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
  14 #include "super.h"
  15 #include "mds_client.h"
  16 #include "cache.h"
  17
  18 static __le32 ceph_flags_sys2wire(u32 flags)
  19 {
  20         u32 wire_flags = 0;
  21
  22         switch (flags & O_ACCMODE) {
  23         case O_RDONLY:
  24                 wire_flags |= CEPH_O_RDONLY;
  25                 break;
  26         case O_WRONLY:
  27                 wire_flags |= CEPH_O_WRONLY;
  28                 break;
  29         case O_RDWR:
  30                 wire_flags |= CEPH_O_RDWR;
  31                 break;
  32         }
  33
  34         flags &= ~O_ACCMODE;
  35
  36 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  37
  38         ceph_sys2wire(O_CREAT);
  39         ceph_sys2wire(O_EXCL);
  40         ceph_sys2wire(O_TRUNC);
  41         ceph_sys2wire(O_DIRECTORY);
  42         ceph_sys2wire(O_NOFOLLOW);
  43
  44 #undef ceph_sys2wire
  45
  46         if (flags)
  47                 dout("unused open flags: %x\n", flags);
  48
  49         return cpu_to_le32(wire_flags);
  50 }
  51
  52 /*
  53  * Ceph file operations
  54  *
  55  * Implement basic open/close functionality, and implement
  56  * read/write.
  57  *
  58  * We implement three modes of file I/O:
  59  *  - buffered uses the generic_file_aio_{read,write} helpers
  60  *
  61  *  - synchronous is used when there is multi-client read/write
  62  *    sharing, avoids the page cache, and synchronously waits for an
  63  *    ack from the OSD.
  64  *
  65  *  - direct io takes the variant of the sync path that references
  66  *    user pages directly.
  67  *
  68  * fsync() flushes and waits on dirty pages, but just queues metadata
  69  * for writeback: since the MDS can recover size and mtime there is no
  70  * need to wait for MDS acknowledgement.
  71  */
  72
  73 /*
  74  * How many pages to get in one call to iov_iter_get_pages().  This
  75  * determines the size of the on-stack array used as a buffer.
  76  */
  77 #define ITER_GET_BVECS_PAGES    64
  78
  79 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  80                                 struct bio_vec *bvecs)
  81 {
  82         size_t size = 0;
  83         int bvec_idx = 0;
  84
  85         if (maxsize > iov_iter_count(iter))
  86                 maxsize = iov_iter_count(iter);
  87
  88         while (size < maxsize) {
  89                 struct page *pages[ITER_GET_BVECS_PAGES];
  90                 ssize_t bytes;
  91                 size_t start;
  92                 int idx = 0;
  93
  94                 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
  95                                            ITER_GET_BVECS_PAGES, &start);
  96                 if (bytes < 0)
  97                         return size ?: bytes;
  98
  99                 iov_iter_advance(iter, bytes);
 100                 size += bytes;
 101
 102                 for ( ; bytes; idx++, bvec_idx++) {
 103                         struct bio_vec bv = {
 104                                 .bv_page = pages[idx],
 105                                 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
 106                                 .bv_offset = start,
 107                         };
 108
 109                         bvecs[bvec_idx] = bv;
 110                         bytes -= bv.bv_len;
 111                         start = 0;
 112                 }
 113         }
 114
 115         return size;
 116 }
 117
 118 /*
 119  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 120  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 121  * page.
 122  *
 123  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 124  * Return the number of bytes in the created bio_vec array, or an error.
 125  */
 126 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 127                                     struct bio_vec **bvecs, int *num_bvecs)
 128 {
 129         struct bio_vec *bv;
 130         size_t orig_count = iov_iter_count(iter);
 131         ssize_t bytes;
 132         int npages;
 133
 134         iov_iter_truncate(iter, maxsize);
 135         npages = iov_iter_npages(iter, INT_MAX);
 136         iov_iter_reexpand(iter, orig_count);
 137
 138         /*
 139          * __iter_get_bvecs() may populate only part of the array -- zero it
 140          * out.
 141          */
 142         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 143         if (!bv)
 144                 return -ENOMEM;
 145
 146         bytes = __iter_get_bvecs(iter, maxsize, bv);
 147         if (bytes < 0) {
 148                 /*
 149                  * No pages were pinned -- just free the array.
 150                  */
 151                 kvfree(bv);
 152                 return bytes;
 153         }
 154
 155         *bvecs = bv;
 156         *num_bvecs = npages;
 157         return bytes;
 158 }
 159
 160 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 161 {
 162         int i;
 163
 164         for (i = 0; i < num_bvecs; i++) {
 165                 if (bvecs[i].bv_page) {
 166                         if (should_dirty)
 167                                 set_page_dirty_lock(bvecs[i].bv_page);
 168                         put_page(bvecs[i].bv_page);
 169                 }
 170         }
 171         kvfree(bvecs);
 172 }
 173
 174 /*
 175  * Prepare an open request.  Preallocate ceph_cap to avoid an
 176  * inopportune ENOMEM later.
 177  */
 178 static struct ceph_mds_request *
 179 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 180 {
 181         struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
 182         struct ceph_mds_client *mdsc = fsc->mdsc;
 183         struct ceph_mds_request *req;
 184         int want_auth = USE_ANY_MDS;
 185         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 186
 187         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 188                 want_auth = USE_AUTH_MDS;
 189
 190         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 191         if (IS_ERR(req))
 192                 goto out;
 193         req->r_fmode = ceph_flags_to_mode(flags);
 194         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 195         req->r_args.open.mode = cpu_to_le32(create_mode);
 196 out:
 197         return req;
 198 }
 199
 200 static int ceph_init_file_info(struct inode *inode, struct file *file,
 201                                         int fmode, bool isdir)
 202 {
 203         struct ceph_file_info *fi;
 204
 205         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 206                         inode->i_mode, isdir ? "dir" : "regular");
 207         BUG_ON(inode->i_fop->release != ceph_release);
 208
 209         if (isdir) {
 210                 struct ceph_dir_file_info *dfi =
 211                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 212                 if (!dfi) {
 213                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 214                         return -ENOMEM;
 215                 }
 216
 217                 file->private_data = dfi;
 218                 fi = &dfi->file_info;
 219                 dfi->next_offset = 2;
 220                 dfi->readdir_cache_idx = -1;
 221         } else {
 222                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 223                 if (!fi) {
 224                         ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 225                         return -ENOMEM;
 226                 }
 227
 228                 file->private_data = fi;
 229         }
 230
 231         fi->fmode = fmode;
 232         spin_lock_init(&fi->rw_contexts_lock);
 233         INIT_LIST_HEAD(&fi->rw_contexts);
 234
 235         return 0;
 236 }
 237
 238 /*
 239  * initialize private struct file data.
 240  * if we fail, clean up by dropping fmode reference on the ceph_inode
 241  */
 242 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 243 {
 244         int ret = 0;
 245
 246         switch (inode->i_mode & S_IFMT) {
 247         case S_IFREG:
 248                 ceph_fscache_register_inode_cookie(inode);
 249                 ceph_fscache_file_set_cookie(inode, file);
 250         case S_IFDIR:
 251                 ret = ceph_init_file_info(inode, file, fmode,
 252                                                 S_ISDIR(inode->i_mode));
 253                 if (ret)
 254                         return ret;
 255                 break;
 256
 257         case S_IFLNK:
 258                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 259                      inode->i_mode);
 260                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 261                 break;
 262
 263         default:
 264                 dout("init_file %p %p 0%o (special)\n", inode, file,
 265                      inode->i_mode);
 266                 /*
 267                  * we need to drop the open ref now, since we don't
 268                  * have .release set to ceph_release.
