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_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 186         struct ceph_mds_request *req;
 187         int want_auth = USE_ANY_MDS;
 188         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 189
 190         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 191                 want_auth = USE_AUTH_MDS;
 192
 193         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 194         if (IS_ERR(req))
 195                 goto out;
 196         req->r_fmode = ceph_flags_to_mode(flags);
 197         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 198         req->r_args.open.mode = cpu_to_le32(create_mode);
 199 out:
 200         return req;
 201 }
 202
 203 static int ceph_init_file_info(struct inode *inode, struct file *file,
 204                                         int fmode, bool isdir)
 205 {
 206         struct ceph_inode_info *ci = ceph_inode(inode);
 207         struct ceph_mount_options *opt =
 208                 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
 209         struct ceph_file_info *fi;
 210         int ret;
 211
 212         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 213                         inode->i_mode, isdir ? "dir" : "regular");
 214         BUG_ON(inode->i_fop->release != ceph_release);
 215
 216         if (isdir) {
 217                 struct ceph_dir_file_info *dfi =
 218                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 219                 if (!dfi)
 220                         return -ENOMEM;
 221
 222                 file->private_data = dfi;
 223                 fi = &dfi->file_info;
 224                 dfi->next_offset = 2;
 225                 dfi->readdir_cache_idx = -1;
 226         } else {
 227                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 228                 if (!fi)
 229                         return -ENOMEM;
 230
 231                 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
 232                         fi->flags |= CEPH_F_SYNC;
 233
 234                 file->private_data = fi;
 235         }
 236
 237         ceph_get_fmode(ci, fmode, 1);
 238         fi->fmode = fmode;
 239
 240         spin_lock_init(&fi->rw_contexts_lock);
 241         INIT_LIST_HEAD(&fi->rw_contexts);
 242         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 243
 244         if ((file->f_mode & FMODE_WRITE) &&
 245             ci->i_inline_version != CEPH_INLINE_NONE) {
 246                 ret = ceph_uninline_data(file);
 247                 if (ret < 0)
 248                         goto error;
 249         }
 250
 251         return 0;
 252
 253 error:
 254         ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 255         ceph_put_fmode(ci, fi->fmode, 1);
 256         kmem_cache_free(ceph_file_cachep, fi);
 257         /* wake up anyone waiting for caps on this inode */
 258         wake_up_all(&ci->i_cap_wq);
 259         return ret;
 260 }
 261
 262 /*
 263  * initialize private struct file data.
 264  * if we fail, clean up by dropping fmode reference on the ceph_inode
 265  */
 266 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 267 {
 268         int ret = 0;
 269
 270         switch (inode->i_mode & S_IFMT) {
 271         case S_IFREG:
 272                 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
 273                 fallthrough;
 274         case S_IFDIR:
 275                 ret = ceph_init_file_info(inode, file, fmode,
 276                                                 S_ISDIR(inode->i_mode));
 277                 break;
 278
 279         case S_IFLNK:
 280                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 281                      inode->i_mode);
 282                 break;
 283
 284         default:
 285                 dout("init_file %p %p 0%o (special)\n", inode, file,
 286                      inode->i_mode);
 287                 /*
 288                  * we need to drop the open ref now, since we don't
 289                  * have .release set to ceph_release.
 290                  */
 291                 BUG_ON(inode->i_fop->release == ceph_release);
 292
 293                 /* call the proper open fop */
 294                 ret = inode->i_fop->open(inode, file);
 295         }
 296         return ret;
 297 }
 298
 299 /*
 300  * try renew caps after session gets killed.
 301  */
 302 int ceph_renew_caps(struct inode *inode, int fmode)
 303 {
 304         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 305         struct ceph_inode_info *ci = ceph_inode(inode);
 306         struct ceph_mds_request *req;
 307         int err, flags, wanted;
 308
 309         spin_lock(&ci->i_ceph_lock);
 310         __ceph_touch_fmode(ci, mdsc, fmode);
 311         wanted = __ceph_caps_file_wanted(ci);
 312         if (__ceph_is_any_real_caps(ci) &&
 313             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 314                 int issued = __ceph_caps_issued(ci, NULL);
 315                 spin_unlock(&ci->i_ceph_lock);
 316                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 317                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 318                 ceph_check_caps(ci, 0, NULL);
 319                 return 0;
 320         }
 321         spin_unlock(&ci->i_ceph_lock);
 322
 323         flags = 0;
 324         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 325                 flags = O_RDWR;
 326         else if (wanted & CEPH_CAP_FILE_RD)
 327                 flags = O_RDONLY;
 328         else if (wanted & CEPH_CAP_FILE_WR)
 329                 flags = O_WRONLY;
 330 #ifdef O_LAZY
 331         if (wanted & CEPH_CAP_FILE_LAZYIO)
 332                 flags |= O_LAZY;
 333 #endif
 334
 335         req = prepare_open_request(inode->i_sb, flags, 0);
 336         if (IS_ERR(req)) {
 337                 err = PTR_ERR(req);
 338                 goto out;
 339         }
 340
 341         req->r_inode = inode;
 342         ihold(inode);
 343         req->r_num_caps = 1;
 344
 345         err = ceph_mdsc_do_request(mdsc, NULL, req);
 346         ceph_mdsc_put_request(req);
 347 out:
 348         dout("renew caps %p open result=%d\n", inode, err);
 349         return err < 0 ? err : 0;
 350 }
 351
 352 /*
 353  * If we already have the requisite capabilities, we can satisfy
 354  * the open request locally (no need to request new caps from the
 355  * MDS).  We do, however, need to inform the MDS (asynchronously)
 356  * if our wanted caps set expands.
 357  */
 358 int ceph_open(struct inode *inode, struct file *file)
 359 {
 360         struct ceph_inode_info *ci = ceph_inode(inode);
 361         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 362         struct ceph_mds_client *mdsc = fsc->mdsc;
 363         struct ceph_mds_request *req;
 364         struct ceph_file_info *fi = file->private_data;
 365         int err;
 366         int flags, fmode, wanted;
 367
 368         if (fi) {
 369                 dout("open file %p is already opened\n", file);
 370                 return 0;
 371         }
 372
 373         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 374         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 375         if (S_ISDIR(inode->i_mode))
 376                 flags = O_DIRECTORY;  /* mds likes to know */
 377
 378         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 379              ceph_vinop(inode), file, flags, file->f_flags);
 380         fmode = ceph_flags_to_mode(flags);
 381         wanted = ceph_caps_for_mode(fmode);
 382
 383         /* snapped files are read-only */
 384         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 385                 return -EROFS;
 386
 387         /* trivially open snapdir */
 388         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 389                 return ceph_init_file(inode, file, fmode);
 390         }
 391
 392         /*
 393          * No need to block if we have caps on the auth MDS (for
 394          * write) or any MDS (for read).  Update wanted set
 395          * asynchronously.
 396          */
 397         spin_lock(&ci->i_ceph_lock);
 398         if (__ceph_is_any_real_caps(ci) &&
 399             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 400                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 401                 int issued = __ceph_caps_issued(ci, NULL);
 402
 403                 dout("open %p fmode %d want %s issued %s using existing\n",
 404                      inode, fmode, ceph_cap_string(wanted),
 405                      ceph_cap_string(issued));
 406                 __ceph_touch_fmode(ci, mdsc, fmode);
 407                 spin_unlock(&ci->i_ceph_lock);
 408
 409                 /* adjust wanted? */
 410                 if ((issued & wanted) != wanted &&
 411                     (mds_wanted & wanted) != wanted &&
 412                     ceph_snap(inode) != CEPH_SNAPDIR)
 413                         ceph_check_caps(ci, 0, NULL);
 414
 415                 return ceph_init_file(inode, file, fmode);
 416         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 417                    (ci->i_snap_caps & wanted) == wanted) {
 418                 __ceph_touch_fmode(ci, mdsc, fmode);
 419                 spin_unlock(&ci->i_ceph_lock);
 420                 return ceph_init_file(inode, file, fmode);
 421         }
 422
 423         spin_unlock(&ci->i_ceph_lock);
 424
 425         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 426         req = prepare_open_request(inode->i_sb, flags, 0);
 427         if (IS_ERR(req)) {
 428                 err = PTR_ERR(req);
 429                 goto out;
 430         }
 431         req->r_inode = inode;
 432         ihold(inode);
 433
 434         req->r_num_caps = 1;
 435         err = ceph_mdsc_do_request(mdsc, NULL, req);
 436         if (!err)
 437                 err = ceph_init_file(inode, file, req->r_fmode);
 438         ceph_mdsc_put_request(req);
 439         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 440 out:
 441         return err;
 442 }
 443
 444 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 445 static void
 446 cache_file_layout(struct inode *dst, struct inode *src)
 447 {
 448         struct ceph_inode_info *cdst = ceph_inode(dst);
 449         struct ceph_inode_info *csrc = ceph_inode(src);
 450
 451         spin_lock(&cdst->i_ceph_lock);
 452         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 453             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 454                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 455                         sizeof(cdst->i_cached_layout));
 456                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 457                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 458         }
 459         spin_unlock(&cdst->i_ceph_lock);
 460 }
 461
 462 /*
 463  * Try to set up an async create. We need caps, a file layout, and inode number,
 464  * and either a lease on the dentry or complete dir info. If any of those
 465  * criteria are not satisfied, then return false and the caller can go
 466  * synchronous.
