fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "io.h"
  20 #include "metric.h"
  21
  22 static __le32 ceph_flags_sys2wire(u32 flags)
  23 {
  24         u32 wire_flags = 0;
  25
  26         switch (flags & O_ACCMODE) {
  27         case O_RDONLY:
  28                 wire_flags |= CEPH_O_RDONLY;
  29                 break;
  30         case O_WRONLY:
  31                 wire_flags |= CEPH_O_WRONLY;
  32                 break;
  33         case O_RDWR:
  34                 wire_flags |= CEPH_O_RDWR;
  35                 break;
  36         }
  37
  38         flags &= ~O_ACCMODE;
  39
  40 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  41
  42         ceph_sys2wire(O_CREAT);
  43         ceph_sys2wire(O_EXCL);
  44         ceph_sys2wire(O_TRUNC);
  45         ceph_sys2wire(O_DIRECTORY);
  46         ceph_sys2wire(O_NOFOLLOW);
  47
  48 #undef ceph_sys2wire
  49
  50         if (flags)
  51                 dout("unused open flags: %x\n", flags);
  52
  53         return cpu_to_le32(wire_flags);
  54 }
  55
  56 /*
  57  * Ceph file operations
  58  *
  59  * Implement basic open/close functionality, and implement
  60  * read/write.
  61  *
  62  * We implement three modes of file I/O:
  63  *  - buffered uses the generic_file_aio_{read,write} helpers
  64  *
  65  *  - synchronous is used when there is multi-client read/write
  66  *    sharing, avoids the page cache, and synchronously waits for an
  67  *    ack from the OSD.
  68  *
  69  *  - direct io takes the variant of the sync path that references
  70  *    user pages directly.
  71  *
  72  * fsync() flushes and waits on dirty pages, but just queues metadata
  73  * for writeback: since the MDS can recover size and mtime there is no
  74  * need to wait for MDS acknowledgement.
  75  */
  76
  77 /*
  78  * How many pages to get in one call to iov_iter_get_pages().  This
  79  * determines the size of the on-stack array used as a buffer.
  80  */
  81 #define ITER_GET_BVECS_PAGES    64
  82
  83 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  84                                 struct bio_vec *bvecs)
  85 {
  86         size_t size = 0;
  87         int bvec_idx = 0;
  88
  89         if (maxsize > iov_iter_count(iter))
  90                 maxsize = iov_iter_count(iter);
  91
  92         while (size < maxsize) {
  93                 struct page *pages[ITER_GET_BVECS_PAGES];
  94                 ssize_t bytes;
  95                 size_t start;
  96                 int idx = 0;
  97
  98                 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
  99                                            ITER_GET_BVECS_PAGES, &start);
 100                 if (bytes < 0)
 101                         return size ?: bytes;
 102
 103                 iov_iter_advance(iter, bytes);
 104                 size += bytes;
 105
 106                 for ( ; bytes; idx++, bvec_idx++) {
 107                         struct bio_vec bv = {
 108                                 .bv_page = pages[idx],
 109                                 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
 110                                 .bv_offset = start,
 111                         };
 112
 113                         bvecs[bvec_idx] = bv;
 114                         bytes -= bv.bv_len;
 115                         start = 0;
 116                 }
 117         }
 118
 119         return size;
 120 }
 121
 122 /*
 123  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 124  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 125  * page.
 126  *
 127  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 128  * Return the number of bytes in the created bio_vec array, or an error.
 129  */
 130 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 131                                     struct bio_vec **bvecs, int *num_bvecs)
 132 {
 133         struct bio_vec *bv;
 134         size_t orig_count = iov_iter_count(iter);
 135         ssize_t bytes;
 136         int npages;
 137
 138         iov_iter_truncate(iter, maxsize);
 139         npages = iov_iter_npages(iter, INT_MAX);
 140         iov_iter_reexpand(iter, orig_count);
 141
 142         /*
 143          * __iter_get_bvecs() may populate only part of the array -- zero it
 144          * out.
 145          */
 146         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 147         if (!bv)
 148                 return -ENOMEM;
 149
 150         bytes = __iter_get_bvecs(iter, maxsize, bv);
 151         if (bytes < 0) {
 152                 /*
 153                  * No pages were pinned -- just free the array.
 154                  */
 155                 kvfree(bv);
 156                 return bytes;
 157         }
 158
 159         *bvecs = bv;
 160         *num_bvecs = npages;
 161         return bytes;
 162 }
 163
 164 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 165 {
 166         int i;
 167
 168         for (i = 0; i < num_bvecs; i++) {
 169                 if (bvecs[i].bv_page) {
 170                         if (should_dirty)
 171                                 set_page_dirty_lock(bvecs[i].bv_page);
 172                         put_page(bvecs[i].bv_page);
 173                 }
 174         }
 175         kvfree(bvecs);
 176 }
 177
 178 /*
 179  * Prepare an open request.  Preallocate ceph_cap to avoid an
 180  * inopportune ENOMEM later.
 181  */
 182 static struct ceph_mds_request *
 183 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 184 {
 185         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 186         struct ceph_mds_request *req;
 187         int want_auth = USE_ANY_MDS;
 188         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 189
 190         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 191                 want_auth = USE_AUTH_MDS;
 192
 193         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 194         if (IS_ERR(req))
 195                 goto out;
 196         req->r_fmode = ceph_flags_to_mode(flags);
 197         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 198         req->r_args.open.mode = cpu_to_le32(create_mode);
 199 out:
 200         return req;
 201 }
 202
 203 static int ceph_init_file_info(struct inode *inode, struct file *file,
 204                                         int fmode, bool isdir)
 205 {
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         struct ceph_file_info *fi;
 208
 209         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 210                         inode->i_mode, isdir ? "dir" : "regular");
 211         BUG_ON(inode->i_fop->release != ceph_release);
 212
 213         if (isdir) {
 214                 struct ceph_dir_file_info *dfi =
 215                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 216                 if (!dfi)
 217                         return -ENOMEM;
 218
 219                 file->private_data = dfi;
 220                 fi = &dfi->file_info;
 221                 dfi->next_offset = 2;
 222                 dfi->readdir_cache_idx = -1;
 223         } else {
 224                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 225                 if (!fi)
 226                         return -ENOMEM;
 227
 228                 file->private_data = fi;
 229         }
 230
 231         ceph_get_fmode(ci, fmode, 1);
 232         fi->fmode = fmode;
 233
 234         spin_lock_init(&fi->rw_contexts_lock);
 235         INIT_LIST_HEAD(&fi->rw_contexts);
 236         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 237
 238         return 0;
 239 }
 240
 241 /*
 242  * initialize private struct file data.
 243  * if we fail, clean up by dropping fmode reference on the ceph_inode
 244  */
 245 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 246 {
 247         int ret = 0;
 248
 249         switch (inode->i_mode & S_IFMT) {
 250         case S_IFREG:
 251                 ceph_fscache_register_inode_cookie(inode);
 252                 ceph_fscache_file_set_cookie(inode, file);
 253                 fallthrough;
 254         case S_IFDIR:
 255                 ret = ceph_init_file_info(inode, file, fmode,
 256                                                 S_ISDIR(inode->i_mode));
 257                 break;
 258
 259         case S_IFLNK:
 260                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 261                      inode->i_mode);
 262                 break;
 263
 264         default:
 265                 dout("init_file %p %p 0%o (special)\n", inode, file,
 266                      inode->i_mode);
 267                 /*
 268                  * we need to drop the open ref now, since we don't
 269                  * have .release set to ceph_release.
 270                  */
 271                 BUG_ON(inode->i_fop->release == ceph_release);
 272
 273                 /* call the proper open fop */
 274                 ret = inode->i_fop->open(inode, file);
 275         }
 276         return ret;
 277 }
 278
 279 /*
 280  * try renew caps after session gets killed.
 281  */
 282 int ceph_renew_caps(struct inode *inode, int fmode)
 283 {
 284         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 285         struct ceph_inode_info *ci = ceph_inode(inode);
 286         struct ceph_mds_request *req;
 287         int err, flags, wanted;
 288
 289         spin_lock(&ci->i_ceph_lock);
 290         __ceph_touch_fmode(ci, mdsc, fmode);
 291         wanted = __ceph_caps_file_wanted(ci);
 292         if (__ceph_is_any_real_caps(ci) &&
 293             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 294                 int issued = __ceph_caps_issued(ci, NULL);
 295                 spin_unlock(&ci->i_ceph_lock);
 296                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 297                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 298                 ceph_check_caps(ci, 0, NULL);
 299                 return 0;
 300         }
 301         spin_unlock(&ci->i_ceph_lock);
 302
 303         flags = 0;
 304         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 305                 flags = O_RDWR;
 306         else if (wanted & CEPH_CAP_FILE_RD)
 307                 flags = O_RDONLY;
 308         else if (wanted & CEPH_CAP_FILE_WR)
 309                 flags = O_WRONLY;
 310 #ifdef O_LAZY
 311         if (wanted & CEPH_CAP_FILE_LAZYIO)
 312                 flags |= O_LAZY;
 313 #endif
 314
 315         req = prepare_open_request(inode->i_sb, flags, 0);
 316         if (IS_ERR(req)) {
 317                 err = PTR_ERR(req);
 318                 goto out;
 319         }
 320
 321         req->r_inode = inode;
 322         ihold(inode);
 323         req->r_num_caps = 1;
 324
 325         err = ceph_mdsc_do_request(mdsc, NULL, req);
 326         ceph_mdsc_put_request(req);
 327 out:
 328         dout("renew caps %p open result=%d\n", inode, err);
 329         return err < 0 ? err : 0;
 330 }
 331
 332 /*
 333  * If we already have the requisite capabilities, we can satisfy
 334  * the open request locally (no need to request new caps from the
 335  * MDS).  We do, however, need to inform the MDS (asynchronously)
 336  * if our wanted caps set expands.
