40  * If there is a MDS failure, we find out about it when we (possibly
  41  * request and) receive a new MDS map, and can resubmit affected
  42  * requests.
  43  *
  44  * For the most part, though, we take advantage of a lossless
  45  * communications channel to the MDS, and do not need to worry about
  46  * timing out or resubmitting requests.
  47  *
  48  * We maintain a stateful "session" with each MDS we interact with.
  49  * Within each session, we sent periodic heartbeat messages to ensure
  50  * any capabilities or leases we have been issues remain valid.  If
  51  * the session times out and goes stale, our leases and capabilities
  52  * are no longer valid.
  53  */
  54
  55 struct ceph_reconnect_state {
  56         struct ceph_mds_session *session;
  57         int nr_caps, nr_realms;
  58         struct ceph_pagelist *pagelist;
  59         unsigned msg_version;
  60         bool allow_multi;
  61 };
  62
  63 static void __wake_requests(struct ceph_mds_client *mdsc,
  64                             struct list_head *head);
  65 static void ceph_cap_release_work(struct work_struct *work);
  66 static void ceph_cap_reclaim_work(struct work_struct *work);
  67
  68 static const struct ceph_connection_operations mds_con_ops;
  69
  70
  71 /*
  72  * mds reply parsing
  73  */
  74
  75 static int parse_reply_info_quota(void **p, void *end,
  76                                   struct ceph_mds_reply_info_in *info)
  77 {
  78         u8 struct_v, struct_compat;
  79         u32 struct_len;
  80
  81         ceph_decode_8_safe(p, end, struct_v, bad);
  82         ceph_decode_8_safe(p, end, struct_compat, bad);
  83         /* struct_v is expected to be >= 1. we only
  84          * understand encoding with struct_compat == 1. */
  85         if (!struct_v || struct_compat != 1)
  86                 goto bad;
  87         ceph_decode_32_safe(p, end, struct_len, bad);
  88         ceph_decode_need(p, end, struct_len, bad);
  89         end = *p + struct_len;
  90         ceph_decode_64_safe(p, end, info->max_bytes, bad);
  91         ceph_decode_64_safe(p, end, info->max_files, bad);
  92         *p = end;
  93         return 0;
  94 bad:
  95         return -EIO;
  96 }
  97
  98 /*
  99  * parse individual inode info
 100  */
 101 static int parse_reply_info_in(void **p, void *end,
 102                                struct ceph_mds_reply_info_in *info,
 103                                u64 features)
 104 {
 105         int err = 0;
 106         u8 struct_v = 0;
 107
 108         if (features == (u64)-1) {
 109                 u32 struct_len;
 110                 u8 struct_compat;
 111                 ceph_decode_8_safe(p, end, struct_v, bad);
 112                 ceph_decode_8_safe(p, end, struct_compat, bad);
 113                 /* struct_v is expected to be >= 1. we only understand
 114                  * encoding with struct_compat == 1. */
 115                 if (!struct_v || struct_compat != 1)
 116                         goto bad;
 117                 ceph_decode_32_safe(p, end, struct_len, bad);
 118                 ceph_decode_need(p, end, struct_len, bad);
 119                 end = *p + struct_len;
 120         }
 121
 122         ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
 123         info->in = *p;
 124         *p += sizeof(struct ceph_mds_reply_inode) +
 125                 sizeof(*info->in->fragtree.splits) *
 126                 le32_to_cpu(info->in->fragtree.nsplits);
 127
 128         ceph_decode_32_safe(p, end, info->symlink_len, bad);
 129         ceph_decode_need(p, end, info->symlink_len, bad);
 130         info->symlink = *p;
 131         *p += info->symlink_len;
 132
 133         ceph_decode_copy_safe(p, end, &info->dir_layout,
 134                               sizeof(info->dir_layout), bad);
 135         ceph_decode_32_safe(p, end, info->xattr_len, bad);
 136         ceph_decode_need(p, end, info->xattr_len, bad);
 137         info->xattr_data = *p;
 138         *p += info->xattr_len;
 139
 140         if (features == (u64)-1) {
 141                 /* inline data */
 142                 ceph_decode_64_safe(p, end, info->inline_version, bad);
 143                 ceph_decode_32_safe(p, end, info->inline_len, bad);
 144                 ceph_decode_need(p, end, info->inline_len, bad);
 145                 info->inline_data = *p;
 146                 *p += info->inline_len;
 147                 /* quota */
 148                 err = parse_reply_info_quota(p, end, info);
 149                 if (err < 0)
 150                         goto out_bad;
 151                 /* pool namespace */
 152                 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 153                 if (info->pool_ns_len > 0) {
 154                         ceph_decode_need(p, end, info->pool_ns_len, bad);
 155                         info->pool_ns_data = *p;
 156                         *p += info->pool_ns_len;
 157                 }
 158
 159                 /* btime */
 160                 ceph_decode_need(p, end, sizeof(info->btime), bad);
 161                 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 162
 163                 /* change attribute */
 164                 ceph_decode_64_safe(p, end, info->change_attr, bad);
 165
 166                 /* dir pin */
 167                 if (struct_v >= 2) {
 168                         ceph_decode_32_safe(p, end, info->dir_pin, bad);
 169                 } else {
 170                         info->dir_pin = -ENODATA;
 171                 }
 172
 173                 /* snapshot birth time, remains zero for v<=2 */
 174                 if (struct_v >= 3) {
 175                         ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
 176                         ceph_decode_copy(p, &info->snap_btime,
 177                                          sizeof(info->snap_btime));
 178                 } else {
 179                         memset(&info->snap_btime, 0, sizeof(info->snap_btime));
 180                 }
 181
 182                 /* snapshot count, remains zero for v<=3 */
 183                 if (struct_v >= 4) {
 184                         ceph_decode_64_safe(p, end, info->rsnaps, bad);
 185                 } else {
 186                         info->rsnaps = 0;
 187                 }
 188
 189                 if (struct_v >= 5) {
 190                         u32 alen;
 191
 192                         ceph_decode_32_safe(p, end, alen, bad);
 193
 194                         while (alen--) {
 195                                 u32 len;
 196
 197                                 /* key */
 198                                 ceph_decode_32_safe(p, end, len, bad);
 199                                 ceph_decode_skip_n(p, end, len, bad);
 200                                 /* value */
 201                                 ceph_decode_32_safe(p, end, len, bad);
 202                                 ceph_decode_skip_n(p, end, len, bad);
 203                         }
 204                 }
 205
 206                 /* fscrypt flag -- ignore */
 207                 if (struct_v >= 6)
 208                         ceph_decode_skip_8(p, end, bad);
 209
 210                 info->fscrypt_auth = NULL;
 211                 info->fscrypt_auth_len = 0;
 212                 info->fscrypt_file = NULL;
 213                 info->fscrypt_file_len = 0;
 214                 if (struct_v >= 7) {
 215                         ceph_decode_32_safe(p, end, info->fscrypt_auth_len, bad);
 216                         if (info->fscrypt_auth_len) {
 217                                 info->fscrypt_auth = kmalloc(info->fscrypt_auth_len,
 218                                                              GFP_KERNEL);
 219                                 if (!info->fscrypt_auth)
 220                                         return -ENOMEM;
 221                                 ceph_decode_copy_safe(p, end, info->fscrypt_auth,
 222                                                       info->fscrypt_auth_len, bad);
 223                         }
 224                         ceph_decode_32_safe(p, end, info->fscrypt_file_len, bad);
 225                         if (info->fscrypt_file_len) {
 226                                 info->fscrypt_file = kmalloc(info->fscrypt_file_len,
 227                                                              GFP_KERNEL);
 228                                 if (!info->fscrypt_file)
 229                                         return -ENOMEM;
 230                                 ceph_decode_copy_safe(p, end, info->fscrypt_file,
 231                                                       info->fscrypt_file_len, bad);
 232                         }
 233                 }
 234                 *p = end;
 235         } else {
 236                 /* legacy (unversioned) struct */
 237                 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
 238                         ceph_decode_64_safe(p, end, info->inline_version, bad);
 239                         ceph_decode_32_safe(p, end, info->inline_len, bad);
 240                         ceph_decode_need(p, end, info->inline_len, bad);
 241                         info->inline_data = *p;
 242                         *p += info->inline_len;
 243                 } else
 244                         info->inline_version = CEPH_INLINE_NONE;
 245
 246                 if (features & CEPH_FEATURE_MDS_QUOTA) {
 247                         err = parse_reply_info_quota(p, end, info);
 248                         if (err < 0)
 249                                 goto out_bad;
 250                 } else {
 251                         info->max_bytes = 0;
 252                         info->max_files = 0;
 253                 }
 254
 255                 info->pool_ns_len = 0;
 256                 info->pool_ns_data = NULL;
 257                 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
 258                         ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
 259                         if (info->pool_ns_len > 0) {
 260                                 ceph_decode_need(p, end, info->pool_ns_len, bad);
 261                                 info->pool_ns_data = *p;
 262                                 *p += info->pool_ns_len;
 263                         }
 264                 }
 265
 266                 if (features & CEPH_FEATURE_FS_BTIME) {
 267                         ceph_decode_need(p, end, sizeof(info->btime), bad);
 268                         ceph_decode_copy(p, &info->btime, sizeof(info->btime));
 269                         ceph_decode_64_safe(p, end, info->change_attr, bad);
 270                 }
 271
 272                 info->dir_pin = -ENODATA;
 273                 /* info->snap_btime and info->rsnaps remain zero */
 274         }
 275         return 0;
 276 bad:
 277         err = -EIO;
 278 out_bad:
 279         return err;
 280 }
 281
 282 static int parse_reply_info_dir(void **p, void *end,
 283                                 struct ceph_mds_reply_dirfrag **dirfrag,
 284                                 u64 features)
 285 {
 286         if (features == (u64)-1) {
 287                 u8 struct_v, struct_compat;
 288                 u32 struct_len;
 289                 ceph_decode_8_safe(p, end, struct_v, bad);
 290                 ceph_decode_8_safe(p, end, struct_compat, bad);
 291                 /* struct_v is expected to be >= 1. we only understand
 292                  * encoding whose struct_compat == 1. */
 293                 if (!struct_v || struct_compat != 1)
 294                         goto bad;
 295                 ceph_decode_32_safe(p, end, struct_len, bad);
 296                 ceph_decode_need(p, end, struct_len, bad);
 297                 end = *p + struct_len;
 298         }
 299
 300         ceph_decode_need(p, end, sizeof(**dirfrag), bad);
 301         *dirfrag = *p;
 302         *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
 303         if (unlikely(*p > end))
 304                 goto bad;
 305         if (features == (u64)-1)
 306                 *p = end;
 307         return 0;
 308 bad:
 309         return -EIO;
 310 }
 311
 312 static int parse_reply_info_lease(void **p, void *end,
 313                                   struct ceph_mds_reply_lease **lease,
 314                                   u64 features, u32 *altname_len, u8 **altname)
 315 {
 316         u8 struct_v;
 317         u32 struct_len;
 318         void *lend;
 319
 320         if (features == (u64)-1) {
 321                 u8 struct_compat;
 322
 323                 ceph_decode_8_safe(p, end, struct_v, bad);
 324                 ceph_decode_8_safe(p, end, struct_compat, bad);
 325
 326                 /* struct_v is expected to be >= 1. we only understand
 327                  * encoding whose struct_compat == 1. */
 328                 if (!struct_v || struct_compat != 1)
 329                         goto bad;
 330
 331                 ceph_decode_32_safe(p, end, struct_len, bad);
 332         } else {
 333                 struct_len = sizeof(**lease);
 334                 *altname_len = 0;
 335                 *altname = NULL;
 336         }
 337
 338         lend = *p + struct_len;
 339         ceph_decode_need(p, end, struct_len, bad);
 340         *lease = *p;
 341         *p += sizeof(**lease);
 342
 343         if (features == (u64)-1) {
 344                 if (struct_v >= 2) {
 345                         ceph_decode_32_safe(p, end, *altname_len, bad);
 346                         ceph_decode_need(p, end, *altname_len, bad);
 347                         *altname = *p;
 348                         *p += *altname_len;
 349                 } else {
 350                         *altname = NULL;
 351                         *altname_len = 0;
 352                 }
 353         }
 354         *p = lend;
 355         return 0;
 356 bad:
 357         return -EIO;
 358 }
 359
 360 /*
 361  * parse a normal reply, which may contain a (dir+)dentry and/or a
 362  * target inode.
 363  */
 364 static int parse_reply_info_trace(void **p, void *end,
 365                                   struct ceph_mds_reply_info_parsed *info,
 366                                   u64 features)
 367 {
 368         int err;
 369
 370         if (info->head->is_dentry) {
 371                 err = parse_reply_info_in(p, end, &info->diri, features);
 372                 if (err < 0)
 373                         goto out_bad;
 374
 375                 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
 376                 if (err < 0)
 377                         goto out_bad;
 378
 379                 ceph_decode_32_safe(p, end, info->dname_len, bad);
 380                 ceph_decode_need(p, end, info->dname_len, bad);
 381                 info->dname = *p;
 382                 *p += info->dname_len;
 383
 384                 err = parse_reply_info_lease(p, end, &info->dlease, features,
 385                                              &info->altname_len, &info->altname);
 386                 if (err < 0)
 387                         goto out_bad;
 388         }
 389
 390         if (info->head->is_target) {
 391                 err = parse_reply_info_in(p, end, &info->targeti, features);
 392                 if (err < 0)
 393                         goto out_bad;
 394         }
 395
 396         if (unlikely(*p != end))
 397                 goto bad;
 398         return 0;
 399
 400 bad:
 401         err = -EIO;
 402 out_bad:
 403         pr_err("problem parsing mds trace %d\n", err);
 404         return err;
 405 }
 406
 407 /*
 408  * parse readdir results
 409  */
 410 static int parse_reply_info_readdir(void **p, void *end,
 411                                     struct ceph_mds_request *req,
 412                                     u64 features)
 413 {
 414         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 415         struct ceph_client *cl = req->r_mdsc->fsc->client;
 416         u32 num, i = 0;
 417         int err;
 418
 419         err = parse_reply_info_dir(p, end, &info->dir_dir, features);
 420         if (err < 0)
 421                 goto out_bad;
 422
 423         ceph_decode_need(p, end, sizeof(num) + 2, bad);
 424         num = ceph_decode_32(p);
 425         {
 426                 u16 flags = ceph_decode_16(p);
 427                 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
 428                 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
 429                 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
 430                 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
 431         }
 432         if (num == 0)
 433                 goto done;
 434
 435         BUG_ON(!info->dir_entries);
 436         if ((unsigned long)(info->dir_entries + num) >
 437             (unsigned long)info->dir_entries + info->dir_buf_size) {
 438                 pr_err_client(cl, "dir contents are larger than expected\n");
 439                 WARN_ON(1);
 440                 goto bad;
 441         }
 442
 443         info->dir_nr = num;
 444         while (num) {
 445                 struct inode *inode = d_inode(req->r_dentry);
 446                 struct ceph_inode_info *ci = ceph_inode(inode);
 447                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 448                 struct fscrypt_str tname = FSTR_INIT(NULL, 0);
 449                 struct fscrypt_str oname = FSTR_INIT(NULL, 0);
 450                 struct ceph_fname fname;
 451                 u32 altname_len, _name_len;
 452                 u8 *altname, *_name;
 453
 454                 /* dentry */
 455                 ceph_decode_32_safe(p, end, _name_len, bad);
 456                 ceph_decode_need(p, end, _name_len, bad);
 457                 _name = *p;
 458                 *p += _name_len;
 459                 doutc(cl, "parsed dir dname '%.*s'\n", _name_len, _name);
 460
 461                 if (info->hash_order)
 462                         rde->raw_hash = ceph_str_hash(ci->i_dir_layout.dl_dir_hash,
 463                                                       _name, _name_len);
 464
 465                 /* dentry lease */
 466                 err = parse_reply_info_lease(p, end, &rde->lease, features,
 467                                              &altname_len, &altname);
 468                 if (err)
 469                         goto out_bad;
 470
 471                 /*
 472                  * Try to dencrypt the dentry names and update them
 473                  * in the ceph_mds_reply_dir_entry struct.
 474                  */
 475                 fname.dir = inode;
 476                 fname.name = _name;
 477                 fname.name_len = _name_len;
 478                 fname.ctext = altname;
 479                 fname.ctext_len = altname_len;
 480                 /*
 481                  * The _name_len maybe larger than altname_len, such as
 482                  * when the human readable name length is in range of
 483                  * (CEPH_NOHASH_NAME_MAX, CEPH_NOHASH_NAME_MAX + SHA256_DIGEST_SIZE),
 484                  * then the copy in ceph_fname_to_usr will corrupt the
 485                  * data if there has no encryption key.
 486                  *
 487                  * Just set the no_copy flag and then if there has no
 488                  * encryption key the oname.name will be assigned to
 489                  * _name always.
 490                  */
 491                 fname.no_copy = true;
 492                 if (altname_len == 0) {
 493                         /*
 494                          * Set tname to _name, and this will be used
 495                          * to do the base64_decode in-place. It's
 496                          * safe because the decoded string should
 497                          * always be shorter, which is 3/4 of origin
 498                          * string.
 499                          */
 500                         tname.name = _name;
 501
 502                         /*
 503                          * Set oname to _name too, and this will be
 504                          * used to do the dencryption in-place.
 505                          */
 506                         oname.name = _name;
 507                         oname.len = _name_len;
 508                 } else {
 509                         /*
 510                          * This will do the decryption only in-place
 511                          * from altname cryptext directly.
 512                          */
 513                         oname.name = altname;
 514                         oname.len = altname_len;
 515                 }
 516                 rde->is_nokey = false;
 517                 err = ceph_fname_to_usr(&fname, &tname, &oname, &rde->is_nokey);
 518                 if (err) {
 519                         pr_err_client(cl, "unable to decode %.*s, got %d\n",
 520                                       _name_len, _name, err);
 521                         goto out_bad;
 522                 }
 523                 rde->name = oname.name;
 524                 rde->name_len = oname.len;
 525
 526                 /* inode */
 527                 err = parse_reply_info_in(p, end, &rde->inode, features);
 528                 if (err < 0)
 529                         goto out_bad;
 530                 /* ceph_readdir_prepopulate() will update it */
 531                 rde->offset = 0;
 532                 i++;
 533                 num--;
 534         }
 535
 536 done:
 537         /* Skip over any unrecognized fields */
 538         *p = end;
 539         return 0;
 540
 541 bad:
 542         err = -EIO;
 543 out_bad:
 544         pr_err_client(cl, "problem parsing dir contents %d\n", err);
 545         return err;
 546 }
 547
 548 /*
 549  * parse fcntl F_GETLK results
 550  */
 551 static int parse_reply_info_filelock(void **p, void *end,
 552                                      struct ceph_mds_reply_info_parsed *info,
 553                                      u64 features)
 554 {
 555         if (*p + sizeof(*info->filelock_reply) > end)
 556                 goto bad;
 557
 558         info->filelock_reply = *p;
 559
 560         /* Skip over any unrecognized fields */
 561         *p = end;
 562         return 0;
 563 bad:
 564         return -EIO;
 565 }
 566
 567
 568 #if BITS_PER_LONG == 64
 569
 570 #define DELEGATED_INO_AVAILABLE         xa_mk_value(1)
 571
 572 static int ceph_parse_deleg_inos(void **p, void *end,
 573                                  struct ceph_mds_session *s)
 574 {
 575         struct ceph_client *cl = s->s_mdsc->fsc->client;
 576         u32 sets;
 577
 578         ceph_decode_32_safe(p, end, sets, bad);
 579         doutc(cl, "got %u sets of delegated inodes\n", sets);
 580         while (sets--) {
 581                 u64 start, len;
 582
 583                 ceph_decode_64_safe(p, end, start, bad);
 584                 ceph_decode_64_safe(p, end, len, bad);
 585
 586                 /* Don't accept a delegation of system inodes */
 587                 if (start < CEPH_INO_SYSTEM_BASE) {
 588                         pr_warn_ratelimited_client(cl,
 589                                 "ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
 590                                 start, len);
 591                         continue;
 592                 }
 593                 while (len--) {
 594                         int err = xa_insert(&s->s_delegated_inos, start++,
 595                                             DELEGATED_INO_AVAILABLE,
 596                                             GFP_KERNEL);
 597                         if (!err) {
 598                                 doutc(cl, "added delegated inode 0x%llx\n", start - 1);
 599                         } else if (err == -EBUSY) {
 600                                 pr_warn_client(cl,
 601                                         "MDS delegated inode 0x%llx more than once.\n",
 602                                         start - 1);
 603                         } else {
 604                                 return err;
 605                         }
 606                 }
 607         }
 608         return 0;
 609 bad:
 610         return -EIO;
 611 }
 612
 613 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 614 {
 615         unsigned long ino;
 616         void *val;
 617
 618         xa_for_each(&s->s_delegated_inos, ino, val) {
 619                 val = xa_erase(&s->s_delegated_inos, ino);
 620                 if (val == DELEGATED_INO_AVAILABLE)
 621                         return ino;
 622         }
 623         return 0;
 624 }
 625
 626 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 627 {
 628         return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
 629                          GFP_KERNEL);
 630 }
 631 #else /* BITS_PER_LONG == 64 */
 632 /*
 633  * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
 634  * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
 635  * and bottom words?
 636  */
 637 static int ceph_parse_deleg_inos(void **p, void *end,
 638                                  struct ceph_mds_session *s)
 639 {
 640         u32 sets;
 641
 642         ceph_decode_32_safe(p, end, sets, bad);
 643         if (sets)
 644                 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
 645         return 0;
 646 bad:
 647         return -EIO;
 648 }
 649
 650 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
 651 {
 652         return 0;
 653 }
 654
 655 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
 656 {
 657         return 0;
 658 }
 659 #endif /* BITS_PER_LONG == 64 */
 660
 661 /*
 662  * parse create results
 663  */
 664 static int parse_reply_info_create(void **p, void *end,
 665                                   struct ceph_mds_reply_info_parsed *info,
 666                                   u64 features, struct ceph_mds_session *s)
 667 {
 668         int ret;
 669
 670         if (features == (u64)-1 ||
 671             (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
 672                 if (*p == end) {
 673                         /* Malformed reply? */
 674                         info->has_create_ino = false;
 675                 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
 676                         info->has_create_ino = true;
 677                         /* struct_v, struct_compat, and len */
 678                         ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
 679                         ceph_decode_64_safe(p, end, info->ino, bad);
 680                         ret = ceph_parse_deleg_inos(p, end, s);
 681                         if (ret)
 682                                 return ret;
 683                 } else {
 684                         /* legacy */
 685                         ceph_decode_64_safe(p, end, info->ino, bad);
 686                         info->has_create_ino = true;
 687                 }
 688         } else {
 689                 if (*p != end)
 690                         goto bad;
 691         }
 692
 693         /* Skip over any unrecognized fields */
 694         *p = end;
 695         return 0;
 696 bad:
 697         return -EIO;
 698 }
 699
 700 static int parse_reply_info_getvxattr(void **p, void *end,
 701                                       struct ceph_mds_reply_info_parsed *info,
 702                                       u64 features)
 703 {
 704         u32 value_len;
 705
 706         ceph_decode_skip_8(p, end, bad); /* skip current version: 1 */
 707         ceph_decode_skip_8(p, end, bad); /* skip first version: 1 */
 708         ceph_decode_skip_32(p, end, bad); /* skip payload length */
 709
 710         ceph_decode_32_safe(p, end, value_len, bad);
 711
 712         if (value_len == end - *p) {
 713           info->xattr_info.xattr_value = *p;
 714           info->xattr_info.xattr_value_len = value_len;
 715           *p = end;
 716           return value_len;
 717         }
 718 bad:
 719         return -EIO;
 720 }
 721
 722 /*
 723  * parse extra results
 724  */
 725 static int parse_reply_info_extra(void **p, void *end,
 726                                   struct ceph_mds_request *req,
 727                                   u64 features, struct ceph_mds_session *s)
 728 {
 729         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 730         u32 op = le32_to_cpu(info->head->op);
 731
 732         if (op == CEPH_MDS_OP_GETFILELOCK)
 733                 return parse_reply_info_filelock(p, end, info, features);
 734         else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
 735                 return parse_reply_info_readdir(p, end, req, features);
 736         else if (op == CEPH_MDS_OP_CREATE)
 737                 return parse_reply_info_create(p, end, info, features, s);
 738         else if (op == CEPH_MDS_OP_GETVXATTR)
 739                 return parse_reply_info_getvxattr(p, end, info, features);
 740         else
 741                 return -EIO;
 742 }
 743
 744 /*
 745  * parse entire mds reply
 746  */
 747 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
 748                             struct ceph_mds_request *req, u64 features)
 749 {
 750         struct ceph_mds_reply_info_parsed *info = &req->r_reply_info;
 751         struct ceph_client *cl = s->s_mdsc->fsc->client;
 752         void *p, *end;
 753         u32 len;
 754         int err;
 755
 756         info->head = msg->front.iov_base;
 757         p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
 758         end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
 759
 760         /* trace */
 761         ceph_decode_32_safe(&p, end, len, bad);
 762         if (len > 0) {
 763                 ceph_decode_need(&p, end, len, bad);
 764                 err = parse_reply_info_trace(&p, p+len, info, features);
 765                 if (err < 0)
 766                         goto out_bad;
 767         }
 768
 769         /* extra */
 770         ceph_decode_32_safe(&p, end, len, bad);
 771         if (len > 0) {
 772                 ceph_decode_need(&p, end, len, bad);
 773                 err = parse_reply_info_extra(&p, p+len, req, features, s);
 774                 if (err < 0)
 775                         goto out_bad;
 776         }
 777
 778         /* snap blob */
 779         ceph_decode_32_safe(&p, end, len, bad);
 780         info->snapblob_len = len;
 781         info->snapblob = p;
 782         p += len;
 783
 784         if (p != end)
 785                 goto bad;
 786         return 0;
 787
 788 bad:
 789         err = -EIO;
 790 out_bad:
 791         pr_err_client(cl, "mds parse_reply err %d\n", err);
 792         ceph_msg_dump(msg);
 793         return err;
 794 }
 795
 796 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
 797 {
 798         int i;
 799
 800         kfree(info->diri.fscrypt_auth);
 801         kfree(info->diri.fscrypt_file);
 802         kfree(info->targeti.fscrypt_auth);
 803         kfree(info->targeti.fscrypt_file);
 804         if (!info->dir_entries)
 805                 return;
 806
 807         for (i = 0; i < info->dir_nr; i++) {
 808                 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
 809
 810                 kfree(rde->inode.fscrypt_auth);
 811                 kfree(rde->inode.fscrypt_file);
 812         }
 813         free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
 814 }
 815
 816 /*
 817  * In async unlink case the kclient won't wait for the first reply
 818  * from MDS and just drop all the links and unhash the dentry and then
 819  * succeeds immediately.
 820  *
 821  * For any new create/link/rename,etc requests followed by using the
 822  * same file names we must wait for the first reply of the inflight
 823  * unlink request, or the MDS possibly will fail these following
 824  * requests with -EEXIST if the inflight async unlink request was
 825  * delayed for some reasons.
 826  *
 827  * And the worst case is that for the none async openc request it will
 828  * successfully open the file if the CDentry hasn't been unlinked yet,
 829  * but later the previous delayed async unlink request will remove the
 830  * CDentry. That means the just created file is possibly deleted later
 831  * by accident.
 832  *
 833  * We need to wait for the inflight async unlink requests to finish
 834  * when creating new files/directories by using the same file names.
