fs/ceph/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3 #include <linux/ceph/striper.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/file.h>
   9 #include <linux/mount.h>
  10 #include <linux/namei.h>
  11 #include <linux/writeback.h>
  12 #include <linux/falloc.h>
  13 #include <linux/iversion.h>
  14 #include <linux/ktime.h>
  15
  16 #include "super.h"
  17 #include "mds_client.h"
  18 #include "cache.h"
  19 #include "io.h"
  20 #include "metric.h"
  21
  22 static __le32 ceph_flags_sys2wire(u32 flags)
  23 {
  24         u32 wire_flags = 0;
  25
  26         switch (flags & O_ACCMODE) {
  27         case O_RDONLY:
  28                 wire_flags |= CEPH_O_RDONLY;
  29                 break;
  30         case O_WRONLY:
  31                 wire_flags |= CEPH_O_WRONLY;
  32                 break;
  33         case O_RDWR:
  34                 wire_flags |= CEPH_O_RDWR;
  35                 break;
  36         }
  37
  38         flags &= ~O_ACCMODE;
  39
  40 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
  41
  42         ceph_sys2wire(O_CREAT);
  43         ceph_sys2wire(O_EXCL);
  44         ceph_sys2wire(O_TRUNC);
  45         ceph_sys2wire(O_DIRECTORY);
  46         ceph_sys2wire(O_NOFOLLOW);
  47
  48 #undef ceph_sys2wire
  49
  50         if (flags)
  51                 dout("unused open flags: %x\n", flags);
  52
  53         return cpu_to_le32(wire_flags);
  54 }
  55
  56 /*
  57  * Ceph file operations
  58  *
  59  * Implement basic open/close functionality, and implement
  60  * read/write.
  61  *
  62  * We implement three modes of file I/O:
  63  *  - buffered uses the generic_file_aio_{read,write} helpers
  64  *
  65  *  - synchronous is used when there is multi-client read/write
  66  *    sharing, avoids the page cache, and synchronously waits for an
  67  *    ack from the OSD.
  68  *
  69  *  - direct io takes the variant of the sync path that references
  70  *    user pages directly.
  71  *
  72  * fsync() flushes and waits on dirty pages, but just queues metadata
  73  * for writeback: since the MDS can recover size and mtime there is no
  74  * need to wait for MDS acknowledgement.
  75  */
  76
  77 /*
  78  * How many pages to get in one call to iov_iter_get_pages().  This
  79  * determines the size of the on-stack array used as a buffer.
  80  */
  81 #define ITER_GET_BVECS_PAGES    64
  82
  83 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
  84                                 struct bio_vec *bvecs)
  85 {
  86         size_t size = 0;
  87         int bvec_idx = 0;
  88
  89         if (maxsize > iov_iter_count(iter))
  90                 maxsize = iov_iter_count(iter);
  91
  92         while (size < maxsize) {
  93                 struct page *pages[ITER_GET_BVECS_PAGES];
  94                 ssize_t bytes;
  95                 size_t start;
  96                 int idx = 0;
  97
  98                 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
  99                                            ITER_GET_BVECS_PAGES, &start);
 100                 if (bytes < 0)
 101                         return size ?: bytes;
 102
 103                 size += bytes;
 104
 105                 for ( ; bytes; idx++, bvec_idx++) {
 106                         int len = min_t(int, bytes, PAGE_SIZE - start);
 107
 108                         bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
 109                         bytes -= len;
 110                         start = 0;
 111                 }
 112         }
 113
 114         return size;
 115 }
 116
 117 /*
 118  * iov_iter_get_pages() only considers one iov_iter segment, no matter
 119  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
 120  * page.
 121  *
 122  * Attempt to get up to @maxsize bytes worth of pages from @iter.
 123  * Return the number of bytes in the created bio_vec array, or an error.
 124  */
 125 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
 126                                     struct bio_vec **bvecs, int *num_bvecs)
 127 {
 128         struct bio_vec *bv;
 129         size_t orig_count = iov_iter_count(iter);
 130         ssize_t bytes;
 131         int npages;
 132
 133         iov_iter_truncate(iter, maxsize);
 134         npages = iov_iter_npages(iter, INT_MAX);
 135         iov_iter_reexpand(iter, orig_count);
 136
 137         /*
 138          * __iter_get_bvecs() may populate only part of the array -- zero it
 139          * out.
 140          */
 141         bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
 142         if (!bv)
 143                 return -ENOMEM;
 144
 145         bytes = __iter_get_bvecs(iter, maxsize, bv);
 146         if (bytes < 0) {
 147                 /*
 148                  * No pages were pinned -- just free the array.
 149                  */
 150                 kvfree(bv);
 151                 return bytes;
 152         }
 153
 154         *bvecs = bv;
 155         *num_bvecs = npages;
 156         return bytes;
 157 }
 158
 159 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
 160 {
 161         int i;
 162
 163         for (i = 0; i < num_bvecs; i++) {
 164                 if (bvecs[i].bv_page) {
 165                         if (should_dirty)
 166                                 set_page_dirty_lock(bvecs[i].bv_page);
 167                         put_page(bvecs[i].bv_page);
 168                 }
 169         }
 170         kvfree(bvecs);
 171 }
 172
 173 /*
 174  * Prepare an open request.  Preallocate ceph_cap to avoid an
 175  * inopportune ENOMEM later.
 176  */
 177 static struct ceph_mds_request *
 178 prepare_open_request(struct super_block *sb, int flags, int create_mode)
 179 {
 180         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
 181         struct ceph_mds_request *req;
 182         int want_auth = USE_ANY_MDS;
 183         int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
 184
 185         if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
 186                 want_auth = USE_AUTH_MDS;
 187
 188         req = ceph_mdsc_create_request(mdsc, op, want_auth);
 189         if (IS_ERR(req))
 190                 goto out;
 191         req->r_fmode = ceph_flags_to_mode(flags);
 192         req->r_args.open.flags = ceph_flags_sys2wire(flags);
 193         req->r_args.open.mode = cpu_to_le32(create_mode);
 194 out:
 195         return req;
 196 }
 197
 198 static int ceph_init_file_info(struct inode *inode, struct file *file,
 199                                         int fmode, bool isdir)
 200 {
 201         struct ceph_inode_info *ci = ceph_inode(inode);
 202         struct ceph_mount_options *opt =
 203                 ceph_inode_to_client(&ci->netfs.inode)->mount_options;
 204         struct ceph_file_info *fi;
 205         int ret;
 206
 207         dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
 208                         inode->i_mode, isdir ? "dir" : "regular");
 209         BUG_ON(inode->i_fop->release != ceph_release);
 210
 211         if (isdir) {
 212                 struct ceph_dir_file_info *dfi =
 213                         kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
 214                 if (!dfi)
 215                         return -ENOMEM;
 216
 217                 file->private_data = dfi;
 218                 fi = &dfi->file_info;
 219                 dfi->next_offset = 2;
 220                 dfi->readdir_cache_idx = -1;
 221         } else {
 222                 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
 223                 if (!fi)
 224                         return -ENOMEM;
 225
 226                 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
 227                         fi->flags |= CEPH_F_SYNC;
 228
 229                 file->private_data = fi;
 230         }
 231
 232         ceph_get_fmode(ci, fmode, 1);
 233         fi->fmode = fmode;
 234
 235         spin_lock_init(&fi->rw_contexts_lock);
 236         INIT_LIST_HEAD(&fi->rw_contexts);
 237         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 238
 239         if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
 240                 ret = ceph_uninline_data(file);
 241                 if (ret < 0)
 242                         goto error;
 243         }
 244
 245         return 0;
 246
 247 error:
 248         ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 249         ceph_put_fmode(ci, fi->fmode, 1);
 250         kmem_cache_free(ceph_file_cachep, fi);
 251         /* wake up anyone waiting for caps on this inode */
 252         wake_up_all(&ci->i_cap_wq);
 253         return ret;
 254 }
 255
 256 /*
 257  * initialize private struct file data.
 258  * if we fail, clean up by dropping fmode reference on the ceph_inode
 259  */
 260 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 261 {
 262         int ret = 0;
 263
 264         switch (inode->i_mode & S_IFMT) {
 265         case S_IFREG:
 266                 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
 267                 fallthrough;
 268         case S_IFDIR:
 269                 ret = ceph_init_file_info(inode, file, fmode,
 270                                                 S_ISDIR(inode->i_mode));
 271                 break;
 272
 273         case S_IFLNK:
 274                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 275                      inode->i_mode);
 276                 break;
 277
 278         default:
 279                 dout("init_file %p %p 0%o (special)\n", inode, file,
 280                      inode->i_mode);
 281                 /*
 282                  * we need to drop the open ref now, since we don't
 283                  * have .release set to ceph_release.
 284                  */
 285                 BUG_ON(inode->i_fop->release == ceph_release);
 286
 287                 /* call the proper open fop */
 288                 ret = inode->i_fop->open(inode, file);
 289         }
 290         return ret;
 291 }
 292
 293 /*
 294  * try renew caps after session gets killed.
 295  */
 296 int ceph_renew_caps(struct inode *inode, int fmode)
 297 {
 298         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 299         struct ceph_inode_info *ci = ceph_inode(inode);
 300         struct ceph_mds_request *req;
 301         int err, flags, wanted;
 302
 303         spin_lock(&ci->i_ceph_lock);
 304         __ceph_touch_fmode(ci, mdsc, fmode);
 305         wanted = __ceph_caps_file_wanted(ci);
 306         if (__ceph_is_any_real_caps(ci) &&
 307             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 308                 int issued = __ceph_caps_issued(ci, NULL);
 309                 spin_unlock(&ci->i_ceph_lock);
 310                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 311                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 312                 ceph_check_caps(ci, 0);
 313                 return 0;
 314         }
 315         spin_unlock(&ci->i_ceph_lock);
 316
 317         flags = 0;
 318         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 319                 flags = O_RDWR;
 320         else if (wanted & CEPH_CAP_FILE_RD)
 321                 flags = O_RDONLY;
 322         else if (wanted & CEPH_CAP_FILE_WR)
 323                 flags = O_WRONLY;
 324 #ifdef O_LAZY
 325         if (wanted & CEPH_CAP_FILE_LAZYIO)
 326                 flags |= O_LAZY;
 327 #endif
 328
 329         req = prepare_open_request(inode->i_sb, flags, 0);
 330         if (IS_ERR(req)) {
 331                 err = PTR_ERR(req);
 332                 goto out;
 333         }
 334
 335         req->r_inode = inode;
 336         ihold(inode);
 337         req->r_num_caps = 1;
 338
 339         err = ceph_mdsc_do_request(mdsc, NULL, req);
 340         ceph_mdsc_put_request(req);
 341 out:
 342         dout("renew caps %p open result=%d\n", inode, err);
 343         return err < 0 ? err : 0;
 344 }
 345
 346 /*
 347  * If we already have the requisite capabilities, we can satisfy
 348  * the open request locally (no need to request new caps from the
 349  * MDS).  We do, however, need to inform the MDS (asynchronously)
 350  * if our wanted caps set expands.