 269                  */
 270                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 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)
 283 {
 284         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 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         wanted = __ceph_caps_file_wanted(ci);
 291         if (__ceph_is_any_real_caps(ci) &&
 292             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 293                 int issued = __ceph_caps_issued(ci, NULL);
 294                 spin_unlock(&ci->i_ceph_lock);
 295                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 296                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 297                 ceph_check_caps(ci, 0, NULL);
 298                 return 0;
 299         }
 300         spin_unlock(&ci->i_ceph_lock);
 301
 302         flags = 0;
 303         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 304                 flags = O_RDWR;
 305         else if (wanted & CEPH_CAP_FILE_RD)
 306                 flags = O_RDONLY;
 307         else if (wanted & CEPH_CAP_FILE_WR)
 308                 flags = O_WRONLY;
 309 #ifdef O_LAZY
 310         if (wanted & CEPH_CAP_FILE_LAZYIO)
 311                 flags |= O_LAZY;
 312 #endif
 313
 314         req = prepare_open_request(inode->i_sb, flags, 0);
 315         if (IS_ERR(req)) {
 316                 err = PTR_ERR(req);
 317                 goto out;
 318         }
 319
 320         req->r_inode = inode;
 321         ihold(inode);
 322         req->r_num_caps = 1;
 323         req->r_fmode = -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                 spin_lock(&ci->i_ceph_lock);
 370                 __ceph_get_fmode(ci, fmode);
 371                 spin_unlock(&ci->i_ceph_lock);
 372                 return ceph_init_file(inode, file, fmode);
 373         }
 374
 375         /*
 376          * No need to block if we have caps on the auth MDS (for
 377          * write) or any MDS (for read).  Update wanted set
 378          * asynchronously.
 379          */
 380         spin_lock(&ci->i_ceph_lock);
 381         if (__ceph_is_any_real_caps(ci) &&
 382             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 383                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 384                 int issued = __ceph_caps_issued(ci, NULL);
 385
 386                 dout("open %p fmode %d want %s issued %s using existing\n",
 387                      inode, fmode, ceph_cap_string(wanted),
 388                      ceph_cap_string(issued));
 389                 __ceph_get_fmode(ci, fmode);
 390                 spin_unlock(&ci->i_ceph_lock);
 391
 392                 /* adjust wanted? */
 393                 if ((issued & wanted) != wanted &&
 394                     (mds_wanted & wanted) != wanted &&
 395                     ceph_snap(inode) != CEPH_SNAPDIR)
 396                         ceph_check_caps(ci, 0, NULL);
 397
 398                 return ceph_init_file(inode, file, fmode);
 399         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 400                    (ci->i_snap_caps & wanted) == wanted) {
 401                 __ceph_get_fmode(ci, fmode);
 402                 spin_unlock(&ci->i_ceph_lock);
 403                 return ceph_init_file(inode, file, fmode);
 404         }
 405
 406         spin_unlock(&ci->i_ceph_lock);
 407
 408         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 409         req = prepare_open_request(inode->i_sb, flags, 0);
 410         if (IS_ERR(req)) {
 411                 err = PTR_ERR(req);
 412                 goto out;
 413         }
 414         req->r_inode = inode;
 415         ihold(inode);
 416
 417         req->r_num_caps = 1;
 418         err = ceph_mdsc_do_request(mdsc, NULL, req);
 419         if (!err)
 420                 err = ceph_init_file(inode, file, req->r_fmode);
 421         ceph_mdsc_put_request(req);
 422         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 423 out:
 424         return err;
 425 }
 426
 427
 428 /*
 429  * Do a lookup + open with a single request.  If we get a non-existent
 430  * file or symlink, return 1 so the VFS can retry.
 431  */
 432 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 433                      struct file *file, unsigned flags, umode_t mode)
 434 {
 435         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 436         struct ceph_mds_client *mdsc = fsc->mdsc;
 437         struct ceph_mds_request *req;
 438         struct dentry *dn;
 439         struct ceph_acls_info acls = {};
 440         int mask;
 441         int err;
 442
 443         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 444              dir, dentry, dentry,
 445              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 446
 447         if (dentry->d_name.len > NAME_MAX)
 448                 return -ENAMETOOLONG;
 449
 450         if (flags & O_CREAT) {
 451                 if (ceph_quota_is_max_files_exceeded(dir))
 452                         return -EDQUOT;
 453                 err = ceph_pre_init_acls(dir, &mode, &acls);
 454                 if (err < 0)
 455                         return err;
 456         }
 457
 458         /* do the open */
 459         req = prepare_open_request(dir->i_sb, flags, mode);
 460         if (IS_ERR(req)) {
 461                 err = PTR_ERR(req);
 462                 goto out_acl;
 463         }
 464         req->r_dentry = dget(dentry);
 465         req->r_num_caps = 2;
 466         if (flags & O_CREAT) {
 467                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 468                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 469                 if (acls.pagelist) {
 470                         req->r_pagelist = acls.pagelist;
 471                         acls.pagelist = NULL;
 472                 }
 473         }
 474
 475        mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 476        if (ceph_security_xattr_wanted(dir))
 477                mask |= CEPH_CAP_XATTR_SHARED;
 478        req->r_args.open.mask = cpu_to_le32(mask);
 479
 480         req->r_parent = dir;
 481         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 482         err = ceph_mdsc_do_request(mdsc,
 483                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 484                                    req);
 485         err = ceph_handle_snapdir(req, dentry, err);
 486         if (err)
 487                 goto out_req;
 488
 489         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 490                 err = ceph_handle_notrace_create(dir, dentry);
 491
 492         if (d_in_lookup(dentry)) {
 493                 dn = ceph_finish_lookup(req, dentry, err);
 494                 if (IS_ERR(dn))
 495                         err = PTR_ERR(dn);
 496         } else {
 497                 /* we were given a hashed negative dentry */
 498                 dn = NULL;
 499         }
 500         if (err)
 501                 goto out_req;
 502         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 503                 /* make vfs retry on splice, ENOENT, or symlink */
 504                 dout("atomic_open finish_no_open on dn %p\n", dn);
 505                 err = finish_no_open(file, dn);
 506         } else {
 507                 dout("atomic_open finish_open on dn %p\n", dn);
 508                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 509                         ceph_init_inode_acls(d_inode(dentry), &acls);
 510                         file->f_mode |= FMODE_CREATED;
 511                 }
 512                 err = finish_open(file, dentry, ceph_open);
 513         }
 514 out_req:
 515         if (!req->r_err && req->r_target_inode)
 516                 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
 517         ceph_mdsc_put_request(req);
 518 out_acl:
 519         ceph_release_acls_info(&acls);
 520         dout("atomic_open result=%d\n", err);
 521         return err;
 522 }
 523
 524 int ceph_release(struct inode *inode, struct file *file)
 525 {
 526         struct ceph_inode_info *ci = ceph_inode(inode);
 527
 528         if (S_ISDIR(inode->i_mode)) {
 529                 struct ceph_dir_file_info *dfi = file->private_data;
 530                 dout("release inode %p dir file %p\n", inode, file);
 531                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 532
 533                 ceph_put_fmode(ci, dfi->file_info.fmode);
 534
 535                 if (dfi->last_readdir)
 536                         ceph_mdsc_put_request(dfi->last_readdir);
 537                 kfree(dfi->last_name);
 538                 kfree(dfi->dir_info);
 539                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 540         } else {
 541                 struct ceph_file_info *fi = file->private_data;
 542                 dout("release inode %p regular file %p\n", inode, file);
 543                 WARN_ON(!list_empty(&fi->rw_contexts));
 544
 545                 ceph_put_fmode(ci, fi->fmode);
 546                 kmem_cache_free(ceph_file_cachep, fi);
 547         }
 548
 549         /* wake up anyone waiting for caps on this inode */
 550         wake_up_all(&ci->i_cap_wq);
 551         return 0;
 552 }
 553
 554 enum {
 555         HAVE_RETRIED = 1,
 556         CHECK_EOF =    2,
 557         READ_INLINE =  3,
 558 };
 559
 560 /*
 561  * Completely synchronous read and write methods.  Direct from __user
 562  * buffer to osd, or directly to user pages (if O_DIRECT).
 563  *
 564  * If the read spans object boundary, just do multiple reads.  (That's not
 565  * atomic, but good enough for now.)
 566  *
 567  * If we get a short result from the OSD, check against i_size; we need to
 568  * only return a short read to the caller if we hit EOF.
 569  */
 570 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 571                               int *retry_op)
 572 {
 573         struct file *file = iocb->ki_filp;
 574         struct inode *inode = file_inode(file);
 575         struct ceph_inode_info *ci = ceph_inode(inode);
 576         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 577         struct ceph_osd_client *osdc = &fsc->client->osdc;
 578         ssize_t ret;
 579         u64 off = iocb->ki_pos;
 580         u64 len = iov_iter_count(to);
 581
 582         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 583              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 584
 585         if (!len)
 586                 return 0;
 587         /*
 588          * flush any page cache pages in this range.  this
 589          * will make concurrent normal and sync io slow,
 590          * but it will at least behave sensibly when they are
 591          * in sequence.
