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->meta_err = errseq_sample(&ci->i_meta_err);
 238         fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
 239
 240         return 0;
 241 }
 242
 243 /*
 244  * initialize private struct file data.
 245  * if we fail, clean up by dropping fmode reference on the ceph_inode
 246  */
 247 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 248 {
 249         int ret = 0;
 250
 251         switch (inode->i_mode & S_IFMT) {
 252         case S_IFREG:
 253                 ceph_fscache_register_inode_cookie(inode);
 254                 ceph_fscache_file_set_cookie(inode, file);
 255                 /* fall through */
 256         case S_IFDIR:
 257                 ret = ceph_init_file_info(inode, file, fmode,
 258                                                 S_ISDIR(inode->i_mode));
 259                 if (ret)
 260                         return ret;
 261                 break;
 262
 263         case S_IFLNK:
 264                 dout("init_file %p %p 0%o (symlink)\n", inode, file,
 265                      inode->i_mode);
 266                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 267                 break;
 268
 269         default:
 270                 dout("init_file %p %p 0%o (special)\n", inode, file,
 271                      inode->i_mode);
 272                 /*
 273                  * we need to drop the open ref now, since we don't
 274                  * have .release set to ceph_release.
 275                  */
 276                 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
 277                 BUG_ON(inode->i_fop->release == ceph_release);
 278
 279                 /* call the proper open fop */
 280                 ret = inode->i_fop->open(inode, file);
 281         }
 282         return ret;
 283 }
 284
 285 /*
 286  * try renew caps after session gets killed.
 287  */
 288 int ceph_renew_caps(struct inode *inode)
 289 {
 290         struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
 291         struct ceph_inode_info *ci = ceph_inode(inode);
 292         struct ceph_mds_request *req;
 293         int err, flags, wanted;
 294
 295         spin_lock(&ci->i_ceph_lock);
 296         wanted = __ceph_caps_file_wanted(ci);
 297         if (__ceph_is_any_real_caps(ci) &&
 298             (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
 299                 int issued = __ceph_caps_issued(ci, NULL);
 300                 spin_unlock(&ci->i_ceph_lock);
 301                 dout("renew caps %p want %s issued %s updating mds_wanted\n",
 302                      inode, ceph_cap_string(wanted), ceph_cap_string(issued));
 303                 ceph_check_caps(ci, 0, NULL);
 304                 return 0;
 305         }
 306         spin_unlock(&ci->i_ceph_lock);
 307
 308         flags = 0;
 309         if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
 310                 flags = O_RDWR;
 311         else if (wanted & CEPH_CAP_FILE_RD)
 312                 flags = O_RDONLY;
 313         else if (wanted & CEPH_CAP_FILE_WR)
 314                 flags = O_WRONLY;
 315 #ifdef O_LAZY
 316         if (wanted & CEPH_CAP_FILE_LAZYIO)
 317                 flags |= O_LAZY;
 318 #endif
 319
 320         req = prepare_open_request(inode->i_sb, flags, 0);
 321         if (IS_ERR(req)) {
 322                 err = PTR_ERR(req);
 323                 goto out;
 324         }
 325
 326         req->r_inode = inode;
 327         ihold(inode);
 328         req->r_num_caps = 1;
 329         req->r_fmode = -1;
 330
 331         err = ceph_mdsc_do_request(mdsc, NULL, req);
 332         ceph_mdsc_put_request(req);
 333 out:
 334         dout("renew caps %p open result=%d\n", inode, err);
 335         return err < 0 ? err : 0;
 336 }
 337
 338 /*
 339  * If we already have the requisite capabilities, we can satisfy
 340  * the open request locally (no need to request new caps from the
 341  * MDS).  We do, however, need to inform the MDS (asynchronously)
 342  * if our wanted caps set expands.
 343  */
 344 int ceph_open(struct inode *inode, struct file *file)
 345 {
 346         struct ceph_inode_info *ci = ceph_inode(inode);
 347         struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
 348         struct ceph_mds_client *mdsc = fsc->mdsc;
 349         struct ceph_mds_request *req;
 350         struct ceph_file_info *fi = file->private_data;
 351         int err;
 352         int flags, fmode, wanted;
 353
 354         if (fi) {
 355                 dout("open file %p is already opened\n", file);
 356                 return 0;
 357         }
 358
 359         /* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
 360         flags = file->f_flags & ~(O_CREAT|O_EXCL);
 361         if (S_ISDIR(inode->i_mode))
 362                 flags = O_DIRECTORY;  /* mds likes to know */
 363
 364         dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
 365              ceph_vinop(inode), file, flags, file->f_flags);
 366         fmode = ceph_flags_to_mode(flags);
 367         wanted = ceph_caps_for_mode(fmode);
 368
 369         /* snapped files are read-only */
 370         if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
 371                 return -EROFS;
 372
 373         /* trivially open snapdir */
 374         if (ceph_snap(inode) == CEPH_SNAPDIR) {
 375                 spin_lock(&ci->i_ceph_lock);
 376                 __ceph_get_fmode(ci, fmode);
 377                 spin_unlock(&ci->i_ceph_lock);
 378                 return ceph_init_file(inode, file, fmode);
 379         }
 380
 381         /*
 382          * No need to block if we have caps on the auth MDS (for
 383          * write) or any MDS (for read).  Update wanted set
 384          * asynchronously.
 385          */
 386         spin_lock(&ci->i_ceph_lock);
 387         if (__ceph_is_any_real_caps(ci) &&
 388             (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
 389                 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
 390                 int issued = __ceph_caps_issued(ci, NULL);
 391
 392                 dout("open %p fmode %d want %s issued %s using existing\n",
 393                      inode, fmode, ceph_cap_string(wanted),
 394                      ceph_cap_string(issued));
 395                 __ceph_get_fmode(ci, fmode);
 396                 spin_unlock(&ci->i_ceph_lock);
 397
 398                 /* adjust wanted? */
 399                 if ((issued & wanted) != wanted &&
 400                     (mds_wanted & wanted) != wanted &&
 401                     ceph_snap(inode) != CEPH_SNAPDIR)
 402                         ceph_check_caps(ci, 0, NULL);
 403
 404                 return ceph_init_file(inode, file, fmode);
 405         } else if (ceph_snap(inode) != CEPH_NOSNAP &&
 406                    (ci->i_snap_caps & wanted) == wanted) {
 407                 __ceph_get_fmode(ci, fmode);
 408                 spin_unlock(&ci->i_ceph_lock);
 409                 return ceph_init_file(inode, file, fmode);
 410         }
 411
 412         spin_unlock(&ci->i_ceph_lock);
 413
 414         dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
 415         req = prepare_open_request(inode->i_sb, flags, 0);
 416         if (IS_ERR(req)) {
 417                 err = PTR_ERR(req);
 418                 goto out;
 419         }
 420         req->r_inode = inode;
 421         ihold(inode);
 422
 423         req->r_num_caps = 1;
 424         err = ceph_mdsc_do_request(mdsc, NULL, req);
 425         if (!err)
 426                 err = ceph_init_file(inode, file, req->r_fmode);
 427         ceph_mdsc_put_request(req);
 428         dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
 429 out:
 430         return err;
 431 }
 432
 433
 434 /*
 435  * Do a lookup + open with a single request.  If we get a non-existent
 436  * file or symlink, return 1 so the VFS can retry.