 351  */
 352 int ceph_open(struct inode *inode, struct file *file)
 353 {
 354         struct ceph_inode_info *ci = ceph_inode(inode);
 355         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 356         struct ceph_mds_client *mdsc = fsc->mdsc;
 357         struct ceph_mds_request *req;
 358         struct ceph_file_info *fi = file->private_data;
 359         int err;
 360         int flags, fmode, wanted;
 361
 362         if (fi) {
 363                 dout("open file %p is already opened\n", file);
 364                 return 0;
 365         }
 366
 367         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 368         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 369         if (S_ISDIR(inode->i_mode)) {
 370                 flags = O_DIRECTORY;  /* mds likes to know */
 371         } else if (S_ISREG(inode->i_mode)) {
 372                 err = fscrypt_file_open(inode, file);
 373                 if (err)
 374                         return err;
 375         }
 376
 377         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 378              ceph_vinop(inode), file, flags, file->f_flags);
 379         fmode = ceph_flags_to_mode(flags);
 380         wanted = ceph_caps_for_mode(fmode);
 381
 382         /* snapped files are read-only */
 383         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 384                 return -EROFS;
 385
 386         /* trivially open snapdir */
 387         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 388                 return ceph_init_file(inode, file, fmode);
 389         }
 390
 391         /*
 392          * No need to block if we have caps on the auth MDS (for
 393          * write) or any MDS (for read).  Update wanted set
 394          * asynchronously.
 395          */
 396         spin_lock(&ci->i_ceph_lock);
 397         if (__ceph_is_any_real_caps(ci) &&
 398             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 399                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 400                 int issued = __ceph_caps_issued(ci, NULL);
 401
 402                 dout("open %p fmode %d want %s issued %s using existing\n",
 403                      inode, fmode, ceph_cap_string(wanted),
 404                      ceph_cap_string(issued));
 405                 __ceph_touch_fmode(ci, mdsc, fmode);
 406                 spin_unlock(&ci->i_ceph_lock);
 407
 408                 /* adjust wanted? */
 409                 if ((issued & wanted) != wanted &&
 410                     (mds_wanted & wanted) != wanted &&
 411                     ceph_snap(inode) != CEPH_SNAPDIR)
 412                         ceph_check_caps(ci, 0);
 413
 414                 return ceph_init_file(inode, file, fmode);
 415         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 416                    (ci->i_snap_caps & wanted) == wanted) {
 417                 __ceph_touch_fmode(ci, mdsc, fmode);
 418                 spin_unlock(&ci->i_ceph_lock);
 419                 return ceph_init_file(inode, file, fmode);
 420         }
 421
 422         spin_unlock(&ci->i_ceph_lock);
 423
 424         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 425         req = prepare_open_request(inode->i_sb, flags, 0);
 426         if (IS_ERR(req)) {
 427                 err = PTR_ERR(req);
 428                 goto out;
 429         }
 430         req->r_inode = inode;
 431         ihold(inode);
 432
 433         req->r_num_caps = 1;
 434         err = ceph_mdsc_do_request(mdsc, NULL, req);
 435         if (!err)
 436                 err = ceph_init_file(inode, file, req->r_fmode);
 437         ceph_mdsc_put_request(req);
 438         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 439 out:
 440         return err;
 441 }
 442
 443 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
 444 static void
 445 cache_file_layout(struct inode *dst, struct inode *src)
 446 {
 447         struct ceph_inode_info *cdst = ceph_inode(dst);
 448         struct ceph_inode_info *csrc = ceph_inode(src);
 449
 450         spin_lock(&cdst->i_ceph_lock);
 451         if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
 452             !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
 453                 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
 454                         sizeof(cdst->i_cached_layout));
 455                 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
 456                                    ceph_try_get_string(csrc->i_layout.pool_ns));
 457         }
 458         spin_unlock(&cdst->i_ceph_lock);
 459 }
 460
 461 /*
 462  * Try to set up an async create. We need caps, a file layout, and inode number,
 463  * and either a lease on the dentry or complete dir info. If any of those
 464  * criteria are not satisfied, then return false and the caller can go
 465  * synchronous.
 466  */
 467 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
 468                                  struct ceph_file_layout *lo, u64 *pino)
 469 {
 470         struct ceph_inode_info *ci = ceph_inode(dir);
 471         struct ceph_dentry_info *di = ceph_dentry(dentry);
 472         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
 473         u64 ino;
 474
 475         spin_lock(&ci->i_ceph_lock);
 476         /* No auth cap means no chance for Dc caps */
 477         if (!ci->i_auth_cap)
 478                 goto no_async;
 479
 480         /* Any delegated inos? */
 481         if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
 482                 goto no_async;
 483
 484         if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
 485                 goto no_async;
 486
 487         if ((__ceph_caps_issued(ci, NULL) & want) != want)
 488                 goto no_async;
 489
 490         if (d_in_lookup(dentry)) {
 491                 if (!__ceph_dir_is_complete(ci))
 492                         goto no_async;
 493                 spin_lock(&dentry->d_lock);
 494                 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 495                 spin_unlock(&dentry->d_lock);
 496         } else if (atomic_read(&ci->i_shared_gen) !=
 497                    READ_ONCE(di->lease_shared_gen)) {
 498                 goto no_async;
 499         }
 500
 501         ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
 502         if (!ino)
 503                 goto no_async;
 504
 505         *pino = ino;
 506         ceph_take_cap_refs(ci, want, false);
 507         memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
 508         rcu_assign_pointer(lo->pool_ns,
 509                            ceph_try_get_string(ci->i_cached_layout.pool_ns));
 510         got = want;
 511 no_async:
 512         spin_unlock(&ci->i_ceph_lock);
 513         return got;
 514 }
 515
 516 static void restore_deleg_ino(struct inode *dir, u64 ino)
 517 {
 518         struct ceph_inode_info *ci = ceph_inode(dir);
 519         struct ceph_mds_session *s = NULL;
 520
 521         spin_lock(&ci->i_ceph_lock);
 522         if (ci->i_auth_cap)
 523                 s = ceph_get_mds_session(ci->i_auth_cap->session);
 524         spin_unlock(&ci->i_ceph_lock);
 525         if (s) {
 526                 int err = ceph_restore_deleg_ino(s, ino);
 527                 if (err)
 528                         pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
 529                                 ino, err);
 530                 ceph_put_mds_session(s);
 531         }
 532 }
 533
 534 static void wake_async_create_waiters(struct inode *inode,
 535                                       struct ceph_mds_session *session)
 536 {
 537         struct ceph_inode_info *ci = ceph_inode(inode);
 538         bool check_cap = false;
 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                 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
 546                         ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
 547                         check_cap = true;
 548                 }
 549         }
 550         ceph_kick_flushing_inode_caps(session, ci);
 551         spin_unlock(&ci->i_ceph_lock);
 552
 553         if (check_cap)
 554                 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
 555 }
 556
 557 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
 558                                  struct ceph_mds_request *req)
 559 {
 560         struct dentry *dentry = req->r_dentry;
 561         struct inode *dinode = d_inode(dentry);
 562         struct inode *tinode = req->r_target_inode;
 563         int result = req->r_err ? req->r_err :
 564                         le32_to_cpu(req->r_reply_info.head->result);
 565
 566         WARN_ON_ONCE(dinode && tinode && dinode != tinode);
 567
 568         /* MDS changed -- caller must resubmit */
 569         if (result == -EJUKEBOX)
 570                 goto out;
 571
 572         mapping_set_error(req->r_parent->i_mapping, result);
 573
 574         if (result) {
 575                 int pathlen = 0;
 576                 u64 base = 0;
 577                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
 578                                                   &base, 0);
 579
 580                 pr_warn("async create failure path=(%llx)%s result=%d!\n",
 581                         base, IS_ERR(path) ? "<<bad>>" : path, result);
 582                 ceph_mdsc_free_path(path, pathlen);
 583
 584                 ceph_dir_clear_complete(req->r_parent);
 585                 if (!d_unhashed(dentry))
 586                         d_drop(dentry);
 587
 588                 if (dinode) {
 589                         mapping_set_error(dinode->i_mapping, result);
 590                         ceph_inode_shutdown(dinode);
 591                         wake_async_create_waiters(dinode, req->r_session);
 592                 }
 593         }
 594
 595         if (tinode) {
 596                 u64 ino = ceph_vino(tinode).ino;
 597
 598                 if (req->r_deleg_ino != ino)
 599                         pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
 600                                 __func__, req->r_err, req->r_deleg_ino, ino);
 601
 602                 mapping_set_error(tinode->i_mapping, result);
 603                 wake_async_create_waiters(tinode, req->r_session);
 604         } else if (!result) {
 605                 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
 606                         req->r_deleg_ino);
 607         }
 608 out:
 609         ceph_mdsc_release_dir_caps(req);
 610 }
 611
 612 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
 613                                     struct dentry *dentry,
 614                                     struct file *file, umode_t mode,
 615                                     struct ceph_mds_request *req,
 616                                     struct ceph_acl_sec_ctx *as_ctx,
 617                                     struct ceph_file_layout *lo)
 618 {
 619         int ret;
 620         char xattr_buf[4];
 621         struct ceph_mds_reply_inode in = { };
 622         struct ceph_mds_reply_info_in iinfo = { .in = &in };
 623         struct ceph_inode_info *ci = ceph_inode(dir);
 624         struct ceph_dentry_info *di = ceph_dentry(dentry);
 625         struct timespec64 now;
 626         struct ceph_string *pool_ns;
 627         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 628         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 629                                   .snap = CEPH_NOSNAP };
 630
 631         ktime_get_real_ts64(&now);
 632
 633         iinfo.inline_version = CEPH_INLINE_NONE;
 634         iinfo.change_attr = 1;
 635         ceph_encode_timespec64(&iinfo.btime, &now);
 636
 637         if (req->r_pagelist) {
 638                 iinfo.xattr_len = req->r_pagelist->length;
 639                 iinfo.xattr_data = req->r_pagelist->mapped_tail;
 640         } else {
 641                 /* fake it */
 642                 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
 643                 iinfo.xattr_data = xattr_buf;
 644                 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
 645         }
 646
 647         in.ino = cpu_to_le64(vino.ino);
 648         in.snapid = cpu_to_le64(CEPH_NOSNAP);
 649         in.version = cpu_to_le64(1);    // ???
 650         in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
 651         in.cap.cap_id = cpu_to_le64(1);
 652         in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
 653         in.cap.flags = CEPH_CAP_FLAG_AUTH;
 654         in.ctime = in.mtime = in.atime = iinfo.btime;
 655         in.truncate_seq = cpu_to_le32(1);
 656         in.truncate_size = cpu_to_le64(-1ULL);
 657         in.xattr_version = cpu_to_le64(1);
 658         in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
 659         if (dir->i_mode & S_ISGID) {
 660                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
 661
 662                 /* Directories always inherit the setgid bit. */
 663                 if (S_ISDIR(mode))
 664                         mode |= S_ISGID;
 665         } else {
 666                 in.gid = cpu_to_le32(from_kgid(&init_user_ns, current_fsgid()));
 667         }
 668         in.mode = cpu_to_le32((u32)mode);
 669
 670         in.nlink = cpu_to_le32(1);
 671         in.max_size = cpu_to_le64(lo->stripe_unit);
 672
 673         ceph_file_layout_to_legacy(lo, &in.layout);
 674         /* lo is private, so pool_ns can't change */
 675         pool_ns = rcu_dereference_raw(lo->pool_ns);
 676         if (pool_ns) {
 677                 iinfo.pool_ns_len = pool_ns->len;
 678                 iinfo.pool_ns_data = pool_ns->str;
 679         }
 680
 681         down_read(&mdsc->snap_rwsem);
 682         ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
 683                               req->r_fmode, NULL);
 684         up_read(&mdsc->snap_rwsem);
 685         if (ret) {
 686                 dout("%s failed to fill inode: %d\n", __func__, ret);
 687                 ceph_dir_clear_complete(dir);
 688                 if (!d_unhashed(dentry))
 689                         d_drop(dentry);
 690                 discard_new_inode(inode);
 691         } else {
 692                 struct dentry *dn;
 693
 694                 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
 695                         vino.ino, ceph_ino(dir), dentry->d_name.name);
 696                 ceph_dir_clear_ordered(dir);
 697                 ceph_init_inode_acls(inode, as_ctx);
 698                 if (inode->i_state & I_NEW) {
 699                         /*
 700                          * If it's not I_NEW, then someone created this before
 701                          * we got here. Assume the server is aware of it at
 702                          * that point and don't worry about setting
 703                          * CEPH_I_ASYNC_CREATE.
