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