 835  */
 836 int ceph_wait_on_conflict_unlink(struct dentry *dentry)
 837 {
 838         struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dentry->d_sb);
 839         struct ceph_client *cl = fsc->client;
 840         struct dentry *pdentry = dentry->d_parent;
 841         struct dentry *udentry, *found = NULL;
 842         struct ceph_dentry_info *di;
 843         struct qstr dname;
 844         u32 hash = dentry->d_name.hash;
 845         int err;
 846
 847         dname.name = dentry->d_name.name;
 848         dname.len = dentry->d_name.len;
 849
 850         rcu_read_lock();
 851         hash_for_each_possible_rcu(fsc->async_unlink_conflict, di,
 852                                    hnode, hash) {
 853                 udentry = di->dentry;
 854
 855                 spin_lock(&udentry->d_lock);
 856                 if (udentry->d_name.hash != hash)
 857                         goto next;
 858                 if (unlikely(udentry->d_parent != pdentry))
 859                         goto next;
 860                 if (!hash_hashed(&di->hnode))
 861                         goto next;
 862
 863                 if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
 864                         pr_warn_client(cl, "dentry %p:%pd async unlink bit is not set\n",
 865                                        dentry, dentry);
 866
 867                 if (!d_same_name(udentry, pdentry, &dname))
 868                         goto next;
 869
 870                 found = dget_dlock(udentry);
 871                 spin_unlock(&udentry->d_lock);
 872                 break;
 873 next:
 874                 spin_unlock(&udentry->d_lock);
 875         }
 876         rcu_read_unlock();
 877
 878         if (likely(!found))
 879                 return 0;
 880
 881         doutc(cl, "dentry %p:%pd conflict with old %p:%pd\n", dentry, dentry,
 882               found, found);
 883
 884         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT,
 885                           TASK_KILLABLE);
 886         dput(found);
 887         return err;
 888 }
 889
 890
 891 /*
 892  * sessions
 893  */
 894 const char *ceph_session_state_name(int s)
 895 {
 896         switch (s) {
 897         case CEPH_MDS_SESSION_NEW: return "new";
 898         case CEPH_MDS_SESSION_OPENING: return "opening";
 899         case CEPH_MDS_SESSION_OPEN: return "open";
 900         case CEPH_MDS_SESSION_HUNG: return "hung";
 901         case CEPH_MDS_SESSION_CLOSING: return "closing";
 902         case CEPH_MDS_SESSION_CLOSED: return "closed";
 903         case CEPH_MDS_SESSION_RESTARTING: return "restarting";
 904         case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
 905         case CEPH_MDS_SESSION_REJECTED: return "rejected";
 906         default: return "???";
 907         }
 908 }
 909
 910 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
 911 {
 912         if (refcount_inc_not_zero(&s->s_ref))
 913                 return s;
 914         return NULL;
 915 }
 916
 917 void ceph_put_mds_session(struct ceph_mds_session *s)
 918 {
 919         if (IS_ERR_OR_NULL(s))
 920                 return;
 921
 922         if (refcount_dec_and_test(&s->s_ref)) {
 923                 if (s->s_auth.authorizer)
 924                         ceph_auth_destroy_authorizer(s->s_auth.authorizer);
 925                 WARN_ON(mutex_is_locked(&s->s_mutex));
 926                 xa_destroy(&s->s_delegated_inos);
 927                 kfree(s);
 928         }
 929 }
 930
 931 /*
 932  * called under mdsc->mutex
 933  */
 934 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
 935                                                    int mds)
 936 {
 937         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 938                 return NULL;
 939         return ceph_get_mds_session(mdsc->sessions[mds]);
 940 }
 941
 942 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
 943 {
 944         if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
 945                 return false;
 946         else
 947                 return true;
 948 }
 949
 950 static int __verify_registered_session(struct ceph_mds_client *mdsc,
 951                                        struct ceph_mds_session *s)
 952 {
 953         if (s->s_mds >= mdsc->max_sessions ||
 954             mdsc->sessions[s->s_mds] != s)
 955                 return -ENOENT;
 956         return 0;
 957 }
 958
 959 /*
 960  * create+register a new session for given mds.
 961  * called under mdsc->mutex.
 962  */
 963 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 964                                                  int mds)
 965 {
 966         struct ceph_client *cl = mdsc->fsc->client;
 967         struct ceph_mds_session *s;
 968
 969         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
 970                 return ERR_PTR(-EIO);
 971
 972         if (mds >= mdsc->mdsmap->possible_max_rank)
 973                 return ERR_PTR(-EINVAL);
 974
 975         s = kzalloc(sizeof(*s), GFP_NOFS);
 976         if (!s)
 977                 return ERR_PTR(-ENOMEM);
 978
 979         if (mds >= mdsc->max_sessions) {
 980                 int newmax = 1 << get_count_order(mds + 1);
 981                 struct ceph_mds_session **sa;
 982
 983                 doutc(cl, "realloc to %d\n", newmax);
 984                 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
 985                 if (!sa)
 986                         goto fail_realloc;
 987                 if (mdsc->sessions) {
 988                         memcpy(sa, mdsc->sessions,
 989                                mdsc->max_sessions * sizeof(void *));
 990                         kfree(mdsc->sessions);
 991                 }
 992                 mdsc->sessions = sa;
 993                 mdsc->max_sessions = newmax;
 994         }
 995
 996         doutc(cl, "mds%d\n", mds);
 997         s->s_mdsc = mdsc;
 998         s->s_mds = mds;
 999         s->s_state = CEPH_MDS_SESSION_NEW;
1000         mutex_init(&s->s_mutex);
1001
1002         ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
1003
1004         atomic_set(&s->s_cap_gen, 1);
1005         s->s_cap_ttl = jiffies - 1;
1006
1007         spin_lock_init(&s->s_cap_lock);
1008         INIT_LIST_HEAD(&s->s_caps);
1009         refcount_set(&s->s_ref, 1);
1010         INIT_LIST_HEAD(&s->s_waiting);
1011         INIT_LIST_HEAD(&s->s_unsafe);
1012         xa_init(&s->s_delegated_inos);
1013         INIT_LIST_HEAD(&s->s_cap_releases);
1014         INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
1015
1016         INIT_LIST_HEAD(&s->s_cap_dirty);
1017         INIT_LIST_HEAD(&s->s_cap_flushing);
1018
1019         mdsc->sessions[mds] = s;
1020         atomic_inc(&mdsc->num_sessions);
1021         refcount_inc(&s->s_ref);  /* one ref to sessions[], one to caller */
1022
1023         ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
1024                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
1025
1026         return s;
1027
1028 fail_realloc:
1029         kfree(s);
1030         return ERR_PTR(-ENOMEM);
1031 }
1032
1033 /*
1034  * called under mdsc->mutex
1035  */
1036 static void __unregister_session(struct ceph_mds_client *mdsc,
1037                                struct ceph_mds_session *s)
1038 {
1039         doutc(mdsc->fsc->client, "mds%d %p\n", s->s_mds, s);
1040         BUG_ON(mdsc->sessions[s->s_mds] != s);
1041         mdsc->sessions[s->s_mds] = NULL;
1042         ceph_con_close(&s->s_con);
1043         ceph_put_mds_session(s);
1044         atomic_dec(&mdsc->num_sessions);
1045 }
1046
1047 /*
1048  * drop session refs in request.
1049  *
1050  * should be last request ref, or hold mdsc->mutex
1051  */
1052 static void put_request_session(struct ceph_mds_request *req)
1053 {
1054         if (req->r_session) {
1055                 ceph_put_mds_session(req->r_session);
1056                 req->r_session = NULL;
1057         }
1058 }
1059
1060 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
1061                                 void (*cb)(struct ceph_mds_session *),
1062                                 bool check_state)
1063 {
1064         int mds;
1065
1066         mutex_lock(&mdsc->mutex);
1067         for (mds = 0; mds < mdsc->max_sessions; ++mds) {
1068                 struct ceph_mds_session *s;
1069
1070                 s = __ceph_lookup_mds_session(mdsc, mds);
1071                 if (!s)
1072                         continue;
1073
1074                 if (check_state && !check_session_state(s)) {
1075                         ceph_put_mds_session(s);
1076                         continue;
1077                 }
1078
1079                 mutex_unlock(&mdsc->mutex);
1080                 cb(s);
1081                 ceph_put_mds_session(s);
1082                 mutex_lock(&mdsc->mutex);
1083         }
1084         mutex_unlock(&mdsc->mutex);
1085 }
1086
1087 void ceph_mdsc_release_request(struct kref *kref)
1088 {
1089         struct ceph_mds_request *req = container_of(kref,
1090                                                     struct ceph_mds_request,
1091                                                     r_kref);
1092         ceph_mdsc_release_dir_caps_async(req);
1093         destroy_reply_info(&req->r_reply_info);
1094         if (req->r_request)
1095                 ceph_msg_put(req->r_request);
1096         if (req->r_reply)
1097                 ceph_msg_put(req->r_reply);
1098         if (req->r_inode) {
1099                 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1100                 iput(req->r_inode);
1101         }
1102         if (req->r_parent) {
1103                 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
1104                 iput(req->r_parent);
1105         }
1106         iput(req->r_target_inode);
1107         iput(req->r_new_inode);
1108         if (req->r_dentry)
1109                 dput(req->r_dentry);
1110         if (req->r_old_dentry)
1111                 dput(req->r_old_dentry);
1112         if (req->r_old_dentry_dir) {
1113                 /*
1114                  * track (and drop pins for) r_old_dentry_dir
1115                  * separately, since r_old_dentry's d_parent may have
1116                  * changed between the dir mutex being dropped and
1117                  * this request being freed.
1118                  */
1119                 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
1120                                   CEPH_CAP_PIN);
1121                 iput(req->r_old_dentry_dir);
1122         }
1123         kfree(req->r_path1);
1124         kfree(req->r_path2);
1125         put_cred(req->r_cred);
1126         if (req->r_mnt_idmap)
1127                 mnt_idmap_put(req->r_mnt_idmap);
1128         if (req->r_pagelist)
1129                 ceph_pagelist_release(req->r_pagelist);
1130         kfree(req->r_fscrypt_auth);
1131         kfree(req->r_altname);
1132         put_request_session(req);
1133         ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
1134         WARN_ON_ONCE(!list_empty(&req->r_wait));
1135         kmem_cache_free(ceph_mds_request_cachep, req);
1136 }
1137
1138 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
1139
1140 /*
1141  * lookup session, bump ref if found.
1142  *
1143  * called under mdsc->mutex.
1144  */
1145 static struct ceph_mds_request *
1146 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
1147 {
1148         struct ceph_mds_request *req;
1149
1150         req = lookup_request(&mdsc->request_tree, tid);
1151         if (req)
1152                 ceph_mdsc_get_request(req);
1153
1154         return req;
1155 }
1156
1157 /*
1158  * Register an in-flight request, and assign a tid.  Link to directory
1159  * are modifying (if any).
1160  *
1161  * Called under mdsc->mutex.
1162  */
1163 static void __register_request(struct ceph_mds_client *mdsc,
1164                                struct ceph_mds_request *req,
1165                                struct inode *dir)
1166 {
1167         struct ceph_client *cl = mdsc->fsc->client;
1168         int ret = 0;
1169
1170         req->r_tid = ++mdsc->last_tid;
1171         if (req->r_num_caps) {
1172                 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1173                                         req->r_num_caps);
1174                 if (ret < 0) {
1175                         pr_err_client(cl, "%p failed to reserve caps: %d\n",
1176                                       req, ret);
1177                         /* set req->r_err to fail early from __do_request */
1178                         req->r_err = ret;
1179                         return;
1180                 }
1181         }
1182         doutc(cl, "%p tid %lld\n", req, req->r_tid);
1183         ceph_mdsc_get_request(req);
1184         insert_request(&mdsc->request_tree, req);
1185
1186         req->r_cred = get_current_cred();
1187         if (!req->r_mnt_idmap)
1188                 req->r_mnt_idmap = &nop_mnt_idmap;
1189
1190         if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1191                 mdsc->oldest_tid = req->r_tid;
1192
1193         if (dir) {
1194                 struct ceph_inode_info *ci = ceph_inode(dir);
1195
1196                 ihold(dir);
1197                 req->r_unsafe_dir = dir;
1198                 spin_lock(&ci->i_unsafe_lock);
1199                 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1200                 spin_unlock(&ci->i_unsafe_lock);
1201         }
1202 }
1203
1204 static void __unregister_request(struct ceph_mds_client *mdsc,
1205                                  struct ceph_mds_request *req)
1206 {
1207         doutc(mdsc->fsc->client, "%p tid %lld\n", req, req->r_tid);
1208
1209         /* Never leave an unregistered request on an unsafe list! */
1210         list_del_init(&req->r_unsafe_item);
1211
1212         if (req->r_tid == mdsc->oldest_tid) {
1213                 struct rb_node *p = rb_next(&req->r_node);
1214                 mdsc->oldest_tid = 0;
1215                 while (p) {
1216                         struct ceph_mds_request *next_req =
1217                                 rb_entry(p, struct ceph_mds_request, r_node);
1218                         if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1219                                 mdsc->oldest_tid = next_req->r_tid;
1220                                 break;
1221                         }
1222                         p = rb_next(p);
1223                 }
1224         }
1225
1226         erase_request(&mdsc->request_tree, req);
1227
1228         if (req->r_unsafe_dir) {
1229                 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1230                 spin_lock(&ci->i_unsafe_lock);
1231                 list_del_init(&req->r_unsafe_dir_item);
1232                 spin_unlock(&ci->i_unsafe_lock);
1233         }
1234         if (req->r_target_inode &&
1235             test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1236                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1237                 spin_lock(&ci->i_unsafe_lock);
1238                 list_del_init(&req->r_unsafe_target_item);
1239                 spin_unlock(&ci->i_unsafe_lock);
1240         }
1241
1242         if (req->r_unsafe_dir) {
1243                 iput(req->r_unsafe_dir);
1244                 req->r_unsafe_dir = NULL;
1245         }
1246
1247         complete_all(&req->r_safe_completion);
1248
1249         ceph_mdsc_put_request(req);
1250 }
1251
1252 /*
1253  * Walk back up the dentry tree until we hit a dentry representing a
1254  * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1255  * when calling this) to ensure that the objects won't disappear while we're
1256  * working with them. Once we hit a candidate dentry, we attempt to take a
1257  * reference to it, and return that as the result.
1258  */
1259 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1260 {
1261         struct inode *inode = NULL;
1262
1263         while (dentry && !IS_ROOT(dentry)) {
1264                 inode = d_inode_rcu(dentry);
1265                 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1266                         break;
1267                 dentry = dentry->d_parent;
1268         }
1269         if (inode)
1270                 inode = igrab(inode);
1271         return inode;
1272 }
1273
1274 /*
1275  * Choose mds to send request to next.  If there is a hint set in the
1276  * request (e.g., due to a prior forward hint from the mds), use that.
1277  * Otherwise, consult frag tree and/or caps to identify the
1278  * appropriate mds.  If all else fails, choose randomly.
1279  *
1280  * Called under mdsc->mutex.
1281  */
1282 static int __choose_mds(struct ceph_mds_client *mdsc,
1283                         struct ceph_mds_request *req,
1284                         bool *random)
1285 {
1286         struct inode *inode;
1287         struct ceph_inode_info *ci;
1288         struct ceph_cap *cap;
1289         int mode = req->r_direct_mode;
1290         int mds = -1;
1291         u32 hash = req->r_direct_hash;
1292         bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1293         struct ceph_client *cl = mdsc->fsc->client;
1294
1295         if (random)
1296                 *random = false;
1297
1298         /*
1299          * is there a specific mds we should try?  ignore hint if we have
1300          * no session and the mds is not up (active or recovering).
1301          */
1302         if (req->r_resend_mds >= 0 &&
1303             (__have_session(mdsc, req->r_resend_mds) ||
1304              ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1305                 doutc(cl, "using resend_mds mds%d\n", req->r_resend_mds);
1306                 return req->r_resend_mds;
1307         }
1308
1309         if (mode == USE_RANDOM_MDS)
1310                 goto random;
1311
1312         inode = NULL;
1313         if (req->r_inode) {
1314                 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1315                         inode = req->r_inode;
1316                         ihold(inode);
1317                 } else {
1318                         /* req->r_dentry is non-null for LSSNAP request */
1319                         rcu_read_lock();
1320                         inode = get_nonsnap_parent(req->r_dentry);
1321                         rcu_read_unlock();
1322                         doutc(cl, "using snapdir's parent %p %llx.%llx\n",
1323                               inode, ceph_vinop(inode));
1324                 }
1325         } else if (req->r_dentry) {
1326                 /* ignore race with rename; old or new d_parent is okay */
1327                 struct dentry *parent;
1328                 struct inode *dir;
1329
1330                 rcu_read_lock();
1331                 parent = READ_ONCE(req->r_dentry->d_parent);
1332                 dir = req->r_parent ? : d_inode_rcu(parent);
1333
1334                 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1335                         /*  not this fs or parent went negative */
1336                         inode = d_inode(req->r_dentry);
1337                         if (inode)
1338                                 ihold(inode);
1339                 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1340                         /* direct snapped/virtual snapdir requests
1341                          * based on parent dir inode */
1342                         inode = get_nonsnap_parent(parent);
1343                         doutc(cl, "using nonsnap parent %p %llx.%llx\n",
1344                               inode, ceph_vinop(inode));
1345                 } else {
1346                         /* dentry target */
1347                         inode = d_inode(req->r_dentry);
1348                         if (!inode || mode == USE_AUTH_MDS) {
1349                                 /* dir + name */
1350                                 inode = igrab(dir);
1351                                 hash = ceph_dentry_hash(dir, req->r_dentry);
1352                                 is_hash = true;
1353                         } else {
1354                                 ihold(inode);
1355                         }
1356                 }
1357                 rcu_read_unlock();
1358         }
1359
1360         if (!inode)
1361                 goto random;
1362
1363         doutc(cl, "%p %llx.%llx is_hash=%d (0x%x) mode %d\n", inode,
1364               ceph_vinop(inode), (int)is_hash, hash, mode);
1365         ci = ceph_inode(inode);
1366
1367         if (is_hash && S_ISDIR(inode->i_mode)) {
1368                 struct ceph_inode_frag frag;
1369                 int found;
1370
1371                 ceph_choose_frag(ci, hash, &frag, &found);
1372                 if (found) {
1373                         if (mode == USE_ANY_MDS && frag.ndist > 0) {
1374                                 u8 r;
1375
1376                                 /* choose a random replica */
1377                                 get_random_bytes(&r, 1);
1378                                 r %= frag.ndist;
1379                                 mds = frag.dist[r];
1380                                 doutc(cl, "%p %llx.%llx frag %u mds%d (%d/%d)\n",
1381                                       inode, ceph_vinop(inode), frag.frag,
1382                                       mds, (int)r, frag.ndist);
1383                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1384                                     CEPH_MDS_STATE_ACTIVE &&
1385                                     !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1386                                         goto out;
1387                         }
1388
1389                         /* since this file/dir wasn't known to be
1390                          * replicated, then we want to look for the
1391                          * authoritative mds. */
1392                         if (frag.mds >= 0) {
1393                                 /* choose auth mds */
1394                                 mds = frag.mds;
1395                                 doutc(cl, "%p %llx.%llx frag %u mds%d (auth)\n",
1396                                       inode, ceph_vinop(inode), frag.frag, mds);
1397                                 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1398                                     CEPH_MDS_STATE_ACTIVE) {
1399                                         if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1400                                                                   mds))
1401                                                 goto out;
1402                                 }
1403                         }
1404                         mode = USE_AUTH_MDS;
1405                 }
1406         }
1407
1408         spin_lock(&ci->i_ceph_lock);
1409         cap = NULL;
1410         if (mode == USE_AUTH_MDS)
1411                 cap = ci->i_auth_cap;
1412         if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1413                 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1414         if (!cap) {
1415                 spin_unlock(&ci->i_ceph_lock);
1416                 iput(inode);
1417                 goto random;
1418         }
1419         mds = cap->session->s_mds;
1420         doutc(cl, "%p %llx.%llx mds%d (%scap %p)\n", inode,
1421               ceph_vinop(inode), mds,
1422               cap == ci->i_auth_cap ? "auth " : "", cap);
1423         spin_unlock(&ci->i_ceph_lock);
1424 out:
1425         iput(inode);
1426         return mds;
1427
1428 random:
1429         if (random)
1430                 *random = true;
1431
1432         mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1433         doutc(cl, "chose random mds%d\n", mds);
1434         return mds;
1435 }
1436
1437
1438 /*
1439  * session messages
1440  */
1441 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1442 {
1443         struct ceph_msg *msg;
1444         struct ceph_mds_session_head *h;
1445
1446         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1447                            false);
1448         if (!msg) {
1449                 pr_err("ENOMEM creating session %s msg\n",
1450                        ceph_session_op_name(op));
1451                 return NULL;
1452         }
1453         h = msg->front.iov_base;
1454         h->op = cpu_to_le32(op);
1455         h->seq = cpu_to_le64(seq);
1456
1457         return msg;
1458 }
1459
1460 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1461 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1462 static int encode_supported_features(void **p, void *end)
1463 {
1464         static const size_t count = ARRAY_SIZE(feature_bits);
1465
1466         if (count > 0) {
1467                 size_t i;
1468                 size_t size = FEATURE_BYTES(count);
1469                 unsigned long bit;
1470
1471                 if (WARN_ON_ONCE(*p + 4 + size > end))
1472                         return -ERANGE;
1473
1474                 ceph_encode_32(p, size);
1475                 memset(*p, 0, size);
1476                 for (i = 0; i < count; i++) {
1477                         bit = feature_bits[i];
1478                         ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1479                 }
1480                 *p += size;
1481         } else {
1482                 if (WARN_ON_ONCE(*p + 4 > end))
1483                         return -ERANGE;
1484
1485                 ceph_encode_32(p, 0);
1486         }
1487
1488         return 0;
1489 }
1490
1491 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1492 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1493 static int encode_metric_spec(void **p, void *end)
1494 {
1495         static const size_t count = ARRAY_SIZE(metric_bits);
1496
1497         /* header */
1498         if (WARN_ON_ONCE(*p + 2 > end))
1499                 return -ERANGE;
1500
1501         ceph_encode_8(p, 1); /* version */
1502         ceph_encode_8(p, 1); /* compat */
1503
1504         if (count > 0) {
1505                 size_t i;
1506                 size_t size = METRIC_BYTES(count);
1507
1508                 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1509                         return -ERANGE;
1510
1511                 /* metric spec info length */
1512                 ceph_encode_32(p, 4 + size);
1513
1514                 /* metric spec */
1515                 ceph_encode_32(p, size);
1516                 memset(*p, 0, size);
1517                 for (i = 0; i < count; i++)
1518                         ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1519                 *p += size;
1520         } else {
1521                 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1522                         return -ERANGE;
1523
1524                 /* metric spec info length */
1525                 ceph_encode_32(p, 4);
1526                 /* metric spec */
1527                 ceph_encode_32(p, 0);
1528         }
1529
1530         return 0;
1531 }
1532
1533 /*
1534  * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1535  * to include additional client metadata fields.
1536  */
1537 static struct ceph_msg *
1538 create_session_full_msg(struct ceph_mds_client *mdsc, int op, u64 seq)
1539 {
1540         struct ceph_msg *msg;
1541         struct ceph_mds_session_head *h;
1542         int i;
1543         int extra_bytes = 0;
1544         int metadata_key_count = 0;
1545         struct ceph_options *opt = mdsc->fsc->client->options;
1546         struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1547         struct ceph_client *cl = mdsc->fsc->client;
1548         size_t size, count;
1549         void *p, *end;
1550         int ret;
1551
1552         const char* metadata[][2] = {
1553                 {"hostname", mdsc->nodename},
1554                 {"kernel_version", init_utsname()->release},
1555                 {"entity_id", opt->name ? : ""},
1556                 {"root", fsopt->server_path ? : "/"},
1557                 {NULL, NULL}
1558         };
1559
1560         /* Calculate serialized length of metadata */
1561         extra_bytes = 4;  /* map length */
1562         for (i = 0; metadata[i][0]; ++i) {
1563                 extra_bytes += 8 + strlen(metadata[i][0]) +
1564                         strlen(metadata[i][1]);
1565                 metadata_key_count++;
1566         }
1567
1568         /* supported feature */
1569         size = 0;
1570         count = ARRAY_SIZE(feature_bits);
1571         if (count > 0)
1572                 size = FEATURE_BYTES(count);
1573         extra_bytes += 4 + size;
1574
1575         /* metric spec */
1576         size = 0;
1577         count = ARRAY_SIZE(metric_bits);
1578         if (count > 0)
1579                 size = METRIC_BYTES(count);
1580         extra_bytes += 2 + 4 + 4 + size;
1581
1582         /* flags, mds auth caps and oldest_client_tid */
1583         extra_bytes += 4 + 4 + 8;
1584
1585         /* Allocate the message */
1586         msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1587                            GFP_NOFS, false);
1588         if (!msg) {
1589                 pr_err_client(cl, "ENOMEM creating session open msg\n");
1590                 return ERR_PTR(-ENOMEM);
1591         }
1592         p = msg->front.iov_base;
1593         end = p + msg->front.iov_len;
1594
1595         h = p;
1596         h->op = cpu_to_le32(op);
1597         h->seq = cpu_to_le64(seq);
1598
1599         /*
1600          * Serialize client metadata into waiting buffer space, using
1601          * the format that userspace expects for map<string, string>
1602          *
1603          * ClientSession messages with metadata are v7
1604          */
1605         msg->hdr.version = cpu_to_le16(7);
1606         msg->hdr.compat_version = cpu_to_le16(1);
1607
1608         /* The write pointer, following the session_head structure */
1609         p += sizeof(*h);
1610
1611         /* Number of entries in the map */
1612         ceph_encode_32(&p, metadata_key_count);
1613
1614         /* Two length-prefixed strings for each entry in the map */
1615         for (i = 0; metadata[i][0]; ++i) {
1616                 size_t const key_len = strlen(metadata[i][0]);
1617                 size_t const val_len = strlen(metadata[i][1]);
1618
1619                 ceph_encode_32(&p, key_len);
1620                 memcpy(p, metadata[i][0], key_len);
1621                 p += key_len;
1622                 ceph_encode_32(&p, val_len);
1623                 memcpy(p, metadata[i][1], val_len);
1624                 p += val_len;
1625         }
1626
1627         ret = encode_supported_features(&p, end);
1628         if (ret) {
1629                 pr_err_client(cl, "encode_supported_features failed!\n");
1630                 ceph_msg_put(msg);
1631                 return ERR_PTR(ret);
1632         }
1633
1634         ret = encode_metric_spec(&p, end);
1635         if (ret) {
1636                 pr_err_client(cl, "encode_metric_spec failed!\n");
1637                 ceph_msg_put(msg);
1638                 return ERR_PTR(ret);
1639         }
1640
1641         /* version == 5, flags */
1642         ceph_encode_32(&p, 0);
1643
1644         /* version == 6, mds auth caps */
1645         ceph_encode_32(&p, 0);
1646
1647         /* version == 7, oldest_client_tid */
1648         ceph_encode_64(&p, mdsc->oldest_tid);
1649
1650         msg->front.iov_len = p - msg->front.iov_base;
1651         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1652
1653         return msg;
1654 }
1655
1656 /*
1657  * send session open request.
