fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "io.h"
  20 #include "metric.h"
  21
  22 static __le32 ceph_flags_sys2wire(u32 flags)
  23 {
  24         u32 wire_flags = 0;
  25
  26         switch (flags & O_ACCMODE) {
  27         case O_RDONLY:
  28                 wire_flags |= CEPH_O_RDONLY;
  29                 break;
  30         case O_WRONLY:
  31                 wire_flags |= CEPH_O_WRONLY;
  32                 break;
  33         case O_RDWR:
  34                 wire_flags |= CEPH_O_RDWR;
  35                 break;
  36         }
  37
  38         flags &= ~O_ACCMODE;
  39
  40 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  41
  42         ceph_sys2wire(O_CREAT);
  43         ceph_sys2wire(O_EXCL);
  44         ceph_sys2wire(O_TRUNC);
  45         ceph_sys2wire(O_DIRECTORY);
  46         ceph_sys2wire(O_NOFOLLOW);
  47
  48 #undef ceph_sys2wire
  49
  50         if (flags)
  51                 dout("unused open flags: %x\n", flags);
  52
  53         return cpu_to_le32(wire_flags);
  54 }
  55
  56 /*
  57  * Ceph file operations
  58  *
  59  * Implement basic open/close functionality, and implement
  60  * read/write.
  61  *
  62  * We implement three modes of file I/O:
  63  *  - buffered uses the generic_file_aio_{read,write} helpers
  64  *
  65  *  - synchronous is used when there is multi-client read/write
  66  *    sharing, avoids the page cache, and synchronously waits for an
  67  *    ack from the OSD.
  68  *
  69  *  - direct io takes the variant of the sync path that references
  70  *    user pages directly.
  71  *
  72  * fsync() flushes and waits on dirty pages, but just queues metadata
  73  * for writeback: since the MDS can recover size and mtime there is no
  74  * need to wait for MDS acknowledgement.
  75  */
  76
  77 /*
  78  * How many pages to get in one call to iov_iter_get_pages().  This
  79  * determines the size of the on-stack array used as a buffer.
  80  */
  81 #define ITER_GET_BVECS_PAGES    64
  82
  83 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  84                                 struct bio_vec *bvecs)
  85 {
  86         size_t size = 0;
  87         int bvec_idx = 0;
  88
  89         if (maxsize > iov_iter_count(iter))
  90                 maxsize = iov_iter_count(iter);
  91
  92         while (size < maxsize) {
  93                 struct page *pages[ITER_GET_BVECS_PAGES];
  94                 ssize_t bytes;
  95                 size_t start;
  96                 int idx = 0;
  97
  98                 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
  99                                            ITER_GET_BVECS_PAGES, &start);
 100                 if (bytes < 0)
 101                         return size ?: bytes;
 102
 103                 size += bytes;
 104
 105                 for ( ; bytes; idx++, bvec_idx++) {
 106                         int len = min_t(int, bytes, PAGE_SIZE - start);
 107
 108                         bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
 109                         bytes -= len;
 110                         start = 0;
 111                 }
 112         }
 113
 114         return size;
 115 }
 116
 117 /*
 118  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 119  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 120  * page.
 121  *
 122  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 123  * Return the number of bytes in the created bio_vec array, or an error.
 124  */
 125 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 126                                     struct bio_vec **bvecs, int *num_bvecs)
 127 {
 128         struct bio_vec *bv;
 129         size_t orig_count = iov_iter_count(iter);
 130         ssize_t bytes;
 131         int npages;
 132
 133         iov_iter_truncate(iter, maxsize);
 134         npages = iov_iter_npages(iter, INT_MAX);
 135         iov_iter_reexpand(iter, orig_count);
 136
 137         /*
 138          * __iter_get_bvecs() may populate only part of the array -- zero it
 139          * out.
 140          */
 141         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 142         if (!bv)
 143                 return -ENOMEM;
 144
 145         bytes = __iter_get_bvecs(iter, maxsize, bv);
 146         if (bytes < 0) {
 147                 /*
 148                  * No pages were pinned -- just free the array.
 149                  */
 150                 kvfree(bv);
 151                 return bytes;
 152         }
 153
 154         *bvecs = bv;
 155         *num_bvecs = npages;
 156         return bytes;
 157 }
 158
 159 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 160 {
 161         int i;
 162
 163         for (i = 0; i < num_bvecs; i++) {
 164                 if (bvecs[i].bv_page) {
 165                         if (should_dirty)
 166                                 set_page_dirty_lock(bvecs[i].bv_page);
 167                         put_page(bvecs[i].bv_page);
 168                 }
 169         }
 170         kvfree(bvecs);
 171 }
 172
 173 /*
 174  * Prepare an open request.  Preallocate ceph_cap to avoid an
 175  * inopportune ENOMEM later.
 176  */
 177 static struct ceph_mds_request *
 178 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 179 {
 180         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 181         struct ceph_mds_request *req;
 182         int want_auth = USE_ANY_MDS;
 183         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 184
 185         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 186                 want_auth = USE_AUTH_MDS;
 187
 188         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 189         if (IS_ERR(req))
 190                 goto out;
 191         req->r_fmode = ceph_flags_to_mode(flags);
 192         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 193         req->r_args.open.mode = cpu_to_le32(create_mode);
 194 out:
 195         return req;
 196 }
 197
 198 static int ceph_init_file_info(struct inode *inode, struct file *file,
 199                                         int fmode, bool isdir)
 200 {
 201         struct ceph_inode_info *ci = ceph_inode(inode);
 202         struct ceph_mount_options *opt =
 203                 ceph_inode_to_client(&ci->netfs.inode)->mount_options;
 204         struct ceph_file_info *fi;
 205         int ret;
 206
 207         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 208                         inode->i_mode, isdir ? "dir" : "regular");
 209         BUG_ON(inode->i_fop->release != ceph_release);
 210
 211         if (isdir) {
 212                 struct ceph_dir_file_info *dfi =
 213                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 214                 if (!dfi)
 215                         return -ENOMEM;
 216
 217                 file->private_data = dfi;
 218                 fi = &dfi->file_info;
 219                 dfi->next_offset = 2;
 220                 dfi->readdir_cache_idx = -1;
 221         } else {
 222                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 223                 if (!fi)
 224                         return -ENOMEM;
 225
 226                 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
 227                         fi->flags |= CEPH_F_SYNC;
 228
 229                 file->private_data = fi;
 230         }
 231
 232         ceph_get_fmode(ci, fmode, 1);
 233         fi->fmode = fmode;
 234
 235         spin_lock_init(&fi->rw_contexts_lock);
 236         INIT_LIST_HEAD(&fi->rw_contexts);
 237         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 238
 239         if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
 240                 ret = ceph_uninline_data(file);
 241                 if (ret < 0)
 242                         goto error;
 243         }
 244
 245         return 0;
 246
 247 error:
 248         ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 249         ceph_put_fmode(ci, fi->fmode, 1);
 250         kmem_cache_free(ceph_file_cachep, fi);
 251         /* wake up anyone waiting for caps on this inode */
 252         wake_up_all(&ci->i_cap_wq);
 253         return ret;
 254 }
 255
 256 /*
 257  * initialize private struct file data.
 258  * if we fail, clean up by dropping fmode reference on the ceph_inode
 259  */
 260 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 261 {
 262         int ret = 0;
 263
 264         switch (inode->i_mode & S_IFMT) {
 265         case S_IFREG:
 266                 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
 267                 fallthrough;
 268         case S_IFDIR:
 269                 ret = ceph_init_file_info(inode, file, fmode,
 270                                                 S_ISDIR(inode->i_mode));
 271                 break;
 272
 273         case S_IFLNK:
 274                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 275                      inode->i_mode);
 276                 break;
 277
 278         default:
 279                 dout("init_file %p %p 0%o (special)\n", inode, file,
 280                      inode->i_mode);
 281                 /*
 282                  * we need to drop the open ref now, since we don't
 283                  * have .release set to ceph_release.
 284                  */
 285                 BUG_ON(inode->i_fop->release == ceph_release);
 286
 287                 /* call the proper open fop */
 288                 ret = inode->i_fop->open(inode, file);
 289         }
 290         return ret;
 291 }
 292
 293 /*
 294  * try renew caps after session gets killed.
 295  */
 296 int ceph_renew_caps(struct inode *inode, int fmode)
 297 {
 298         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 299         struct ceph_inode_info *ci = ceph_inode(inode);
 300         struct ceph_mds_request *req;
 301         int err, flags, wanted;
 302
 303         spin_lock(&ci->i_ceph_lock);
 304         __ceph_touch_fmode(ci, mdsc, fmode);
 305         wanted = __ceph_caps_file_wanted(ci);
 306         if (__ceph_is_any_real_caps(ci) &&
 307             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 308                 int issued = __ceph_caps_issued(ci, NULL);
 309                 spin_unlock(&ci->i_ceph_lock);
 310                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 311                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 312                 ceph_check_caps(ci, 0);
 313                 return 0;
 314         }
 315         spin_unlock(&ci->i_ceph_lock);
 316
 317         flags = 0;
 318         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 319                 flags = O_RDWR;
 320         else if (wanted & CEPH_CAP_FILE_RD)
 321                 flags = O_RDONLY;
 322         else if (wanted & CEPH_CAP_FILE_WR)
 323                 flags = O_WRONLY;
 324 #ifdef O_LAZY
 325         if (wanted & CEPH_CAP_FILE_LAZYIO)
 326                 flags |= O_LAZY;
 327 #endif
 328
 329         req = prepare_open_request(inode->i_sb, flags, 0);
 330         if (IS_ERR(req)) {
 331                 err = PTR_ERR(req);
 332                 goto out;
 333         }
 334
 335         req->r_inode = inode;
 336         ihold(inode);
 337         req->r_num_caps = 1;
 338
 339         err = ceph_mdsc_do_request(mdsc, NULL, req);
 340         ceph_mdsc_put_request(req);
 341 out:
 342         dout("renew caps %p open result=%d\n", inode, err);
 343         return err < 0 ? err : 0;
 344 }
 345
 346 /*
 347  * If we already have the requisite capabilities, we can satisfy
 348  * the open request locally (no need to request new caps from the
 349  * MDS).  We do, however, need to inform the MDS (asynchronously)
 350  * if our wanted caps set expands.
