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 if (err) 279 goto out_req; 280 281 err = ceph_handle_snapdir(req, dentry, err); 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 || S_ISLNK(dentry->d_inode->i_mode)) { 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 u64 tmp = min(this_len - ret, 396 inode->i_size - pos - ret); 397 dout(" zero gap %llu to %llu\n", 398 pos + ret, pos + ret + tmp); 399 ceph_zero_page_vector_range(page_align + read + ret, 400 tmp, pages); 401 ret += tmp; 402 } 403 404 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT; 405 pos += ret; 406 read = pos - off; 407 left -= ret; 408 page_pos += didpages; 409 pages_left -= didpages; 410 411 /* hit stripe and need continue*/ 412 if (left && hit_stripe && pos < inode->i_size) 413 goto more; 414 } 415 416 if (read > 0) { 417 ret = read; 418 /* did we bounce off eof? */ 419 if (pos + left > inode->i_size) 420 *checkeof = CHECK_EOF; 421 } 422 423 dout("striped_read returns %d\n", ret); 424 return ret; 425 } 426 427 /* 428 * Completely synchronous read and write methods. Direct from __user 429 * buffer to osd, or directly to user pages (if O_DIRECT). 430 * 431 * If the read spans object boundary, just do multiple reads. 432 */ 433 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i, 434 int *checkeof) 435 { 436 struct file *file = iocb->ki_filp; 437 struct inode *inode = file_inode(file); 438 struct page **pages; 439 u64 off = iocb->ki_pos; 440 int num_pages, ret; 441 size_t len = iov_iter_count(i); 442 443 dout("sync_read on file %p %llu~%u %s\n", file, off, 444 (unsigned)len, 445 (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); 446 447 if (!len) 448 return 0; 449 /* 450 * flush any page cache pages in this range. this 451 * will make concurrent normal and sync io slow, 452 * but it will at least behave sensibly when they are 453 * in sequence. 454 */ 455 ret = filemap_write_and_wait_range(inode->i_mapping, off, 456 off + len); 457 if (ret < 0) 458 return ret; 459 460 if (file->f_flags & O_DIRECT) { 461 while (iov_iter_count(i)) { 462 size_t start; 463 ssize_t n; 464 465 n = iov_iter_get_pages_alloc(i, &pages, INT_MAX, &start); 466 if (n < 0) 467 return n; 468 469 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE; 470 471 ret = striped_read(inode, off, n, 472 pages, num_pages, checkeof, 473 1, start); 474 475 ceph_put_page_vector(pages, num_pages, true); 476 477 if (ret <= 0) 478 break; 479 off += ret; 480 iov_iter_advance(i, ret); 481 if (ret < n) 482 break; 483 } 484 } else { 485 num_pages = calc_pages_for(off, len); 486 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 487 if (IS_ERR(pages)) 488 return PTR_ERR(pages); 489 ret = striped_read(inode, off, len, pages, 490 num_pages, checkeof, 0, 0); 491 if (ret > 0) { 492 int l, k = 0; 493 size_t left = ret; 494 495 while (left) { 496 size_t page_off = off & ~PAGE_MASK; 497 size_t copy = min_t(size_t, 498 PAGE_SIZE - page_off, left); 499 l = copy_page_to_iter(pages[k++], page_off, 500 copy, i); 501 off += l; 502 left -= l; 503 if (l < copy) 504 break; 505 } 506 } 507 ceph_release_page_vector(pages, num_pages); 508 } 509 510 if (off > iocb->ki_pos) { 511 ret = off - iocb->ki_pos; 512 iocb->ki_pos = off; 513 } 514 515 dout("sync_read result %d\n", ret); 516 return ret; 517 } 518 519 /* 520 * Write commit request unsafe callback, called to tell us when a 521 * request is unsafe (that is, in flight--has been handed to the 522 * messenger to send to its target osd). It is called again when 523 * we've received a response message indicating the request is 524 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request 525 * is completed early (and unsuccessfully) due to a timeout or 526 * interrupt. 527 * 528 * This is used if we requested both an ACK and ONDISK commit reply 529 * from the OSD. 530 */ 531 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe) 532 { 533 struct ceph_inode_info *ci = ceph_inode(req->r_inode); 534 535 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid, 536 unsafe ? "un" : ""); 537 if (unsafe) { 538 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); 539 spin_lock(&ci->i_unsafe_lock); 540 list_add_tail(&req->r_unsafe_item, 541 &ci->i_unsafe_writes); 542 spin_unlock(&ci->i_unsafe_lock); 543 } else { 544 spin_lock(&ci->i_unsafe_lock); 545 list_del_init(&req->r_unsafe_item); 546 spin_unlock(&ci->i_unsafe_lock); 547 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); 548 } 549 } 550 551 552 /* 553 * Synchronous write, straight from __user pointer or user pages. 