1658  *
1659  * called under mdsc->mutex
1660  */
1661 static int __open_session(struct ceph_mds_client *mdsc,
1662                           struct ceph_mds_session *session)
1663 {
1664         struct ceph_msg *msg;
1665         int mstate;
1666         int mds = session->s_mds;
1667
1668         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1669                 return -EIO;
1670
1671         /* wait for mds to go active? */
1672         mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1673         doutc(mdsc->fsc->client, "open_session to mds%d (%s)\n", mds,
1674               ceph_mds_state_name(mstate));
1675         session->s_state = CEPH_MDS_SESSION_OPENING;
1676         session->s_renew_requested = jiffies;
1677
1678         /* send connect message */
1679         msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_OPEN,
1680                                       session->s_seq);
1681         if (IS_ERR(msg))
1682                 return PTR_ERR(msg);
1683         ceph_con_send(&session->s_con, msg);
1684         return 0;
1685 }
1686
1687 /*
1688  * open sessions for any export targets for the given mds
1689  *
1690  * called under mdsc->mutex
1691  */
1692 static struct ceph_mds_session *
1693 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1694 {
1695         struct ceph_mds_session *session;
1696         int ret;
1697
1698         session = __ceph_lookup_mds_session(mdsc, target);
1699         if (!session) {
1700                 session = register_session(mdsc, target);
1701                 if (IS_ERR(session))
1702                         return session;
1703         }
1704         if (session->s_state == CEPH_MDS_SESSION_NEW ||
1705             session->s_state == CEPH_MDS_SESSION_CLOSING) {
1706                 ret = __open_session(mdsc, session);
1707                 if (ret)
1708                         return ERR_PTR(ret);
1709         }
1710
1711         return session;
1712 }
1713
1714 struct ceph_mds_session *
1715 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1716 {
1717         struct ceph_mds_session *session;
1718         struct ceph_client *cl = mdsc->fsc->client;
1719
1720         doutc(cl, "to mds%d\n", target);
1721
1722         mutex_lock(&mdsc->mutex);
1723         session = __open_export_target_session(mdsc, target);
1724         mutex_unlock(&mdsc->mutex);
1725
1726         return session;
1727 }
1728
1729 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1730                                           struct ceph_mds_session *session)
1731 {
1732         struct ceph_mds_info *mi;
1733         struct ceph_mds_session *ts;
1734         int i, mds = session->s_mds;
1735         struct ceph_client *cl = mdsc->fsc->client;
1736
1737         if (mds >= mdsc->mdsmap->possible_max_rank)
1738                 return;
1739
1740         mi = &mdsc->mdsmap->m_info[mds];
1741         doutc(cl, "for mds%d (%d targets)\n", session->s_mds,
1742               mi->num_export_targets);
1743
1744         for (i = 0; i < mi->num_export_targets; i++) {
1745                 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1746                 ceph_put_mds_session(ts);
1747         }
1748 }
1749
1750 /*
1751  * session caps
1752  */
1753
1754 static void detach_cap_releases(struct ceph_mds_session *session,
1755                                 struct list_head *target)
1756 {
1757         struct ceph_client *cl = session->s_mdsc->fsc->client;
1758
1759         lockdep_assert_held(&session->s_cap_lock);
1760
1761         list_splice_init(&session->s_cap_releases, target);
1762         session->s_num_cap_releases = 0;
1763         doutc(cl, "mds%d\n", session->s_mds);
1764 }
1765
1766 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1767                                  struct list_head *dispose)
1768 {
1769         while (!list_empty(dispose)) {
1770                 struct ceph_cap *cap;
1771                 /* zero out the in-progress message */
1772                 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1773                 list_del(&cap->session_caps);
1774                 ceph_put_cap(mdsc, cap);
1775         }
1776 }
1777
1778 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1779                                      struct ceph_mds_session *session)
1780 {
1781         struct ceph_client *cl = mdsc->fsc->client;
1782         struct ceph_mds_request *req;
1783         struct rb_node *p;
1784
1785         doutc(cl, "mds%d\n", session->s_mds);
1786         mutex_lock(&mdsc->mutex);
1787         while (!list_empty(&session->s_unsafe)) {
1788                 req = list_first_entry(&session->s_unsafe,
1789                                        struct ceph_mds_request, r_unsafe_item);
1790                 pr_warn_ratelimited_client(cl, " dropping unsafe request %llu\n",
1791                                            req->r_tid);
1792                 if (req->r_target_inode)
1793                         mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1794                 if (req->r_unsafe_dir)
1795                         mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1796                 __unregister_request(mdsc, req);
1797         }
1798         /* zero r_attempts, so kick_requests() will re-send requests */
1799         p = rb_first(&mdsc->request_tree);
1800         while (p) {
1801                 req = rb_entry(p, struct ceph_mds_request, r_node);
1802                 p = rb_next(p);
1803                 if (req->r_session &&
1804                     req->r_session->s_mds == session->s_mds)
1805                         req->r_attempts = 0;
1806         }
1807         mutex_unlock(&mdsc->mutex);
1808 }
1809
1810 /*
1811  * Helper to safely iterate over all caps associated with a session, with
1812  * special care taken to handle a racing __ceph_remove_cap().
1813  *
1814  * Caller must hold session s_mutex.
1815  */
1816 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1817                               int (*cb)(struct inode *, int mds, void *),
1818                               void *arg)
1819 {
1820         struct ceph_client *cl = session->s_mdsc->fsc->client;
1821         struct list_head *p;
1822         struct ceph_cap *cap;
1823         struct inode *inode, *last_inode = NULL;
1824         struct ceph_cap *old_cap = NULL;
1825         int ret;
1826
1827         doutc(cl, "%p mds%d\n", session, session->s_mds);
1828         spin_lock(&session->s_cap_lock);
1829         p = session->s_caps.next;
1830         while (p != &session->s_caps) {
1831                 int mds;
1832
1833                 cap = list_entry(p, struct ceph_cap, session_caps);
1834                 inode = igrab(&cap->ci->netfs.inode);
1835                 if (!inode) {
1836                         p = p->next;
1837                         continue;
1838                 }
1839                 session->s_cap_iterator = cap;
1840                 mds = cap->mds;
1841                 spin_unlock(&session->s_cap_lock);
1842
1843                 if (last_inode) {
1844                         iput(last_inode);
1845                         last_inode = NULL;
1846                 }
1847                 if (old_cap) {
1848                         ceph_put_cap(session->s_mdsc, old_cap);
1849                         old_cap = NULL;
1850                 }
1851
1852                 ret = cb(inode, mds, arg);
1853                 last_inode = inode;
1854
1855                 spin_lock(&session->s_cap_lock);
1856                 p = p->next;
1857                 if (!cap->ci) {
1858                         doutc(cl, "finishing cap %p removal\n", cap);
1859                         BUG_ON(cap->session != session);
1860                         cap->session = NULL;
1861                         list_del_init(&cap->session_caps);
1862                         session->s_nr_caps--;
1863                         atomic64_dec(&session->s_mdsc->metric.total_caps);
1864                         if (cap->queue_release)
1865                                 __ceph_queue_cap_release(session, cap);
1866                         else
1867                                 old_cap = cap;  /* put_cap it w/o locks held */
1868                 }
1869                 if (ret < 0)
1870                         goto out;
1871         }
1872         ret = 0;
1873 out:
1874         session->s_cap_iterator = NULL;
1875         spin_unlock(&session->s_cap_lock);
1876
1877         iput(last_inode);
1878         if (old_cap)
1879                 ceph_put_cap(session->s_mdsc, old_cap);
1880
1881         return ret;
1882 }
1883
1884 static int remove_session_caps_cb(struct inode *inode, int mds, void *arg)
1885 {
1886         struct ceph_inode_info *ci = ceph_inode(inode);
1887         struct ceph_client *cl = ceph_inode_to_client(inode);
1888         bool invalidate = false;
1889         struct ceph_cap *cap;
1890         int iputs = 0;
1891
1892         spin_lock(&ci->i_ceph_lock);
1893         cap = __get_cap_for_mds(ci, mds);
1894         if (cap) {
1895                 doutc(cl, " removing cap %p, ci is %p, inode is %p\n",
1896                       cap, ci, &ci->netfs.inode);
1897
1898                 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1899         }
1900         spin_unlock(&ci->i_ceph_lock);
1901
1902         if (cap)
1903                 wake_up_all(&ci->i_cap_wq);
1904         if (invalidate)
1905                 ceph_queue_invalidate(inode);
1906         while (iputs--)
1907                 iput(inode);
1908         return 0;
1909 }
1910
1911 /*
1912  * caller must hold session s_mutex
1913  */
1914 static void remove_session_caps(struct ceph_mds_session *session)
1915 {
1916         struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1917         struct super_block *sb = fsc->sb;
1918         LIST_HEAD(dispose);
1919
1920         doutc(fsc->client, "on %p\n", session);
1921         ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1922
1923         wake_up_all(&fsc->mdsc->cap_flushing_wq);
1924
1925         spin_lock(&session->s_cap_lock);
1926         if (session->s_nr_caps > 0) {
1927                 struct inode *inode;
1928                 struct ceph_cap *cap, *prev = NULL;
1929                 struct ceph_vino vino;
1930                 /*
1931                  * iterate_session_caps() skips inodes that are being
1932                  * deleted, we need to wait until deletions are complete.
1933                  * __wait_on_freeing_inode() is designed for the job,
1934                  * but it is not exported, so use lookup inode function
1935                  * to access it.
1936                  */
1937                 while (!list_empty(&session->s_caps)) {
1938                         cap = list_entry(session->s_caps.next,
1939                                          struct ceph_cap, session_caps);
1940                         if (cap == prev)
1941                                 break;
1942                         prev = cap;
1943                         vino = cap->ci->i_vino;
1944                         spin_unlock(&session->s_cap_lock);
1945
1946                         inode = ceph_find_inode(sb, vino);
1947                         iput(inode);
1948
1949                         spin_lock(&session->s_cap_lock);
1950                 }
1951         }
1952
1953         // drop cap expires and unlock s_cap_lock
1954         detach_cap_releases(session, &dispose);
1955
1956         BUG_ON(session->s_nr_caps > 0);
1957         BUG_ON(!list_empty(&session->s_cap_flushing));
1958         spin_unlock(&session->s_cap_lock);
1959         dispose_cap_releases(session->s_mdsc, &dispose);
1960 }
1961
1962 enum {
1963         RECONNECT,
1964         RENEWCAPS,
1965         FORCE_RO,
1966 };
1967
1968 /*
1969  * wake up any threads waiting on this session's caps.  if the cap is
1970  * old (didn't get renewed on the client reconnect), remove it now.
1971  *
1972  * caller must hold s_mutex.
1973  */
1974 static int wake_up_session_cb(struct inode *inode, int mds, void *arg)
1975 {
1976         struct ceph_inode_info *ci = ceph_inode(inode);
1977         unsigned long ev = (unsigned long)arg;
1978
1979         if (ev == RECONNECT) {
1980                 spin_lock(&ci->i_ceph_lock);
1981                 ci->i_wanted_max_size = 0;
1982                 ci->i_requested_max_size = 0;
1983                 spin_unlock(&ci->i_ceph_lock);
1984         } else if (ev == RENEWCAPS) {
1985                 struct ceph_cap *cap;
1986
1987                 spin_lock(&ci->i_ceph_lock);
1988                 cap = __get_cap_for_mds(ci, mds);
1989                 /* mds did not re-issue stale cap */
1990                 if (cap && cap->cap_gen < atomic_read(&cap->session->s_cap_gen))
1991                         cap->issued = cap->implemented = CEPH_CAP_PIN;
1992                 spin_unlock(&ci->i_ceph_lock);
1993         } else if (ev == FORCE_RO) {
1994         }
1995         wake_up_all(&ci->i_cap_wq);
1996         return 0;
1997 }
1998
1999 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
2000 {
2001         struct ceph_client *cl = session->s_mdsc->fsc->client;
2002
2003         doutc(cl, "session %p mds%d\n", session, session->s_mds);
2004         ceph_iterate_session_caps(session, wake_up_session_cb,
2005                                   (void *)(unsigned long)ev);
2006 }
2007
2008 /*
2009  * Send periodic message to MDS renewing all currently held caps.  The
2010  * ack will reset the expiration for all caps from this session.
2011  *
2012  * caller holds s_mutex
2013  */
2014 static int send_renew_caps(struct ceph_mds_client *mdsc,
2015                            struct ceph_mds_session *session)
2016 {
2017         struct ceph_client *cl = mdsc->fsc->client;
2018         struct ceph_msg *msg;
2019         int state;
2020
2021         if (time_after_eq(jiffies, session->s_cap_ttl) &&
2022             time_after_eq(session->s_cap_ttl, session->s_renew_requested))
2023                 pr_info_client(cl, "mds%d caps stale\n", session->s_mds);
2024         session->s_renew_requested = jiffies;
2025
2026         /* do not try to renew caps until a recovering mds has reconnected
2027          * with its clients. */
2028         state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
2029         if (state < CEPH_MDS_STATE_RECONNECT) {
2030                 doutc(cl, "ignoring mds%d (%s)\n", session->s_mds,
2031                       ceph_mds_state_name(state));
2032                 return 0;
2033         }
2034
2035         doutc(cl, "to mds%d (%s)\n", session->s_mds,
2036               ceph_mds_state_name(state));
2037         msg = create_session_full_msg(mdsc, CEPH_SESSION_REQUEST_RENEWCAPS,
2038                                       ++session->s_renew_seq);
2039         if (IS_ERR(msg))
2040                 return PTR_ERR(msg);
2041         ceph_con_send(&session->s_con, msg);
2042         return 0;
2043 }
2044
2045 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
2046                              struct ceph_mds_session *session, u64 seq)
2047 {
2048         struct ceph_client *cl = mdsc->fsc->client;
2049         struct ceph_msg *msg;
2050
2051         doutc(cl, "to mds%d (%s)s seq %lld\n", session->s_mds,
2052               ceph_session_state_name(session->s_state), seq);
2053         msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
2054         if (!msg)
2055                 return -ENOMEM;
2056         ceph_con_send(&session->s_con, msg);
2057         return 0;
2058 }
2059
2060
2061 /*
2062  * Note new cap ttl, and any transition from stale -> not stale (fresh?).
2063  *
2064  * Called under session->s_mutex
2065  */
2066 static void renewed_caps(struct ceph_mds_client *mdsc,
2067                          struct ceph_mds_session *session, int is_renew)
2068 {
2069         struct ceph_client *cl = mdsc->fsc->client;
2070         int was_stale;
2071         int wake = 0;
2072
2073         spin_lock(&session->s_cap_lock);
2074         was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
2075
2076         session->s_cap_ttl = session->s_renew_requested +
2077                 mdsc->mdsmap->m_session_timeout*HZ;
2078
2079         if (was_stale) {
2080                 if (time_before(jiffies, session->s_cap_ttl)) {
2081                         pr_info_client(cl, "mds%d caps renewed\n",
2082                                        session->s_mds);
2083                         wake = 1;
2084                 } else {
2085                         pr_info_client(cl, "mds%d caps still stale\n",
2086                                        session->s_mds);
2087                 }
2088         }
2089         doutc(cl, "mds%d ttl now %lu, was %s, now %s\n", session->s_mds,
2090               session->s_cap_ttl, was_stale ? "stale" : "fresh",
2091               time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
2092         spin_unlock(&session->s_cap_lock);
2093
2094         if (wake)
2095                 wake_up_session_caps(session, RENEWCAPS);
2096 }
2097
2098 /*
2099  * send a session close request
2100  */
2101 static int request_close_session(struct ceph_mds_session *session)
2102 {
2103         struct ceph_client *cl = session->s_mdsc->fsc->client;
2104         struct ceph_msg *msg;
2105
2106         doutc(cl, "mds%d state %s seq %lld\n", session->s_mds,
2107               ceph_session_state_name(session->s_state), session->s_seq);
2108         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
2109                                       session->s_seq);
2110         if (!msg)
2111                 return -ENOMEM;
2112         ceph_con_send(&session->s_con, msg);
2113         return 1;
2114 }
2115
2116 /*
2117  * Called with s_mutex held.
2118  */
2119 static int __close_session(struct ceph_mds_client *mdsc,
2120                          struct ceph_mds_session *session)
2121 {
2122         if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
2123                 return 0;
2124         session->s_state = CEPH_MDS_SESSION_CLOSING;
2125         return request_close_session(session);
2126 }
2127
2128 static bool drop_negative_children(struct dentry *dentry)
2129 {
2130         struct dentry *child;
2131         bool all_negative = true;
2132
2133         if (!d_is_dir(dentry))
2134                 goto out;
2135
2136         spin_lock(&dentry->d_lock);
2137         hlist_for_each_entry(child, &dentry->d_children, d_sib) {
2138                 if (d_really_is_positive(child)) {
2139                         all_negative = false;
2140                         break;
2141                 }
2142         }
2143         spin_unlock(&dentry->d_lock);
2144
2145         if (all_negative)
2146                 shrink_dcache_parent(dentry);
2147 out:
2148         return all_negative;
2149 }
2150
2151 /*
2152  * Trim old(er) caps.
2153  *
2154  * Because we can't cache an inode without one or more caps, we do
2155  * this indirectly: if a cap is unused, we prune its aliases, at which
2156  * point the inode will hopefully get dropped to.
2157  *
2158  * Yes, this is a bit sloppy.  Our only real goal here is to respond to
2159  * memory pressure from the MDS, though, so it needn't be perfect.
2160  */
2161 static int trim_caps_cb(struct inode *inode, int mds, void *arg)
2162 {
2163         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2164         struct ceph_client *cl = mdsc->fsc->client;
2165         int *remaining = arg;
2166         struct ceph_inode_info *ci = ceph_inode(inode);
2167         int used, wanted, oissued, mine;
2168         struct ceph_cap *cap;
2169
2170         if (*remaining <= 0)
2171                 return -1;
2172
2173         spin_lock(&ci->i_ceph_lock);
2174         cap = __get_cap_for_mds(ci, mds);
2175         if (!cap) {
2176                 spin_unlock(&ci->i_ceph_lock);
2177                 return 0;
2178         }
2179         mine = cap->issued | cap->implemented;
2180         used = __ceph_caps_used(ci);
2181         wanted = __ceph_caps_file_wanted(ci);
2182         oissued = __ceph_caps_issued_other(ci, cap);
2183
2184         doutc(cl, "%p %llx.%llx cap %p mine %s oissued %s used %s wanted %s\n",
2185               inode, ceph_vinop(inode), cap, ceph_cap_string(mine),
2186               ceph_cap_string(oissued), ceph_cap_string(used),
2187               ceph_cap_string(wanted));
2188         if (cap == ci->i_auth_cap) {
2189                 if (ci->i_dirty_caps || ci->i_flushing_caps ||
2190                     !list_empty(&ci->i_cap_snaps))
2191                         goto out;
2192                 if ((used | wanted) & CEPH_CAP_ANY_WR)
2193                         goto out;
2194                 /* Note: it's possible that i_filelock_ref becomes non-zero
2195                  * after dropping auth caps. It doesn't hurt because reply
2196                  * of lock mds request will re-add auth caps. */
2197                 if (atomic_read(&ci->i_filelock_ref) > 0)
2198                         goto out;
2199         }
2200         /* The inode has cached pages, but it's no longer used.
2201          * we can safely drop it */
2202         if (S_ISREG(inode->i_mode) &&
2203             wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
2204             !(oissued & CEPH_CAP_FILE_CACHE)) {
2205           used = 0;
2206           oissued = 0;
2207         }
2208         if ((used | wanted) & ~oissued & mine)
2209                 goto out;   /* we need these caps */
2210
2211         if (oissued) {
2212                 /* we aren't the only cap.. just remove us */
2213                 ceph_remove_cap(mdsc, cap, true);
2214                 (*remaining)--;
2215         } else {
2216                 struct dentry *dentry;
2217                 /* try dropping referring dentries */
2218                 spin_unlock(&ci->i_ceph_lock);
2219                 dentry = d_find_any_alias(inode);
2220                 if (dentry && drop_negative_children(dentry)) {
2221                         int count;
2222                         dput(dentry);
2223                         d_prune_aliases(inode);
2224                         count = atomic_read(&inode->i_count);
2225                         if (count == 1)
2226                                 (*remaining)--;
2227                         doutc(cl, "%p %llx.%llx cap %p pruned, count now %d\n",
2228                               inode, ceph_vinop(inode), cap, count);
2229                 } else {
2230                         dput(dentry);
2231                 }
2232                 return 0;
2233         }
2234
2235 out:
2236         spin_unlock(&ci->i_ceph_lock);
2237         return 0;
2238 }
2239
2240 /*
2241  * Trim session cap count down to some max number.
2242  */
2243 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2244                    struct ceph_mds_session *session,
2245                    int max_caps)
2246 {
2247         struct ceph_client *cl = mdsc->fsc->client;
2248         int trim_caps = session->s_nr_caps - max_caps;
2249
2250         doutc(cl, "mds%d start: %d / %d, trim %d\n", session->s_mds,
2251               session->s_nr_caps, max_caps, trim_caps);
2252         if (trim_caps > 0) {
2253                 int remaining = trim_caps;
2254
2255                 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2256                 doutc(cl, "mds%d done: %d / %d, trimmed %d\n",
2257                       session->s_mds, session->s_nr_caps, max_caps,
2258                       trim_caps - remaining);
2259         }
2260
2261         ceph_flush_session_cap_releases(mdsc, session);
2262         return 0;
2263 }
2264
2265 static int check_caps_flush(struct ceph_mds_client *mdsc,
2266                             u64 want_flush_tid)
2267 {
2268         struct ceph_client *cl = mdsc->fsc->client;
2269         int ret = 1;
2270
2271         spin_lock(&mdsc->cap_dirty_lock);
2272         if (!list_empty(&mdsc->cap_flush_list)) {
2273                 struct ceph_cap_flush *cf =
2274                         list_first_entry(&mdsc->cap_flush_list,
2275                                          struct ceph_cap_flush, g_list);
2276                 if (cf->tid <= want_flush_tid) {
2277                         doutc(cl, "still flushing tid %llu <= %llu\n",
2278                               cf->tid, want_flush_tid);
2279                         ret = 0;
2280                 }
2281         }
2282         spin_unlock(&mdsc->cap_dirty_lock);
2283         return ret;
2284 }
2285
2286 /*
2287  * flush all dirty inode data to disk.
2288  *
2289  * returns true if we've flushed through want_flush_tid
2290  */
2291 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2292                             u64 want_flush_tid)
2293 {
2294         struct ceph_client *cl = mdsc->fsc->client;
2295
2296         doutc(cl, "want %llu\n", want_flush_tid);
2297
2298         wait_event(mdsc->cap_flushing_wq,
2299                    check_caps_flush(mdsc, want_flush_tid));
2300
2301         doutc(cl, "ok, flushed thru %llu\n", want_flush_tid);
2302 }
2303
2304 /*
2305  * called under s_mutex
2306  */
2307 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2308                                    struct ceph_mds_session *session)
2309 {
2310         struct ceph_client *cl = mdsc->fsc->client;
2311         struct ceph_msg *msg = NULL;
2312         struct ceph_mds_cap_release *head;
2313         struct ceph_mds_cap_item *item;
2314         struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2315         struct ceph_cap *cap;
2316         LIST_HEAD(tmp_list);
2317         int num_cap_releases;
2318         __le32  barrier, *cap_barrier;
2319
2320         down_read(&osdc->lock);
2321         barrier = cpu_to_le32(osdc->epoch_barrier);
2322         up_read(&osdc->lock);
2323
2324         spin_lock(&session->s_cap_lock);
2325 again:
2326         list_splice_init(&session->s_cap_releases, &tmp_list);
2327         num_cap_releases = session->s_num_cap_releases;
2328         session->s_num_cap_releases = 0;
2329         spin_unlock(&session->s_cap_lock);
2330
2331         while (!list_empty(&tmp_list)) {
2332                 if (!msg) {
2333                         msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2334                                         PAGE_SIZE, GFP_NOFS, false);
2335                         if (!msg)
2336                                 goto out_err;
2337                         head = msg->front.iov_base;
2338                         head->num = cpu_to_le32(0);
2339                         msg->front.iov_len = sizeof(*head);
2340
2341                         msg->hdr.version = cpu_to_le16(2);
2342                         msg->hdr.compat_version = cpu_to_le16(1);
2343                 }
2344
2345                 cap = list_first_entry(&tmp_list, struct ceph_cap,
2346                                         session_caps);
2347                 list_del(&cap->session_caps);
2348                 num_cap_releases--;
2349
2350                 head = msg->front.iov_base;
2351                 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2352                                    &head->num);
2353                 item = msg->front.iov_base + msg->front.iov_len;
2354                 item->ino = cpu_to_le64(cap->cap_ino);
2355                 item->cap_id = cpu_to_le64(cap->cap_id);
2356                 item->migrate_seq = cpu_to_le32(cap->mseq);
2357                 item->issue_seq = cpu_to_le32(cap->issue_seq);
2358                 msg->front.iov_len += sizeof(*item);
2359
2360                 ceph_put_cap(mdsc, cap);
2361
2362                 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2363                         // Append cap_barrier field
2364                         cap_barrier = msg->front.iov_base + msg->front.iov_len;
2365                         *cap_barrier = barrier;
2366                         msg->front.iov_len += sizeof(*cap_barrier);
2367
2368                         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2369                         doutc(cl, "mds%d %p\n", session->s_mds, msg);
2370                         ceph_con_send(&session->s_con, msg);
2371                         msg = NULL;
2372                 }
2373         }
2374
2375         BUG_ON(num_cap_releases != 0);
2376
2377         spin_lock(&session->s_cap_lock);
2378         if (!list_empty(&session->s_cap_releases))
2379                 goto again;
2380         spin_unlock(&session->s_cap_lock);
2381
2382         if (msg) {
2383                 // Append cap_barrier field
2384                 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2385                 *cap_barrier = barrier;
2386                 msg->front.iov_len += sizeof(*cap_barrier);
2387
2388                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2389                 doutc(cl, "mds%d %p\n", session->s_mds, msg);
2390                 ceph_con_send(&session->s_con, msg);
2391         }
2392         return;
2393 out_err:
2394         pr_err_client(cl, "mds%d, failed to allocate message\n",
2395                       session->s_mds);
2396         spin_lock(&session->s_cap_lock);
2397         list_splice(&tmp_list, &session->s_cap_releases);
2398         session->s_num_cap_releases += num_cap_releases;
2399         spin_unlock(&session->s_cap_lock);
2400 }
2401
2402 static void ceph_cap_release_work(struct work_struct *work)
2403 {
2404         struct ceph_mds_session *session =
2405                 container_of(work, struct ceph_mds_session, s_cap_release_work);
2406
2407         mutex_lock(&session->s_mutex);
2408         if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2409             session->s_state == CEPH_MDS_SESSION_HUNG)
2410                 ceph_send_cap_releases(session->s_mdsc, session);
2411         mutex_unlock(&session->s_mutex);
2412         ceph_put_mds_session(session);
2413 }
2414
2415 void ceph_flush_session_cap_releases(struct ceph_mds_client *mdsc,
2416                              struct ceph_mds_session *session)
2417 {
2418         struct ceph_client *cl = mdsc->fsc->client;
2419         if (mdsc->stopping)
2420                 return;
2421
2422         ceph_get_mds_session(session);
2423         if (queue_work(mdsc->fsc->cap_wq,
2424                        &session->s_cap_release_work)) {
2425                 doutc(cl, "cap release work queued\n");
2426         } else {
2427                 ceph_put_mds_session(session);
2428                 doutc(cl, "failed to queue cap release work\n");
2429         }
2430 }
2431
2432 /*
2433  * caller holds session->s_cap_lock
2434  */
2435 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2436                               struct ceph_cap *cap)
2437 {
2438         list_add_tail(&cap->session_caps, &session->s_cap_releases);
2439         session->s_num_cap_releases++;
2440
2441         if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2442                 ceph_flush_session_cap_releases(session->s_mdsc, session);
2443 }
2444
2445 static void ceph_cap_reclaim_work(struct work_struct *work)
2446 {
2447         struct ceph_mds_client *mdsc =
2448                 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2449         int ret = ceph_trim_dentries(mdsc);
2450         if (ret == -EAGAIN)
2451                 ceph_queue_cap_reclaim_work(mdsc);
2452 }
2453
2454 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2455 {
2456         struct ceph_client *cl = mdsc->fsc->client;
2457         if (mdsc->stopping)
2458                 return;
2459
2460         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2461                 doutc(cl, "caps reclaim work queued\n");
2462         } else {
2463                 doutc(cl, "failed to queue caps release work\n");
2464         }
2465 }
2466
2467 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2468 {
2469         int val;
2470         if (!nr)
2471                 return;
2472         val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2473         if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2474                 atomic_set(&mdsc->cap_reclaim_pending, 0);
2475                 ceph_queue_cap_reclaim_work(mdsc);
2476         }
2477 }
2478
2479 void ceph_queue_cap_unlink_work(struct ceph_mds_client *mdsc)
2480 {
2481         struct ceph_client *cl = mdsc->fsc->client;
2482         if (mdsc->stopping)
2483                 return;
2484
2485         if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_unlink_work)) {
2486                 doutc(cl, "caps unlink work queued\n");
2487         } else {
2488                 doutc(cl, "failed to queue caps unlink work\n");
2489         }
2490 }
2491
2492 static void ceph_cap_unlink_work(struct work_struct *work)
2493 {
2494         struct ceph_mds_client *mdsc =
2495                 container_of(work, struct ceph_mds_client, cap_unlink_work);
2496         struct ceph_client *cl = mdsc->fsc->client;
2497
2498         doutc(cl, "begin\n");
2499         spin_lock(&mdsc->cap_delay_lock);
2500         while (!list_empty(&mdsc->cap_unlink_delay_list)) {
2501                 struct ceph_inode_info *ci;
2502                 struct inode *inode;
2503
2504                 ci = list_first_entry(&mdsc->cap_unlink_delay_list,
2505                                       struct ceph_inode_info,
2506                                       i_cap_delay_list);
2507                 list_del_init(&ci->i_cap_delay_list);
2508
2509                 inode = igrab(&ci->netfs.inode);
2510                 if (inode) {
2511                         spin_unlock(&mdsc->cap_delay_lock);
2512                         doutc(cl, "on %p %llx.%llx\n", inode,
2513                               ceph_vinop(inode));
2514                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2515                         iput(inode);
2516                         spin_lock(&mdsc->cap_delay_lock);
2517                 }
2518         }
2519         spin_unlock(&mdsc->cap_delay_lock);
2520         doutc(cl, "done\n");
2521 }
2522
2523 /*
2524  * requests
2525  */
2526
2527 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2528                                     struct inode *dir)
2529 {
2530         struct ceph_inode_info *ci = ceph_inode(dir);
2531         struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2532         struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2533         size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2534         unsigned int num_entries;
2535         int order;
2536
2537         spin_lock(&ci->i_ceph_lock);
2538         num_entries = ci->i_files + ci->i_subdirs;
2539         spin_unlock(&ci->i_ceph_lock);
2540         num_entries = max(num_entries, 1U);
2541         num_entries = min(num_entries, opt->max_readdir);
2542
2543         order = get_order(size * num_entries);
2544         while (order >= 0) {
2545                 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2546                                                              __GFP_NOWARN |
2547                                                              __GFP_ZERO,
2548                                                              order);
2549                 if (rinfo->dir_entries)
2550                         break;
2551                 order--;
2552         }
2553         if (!rinfo->dir_entries)
2554                 return -ENOMEM;
2555
2556         num_entries = (PAGE_SIZE << order) / size;
2557         num_entries = min(num_entries, opt->max_readdir);
2558
2559         rinfo->dir_buf_size = PAGE_SIZE << order;
2560         req->r_num_caps = num_entries + 1;
2561         req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2562         req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2563         return 0;
2564 }
2565
2566 /*
2567  * Create an mds request.
