1 #include <linux/ceph/ceph_debug.h> 2 3 #include <linux/module.h> 4 #include <linux/sched.h> 5 #include <linux/slab.h> 6 #include <linux/file.h> 7 #include <linux/mount.h> 8 #include <linux/namei.h> 9 #include <linux/writeback.h> 10 #include <linux/aio.h> 11 #include <linux/falloc.h> 12 13 #include "super.h" 14 #include "mds_client.h" 15 #include "cache.h" 16 17 /* 18 * Ceph file operations 19 * 20 * Implement basic open/close functionality, and implement 21 * read/write. 22 * 23 * We implement three modes of file I/O: 24 * - buffered uses the generic_file_aio_{read,write} helpers 25 * 26 * - synchronous is used when there is multi-client read/write 27 * sharing, avoids the page cache, and synchronously waits for an 28 * ack from the OSD. 29 * 30 * - direct io takes the variant of the sync path that references 31 * user pages directly. 32 * 33 * fsync() flushes and waits on dirty pages, but just queues metadata 34 * for writeback: since the MDS can recover size and mtime there is no 35 * need to wait for MDS acknowledgement. 36 */ 37 38 39 /* 40 * Prepare an open request. Preallocate ceph_cap to avoid an 41 * inopportune ENOMEM later. 42 */ 43 static struct ceph_mds_request * 44 prepare_open_request(struct super_block *sb, int flags, int create_mode) 45 { 46 struct ceph_fs_client *fsc = ceph_sb_to_client(sb); 47 struct ceph_mds_client *mdsc = fsc->mdsc; 48 struct ceph_mds_request *req; 49 int want_auth = USE_ANY_MDS; 50 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN; 51 52 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC)) 53 want_auth = USE_AUTH_MDS; 54 55 req = ceph_mdsc_create_request(mdsc, op, want_auth); 56 if (IS_ERR(req)) 57 goto out; 58 req->r_fmode = ceph_flags_to_mode(flags); 59 req->r_args.open.flags = cpu_to_le32(flags); 60 req->r_args.open.mode = cpu_to_le32(create_mode); 61 out: 62 return req; 63 } 64 65 /* 66 * initialize private struct file data. 67 * if we fail, clean up by dropping fmode reference on the ceph_inode 68 */ 69 static int ceph_init_file(struct inode *inode, struct file *file, int fmode) 70 { 71 struct ceph_file_info *cf; 72 int ret = 0; 73 struct ceph_inode_info *ci = ceph_inode(inode); 74 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); 75 struct ceph_mds_client *mdsc = fsc->mdsc; 76 77 switch (inode->i_mode & S_IFMT) { 78 case S_IFREG: 79 /* First file open request creates the cookie, we want to keep 80 * this cookie around for the filetime of the inode as not to 81 * have to worry about fscache register / revoke / operation 82 * races. 83 * 84 * Also, if we know the operation is going to invalidate data 85 * (non readonly) just nuke the cache right away. 86 */ 87 ceph_fscache_register_inode_cookie(mdsc->fsc, ci); 88 if ((fmode & CEPH_FILE_MODE_WR)) 89 ceph_fscache_invalidate(inode); 90 case S_IFDIR: 91 dout("init_file %p %p 0%o (regular)\n", inode, file, 92 inode->i_mode); 93 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); 94 if (cf == NULL) { 95 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 96 return -ENOMEM; 97 } 98 cf->fmode = fmode; 99 cf->next_offset = 2; 100 file->private_data = cf; 101 BUG_ON(inode->i_fop->release != ceph_release); 102 break; 103 104 case S_IFLNK: 105 dout("init_file %p %p 0%o (symlink)\n", inode, file, 106 inode->i_mode); 107 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 108 break; 109 110 default: 111 dout("init_file %p %p 0%o (special)\n", inode, file, 112 inode->i_mode); 113 /* 114 * we need to drop the open ref now, since we don't 115 * have .release set to ceph_release. 116 */ 117 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 118 BUG_ON(inode->i_fop->release == ceph_release); 119 120 /* call the proper open fop */ 121 ret = inode->i_fop->open(inode, file); 122 } 123 return ret; 124 } 125 126 /* 127 * If we already have the requisite capabilities, we can satisfy 128 * the open request locally (no need to request new caps from the 129 * MDS). We do, however, need to inform the MDS (asynchronously) 130 * if our wanted caps set expands. 131 */ 132 int ceph_open(struct inode *inode, struct file *file) 133 { 134 struct ceph_inode_info *ci = ceph_inode(inode); 135 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); 136 struct ceph_mds_client *mdsc = fsc->mdsc; 137 struct ceph_mds_request *req; 138 struct ceph_file_info *cf = file->private_data; 139 struct inode *parent_inode = NULL; 140 int err; 141 int flags, fmode, wanted; 142 143 if (cf) { 144 dout("open file %p is already opened\n", file); 145 return 0; 146 } 147 148 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */ 149 flags = file->f_flags & ~(O_CREAT|O_EXCL); 150 if (S_ISDIR(inode->i_mode)) 151 flags = O_DIRECTORY; /* mds likes to know */ 152 153 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode, 154 ceph_vinop(inode), file, flags, file->f_flags); 155 fmode = ceph_flags_to_mode(flags); 156 wanted = ceph_caps_for_mode(fmode); 157 158 /* snapped files are read-only */ 159 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE)) 160 return -EROFS; 161 162 /* trivially open snapdir */ 163 if (ceph_snap(inode) == CEPH_SNAPDIR) { 164 spin_lock(&ci->i_ceph_lock); 165 __ceph_get_fmode(ci, fmode); 166 spin_unlock(&ci->i_ceph_lock); 167 return ceph_init_file(inode, file, fmode); 168 } 169 170 /* 171 * No need to block if we have caps on the auth MDS (for 172 * write) or any MDS (for read). Update wanted set 173 * asynchronously. 