 704                          */
 705                         ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
 706                         unlock_new_inode(inode);
 707                 }
 708                 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
 709                         if (!d_unhashed(dentry))
 710                                 d_drop(dentry);
 711                         dn = d_splice_alias(inode, dentry);
 712                         WARN_ON_ONCE(dn && dn != dentry);
 713                 }
 714                 file->f_mode |= FMODE_CREATED;
 715                 ret = finish_open(file, dentry, ceph_open);
 716         }
 717
 718         spin_lock(&dentry->d_lock);
 719         di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
 720         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
 721         spin_unlock(&dentry->d_lock);
 722
 723         return ret;
 724 }
 725
 726 /*
 727  * Do a lookup + open with a single request.  If we get a non-existent
 728  * file or symlink, return 1 so the VFS can retry.
 729  */
 730 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 731                      struct file *file, unsigned flags, umode_t mode)
 732 {
 733         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 734         struct ceph_mds_client *mdsc = fsc->mdsc;
 735         struct ceph_mds_request *req;
 736         struct inode *new_inode = NULL;
 737         struct dentry *dn;
 738         struct ceph_acl_sec_ctx as_ctx = {};
 739         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
 740         int mask;
 741         int err;
 742
 743         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 744              dir, dentry, dentry,
 745              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 746
 747         if (dentry->d_name.len > NAME_MAX)
 748                 return -ENAMETOOLONG;
 749
 750         err = ceph_wait_on_conflict_unlink(dentry);
 751         if (err)
 752                 return err;
 753         /*
 754          * Do not truncate the file, since atomic_open is called before the
 755          * permission check. The caller will do the truncation afterward.
 756          */
 757         flags &= ~O_TRUNC;
 758
 759 retry:
 760         if (flags & O_CREAT) {
 761                 if (ceph_quota_is_max_files_exceeded(dir))
 762                         return -EDQUOT;
 763
 764                 new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 765                 if (IS_ERR(new_inode)) {
 766                         err = PTR_ERR(new_inode);
 767                         goto out_ctx;
 768                 }
 769                 /* Async create can't handle more than a page of xattrs */
 770                 if (as_ctx.pagelist &&
 771                     !list_is_singular(&as_ctx.pagelist->head))
 772                         try_async = false;
 773         } else if (!d_in_lookup(dentry)) {
 774                 /* If it's not being looked up, it's negative */
 775                 return -ENOENT;
 776         }
 777
 778         /* do the open */
 779         req = prepare_open_request(dir->i_sb, flags, mode);
 780         if (IS_ERR(req)) {
 781                 err = PTR_ERR(req);
 782                 goto out_ctx;
 783         }
 784         req->r_dentry = dget(dentry);
 785         req->r_num_caps = 2;
 786         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 787         if (ceph_security_xattr_wanted(dir))
 788                 mask |= CEPH_CAP_XATTR_SHARED;
 789         req->r_args.open.mask = cpu_to_le32(mask);
 790         req->r_parent = dir;
 791         ihold(dir);
 792         if (IS_ENCRYPTED(dir)) {
 793                 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
 794                 if (!fscrypt_has_encryption_key(dir)) {
 795                         spin_lock(&dentry->d_lock);
 796                         dentry->d_flags |= DCACHE_NOKEY_NAME;
 797                         spin_unlock(&dentry->d_lock);
 798                 }
 799         }
 800
 801         if (flags & O_CREAT) {
 802                 struct ceph_file_layout lo;
 803
 804                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
 805                                      CEPH_CAP_XATTR_EXCL;
 806                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 807
 808                 ceph_as_ctx_to_req(req, &as_ctx);
 809
 810                 if (try_async && (req->r_dir_caps =
 811                                   try_prep_async_create(dir, dentry, &lo,
 812                                                         &req->r_deleg_ino))) {
 813                         struct ceph_vino vino = { .ino = req->r_deleg_ino,
 814                                                   .snap = CEPH_NOSNAP };
 815                         struct ceph_dentry_info *di = ceph_dentry(dentry);
 816
 817                         set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
 818                         req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
 819                         req->r_callback = ceph_async_create_cb;
 820
 821                         /* Hash inode before RPC */
 822                         new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
 823                         if (IS_ERR(new_inode)) {
 824                                 err = PTR_ERR(new_inode);
 825                                 new_inode = NULL;
 826                                 goto out_req;
 827                         }
 828                         WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
 829
 830                         spin_lock(&dentry->d_lock);
 831                         di->flags |= CEPH_DENTRY_ASYNC_CREATE;
 832                         spin_unlock(&dentry->d_lock);
 833
 834                         err = ceph_mdsc_submit_request(mdsc, dir, req);
 835                         if (!err) {
 836                                 err = ceph_finish_async_create(dir, new_inode,
 837                                                                dentry, file,
 838                                                                mode, req,
 839                                                                &as_ctx, &lo);
 840                                 new_inode = NULL;
 841                         } else if (err == -EJUKEBOX) {
 842                                 restore_deleg_ino(dir, req->r_deleg_ino);
 843                                 ceph_mdsc_put_request(req);
 844                                 discard_new_inode(new_inode);
 845                                 ceph_release_acl_sec_ctx(&as_ctx);
 846                                 memset(&as_ctx, 0, sizeof(as_ctx));
 847                                 new_inode = NULL;
 848                                 try_async = false;
 849                                 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 850                                 goto retry;
 851                         }
 852                         ceph_put_string(rcu_dereference_raw(lo.pool_ns));
 853                         goto out_req;
 854                 }
 855         }
 856
 857         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 858         req->r_new_inode = new_inode;
 859         new_inode = NULL;
 860         err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
 861         if (err == -ENOENT) {
 862                 dentry = ceph_handle_snapdir(req, dentry);
 863                 if (IS_ERR(dentry)) {
 864                         err = PTR_ERR(dentry);
 865                         goto out_req;
 866                 }
 867                 err = 0;
 868         }
 869
 870         if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 871                 err = ceph_handle_notrace_create(dir, dentry);
 872
 873         if (d_in_lookup(dentry)) {
 874                 dn = ceph_finish_lookup(req, dentry, err);
 875                 if (IS_ERR(dn))
 876                         err = PTR_ERR(dn);
 877         } else {
 878                 /* we were given a hashed negative dentry */
 879                 dn = NULL;
 880         }
 881         if (err)
 882                 goto out_req;
 883         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 884                 /* make vfs retry on splice, ENOENT, or symlink */
 885                 dout("atomic_open finish_no_open on dn %p\n", dn);
 886                 err = finish_no_open(file, dn);
 887         } else {
 888                 if (IS_ENCRYPTED(dir) &&
 889                     !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
 890                         pr_warn("Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
 891                                 ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
 892                         goto out_req;
 893                 }
 894
 895                 dout("atomic_open finish_open on dn %p\n", dn);
 896                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 897                         struct inode *newino = d_inode(dentry);
 898
 899                         cache_file_layout(dir, newino);
 900                         ceph_init_inode_acls(newino, &as_ctx);
 901                         file->f_mode |= FMODE_CREATED;
 902                 }
 903                 err = finish_open(file, dentry, ceph_open);
 904         }
 905 out_req:
 906         ceph_mdsc_put_request(req);
 907         iput(new_inode);
 908 out_ctx:
 909         ceph_release_acl_sec_ctx(&as_ctx);
 910         dout("atomic_open result=%d\n", err);
 911         return err;
 912 }
 913
 914 int ceph_release(struct inode *inode, struct file *file)
 915 {
 916         struct ceph_inode_info *ci = ceph_inode(inode);
 917
 918         if (S_ISDIR(inode->i_mode)) {
 919                 struct ceph_dir_file_info *dfi = file->private_data;
 920                 dout("release inode %p dir file %p\n", inode, file);
 921                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 922
 923                 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
 924
 925                 if (dfi->last_readdir)
 926                         ceph_mdsc_put_request(dfi->last_readdir);
 927                 kfree(dfi->last_name);
 928                 kfree(dfi->dir_info);
 929                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 930         } else {
 931                 struct ceph_file_info *fi = file->private_data;
 932                 dout("release inode %p regular file %p\n", inode, file);
 933                 WARN_ON(!list_empty(&fi->rw_contexts));
 934
 935                 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
 936                 ceph_put_fmode(ci, fi->fmode, 1);
 937
 938                 kmem_cache_free(ceph_file_cachep, fi);
 939         }
 940
 941         /* wake up anyone waiting for caps on this inode */
 942         wake_up_all(&ci->i_cap_wq);
 943         return 0;
 944 }
 945
 946 enum {
 947         HAVE_RETRIED = 1,
 948         CHECK_EOF =    2,
 949         READ_INLINE =  3,
 950 };
 951
 952 /*
 953  * Completely synchronous read and write methods.  Direct from __user
 954  * buffer to osd, or directly to user pages (if O_DIRECT).
 955  *
 956  * If the read spans object boundary, just do multiple reads.  (That's not
 957  * atomic, but good enough for now.)
 958  *
 959  * If we get a short result from the OSD, check against i_size; we need to
 960  * only return a short read to the caller if we hit EOF.
 961  */
 962 ssize_t __ceph_sync_read(struct inode *inode, loff_t *ki_pos,
 963                          struct iov_iter *to, int *retry_op,
 964                          u64 *last_objver)
 965 {
 966         struct ceph_inode_info *ci = ceph_inode(inode);
 967         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 968         struct ceph_osd_client *osdc = &fsc->client->osdc;
 969         ssize_t ret;
 970         u64 off = *ki_pos;
 971         u64 len = iov_iter_count(to);
 972         u64 i_size = i_size_read(inode);
 973         bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
 974         u64 objver = 0;
 975
 976         dout("sync_read on inode %p %llx~%llx\n", inode, *ki_pos, len);
 977
 978         if (ceph_inode_is_shutdown(inode))
 979                 return -EIO;
 980
 981         if (!len)
 982                 return 0;
 983         /*
 984          * flush any page cache pages in this range.  this
 985          * will make concurrent normal and sync io slow,
 986          * but it will at least behave sensibly when they are
 987          * in sequence.