 467  */
 468 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 469                                  struct ceph_file_layout *lo, u64 *pino)
 470 {
 471         struct ceph_inode_info *ci = ceph_inode(dir);
 472         struct ceph_dentry_info *di = ceph_dentry(dentry);
 473         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 474         u64 ino;
 475
 476         spin_lock(&ci->i_ceph_lock);
 477         /* No auth cap means no chance for Dc caps */
 478         if (!ci->i_auth_cap)
 479                 goto no_async;
 480
 481         /* Any delegated inos? */
 482         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 483                 goto no_async;
 484
 485         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 486                 goto no_async;
 487
 488         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 489                 goto no_async;
 490
 491         if (d_in_lookup(dentry)) {
 492                 if (!__ceph_dir_is_complete(ci))
 493                         goto no_async;
 494                 spin_lock(&dentry->d_lock);
 495                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 496                 spin_unlock(&dentry->d_lock);
 497         } else if (atomic_read(&ci->i_shared_gen) !=
 498                    READ_ONCE(di->lease_shared_gen)) {
 499                 goto no_async;
 500         }
 501
 502         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 503         if (!ino)
 504                 goto no_async;
 505
 506         *pino = ino;
 507         ceph_take_cap_refs(ci, want, false);
 508         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 509         rcu_assign_pointer(lo->pool_ns,
 510                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 511         got = want;
 512 no_async:
 513         spin_unlock(&ci->i_ceph_lock);
 514         return got;
 515 }
 516
 517 static void restore_deleg_ino(struct inode *dir, u64 ino)
 518 {
 519         struct ceph_inode_info *ci = ceph_inode(dir);
 520         struct ceph_mds_session *s = NULL;
 521
 522         spin_lock(&ci->i_ceph_lock);
 523         if (ci->i_auth_cap)
 524                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 525         spin_unlock(&ci->i_ceph_lock);
 526         if (s) {
 527                 int err = ceph_restore_deleg_ino(s, ino);
 528                 if (err)
 529                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 530                                 ino, err);
 531                 ceph_put_mds_session(s);
 532         }
 533 }
 534
 535 static void wake_async_create_waiters(struct inode *inode,
 536                                       struct ceph_mds_session *session)
 537 {
 538         struct ceph_inode_info *ci = ceph_inode(inode);
 539
 540         spin_lock(&ci->i_ceph_lock);
 541         if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
 542                 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
 543                 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
 544         }
 545         ceph_kick_flushing_inode_caps(session, ci);
 546         spin_unlock(&ci->i_ceph_lock);
 547 }
 548
 549 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 550                                  struct ceph_mds_request *req)
 551 {
 552         struct dentry *dentry = req->r_dentry;
 553         struct inode *dinode = d_inode(dentry);
 554         struct inode *tinode = req->r_target_inode;
 555         int result = req->r_err ? req->r_err :
 556                         le32_to_cpu(req->r_reply_info.head->result);
 557
 558         WARN_ON_ONCE(dinode && tinode && dinode != tinode);
 559
 560         /* MDS changed -- caller must resubmit */
 561         if (result == -EJUKEBOX)
 562                 goto out;
 563
 564         mapping_set_error(req->r_parent->i_mapping, result);
 565
 566         if (result) {
 567                 int pathlen = 0;
 568                 u64 base = 0;
 569                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 570                                                   &base, 0);
 571
 572                 pr_warn("ceph: async create failure path=(%llx)%s result=%d!\n",
 573                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 574                 ceph_mdsc_free_path(path, pathlen);
 575
 576                 ceph_dir_clear_complete(req->r_parent);
 577                 if (!d_unhashed(dentry))
 578                         d_drop(dentry);
 579
 580                 if (dinode) {
 581                         mapping_set_error(dinode->i_mapping, result);
 582                         ceph_inode_shutdown(dinode);
 583                         wake_async_create_waiters(dinode, req->r_session);
 584                 }
 585         }
 586
 587         if (tinode) {
 588                 u64 ino = ceph_vino(tinode).ino;
 589
 590                 if (req->r_deleg_ino != ino)
 591                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 592                                 __func__, req->r_err, req->r_deleg_ino, ino);
 593
 594                 mapping_set_error(tinode->i_mapping, result);
 595                 wake_async_create_waiters(tinode, req->r_session);
 596         } else if (!result) {
 597                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 598                         req->r_deleg_ino);
 599         }
 600 out:
 601         ceph_mdsc_release_dir_caps(req);
 602 }
 603
 604 static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
 605                                     struct file *file, umode_t mode,
 606                                     struct ceph_mds_request *req,
 607                                     struct ceph_acl_sec_ctx *as_ctx,
 608                                     struct ceph_file_layout *lo)
 609 {
 610         int ret;
 611         char xattr_buf[4];
 612         struct ceph_mds_reply_inode in = { };
 613         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 614         struct ceph_inode_info *ci = ceph_inode(dir);
 615         struct inode *inode;
 616         struct timespec64 now;
 617         struct ceph_string *pool_ns;
 618         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 619         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 620                                   .snap = CEPH_NOSNAP };
 621
 622         ktime_get_real_ts64(&now);
 623
 624         inode = ceph_get_inode(dentry->d_sb, vino);
 625         if (IS_ERR(inode))
 626                 return PTR_ERR(inode);
 627
 628         iinfo.inline_version = CEPH_INLINE_NONE;
 629         iinfo.change_attr = 1;
 630         ceph_encode_timespec64(&iinfo.btime, &now);
 631
 632         if (req->r_pagelist) {
 633                 iinfo.xattr_len = req->r_pagelist->length;
 634                 iinfo.xattr_data = req->r_pagelist->mapped_tail;
 635         } else {
 636                 /* fake it */
 637                 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 638                 iinfo.xattr_data = xattr_buf;
 639                 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 640         }
 641
 642         in.ino = cpu_to_le64(vino.ino);
 643         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 644         in.version = cpu_to_le64(1);    // ???
 645         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 646         in.cap.cap_id = cpu_to_le64(1);
 647         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 648         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 649         in.ctime = in.mtime = in.atime = iinfo.btime;
 650         in.truncate_seq = cpu_to_le32(1);
 651         in.truncate_size = cpu_to_le64(-1ULL);
 652         in.xattr_version = cpu_to_le64(1);
 653         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 654         if (dir->i_mode & S_ISGID) {
 655                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
 656
 657                 /* Directories always inherit the setgid bit. */
 658                 if (S_ISDIR(mode))
 659                         mode |= S_ISGID;
 660                 else if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP) &&
 661                          !in_group_p(dir->i_gid) &&
 662                          !capable_wrt_inode_uidgid(&init_user_ns, dir, CAP_FSETID))
 663                         mode &= ~S_ISGID;
 664         } else {
 665                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, current_fsgid()));
 666         }
 667         in.mode = cpu_to_le32((u32)mode);
 668
 669         in.nlink = cpu_to_le32(1);
 670         in.max_size = cpu_to_le64(lo->stripe_unit);
 671
 672         ceph_file_layout_to_legacy(lo, &in.layout);
 673         /* lo is private, so pool_ns can't change */
 674         pool_ns = rcu_dereference_raw(lo->pool_ns);
 675         if (pool_ns) {
 676                 iinfo.pool_ns_len = pool_ns->len;
 677                 iinfo.pool_ns_data = pool_ns->str;
 678         }
 679
 680         down_read(&mdsc->snap_rwsem);
 681         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 682                               req->r_fmode, NULL);
 683         up_read(&mdsc->snap_rwsem);
 684         if (ret) {
 685                 dout("%s failed to fill inode: %d\n", __func__, ret);
 686                 ceph_dir_clear_complete(dir);
 687                 if (!d_unhashed(dentry))
 688                         d_drop(dentry);
 689                 if (inode->i_state & I_NEW)
 690                         discard_new_inode(inode);
 691         } else {
 692                 struct dentry *dn;
 693
 694                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 695                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 696                 ceph_dir_clear_ordered(dir);
 697                 ceph_init_inode_acls(inode, as_ctx);
 698                 if (inode->i_state & I_NEW) {
 699                         /*
 700                          * If it's not I_NEW, then someone created this before
 701                          * we got here. Assume the server is aware of it at
 702                          * that point and don't worry about setting
 703                          * CEPH_I_ASYNC_CREATE.
 704                          */
 705                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 706                         unlock_new_inode(inode);
 707                 }
 708                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 709                         if (!d_unhashed(dentry))
 710                                 d_drop(dentry);
 711                         dn = d_splice_alias(inode, dentry);
 712                         WARN_ON_ONCE(dn && dn != dentry);
 713                 }
 714                 file->f_mode |= FMODE_CREATED;
 715                 ret = finish_open(file, dentry, ceph_open);
 716         }
 717         return ret;
 718 }
 719
 720 /*
 721  * Do a lookup + open with a single request.  If we get a non-existent
 722  * file or symlink, return 1 so the VFS can retry.
