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