 592          */
 593         ret = filemap_write_and_wait_range(inode->i_mapping, off, off + len);
 594         if (ret < 0)
 595                 return ret;
 596
 597         ret = 0;
 598         while ((len = iov_iter_count(to)) > 0) {
 599                 struct ceph_osd_request *req;
 600                 struct page **pages;
 601                 int num_pages;
 602                 size_t page_off;
 603                 u64 i_size;
 604                 bool more;
 605
 606                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 607                                         ci->i_vino, off, &len, 0, 1,
 608                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 609                                         NULL, ci->i_truncate_seq,
 610                                         ci->i_truncate_size, false);
 611                 if (IS_ERR(req)) {
 612                         ret = PTR_ERR(req);
 613                         break;
 614                 }
 615
 616                 more = len < iov_iter_count(to);
 617
 618                 if (unlikely(iov_iter_is_pipe(to))) {
 619                         ret = iov_iter_get_pages_alloc(to, &pages, len,
 620                                                        &page_off);
 621                         if (ret <= 0) {
 622                                 ceph_osdc_put_request(req);
 623                                 ret = -ENOMEM;
 624                                 break;
 625                         }
 626                         num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
 627                         if (ret < len) {
 628                                 len = ret;
 629                                 osd_req_op_extent_update(req, 0, len);
 630                                 more = false;
 631                         }
 632                 } else {
 633                         num_pages = calc_pages_for(off, len);
 634                         page_off = off & ~PAGE_MASK;
 635                         pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 636                         if (IS_ERR(pages)) {
 637                                 ceph_osdc_put_request(req);
 638                                 ret = PTR_ERR(pages);
 639                                 break;
 640                         }
 641                 }
 642
 643                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 644                                                  false, false);
 645                 ret = ceph_osdc_start_request(osdc, req, false);
 646                 if (!ret)
 647                         ret = ceph_osdc_wait_request(osdc, req);
 648                 ceph_osdc_put_request(req);
 649
 650                 i_size = i_size_read(inode);
 651                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 652                      off, len, ret, i_size, (more ? " MORE" : ""));
 653
 654                 if (ret == -ENOENT)
 655                         ret = 0;
 656                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 657                         int zlen = min(len - ret, i_size - off - ret);
 658                         int zoff = page_off + ret;
 659                         dout("sync_read zero gap %llu~%llu\n",
 660                              off + ret, off + ret + zlen);
 661                         ceph_zero_page_vector_range(zoff, zlen, pages);
 662                         ret += zlen;
 663                 }
 664
 665                 if (unlikely(iov_iter_is_pipe(to))) {
 666                         if (ret > 0) {
 667                                 iov_iter_advance(to, ret);
 668                                 off += ret;
 669                         } else {
 670                                 iov_iter_advance(to, 0);
 671                         }
 672                         ceph_put_page_vector(pages, num_pages, false);
 673                 } else {
 674                         int idx = 0;
 675                         size_t left = ret > 0 ? ret : 0;
 676                         while (left > 0) {
 677                                 size_t len, copied;
 678                                 page_off = off & ~PAGE_MASK;
 679                                 len = min_t(size_t, left, PAGE_SIZE - page_off);
 680                                 copied = copy_page_to_iter(pages[idx++],
 681                                                            page_off, len, to);
 682                                 off += copied;
 683                                 left -= copied;
 684                                 if (copied < len) {
 685                                         ret = -EFAULT;
 686                                         break;
 687                                 }
 688                         }
 689                         ceph_release_page_vector(pages, num_pages);
 690                 }
 691
 692                 if (ret <= 0 || off >= i_size || !more)
 693                         break;
 694         }
 695
 696         if (off > iocb->ki_pos) {
 697                 if (ret >= 0 &&
 698                     iov_iter_count(to) > 0 && off >= i_size_read(inode))
 699                         *retry_op = CHECK_EOF;
 700                 ret = off - iocb->ki_pos;
 701                 iocb->ki_pos = off;
 702         }
 703
 704         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 705         return ret;
 706 }
 707
 708 struct ceph_aio_request {
 709         struct kiocb *iocb;
 710         size_t total_len;
 711         bool write;
 712         bool should_dirty;
 713         int error;
 714         struct list_head osd_reqs;
 715         unsigned num_reqs;
 716         atomic_t pending_reqs;
 717         struct timespec64 mtime;
 718         struct ceph_cap_flush *prealloc_cf;
 719 };
 720
 721 struct ceph_aio_work {
 722         struct work_struct work;
 723         struct ceph_osd_request *req;
 724 };
 725
 726 static void ceph_aio_retry_work(struct work_struct *work);
 727
 728 static void ceph_aio_complete(struct inode *inode,
 729                               struct ceph_aio_request *aio_req)
 730 {
 731         struct ceph_inode_info *ci = ceph_inode(inode);
 732         int ret;
 733
 734         if (!atomic_dec_and_test(&aio_req->pending_reqs))
 735                 return;
 736
 737         ret = aio_req->error;
 738         if (!ret)
 739                 ret = aio_req->total_len;
 740
 741         dout("ceph_aio_complete %p rc %d\n", inode, ret);
 742
 743         if (ret >= 0 && aio_req->write) {
 744                 int dirty;
 745
 746                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
 747                 if (endoff > i_size_read(inode)) {
 748                         if (ceph_inode_set_size(inode, endoff))
 749                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
 750                 }
 751
 752                 spin_lock(&ci->i_ceph_lock);
 753                 ci->i_inline_version = CEPH_INLINE_NONE;
 754                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
 755                                                &aio_req->prealloc_cf);
 756                 spin_unlock(&ci->i_ceph_lock);
 757                 if (dirty)
 758                         __mark_inode_dirty(inode, dirty);
 759
 760         }
 761
 762         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
 763                                                 CEPH_CAP_FILE_RD));
 764
 765         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
 766
 767         ceph_free_cap_flush(aio_req->prealloc_cf);
 768         kfree(aio_req);
 769 }
 770
 771 static void ceph_aio_complete_req(struct ceph_osd_request *req)
 772 {
 773         int rc = req->r_result;
 774         struct inode *inode = req->r_inode;
 775         struct ceph_aio_request *aio_req = req->r_priv;
 776         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 777
 778         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
 779         BUG_ON(!osd_data->num_bvecs);
 780
 781         dout("ceph_aio_complete_req %p rc %d bytes %u\n",
 782              inode, rc, osd_data->bvec_pos.iter.bi_size);
 783
 784         if (rc == -EOLDSNAPC) {
 785                 struct ceph_aio_work *aio_work;
 786                 BUG_ON(!aio_req->write);
 787
 788                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
 789                 if (aio_work) {
 790                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
 791                         aio_work->req = req;
 792                         queue_work(ceph_inode_to_client(inode)->wb_wq,
 793                                    &aio_work->work);
 794                         return;
 795                 }
 796                 rc = -ENOMEM;
 797         } else if (!aio_req->write) {
 798                 if (rc == -ENOENT)
 799                         rc = 0;
 800                 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
 801                         struct iov_iter i;
 802                         int zlen = osd_data->bvec_pos.iter.bi_size - rc;
 803
 804                         /*
 805                          * If read is satisfied by single OSD request,
 806                          * it can pass EOF. Otherwise read is within
 807                          * i_size.