 723  */
 724 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 725                      struct file *file, unsigned flags, umode_t mode)
 726 {
 727         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 728         struct ceph_mds_client *mdsc = fsc->mdsc;
 729         struct ceph_mds_request *req;
 730         struct dentry *dn;
 731         struct ceph_acl_sec_ctx as_ctx = {};
 732         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 733         int mask;
 734         int err;
 735
 736         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 737              dir, dentry, dentry,
 738              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 739
 740         if (dentry->d_name.len > NAME_MAX)
 741                 return -ENAMETOOLONG;
 742
 743         if (flags & O_CREAT) {
 744                 if (ceph_quota_is_max_files_exceeded(dir))
 745                         return -EDQUOT;
 746                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 747                 if (err < 0)
 748                         return err;
 749                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 750                 if (err < 0)
 751                         goto out_ctx;
 752                 /* Async create can't handle more than a page of xattrs */
 753                 if (as_ctx.pagelist &&
 754                     !list_is_singular(&as_ctx.pagelist->head))
 755                         try_async = false;
 756         } else if (!d_in_lookup(dentry)) {
 757                 /* If it's not being looked up, it's negative */
 758                 return -ENOENT;
 759         }
 760 retry:
 761         /* do the open */
 762         req = prepare_open_request(dir->i_sb, flags, mode);
 763         if (IS_ERR(req)) {
 764                 err = PTR_ERR(req);
 765                 goto out_ctx;
 766         }
 767         req->r_dentry = dget(dentry);
 768         req->r_num_caps = 2;
 769         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 770         if (ceph_security_xattr_wanted(dir))
 771                 mask |= CEPH_CAP_XATTR_SHARED;
 772         req->r_args.open.mask = cpu_to_le32(mask);
 773         req->r_parent = dir;
 774         ihold(dir);
 775
 776         if (flags & O_CREAT) {
 777                 struct ceph_file_layout lo;
 778
 779                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 780                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 781                 if (as_ctx.pagelist) {
 782                         req->r_pagelist = as_ctx.pagelist;
 783                         as_ctx.pagelist = NULL;
 784                 }
 785                 if (try_async &&
 786                     (req->r_dir_caps =
 787                       try_prep_async_create(dir, dentry, &lo,
 788                                             &req->r_deleg_ino))) {
 789                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 790                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 791                         req->r_callback = ceph_async_create_cb;
 792                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 793                         if (!err) {
 794                                 err = ceph_finish_async_create(dir, dentry,
 795                                                         file, mode, req,
 796                                                         &as_ctx, &lo);
 797                         } else if (err == -EJUKEBOX) {
 798                                 restore_deleg_ino(dir, req->r_deleg_ino);
 799                                 ceph_mdsc_put_request(req);
 800                                 try_async = false;
 801                                 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 802                                 goto retry;
 803                         }
 804                         ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 805                         goto out_req;
 806                 }
 807         }
 808
 809         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 810         err = ceph_mdsc_do_request(mdsc,
 811                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 812                                    req);
 813         if (err == -ENOENT) {
 814                 dentry = ceph_handle_snapdir(req, dentry);
 815                 if (IS_ERR(dentry)) {
 816                         err = PTR_ERR(dentry);
 817                         goto out_req;
 818                 }
 819                 err = 0;
 820         }
 821
 822         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 823                 err = ceph_handle_notrace_create(dir, dentry);
 824
 825         if (d_in_lookup(dentry)) {
 826                 dn = ceph_finish_lookup(req, dentry, err);
 827                 if (IS_ERR(dn))
 828                         err = PTR_ERR(dn);
 829         } else {
 830                 /* we were given a hashed negative dentry */
 831                 dn = NULL;
 832         }
 833         if (err)
 834                 goto out_req;
 835         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 836                 /* make vfs retry on splice, ENOENT, or symlink */
 837                 dout("atomic_open finish_no_open on dn %p\n", dn);
 838                 err = finish_no_open(file, dn);
 839         } else {
 840                 dout("atomic_open finish_open on dn %p\n", dn);
 841                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 842                         struct inode *newino = d_inode(dentry);
 843
 844                         cache_file_layout(dir, newino);
 845                         ceph_init_inode_acls(newino, &as_ctx);
 846                         file->f_mode |= FMODE_CREATED;
 847                 }
 848                 err = finish_open(file, dentry, ceph_open);
 849         }
 850 out_req:
 851         ceph_mdsc_put_request(req);
 852 out_ctx:
 853         ceph_release_acl_sec_ctx(&as_ctx);
 854         dout("atomic_open result=%d\n", err);
 855         return err;
 856 }
 857
 858 int ceph_release(struct inode *inode, struct file *file)
 859 {
 860         struct ceph_inode_info *ci = ceph_inode(inode);
 861
 862         if (S_ISDIR(inode->i_mode)) {
 863                 struct ceph_dir_file_info *dfi = file->private_data;
 864                 dout("release inode %p dir file %p\n", inode, file);
 865                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 866
 867                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 868
 869                 if (dfi->last_readdir)
 870                         ceph_mdsc_put_request(dfi->last_readdir);
 871                 kfree(dfi->last_name);
 872                 kfree(dfi->dir_info);
 873                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 874         } else {
 875                 struct ceph_file_info *fi = file->private_data;
 876                 dout("release inode %p regular file %p\n", inode, file);
 877                 WARN_ON(!list_empty(&fi->rw_contexts));
 878
 879                 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 880                 ceph_put_fmode(ci, fi->fmode, 1);
 881
 882                 kmem_cache_free(ceph_file_cachep, fi);
 883         }
 884
 885         /* wake up anyone waiting for caps on this inode */
 886         wake_up_all(&ci->i_cap_wq);
 887         return 0;
 888 }
 889
 890 enum {
 891         HAVE_RETRIED = 1,
 892         CHECK_EOF =    2,
 893         READ_INLINE =  3,
 894 };
 895
 896 /*
 897  * Completely synchronous read and write methods.  Direct from __user
 898  * buffer to osd, or directly to user pages (if O_DIRECT).
 899  *
 900  * If the read spans object boundary, just do multiple reads.  (That's not
 901  * atomic, but good enough for now.)
 902  *
 903  * If we get a short result from the OSD, check against i_size; we need to
 904  * only return a short read to the caller if we hit EOF.
 905  */
 906 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 907                               int *retry_op)
 908 {
 909         struct file *file = iocb->ki_filp;
 910         struct inode *inode = file_inode(file);
 911         struct ceph_inode_info *ci = ceph_inode(inode);
 912         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 913         struct ceph_osd_client *osdc = &fsc->client->osdc;
 914         ssize_t ret;
 915         u64 off = iocb->ki_pos;
 916         u64 len = iov_iter_count(to);
 917         u64 i_size = i_size_read(inode);
 918
 919         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 920              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 921
 922         if (!len)
 923                 return 0;
 924         /*
 925          * flush any page cache pages in this range.  this
 926          * will make concurrent normal and sync io slow,
 927          * but it will at least behave sensibly when they are
 928          * in sequence.
 929          */
 930         ret = filemap_write_and_wait_range(inode->i_mapping,
 931                                            off, off + len - 1);
 932         if (ret < 0)
 933                 return ret;
 934
 935         ret = 0;
 936         while ((len = iov_iter_count(to)) > 0) {
 937                 struct ceph_osd_request *req;
 938                 struct page **pages;
 939                 int num_pages;
 940                 size_t page_off;
 941                 bool more;
 942                 int idx;
 943                 size_t left;
 944
 945                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 946                                         ci->i_vino, off, &len, 0, 1,
 947                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 948                                         NULL, ci->i_truncate_seq,
 949                                         ci->i_truncate_size, false);
 950                 if (IS_ERR(req)) {
 951                         ret = PTR_ERR(req);
 952                         break;
 953                 }
 954
 955                 more = len < iov_iter_count(to);
 956
 957                 num_pages = calc_pages_for(off, len);
 958                 page_off = off & ~PAGE_MASK;
 959                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 960                 if (IS_ERR(pages)) {
 961                         ceph_osdc_put_request(req);
 962                         ret = PTR_ERR(pages);
 963                         break;
 964                 }
 965
 966                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 967                                                  false, false);
 968                 ret = ceph_osdc_start_request(osdc, req, false);
 969                 if (!ret)
 970                         ret = ceph_osdc_wait_request(osdc, req);
 971
 972                 ceph_update_read_metrics(&fsc->mdsc->metric,
 973                                          req->r_start_latency,
 974                                          req->r_end_latency,
 975                                          len, ret);
 976
 977                 ceph_osdc_put_request(req);
 978
 979                 i_size = i_size_read(inode);
 980                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 981                      off, len, ret, i_size, (more ? " MORE" : ""));
 982
 983                 if (ret == -ENOENT)
 984                         ret = 0;
 985                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 986                         int zlen = min(len - ret, i_size - off - ret);
 987                         int zoff = page_off + ret;
 988                         dout("sync_read zero gap %llu~%llu\n",
 989                              off + ret, off + ret + zlen);
 990                         ceph_zero_page_vector_range(zoff, zlen, pages);
 991                         ret += zlen;
 992                 }
 993
 994                 idx = 0;
 995                 left = ret > 0 ? ret : 0;
 996                 while (left > 0) {
 997                         size_t len, copied;
 998                         page_off = off & ~PAGE_MASK;
 999                         len = min_t(size_t, left, PAGE_SIZE - page_off);
1000                         SetPageUptodate(pages[idx]);
1001                         copied = copy_page_to_iter(pages[idx++],
1002                                                    page_off, len, to);
1003                         off += copied;
1004                         left -= copied;
1005                         if (copied < len) {
1006                                 ret = -EFAULT;
1007                                 break;
1008                         }
1009                 }
1010                 ceph_release_page_vector(pages, num_pages);
1011
1012                 if (ret < 0) {
1013                         if (ret == -EBLOCKLISTED)
1014                                 fsc->blocklisted = true;
1015                         break;
1016                 }
1017
1018                 if (off >= i_size || !more)
1019                         break;
1020         }
1021
1022         if (off > iocb->ki_pos) {
1023                 if (off >= i_size) {
1024                         *retry_op = CHECK_EOF;
1025                         ret = i_size - iocb->ki_pos;
1026                         iocb->ki_pos = i_size;
1027                 } else {
1028                         ret = off - iocb->ki_pos;
1029                         iocb->ki_pos = off;
1030                 }
1031         }
1032
1033         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
1034         return ret;
1035 }
1036
1037 struct ceph_aio_request {
1038         struct kiocb *iocb;
1039         size_t total_len;
1040         bool write;
1041         bool should_dirty;
1042         int error;
1043         struct list_head osd_reqs;
1044         unsigned num_reqs;
1045         atomic_t pending_reqs;
1046         struct timespec64 mtime;
1047         struct ceph_cap_flush *prealloc_cf;
1048 };
1049
1050 struct ceph_aio_work {
1051         struct work_struct work;
1052         struct ceph_osd_request *req;
1053 };
1054
1055 static void ceph_aio_retry_work(struct work_struct *work);
1056
1057 static void ceph_aio_complete(struct inode *inode,
1058                               struct ceph_aio_request *aio_req)
1059 {
1060         struct ceph_inode_info *ci = ceph_inode(inode);
1061         int ret;
1062
1063         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1064                 return;
1065
1066         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1067                 inode_dio_end(inode);
1068
1069         ret = aio_req->error;
1070         if (!ret)
1071                 ret = aio_req->total_len;
1072
1073         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1074
1075         if (ret >= 0 && aio_req->write) {
1076                 int dirty;
1077
1078                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1079                 if (endoff > i_size_read(inode)) {
1080                         if (ceph_inode_set_size(inode, endoff))
1081                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1082                 }
1083
1084                 spin_lock(&ci->i_ceph_lock);
1085                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1086                                                &aio_req->prealloc_cf);
1087                 spin_unlock(&ci->i_ceph_lock);
1088                 if (dirty)
1089                         __mark_inode_dirty(inode, dirty);
1090
1091         }
1092
1093         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1094                                                 CEPH_CAP_FILE_RD));
1095
1096         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1097
1098         ceph_free_cap_flush(aio_req->prealloc_cf);
1099         kfree(aio_req);
1100 }
1101
1102 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1103 {
1104         int rc = req->r_result;
1105         struct inode *inode = req->r_inode;
1106         struct ceph_aio_request *aio_req = req->r_priv;
1107         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1108         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1109         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1110
1111         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1112         BUG_ON(!osd_data->num_bvecs);
1113
1114         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1115
1116         if (rc == -EOLDSNAPC) {
1117                 struct ceph_aio_work *aio_work;
1118                 BUG_ON(!aio_req->write);
1119
1120                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1121                 if (aio_work) {
1122                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1123                         aio_work->req = req;
1124                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1125                                    &aio_work->work);
1126                         return;
1127                 }
1128                 rc = -ENOMEM;
1129         } else if (!aio_req->write) {
1130                 if (rc == -ENOENT)
1131                         rc = 0;
1132                 if (rc >= 0 && len > rc) {
1133                         struct iov_iter i;
1134                         int zlen = len - rc;
1135
1136                         /*
1137                          * If read is satisfied by single OSD request,
1138                          * it can pass EOF. Otherwise read is within
1139                          * i_size.