174 */ 175 spin_lock(&ci->i_ceph_lock); 176 if (__ceph_is_any_real_caps(ci) && 177 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) { 178 int mds_wanted = __ceph_caps_mds_wanted(ci); 179 int issued = __ceph_caps_issued(ci, NULL); 180 181 dout("open %p fmode %d want %s issued %s using existing\n", 182 inode, fmode, ceph_cap_string(wanted), 183 ceph_cap_string(issued)); 184 __ceph_get_fmode(ci, fmode); 185 spin_unlock(&ci->i_ceph_lock); 186 187 /* adjust wanted? */ 188 if ((issued & wanted) != wanted && 189 (mds_wanted & wanted) != wanted && 190 ceph_snap(inode) != CEPH_SNAPDIR) 191 ceph_check_caps(ci, 0, NULL); 192 193 return ceph_init_file(inode, file, fmode); 194 } else if (ceph_snap(inode) != CEPH_NOSNAP && 195 (ci->i_snap_caps & wanted) == wanted) { 196 __ceph_get_fmode(ci, fmode); 197 spin_unlock(&ci->i_ceph_lock); 198 return ceph_init_file(inode, file, fmode); 199 } 200 201 spin_unlock(&ci->i_ceph_lock); 202 203 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted)); 204 req = prepare_open_request(inode->i_sb, flags, 0); 205 if (IS_ERR(req)) { 206 err = PTR_ERR(req); 207 goto out; 208 } 209 req->r_inode = inode; 210 ihold(inode); 211 212 req->r_num_caps = 1; 213 if (flags & O_CREAT) 214 parent_inode = ceph_get_dentry_parent_inode(file->f_path.dentry); 215 err = ceph_mdsc_do_request(mdsc, parent_inode, req); 216 iput(parent_inode); 217 if (!err) 218 err = ceph_init_file(inode, file, req->r_fmode); 219 ceph_mdsc_put_request(req); 220 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode)); 221 out: 222 return err; 223 } 224 225 226 /* 227 * Do a lookup + open with a single request. If we get a non-existent 228 * file or symlink, return 1 so the VFS can retry. 229 */ 230 int ceph_atomic_open(struct inode *dir, struct dentry *dentry, 231 struct file *file, unsigned flags, umode_t mode, 232 int *opened) 233 { 234 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); 235 struct ceph_mds_client *mdsc = fsc->mdsc; 236 struct ceph_mds_request *req; 237 struct dentry *dn; 238 struct ceph_acls_info acls = {}; 239 int err; 240 241 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n", 242 dir, dentry, dentry, 243 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode); 244 245 if (dentry->d_name.len > NAME_MAX) 246 return -ENAMETOOLONG; 247 248 err = ceph_init_dentry(dentry); 249 if (err < 0) 250 return err; 251 252 if (flags & O_CREAT) { 253 err = ceph_pre_init_acls(dir, &mode, &acls); 254 if (err < 0) 255 return err; 256 } 257 258 /* do the open */ 259 req = prepare_open_request(dir->i_sb, flags, mode); 260 if (IS_ERR(req)) { 261 err = PTR_ERR(req); 262 goto out_acl; 263 } 264 req->r_dentry = dget(dentry); 265 req->r_num_caps = 2; 266 if (flags & O_CREAT) { 267 req->r_dentry_drop = CEPH_CAP_FILE_SHARED; 268 req->r_dentry_unless = CEPH_CAP_FILE_EXCL; 269 if (acls.pagelist) { 270 req->r_pagelist = acls.pagelist; 271 acls.pagelist = NULL; 272 } 273 } 274 req->r_locked_dir = dir; /* caller holds dir->i_mutex */ 275 err = ceph_mdsc_do_request(mdsc, 276 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL, 277 req); 278 err = ceph_handle_snapdir(req, dentry, err); 279 if (err) 280 goto out_req; 281 282 if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry) 283 err = ceph_handle_notrace_create(dir, dentry); 284 285 if (d_unhashed(dentry)) { 286 dn = ceph_finish_lookup(req, dentry, err); 287 if (IS_ERR(dn)) 288 err = PTR_ERR(dn); 289 } else { 290 /* we were given a hashed negative dentry */ 291 dn = NULL; 292 } 293 if (err) 294 goto out_req; 295 if (dn || dentry->d_inode == NULL || d_is_symlink(dentry)) { 296 /* make vfs retry on splice, ENOENT, or symlink */ 297 dout("atomic_open finish_no_open on dn %p\n", dn); 298 err = finish_no_open(file, dn); 299 } else { 300 dout("atomic_open finish_open on dn %p\n", dn); 301 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) { 302 ceph_init_inode_acls(dentry->d_inode, &acls); 303 *opened |= FILE_CREATED; 304 } 305 err = finish_open(file, dentry, ceph_open, opened); 306 } 307 out_req: 308 if (!req->r_err && req->r_target_inode) 309 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode); 310 