554 * 555 * If write spans object boundary, just do multiple writes. (For a 556 * correct atomic write, we should e.g. take write locks on all 557 * objects, rollback on failure, etc.) 558 */ 559 static ssize_t 560 ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 561 { 562 struct file *file = iocb->ki_filp; 563 struct inode *inode = file_inode(file); 564 struct ceph_inode_info *ci = ceph_inode(inode); 565 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 566 struct ceph_snap_context *snapc; 567 struct ceph_vino vino; 568 struct ceph_osd_request *req; 569 struct page **pages; 570 int num_pages; 571 int written = 0; 572 int flags; 573 int check_caps = 0; 574 int ret; 575 struct timespec mtime = CURRENT_TIME; 576 size_t count = iov_iter_count(from); 577 578 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 579 return -EROFS; 580 581 dout("sync_direct_write on file %p %lld~%u\n", file, pos, 582 (unsigned)count); 583 584 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 585 if (ret < 0) 586 return ret; 587 588 ret = invalidate_inode_pages2_range(inode->i_mapping, 589 pos >> PAGE_CACHE_SHIFT, 590 (pos + count) >> PAGE_CACHE_SHIFT); 591 if (ret < 0) 592 dout("invalidate_inode_pages2_range returned %d\n", ret); 593 594 flags = CEPH_OSD_FLAG_ORDERSNAP | 595 CEPH_OSD_FLAG_ONDISK | 596 CEPH_OSD_FLAG_WRITE; 597 598 while (iov_iter_count(from) > 0) { 599 u64 len = iov_iter_single_seg_count(from); 600 size_t start; 601 ssize_t n; 602 603 snapc = ci->i_snap_realm->cached_context; 604 vino = ceph_vino(inode); 605 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 606 vino, pos, &len, 0, 607 2,/*include a 'startsync' command*/ 608 CEPH_OSD_OP_WRITE, flags, snapc, 609 ci->i_truncate_seq, 610 ci->i_truncate_size, 611 false); 612 if (IS_ERR(req)) { 613 ret = PTR_ERR(req); 614 break; 615 } 616 617 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 618 619 n = iov_iter_get_pages_alloc(from, &pages, len, &start); 620 if (unlikely(n < 0)) { 621 ret = n; 622 ceph_osdc_put_request(req); 623 break; 624 } 625 626 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE; 627 /* 628 * throw out any page cache pages in this range. this 629 * may block. 630 */ 631 truncate_inode_pages_range(inode->i_mapping, pos, 632 (pos+n) | (PAGE_CACHE_SIZE-1)); 633 osd_req_op_extent_osd_data_pages(req, 0, pages, n, start, 634 false, false); 635 636 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 637 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 638 639 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 640 if (!ret) 641 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 642 643 ceph_put_page_vector(pages, num_pages, false); 644 645 ceph_osdc_put_request(req); 646 if (ret) 647 break; 648 pos += n; 649 written += n; 650 iov_iter_advance(from, n); 651 652 if (pos > i_size_read(inode)) { 653 check_caps = ceph_inode_set_size(inode, pos); 654 if (check_caps) 655 ceph_check_caps(ceph_inode(inode), 656 CHECK_CAPS_AUTHONLY, 657 NULL); 658 } 659 } 660 661 if (ret != -EOLDSNAPC && written > 0) { 662 iocb->ki_pos = pos; 663 ret = written; 664 } 665 return ret; 666 } 667 668 669 /* 670 * Synchronous write, straight from __user pointer or user pages. 