 351  */
 352 int ceph_open(struct inode *inode, struct file *file)
 353 {
 354         struct ceph_inode_info *ci = ceph_inode(inode);
 355         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 356         struct ceph_mds_client *mdsc = fsc->mdsc;
 357         struct ceph_mds_request *req;
 358         struct ceph_file_info *fi = file->private_data;
 359         int err;
 360         int flags, fmode, wanted;
 361
 362         if (fi) {
 363                 dout("open file %p is already opened\n", file);
 364                 return 0;
 365         }
 366
 367         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 368         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 369         if (S_ISDIR(inode->i_mode))
 370                 flags = O_DIRECTORY;  /* mds likes to know */
 371
 372         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 373              ceph_vinop(inode), file, flags, file->f_flags);
 374         fmode = ceph_flags_to_mode(flags);
 375         wanted = ceph_caps_for_mode(fmode);
 376
 377         /* snapped files are read-only */
 378         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 379                 return -EROFS;
 380
 381         /* trivially open snapdir */
 382         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 383                 return ceph_init_file(inode, file, fmode);
 384         }
 385
 386         /*
 387          * No need to block if we have caps on the auth MDS (for
 388          * write) or any MDS (for read).  Update wanted set
 389          * asynchronously.
 390          */
 391         spin_lock(&ci->i_ceph_lock);
 392         if (__ceph_is_any_real_caps(ci) &&
 393             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 394                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 395                 int issued = __ceph_caps_issued(ci, NULL);
 396
 397                 dout("open %p fmode %d want %s issued %s using existing\n",
 398                      inode, fmode, ceph_cap_string(wanted),
 399                      ceph_cap_string(issued));
 400                 __ceph_touch_fmode(ci, mdsc, fmode);
 401                 spin_unlock(&ci->i_ceph_lock);
 402
 403                 /* adjust wanted? */
 404                 if ((issued & wanted) != wanted &&
 405                     (mds_wanted & wanted) != wanted &&
 406                     ceph_snap(inode) != CEPH_SNAPDIR)
 407                         ceph_check_caps(ci, 0);
 408
 409                 return ceph_init_file(inode, file, fmode);
 410         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 411                    (ci->i_snap_caps & wanted) == wanted) {
 412                 __ceph_touch_fmode(ci, mdsc, fmode);
 413                 spin_unlock(&ci->i_ceph_lock);
 414                 return ceph_init_file(inode, file, fmode);
 415         }
 416
 417         spin_unlock(&ci->i_ceph_lock);
 418
 419         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 420         req = prepare_open_request(inode->i_sb, flags, 0);
 421         if (IS_ERR(req)) {
 422                 err = PTR_ERR(req);
 423                 goto out;
 424         }
 425         req->r_inode = inode;
 426         ihold(inode);
 427
 428         req->r_num_caps = 1;
 429         err = ceph_mdsc_do_request(mdsc, NULL, req);
 430         if (!err)
 431                 err = ceph_init_file(inode, file, req->r_fmode);
 432         ceph_mdsc_put_request(req);
 433         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 434 out:
 435         return err;
 436 }
 437
 438 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 439 static void
 440 cache_file_layout(struct inode *dst, struct inode *src)
 441 {
 442         struct ceph_inode_info *cdst = ceph_inode(dst);
 443         struct ceph_inode_info *csrc = ceph_inode(src);
 444
 445         spin_lock(&cdst->i_ceph_lock);
 446         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 447             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 448                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 449                         sizeof(cdst->i_cached_layout));
 450                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 451                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 452         }
 453         spin_unlock(&cdst->i_ceph_lock);
 454 }
 455
 456 /*
 457  * Try to set up an async create. We need caps, a file layout, and inode number,
 458  * and either a lease on the dentry or complete dir info. If any of those
 459  * criteria are not satisfied, then return false and the caller can go
 460  * synchronous.
 461  */
 462 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 463                                  struct ceph_file_layout *lo, u64 *pino)
 464 {
 465         struct ceph_inode_info *ci = ceph_inode(dir);
 466         struct ceph_dentry_info *di = ceph_dentry(dentry);
 467         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 468         u64 ino;
 469
 470         spin_lock(&ci->i_ceph_lock);
 471         /* No auth cap means no chance for Dc caps */
 472         if (!ci->i_auth_cap)
 473                 goto no_async;
 474
 475         /* Any delegated inos? */
 476         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 477                 goto no_async;
 478
 479         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 480                 goto no_async;
 481
 482         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 483                 goto no_async;
 484
 485         if (d_in_lookup(dentry)) {
 486                 if (!__ceph_dir_is_complete(ci))
 487                         goto no_async;
 488                 spin_lock(&dentry->d_lock);
 489                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 490                 spin_unlock(&dentry->d_lock);
 491         } else if (atomic_read(&ci->i_shared_gen) !=
 492                    READ_ONCE(di->lease_shared_gen)) {
 493                 goto no_async;
 494         }
 495
 496         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 497         if (!ino)
 498                 goto no_async;
 499
 500         *pino = ino;
 501         ceph_take_cap_refs(ci, want, false);
 502         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 503         rcu_assign_pointer(lo->pool_ns,
 504                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 505         got = want;
 506 no_async:
 507         spin_unlock(&ci->i_ceph_lock);
 508         return got;
 509 }
 510
 511 static void restore_deleg_ino(struct inode *dir, u64 ino)
 512 {
 513         struct ceph_inode_info *ci = ceph_inode(dir);
 514         struct ceph_mds_session *s = NULL;
 515
 516         spin_lock(&ci->i_ceph_lock);
 517         if (ci->i_auth_cap)
 518                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 519         spin_unlock(&ci->i_ceph_lock);
 520         if (s) {
 521                 int err = ceph_restore_deleg_ino(s, ino);
 522                 if (err)
 523                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 524                                 ino, err);
 525                 ceph_put_mds_session(s);
 526         }
 527 }
 528
 529 static void wake_async_create_waiters(struct inode *inode,
 530                                       struct ceph_mds_session *session)
 531 {
 532         struct ceph_inode_info *ci = ceph_inode(inode);
 533         bool check_cap = false;
 534
 535         spin_lock(&ci->i_ceph_lock);
 536         if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 537                 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 538                 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 539
 540                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
 541                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
 542                         check_cap = true;
 543                 }
 544         }
 545         ceph_kick_flushing_inode_caps(session, ci);
 546         spin_unlock(&ci->i_ceph_lock);
 547
 548         if (check_cap)
 549                 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
 550 }
 551
 552 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 553                                  struct ceph_mds_request *req)
 554 {
 555         struct dentry *dentry = req->r_dentry;
 556         struct inode *dinode = d_inode(dentry);
 557         struct inode *tinode = req->r_target_inode;
 558         int result = req->r_err ? req->r_err :
 559                         le32_to_cpu(req->r_reply_info.head->result);
 560
 561         WARN_ON_ONCE(dinode && tinode && dinode != tinode);
 562
 563         /* MDS changed -- caller must resubmit */
 564         if (result == -EJUKEBOX)
 565                 goto out;
 566
 567         mapping_set_error(req->r_parent->i_mapping, result);
 568
 569         if (result) {
 570                 int pathlen = 0;
 571                 u64 base = 0;
 572                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 573                                                   &base, 0);
 574
 575                 pr_warn("async create failure path=(%llx)%s result=%d!\n",
 576                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 577                 ceph_mdsc_free_path(path, pathlen);
 578
 579                 ceph_dir_clear_complete(req->r_parent);
 580                 if (!d_unhashed(dentry))
 581                         d_drop(dentry);
 582
 583                 if (dinode) {
 584                         mapping_set_error(dinode->i_mapping, result);
 585                         ceph_inode_shutdown(dinode);
 586                         wake_async_create_waiters(dinode, req->r_session);
 587                 }
 588         }
 589
 590         if (tinode) {
 591                 u64 ino = ceph_vino(tinode).ino;
 592
 593                 if (req->r_deleg_ino != ino)
 594                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 595                                 __func__, req->r_err, req->r_deleg_ino, ino);
 596
 597                 mapping_set_error(tinode->i_mapping, result);
 598                 wake_async_create_waiters(tinode, req->r_session);
 599         } else if (!result) {
 600                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 601                         req->r_deleg_ino);
 602         }
 603 out:
 604         ceph_mdsc_release_dir_caps(req);
 605 }
 606
 607 static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
 608                                     struct file *file, umode_t mode,
 609                                     struct ceph_mds_request *req,
 610                                     struct ceph_acl_sec_ctx *as_ctx,
 611                                     struct ceph_file_layout *lo)
 612 {
 613         int ret;
 614         char xattr_buf[4];
 615         struct ceph_mds_reply_inode in = { };
 616         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 617         struct ceph_inode_info *ci = ceph_inode(dir);
 618         struct ceph_dentry_info *di = ceph_dentry(dentry);
 619         struct inode *inode;
 620         struct timespec64 now;
 621         struct ceph_string *pool_ns;
 622         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 623         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 624                                   .snap = CEPH_NOSNAP };
 625
 626         ktime_get_real_ts64(&now);
 627
 628         inode = ceph_get_inode(dentry->d_sb, vino);
 629         if (IS_ERR(inode))
 630                 return PTR_ERR(inode);
 631
 632         iinfo.inline_version = CEPH_INLINE_NONE;
 633         iinfo.change_attr = 1;
 634         ceph_encode_timespec64(&iinfo.btime, &now);
 635
 636         if (req->r_pagelist) {
 637                 iinfo.xattr_len = req->r_pagelist->length;
 638                 iinfo.xattr_data = req->r_pagelist->mapped_tail;
 639         } else {
 640                 /* fake it */
 641                 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 642                 iinfo.xattr_data = xattr_buf;
 643                 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 644         }
 645
 646         in.ino = cpu_to_le64(vino.ino);
 647         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 648         in.version = cpu_to_le64(1);    // ???