 337  */
 338 int ceph_open(struct inode *inode, struct file *file)
 339 {
 340         struct ceph_inode_info *ci = ceph_inode(inode);
 341         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 342         struct ceph_mds_client *mdsc = fsc->mdsc;
 343         struct ceph_mds_request *req;
 344         struct ceph_file_info *fi = file->private_data;
 345         int err;
 346         int flags, fmode, wanted;
 347
 348         if (fi) {
 349                 dout("open file %p is already opened\n", file);
 350                 return 0;
 351         }
 352
 353         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 354         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 355         if (S_ISDIR(inode->i_mode))
 356                 flags = O_DIRECTORY;  /* mds likes to know */
 357
 358         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 359              ceph_vinop(inode), file, flags, file->f_flags);
 360         fmode = ceph_flags_to_mode(flags);
 361         wanted = ceph_caps_for_mode(fmode);
 362
 363         /* snapped files are read-only */
 364         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 365                 return -EROFS;
 366
 367         /* trivially open snapdir */
 368         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 369                 return ceph_init_file(inode, file, fmode);
 370         }
 371
 372         /*
 373          * No need to block if we have caps on the auth MDS (for
 374          * write) or any MDS (for read).  Update wanted set
 375          * asynchronously.
 376          */
 377         spin_lock(&ci->i_ceph_lock);
 378         if (__ceph_is_any_real_caps(ci) &&
 379             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 380                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 381                 int issued = __ceph_caps_issued(ci, NULL);
 382
 383                 dout("open %p fmode %d want %s issued %s using existing\n",
 384                      inode, fmode, ceph_cap_string(wanted),
 385                      ceph_cap_string(issued));
 386                 __ceph_touch_fmode(ci, mdsc, fmode);
 387                 spin_unlock(&ci->i_ceph_lock);
 388
 389                 /* adjust wanted? */
 390                 if ((issued & wanted) != wanted &&
 391                     (mds_wanted & wanted) != wanted &&
 392                     ceph_snap(inode) != CEPH_SNAPDIR)
 393                         ceph_check_caps(ci, 0, NULL);
 394
 395                 return ceph_init_file(inode, file, fmode);
 396         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 397                    (ci->i_snap_caps & wanted) == wanted) {
 398                 __ceph_touch_fmode(ci, mdsc, fmode);
 399                 spin_unlock(&ci->i_ceph_lock);
 400                 return ceph_init_file(inode, file, fmode);
 401         }
 402
 403         spin_unlock(&ci->i_ceph_lock);
 404
 405         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 406         req = prepare_open_request(inode->i_sb, flags, 0);
 407         if (IS_ERR(req)) {
 408                 err = PTR_ERR(req);
 409                 goto out;
 410         }
 411         req->r_inode = inode;
 412         ihold(inode);
 413
 414         req->r_num_caps = 1;
 415         err = ceph_mdsc_do_request(mdsc, NULL, req);
 416         if (!err)
 417                 err = ceph_init_file(inode, file, req->r_fmode);
 418         ceph_mdsc_put_request(req);
 419         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 420 out:
 421         return err;
 422 }
 423
 424 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 425 static void
 426 cache_file_layout(struct inode *dst, struct inode *src)
 427 {
 428         struct ceph_inode_info *cdst = ceph_inode(dst);
 429         struct ceph_inode_info *csrc = ceph_inode(src);
 430
 431         spin_lock(&cdst->i_ceph_lock);
 432         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 433             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 434                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 435                         sizeof(cdst->i_cached_layout));
 436                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 437                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 438         }
 439         spin_unlock(&cdst->i_ceph_lock);
 440 }
 441
 442 /*
 443  * Try to set up an async create. We need caps, a file layout, and inode number,
 444  * and either a lease on the dentry or complete dir info. If any of those
 445  * criteria are not satisfied, then return false and the caller can go
 446  * synchronous.
 447  */
 448 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 449                                  struct ceph_file_layout *lo, u64 *pino)
 450 {
 451         struct ceph_inode_info *ci = ceph_inode(dir);
 452         struct ceph_dentry_info *di = ceph_dentry(dentry);
 453         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 454         u64 ino;
 455
 456         spin_lock(&ci->i_ceph_lock);
 457         /* No auth cap means no chance for Dc caps */
 458         if (!ci->i_auth_cap)
 459                 goto no_async;
 460
 461         /* Any delegated inos? */
 462         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 463                 goto no_async;
 464
 465         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 466                 goto no_async;
 467
 468         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 469                 goto no_async;
 470
 471         if (d_in_lookup(dentry)) {
 472                 if (!__ceph_dir_is_complete(ci))
 473                         goto no_async;
 474                 spin_lock(&dentry->d_lock);
 475                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 476                 spin_unlock(&dentry->d_lock);
 477         } else if (atomic_read(&ci->i_shared_gen) !=
 478                    READ_ONCE(di->lease_shared_gen)) {
 479                 goto no_async;
 480         }
 481
 482         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 483         if (!ino)
 484                 goto no_async;
 485
 486         *pino = ino;
 487         ceph_take_cap_refs(ci, want, false);
 488         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 489         rcu_assign_pointer(lo->pool_ns,
 490                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 491         got = want;
 492 no_async:
 493         spin_unlock(&ci->i_ceph_lock);
 494         return got;
 495 }
 496
 497 static void restore_deleg_ino(struct inode *dir, u64 ino)
 498 {
 499         struct ceph_inode_info *ci = ceph_inode(dir);
 500         struct ceph_mds_session *s = NULL;
 501
 502         spin_lock(&ci->i_ceph_lock);
 503         if (ci->i_auth_cap)
 504                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 505         spin_unlock(&ci->i_ceph_lock);
 506         if (s) {
 507                 int err = ceph_restore_deleg_ino(s, ino);
 508                 if (err)
 509                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 510                                 ino, err);
 511                 ceph_put_mds_session(s);
 512         }
 513 }
 514
 515 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 516                                  struct ceph_mds_request *req)
 517 {
 518         int result = req->r_err ? req->r_err :
 519                         le32_to_cpu(req->r_reply_info.head->result);
 520
 521         if (result == -EJUKEBOX)
 522                 goto out;
 523
 524         mapping_set_error(req->r_parent->i_mapping, result);
 525
 526         if (result) {
 527                 struct dentry *dentry = req->r_dentry;
 528                 struct inode *inode = d_inode(dentry);
 529                 int pathlen = 0;
 530                 u64 base = 0;
 531                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 532                                                   &base, 0);
 533
 534                 ceph_dir_clear_complete(req->r_parent);
 535                 if (!d_unhashed(dentry))
 536                         d_drop(dentry);
 537
 538                 ceph_inode_shutdown(inode);
 539
 540                 pr_warn("ceph: async create failure path=(%llx)%s result=%d!\n",
 541                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 542                 ceph_mdsc_free_path(path, pathlen);
 543         }
 544
 545         if (req->r_target_inode) {
 546                 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
 547                 u64 ino = ceph_vino(req->r_target_inode).ino;
 548
 549                 if (req->r_deleg_ino != ino)
 550                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 551                                 __func__, req->r_err, req->r_deleg_ino, ino);
 552                 mapping_set_error(req->r_target_inode->i_mapping, result);
 553
 554                 spin_lock(&ci->i_ceph_lock);
 555                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 556                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 557                         wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 558                 }
 559                 ceph_kick_flushing_inode_caps(req->r_session, ci);
 560                 spin_unlock(&ci->i_ceph_lock);
 561         } else if (!result) {
 562                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 563                         req->r_deleg_ino);
 564         }
 565 out:
 566         ceph_mdsc_release_dir_caps(req);
 567 }
 568
 569 static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
 570                                     struct file *file, umode_t mode,
 571                                     struct ceph_mds_request *req,
 572                                     struct ceph_acl_sec_ctx *as_ctx,
 573                                     struct ceph_file_layout *lo)
 574 {
 575         int ret;
 576         char xattr_buf[4];
 577         struct ceph_mds_reply_inode in = { };
 578         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 579         struct ceph_inode_info *ci = ceph_inode(dir);
 580         struct inode *inode;
 581         struct timespec64 now;
 582         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 583         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 584                                   .snap = CEPH_NOSNAP };
 585
 586         ktime_get_real_ts64(&now);
 587
 588         inode = ceph_get_inode(dentry->d_sb, vino);
 589         if (IS_ERR(inode))
 590                 return PTR_ERR(inode);
 591
 592         iinfo.inline_version = CEPH_INLINE_NONE;
 593         iinfo.change_attr = 1;
 594         ceph_encode_timespec64(&iinfo.btime, &now);
 595
 596         iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 597         iinfo.xattr_data = xattr_buf;
 598         memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 599
 600         in.ino = cpu_to_le64(vino.ino);
 601         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 602         in.version = cpu_to_le64(1);    // ???
 603         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 604         in.cap.cap_id = cpu_to_le64(1);
 605         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 606         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 607         in.ctime = in.mtime = in.atime = iinfo.btime;
 608         in.mode = cpu_to_le32((u32)mode);
 609         in.truncate_seq = cpu_to_le32(1);
 610         in.truncate_size = cpu_to_le64(-1ULL);
 611         in.xattr_version = cpu_to_le64(1);
 612         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 613         in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_mode & S_ISGID ?
 614                                 dir->i_gid : current_fsgid()));
 615         in.nlink = cpu_to_le32(1);
 616         in.max_size = cpu_to_le64(lo->stripe_unit);
 617
 618         ceph_file_layout_to_legacy(lo, &in.layout);
 619
 620         down_read(&mdsc->snap_rwsem);
 621         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 622                               req->r_fmode, NULL);
 623         up_read(&mdsc->snap_rwsem);
 624         if (ret) {
 625                 dout("%s failed to fill inode: %d\n", __func__, ret);
 626                 ceph_dir_clear_complete(dir);
 627                 if (!d_unhashed(dentry))
 628                         d_drop(dentry);
 629                 if (inode->i_state & I_NEW)
 630                         discard_new_inode(inode);
 631         } else {
 632                 struct dentry *dn;
 633
 634                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 635                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 636                 ceph_dir_clear_ordered(dir);
 637                 ceph_init_inode_acls(inode, as_ctx);
 638                 if (inode->i_state & I_NEW) {
 639                         /*
 640                          * If it's not I_NEW, then someone created this before
 641                          * we got here. Assume the server is aware of it at
 642                          * that point and don't worry about setting
 643                          * CEPH_I_ASYNC_CREATE.