671 * 672 * If write spans object boundary, just do multiple writes. (For a 673 * correct atomic write, we should e.g. take write locks on all 674 * objects, rollback on failure, etc.) 675 */ 676 static ssize_t 677 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 678 { 679 struct file *file = iocb->ki_filp; 680 struct inode *inode = file_inode(file); 681 struct ceph_inode_info *ci = ceph_inode(inode); 682 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 683 struct ceph_snap_context *snapc; 684 struct ceph_vino vino; 685 struct ceph_osd_request *req; 686 struct page **pages; 687 u64 len; 688 int num_pages; 689 int written = 0; 690 int flags; 691 int check_caps = 0; 692 int ret; 693 struct timespec mtime = CURRENT_TIME; 694 size_t count = iov_iter_count(from); 695 696 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP) 697 return -EROFS; 698 699 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count); 700 701 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count); 702 if (ret < 0) 703 return ret; 704 705 ret = invalidate_inode_pages2_range(inode->i_mapping, 706 pos >> PAGE_CACHE_SHIFT, 707 (pos + count) >> PAGE_CACHE_SHIFT); 708 if (ret < 0) 709 dout("invalidate_inode_pages2_range returned %d\n", ret); 710 711 flags = CEPH_OSD_FLAG_ORDERSNAP | 712 CEPH_OSD_FLAG_ONDISK | 713 CEPH_OSD_FLAG_WRITE | 714 CEPH_OSD_FLAG_ACK; 715 716 while ((len = iov_iter_count(from)) > 0) { 717 size_t left; 718 int n; 719 720 snapc = ci->i_snap_realm->cached_context; 721 vino = ceph_vino(inode); 722 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 723 vino, pos, &len, 0, 1, 724 CEPH_OSD_OP_WRITE, flags, snapc, 725 ci->i_truncate_seq, 726 ci->i_truncate_size, 727 false); 728 if (IS_ERR(req)) { 729 ret = PTR_ERR(req); 730 break; 731 } 732 733 /* 734 * write from beginning of first page, 735 * regardless of io alignment 736 */ 737 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; 738 739 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 740 if (IS_ERR(pages)) { 741 ret = PTR_ERR(pages); 742 goto out; 743 } 744 745 left = len; 746 for (n = 0; n < num_pages; n++) { 747 size_t plen = min_t(size_t, left, PAGE_SIZE); 748 ret = copy_page_from_iter(pages[n], 0, plen, from); 749 if (ret != plen) { 750 ret = -EFAULT; 751 break; 752 } 753 left -= ret; 754 } 755 756 if (ret < 0) { 757 ceph_release_page_vector(pages, num_pages); 758 goto out; 759 } 760 761 /* get a second commit callback */ 762 req->r_unsafe_callback = ceph_sync_write_unsafe; 763 req->r_inode = inode; 764 765 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, 766 false, true); 767 768 /* BUG_ON(vino.snap != CEPH_NOSNAP); */ 769 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime); 770 771 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 772 if (!ret) 773 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 774 775 out: 776 ceph_osdc_put_request(req); 777 if (ret == 0) { 778 pos += len; 779 written += len; 780 781 if (pos > i_size_read(inode)) { 782 check_caps = ceph_inode_set_size(inode, pos); 783 if (check_caps) 784 ceph_check_caps(ceph_inode(inode), 785 CHECK_CAPS_AUTHONLY, 786 NULL); 787 } 788 } else 789 break; 790 } 791 792 if (ret != -EOLDSNAPC && written > 0) { 793 ret = written; 794 iocb->ki_pos = pos; 795 } 796 return ret; 797 } 798 799 /* 800 * Wrap generic_file_aio_read with checks for cap bits on the inode. 801 * Atomically grab references, so that those bits are not released 802 * back to the MDS mid-read. 803 * 804 * Hmm, the sync read case isn't actually async... should it be? 805 */ 806 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to) 807 { 808 struct file *filp = iocb->ki_filp; 809 struct ceph_file_info *fi = filp->private_data; 810 size_t len = iocb->ki_nbytes; 811 struct inode *inode = file_inode(filp); 812 struct ceph_inode_info *ci = ceph_inode(inode); 813 struct page *pinned_page = NULL; 814 ssize_t ret; 815 int want, got = 0; 816 int retry_op = 0, read = 0; 817 818 again: 819 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", 820 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode); 821 822 if (fi->fmode & CEPH_FILE_MODE_LAZY) 823 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; 824 else 825 want = CEPH_CAP_FILE_CACHE; 826 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page); 827 if (ret < 0) 828 return ret; 829 830 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 || 831 (iocb->ki_filp->f_flags & O_DIRECT) || 832 (fi->flags & CEPH_F_SYNC)) { 833 834 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n", 835 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, 836 ceph_cap_string(got)); 837 838 if (ci->i_inline_version == CEPH_INLINE_NONE) { 839 /* hmm, this isn't really async... */ 840 ret = ceph_sync_read(iocb, to, &retry_op); 