 437  */
 438 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
 439                      struct file *file, unsigned flags, umode_t mode)
 440 {
 441         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 442         struct ceph_mds_client *mdsc = fsc->mdsc;
 443         struct ceph_mds_request *req;
 444         struct dentry *dn;
 445         struct ceph_acl_sec_ctx as_ctx = {};
 446         int mask;
 447         int err;
 448
 449         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
 450              dir, dentry, dentry,
 451              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
 452
 453         if (dentry->d_name.len > NAME_MAX)
 454                 return -ENAMETOOLONG;
 455
 456         if (flags & O_CREAT) {
 457                 if (ceph_quota_is_max_files_exceeded(dir))
 458                         return -EDQUOT;
 459                 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 460                 if (err < 0)
 461                         return err;
 462                 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 463                 if (err < 0)
 464                         goto out_ctx;
 465         } else if (!d_in_lookup(dentry)) {
 466                 /* If it's not being looked up, it's negative */
 467                 return -ENOENT;
 468         }
 469
 470         /* do the open */
 471         req = prepare_open_request(dir->i_sb, flags, mode);
 472         if (IS_ERR(req)) {
 473                 err = PTR_ERR(req);
 474                 goto out_ctx;
 475         }
 476         req->r_dentry = dget(dentry);
 477         req->r_num_caps = 2;
 478         if (flags & O_CREAT) {
 479                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 480                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 481                 if (as_ctx.pagelist) {
 482                         req->r_pagelist = as_ctx.pagelist;
 483                         as_ctx.pagelist = NULL;
 484                 }
 485         }
 486
 487        mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 488        if (ceph_security_xattr_wanted(dir))
 489                mask |= CEPH_CAP_XATTR_SHARED;
 490        req->r_args.open.mask = cpu_to_le32(mask);
 491
 492         req->r_parent = dir;
 493         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 494         err = ceph_mdsc_do_request(mdsc,
 495                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
 496                                    req);
 497         err = ceph_handle_snapdir(req, dentry, err);
 498         if (err)
 499                 goto out_req;
 500
 501         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
 502                 err = ceph_handle_notrace_create(dir, dentry);
 503
 504         if (d_in_lookup(dentry)) {
 505                 dn = ceph_finish_lookup(req, dentry, err);
 506                 if (IS_ERR(dn))
 507                         err = PTR_ERR(dn);
 508         } else {
 509                 /* we were given a hashed negative dentry */
 510                 dn = NULL;
 511         }
 512         if (err)
 513                 goto out_req;
 514         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
 515                 /* make vfs retry on splice, ENOENT, or symlink */
 516                 dout("atomic_open finish_no_open on dn %p\n", dn);
 517                 err = finish_no_open(file, dn);
 518         } else {
 519                 dout("atomic_open finish_open on dn %p\n", dn);
 520                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
 521                         ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 522                         file->f_mode |= FMODE_CREATED;
 523                 }
 524                 err = finish_open(file, dentry, ceph_open);
 525         }
 526 out_req:
 527         if (!req->r_err && req->r_target_inode)
 528                 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
 529         ceph_mdsc_put_request(req);
 530 out_ctx:
 531         ceph_release_acl_sec_ctx(&as_ctx);
 532         dout("atomic_open result=%d\n", err);
 533         return err;
 534 }
 535
 536 int ceph_release(struct inode *inode, struct file *file)
 537 {
 538         struct ceph_inode_info *ci = ceph_inode(inode);
 539
 540         if (S_ISDIR(inode->i_mode)) {
 541                 struct ceph_dir_file_info *dfi = file->private_data;
 542                 dout("release inode %p dir file %p\n", inode, file);
 543                 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
 544
 545                 ceph_put_fmode(ci, dfi->file_info.fmode);
 546
 547                 if (dfi->last_readdir)
 548                         ceph_mdsc_put_request(dfi->last_readdir);
 549                 kfree(dfi->last_name);
 550                 kfree(dfi->dir_info);
 551                 kmem_cache_free(ceph_dir_file_cachep, dfi);
 552         } else {
 553                 struct ceph_file_info *fi = file->private_data;
 554                 dout("release inode %p regular file %p\n", inode, file);
 555                 WARN_ON(!list_empty(&fi->rw_contexts));
 556
 557                 ceph_put_fmode(ci, fi->fmode);
 558                 kmem_cache_free(ceph_file_cachep, fi);
 559         }
 560
 561         /* wake up anyone waiting for caps on this inode */
 562         wake_up_all(&ci->i_cap_wq);
 563         return 0;
 564 }
 565
 566 enum {
 567         HAVE_RETRIED = 1,
 568         CHECK_EOF =    2,
 569         READ_INLINE =  3,
 570 };
 571
 572 /*
 573  * Completely synchronous read and write methods.  Direct from __user
 574  * buffer to osd, or directly to user pages (if O_DIRECT).
 575  *
 576  * If the read spans object boundary, just do multiple reads.  (That's not
 577  * atomic, but good enough for now.)
 578  *
 579  * If we get a short result from the OSD, check against i_size; we need to
 580  * only return a short read to the caller if we hit EOF.
 581  */
 582 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
 583                               int *retry_op)
 584 {
 585         struct file *file = iocb->ki_filp;
 586         struct inode *inode = file_inode(file);
 587         struct ceph_inode_info *ci = ceph_inode(inode);
 588         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 589         struct ceph_osd_client *osdc = &fsc->client->osdc;
 590         ssize_t ret;
 591         u64 off = iocb->ki_pos;
 592         u64 len = iov_iter_count(to);
 593
 594         dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
 595              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
 596
 597         if (!len)
 598                 return 0;
 599         /*
 600          * flush any page cache pages in this range.  this
 601          * will make concurrent normal and sync io slow,
 602          * but it will at least behave sensibly when they are
 603          * in sequence.
 604          */
 605         ret = filemap_write_and_wait_range(inode->i_mapping,
 606                                            off, off + len - 1);
 607         if (ret < 0)
 608                 return ret;
 609
 610         ret = 0;
 611         while ((len = iov_iter_count(to)) > 0) {
 612                 struct ceph_osd_request *req;
 613                 struct page **pages;
 614                 int num_pages;
 615                 size_t page_off;
 616                 u64 i_size;
 617                 bool more;
 618
 619                 req = ceph_osdc_new_request(osdc, &ci->i_layout,
 620                                         ci->i_vino, off, &len, 0, 1,
 621                                         CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
 622                                         NULL, ci->i_truncate_seq,
 623                                         ci->i_truncate_size, false);
 624                 if (IS_ERR(req)) {
 625                         ret = PTR_ERR(req);
 626                         break;
 627                 }
 628
 629                 more = len < iov_iter_count(to);
 630
 631                 if (unlikely(iov_iter_is_pipe(to))) {
 632                         ret = iov_iter_get_pages_alloc(to, &pages, len,
 633                                                        &page_off);
 634                         if (ret <= 0) {
 635                                 ceph_osdc_put_request(req);
 636                                 ret = -ENOMEM;
 637                                 break;
 638                         }
 639                         num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
 640                         if (ret < len) {
 641                                 len = ret;
 642                                 osd_req_op_extent_update(req, 0, len);
 643                                 more = false;
 644                         }
 645                 } else {
 646                         num_pages = calc_pages_for(off, len);
 647                         page_off = off & ~PAGE_MASK;
 648                         pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
 649                         if (IS_ERR(pages)) {
 650                                 ceph_osdc_put_request(req);
 651                                 ret = PTR_ERR(pages);
 652                                 break;
 653                         }
 654                 }
 655
 656                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
 657                                                  false, false);
 658                 ret = ceph_osdc_start_request(osdc, req, false);
 659                 if (!ret)
 660                         ret = ceph_osdc_wait_request(osdc, req);
 661                 ceph_osdc_put_request(req);
 662
 663                 i_size = i_size_read(inode);
 664                 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
 665                      off, len, ret, i_size, (more ? " MORE" : ""));
 666
 667                 if (ret == -ENOENT)
 668                         ret = 0;
 669                 if (ret >= 0 && ret < len && (off + ret < i_size)) {
 670                         int zlen = min(len - ret, i_size - off - ret);
 671                         int zoff = page_off + ret;
 672                         dout("sync_read zero gap %llu~%llu\n",
 673                              off + ret, off + ret + zlen);
 674                         ceph_zero_page_vector_range(zoff, zlen, pages);
 675                         ret += zlen;
 676                 }
 677
 678                 if (unlikely(iov_iter_is_pipe(to))) {
 679                         if (ret > 0) {
 680                                 iov_iter_advance(to, ret);
 681                                 off += ret;
 682                         } else {
 683                                 iov_iter_advance(to, 0);
 684                         }
 685                         ceph_put_page_vector(pages, num_pages, false);
 686                 } else {
 687                         int idx = 0;
 688                         size_t left = ret > 0 ? ret : 0;
 689                         while (left > 0) {
 690                                 size_t len, copied;
 691                                 page_off = off & ~PAGE_MASK;
 692                                 len = min_t(size_t, left, PAGE_SIZE - page_off);
 693                                 copied = copy_page_to_iter(pages[idx++],
 694                                                            page_off, len, to);
 695                                 off += copied;
 696                                 left -= copied;
 697                                 if (copied < len) {
 698                                         ret = -EFAULT;
 699                                         break;
 700                                 }
 701                         }
 702                         ceph_release_page_vector(pages, num_pages);
 703                 }
 704
 705                 if (ret < 0) {