 649         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 650         in.cap.cap_id = cpu_to_le64(1);
 651         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 652         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 653         in.ctime = in.mtime = in.atime = iinfo.btime;
 654         in.truncate_seq = cpu_to_le32(1);
 655         in.truncate_size = cpu_to_le64(-1ULL);
 656         in.xattr_version = cpu_to_le64(1);
 657         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 658         if (dir->i_mode & S_ISGID) {
 659                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
 660
 661                 /* Directories always inherit the setgid bit. */
 662                 if (S_ISDIR(mode))
 663                         mode |= S_ISGID;
 664         } else {
 665                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, current_fsgid()));
 666         }
 667         in.mode = cpu_to_le32((u32)mode);
 668
 669         in.nlink = cpu_to_le32(1);
 670         in.max_size = cpu_to_le64(lo->stripe_unit);
 671
 672         ceph_file_layout_to_legacy(lo, &in.layout);
 673         /* lo is private, so pool_ns can't change */
 674         pool_ns = rcu_dereference_raw(lo->pool_ns);
 675         if (pool_ns) {
 676                 iinfo.pool_ns_len = pool_ns->len;
 677                 iinfo.pool_ns_data = pool_ns->str;
 678         }
 679
 680         down_read(&mdsc->snap_rwsem);
 681         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 682                               req->r_fmode, NULL);
 683         up_read(&mdsc->snap_rwsem);
 684         if (ret) {
 685                 dout("%s failed to fill inode: %d\n", __func__, ret);
 686                 ceph_dir_clear_complete(dir);
 687                 if (!d_unhashed(dentry))
 688                         d_drop(dentry);
 689                 if (inode->i_state & I_NEW)
 690                         discard_new_inode(inode);
 691         } else {
 692                 struct dentry *dn;
 693
 694                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 695                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 696                 ceph_dir_clear_ordered(dir);
 697                 ceph_init_inode_acls(inode, as_ctx);
 698                 if (inode->i_state & I_NEW) {
 699                         /*
 700                          * If it's not I_NEW, then someone created this before
 701                          * we got here. Assume the server is aware of it at
 702                          * that point and don't worry about setting
 703                          * CEPH_I_ASYNC_CREATE.
 704                          */
 705                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 706                         unlock_new_inode(inode);
 707                 }
 708                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 709                         if (!d_unhashed(dentry))
 710                                 d_drop(dentry);
 711                         dn = d_splice_alias(inode, dentry);
 712                         WARN_ON_ONCE(dn && dn != dentry);
 713                 }
 714                 file->f_mode |= FMODE_CREATED;
 715                 ret = finish_open(file, dentry, ceph_open);
 716         }
 717
 718         spin_lock(&dentry->d_lock);
 719         di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
 720         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
 721         spin_unlock(&dentry->d_lock);
 722
 723         return ret;
 724 }
 725
 726 /*
 727  * Do a lookup + open with a single request.  If we get a non-existent
 728  * file or symlink, return 1 so the VFS can retry.
 729  */
 730 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 731                      struct file *file, unsigned flags, umode_t mode)
 732 {
 733         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 734         struct ceph_mds_client *mdsc = fsc->mdsc;
 735         struct ceph_mds_request *req;
 736         struct dentry *dn;
 737         struct ceph_acl_sec_ctx as_ctx = {};
 738         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 739         int mask;
 740         int err;
 741
 742         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 743              dir, dentry, dentry,
 744              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 745
 746         if (dentry->d_name.len > NAME_MAX)
 747                 return -ENAMETOOLONG;
 748
 749         err = ceph_wait_on_conflict_unlink(dentry);
 750         if (err)
 751                 return err;
 752         /*
 753          * Do not truncate the file, since atomic_open is called before the
 754          * permission check. The caller will do the truncation afterward.
 755          */
 756         flags &= ~O_TRUNC;
 757
 758         if (flags & O_CREAT) {
 759                 if (ceph_quota_is_max_files_exceeded(dir))
 760                         return -EDQUOT;
 761                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 762                 if (err < 0)
 763                         return err;
 764                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 765                 if (err < 0)
 766                         goto out_ctx;
 767                 /* Async create can't handle more than a page of xattrs */
 768                 if (as_ctx.pagelist &&
 769                     !list_is_singular(&as_ctx.pagelist->head))
 770                         try_async = false;
 771         } else if (!d_in_lookup(dentry)) {
 772                 /* If it's not being looked up, it's negative */
 773                 return -ENOENT;
 774         }
 775 retry:
 776         /* do the open */
 777         req = prepare_open_request(dir->i_sb, flags, mode);
 778         if (IS_ERR(req)) {
 779                 err = PTR_ERR(req);
 780                 goto out_ctx;
 781         }
 782         req->r_dentry = dget(dentry);
 783         req->r_num_caps = 2;
 784         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 785         if (ceph_security_xattr_wanted(dir))
 786                 mask |= CEPH_CAP_XATTR_SHARED;
 787         req->r_args.open.mask = cpu_to_le32(mask);
 788         req->r_parent = dir;
 789         ihold(dir);
 790
 791         if (flags & O_CREAT) {
 792                 struct ceph_file_layout lo;
 793
 794                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 795                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 796                 if (as_ctx.pagelist) {
 797                         req->r_pagelist = as_ctx.pagelist;
 798                         as_ctx.pagelist = NULL;
 799                 }
 800                 if (try_async &&
 801                     (req->r_dir_caps =
 802                       try_prep_async_create(dir, dentry, &lo,
 803                                             &req->r_deleg_ino))) {
 804                         struct ceph_dentry_info *di = ceph_dentry(dentry);
 805
 806                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 807                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 808                         req->r_callback = ceph_async_create_cb;
 809
 810                         spin_lock(&dentry->d_lock);
 811                         di->flags |= CEPH_DENTRY_ASYNC_CREATE;
 812                         spin_unlock(&dentry->d_lock);
 813
 814                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 815                         if (!err) {
 816                                 err = ceph_finish_async_create(dir, dentry,
 817                                                         file, mode, req,
 818                                                         &as_ctx, &lo);
 819                         } else if (err == -EJUKEBOX) {
 820                                 restore_deleg_ino(dir, req->r_deleg_ino);
 821                                 ceph_mdsc_put_request(req);
 822                                 try_async = false;
 823                                 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 824                                 goto retry;
 825                         }
 826                         ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 827                         goto out_req;
 828                 }
 829         }
 830
 831         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 832         err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
 833         if (err == -ENOENT) {
 834                 dentry = ceph_handle_snapdir(req, dentry);
 835                 if (IS_ERR(dentry)) {
 836                         err = PTR_ERR(dentry);
 837                         goto out_req;
 838                 }
 839                 err = 0;
 840         }
 841
 842         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 843                 err = ceph_handle_notrace_create(dir, dentry);
 844
 845         if (d_in_lookup(dentry)) {
 846                 dn = ceph_finish_lookup(req, dentry, err);
 847                 if (IS_ERR(dn))
 848                         err = PTR_ERR(dn);
 849         } else {
 850                 /* we were given a hashed negative dentry */
 851                 dn = NULL;
 852         }
 853         if (err)
 854                 goto out_req;
 855         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 856                 /* make vfs retry on splice, ENOENT, or symlink */
 857                 dout("atomic_open finish_no_open on dn %p\n", dn);
 858                 err = finish_no_open(file, dn);
 859         } else {
 860                 dout("atomic_open finish_open on dn %p\n", dn);
 861                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 862                         struct inode *newino = d_inode(dentry);
 863
 864                         cache_file_layout(dir, newino);
 865                         ceph_init_inode_acls(newino, &as_ctx);
 866                         file->f_mode |= FMODE_CREATED;
 867                 }
 868                 err = finish_open(file, dentry, ceph_open);
 869         }
 870 out_req:
 871         ceph_mdsc_put_request(req);
 872 out_ctx:
 873         ceph_release_acl_sec_ctx(&as_ctx);
 874         dout("atomic_open result=%d\n", err);
 875         return err;
 876 }
 877
 878 int ceph_release(struct inode *inode, struct file *file)
 879 {
 880         struct ceph_inode_info *ci = ceph_inode(inode);
 881
 882         if (S_ISDIR(inode->i_mode)) {
 883                 struct ceph_dir_file_info *dfi = file->private_data;
 884                 dout("release inode %p dir file %p\n", inode, file);
 885                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 886
 887                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 888
 889                 if (dfi->last_readdir)
 890                         ceph_mdsc_put_request(dfi->last_readdir);
 891                 kfree(dfi->last_name);
 892                 kfree(dfi->dir_info);
 893                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 894         } else {
 895                 struct ceph_file_info *fi = file->private_data;
 896                 dout("release inode %p regular file %p\n", inode, file);
 897                 WARN_ON(!list_empty(&fi->rw_contexts));
 898
 899                 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 900                 ceph_put_fmode(ci, fi->fmode, 1);
 901
 902                 kmem_cache_free(ceph_file_cachep, fi);
 903         }
 904
 905         /* wake up anyone waiting for caps on this inode */
 906         wake_up_all(&ci->i_cap_wq);
 907         return 0;
 908 }
 909
 910 enum {
 911         HAVE_RETRIED = 1,
 912         CHECK_EOF =    2,
 913         READ_INLINE =  3,
 914 };
 915
 916 /*
 917  * Completely synchronous read and write methods.  Direct from __user
 918  * buffer to osd, or directly to user pages (if O_DIRECT).
 919  *
 920  * If the read spans object boundary, just do multiple reads.  (That's not
 921  * atomic, but good enough for now.)
 922  *
 923  * If we get a short result from the OSD, check against i_size; we need to
 924  * only return a short read to the caller if we hit EOF.
 925  */
 926 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 927                               int *retry_op)
 928 {
 929         struct file *file = iocb->ki_filp;
 930         struct inode *inode = file_inode(file);
 931         struct ceph_inode_info *ci = ceph_inode(inode);
 932         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 933         struct ceph_osd_client *osdc = &fsc->client->osdc;
 934         ssize_t ret;
 935         u64 off = iocb->ki_pos;
 936         u64 len = iov_iter_count(to);
 937         u64 i_size = i_size_read(inode);
 938
 939         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 940              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 941
 942         if (!len)
 943                 return 0;
 944         /*
 945          * flush any page cache pages in this range.  this
 946          * will make concurrent normal and sync io slow,
 947          * but it will at least behave sensibly when they are
 948          * in sequence.