ceph_mdsc_put_request(req); 311 out_acl: 312 ceph_release_acls_info(&acls); 313 dout("atomic_open result=%d\n", err); 314 return err; 315 } 316 317 int ceph_release(struct inode *inode, struct file *file) 318 { 319 struct ceph_inode_info *ci = ceph_inode(inode); 320 struct ceph_file_info *cf = file->private_data; 321 322 dout("release inode %p file %p\n", inode, file); 323 ceph_put_fmode(ci, cf->fmode); 324 if (cf->last_readdir) 325 ceph_mdsc_put_request(cf->last_readdir); 326 kfree(cf->last_name); 327 kfree(cf->dir_info); 328 dput(cf->dentry); 329 kmem_cache_free(ceph_file_cachep, cf); 330 331 /* wake up anyone waiting for caps on this inode */ 332 wake_up_all(&ci->i_cap_wq); 333 return 0; 334 } 335 336 enum { 337 CHECK_EOF = 1, 338 READ_INLINE = 2, 339 }; 340 341 /* 342 * Read a range of bytes striped over one or more objects. Iterate over 343 * objects we stripe over. (That's not atomic, but good enough for now.) 344 * 345 * If we get a short result from the OSD, check against i_size; we need to 346 * only return a short read to the caller if we hit EOF. 347 */ 348 static int striped_read(struct inode *inode, 349 u64 off, u64 len, 350 struct page **pages, int num_pages, 351 int *checkeof, bool o_direct, 352 unsigned long buf_align) 353 { 354 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 355 struct ceph_inode_info *ci = ceph_inode(inode); 356 u64 pos, this_len, left; 357 int io_align, page_align; 358 int pages_left; 359 int read; 360 struct page **page_pos; 361 int ret; 362 bool hit_stripe, was_short; 363 364 /* 365 * we may need to do multiple reads. not atomic, unfortunately. 366 */ 367 pos = off; 368 left = len; 369 page_pos = pages; 370 pages_left = num_pages; 371 read = 0; 372 io_align = off & ~PAGE_MASK; 373 374 more: 375 if (o_direct) 376 page_align = (pos - io_align + buf_align) & ~PAGE_MASK; 377 else 378 page_align = pos & ~PAGE_MASK; 379 this_len = left; 380 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode), 381 &ci->i_layout, pos, &this_len, 382 ci->i_truncate_seq, 383 ci->i_truncate_size, 384 page_pos, pages_left, page_align); 385 if (ret == -ENOENT) 386 ret = 0; 387 hit_stripe = this_len < left; 388 was_short = ret >= 0 && ret < this_len; 389 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read, 390 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); 391 392 if (ret >= 0) { 393 int didpages; 394 if (was_short && (pos + ret < inode->i_size)) { 395 int zlen = min(this_len - ret, 396 inode->i_size - pos - ret); 397 int zoff = (o_direct ? buf_align : io_align) + 398 read + ret; 399 dout(" zero gap %llu to %llu\n", 400 pos + ret, pos + ret + zlen); 401 ceph_zero_page_vector_range(zoff, zlen, pages); 402 ret += zlen; 403 } 404 405 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT; 406 pos += ret; 407 read = pos - off; 408 left -= ret; 409 page_pos += didpages; 410 pages_left -= didpages; 411 412 /* hit stripe and need continue*/ 413 if (left && hit_stripe && pos < inode->i_size) 414 goto more; 415 } 416 417 if (read > 0) { 418 ret = read; 419 /* did we bounce off eof? */ 420 if (pos + left > inode->i_size) 421 *checkeof = CHECK_EOF; 422 } 423 424 dout("striped_read returns %d\n", ret); 425 return ret; 426 } 427 428 /* 429 * Completely synchronous read and write methods. Direct from __user 430 * buffer to osd, or directly to user pages (if O_DIRECT). 431 * 432 * If the read spans object boundary, just do multiple reads. 433 */ 434 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i, 435 int *checkeof) 436 { 437 struct file *file = iocb->ki_filp; 438 struct inode *inode = file_inode(file); 439 struct page **pages; 440 u64 off = iocb->ki_pos; 441 int num_pages, ret; 442 size_t len = iov_iter_count(i); 443 444 dout("sync_read on file %p %llu~%u %s\n", file, off, 445 (unsigned)len, 446 (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); 447 448 if (!len) 449 return 0; 450 /* 451 * flush any page cache pages in this range. this 452 * will make concurrent normal and sync io slow, 453 * but it will at least behave sensibly when they are 454 * in sequence. 