2568  */
2569 struct ceph_mds_request *
2570 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2571 {
2572         struct ceph_mds_request *req;
2573
2574         req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2575         if (!req)
2576                 return ERR_PTR(-ENOMEM);
2577
2578         mutex_init(&req->r_fill_mutex);
2579         req->r_mdsc = mdsc;
2580         req->r_started = jiffies;
2581         req->r_start_latency = ktime_get();
2582         req->r_resend_mds = -1;
2583         INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2584         INIT_LIST_HEAD(&req->r_unsafe_target_item);
2585         req->r_fmode = -1;
2586         req->r_feature_needed = -1;
2587         kref_init(&req->r_kref);
2588         RB_CLEAR_NODE(&req->r_node);
2589         INIT_LIST_HEAD(&req->r_wait);
2590         init_completion(&req->r_completion);
2591         init_completion(&req->r_safe_completion);
2592         INIT_LIST_HEAD(&req->r_unsafe_item);
2593
2594         ktime_get_coarse_real_ts64(&req->r_stamp);
2595
2596         req->r_op = op;
2597         req->r_direct_mode = mode;
2598         return req;
2599 }
2600
2601 /*
2602  * return oldest (lowest) request, tid in request tree, 0 if none.
2603  *
2604  * called under mdsc->mutex.
2605  */
2606 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2607 {
2608         if (RB_EMPTY_ROOT(&mdsc->request_tree))
2609                 return NULL;
2610         return rb_entry(rb_first(&mdsc->request_tree),
2611                         struct ceph_mds_request, r_node);
2612 }
2613
2614 static inline  u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2615 {
2616         return mdsc->oldest_tid;
2617 }
2618
2619 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
2620 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2621 {
2622         struct inode *dir = req->r_parent;
2623         struct dentry *dentry = req->r_dentry;
2624         u8 *cryptbuf = NULL;
2625         u32 len = 0;
2626         int ret = 0;
2627
2628         /* only encode if we have parent and dentry */
2629         if (!dir || !dentry)
2630                 goto success;
2631
2632         /* No-op unless this is encrypted */
2633         if (!IS_ENCRYPTED(dir))
2634                 goto success;
2635
2636         ret = ceph_fscrypt_prepare_readdir(dir);
2637         if (ret < 0)
2638                 return ERR_PTR(ret);
2639
2640         /* No key? Just ignore it. */
2641         if (!fscrypt_has_encryption_key(dir))
2642                 goto success;
2643
2644         if (!fscrypt_fname_encrypted_size(dir, dentry->d_name.len, NAME_MAX,
2645                                           &len)) {
2646                 WARN_ON_ONCE(1);
2647                 return ERR_PTR(-ENAMETOOLONG);
2648         }
2649
2650         /* No need to append altname if name is short enough */
2651         if (len <= CEPH_NOHASH_NAME_MAX) {
2652                 len = 0;
2653                 goto success;
2654         }
2655
2656         cryptbuf = kmalloc(len, GFP_KERNEL);
2657         if (!cryptbuf)
2658                 return ERR_PTR(-ENOMEM);
2659
2660         ret = fscrypt_fname_encrypt(dir, &dentry->d_name, cryptbuf, len);
2661         if (ret) {
2662                 kfree(cryptbuf);
2663                 return ERR_PTR(ret);
2664         }
2665 success:
2666         *plen = len;
2667         return cryptbuf;
2668 }
2669 #else
2670 static u8 *get_fscrypt_altname(const struct ceph_mds_request *req, u32 *plen)
2671 {
2672         *plen = 0;
2673         return NULL;
2674 }
2675 #endif
2676
2677 /**
2678  * ceph_mdsc_build_path - build a path string to a given dentry
2679  * @mdsc: mds client
2680  * @dentry: dentry to which path should be built
2681  * @plen: returned length of string
2682  * @pbase: returned base inode number
2683  * @for_wire: is this path going to be sent to the MDS?
2684  *
2685  * Build a string that represents the path to the dentry. This is mostly called
2686  * for two different purposes:
2687  *
2688  * 1) we need to build a path string to send to the MDS (for_wire == true)
2689  * 2) we need a path string for local presentation (e.g. debugfs)
2690  *    (for_wire == false)
2691  *
2692  * The path is built in reverse, starting with the dentry. Walk back up toward
2693  * the root, building the path until the first non-snapped inode is reached
2694  * (for_wire) or the root inode is reached (!for_wire).
2695  *
2696  * Encode hidden .snap dirs as a double /, i.e.
2697  *   foo/.snap/bar -> foo//bar
2698  */
2699 char *ceph_mdsc_build_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2700                            int *plen, u64 *pbase, int for_wire)
2701 {
2702         struct ceph_client *cl = mdsc->fsc->client;
2703         struct dentry *cur;
2704         struct inode *inode;
2705         char *path;
2706         int pos;
2707         unsigned seq;
2708         u64 base;
2709
2710         if (!dentry)
2711                 return ERR_PTR(-EINVAL);
2712
2713         path = __getname();
2714         if (!path)
2715                 return ERR_PTR(-ENOMEM);
2716 retry:
2717         pos = PATH_MAX - 1;
2718         path[pos] = '\0';
2719
2720         seq = read_seqbegin(&rename_lock);
2721         cur = dget(dentry);
2722         for (;;) {
2723                 struct dentry *parent;
2724
2725                 spin_lock(&cur->d_lock);
2726                 inode = d_inode(cur);
2727                 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2728                         doutc(cl, "path+%d: %p SNAPDIR\n", pos, cur);
2729                         spin_unlock(&cur->d_lock);
2730                         parent = dget_parent(cur);
2731                 } else if (for_wire && inode && dentry != cur &&
2732                            ceph_snap(inode) == CEPH_NOSNAP) {
2733                         spin_unlock(&cur->d_lock);
2734                         pos++; /* get rid of any prepended '/' */
2735                         break;
2736                 } else if (!for_wire || !IS_ENCRYPTED(d_inode(cur->d_parent))) {
2737                         pos -= cur->d_name.len;
2738                         if (pos < 0) {
2739                                 spin_unlock(&cur->d_lock);
2740                                 break;
2741                         }
2742                         memcpy(path + pos, cur->d_name.name, cur->d_name.len);
2743                         spin_unlock(&cur->d_lock);
2744                         parent = dget_parent(cur);
2745                 } else {
2746                         int len, ret;
2747                         char buf[NAME_MAX];
2748
2749                         /*
2750                          * Proactively copy name into buf, in case we need to
2751                          * present it as-is.
2752                          */
2753                         memcpy(buf, cur->d_name.name, cur->d_name.len);
2754                         len = cur->d_name.len;
2755                         spin_unlock(&cur->d_lock);
2756                         parent = dget_parent(cur);
2757
2758                         ret = ceph_fscrypt_prepare_readdir(d_inode(parent));
2759                         if (ret < 0) {
2760                                 dput(parent);
2761                                 dput(cur);
2762                                 return ERR_PTR(ret);
2763                         }
2764
2765                         if (fscrypt_has_encryption_key(d_inode(parent))) {
2766                                 len = ceph_encode_encrypted_fname(d_inode(parent),
2767                                                                   cur, buf);
2768                                 if (len < 0) {
2769                                         dput(parent);
2770                                         dput(cur);
2771                                         return ERR_PTR(len);
2772                                 }
2773                         }
2774                         pos -= len;
2775                         if (pos < 0) {
2776                                 dput(parent);
2777                                 break;
2778                         }
2779                         memcpy(path + pos, buf, len);
2780                 }
2781                 dput(cur);
2782                 cur = parent;
2783
2784                 /* Are we at the root? */
2785                 if (IS_ROOT(cur))
2786                         break;
2787
2788                 /* Are we out of buffer? */
2789                 if (--pos < 0)
2790                         break;
2791
2792                 path[pos] = '/';
2793         }
2794         inode = d_inode(cur);
2795         base = inode ? ceph_ino(inode) : 0;
2796         dput(cur);
2797
2798         if (read_seqretry(&rename_lock, seq))
2799                 goto retry;
2800
2801         if (pos < 0) {
2802                 /*
2803                  * The path is longer than PATH_MAX and this function
2804                  * cannot ever succeed.  Creating paths that long is
2805                  * possible with Ceph, but Linux cannot use them.
2806                  */
2807                 return ERR_PTR(-ENAMETOOLONG);
2808         }
2809
2810         *pbase = base;
2811         *plen = PATH_MAX - 1 - pos;
2812         doutc(cl, "on %p %d built %llx '%.*s'\n", dentry, d_count(dentry),
2813               base, *plen, path + pos);
2814         return path + pos;
2815 }
2816
2817 static int build_dentry_path(struct ceph_mds_client *mdsc, struct dentry *dentry,
2818                              struct inode *dir, const char **ppath, int *ppathlen,
2819                              u64 *pino, bool *pfreepath, bool parent_locked)
2820 {
2821         char *path;
2822
2823         rcu_read_lock();
2824         if (!dir)
2825                 dir = d_inode_rcu(dentry->d_parent);
2826         if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP &&
2827             !IS_ENCRYPTED(dir)) {
2828                 *pino = ceph_ino(dir);
2829                 rcu_read_unlock();
2830                 *ppath = dentry->d_name.name;
2831                 *ppathlen = dentry->d_name.len;
2832                 return 0;
2833         }
2834         rcu_read_unlock();
2835         path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2836         if (IS_ERR(path))
2837                 return PTR_ERR(path);
2838         *ppath = path;
2839         *pfreepath = true;
2840         return 0;
2841 }
2842
2843 static int build_inode_path(struct inode *inode,
2844                             const char **ppath, int *ppathlen, u64 *pino,
2845                             bool *pfreepath)
2846 {
2847         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
2848         struct dentry *dentry;
2849         char *path;
2850
2851         if (ceph_snap(inode) == CEPH_NOSNAP) {
2852                 *pino = ceph_ino(inode);
2853                 *ppathlen = 0;
2854                 return 0;
2855         }
2856         dentry = d_find_alias(inode);
2857         path = ceph_mdsc_build_path(mdsc, dentry, ppathlen, pino, 1);
2858         dput(dentry);
2859         if (IS_ERR(path))
2860                 return PTR_ERR(path);
2861         *ppath = path;
2862         *pfreepath = true;
2863         return 0;
2864 }
2865
2866 /*
2867  * request arguments may be specified via an inode *, a dentry *, or
2868  * an explicit ino+path.
2869  */
2870 static int set_request_path_attr(struct ceph_mds_client *mdsc, struct inode *rinode,
2871                                  struct dentry *rdentry, struct inode *rdiri,
2872                                  const char *rpath, u64 rino, const char **ppath,
2873                                  int *pathlen, u64 *ino, bool *freepath,
2874                                  bool parent_locked)
2875 {
2876         struct ceph_client *cl = mdsc->fsc->client;
2877         int r = 0;
2878
2879         if (rinode) {
2880                 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2881                 doutc(cl, " inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2882                       ceph_snap(rinode));
2883         } else if (rdentry) {
2884                 r = build_dentry_path(mdsc, rdentry, rdiri, ppath, pathlen, ino,
2885                                         freepath, parent_locked);
2886                 doutc(cl, " dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen, *ppath);
2887         } else if (rpath || rino) {
2888                 *ino = rino;
2889                 *ppath = rpath;
2890                 *pathlen = rpath ? strlen(rpath) : 0;
2891                 doutc(cl, " path %.*s\n", *pathlen, rpath);
2892         }
2893
2894         return r;
2895 }
2896
2897 static void encode_mclientrequest_tail(void **p,
2898                                        const struct ceph_mds_request *req)
2899 {
2900         struct ceph_timespec ts;
2901         int i;
2902
2903         ceph_encode_timespec64(&ts, &req->r_stamp);
2904         ceph_encode_copy(p, &ts, sizeof(ts));
2905
2906         /* v4: gid_list */
2907         ceph_encode_32(p, req->r_cred->group_info->ngroups);
2908         for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2909                 ceph_encode_64(p, from_kgid(&init_user_ns,
2910                                             req->r_cred->group_info->gid[i]));
2911
2912         /* v5: altname */
2913         ceph_encode_32(p, req->r_altname_len);
2914         ceph_encode_copy(p, req->r_altname, req->r_altname_len);
2915
2916         /* v6: fscrypt_auth and fscrypt_file */
2917         if (req->r_fscrypt_auth) {
2918                 u32 authlen = ceph_fscrypt_auth_len(req->r_fscrypt_auth);
2919
2920                 ceph_encode_32(p, authlen);
2921                 ceph_encode_copy(p, req->r_fscrypt_auth, authlen);
2922         } else {
2923                 ceph_encode_32(p, 0);
2924         }
2925         if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags)) {
2926                 ceph_encode_32(p, sizeof(__le64));
2927                 ceph_encode_64(p, req->r_fscrypt_file);
2928         } else {
2929                 ceph_encode_32(p, 0);
2930         }
2931 }
2932
2933 static inline u16 mds_supported_head_version(struct ceph_mds_session *session)
2934 {
2935         if (!test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD, &session->s_features))
2936                 return 1;
2937
2938         if (!test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features))
2939                 return 2;
2940
2941         return CEPH_MDS_REQUEST_HEAD_VERSION;
2942 }
2943
2944 static struct ceph_mds_request_head_legacy *
2945 find_legacy_request_head(void *p, u64 features)
2946 {
2947         bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2948         struct ceph_mds_request_head_old *ohead;
2949
2950         if (legacy)
2951                 return (struct ceph_mds_request_head_legacy *)p;
2952         ohead = (struct ceph_mds_request_head_old *)p;
2953         return (struct ceph_mds_request_head_legacy *)&ohead->oldest_client_tid;
2954 }
2955
2956 /*
2957  * called under mdsc->mutex
2958  */
2959 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2960                                                struct ceph_mds_request *req,
2961                                                bool drop_cap_releases)
2962 {
2963         int mds = session->s_mds;
2964         struct ceph_mds_client *mdsc = session->s_mdsc;
2965         struct ceph_client *cl = mdsc->fsc->client;
2966         struct ceph_msg *msg;
2967         struct ceph_mds_request_head_legacy *lhead;
2968         const char *path1 = NULL;
2969         const char *path2 = NULL;
2970         u64 ino1 = 0, ino2 = 0;
2971         int pathlen1 = 0, pathlen2 = 0;
2972         bool freepath1 = false, freepath2 = false;
2973         struct dentry *old_dentry = NULL;
2974         int len;
2975         u16 releases;
2976         void *p, *end;
2977         int ret;
2978         bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2979         u16 request_head_version = mds_supported_head_version(session);
2980         kuid_t caller_fsuid = req->r_cred->fsuid;
2981         kgid_t caller_fsgid = req->r_cred->fsgid;
2982
2983         ret = set_request_path_attr(mdsc, req->r_inode, req->r_dentry,
2984                               req->r_parent, req->r_path1, req->r_ino1.ino,
2985                               &path1, &pathlen1, &ino1, &freepath1,
2986                               test_bit(CEPH_MDS_R_PARENT_LOCKED,
2987                                         &req->r_req_flags));
2988         if (ret < 0) {
2989                 msg = ERR_PTR(ret);
2990                 goto out;
2991         }
2992
2993         /* If r_old_dentry is set, then assume that its parent is locked */
2994         if (req->r_old_dentry &&
2995             !(req->r_old_dentry->d_flags & DCACHE_DISCONNECTED))
2996                 old_dentry = req->r_old_dentry;
2997         ret = set_request_path_attr(mdsc, NULL, old_dentry,
2998                               req->r_old_dentry_dir,
2999                               req->r_path2, req->r_ino2.ino,
3000                               &path2, &pathlen2, &ino2, &freepath2, true);
3001         if (ret < 0) {
3002                 msg = ERR_PTR(ret);
3003                 goto out_free1;
3004         }
3005
3006         req->r_altname = get_fscrypt_altname(req, &req->r_altname_len);
3007         if (IS_ERR(req->r_altname)) {
3008                 msg = ERR_CAST(req->r_altname);
3009                 req->r_altname = NULL;
3010                 goto out_free2;
3011         }
3012
3013         /*
3014          * For old cephs without supporting the 32bit retry/fwd feature
3015          * it will copy the raw memories directly when decoding the
3016          * requests. While new cephs will decode the head depending the
3017          * version member, so we need to make sure it will be compatible
3018          * with them both.
3019          */
3020         if (legacy)
3021                 len = sizeof(struct ceph_mds_request_head_legacy);
3022         else if (request_head_version == 1)
3023                 len = sizeof(struct ceph_mds_request_head_old);
3024         else if (request_head_version == 2)
3025                 len = offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3026         else
3027                 len = sizeof(struct ceph_mds_request_head);
3028
3029         /* filepaths */
3030         len += 2 * (1 + sizeof(u32) + sizeof(u64));
3031         len += pathlen1 + pathlen2;
3032
3033         /* cap releases */
3034         len += sizeof(struct ceph_mds_request_release) *
3035                 (!!req->r_inode_drop + !!req->r_dentry_drop +
3036                  !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
3037
3038         if (req->r_dentry_drop)
3039                 len += pathlen1;
3040         if (req->r_old_dentry_drop)
3041                 len += pathlen2;
3042
3043         /* MClientRequest tail */
3044
3045         /* req->r_stamp */
3046         len += sizeof(struct ceph_timespec);
3047
3048         /* gid list */
3049         len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
3050
3051         /* alternate name */
3052         len += sizeof(u32) + req->r_altname_len;
3053
3054         /* fscrypt_auth */
3055         len += sizeof(u32); // fscrypt_auth
3056         if (req->r_fscrypt_auth)
3057                 len += ceph_fscrypt_auth_len(req->r_fscrypt_auth);
3058
3059         /* fscrypt_file */
3060         len += sizeof(u32);
3061         if (test_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags))
3062                 len += sizeof(__le64);
3063
3064         msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
3065         if (!msg) {
3066                 msg = ERR_PTR(-ENOMEM);
3067                 goto out_free2;
3068         }
3069
3070         msg->hdr.tid = cpu_to_le64(req->r_tid);
3071
3072         lhead = find_legacy_request_head(msg->front.iov_base,
3073                                          session->s_con.peer_features);
3074
3075         if ((req->r_mnt_idmap != &nop_mnt_idmap) &&
3076             !test_bit(CEPHFS_FEATURE_HAS_OWNER_UIDGID, &session->s_features)) {
3077                 WARN_ON_ONCE(!IS_CEPH_MDS_OP_NEWINODE(req->r_op));
3078
3079                 if (enable_unsafe_idmap) {
3080                         pr_warn_once_client(cl,
3081                                 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3082                                 " is not supported by MDS. UID/GID-based restrictions may"
3083                                 " not work properly.\n");
3084
3085                         caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3086                                                    VFSUIDT_INIT(req->r_cred->fsuid));
3087                         caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3088                                                    VFSGIDT_INIT(req->r_cred->fsgid));
3089                 } else {
3090                         pr_err_ratelimited_client(cl,
3091                                 "idmapped mount is used and CEPHFS_FEATURE_HAS_OWNER_UIDGID"
3092                                 " is not supported by MDS. Fail request with -EIO.\n");
3093
3094                         ret = -EIO;
3095                         goto out_err;
3096                 }
3097         }
3098
3099         /*
3100          * The ceph_mds_request_head_legacy didn't contain a version field, and
3101          * one was added when we moved the message version from 3->4.
3102          */
3103         if (legacy) {
3104                 msg->hdr.version = cpu_to_le16(3);
3105                 p = msg->front.iov_base + sizeof(*lhead);
3106         } else if (request_head_version == 1) {
3107                 struct ceph_mds_request_head_old *ohead = msg->front.iov_base;
3108
3109                 msg->hdr.version = cpu_to_le16(4);
3110                 ohead->version = cpu_to_le16(1);
3111                 p = msg->front.iov_base + sizeof(*ohead);
3112         } else if (request_head_version == 2) {
3113                 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3114
3115                 msg->hdr.version = cpu_to_le16(6);
3116                 nhead->version = cpu_to_le16(2);
3117
3118                 p = msg->front.iov_base + offsetofend(struct ceph_mds_request_head, ext_num_fwd);
3119         } else {
3120                 struct ceph_mds_request_head *nhead = msg->front.iov_base;
3121                 kuid_t owner_fsuid;
3122                 kgid_t owner_fsgid;
3123
3124                 msg->hdr.version = cpu_to_le16(6);
3125                 nhead->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
3126                 nhead->struct_len = cpu_to_le32(sizeof(struct ceph_mds_request_head));
3127
3128                 if (IS_CEPH_MDS_OP_NEWINODE(req->r_op)) {
3129                         owner_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
3130                                                 VFSUIDT_INIT(req->r_cred->fsuid));
3131                         owner_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
3132                                                 VFSGIDT_INIT(req->r_cred->fsgid));
3133                         nhead->owner_uid = cpu_to_le32(from_kuid(&init_user_ns, owner_fsuid));
3134                         nhead->owner_gid = cpu_to_le32(from_kgid(&init_user_ns, owner_fsgid));
3135                 } else {
3136                         nhead->owner_uid = cpu_to_le32(-1);
3137                         nhead->owner_gid = cpu_to_le32(-1);
3138                 }
3139
3140                 p = msg->front.iov_base + sizeof(*nhead);
3141         }
3142
3143         end = msg->front.iov_base + msg->front.iov_len;
3144
3145         lhead->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
3146         lhead->op = cpu_to_le32(req->r_op);
3147         lhead->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
3148                                                   caller_fsuid));
3149         lhead->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
3150                                                   caller_fsgid));
3151         lhead->ino = cpu_to_le64(req->r_deleg_ino);
3152         lhead->args = req->r_args;
3153
3154         ceph_encode_filepath(&p, end, ino1, path1);
3155         ceph_encode_filepath(&p, end, ino2, path2);
3156
3157         /* make note of release offset, in case we need to replay */
3158         req->r_request_release_offset = p - msg->front.iov_base;
3159
3160         /* cap releases */
3161         releases = 0;
3162         if (req->r_inode_drop)
3163                 releases += ceph_encode_inode_release(&p,
3164                       req->r_inode ? req->r_inode : d_inode(req->r_dentry),
3165                       mds, req->r_inode_drop, req->r_inode_unless,
3166                       req->r_op == CEPH_MDS_OP_READDIR);
3167         if (req->r_dentry_drop) {
3168                 ret = ceph_encode_dentry_release(&p, req->r_dentry,
3169                                 req->r_parent, mds, req->r_dentry_drop,
3170                                 req->r_dentry_unless);
3171                 if (ret < 0)
3172                         goto out_err;
3173                 releases += ret;
3174         }
3175         if (req->r_old_dentry_drop) {
3176                 ret = ceph_encode_dentry_release(&p, req->r_old_dentry,
3177                                 req->r_old_dentry_dir, mds,
3178                                 req->r_old_dentry_drop,
3179                                 req->r_old_dentry_unless);
3180                 if (ret < 0)
3181                         goto out_err;
3182                 releases += ret;
3183         }
3184         if (req->r_old_inode_drop)
3185                 releases += ceph_encode_inode_release(&p,
3186                       d_inode(req->r_old_dentry),
3187                       mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
3188
3189         if (drop_cap_releases) {
3190                 releases = 0;
3191                 p = msg->front.iov_base + req->r_request_release_offset;
3192         }
3193
3194         lhead->num_releases = cpu_to_le16(releases);
3195
3196         encode_mclientrequest_tail(&p, req);
3197
3198         if (WARN_ON_ONCE(p > end)) {
3199                 ceph_msg_put(msg);
3200                 msg = ERR_PTR(-ERANGE);
3201                 goto out_free2;
3202         }
3203
3204         msg->front.iov_len = p - msg->front.iov_base;
3205         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3206
3207         if (req->r_pagelist) {
3208                 struct ceph_pagelist *pagelist = req->r_pagelist;
3209                 ceph_msg_data_add_pagelist(msg, pagelist);
3210                 msg->hdr.data_len = cpu_to_le32(pagelist->length);
3211         } else {
3212                 msg->hdr.data_len = 0;
3213         }
3214
3215         msg->hdr.data_off = cpu_to_le16(0);
3216
3217 out_free2:
3218         if (freepath2)
3219                 ceph_mdsc_free_path((char *)path2, pathlen2);
3220 out_free1:
3221         if (freepath1)
3222                 ceph_mdsc_free_path((char *)path1, pathlen1);
3223 out:
3224         return msg;
3225 out_err:
3226         ceph_msg_put(msg);
3227         msg = ERR_PTR(ret);
3228         goto out_free2;
3229 }
3230
3231 /*
3232  * called under mdsc->mutex if error, under no mutex if
3233  * success.
3234  */
3235 static void complete_request(struct ceph_mds_client *mdsc,
3236                              struct ceph_mds_request *req)
3237 {
3238         req->r_end_latency = ktime_get();
3239
3240         if (req->r_callback)
3241                 req->r_callback(mdsc, req);
3242         complete_all(&req->r_completion);
3243 }
3244
3245 /*
3246  * called under mdsc->mutex
3247  */
3248 static int __prepare_send_request(struct ceph_mds_session *session,
3249                                   struct ceph_mds_request *req,
3250                                   bool drop_cap_releases)
3251 {
3252         int mds = session->s_mds;
3253         struct ceph_mds_client *mdsc = session->s_mdsc;
3254         struct ceph_client *cl = mdsc->fsc->client;
3255         struct ceph_mds_request_head_legacy *lhead;
3256         struct ceph_mds_request_head *nhead;
3257         struct ceph_msg *msg;
3258         int flags = 0, old_max_retry;
3259         bool old_version = !test_bit(CEPHFS_FEATURE_32BITS_RETRY_FWD,
3260                                      &session->s_features);
3261
3262         /*
3263          * Avoid infinite retrying after overflow. The client will
3264          * increase the retry count and if the MDS is old version,
3265          * so we limit to retry at most 256 times.