1140                          */
1141                         if (aio_req->num_reqs == 1) {
1142                                 loff_t i_size = i_size_read(inode);
1143                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1144                                 if (endoff < i_size)
1145                                         zlen = min_t(size_t, zlen,
1146                                                      i_size - endoff);
1147                                 aio_req->total_len = rc + zlen;
1148                         }
1149
1150                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1151                                       osd_data->num_bvecs, len);
1152                         iov_iter_advance(&i, rc);
1153                         iov_iter_zero(zlen, &i);
1154                 }
1155         }
1156
1157         /* r_start_latency == 0 means the request was not submitted */
1158         if (req->r_start_latency) {
1159                 if (aio_req->write)
1160                         ceph_update_write_metrics(metric, req->r_start_latency,
1161                                                   req->r_end_latency, len, rc);
1162                 else
1163                         ceph_update_read_metrics(metric, req->r_start_latency,
1164                                                  req->r_end_latency, len, rc);
1165         }
1166
1167         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1168                   aio_req->should_dirty);
1169         ceph_osdc_put_request(req);
1170
1171         if (rc < 0)
1172                 cmpxchg(&aio_req->error, 0, rc);
1173
1174         ceph_aio_complete(inode, aio_req);
1175         return;
1176 }
1177
1178 static void ceph_aio_retry_work(struct work_struct *work)
1179 {
1180         struct ceph_aio_work *aio_work =
1181                 container_of(work, struct ceph_aio_work, work);
1182         struct ceph_osd_request *orig_req = aio_work->req;
1183         struct ceph_aio_request *aio_req = orig_req->r_priv;
1184         struct inode *inode = orig_req->r_inode;
1185         struct ceph_inode_info *ci = ceph_inode(inode);
1186         struct ceph_snap_context *snapc;
1187         struct ceph_osd_request *req;
1188         int ret;
1189
1190         spin_lock(&ci->i_ceph_lock);
1191         if (__ceph_have_pending_cap_snap(ci)) {
1192                 struct ceph_cap_snap *capsnap =
1193                         list_last_entry(&ci->i_cap_snaps,
1194                                         struct ceph_cap_snap,
1195                                         ci_item);
1196                 snapc = ceph_get_snap_context(capsnap->context);
1197         } else {
1198                 BUG_ON(!ci->i_head_snapc);
1199                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1200         }
1201         spin_unlock(&ci->i_ceph_lock);
1202
1203         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1204                         false, GFP_NOFS);
1205         if (!req) {
1206                 ret = -ENOMEM;
1207                 req = orig_req;
1208                 goto out;
1209         }
1210
1211         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1212         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1213         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1214
1215         req->r_ops[0] = orig_req->r_ops[0];
1216
1217         req->r_mtime = aio_req->mtime;
1218         req->r_data_offset = req->r_ops[0].extent.offset;
1219
1220         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1221         if (ret) {
1222                 ceph_osdc_put_request(req);
1223                 req = orig_req;
1224                 goto out;
1225         }
1226
1227         ceph_osdc_put_request(orig_req);
1228
1229         req->r_callback = ceph_aio_complete_req;
1230         req->r_inode = inode;
1231         req->r_priv = aio_req;
1232
1233         ret = ceph_osdc_start_request(req->r_osdc, req, false);
1234 out:
1235         if (ret < 0) {
1236                 req->r_result = ret;
1237                 ceph_aio_complete_req(req);
1238         }
1239
1240         ceph_put_snap_context(snapc);
1241         kfree(aio_work);
1242 }
1243
1244 static ssize_t
1245 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1246                        struct ceph_snap_context *snapc,
1247                        struct ceph_cap_flush **pcf)
1248 {
1249         struct file *file = iocb->ki_filp;
1250         struct inode *inode = file_inode(file);
1251         struct ceph_inode_info *ci = ceph_inode(inode);
1252         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1253         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1254         struct ceph_vino vino;
1255         struct ceph_osd_request *req;
1256         struct bio_vec *bvecs;
1257         struct ceph_aio_request *aio_req = NULL;
1258         int num_pages = 0;
1259         int flags;
1260         int ret = 0;
1261         struct timespec64 mtime = current_time(inode);
1262         size_t count = iov_iter_count(iter);
1263         loff_t pos = iocb->ki_pos;
1264         bool write = iov_iter_rw(iter) == WRITE;
1265         bool should_dirty = !write && iter_is_iovec(iter);
1266
1267         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1268                 return -EROFS;
1269
1270         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1271              (write ? "write" : "read"), file, pos, (unsigned)count,
1272              snapc, snapc ? snapc->seq : 0);
1273
1274         if (write) {
1275                 int ret2;
1276
1277                 ceph_fscache_invalidate(inode, true);
1278
1279                 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1280                                         pos >> PAGE_SHIFT,
1281                                         (pos + count - 1) >> PAGE_SHIFT);
1282                 if (ret2 < 0)
1283                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1284
1285                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1286         } else {
1287                 flags = CEPH_OSD_FLAG_READ;
1288         }
1289
1290         while (iov_iter_count(iter) > 0) {
1291                 u64 size = iov_iter_count(iter);
1292                 ssize_t len;
1293
1294                 if (write)
1295                         size = min_t(u64, size, fsc->mount_options->wsize);
1296                 else
1297                         size = min_t(u64, size, fsc->mount_options->rsize);
1298
1299                 vino = ceph_vino(inode);
1300                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1301                                             vino, pos, &size, 0,
1302                                             1,
1303                                             write ? CEPH_OSD_OP_WRITE :
1304                                                     CEPH_OSD_OP_READ,
1305                                             flags, snapc,
1306                                             ci->i_truncate_seq,
1307                                             ci->i_truncate_size,
1308                                             false);
1309                 if (IS_ERR(req)) {
1310                         ret = PTR_ERR(req);
1311                         break;
1312                 }
1313
1314                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1315                 if (len < 0) {
1316                         ceph_osdc_put_request(req);
1317                         ret = len;
1318                         break;
1319                 }
1320                 if (len != size)
1321                         osd_req_op_extent_update(req, 0, len);
1322
1323                 /*
1324                  * To simplify error handling, allow AIO when IO within i_size
1325                  * or IO can be satisfied by single OSD request.
1326                  */
1327                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1328                     (len == count || pos + count <= i_size_read(inode))) {
1329                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1330                         if (aio_req) {
1331                                 aio_req->iocb = iocb;
1332                                 aio_req->write = write;
1333                                 aio_req->should_dirty = should_dirty;
1334                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1335                                 if (write) {
1336                                         aio_req->mtime = mtime;
1337                                         swap(aio_req->prealloc_cf, *pcf);
1338                                 }
1339                         }
1340                         /* ignore error */
1341                 }
1342
1343                 if (write) {
1344                         /*
1345                          * throw out any page cache pages in this range. this
1346                          * may block.