 706                         if (ret == -EBLACKLISTED)
 707                                 fsc->blacklisted = true;
 708                         break;
 709                 }
 710
 711                 if (off >= i_size || !more)
 712                         break;
 713         }
 714
 715         if (off > iocb->ki_pos) {
 716                 if (ret >= 0 &&
 717                     iov_iter_count(to) > 0 && off >= i_size_read(inode))
 718                         *retry_op = CHECK_EOF;
 719                 ret = off - iocb->ki_pos;
 720                 iocb->ki_pos = off;
 721         }
 722
 723         dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
 724         return ret;
 725 }
 726
 727 struct ceph_aio_request {
 728         struct kiocb *iocb;
 729         size_t total_len;
 730         bool write;
 731         bool should_dirty;
 732         int error;
 733         struct list_head osd_reqs;
 734         unsigned num_reqs;
 735         atomic_t pending_reqs;
 736         struct timespec64 mtime;
 737         struct ceph_cap_flush *prealloc_cf;
 738 };
 739
 740 struct ceph_aio_work {
 741         struct work_struct work;
 742         struct ceph_osd_request *req;
 743 };
 744
 745 static void ceph_aio_retry_work(struct work_struct *work);
 746
 747 static void ceph_aio_complete(struct inode *inode,
 748                               struct ceph_aio_request *aio_req)
 749 {
 750         struct ceph_inode_info *ci = ceph_inode(inode);
 751         int ret;
 752
 753         if (!atomic_dec_and_test(&aio_req->pending_reqs))
 754                 return;
 755
 756         if (aio_req->iocb->ki_flags & IOCB_DIRECT)
 757                 inode_dio_end(inode);
 758
 759         ret = aio_req->error;
 760         if (!ret)
 761                 ret = aio_req->total_len;
 762
 763         dout("ceph_aio_complete %p rc %d\n", inode, ret);
 764
 765         if (ret >= 0 && aio_req->write) {
 766                 int dirty;
 767
 768                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
 769                 if (endoff > i_size_read(inode)) {
 770                         if (ceph_inode_set_size(inode, endoff))
 771                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
 772                 }
 773
 774                 spin_lock(&ci->i_ceph_lock);
 775                 ci->i_inline_version = CEPH_INLINE_NONE;
 776                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
 777                                                &aio_req->prealloc_cf);
 778                 spin_unlock(&ci->i_ceph_lock);
 779                 if (dirty)
 780                         __mark_inode_dirty(inode, dirty);
 781
 782         }
 783
 784         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
 785                                                 CEPH_CAP_FILE_RD));
 786
 787         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
 788
 789         ceph_free_cap_flush(aio_req->prealloc_cf);
 790         kfree(aio_req);
 791 }
 792
 793 static void ceph_aio_complete_req(struct ceph_osd_request *req)
 794 {
 795         int rc = req->r_result;
 796         struct inode *inode = req->r_inode;
 797         struct ceph_aio_request *aio_req = req->r_priv;
 798         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
 799
 800         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
 801         BUG_ON(!osd_data->num_bvecs);
 802
 803         dout("ceph_aio_complete_req %p rc %d bytes %u\n",
 804              inode, rc, osd_data->bvec_pos.iter.bi_size);
 805
 806         if (rc == -EOLDSNAPC) {
 807                 struct ceph_aio_work *aio_work;
 808                 BUG_ON(!aio_req->write);
 809
 810                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
 811                 if (aio_work) {
 812                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
 813                         aio_work->req = req;
 814                         queue_work(ceph_inode_to_client(inode)->inode_wq,
 815                                    &aio_work->work);
 816                         return;
 817                 }
 818                 rc = -ENOMEM;
 819         } else if (!aio_req->write) {
 820                 if (rc == -ENOENT)
 821                         rc = 0;
 822                 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
 823                         struct iov_iter i;
 824                         int zlen = osd_data->bvec_pos.iter.bi_size - rc;
 825
 826                         /*
 827                          * If read is satisfied by single OSD request,
 828                          * it can pass EOF. Otherwise read is within
 829                          * i_size.
 830                          */
 831                         if (aio_req->num_reqs == 1) {
 832                                 loff_t i_size = i_size_read(inode);
 833                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
 834                                 if (endoff < i_size)
 835                                         zlen = min_t(size_t, zlen,
 836                                                      i_size - endoff);
 837                                 aio_req->total_len = rc + zlen;
 838                         }
 839
 840                         iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
 841                                       osd_data->num_bvecs,
 842                                       osd_data->bvec_pos.iter.bi_size);
 843                         iov_iter_advance(&i, rc);
 844                         iov_iter_zero(zlen, &i);
 845                 }
 846         }
 847
 848         put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
 849                   aio_req->should_dirty);
 850         ceph_osdc_put_request(req);
 851
 852         if (rc < 0)
 853                 cmpxchg(&aio_req->error, 0, rc);
 854
 855         ceph_aio_complete(inode, aio_req);
 856         return;
 857 }
 858
 859 static void ceph_aio_retry_work(struct work_struct *work)
 860 {
 861         struct ceph_aio_work *aio_work =
 862                 container_of(work, struct ceph_aio_work, work);
 863         struct ceph_osd_request *orig_req = aio_work->req;
 864         struct ceph_aio_request *aio_req = orig_req->r_priv;
 865         struct inode *inode = orig_req->r_inode;
 866         struct ceph_inode_info *ci = ceph_inode(inode);
 867         struct ceph_snap_context *snapc;
 868         struct ceph_osd_request *req;
 869         int ret;
 870
 871         spin_lock(&ci->i_ceph_lock);
 872         if (__ceph_have_pending_cap_snap(ci)) {
 873                 struct ceph_cap_snap *capsnap =
 874                         list_last_entry(&ci->i_cap_snaps,
 875                                         struct ceph_cap_snap,
 876                                         ci_item);
 877                 snapc = ceph_get_snap_context(capsnap->context);
 878         } else {
 879                 BUG_ON(!ci->i_head_snapc);
 880                 snapc = ceph_get_snap_context(ci->i_head_snapc);
 881         }
 882         spin_unlock(&ci->i_ceph_lock);
 883
 884         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
 885                         false, GFP_NOFS);
 886         if (!req) {
 887                 ret = -ENOMEM;
 888                 req = orig_req;
 889                 goto out;
 890         }
 891
 892         req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 893         ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
 894         ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
 895
 896         req->r_ops[0] = orig_req->r_ops[0];
 897
 898         req->r_mtime = aio_req->mtime;
 899         req->r_data_offset = req->r_ops[0].extent.offset;
 900
 901         ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
 902         if (ret) {
 903                 ceph_osdc_put_request(req);
 904                 req = orig_req;
 905                 goto out;
 906         }
 907
 908         ceph_osdc_put_request(orig_req);
 909
 910         req->r_callback = ceph_aio_complete_req;
 911         req->r_inode = inode;
 912         req->r_priv = aio_req;
 913
 914         ret = ceph_osdc_start_request(req->r_osdc, req, false);
 915 out:
 916         if (ret < 0) {
 917                 req->r_result = ret;
 918                 ceph_aio_complete_req(req);
 919         }
 920
 921         ceph_put_snap_context(snapc);
 922         kfree(aio_work);
 923 }
 924
 925 static ssize_t
 926 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
 927                        struct ceph_snap_context *snapc,
 928                        struct ceph_cap_flush **pcf)
 929 {
 930         struct file *file = iocb->ki_filp;
 931         struct inode *inode = file_inode(file);
 932         struct ceph_inode_info *ci = ceph_inode(inode);
 933         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 934         struct ceph_vino vino;
 935         struct ceph_osd_request *req;
 936         struct bio_vec *bvecs;
 937         struct ceph_aio_request *aio_req = NULL;
 938         int num_pages = 0;
 939         int flags;
 940         int ret = 0;
 941         struct timespec64 mtime = current_time(inode);
 942         size_t count = iov_iter_count(iter);
 943         loff_t pos = iocb->ki_pos;
 944         bool write = iov_iter_rw(iter) == WRITE;
 945         bool should_dirty = !write && iter_is_iovec(iter);
 946
 947         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
 948                 return -EROFS;
 949
 950         dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
 951              (write ? "write" : "read"), file, pos, (unsigned)count,
 952              snapc, snapc ? snapc->seq : 0);
 953
 954         if (write) {
 955                 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
 956                                         pos >> PAGE_SHIFT,
 957                                         (pos + count - 1) >> PAGE_SHIFT);
 958                 if (ret2 < 0)
 959                         dout("invalidate_inode_pages2_range returned %d\n", ret2);
 960
 961                 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
 962         } else {
 963                 flags = CEPH_OSD_FLAG_READ;
 964         }
 965
 966         while (iov_iter_count(iter) > 0) {
 967                 u64 size = iov_iter_count(iter);
 968                 ssize_t len;
 969
 970                 if (write)
 971                         size = min_t(u64, size, fsc->mount_options->wsize);
 972                 else
 973                         size = min_t(u64, size, fsc->mount_options->rsize);
 974
 975                 vino = ceph_vino(inode);
 976                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
 977                                             vino, pos, &size, 0,
 978                                             1,
 979                                             write ? CEPH_OSD_OP_WRITE :
 980                                                     CEPH_OSD_OP_READ,
 981                                             flags, snapc,
 982                                             ci->i_truncate_seq,
 983                                             ci->i_truncate_size,
 984                                             false);
 985                 if (IS_ERR(req)) {
 986                         ret = PTR_ERR(req);
 987                         break;
 988                 }
 989
 990                 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
 991                 if (len < 0) {
 992                         ceph_osdc_put_request(req);
 993                         ret = len;
 994                         break;
 995                 }
 996                 if (len != size)
 997                         osd_req_op_extent_update(req, 0, len);
 998
 999                 /*
1000                  * To simplify error handling, allow AIO when IO within i_size
1001                  * or IO can be satisfied by single OSD request.