 949          */
 950         ret = filemap_write_and_wait_range(inode->i_mapping,
 951                                            off, off + len - 1);
 952         if (ret < 0)
 953                 return ret;
 954
 955         ret = 0;
 956         while ((len = iov_iter_count(to)) > 0) {
 957                 struct ceph_osd_request *req;
 958                 struct page **pages;
 959                 int num_pages;
 960                 size_t page_off;
 961                 bool more;
 962                 int idx;
 963                 size_t left;
 964
 965                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 966                                         ci->i_vino, off, &len, 0, 1,
 967                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 968                                         NULL, ci->i_truncate_seq,
 969                                         ci->i_truncate_size, false);
 970                 if (IS_ERR(req)) {
 971                         ret = PTR_ERR(req);
 972                         break;
 973                 }
 974
 975                 more = len < iov_iter_count(to);
 976
 977                 num_pages = calc_pages_for(off, len);
 978                 page_off = off & ~PAGE_MASK;
 979                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 980                 if (IS_ERR(pages)) {
 981                         ceph_osdc_put_request(req);
 982                         ret = PTR_ERR(pages);
 983                         break;
 984                 }
 985
 986                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 987                                                  false, false);
 988                 ceph_osdc_start_request(osdc, req);
 989                 ret = ceph_osdc_wait_request(osdc, req);
 990
 991                 ceph_update_read_metrics(&fsc->mdsc->metric,
 992                                          req->r_start_latency,
 993                                          req->r_end_latency,
 994                                          len, ret);
 995
 996                 ceph_osdc_put_request(req);
 997
 998                 i_size = i_size_read(inode);
 999                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
1000                      off, len, ret, i_size, (more ? " MORE" : ""));
1001
1002                 if (ret == -ENOENT)
1003                         ret = 0;
1004                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
1005                         int zlen = min(len - ret, i_size - off - ret);
1006                         int zoff = page_off + ret;
1007                         dout("sync_read zero gap %llu~%llu\n",
1008                              off + ret, off + ret + zlen);
1009                         ceph_zero_page_vector_range(zoff, zlen, pages);
1010                         ret += zlen;
1011                 }
1012
1013                 idx = 0;
1014                 left = ret > 0 ? ret : 0;
1015                 while (left > 0) {
1016                         size_t len, copied;
1017                         page_off = off & ~PAGE_MASK;
1018                         len = min_t(size_t, left, PAGE_SIZE - page_off);
1019                         SetPageUptodate(pages[idx]);
1020                         copied = copy_page_to_iter(pages[idx++],
1021                                                    page_off, len, to);
1022                         off += copied;
1023                         left -= copied;
1024                         if (copied < len) {
1025                                 ret = -EFAULT;
1026                                 break;
1027                         }
1028                 }
1029                 ceph_release_page_vector(pages, num_pages);
1030
1031                 if (ret < 0) {
1032                         if (ret == -EBLOCKLISTED)
1033                                 fsc->blocklisted = true;
1034                         break;
1035                 }
1036
1037                 if (off >= i_size || !more)
1038                         break;
1039         }
1040
1041         if (off > iocb->ki_pos) {
1042                 if (off >= i_size) {
1043                         *retry_op = CHECK_EOF;
1044                         ret = i_size - iocb->ki_pos;
1045                         iocb->ki_pos = i_size;
1046                 } else {
1047                         ret = off - iocb->ki_pos;
1048                         iocb->ki_pos = off;
1049                 }
1050         }
1051
1052         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
1053         return ret;
1054 }
1055
1056 struct ceph_aio_request {
1057         struct kiocb *iocb;
1058         size_t total_len;
1059         bool write;
1060         bool should_dirty;
1061         int error;
1062         struct list_head osd_reqs;
1063         unsigned num_reqs;
1064         atomic_t pending_reqs;
1065         struct timespec64 mtime;
1066         struct ceph_cap_flush *prealloc_cf;
1067 };
1068
1069 struct ceph_aio_work {
1070         struct work_struct work;
1071         struct ceph_osd_request *req;
1072 };
1073
1074 static void ceph_aio_retry_work(struct work_struct *work);
1075
1076 static void ceph_aio_complete(struct inode *inode,
1077                               struct ceph_aio_request *aio_req)
1078 {
1079         struct ceph_inode_info *ci = ceph_inode(inode);
1080         int ret;
1081
1082         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1083                 return;
1084
1085         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1086                 inode_dio_end(inode);
1087
1088         ret = aio_req->error;
1089         if (!ret)
1090                 ret = aio_req->total_len;
1091
1092         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1093
1094         if (ret >= 0 && aio_req->write) {
1095                 int dirty;
1096
1097                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1098                 if (endoff > i_size_read(inode)) {
1099                         if (ceph_inode_set_size(inode, endoff))
1100                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1101                 }
1102
1103                 spin_lock(&ci->i_ceph_lock);
1104                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1105                                                &aio_req->prealloc_cf);
1106                 spin_unlock(&ci->i_ceph_lock);
1107                 if (dirty)
1108                         __mark_inode_dirty(inode, dirty);
1109
1110         }
1111
1112         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1113                                                 CEPH_CAP_FILE_RD));
1114
1115         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1116
1117         ceph_free_cap_flush(aio_req->prealloc_cf);
1118         kfree(aio_req);
1119 }
1120
1121 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1122 {
1123         int rc = req->r_result;
1124         struct inode *inode = req->r_inode;
1125         struct ceph_aio_request *aio_req = req->r_priv;
1126         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1127         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1128         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1129
1130         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1131         BUG_ON(!osd_data->num_bvecs);
1132
1133         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1134
1135         if (rc == -EOLDSNAPC) {
1136                 struct ceph_aio_work *aio_work;
1137                 BUG_ON(!aio_req->write);
1138
1139                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1140                 if (aio_work) {
1141                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1142                         aio_work->req = req;
1143                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1144                                    &aio_work->work);
1145                         return;
1146                 }
1147                 rc = -ENOMEM;
1148         } else if (!aio_req->write) {
1149                 if (rc == -ENOENT)
1150                         rc = 0;
1151                 if (rc >= 0 && len > rc) {
1152                         struct iov_iter i;
1153                         int zlen = len - rc;
1154
1155                         /*
1156                          * If read is satisfied by single OSD request,
1157                          * it can pass EOF. Otherwise read is within
1158                          * i_size.
1159                          */
1160                         if (aio_req->num_reqs == 1) {
1161                                 loff_t i_size = i_size_read(inode);
1162                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1163                                 if (endoff < i_size)
1164                                         zlen = min_t(size_t, zlen,
1165                                                      i_size - endoff);
1166                                 aio_req->total_len = rc + zlen;
1167                         }
1168
1169                         iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1170                                       osd_data->num_bvecs, len);
1171                         iov_iter_advance(&i, rc);
1172                         iov_iter_zero(zlen, &i);
1173                 }
1174         }
1175
1176         /* r_start_latency == 0 means the request was not submitted */
1177         if (req->r_start_latency) {
1178                 if (aio_req->write)
1179                         ceph_update_write_metrics(metric, req->r_start_latency,
1180                                                   req->r_end_latency, len, rc);
1181                 else
1182                         ceph_update_read_metrics(metric, req->r_start_latency,
1183                                                  req->r_end_latency, len, rc);
1184         }
1185
1186         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1187                   aio_req->should_dirty);
1188         ceph_osdc_put_request(req);
1189
1190         if (rc < 0)
1191                 cmpxchg(&aio_req->error, 0, rc);
1192
1193         ceph_aio_complete(inode, aio_req);
1194         return;
1195 }
1196
1197 static void ceph_aio_retry_work(struct work_struct *work)
1198 {
1199         struct ceph_aio_work *aio_work =
1200                 container_of(work, struct ceph_aio_work, work);
1201         struct ceph_osd_request *orig_req = aio_work->req;
1202         struct ceph_aio_request *aio_req = orig_req->r_priv;
1203         struct inode *inode = orig_req->r_inode;
1204         struct ceph_inode_info *ci = ceph_inode(inode);
1205         struct ceph_snap_context *snapc;
1206         struct ceph_osd_request *req;
1207         int ret;
1208
1209         spin_lock(&ci->i_ceph_lock);
1210         if (__ceph_have_pending_cap_snap(ci)) {
1211                 struct ceph_cap_snap *capsnap =
1212                         list_last_entry(&ci->i_cap_snaps,
1213                                         struct ceph_cap_snap,
1214                                         ci_item);
1215                 snapc = ceph_get_snap_context(capsnap->context);
1216         } else {
1217                 BUG_ON(!ci->i_head_snapc);
1218                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1219         }
1220         spin_unlock(&ci->i_ceph_lock);
1221
1222         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1223                         false, GFP_NOFS);
1224         if (!req) {
1225                 ret = -ENOMEM;
1226                 req = orig_req;
1227                 goto out;
1228         }
1229
1230         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1231         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1232         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1233
1234         req->r_ops[0] = orig_req->r_ops[0];
1235
1236         req->r_mtime = aio_req->mtime;
1237         req->r_data_offset = req->r_ops[0].extent.offset;
1238
1239         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1240         if (ret) {
1241                 ceph_osdc_put_request(req);
1242                 req = orig_req;
1243                 goto out;
1244         }
1245
1246         ceph_osdc_put_request(orig_req);
1247
1248         req->r_callback = ceph_aio_complete_req;
1249         req->r_inode = inode;
1250         req->r_priv = aio_req;
1251
1252         ceph_osdc_start_request(req->r_osdc, req);
1253 out:
1254         if (ret < 0) {
1255                 req->r_result = ret;
1256                 ceph_aio_complete_req(req);
1257         }
1258
1259         ceph_put_snap_context(snapc);
1260         kfree(aio_work);
1261 }
1262
1263 static ssize_t
1264 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1265                        struct ceph_snap_context *snapc,
1266                        struct ceph_cap_flush **pcf)
1267 {
1268         struct file *file = iocb->ki_filp;
1269         struct inode *inode = file_inode(file);
1270         struct ceph_inode_info *ci = ceph_inode(inode);
1271         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1272         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1273         struct ceph_vino vino;
1274         struct ceph_osd_request *req;
1275         struct bio_vec *bvecs;
1276         struct ceph_aio_request *aio_req = NULL;
1277         int num_pages = 0;
1278         int flags;
1279         int ret = 0;
1280         struct timespec64 mtime = current_time(inode);
1281         size_t count = iov_iter_count(iter);
1282         loff_t pos = iocb->ki_pos;
1283         bool write = iov_iter_rw(iter) == WRITE;
1284         bool should_dirty = !write && user_backed_iter(iter);
1285
1286         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1287                 return -EROFS;
1288
1289         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1290              (write ? "write" : "read"), file, pos, (unsigned)count,
1291              snapc, snapc ? snapc->seq : 0);
1292
1293         if (write) {
1294                 int ret2;
1295
1296                 ceph_fscache_invalidate(inode, true);
1297
1298                 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1299                                         pos >> PAGE_SHIFT,
1300                                         (pos + count - 1) >> PAGE_SHIFT);
1301                 if (ret2 < 0)
1302                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1303
1304                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1305         } else {
1306                 flags = CEPH_OSD_FLAG_READ;
1307         }
1308
1309         while (iov_iter_count(iter) > 0) {
1310                 u64 size = iov_iter_count(iter);
1311                 ssize_t len;
1312
1313                 if (write)
1314                         size = min_t(u64, size, fsc->mount_options->wsize);
1315                 else
1316                         size = min_t(u64, size, fsc->mount_options->rsize);
1317
1318                 vino = ceph_vino(inode);
1319                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1320                                             vino, pos, &size, 0,
1321                                             1,
1322                                             write ? CEPH_OSD_OP_WRITE :
1323                                                     CEPH_OSD_OP_READ,
1324                                             flags, snapc,
1325                                             ci->i_truncate_seq,
1326                                             ci->i_truncate_size,
1327                                             false);
1328                 if (IS_ERR(req)) {
1329                         ret = PTR_ERR(req);
1330                         break;
1331                 }
1332
1333                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1334                 if (len < 0) {
1335                         ceph_osdc_put_request(req);
1336                         ret = len;
1337                         break;
1338                 }
1339                 if (len != size)
1340                         osd_req_op_extent_update(req, 0, len);
1341
1342                 /*
1343                  * To simplify error handling, allow AIO when IO within i_size
1344                  * or IO can be satisfied by single OSD request.