455 */ 456 ret = filemap_write_and_wait_range(inode->i_mapping, off, 457 off + len); 458 if (ret < 0) 459 return ret; 460 461 if (file->f_flags & O_DIRECT) { 462 while (iov_iter_count(i)) { 463 size_t start; 464 ssize_t n; 465 466 n = iov_iter_get_pages_alloc(i, &pages, INT_MAX, &start); 467 if (n < 0) 468 return n; 469 470 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE; 471 472 ret = striped_read(inode, off, n, 473 pages, num_pages, checkeof, 474 1, start); 475 476 ceph_put_page_vector(pages, num_pages, true); 477 478 if (ret <= 0) 479 break; 480 off += ret; 481 iov_iter_advance(i, ret); 482 if (ret < n) 483 break; 484 } 485 } else { 486 num_pages = calc_pages_for(off, len); 487 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 488 if (IS_ERR(pages)) 489 return PTR_ERR(pages); 490 ret = striped_read(inode, off, len, pages, 491 num_pages, checkeof, 0, 0); 492 if (ret > 0) { 493 int l, k = 0; 494 size_t left = ret; 495 496 while (left) { 497 size_t page_off = off & ~PAGE_MASK; 498 size_t copy = min_t(size_t, 499 PAGE_SIZE - page_off, left); 500 l = copy_page_to_iter(pages[k++], page_off, 501 copy, i); 502 off += l; 503 left -= l; 504 if (l < copy) 505 break; 506 } 507 } 508 ceph_release_page_vector(pages, num_pages); 509 } 510 511 if (off > iocb->ki_pos) { 512 ret = off - iocb->ki_pos; 513 iocb->ki_pos = off; 514 } 515 516 dout("sync_read result %d\n", ret); 517 return ret; 518 } 519 520 /* 521 * Write commit request unsafe callback, called to tell us when a 522 * request is unsafe (that is, in flight--has been handed to the 523 * messenger to send to its target osd). It is called again when 524 * we've received a response message indicating the request is 525 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request 526 * is completed early (and unsuccessfully) due to a timeout or 527 * interrupt. 528 * 529 * This is used if we requested both an ACK and ONDISK commit reply 530 * from the OSD. 531 */ 532 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe) 533 { 534 struct ceph_inode_info *ci = ceph_inode(req->r_inode); 535 536 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid, 537 unsafe ? "un" : ""); 538 if (unsafe) { 539 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); 540 spin_lock(&ci->i_unsafe_lock); 541 list_add_tail(&req->r_unsafe_item, 542 &ci->i_unsafe_writes); 543 spin_unlock(&ci->i_unsafe_lock); 544 } else { 545 spin_lock(&ci->i_unsafe_lock); 546 list_del_init(&req->r_unsafe_item); 547 spin_unlock(&ci->i_unsafe_lock); 548 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); 549 } 550 } 551 552 553 /* 554 * Synchronous write, straight from __user pointer or user pages. 555 * 556 * If write spans object boundary, just do multiple writes. (For a 557 * correct atomic write, we should e.g. take write locks on all 558 * objects, rollback on failure, etc.) 559 */ 560 static ssize_t 561 ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 562 { 563 struct file *file = iocb->ki_filp; 564 struct inode *inode = file_inode(file); 565 struct ceph_inode_info *ci = ceph_inode(inode); 566 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 567 struct ceph_snap_context *snapc; 568 struct ceph_vino vino; 569 struct ceph_osd_request *req; 570 struct page **pages; 571 int num_pages; 572 int written = 0; 573 int flags; 574 int check_caps = 0; 575 int ret; 576 struct timespec mtime = CURRENT_TIME; 577 size_t count = iov_iter_count(from); 578 579 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 580 return -EROFS; 581 582 dout("sync_direct_write on file %p %lld~%u\n", file, pos, 583 (unsigned)count); 584 585 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 586 if (ret < 0) 587 return ret; 588 589 ret = invalidate_inode_pages2_range(inode->i_mapping, 590 pos >> PAGE_CACHE_SHIFT, 591 (pos + count) >> PAGE_CACHE_SHIFT); 592 if (ret < 0) 593 dout("invalidate_inode_pages2_range returned %d\n", ret); 594 595 flags = CEPH_OSD_FLAG_ORDERSNAP | 596 CEPH_OSD_FLAG_ONDISK | 597 CEPH_OSD_FLAG_WRITE; 598 599 while (iov_iter_count(from) > 0) { 600 u64 len = iov_iter_single_seg_count(from); 601 size_t start; 602 ssize_t n; 603 604 snapc = ci->i_snap_realm->cached_context; 605 vino = ceph_vino(inode); 606 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 607 vino, pos, &len, 0, 608 2,/*include a 'startsync' command*/ 609 CEPH_OSD_OP_WRITE, flags, snapc, 610 ci->i_truncate_seq, 611 ci->i_truncate_size, 612 false); 613 if (IS_ERR(req)) { 614 ret = PTR_ERR(req); 615 break; 616 } 617 618 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 619 620 n = iov_iter_get_pages_alloc(from, &pages, len, &start); 621 if (unlikely(n < 0)) { 622 ret = n; 623 ceph_osdc_put_request(req); 624 break; 625 } 626 627 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE; 628 /* 629 * throw out any page cache pages in this range. this 630 * may block. 631 */ 632 truncate_inode_pages_range(inode->i_mapping, pos, 633 (pos+n) | (PAGE_CACHE_SIZE-1)); 634 osd_req_op_extent_osd_data_pages(req, 0, pages, n, start, 635 false, false); 636 637 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 638 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 639 640 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 641 if (!ret) 642 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 643 644 ceph_put_page_vector(pages, num_pages, false); 645 646 ceph_osdc_put_request(req); 647 if (ret) 648 break; 649 pos += n; 650 written += n; 651 iov_iter_advance(from, n); 652 653 if (pos > i_size_read(inode)) { 654 check_caps = ceph_inode_set_size(inode, pos); 655 if (check_caps) 656 ceph_check_caps(ceph_inode(inode), 657 CHECK_CAPS_AUTHONLY, 658 NULL); 659 } 660 } 661 662 if (ret != -EOLDSNAPC && written > 0) { 663 iocb->ki_pos = pos; 664 ret = written; 665 } 666 return ret; 667 } 668 669 670 /* 671 * Synchronous write, straight from __user pointer or user pages. 