 644                          */
 645                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 646                         unlock_new_inode(inode);
 647                 }
 648                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 649                         if (!d_unhashed(dentry))
 650                                 d_drop(dentry);
 651                         dn = d_splice_alias(inode, dentry);
 652                         WARN_ON_ONCE(dn && dn != dentry);
 653                 }
 654                 file->f_mode |= FMODE_CREATED;
 655                 ret = finish_open(file, dentry, ceph_open);
 656         }
 657         return ret;
 658 }
 659
 660 /*
 661  * Do a lookup + open with a single request.  If we get a non-existent
 662  * file or symlink, return 1 so the VFS can retry.
 663  */
 664 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 665                      struct file *file, unsigned flags, umode_t mode)
 666 {
 667         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 668         struct ceph_mds_client *mdsc = fsc->mdsc;
 669         struct ceph_mds_request *req;
 670         struct dentry *dn;
 671         struct ceph_acl_sec_ctx as_ctx = {};
 672         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 673         int mask;
 674         int err;
 675
 676         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 677              dir, dentry, dentry,
 678              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 679
 680         if (dentry->d_name.len > NAME_MAX)
 681                 return -ENAMETOOLONG;
 682
 683         if (flags & O_CREAT) {
 684                 if (ceph_quota_is_max_files_exceeded(dir))
 685                         return -EDQUOT;
 686                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 687                 if (err < 0)
 688                         return err;
 689                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 690                 if (err < 0)
 691                         goto out_ctx;
 692         } else if (!d_in_lookup(dentry)) {
 693                 /* If it's not being looked up, it's negative */
 694                 return -ENOENT;
 695         }
 696 retry:
 697         /* do the open */
 698         req = prepare_open_request(dir->i_sb, flags, mode);
 699         if (IS_ERR(req)) {
 700                 err = PTR_ERR(req);
 701                 goto out_ctx;
 702         }
 703         req->r_dentry = dget(dentry);
 704         req->r_num_caps = 2;
 705         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 706         if (ceph_security_xattr_wanted(dir))
 707                 mask |= CEPH_CAP_XATTR_SHARED;
 708         req->r_args.open.mask = cpu_to_le32(mask);
 709         req->r_parent = dir;
 710         ihold(dir);
 711
 712         if (flags & O_CREAT) {
 713                 struct ceph_file_layout lo;
 714
 715                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 716                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 717                 if (as_ctx.pagelist) {
 718                         req->r_pagelist = as_ctx.pagelist;
 719                         as_ctx.pagelist = NULL;
 720                 }
 721                 if (try_async &&
 722                     (req->r_dir_caps =
 723                       try_prep_async_create(dir, dentry, &lo,
 724                                             &req->r_deleg_ino))) {
 725                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 726                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 727                         req->r_callback = ceph_async_create_cb;
 728                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 729                         if (!err) {
 730                                 err = ceph_finish_async_create(dir, dentry,
 731                                                         file, mode, req,
 732                                                         &as_ctx, &lo);
 733                         } else if (err == -EJUKEBOX) {
 734                                 restore_deleg_ino(dir, req->r_deleg_ino);
 735                                 ceph_mdsc_put_request(req);
 736                                 try_async = false;
 737                                 goto retry;
 738                         }
 739                         goto out_req;
 740                 }
 741         }
 742
 743         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 744         err = ceph_mdsc_do_request(mdsc,
 745                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 746                                    req);
 747         if (err == -ENOENT) {
 748                 dentry = ceph_handle_snapdir(req, dentry);
 749                 if (IS_ERR(dentry)) {
 750                         err = PTR_ERR(dentry);
 751                         goto out_req;
 752                 }
 753                 err = 0;
 754         }
 755
 756         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 757                 err = ceph_handle_notrace_create(dir, dentry);
 758
 759         if (d_in_lookup(dentry)) {
 760                 dn = ceph_finish_lookup(req, dentry, err);
 761                 if (IS_ERR(dn))
 762                         err = PTR_ERR(dn);
 763         } else {
 764                 /* we were given a hashed negative dentry */
 765                 dn = NULL;
 766         }
 767         if (err)
 768                 goto out_req;
 769         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 770                 /* make vfs retry on splice, ENOENT, or symlink */
 771                 dout("atomic_open finish_no_open on dn %p\n", dn);
 772                 err = finish_no_open(file, dn);
 773         } else {
 774                 dout("atomic_open finish_open on dn %p\n", dn);
 775                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 776                         struct inode *newino = d_inode(dentry);
 777
 778                         cache_file_layout(dir, newino);
 779                         ceph_init_inode_acls(newino, &as_ctx);
 780                         file->f_mode |= FMODE_CREATED;
 781                 }
 782                 err = finish_open(file, dentry, ceph_open);
 783         }
 784 out_req:
 785         ceph_mdsc_put_request(req);
 786 out_ctx:
 787         ceph_release_acl_sec_ctx(&as_ctx);
 788         dout("atomic_open result=%d\n", err);
 789         return err;
 790 }
 791
 792 int ceph_release(struct inode *inode, struct file *file)
 793 {
 794         struct ceph_inode_info *ci = ceph_inode(inode);
 795
 796         if (S_ISDIR(inode->i_mode)) {
 797                 struct ceph_dir_file_info *dfi = file->private_data;
 798                 dout("release inode %p dir file %p\n", inode, file);
 799                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 800
 801                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 802
 803                 if (dfi->last_readdir)
 804                         ceph_mdsc_put_request(dfi->last_readdir);
 805                 kfree(dfi->last_name);
 806                 kfree(dfi->dir_info);
 807                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 808         } else {
 809                 struct ceph_file_info *fi = file->private_data;
 810                 dout("release inode %p regular file %p\n", inode, file);
 811                 WARN_ON(!list_empty(&fi->rw_contexts));
 812
 813                 ceph_put_fmode(ci, fi->fmode, 1);
 814
 815                 kmem_cache_free(ceph_file_cachep, fi);
 816         }
 817
 818         /* wake up anyone waiting for caps on this inode */
 819         wake_up_all(&ci->i_cap_wq);
 820         return 0;
 821 }
 822
 823 enum {
 824         HAVE_RETRIED = 1,
 825         CHECK_EOF =    2,
 826         READ_INLINE =  3,
 827 };
 828
 829 /*
 830  * Completely synchronous read and write methods.  Direct from __user
 831  * buffer to osd, or directly to user pages (if O_DIRECT).
 832  *
 833  * If the read spans object boundary, just do multiple reads.  (That's not
 834  * atomic, but good enough for now.)
 835  *
 836  * If we get a short result from the OSD, check against i_size; we need to
 837  * only return a short read to the caller if we hit EOF.
 838  */
 839 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 840                               int *retry_op)
 841 {
 842         struct file *file = iocb->ki_filp;
 843         struct inode *inode = file_inode(file);
 844         struct ceph_inode_info *ci = ceph_inode(inode);
 845         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 846         struct ceph_osd_client *osdc = &fsc->client->osdc;
 847         ssize_t ret;
 848         u64 off = iocb->ki_pos;
 849         u64 len = iov_iter_count(to);
 850         u64 i_size;
 851
 852         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 853              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 854
 855         if (!len)
 856                 return 0;
 857         /*
 858          * flush any page cache pages in this range.  this
 859          * will make concurrent normal and sync io slow,
 860          * but it will at least behave sensibly when they are
 861          * in sequence.