1002                  */
1003                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1004                     (len == count || pos + count <= i_size_read(inode))) {
1005                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1006                         if (aio_req) {
1007                                 aio_req->iocb = iocb;
1008                                 aio_req->write = write;
1009                                 aio_req->should_dirty = should_dirty;
1010                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
1011                                 if (write) {
1012                                         aio_req->mtime = mtime;
1013                                         swap(aio_req->prealloc_cf, *pcf);
1014                                 }
1015                         }
1016                         /* ignore error */
1017                 }
1018
1019                 if (write) {
1020                         /*
1021                          * throw out any page cache pages in this range. this
1022                          * may block.
1023                          */
1024                         truncate_inode_pages_range(inode->i_mapping, pos,
1025                                                    PAGE_ALIGN(pos + len) - 1);
1026
1027                         req->r_mtime = mtime;
1028                 }
1029
1030                 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1031
1032                 if (aio_req) {
1033                         aio_req->total_len += len;
1034                         aio_req->num_reqs++;
1035                         atomic_inc(&aio_req->pending_reqs);
1036
1037                         req->r_callback = ceph_aio_complete_req;
1038                         req->r_inode = inode;
1039                         req->r_priv = aio_req;
1040                         list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1041
1042                         pos += len;
1043                         continue;
1044                 }
1045
1046                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1047                 if (!ret)
1048                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1049
1050                 size = i_size_read(inode);
1051                 if (!write) {
1052                         if (ret == -ENOENT)
1053                                 ret = 0;
1054                         if (ret >= 0 && ret < len && pos + ret < size) {
1055                                 struct iov_iter i;
1056                                 int zlen = min_t(size_t, len - ret,
1057                                                  size - pos - ret);
1058
1059                                 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1060                                 iov_iter_advance(&i, ret);
1061                                 iov_iter_zero(zlen, &i);
1062                                 ret += zlen;
1063                         }
1064                         if (ret >= 0)
1065                                 len = ret;
1066                 }
1067
1068                 put_bvecs(bvecs, num_pages, should_dirty);
1069                 ceph_osdc_put_request(req);
1070                 if (ret < 0)
1071                         break;
1072
1073                 pos += len;
1074                 if (!write && pos >= size)
1075                         break;
1076
1077                 if (write && pos > size) {
1078                         if (ceph_inode_set_size(inode, pos))
1079                                 ceph_check_caps(ceph_inode(inode),
1080                                                 CHECK_CAPS_AUTHONLY,
1081                                                 NULL);
1082                 }
1083         }
1084
1085         if (aio_req) {
1086                 LIST_HEAD(osd_reqs);
1087
1088                 if (aio_req->num_reqs == 0) {
1089                         kfree(aio_req);
1090                         return ret;
1091                 }
1092
1093                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1094                                               CEPH_CAP_FILE_RD);
1095
1096                 list_splice(&aio_req->osd_reqs, &osd_reqs);
1097                 inode_dio_begin(inode);
1098                 while (!list_empty(&osd_reqs)) {
1099                         req = list_first_entry(&osd_reqs,
1100                                                struct ceph_osd_request,
1101                                                r_private_item);
1102                         list_del_init(&req->r_private_item);
1103                         if (ret >= 0)
1104                                 ret = ceph_osdc_start_request(req->r_osdc,
1105                                                               req, false);
1106                         if (ret < 0) {
1107                                 req->r_result = ret;
1108                                 ceph_aio_complete_req(req);
1109                         }
1110                 }
1111                 return -EIOCBQUEUED;
1112         }
1113
1114         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1115                 ret = pos - iocb->ki_pos;
1116                 iocb->ki_pos = pos;
1117         }
1118         return ret;
1119 }
1120
1121 /*
1122  * Synchronous write, straight from __user pointer or user pages.
1123  *
1124  * If write spans object boundary, just do multiple writes.  (For a
1125  * correct atomic write, we should e.g. take write locks on all
1126  * objects, rollback on failure, etc.)
1127  */
1128 static ssize_t
1129 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1130                 struct ceph_snap_context *snapc)
1131 {
1132         struct file *file = iocb->ki_filp;
1133         struct inode *inode = file_inode(file);
1134         struct ceph_inode_info *ci = ceph_inode(inode);
1135         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1136         struct ceph_vino vino;
1137         struct ceph_osd_request *req;
1138         struct page **pages;
1139         u64 len;
1140         int num_pages;
1141         int written = 0;
1142         int flags;
1143         int ret;
1144         bool check_caps = false;
1145         struct timespec64 mtime = current_time(inode);
1146         size_t count = iov_iter_count(from);
1147
1148         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1149                 return -EROFS;
1150
1151         dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1152              file, pos, (unsigned)count, snapc, snapc->seq);
1153
1154         ret = filemap_write_and_wait_range(inode->i_mapping,
1155                                            pos, pos + count - 1);
1156         if (ret < 0)
1157                 return ret;
1158
1159         ret = invalidate_inode_pages2_range(inode->i_mapping,
1160                                             pos >> PAGE_SHIFT,
1161                                             (pos + count - 1) >> PAGE_SHIFT);
1162         if (ret < 0)
1163                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1164
1165         flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1166
1167         while ((len = iov_iter_count(from)) > 0) {
1168                 size_t left;
1169                 int n;
1170
1171                 vino = ceph_vino(inode);
1172                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1173                                             vino, pos, &len, 0, 1,
1174                                             CEPH_OSD_OP_WRITE, flags, snapc,
1175                                             ci->i_truncate_seq,
1176                                             ci->i_truncate_size,
1177                                             false);
1178                 if (IS_ERR(req)) {
1179                         ret = PTR_ERR(req);
1180                         break;
1181                 }
1182
1183                 /*
1184                  * write from beginning of first page,
1185                  * regardless of io alignment
1186                  */
1187                 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1188
1189                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1190                 if (IS_ERR(pages)) {
1191                         ret = PTR_ERR(pages);
1192                         goto out;
1193                 }
1194
1195                 left = len;
1196                 for (n = 0; n < num_pages; n++) {
1197                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1198                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1199                         if (ret != plen) {
1200                                 ret = -EFAULT;
1201                                 break;
1202                         }
1203                         left -= ret;
1204                 }
1205
1206                 if (ret < 0) {
1207                         ceph_release_page_vector(pages, num_pages);
1208                         goto out;
1209                 }
1210
1211                 req->r_inode = inode;
1212
1213                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1214                                                 false, true);
1215
1216                 req->r_mtime = mtime;
1217                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1218                 if (!ret)
1219                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1220
1221 out:
1222                 ceph_osdc_put_request(req);
1223                 if (ret != 0) {
1224                         ceph_set_error_write(ci);
1225                         break;
1226                 }
1227
1228                 ceph_clear_error_write(ci);
1229                 pos += len;
1230                 written += len;
1231                 if (pos > i_size_read(inode)) {
1232                         check_caps = ceph_inode_set_size(inode, pos);
1233                         if (check_caps)
1234                                 ceph_check_caps(ceph_inode(inode),
1235                                                 CHECK_CAPS_AUTHONLY,
1236                                                 NULL);
1237                 }
1238
1239         }
1240
1241         if (ret != -EOLDSNAPC && written > 0) {
1242                 ret = written;
1243                 iocb->ki_pos = pos;
1244         }
1245         return ret;
1246 }
1247
1248 /*
1249  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1250  * Atomically grab references, so that those bits are not released
1251  * back to the MDS mid-read.