841 } else { 842 retry_op = READ_INLINE; 843 } 844 } else { 845 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", 846 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, 847 ceph_cap_string(got)); 848 849 ret = generic_file_read_iter(iocb, to); 850 } 851 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", 852 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); 853 if (pinned_page) { 854 page_cache_release(pinned_page); 855 pinned_page = NULL; 856 } 857 ceph_put_cap_refs(ci, got); 858 if (retry_op && ret >= 0) { 859 int statret; 860 struct page *page = NULL; 861 loff_t i_size; 862 if (retry_op == READ_INLINE) { 863 page = __page_cache_alloc(GFP_NOFS); 864 if (!page) 865 return -ENOMEM; 866 } 867 868 statret = __ceph_do_getattr(inode, page, 869 CEPH_STAT_CAP_INLINE_DATA, !!page); 870 if (statret < 0) { 871 __free_page(page); 872 if (statret == -ENODATA) { 873 BUG_ON(retry_op != READ_INLINE); 874 goto again; 875 } 876 return statret; 877 } 878 879 i_size = i_size_read(inode); 880 if (retry_op == READ_INLINE) { 881 /* does not support inline data > PAGE_SIZE */ 882 if (i_size > PAGE_CACHE_SIZE) { 883 ret = -EIO; 884 } else if (iocb->ki_pos < i_size) { 885 loff_t end = min_t(loff_t, i_size, 886 iocb->ki_pos + len); 887 if (statret < end) 888 zero_user_segment(page, statret, end); 889 ret = copy_page_to_iter(page, 890 iocb->ki_pos & ~PAGE_MASK, 891 end - iocb->ki_pos, to); 892 iocb->ki_pos += ret; 893 } else { 894 ret = 0; 895 } 896 __free_pages(page, 0); 897 return ret; 898 } 899 900 /* hit EOF or hole? */ 901 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size && 902 ret < len) { 903 dout("sync_read hit hole, ppos %lld < size %lld" 904 ", reading more\n", iocb->ki_pos, 905 inode->i_size); 906 907 read += ret; 908 len -= ret; 909 retry_op = 0; 910 goto again; 911 } 912 } 913 914 if (ret >= 0) 915 ret += read; 916 917 return ret; 918 } 919 920 /* 921 * Take cap references to avoid releasing caps to MDS mid-write. 922 * 923 * If we are synchronous, and write with an old snap context, the OSD 924 * may return EOLDSNAPC. In that case, retry the write.. _after_ 925 * dropping our cap refs and allowing the pending snap to logically 926 * complete _before_ this write occurs. 927 * 928 * If we are near ENOSPC, write synchronously. 929 */ 930 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from) 931 { 932 struct file *file = iocb->ki_filp; 933 struct ceph_file_info *fi = file->private_data; 934 struct inode *inode = file_inode(file); 935 struct ceph_inode_info *ci = ceph_inode(inode); 936 struct ceph_osd_client *osdc = 937 &ceph_sb_to_client(inode->i_sb)->client->osdc; 938 ssize_t count = iov_iter_count(from), written = 0; 939 int err, want, got; 940 loff_t pos = iocb->ki_pos; 941 942 if (ceph_snap(inode) != CEPH_NOSNAP) 943 return -EROFS; 944 945 mutex_lock(&inode->i_mutex); 946 947 /* We can write back this queue in page reclaim */ 948 current->backing_dev_info = file->f_mapping->backing_dev_info; 949 950 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode)); 951 if (err) 952 goto out; 953 954 if (count == 0) 955 goto out; 956 iov_iter_truncate(from, count); 957 958 err = file_remove_suid(file); 959 if (err) 960 goto out; 961 962 err = file_update_time(file); 963 if (err) 964 goto out; 965 966 if (ci->i_inline_version != CEPH_INLINE_NONE) { 967 err = ceph_uninline_data(file, NULL); 968 if (err < 0) 969 goto out; 970 } 971 972 retry_snap: 973 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) { 974 err = -ENOSPC; 975 goto out; 976 } 977 978 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n", 979 inode, ceph_vinop(inode), pos, count, inode->i_size); 980 if (fi->fmode & CEPH_FILE_MODE_LAZY) 981 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 982 else 983 want = CEPH_CAP_FILE_BUFFER; 984 got = 0; 985 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count, 986 &got, NULL); 987 if (err < 0) 988 goto out; 989 990 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n", 991 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got)); 992 993 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || 994 (file->f_flags & O_DIRECT) || (fi->flags & CEPH_F_SYNC)) { 995 struct iov_iter data; 996 mutex_unlock(&inode->i_mutex); 997 /* we might need to revert back to that point */ 998 data = *from; 999 if (file->f_flags & O_DIRECT) 1000 written = ceph_sync_direct_write(iocb, &data, pos); 1001 else 1002 written = ceph_sync_write(iocb, &data, pos); 1003 if (written == -EOLDSNAPC) { 1004 dout("aio_write %p %llx.