 808                          */
 809                         if (aio_req->num_reqs == 1) {
 810                                 loff_t i_size = i_size_read(inode);
 811                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
 812                                 if (endoff < i_size)
 813                                         zlen = min_t(size_t, zlen,
 814                                                      i_size - endoff);
 815                                 aio_req->total_len = rc + zlen;
 816                         }
 817
 818                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
 819                                       osd_data->num_bvecs,
 820                                       osd_data->bvec_pos.iter.bi_size);
 821                         iov_iter_advance(&i, rc);
 822                         iov_iter_zero(zlen, &i);
 823                 }
 824         }
 825
 826         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
 827                   aio_req->should_dirty);
 828         ceph_osdc_put_request(req);
 829
 830         if (rc < 0)
 831                 cmpxchg(&aio_req->error, 0, rc);
 832
 833         ceph_aio_complete(inode, aio_req);
 834         return;
 835 }
 836
 837 static void ceph_aio_retry_work(struct work_struct *work)
 838 {
 839         struct ceph_aio_work *aio_work =
 840                 container_of(work, struct ceph_aio_work, work);
 841         struct ceph_osd_request *orig_req = aio_work->req;
 842         struct ceph_aio_request *aio_req = orig_req->r_priv;
 843         struct inode *inode = orig_req->r_inode;
 844         struct ceph_inode_info *ci = ceph_inode(inode);
 845         struct ceph_snap_context *snapc;
 846         struct ceph_osd_request *req;
 847         int ret;
 848
 849         spin_lock(&ci->i_ceph_lock);
 850         if (__ceph_have_pending_cap_snap(ci)) {
 851                 struct ceph_cap_snap *capsnap =
 852                         list_last_entry(&ci->i_cap_snaps,
 853                                         struct ceph_cap_snap,
 854                                         ci_item);
 855                 snapc = ceph_get_snap_context(capsnap->context);
 856         } else {
 857                 BUG_ON(!ci->i_head_snapc);
 858                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 859         }
 860         spin_unlock(&ci->i_ceph_lock);
 861
 862         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
 863                         false, GFP_NOFS);
 864         if (!req) {
 865                 ret = -ENOMEM;
 866                 req = orig_req;
 867                 goto out;
 868         }
 869
 870         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 871         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
 872         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
 873
 874         req->r_ops[0] = orig_req->r_ops[0];
 875
 876         req->r_mtime = aio_req->mtime;
 877         req->r_data_offset = req->r_ops[0].extent.offset;
 878
 879         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
 880         if (ret) {
 881                 ceph_osdc_put_request(req);
 882                 req = orig_req;
 883                 goto out;
 884         }
 885
 886         ceph_osdc_put_request(orig_req);
 887
 888         req->r_callback = ceph_aio_complete_req;
 889         req->r_inode = inode;
 890         req->r_priv = aio_req;
 891
 892         ret = ceph_osdc_start_request(req->r_osdc, req, false);
 893 out:
 894         if (ret < 0) {
 895                 req->r_result = ret;
 896                 ceph_aio_complete_req(req);
 897         }
 898
 899         ceph_put_snap_context(snapc);
 900         kfree(aio_work);
 901 }
 902
 903 static ssize_t
 904 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
 905                        struct ceph_snap_context *snapc,
 906                        struct ceph_cap_flush **pcf)
 907 {
 908         struct file *file = iocb->ki_filp;
 909         struct inode *inode = file_inode(file);
 910         struct ceph_inode_info *ci = ceph_inode(inode);
 911         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 912         struct ceph_vino vino;
 913         struct ceph_osd_request *req;
 914         struct bio_vec *bvecs;
 915         struct ceph_aio_request *aio_req = NULL;
 916         int num_pages = 0;
 917         int flags;
 918         int ret;
 919         struct timespec64 mtime = current_time(inode);
 920         size_t count = iov_iter_count(iter);
 921         loff_t pos = iocb->ki_pos;
 922         bool write = iov_iter_rw(iter) == WRITE;
 923         bool should_dirty = !write && iter_is_iovec(iter);
 924
 925         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
 926                 return -EROFS;
 927
 928         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
 929              (write ? "write" : "read"), file, pos, (unsigned)count,
 930              snapc, snapc->seq);
 931
 932         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
 933         if (ret < 0)
 934                 return ret;
 935
 936         if (write) {
 937                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
 938                                         pos >> PAGE_SHIFT,
 939                                         (pos + count) >> PAGE_SHIFT);
 940                 if (ret2 < 0)
 941                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
 942
 943                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 944         } else {
 945                 flags = CEPH_OSD_FLAG_READ;
 946         }
 947
 948         while (iov_iter_count(iter) > 0) {
 949                 u64 size = iov_iter_count(iter);
 950                 ssize_t len;
 951
 952                 if (write)
 953                         size = min_t(u64, size, fsc->mount_options->wsize);
 954                 else
 955                         size = min_t(u64, size, fsc->mount_options->rsize);
 956
 957                 vino = ceph_vino(inode);
 958                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 959                                             vino, pos, &size, 0,
 960                                             1,
 961                                             write ? CEPH_OSD_OP_WRITE :
 962                                                     CEPH_OSD_OP_READ,
 963                                             flags, snapc,
 964                                             ci->i_truncate_seq,
 965                                             ci->i_truncate_size,
 966                                             false);
 967                 if (IS_ERR(req)) {
 968                         ret = PTR_ERR(req);
 969                         break;
 970                 }
 971
 972                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
 973                 if (len < 0) {
 974                         ceph_osdc_put_request(req);
 975                         ret = len;
 976                         break;
 977                 }
 978                 if (len != size)
 979                         osd_req_op_extent_update(req, 0, len);
 980
 981                 /*
 982                  * To simplify error handling, allow AIO when IO within i_size
 983                  * or IO can be satisfied by single OSD request.
 984                  */
 985                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
 986                     (len == count || pos + count <= i_size_read(inode))) {
 987                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
 988                         if (aio_req) {
 989                                 aio_req->iocb = iocb;
 990                                 aio_req->write = write;
 991                                 aio_req->should_dirty = should_dirty;
 992                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
 993                                 if (write) {
 994                                         aio_req->mtime = mtime;
 995                                         swap(aio_req->prealloc_cf, *pcf);
 996                                 }
 997                         }
 998                         /* ignore error */
 999                 }
1000
1001                 if (write) {
1002                         /*
1003                          * throw out any page cache pages in this range. this
1004                          * may block.
1005                          */
1006                         truncate_inode_pages_range(inode->i_mapping, pos,
1007                                         (pos+len) | (PAGE_SIZE - 1));
1008
1009                         req->r_mtime = mtime;
1010                 }
1011
1012                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1013
1014                 if (aio_req) {
1015                         aio_req->total_len += len;
1016                         aio_req->num_reqs++;
1017                         atomic_inc(&aio_req->pending_reqs);
1018
1019                         req->r_callback = ceph_aio_complete_req;
1020                         req->r_inode = inode;
1021                         req->r_priv = aio_req;
1022                         list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
1023
1024                         pos += len;
1025                         continue;
1026                 }
1027
1028                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1029                 if (!ret)
1030                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1031
1032                 size = i_size_read(inode);
1033                 if (!write) {
1034                         if (ret == -ENOENT)
1035                                 ret = 0;
1036                         if (ret >= 0 && ret < len && pos + ret < size) {
1037                                 struct iov_iter i;
1038                                 int zlen = min_t(size_t, len - ret,
1039                                                  size - pos - ret);
1040
1041                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1042                                 iov_iter_advance(&i, ret);
1043                                 iov_iter_zero(zlen, &i);
1044                                 ret += zlen;
1045                         }
1046                         if (ret >= 0)
1047                                 len = ret;
1048                 }
1049
1050                 put_bvecs(bvecs, num_pages, should_dirty);
1051                 ceph_osdc_put_request(req);
1052                 if (ret < 0)
1053                         break;
1054
1055                 pos += len;
1056                 if (!write && pos >= size)
1057                         break;
1058
1059                 if (write && pos > size) {
1060                         if (ceph_inode_set_size(inode, pos))
1061                                 ceph_check_caps(ceph_inode(inode),
1062                                                 CHECK_CAPS_AUTHONLY,
1063                                                 NULL);
1064                 }
1065         }
1066
1067         if (aio_req) {
1068                 LIST_HEAD(osd_reqs);
1069
1070                 if (aio_req->num_reqs == 0) {
1071                         kfree(aio_req);
1072                         return ret;
1073                 }
1074
1075                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1076                                               CEPH_CAP_FILE_RD);
1077
1078                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1079                 while (!list_empty(&osd_reqs)) {
1080                         req = list_first_entry(&osd_reqs,
1081                                                struct ceph_osd_request,
1082                                                r_unsafe_item);
1083                         list_del_init(&req->r_unsafe_item);
1084                         if (ret >= 0)
1085                                 ret = ceph_osdc_start_request(req->r_osdc,
1086                                                               req, false);
1087                         if (ret < 0) {
1088                                 req->r_result = ret;
1089                                 ceph_aio_complete_req(req);
1090                         }
1091                 }
1092                 return -EIOCBQUEUED;
1093         }
1094
1095         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1096                 ret = pos - iocb->ki_pos;
1097                 iocb->ki_pos = pos;
1098         }
1099         return ret;
1100 }
1101
1102 /*
1103  * Synchronous write, straight from __user pointer or user pages.