672 * 673 * If write spans object boundary, just do multiple writes. (For a 674 * correct atomic write, we should e.g. take write locks on all 675 * objects, rollback on failure, etc.) 676 */ 677 static ssize_t 678 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 679 { 680 struct file *file = iocb->ki_filp; 681 struct inode *inode = file_inode(file); 682 struct ceph_inode_info *ci = ceph_inode(inode); 683 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 684 struct ceph_snap_context *snapc; 685 struct ceph_vino vino; 686 struct ceph_osd_request *req; 687 struct page **pages; 688 u64 len; 689 int num_pages; 690 int written = 0; 691 int flags; 692 int check_caps = 0; 693 int ret; 694 struct timespec mtime = CURRENT_TIME; 695 size_t count = iov_iter_count(from); 696 697 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 698 return -EROFS; 699 700 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count); 701 702 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 703 if (ret < 0) 704 return ret; 705 706 ret = invalidate_inode_pages2_range(inode->i_mapping, 707 pos >> PAGE_CACHE_SHIFT, 708 (pos + count) >> PAGE_CACHE_SHIFT); 709 if (ret < 0) 710 dout("invalidate_inode_pages2_range returned %d\n", ret); 711 712 flags = CEPH_OSD_FLAG_ORDERSNAP | 713 CEPH_OSD_FLAG_ONDISK | 714 CEPH_OSD_FLAG_WRITE | 715 CEPH_OSD_FLAG_ACK; 716 717 while ((len = iov_iter_count(from)) > 0) { 718 size_t left; 719 int n; 720 721 snapc = ci->i_snap_realm->cached_context; 722 vino = ceph_vino(inode); 723 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 724 vino, pos, &len, 0, 1, 725 CEPH_OSD_OP_WRITE, flags, snapc, 726 ci->i_truncate_seq, 727 ci->i_truncate_size, 728 false); 729 if (IS_ERR(req)) { 730 ret = PTR_ERR(req); 731 break; 732 } 733 734 /* 735 * write from beginning of first page, 736 * regardless of io alignment 737 */ 738 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; 739 740 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 741 if (IS_ERR(pages)) { 742 ret = PTR_ERR(pages); 743 goto out; 744 } 745 746 left = len; 747 for (n = 0; n < num_pages; n++) { 748 size_t plen = min_t(size_t, left, PAGE_SIZE); 749 ret = copy_page_from_iter(pages[n], 0, plen, from); 750 if (ret != plen) { 751 ret = -EFAULT; 752 break; 753 } 754 left -= ret; 755 } 756 757 if (ret < 0) { 758 ceph_release_page_vector(pages, num_pages); 759 goto out; 760 } 761 762 /* get a second commit callback */ 763 req->r_unsafe_callback = ceph_sync_write_unsafe; 764 req->r_inode = inode; 765 766 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, 767 false, true); 768 769 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 770 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 771 772 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 773 if (!ret) 774 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 775 776 out: 777 ceph_osdc_put_request(req); 778 if (ret == 0) { 779 pos += len; 780 written += len; 781 782 if (pos > i_size_read(inode)) { 783 check_caps = ceph_inode_set_size(inode, pos); 784 if (check_caps) 785 ceph_check_caps(ceph_inode(inode), 786 CHECK_CAPS_AUTHONLY, 787 NULL); 788 } 789 } else 790 break; 791 } 792 793 if (ret != -EOLDSNAPC && written > 0) { 794 ret = written; 795 iocb->ki_pos = pos; 796 } 797 return ret; 798 } 799 800 /* 801 * Wrap generic_file_aio_read with checks for cap bits on the inode. 802 * Atomically grab references, so that those bits are not released 803 * back to the MDS mid-read. 804 * 805 * Hmm, the sync read case isn't actually async... should it be? 806 */ 807 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to) 808 { 809 struct file *filp = iocb->ki_filp; 810 struct ceph_file_info *fi = filp->private_data; 811 size_t len = iocb->ki_nbytes; 812 struct inode *inode = file_inode(filp); 813 struct ceph_inode_info *ci = ceph_inode(inode); 814 struct page *pinned_page = NULL; 815 ssize_t ret; 816 int want, got = 0; 817 int retry_op = 0, read = 0; 818 819 again: 820 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", 821 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode); 822 823 if (fi->fmode & CEPH_FILE_MODE_LAZY) 824 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; 825 else 826 want = CEPH_CAP_FILE_CACHE; 827 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page); 828 if (ret < 0) 829 return ret; 830 831 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 || 832 (iocb->ki_filp->f_flags & O_DIRECT) || 833 (fi->flags & CEPH_F_SYNC)) { 834 835 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n", 836 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, 837 ceph_cap_string(got)); 838 839 if (ci->i_inline_version == CEPH_INLINE_NONE) { 840 /* hmm, this isn't really async... */ 841 ret = ceph_sync_read(iocb, to, &retry_op); 842 } else { 843 retry_op = READ_INLINE; 844 } 845 } else { 846 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", 847 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, 848 ceph_cap_string(got)); 849 850 ret = generic_file_read_iter(iocb, to); 851 } 852 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", 853 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); 854 if (pinned_page) { 855 page_cache_release(pinned_page); 856 pinned_page = NULL; 857 } 858 ceph_put_cap_refs(ci, got); 859 if (retry_op && ret >= 0) { 860 int statret; 861 struct page *page = NULL; 862 loff_t i_size; 863 if (retry_op == READ_INLINE) { 864 page = __page_cache_alloc(GFP_NOFS); 865 if (!page) 866 return -ENOMEM; 867 } 868 869 statret = __ceph_do_getattr(inode, page, 870 CEPH_STAT_CAP_INLINE_DATA, !!page); 871 if (statret < 0) { 872 __free_page(page); 873 if (statret == -ENODATA) { 874 BUG_ON(retry_op != READ_INLINE); 875 goto again; 876 } 877 return statret; 878 } 879 880 i_size = i_size_read(inode); 881 if (retry_op == READ_INLINE) { 882 BUG_ON(ret > 0 || read > 0); 883 if (iocb->ki_pos < i_size && 884 iocb->ki_pos < PAGE_CACHE_SIZE) { 885 loff_t end = min_t(loff_t, i_size, 886 iocb->ki_pos + len); 887 end = min_t(loff_t, end, PAGE_CACHE_SIZE); 888 if (statret < end) 889 zero_user_segment(page, statret, end); 890 ret = copy_page_to_iter(page, 891 iocb->ki_pos & ~PAGE_MASK, 892 end - iocb->ki_pos, to); 893 iocb->ki_pos += ret; 894 read += ret; 895 } 896 if (iocb->ki_pos < i_size && read < len) { 897 size_t zlen = min_t(size_t, len - read, 898 i_size - iocb->ki_pos); 899 ret = iov_iter_zero(zlen, to); 900 iocb->ki_pos += ret; 901 read += ret; 902 } 903 __free_pages(page, 0); 904 return read; 905 } 906 907 /* hit EOF or hole? */ 908 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size && 909 ret < len) { 910 dout("sync_read hit hole, ppos %lld < size %lld" 911 ", reading more\n", iocb->ki_pos, 912 inode->i_size); 913 914 read += ret; 915 len -= ret; 916 retry_op = 0; 917 goto again; 918 } 919 } 920 921 if (ret >= 0) 922 ret += read; 923 924 return ret; 925 } 926 927 /* 928 * Take cap references to avoid releasing caps to MDS mid-write. 929 * 930 * If we are synchronous, and write with an old snap context, the OSD 931 * may return EOLDSNAPC. In that case, retry the write.. _after_ 932 * dropping our cap refs and allowing the pending snap to logically 933 * complete _before_ this write occurs. 934 * 935 * If we are near ENOSPC, write synchronously. 936 */ 937 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from) 938 { 939 struct file *file = iocb->ki_filp; 940 struct ceph_file_info *fi = file->private_data; 941 struct inode *inode = file_inode(file); 942 struct ceph_inode_info *ci = ceph_inode(inode); 943 struct ceph_osd_client *osdc = 944 &ceph_sb_to_client(inode->i_sb)->client->osdc; 945 ssize_t count = iov_iter_count(from), written = 0; 946 int err, want, got; 947 loff_t pos = iocb->ki_pos; 948 949 if (ceph_snap(inode) != CEPH_NOSNAP) 950 return -EROFS; 951 952 mutex_lock(&inode->i_mutex); 953 954 /* We can write back this queue in page reclaim */ 955 current->backing_dev_info = inode_to_bdi(inode); 956 957 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode)); 958 if (err) 959 goto out; 960 961 if (count == 0) 962 goto out; 963 iov_iter_truncate(from, count); 964 965 err = file_remove_suid(file); 966 if (err) 967 goto out; 968 969 err = file_update_time(file); 970 if (err) 971 goto out; 972 973 if (ci->i_inline_version != CEPH_INLINE_NONE) { 974 err = ceph_uninline_data(file, NULL); 975 if (err < 0) 976 goto out; 977 } 978 979 retry_snap: 980 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) { 981 err = -ENOSPC; 982 goto out; 983 } 984 985 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n", 986 inode, ceph_vinop(inode), pos, count, inode->i_size); 987 if (fi->fmode & CEPH_FILE_MODE_LAZY) 988 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 989 else 990 want = CEPH_CAP_FILE_BUFFER; 991 got = 0; 992 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count, 993 &got, NULL); 994 if (err < 0) 995 goto out; 996 997 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n", 998 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got)); 999 1000 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || 1001 (file->f_flags & O_DIRECT) || (fi->flags & CEPH_F_SYNC)) { 1002 struct iov_iter data; 1003 mutex_unlock(&inode->i_mutex); 1004 /* we might need to revert back to that point */ 1005 data = *from; 1006 if (file->f_flags & O_DIRECT) 1007 written = ceph_sync_direct_write(iocb, &data, pos); 1008 else 1009 written = ceph_sync_write(iocb, &data, pos); 1010 if (written == -EOLDSNAPC) { 1011 dout("aio_write %p %llx.