 862          */
 863         ret = filemap_write_and_wait_range(inode->i_mapping,
 864                                            off, off + len - 1);
 865         if (ret < 0)
 866                 return ret;
 867
 868         ret = 0;
 869         while ((len = iov_iter_count(to)) > 0) {
 870                 struct ceph_osd_request *req;
 871                 struct page **pages;
 872                 int num_pages;
 873                 size_t page_off;
 874                 bool more;
 875                 int idx;
 876                 size_t left;
 877
 878                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 879                                         ci->i_vino, off, &len, 0, 1,
 880                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 881                                         NULL, ci->i_truncate_seq,
 882                                         ci->i_truncate_size, false);
 883                 if (IS_ERR(req)) {
 884                         ret = PTR_ERR(req);
 885                         break;
 886                 }
 887
 888                 more = len < iov_iter_count(to);
 889
 890                 num_pages = calc_pages_for(off, len);
 891                 page_off = off & ~PAGE_MASK;
 892                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 893                 if (IS_ERR(pages)) {
 894                         ceph_osdc_put_request(req);
 895                         ret = PTR_ERR(pages);
 896                         break;
 897                 }
 898
 899                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 900                                                  false, false);
 901                 ret = ceph_osdc_start_request(osdc, req, false);
 902                 if (!ret)
 903                         ret = ceph_osdc_wait_request(osdc, req);
 904
 905                 ceph_update_read_metrics(&fsc->mdsc->metric,
 906                                          req->r_start_latency,
 907                                          req->r_end_latency,
 908                                          len, ret);
 909
 910                 ceph_osdc_put_request(req);
 911
 912                 i_size = i_size_read(inode);
 913                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 914                      off, len, ret, i_size, (more ? " MORE" : ""));
 915
 916                 if (ret == -ENOENT)
 917                         ret = 0;
 918                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 919                         int zlen = min(len - ret, i_size - off - ret);
 920                         int zoff = page_off + ret;
 921                         dout("sync_read zero gap %llu~%llu\n",
 922                              off + ret, off + ret + zlen);
 923                         ceph_zero_page_vector_range(zoff, zlen, pages);
 924                         ret += zlen;
 925                 }
 926
 927                 idx = 0;
 928                 left = ret > 0 ? ret : 0;
 929                 while (left > 0) {
 930                         size_t len, copied;
 931                         page_off = off & ~PAGE_MASK;
 932                         len = min_t(size_t, left, PAGE_SIZE - page_off);
 933                         SetPageUptodate(pages[idx]);
 934                         copied = copy_page_to_iter(pages[idx++],
 935                                                    page_off, len, to);
 936                         off += copied;
 937                         left -= copied;
 938                         if (copied < len) {
 939                                 ret = -EFAULT;
 940                                 break;
 941                         }
 942                 }
 943                 ceph_release_page_vector(pages, num_pages);
 944
 945                 if (ret < 0) {
 946                         if (ret == -EBLOCKLISTED)
 947                                 fsc->blocklisted = true;
 948                         break;
 949                 }
 950
 951                 if (off >= i_size || !more)
 952                         break;
 953         }
 954
 955         if (off > iocb->ki_pos) {
 956                 if (off >= i_size) {
 957                         *retry_op = CHECK_EOF;
 958                         ret = i_size - iocb->ki_pos;
 959                         iocb->ki_pos = i_size;
 960                 } else {
 961                         ret = off - iocb->ki_pos;
 962                         iocb->ki_pos = off;
 963                 }
 964         }
 965
 966         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 967         return ret;
 968 }
 969
 970 struct ceph_aio_request {
 971         struct kiocb *iocb;
 972         size_t total_len;
 973         bool write;
 974         bool should_dirty;
 975         int error;
 976         struct list_head osd_reqs;
 977         unsigned num_reqs;
 978         atomic_t pending_reqs;
 979         struct timespec64 mtime;
 980         struct ceph_cap_flush *prealloc_cf;
 981 };
 982
 983 struct ceph_aio_work {
 984         struct work_struct work;
 985         struct ceph_osd_request *req;
 986 };
 987
 988 static void ceph_aio_retry_work(struct work_struct *work);
 989
 990 static void ceph_aio_complete(struct inode *inode,
 991                               struct ceph_aio_request *aio_req)
 992 {
 993         struct ceph_inode_info *ci = ceph_inode(inode);
 994         int ret;
 995
 996         if (!atomic_dec_and_test(&aio_req->pending_reqs))
 997                 return;
 998
 999         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1000                 inode_dio_end(inode);
1001
1002         ret = aio_req->error;
1003         if (!ret)
1004                 ret = aio_req->total_len;
1005
1006         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1007
1008         if (ret >= 0 && aio_req->write) {
1009                 int dirty;
1010
1011                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1012                 if (endoff > i_size_read(inode)) {
1013                         if (ceph_inode_set_size(inode, endoff))
1014                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1015                 }
1016
1017                 spin_lock(&ci->i_ceph_lock);
1018                 ci->i_inline_version = CEPH_INLINE_NONE;
1019                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1020                                                &aio_req->prealloc_cf);
1021                 spin_unlock(&ci->i_ceph_lock);
1022                 if (dirty)
1023                         __mark_inode_dirty(inode, dirty);
1024
1025         }
1026
1027         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1028                                                 CEPH_CAP_FILE_RD));
1029
1030         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1031
1032         ceph_free_cap_flush(aio_req->prealloc_cf);
1033         kfree(aio_req);
1034 }
1035
1036 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1037 {
1038         int rc = req->r_result;
1039         struct inode *inode = req->r_inode;
1040         struct ceph_aio_request *aio_req = req->r_priv;
1041         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1042         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1043         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1044
1045         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1046         BUG_ON(!osd_data->num_bvecs);
1047
1048         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1049
1050         if (rc == -EOLDSNAPC) {
1051                 struct ceph_aio_work *aio_work;
1052                 BUG_ON(!aio_req->write);
1053
1054                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1055                 if (aio_work) {
1056                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1057                         aio_work->req = req;
1058                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1059                                    &aio_work->work);
1060                         return;
1061                 }
1062                 rc = -ENOMEM;
1063         } else if (!aio_req->write) {
1064                 if (rc == -ENOENT)
1065                         rc = 0;
1066                 if (rc >= 0 && len > rc) {
1067                         struct iov_iter i;
1068                         int zlen = len - rc;
1069
1070                         /*
1071                          * If read is satisfied by single OSD request,
1072                          * it can pass EOF. Otherwise read is within
1073                          * i_size.
1074                          */
1075                         if (aio_req->num_reqs == 1) {
1076                                 loff_t i_size = i_size_read(inode);
1077                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1078                                 if (endoff < i_size)
1079                                         zlen = min_t(size_t, zlen,
1080                                                      i_size - endoff);
1081                                 aio_req->total_len = rc + zlen;
1082                         }
1083
1084                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1085                                       osd_data->num_bvecs, len);
1086                         iov_iter_advance(&i, rc);
1087                         iov_iter_zero(zlen, &i);
1088                 }
1089         }
1090
1091         /* r_start_latency == 0 means the request was not submitted */
1092         if (req->r_start_latency) {
1093                 if (aio_req->write)
1094                         ceph_update_write_metrics(metric, req->r_start_latency,
1095                                                   req->r_end_latency, len, rc);
1096                 else
1097                         ceph_update_read_metrics(metric, req->r_start_latency,
1098                                                  req->r_end_latency, len, rc);
1099         }
1100
1101         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1102                   aio_req->should_dirty);
1103         ceph_osdc_put_request(req);
1104
1105         if (rc < 0)
1106                 cmpxchg(&aio_req->error, 0, rc);
1107
1108         ceph_aio_complete(inode, aio_req);
1109         return;
1110 }
1111
1112 static void ceph_aio_retry_work(struct work_struct *work)
1113 {
1114         struct ceph_aio_work *aio_work =
1115                 container_of(work, struct ceph_aio_work, work);
1116         struct ceph_osd_request *orig_req = aio_work->req;
1117         struct ceph_aio_request *aio_req = orig_req->r_priv;
1118         struct inode *inode = orig_req->r_inode;
1119         struct ceph_inode_info *ci = ceph_inode(inode);
1120         struct ceph_snap_context *snapc;
1121         struct ceph_osd_request *req;
1122         int ret;
1123
1124         spin_lock(&ci->i_ceph_lock);
1125         if (__ceph_have_pending_cap_snap(ci)) {
1126                 struct ceph_cap_snap *capsnap =
1127                         list_last_entry(&ci->i_cap_snaps,
1128                                         struct ceph_cap_snap,
1129                                         ci_item);
1130                 snapc = ceph_get_snap_context(capsnap->context);
1131         } else {
1132                 BUG_ON(!ci->i_head_snapc);
1133                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1134         }
1135         spin_unlock(&ci->i_ceph_lock);
1136
1137         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1138                         false, GFP_NOFS);
1139         if (!req) {
1140                 ret = -ENOMEM;
1141                 req = orig_req;
1142                 goto out;
1143         }
1144
1145         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1146         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1147         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1148
1149         req->r_ops[0] = orig_req->r_ops[0];
1150
1151         req->r_mtime = aio_req->mtime;
1152         req->r_data_offset = req->r_ops[0].extent.offset;
1153
1154         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1155         if (ret) {
1156                 ceph_osdc_put_request(req);
1157                 req = orig_req;
1158                 goto out;
1159         }
1160
1161         ceph_osdc_put_request(orig_req);
1162
1163         req->r_callback = ceph_aio_complete_req;
1164         req->r_inode = inode;
1165         req->r_priv = aio_req;
1166
1167         ret = ceph_osdc_start_request(req->r_osdc, req, false);
1168 out:
1169         if (ret < 0) {
1170                 req->r_result = ret;
1171                 ceph_aio_complete_req(req);
1172         }
1173
1174         ceph_put_snap_context(snapc);
1175         kfree(aio_work);
1176 }
1177
1178 static ssize_t
1179 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1180                        struct ceph_snap_context *snapc,
1181                        struct ceph_cap_flush **pcf)
1182 {
1183         struct file *file = iocb->ki_filp;
1184         struct inode *inode = file_inode(file);
1185         struct ceph_inode_info *ci = ceph_inode(inode);
1186         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1187         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1188         struct ceph_vino vino;
1189         struct ceph_osd_request *req;
1190         struct bio_vec *bvecs;
1191         struct ceph_aio_request *aio_req = NULL;
1192         int num_pages = 0;
1193         int flags;
1194         int ret = 0;
1195         struct timespec64 mtime = current_time(inode);
1196         size_t count = iov_iter_count(iter);
1197         loff_t pos = iocb->ki_pos;
1198         bool write = iov_iter_rw(iter) == WRITE;
1199         bool should_dirty = !write && iter_is_iovec(iter);
1200
1201         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1202                 return -EROFS;
1203
1204         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1205              (write ? "write" : "read"), file, pos, (unsigned)count,
1206              snapc, snapc ? snapc->seq : 0);
1207
1208         if (write) {
1209                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1210                                         pos >> PAGE_SHIFT,
1211                                         (pos + count - 1) >> PAGE_SHIFT);
1212                 if (ret2 < 0)
1213                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1214
1215                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1216         } else {
1217                 flags = CEPH_OSD_FLAG_READ;
1218         }
1219
1220         while (iov_iter_count(iter) > 0) {
1221                 u64 size = iov_iter_count(iter);
1222                 ssize_t len;
1223
1224                 if (write)
1225                         size = min_t(u64, size, fsc->mount_options->wsize);
1226                 else
1227                         size = min_t(u64, size, fsc->mount_options->rsize);
1228
1229                 vino = ceph_vino(inode);
1230                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1231                                             vino, pos, &size, 0,
1232                                             1,
1233                                             write ? CEPH_OSD_OP_WRITE :
1234                                                     CEPH_OSD_OP_READ,
1235                                             flags, snapc,
1236                                             ci->i_truncate_seq,
1237                                             ci->i_truncate_size,
1238                                             false);
1239                 if (IS_ERR(req)) {
1240                         ret = PTR_ERR(req);
1241                         break;
1242                 }
1243
1244                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1245                 if (len < 0) {
1246                         ceph_osdc_put_request(req);
1247                         ret = len;
1248                         break;
1249                 }
1250                 if (len != size)
1251                         osd_req_op_extent_update(req, 0, len);
1252
1253                 /*
1254                  * To simplify error handling, allow AIO when IO within i_size
1255                  * or IO can be satisfied by single OSD request.