1345                  */
1346                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1347                     (len == count || pos + count <= i_size_read(inode))) {
1348                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1349                         if (aio_req) {
1350                                 aio_req->iocb = iocb;
1351                                 aio_req->write = write;
1352                                 aio_req->should_dirty = should_dirty;
1353                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1354                                 if (write) {
1355                                         aio_req->mtime = mtime;
1356                                         swap(aio_req->prealloc_cf, *pcf);
1357                                 }
1358                         }
1359                         /* ignore error */
1360                 }
1361
1362                 if (write) {
1363                         /*
1364                          * throw out any page cache pages in this range. this
1365                          * may block.
1366                          */
1367                         truncate_inode_pages_range(inode->i_mapping, pos,
1368                                                    PAGE_ALIGN(pos + len) - 1);
1369
1370                         req->r_mtime = mtime;
1371                 }
1372
1373                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1374
1375                 if (aio_req) {
1376                         aio_req->total_len += len;
1377                         aio_req->num_reqs++;
1378                         atomic_inc(&aio_req->pending_reqs);
1379
1380                         req->r_callback = ceph_aio_complete_req;
1381                         req->r_inode = inode;
1382                         req->r_priv = aio_req;
1383                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1384
1385                         pos += len;
1386                         continue;
1387                 }
1388
1389                 ceph_osdc_start_request(req->r_osdc, req);
1390                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1391
1392                 if (write)
1393                         ceph_update_write_metrics(metric, req->r_start_latency,
1394                                                   req->r_end_latency, len, ret);
1395                 else
1396                         ceph_update_read_metrics(metric, req->r_start_latency,
1397                                                  req->r_end_latency, len, ret);
1398
1399                 size = i_size_read(inode);
1400                 if (!write) {
1401                         if (ret == -ENOENT)
1402                                 ret = 0;
1403                         if (ret >= 0 && ret < len && pos + ret < size) {
1404                                 struct iov_iter i;
1405                                 int zlen = min_t(size_t, len - ret,
1406                                                  size - pos - ret);
1407
1408                                 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1409                                 iov_iter_advance(&i, ret);
1410                                 iov_iter_zero(zlen, &i);
1411                                 ret += zlen;
1412                         }
1413                         if (ret >= 0)
1414                                 len = ret;
1415                 }
1416
1417                 put_bvecs(bvecs, num_pages, should_dirty);
1418                 ceph_osdc_put_request(req);
1419                 if (ret < 0)
1420                         break;
1421
1422                 pos += len;
1423                 if (!write && pos >= size)
1424                         break;
1425
1426                 if (write && pos > size) {
1427                         if (ceph_inode_set_size(inode, pos))
1428                                 ceph_check_caps(ceph_inode(inode),
1429                                                 CHECK_CAPS_AUTHONLY);
1430                 }
1431         }
1432
1433         if (aio_req) {
1434                 LIST_HEAD(osd_reqs);
1435
1436                 if (aio_req->num_reqs == 0) {
1437                         kfree(aio_req);
1438                         return ret;
1439                 }
1440
1441                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1442                                               CEPH_CAP_FILE_RD);
1443
1444                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1445                 inode_dio_begin(inode);
1446                 while (!list_empty(&osd_reqs)) {
1447                         req = list_first_entry(&osd_reqs,
1448                                                struct ceph_osd_request,
1449                                                r_private_item);
1450                         list_del_init(&req->r_private_item);
1451                         if (ret >= 0)
1452                                 ceph_osdc_start_request(req->r_osdc, req);
1453                         if (ret < 0) {
1454                                 req->r_result = ret;
1455                                 ceph_aio_complete_req(req);
1456                         }
1457                 }
1458                 return -EIOCBQUEUED;
1459         }
1460
1461         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1462                 ret = pos - iocb->ki_pos;
1463                 iocb->ki_pos = pos;
1464         }
1465         return ret;
1466 }
1467
1468 /*
1469  * Synchronous write, straight from __user pointer or user pages.
1470  *
1471  * If write spans object boundary, just do multiple writes.  (For a
1472  * correct atomic write, we should e.g. take write locks on all
1473  * objects, rollback on failure, etc.)
1474  */
1475 static ssize_t
1476 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1477                 struct ceph_snap_context *snapc)
1478 {
1479         struct file *file = iocb->ki_filp;
1480         struct inode *inode = file_inode(file);
1481         struct ceph_inode_info *ci = ceph_inode(inode);
1482         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1483         struct ceph_vino vino;
1484         struct ceph_osd_request *req;
1485         struct page **pages;
1486         u64 len;
1487         int num_pages;
1488         int written = 0;
1489         int flags;
1490         int ret;
1491         bool check_caps = false;
1492         struct timespec64 mtime = current_time(inode);
1493         size_t count = iov_iter_count(from);
1494
1495         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1496                 return -EROFS;
1497
1498         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1499              file, pos, (unsigned)count, snapc, snapc->seq);
1500
1501         ret = filemap_write_and_wait_range(inode->i_mapping,
1502                                            pos, pos + count - 1);
1503         if (ret < 0)
1504                 return ret;
1505
1506         ceph_fscache_invalidate(inode, false);
1507         ret = invalidate_inode_pages2_range(inode->i_mapping,
1508                                             pos >> PAGE_SHIFT,
1509                                             (pos + count - 1) >> PAGE_SHIFT);
1510         if (ret < 0)
1511                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1512
1513         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1514
1515         while ((len = iov_iter_count(from)) > 0) {
1516                 size_t left;
1517                 int n;
1518
1519                 vino = ceph_vino(inode);
1520                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1521                                             vino, pos, &len, 0, 1,
1522                                             CEPH_OSD_OP_WRITE, flags, snapc,
1523                                             ci->i_truncate_seq,
1524                                             ci->i_truncate_size,
1525                                             false);
1526                 if (IS_ERR(req)) {
1527                         ret = PTR_ERR(req);
1528                         break;
1529                 }
1530
1531                 /*
1532                  * write from beginning of first page,
1533                  * regardless of io alignment
1534                  */
1535                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1536
1537                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1538                 if (IS_ERR(pages)) {
1539                         ret = PTR_ERR(pages);
1540                         goto out;
1541                 }
1542
1543                 left = len;
1544                 for (n = 0; n < num_pages; n++) {
1545                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1546                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1547                         if (ret != plen) {
1548                                 ret = -EFAULT;
1549                                 break;
1550                         }
1551                         left -= ret;
1552                 }
1553
1554                 if (ret < 0) {
1555                         ceph_release_page_vector(pages, num_pages);
1556                         goto out;
1557                 }
1558
1559                 req->r_inode = inode;
1560
1561                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1562                                                 false, true);
1563
1564                 req->r_mtime = mtime;
1565                 ceph_osdc_start_request(&fsc->client->osdc, req);
1566                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1567
1568                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1569                                           req->r_end_latency, len, ret);
1570 out:
1571                 ceph_osdc_put_request(req);
1572                 if (ret != 0) {
1573                         ceph_set_error_write(ci);
1574                         break;
1575                 }
1576
1577                 ceph_clear_error_write(ci);
1578                 pos += len;
1579                 written += len;
1580                 if (pos > i_size_read(inode)) {
1581                         check_caps = ceph_inode_set_size(inode, pos);
1582                         if (check_caps)
1583                                 ceph_check_caps(ceph_inode(inode),
1584                                                 CHECK_CAPS_AUTHONLY);
1585                 }
1586
1587         }
1588
1589         if (ret != -EOLDSNAPC && written > 0) {
1590                 ret = written;
1591                 iocb->ki_pos = pos;
1592         }
1593         return ret;
1594 }
1595
1596 /*
1597  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1598  * Atomically grab references, so that those bits are not released
1599  * back to the MDS mid-read.
1600  *
1601  * Hmm, the sync read case isn't actually async... should it be?