3266          */
3267         if (req->r_attempts) {
3268                old_max_retry = sizeof_field(struct ceph_mds_request_head_old,
3269                                             num_retry);
3270                old_max_retry = 1 << (old_max_retry * BITS_PER_BYTE);
3271                if ((old_version && req->r_attempts >= old_max_retry) ||
3272                    ((uint32_t)req->r_attempts >= U32_MAX)) {
3273                         pr_warn_ratelimited_client(cl, "request tid %llu seq overflow\n",
3274                                                    req->r_tid);
3275                         return -EMULTIHOP;
3276                }
3277         }
3278
3279         req->r_attempts++;
3280         if (req->r_inode) {
3281                 struct ceph_cap *cap =
3282                         ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
3283
3284                 if (cap)
3285                         req->r_sent_on_mseq = cap->mseq;
3286                 else
3287                         req->r_sent_on_mseq = -1;
3288         }
3289         doutc(cl, "%p tid %lld %s (attempt %d)\n", req, req->r_tid,
3290               ceph_mds_op_name(req->r_op), req->r_attempts);
3291
3292         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3293                 void *p;
3294
3295                 /*
3296                  * Replay.  Do not regenerate message (and rebuild
3297                  * paths, etc.); just use the original message.
3298                  * Rebuilding paths will break for renames because
3299                  * d_move mangles the src name.
3300                  */
3301                 msg = req->r_request;
3302                 lhead = find_legacy_request_head(msg->front.iov_base,
3303                                                  session->s_con.peer_features);
3304
3305                 flags = le32_to_cpu(lhead->flags);
3306                 flags |= CEPH_MDS_FLAG_REPLAY;
3307                 lhead->flags = cpu_to_le32(flags);
3308
3309                 if (req->r_target_inode)
3310                         lhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
3311
3312                 lhead->num_retry = req->r_attempts - 1;
3313                 if (!old_version) {
3314                         nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3315                         nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3316                 }
3317
3318                 /* remove cap/dentry releases from message */
3319                 lhead->num_releases = 0;
3320
3321                 p = msg->front.iov_base + req->r_request_release_offset;
3322                 encode_mclientrequest_tail(&p, req);
3323
3324                 msg->front.iov_len = p - msg->front.iov_base;
3325                 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
3326                 return 0;
3327         }
3328
3329         if (req->r_request) {
3330                 ceph_msg_put(req->r_request);
3331                 req->r_request = NULL;
3332         }
3333         msg = create_request_message(session, req, drop_cap_releases);
3334         if (IS_ERR(msg)) {
3335                 req->r_err = PTR_ERR(msg);
3336                 return PTR_ERR(msg);
3337         }
3338         req->r_request = msg;
3339
3340         lhead = find_legacy_request_head(msg->front.iov_base,
3341                                          session->s_con.peer_features);
3342         lhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
3343         if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3344                 flags |= CEPH_MDS_FLAG_REPLAY;
3345         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
3346                 flags |= CEPH_MDS_FLAG_ASYNC;
3347         if (req->r_parent)
3348                 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
3349         lhead->flags = cpu_to_le32(flags);
3350         lhead->num_fwd = req->r_num_fwd;
3351         lhead->num_retry = req->r_attempts - 1;
3352         if (!old_version) {
3353                 nhead = (struct ceph_mds_request_head*)msg->front.iov_base;
3354                 nhead->ext_num_fwd = cpu_to_le32(req->r_num_fwd);
3355                 nhead->ext_num_retry = cpu_to_le32(req->r_attempts - 1);
3356         }
3357
3358         doutc(cl, " r_parent = %p\n", req->r_parent);
3359         return 0;
3360 }
3361
3362 /*
3363  * called under mdsc->mutex
3364  */
3365 static int __send_request(struct ceph_mds_session *session,
3366                           struct ceph_mds_request *req,
3367                           bool drop_cap_releases)
3368 {
3369         int err;
3370
3371         err = __prepare_send_request(session, req, drop_cap_releases);
3372         if (!err) {
3373                 ceph_msg_get(req->r_request);
3374                 ceph_con_send(&session->s_con, req->r_request);
3375         }
3376
3377         return err;
3378 }
3379
3380 /*
3381  * send request, or put it on the appropriate wait list.
3382  */
3383 static void __do_request(struct ceph_mds_client *mdsc,
3384                         struct ceph_mds_request *req)
3385 {
3386         struct ceph_client *cl = mdsc->fsc->client;
3387         struct ceph_mds_session *session = NULL;
3388         int mds = -1;
3389         int err = 0;
3390         bool random;
3391
3392         if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3393                 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
3394                         __unregister_request(mdsc, req);
3395                 return;
3396         }
3397
3398         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
3399                 doutc(cl, "metadata corrupted\n");
3400                 err = -EIO;
3401                 goto finish;
3402         }
3403         if (req->r_timeout &&
3404             time_after_eq(jiffies, req->r_started + req->r_timeout)) {
3405                 doutc(cl, "timed out\n");
3406                 err = -ETIMEDOUT;
3407                 goto finish;
3408         }
3409         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
3410                 doutc(cl, "forced umount\n");
3411                 err = -EIO;
3412                 goto finish;
3413         }
3414         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
3415                 if (mdsc->mdsmap_err) {
3416                         err = mdsc->mdsmap_err;
3417                         doutc(cl, "mdsmap err %d\n", err);
3418                         goto finish;
3419                 }
3420                 if (mdsc->mdsmap->m_epoch == 0) {
3421                         doutc(cl, "no mdsmap, waiting for map\n");
3422                         list_add(&req->r_wait, &mdsc->waiting_for_map);
3423                         return;
3424                 }
3425                 if (!(mdsc->fsc->mount_options->flags &
3426                       CEPH_MOUNT_OPT_MOUNTWAIT) &&
3427                     !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
3428                         err = -EHOSTUNREACH;
3429                         goto finish;
3430                 }
3431         }
3432
3433         put_request_session(req);
3434
3435         mds = __choose_mds(mdsc, req, &random);
3436         if (mds < 0 ||
3437             ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
3438                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3439                         err = -EJUKEBOX;
3440                         goto finish;
3441                 }
3442                 doutc(cl, "no mds or not active, waiting for map\n");
3443                 list_add(&req->r_wait, &mdsc->waiting_for_map);
3444                 return;
3445         }
3446
3447         /* get, open session */
3448         session = __ceph_lookup_mds_session(mdsc, mds);
3449         if (!session) {
3450                 session = register_session(mdsc, mds);
3451                 if (IS_ERR(session)) {
3452                         err = PTR_ERR(session);
3453                         goto finish;
3454                 }
3455         }
3456         req->r_session = ceph_get_mds_session(session);
3457
3458         doutc(cl, "mds%d session %p state %s\n", mds, session,
3459               ceph_session_state_name(session->s_state));
3460
3461         /*
3462          * The old ceph will crash the MDSs when see unknown OPs
3463          */
3464         if (req->r_feature_needed > 0 &&
3465             !test_bit(req->r_feature_needed, &session->s_features)) {
3466                 err = -EOPNOTSUPP;
3467                 goto out_session;
3468         }
3469
3470         if (session->s_state != CEPH_MDS_SESSION_OPEN &&
3471             session->s_state != CEPH_MDS_SESSION_HUNG) {
3472                 /*
3473                  * We cannot queue async requests since the caps and delegated
3474                  * inodes are bound to the session. Just return -EJUKEBOX and
3475                  * let the caller retry a sync request in that case.
3476                  */
3477                 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
3478                         err = -EJUKEBOX;
3479                         goto out_session;
3480                 }
3481
3482                 /*
3483                  * If the session has been REJECTED, then return a hard error,
3484                  * unless it's a CLEANRECOVER mount, in which case we'll queue
3485                  * it to the mdsc queue.
3486                  */
3487                 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
3488                         if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
3489                                 list_add(&req->r_wait, &mdsc->waiting_for_map);
3490                         else
3491                                 err = -EACCES;
3492                         goto out_session;
3493                 }
3494
3495                 if (session->s_state == CEPH_MDS_SESSION_NEW ||
3496                     session->s_state == CEPH_MDS_SESSION_CLOSING) {
3497                         err = __open_session(mdsc, session);
3498                         if (err)
3499                                 goto out_session;
3500                         /* retry the same mds later */
3501                         if (random)
3502                                 req->r_resend_mds = mds;
3503                 }
3504                 list_add(&req->r_wait, &session->s_waiting);
3505                 goto out_session;
3506         }
3507
3508         /* send request */
3509         req->r_resend_mds = -1;   /* forget any previous mds hint */
3510
3511         if (req->r_request_started == 0)   /* note request start time */
3512                 req->r_request_started = jiffies;
3513
3514         /*
3515          * For async create we will choose the auth MDS of frag in parent
3516          * directory to send the request and usually this works fine, but
3517          * if the migrated the dirtory to another MDS before it could handle
3518          * it the request will be forwarded.
3519          *
3520          * And then the auth cap will be changed.
3521          */
3522         if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
3523                 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
3524                 struct ceph_inode_info *ci;
3525                 struct ceph_cap *cap;
3526
3527                 /*
3528                  * The request maybe handled very fast and the new inode
3529                  * hasn't been linked to the dentry yet. We need to wait
3530                  * for the ceph_finish_async_create(), which shouldn't be
3531                  * stuck too long or fail in thoery, to finish when forwarding
3532                  * the request.
3533                  */
3534                 if (!d_inode(req->r_dentry)) {
3535                         err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3536                                           TASK_KILLABLE);
3537                         if (err) {
3538                                 mutex_lock(&req->r_fill_mutex);
3539                                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3540                                 mutex_unlock(&req->r_fill_mutex);
3541                                 goto out_session;
3542                         }
3543                 }
3544
3545                 ci = ceph_inode(d_inode(req->r_dentry));
3546
3547                 spin_lock(&ci->i_ceph_lock);
3548                 cap = ci->i_auth_cap;
3549                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3550                         doutc(cl, "session changed for auth cap %d -> %d\n",
3551                               cap->session->s_mds, session->s_mds);
3552
3553                         /* Remove the auth cap from old session */
3554                         spin_lock(&cap->session->s_cap_lock);
3555                         cap->session->s_nr_caps--;
3556                         list_del_init(&cap->session_caps);
3557                         spin_unlock(&cap->session->s_cap_lock);
3558
3559                         /* Add the auth cap to the new session */
3560                         cap->mds = mds;
3561                         cap->session = session;
3562                         spin_lock(&session->s_cap_lock);
3563                         session->s_nr_caps++;
3564                         list_add_tail(&cap->session_caps, &session->s_caps);
3565                         spin_unlock(&session->s_cap_lock);
3566
3567                         change_auth_cap_ses(ci, session);
3568                 }
3569                 spin_unlock(&ci->i_ceph_lock);
3570         }
3571
3572         err = __send_request(session, req, false);
3573
3574 out_session:
3575         ceph_put_mds_session(session);
3576 finish:
3577         if (err) {
3578                 doutc(cl, "early error %d\n", err);
3579                 req->r_err = err;
3580                 complete_request(mdsc, req);
3581                 __unregister_request(mdsc, req);
3582         }
3583         return;
3584 }
3585
3586 /*
3587  * called under mdsc->mutex
3588  */
3589 static void __wake_requests(struct ceph_mds_client *mdsc,
3590                             struct list_head *head)
3591 {
3592         struct ceph_client *cl = mdsc->fsc->client;
3593         struct ceph_mds_request *req;
3594         LIST_HEAD(tmp_list);
3595
3596         list_splice_init(head, &tmp_list);
3597
3598         while (!list_empty(&tmp_list)) {
3599                 req = list_entry(tmp_list.next,
3600                                  struct ceph_mds_request, r_wait);
3601                 list_del_init(&req->r_wait);
3602                 doutc(cl, " wake request %p tid %llu\n", req,
3603                       req->r_tid);
3604                 __do_request(mdsc, req);
3605         }
3606 }
3607
3608 /*
3609  * Wake up threads with requests pending for @mds, so that they can
3610  * resubmit their requests to a possibly different mds.
3611  */
3612 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3613 {
3614         struct ceph_client *cl = mdsc->fsc->client;
3615         struct ceph_mds_request *req;
3616         struct rb_node *p = rb_first(&mdsc->request_tree);
3617
3618         doutc(cl, "kick_requests mds%d\n", mds);
3619         while (p) {
3620                 req = rb_entry(p, struct ceph_mds_request, r_node);
3621                 p = rb_next(p);
3622                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3623                         continue;
3624                 if (req->r_attempts > 0)
3625                         continue; /* only new requests */
3626                 if (req->r_session &&
3627                     req->r_session->s_mds == mds) {
3628                         doutc(cl, " kicking tid %llu\n", req->r_tid);
3629                         list_del_init(&req->r_wait);
3630                         __do_request(mdsc, req);
3631                 }
3632         }
3633 }
3634
3635 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3636                               struct ceph_mds_request *req)
3637 {
3638         struct ceph_client *cl = mdsc->fsc->client;
3639         int err = 0;
3640
3641         /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3642         if (req->r_inode)
3643                 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3644         if (req->r_parent) {
3645                 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3646                 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3647                             CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3648                 spin_lock(&ci->i_ceph_lock);
3649                 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3650                 __ceph_touch_fmode(ci, mdsc, fmode);
3651                 spin_unlock(&ci->i_ceph_lock);
3652         }
3653         if (req->r_old_dentry_dir)
3654                 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3655                                   CEPH_CAP_PIN);
3656
3657         if (req->r_inode) {
3658                 err = ceph_wait_on_async_create(req->r_inode);
3659                 if (err) {
3660                         doutc(cl, "wait for async create returned: %d\n", err);
3661                         return err;
3662                 }
3663         }
3664
3665         if (!err && req->r_old_inode) {
3666                 err = ceph_wait_on_async_create(req->r_old_inode);
3667                 if (err) {
3668                         doutc(cl, "wait for async create returned: %d\n", err);
3669                         return err;
3670                 }
3671         }
3672
3673         doutc(cl, "submit_request on %p for inode %p\n", req, dir);
3674         mutex_lock(&mdsc->mutex);
3675         __register_request(mdsc, req, dir);
3676         __do_request(mdsc, req);
3677         err = req->r_err;
3678         mutex_unlock(&mdsc->mutex);
3679         return err;
3680 }
3681
3682 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3683                            struct ceph_mds_request *req,
3684                            ceph_mds_request_wait_callback_t wait_func)
3685 {
3686         struct ceph_client *cl = mdsc->fsc->client;
3687         int err;
3688
3689         /* wait */
3690         doutc(cl, "do_request waiting\n");
3691         if (wait_func) {
3692                 err = wait_func(mdsc, req);
3693         } else {
3694                 long timeleft = wait_for_completion_killable_timeout(
3695                                         &req->r_completion,
3696                                         ceph_timeout_jiffies(req->r_timeout));
3697                 if (timeleft > 0)
3698                         err = 0;
3699                 else if (!timeleft)
3700                         err = -ETIMEDOUT;  /* timed out */
3701                 else
3702                         err = timeleft;  /* killed */
3703         }
3704         doutc(cl, "do_request waited, got %d\n", err);
3705         mutex_lock(&mdsc->mutex);
3706
3707         /* only abort if we didn't race with a real reply */
3708         if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3709                 err = le32_to_cpu(req->r_reply_info.head->result);
3710         } else if (err < 0) {
3711                 doutc(cl, "aborted request %lld with %d\n", req->r_tid, err);
3712
3713                 /*
3714                  * ensure we aren't running concurrently with
3715                  * ceph_fill_trace or ceph_readdir_prepopulate, which
3716                  * rely on locks (dir mutex) held by our caller.
3717                  */
3718                 mutex_lock(&req->r_fill_mutex);
3719                 req->r_err = err;
3720                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3721                 mutex_unlock(&req->r_fill_mutex);
3722
3723                 if (req->r_parent &&
3724                     (req->r_op & CEPH_MDS_OP_WRITE))
3725                         ceph_invalidate_dir_request(req);
3726         } else {
3727                 err = req->r_err;
3728         }
3729
3730         mutex_unlock(&mdsc->mutex);
3731         return err;
3732 }
3733
3734 /*
3735  * Synchrously perform an mds request.  Take care of all of the
3736  * session setup, forwarding, retry details.
3737  */
3738 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3739                          struct inode *dir,
3740                          struct ceph_mds_request *req)
3741 {
3742         struct ceph_client *cl = mdsc->fsc->client;
3743         int err;
3744
3745         doutc(cl, "do_request on %p\n", req);
3746
3747         /* issue */
3748         err = ceph_mdsc_submit_request(mdsc, dir, req);
3749         if (!err)
3750                 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3751         doutc(cl, "do_request %p done, result %d\n", req, err);
3752         return err;
3753 }
3754
3755 /*
3756  * Invalidate dir's completeness, dentry lease state on an aborted MDS
3757  * namespace request.
3758  */
3759 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3760 {
3761         struct inode *dir = req->r_parent;
3762         struct inode *old_dir = req->r_old_dentry_dir;
3763         struct ceph_client *cl = req->r_mdsc->fsc->client;
3764
3765         doutc(cl, "invalidate_dir_request %p %p (complete, lease(s))\n",
3766               dir, old_dir);
3767
3768         ceph_dir_clear_complete(dir);
3769         if (old_dir)
3770                 ceph_dir_clear_complete(old_dir);
3771         if (req->r_dentry)
3772                 ceph_invalidate_dentry_lease(req->r_dentry);
3773         if (req->r_old_dentry)
3774                 ceph_invalidate_dentry_lease(req->r_old_dentry);
3775 }
3776
3777 /*
3778  * Handle mds reply.
3779  *
3780  * We take the session mutex and parse and process the reply immediately.
3781  * This preserves the logical ordering of replies, capabilities, etc., sent
3782  * by the MDS as they are applied to our local cache.
3783  */
3784 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3785 {
3786         struct ceph_mds_client *mdsc = session->s_mdsc;
3787         struct ceph_client *cl = mdsc->fsc->client;
3788         struct ceph_mds_request *req;
3789         struct ceph_mds_reply_head *head = msg->front.iov_base;
3790         struct ceph_mds_reply_info_parsed *rinfo;  /* parsed reply info */
3791         struct ceph_snap_realm *realm;
3792         u64 tid;
3793         int err, result;
3794         int mds = session->s_mds;
3795         bool close_sessions = false;
3796
3797         if (msg->front.iov_len < sizeof(*head)) {
3798                 pr_err_client(cl, "got corrupt (short) reply\n");
3799                 ceph_msg_dump(msg);
3800                 return;
3801         }
3802
3803         /* get request, session */
3804         tid = le64_to_cpu(msg->hdr.tid);
3805         mutex_lock(&mdsc->mutex);
3806         req = lookup_get_request(mdsc, tid);
3807         if (!req) {
3808                 doutc(cl, "on unknown tid %llu\n", tid);
3809                 mutex_unlock(&mdsc->mutex);
3810                 return;
3811         }
3812         doutc(cl, "handle_reply %p\n", req);
3813
3814         /* correct session? */
3815         if (req->r_session != session) {
3816                 pr_err_client(cl, "got %llu on session mds%d not mds%d\n",
3817                               tid, session->s_mds,
3818                               req->r_session ? req->r_session->s_mds : -1);
3819                 mutex_unlock(&mdsc->mutex);
3820                 goto out;
3821         }
3822
3823         /* dup? */
3824         if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3825             (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3826                 pr_warn_client(cl, "got a dup %s reply on %llu from mds%d\n",
3827                                head->safe ? "safe" : "unsafe", tid, mds);
3828                 mutex_unlock(&mdsc->mutex);
3829                 goto out;
3830         }
3831         if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3832                 pr_warn_client(cl, "got unsafe after safe on %llu from mds%d\n",
3833                                tid, mds);
3834                 mutex_unlock(&mdsc->mutex);
3835                 goto out;
3836         }
3837
3838         result = le32_to_cpu(head->result);
3839
3840         if (head->safe) {
3841                 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3842                 __unregister_request(mdsc, req);
3843
3844                 /* last request during umount? */
3845                 if (mdsc->stopping && !__get_oldest_req(mdsc))
3846                         complete_all(&mdsc->safe_umount_waiters);
3847
3848                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3849                         /*
3850                          * We already handled the unsafe response, now do the
3851                          * cleanup.  No need to examine the response; the MDS
3852                          * doesn't include any result info in the safe
3853                          * response.  And even if it did, there is nothing
3854                          * useful we could do with a revised return value.
3855                          */
3856                         doutc(cl, "got safe reply %llu, mds%d\n", tid, mds);
3857
3858                         mutex_unlock(&mdsc->mutex);
3859                         goto out;
3860                 }
3861         } else {
3862                 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3863                 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3864         }
3865
3866         doutc(cl, "tid %lld result %d\n", tid, result);
3867         if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3868                 err = parse_reply_info(session, msg, req, (u64)-1);
3869         else
3870                 err = parse_reply_info(session, msg, req,
3871                                        session->s_con.peer_features);
3872         mutex_unlock(&mdsc->mutex);
3873
3874         /* Must find target inode outside of mutexes to avoid deadlocks */
3875         rinfo = &req->r_reply_info;
3876         if ((err >= 0) && rinfo->head->is_target) {
3877                 struct inode *in = xchg(&req->r_new_inode, NULL);
3878                 struct ceph_vino tvino = {
3879                         .ino  = le64_to_cpu(rinfo->targeti.in->ino),
3880                         .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3881                 };
3882
3883                 /*
3884                  * If we ended up opening an existing inode, discard
3885                  * r_new_inode
3886                  */
3887                 if (req->r_op == CEPH_MDS_OP_CREATE &&
3888                     !req->r_reply_info.has_create_ino) {
3889                         /* This should never happen on an async create */
3890                         WARN_ON_ONCE(req->r_deleg_ino);
3891                         iput(in);
3892                         in = NULL;
3893                 }
3894
3895                 in = ceph_get_inode(mdsc->fsc->sb, tvino, in);
3896                 if (IS_ERR(in)) {
3897                         err = PTR_ERR(in);
3898                         mutex_lock(&session->s_mutex);
3899                         goto out_err;
3900                 }
3901                 req->r_target_inode = in;
3902         }
3903
3904         mutex_lock(&session->s_mutex);
3905         if (err < 0) {
3906                 pr_err_client(cl, "got corrupt reply mds%d(tid:%lld)\n",
3907                               mds, tid);
3908                 ceph_msg_dump(msg);
3909                 goto out_err;
3910         }
3911
3912         /* snap trace */
3913         realm = NULL;
3914         if (rinfo->snapblob_len) {
3915                 down_write(&mdsc->snap_rwsem);
3916                 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3917                                 rinfo->snapblob + rinfo->snapblob_len,
3918                                 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3919                                 &realm);
3920                 if (err) {
3921                         up_write(&mdsc->snap_rwsem);
3922                         close_sessions = true;
3923                         if (err == -EIO)
3924                                 ceph_msg_dump(msg);
3925                         goto out_err;
3926                 }
3927                 downgrade_write(&mdsc->snap_rwsem);
3928         } else {
3929                 down_read(&mdsc->snap_rwsem);
3930         }
3931
3932         /* insert trace into our cache */
3933         mutex_lock(&req->r_fill_mutex);
3934         current->journal_info = req;
3935         err = ceph_fill_trace(mdsc->fsc->sb, req);
3936         if (err == 0) {
3937                 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3938                                     req->r_op == CEPH_MDS_OP_LSSNAP))
3939                         err = ceph_readdir_prepopulate(req, req->r_session);
3940         }
3941         current->journal_info = NULL;
3942         mutex_unlock(&req->r_fill_mutex);
3943
3944         up_read(&mdsc->snap_rwsem);
3945         if (realm)
3946                 ceph_put_snap_realm(mdsc, realm);
3947
3948         if (err == 0) {
3949                 if (req->r_target_inode &&
3950                     test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3951                         struct ceph_inode_info *ci =
3952                                 ceph_inode(req->r_target_inode);
3953                         spin_lock(&ci->i_unsafe_lock);
3954                         list_add_tail(&req->r_unsafe_target_item,
3955                                       &ci->i_unsafe_iops);
3956                         spin_unlock(&ci->i_unsafe_lock);
3957                 }
3958
3959                 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3960         }
3961 out_err:
3962         mutex_lock(&mdsc->mutex);
3963         if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3964                 if (err) {
3965                         req->r_err = err;
3966                 } else {
3967                         req->r_reply =  ceph_msg_get(msg);
3968                         set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3969                 }
3970         } else {
3971                 doutc(cl, "reply arrived after request %lld was aborted\n", tid);
3972         }
3973         mutex_unlock(&mdsc->mutex);
3974
3975         mutex_unlock(&session->s_mutex);
3976
3977         /* kick calling process */
3978         complete_request(mdsc, req);
3979
3980         ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3981                                      req->r_end_latency, err);
3982 out:
3983         ceph_mdsc_put_request(req);
3984
3985         /* Defer closing the sessions after s_mutex lock being released */
3986         if (close_sessions)
3987                 ceph_mdsc_close_sessions(mdsc);
3988         return;
3989 }
3990
3991
3992
3993 /*
3994  * handle mds notification that our request has been forwarded.
3995  */
3996 static void handle_forward(struct ceph_mds_client *mdsc,
3997                            struct ceph_mds_session *session,
3998                            struct ceph_msg *msg)
3999 {
4000         struct ceph_client *cl = mdsc->fsc->client;
4001         struct ceph_mds_request *req;
4002         u64 tid = le64_to_cpu(msg->hdr.tid);
4003         u32 next_mds;
4004         u32 fwd_seq;
4005         int err = -EINVAL;
4006         void *p = msg->front.iov_base;
4007         void *end = p + msg->front.iov_len;
4008         bool aborted = false;
4009
4010         ceph_decode_need(&p, end, 2*sizeof(u32), bad);
4011         next_mds = ceph_decode_32(&p);
4012         fwd_seq = ceph_decode_32(&p);
4013
4014         mutex_lock(&mdsc->mutex);
4015         req = lookup_get_request(mdsc, tid);
4016         if (!req) {
4017                 mutex_unlock(&mdsc->mutex);
4018                 doutc(cl, "forward tid %llu to mds%d - req dne\n", tid, next_mds);
4019                 return;  /* dup reply? */
4020         }
4021
4022         if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
4023                 doutc(cl, "forward tid %llu aborted, unregistering\n", tid);
4024                 __unregister_request(mdsc, req);
4025         } else if (fwd_seq <= req->r_num_fwd || (uint32_t)fwd_seq >= U32_MAX) {
4026                 /*
4027                  * Avoid infinite retrying after overflow.
4028                  *
4029                  * The MDS will increase the fwd count and in client side
4030                  * if the num_fwd is less than the one saved in request
4031                  * that means the MDS is an old version and overflowed of
4032                  * 8 bits.
4033                  */
4034                 mutex_lock(&req->r_fill_mutex);
4035                 req->r_err = -EMULTIHOP;
4036                 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
4037                 mutex_unlock(&req->r_fill_mutex);
4038                 aborted = true;
4039                 pr_warn_ratelimited_client(cl, "forward tid %llu seq overflow\n",
4040                                            tid);
4041         } else {
4042                 /* resend. forward race not possible; mds would drop */
4043                 doutc(cl, "forward tid %llu to mds%d (we resend)\n", tid, next_mds);
4044                 BUG_ON(req->r_err);
4045                 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
4046                 req->r_attempts = 0;
4047                 req->r_num_fwd = fwd_seq;
4048                 req->r_resend_mds = next_mds;
4049                 put_request_session(req);
4050                 __do_request(mdsc, req);
4051         }
4052         mutex_unlock(&mdsc->mutex);
4053
4054         /* kick calling process */
4055         if (aborted)
4056                 complete_request(mdsc, req);
4057         ceph_mdsc_put_request(req);
4058         return;
4059
4060 bad:
4061         pr_err_client(cl, "decode error err=%d\n", err);
4062         ceph_msg_dump(msg);
4063 }
4064
4065 static int __decode_session_metadata(void **p, void *end,
4066                                      bool *blocklisted)
4067 {
4068         /* map<string,string> */
4069         u32 n;
4070         bool err_str;
4071         ceph_decode_32_safe(p, end, n, bad);
4072         while (n-- > 0) {
4073                 u32 len;
4074                 ceph_decode_32_safe(p, end, len, bad);
4075                 ceph_decode_need(p, end, len, bad);
4076                 err_str = !strncmp(*p, "error_string", len);
4077                 *p += len;
4078                 ceph_decode_32_safe(p, end, len, bad);
4079                 ceph_decode_need(p, end, len, bad);
4080                 /*
4081                  * Match "blocklisted (blacklisted)" from newer MDSes,
4082                  * or "blacklisted" from older MDSes.