1256                  */
1257                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1258                     (len == count || pos + count <= i_size_read(inode))) {
1259                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1260                         if (aio_req) {
1261                                 aio_req->iocb = iocb;
1262                                 aio_req->write = write;
1263                                 aio_req->should_dirty = should_dirty;
1264                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1265                                 if (write) {
1266                                         aio_req->mtime = mtime;
1267                                         swap(aio_req->prealloc_cf, *pcf);
1268                                 }
1269                         }
1270                         /* ignore error */
1271                 }
1272
1273                 if (write) {
1274                         /*
1275                          * throw out any page cache pages in this range. this
1276                          * may block.
1277                          */
1278                         truncate_inode_pages_range(inode->i_mapping, pos,
1279                                                    PAGE_ALIGN(pos + len) - 1);
1280
1281                         req->r_mtime = mtime;
1282                 }
1283
1284                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1285
1286                 if (aio_req) {
1287                         aio_req->total_len += len;
1288                         aio_req->num_reqs++;
1289                         atomic_inc(&aio_req->pending_reqs);
1290
1291                         req->r_callback = ceph_aio_complete_req;
1292                         req->r_inode = inode;
1293                         req->r_priv = aio_req;
1294                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1295
1296                         pos += len;
1297                         continue;
1298                 }
1299
1300                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1301                 if (!ret)
1302                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1303
1304                 if (write)
1305                         ceph_update_write_metrics(metric, req->r_start_latency,
1306                                                   req->r_end_latency, len, ret);
1307                 else
1308                         ceph_update_read_metrics(metric, req->r_start_latency,
1309                                                  req->r_end_latency, len, ret);
1310
1311                 size = i_size_read(inode);
1312                 if (!write) {
1313                         if (ret == -ENOENT)
1314                                 ret = 0;
1315                         if (ret >= 0 && ret < len && pos + ret < size) {
1316                                 struct iov_iter i;
1317                                 int zlen = min_t(size_t, len - ret,
1318                                                  size - pos - ret);
1319
1320                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1321                                 iov_iter_advance(&i, ret);
1322                                 iov_iter_zero(zlen, &i);
1323                                 ret += zlen;
1324                         }
1325                         if (ret >= 0)
1326                                 len = ret;
1327                 }
1328
1329                 put_bvecs(bvecs, num_pages, should_dirty);
1330                 ceph_osdc_put_request(req);
1331                 if (ret < 0)
1332                         break;
1333
1334                 pos += len;
1335                 if (!write && pos >= size)
1336                         break;
1337
1338                 if (write && pos > size) {
1339                         if (ceph_inode_set_size(inode, pos))
1340                                 ceph_check_caps(ceph_inode(inode),
1341                                                 CHECK_CAPS_AUTHONLY,
1342                                                 NULL);
1343                 }
1344         }
1345
1346         if (aio_req) {
1347                 LIST_HEAD(osd_reqs);
1348
1349                 if (aio_req->num_reqs == 0) {
1350                         kfree(aio_req);
1351                         return ret;
1352                 }
1353
1354                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1355                                               CEPH_CAP_FILE_RD);
1356
1357                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1358                 inode_dio_begin(inode);
1359                 while (!list_empty(&osd_reqs)) {
1360                         req = list_first_entry(&osd_reqs,
1361                                                struct ceph_osd_request,
1362                                                r_private_item);
1363                         list_del_init(&req->r_private_item);
1364                         if (ret >= 0)
1365                                 ret = ceph_osdc_start_request(req->r_osdc,
1366                                                               req, false);
1367                         if (ret < 0) {
1368                                 req->r_result = ret;
1369                                 ceph_aio_complete_req(req);
1370                         }
1371                 }
1372                 return -EIOCBQUEUED;
1373         }
1374
1375         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1376                 ret = pos - iocb->ki_pos;
1377                 iocb->ki_pos = pos;
1378         }
1379         return ret;
1380 }
1381
1382 /*
1383  * Synchronous write, straight from __user pointer or user pages.
1384  *
1385  * If write spans object boundary, just do multiple writes.  (For a
1386  * correct atomic write, we should e.g. take write locks on all
1387  * objects, rollback on failure, etc.)
1388  */
1389 static ssize_t
1390 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1391                 struct ceph_snap_context *snapc)
1392 {
1393         struct file *file = iocb->ki_filp;
1394         struct inode *inode = file_inode(file);
1395         struct ceph_inode_info *ci = ceph_inode(inode);
1396         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1397         struct ceph_vino vino;
1398         struct ceph_osd_request *req;
1399         struct page **pages;
1400         u64 len;
1401         int num_pages;
1402         int written = 0;
1403         int flags;
1404         int ret;
1405         bool check_caps = false;
1406         struct timespec64 mtime = current_time(inode);
1407         size_t count = iov_iter_count(from);
1408
1409         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1410                 return -EROFS;
1411
1412         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1413              file, pos, (unsigned)count, snapc, snapc->seq);
1414
1415         ret = filemap_write_and_wait_range(inode->i_mapping,
1416                                            pos, pos + count - 1);
1417         if (ret < 0)
1418                 return ret;
1419
1420         ret = invalidate_inode_pages2_range(inode->i_mapping,
1421                                             pos >> PAGE_SHIFT,
1422                                             (pos + count - 1) >> PAGE_SHIFT);
1423         if (ret < 0)
1424                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1425
1426         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1427
1428         while ((len = iov_iter_count(from)) > 0) {
1429                 size_t left;
1430                 int n;
1431
1432                 vino = ceph_vino(inode);
1433                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1434                                             vino, pos, &len, 0, 1,
1435                                             CEPH_OSD_OP_WRITE, flags, snapc,
1436                                             ci->i_truncate_seq,
1437                                             ci->i_truncate_size,
1438                                             false);
1439                 if (IS_ERR(req)) {
1440                         ret = PTR_ERR(req);
1441                         break;
1442                 }
1443
1444                 /*
1445                  * write from beginning of first page,
1446                  * regardless of io alignment
1447                  */
1448                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1449
1450                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1451                 if (IS_ERR(pages)) {
1452                         ret = PTR_ERR(pages);
1453                         goto out;
1454                 }
1455
1456                 left = len;
1457                 for (n = 0; n < num_pages; n++) {
1458                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1459                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1460                         if (ret != plen) {
1461                                 ret = -EFAULT;
1462                                 break;
1463                         }
1464                         left -= ret;
1465                 }
1466
1467                 if (ret < 0) {
1468                         ceph_release_page_vector(pages, num_pages);
1469                         goto out;
1470                 }
1471
1472                 req->r_inode = inode;
1473
1474                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1475                                                 false, true);
1476
1477                 req->r_mtime = mtime;
1478                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1479                 if (!ret)
1480                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1481
1482                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1483                                           req->r_end_latency, len, ret);
1484 out:
1485                 ceph_osdc_put_request(req);
1486                 if (ret != 0) {
1487                         ceph_set_error_write(ci);
1488                         break;
1489                 }
1490
1491                 ceph_clear_error_write(ci);
1492                 pos += len;
1493                 written += len;
1494                 if (pos > i_size_read(inode)) {
1495                         check_caps = ceph_inode_set_size(inode, pos);
1496                         if (check_caps)
1497                                 ceph_check_caps(ceph_inode(inode),
1498                                                 CHECK_CAPS_AUTHONLY,
1499                                                 NULL);
1500                 }
1501
1502         }
1503
1504         if (ret != -EOLDSNAPC && written > 0) {
1505                 ret = written;
1506                 iocb->ki_pos = pos;
1507         }
1508         return ret;
1509 }
1510
1511 /*
1512  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1513  * Atomically grab references, so that those bits are not released
1514  * back to the MDS mid-read.
1515  *
1516  * Hmm, the sync read case isn't actually async... should it be?
1517  */
1518 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1519 {
1520         struct file *filp = iocb->ki_filp;
1521         struct ceph_file_info *fi = filp->private_data;
1522         size_t len = iov_iter_count(to);
1523         struct inode *inode = file_inode(filp);
1524         struct ceph_inode_info *ci = ceph_inode(inode);
1525         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1526         ssize_t ret;
1527         int want, got = 0;
1528         int retry_op = 0, read = 0;
1529
1530 again:
1531         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1532              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1533
1534         if (ceph_inode_is_shutdown(inode))
1535                 return -ESTALE;
1536
1537         if (direct_lock)
1538                 ceph_start_io_direct(inode);
1539         else
1540                 ceph_start_io_read(inode);
1541
1542         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1543                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1544         else
1545                 want = CEPH_CAP_FILE_CACHE;
1546         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1547         if (ret < 0) {
1548                 if (iocb->ki_flags & IOCB_DIRECT)
1549                         ceph_end_io_direct(inode);
1550                 else
1551                         ceph_end_io_read(inode);
1552                 return ret;
1553         }
1554
1555         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1556             (iocb->ki_flags & IOCB_DIRECT) ||
1557             (fi->flags & CEPH_F_SYNC)) {
1558
1559                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1560                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1561                      ceph_cap_string(got));
1562
1563                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1564                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1565                                 ret = ceph_direct_read_write(iocb, to,
1566                                                              NULL, NULL);
1567                                 if (ret >= 0 && ret < len)
1568                                         retry_op = CHECK_EOF;
1569                         } else {
1570                                 ret = ceph_sync_read(iocb, to, &retry_op);
1571                         }
1572                 } else {
1573                         retry_op = READ_INLINE;
1574                 }
1575         } else {
1576                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1577                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1578                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1579                      ceph_cap_string(got));
1580                 ceph_add_rw_context(fi, &rw_ctx);
1581                 ret = generic_file_read_iter(iocb, to);
1582                 ceph_del_rw_context(fi, &rw_ctx);
1583         }
1584
1585         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1586              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1587         ceph_put_cap_refs(ci, got);
1588
1589         if (direct_lock)
1590                 ceph_end_io_direct(inode);
1591         else
1592                 ceph_end_io_read(inode);
1593
1594         if (retry_op > HAVE_RETRIED && ret >= 0) {
1595                 int statret;
1596                 struct page *page = NULL;
1597                 loff_t i_size;
1598                 if (retry_op == READ_INLINE) {
1599                         page = __page_cache_alloc(GFP_KERNEL);
1600                         if (!page)
1601                                 return -ENOMEM;
1602                 }
1603
1604                 statret = __ceph_do_getattr(inode, page,
1605                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1606                 if (statret < 0) {
1607                         if (page)
1608                                 __free_page(page);
1609                         if (statret == -ENODATA) {
1610                                 BUG_ON(retry_op != READ_INLINE);
1611                                 goto again;
1612                         }
1613                         return statret;
1614                 }
1615
1616                 i_size = i_size_read(inode);
1617                 if (retry_op == READ_INLINE) {
1618                         BUG_ON(ret > 0 || read > 0);
1619                         if (iocb->ki_pos < i_size &&
1620                             iocb->ki_pos < PAGE_SIZE) {
1621                                 loff_t end = min_t(loff_t, i_size,
1622                                                    iocb->ki_pos + len);
1623                                 end = min_t(loff_t, end, PAGE_SIZE);
1624                                 if (statret < end)
1625                                         zero_user_segment(page, statret, end);
1626                                 ret = copy_page_to_iter(page,
1627                                                 iocb->ki_pos & ~PAGE_MASK,
1628                                                 end - iocb->ki_pos, to);
1629                                 iocb->ki_pos += ret;
1630                                 read += ret;
1631                         }
1632                         if (iocb->ki_pos < i_size && read < len) {
1633                                 size_t zlen = min_t(size_t, len - read,
1634                                                     i_size - iocb->ki_pos);
1635                                 ret = iov_iter_zero(zlen, to);
1636                                 iocb->ki_pos += ret;
1637                                 read += ret;
1638                         }
1639                         __free_pages(page, 0);
1640                         return read;
1641                 }
1642
1643                 /* hit EOF or hole? */
1644                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1645                     ret < len) {
1646                         dout("sync_read hit hole, ppos %lld < size %lld"
1647                              ", reading more\n", iocb->ki_pos, i_size);
1648
1649                         read += ret;
1650                         len -= ret;
1651                         retry_op = HAVE_RETRIED;
1652                         goto again;
1653                 }
1654         }
1655
1656         if (ret >= 0)
1657                 ret += read;
1658
1659         return ret;
1660 }
1661
1662 /*
1663  * Take cap references to avoid releasing caps to MDS mid-write.