1104  *
1105  * If write spans object boundary, just do multiple writes.  (For a
1106  * correct atomic write, we should e.g. take write locks on all
1107  * objects, rollback on failure, etc.)
1108  */
1109 static ssize_t
1110 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1111                 struct ceph_snap_context *snapc)
1112 {
1113         struct file *file = iocb->ki_filp;
1114         struct inode *inode = file_inode(file);
1115         struct ceph_inode_info *ci = ceph_inode(inode);
1116         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1117         struct ceph_vino vino;
1118         struct ceph_osd_request *req;
1119         struct page **pages;
1120         u64 len;
1121         int num_pages;
1122         int written = 0;
1123         int flags;
1124         int ret;
1125         bool check_caps = false;
1126         struct timespec64 mtime = current_time(inode);
1127         size_t count = iov_iter_count(from);
1128
1129         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1130                 return -EROFS;
1131
1132         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1133              file, pos, (unsigned)count, snapc, snapc->seq);
1134
1135         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1136         if (ret < 0)
1137                 return ret;
1138
1139         ret = invalidate_inode_pages2_range(inode->i_mapping,
1140                                             pos >> PAGE_SHIFT,
1141                                             (pos + count) >> PAGE_SHIFT);
1142         if (ret < 0)
1143                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1144
1145         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1146
1147         while ((len = iov_iter_count(from)) > 0) {
1148                 size_t left;
1149                 int n;
1150
1151                 vino = ceph_vino(inode);
1152                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1153                                             vino, pos, &len, 0, 1,
1154                                             CEPH_OSD_OP_WRITE, flags, snapc,
1155                                             ci->i_truncate_seq,
1156                                             ci->i_truncate_size,
1157                                             false);
1158                 if (IS_ERR(req)) {
1159                         ret = PTR_ERR(req);
1160                         break;
1161                 }
1162
1163                 /*
1164                  * write from beginning of first page,
1165                  * regardless of io alignment
1166                  */
1167                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1168
1169                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1170                 if (IS_ERR(pages)) {
1171                         ret = PTR_ERR(pages);
1172                         goto out;
1173                 }
1174
1175                 left = len;
1176                 for (n = 0; n < num_pages; n++) {
1177                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1178                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1179                         if (ret != plen) {
1180                                 ret = -EFAULT;
1181                                 break;
1182                         }
1183                         left -= ret;
1184                 }
1185
1186                 if (ret < 0) {
1187                         ceph_release_page_vector(pages, num_pages);
1188                         goto out;
1189                 }
1190
1191                 req->r_inode = inode;
1192
1193                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1194                                                 false, true);
1195
1196                 req->r_mtime = mtime;
1197                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1198                 if (!ret)
1199                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1200
1201 out:
1202                 ceph_osdc_put_request(req);
1203                 if (ret != 0) {
1204                         ceph_set_error_write(ci);
1205                         break;
1206                 }
1207
1208                 ceph_clear_error_write(ci);
1209                 pos += len;
1210                 written += len;
1211                 if (pos > i_size_read(inode)) {
1212                         check_caps = ceph_inode_set_size(inode, pos);
1213                         if (check_caps)
1214                                 ceph_check_caps(ceph_inode(inode),
1215                                                 CHECK_CAPS_AUTHONLY,
1216                                                 NULL);
1217                 }
1218
1219         }
1220
1221         if (ret != -EOLDSNAPC && written > 0) {
1222                 ret = written;
1223                 iocb->ki_pos = pos;
1224         }
1225         return ret;
1226 }
1227
1228 /*
1229  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1230  * Atomically grab references, so that those bits are not released
1231  * back to the MDS mid-read.
1232  *
1233  * Hmm, the sync read case isn't actually async... should it be?
1234  */
1235 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1236 {
1237         struct file *filp = iocb->ki_filp;
1238         struct ceph_file_info *fi = filp->private_data;
1239         size_t len = iov_iter_count(to);
1240         struct inode *inode = file_inode(filp);
1241         struct ceph_inode_info *ci = ceph_inode(inode);
1242         struct page *pinned_page = NULL;
1243         ssize_t ret;
1244         int want, got = 0;
1245         int retry_op = 0, read = 0;
1246
1247 again:
1248         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1249              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1250
1251         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1252                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1253         else
1254                 want = CEPH_CAP_FILE_CACHE;
1255         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1256         if (ret < 0)
1257                 return ret;
1258
1259         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1260             (iocb->ki_flags & IOCB_DIRECT) ||
1261             (fi->flags & CEPH_F_SYNC)) {
1262
1263                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1264                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1265                      ceph_cap_string(got));
1266
1267                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1268                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1269                                 ret = ceph_direct_read_write(iocb, to,
1270                                                              NULL, NULL);
1271                                 if (ret >= 0 && ret < len)
1272                                         retry_op = CHECK_EOF;
1273                         } else {
1274                                 ret = ceph_sync_read(iocb, to, &retry_op);
1275                         }
1276                 } else {
1277                         retry_op = READ_INLINE;
1278                 }
1279         } else {
1280                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1281                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1282                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1283                      ceph_cap_string(got));
1284                 ceph_add_rw_context(fi, &rw_ctx);
1285                 ret = generic_file_read_iter(iocb, to);
1286                 ceph_del_rw_context(fi, &rw_ctx);
1287         }
1288         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1289              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1290         if (pinned_page) {
1291                 put_page(pinned_page);
1292                 pinned_page = NULL;
1293         }
1294         ceph_put_cap_refs(ci, got);
1295         if (retry_op > HAVE_RETRIED && ret >= 0) {
1296                 int statret;
1297                 struct page *page = NULL;
1298                 loff_t i_size;
1299                 if (retry_op == READ_INLINE) {
1300                         page = __page_cache_alloc(GFP_KERNEL);
1301                         if (!page)
1302                                 return -ENOMEM;
1303                 }
1304
1305                 statret = __ceph_do_getattr(inode, page,
1306                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1307                 if (statret < 0) {
1308                         if (page)
1309                                 __free_page(page);
1310                         if (statret == -ENODATA) {
1311                                 BUG_ON(retry_op != READ_INLINE);
1312                                 goto again;
1313                         }
1314                         return statret;
1315                 }
1316
1317                 i_size = i_size_read(inode);
1318                 if (retry_op == READ_INLINE) {
1319                         BUG_ON(ret > 0 || read > 0);
1320                         if (iocb->ki_pos < i_size &&
1321                             iocb->ki_pos < PAGE_SIZE) {
1322                                 loff_t end = min_t(loff_t, i_size,
1323                                                    iocb->ki_pos + len);
1324                                 end = min_t(loff_t, end, PAGE_SIZE);
1325                                 if (statret < end)
1326                                         zero_user_segment(page, statret, end);
1327                                 ret = copy_page_to_iter(page,
1328                                                 iocb->ki_pos & ~PAGE_MASK,
1329                                                 end - iocb->ki_pos, to);
1330                                 iocb->ki_pos += ret;
1331                                 read += ret;
1332                         }
1333                         if (iocb->ki_pos < i_size && read < len) {
1334                                 size_t zlen = min_t(size_t, len - read,
1335                                                     i_size - iocb->ki_pos);
1336                                 ret = iov_iter_zero(zlen, to);
1337                                 iocb->ki_pos += ret;
1338                                 read += ret;
1339                         }
1340                         __free_pages(page, 0);
1341                         return read;
1342                 }
1343
1344                 /* hit EOF or hole? */
1345                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1346                     ret < len) {
1347                         dout("sync_read hit hole, ppos %lld < size %lld"
1348                              ", reading more\n", iocb->ki_pos, i_size);
1349
1350                         read += ret;
1351                         len -= ret;
1352                         retry_op = HAVE_RETRIED;
1353                         goto again;
1354                 }
1355         }
1356
1357         if (ret >= 0)
1358                 ret += read;
1359
1360         return ret;
1361 }
1362
1363 /*
1364  * Take cap references to avoid releasing caps to MDS mid-write.