1347                          */
1348                         truncate_inode_pages_range(inode->i_mapping, pos,
1349                                                    PAGE_ALIGN(pos + len) - 1);
1350
1351                         req->r_mtime = mtime;
1352                 }
1353
1354                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1355
1356                 if (aio_req) {
1357                         aio_req->total_len += len;
1358                         aio_req->num_reqs++;
1359                         atomic_inc(&aio_req->pending_reqs);
1360
1361                         req->r_callback = ceph_aio_complete_req;
1362                         req->r_inode = inode;
1363                         req->r_priv = aio_req;
1364                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1365
1366                         pos += len;
1367                         continue;
1368                 }
1369
1370                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1371                 if (!ret)
1372                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1373
1374                 if (write)
1375                         ceph_update_write_metrics(metric, req->r_start_latency,
1376                                                   req->r_end_latency, len, ret);
1377                 else
1378                         ceph_update_read_metrics(metric, req->r_start_latency,
1379                                                  req->r_end_latency, len, ret);
1380
1381                 size = i_size_read(inode);
1382                 if (!write) {
1383                         if (ret == -ENOENT)
1384                                 ret = 0;
1385                         if (ret >= 0 && ret < len && pos + ret < size) {
1386                                 struct iov_iter i;
1387                                 int zlen = min_t(size_t, len - ret,
1388                                                  size - pos - ret);
1389
1390                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1391                                 iov_iter_advance(&i, ret);
1392                                 iov_iter_zero(zlen, &i);
1393                                 ret += zlen;
1394                         }
1395                         if (ret >= 0)
1396                                 len = ret;
1397                 }
1398
1399                 put_bvecs(bvecs, num_pages, should_dirty);
1400                 ceph_osdc_put_request(req);
1401                 if (ret < 0)
1402                         break;
1403
1404                 pos += len;
1405                 if (!write && pos >= size)
1406                         break;
1407
1408                 if (write && pos > size) {
1409                         if (ceph_inode_set_size(inode, pos))
1410                                 ceph_check_caps(ceph_inode(inode),
1411                                                 CHECK_CAPS_AUTHONLY,
1412                                                 NULL);
1413                 }
1414         }
1415
1416         if (aio_req) {
1417                 LIST_HEAD(osd_reqs);
1418
1419                 if (aio_req->num_reqs == 0) {
1420                         kfree(aio_req);
1421                         return ret;
1422                 }
1423
1424                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1425                                               CEPH_CAP_FILE_RD);
1426
1427                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1428                 inode_dio_begin(inode);
1429                 while (!list_empty(&osd_reqs)) {
1430                         req = list_first_entry(&osd_reqs,
1431                                                struct ceph_osd_request,
1432                                                r_private_item);
1433                         list_del_init(&req->r_private_item);
1434                         if (ret >= 0)
1435                                 ret = ceph_osdc_start_request(req->r_osdc,
1436                                                               req, false);
1437                         if (ret < 0) {
1438                                 req->r_result = ret;
1439                                 ceph_aio_complete_req(req);
1440                         }
1441                 }
1442                 return -EIOCBQUEUED;
1443         }
1444
1445         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1446                 ret = pos - iocb->ki_pos;
1447                 iocb->ki_pos = pos;
1448         }
1449         return ret;
1450 }
1451
1452 /*
1453  * Synchronous write, straight from __user pointer or user pages.
1454  *
1455  * If write spans object boundary, just do multiple writes.  (For a
1456  * correct atomic write, we should e.g. take write locks on all
1457  * objects, rollback on failure, etc.)
1458  */
1459 static ssize_t
1460 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1461                 struct ceph_snap_context *snapc)
1462 {
1463         struct file *file = iocb->ki_filp;
1464         struct inode *inode = file_inode(file);
1465         struct ceph_inode_info *ci = ceph_inode(inode);
1466         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1467         struct ceph_vino vino;
1468         struct ceph_osd_request *req;
1469         struct page **pages;
1470         u64 len;
1471         int num_pages;
1472         int written = 0;
1473         int flags;
1474         int ret;
1475         bool check_caps = false;
1476         struct timespec64 mtime = current_time(inode);
1477         size_t count = iov_iter_count(from);
1478
1479         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1480                 return -EROFS;
1481
1482         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1483              file, pos, (unsigned)count, snapc, snapc->seq);
1484
1485         ret = filemap_write_and_wait_range(inode->i_mapping,
1486                                            pos, pos + count - 1);
1487         if (ret < 0)
1488                 return ret;
1489
1490         ceph_fscache_invalidate(inode, false);
1491         ret = invalidate_inode_pages2_range(inode->i_mapping,
1492                                             pos >> PAGE_SHIFT,
1493                                             (pos + count - 1) >> PAGE_SHIFT);
1494         if (ret < 0)
1495                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1496
1497         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1498
1499         while ((len = iov_iter_count(from)) > 0) {
1500                 size_t left;
1501                 int n;
1502
1503                 vino = ceph_vino(inode);
1504                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1505                                             vino, pos, &len, 0, 1,
1506                                             CEPH_OSD_OP_WRITE, flags, snapc,
1507                                             ci->i_truncate_seq,
1508                                             ci->i_truncate_size,
1509                                             false);
1510                 if (IS_ERR(req)) {
1511                         ret = PTR_ERR(req);
1512                         break;
1513                 }
1514
1515                 /*
1516                  * write from beginning of first page,
1517                  * regardless of io alignment
1518                  */
1519                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1520
1521                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1522                 if (IS_ERR(pages)) {
1523                         ret = PTR_ERR(pages);
1524                         goto out;
1525                 }
1526
1527                 left = len;
1528                 for (n = 0; n < num_pages; n++) {
1529                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1530                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1531                         if (ret != plen) {
1532                                 ret = -EFAULT;
1533                                 break;
1534                         }
1535                         left -= ret;
1536                 }
1537
1538                 if (ret < 0) {
1539                         ceph_release_page_vector(pages, num_pages);
1540                         goto out;
1541                 }
1542
1543                 req->r_inode = inode;
1544
1545                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1546                                                 false, true);
1547
1548                 req->r_mtime = mtime;
1549                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1550                 if (!ret)
1551                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1552
1553                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1554                                           req->r_end_latency, len, ret);
1555 out:
1556                 ceph_osdc_put_request(req);
1557                 if (ret != 0) {
1558                         ceph_set_error_write(ci);
1559                         break;
1560                 }
1561
1562                 ceph_clear_error_write(ci);
1563                 pos += len;
1564                 written += len;
1565                 if (pos > i_size_read(inode)) {
1566                         check_caps = ceph_inode_set_size(inode, pos);
1567                         if (check_caps)
1568                                 ceph_check_caps(ceph_inode(inode),
1569                                                 CHECK_CAPS_AUTHONLY,
1570                                                 NULL);
1571                 }
1572
1573         }
1574
1575         if (ret != -EOLDSNAPC && written > 0) {
1576                 ret = written;
1577                 iocb->ki_pos = pos;
1578         }
1579         return ret;
1580 }
1581
1582 /*
1583  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1584  * Atomically grab references, so that those bits are not released
1585  * back to the MDS mid-read.
1586  *
1587  * Hmm, the sync read case isn't actually async... should it be?
1588  */
1589 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1590 {
1591         struct file *filp = iocb->ki_filp;
1592         struct ceph_file_info *fi = filp->private_data;
1593         size_t len = iov_iter_count(to);
1594         struct inode *inode = file_inode(filp);
1595         struct ceph_inode_info *ci = ceph_inode(inode);
1596         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1597         ssize_t ret;
1598         int want = 0, got = 0;
1599         int retry_op = 0, read = 0;
1600
1601 again:
1602         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1603              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1604
1605         if (ceph_inode_is_shutdown(inode))
1606                 return -ESTALE;
1607
1608         if (direct_lock)
1609                 ceph_start_io_direct(inode);
1610         else
1611                 ceph_start_io_read(inode);
1612
1613         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1614                 want |= CEPH_CAP_FILE_CACHE;
1615         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1616                 want |= CEPH_CAP_FILE_LAZYIO;
1617
1618         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1619         if (ret < 0) {
1620                 if (direct_lock)
1621                         ceph_end_io_direct(inode);
1622                 else
1623                         ceph_end_io_read(inode);
1624                 return ret;
1625         }
1626
1627         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1628             (iocb->ki_flags & IOCB_DIRECT) ||
1629             (fi->flags & CEPH_F_SYNC)) {
1630
1631                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1632                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1633                      ceph_cap_string(got));
1634
1635                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1636                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1637                                 ret = ceph_direct_read_write(iocb, to,
1638                                                              NULL, NULL);
1639                                 if (ret >= 0 && ret < len)
1640                                         retry_op = CHECK_EOF;
1641                         } else {
1642                                 ret = ceph_sync_read(iocb, to, &retry_op);
1643                         }
1644                 } else {
1645                         retry_op = READ_INLINE;
1646                 }
1647         } else {
1648                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1649                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1650                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1651                      ceph_cap_string(got));
1652                 ceph_add_rw_context(fi, &rw_ctx);
1653                 ret = generic_file_read_iter(iocb, to);
1654                 ceph_del_rw_context(fi, &rw_ctx);
1655         }
1656
1657         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1658              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1659         ceph_put_cap_refs(ci, got);
1660
1661         if (direct_lock)
1662                 ceph_end_io_direct(inode);
1663         else
1664                 ceph_end_io_read(inode);
1665
1666         if (retry_op > HAVE_RETRIED && ret >= 0) {
1667                 int statret;
1668                 struct page *page = NULL;
1669                 loff_t i_size;
1670                 if (retry_op == READ_INLINE) {
1671                         page = __page_cache_alloc(GFP_KERNEL);
1672                         if (!page)
1673                                 return -ENOMEM;
1674                 }
1675
1676                 statret = __ceph_do_getattr(inode, page,
1677                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1678                 if (statret < 0) {
1679                         if (page)
1680                                 __free_page(page);
1681                         if (statret == -ENODATA) {
1682                                 BUG_ON(retry_op != READ_INLINE);
1683                                 goto again;
1684                         }
1685                         return statret;
1686                 }
1687
1688                 i_size = i_size_read(inode);
1689                 if (retry_op == READ_INLINE) {
1690                         BUG_ON(ret > 0 || read > 0);
1691                         if (iocb->ki_pos < i_size &&
1692                             iocb->ki_pos < PAGE_SIZE) {
1693                                 loff_t end = min_t(loff_t, i_size,
1694                                                    iocb->ki_pos + len);
1695                                 end = min_t(loff_t, end, PAGE_SIZE);
1696                                 if (statret < end)
1697                                         zero_user_segment(page, statret, end);
1698                                 ret = copy_page_to_iter(page,
1699                                                 iocb->ki_pos & ~PAGE_MASK,
1700                                                 end - iocb->ki_pos, to);
1701                                 iocb->ki_pos += ret;
1702                                 read += ret;
1703                         }
1704                         if (iocb->ki_pos < i_size && read < len) {
1705                                 size_t zlen = min_t(size_t, len - read,
1706                                                     i_size - iocb->ki_pos);
1707                                 ret = iov_iter_zero(zlen, to);
1708                                 iocb->ki_pos += ret;
1709                                 read += ret;
1710                         }
1711                         __free_pages(page, 0);
1712                         return read;
1713                 }
1714
1715                 /* hit EOF or hole? */
1716                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1717                     ret < len) {
1718                         dout("sync_read hit hole, ppos %lld < size %lld"
1719                              ", reading more\n", iocb->ki_pos, i_size);
1720
1721                         read += ret;
1722                         len -= ret;
1723                         retry_op = HAVE_RETRIED;
1724                         goto again;
1725                 }
1726         }
1727
1728         if (ret >= 0)
1729                 ret += read;
1730
1731         return ret;
1732 }
1733
1734 /*
1735  * Take cap references to avoid releasing caps to MDS mid-write.