 988          */
 989         ret = filemap_write_and_wait_range(inode->i_mapping,
 990                                            off, off + len - 1);
 991         if (ret < 0)
 992                 return ret;
 993
 994         ret = 0;
 995         while ((len = iov_iter_count(to)) > 0) {
 996                 struct ceph_osd_request *req;
 997                 struct page **pages;
 998                 int num_pages;
 999                 size_t page_off;
1000                 bool more;
1001                 int idx;
1002                 size_t left;
1003                 struct ceph_osd_req_op *op;
1004
1005                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1006                                         ci->i_vino, off, &len, 0, 1,
1007                                         sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
1008                                         CEPH_OSD_FLAG_READ,
1009                                         NULL, ci->i_truncate_seq,
1010                                         ci->i_truncate_size, false);
1011                 if (IS_ERR(req)) {
1012                         ret = PTR_ERR(req);
1013                         break;
1014                 }
1015
1016                 more = len < iov_iter_count(to);
1017
1018                 num_pages = calc_pages_for(off, len);
1019                 page_off = off & ~PAGE_MASK;
1020                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1021                 if (IS_ERR(pages)) {
1022                         ceph_osdc_put_request(req);
1023                         ret = PTR_ERR(pages);
1024                         break;
1025                 }
1026
1027                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
1028                                                  false, false);
1029
1030                 op = &req->r_ops[0];
1031                 if (sparse) {
1032                         ret = ceph_alloc_sparse_ext_map(op);
1033                         if (ret) {
1034                                 ceph_osdc_put_request(req);
1035                                 break;
1036                         }
1037                 }
1038
1039                 ceph_osdc_start_request(osdc, req);
1040                 ret = ceph_osdc_wait_request(osdc, req);
1041
1042                 ceph_update_read_metrics(&fsc->mdsc->metric,
1043                                          req->r_start_latency,
1044                                          req->r_end_latency,
1045                                          len, ret);
1046
1047                 if (ret > 0)
1048                         objver = req->r_version;
1049
1050                 i_size = i_size_read(inode);
1051                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
1052                      off, len, ret, i_size, (more ? " MORE" : ""));
1053
1054                 /* Fix it to go to end of extent map */
1055                 if (sparse && ret >= 0)
1056                         ret = ceph_sparse_ext_map_end(op);
1057                 else if (ret == -ENOENT)
1058                         ret = 0;
1059
1060                 ceph_osdc_put_request(req);
1061
1062                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
1063                         int zlen = min(len - ret, i_size - off - ret);
1064                         int zoff = page_off + ret;
1065
1066                         dout("sync_read zero gap %llu~%llu\n",
1067                                 off + ret, off + ret + zlen);
1068                         ceph_zero_page_vector_range(zoff, zlen, pages);
1069                         ret += zlen;
1070                 }
1071
1072                 idx = 0;
1073                 left = ret > 0 ? ret : 0;
1074                 while (left > 0) {
1075                         size_t len, copied;
1076                         page_off = off & ~PAGE_MASK;
1077                         len = min_t(size_t, left, PAGE_SIZE - page_off);
1078                         SetPageUptodate(pages[idx]);
1079                         copied = copy_page_to_iter(pages[idx++],
1080                                                    page_off, len, to);
1081                         off += copied;
1082                         left -= copied;
1083                         if (copied < len) {
1084                                 ret = -EFAULT;
1085                                 break;
1086                         }
1087                 }
1088                 ceph_release_page_vector(pages, num_pages);
1089
1090                 if (ret < 0) {
1091                         if (ret == -EBLOCKLISTED)
1092                                 fsc->blocklisted = true;
1093                         break;
1094                 }
1095
1096                 if (off >= i_size || !more)
1097                         break;
1098         }
1099
1100         if (off > *ki_pos) {
1101                 if (off >= i_size) {
1102                         *retry_op = CHECK_EOF;
1103                         ret = i_size - *ki_pos;
1104                         *ki_pos = i_size;
1105                 } else {
1106                         ret = off - *ki_pos;
1107                         *ki_pos = off;
1108                 }
1109         }
1110
1111         if (last_objver && ret > 0)
1112                 *last_objver = objver;
1113
1114         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
1115         return ret;
1116 }
1117
1118 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
1119                               int *retry_op)
1120 {
1121         struct file *file = iocb->ki_filp;
1122         struct inode *inode = file_inode(file);
1123
1124         dout("sync_read on file %p %llx~%zx %s\n", file, iocb->ki_pos,
1125              iov_iter_count(to), (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
1126
1127         return __ceph_sync_read(inode, &iocb->ki_pos, to, retry_op, NULL);
1128 }
1129
1130 struct ceph_aio_request {
1131         struct kiocb *iocb;
1132         size_t total_len;
1133         bool write;
1134         bool should_dirty;
1135         int error;
1136         struct list_head osd_reqs;
1137         unsigned num_reqs;
1138         atomic_t pending_reqs;
1139         struct timespec64 mtime;
1140         struct ceph_cap_flush *prealloc_cf;
1141 };
1142
1143 struct ceph_aio_work {
1144         struct work_struct work;
1145         struct ceph_osd_request *req;
1146 };
1147
1148 static void ceph_aio_retry_work(struct work_struct *work);
1149
1150 static void ceph_aio_complete(struct inode *inode,
1151                               struct ceph_aio_request *aio_req)
1152 {
1153         struct ceph_inode_info *ci = ceph_inode(inode);
1154         int ret;
1155
1156         if (!atomic_dec_and_test(&aio_req->pending_reqs))
1157                 return;
1158
1159         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1160                 inode_dio_end(inode);
1161
1162         ret = aio_req->error;
1163         if (!ret)
1164                 ret = aio_req->total_len;
1165
1166         dout("ceph_aio_complete %p rc %d\n", inode, ret);
1167
1168         if (ret >= 0 && aio_req->write) {
1169                 int dirty;
1170
1171                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1172                 if (endoff > i_size_read(inode)) {
1173                         if (ceph_inode_set_size(inode, endoff))
1174                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1175                 }
1176
1177                 spin_lock(&ci->i_ceph_lock);
1178                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1179                                                &aio_req->prealloc_cf);
1180                 spin_unlock(&ci->i_ceph_lock);
1181                 if (dirty)
1182                         __mark_inode_dirty(inode, dirty);
1183
1184         }
1185
1186         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1187                                                 CEPH_CAP_FILE_RD));
1188
1189         aio_req->iocb->ki_complete(aio_req->iocb, ret);
1190
1191         ceph_free_cap_flush(aio_req->prealloc_cf);
1192         kfree(aio_req);
1193 }
1194
1195 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1196 {
1197         int rc = req->r_result;
1198         struct inode *inode = req->r_inode;
1199         struct ceph_aio_request *aio_req = req->r_priv;
1200         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1201         struct ceph_osd_req_op *op = &req->r_ops[0];
1202         struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1203         unsigned int len = osd_data->bvec_pos.iter.bi_size;
1204         bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1205
1206         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1207         BUG_ON(!osd_data->num_bvecs);
1208
1209         dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1210
1211         if (rc == -EOLDSNAPC) {
1212                 struct ceph_aio_work *aio_work;
1213                 BUG_ON(!aio_req->write);
1214
1215                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1216                 if (aio_work) {
1217                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1218                         aio_work->req = req;
1219                         queue_work(ceph_inode_to_client(inode)->inode_wq,
1220                                    &aio_work->work);
1221                         return;
1222                 }
1223                 rc = -ENOMEM;
1224         } else if (!aio_req->write) {
1225                 if (sparse && rc >= 0)
1226                         rc = ceph_sparse_ext_map_end(op);
1227                 if (rc == -ENOENT)
1228                         rc = 0;
1229                 if (rc >= 0 && len > rc) {
1230                         struct iov_iter i;
1231                         int zlen = len - rc;
1232
1233                         /*
1234                          * If read is satisfied by single OSD request,
1235                          * it can pass EOF. Otherwise read is within
1236                          * i_size.
1237                          */
1238                         if (aio_req->num_reqs == 1) {
1239                                 loff_t i_size = i_size_read(inode);
1240                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
1241                                 if (endoff < i_size)
1242                                         zlen = min_t(size_t, zlen,
1243                                                      i_size - endoff);
1244                                 aio_req->total_len = rc + zlen;
1245                         }
1246
1247                         iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1248                                       osd_data->num_bvecs, len);
1249                         iov_iter_advance(&i, rc);
1250                         iov_iter_zero(zlen, &i);
1251                 }
1252         }
1253
1254         /* r_start_latency == 0 means the request was not submitted */
1255         if (req->r_start_latency) {
1256                 if (aio_req->write)
1257                         ceph_update_write_metrics(metric, req->r_start_latency,
1258                                                   req->r_end_latency, len, rc);
1259                 else
1260                         ceph_update_read_metrics(metric, req->r_start_latency,
1261                                                  req->r_end_latency, len, rc);
1262         }
1263
1264         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1265                   aio_req->should_dirty);
1266         ceph_osdc_put_request(req);
1267
1268         if (rc < 0)
1269                 cmpxchg(&aio_req->error, 0, rc);
1270
1271         ceph_aio_complete(inode, aio_req);
1272         return;
1273 }
1274
1275 static void ceph_aio_retry_work(struct work_struct *work)
1276 {
1277         struct ceph_aio_work *aio_work =
1278                 container_of(work, struct ceph_aio_work, work);
1279         struct ceph_osd_request *orig_req = aio_work->req;
1280         struct ceph_aio_request *aio_req = orig_req->r_priv;
1281         struct inode *inode = orig_req->r_inode;
1282         struct ceph_inode_info *ci = ceph_inode(inode);
1283         struct ceph_snap_context *snapc;
1284         struct ceph_osd_request *req;
1285         int ret;
1286
1287         spin_lock(&ci->i_ceph_lock);
1288         if (__ceph_have_pending_cap_snap(ci)) {
1289                 struct ceph_cap_snap *capsnap =
1290                         list_last_entry(&ci->i_cap_snaps,
1291                                         struct ceph_cap_snap,
1292                                         ci_item);
1293                 snapc = ceph_get_snap_context(capsnap->context);
1294         } else {
1295                 BUG_ON(!ci->i_head_snapc);
1296                 snapc = ceph_get_snap_context(ci->i_head_snapc);
1297         }
1298         spin_unlock(&ci->i_ceph_lock);
1299
1300         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1301                         false, GFP_NOFS);
1302         if (!req) {
1303                 ret = -ENOMEM;
1304                 req = orig_req;
1305                 goto out;
1306         }
1307
1308         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1309         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1310         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1311
1312         req->r_ops[0] = orig_req->r_ops[0];
1313
1314         req->r_mtime = aio_req->mtime;
1315         req->r_data_offset = req->r_ops[0].extent.offset;
1316
1317         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1318         if (ret) {
1319                 ceph_osdc_put_request(req);
1320                 req = orig_req;
1321                 goto out;
1322         }
1323
1324         ceph_osdc_put_request(orig_req);
1325
1326         req->r_callback = ceph_aio_complete_req;
1327         req->r_inode = inode;
1328         req->r_priv = aio_req;
1329
1330         ceph_osdc_start_request(req->r_osdc, req);
1331 out:
1332         if (ret < 0) {
1333                 req->r_result = ret;
1334                 ceph_aio_complete_req(req);
1335         }
1336
1337         ceph_put_snap_context(snapc);
1338         kfree(aio_work);
1339 }
1340
1341 static ssize_t
1342 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1343                        struct ceph_snap_context *snapc,
1344                        struct ceph_cap_flush **pcf)
1345 {
1346         struct file *file = iocb->ki_filp;
1347         struct inode *inode = file_inode(file);
1348         struct ceph_inode_info *ci = ceph_inode(inode);
1349         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1350         struct ceph_client_metric *metric = &fsc->mdsc->metric;
1351         struct ceph_vino vino;
1352         struct ceph_osd_request *req;
1353         struct bio_vec *bvecs;
1354         struct ceph_aio_request *aio_req = NULL;
1355         int num_pages = 0;
1356         int flags;
1357         int ret = 0;
1358         struct timespec64 mtime = current_time(inode);
1359         size_t count = iov_iter_count(iter);
1360         loff_t pos = iocb->ki_pos;
1361         bool write = iov_iter_rw(iter) == WRITE;
1362         bool should_dirty = !write && user_backed_iter(iter);
1363         bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1364
1365         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1366                 return -EROFS;
1367
1368         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1369              (write ? "write" : "read"), file, pos, (unsigned)count,
1370              snapc, snapc ? snapc->seq : 0);
1371
1372         if (write) {
1373                 int ret2;
1374
1375                 ceph_fscache_invalidate(inode, true);
1376
1377                 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1378                                         pos >> PAGE_SHIFT,
1379                                         (pos + count - 1) >> PAGE_SHIFT);
1380                 if (ret2 < 0)
1381                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
1382
1383                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1384         } else {
1385                 flags = CEPH_OSD_FLAG_READ;
1386         }
1387
1388         while (iov_iter_count(iter) > 0) {
1389                 u64 size = iov_iter_count(iter);
1390                 ssize_t len;
1391                 struct ceph_osd_req_op *op;
1392                 int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1393
1394                 if (write)
1395                         size = min_t(u64, size, fsc->mount_options->wsize);
1396                 else
1397                         size = min_t(u64, size, fsc->mount_options->rsize);
1398
1399                 vino = ceph_vino(inode);
1400                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1401                                             vino, pos, &size, 0,
1402                                             1,
1403                                             write ? CEPH_OSD_OP_WRITE : readop,
1404                                             flags, snapc,
1405                                             ci->i_truncate_seq,
1406                                             ci->i_truncate_size,
1407                                             false);
1408                 if (IS_ERR(req)) {
1409                         ret = PTR_ERR(req);
1410                         break;
1411                 }
1412
1413                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1414                 if (len < 0) {
1415                         ceph_osdc_put_request(req);
1416                         ret = len;
1417                         break;
1418                 }
1419                 if (len != size)
1420                         osd_req_op_extent_update(req, 0, len);
1421
1422                 /*
1423                  * To simplify error handling, allow AIO when IO within i_size
1424                  * or IO can be satisfied by single OSD request.