1365  *
1366  * If we are synchronous, and write with an old snap context, the OSD
1367  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1368  * dropping our cap refs and allowing the pending snap to logically
1369  * complete _before_ this write occurs.
1370  *
1371  * If we are near ENOSPC, write synchronously.
1372  */
1373 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1374 {
1375         struct file *file = iocb->ki_filp;
1376         struct ceph_file_info *fi = file->private_data;
1377         struct inode *inode = file_inode(file);
1378         struct ceph_inode_info *ci = ceph_inode(inode);
1379         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1380         struct ceph_cap_flush *prealloc_cf;
1381         ssize_t count, written = 0;
1382         int err, want, got;
1383         loff_t pos;
1384         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1385
1386         if (ceph_snap(inode) != CEPH_NOSNAP)
1387                 return -EROFS;
1388
1389         prealloc_cf = ceph_alloc_cap_flush();
1390         if (!prealloc_cf)
1391                 return -ENOMEM;
1392
1393 retry_snap:
1394         inode_lock(inode);
1395
1396         /* We can write back this queue in page reclaim */
1397         current->backing_dev_info = inode_to_bdi(inode);
1398
1399         if (iocb->ki_flags & IOCB_APPEND) {
1400                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1401                 if (err < 0)
1402                         goto out;
1403         }
1404
1405         err = generic_write_checks(iocb, from);
1406         if (err <= 0)
1407                 goto out;
1408
1409         pos = iocb->ki_pos;
1410         if (unlikely(pos >= limit)) {
1411                 err = -EFBIG;
1412                 goto out;
1413         } else {
1414                 iov_iter_truncate(from, limit - pos);
1415         }
1416
1417         count = iov_iter_count(from);
1418         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1419                 err = -EDQUOT;
1420                 goto out;
1421         }
1422
1423         err = file_remove_privs(file);
1424         if (err)
1425                 goto out;
1426
1427         err = file_update_time(file);
1428         if (err)
1429                 goto out;
1430
1431         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1432                 err = ceph_uninline_data(file, NULL);
1433                 if (err < 0)
1434                         goto out;
1435         }
1436
1437         /* FIXME: not complete since it doesn't account for being at quota */
1438         if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1439                 err = -ENOSPC;
1440                 goto out;
1441         }
1442
1443         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1444              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1445         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1446                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1447         else
1448                 want = CEPH_CAP_FILE_BUFFER;
1449         got = 0;
1450         err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1451                             &got, NULL);
1452         if (err < 0)
1453                 goto out;
1454
1455         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1456              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1457
1458         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1459             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1460             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1461                 struct ceph_snap_context *snapc;
1462                 struct iov_iter data;
1463                 inode_unlock(inode);
1464
1465                 spin_lock(&ci->i_ceph_lock);
1466                 if (__ceph_have_pending_cap_snap(ci)) {
1467                         struct ceph_cap_snap *capsnap =
1468                                         list_last_entry(&ci->i_cap_snaps,
1469                                                         struct ceph_cap_snap,
1470                                                         ci_item);
1471                         snapc = ceph_get_snap_context(capsnap->context);
1472                 } else {
1473                         BUG_ON(!ci->i_head_snapc);
1474                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1475                 }
1476                 spin_unlock(&ci->i_ceph_lock);
1477
1478                 /* we might need to revert back to that point */
1479                 data = *from;
1480                 if (iocb->ki_flags & IOCB_DIRECT)
1481                         written = ceph_direct_read_write(iocb, &data, snapc,
1482                                                          &prealloc_cf);
1483                 else
1484                         written = ceph_sync_write(iocb, &data, pos, snapc);
1485                 if (written > 0)
1486                         iov_iter_advance(from, written);
1487                 ceph_put_snap_context(snapc);
1488         } else {
1489                 /*
1490                  * No need to acquire the i_truncate_mutex. Because
1491                  * the MDS revokes Fwb caps before sending truncate
1492                  * message to us. We can't get Fwb cap while there
1493                  * are pending vmtruncate. So write and vmtruncate
1494                  * can not run at the same time
1495                  */
1496                 written = generic_perform_write(file, from, pos);
1497                 if (likely(written >= 0))
1498                         iocb->ki_pos = pos + written;
1499                 inode_unlock(inode);
1500         }
1501
1502         if (written >= 0) {
1503                 int dirty;
1504
1505                 spin_lock(&ci->i_ceph_lock);
1506                 ci->i_inline_version = CEPH_INLINE_NONE;
1507                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1508                                                &prealloc_cf);
1509                 spin_unlock(&ci->i_ceph_lock);
1510                 if (dirty)
1511                         __mark_inode_dirty(inode, dirty);
1512                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1513                         ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1514         }
1515
1516         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1517              inode, ceph_vinop(inode), pos, (unsigned)count,
1518              ceph_cap_string(got));
1519         ceph_put_cap_refs(ci, got);
1520
1521         if (written == -EOLDSNAPC) {
1522                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1523                      inode, ceph_vinop(inode), pos, (unsigned)count);
1524                 goto retry_snap;
1525         }
1526
1527         if (written >= 0) {
1528                 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1529                         iocb->ki_flags |= IOCB_DSYNC;
1530                 written = generic_write_sync(iocb, written);
1531         }
1532
1533         goto out_unlocked;
1534
1535 out:
1536         inode_unlock(inode);
1537 out_unlocked:
1538         ceph_free_cap_flush(prealloc_cf);
1539         current->backing_dev_info = NULL;
1540         return written ? written : err;
1541 }
1542
1543 /*
1544  * llseek.  be sure to verify file size on SEEK_END.
1545  */
1546 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1547 {
1548         struct inode *inode = file->f_mapping->host;
1549         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1550         loff_t i_size;
1551         loff_t ret;
1552
1553         inode_lock(inode);
1554
1555         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1556                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1557                 if (ret < 0)
1558                         goto out;
1559         }
1560
1561         i_size = i_size_read(inode);
1562         switch (whence) {
1563         case SEEK_END:
1564                 offset += i_size;
1565                 break;
1566         case SEEK_CUR:
1567                 /*
1568                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1569                  * position-querying operation.  Avoid rewriting the "same"
1570                  * f_pos value back to the file because a concurrent read(),
1571                  * write() or lseek() might have altered it
1572                  */
1573                 if (offset == 0) {
1574                         ret = file->f_pos;
1575                         goto out;
1576                 }
1577                 offset += file->f_pos;
1578                 break;
1579         case SEEK_DATA:
1580                 if (offset < 0 || offset >= i_size) {
1581                         ret = -ENXIO;
1582                         goto out;
1583                 }
1584                 break;
1585         case SEEK_HOLE:
1586                 if (offset < 0 || offset >= i_size) {
1587                         ret = -ENXIO;
1588                         goto out;
1589                 }
1590                 offset = i_size;
1591                 break;
1592         }
1593
1594         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1595
1596 out:
1597         inode_unlock(inode);
1598         return ret;
1599 }
1600
1601 static inline void ceph_zero_partial_page(
1602         struct inode *inode, loff_t offset, unsigned size)
1603 {
1604         struct page *page;
1605         pgoff_t index = offset >> PAGE_SHIFT;
1606
1607         page = find_lock_page(inode->i_mapping, index);
1608         if (page) {
1609                 wait_on_page_writeback(page);
1610                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1611                 unlock_page(page);
1612                 put_page(page);
1613         }
1614 }
1615