1736  *
1737  * If we are synchronous, and write with an old snap context, the OSD
1738  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1739  * dropping our cap refs and allowing the pending snap to logically
1740  * complete _before_ this write occurs.
1741  *
1742  * If we are near ENOSPC, write synchronously.
1743  */
1744 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1745 {
1746         struct file *file = iocb->ki_filp;
1747         struct ceph_file_info *fi = file->private_data;
1748         struct inode *inode = file_inode(file);
1749         struct ceph_inode_info *ci = ceph_inode(inode);
1750         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1751         struct ceph_osd_client *osdc = &fsc->client->osdc;
1752         struct ceph_cap_flush *prealloc_cf;
1753         ssize_t count, written = 0;
1754         int err, want = 0, got;
1755         bool direct_lock = false;
1756         u32 map_flags;
1757         u64 pool_flags;
1758         loff_t pos;
1759         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1760
1761         if (ceph_inode_is_shutdown(inode))
1762                 return -ESTALE;
1763
1764         if (ceph_snap(inode) != CEPH_NOSNAP)
1765                 return -EROFS;
1766
1767         prealloc_cf = ceph_alloc_cap_flush();
1768         if (!prealloc_cf)
1769                 return -ENOMEM;
1770
1771         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1772                 direct_lock = true;
1773
1774 retry_snap:
1775         if (direct_lock)
1776                 ceph_start_io_direct(inode);
1777         else
1778                 ceph_start_io_write(inode);
1779
1780         /* We can write back this queue in page reclaim */
1781         current->backing_dev_info = inode_to_bdi(inode);
1782
1783         if (iocb->ki_flags & IOCB_APPEND) {
1784                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1785                 if (err < 0)
1786                         goto out;
1787         }
1788
1789         err = generic_write_checks(iocb, from);
1790         if (err <= 0)
1791                 goto out;
1792
1793         pos = iocb->ki_pos;
1794         if (unlikely(pos >= limit)) {
1795                 err = -EFBIG;
1796                 goto out;
1797         } else {
1798                 iov_iter_truncate(from, limit - pos);
1799         }
1800
1801         count = iov_iter_count(from);
1802         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1803                 err = -EDQUOT;
1804                 goto out;
1805         }
1806
1807         down_read(&osdc->lock);
1808         map_flags = osdc->osdmap->flags;
1809         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1810         up_read(&osdc->lock);
1811         if ((map_flags & CEPH_OSDMAP_FULL) ||
1812             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1813                 err = -ENOSPC;
1814                 goto out;
1815         }
1816
1817         err = file_remove_privs(file);
1818         if (err)
1819                 goto out;
1820
1821         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1822              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1823         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1824                 want |= CEPH_CAP_FILE_BUFFER;
1825         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1826                 want |= CEPH_CAP_FILE_LAZYIO;
1827         got = 0;
1828         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1829         if (err < 0)
1830                 goto out;
1831
1832         err = file_update_time(file);
1833         if (err)
1834                 goto out_caps;
1835
1836         inode_inc_iversion_raw(inode);
1837
1838         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1839              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1840
1841         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1842             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1843             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1844                 struct ceph_snap_context *snapc;
1845                 struct iov_iter data;
1846
1847                 spin_lock(&ci->i_ceph_lock);
1848                 if (__ceph_have_pending_cap_snap(ci)) {
1849                         struct ceph_cap_snap *capsnap =
1850                                         list_last_entry(&ci->i_cap_snaps,
1851                                                         struct ceph_cap_snap,
1852                                                         ci_item);
1853                         snapc = ceph_get_snap_context(capsnap->context);
1854                 } else {
1855                         BUG_ON(!ci->i_head_snapc);
1856                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1857                 }
1858                 spin_unlock(&ci->i_ceph_lock);
1859
1860                 /* we might need to revert back to that point */
1861                 data = *from;
1862                 if (iocb->ki_flags & IOCB_DIRECT)
1863                         written = ceph_direct_read_write(iocb, &data, snapc,
1864                                                          &prealloc_cf);
1865                 else
1866                         written = ceph_sync_write(iocb, &data, pos, snapc);
1867                 if (direct_lock)
1868                         ceph_end_io_direct(inode);
1869                 else
1870                         ceph_end_io_write(inode);
1871                 if (written > 0)
1872                         iov_iter_advance(from, written);
1873                 ceph_put_snap_context(snapc);
1874         } else {
1875                 /*
1876                  * No need to acquire the i_truncate_mutex. Because
1877                  * the MDS revokes Fwb caps before sending truncate
1878                  * message to us. We can't get Fwb cap while there
1879                  * are pending vmtruncate. So write and vmtruncate
1880                  * can not run at the same time
1881                  */
1882                 written = generic_perform_write(iocb, from);
1883                 if (likely(written >= 0))
1884                         iocb->ki_pos = pos + written;
1885                 ceph_end_io_write(inode);
1886         }
1887
1888         if (written >= 0) {
1889                 int dirty;
1890
1891                 spin_lock(&ci->i_ceph_lock);
1892                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1893                                                &prealloc_cf);
1894                 spin_unlock(&ci->i_ceph_lock);
1895                 if (dirty)
1896                         __mark_inode_dirty(inode, dirty);
1897                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1898                         ceph_check_caps(ci, 0, NULL);
1899         }
1900
1901         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1902              inode, ceph_vinop(inode), pos, (unsigned)count,
1903              ceph_cap_string(got));
1904         ceph_put_cap_refs(ci, got);
1905
1906         if (written == -EOLDSNAPC) {
1907                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1908                      inode, ceph_vinop(inode), pos, (unsigned)count);
1909                 goto retry_snap;
1910         }
1911
1912         if (written >= 0) {
1913                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1914                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1915                         iocb->ki_flags |= IOCB_DSYNC;
1916                 written = generic_write_sync(iocb, written);
1917         }
1918
1919         goto out_unlocked;
1920 out_caps:
1921         ceph_put_cap_refs(ci, got);
1922 out:
1923         if (direct_lock)
1924                 ceph_end_io_direct(inode);
1925         else
1926                 ceph_end_io_write(inode);
1927 out_unlocked:
1928         ceph_free_cap_flush(prealloc_cf);
1929         current->backing_dev_info = NULL;
1930         return written ? written : err;
1931 }
1932
1933 /*
1934  * llseek.  be sure to verify file size on SEEK_END.