1425                  */
1426                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1427                     (len == count || pos + count <= i_size_read(inode))) {
1428                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1429                         if (aio_req) {
1430                                 aio_req->iocb = iocb;
1431                                 aio_req->write = write;
1432                                 aio_req->should_dirty = should_dirty;
1433                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1434                                 if (write) {
1435                                         aio_req->mtime = mtime;
1436                                         swap(aio_req->prealloc_cf, *pcf);
1437                                 }
1438                         }
1439                         /* ignore error */
1440                 }
1441
1442                 if (write) {
1443                         /*
1444                          * throw out any page cache pages in this range. this
1445                          * may block.
1446                          */
1447                         truncate_inode_pages_range(inode->i_mapping, pos,
1448                                                    PAGE_ALIGN(pos + len) - 1);
1449
1450                         req->r_mtime = mtime;
1451                 }
1452
1453                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1454                 op = &req->r_ops[0];
1455                 if (sparse) {
1456                         ret = ceph_alloc_sparse_ext_map(op);
1457                         if (ret) {
1458                                 ceph_osdc_put_request(req);
1459                                 break;
1460                         }
1461                 }
1462
1463                 if (aio_req) {
1464                         aio_req->total_len += len;
1465                         aio_req->num_reqs++;
1466                         atomic_inc(&aio_req->pending_reqs);
1467
1468                         req->r_callback = ceph_aio_complete_req;
1469                         req->r_inode = inode;
1470                         req->r_priv = aio_req;
1471                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1472
1473                         pos += len;
1474                         continue;
1475                 }
1476
1477                 ceph_osdc_start_request(req->r_osdc, req);
1478                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1479
1480                 if (write)
1481                         ceph_update_write_metrics(metric, req->r_start_latency,
1482                                                   req->r_end_latency, len, ret);
1483                 else
1484                         ceph_update_read_metrics(metric, req->r_start_latency,
1485                                                  req->r_end_latency, len, ret);
1486
1487                 size = i_size_read(inode);
1488                 if (!write) {
1489                         if (sparse && ret >= 0)
1490                                 ret = ceph_sparse_ext_map_end(op);
1491                         else if (ret == -ENOENT)
1492                                 ret = 0;
1493
1494                         if (ret >= 0 && ret < len && pos + ret < size) {
1495                                 struct iov_iter i;
1496                                 int zlen = min_t(size_t, len - ret,
1497                                                  size - pos - ret);
1498
1499                                 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1500                                 iov_iter_advance(&i, ret);
1501                                 iov_iter_zero(zlen, &i);
1502                                 ret += zlen;
1503                         }
1504                         if (ret >= 0)
1505                                 len = ret;
1506                 }
1507
1508                 put_bvecs(bvecs, num_pages, should_dirty);
1509                 ceph_osdc_put_request(req);
1510                 if (ret < 0)
1511                         break;
1512
1513                 pos += len;
1514                 if (!write && pos >= size)
1515                         break;
1516
1517                 if (write && pos > size) {
1518                         if (ceph_inode_set_size(inode, pos))
1519                                 ceph_check_caps(ceph_inode(inode),
1520                                                 CHECK_CAPS_AUTHONLY);
1521                 }
1522         }
1523
1524         if (aio_req) {
1525                 LIST_HEAD(osd_reqs);
1526
1527                 if (aio_req->num_reqs == 0) {
1528                         kfree(aio_req);
1529                         return ret;
1530                 }
1531
1532                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1533                                               CEPH_CAP_FILE_RD);
1534
1535                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1536                 inode_dio_begin(inode);
1537                 while (!list_empty(&osd_reqs)) {
1538                         req = list_first_entry(&osd_reqs,
1539                                                struct ceph_osd_request,
1540                                                r_private_item);
1541                         list_del_init(&req->r_private_item);
1542                         if (ret >= 0)
1543                                 ceph_osdc_start_request(req->r_osdc, req);
1544                         if (ret < 0) {
1545                                 req->r_result = ret;
1546                                 ceph_aio_complete_req(req);
1547                         }
1548                 }
1549                 return -EIOCBQUEUED;
1550         }
1551
1552         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1553                 ret = pos - iocb->ki_pos;
1554                 iocb->ki_pos = pos;
1555         }
1556         return ret;
1557 }
1558
1559 /*
1560  * Synchronous write, straight from __user pointer or user pages.
1561  *
1562  * If write spans object boundary, just do multiple writes.  (For a
1563  * correct atomic write, we should e.g. take write locks on all
1564  * objects, rollback on failure, etc.)
1565  */
1566 static ssize_t
1567 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1568                 struct ceph_snap_context *snapc)
1569 {
1570         struct file *file = iocb->ki_filp;
1571         struct inode *inode = file_inode(file);
1572         struct ceph_inode_info *ci = ceph_inode(inode);
1573         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1574         struct ceph_vino vino;
1575         struct ceph_osd_request *req;
1576         struct page **pages;
1577         u64 len;
1578         int num_pages;
1579         int written = 0;
1580         int flags;
1581         int ret;
1582         bool check_caps = false;
1583         struct timespec64 mtime = current_time(inode);
1584         size_t count = iov_iter_count(from);
1585
1586         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1587                 return -EROFS;
1588
1589         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1590              file, pos, (unsigned)count, snapc, snapc->seq);
1591
1592         ret = filemap_write_and_wait_range(inode->i_mapping,
1593                                            pos, pos + count - 1);
1594         if (ret < 0)
1595                 return ret;
1596
1597         ceph_fscache_invalidate(inode, false);
1598         ret = invalidate_inode_pages2_range(inode->i_mapping,
1599                                             pos >> PAGE_SHIFT,
1600                                             (pos + count - 1) >> PAGE_SHIFT);
1601         if (ret < 0)
1602                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1603
1604         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1605
1606         while ((len = iov_iter_count(from)) > 0) {
1607                 size_t left;
1608                 int n;
1609
1610                 vino = ceph_vino(inode);
1611                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1612                                             vino, pos, &len, 0, 1,
1613                                             CEPH_OSD_OP_WRITE, flags, snapc,
1614                                             ci->i_truncate_seq,
1615                                             ci->i_truncate_size,
1616                                             false);
1617                 if (IS_ERR(req)) {
1618                         ret = PTR_ERR(req);
1619                         break;
1620                 }
1621
1622                 /*
1623                  * write from beginning of first page,
1624                  * regardless of io alignment
1625                  */
1626                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1627
1628                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1629                 if (IS_ERR(pages)) {
1630                         ret = PTR_ERR(pages);
1631                         goto out;
1632                 }
1633
1634                 left = len;
1635                 for (n = 0; n < num_pages; n++) {
1636                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1637                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1638                         if (ret != plen) {
1639                                 ret = -EFAULT;
1640                                 break;
1641                         }
1642                         left -= ret;
1643                 }
1644
1645                 if (ret < 0) {
1646                         ceph_release_page_vector(pages, num_pages);
1647                         goto out;
1648                 }
1649
1650                 req->r_inode = inode;
1651
1652                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1653                                                 false, true);
1654
1655                 req->r_mtime = mtime;
1656                 ceph_osdc_start_request(&fsc->client->osdc, req);
1657                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1658
1659                 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1660                                           req->r_end_latency, len, ret);
1661 out:
1662                 ceph_osdc_put_request(req);
1663                 if (ret != 0) {
1664                         ceph_set_error_write(ci);
1665                         break;
1666                 }
1667
1668                 ceph_clear_error_write(ci);
1669                 pos += len;
1670                 written += len;
1671                 if (pos > i_size_read(inode)) {
1672                         check_caps = ceph_inode_set_size(inode, pos);
1673                         if (check_caps)
1674                                 ceph_check_caps(ceph_inode(inode),
1675                                                 CHECK_CAPS_AUTHONLY);
1676                 }
1677
1678         }
1679
1680         if (ret != -EOLDSNAPC && written > 0) {
1681                 ret = written;
1682                 iocb->ki_pos = pos;
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1689  * Atomically grab references, so that those bits are not released
1690  * back to the MDS mid-read.
1691  *
1692  * Hmm, the sync read case isn't actually async... should it be?