1616 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1617                                       loff_t length)
1618 {
1619         loff_t nearly = round_up(offset, PAGE_SIZE);
1620         if (offset < nearly) {
1621                 loff_t size = nearly - offset;
1622                 if (length < size)
1623                         size = length;
1624                 ceph_zero_partial_page(inode, offset, size);
1625                 offset += size;
1626                 length -= size;
1627         }
1628         if (length >= PAGE_SIZE) {
1629                 loff_t size = round_down(length, PAGE_SIZE);
1630                 truncate_pagecache_range(inode, offset, offset + size - 1);
1631                 offset += size;
1632                 length -= size;
1633         }
1634         if (length)
1635                 ceph_zero_partial_page(inode, offset, length);
1636 }
1637
1638 static int ceph_zero_partial_object(struct inode *inode,
1639                                     loff_t offset, loff_t *length)
1640 {
1641         struct ceph_inode_info *ci = ceph_inode(inode);
1642         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1643         struct ceph_osd_request *req;
1644         int ret = 0;
1645         loff_t zero = 0;
1646         int op;
1647
1648         if (!length) {
1649                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1650                 length = &zero;
1651         } else {
1652                 op = CEPH_OSD_OP_ZERO;
1653         }
1654
1655         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1656                                         ceph_vino(inode),
1657                                         offset, length,
1658                                         0, 1, op,
1659                                         CEPH_OSD_FLAG_WRITE,
1660                                         NULL, 0, 0, false);
1661         if (IS_ERR(req)) {
1662                 ret = PTR_ERR(req);
1663                 goto out;
1664         }
1665
1666         req->r_mtime = inode->i_mtime;
1667         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1668         if (!ret) {
1669                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1670                 if (ret == -ENOENT)
1671                         ret = 0;
1672         }
1673         ceph_osdc_put_request(req);
1674
1675 out:
1676         return ret;
1677 }
1678
1679 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1680 {
1681         int ret = 0;
1682         struct ceph_inode_info *ci = ceph_inode(inode);
1683         s32 stripe_unit = ci->i_layout.stripe_unit;
1684         s32 stripe_count = ci->i_layout.stripe_count;
1685         s32 object_size = ci->i_layout.object_size;
1686         u64 object_set_size = object_size * stripe_count;
1687         u64 nearly, t;
1688
1689         /* round offset up to next period boundary */
1690         nearly = offset + object_set_size - 1;
1691         t = nearly;
1692         nearly -= do_div(t, object_set_size);
1693
1694         while (length && offset < nearly) {
1695                 loff_t size = length;
1696                 ret = ceph_zero_partial_object(inode, offset, &size);
1697                 if (ret < 0)
1698                         return ret;
1699                 offset += size;
1700                 length -= size;
1701         }
1702         while (length >= object_set_size) {
1703                 int i;
1704                 loff_t pos = offset;
1705                 for (i = 0; i < stripe_count; ++i) {
1706                         ret = ceph_zero_partial_object(inode, pos, NULL);
1707                         if (ret < 0)
1708                                 return ret;
1709                         pos += stripe_unit;
1710                 }
1711                 offset += object_set_size;
1712                 length -= object_set_size;
1713         }
1714         while (length) {
1715                 loff_t size = length;
1716                 ret = ceph_zero_partial_object(inode, offset, &size);
1717                 if (ret < 0)
1718                         return ret;
1719                 offset += size;
1720                 length -= size;
1721         }
1722         return ret;
1723 }
1724
1725 static long ceph_fallocate(struct file *file, int mode,
1726                                 loff_t offset, loff_t length)
1727 {
1728         struct ceph_file_info *fi = file->private_data;
1729         struct inode *inode = file_inode(file);
1730         struct ceph_inode_info *ci = ceph_inode(inode);
1731         struct ceph_cap_flush *prealloc_cf;
1732         int want, got = 0;
1733         int dirty;
1734         int ret = 0;
1735         loff_t endoff = 0;
1736         loff_t size;
1737
1738         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1739                 return -EOPNOTSUPP;
1740
1741         if (!S_ISREG(inode->i_mode))
1742                 return -EOPNOTSUPP;
1743
1744         prealloc_cf = ceph_alloc_cap_flush();
1745         if (!prealloc_cf)
1746                 return -ENOMEM;
1747
1748         inode_lock(inode);
1749
1750         if (ceph_snap(inode) != CEPH_NOSNAP) {
1751                 ret = -EROFS;
1752                 goto unlock;
1753         }
1754
1755         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1756                 ret = ceph_uninline_data(file, NULL);
1757                 if (ret < 0)
1758                         goto unlock;
1759         }
1760
1761         size = i_size_read(inode);
1762
1763         /* Are we punching a hole beyond EOF? */
1764         if (offset >= size)
1765                 goto unlock;
1766         if ((offset + length) > size)
1767                 length = size - offset;
1768
1769         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1770                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1771         else
1772                 want = CEPH_CAP_FILE_BUFFER;
1773
1774         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1775         if (ret < 0)
1776                 goto unlock;
1777
1778         ceph_zero_pagecache_range(inode, offset, length);
1779         ret = ceph_zero_objects(inode, offset, length);
1780
1781         if (!ret) {
1782                 spin_lock(&ci->i_ceph_lock);
1783                 ci->i_inline_version = CEPH_INLINE_NONE;
1784                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1785                                                &prealloc_cf);
1786                 spin_unlock(&ci->i_ceph_lock);
1787                 if (dirty)
1788                         __mark_inode_dirty(inode, dirty);
1789         }
1790
1791         ceph_put_cap_refs(ci, got);
1792 unlock:
1793         inode_unlock(inode);
1794         ceph_free_cap_flush(prealloc_cf);
1795         return ret;
1796 }
1797
1798 /*
1799  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
1800  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
1801  * this fails; zero is returned on success.
1802  */
1803 static int get_rd_wr_caps(struct ceph_inode_info *src_ci,
1804                           loff_t src_endoff, int *src_got,
1805                           struct ceph_inode_info *dst_ci,
1806                           loff_t dst_endoff, int *dst_got)
1807 {
1808         int ret = 0;
1809         bool retrying = false;
1810
1811 retry_caps:
1812         ret = ceph_get_caps(dst_ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
1813                             dst_endoff, dst_got, NULL);
1814         if (ret < 0)
1815                 return ret;
1816
1817         /*
1818          * Since we're already holding the FILE_WR capability for the dst file,
1819          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
1820          * retry dance instead to try to get both capabilities.
1821          */
1822         ret = ceph_try_get_caps(src_ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
1823                                 false, src_got);
1824         if (ret <= 0) {
1825                 /* Start by dropping dst_ci caps and getting src_ci caps */
1826                 ceph_put_cap_refs(dst_ci, *dst_got);
1827                 if (retrying) {
1828                         if (!ret)
1829                                 /* ceph_try_get_caps masks EAGAIN */
1830                                 ret = -EAGAIN;
1831                         return ret;
1832                 }
1833                 ret = ceph_get_caps(src_ci, CEPH_CAP_FILE_RD,
1834                                     CEPH_CAP_FILE_SHARED, src_endoff,
1835                                     src_got, NULL);
1836                 if (ret < 0)
1837                         return ret;
1838                 /*... drop src_ci caps too, and retry */
1839                 ceph_put_cap_refs(src_ci, *src_got);
1840                 retrying = true;
1841                 goto retry_caps;
1842         }
1843         return ret;
1844 }
1845
1846 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
1847                            struct ceph_inode_info *dst_ci, int dst_got)
1848 {
1849         ceph_put_cap_refs(src_ci, src_got);
1850         ceph_put_cap_refs(dst_ci, dst_got);
1851 }
1852
1853 /*
1854  * This function does several size-related checks, returning an error if:
1855  *  - source file is smaller than off+len
1856  *  - destination file size is not OK (inode_newsize_ok())
1857  *  - max bytes quotas is exceeded
1858  */
1859 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
1860                            loff_t src_off, loff_t dst_off, size_t len)
1861 {
1862         loff_t size, endoff;
1863
1864         size = i_size_read(src_inode);
1865         /*
1866          * Don't copy beyond source file EOF.  Instead of simply setting length
1867          * to (size - src_off), just drop to VFS default implementation, as the
1868          * local i_size may be stale due to other clients writing to the source
1869          * inode.
