fs/ceph/file.c

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