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