1870          */
1871         if (src_off + len > size) {
1872                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
1873                      src_off, len, size);
1874                 return -EOPNOTSUPP;
1875         }
1876         size = i_size_read(dst_inode);
1877
1878         endoff = dst_off + len;
1879         if (inode_newsize_ok(dst_inode, endoff))
1880                 return -EOPNOTSUPP;
1881
1882         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
1883                 return -EDQUOT;
1884
1885         return 0;
1886 }
1887
1888 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
1889                                     struct file *dst_file, loff_t dst_off,
1890                                     size_t len, unsigned int flags)
1891 {
1892         struct inode *src_inode = file_inode(src_file);
1893         struct inode *dst_inode = file_inode(dst_file);
1894         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
1895         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
1896         struct ceph_cap_flush *prealloc_cf;
1897         struct ceph_object_locator src_oloc, dst_oloc;
1898         struct ceph_object_id src_oid, dst_oid;
1899         loff_t endoff = 0, size;
1900         ssize_t ret = -EIO;
1901         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
1902         u32 src_objlen, dst_objlen, object_size;
1903         int src_got = 0, dst_got = 0, err, dirty;
1904         bool do_final_copy = false;
1905
1906         if (src_inode == dst_inode)
1907                 return -EINVAL;
1908         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
1909                 return -EROFS;
1910
1911         /*
1912          * Some of the checks below will return -EOPNOTSUPP, which will force a
1913          * fallback to the default VFS copy_file_range implementation.  This is
1914          * desirable in several cases (for ex, the 'len' is smaller than the
1915          * size of the objects, or in cases where that would be more
1916          * efficient).
1917          */
1918
1919         if (ceph_test_mount_opt(ceph_inode_to_client(src_inode), NOCOPYFROM))
1920                 return -EOPNOTSUPP;
1921
1922         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
1923             (src_ci->i_layout.stripe_count != dst_ci->i_layout.stripe_count) ||
1924             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size))
1925                 return -EOPNOTSUPP;
1926
1927         if (len < src_ci->i_layout.object_size)
1928                 return -EOPNOTSUPP; /* no remote copy will be done */
1929
1930         prealloc_cf = ceph_alloc_cap_flush();
1931         if (!prealloc_cf)
1932                 return -ENOMEM;
1933
1934         /* Start by sync'ing the source and destination files */
1935         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
1936         if (ret < 0) {
1937                 dout("failed to write src file (%zd)\n", ret);
1938                 goto out;
1939         }
1940         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
1941         if (ret < 0) {
1942                 dout("failed to write dst file (%zd)\n", ret);
1943                 goto out;
1944         }
1945
1946         /*
1947          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
1948          * clients may have dirty data in their caches.  And OSDs know nothing
1949          * about caps, so they can't safely do the remote object copies.
1950          */
1951         err = get_rd_wr_caps(src_ci, (src_off + len), &src_got,
1952                              dst_ci, (dst_off + len), &dst_got);
1953         if (err < 0) {
1954                 dout("get_rd_wr_caps returned %d\n", err);
1955                 ret = -EOPNOTSUPP;
1956                 goto out;
1957         }
1958
1959         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
1960         if (ret < 0)
1961                 goto out_caps;
1962
1963         size = i_size_read(dst_inode);
1964         endoff = dst_off + len;
1965
1966         /* Drop dst file cached pages */
1967         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
1968                                             dst_off >> PAGE_SHIFT,
1969                                             endoff >> PAGE_SHIFT);
1970         if (ret < 0) {
1971                 dout("Failed to invalidate inode pages (%zd)\n", ret);
1972                 ret = 0; /* XXX */
1973         }
1974         src_oloc.pool = src_ci->i_layout.pool_id;
1975         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
1976         dst_oloc.pool = dst_ci->i_layout.pool_id;
1977         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
1978
1979         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
1980                                       src_ci->i_layout.object_size,
1981                                       &src_objnum, &src_objoff, &src_objlen);
1982         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
1983                                       dst_ci->i_layout.object_size,
1984                                       &dst_objnum, &dst_objoff, &dst_objlen);
1985         /* object-level offsets need to the same */
1986         if (src_objoff != dst_objoff) {
1987                 ret = -EOPNOTSUPP;
1988                 goto out_caps;
1989         }
1990
1991         /*
1992          * Do a manual copy if the object offset isn't object aligned.
1993          * 'src_objlen' contains the bytes left until the end of the object,
1994          * starting at the src_off
1995          */
1996         if (src_objoff) {
1997                 /*
1998                  * we need to temporarily drop all caps as we'll be calling
1999                  * {read,write}_iter, which will get caps again.
2000                  */
2001                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2002                 ret = do_splice_direct(src_file, &src_off, dst_file,
2003                                        &dst_off, src_objlen, flags);
2004                 if (ret < 0) {
2005                         dout("do_splice_direct returned %d\n", err);
2006                         goto out;
2007                 }
2008                 len -= ret;
2009                 err = get_rd_wr_caps(src_ci, (src_off + len),
2010                                      &src_got, dst_ci,
2011                                      (dst_off + len), &dst_got);
2012                 if (err < 0)
2013                         goto out;
2014                 err = is_file_size_ok(src_inode, dst_inode,
2015                                       src_off, dst_off, len);
2016                 if (err < 0)
2017                         goto out_caps;
2018         }
2019         object_size = src_ci->i_layout.object_size;
2020         while (len >= object_size) {
2021                 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2022                                               object_size, &src_objnum,
2023                                               &src_objoff, &src_objlen);
2024                 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2025                                               object_size, &dst_objnum,
2026                                               &dst_objoff, &dst_objlen);
2027                 ceph_oid_init(&src_oid);
2028                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2029                                 src_ci->i_vino.ino, src_objnum);
2030                 ceph_oid_init(&dst_oid);
2031                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2032                                 dst_ci->i_vino.ino, dst_objnum);
2033                 /* Do an object remote copy */
2034                 err = ceph_osdc_copy_from(
2035                         &ceph_inode_to_client(src_inode)->client->osdc,
2036                         src_ci->i_vino.snap, 0,
2037                         &src_oid, &src_oloc,
2038                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2039                         CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2040                         &dst_oid, &dst_oloc,
2041                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2042                         CEPH_OSD_OP_FLAG_FADVISE_DONTNEED, 0);
2043                 if (err) {
2044                         dout("ceph_osdc_copy_from returned %d\n", err);
2045                         if (!ret)
2046                                 ret = err;
2047                         goto out_caps;
2048                 }
2049                 len -= object_size;
2050                 src_off += object_size;
2051                 dst_off += object_size;
2052                 ret += object_size;
2053         }
2054
2055         if (len)
2056                 /* We still need one final local copy */
2057                 do_final_copy = true;
2058
2059         file_update_time(dst_file);
2060         if (endoff > size) {
2061                 int caps_flags = 0;
2062
2063                 /* Let the MDS know about dst file size change */
2064                 if (ceph_quota_is_max_bytes_approaching(dst_inode, endoff))
2065                         caps_flags |= CHECK_CAPS_NODELAY;
2066                 if (ceph_inode_set_size(dst_inode, endoff))
2067                         caps_flags |= CHECK_CAPS_AUTHONLY;
2068                 if (caps_flags)
2069                         ceph_check_caps(dst_ci, caps_flags, NULL);
2070         }
2071         /* Mark Fw dirty */
2072         spin_lock(&dst_ci->i_ceph_lock);
2073         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2074         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2075         spin_unlock(&dst_ci->i_ceph_lock);
2076         if (dirty)
2077                 __mark_inode_dirty(dst_inode, dirty);
2078
2079 out_caps:
2080         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2081
2082         if (do_final_copy) {
2083                 err = do_splice_direct(src_file, &src_off, dst_file,
2084                                        &dst_off, len, flags);
2085                 if (err < 0) {
2086                         dout("do_splice_direct returned %d\n", err);
2087                         goto out;
2088                 }
2089                 len -= err;
2090                 ret += err;
2091         }
2092
2093 out:
2094         ceph_free_cap_flush(prealloc_cf);
2095
2096         return ret;
2097 }
2098
2099 const struct file_operations ceph_file_fops = {
2100         .open = ceph_open,
2101         .release = ceph_release,
2102         .llseek = ceph_llseek,
2103         .read_iter = ceph_read_iter,
2104         .write_iter = ceph_write_iter,
2105         .mmap = ceph_mmap,
2106         .fsync = ceph_fsync,
2107         .lock = ceph_lock,
2108         .flock = ceph_flock,
2109         .splice_read = generic_file_splice_read,
2110         .splice_write = iter_file_splice_write,
2111         .unlocked_ioctl = ceph_ioctl,
2112         .compat_ioctl   = ceph_ioctl,
2113         .fallocate      = ceph_fallocate,
2114         .copy_file_range = ceph_copy_file_range,
2115 };