1664  *
1665  * If we are synchronous, and write with an old snap context, the OSD
1666  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1667  * dropping our cap refs and allowing the pending snap to logically
1668  * complete _before_ this write occurs.
1669  *
1670  * If we are near ENOSPC, write synchronously.
1671  */
1672 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1673 {
1674         struct file *file = iocb->ki_filp;
1675         struct ceph_file_info *fi = file->private_data;
1676         struct inode *inode = file_inode(file);
1677         struct ceph_inode_info *ci = ceph_inode(inode);
1678         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1679         struct ceph_osd_client *osdc = &fsc->client->osdc;
1680         struct ceph_cap_flush *prealloc_cf;
1681         ssize_t count, written = 0;
1682         int err, want, got;
1683         bool direct_lock = false;
1684         u32 map_flags;
1685         u64 pool_flags;
1686         loff_t pos;
1687         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1688
1689         if (ceph_inode_is_shutdown(inode))
1690                 return -ESTALE;
1691
1692         if (ceph_snap(inode) != CEPH_NOSNAP)
1693                 return -EROFS;
1694
1695         prealloc_cf = ceph_alloc_cap_flush();
1696         if (!prealloc_cf)
1697                 return -ENOMEM;
1698
1699         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1700                 direct_lock = true;
1701
1702 retry_snap:
1703         if (direct_lock)
1704                 ceph_start_io_direct(inode);
1705         else
1706                 ceph_start_io_write(inode);
1707
1708         /* We can write back this queue in page reclaim */
1709         current->backing_dev_info = inode_to_bdi(inode);
1710
1711         if (iocb->ki_flags & IOCB_APPEND) {
1712                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1713                 if (err < 0)
1714                         goto out;
1715         }
1716
1717         err = generic_write_checks(iocb, from);
1718         if (err <= 0)
1719                 goto out;
1720
1721         pos = iocb->ki_pos;
1722         if (unlikely(pos >= limit)) {
1723                 err = -EFBIG;
1724                 goto out;
1725         } else {
1726                 iov_iter_truncate(from, limit - pos);
1727         }
1728
1729         count = iov_iter_count(from);
1730         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1731                 err = -EDQUOT;
1732                 goto out;
1733         }
1734
1735         down_read(&osdc->lock);
1736         map_flags = osdc->osdmap->flags;
1737         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1738         up_read(&osdc->lock);
1739         if ((map_flags & CEPH_OSDMAP_FULL) ||
1740             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1741                 err = -ENOSPC;
1742                 goto out;
1743         }
1744
1745         err = file_remove_privs(file);
1746         if (err)
1747                 goto out;
1748
1749         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1750                 err = ceph_uninline_data(file, NULL);
1751                 if (err < 0)
1752                         goto out;
1753         }
1754
1755         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1756              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1757         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1758                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1759         else
1760                 want = CEPH_CAP_FILE_BUFFER;
1761         got = 0;
1762         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1763         if (err < 0)
1764                 goto out;
1765
1766         err = file_update_time(file);
1767         if (err)
1768                 goto out_caps;
1769
1770         inode_inc_iversion_raw(inode);
1771
1772         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1773              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1774
1775         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1776             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1777             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1778                 struct ceph_snap_context *snapc;
1779                 struct iov_iter data;
1780
1781                 spin_lock(&ci->i_ceph_lock);
1782                 if (__ceph_have_pending_cap_snap(ci)) {
1783                         struct ceph_cap_snap *capsnap =
1784                                         list_last_entry(&ci->i_cap_snaps,
1785                                                         struct ceph_cap_snap,
1786                                                         ci_item);
1787                         snapc = ceph_get_snap_context(capsnap->context);
1788                 } else {
1789                         BUG_ON(!ci->i_head_snapc);
1790                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1791                 }
1792                 spin_unlock(&ci->i_ceph_lock);
1793
1794                 /* we might need to revert back to that point */
1795                 data = *from;
1796                 if (iocb->ki_flags & IOCB_DIRECT)
1797                         written = ceph_direct_read_write(iocb, &data, snapc,
1798                                                          &prealloc_cf);
1799                 else
1800                         written = ceph_sync_write(iocb, &data, pos, snapc);
1801                 if (direct_lock)
1802                         ceph_end_io_direct(inode);
1803                 else
1804                         ceph_end_io_write(inode);
1805                 if (written > 0)
1806                         iov_iter_advance(from, written);
1807                 ceph_put_snap_context(snapc);
1808         } else {
1809                 /*
1810                  * No need to acquire the i_truncate_mutex. Because
1811                  * the MDS revokes Fwb caps before sending truncate
1812                  * message to us. We can't get Fwb cap while there
1813                  * are pending vmtruncate. So write and vmtruncate
1814                  * can not run at the same time
1815                  */
1816                 written = generic_perform_write(file, from, pos);
1817                 if (likely(written >= 0))
1818                         iocb->ki_pos = pos + written;
1819                 ceph_end_io_write(inode);
1820         }
1821
1822         if (written >= 0) {
1823                 int dirty;
1824
1825                 spin_lock(&ci->i_ceph_lock);
1826                 ci->i_inline_version = CEPH_INLINE_NONE;
1827                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1828                                                &prealloc_cf);
1829                 spin_unlock(&ci->i_ceph_lock);
1830                 if (dirty)
1831                         __mark_inode_dirty(inode, dirty);
1832                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1833                         ceph_check_caps(ci, 0, NULL);
1834         }
1835
1836         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1837              inode, ceph_vinop(inode), pos, (unsigned)count,
1838              ceph_cap_string(got));
1839         ceph_put_cap_refs(ci, got);
1840
1841         if (written == -EOLDSNAPC) {
1842                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1843                      inode, ceph_vinop(inode), pos, (unsigned)count);
1844                 goto retry_snap;
1845         }
1846
1847         if (written >= 0) {
1848                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1849                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1850                         iocb->ki_flags |= IOCB_DSYNC;
1851                 written = generic_write_sync(iocb, written);
1852         }
1853
1854         goto out_unlocked;
1855 out_caps:
1856         ceph_put_cap_refs(ci, got);
1857 out:
1858         if (direct_lock)
1859                 ceph_end_io_direct(inode);
1860         else
1861                 ceph_end_io_write(inode);
1862 out_unlocked:
1863         ceph_free_cap_flush(prealloc_cf);
1864         current->backing_dev_info = NULL;
1865         return written ? written : err;
1866 }
1867
1868 /*
1869  * llseek.  be sure to verify file size on SEEK_END.