%llx %llu~%u" 1012 "got EOLDSNAPC, retrying\n", 1013 inode, ceph_vinop(inode), 1014 pos, (unsigned)count); 1015 mutex_lock(&inode->i_mutex); 1016 goto retry_snap; 1017 } 1018 if (written > 0) 1019 iov_iter_advance(from, written); 1020 } else { 1021 loff_t old_size = inode->i_size; 1022 /* 1023 * No need to acquire the i_truncate_mutex. Because 1024 * the MDS revokes Fwb caps before sending truncate 1025 * message to us. We can't get Fwb cap while there 1026 * are pending vmtruncate. So write and vmtruncate 1027 * can not run at the same time 1028 */ 1029 written = generic_perform_write(file, from, pos); 1030 if (likely(written >= 0)) 1031 iocb->ki_pos = pos + written; 1032 if (inode->i_size > old_size) 1033 ceph_fscache_update_objectsize(inode); 1034 mutex_unlock(&inode->i_mutex); 1035 } 1036 1037 if (written >= 0) { 1038 int dirty; 1039 spin_lock(&ci->i_ceph_lock); 1040 ci->i_inline_version = CEPH_INLINE_NONE; 1041 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1042 spin_unlock(&ci->i_ceph_lock); 1043 if (dirty) 1044 __mark_inode_dirty(inode, dirty); 1045 } 1046 1047 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", 1048 inode, ceph_vinop(inode), pos, (unsigned)count, 1049 ceph_cap_string(got)); 1050 ceph_put_cap_refs(ci, got); 1051 1052 if (written >= 0 && 1053 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) || 1054 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { 1055 err = vfs_fsync_range(file, pos, pos + written - 1, 1); 1056 if (err < 0) 1057 written = err; 1058 } 1059 1060 goto out_unlocked; 1061 1062 out: 1063 mutex_unlock(&inode->i_mutex); 1064 out_unlocked: 1065 current->backing_dev_info = NULL; 1066 return written ? written : err; 1067 } 1068 1069 /* 1070 * llseek. be sure to verify file size on SEEK_END. 1071 */ 1072 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence) 1073 { 1074 struct inode *inode = file->f_mapping->host; 1075 int ret; 1076 1077 mutex_lock(&inode->i_mutex); 1078 1079 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) { 1080 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false); 1081 if (ret < 0) { 1082 offset = ret; 1083 goto out; 1084 } 1085 } 1086 1087 switch (whence) { 1088 case SEEK_END: 1089 offset += inode->i_size; 1090 break; 1091 case SEEK_CUR: 1092 /* 1093 * Here we special-case the lseek(fd, 0, SEEK_CUR) 1094 * position-querying operation. Avoid rewriting the "same" 1095 * f_pos value back to the file because a concurrent read(), 1096 * write() or lseek() might have altered it 1097 */ 1098 if (offset == 0) { 1099 offset = file->f_pos; 1100 goto out; 1101 } 1102 offset += file->f_pos; 1103 break; 1104 case SEEK_DATA: 1105 if (offset >= inode->i_size) { 1106 ret = -ENXIO; 1107 goto out; 1108 } 1109 break; 1110 case SEEK_HOLE: 1111 if (offset >= inode->i_size) { 1112 ret = -ENXIO; 1113 goto out; 1114 } 1115 offset = inode->i_size; 1116 break; 1117 } 1118 1119 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes); 1120 1121 out: 1122 mutex_unlock(&inode->i_mutex); 1123 return offset; 1124 } 1125 1126 static inline void ceph_zero_partial_page( 1127 struct inode *inode, loff_t offset, unsigned size) 1128 { 1129 struct page *page; 1130 pgoff_t index = offset >> PAGE_CACHE_SHIFT; 1131 1132 page = find_lock_page(inode->i_mapping, index); 1133 if (page) { 1134 wait_on_page_writeback(page); 1135 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size); 1136 unlock_page(page); 1137 page_cache_release(page); 1138 } 1139 } 1140 1141 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset, 1142 loff_t length) 1143 { 1144 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE); 1145 if (offset < nearly) { 1146 loff_t size = nearly - offset; 1147 if (length < size) 1148 size = length; 1149 ceph_zero_partial_page(inode, offset, size); 1150 offset += size; 1151 length -= size; 1152 } 1153 if (length >= PAGE_CACHE_SIZE) { 1154 loff_t size = round_down(length, PAGE_CACHE_SIZE); 1155 truncate_pagecache_range(inode, offset, offset + size - 1); 1156 offset += size; 1157 length -= size; 1158 } 1159 if (length) 1160 ceph_zero_partial_page(inode, offset, length); 1161 } 1162 1163 static int ceph_zero_partial_object(struct inode *inode, 1164 loff_t offset, loff_t *length) 1165 { 1166 struct ceph_inode_info *ci = ceph_inode(inode); 1167 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1168 struct ceph_osd_request *req; 1169 int ret = 0; 1170 loff_t zero = 0; 1171 int op; 1172 1173 if (!length) { 1174 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE; 