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