1870  */
1871 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1872 {
1873         struct inode *inode = file->f_mapping->host;
1874         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1875         loff_t i_size;
1876         loff_t ret;
1877
1878         inode_lock(inode);
1879
1880         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1881                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1882                 if (ret < 0)
1883                         goto out;
1884         }
1885
1886         i_size = i_size_read(inode);
1887         switch (whence) {
1888         case SEEK_END:
1889                 offset += i_size;
1890                 break;
1891         case SEEK_CUR:
1892                 /*
1893                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1894                  * position-querying operation.  Avoid rewriting the "same"
1895                  * f_pos value back to the file because a concurrent read(),
1896                  * write() or lseek() might have altered it
1897                  */
1898                 if (offset == 0) {
1899                         ret = file->f_pos;
1900                         goto out;
1901                 }
1902                 offset += file->f_pos;
1903                 break;
1904         case SEEK_DATA:
1905                 if (offset < 0 || offset >= i_size) {
1906                         ret = -ENXIO;
1907                         goto out;
1908                 }
1909                 break;
1910         case SEEK_HOLE:
1911                 if (offset < 0 || offset >= i_size) {
1912                         ret = -ENXIO;
1913                         goto out;
1914                 }
1915                 offset = i_size;
1916                 break;
1917         }
1918
1919         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1920
1921 out:
1922         inode_unlock(inode);
1923         return ret;
1924 }
1925
1926 static inline void ceph_zero_partial_page(
1927         struct inode *inode, loff_t offset, unsigned size)
1928 {
1929         struct page *page;
1930         pgoff_t index = offset >> PAGE_SHIFT;
1931
1932         page = find_lock_page(inode->i_mapping, index);
1933         if (page) {
1934                 wait_on_page_writeback(page);
1935                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1936                 unlock_page(page);
1937                 put_page(page);
1938         }
1939 }
1940
1941 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1942                                       loff_t length)
1943 {
1944         loff_t nearly = round_up(offset, PAGE_SIZE);
1945         if (offset < nearly) {
1946                 loff_t size = nearly - offset;
1947                 if (length < size)
1948                         size = length;
1949                 ceph_zero_partial_page(inode, offset, size);
1950                 offset += size;
1951                 length -= size;
1952         }
1953         if (length >= PAGE_SIZE) {
1954                 loff_t size = round_down(length, PAGE_SIZE);
1955                 truncate_pagecache_range(inode, offset, offset + size - 1);
1956                 offset += size;
1957                 length -= size;
1958         }
1959         if (length)
1960                 ceph_zero_partial_page(inode, offset, length);
1961 }
1962
1963 static int ceph_zero_partial_object(struct inode *inode,
1964                                     loff_t offset, loff_t *length)
1965 {
1966         struct ceph_inode_info *ci = ceph_inode(inode);
1967         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1968         struct ceph_osd_request *req;
1969         int ret = 0;
1970         loff_t zero = 0;
1971         int op;
1972
1973         if (!length) {
1974                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1975                 length = &zero;
1976         } else {
1977                 op = CEPH_OSD_OP_ZERO;
1978         }
1979
1980         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1981                                         ceph_vino(inode),
1982                                         offset, length,
1983                                         0, 1, op,
1984                                         CEPH_OSD_FLAG_WRITE,
1985                                         NULL, 0, 0, false);
1986         if (IS_ERR(req)) {
1987                 ret = PTR_ERR(req);
1988                 goto out;
1989         }
1990
1991         req->r_mtime = inode->i_mtime;
1992         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1993         if (!ret) {
1994                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1995                 if (ret == -ENOENT)
1996                         ret = 0;
1997         }
1998         ceph_osdc_put_request(req);
1999
2000 out:
2001         return ret;
2002 }
2003
2004 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2005 {
2006         int ret = 0;
2007         struct ceph_inode_info *ci = ceph_inode(inode);
2008         s32 stripe_unit = ci->i_layout.stripe_unit;
2009         s32 stripe_count = ci->i_layout.stripe_count;
2010         s32 object_size = ci->i_layout.object_size;
2011         u64 object_set_size = object_size * stripe_count;
2012         u64 nearly, t;
2013
2014         /* round offset up to next period boundary */
2015         nearly = offset + object_set_size - 1;
2016         t = nearly;
2017         nearly -= do_div(t, object_set_size);
2018
2019         while (length && offset < nearly) {
2020                 loff_t size = length;
2021                 ret = ceph_zero_partial_object(inode, offset, &size);
2022                 if (ret < 0)
2023                         return ret;
2024                 offset += size;
2025                 length -= size;
2026         }
2027         while (length >= object_set_size) {
2028                 int i;
2029                 loff_t pos = offset;
2030                 for (i = 0; i < stripe_count; ++i) {
2031                         ret = ceph_zero_partial_object(inode, pos, NULL);
2032                         if (ret < 0)
2033                                 return ret;
2034                         pos += stripe_unit;
2035                 }
2036                 offset += object_set_size;
2037                 length -= object_set_size;
2038         }
2039         while (length) {
2040                 loff_t size = length;
2041                 ret = ceph_zero_partial_object(inode, offset, &size);
2042                 if (ret < 0)
2043                         return ret;
2044                 offset += size;
2045                 length -= size;
2046         }
2047         return ret;
2048 }
2049
2050 static long ceph_fallocate(struct file *file, int mode,
2051                                 loff_t offset, loff_t length)
2052 {
2053         struct ceph_file_info *fi = file->private_data;
2054         struct inode *inode = file_inode(file);
2055         struct ceph_inode_info *ci = ceph_inode(inode);
2056         struct ceph_cap_flush *prealloc_cf;
2057         int want, got = 0;
2058         int dirty;
2059         int ret = 0;
2060         loff_t endoff = 0;
2061         loff_t size;
2062
2063         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2064                 return -EOPNOTSUPP;
2065
2066         if (!S_ISREG(inode->i_mode))
2067                 return -EOPNOTSUPP;
2068
2069         prealloc_cf = ceph_alloc_cap_flush();
2070         if (!prealloc_cf)
2071                 return -ENOMEM;
2072
2073         inode_lock(inode);
2074
2075         if (ceph_snap(inode) != CEPH_NOSNAP) {
2076                 ret = -EROFS;
2077                 goto unlock;
2078         }
2079
2080         if (ci->i_inline_version != CEPH_INLINE_NONE) {
2081                 ret = ceph_uninline_data(file, NULL);
2082                 if (ret < 0)
2083                         goto unlock;
2084         }
2085
2086         size = i_size_read(inode);
2087
2088         /* Are we punching a hole beyond EOF? */
2089         if (offset >= size)
2090                 goto unlock;
2091         if ((offset + length) > size)
2092                 length = size - offset;
2093
2094         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2095                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2096         else
2097                 want = CEPH_CAP_FILE_BUFFER;
2098
2099         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2100         if (ret < 0)
2101                 goto unlock;
2102
2103         filemap_invalidate_lock(inode->i_mapping);
2104         ceph_zero_pagecache_range(inode, offset, length);
2105         ret = ceph_zero_objects(inode, offset, length);
2106
2107         if (!ret) {
2108                 spin_lock(&ci->i_ceph_lock);
2109                 ci->i_inline_version = CEPH_INLINE_NONE;
2110                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2111                                                &prealloc_cf);
2112                 spin_unlock(&ci->i_ceph_lock);
2113                 if (dirty)
2114                         __mark_inode_dirty(inode, dirty);
2115         }
2116         filemap_invalidate_unlock(inode->i_mapping);
2117
2118         ceph_put_cap_refs(ci, got);
2119 unlock:
2120         inode_unlock(inode);
2121         ceph_free_cap_flush(prealloc_cf);
2122         return ret;
2123 }
2124
2125 /*
2126  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2127  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2128  * this fails; zero is returned on success.
2129  */
2130 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2131                           struct file *dst_filp,
2132                           loff_t dst_endoff, int *dst_got)
2133 {
2134         int ret = 0;
2135         bool retrying = false;
2136
2137 retry_caps:
2138         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2139                             dst_endoff, dst_got);
2140         if (ret < 0)
2141                 return ret;
2142
2143         /*
2144          * Since we're already holding the FILE_WR capability for the dst file,
2145          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2146          * retry dance instead to try to get both capabilities.
2147          */
2148         ret = ceph_try_get_caps(file_inode(src_filp),
2149                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2150                                 false, src_got);
2151         if (ret <= 0) {
2152                 /* Start by dropping dst_ci caps and getting src_ci caps */
2153                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2154                 if (retrying) {
2155                         if (!ret)
2156                                 /* ceph_try_get_caps masks EAGAIN */
2157                                 ret = -EAGAIN;
2158                         return ret;
2159                 }
2160                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2161                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2162                 if (ret < 0)
2163                         return ret;
2164                 /*... drop src_ci caps too, and retry */
2165                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2166                 retrying = true;
2167                 goto retry_caps;
2168         }
2169         return ret;
2170 }
2171
2172 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2173                            struct ceph_inode_info *dst_ci, int dst_got)
2174 {
2175         ceph_put_cap_refs(src_ci, src_got);
2176         ceph_put_cap_refs(dst_ci, dst_got);
2177 }
2178
2179 /*
2180  * This function does several size-related checks, returning an error if:
2181  *  - source file is smaller than off+len
2182  *  - destination file size is not OK (inode_newsize_ok())
2183  *  - max bytes quotas is exceeded
2184  */
2185 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2186                            loff_t src_off, loff_t dst_off, size_t len)
2187 {
2188         loff_t size, endoff;
2189
2190         size = i_size_read(src_inode);
2191         /*
2192          * Don't copy beyond source file EOF.  Instead of simply setting length
2193          * to (size - src_off), just drop to VFS default implementation, as the
2194          * local i_size may be stale due to other clients writing to the source
2195          * inode.