1935  */
1936 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1937 {
1938         struct inode *inode = file->f_mapping->host;
1939         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1940         loff_t i_size;
1941         loff_t ret;
1942
1943         inode_lock(inode);
1944
1945         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1946                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1947                 if (ret < 0)
1948                         goto out;
1949         }
1950
1951         i_size = i_size_read(inode);
1952         switch (whence) {
1953         case SEEK_END:
1954                 offset += i_size;
1955                 break;
1956         case SEEK_CUR:
1957                 /*
1958                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1959                  * position-querying operation.  Avoid rewriting the "same"
1960                  * f_pos value back to the file because a concurrent read(),
1961                  * write() or lseek() might have altered it
1962                  */
1963                 if (offset == 0) {
1964                         ret = file->f_pos;
1965                         goto out;
1966                 }
1967                 offset += file->f_pos;
1968                 break;
1969         case SEEK_DATA:
1970                 if (offset < 0 || offset >= i_size) {
1971                         ret = -ENXIO;
1972                         goto out;
1973                 }
1974                 break;
1975         case SEEK_HOLE:
1976                 if (offset < 0 || offset >= i_size) {
1977                         ret = -ENXIO;
1978                         goto out;
1979                 }
1980                 offset = i_size;
1981                 break;
1982         }
1983
1984         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1985
1986 out:
1987         inode_unlock(inode);
1988         return ret;
1989 }
1990
1991 static inline void ceph_zero_partial_page(
1992         struct inode *inode, loff_t offset, unsigned size)
1993 {
1994         struct page *page;
1995         pgoff_t index = offset >> PAGE_SHIFT;
1996
1997         page = find_lock_page(inode->i_mapping, index);
1998         if (page) {
1999                 wait_on_page_writeback(page);
2000                 zero_user(page, offset & (PAGE_SIZE - 1), size);
2001                 unlock_page(page);
2002                 put_page(page);
2003         }
2004 }
2005
2006 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2007                                       loff_t length)
2008 {
2009         loff_t nearly = round_up(offset, PAGE_SIZE);
2010         if (offset < nearly) {
2011                 loff_t size = nearly - offset;
2012                 if (length < size)
2013                         size = length;
2014                 ceph_zero_partial_page(inode, offset, size);
2015                 offset += size;
2016                 length -= size;
2017         }
2018         if (length >= PAGE_SIZE) {
2019                 loff_t size = round_down(length, PAGE_SIZE);
2020                 truncate_pagecache_range(inode, offset, offset + size - 1);
2021                 offset += size;
2022                 length -= size;
2023         }
2024         if (length)
2025                 ceph_zero_partial_page(inode, offset, length);
2026 }
2027
2028 static int ceph_zero_partial_object(struct inode *inode,
2029                                     loff_t offset, loff_t *length)
2030 {
2031         struct ceph_inode_info *ci = ceph_inode(inode);
2032         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2033         struct ceph_osd_request *req;
2034         int ret = 0;
2035         loff_t zero = 0;
2036         int op;
2037
2038         if (!length) {
2039                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2040                 length = &zero;
2041         } else {
2042                 op = CEPH_OSD_OP_ZERO;
2043         }
2044
2045         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2046                                         ceph_vino(inode),
2047                                         offset, length,
2048                                         0, 1, op,
2049                                         CEPH_OSD_FLAG_WRITE,
2050                                         NULL, 0, 0, false);
2051         if (IS_ERR(req)) {
2052                 ret = PTR_ERR(req);
2053                 goto out;
2054         }
2055
2056         req->r_mtime = inode->i_mtime;
2057         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
2058         if (!ret) {
2059                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2060                 if (ret == -ENOENT)
2061                         ret = 0;
2062         }
2063         ceph_osdc_put_request(req);
2064
2065 out:
2066         return ret;
2067 }
2068
2069 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2070 {
2071         int ret = 0;
2072         struct ceph_inode_info *ci = ceph_inode(inode);
2073         s32 stripe_unit = ci->i_layout.stripe_unit;
2074         s32 stripe_count = ci->i_layout.stripe_count;
2075         s32 object_size = ci->i_layout.object_size;
2076         u64 object_set_size = object_size * stripe_count;
2077         u64 nearly, t;
2078
2079         /* round offset up to next period boundary */
2080         nearly = offset + object_set_size - 1;
2081         t = nearly;
2082         nearly -= do_div(t, object_set_size);
2083
2084         while (length && offset < nearly) {
2085                 loff_t size = length;
2086                 ret = ceph_zero_partial_object(inode, offset, &size);
2087                 if (ret < 0)
2088                         return ret;
2089                 offset += size;
2090                 length -= size;
2091         }
2092         while (length >= object_set_size) {
2093                 int i;
2094                 loff_t pos = offset;
2095                 for (i = 0; i < stripe_count; ++i) {
2096                         ret = ceph_zero_partial_object(inode, pos, NULL);
2097                         if (ret < 0)
2098                                 return ret;
2099                         pos += stripe_unit;
2100                 }
2101                 offset += object_set_size;
2102                 length -= object_set_size;
2103         }
2104         while (length) {
2105                 loff_t size = length;
2106                 ret = ceph_zero_partial_object(inode, offset, &size);
2107                 if (ret < 0)
2108                         return ret;
2109                 offset += size;
2110                 length -= size;
2111         }
2112         return ret;
2113 }
2114
2115 static long ceph_fallocate(struct file *file, int mode,
2116                                 loff_t offset, loff_t length)
2117 {
2118         struct ceph_file_info *fi = file->private_data;
2119         struct inode *inode = file_inode(file);
2120         struct ceph_inode_info *ci = ceph_inode(inode);
2121         struct ceph_cap_flush *prealloc_cf;
2122         int want, got = 0;
2123         int dirty;
2124         int ret = 0;
2125         loff_t endoff = 0;
2126         loff_t size;
2127
2128         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2129                 return -EOPNOTSUPP;
2130
2131         if (!S_ISREG(inode->i_mode))
2132                 return -EOPNOTSUPP;
2133
2134         prealloc_cf = ceph_alloc_cap_flush();
2135         if (!prealloc_cf)
2136                 return -ENOMEM;
2137
2138         inode_lock(inode);
2139
2140         if (ceph_snap(inode) != CEPH_NOSNAP) {
2141                 ret = -EROFS;
2142                 goto unlock;
2143         }
2144
2145         size = i_size_read(inode);
2146
2147         /* Are we punching a hole beyond EOF? */
2148         if (offset >= size)
2149                 goto unlock;
2150         if ((offset + length) > size)
2151                 length = size - offset;
2152
2153         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2154                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2155         else
2156                 want = CEPH_CAP_FILE_BUFFER;
2157
2158         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2159         if (ret < 0)
2160                 goto unlock;
2161
2162         filemap_invalidate_lock(inode->i_mapping);
2163         ceph_fscache_invalidate(inode, false);
2164         ceph_zero_pagecache_range(inode, offset, length);
2165         ret = ceph_zero_objects(inode, offset, length);
2166
2167         if (!ret) {
2168                 spin_lock(&ci->i_ceph_lock);
2169                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2170                                                &prealloc_cf);
2171                 spin_unlock(&ci->i_ceph_lock);
2172                 if (dirty)
2173                         __mark_inode_dirty(inode, dirty);
2174         }
2175         filemap_invalidate_unlock(inode->i_mapping);
2176
2177         ceph_put_cap_refs(ci, got);
2178 unlock:
2179         inode_unlock(inode);
2180         ceph_free_cap_flush(prealloc_cf);
2181         return ret;
2182 }
2183
2184 /*
2185  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2186  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2187  * this fails; zero is returned on success.
2188  */
2189 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2190                           struct file *dst_filp,
2191                           loff_t dst_endoff, int *dst_got)
2192 {
2193         int ret = 0;
2194         bool retrying = false;
2195
2196 retry_caps:
2197         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2198                             dst_endoff, dst_got);
2199         if (ret < 0)
2200                 return ret;
2201
2202         /*
2203          * Since we're already holding the FILE_WR capability for the dst file,
2204          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2205          * retry dance instead to try to get both capabilities.
2206          */
2207         ret = ceph_try_get_caps(file_inode(src_filp),
2208                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2209                                 false, src_got);
2210         if (ret <= 0) {
2211                 /* Start by dropping dst_ci caps and getting src_ci caps */
2212                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2213                 if (retrying) {
2214                         if (!ret)
2215                                 /* ceph_try_get_caps masks EAGAIN */
2216                                 ret = -EAGAIN;
2217                         return ret;
2218                 }
2219                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2220                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2221                 if (ret < 0)
2222                         return ret;
2223                 /*... drop src_ci caps too, and retry */
2224                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2225                 retrying = true;
2226                 goto retry_caps;
2227         }
2228         return ret;
2229 }
2230
2231 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2232                            struct ceph_inode_info *dst_ci, int dst_got)
2233 {
2234         ceph_put_cap_refs(src_ci, src_got);
2235         ceph_put_cap_refs(dst_ci, dst_got);
2236 }
2237
2238 /*
2239  * This function does several size-related checks, returning an error if:
2240  *  - source file is smaller than off+len
2241  *  - destination file size is not OK (inode_newsize_ok())
2242  *  - max bytes quotas is exceeded
2243  */
2244 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2245                            loff_t src_off, loff_t dst_off, size_t len)
2246 {
2247         loff_t size, endoff;
2248
2249         size = i_size_read(src_inode);
2250         /*
2251          * Don't copy beyond source file EOF.  Instead of simply setting length
2252          * to (size - src_off), just drop to VFS default implementation, as the
2253          * local i_size may be stale due to other clients writing to the source
2254          * inode.