1175 length = &zero; 1176 } else { 1177 op = CEPH_OSD_OP_ZERO; 1178 } 1179 1180 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 1181 ceph_vino(inode), 1182 offset, length, 1183 0, 1, op, 1184 CEPH_OSD_FLAG_WRITE | 1185 CEPH_OSD_FLAG_ONDISK, 1186 NULL, 0, 0, false); 1187 if (IS_ERR(req)) { 1188 ret = PTR_ERR(req); 1189 goto out; 1190 } 1191 1192 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap, 1193 &inode->i_mtime); 1194 1195 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 1196 if (!ret) { 1197 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 1198 if (ret == -ENOENT) 1199 ret = 0; 1200 } 1201 ceph_osdc_put_request(req); 1202 1203 out: 1204 return ret; 1205 } 1206 1207 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length) 1208 { 1209 int ret = 0; 1210 struct ceph_inode_info *ci = ceph_inode(inode); 1211 s32 stripe_unit = ceph_file_layout_su(ci->i_layout); 1212 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout); 1213 s32 object_size = ceph_file_layout_object_size(ci->i_layout); 1214 u64 object_set_size = object_size * stripe_count; 1215 u64 nearly, t; 1216 1217 /* round offset up to next period boundary */ 1218 nearly = offset + object_set_size - 1; 1219 t = nearly; 1220 nearly -= do_div(t, object_set_size); 1221 1222 while (length && offset < nearly) { 1223 loff_t size = length; 1224 ret = ceph_zero_partial_object(inode, offset, &size); 1225 if (ret < 0) 1226 return ret; 1227 offset += size; 1228 length -= size; 1229 } 1230 while (length >= object_set_size) { 1231 int i; 1232 loff_t pos = offset; 1233 for (i = 0; i < stripe_count; ++i) { 1234 ret = ceph_zero_partial_object(inode, pos, NULL); 1235 if (ret < 0) 1236 return ret; 1237 pos += stripe_unit; 1238 } 1239 offset += object_set_size; 1240 length -= object_set_size; 1241 } 1242 while (length) { 1243 loff_t size = length; 1244 ret = ceph_zero_partial_object(inode, offset, &size); 1245 if (ret < 0) 1246 return ret; 1247 offset += size; 1248 length -= size; 1249 } 1250 return ret; 1251 } 1252 1253 static long ceph_fallocate(struct file *file, int mode, 1254 loff_t offset, loff_t length) 1255 { 1256 struct ceph_file_info *fi = file->private_data; 1257 struct inode *inode = file_inode(file); 1258 struct ceph_inode_info *ci = ceph_inode(inode); 1259 struct ceph_osd_client *osdc = 1260 &ceph_inode_to_client(inode)->client->osdc; 1261 int want, got = 0; 1262 int dirty; 1263 int ret = 0; 1264 loff_t endoff = 0; 1265 loff_t size; 1266 1267 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE)) 1268 return -EOPNOTSUPP; 1269 1270 if (!S_ISREG(inode->i_mode)) 1271 return -EOPNOTSUPP; 1272 1273 mutex_lock(&inode->i_mutex); 1274 1275 if (ceph_snap(inode) != CEPH_NOSNAP) { 1276 ret = -EROFS; 1277 goto unlock; 1278 } 1279 1280 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) && 1281 !(mode & FALLOC_FL_PUNCH_HOLE)) { 1282 ret = -ENOSPC; 1283 goto unlock; 1284 } 1285 1286 if (ci->i_inline_version != CEPH_INLINE_NONE) { 1287 ret = ceph_uninline_data(file, NULL); 1288 if (ret < 0) 1289 goto unlock; 1290 } 1291 1292 size = i_size_read(inode); 1293 if (!(mode & FALLOC_FL_KEEP_SIZE)) 1294 endoff = offset + length; 1295 1296 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1297 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 1298 else 1299 want = CEPH_CAP_FILE_BUFFER; 1300 1301 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL); 1302 if (ret < 0) 1303 goto unlock; 1304 1305 if (mode & FALLOC_FL_PUNCH_HOLE) { 1306 if (offset < size) 1307 ceph_zero_pagecache_range(inode, offset, length); 1308 ret = ceph_zero_objects(inode, offset, length); 1309 } else if (endoff > size) { 1310 truncate_pagecache_range(inode, size, -1); 1311 if (ceph_inode_set_size(inode, endoff)) 1312 ceph_check_caps(ceph_inode(inode), 1313 CHECK_CAPS_AUTHONLY, NULL); 1314 } 1315 1316 if (!ret) { 1317 spin_lock(&ci->i_ceph_lock); 1318 ci->i_inline_version = CEPH_INLINE_NONE; 1319 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1320 spin_unlock(&ci->i_ceph_lock); 1321 if (dirty) 1322 __mark_inode_dirty(inode, dirty); 1323 } 1324 1325 ceph_put_cap_refs(ci, got); 1326 unlock: 1327 mutex_unlock(&inode->i_mutex); 1328 return ret; 1329 } 1330 1331 const struct file_operations ceph_file_fops = { 1332 .open = ceph_open, 1333 .release = ceph_release, 1334 .llseek = ceph_llseek, 1335 .read = new_sync_read, 1336 .write = new_sync_write, 1337 .read_iter = ceph_read_iter, 1338 .write_iter = ceph_write_iter, 1339 .mmap = ceph_mmap, 1340 .fsync = ceph_fsync, 1341 .lock = ceph_lock, 1342 .flock = ceph_flock, 1343 .splice_read = generic_file_splice_read, 1344 .splice_write = iter_file_splice_write, 1345 .unlocked_ioctl = ceph_ioctl, 1346 .compat_ioctl = ceph_ioctl, 1347 .fallocate = ceph_fallocate, 1348 }; 1349 1350