4083                  */
4084                 if (err_str && strnstr(*p, "blacklisted", len))
4085                         *blocklisted = true;
4086                 *p += len;
4087         }
4088         return 0;
4089 bad:
4090         return -1;
4091 }
4092
4093 /*
4094  * handle a mds session control message
4095  */
4096 static void handle_session(struct ceph_mds_session *session,
4097                            struct ceph_msg *msg)
4098 {
4099         struct ceph_mds_client *mdsc = session->s_mdsc;
4100         struct ceph_client *cl = mdsc->fsc->client;
4101         int mds = session->s_mds;
4102         int msg_version = le16_to_cpu(msg->hdr.version);
4103         void *p = msg->front.iov_base;
4104         void *end = p + msg->front.iov_len;
4105         struct ceph_mds_session_head *h;
4106         struct ceph_mds_cap_auth *cap_auths = NULL;
4107         u32 op, cap_auths_num = 0;
4108         u64 seq, features = 0;
4109         int wake = 0;
4110         bool blocklisted = false;
4111         u32 i;
4112
4113
4114         /* decode */
4115         ceph_decode_need(&p, end, sizeof(*h), bad);
4116         h = p;
4117         p += sizeof(*h);
4118
4119         op = le32_to_cpu(h->op);
4120         seq = le64_to_cpu(h->seq);
4121
4122         if (msg_version >= 3) {
4123                 u32 len;
4124                 /* version >= 2 and < 5, decode metadata, skip otherwise
4125                  * as it's handled via flags.
4126                  */
4127                 if (msg_version >= 5)
4128                         ceph_decode_skip_map(&p, end, string, string, bad);
4129                 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
4130                         goto bad;
4131
4132                 /* version >= 3, feature bits */
4133                 ceph_decode_32_safe(&p, end, len, bad);
4134                 if (len) {
4135                         ceph_decode_64_safe(&p, end, features, bad);
4136                         p += len - sizeof(features);
4137                 }
4138         }
4139
4140         if (msg_version >= 5) {
4141                 u32 flags, len;
4142
4143                 /* version >= 4 */
4144                 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
4145                 ceph_decode_32_safe(&p, end, len, bad); /* len */
4146                 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
4147
4148                 /* version >= 5, flags   */
4149                 ceph_decode_32_safe(&p, end, flags, bad);
4150                 if (flags & CEPH_SESSION_BLOCKLISTED) {
4151                         pr_warn_client(cl, "mds%d session blocklisted\n",
4152                                        session->s_mds);
4153                         blocklisted = true;
4154                 }
4155         }
4156
4157         if (msg_version >= 6) {
4158                 ceph_decode_32_safe(&p, end, cap_auths_num, bad);
4159                 doutc(cl, "cap_auths_num %d\n", cap_auths_num);
4160
4161                 if (cap_auths_num && op != CEPH_SESSION_OPEN) {
4162                         WARN_ON_ONCE(op != CEPH_SESSION_OPEN);
4163                         goto skip_cap_auths;
4164                 }
4165
4166                 cap_auths = kcalloc(cap_auths_num,
4167                                     sizeof(struct ceph_mds_cap_auth),
4168                                     GFP_KERNEL);
4169                 if (!cap_auths) {
4170                         pr_err_client(cl, "No memory for cap_auths\n");
4171                         return;
4172                 }
4173
4174                 for (i = 0; i < cap_auths_num; i++) {
4175                         u32 _len, j;
4176
4177                         /* struct_v, struct_compat, and struct_len in MDSCapAuth */
4178                         ceph_decode_skip_n(&p, end, 2 + sizeof(u32), bad);
4179
4180                         /* struct_v, struct_compat, and struct_len in MDSCapMatch */
4181                         ceph_decode_skip_n(&p, end, 2 + sizeof(u32), bad);
4182                         ceph_decode_64_safe(&p, end, cap_auths[i].match.uid, bad);
4183                         ceph_decode_32_safe(&p, end, _len, bad);
4184                         if (_len) {
4185                                 cap_auths[i].match.gids = kcalloc(_len, sizeof(u32),
4186                                                                   GFP_KERNEL);
4187                                 if (!cap_auths[i].match.gids) {
4188                                         pr_err_client(cl, "No memory for gids\n");
4189                                         goto fail;
4190                                 }
4191
4192                                 cap_auths[i].match.num_gids = _len;
4193                                 for (j = 0; j < _len; j++)
4194                                         ceph_decode_32_safe(&p, end,
4195                                                             cap_auths[i].match.gids[j],
4196                                                             bad);
4197                         }
4198
4199                         ceph_decode_32_safe(&p, end, _len, bad);
4200                         if (_len) {
4201                                 cap_auths[i].match.path = kcalloc(_len + 1, sizeof(char),
4202                                                                   GFP_KERNEL);
4203                                 if (!cap_auths[i].match.path) {
4204                                         pr_err_client(cl, "No memory for path\n");
4205                                         goto fail;
4206                                 }
4207                                 ceph_decode_copy(&p, cap_auths[i].match.path, _len);
4208
4209                                 /* Remove the tailing '/' */
4210                                 while (_len && cap_auths[i].match.path[_len - 1] == '/') {
4211                                         cap_auths[i].match.path[_len - 1] = '\0';
4212                                         _len -= 1;
4213                                 }
4214                         }
4215
4216                         ceph_decode_32_safe(&p, end, _len, bad);
4217                         if (_len) {
4218                                 cap_auths[i].match.fs_name = kcalloc(_len + 1, sizeof(char),
4219                                                                      GFP_KERNEL);
4220                                 if (!cap_auths[i].match.fs_name) {
4221                                         pr_err_client(cl, "No memory for fs_name\n");
4222                                         goto fail;
4223                                 }
4224                                 ceph_decode_copy(&p, cap_auths[i].match.fs_name, _len);
4225                         }
4226
4227                         ceph_decode_8_safe(&p, end, cap_auths[i].match.root_squash, bad);
4228                         ceph_decode_8_safe(&p, end, cap_auths[i].readable, bad);
4229                         ceph_decode_8_safe(&p, end, cap_auths[i].writeable, bad);
4230                         doutc(cl, "uid %lld, num_gids %u, path %s, fs_name %s, root_squash %d, readable %d, writeable %d\n",
4231                               cap_auths[i].match.uid, cap_auths[i].match.num_gids,
4232                               cap_auths[i].match.path, cap_auths[i].match.fs_name,
4233                               cap_auths[i].match.root_squash,
4234                               cap_auths[i].readable, cap_auths[i].writeable);
4235                 }
4236         }
4237
4238 skip_cap_auths:
4239         mutex_lock(&mdsc->mutex);
4240         if (op == CEPH_SESSION_OPEN) {
4241                 if (mdsc->s_cap_auths) {
4242                         for (i = 0; i < mdsc->s_cap_auths_num; i++) {
4243                                 kfree(mdsc->s_cap_auths[i].match.gids);
4244                                 kfree(mdsc->s_cap_auths[i].match.path);
4245                                 kfree(mdsc->s_cap_auths[i].match.fs_name);
4246                         }
4247                         kfree(mdsc->s_cap_auths);
4248                 }
4249                 mdsc->s_cap_auths_num = cap_auths_num;
4250                 mdsc->s_cap_auths = cap_auths;
4251         }
4252         if (op == CEPH_SESSION_CLOSE) {
4253                 ceph_get_mds_session(session);
4254                 __unregister_session(mdsc, session);
4255         }
4256         /* FIXME: this ttl calculation is generous */
4257         session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
4258         mutex_unlock(&mdsc->mutex);
4259
4260         mutex_lock(&session->s_mutex);
4261
4262         doutc(cl, "mds%d %s %p state %s seq %llu\n", mds,
4263               ceph_session_op_name(op), session,
4264               ceph_session_state_name(session->s_state), seq);
4265
4266         if (session->s_state == CEPH_MDS_SESSION_HUNG) {
4267                 session->s_state = CEPH_MDS_SESSION_OPEN;
4268                 pr_info_client(cl, "mds%d came back\n", session->s_mds);
4269         }
4270
4271         switch (op) {
4272         case CEPH_SESSION_OPEN:
4273                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4274                         pr_info_client(cl, "mds%d reconnect success\n",
4275                                        session->s_mds);
4276
4277                 session->s_features = features;
4278                 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
4279                         pr_notice_client(cl, "mds%d is already opened\n",
4280                                          session->s_mds);
4281                 } else {
4282                         session->s_state = CEPH_MDS_SESSION_OPEN;
4283                         renewed_caps(mdsc, session, 0);
4284                         if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
4285                                      &session->s_features))
4286                                 metric_schedule_delayed(&mdsc->metric);
4287                 }
4288
4289                 /*
4290                  * The connection maybe broken and the session in client
4291                  * side has been reinitialized, need to update the seq
4292                  * anyway.
4293                  */
4294                 if (!session->s_seq && seq)
4295                         session->s_seq = seq;
4296
4297                 wake = 1;
4298                 if (mdsc->stopping)
4299                         __close_session(mdsc, session);
4300                 break;
4301
4302         case CEPH_SESSION_RENEWCAPS:
4303                 if (session->s_renew_seq == seq)
4304                         renewed_caps(mdsc, session, 1);
4305                 break;
4306
4307         case CEPH_SESSION_CLOSE:
4308                 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
4309                         pr_info_client(cl, "mds%d reconnect denied\n",
4310                                        session->s_mds);
4311                 session->s_state = CEPH_MDS_SESSION_CLOSED;
4312                 cleanup_session_requests(mdsc, session);
4313                 remove_session_caps(session);
4314                 wake = 2; /* for good measure */
4315                 wake_up_all(&mdsc->session_close_wq);
4316                 break;
4317
4318         case CEPH_SESSION_STALE:
4319                 pr_info_client(cl, "mds%d caps went stale, renewing\n",
4320                                session->s_mds);
4321                 atomic_inc(&session->s_cap_gen);
4322                 session->s_cap_ttl = jiffies - 1;
4323                 send_renew_caps(mdsc, session);
4324                 break;
4325
4326         case CEPH_SESSION_RECALL_STATE:
4327                 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
4328                 break;
4329
4330         case CEPH_SESSION_FLUSHMSG:
4331                 /* flush cap releases */
4332                 spin_lock(&session->s_cap_lock);
4333                 if (session->s_num_cap_releases)
4334                         ceph_flush_session_cap_releases(mdsc, session);
4335                 spin_unlock(&session->s_cap_lock);
4336
4337                 send_flushmsg_ack(mdsc, session, seq);
4338                 break;
4339
4340         case CEPH_SESSION_FORCE_RO:
4341                 doutc(cl, "force_session_readonly %p\n", session);
4342                 spin_lock(&session->s_cap_lock);
4343                 session->s_readonly = true;
4344                 spin_unlock(&session->s_cap_lock);
4345                 wake_up_session_caps(session, FORCE_RO);
4346                 break;
4347
4348         case CEPH_SESSION_REJECT:
4349                 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
4350                 pr_info_client(cl, "mds%d rejected session\n",
4351                                session->s_mds);
4352                 session->s_state = CEPH_MDS_SESSION_REJECTED;
4353                 cleanup_session_requests(mdsc, session);
4354                 remove_session_caps(session);
4355                 if (blocklisted)
4356                         mdsc->fsc->blocklisted = true;
4357                 wake = 2; /* for good measure */
4358                 break;
4359
4360         default:
4361                 pr_err_client(cl, "bad op %d mds%d\n", op, mds);
4362                 WARN_ON(1);
4363         }
4364
4365         mutex_unlock(&session->s_mutex);
4366         if (wake) {
4367                 mutex_lock(&mdsc->mutex);
4368                 __wake_requests(mdsc, &session->s_waiting);
4369                 if (wake == 2)
4370                         kick_requests(mdsc, mds);
4371                 mutex_unlock(&mdsc->mutex);
4372         }
4373         if (op == CEPH_SESSION_CLOSE)
4374                 ceph_put_mds_session(session);
4375         return;
4376
4377 bad:
4378         pr_err_client(cl, "corrupt message mds%d len %d\n", mds,
4379                       (int)msg->front.iov_len);
4380         ceph_msg_dump(msg);
4381 fail:
4382         for (i = 0; i < cap_auths_num; i++) {
4383                 kfree(cap_auths[i].match.gids);
4384                 kfree(cap_auths[i].match.path);
4385                 kfree(cap_auths[i].match.fs_name);
4386         }
4387         kfree(cap_auths);
4388         return;
4389 }
4390
4391 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
4392 {
4393         struct ceph_client *cl = req->r_mdsc->fsc->client;
4394         int dcaps;
4395
4396         dcaps = xchg(&req->r_dir_caps, 0);
4397         if (dcaps) {
4398                 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4399                 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
4400         }
4401 }
4402
4403 void ceph_mdsc_release_dir_caps_async(struct ceph_mds_request *req)
4404 {
4405         struct ceph_client *cl = req->r_mdsc->fsc->client;
4406         int dcaps;
4407
4408         dcaps = xchg(&req->r_dir_caps, 0);
4409         if (dcaps) {
4410                 doutc(cl, "releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
4411                 ceph_put_cap_refs_async(ceph_inode(req->r_parent), dcaps);
4412         }
4413 }
4414
4415 /*
4416  * called under session->mutex.
4417  */
4418 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
4419                                    struct ceph_mds_session *session)
4420 {
4421         struct ceph_mds_request *req, *nreq;
4422         struct rb_node *p;
4423
4424         doutc(mdsc->fsc->client, "mds%d\n", session->s_mds);
4425
4426         mutex_lock(&mdsc->mutex);
4427         list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
4428                 __send_request(session, req, true);
4429
4430         /*
4431          * also re-send old requests when MDS enters reconnect stage. So that MDS
4432          * can process completed request in clientreplay stage.
4433          */
4434         p = rb_first(&mdsc->request_tree);
4435         while (p) {
4436                 req = rb_entry(p, struct ceph_mds_request, r_node);
4437                 p = rb_next(p);
4438                 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
4439                         continue;
4440                 if (req->r_attempts == 0)
4441                         continue; /* only old requests */
4442                 if (!req->r_session)
4443                         continue;
4444                 if (req->r_session->s_mds != session->s_mds)
4445                         continue;
4446
4447                 ceph_mdsc_release_dir_caps_async(req);
4448
4449                 __send_request(session, req, true);
4450         }
4451         mutex_unlock(&mdsc->mutex);
4452 }
4453
4454 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
4455 {
4456         struct ceph_msg *reply;
4457         struct ceph_pagelist *_pagelist;
4458         struct page *page;
4459         __le32 *addr;
4460         int err = -ENOMEM;
4461
4462         if (!recon_state->allow_multi)
4463                 return -ENOSPC;
4464
4465         /* can't handle message that contains both caps and realm */
4466         BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
4467
4468         /* pre-allocate new pagelist */
4469         _pagelist = ceph_pagelist_alloc(GFP_NOFS);
4470         if (!_pagelist)
4471                 return -ENOMEM;
4472
4473         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4474         if (!reply)
4475                 goto fail_msg;
4476
4477         /* placeholder for nr_caps */
4478         err = ceph_pagelist_encode_32(_pagelist, 0);
4479         if (err < 0)
4480                 goto fail;
4481
4482         if (recon_state->nr_caps) {
4483                 /* currently encoding caps */
4484                 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
4485                 if (err)
4486                         goto fail;
4487         } else {
4488                 /* placeholder for nr_realms (currently encoding relams) */
4489                 err = ceph_pagelist_encode_32(_pagelist, 0);
4490                 if (err < 0)
4491                         goto fail;
4492         }
4493
4494         err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
4495         if (err)
4496                 goto fail;
4497
4498         page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
4499         addr = kmap_atomic(page);
4500         if (recon_state->nr_caps) {
4501                 /* currently encoding caps */
4502                 *addr = cpu_to_le32(recon_state->nr_caps);
4503         } else {
4504                 /* currently encoding relams */
4505                 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
4506         }
4507         kunmap_atomic(addr);
4508
4509         reply->hdr.version = cpu_to_le16(5);
4510         reply->hdr.compat_version = cpu_to_le16(4);
4511
4512         reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
4513         ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
4514
4515         ceph_con_send(&recon_state->session->s_con, reply);
4516         ceph_pagelist_release(recon_state->pagelist);
4517
4518         recon_state->pagelist = _pagelist;
4519         recon_state->nr_caps = 0;
4520         recon_state->nr_realms = 0;
4521         recon_state->msg_version = 5;
4522         return 0;
4523 fail:
4524         ceph_msg_put(reply);
4525 fail_msg:
4526         ceph_pagelist_release(_pagelist);
4527         return err;
4528 }
4529
4530 static struct dentry* d_find_primary(struct inode *inode)
4531 {
4532         struct dentry *alias, *dn = NULL;
4533
4534         if (hlist_empty(&inode->i_dentry))
4535                 return NULL;
4536
4537         spin_lock(&inode->i_lock);
4538         if (hlist_empty(&inode->i_dentry))
4539                 goto out_unlock;
4540
4541         if (S_ISDIR(inode->i_mode)) {
4542                 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
4543                 if (!IS_ROOT(alias))
4544                         dn = dget(alias);
4545                 goto out_unlock;
4546         }
4547
4548         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
4549                 spin_lock(&alias->d_lock);
4550                 if (!d_unhashed(alias) &&
4551                     (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
4552                         dn = dget_dlock(alias);
4553                 }
4554                 spin_unlock(&alias->d_lock);
4555                 if (dn)
4556                         break;
4557         }
4558 out_unlock:
4559         spin_unlock(&inode->i_lock);
4560         return dn;
4561 }
4562
4563 /*
4564  * Encode information about a cap for a reconnect with the MDS.
4565  */
4566 static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
4567 {
4568         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
4569         struct ceph_client *cl = ceph_inode_to_client(inode);
4570         union {
4571                 struct ceph_mds_cap_reconnect v2;
4572                 struct ceph_mds_cap_reconnect_v1 v1;
4573         } rec;
4574         struct ceph_inode_info *ci = ceph_inode(inode);
4575         struct ceph_reconnect_state *recon_state = arg;
4576         struct ceph_pagelist *pagelist = recon_state->pagelist;
4577         struct dentry *dentry;
4578         struct ceph_cap *cap;
4579         char *path;
4580         int pathlen = 0, err;
4581         u64 pathbase;
4582         u64 snap_follows;
4583
4584         dentry = d_find_primary(inode);
4585         if (dentry) {
4586                 /* set pathbase to parent dir when msg_version >= 2 */
4587                 path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase,
4588                                             recon_state->msg_version >= 2);
4589                 dput(dentry);
4590                 if (IS_ERR(path)) {
4591                         err = PTR_ERR(path);
4592                         goto out_err;
4593                 }
4594         } else {
4595                 path = NULL;
4596                 pathbase = 0;
4597         }
4598
4599         spin_lock(&ci->i_ceph_lock);
4600         cap = __get_cap_for_mds(ci, mds);
4601         if (!cap) {
4602                 spin_unlock(&ci->i_ceph_lock);
4603                 err = 0;
4604                 goto out_err;
4605         }
4606         doutc(cl, " adding %p ino %llx.%llx cap %p %lld %s\n", inode,
4607               ceph_vinop(inode), cap, cap->cap_id,
4608               ceph_cap_string(cap->issued));
4609
4610         cap->seq = 0;        /* reset cap seq */
4611         cap->issue_seq = 0;  /* and issue_seq */
4612         cap->mseq = 0;       /* and migrate_seq */
4613         cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
4614
4615         /* These are lost when the session goes away */
4616         if (S_ISDIR(inode->i_mode)) {
4617                 if (cap->issued & CEPH_CAP_DIR_CREATE) {
4618                         ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
4619                         memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
4620                 }
4621                 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
4622         }
4623
4624         if (recon_state->msg_version >= 2) {
4625                 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
4626                 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4627                 rec.v2.issued = cpu_to_le32(cap->issued);
4628                 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4629                 rec.v2.pathbase = cpu_to_le64(pathbase);
4630                 rec.v2.flock_len = (__force __le32)
4631                         ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
4632         } else {
4633                 struct timespec64 ts;
4634
4635                 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
4636                 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
4637                 rec.v1.issued = cpu_to_le32(cap->issued);
4638                 rec.v1.size = cpu_to_le64(i_size_read(inode));
4639                 ts = inode_get_mtime(inode);
4640                 ceph_encode_timespec64(&rec.v1.mtime, &ts);
4641                 ts = inode_get_atime(inode);
4642                 ceph_encode_timespec64(&rec.v1.atime, &ts);
4643                 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
4644                 rec.v1.pathbase = cpu_to_le64(pathbase);
4645         }
4646
4647         if (list_empty(&ci->i_cap_snaps)) {
4648                 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
4649         } else {
4650                 struct ceph_cap_snap *capsnap =
4651                         list_first_entry(&ci->i_cap_snaps,
4652                                          struct ceph_cap_snap, ci_item);
4653                 snap_follows = capsnap->follows;
4654         }
4655         spin_unlock(&ci->i_ceph_lock);
4656
4657         if (recon_state->msg_version >= 2) {
4658                 int num_fcntl_locks, num_flock_locks;
4659                 struct ceph_filelock *flocks = NULL;
4660                 size_t struct_len, total_len = sizeof(u64);
4661                 u8 struct_v = 0;
4662
4663 encode_again:
4664                 if (rec.v2.flock_len) {
4665                         ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4666                 } else {
4667                         num_fcntl_locks = 0;
4668                         num_flock_locks = 0;
4669                 }
4670                 if (num_fcntl_locks + num_flock_locks > 0) {
4671                         flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4672                                                sizeof(struct ceph_filelock),
4673                                                GFP_NOFS);
4674                         if (!flocks) {
4675                                 err = -ENOMEM;
4676                                 goto out_err;
4677                         }
4678                         err = ceph_encode_locks_to_buffer(inode, flocks,
4679                                                           num_fcntl_locks,
4680                                                           num_flock_locks);
4681                         if (err) {
4682                                 kfree(flocks);
4683                                 flocks = NULL;
4684                                 if (err == -ENOSPC)
4685                                         goto encode_again;
4686                                 goto out_err;
4687                         }
4688                 } else {
4689                         kfree(flocks);
4690                         flocks = NULL;
4691                 }
4692
4693                 if (recon_state->msg_version >= 3) {
4694                         /* version, compat_version and struct_len */
4695                         total_len += 2 * sizeof(u8) + sizeof(u32);
4696                         struct_v = 2;
4697                 }
4698                 /*
4699                  * number of encoded locks is stable, so copy to pagelist
4700                  */
4701                 struct_len = 2 * sizeof(u32) +
4702                             (num_fcntl_locks + num_flock_locks) *
4703                             sizeof(struct ceph_filelock);
4704                 rec.v2.flock_len = cpu_to_le32(struct_len);
4705
4706                 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4707
4708                 if (struct_v >= 2)
4709                         struct_len += sizeof(u64); /* snap_follows */
4710
4711                 total_len += struct_len;
4712
4713                 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4714                         err = send_reconnect_partial(recon_state);
4715                         if (err)
4716                                 goto out_freeflocks;
4717                         pagelist = recon_state->pagelist;
4718                 }
4719
4720                 err = ceph_pagelist_reserve(pagelist, total_len);
4721                 if (err)
4722                         goto out_freeflocks;
4723
4724                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4725                 if (recon_state->msg_version >= 3) {
4726                         ceph_pagelist_encode_8(pagelist, struct_v);
4727                         ceph_pagelist_encode_8(pagelist, 1);
4728                         ceph_pagelist_encode_32(pagelist, struct_len);
4729                 }
4730                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4731                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4732                 ceph_locks_to_pagelist(flocks, pagelist,
4733                                        num_fcntl_locks, num_flock_locks);
4734                 if (struct_v >= 2)
4735                         ceph_pagelist_encode_64(pagelist, snap_follows);
4736 out_freeflocks:
4737                 kfree(flocks);
4738         } else {
4739                 err = ceph_pagelist_reserve(pagelist,
4740                                             sizeof(u64) + sizeof(u32) +
4741                                             pathlen + sizeof(rec.v1));
4742                 if (err)
4743                         goto out_err;
4744
4745                 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4746                 ceph_pagelist_encode_string(pagelist, path, pathlen);
4747                 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4748         }
4749
4750 out_err:
4751         ceph_mdsc_free_path(path, pathlen);
4752         if (!err)
4753                 recon_state->nr_caps++;
4754         return err;
4755 }
4756
4757 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4758                               struct ceph_reconnect_state *recon_state)
4759 {
4760         struct rb_node *p;
4761         struct ceph_pagelist *pagelist = recon_state->pagelist;
4762         struct ceph_client *cl = mdsc->fsc->client;
4763         int err = 0;
4764
4765         if (recon_state->msg_version >= 4) {
4766                 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4767                 if (err < 0)
4768                         goto fail;
4769         }
4770
4771         /*
4772          * snaprealms.  we provide mds with the ino, seq (version), and
4773          * parent for all of our realms.  If the mds has any newer info,
4774          * it will tell us.
4775          */
4776         for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4777                 struct ceph_snap_realm *realm =
4778                        rb_entry(p, struct ceph_snap_realm, node);
4779                 struct ceph_mds_snaprealm_reconnect sr_rec;
4780
4781                 if (recon_state->msg_version >= 4) {
4782                         size_t need = sizeof(u8) * 2 + sizeof(u32) +
4783                                       sizeof(sr_rec);
4784
4785                         if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4786                                 err = send_reconnect_partial(recon_state);
4787                                 if (err)
4788                                         goto fail;
4789                                 pagelist = recon_state->pagelist;
4790                         }
4791
4792                         err = ceph_pagelist_reserve(pagelist, need);
4793                         if (err)
4794                                 goto fail;
4795
4796                         ceph_pagelist_encode_8(pagelist, 1);
4797                         ceph_pagelist_encode_8(pagelist, 1);
4798                         ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4799                 }
4800
4801                 doutc(cl, " adding snap realm %llx seq %lld parent %llx\n",
4802                       realm->ino, realm->seq, realm->parent_ino);
4803                 sr_rec.ino = cpu_to_le64(realm->ino);
4804                 sr_rec.seq = cpu_to_le64(realm->seq);
4805                 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4806
4807                 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4808                 if (err)
4809                         goto fail;
4810
4811                 recon_state->nr_realms++;
4812         }
4813 fail:
4814         return err;
4815 }
4816
4817
4818 /*
4819  * If an MDS fails and recovers, clients need to reconnect in order to
4820  * reestablish shared state.  This includes all caps issued through
4821  * this session _and_ the snap_realm hierarchy.  Because it's not
4822  * clear which snap realms the mds cares about, we send everything we
4823  * know about.. that ensures we'll then get any new info the
4824  * recovering MDS might have.
4825  *
4826  * This is a relatively heavyweight operation, but it's rare.
4827  */
4828 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4829                                struct ceph_mds_session *session)
4830 {
4831         struct ceph_client *cl = mdsc->fsc->client;
4832         struct ceph_msg *reply;
4833         int mds = session->s_mds;
4834         int err = -ENOMEM;
4835         struct ceph_reconnect_state recon_state = {
4836                 .session = session,
4837         };
4838         LIST_HEAD(dispose);
4839
4840         pr_info_client(cl, "mds%d reconnect start\n", mds);
4841
4842         recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4843         if (!recon_state.pagelist)
4844                 goto fail_nopagelist;
4845
4846         reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4847         if (!reply)
4848                 goto fail_nomsg;
4849
4850         xa_destroy(&session->s_delegated_inos);
4851
4852         mutex_lock(&session->s_mutex);
4853         session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4854         session->s_seq = 0;
4855
4856         doutc(cl, "session %p state %s\n", session,
4857               ceph_session_state_name(session->s_state));
4858
4859         atomic_inc(&session->s_cap_gen);
4860
4861         spin_lock(&session->s_cap_lock);
4862         /* don't know if session is readonly */
4863         session->s_readonly = 0;
4864         /*
4865          * notify __ceph_remove_cap() that we are composing cap reconnect.
4866          * If a cap get released before being added to the cap reconnect,
4867          * __ceph_remove_cap() should skip queuing cap release.
