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