fs/ceph/file.c

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