1252  *
1253  * Hmm, the sync read case isn't actually async... should it be?
1254  */
1255 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1256 {
1257         struct file *filp = iocb->ki_filp;
1258         struct ceph_file_info *fi = filp->private_data;
1259         size_t len = iov_iter_count(to);
1260         struct inode *inode = file_inode(filp);
1261         struct ceph_inode_info *ci = ceph_inode(inode);
1262         struct page *pinned_page = NULL;
1263         ssize_t ret;
1264         int want, got = 0;
1265         int retry_op = 0, read = 0;
1266
1267 again:
1268         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1269              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1270
1271         if (iocb->ki_flags & IOCB_DIRECT)
1272                 ceph_start_io_direct(inode);
1273         else
1274                 ceph_start_io_read(inode);
1275
1276         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1277                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1278         else
1279                 want = CEPH_CAP_FILE_CACHE;
1280         ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1,
1281                             &got, &pinned_page);
1282         if (ret < 0) {
1283                 if (iocb->ki_flags & IOCB_DIRECT)
1284                         ceph_end_io_direct(inode);
1285                 else
1286                         ceph_end_io_read(inode);
1287                 return ret;
1288         }
1289
1290         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1291             (iocb->ki_flags & IOCB_DIRECT) ||
1292             (fi->flags & CEPH_F_SYNC)) {
1293
1294                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1295                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1296                      ceph_cap_string(got));
1297
1298                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1299                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1300                                 ret = ceph_direct_read_write(iocb, to,
1301                                                              NULL, NULL);
1302                                 if (ret >= 0 && ret < len)
1303                                         retry_op = CHECK_EOF;
1304                         } else {
1305                                 ret = ceph_sync_read(iocb, to, &retry_op);
1306                         }
1307                 } else {
1308                         retry_op = READ_INLINE;
1309                 }
1310         } else {
1311                 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1312                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1313                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1314                      ceph_cap_string(got));
1315                 ceph_add_rw_context(fi, &rw_ctx);
1316                 ret = generic_file_read_iter(iocb, to);
1317                 ceph_del_rw_context(fi, &rw_ctx);
1318         }
1319
1320         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1321              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1322         if (pinned_page) {
1323                 put_page(pinned_page);
1324                 pinned_page = NULL;
1325         }
1326         ceph_put_cap_refs(ci, got);
1327
1328         if (iocb->ki_flags & IOCB_DIRECT)
1329                 ceph_end_io_direct(inode);
1330         else
1331                 ceph_end_io_read(inode);
1332
1333         if (retry_op > HAVE_RETRIED && ret >= 0) {
1334                 int statret;
1335                 struct page *page = NULL;
1336                 loff_t i_size;
1337                 if (retry_op == READ_INLINE) {
1338                         page = __page_cache_alloc(GFP_KERNEL);
1339                         if (!page)
1340                                 return -ENOMEM;
1341                 }
1342
1343                 statret = __ceph_do_getattr(inode, page,
1344                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1345                 if (statret < 0) {
1346                         if (page)
1347                                 __free_page(page);
1348                         if (statret == -ENODATA) {
1349                                 BUG_ON(retry_op != READ_INLINE);
1350                                 goto again;
1351                         }
1352                         return statret;
1353                 }
1354
1355                 i_size = i_size_read(inode);
1356                 if (retry_op == READ_INLINE) {
1357                         BUG_ON(ret > 0 || read > 0);
1358                         if (iocb->ki_pos < i_size &&
1359                             iocb->ki_pos < PAGE_SIZE) {
1360                                 loff_t end = min_t(loff_t, i_size,
1361                                                    iocb->ki_pos + len);
1362                                 end = min_t(loff_t, end, PAGE_SIZE);
1363                                 if (statret < end)
1364                                         zero_user_segment(page, statret, end);
1365                                 ret = copy_page_to_iter(page,
1366                                                 iocb->ki_pos & ~PAGE_MASK,
1367                                                 end - iocb->ki_pos, to);
1368                                 iocb->ki_pos += ret;
1369                                 read += ret;
1370                         }
1371                         if (iocb->ki_pos < i_size && read < len) {
1372                                 size_t zlen = min_t(size_t, len - read,
1373                                                     i_size - iocb->ki_pos);
1374                                 ret = iov_iter_zero(zlen, to);
1375                                 iocb->ki_pos += ret;
1376                                 read += ret;
1377                         }
1378                         __free_pages(page, 0);
1379                         return read;
1380                 }
1381
1382                 /* hit EOF or hole? */
1383                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1384                     ret < len) {
1385                         dout("sync_read hit hole, ppos %lld < size %lld"
1386                              ", reading more\n", iocb->ki_pos, i_size);
1387
1388                         read += ret;
1389                         len -= ret;
1390                         retry_op = HAVE_RETRIED;
1391                         goto again;
1392                 }
1393         }
1394
1395         if (ret >= 0)
1396                 ret += read;
1397
1398         return ret;
1399 }
1400
1401 /*
1402  * Take cap references to avoid releasing caps to MDS mid-write.
1403  *
1404  * If we are synchronous, and write with an old snap context, the OSD
1405  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1406  * dropping our cap refs and allowing the pending snap to logically
1407  * complete _before_ this write occurs.
1408  *
1409  * If we are near ENOSPC, write synchronously.