1602  */
1603 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1604 {
1605         struct file *filp = iocb->ki_filp;
1606         struct ceph_file_info *fi = filp->private_data;
1607         size_t len = iov_iter_count(to);
1608         struct inode *inode = file_inode(filp);
1609         struct ceph_inode_info *ci = ceph_inode(inode);
1610         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1611         ssize_t ret;
1612         int want = 0, got = 0;
1613         int retry_op = 0, read = 0;
1614
1615 again:
1616         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1617              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1618
1619         if (ceph_inode_is_shutdown(inode))
1620                 return -ESTALE;
1621
1622         if (direct_lock)
1623                 ceph_start_io_direct(inode);
1624         else
1625                 ceph_start_io_read(inode);
1626
1627         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1628                 want |= CEPH_CAP_FILE_CACHE;
1629         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1630                 want |= CEPH_CAP_FILE_LAZYIO;
1631
1632         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1633         if (ret < 0) {
1634                 if (direct_lock)
1635                         ceph_end_io_direct(inode);
1636                 else
1637                         ceph_end_io_read(inode);
1638                 return ret;
1639         }
1640
1641         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1642             (iocb->ki_flags & IOCB_DIRECT) ||
1643             (fi->flags & CEPH_F_SYNC)) {
1644
1645                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1646                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1647                      ceph_cap_string(got));
1648
1649                 if (!ceph_has_inline_data(ci)) {
1650                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1651                                 ret = ceph_direct_read_write(iocb, to,
1652                                                              NULL, NULL);
1653                                 if (ret >= 0 && ret < len)
1654                                         retry_op = CHECK_EOF;
1655                         } else {
1656                                 ret = ceph_sync_read(iocb, to, &retry_op);
1657                         }
1658                 } else {
1659                         retry_op = READ_INLINE;
1660                 }
1661         } else {
1662                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1663                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1664                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1665                      ceph_cap_string(got));
1666                 ceph_add_rw_context(fi, &rw_ctx);
1667                 ret = generic_file_read_iter(iocb, to);
1668                 ceph_del_rw_context(fi, &rw_ctx);
1669         }
1670
1671         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1672              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1673         ceph_put_cap_refs(ci, got);
1674
1675         if (direct_lock)
1676                 ceph_end_io_direct(inode);
1677         else
1678                 ceph_end_io_read(inode);
1679
1680         if (retry_op > HAVE_RETRIED && ret >= 0) {
1681                 int statret;
1682                 struct page *page = NULL;
1683                 loff_t i_size;
1684                 if (retry_op == READ_INLINE) {
1685                         page = __page_cache_alloc(GFP_KERNEL);
1686                         if (!page)
1687                                 return -ENOMEM;
1688                 }
1689
1690                 statret = __ceph_do_getattr(inode, page,
1691                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1692                 if (statret < 0) {
1693                         if (page)
1694                                 __free_page(page);
1695                         if (statret == -ENODATA) {
1696                                 BUG_ON(retry_op != READ_INLINE);
1697                                 goto again;
1698                         }
1699                         return statret;
1700                 }
1701
1702                 i_size = i_size_read(inode);
1703                 if (retry_op == READ_INLINE) {
1704                         BUG_ON(ret > 0 || read > 0);
1705                         if (iocb->ki_pos < i_size &&
1706                             iocb->ki_pos < PAGE_SIZE) {
1707                                 loff_t end = min_t(loff_t, i_size,
1708                                                    iocb->ki_pos + len);
1709                                 end = min_t(loff_t, end, PAGE_SIZE);
1710                                 if (statret < end)
1711                                         zero_user_segment(page, statret, end);
1712                                 ret = copy_page_to_iter(page,
1713                                                 iocb->ki_pos & ~PAGE_MASK,
1714                                                 end - iocb->ki_pos, to);
1715                                 iocb->ki_pos += ret;
1716                                 read += ret;
1717                         }
1718                         if (iocb->ki_pos < i_size && read < len) {
1719                                 size_t zlen = min_t(size_t, len - read,
1720                                                     i_size - iocb->ki_pos);
1721                                 ret = iov_iter_zero(zlen, to);
1722                                 iocb->ki_pos += ret;
1723                                 read += ret;
1724                         }
1725                         __free_pages(page, 0);
1726                         return read;
1727                 }
1728
1729                 /* hit EOF or hole? */
1730                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1731                     ret < len) {
1732                         dout("sync_read hit hole, ppos %lld < size %lld"
1733                              ", reading more\n", iocb->ki_pos, i_size);
1734
1735                         read += ret;
1736                         len -= ret;
1737                         retry_op = HAVE_RETRIED;
1738                         goto again;
1739                 }
1740         }
1741
1742         if (ret >= 0)
1743                 ret += read;
1744
1745         return ret;
1746 }
1747
1748 /*
1749  * Take cap references to avoid releasing caps to MDS mid-write.
1750  *
1751  * If we are synchronous, and write with an old snap context, the OSD
1752  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1753  * dropping our cap refs and allowing the pending snap to logically
1754  * complete _before_ this write occurs.
1755  *
1756  * If we are near ENOSPC, write synchronously.
1757  */
1758 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1759 {
1760         struct file *file = iocb->ki_filp;
1761         struct ceph_file_info *fi = file->private_data;
1762         struct inode *inode = file_inode(file);
1763         struct ceph_inode_info *ci = ceph_inode(inode);
1764         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1765         struct ceph_osd_client *osdc = &fsc->client->osdc;
1766         struct ceph_cap_flush *prealloc_cf;
1767         ssize_t count, written = 0;
1768         int err, want = 0, got;
1769         bool direct_lock = false;
1770         u32 map_flags;
1771         u64 pool_flags;
1772         loff_t pos;
1773         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1774
1775         if (ceph_inode_is_shutdown(inode))
1776                 return -ESTALE;
1777
1778         if (ceph_snap(inode) != CEPH_NOSNAP)
1779                 return -EROFS;
1780
1781         prealloc_cf = ceph_alloc_cap_flush();
1782         if (!prealloc_cf)
1783                 return -ENOMEM;
1784
1785         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1786                 direct_lock = true;
1787
1788 retry_snap:
1789         if (direct_lock)
1790                 ceph_start_io_direct(inode);
1791         else
1792                 ceph_start_io_write(inode);
1793
1794         /* We can write back this queue in page reclaim */
1795         current->backing_dev_info = inode_to_bdi(inode);
1796
1797         if (iocb->ki_flags & IOCB_APPEND) {
1798                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1799                 if (err < 0)
1800                         goto out;
1801         }
1802
1803         err = generic_write_checks(iocb, from);
1804         if (err <= 0)
1805                 goto out;
1806
1807         pos = iocb->ki_pos;
1808         if (unlikely(pos >= limit)) {
1809                 err = -EFBIG;
1810                 goto out;
1811         } else {
1812                 iov_iter_truncate(from, limit - pos);
1813         }
1814
1815         count = iov_iter_count(from);
1816         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1817                 err = -EDQUOT;
1818                 goto out;
1819         }
1820
1821         down_read(&osdc->lock);
1822         map_flags = osdc->osdmap->flags;
1823         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1824         up_read(&osdc->lock);
1825         if ((map_flags & CEPH_OSDMAP_FULL) ||
1826             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1827                 err = -ENOSPC;
1828                 goto out;
1829         }
1830
1831         err = file_remove_privs(file);
1832         if (err)
1833                 goto out;
1834
1835         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1836              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1837         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1838                 want |= CEPH_CAP_FILE_BUFFER;
1839         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1840                 want |= CEPH_CAP_FILE_LAZYIO;
1841         got = 0;
1842         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1843         if (err < 0)
1844                 goto out;
1845
1846         err = file_update_time(file);
1847         if (err)
1848                 goto out_caps;
1849
1850         inode_inc_iversion_raw(inode);
1851
1852         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1853              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1854
1855         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1856             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1857             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1858                 struct ceph_snap_context *snapc;
1859                 struct iov_iter data;
1860
1861                 spin_lock(&ci->i_ceph_lock);
1862                 if (__ceph_have_pending_cap_snap(ci)) {
1863                         struct ceph_cap_snap *capsnap =
1864                                         list_last_entry(&ci->i_cap_snaps,
1865                                                         struct ceph_cap_snap,
1866                                                         ci_item);
1867                         snapc = ceph_get_snap_context(capsnap->context);
1868                 } else {
1869                         BUG_ON(!ci->i_head_snapc);
1870                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1871                 }
1872                 spin_unlock(&ci->i_ceph_lock);
1873
1874                 /* we might need to revert back to that point */
1875                 data = *from;
1876                 if (iocb->ki_flags & IOCB_DIRECT)
1877                         written = ceph_direct_read_write(iocb, &data, snapc,
1878                                                          &prealloc_cf);
1879                 else
1880                         written = ceph_sync_write(iocb, &data, pos, snapc);
1881                 if (direct_lock)
1882                         ceph_end_io_direct(inode);
1883                 else
1884                         ceph_end_io_write(inode);
1885                 if (written > 0)
1886                         iov_iter_advance(from, written);
1887                 ceph_put_snap_context(snapc);
1888         } else {
1889                 /*
1890                  * No need to acquire the i_truncate_mutex. Because
1891                  * the MDS revokes Fwb caps before sending truncate
1892                  * message to us. We can't get Fwb cap while there
1893                  * are pending vmtruncate. So write and vmtruncate
1894                  * can not run at the same time
1895                  */
1896                 written = generic_perform_write(iocb, from);
1897                 if (likely(written >= 0))
1898                         iocb->ki_pos = pos + written;
1899                 ceph_end_io_write(inode);
1900         }
1901
1902         if (written >= 0) {
1903                 int dirty;
1904
1905                 spin_lock(&ci->i_ceph_lock);
1906                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1907                                                &prealloc_cf);
1908                 spin_unlock(&ci->i_ceph_lock);
1909                 if (dirty)
1910                         __mark_inode_dirty(inode, dirty);
1911                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1912                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
1913         }
1914
1915         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1916              inode, ceph_vinop(inode), pos, (unsigned)count,
1917              ceph_cap_string(got));
1918         ceph_put_cap_refs(ci, got);
1919
1920         if (written == -EOLDSNAPC) {
1921                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1922                      inode, ceph_vinop(inode), pos, (unsigned)count);
1923                 goto retry_snap;
1924         }
1925
1926         if (written >= 0) {
1927                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1928                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1929                         iocb->ki_flags |= IOCB_DSYNC;
1930                 written = generic_write_sync(iocb, written);
1931         }
1932
1933         goto out_unlocked;
1934 out_caps:
1935         ceph_put_cap_refs(ci, got);
1936 out:
1937         if (direct_lock)
1938                 ceph_end_io_direct(inode);
1939         else
1940                 ceph_end_io_write(inode);
1941 out_unlocked:
1942         ceph_free_cap_flush(prealloc_cf);
1943         current->backing_dev_info = NULL;
1944         return written ? written : err;
1945 }
1946
1947 /*
1948  * llseek.  be sure to verify file size on SEEK_END.