2196          */
2197         if (src_off + len > size) {
2198                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2199                      src_off, len, size);
2200                 return -EOPNOTSUPP;
2201         }
2202         size = i_size_read(dst_inode);
2203
2204         endoff = dst_off + len;
2205         if (inode_newsize_ok(dst_inode, endoff))
2206                 return -EOPNOTSUPP;
2207
2208         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2209                 return -EDQUOT;
2210
2211         return 0;
2212 }
2213
2214 static struct ceph_osd_request *
2215 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2216                             u64 src_snapid,
2217                             struct ceph_object_id *src_oid,
2218                             struct ceph_object_locator *src_oloc,
2219                             struct ceph_object_id *dst_oid,
2220                             struct ceph_object_locator *dst_oloc,
2221                             u32 truncate_seq, u64 truncate_size)
2222 {
2223         struct ceph_osd_request *req;
2224         int ret;
2225         u32 src_fadvise_flags =
2226                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2227                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2228         u32 dst_fadvise_flags =
2229                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2230                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2231
2232         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2233         if (!req)
2234                 return ERR_PTR(-ENOMEM);
2235
2236         req->r_flags = CEPH_OSD_FLAG_WRITE;
2237
2238         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2239         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2240
2241         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2242                                         src_oid, src_oloc,
2243                                         src_fadvise_flags,
2244                                         dst_fadvise_flags,
2245                                         truncate_seq,
2246                                         truncate_size,
2247                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2248         if (ret)
2249                 goto out;
2250
2251         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2252         if (ret)
2253                 goto out;
2254
2255         return req;
2256
2257 out:
2258         ceph_osdc_put_request(req);
2259         return ERR_PTR(ret);
2260 }
2261
2262 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2263                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2264                                     struct ceph_fs_client *fsc,
2265                                     size_t len, unsigned int flags)
2266 {
2267         struct ceph_object_locator src_oloc, dst_oloc;
2268         struct ceph_object_id src_oid, dst_oid;
2269         struct ceph_osd_client *osdc;
2270         struct ceph_osd_request *req;
2271         size_t bytes = 0;
2272         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2273         u32 src_objlen, dst_objlen;
2274         u32 object_size = src_ci->i_layout.object_size;
2275         int ret;
2276
2277         src_oloc.pool = src_ci->i_layout.pool_id;
2278         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2279         dst_oloc.pool = dst_ci->i_layout.pool_id;
2280         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2281         osdc = &fsc->client->osdc;
2282
2283         while (len >= object_size) {
2284                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2285                                               object_size, &src_objnum,
2286                                               &src_objoff, &src_objlen);
2287                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2288                                               object_size, &dst_objnum,
2289                                               &dst_objoff, &dst_objlen);
2290                 ceph_oid_init(&src_oid);
2291                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2292                                 src_ci->i_vino.ino, src_objnum);
2293                 ceph_oid_init(&dst_oid);
2294                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2295                                 dst_ci->i_vino.ino, dst_objnum);
2296                 /* Do an object remote copy */
2297                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2298                                                   &src_oid, &src_oloc,
2299                                                   &dst_oid, &dst_oloc,
2300                                                   dst_ci->i_truncate_seq,
2301                                                   dst_ci->i_truncate_size);
2302                 if (IS_ERR(req))
2303                         ret = PTR_ERR(req);
2304                 else {
2305                         ceph_osdc_start_request(osdc, req, false);
2306                         ret = ceph_osdc_wait_request(osdc, req);
2307                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2308                                                      req->r_start_latency,
2309                                                      req->r_end_latency,
2310                                                      object_size, ret);
2311                         ceph_osdc_put_request(req);
2312                 }
2313                 if (ret) {
2314                         if (ret == -EOPNOTSUPP) {
2315                                 fsc->have_copy_from2 = false;
2316                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2317                         }
2318                         dout("ceph_osdc_copy_from returned %d\n", ret);
2319                         if (!bytes)
2320                                 bytes = ret;
2321                         goto out;
2322                 }
2323                 len -= object_size;
2324                 bytes += object_size;
2325                 *src_off += object_size;
2326                 *dst_off += object_size;
2327         }
2328
2329 out:
2330         ceph_oloc_destroy(&src_oloc);
2331         ceph_oloc_destroy(&dst_oloc);
2332         return bytes;
2333 }
2334
2335 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2336                                       struct file *dst_file, loff_t dst_off,
2337                                       size_t len, unsigned int flags)
2338 {
2339         struct inode *src_inode = file_inode(src_file);
2340         struct inode *dst_inode = file_inode(dst_file);
2341         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2342         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2343         struct ceph_cap_flush *prealloc_cf;
2344         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2345         loff_t size;
2346         ssize_t ret = -EIO, bytes;
2347         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2348         u32 src_objlen, dst_objlen;
2349         int src_got = 0, dst_got = 0, err, dirty;
2350
2351         if (src_inode->i_sb != dst_inode->i_sb) {
2352                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2353
2354                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2355                                       &dst_fsc->client->fsid)) {
2356                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2357                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2358                         return -EXDEV;
2359                 }
2360         }
2361         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2362                 return -EROFS;
2363
2364         /*
2365          * Some of the checks below will return -EOPNOTSUPP, which will force a
2366          * fallback to the default VFS copy_file_range implementation.  This is
2367          * desirable in several cases (for ex, the 'len' is smaller than the
2368          * size of the objects, or in cases where that would be more
2369          * efficient).
2370          */
2371
2372         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2373                 return -EOPNOTSUPP;
2374
2375         if (!src_fsc->have_copy_from2)
2376                 return -EOPNOTSUPP;
2377
2378         /*
2379          * Striped file layouts require that we copy partial objects, but the
2380          * OSD copy-from operation only supports full-object copies.  Limit
2381          * this to non-striped file layouts for now.
2382          */
2383         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2384             (src_ci->i_layout.stripe_count != 1) ||
2385             (dst_ci->i_layout.stripe_count != 1) ||
2386             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2387                 dout("Invalid src/dst files layout\n");
2388                 return -EOPNOTSUPP;
2389         }
2390
2391         if (len < src_ci->i_layout.object_size)
2392                 return -EOPNOTSUPP; /* no remote copy will be done */
2393
2394         prealloc_cf = ceph_alloc_cap_flush();
2395         if (!prealloc_cf)
2396                 return -ENOMEM;
2397
2398         /* Start by sync'ing the source and destination files */
2399         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2400         if (ret < 0) {
2401                 dout("failed to write src file (%zd)\n", ret);
2402                 goto out;
2403         }
2404         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2405         if (ret < 0) {
2406                 dout("failed to write dst file (%zd)\n", ret);
2407                 goto out;
2408         }
2409
2410         /*
2411          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2412          * clients may have dirty data in their caches.  And OSDs know nothing
2413          * about caps, so they can't safely do the remote object copies.
2414          */
2415         err = get_rd_wr_caps(src_file, &src_got,
2416                              dst_file, (dst_off + len), &dst_got);
2417         if (err < 0) {
2418                 dout("get_rd_wr_caps returned %d\n", err);
2419                 ret = -EOPNOTSUPP;
2420                 goto out;
2421         }
2422
2423         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2424         if (ret < 0)
2425                 goto out_caps;
2426
2427         /* Drop dst file cached pages */
2428         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2429                                             dst_off >> PAGE_SHIFT,
2430                                             (dst_off + len) >> PAGE_SHIFT);
2431         if (ret < 0) {
2432                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2433                 ret = 0; /* XXX */
2434         }
2435         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2436                                       src_ci->i_layout.object_size,
2437                                       &src_objnum, &src_objoff, &src_objlen);
2438         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2439                                       dst_ci->i_layout.object_size,
2440                                       &dst_objnum, &dst_objoff, &dst_objlen);
2441         /* object-level offsets need to the same */
2442         if (src_objoff != dst_objoff) {
2443                 ret = -EOPNOTSUPP;
2444                 goto out_caps;
2445         }
2446
2447         /*
2448          * Do a manual copy if the object offset isn't object aligned.
2449          * 'src_objlen' contains the bytes left until the end of the object,
2450          * starting at the src_off
2451          */
2452         if (src_objoff) {
2453                 dout("Initial partial copy of %u bytes\n", src_objlen);
2454
2455                 /*
2456                  * we need to temporarily drop all caps as we'll be calling
2457                  * {read,write}_iter, which will get caps again.
2458                  */
2459                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2460                 ret = do_splice_direct(src_file, &src_off, dst_file,
2461                                        &dst_off, src_objlen, flags);
2462                 /* Abort on short copies or on error */
2463                 if (ret < src_objlen) {
2464                         dout("Failed partial copy (%zd)\n", ret);
2465                         goto out;
2466                 }
2467                 len -= ret;
2468                 err = get_rd_wr_caps(src_file, &src_got,
2469                                      dst_file, (dst_off + len), &dst_got);
2470                 if (err < 0)
2471                         goto out;
2472                 err = is_file_size_ok(src_inode, dst_inode,
2473                                       src_off, dst_off, len);
2474                 if (err < 0)
2475                         goto out_caps;
2476         }
2477
2478         size = i_size_read(dst_inode);
2479         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2480                                      src_fsc, len, flags);
2481         if (bytes <= 0) {
2482                 if (!ret)
2483                         ret = bytes;
2484                 goto out_caps;
2485         }
2486         dout("Copied %zu bytes out of %zu\n", bytes, len);
2487         len -= bytes;
2488         ret += bytes;
2489
2490         file_update_time(dst_file);
2491         inode_inc_iversion_raw(dst_inode);
2492
2493         if (dst_off > size) {
2494                 /* Let the MDS know about dst file size change */
2495                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2496                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2497                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2498         }
2499         /* Mark Fw dirty */
2500         spin_lock(&dst_ci->i_ceph_lock);
2501         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2502         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2503         spin_unlock(&dst_ci->i_ceph_lock);
2504         if (dirty)
2505                 __mark_inode_dirty(dst_inode, dirty);
2506
2507 out_caps:
2508         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2509
2510         /*
2511          * Do the final manual copy if we still have some bytes left, unless
2512          * there were errors in remote object copies (len >= object_size).
2513          */
2514         if (len && (len < src_ci->i_layout.object_size)) {
2515                 dout("Final partial copy of %zu bytes\n", len);
2516                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2517                                          &dst_off, len, flags);
2518                 if (bytes > 0)
2519                         ret += bytes;
2520                 else
2521                         dout("Failed partial copy (%zd)\n", bytes);
2522         }
2523
2524 out:
2525         ceph_free_cap_flush(prealloc_cf);
2526
2527         return ret;
2528 }
2529
2530 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2531                                     struct file *dst_file, loff_t dst_off,
2532                                     size_t len, unsigned int flags)
2533 {
2534         ssize_t ret;
2535
2536         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2537                                      len, flags);
2538
2539         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2540                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2541                                               dst_off, len, flags);
2542         return ret;
2543 }
2544
2545 const struct file_operations ceph_file_fops = {
2546         .open = ceph_open,
2547         .release = ceph_release,
2548         .llseek = ceph_llseek,
2549         .read_iter = ceph_read_iter,
2550         .write_iter = ceph_write_iter,
2551         .mmap = ceph_mmap,
2552         .fsync = ceph_fsync,
2553         .lock = ceph_lock,
2554         .setlease = simple_nosetlease,
2555         .flock = ceph_flock,
2556         .splice_read = generic_file_splice_read,
2557         .splice_write = iter_file_splice_write,
2558         .unlocked_ioctl = ceph_ioctl,
2559         .compat_ioctl = compat_ptr_ioctl,
2560         .fallocate      = ceph_fallocate,
2561         .copy_file_range = ceph_copy_file_range,
2562 };