1410  */
1411 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1412 {
1413         struct file *file = iocb->ki_filp;
1414         struct ceph_file_info *fi = file->private_data;
1415         struct inode *inode = file_inode(file);
1416         struct ceph_inode_info *ci = ceph_inode(inode);
1417         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1418         struct ceph_cap_flush *prealloc_cf;
1419         ssize_t count, written = 0;
1420         int err, want, got;
1421         loff_t pos;
1422         loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1423
1424         if (ceph_snap(inode) != CEPH_NOSNAP)
1425                 return -EROFS;
1426
1427         prealloc_cf = ceph_alloc_cap_flush();
1428         if (!prealloc_cf)
1429                 return -ENOMEM;
1430
1431 retry_snap:
1432         if (iocb->ki_flags & IOCB_DIRECT)
1433                 ceph_start_io_direct(inode);
1434         else
1435                 ceph_start_io_write(inode);
1436
1437         /* We can write back this queue in page reclaim */
1438         current->backing_dev_info = inode_to_bdi(inode);
1439
1440         if (iocb->ki_flags & IOCB_APPEND) {
1441                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1442                 if (err < 0)
1443                         goto out;
1444         }
1445
1446         err = generic_write_checks(iocb, from);
1447         if (err <= 0)
1448                 goto out;
1449
1450         pos = iocb->ki_pos;
1451         if (unlikely(pos >= limit)) {
1452                 err = -EFBIG;
1453                 goto out;
1454         } else {
1455                 iov_iter_truncate(from, limit - pos);
1456         }
1457
1458         count = iov_iter_count(from);
1459         if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1460                 err = -EDQUOT;
1461                 goto out;
1462         }
1463
1464         err = file_remove_privs(file);
1465         if (err)
1466                 goto out;
1467
1468         err = file_update_time(file);
1469         if (err)
1470                 goto out;
1471
1472         inode_inc_iversion_raw(inode);
1473
1474         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1475                 err = ceph_uninline_data(file, NULL);
1476                 if (err < 0)
1477                         goto out;
1478         }
1479
1480         /* FIXME: not complete since it doesn't account for being at quota */
1481         if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1482                 err = -ENOSPC;
1483                 goto out;
1484         }
1485
1486         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1487              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1488         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1489                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1490         else
1491                 want = CEPH_CAP_FILE_BUFFER;
1492         got = 0;
1493         err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count,
1494                             &got, NULL);
1495         if (err < 0)
1496                 goto out;
1497
1498         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1499              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1500
1501         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1502             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1503             (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1504                 struct ceph_snap_context *snapc;
1505                 struct iov_iter data;
1506
1507                 spin_lock(&ci->i_ceph_lock);
1508                 if (__ceph_have_pending_cap_snap(ci)) {
1509                         struct ceph_cap_snap *capsnap =
1510                                         list_last_entry(&ci->i_cap_snaps,
1511                                                         struct ceph_cap_snap,
1512                                                         ci_item);
1513                         snapc = ceph_get_snap_context(capsnap->context);
1514                 } else {
1515                         BUG_ON(!ci->i_head_snapc);
1516                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1517                 }
1518                 spin_unlock(&ci->i_ceph_lock);
1519
1520                 /* we might need to revert back to that point */
1521                 data = *from;
1522                 if (iocb->ki_flags & IOCB_DIRECT) {
1523                         written = ceph_direct_read_write(iocb, &data, snapc,
1524                                                          &prealloc_cf);
1525                         ceph_end_io_direct(inode);
1526                 } else {
1527                         written = ceph_sync_write(iocb, &data, pos, snapc);
1528                         ceph_end_io_write(inode);
1529                 }
1530                 if (written > 0)
1531                         iov_iter_advance(from, written);
1532                 ceph_put_snap_context(snapc);
1533         } else {
1534                 /*
1535                  * No need to acquire the i_truncate_mutex. Because
1536                  * the MDS revokes Fwb caps before sending truncate
1537                  * message to us. We can't get Fwb cap while there
1538                  * are pending vmtruncate. So write and vmtruncate
1539                  * can not run at the same time
1540                  */
1541                 written = generic_perform_write(file, from, pos);
1542                 if (likely(written >= 0))
1543                         iocb->ki_pos = pos + written;
1544                 ceph_end_io_write(inode);
1545         }
1546
1547         if (written >= 0) {
1548                 int dirty;
1549
1550                 spin_lock(&ci->i_ceph_lock);
1551                 ci->i_inline_version = CEPH_INLINE_NONE;
1552                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1553                                                &prealloc_cf);
1554                 spin_unlock(&ci->i_ceph_lock);
1555                 if (dirty)
1556                         __mark_inode_dirty(inode, dirty);
1557                 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1558                         ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1559         }
1560
1561         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1562              inode, ceph_vinop(inode), pos, (unsigned)count,
1563              ceph_cap_string(got));
1564         ceph_put_cap_refs(ci, got);
1565
1566         if (written == -EOLDSNAPC) {
1567                 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1568                      inode, ceph_vinop(inode), pos, (unsigned)count);
1569                 goto retry_snap;
1570         }
1571
1572         if (written >= 0) {
1573                 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1574                         iocb->ki_flags |= IOCB_DSYNC;
1575                 written = generic_write_sync(iocb, written);
1576         }
1577
1578         goto out_unlocked;
1579 out:
1580         if (iocb->ki_flags & IOCB_DIRECT)
1581                 ceph_end_io_direct(inode);
1582         else
1583                 ceph_end_io_write(inode);
1584 out_unlocked:
1585         ceph_free_cap_flush(prealloc_cf);
1586         current->backing_dev_info = NULL;
1587         return written ? written : err;
1588 }
1589
1590 /*
1591  * llseek.  be sure to verify file size on SEEK_END.
1592  */
1593 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1594 {
1595         struct inode *inode = file->f_mapping->host;
1596         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1597         loff_t i_size;
1598         loff_t ret;
1599
1600         inode_lock(inode);
1601
1602         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1603                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1604                 if (ret < 0)
1605                         goto out;
1606         }
1607
1608         i_size = i_size_read(inode);
1609         switch (whence) {
1610         case SEEK_END:
1611                 offset += i_size;
1612                 break;
1613         case SEEK_CUR:
1614                 /*
1615                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1616                  * position-querying operation.  Avoid rewriting the "same"
1617                  * f_pos value back to the file because a concurrent read(),
1618                  * write() or lseek() might have altered it
1619                  */
1620                 if (offset == 0) {
1621                         ret = file->f_pos;
1622                         goto out;
1623                 }
1624                 offset += file->f_pos;
1625                 break;
1626         case SEEK_DATA:
1627                 if (offset < 0 || offset >= i_size) {
1628                         ret = -ENXIO;
1629                         goto out;
1630                 }
1631                 break;
1632         case SEEK_HOLE:
1633                 if (offset < 0 || offset >= i_size) {
1634                         ret = -ENXIO;
1635                         goto out;
1636                 }
1637                 offset = i_size;
1638                 break;
1639         }
1640
1641         ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1642
1643 out:
1644         inode_unlock(inode);
1645         return ret;
1646 }
1647
1648 static inline void ceph_zero_partial_page(
1649         struct inode *inode, loff_t offset, unsigned size)
1650 {
1651         struct page *page;
1652         pgoff_t index = offset >> PAGE_SHIFT;
1653
1654         page = find_lock_page(inode->i_mapping, index);
1655         if (page) {
1656                 wait_on_page_writeback(page);
1657                 zero_user(page, offset & (PAGE_SIZE - 1), size);
1658                 unlock_page(page);
1659                 put_page(page);
1660         }
1661 }
1662
1663 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1664                                       loff_t length)
1665 {
1666         loff_t nearly = round_up(offset, PAGE_SIZE);
1667         if (offset < nearly) {
1668                 loff_t size = nearly - offset;
1669                 if (length < size)
1670                         size = length;
1671                 ceph_zero_partial_page(inode, offset, size);
1672                 offset += size;
1673                 length -= size;
1674         }
1675         if (length >= PAGE_SIZE) {
1676                 loff_t size = round_down(length, PAGE_SIZE);
1677                 truncate_pagecache_range(inode, offset, offset + size - 1);
1678                 offset += size;
1679                 length -= size;
1680         }
1681         if (length)
1682                 ceph_zero_partial_page(inode, offset, length);
1683 }
1684
1685 static int ceph_zero_partial_object(struct inode *inode,
1686                                     loff_t offset, loff_t *length)
1687 {
1688         struct ceph_inode_info *ci = ceph_inode(inode);
1689         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1690         struct ceph_osd_request *req;
1691         int ret = 0;
1692         loff_t zero = 0;