1693  */
1694 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1695 {
1696         struct file *filp = iocb->ki_filp;
1697         struct ceph_file_info *fi = filp->private_data;
1698         size_t len = iov_iter_count(to);
1699         struct inode *inode = file_inode(filp);
1700         struct ceph_inode_info *ci = ceph_inode(inode);
1701         bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1702         ssize_t ret;
1703         int want = 0, got = 0;
1704         int retry_op = 0, read = 0;
1705
1706 again:
1707         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1708              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1709
1710         if (ceph_inode_is_shutdown(inode))
1711                 return -ESTALE;
1712
1713         if (direct_lock)
1714                 ceph_start_io_direct(inode);
1715         else
1716                 ceph_start_io_read(inode);
1717
1718         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1719                 want |= CEPH_CAP_FILE_CACHE;
1720         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1721                 want |= CEPH_CAP_FILE_LAZYIO;
1722
1723         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1724         if (ret < 0) {
1725                 if (direct_lock)
1726                         ceph_end_io_direct(inode);
1727                 else
1728                         ceph_end_io_read(inode);
1729                 return ret;
1730         }
1731
1732         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1733             (iocb->ki_flags & IOCB_DIRECT) ||
1734             (fi->flags & CEPH_F_SYNC)) {
1735
1736                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1737                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1738                      ceph_cap_string(got));
1739
1740                 if (!ceph_has_inline_data(ci)) {
1741                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1742                                 ret = ceph_direct_read_write(iocb, to,
1743                                                              NULL, NULL);
1744                                 if (ret >= 0 && ret < len)
1745                                         retry_op = CHECK_EOF;
1746                         } else {
1747                                 ret = ceph_sync_read(iocb, to, &retry_op);
1748                         }
1749                 } else {
1750                         retry_op = READ_INLINE;
1751                 }
1752         } else {
1753                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1754                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1755                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1756                      ceph_cap_string(got));
1757                 ceph_add_rw_context(fi, &rw_ctx);
1758                 ret = generic_file_read_iter(iocb, to);
1759                 ceph_del_rw_context(fi, &rw_ctx);
1760         }
1761
1762         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1763              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1764         ceph_put_cap_refs(ci, got);
1765
1766         if (direct_lock)
1767                 ceph_end_io_direct(inode);
1768         else
1769                 ceph_end_io_read(inode);
1770
1771         if (retry_op > HAVE_RETRIED && ret >= 0) {
1772                 int statret;
1773                 struct page *page = NULL;
1774                 loff_t i_size;
1775                 if (retry_op == READ_INLINE) {
1776                         page = __page_cache_alloc(GFP_KERNEL);
1777                         if (!page)
1778                                 return -ENOMEM;
1779                 }
1780
1781                 statret = __ceph_do_getattr(inode, page,
1782                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1783                 if (statret < 0) {
1784                         if (page)
1785                                 __free_page(page);
1786                         if (statret == -ENODATA) {
1787                                 BUG_ON(retry_op != READ_INLINE);
1788                                 goto again;
1789                         }
1790                         return statret;
1791                 }
1792
1793                 i_size = i_size_read(inode);
1794                 if (retry_op == READ_INLINE) {
1795                         BUG_ON(ret > 0 || read > 0);
1796                         if (iocb->ki_pos < i_size &&
1797                             iocb->ki_pos < PAGE_SIZE) {
1798                                 loff_t end = min_t(loff_t, i_size,
1799                                                    iocb->ki_pos + len);
1800                                 end = min_t(loff_t, end, PAGE_SIZE);
1801                                 if (statret < end)
1802                                         zero_user_segment(page, statret, end);
1803                                 ret = copy_page_to_iter(page,
1804                                                 iocb->ki_pos & ~PAGE_MASK,
1805                                                 end - iocb->ki_pos, to);
1806                                 iocb->ki_pos += ret;
1807                                 read += ret;
1808                         }
1809                         if (iocb->ki_pos < i_size && read < len) {
1810                                 size_t zlen = min_t(size_t, len - read,
1811                                                     i_size - iocb->ki_pos);
1812                                 ret = iov_iter_zero(zlen, to);
1813                                 iocb->ki_pos += ret;
1814                                 read += ret;
1815                         }
1816                         __free_pages(page, 0);
1817                         return read;
1818                 }
1819
1820                 /* hit EOF or hole? */
1821                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1822                     ret < len) {
1823                         dout("sync_read hit hole, ppos %lld < size %lld"
1824                              ", reading more\n", iocb->ki_pos, i_size);
1825
1826                         read += ret;
1827                         len -= ret;
1828                         retry_op = HAVE_RETRIED;
1829                         goto again;
1830                 }
1831         }
1832
1833         if (ret >= 0)
1834                 ret += read;
1835
1836         return ret;
1837 }
1838
1839 /*
1840  * Wrap filemap_splice_read with checks for cap bits on the inode.
1841  * Atomically grab references, so that those bits are not released
1842  * back to the MDS mid-read.
1843  */
1844 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
1845                                 struct pipe_inode_info *pipe,
1846                                 size_t len, unsigned int flags)
1847 {
1848         struct ceph_file_info *fi = in->private_data;
1849         struct inode *inode = file_inode(in);
1850         struct ceph_inode_info *ci = ceph_inode(inode);
1851         ssize_t ret;
1852         int want = 0, got = 0;
1853         CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
1854
1855         dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
1856              inode, ceph_vinop(inode), *ppos, len, inode);
1857
1858         if (ceph_inode_is_shutdown(inode))
1859                 return -ESTALE;
1860
1861         if (ceph_has_inline_data(ci) ||
1862             (fi->flags & CEPH_F_SYNC))
1863                 return copy_splice_read(in, ppos, pipe, len, flags);
1864
1865         ceph_start_io_read(inode);
1866
1867         want = CEPH_CAP_FILE_CACHE;
1868         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1869                 want |= CEPH_CAP_FILE_LAZYIO;
1870
1871         ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
1872         if (ret < 0)
1873                 goto out_end;
1874
1875         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
1876                 dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
1877                      inode, ceph_vinop(inode), *ppos, len,
1878                      ceph_cap_string(got));
1879
1880                 ceph_put_cap_refs(ci, got);
1881                 ceph_end_io_read(inode);
1882                 return copy_splice_read(in, ppos, pipe, len, flags);
1883         }
1884
1885         dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
1886              inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
1887
1888         rw_ctx.caps = got;
1889         ceph_add_rw_context(fi, &rw_ctx);
1890         ret = filemap_splice_read(in, ppos, pipe, len, flags);
1891         ceph_del_rw_context(fi, &rw_ctx);
1892
1893         dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
1894              inode, ceph_vinop(inode), ceph_cap_string(got), ret);
1895
1896         ceph_put_cap_refs(ci, got);
1897 out_end:
1898         ceph_end_io_read(inode);
1899         return ret;
1900 }
1901
1902 /*
1903  * Take cap references to avoid releasing caps to MDS mid-write.
1904  *
1905  * If we are synchronous, and write with an old snap context, the OSD
1906  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1907  * dropping our cap refs and allowing the pending snap to logically
1908  * complete _before_ this write occurs.
1909  *
1910  * If we are near ENOSPC, write synchronously.
1911  */
1912 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1913 {
1914         struct file *file = iocb->ki_filp;
1915         struct ceph_file_info *fi = file->private_data;
1916         struct inode *inode = file_inode(file);
1917         struct ceph_inode_info *ci = ceph_inode(inode);
1918         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1919         struct ceph_osd_client *osdc = &fsc->client->osdc;
1920         struct ceph_cap_flush *prealloc_cf;
1921         ssize_t count, written = 0;
1922         int err, want = 0, got;
1923         bool direct_lock = false;
1924         u32 map_flags;
1925         u64 pool_flags;
1926         loff_t pos;
1927         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1928
1929         if (ceph_inode_is_shutdown(inode))
1930                 return -ESTALE;
1931
1932         if (ceph_snap(inode) != CEPH_NOSNAP)
1933                 return -EROFS;
1934
1935         prealloc_cf = ceph_alloc_cap_flush();
1936         if (!prealloc_cf)
1937                 return -ENOMEM;
1938
1939         if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1940                 direct_lock = true;
1941
1942 retry_snap:
1943         if (direct_lock)
1944                 ceph_start_io_direct(inode);
1945         else
1946                 ceph_start_io_write(inode);
1947
1948         if (iocb->ki_flags & IOCB_APPEND) {
1949                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1950                 if (err < 0)
1951                         goto out;
1952         }
1953
1954         err = generic_write_checks(iocb, from);
1955         if (err <= 0)
1956                 goto out;
1957
1958         pos = iocb->ki_pos;
1959         if (unlikely(pos >= limit)) {
1960                 err = -EFBIG;
1961                 goto out;
1962         } else {
1963                 iov_iter_truncate(from, limit - pos);
1964         }
1965
1966         count = iov_iter_count(from);
1967         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1968                 err = -EDQUOT;
1969                 goto out;
1970         }
1971
1972         down_read(&osdc->lock);
1973         map_flags = osdc->osdmap->flags;
1974         pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1975         up_read(&osdc->lock);
1976         if ((map_flags & CEPH_OSDMAP_FULL) ||
1977             (pool_flags & CEPH_POOL_FLAG_FULL)) {
1978                 err = -ENOSPC;
1979                 goto out;
1980         }
1981
1982         err = file_remove_privs(file);
1983         if (err)
1984                 goto out;
1985
1986         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1987              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1988         if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1989                 want |= CEPH_CAP_FILE_BUFFER;
1990         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1991                 want |= CEPH_CAP_FILE_LAZYIO;
1992         got = 0;
1993         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1994         if (err < 0)
1995                 goto out;
1996
1997         err = file_update_time(file);
1998         if (err)
1999                 goto out_caps;
2000
2001         inode_inc_iversion_raw(inode);
2002
2003         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
2004              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
2005
2006         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2007             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
2008             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
2009                 struct ceph_snap_context *snapc;
2010                 struct iov_iter data;
2011
2012                 spin_lock(&ci->i_ceph_lock);
2013                 if (__ceph_have_pending_cap_snap(ci)) {
2014                         struct ceph_cap_snap *capsnap =
2015                                         list_last_entry(&ci->i_cap_snaps,
2016                                                         struct ceph_cap_snap,
2017                                                         ci_item);
2018                         snapc = ceph_get_snap_context(capsnap->context);
2019                 } else {
2020                         BUG_ON(!ci->i_head_snapc);
2021                         snapc = ceph_get_snap_context(ci->i_head_snapc);
2022                 }
2023                 spin_unlock(&ci->i_ceph_lock);
2024
2025                 /* we might need to revert back to that point */
2026                 data = *from;
2027                 if (iocb->ki_flags & IOCB_DIRECT)
2028                         written = ceph_direct_read_write(iocb, &data, snapc,
2029                                                          &prealloc_cf);
2030                 else
2031                         written = ceph_sync_write(iocb, &data, pos, snapc);
2032                 if (direct_lock)
2033                         ceph_end_io_direct(inode);
2034                 else
2035                         ceph_end_io_write(inode);
2036                 if (written > 0)
2037                         iov_iter_advance(from, written);
2038                 ceph_put_snap_context(snapc);
2039         } else {
2040                 /*
2041                  * No need to acquire the i_truncate_mutex. Because
2042                  * the MDS revokes Fwb caps before sending truncate
2043                  * message to us. We can't get Fwb cap while there
2044                  * are pending vmtruncate. So write and vmtruncate
2045                  * can not run at the same time
2046                  */
2047                 written = generic_perform_write(iocb, from);
2048                 ceph_end_io_write(inode);
2049         }
2050
2051         if (written >= 0) {
2052                 int dirty;
2053
2054                 spin_lock(&ci->i_ceph_lock);
2055                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2056                                                &prealloc_cf);
2057                 spin_unlock(&ci->i_ceph_lock);
2058                 if (dirty)
2059                         __mark_inode_dirty(inode, dirty);
2060                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2061                         ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2062         }
2063
2064         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
2065              inode, ceph_vinop(inode), pos, (unsigned)count,
2066              ceph_cap_string(got));
2067         ceph_put_cap_refs(ci, got);
2068
2069         if (written == -EOLDSNAPC) {
2070                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2071                      inode, ceph_vinop(inode), pos, (unsigned)count);
2072                 goto retry_snap;
2073         }
2074
2075         if (written >= 0) {
2076                 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2077                     (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2078                         iocb->ki_flags |= IOCB_DSYNC;
2079                 written = generic_write_sync(iocb, written);
2080         }
2081
2082         goto out_unlocked;
2083 out_caps:
2084         ceph_put_cap_refs(ci, got);
2085 out:
2086         if (direct_lock)
2087                 ceph_end_io_direct(inode);
2088         else
2089                 ceph_end_io_write(inode);
2090 out_unlocked:
2091         ceph_free_cap_flush(prealloc_cf);
2092         return written ? written : err;
2093 }
2094
2095 /*
2096  * llseek.  be sure to verify file size on SEEK_END.