1949  */
1950 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1951 {
1952         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1953                 struct inode *inode = file_inode(file);
1954                 int ret;
1955
1956                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1957                 if (ret < 0)
1958                         return ret;
1959         }
1960         return generic_file_llseek(file, offset, whence);
1961 }
1962
1963 static inline void ceph_zero_partial_page(
1964         struct inode *inode, loff_t offset, unsigned size)
1965 {
1966         struct page *page;
1967         pgoff_t index = offset >> PAGE_SHIFT;
1968
1969         page = find_lock_page(inode->i_mapping, index);
1970         if (page) {
1971                 wait_on_page_writeback(page);
1972                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1973                 unlock_page(page);
1974                 put_page(page);
1975         }
1976 }
1977
1978 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1979                                       loff_t length)
1980 {
1981         loff_t nearly = round_up(offset, PAGE_SIZE);
1982         if (offset < nearly) {
1983                 loff_t size = nearly - offset;
1984                 if (length < size)
1985                         size = length;
1986                 ceph_zero_partial_page(inode, offset, size);
1987                 offset += size;
1988                 length -= size;
1989         }
1990         if (length >= PAGE_SIZE) {
1991                 loff_t size = round_down(length, PAGE_SIZE);
1992                 truncate_pagecache_range(inode, offset, offset + size - 1);
1993                 offset += size;
1994                 length -= size;
1995         }
1996         if (length)
1997                 ceph_zero_partial_page(inode, offset, length);
1998 }
1999
2000 static int ceph_zero_partial_object(struct inode *inode,
2001                                     loff_t offset, loff_t *length)
2002 {
2003         struct ceph_inode_info *ci = ceph_inode(inode);
2004         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2005         struct ceph_osd_request *req;
2006         int ret = 0;
2007         loff_t zero = 0;
2008         int op;
2009
2010         if (ceph_inode_is_shutdown(inode))
2011                 return -EIO;
2012
2013         if (!length) {
2014                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2015                 length = &zero;
2016         } else {
2017                 op = CEPH_OSD_OP_ZERO;
2018         }
2019
2020         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2021                                         ceph_vino(inode),
2022                                         offset, length,
2023                                         0, 1, op,
2024                                         CEPH_OSD_FLAG_WRITE,
2025                                         NULL, 0, 0, false);
2026         if (IS_ERR(req)) {
2027                 ret = PTR_ERR(req);
2028                 goto out;
2029         }
2030
2031         req->r_mtime = inode->i_mtime;
2032         ceph_osdc_start_request(&fsc->client->osdc, req);
2033         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2034         if (ret == -ENOENT)
2035                 ret = 0;
2036         ceph_osdc_put_request(req);
2037
2038 out:
2039         return ret;
2040 }
2041
2042 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2043 {
2044         int ret = 0;
2045         struct ceph_inode_info *ci = ceph_inode(inode);
2046         s32 stripe_unit = ci->i_layout.stripe_unit;
2047         s32 stripe_count = ci->i_layout.stripe_count;
2048         s32 object_size = ci->i_layout.object_size;
2049         u64 object_set_size = object_size * stripe_count;
2050         u64 nearly, t;
2051
2052         /* round offset up to next period boundary */
2053         nearly = offset + object_set_size - 1;
2054         t = nearly;
2055         nearly -= do_div(t, object_set_size);
2056
2057         while (length && offset < nearly) {
2058                 loff_t size = length;
2059                 ret = ceph_zero_partial_object(inode, offset, &size);
2060                 if (ret < 0)
2061                         return ret;
2062                 offset += size;
2063                 length -= size;
2064         }
2065         while (length >= object_set_size) {
2066                 int i;
2067                 loff_t pos = offset;
2068                 for (i = 0; i < stripe_count; ++i) {
2069                         ret = ceph_zero_partial_object(inode, pos, NULL);
2070                         if (ret < 0)
2071                                 return ret;
2072                         pos += stripe_unit;
2073                 }
2074                 offset += object_set_size;
2075                 length -= object_set_size;
2076         }
2077         while (length) {
2078                 loff_t size = length;
2079                 ret = ceph_zero_partial_object(inode, offset, &size);
2080                 if (ret < 0)
2081                         return ret;
2082                 offset += size;
2083                 length -= size;
2084         }
2085         return ret;
2086 }
2087
2088 static long ceph_fallocate(struct file *file, int mode,
2089                                 loff_t offset, loff_t length)
2090 {
2091         struct ceph_file_info *fi = file->private_data;
2092         struct inode *inode = file_inode(file);
2093         struct ceph_inode_info *ci = ceph_inode(inode);
2094         struct ceph_cap_flush *prealloc_cf;
2095         int want, got = 0;
2096         int dirty;
2097         int ret = 0;
2098         loff_t endoff = 0;
2099         loff_t size;
2100
2101         dout("%s %p %llx.%llx mode %x, offset %llu length %llu\n", __func__,
2102              inode, ceph_vinop(inode), mode, offset, length);
2103
2104         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2105                 return -EOPNOTSUPP;
2106
2107         if (!S_ISREG(inode->i_mode))
2108                 return -EOPNOTSUPP;
2109
2110         prealloc_cf = ceph_alloc_cap_flush();
2111         if (!prealloc_cf)
2112                 return -ENOMEM;
2113
2114         inode_lock(inode);
2115
2116         if (ceph_snap(inode) != CEPH_NOSNAP) {
2117                 ret = -EROFS;
2118                 goto unlock;
2119         }
2120
2121         size = i_size_read(inode);
2122
2123         /* Are we punching a hole beyond EOF? */
2124         if (offset >= size)
2125                 goto unlock;
2126         if ((offset + length) > size)
2127                 length = size - offset;
2128
2129         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2130                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2131         else
2132                 want = CEPH_CAP_FILE_BUFFER;
2133
2134         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2135         if (ret < 0)
2136                 goto unlock;
2137
2138         ret = file_modified(file);
2139         if (ret)
2140                 goto put_caps;
2141
2142         filemap_invalidate_lock(inode->i_mapping);
2143         ceph_fscache_invalidate(inode, false);
2144         ceph_zero_pagecache_range(inode, offset, length);
2145         ret = ceph_zero_objects(inode, offset, length);
2146
2147         if (!ret) {
2148                 spin_lock(&ci->i_ceph_lock);
2149                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2150                                                &prealloc_cf);
2151                 spin_unlock(&ci->i_ceph_lock);
2152                 if (dirty)
2153                         __mark_inode_dirty(inode, dirty);
2154         }
2155         filemap_invalidate_unlock(inode->i_mapping);
2156
2157 put_caps:
2158         ceph_put_cap_refs(ci, got);
2159 unlock:
2160         inode_unlock(inode);
2161         ceph_free_cap_flush(prealloc_cf);
2162         return ret;
2163 }
2164
2165 /*
2166  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2167  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2168  * this fails; zero is returned on success.
2169  */
2170 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2171                           struct file *dst_filp,
2172                           loff_t dst_endoff, int *dst_got)
2173 {
2174         int ret = 0;
2175         bool retrying = false;
2176
2177 retry_caps:
2178         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2179                             dst_endoff, dst_got);
2180         if (ret < 0)
2181                 return ret;
2182
2183         /*
2184          * Since we're already holding the FILE_WR capability for the dst file,
2185          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2186          * retry dance instead to try to get both capabilities.
2187          */
2188         ret = ceph_try_get_caps(file_inode(src_filp),
2189                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2190                                 false, src_got);
2191         if (ret <= 0) {
2192                 /* Start by dropping dst_ci caps and getting src_ci caps */
2193                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2194                 if (retrying) {
2195                         if (!ret)
2196                                 /* ceph_try_get_caps masks EAGAIN */
2197                                 ret = -EAGAIN;
2198                         return ret;
2199                 }
2200                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2201                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2202                 if (ret < 0)
2203                         return ret;
2204                 /*... drop src_ci caps too, and retry */
2205                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2206                 retrying = true;
2207                 goto retry_caps;
2208         }
2209         return ret;
2210 }
2211
2212 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2213                            struct ceph_inode_info *dst_ci, int dst_got)
2214 {
2215         ceph_put_cap_refs(src_ci, src_got);
2216         ceph_put_cap_refs(dst_ci, dst_got);
2217 }
2218
2219 /*
2220  * This function does several size-related checks, returning an error if:
2221  *  - source file is smaller than off+len
2222  *  - destination file size is not OK (inode_newsize_ok())
2223  *  - max bytes quotas is exceeded
2224  */
2225 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2226                            loff_t src_off, loff_t dst_off, size_t len)
2227 {
2228         loff_t size, endoff;
2229
2230         size = i_size_read(src_inode);
2231         /*
2232          * Don't copy beyond source file EOF.  Instead of simply setting length
2233          * to (size - src_off), just drop to VFS default implementation, as the
2234          * local i_size may be stale due to other clients writing to the source
2235          * inode.