1693         int op;
1694
1695         if (!length) {
1696                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1697                 length = &zero;
1698         } else {
1699                 op = CEPH_OSD_OP_ZERO;
1700         }
1701
1702         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1703                                         ceph_vino(inode),
1704                                         offset, length,
1705                                         0, 1, op,
1706                                         CEPH_OSD_FLAG_WRITE,
1707                                         NULL, 0, 0, false);
1708         if (IS_ERR(req)) {
1709                 ret = PTR_ERR(req);
1710                 goto out;
1711         }
1712
1713         req->r_mtime = inode->i_mtime;
1714         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1715         if (!ret) {
1716                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1717                 if (ret == -ENOENT)
1718                         ret = 0;
1719         }
1720         ceph_osdc_put_request(req);
1721
1722 out:
1723         return ret;
1724 }
1725
1726 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1727 {
1728         int ret = 0;
1729         struct ceph_inode_info *ci = ceph_inode(inode);
1730         s32 stripe_unit = ci->i_layout.stripe_unit;
1731         s32 stripe_count = ci->i_layout.stripe_count;
1732         s32 object_size = ci->i_layout.object_size;
1733         u64 object_set_size = object_size * stripe_count;
1734         u64 nearly, t;
1735
1736         /* round offset up to next period boundary */
1737         nearly = offset + object_set_size - 1;
1738         t = nearly;
1739         nearly -= do_div(t, object_set_size);
1740
1741         while (length && offset < nearly) {
1742                 loff_t size = length;
1743                 ret = ceph_zero_partial_object(inode, offset, &size);
1744                 if (ret < 0)
1745                         return ret;
1746                 offset += size;
1747                 length -= size;
1748         }
1749         while (length >= object_set_size) {
1750                 int i;
1751                 loff_t pos = offset;
1752                 for (i = 0; i < stripe_count; ++i) {
1753                         ret = ceph_zero_partial_object(inode, pos, NULL);
1754                         if (ret < 0)
1755                                 return ret;
1756                         pos += stripe_unit;
1757                 }
1758                 offset += object_set_size;
1759                 length -= object_set_size;
1760         }
1761         while (length) {
1762                 loff_t size = length;
1763                 ret = ceph_zero_partial_object(inode, offset, &size);
1764                 if (ret < 0)
1765                         return ret;
1766                 offset += size;
1767                 length -= size;
1768         }
1769         return ret;
1770 }
1771
1772 static long ceph_fallocate(struct file *file, int mode,
1773                                 loff_t offset, loff_t length)
1774 {
1775         struct ceph_file_info *fi = file->private_data;
1776         struct inode *inode = file_inode(file);
1777         struct ceph_inode_info *ci = ceph_inode(inode);
1778         struct ceph_cap_flush *prealloc_cf;
1779         int want, got = 0;
1780         int dirty;
1781         int ret = 0;
1782         loff_t endoff = 0;
1783         loff_t size;
1784
1785         if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1786                 return -EOPNOTSUPP;
1787
1788         if (!S_ISREG(inode->i_mode))
1789                 return -EOPNOTSUPP;
1790
1791         prealloc_cf = ceph_alloc_cap_flush();
1792         if (!prealloc_cf)
1793                 return -ENOMEM;
1794
1795         inode_lock(inode);
1796
1797         if (ceph_snap(inode) != CEPH_NOSNAP) {
1798                 ret = -EROFS;
1799                 goto unlock;
1800         }
1801
1802         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1803                 ret = ceph_uninline_data(file, NULL);
1804                 if (ret < 0)
1805                         goto unlock;
1806         }
1807
1808         size = i_size_read(inode);
1809
1810         /* Are we punching a hole beyond EOF? */
1811         if (offset >= size)
1812                 goto unlock;
1813         if ((offset + length) > size)
1814                 length = size - offset;
1815
1816         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1817                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1818         else
1819                 want = CEPH_CAP_FILE_BUFFER;
1820
1821         ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1822         if (ret < 0)
1823                 goto unlock;
1824
1825         ceph_zero_pagecache_range(inode, offset, length);
1826         ret = ceph_zero_objects(inode, offset, length);
1827
1828         if (!ret) {
1829                 spin_lock(&ci->i_ceph_lock);
1830                 ci->i_inline_version = CEPH_INLINE_NONE;
1831                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1832                                                &prealloc_cf);
1833                 spin_unlock(&ci->i_ceph_lock);
1834                 if (dirty)
1835                         __mark_inode_dirty(inode, dirty);
1836         }
1837
1838         ceph_put_cap_refs(ci, got);
1839 unlock:
1840         inode_unlock(inode);
1841         ceph_free_cap_flush(prealloc_cf);
1842         return ret;
1843 }
1844
1845 /*
1846  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
1847  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
1848  * this fails; zero is returned on success.
1849  */
1850 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
1851                           struct file *dst_filp,
1852                           loff_t dst_endoff, int *dst_got)
1853 {
1854         int ret = 0;
1855         bool retrying = false;
1856
1857 retry_caps:
1858         ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
1859                             dst_endoff, dst_got, NULL);
1860         if (ret < 0)
1861                 return ret;
1862
1863         /*
1864          * Since we're already holding the FILE_WR capability for the dst file,
1865          * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
1866          * retry dance instead to try to get both capabilities.
1867          */
1868         ret = ceph_try_get_caps(file_inode(src_filp),
1869                                 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
1870                                 false, src_got);
1871         if (ret <= 0) {
1872                 /* Start by dropping dst_ci caps and getting src_ci caps */
1873                 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
1874                 if (retrying) {
1875                         if (!ret)
1876                                 /* ceph_try_get_caps masks EAGAIN */
1877                                 ret = -EAGAIN;
1878                         return ret;
1879                 }
1880                 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
1881                                     CEPH_CAP_FILE_SHARED, -1, src_got, NULL);
1882                 if (ret < 0)
1883                         return ret;
1884                 /*... drop src_ci caps too, and retry */
1885                 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
1886                 retrying = true;
1887                 goto retry_caps;
1888         }
1889         return ret;
1890 }
1891
1892 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
1893                            struct ceph_inode_info *dst_ci, int dst_got)
1894 {
1895         ceph_put_cap_refs(src_ci, src_got);
1896         ceph_put_cap_refs(dst_ci, dst_got);
1897 }
1898
1899 /*
1900  * This function does several size-related checks, returning an error if:
1901  *  - source file is smaller than off+len
1902  *  - destination file size is not OK (inode_newsize_ok())
1903  *  - max bytes quotas is exceeded
1904  */
1905 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
1906                            loff_t src_off, loff_t dst_off, size_t len)
1907 {
1908         loff_t size, endoff;
1909
1910         size = i_size_read(src_inode);
1911         /*
1912          * Don't copy beyond source file EOF.  Instead of simply setting length
1913          * to (size - src_off), just drop to VFS default implementation, as the
1914          * local i_size may be stale due to other clients writing to the source
1915          * inode.
1916          */
1917         if (src_off + len > size) {
1918                 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
1919                      src_off, len, size);
1920                 return -EOPNOTSUPP;
1921         }
1922         size = i_size_read(dst_inode);
1923
1924         endoff = dst_off + len;
1925         if (inode_newsize_ok(dst_inode, endoff))
1926                 return -EOPNOTSUPP;
1927
1928         if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
1929                 return -EDQUOT;
1930
1931         return 0;
1932 }
1933
1934 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
1935                                       struct file *dst_file, loff_t dst_off,
1936                                       size_t len, unsigned int flags)
1937 {
1938         struct inode *src_inode = file_inode(src_file);
1939         struct inode *dst_inode = file_inode(dst_file);
1940         struct ceph_inode_info *src_ci = ceph_inode(src_inode);
1941         struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
1942         struct ceph_cap_flush *prealloc_cf;
1943         struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
1944         struct ceph_object_locator src_oloc, dst_oloc;
1945         struct ceph_object_id src_oid, dst_oid;
1946         loff_t endoff = 0, size;
1947         ssize_t ret = -EIO;
1948         u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
1949         u32 src_objlen, dst_objlen, object_size;
1950         int src_got = 0, dst_got = 0, err, dirty;
1951         bool do_final_copy = false;
1952
1953         if (src_inode->i_sb != dst_inode->i_sb) {
1954                 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
1955
1956                 if (ceph_fsid_compare(&src_fsc->client->fsid,
1957                                       &dst_fsc->client->fsid)) {
1958                         dout("Copying files across clusters: src: %pU dst: %pU\n",
1959                              &src_fsc->client->fsid, &dst_fsc->client->fsid);
1960                         return -EXDEV;
1961                 }
1962         }
1963         if (ceph_snap(dst_inode) != CEPH_NOSNAP)
1964                 return -EROFS;
1965
1966         /*
1967          * Some of the checks below will return -EOPNOTSUPP, which will force a
1968          * fallback to the default VFS copy_file_range implementation.  This is
1969          * desirable in several cases (for ex, the 'len' is smaller than the
1970          * size of the objects, or in cases where that would be more
1971          * efficient).