2097  */
2098 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2099 {
2100         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2101                 struct inode *inode = file_inode(file);
2102                 int ret;
2103
2104                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2105                 if (ret < 0)
2106                         return ret;
2107         }
2108         return generic_file_llseek(file, offset, whence);
2109 }
2110
2111 static inline void ceph_zero_partial_page(
2112         struct inode *inode, loff_t offset, unsigned size)
2113 {
2114         struct page *page;
2115         pgoff_t index = offset >> PAGE_SHIFT;
2116
2117         page = find_lock_page(inode->i_mapping, index);
2118         if (page) {
2119                 wait_on_page_writeback(page);
2120                 zero_user(page, offset & (PAGE_SIZE - 1), size);
2121                 unlock_page(page);
2122                 put_page(page);
2123         }
2124 }
2125
2126 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2127                                       loff_t length)
2128 {
2129         loff_t nearly = round_up(offset, PAGE_SIZE);
2130         if (offset < nearly) {
2131                 loff_t size = nearly - offset;
2132                 if (length < size)
2133                         size = length;
2134                 ceph_zero_partial_page(inode, offset, size);
2135                 offset += size;
2136                 length -= size;
2137         }
2138         if (length >= PAGE_SIZE) {
2139                 loff_t size = round_down(length, PAGE_SIZE);
2140                 truncate_pagecache_range(inode, offset, offset + size - 1);
2141                 offset += size;
2142                 length -= size;
2143         }
2144         if (length)
2145                 ceph_zero_partial_page(inode, offset, length);
2146 }
2147
2148 static int ceph_zero_partial_object(struct inode *inode,
2149                                     loff_t offset, loff_t *length)
2150 {
2151         struct ceph_inode_info *ci = ceph_inode(inode);
2152         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2153         struct ceph_osd_request *req;
2154         int ret = 0;
2155         loff_t zero = 0;
2156         int op;
2157
2158         if (ceph_inode_is_shutdown(inode))
2159                 return -EIO;
2160
2161         if (!length) {
2162                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2163                 length = &zero;
2164         } else {
2165                 op = CEPH_OSD_OP_ZERO;
2166         }
2167
2168         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2169                                         ceph_vino(inode),
2170                                         offset, length,
2171                                         0, 1, op,
2172                                         CEPH_OSD_FLAG_WRITE,
2173                                         NULL, 0, 0, false);
2174         if (IS_ERR(req)) {
2175                 ret = PTR_ERR(req);
2176                 goto out;
2177         }
2178
2179         req->r_mtime = inode->i_mtime;
2180         ceph_osdc_start_request(&fsc->client->osdc, req);
2181         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2182         if (ret == -ENOENT)
2183                 ret = 0;
2184         ceph_osdc_put_request(req);
2185
2186 out:
2187         return ret;
2188 }
2189
2190 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2191 {
2192         int ret = 0;
2193         struct ceph_inode_info *ci = ceph_inode(inode);
2194         s32 stripe_unit = ci->i_layout.stripe_unit;
2195         s32 stripe_count = ci->i_layout.stripe_count;
2196         s32 object_size = ci->i_layout.object_size;
2197         u64 object_set_size = object_size * stripe_count;
2198         u64 nearly, t;
2199
2200         /* round offset up to next period boundary */
2201         nearly = offset + object_set_size - 1;
2202         t = nearly;
2203         nearly -= do_div(t, object_set_size);
2204
2205         while (length && offset < nearly) {
2206                 loff_t size = length;
2207                 ret = ceph_zero_partial_object(inode, offset, &size);
2208                 if (ret < 0)
2209                         return ret;
2210                 offset += size;
2211                 length -= size;
2212         }
2213         while (length >= object_set_size) {
2214                 int i;
2215                 loff_t pos = offset;
2216                 for (i = 0; i < stripe_count; ++i) {
2217                         ret = ceph_zero_partial_object(inode, pos, NULL);
2218                         if (ret < 0)
2219                                 return ret;
2220                         pos += stripe_unit;
2221                 }
2222                 offset += object_set_size;
2223                 length -= object_set_size;
2224         }
2225         while (length) {
2226                 loff_t size = length;
2227                 ret = ceph_zero_partial_object(inode, offset, &size);
2228                 if (ret < 0)
2229                         return ret;
2230                 offset += size;
2231                 length -= size;
2232         }
2233         return ret;
2234 }
2235
2236 static long ceph_fallocate(struct file *file, int mode,
2237                                 loff_t offset, loff_t length)
2238 {
2239         struct ceph_file_info *fi = file->private_data;
2240         struct inode *inode = file_inode(file);
2241         struct ceph_inode_info *ci = ceph_inode(inode);
2242         struct ceph_cap_flush *prealloc_cf;
2243         int want, got = 0;
2244         int dirty;
2245         int ret = 0;
2246         loff_t endoff = 0;
2247         loff_t size;
2248
2249         dout("%s %p %llx.%llx mode %x, offset %llu length %llu\n", __func__,
2250              inode, ceph_vinop(inode), mode, offset, length);
2251
2252         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2253                 return -EOPNOTSUPP;
2254
2255         if (!S_ISREG(inode->i_mode))
2256                 return -EOPNOTSUPP;
2257
2258         if (IS_ENCRYPTED(inode))
2259                 return -EOPNOTSUPP;
2260
2261         prealloc_cf = ceph_alloc_cap_flush();
2262         if (!prealloc_cf)
2263                 return -ENOMEM;
2264
2265         inode_lock(inode);
2266
2267         if (ceph_snap(inode) != CEPH_NOSNAP) {
2268                 ret = -EROFS;
2269                 goto unlock;
2270         }
2271
2272         size = i_size_read(inode);
2273
2274         /* Are we punching a hole beyond EOF? */
2275         if (offset >= size)
2276                 goto unlock;
2277         if ((offset + length) > size)
2278                 length = size - offset;
2279
2280         if (fi->fmode & CEPH_FILE_MODE_LAZY)
2281                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2282         else
2283                 want = CEPH_CAP_FILE_BUFFER;
2284
2285         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2286         if (ret < 0)
2287                 goto unlock;
2288
2289         ret = file_modified(file);
2290         if (ret)
2291                 goto put_caps;
2292
2293         filemap_invalidate_lock(inode->i_mapping);
2294         ceph_fscache_invalidate(inode, false);
2295         ceph_zero_pagecache_range(inode, offset, length);
2296         ret = ceph_zero_objects(inode, offset, length);
2297
2298         if (!ret) {
2299                 spin_lock(&ci->i_ceph_lock);
2300                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2301                                                &prealloc_cf);
2302                 spin_unlock(&ci->i_ceph_lock);
2303                 if (dirty)
2304                         __mark_inode_dirty(inode, dirty);
2305         }
2306         filemap_invalidate_unlock(inode->i_mapping);
2307
2308 put_caps:
2309         ceph_put_cap_refs(ci, got);
2310 unlock:
2311         inode_unlock(inode);
2312         ceph_free_cap_flush(prealloc_cf);
2313         return ret;
2314 }
2315
2316 /*
2317  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2318  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2319  * this fails; zero is returned on success.
2320  */
2321 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2322                           struct file *dst_filp,
2323                           loff_t dst_endoff, int *dst_got)
2324 {
2325         int ret = 0;
2326         bool retrying = false;
2327
2328 retry_caps:
2329         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2330                             dst_endoff, dst_got);
2331         if (ret < 0)
2332                 return ret;
2333
2334         /*
2335          * Since we're already holding the FILE_WR capability for the dst file,
2336          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2337          * retry dance instead to try to get both capabilities.
2338          */
2339         ret = ceph_try_get_caps(file_inode(src_filp),
2340                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2341                                 false, src_got);
2342         if (ret <= 0) {
2343                 /* Start by dropping dst_ci caps and getting src_ci caps */
2344                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2345                 if (retrying) {
2346                         if (!ret)
2347                                 /* ceph_try_get_caps masks EAGAIN */
2348                                 ret = -EAGAIN;
2349                         return ret;
2350                 }
2351                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2352                                     CEPH_CAP_FILE_SHARED, -1, src_got);
2353                 if (ret < 0)
2354                         return ret;
2355                 /*... drop src_ci caps too, and retry */
2356                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2357                 retrying = true;
2358                 goto retry_caps;
2359         }
2360         return ret;
2361 }
2362
2363 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2364                            struct ceph_inode_info *dst_ci, int dst_got)
2365 {
2366         ceph_put_cap_refs(src_ci, src_got);
2367         ceph_put_cap_refs(dst_ci, dst_got);
2368 }
2369
2370 /*
2371  * This function does several size-related checks, returning an error if:
2372  *  - source file is smaller than off+len
2373  *  - destination file size is not OK (inode_newsize_ok())
2374  *  - max bytes quotas is exceeded
2375  */
2376 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2377                            loff_t src_off, loff_t dst_off, size_t len)
2378 {
2379         loff_t size, endoff;
2380
2381         size = i_size_read(src_inode);
2382         /*
2383          * Don't copy beyond source file EOF.  Instead of simply setting length
2384          * to (size - src_off), just drop to VFS default implementation, as the
2385          * local i_size may be stale due to other clients writing to the source
2386          * inode.