2236          */
2237         if (src_off + len > size) {
2238                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2239                      src_off, len, size);
2240                 return -EOPNOTSUPP;
2241         }
2242         size = i_size_read(dst_inode);
2243
2244         endoff = dst_off + len;
2245         if (inode_newsize_ok(dst_inode, endoff))
2246                 return -EOPNOTSUPP;
2247
2248         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2249                 return -EDQUOT;
2250
2251         return 0;
2252 }
2253
2254 static struct ceph_osd_request *
2255 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2256                             u64 src_snapid,
2257                             struct ceph_object_id *src_oid,
2258                             struct ceph_object_locator *src_oloc,
2259                             struct ceph_object_id *dst_oid,
2260                             struct ceph_object_locator *dst_oloc,
2261                             u32 truncate_seq, u64 truncate_size)
2262 {
2263         struct ceph_osd_request *req;
2264         int ret;
2265         u32 src_fadvise_flags =
2266                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2267                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2268         u32 dst_fadvise_flags =
2269                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2270                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2271
2272         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2273         if (!req)
2274                 return ERR_PTR(-ENOMEM);
2275
2276         req->r_flags = CEPH_OSD_FLAG_WRITE;
2277
2278         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2279         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2280
2281         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2282                                         src_oid, src_oloc,
2283                                         src_fadvise_flags,
2284                                         dst_fadvise_flags,
2285                                         truncate_seq,
2286                                         truncate_size,
2287                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2288         if (ret)
2289                 goto out;
2290
2291         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2292         if (ret)
2293                 goto out;
2294
2295         return req;
2296
2297 out:
2298         ceph_osdc_put_request(req);
2299         return ERR_PTR(ret);
2300 }
2301
2302 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2303                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2304                                     struct ceph_fs_client *fsc,
2305                                     size_t len, unsigned int flags)
2306 {
2307         struct ceph_object_locator src_oloc, dst_oloc;
2308         struct ceph_object_id src_oid, dst_oid;
2309         struct ceph_osd_client *osdc;
2310         struct ceph_osd_request *req;
2311         size_t bytes = 0;
2312         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2313         u32 src_objlen, dst_objlen;
2314         u32 object_size = src_ci->i_layout.object_size;
2315         int ret;
2316
2317         src_oloc.pool = src_ci->i_layout.pool_id;
2318         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2319         dst_oloc.pool = dst_ci->i_layout.pool_id;
2320         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2321         osdc = &fsc->client->osdc;
2322
2323         while (len >= object_size) {
2324                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2325                                               object_size, &src_objnum,
2326                                               &src_objoff, &src_objlen);
2327                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2328                                               object_size, &dst_objnum,
2329                                               &dst_objoff, &dst_objlen);
2330                 ceph_oid_init(&src_oid);
2331                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2332                                 src_ci->i_vino.ino, src_objnum);
2333                 ceph_oid_init(&dst_oid);
2334                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2335                                 dst_ci->i_vino.ino, dst_objnum);
2336                 /* Do an object remote copy */
2337                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2338                                                   &src_oid, &src_oloc,
2339                                                   &dst_oid, &dst_oloc,
2340                                                   dst_ci->i_truncate_seq,
2341                                                   dst_ci->i_truncate_size);
2342                 if (IS_ERR(req))
2343                         ret = PTR_ERR(req);
2344                 else {
2345                         ceph_osdc_start_request(osdc, req);
2346                         ret = ceph_osdc_wait_request(osdc, req);
2347                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2348                                                      req->r_start_latency,
2349                                                      req->r_end_latency,
2350                                                      object_size, ret);
2351                         ceph_osdc_put_request(req);
2352                 }
2353                 if (ret) {
2354                         if (ret == -EOPNOTSUPP) {
2355                                 fsc->have_copy_from2 = false;
2356                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2357                         }
2358                         dout("ceph_osdc_copy_from returned %d\n", ret);
2359                         if (!bytes)
2360                                 bytes = ret;
2361                         goto out;
2362                 }
2363                 len -= object_size;
2364                 bytes += object_size;
2365                 *src_off += object_size;
2366                 *dst_off += object_size;
2367         }
2368
2369 out:
2370         ceph_oloc_destroy(&src_oloc);
2371         ceph_oloc_destroy(&dst_oloc);
2372         return bytes;
2373 }
2374
2375 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2376                                       struct file *dst_file, loff_t dst_off,
2377                                       size_t len, unsigned int flags)
2378 {
2379         struct inode *src_inode = file_inode(src_file);
2380         struct inode *dst_inode = file_inode(dst_file);
2381         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2382         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2383         struct ceph_cap_flush *prealloc_cf;
2384         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2385         loff_t size;
2386         ssize_t ret = -EIO, bytes;
2387         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2388         u32 src_objlen, dst_objlen;
2389         int src_got = 0, dst_got = 0, err, dirty;
2390
2391         if (src_inode->i_sb != dst_inode->i_sb) {
2392                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2393
2394                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2395                                       &dst_fsc->client->fsid)) {
2396                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2397                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2398                         return -EXDEV;
2399                 }
2400         }
2401         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2402                 return -EROFS;
2403
2404         /*
2405          * Some of the checks below will return -EOPNOTSUPP, which will force a
2406          * fallback to the default VFS copy_file_range implementation.  This is
2407          * desirable in several cases (for ex, the 'len' is smaller than the
2408          * size of the objects, or in cases where that would be more
2409          * efficient).
2410          */
2411
2412         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2413                 return -EOPNOTSUPP;
2414
2415         if (!src_fsc->have_copy_from2)
2416                 return -EOPNOTSUPP;
2417
2418         /*
2419          * Striped file layouts require that we copy partial objects, but the
2420          * OSD copy-from operation only supports full-object copies.  Limit
2421          * this to non-striped file layouts for now.
2422          */
2423         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2424             (src_ci->i_layout.stripe_count != 1) ||
2425             (dst_ci->i_layout.stripe_count != 1) ||
2426             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2427                 dout("Invalid src/dst files layout\n");
2428                 return -EOPNOTSUPP;
2429         }
2430
2431         if (len < src_ci->i_layout.object_size)
2432                 return -EOPNOTSUPP; /* no remote copy will be done */
2433
2434         prealloc_cf = ceph_alloc_cap_flush();
2435         if (!prealloc_cf)
2436                 return -ENOMEM;
2437
2438         /* Start by sync'ing the source and destination files */
2439         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2440         if (ret < 0) {
2441                 dout("failed to write src file (%zd)\n", ret);
2442                 goto out;
2443         }
2444         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2445         if (ret < 0) {
2446                 dout("failed to write dst file (%zd)\n", ret);
2447                 goto out;
2448         }
2449
2450         /*
2451          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2452          * clients may have dirty data in their caches.  And OSDs know nothing
2453          * about caps, so they can't safely do the remote object copies.
2454          */
2455         err = get_rd_wr_caps(src_file, &src_got,
2456                              dst_file, (dst_off + len), &dst_got);
2457         if (err < 0) {
2458                 dout("get_rd_wr_caps returned %d\n", err);
2459                 ret = -EOPNOTSUPP;
2460                 goto out;
2461         }
2462
2463         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2464         if (ret < 0)
2465                 goto out_caps;
2466
2467         /* Drop dst file cached pages */
2468         ceph_fscache_invalidate(dst_inode, false);
2469         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2470                                             dst_off >> PAGE_SHIFT,
2471                                             (dst_off + len) >> PAGE_SHIFT);
2472         if (ret < 0) {
2473                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2474                 ret = 0; /* XXX */
2475         }
2476         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2477                                       src_ci->i_layout.object_size,
2478                                       &src_objnum, &src_objoff, &src_objlen);
2479         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2480                                       dst_ci->i_layout.object_size,
2481                                       &dst_objnum, &dst_objoff, &dst_objlen);
2482         /* object-level offsets need to the same */
2483         if (src_objoff != dst_objoff) {
2484                 ret = -EOPNOTSUPP;
2485                 goto out_caps;
2486         }
2487
2488         /*
2489          * Do a manual copy if the object offset isn't object aligned.
2490          * 'src_objlen' contains the bytes left until the end of the object,
2491          * starting at the src_off
2492          */
2493         if (src_objoff) {
2494                 dout("Initial partial copy of %u bytes\n", src_objlen);
2495
2496                 /*
2497                  * we need to temporarily drop all caps as we'll be calling
2498                  * {read,write}_iter, which will get caps again.
2499                  */
2500                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2501                 ret = do_splice_direct(src_file, &src_off, dst_file,
2502                                        &dst_off, src_objlen, flags);
2503                 /* Abort on short copies or on error */
2504                 if (ret < src_objlen) {
2505                         dout("Failed partial copy (%zd)\n", ret);
2506                         goto out;
2507                 }
2508                 len -= ret;
2509                 err = get_rd_wr_caps(src_file, &src_got,
2510                                      dst_file, (dst_off + len), &dst_got);
2511                 if (err < 0)
2512                         goto out;
2513                 err = is_file_size_ok(src_inode, dst_inode,
2514                                       src_off, dst_off, len);
2515                 if (err < 0)
2516                         goto out_caps;
2517         }
2518
2519         size = i_size_read(dst_inode);
2520         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2521                                      src_fsc, len, flags);
2522         if (bytes <= 0) {
2523                 if (!ret)
2524                         ret = bytes;
2525                 goto out_caps;
2526         }
2527         dout("Copied %zu bytes out of %zu\n", bytes, len);
2528         len -= bytes;
2529         ret += bytes;
2530
2531         file_update_time(dst_file);
2532         inode_inc_iversion_raw(dst_inode);
2533
2534         if (dst_off > size) {
2535                 /* Let the MDS know about dst file size change */
2536                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2537                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2538                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
2539         }
2540         /* Mark Fw dirty */
2541         spin_lock(&dst_ci->i_ceph_lock);
2542         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2543         spin_unlock(&dst_ci->i_ceph_lock);
2544         if (dirty)
2545                 __mark_inode_dirty(dst_inode, dirty);
2546
2547 out_caps:
2548         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2549
2550         /*
2551          * Do the final manual copy if we still have some bytes left, unless
2552          * there were errors in remote object copies (len >= object_size).
2553          */
2554         if (len && (len < src_ci->i_layout.object_size)) {
2555                 dout("Final partial copy of %zu bytes\n", len);
2556                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2557                                          &dst_off, len, flags);
2558                 if (bytes > 0)
2559                         ret += bytes;
2560                 else
2561                         dout("Failed partial copy (%zd)\n", bytes);
2562         }
2563
2564 out:
2565         ceph_free_cap_flush(prealloc_cf);
2566
2567         return ret;
2568 }
2569
2570 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2571                                     struct file *dst_file, loff_t dst_off,
2572                                     size_t len, unsigned int flags)
2573 {
2574         ssize_t ret;
2575
2576         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2577                                      len, flags);
2578
2579         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2580                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2581                                               dst_off, len, flags);
2582         return ret;
2583 }
2584
2585 const struct file_operations ceph_file_fops = {
2586         .open = ceph_open,
2587         .release = ceph_release,
2588         .llseek = ceph_llseek,
2589         .read_iter = ceph_read_iter,
2590         .write_iter = ceph_write_iter,
2591         .mmap = ceph_mmap,
2592         .fsync = ceph_fsync,
2593         .lock = ceph_lock,
2594         .setlease = simple_nosetlease,
2595         .flock = ceph_flock,
2596         .splice_read = generic_file_splice_read,
2597         .splice_write = iter_file_splice_write,
2598         .unlocked_ioctl = ceph_ioctl,
2599         .compat_ioctl = compat_ptr_ioctl,
2600         .fallocate      = ceph_fallocate,
2601         .copy_file_range = ceph_copy_file_range,
2602 };