1972          */
1973
1974         if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
1975                 return -EOPNOTSUPP;
1976
1977         /*
1978          * Striped file layouts require that we copy partial objects, but the
1979          * OSD copy-from operation only supports full-object copies.  Limit
1980          * this to non-striped file layouts for now.
1981          */
1982         if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
1983             (src_ci->i_layout.stripe_count != 1) ||
1984             (dst_ci->i_layout.stripe_count != 1) ||
1985             (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
1986                 dout("Invalid src/dst files layout\n");
1987                 return -EOPNOTSUPP;
1988         }
1989
1990         if (len < src_ci->i_layout.object_size)
1991                 return -EOPNOTSUPP; /* no remote copy will be done */
1992
1993         prealloc_cf = ceph_alloc_cap_flush();
1994         if (!prealloc_cf)
1995                 return -ENOMEM;
1996
1997         /* Start by sync'ing the source and destination files */
1998         ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
1999         if (ret < 0) {
2000                 dout("failed to write src file (%zd)\n", ret);
2001                 goto out;
2002         }
2003         ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2004         if (ret < 0) {
2005                 dout("failed to write dst file (%zd)\n", ret);
2006                 goto out;
2007         }
2008
2009         /*
2010          * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2011          * clients may have dirty data in their caches.  And OSDs know nothing
2012          * about caps, so they can't safely do the remote object copies.
2013          */
2014         err = get_rd_wr_caps(src_file, &src_got,
2015                              dst_file, (dst_off + len), &dst_got);
2016         if (err < 0) {
2017                 dout("get_rd_wr_caps returned %d\n", err);
2018                 ret = -EOPNOTSUPP;
2019                 goto out;
2020         }
2021
2022         ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2023         if (ret < 0)
2024                 goto out_caps;
2025
2026         size = i_size_read(dst_inode);
2027         endoff = dst_off + len;
2028
2029         /* Drop dst file cached pages */
2030         ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2031                                             dst_off >> PAGE_SHIFT,
2032                                             endoff >> PAGE_SHIFT);
2033         if (ret < 0) {
2034                 dout("Failed to invalidate inode pages (%zd)\n", ret);
2035                 ret = 0; /* XXX */
2036         }
2037         src_oloc.pool = src_ci->i_layout.pool_id;
2038         src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2039         dst_oloc.pool = dst_ci->i_layout.pool_id;
2040         dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2041
2042         ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2043                                       src_ci->i_layout.object_size,
2044                                       &src_objnum, &src_objoff, &src_objlen);
2045         ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2046                                       dst_ci->i_layout.object_size,
2047                                       &dst_objnum, &dst_objoff, &dst_objlen);
2048         /* object-level offsets need to the same */
2049         if (src_objoff != dst_objoff) {
2050                 ret = -EOPNOTSUPP;
2051                 goto out_caps;
2052         }
2053
2054         /*
2055          * Do a manual copy if the object offset isn't object aligned.
2056          * 'src_objlen' contains the bytes left until the end of the object,
2057          * starting at the src_off
2058          */
2059         if (src_objoff) {
2060                 /*
2061                  * we need to temporarily drop all caps as we'll be calling
2062                  * {read,write}_iter, which will get caps again.
2063                  */
2064                 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2065                 ret = do_splice_direct(src_file, &src_off, dst_file,
2066                                        &dst_off, src_objlen, flags);
2067                 if (ret < 0) {
2068                         dout("do_splice_direct returned %d\n", err);
2069                         goto out;
2070                 }
2071                 len -= ret;
2072                 err = get_rd_wr_caps(src_file, &src_got,
2073                                      dst_file, (dst_off + len), &dst_got);
2074                 if (err < 0)
2075                         goto out;
2076                 err = is_file_size_ok(src_inode, dst_inode,
2077                                       src_off, dst_off, len);
2078                 if (err < 0)
2079                         goto out_caps;
2080         }
2081         object_size = src_ci->i_layout.object_size;
2082         while (len >= object_size) {
2083                 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2084                                               object_size, &src_objnum,
2085                                               &src_objoff, &src_objlen);
2086                 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2087                                               object_size, &dst_objnum,
2088                                               &dst_objoff, &dst_objlen);
2089                 ceph_oid_init(&src_oid);
2090                 ceph_oid_printf(&src_oid, "%llx.%08llx",
2091                                 src_ci->i_vino.ino, src_objnum);
2092                 ceph_oid_init(&dst_oid);
2093                 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2094                                 dst_ci->i_vino.ino, dst_objnum);
2095                 /* Do an object remote copy */
2096                 err = ceph_osdc_copy_from(
2097                         &src_fsc->client->osdc,
2098                         src_ci->i_vino.snap, 0,
2099                         &src_oid, &src_oloc,
2100                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2101                         CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2102                         &dst_oid, &dst_oloc,
2103                         CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2104                         CEPH_OSD_OP_FLAG_FADVISE_DONTNEED, 0);
2105                 if (err) {
2106                         dout("ceph_osdc_copy_from returned %d\n", err);
2107                         if (!ret)
2108                                 ret = err;
2109                         goto out_caps;
2110                 }
2111                 len -= object_size;
2112                 src_off += object_size;
2113                 dst_off += object_size;
2114                 ret += object_size;
2115         }
2116
2117         if (len)
2118                 /* We still need one final local copy */
2119                 do_final_copy = true;
2120
2121         file_update_time(dst_file);
2122         inode_inc_iversion_raw(dst_inode);
2123
2124         if (endoff > size) {
2125                 int caps_flags = 0;
2126
2127                 /* Let the MDS know about dst file size change */
2128                 if (ceph_quota_is_max_bytes_approaching(dst_inode, endoff))
2129                         caps_flags |= CHECK_CAPS_NODELAY;
2130                 if (ceph_inode_set_size(dst_inode, endoff))
2131                         caps_flags |= CHECK_CAPS_AUTHONLY;
2132                 if (caps_flags)
2133                         ceph_check_caps(dst_ci, caps_flags, NULL);
2134         }
2135         /* Mark Fw dirty */
2136         spin_lock(&dst_ci->i_ceph_lock);
2137         dst_ci->i_inline_version = CEPH_INLINE_NONE;
2138         dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2139         spin_unlock(&dst_ci->i_ceph_lock);
2140         if (dirty)
2141                 __mark_inode_dirty(dst_inode, dirty);
2142
2143 out_caps:
2144         put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2145
2146         if (do_final_copy) {
2147                 err = do_splice_direct(src_file, &src_off, dst_file,
2148                                        &dst_off, len, flags);
2149                 if (err < 0) {
2150                         dout("do_splice_direct returned %d\n", err);
2151                         goto out;
2152                 }
2153                 len -= err;
2154                 ret += err;
2155         }
2156
2157 out:
2158         ceph_free_cap_flush(prealloc_cf);
2159
2160         return ret;
2161 }
2162
2163 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2164                                     struct file *dst_file, loff_t dst_off,
2165                                     size_t len, unsigned int flags)
2166 {
2167         ssize_t ret;
2168
2169         ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2170                                      len, flags);
2171
2172         if (ret == -EOPNOTSUPP || ret == -EXDEV)
2173                 ret = generic_copy_file_range(src_file, src_off, dst_file,
2174                                               dst_off, len, flags);
2175         return ret;
2176 }
2177
2178 const struct file_operations ceph_file_fops = {
2179         .open = ceph_open,
2180         .release = ceph_release,
2181         .llseek = ceph_llseek,
2182         .read_iter = ceph_read_iter,
2183         .write_iter = ceph_write_iter,
2184         .mmap = ceph_mmap,
2185         .fsync = ceph_fsync,
2186         .lock = ceph_lock,
2187         .flock = ceph_flock,
2188         .splice_read = generic_file_splice_read,
2189         .splice_write = iter_file_splice_write,
2190         .unlocked_ioctl = ceph_ioctl,
2191         .compat_ioctl = compat_ptr_ioctl,
2192         .fallocate      = ceph_fallocate,
2193         .copy_file_range = ceph_copy_file_range,
2194 };