2255          */
2256         if (src_off + len > size) {
2257                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2258                      src_off, len, size);
2259                 return -EOPNOTSUPP;
2260         }
2261         size = i_size_read(dst_inode);
2262
2263         endoff = dst_off + len;
2264         if (inode_newsize_ok(dst_inode, endoff))
2265                 return -EOPNOTSUPP;
2266
2267         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2268                 return -EDQUOT;
2269
2270         return 0;
2271 }
2272
2273 static struct ceph_osd_request *
2274 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2275                             u64 src_snapid,
2276                             struct ceph_object_id *src_oid,
2277                             struct ceph_object_locator *src_oloc,
2278                             struct ceph_object_id *dst_oid,
2279                             struct ceph_object_locator *dst_oloc,
2280                             u32 truncate_seq, u64 truncate_size)
2281 {
2282         struct ceph_osd_request *req;
2283         int ret;
2284         u32 src_fadvise_flags =
2285                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2286                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2287         u32 dst_fadvise_flags =
2288                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2289                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2290
2291         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2292         if (!req)
2293                 return ERR_PTR(-ENOMEM);
2294
2295         req->r_flags = CEPH_OSD_FLAG_WRITE;
2296
2297         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2298         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2299
2300         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2301                                         src_oid, src_oloc,
2302                                         src_fadvise_flags,
2303                                         dst_fadvise_flags,
2304                                         truncate_seq,
2305                                         truncate_size,
2306                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2307         if (ret)
2308                 goto out;
2309
2310         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2311         if (ret)
2312                 goto out;
2313
2314         return req;
2315
2316 out:
2317         ceph_osdc_put_request(req);
2318         return ERR_PTR(ret);
2319 }
2320
2321 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2322                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2323                                     struct ceph_fs_client *fsc,
2324                                     size_t len, unsigned int flags)
2325 {
2326         struct ceph_object_locator src_oloc, dst_oloc;
2327         struct ceph_object_id src_oid, dst_oid;
2328         struct ceph_osd_client *osdc;
2329         struct ceph_osd_request *req;
2330         size_t bytes = 0;
2331         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2332         u32 src_objlen, dst_objlen;
2333         u32 object_size = src_ci->i_layout.object_size;
2334         int ret;
2335
2336         src_oloc.pool = src_ci->i_layout.pool_id;
2337         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2338         dst_oloc.pool = dst_ci->i_layout.pool_id;
2339         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2340         osdc = &fsc->client->osdc;
2341
2342         while (len >= object_size) {
2343                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2344                                               object_size, &src_objnum,
2345                                               &src_objoff, &src_objlen);
2346                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2347                                               object_size, &dst_objnum,
2348                                               &dst_objoff, &dst_objlen);
2349                 ceph_oid_init(&src_oid);
2350                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2351                                 src_ci->i_vino.ino, src_objnum);
2352                 ceph_oid_init(&dst_oid);
2353                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2354                                 dst_ci->i_vino.ino, dst_objnum);
2355                 /* Do an object remote copy */
2356                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2357                                                   &src_oid, &src_oloc,
2358                                                   &dst_oid, &dst_oloc,
2359                                                   dst_ci->i_truncate_seq,
2360                                                   dst_ci->i_truncate_size);
2361                 if (IS_ERR(req))
2362                         ret = PTR_ERR(req);
2363                 else {
2364                         ceph_osdc_start_request(osdc, req, false);
2365                         ret = ceph_osdc_wait_request(osdc, req);
2366                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2367                                                      req->r_start_latency,
2368                                                      req->r_end_latency,
2369                                                      object_size, ret);
2370                         ceph_osdc_put_request(req);
2371                 }
2372                 if (ret) {
2373                         if (ret == -EOPNOTSUPP) {
2374                                 fsc->have_copy_from2 = false;
2375                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2376                         }
2377                         dout("ceph_osdc_copy_from returned %d\n", ret);
2378                         if (!bytes)
2379                                 bytes = ret;
2380                         goto out;
2381                 }
2382                 len -= object_size;
2383                 bytes += object_size;
2384                 *src_off += object_size;
2385                 *dst_off += object_size;
2386         }
2387
2388 out:
2389         ceph_oloc_destroy(&src_oloc);
2390         ceph_oloc_destroy(&dst_oloc);
2391         return bytes;
2392 }
2393
2394 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2395                                       struct file *dst_file, loff_t dst_off,
2396                                       size_t len, unsigned int flags)
2397 {
2398         struct inode *src_inode = file_inode(src_file);
2399         struct inode *dst_inode = file_inode(dst_file);
2400         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2401         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2402         struct ceph_cap_flush *prealloc_cf;
2403         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2404         loff_t size;
2405         ssize_t ret = -EIO, bytes;
2406         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2407         u32 src_objlen, dst_objlen;
2408         int src_got = 0, dst_got = 0, err, dirty;
2409
2410         if (src_inode->i_sb != dst_inode->i_sb) {
2411                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2412
2413                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2414                                       &dst_fsc->client->fsid)) {
2415                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2416                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2417                         return -EXDEV;
2418                 }
2419         }
2420         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2421                 return -EROFS;
2422
2423         /*
2424          * Some of the checks below will return -EOPNOTSUPP, which will force a
2425          * fallback to the default VFS copy_file_range implementation.  This is
2426          * desirable in several cases (for ex, the 'len' is smaller than the
2427          * size of the objects, or in cases where that would be more
2428          * efficient).
2429          */
2430
2431         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2432                 return -EOPNOTSUPP;
2433
2434         if (!src_fsc->have_copy_from2)
2435                 return -EOPNOTSUPP;
2436
2437         /*
2438          * Striped file layouts require that we copy partial objects, but the
2439          * OSD copy-from operation only supports full-object copies.  Limit
2440          * this to non-striped file layouts for now.
2441          */
2442         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2443             (src_ci->i_layout.stripe_count != 1) ||
2444             (dst_ci->i_layout.stripe_count != 1) ||
2445             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2446                 dout("Invalid src/dst files layout\n");
2447                 return -EOPNOTSUPP;
2448         }
2449
2450         if (len < src_ci->i_layout.object_size)
2451                 return -EOPNOTSUPP; /* no remote copy will be done */
2452
2453         prealloc_cf = ceph_alloc_cap_flush();
2454         if (!prealloc_cf)
2455                 return -ENOMEM;
2456
2457         /* Start by sync'ing the source and destination files */
2458         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2459         if (ret < 0) {
2460                 dout("failed to write src file (%zd)\n", ret);
2461                 goto out;
2462         }
2463         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2464         if (ret < 0) {
2465                 dout("failed to write dst file (%zd)\n", ret);
2466                 goto out;
2467         }
2468
2469         /*
2470          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2471          * clients may have dirty data in their caches.  And OSDs know nothing
2472          * about caps, so they can't safely do the remote object copies.
2473          */
2474         err = get_rd_wr_caps(src_file, &src_got,
2475                              dst_file, (dst_off + len), &dst_got);
2476         if (err < 0) {
2477                 dout("get_rd_wr_caps returned %d\n", err);
2478                 ret = -EOPNOTSUPP;
2479                 goto out;
2480         }
2481
2482         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2483         if (ret < 0)
2484                 goto out_caps;
2485
2486         /* Drop dst file cached pages */
2487         ceph_fscache_invalidate(dst_inode, false);
2488         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2489                                             dst_off >> PAGE_SHIFT,
2490                                             (dst_off + len) >> PAGE_SHIFT);
2491         if (ret < 0) {
2492                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2493                 ret = 0; /* XXX */
2494         }
2495         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2496                                       src_ci->i_layout.object_size,
2497                                       &src_objnum, &src_objoff, &src_objlen);
2498         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2499                                       dst_ci->i_layout.object_size,
2500                                       &dst_objnum, &dst_objoff, &dst_objlen);
2501         /* object-level offsets need to the same */
2502         if (src_objoff != dst_objoff) {
2503                 ret = -EOPNOTSUPP;
2504                 goto out_caps;
2505         }
2506
2507         /*
2508          * Do a manual copy if the object offset isn't object aligned.
2509          * 'src_objlen' contains the bytes left until the end of the object,
2510          * starting at the src_off
2511          */
2512         if (src_objoff) {
2513                 dout("Initial partial copy of %u bytes\n", src_objlen);
2514
2515                 /*
2516                  * we need to temporarily drop all caps as we'll be calling
2517                  * {read,write}_iter, which will get caps again.
2518                  */
2519                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2520                 ret = do_splice_direct(src_file, &src_off, dst_file,
2521                                        &dst_off, src_objlen, flags);
2522                 /* Abort on short copies or on error */
2523                 if (ret < src_objlen) {
2524                         dout("Failed partial copy (%zd)\n", ret);
2525                         goto out;
2526                 }
2527                 len -= ret;
2528                 err = get_rd_wr_caps(src_file, &src_got,
2529                                      dst_file, (dst_off + len), &dst_got);
2530                 if (err < 0)
2531                         goto out;
2532                 err = is_file_size_ok(src_inode, dst_inode,
2533                                       src_off, dst_off, len);
2534                 if (err < 0)
2535                         goto out_caps;
2536         }
2537
2538         size = i_size_read(dst_inode);
2539         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2540                                      src_fsc, len, flags);
2541         if (bytes <= 0) {
2542                 if (!ret)
2543                         ret = bytes;
2544                 goto out_caps;
2545         }
2546         dout("Copied %zu bytes out of %zu\n", bytes, len);
2547         len -= bytes;
2548         ret += bytes;
2549
2550         file_update_time(dst_file);
2551         inode_inc_iversion_raw(dst_inode);
2552
2553         if (dst_off > size) {
2554                 /* Let the MDS know about dst file size change */
2555                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2556                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2557                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2558         }
2559         /* Mark Fw dirty */
2560         spin_lock(&dst_ci->i_ceph_lock);
2561         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2562         spin_unlock(&dst_ci->i_ceph_lock);
2563         if (dirty)
2564                 __mark_inode_dirty(dst_inode, dirty);
2565
2566 out_caps:
2567         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2568
2569         /*
2570          * Do the final manual copy if we still have some bytes left, unless
2571          * there were errors in remote object copies (len >= object_size).
2572          */
2573         if (len && (len < src_ci->i_layout.object_size)) {
2574                 dout("Final partial copy of %zu bytes\n", len);
2575                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2576                                          &dst_off, len, flags);
2577                 if (bytes > 0)
2578                         ret += bytes;
2579                 else
2580                         dout("Failed partial copy (%zd)\n", bytes);
2581         }
2582
2583 out:
2584         ceph_free_cap_flush(prealloc_cf);
2585
2586         return ret;
2587 }
2588
2589 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2590                                     struct file *dst_file, loff_t dst_off,
2591                                     size_t len, unsigned int flags)
2592 {
2593         ssize_t ret;
2594
2595         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2596                                      len, flags);
2597
2598         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2599                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2600                                               dst_off, len, flags);
2601         return ret;
2602 }
2603
2604 const struct file_operations ceph_file_fops = {
2605         .open = ceph_open,
2606         .release = ceph_release,
2607         .llseek = ceph_llseek,
2608         .read_iter = ceph_read_iter,
2609         .write_iter = ceph_write_iter,
2610         .mmap = ceph_mmap,
2611         .fsync = ceph_fsync,
2612         .lock = ceph_lock,
2613         .setlease = simple_nosetlease,
2614         .flock = ceph_flock,
2615         .splice_read = generic_file_splice_read,
2616         .splice_write = iter_file_splice_write,
2617         .unlocked_ioctl = ceph_ioctl,
2618         .compat_ioctl = compat_ptr_ioctl,
2619         .fallocate      = ceph_fallocate,
2620         .copy_file_range = ceph_copy_file_range,
2621 };