2387          */
2388         if (src_off + len > size) {
2389                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2390                      src_off, len, size);
2391                 return -EOPNOTSUPP;
2392         }
2393         size = i_size_read(dst_inode);
2394
2395         endoff = dst_off + len;
2396         if (inode_newsize_ok(dst_inode, endoff))
2397                 return -EOPNOTSUPP;
2398
2399         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2400                 return -EDQUOT;
2401
2402         return 0;
2403 }
2404
2405 static struct ceph_osd_request *
2406 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2407                             u64 src_snapid,
2408                             struct ceph_object_id *src_oid,
2409                             struct ceph_object_locator *src_oloc,
2410                             struct ceph_object_id *dst_oid,
2411                             struct ceph_object_locator *dst_oloc,
2412                             u32 truncate_seq, u64 truncate_size)
2413 {
2414         struct ceph_osd_request *req;
2415         int ret;
2416         u32 src_fadvise_flags =
2417                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2418                 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2419         u32 dst_fadvise_flags =
2420                 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2421                 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2422
2423         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2424         if (!req)
2425                 return ERR_PTR(-ENOMEM);
2426
2427         req->r_flags = CEPH_OSD_FLAG_WRITE;
2428
2429         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2430         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2431
2432         ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2433                                         src_oid, src_oloc,
2434                                         src_fadvise_flags,
2435                                         dst_fadvise_flags,
2436                                         truncate_seq,
2437                                         truncate_size,
2438                                         CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2439         if (ret)
2440                 goto out;
2441
2442         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2443         if (ret)
2444                 goto out;
2445
2446         return req;
2447
2448 out:
2449         ceph_osdc_put_request(req);
2450         return ERR_PTR(ret);
2451 }
2452
2453 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2454                                     struct ceph_inode_info *dst_ci, u64 *dst_off,
2455                                     struct ceph_fs_client *fsc,
2456                                     size_t len, unsigned int flags)
2457 {
2458         struct ceph_object_locator src_oloc, dst_oloc;
2459         struct ceph_object_id src_oid, dst_oid;
2460         struct ceph_osd_client *osdc;
2461         struct ceph_osd_request *req;
2462         size_t bytes = 0;
2463         u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2464         u32 src_objlen, dst_objlen;
2465         u32 object_size = src_ci->i_layout.object_size;
2466         int ret;
2467
2468         src_oloc.pool = src_ci->i_layout.pool_id;
2469         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2470         dst_oloc.pool = dst_ci->i_layout.pool_id;
2471         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2472         osdc = &fsc->client->osdc;
2473
2474         while (len >= object_size) {
2475                 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2476                                               object_size, &src_objnum,
2477                                               &src_objoff, &src_objlen);
2478                 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2479                                               object_size, &dst_objnum,
2480                                               &dst_objoff, &dst_objlen);
2481                 ceph_oid_init(&src_oid);
2482                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2483                                 src_ci->i_vino.ino, src_objnum);
2484                 ceph_oid_init(&dst_oid);
2485                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2486                                 dst_ci->i_vino.ino, dst_objnum);
2487                 /* Do an object remote copy */
2488                 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2489                                                   &src_oid, &src_oloc,
2490                                                   &dst_oid, &dst_oloc,
2491                                                   dst_ci->i_truncate_seq,
2492                                                   dst_ci->i_truncate_size);
2493                 if (IS_ERR(req))
2494                         ret = PTR_ERR(req);
2495                 else {
2496                         ceph_osdc_start_request(osdc, req);
2497                         ret = ceph_osdc_wait_request(osdc, req);
2498                         ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2499                                                      req->r_start_latency,
2500                                                      req->r_end_latency,
2501                                                      object_size, ret);
2502                         ceph_osdc_put_request(req);
2503                 }
2504                 if (ret) {
2505                         if (ret == -EOPNOTSUPP) {
2506                                 fsc->have_copy_from2 = false;
2507                                 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2508                         }
2509                         dout("ceph_osdc_copy_from returned %d\n", ret);
2510                         if (!bytes)
2511                                 bytes = ret;
2512                         goto out;
2513                 }
2514                 len -= object_size;
2515                 bytes += object_size;
2516                 *src_off += object_size;
2517                 *dst_off += object_size;
2518         }
2519
2520 out:
2521         ceph_oloc_destroy(&src_oloc);
2522         ceph_oloc_destroy(&dst_oloc);
2523         return bytes;
2524 }
2525
2526 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2527                                       struct file *dst_file, loff_t dst_off,
2528                                       size_t len, unsigned int flags)
2529 {
2530         struct inode *src_inode = file_inode(src_file);
2531         struct inode *dst_inode = file_inode(dst_file);
2532         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2533         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2534         struct ceph_cap_flush *prealloc_cf;
2535         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2536         loff_t size;
2537         ssize_t ret = -EIO, bytes;
2538         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2539         u32 src_objlen, dst_objlen;
2540         int src_got = 0, dst_got = 0, err, dirty;
2541
2542         if (src_inode->i_sb != dst_inode->i_sb) {
2543                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2544
2545                 if (ceph_fsid_compare(&src_fsc->client->fsid,
2546                                       &dst_fsc->client->fsid)) {
2547                         dout("Copying files across clusters: src: %pU dst: %pU\n",
2548                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
2549                         return -EXDEV;
2550                 }
2551         }
2552         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2553                 return -EROFS;
2554
2555         /*
2556          * Some of the checks below will return -EOPNOTSUPP, which will force a
2557          * fallback to the default VFS copy_file_range implementation.  This is
2558          * desirable in several cases (for ex, the 'len' is smaller than the
2559          * size of the objects, or in cases where that would be more
2560          * efficient).
2561          */
2562
2563         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2564                 return -EOPNOTSUPP;
2565
2566         if (!src_fsc->have_copy_from2)
2567                 return -EOPNOTSUPP;
2568
2569         /*
2570          * Striped file layouts require that we copy partial objects, but the
2571          * OSD copy-from operation only supports full-object copies.  Limit
2572          * this to non-striped file layouts for now.
2573          */
2574         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2575             (src_ci->i_layout.stripe_count != 1) ||
2576             (dst_ci->i_layout.stripe_count != 1) ||
2577             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2578                 dout("Invalid src/dst files layout\n");
2579                 return -EOPNOTSUPP;
2580         }
2581
2582         /* Every encrypted inode gets its own key, so we can't offload them */
2583         if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
2584                 return -EOPNOTSUPP;
2585
2586         if (len < src_ci->i_layout.object_size)
2587                 return -EOPNOTSUPP; /* no remote copy will be done */
2588
2589         prealloc_cf = ceph_alloc_cap_flush();
2590         if (!prealloc_cf)
2591                 return -ENOMEM;
2592
2593         /* Start by sync'ing the source and destination files */
2594         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2595         if (ret < 0) {
2596                 dout("failed to write src file (%zd)\n", ret);
2597                 goto out;
2598         }
2599         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2600         if (ret < 0) {
2601                 dout("failed to write dst file (%zd)\n", ret);
2602                 goto out;
2603         }
2604
2605         /*
2606          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2607          * clients may have dirty data in their caches.  And OSDs know nothing
2608          * about caps, so they can't safely do the remote object copies.
2609          */
2610         err = get_rd_wr_caps(src_file, &src_got,
2611                              dst_file, (dst_off + len), &dst_got);
2612         if (err < 0) {
2613                 dout("get_rd_wr_caps returned %d\n", err);
2614                 ret = -EOPNOTSUPP;
2615                 goto out;
2616         }
2617
2618         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2619         if (ret < 0)
2620                 goto out_caps;
2621
2622         /* Drop dst file cached pages */
2623         ceph_fscache_invalidate(dst_inode, false);
2624         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2625                                             dst_off >> PAGE_SHIFT,
2626                                             (dst_off + len) >> PAGE_SHIFT);
2627         if (ret < 0) {
2628                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2629                 ret = 0; /* XXX */
2630         }
2631         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2632                                       src_ci->i_layout.object_size,
2633                                       &src_objnum, &src_objoff, &src_objlen);
2634         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2635                                       dst_ci->i_layout.object_size,
2636                                       &dst_objnum, &dst_objoff, &dst_objlen);
2637         /* object-level offsets need to the same */
2638         if (src_objoff != dst_objoff) {
2639                 ret = -EOPNOTSUPP;
2640                 goto out_caps;
2641         }
2642
2643         /*
2644          * Do a manual copy if the object offset isn't object aligned.
2645          * 'src_objlen' contains the bytes left until the end of the object,
2646          * starting at the src_off
2647          */
2648         if (src_objoff) {
2649                 dout("Initial partial copy of %u bytes\n", src_objlen);
2650
2651                 /*
2652                  * we need to temporarily drop all caps as we'll be calling
2653                  * {read,write}_iter, which will get caps again.
2654                  */
2655                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2656                 ret = do_splice_direct(src_file, &src_off, dst_file,
2657                                        &dst_off, src_objlen, flags);
2658                 /* Abort on short copies or on error */
2659                 if (ret < src_objlen) {
2660                         dout("Failed partial copy (%zd)\n", ret);
2661                         goto out;
2662                 }
2663                 len -= ret;
2664                 err = get_rd_wr_caps(src_file, &src_got,
2665                                      dst_file, (dst_off + len), &dst_got);
2666                 if (err < 0)
2667                         goto out;
2668                 err = is_file_size_ok(src_inode, dst_inode,
2669                                       src_off, dst_off, len);
2670                 if (err < 0)
2671                         goto out_caps;
2672         }
2673
2674         size = i_size_read(dst_inode);
2675         bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2676                                      src_fsc, len, flags);
2677         if (bytes <= 0) {
2678                 if (!ret)
2679                         ret = bytes;
2680                 goto out_caps;
2681         }
2682         dout("Copied %zu bytes out of %zu\n", bytes, len);
2683         len -= bytes;
2684         ret += bytes;
2685
2686         file_update_time(dst_file);
2687         inode_inc_iversion_raw(dst_inode);
2688
2689         if (dst_off > size) {
2690                 /* Let the MDS know about dst file size change */
2691                 if (ceph_inode_set_size(dst_inode, dst_off) ||
2692                     ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2693                         ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
2694         }
2695         /* Mark Fw dirty */
2696         spin_lock(&dst_ci->i_ceph_lock);
2697         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2698         spin_unlock(&dst_ci->i_ceph_lock);
2699         if (dirty)
2700                 __mark_inode_dirty(dst_inode, dirty);
2701
2702 out_caps:
2703         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2704
2705         /*
2706          * Do the final manual copy if we still have some bytes left, unless
2707          * there were errors in remote object copies (len >= object_size).
2708          */
2709         if (len && (len < src_ci->i_layout.object_size)) {
2710                 dout("Final partial copy of %zu bytes\n", len);
2711                 bytes = do_splice_direct(src_file, &src_off, dst_file,
2712                                          &dst_off, len, flags);
2713                 if (bytes > 0)
2714                         ret += bytes;
2715                 else
2716                         dout("Failed partial copy (%zd)\n", bytes);
2717         }
2718
2719 out:
2720         ceph_free_cap_flush(prealloc_cf);
2721
2722         return ret;
2723 }
2724
2725 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2726                                     struct file *dst_file, loff_t dst_off,
2727                                     size_t len, unsigned int flags)
2728 {
2729         ssize_t ret;
2730
2731         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2732                                      len, flags);
2733
2734         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2735                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2736                                               dst_off, len, flags);
2737         return ret;
2738 }
2739
2740 const struct file_operations ceph_file_fops = {
2741         .open = ceph_open,
2742         .release = ceph_release,
2743         .llseek = ceph_llseek,
2744         .read_iter = ceph_read_iter,
2745         .write_iter = ceph_write_iter,
2746         .mmap = ceph_mmap,
2747         .fsync = ceph_fsync,
2748         .lock = ceph_lock,
2749         .setlease = simple_nosetlease,
2750         .flock = ceph_flock,
2751         .splice_read = ceph_splice_read,
2752         .splice_write = iter_file_splice_write,
2753         .unlocked_ioctl = ceph_ioctl,
2754         .compat_ioctl = compat_ptr_ioctl,
2755         .fallocate      = ceph_fallocate,
2756         .copy_file_range = ceph_copy_file_range,
2757 };