4868          */
4869         session->s_cap_reconnect = 1;
4870         /* drop old cap expires; we're about to reestablish that state */
4871         detach_cap_releases(session, &dispose);
4872         spin_unlock(&session->s_cap_lock);
4873         dispose_cap_releases(mdsc, &dispose);
4874
4875         /* trim unused caps to reduce MDS's cache rejoin time */
4876         if (mdsc->fsc->sb->s_root)
4877                 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4878
4879         ceph_con_close(&session->s_con);
4880         ceph_con_open(&session->s_con,
4881                       CEPH_ENTITY_TYPE_MDS, mds,
4882                       ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4883
4884         /* replay unsafe requests */
4885         replay_unsafe_requests(mdsc, session);
4886
4887         ceph_early_kick_flushing_caps(mdsc, session);
4888
4889         down_read(&mdsc->snap_rwsem);
4890
4891         /* placeholder for nr_caps */
4892         err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4893         if (err)
4894                 goto fail;
4895
4896         if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4897                 recon_state.msg_version = 3;
4898                 recon_state.allow_multi = true;
4899         } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4900                 recon_state.msg_version = 3;
4901         } else {
4902                 recon_state.msg_version = 2;
4903         }
4904         /* traverse this session's caps */
4905         err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4906
4907         spin_lock(&session->s_cap_lock);
4908         session->s_cap_reconnect = 0;
4909         spin_unlock(&session->s_cap_lock);
4910
4911         if (err < 0)
4912                 goto fail;
4913
4914         /* check if all realms can be encoded into current message */
4915         if (mdsc->num_snap_realms) {
4916                 size_t total_len =
4917                         recon_state.pagelist->length +
4918                         mdsc->num_snap_realms *
4919                         sizeof(struct ceph_mds_snaprealm_reconnect);
4920                 if (recon_state.msg_version >= 4) {
4921                         /* number of realms */
4922                         total_len += sizeof(u32);
4923                         /* version, compat_version and struct_len */
4924                         total_len += mdsc->num_snap_realms *
4925                                      (2 * sizeof(u8) + sizeof(u32));
4926                 }
4927                 if (total_len > RECONNECT_MAX_SIZE) {
4928                         if (!recon_state.allow_multi) {
4929                                 err = -ENOSPC;
4930                                 goto fail;
4931                         }
4932                         if (recon_state.nr_caps) {
4933                                 err = send_reconnect_partial(&recon_state);
4934                                 if (err)
4935                                         goto fail;
4936                         }
4937                         recon_state.msg_version = 5;
4938                 }
4939         }
4940
4941         err = encode_snap_realms(mdsc, &recon_state);
4942         if (err < 0)
4943                 goto fail;
4944
4945         if (recon_state.msg_version >= 5) {
4946                 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4947                 if (err < 0)
4948                         goto fail;
4949         }
4950
4951         if (recon_state.nr_caps || recon_state.nr_realms) {
4952                 struct page *page =
4953                         list_first_entry(&recon_state.pagelist->head,
4954                                         struct page, lru);
4955                 __le32 *addr = kmap_atomic(page);
4956                 if (recon_state.nr_caps) {
4957                         WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4958                         *addr = cpu_to_le32(recon_state.nr_caps);
4959                 } else if (recon_state.msg_version >= 4) {
4960                         *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4961                 }
4962                 kunmap_atomic(addr);
4963         }
4964
4965         reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4966         if (recon_state.msg_version >= 4)
4967                 reply->hdr.compat_version = cpu_to_le16(4);
4968
4969         reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4970         ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4971
4972         ceph_con_send(&session->s_con, reply);
4973
4974         mutex_unlock(&session->s_mutex);
4975
4976         mutex_lock(&mdsc->mutex);
4977         __wake_requests(mdsc, &session->s_waiting);
4978         mutex_unlock(&mdsc->mutex);
4979
4980         up_read(&mdsc->snap_rwsem);
4981         ceph_pagelist_release(recon_state.pagelist);
4982         return;
4983
4984 fail:
4985         ceph_msg_put(reply);
4986         up_read(&mdsc->snap_rwsem);
4987         mutex_unlock(&session->s_mutex);
4988 fail_nomsg:
4989         ceph_pagelist_release(recon_state.pagelist);
4990 fail_nopagelist:
4991         pr_err_client(cl, "error %d preparing reconnect for mds%d\n",
4992                       err, mds);
4993         return;
4994 }
4995
4996
4997 /*
4998  * compare old and new mdsmaps, kicking requests
4999  * and closing out old connections as necessary
5000  *
5001  * called under mdsc->mutex.
5002  */
5003 static void check_new_map(struct ceph_mds_client *mdsc,
5004                           struct ceph_mdsmap *newmap,
5005                           struct ceph_mdsmap *oldmap)
5006 {
5007         int i, j, err;
5008         int oldstate, newstate;
5009         struct ceph_mds_session *s;
5010         unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
5011         struct ceph_client *cl = mdsc->fsc->client;
5012
5013         doutc(cl, "new %u old %u\n", newmap->m_epoch, oldmap->m_epoch);
5014
5015         if (newmap->m_info) {
5016                 for (i = 0; i < newmap->possible_max_rank; i++) {
5017                         for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
5018                                 set_bit(newmap->m_info[i].export_targets[j], targets);
5019                 }
5020         }
5021
5022         for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
5023                 if (!mdsc->sessions[i])
5024                         continue;
5025                 s = mdsc->sessions[i];
5026                 oldstate = ceph_mdsmap_get_state(oldmap, i);
5027                 newstate = ceph_mdsmap_get_state(newmap, i);
5028
5029                 doutc(cl, "mds%d state %s%s -> %s%s (session %s)\n",
5030                       i, ceph_mds_state_name(oldstate),
5031                       ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
5032                       ceph_mds_state_name(newstate),
5033                       ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
5034                       ceph_session_state_name(s->s_state));
5035
5036                 if (i >= newmap->possible_max_rank) {
5037                         /* force close session for stopped mds */
5038                         ceph_get_mds_session(s);
5039                         __unregister_session(mdsc, s);
5040                         __wake_requests(mdsc, &s->s_waiting);
5041                         mutex_unlock(&mdsc->mutex);
5042
5043                         mutex_lock(&s->s_mutex);
5044                         cleanup_session_requests(mdsc, s);
5045                         remove_session_caps(s);
5046                         mutex_unlock(&s->s_mutex);
5047
5048                         ceph_put_mds_session(s);
5049
5050                         mutex_lock(&mdsc->mutex);
5051                         kick_requests(mdsc, i);
5052                         continue;
5053                 }
5054
5055                 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
5056                            ceph_mdsmap_get_addr(newmap, i),
5057                            sizeof(struct ceph_entity_addr))) {
5058                         /* just close it */
5059                         mutex_unlock(&mdsc->mutex);
5060                         mutex_lock(&s->s_mutex);
5061                         mutex_lock(&mdsc->mutex);
5062                         ceph_con_close(&s->s_con);
5063                         mutex_unlock(&s->s_mutex);
5064                         s->s_state = CEPH_MDS_SESSION_RESTARTING;
5065                 } else if (oldstate == newstate) {
5066                         continue;  /* nothing new with this mds */
5067                 }
5068
5069                 /*
5070                  * send reconnect?
5071                  */
5072                 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
5073                     newstate >= CEPH_MDS_STATE_RECONNECT) {
5074                         mutex_unlock(&mdsc->mutex);
5075                         clear_bit(i, targets);
5076                         send_mds_reconnect(mdsc, s);
5077                         mutex_lock(&mdsc->mutex);
5078                 }
5079
5080                 /*
5081                  * kick request on any mds that has gone active.
5082                  */
5083                 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
5084                     newstate >= CEPH_MDS_STATE_ACTIVE) {
5085                         if (oldstate != CEPH_MDS_STATE_CREATING &&
5086                             oldstate != CEPH_MDS_STATE_STARTING)
5087                                 pr_info_client(cl, "mds%d recovery completed\n",
5088                                                s->s_mds);
5089                         kick_requests(mdsc, i);
5090                         mutex_unlock(&mdsc->mutex);
5091                         mutex_lock(&s->s_mutex);
5092                         mutex_lock(&mdsc->mutex);
5093                         ceph_kick_flushing_caps(mdsc, s);
5094                         mutex_unlock(&s->s_mutex);
5095                         wake_up_session_caps(s, RECONNECT);
5096                 }
5097         }
5098
5099         /*
5100          * Only open and reconnect sessions that don't exist yet.
5101          */
5102         for (i = 0; i < newmap->possible_max_rank; i++) {
5103                 /*
5104                  * In case the import MDS is crashed just after
5105                  * the EImportStart journal is flushed, so when
5106                  * a standby MDS takes over it and is replaying
5107                  * the EImportStart journal the new MDS daemon
5108                  * will wait the client to reconnect it, but the
5109                  * client may never register/open the session yet.
5110                  *
5111                  * Will try to reconnect that MDS daemon if the
5112                  * rank number is in the export targets array and
5113                  * is the up:reconnect state.
5114                  */
5115                 newstate = ceph_mdsmap_get_state(newmap, i);
5116                 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
5117                         continue;
5118
5119                 /*
5120                  * The session maybe registered and opened by some
5121                  * requests which were choosing random MDSes during
5122                  * the mdsc->mutex's unlock/lock gap below in rare
5123                  * case. But the related MDS daemon will just queue
5124                  * that requests and be still waiting for the client's
5125                  * reconnection request in up:reconnect state.
5126                  */
5127                 s = __ceph_lookup_mds_session(mdsc, i);
5128                 if (likely(!s)) {
5129                         s = __open_export_target_session(mdsc, i);
5130                         if (IS_ERR(s)) {
5131                                 err = PTR_ERR(s);
5132                                 pr_err_client(cl,
5133                                               "failed to open export target session, err %d\n",
5134                                               err);
5135                                 continue;
5136                         }
5137                 }
5138                 doutc(cl, "send reconnect to export target mds.%d\n", i);
5139                 mutex_unlock(&mdsc->mutex);
5140                 send_mds_reconnect(mdsc, s);
5141                 ceph_put_mds_session(s);
5142                 mutex_lock(&mdsc->mutex);
5143         }
5144
5145         for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
5146                 s = mdsc->sessions[i];
5147                 if (!s)
5148                         continue;
5149                 if (!ceph_mdsmap_is_laggy(newmap, i))
5150                         continue;
5151                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5152                     s->s_state == CEPH_MDS_SESSION_HUNG ||
5153                     s->s_state == CEPH_MDS_SESSION_CLOSING) {
5154                         doutc(cl, " connecting to export targets of laggy mds%d\n", i);
5155                         __open_export_target_sessions(mdsc, s);
5156                 }
5157         }
5158 }
5159
5160
5161
5162 /*
5163  * leases
5164  */
5165
5166 /*
5167  * caller must hold session s_mutex, dentry->d_lock
5168  */
5169 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
5170 {
5171         struct ceph_dentry_info *di = ceph_dentry(dentry);
5172
5173         ceph_put_mds_session(di->lease_session);
5174         di->lease_session = NULL;
5175 }
5176
5177 static void handle_lease(struct ceph_mds_client *mdsc,
5178                          struct ceph_mds_session *session,
5179                          struct ceph_msg *msg)
5180 {
5181         struct ceph_client *cl = mdsc->fsc->client;
5182         struct super_block *sb = mdsc->fsc->sb;
5183         struct inode *inode;
5184         struct dentry *parent, *dentry;
5185         struct ceph_dentry_info *di;
5186         int mds = session->s_mds;
5187         struct ceph_mds_lease *h = msg->front.iov_base;
5188         u32 seq;
5189         struct ceph_vino vino;
5190         struct qstr dname;
5191         int release = 0;
5192
5193         doutc(cl, "from mds%d\n", mds);
5194
5195         if (!ceph_inc_mds_stopping_blocker(mdsc, session))
5196                 return;
5197
5198         /* decode */
5199         if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
5200                 goto bad;
5201         vino.ino = le64_to_cpu(h->ino);
5202         vino.snap = CEPH_NOSNAP;
5203         seq = le32_to_cpu(h->seq);
5204         dname.len = get_unaligned_le32(h + 1);
5205         if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
5206                 goto bad;
5207         dname.name = (void *)(h + 1) + sizeof(u32);
5208
5209         /* lookup inode */
5210         inode = ceph_find_inode(sb, vino);
5211         doutc(cl, "%s, ino %llx %p %.*s\n", ceph_lease_op_name(h->action),
5212               vino.ino, inode, dname.len, dname.name);
5213
5214         mutex_lock(&session->s_mutex);
5215         if (!inode) {
5216                 doutc(cl, "no inode %llx\n", vino.ino);
5217                 goto release;
5218         }
5219
5220         /* dentry */
5221         parent = d_find_alias(inode);
5222         if (!parent) {
5223                 doutc(cl, "no parent dentry on inode %p\n", inode);
5224                 WARN_ON(1);
5225                 goto release;  /* hrm... */
5226         }
5227         dname.hash = full_name_hash(parent, dname.name, dname.len);
5228         dentry = d_lookup(parent, &dname);
5229         dput(parent);
5230         if (!dentry)
5231                 goto release;
5232
5233         spin_lock(&dentry->d_lock);
5234         di = ceph_dentry(dentry);
5235         switch (h->action) {
5236         case CEPH_MDS_LEASE_REVOKE:
5237                 if (di->lease_session == session) {
5238                         if (ceph_seq_cmp(di->lease_seq, seq) > 0)
5239                                 h->seq = cpu_to_le32(di->lease_seq);
5240                         __ceph_mdsc_drop_dentry_lease(dentry);
5241                 }
5242                 release = 1;
5243                 break;
5244
5245         case CEPH_MDS_LEASE_RENEW:
5246                 if (di->lease_session == session &&
5247                     di->lease_gen == atomic_read(&session->s_cap_gen) &&
5248                     di->lease_renew_from &&
5249                     di->lease_renew_after == 0) {
5250                         unsigned long duration =
5251                                 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
5252
5253                         di->lease_seq = seq;
5254                         di->time = di->lease_renew_from + duration;
5255                         di->lease_renew_after = di->lease_renew_from +
5256                                 (duration >> 1);
5257                         di->lease_renew_from = 0;
5258                 }
5259                 break;
5260         }
5261         spin_unlock(&dentry->d_lock);
5262         dput(dentry);
5263
5264         if (!release)
5265                 goto out;
5266
5267 release:
5268         /* let's just reuse the same message */
5269         h->action = CEPH_MDS_LEASE_REVOKE_ACK;
5270         ceph_msg_get(msg);
5271         ceph_con_send(&session->s_con, msg);
5272
5273 out:
5274         mutex_unlock(&session->s_mutex);
5275         iput(inode);
5276
5277         ceph_dec_mds_stopping_blocker(mdsc);
5278         return;
5279
5280 bad:
5281         ceph_dec_mds_stopping_blocker(mdsc);
5282
5283         pr_err_client(cl, "corrupt lease message\n");
5284         ceph_msg_dump(msg);
5285 }
5286
5287 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
5288                               struct dentry *dentry, char action,
5289                               u32 seq)
5290 {
5291         struct ceph_client *cl = session->s_mdsc->fsc->client;
5292         struct ceph_msg *msg;
5293         struct ceph_mds_lease *lease;
5294         struct inode *dir;
5295         int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
5296
5297         doutc(cl, "identry %p %s to mds%d\n", dentry, ceph_lease_op_name(action),
5298               session->s_mds);
5299
5300         msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
5301         if (!msg)
5302                 return;
5303         lease = msg->front.iov_base;
5304         lease->action = action;
5305         lease->seq = cpu_to_le32(seq);
5306
5307         spin_lock(&dentry->d_lock);
5308         dir = d_inode(dentry->d_parent);
5309         lease->ino = cpu_to_le64(ceph_ino(dir));
5310         lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
5311
5312         put_unaligned_le32(dentry->d_name.len, lease + 1);
5313         memcpy((void *)(lease + 1) + 4,
5314                dentry->d_name.name, dentry->d_name.len);
5315         spin_unlock(&dentry->d_lock);
5316
5317         ceph_con_send(&session->s_con, msg);
5318 }
5319
5320 /*
5321  * lock unlock the session, to wait ongoing session activities
5322  */
5323 static void lock_unlock_session(struct ceph_mds_session *s)
5324 {
5325         mutex_lock(&s->s_mutex);
5326         mutex_unlock(&s->s_mutex);
5327 }
5328
5329 static void maybe_recover_session(struct ceph_mds_client *mdsc)
5330 {
5331         struct ceph_client *cl = mdsc->fsc->client;
5332         struct ceph_fs_client *fsc = mdsc->fsc;
5333
5334         if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
5335                 return;
5336
5337         if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
5338                 return;
5339
5340         if (!READ_ONCE(fsc->blocklisted))
5341                 return;
5342
5343         pr_info_client(cl, "auto reconnect after blocklisted\n");
5344         ceph_force_reconnect(fsc->sb);
5345 }
5346
5347 bool check_session_state(struct ceph_mds_session *s)
5348 {
5349         struct ceph_client *cl = s->s_mdsc->fsc->client;
5350
5351         switch (s->s_state) {
5352         case CEPH_MDS_SESSION_OPEN:
5353                 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
5354                         s->s_state = CEPH_MDS_SESSION_HUNG;
5355                         pr_info_client(cl, "mds%d hung\n", s->s_mds);
5356                 }
5357                 break;
5358         case CEPH_MDS_SESSION_CLOSING:
5359         case CEPH_MDS_SESSION_NEW:
5360         case CEPH_MDS_SESSION_RESTARTING:
5361         case CEPH_MDS_SESSION_CLOSED:
5362         case CEPH_MDS_SESSION_REJECTED:
5363                 return false;
5364         }
5365
5366         return true;
5367 }
5368
5369 /*
5370  * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
5371  * then we need to retransmit that request.
5372  */
5373 void inc_session_sequence(struct ceph_mds_session *s)
5374 {
5375         struct ceph_client *cl = s->s_mdsc->fsc->client;
5376
5377         lockdep_assert_held(&s->s_mutex);
5378
5379         s->s_seq++;
5380
5381         if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
5382                 int ret;
5383
5384                 doutc(cl, "resending session close request for mds%d\n", s->s_mds);
5385                 ret = request_close_session(s);
5386                 if (ret < 0)
5387                         pr_err_client(cl, "unable to close session to mds%d: %d\n",
5388                                       s->s_mds, ret);
5389         }
5390 }
5391
5392 /*
5393  * delayed work -- periodically trim expired leases, renew caps with mds.  If
5394  * the @delay parameter is set to 0 or if it's more than 5 secs, the default
5395  * workqueue delay value of 5 secs will be used.
5396  */
5397 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
5398 {
5399         unsigned long max_delay = HZ * 5;
5400
5401         /* 5 secs default delay */
5402         if (!delay || (delay > max_delay))
5403                 delay = max_delay;
5404         schedule_delayed_work(&mdsc->delayed_work,
5405                               round_jiffies_relative(delay));
5406 }
5407
5408 static void delayed_work(struct work_struct *work)
5409 {
5410         struct ceph_mds_client *mdsc =
5411                 container_of(work, struct ceph_mds_client, delayed_work.work);
5412         unsigned long delay;
5413         int renew_interval;
5414         int renew_caps;
5415         int i;
5416
5417         doutc(mdsc->fsc->client, "mdsc delayed_work\n");
5418
5419         if (mdsc->stopping >= CEPH_MDSC_STOPPING_FLUSHED)
5420                 return;
5421
5422         mutex_lock(&mdsc->mutex);
5423         renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
5424         renew_caps = time_after_eq(jiffies, HZ*renew_interval +
5425                                    mdsc->last_renew_caps);
5426         if (renew_caps)
5427                 mdsc->last_renew_caps = jiffies;
5428
5429         for (i = 0; i < mdsc->max_sessions; i++) {
5430                 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
5431                 if (!s)
5432                         continue;
5433
5434                 if (!check_session_state(s)) {
5435                         ceph_put_mds_session(s);
5436                         continue;
5437                 }
5438                 mutex_unlock(&mdsc->mutex);
5439
5440                 ceph_flush_session_cap_releases(mdsc, s);
5441
5442                 mutex_lock(&s->s_mutex);
5443                 if (renew_caps)
5444                         send_renew_caps(mdsc, s);
5445                 else
5446                         ceph_con_keepalive(&s->s_con);
5447                 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
5448                     s->s_state == CEPH_MDS_SESSION_HUNG)
5449                         ceph_send_cap_releases(mdsc, s);
5450                 mutex_unlock(&s->s_mutex);
5451                 ceph_put_mds_session(s);
5452
5453                 mutex_lock(&mdsc->mutex);
5454         }
5455         mutex_unlock(&mdsc->mutex);
5456
5457         delay = ceph_check_delayed_caps(mdsc);
5458
5459         ceph_queue_cap_reclaim_work(mdsc);
5460
5461         ceph_trim_snapid_map(mdsc);
5462
5463         maybe_recover_session(mdsc);
5464
5465         schedule_delayed(mdsc, delay);
5466 }
5467
5468 int ceph_mdsc_init(struct ceph_fs_client *fsc)
5469
5470 {
5471         struct ceph_mds_client *mdsc;
5472         int err;
5473
5474         mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
5475         if (!mdsc)
5476                 return -ENOMEM;
5477         mdsc->fsc = fsc;
5478         mutex_init(&mdsc->mutex);
5479         mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
5480         if (!mdsc->mdsmap) {
5481                 err = -ENOMEM;
5482                 goto err_mdsc;
5483         }
5484
5485         init_completion(&mdsc->safe_umount_waiters);
5486         spin_lock_init(&mdsc->stopping_lock);
5487         atomic_set(&mdsc->stopping_blockers, 0);
5488         init_completion(&mdsc->stopping_waiter);
5489         init_waitqueue_head(&mdsc->session_close_wq);
5490         INIT_LIST_HEAD(&mdsc->waiting_for_map);
5491         mdsc->quotarealms_inodes = RB_ROOT;
5492         mutex_init(&mdsc->quotarealms_inodes_mutex);
5493         init_rwsem(&mdsc->snap_rwsem);
5494         mdsc->snap_realms = RB_ROOT;
5495         INIT_LIST_HEAD(&mdsc->snap_empty);
5496         spin_lock_init(&mdsc->snap_empty_lock);
5497         mdsc->request_tree = RB_ROOT;
5498         INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
5499         mdsc->last_renew_caps = jiffies;
5500         INIT_LIST_HEAD(&mdsc->cap_delay_list);
5501 #ifdef CONFIG_DEBUG_FS
5502         INIT_LIST_HEAD(&mdsc->cap_wait_list);
5503 #endif
5504         spin_lock_init(&mdsc->cap_delay_lock);
5505         INIT_LIST_HEAD(&mdsc->cap_unlink_delay_list);
5506         INIT_LIST_HEAD(&mdsc->snap_flush_list);
5507         spin_lock_init(&mdsc->snap_flush_lock);
5508         mdsc->last_cap_flush_tid = 1;
5509         INIT_LIST_HEAD(&mdsc->cap_flush_list);
5510         INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
5511         spin_lock_init(&mdsc->cap_dirty_lock);
5512         init_waitqueue_head(&mdsc->cap_flushing_wq);
5513         INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
5514         INIT_WORK(&mdsc->cap_unlink_work, ceph_cap_unlink_work);
5515         err = ceph_metric_init(&mdsc->metric);
5516         if (err)
5517                 goto err_mdsmap;
5518
5519         spin_lock_init(&mdsc->dentry_list_lock);
5520         INIT_LIST_HEAD(&mdsc->dentry_leases);
5521         INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
5522
5523         ceph_caps_init(mdsc);
5524         ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
5525
5526         spin_lock_init(&mdsc->snapid_map_lock);
5527         mdsc->snapid_map_tree = RB_ROOT;
5528         INIT_LIST_HEAD(&mdsc->snapid_map_lru);
5529
5530         init_rwsem(&mdsc->pool_perm_rwsem);
5531         mdsc->pool_perm_tree = RB_ROOT;
5532
5533         strscpy(mdsc->nodename, utsname()->nodename,
5534                 sizeof(mdsc->nodename));
5535
5536         fsc->mdsc = mdsc;
5537         return 0;
5538
5539 err_mdsmap:
5540         kfree(mdsc->mdsmap);
5541 err_mdsc:
5542         kfree(mdsc);
5543         return err;
5544 }
5545
5546 /*
5547  * Wait for safe replies on open mds requests.  If we time out, drop
5548  * all requests from the tree to avoid dangling dentry refs.
5549  */
5550 static void wait_requests(struct ceph_mds_client *mdsc)
5551 {
5552         struct ceph_client *cl = mdsc->fsc->client;
5553         struct ceph_options *opts = mdsc->fsc->client->options;
5554         struct ceph_mds_request *req;
5555
5556         mutex_lock(&mdsc->mutex);
5557         if (__get_oldest_req(mdsc)) {
5558                 mutex_unlock(&mdsc->mutex);
5559
5560                 doutc(cl, "waiting for requests\n");
5561                 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
5562                                     ceph_timeout_jiffies(opts->mount_timeout));
5563
5564                 /* tear down remaining requests */
5565                 mutex_lock(&mdsc->mutex);
5566                 while ((req = __get_oldest_req(mdsc))) {
5567                         doutc(cl, "timed out on tid %llu\n", req->r_tid);
5568                         list_del_init(&req->r_wait);
5569                         __unregister_request(mdsc, req);
5570                 }
5571         }
5572         mutex_unlock(&mdsc->mutex);
5573         doutc(cl, "done\n");
5574 }
5575
5576 void send_flush_mdlog(struct ceph_mds_session *s)
5577 {
5578         struct ceph_client *cl = s->s_mdsc->fsc->client;
5579         struct ceph_msg *msg;
5580
5581         /*
5582          * Pre-luminous MDS crashes when it sees an unknown session request
5583          */
5584         if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
5585                 return;
5586
5587         mutex_lock(&s->s_mutex);
5588         doutc(cl, "request mdlog flush to mds%d (%s)s seq %lld\n",
5589               s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5590         msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
5591                                       s->s_seq);
5592         if (!msg) {
5593                 pr_err_client(cl, "failed to request mdlog flush to mds%d (%s) seq %lld\n",
5594                               s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
5595         } else {
5596                 ceph_con_send(&s->s_con, msg);
5597         }
5598         mutex_unlock(&s->s_mutex);
5599 }
5600
5601 static int ceph_mds_auth_match(struct ceph_mds_client *mdsc,
5602                                struct ceph_mds_cap_auth *auth,
5603                                const struct cred *cred,
5604                                char *tpath)
5605 {
5606         u32 caller_uid = from_kuid(&init_user_ns, cred->fsuid);
5607         u32 caller_gid = from_kgid(&init_user_ns, cred->fsgid);
5608         struct ceph_client *cl = mdsc->fsc->client;
5609         const char *spath = mdsc->fsc->mount_options->server_path;
5610         bool gid_matched = false;
5611         u32 gid, tlen, len;
5612         int i, j;
5613
5614         doutc(cl, "match.uid %lld\n", auth->match.uid);
5615         if (auth->match.uid != MDS_AUTH_UID_ANY) {
5616                 if (auth->match.uid != caller_uid)
5617                         return 0;
5618                 if (auth->match.num_gids) {
5619                         for (i = 0; i < auth->match.num_gids; i++) {
5620                                 if (caller_gid == auth->match.gids[i])
5621                                         gid_matched = true;
5622                         }
5623                         if (!gid_matched && cred->group_info->ngroups) {
5624                                 for (i = 0; i < cred->group_info->ngroups; i++) {
5625                                         gid = from_kgid(&init_user_ns,
5626                                                         cred->group_info->gid[i]);
5627                                         for (j = 0; j < auth->match.num_gids; j++) {
5628                                                 if (gid == auth->match.gids[j]) {
5629                                                         gid_matched = true;
5630                                                         break;
5631                                                 }
5632                                         }
5633                                         if (gid_matched)
5634                                                 break;
5635                                 }
5636                         }
5637                         if (!gid_matched)
5638                                 return 0;
5639                 }
5640         }
5641
5642         /* path match */
5643         if (auth->match.path) {
5644                 if (!tpath)
5645                         return 0;
5646
5647                 tlen = strlen(tpath);
5648                 len = strlen(auth->match.path);
5649                 if (len) {
5650                         char *_tpath = tpath;
5651                         bool free_tpath = false;
5652                         int m, n;
5653
5654                         doutc(cl, "server path %s, tpath %s, match.path %s\n",
5655                               spath, tpath, auth->match.path);
5656                         if (spath && (m = strlen(spath)) != 1) {
5657                                 /* mount path + '/' + tpath + an extra space */
5658                                 n = m + 1 + tlen + 1;
5659                                 _tpath = kmalloc(n, GFP_NOFS);
5660                                 if (!_tpath)
5661                                         return -ENOMEM;
5662                                 /* remove the leading '/' */
5663                                 snprintf(_tpath, n, "%s/%s", spath + 1, tpath);
5664                                 free_tpath = true;
5665                                 tlen = strlen(_tpath);
5666                         }
5667
5668                         /*
5669                          * Please note the tailing '/' for match.path has already
5670                          * been removed when parsing.
5671                          *
5672                          * Remove the tailing '/' for the target path.
5673                          */
5674                         while (tlen && _tpath[tlen - 1] == '/') {
5675                                 _tpath[tlen - 1] = '\0';
5676                                 tlen -= 1;
5677                         }
5678                         doutc(cl, "_tpath %s\n", _tpath);
5679
5680                         /*
5681                          * In case first == _tpath && tlen == len:
5682                          *  match.path=/foo  --> /foo _path=/foo     --> match
5683                          *  match.path=/foo/ --> /foo _path=/foo     --> match
5684                          *
5685                          * In case first == _tmatch.path && tlen > len:
5686                          *  match.path=/foo/ --> /foo _path=/foo/    --> match
5687                          *  match.path=/foo  --> /foo _path=/foo/    --> match
5688                          *  match.path=/foo/ --> /foo _path=/foo/d   --> match
5689                          *  match.path=/foo  --> /foo _path=/food    --> mismatch
5690                          *
5691                          * All the other cases                       --> mismatch
5692                          */
5693                         char *first = strstr(_tpath, auth->match.path);
5694                         if (first != _tpath) {
5695                                 if (free_tpath)
5696                                         kfree(_tpath);
5697                                 return 0;
5698                         }
5699
5700                         if (tlen > len && _tpath[len] != '/') {
5701                                 if (free_tpath)
5702                                         kfree(_tpath);
5703                                 return 0;
5704                         }
5705                 }
5706         }
5707
5708         doutc(cl, "matched\n");
5709         return 1;
5710 }
5711
5712 int ceph_mds_check_access(struct ceph_mds_client *mdsc, char *tpath, int mask)
5713 {
5714         const struct cred *cred = get_current_cred();
5715         u32 caller_uid = from_kuid(&init_user_ns, cred->fsuid);
5716         u32 caller_gid = from_kgid(&init_user_ns, cred->fsgid);
5717         struct ceph_mds_cap_auth *rw_perms_s = NULL;
5718         struct ceph_client *cl = mdsc->fsc->client;
5719         bool root_squash_perms = true;
5720         int i, err;
5721
5722         doutc(cl, "tpath '%s', mask %d, caller_uid %d, caller_gid %d\n",
5723               tpath, mask, caller_uid, caller_gid);
5724
5725         for (i = 0; i < mdsc->s_cap_auths_num; i++) {
5726                 struct ceph_mds_cap_auth *s = &mdsc->s_cap_auths[i];
5727
5728                 err = ceph_mds_auth_match(mdsc, s, cred, tpath);
5729                 if (err < 0) {
5730                         put_cred(cred);
5731                         return err;
5732                 } else if (err > 0) {
5733                         /* always follow the last auth caps' permission */
5734                         root_squash_perms = true;
5735                         rw_perms_s = NULL;
5736                         if ((mask & MAY_WRITE) && s->writeable &&
5737                             s->match.root_squash && (!caller_uid || !caller_gid))
5738                                 root_squash_perms = false;
5739
5740                         if (((mask & MAY_WRITE) && !s->writeable) ||
5741                             ((mask & MAY_READ) && !s->readable))
5742                                 rw_perms_s = s;
5743                 }
5744         }
5745
5746         put_cred(cred);
5747
5748         doutc(cl, "root_squash_perms %d, rw_perms_s %p\n", root_squash_perms,
5749               rw_perms_s);
5750         if (root_squash_perms && rw_perms_s == NULL) {
5751                 doutc(cl, "access allowed\n");
5752                 return 0;
5753         }
5754
5755         if (!root_squash_perms) {
5756                 doutc(cl, "root_squash is enabled and user(%d %d) isn't allowed to write",
5757                       caller_uid, caller_gid);
5758         }
5759         if (rw_perms_s) {
5760                 doutc(cl, "mds auth caps readable/writeable %d/%d while request r/w %d/%d",
5761                       rw_perms_s->readable, rw_perms_s->writeable,
5762                       !!(mask & MAY_READ), !!(mask & MAY_WRITE));
5763         }
5764         doutc(cl, "access denied\n");
5765         return -EACCES;
5766 }
5767
5768 /*
5769  * called before mount is ro, and before dentries are torn down.
5770  * (hmm, does this still race with new lookups?)
5771  */
5772 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
5773 {
5774         doutc(mdsc->fsc->client, "begin\n");
5775         mdsc->stopping = CEPH_MDSC_STOPPING_BEGIN;
5776
5777         ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
5778         ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
5779         ceph_flush_dirty_caps(mdsc);
5780         wait_requests(mdsc);
5781
5782         /*
5783          * wait for reply handlers to drop their request refs and
5784          * their inode/dcache refs
5785          */
5786         ceph_msgr_flush();
5787
5788         ceph_cleanup_quotarealms_inodes(mdsc);
5789         doutc(mdsc->fsc->client, "done\n");
5790 }
5791
5792 /*
5793  * flush the mdlog and wait for all write mds requests to flush.
5794  */
5795 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
5796                                                  u64 want_tid)
5797 {
5798         struct ceph_client *cl = mdsc->fsc->client;
5799         struct ceph_mds_request *req = NULL, *nextreq;
5800         struct ceph_mds_session *last_session = NULL;
5801         struct rb_node *n;
5802
5803         mutex_lock(&mdsc->mutex);
5804         doutc(cl, "want %lld\n", want_tid);
5805 restart:
5806         req = __get_oldest_req(mdsc);
5807         while (req && req->r_tid <= want_tid) {
5808                 /* find next request */
5809                 n = rb_next(&req->r_node);
5810                 if (n)
5811                         nextreq = rb_entry(n, struct ceph_mds_request, r_node);
5812                 else
5813                         nextreq = NULL;
5814                 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
5815                     (req->r_op & CEPH_MDS_OP_WRITE)) {
5816                         struct ceph_mds_session *s = req->r_session;
5817
5818                         if (!s) {
5819                                 req = nextreq;
5820                                 continue;
5821                         }
5822
5823                         /* write op */
5824                         ceph_mdsc_get_request(req);
5825                         if (nextreq)
5826                                 ceph_mdsc_get_request(nextreq);
5827                         s = ceph_get_mds_session(s);
5828                         mutex_unlock(&mdsc->mutex);
5829
5830                         /* send flush mdlog request to MDS */
5831                         if (last_session != s) {
5832                                 send_flush_mdlog(s);
5833                                 ceph_put_mds_session(last_session);
5834                                 last_session = s;
5835                         } else {
5836                                 ceph_put_mds_session(s);
5837                         }
5838                         doutc(cl, "wait on %llu (want %llu)\n",
5839                               req->r_tid, want_tid);
5840                         wait_for_completion(&req->r_safe_completion);
5841
5842                         mutex_lock(&mdsc->mutex);
5843                         ceph_mdsc_put_request(req);
5844                         if (!nextreq)
5845                                 break;  /* next dne before, so we're done! */
5846                         if (RB_EMPTY_NODE(&nextreq->r_node)) {
5847                                 /* next request was removed from tree */
5848                                 ceph_mdsc_put_request(nextreq);
5849                                 goto restart;
5850                         }
5851                         ceph_mdsc_put_request(nextreq);  /* won't go away */
5852                 }
5853                 req = nextreq;
5854         }
5855         mutex_unlock(&mdsc->mutex);
5856         ceph_put_mds_session(last_session);
5857         doutc(cl, "done\n");
5858 }
5859
5860 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5861 {
5862         struct ceph_client *cl = mdsc->fsc->client;
5863         u64 want_tid, want_flush;
5864
5865         if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5866                 return;
5867
5868         doutc(cl, "sync\n");
5869         mutex_lock(&mdsc->mutex);
5870         want_tid = mdsc->last_tid;
5871         mutex_unlock(&mdsc->mutex);
5872
5873         ceph_flush_dirty_caps(mdsc);
5874         ceph_flush_cap_releases(mdsc);
5875         spin_lock(&mdsc->cap_dirty_lock);
5876         want_flush = mdsc->last_cap_flush_tid;
5877         if (!list_empty(&mdsc->cap_flush_list)) {
5878                 struct ceph_cap_flush *cf =
5879                         list_last_entry(&mdsc->cap_flush_list,
5880                                         struct ceph_cap_flush, g_list);
5881                 cf->wake = true;
5882         }
5883         spin_unlock(&mdsc->cap_dirty_lock);
5884
5885         doutc(cl, "sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
5886
5887         flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5888         wait_caps_flush(mdsc, want_flush);
5889 }
5890
5891 /*
5892  * true if all sessions are closed, or we force unmount
5893  */
5894 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5895 {
5896         if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5897                 return true;
5898         return atomic_read(&mdsc->num_sessions) <= skipped;
5899 }
5900
5901 /*
5902  * called after sb is ro or when metadata corrupted.
5903  */
5904 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5905 {
5906         struct ceph_options *opts = mdsc->fsc->client->options;
5907         struct ceph_client *cl = mdsc->fsc->client;
5908         struct ceph_mds_session *session;
5909         int i;
5910         int skipped = 0;
5911
5912         doutc(cl, "begin\n");
5913
5914         /* close sessions */
5915         mutex_lock(&mdsc->mutex);
5916         for (i = 0; i < mdsc->max_sessions; i++) {
5917                 session = __ceph_lookup_mds_session(mdsc, i);
5918                 if (!session)
5919                         continue;
5920                 mutex_unlock(&mdsc->mutex);
5921                 mutex_lock(&session->s_mutex);
5922                 if (__close_session(mdsc, session) <= 0)
5923                         skipped++;
5924                 mutex_unlock(&session->s_mutex);
5925                 ceph_put_mds_session(session);
5926                 mutex_lock(&mdsc->mutex);
5927         }
5928         mutex_unlock(&mdsc->mutex);
5929
5930         doutc(cl, "waiting for sessions to close\n");
5931         wait_event_timeout(mdsc->session_close_wq,
5932                            done_closing_sessions(mdsc, skipped),
5933                            ceph_timeout_jiffies(opts->mount_timeout));
5934
5935         /* tear down remaining sessions */
5936         mutex_lock(&mdsc->mutex);
5937         for (i = 0; i < mdsc->max_sessions; i++) {
5938                 if (mdsc->sessions[i]) {
5939                         session = ceph_get_mds_session(mdsc->sessions[i]);
5940                         __unregister_session(mdsc, session);
5941                         mutex_unlock(&mdsc->mutex);
5942                         mutex_lock(&session->s_mutex);
5943                         remove_session_caps(session);
5944                         mutex_unlock(&session->s_mutex);
5945                         ceph_put_mds_session(session);
5946                         mutex_lock(&mdsc->mutex);
5947                 }
5948         }
5949         WARN_ON(!list_empty(&mdsc->cap_delay_list));
5950         mutex_unlock(&mdsc->mutex);
5951
5952         ceph_cleanup_snapid_map(mdsc);
5953         ceph_cleanup_global_and_empty_realms(mdsc);
5954
5955         cancel_work_sync(&mdsc->cap_reclaim_work);
5956         cancel_work_sync(&mdsc->cap_unlink_work);
5957         cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5958
5959         doutc(cl, "done\n");
5960 }
5961
5962 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5963 {
5964         struct ceph_mds_session *session;
5965         int mds;
5966
5967         doutc(mdsc->fsc->client, "force umount\n");
5968
5969         mutex_lock(&mdsc->mutex);
5970         for (mds = 0; mds < mdsc->max_sessions; mds++) {
5971                 session = __ceph_lookup_mds_session(mdsc, mds);
5972                 if (!session)
5973                         continue;
5974
5975                 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5976                         __unregister_session(mdsc, session);
5977                 __wake_requests(mdsc, &session->s_waiting);
5978                 mutex_unlock(&mdsc->mutex);
5979
5980                 mutex_lock(&session->s_mutex);
5981                 __close_session(mdsc, session);
5982                 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5983                         cleanup_session_requests(mdsc, session);
5984                         remove_session_caps(session);
5985                 }
5986                 mutex_unlock(&session->s_mutex);
5987                 ceph_put_mds_session(session);
5988
5989                 mutex_lock(&mdsc->mutex);
5990                 kick_requests(mdsc, mds);
5991         }
5992         __wake_requests(mdsc, &mdsc->waiting_for_map);
5993         mutex_unlock(&mdsc->mutex);
5994 }
5995
5996 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5997 {
5998         doutc(mdsc->fsc->client, "stop\n");
5999         /*
6000          * Make sure the delayed work stopped before releasing
6001          * the resources.
6002          *
6003          * Because the cancel_delayed_work_sync() will only
6004          * guarantee that the work finishes executing. But the
6005          * delayed work will re-arm itself again after that.
6006          */
6007         flush_delayed_work(&mdsc->delayed_work);
6008
6009         if (mdsc->mdsmap)
6010                 ceph_mdsmap_destroy(mdsc->mdsmap);
6011         kfree(mdsc->sessions);
6012         ceph_caps_finalize(mdsc);
6013
6014         if (mdsc->s_cap_auths) {
6015                 int i;
6016
6017                 for (i = 0; i < mdsc->s_cap_auths_num; i++) {
6018                         kfree(mdsc->s_cap_auths[i].match.gids);
6019                         kfree(mdsc->s_cap_auths[i].match.path);
6020                         kfree(mdsc->s_cap_auths[i].match.fs_name);
6021                 }
6022                 kfree(mdsc->s_cap_auths);
6023         }
6024
6025         ceph_pool_perm_destroy(mdsc);
6026 }
6027
6028 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
6029 {
6030         struct ceph_mds_client *mdsc = fsc->mdsc;
6031         doutc(fsc->client, "%p\n", mdsc);
6032
6033         if (!mdsc)
6034                 return;
6035
6036         /* flush out any connection work with references to us */
6037         ceph_msgr_flush();
6038
6039         ceph_mdsc_stop(mdsc);
6040
6041         ceph_metric_destroy(&mdsc->metric);
6042
6043         fsc->mdsc = NULL;
6044         kfree(mdsc);
6045         doutc(fsc->client, "%p done\n", mdsc);
6046 }
6047
6048 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
6049 {
6050         struct ceph_fs_client *fsc = mdsc->fsc;
6051         struct ceph_client *cl = fsc->client;
6052         const char *mds_namespace = fsc->mount_options->mds_namespace;
6053         void *p = msg->front.iov_base;
6054         void *end = p + msg->front.iov_len;
6055         u32 epoch;
6056         u32 num_fs;
6057         u32 mount_fscid = (u32)-1;
6058         int err = -EINVAL;
6059
6060         ceph_decode_need(&p, end, sizeof(u32), bad);
6061         epoch = ceph_decode_32(&p);
6062
6063         doutc(cl, "epoch %u\n", epoch);
6064
6065         /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
6066         ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
6067
6068         ceph_decode_32_safe(&p, end, num_fs, bad);
6069         while (num_fs-- > 0) {
6070                 void *info_p, *info_end;
6071                 u32 info_len;
6072                 u32 fscid, namelen;
6073
6074                 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
6075                 p += 2;         // info_v, info_cv
6076                 info_len = ceph_decode_32(&p);
6077                 ceph_decode_need(&p, end, info_len, bad);
6078                 info_p = p;
6079                 info_end = p + info_len;
6080                 p = info_end;
6081
6082                 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
6083                 fscid = ceph_decode_32(&info_p);
6084                 namelen = ceph_decode_32(&info_p);
6085                 ceph_decode_need(&info_p, info_end, namelen, bad);
6086
6087                 if (mds_namespace &&
6088                     strlen(mds_namespace) == namelen &&
6089                     !strncmp(mds_namespace, (char *)info_p, namelen)) {
6090                         mount_fscid = fscid;
6091                         break;
6092                 }
6093         }
6094
6095         ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
6096         if (mount_fscid != (u32)-1) {
6097                 fsc->client->monc.fs_cluster_id = mount_fscid;
6098                 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
6099                                    0, true);
6100                 ceph_monc_renew_subs(&fsc->client->monc);
6101         } else {
6102                 err = -ENOENT;
6103                 goto err_out;
6104         }
6105         return;
6106
6107 bad:
6108         pr_err_client(cl, "error decoding fsmap %d. Shutting down mount.\n",
6109                       err);
6110         ceph_umount_begin(mdsc->fsc->sb);
6111         ceph_msg_dump(msg);
6112 err_out:
6113         mutex_lock(&mdsc->mutex);
6114         mdsc->mdsmap_err = err;
6115         __wake_requests(mdsc, &mdsc->waiting_for_map);
6116         mutex_unlock(&mdsc->mutex);
6117 }
6118
6119 /*
6120  * handle mds map update.
6121  */
6122 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
6123 {
6124         struct ceph_client *cl = mdsc->fsc->client;
6125         u32 epoch;
6126         u32 maplen;
6127         void *p = msg->front.iov_base;
6128         void *end = p + msg->front.iov_len;
6129         struct ceph_mdsmap *newmap, *oldmap;
6130         struct ceph_fsid fsid;
6131         int err = -EINVAL;
6132
6133         ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
6134         ceph_decode_copy(&p, &fsid, sizeof(fsid));
6135         if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
6136                 return;
6137         epoch = ceph_decode_32(&p);
6138         maplen = ceph_decode_32(&p);
6139         doutc(cl, "epoch %u len %d\n", epoch, (int)maplen);
6140
6141         /* do we need it? */
6142         mutex_lock(&mdsc->mutex);
6143         if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
6144                 doutc(cl, "epoch %u <= our %u\n", epoch, mdsc->mdsmap->m_epoch);
6145                 mutex_unlock(&mdsc->mutex);
6146                 return;
6147         }
6148
6149         newmap = ceph_mdsmap_decode(mdsc, &p, end, ceph_msgr2(mdsc->fsc->client));
6150         if (IS_ERR(newmap)) {
6151                 err = PTR_ERR(newmap);
6152                 goto bad_unlock;
6153         }
6154
6155         /* swap into place */
6156         if (mdsc->mdsmap) {
6157                 oldmap = mdsc->mdsmap;
6158                 mdsc->mdsmap = newmap;
6159                 check_new_map(mdsc, newmap, oldmap);
6160                 ceph_mdsmap_destroy(oldmap);
6161         } else {
6162                 mdsc->mdsmap = newmap;  /* first mds map */
6163         }
6164         mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
6165                                         MAX_LFS_FILESIZE);
6166
6167         __wake_requests(mdsc, &mdsc->waiting_for_map);
6168         ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
6169                           mdsc->mdsmap->m_epoch);
6170
6171         mutex_unlock(&mdsc->mutex);
6172         schedule_delayed(mdsc, 0);
6173         return;
6174
6175 bad_unlock:
6176         mutex_unlock(&mdsc->mutex);
6177 bad:
6178         pr_err_client(cl, "error decoding mdsmap %d. Shutting down mount.\n",
6179                       err);
6180         ceph_umount_begin(mdsc->fsc->sb);
6181         ceph_msg_dump(msg);
6182         return;
6183 }
6184
6185 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
6186 {
6187         struct ceph_mds_session *s = con->private;
6188
6189         if (ceph_get_mds_session(s))
6190                 return con;
6191         return NULL;
6192 }
6193
6194 static void mds_put_con(struct ceph_connection *con)
6195 {
6196         struct ceph_mds_session *s = con->private;
6197
6198         ceph_put_mds_session(s);
6199 }
6200
6201 /*
6202  * if the client is unresponsive for long enough, the mds will kill
6203  * the session entirely.
6204  */
6205 static void mds_peer_reset(struct ceph_connection *con)
6206 {
6207         struct ceph_mds_session *s = con->private;
6208         struct ceph_mds_client *mdsc = s->s_mdsc;
6209
6210         pr_warn_client(mdsc->fsc->client, "mds%d closed our session\n",
6211                        s->s_mds);
6212         if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO &&
6213             ceph_mdsmap_get_state(mdsc->mdsmap, s->s_mds) >= CEPH_MDS_STATE_RECONNECT)
6214                 send_mds_reconnect(mdsc, s);
6215 }
6216
6217 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
6218 {
6219         struct ceph_mds_session *s = con->private;
6220         struct ceph_mds_client *mdsc = s->s_mdsc;
6221         struct ceph_client *cl = mdsc->fsc->client;
6222         int type = le16_to_cpu(msg->hdr.type);
6223
6224         mutex_lock(&mdsc->mutex);
6225         if (__verify_registered_session(mdsc, s) < 0) {
6226                 mutex_unlock(&mdsc->mutex);
6227                 goto out;
6228         }
6229         mutex_unlock(&mdsc->mutex);
6230
6231         switch (type) {
6232         case CEPH_MSG_MDS_MAP:
6233                 ceph_mdsc_handle_mdsmap(mdsc, msg);
6234                 break;
6235         case CEPH_MSG_FS_MAP_USER:
6236                 ceph_mdsc_handle_fsmap(mdsc, msg);
6237                 break;
6238         case CEPH_MSG_CLIENT_SESSION:
6239                 handle_session(s, msg);
6240                 break;
6241         case CEPH_MSG_CLIENT_REPLY:
6242                 handle_reply(s, msg);
6243                 break;
6244         case CEPH_MSG_CLIENT_REQUEST_FORWARD:
6245                 handle_forward(mdsc, s, msg);
6246                 break;
6247         case CEPH_MSG_CLIENT_CAPS:
6248                 ceph_handle_caps(s, msg);
6249                 break;
6250         case CEPH_MSG_CLIENT_SNAP:
6251                 ceph_handle_snap(mdsc, s, msg);
6252                 break;
6253         case CEPH_MSG_CLIENT_LEASE:
6254                 handle_lease(mdsc, s, msg);
6255                 break;
6256         case CEPH_MSG_CLIENT_QUOTA:
6257                 ceph_handle_quota(mdsc, s, msg);
6258                 break;
6259
6260         default:
6261                 pr_err_client(cl, "received unknown message type %d %s\n",
6262                               type, ceph_msg_type_name(type));
6263         }
6264 out:
6265         ceph_msg_put(msg);
6266 }
6267
6268 /*
6269  * authentication
6270  */
6271
6272 /*
6273  * Note: returned pointer is the address of a structure that's
6274  * managed separately.  Caller must *not* attempt to free it.
6275  */
6276 static struct ceph_auth_handshake *
6277 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
6278 {
6279         struct ceph_mds_session *s = con->private;
6280         struct ceph_mds_client *mdsc = s->s_mdsc;
6281         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6282         struct ceph_auth_handshake *auth = &s->s_auth;
6283         int ret;
6284
6285         ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6286                                          force_new, proto, NULL, NULL);
6287         if (ret)
6288                 return ERR_PTR(ret);
6289
6290         return auth;
6291 }
6292
6293 static int mds_add_authorizer_challenge(struct ceph_connection *con,
6294                                     void *challenge_buf, int challenge_buf_len)
6295 {
6296         struct ceph_mds_session *s = con->private;
6297         struct ceph_mds_client *mdsc = s->s_mdsc;
6298         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6299
6300         return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
6301                                             challenge_buf, challenge_buf_len);
6302 }
6303
6304 static int mds_verify_authorizer_reply(struct ceph_connection *con)
6305 {
6306         struct ceph_mds_session *s = con->private;
6307         struct ceph_mds_client *mdsc = s->s_mdsc;
6308         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6309         struct ceph_auth_handshake *auth = &s->s_auth;
6310
6311         return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
6312                 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
6313                 NULL, NULL, NULL, NULL);
6314 }
6315
6316 static int mds_invalidate_authorizer(struct ceph_connection *con)
6317 {
6318         struct ceph_mds_session *s = con->private;
6319         struct ceph_mds_client *mdsc = s->s_mdsc;
6320         struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
6321
6322         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
6323
6324         return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
6325 }
6326
6327 static int mds_get_auth_request(struct ceph_connection *con,
6328                                 void *buf, int *buf_len,
6329                                 void **authorizer, int *authorizer_len)
6330 {
6331         struct ceph_mds_session *s = con->private;
6332         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6333         struct ceph_auth_handshake *auth = &s->s_auth;
6334         int ret;
6335
6336         ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
6337                                        buf, buf_len);
6338         if (ret)
6339                 return ret;
6340
6341         *authorizer = auth->authorizer_buf;
6342         *authorizer_len = auth->authorizer_buf_len;
6343         return 0;
6344 }
6345
6346 static int mds_handle_auth_reply_more(struct ceph_connection *con,
6347                                       void *reply, int reply_len,
6348                                       void *buf, int *buf_len,
6349                                       void **authorizer, int *authorizer_len)
6350 {
6351         struct ceph_mds_session *s = con->private;
6352         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6353         struct ceph_auth_handshake *auth = &s->s_auth;
6354         int ret;
6355
6356         ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
6357                                               buf, buf_len);
6358         if (ret)
6359                 return ret;
6360
6361         *authorizer = auth->authorizer_buf;
6362         *authorizer_len = auth->authorizer_buf_len;
6363         return 0;
6364 }
6365
6366 static int mds_handle_auth_done(struct ceph_connection *con,
6367                                 u64 global_id, void *reply, int reply_len,
6368                                 u8 *session_key, int *session_key_len,
6369                                 u8 *con_secret, int *con_secret_len)
6370 {
6371         struct ceph_mds_session *s = con->private;
6372         struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
6373         struct ceph_auth_handshake *auth = &s->s_auth;
6374
6375         return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
6376                                                session_key, session_key_len,
6377                                                con_secret, con_secret_len);
6378 }
6379
6380 static int mds_handle_auth_bad_method(struct ceph_connection *con,
6381                                       int used_proto, int result,
6382                                       const int *allowed_protos, int proto_cnt,
6383                                       const int *allowed_modes, int mode_cnt)
6384 {
6385         struct ceph_mds_session *s = con->private;
6386         struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
6387         int ret;
6388
6389         if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
6390                                             used_proto, result,
6391                                             allowed_protos, proto_cnt,
6392                                             allowed_modes, mode_cnt)) {
6393                 ret = ceph_monc_validate_auth(monc);
6394                 if (ret)
6395                         return ret;
6396         }
6397
6398         return -EACCES;
6399 }
6400
6401 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
6402                                 struct ceph_msg_header *hdr, int *skip)
6403 {
6404         struct ceph_msg *msg;
6405         int type = (int) le16_to_cpu(hdr->type);
6406         int front_len = (int) le32_to_cpu(hdr->front_len);
6407
6408         if (con->in_msg)
6409                 return con->in_msg;
6410
6411         *skip = 0;
6412         msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
6413         if (!msg) {
6414                 pr_err("unable to allocate msg type %d len %d\n",
6415                        type, front_len);
6416                 return NULL;
6417         }
6418
6419         return msg;
6420 }
6421
6422 static int mds_sign_message(struct ceph_msg *msg)
6423 {
6424        struct ceph_mds_session *s = msg->con->private;
6425        struct ceph_auth_handshake *auth = &s->s_auth;
6426
6427        return ceph_auth_sign_message(auth, msg);
6428 }
6429
6430 static int mds_check_message_signature(struct ceph_msg *msg)
6431 {
6432        struct ceph_mds_session *s = msg->con->private;
6433        struct ceph_auth_handshake *auth = &s->s_auth;
6434
6435        return ceph_auth_check_message_signature(auth, msg);
6436 }
6437
6438 static const struct ceph_connection_operations mds_con_ops = {
6439         .get = mds_get_con,
6440         .put = mds_put_con,
6441         .alloc_msg = mds_alloc_msg,
6442         .dispatch = mds_dispatch,
6443         .peer_reset = mds_peer_reset,
6444         .get_authorizer = mds_get_authorizer,
6445         .add_authorizer_challenge = mds_add_authorizer_challenge,
6446         .verify_authorizer_reply = mds_verify_authorizer_reply,
6447         .invalidate_authorizer = mds_invalidate_authorizer,
6448         .sign_message = mds_sign_message,
6449         .check_message_signature = mds_check_message_signature,
6450         .get_auth_request = mds_get_auth_request,
6451         .handle_auth_reply_more = mds_handle_auth_reply_more,
6452         .handle_auth_done = mds_handle_auth_done,
6453         .handle_auth_bad_method = mds_handle_auth_bad_method,
6454 };
6455
6456 /* eof */