%llx %llu~%u" 1005 "got EOLDSNAPC, retrying\n", 1006 inode, ceph_vinop(inode), 1007 pos, (unsigned)count); 1008 mutex_lock(&inode->i_mutex); 1009 goto retry_snap; 1010 } 1011 if (written > 0) 1012 iov_iter_advance(from, written); 1013 } else { 1014 loff_t old_size = inode->i_size; 1015 /* 1016 * No need to acquire the i_truncate_mutex. Because 1017 * the MDS revokes Fwb caps before sending truncate 1018 * message to us. We can't get Fwb cap while there 1019 * are pending vmtruncate. So write and vmtruncate 1020 * can not run at the same time 1021 */ 1022 written = generic_perform_write(file, from, pos); 1023 if (likely(written >= 0)) 1024 iocb->ki_pos = pos + written; 1025 if (inode->i_size > old_size) 1026 ceph_fscache_update_objectsize(inode); 1027 mutex_unlock(&inode->i_mutex); 1028 } 1029 1030 if (written >= 0) { 1031 int dirty; 1032 spin_lock(&ci->i_ceph_lock); 1033 ci->i_inline_version = CEPH_INLINE_NONE; 1034 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1035 spin_unlock(&ci->i_ceph_lock); 1036 if (dirty) 1037 __mark_inode_dirty(inode, dirty); 1038 } 1039 1040 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", 1041 inode, ceph_vinop(inode), pos, (unsigned)count, 1042 ceph_cap_string(got)); 1043 ceph_put_cap_refs(ci, got); 1044 1045 if (written >= 0 && 1046 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) || 1047 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { 1048 err = vfs_fsync_range(file, pos, pos + written - 1, 1); 1049 if (err < 0) 1050 written = err; 1051 } 1052 1053 goto out_unlocked; 1054 1055 out: 1056 mutex_unlock(&inode->i_mutex); 1057 out_unlocked: 1058 current->backing_dev_info = NULL; 1059 return written ? written : err; 1060 } 1061 1062 /* 1063 * llseek. be sure to verify file size on SEEK_END. 1064 */ 1065 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence) 1066 { 1067 struct inode *inode = file->f_mapping->host; 1068 int ret; 1069 1070 mutex_lock(&inode->i_mutex); 1071 1072 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) { 1073 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false); 1074 if (ret < 0) { 1075 offset = ret; 1076 goto out; 1077 } 1078 } 1079 1080 switch (whence) { 1081 case SEEK_END: 1082 offset += inode->i_size; 1083 break; 1084 case SEEK_CUR: 1085 /* 1086 * Here we special-case the lseek(fd, 0, SEEK_CUR) 1087 * position-querying operation. Avoid rewriting the "same" 1088 * f_pos value back to the file because a concurrent read(), 1089 * write() or lseek() might have altered it 1090 */ 1091 if (offset == 0) { 1092 offset = file->f_pos; 1093 goto out; 1094 } 1095 offset += file->f_pos; 1096 break; 1097 case SEEK_DATA: 1098 if (offset >= inode->i_size) { 1099 ret = -ENXIO; 1100 goto out; 1101 } 1102 break; 1103 case SEEK_HOLE: 1104 if (offset >= inode->i_size) { 1105 ret = -ENXIO; 1106 goto out; 1107 } 1108 offset = inode->i_size; 1109 break; 1110 } 1111 1112 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes); 1113 1114 out: 1115 mutex_unlock(&inode->i_mutex); 1116 return offset; 1117 } 1118 1119 static inline void ceph_zero_partial_page( 1120 struct inode *inode, loff_t offset, unsigned size) 1121 { 1122 struct page *page; 1123 pgoff_t index = offset >> PAGE_CACHE_SHIFT; 1124 1125 page = find_lock_page(inode->i_mapping, index); 1126 if (page) { 1127 wait_on_page_writeback(page); 1128 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size); 1129 unlock_page(page); 1130 page_cache_release(page); 1131 } 1132 } 1133 1134 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset, 1135 loff_t length) 1136 { 1137 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE); 1138 if (offset < nearly) { 1139 loff_t size = nearly - offset; 1140 if (length < size) 1141 size = length; 1142 ceph_zero_partial_page(inode, offset, size); 1143 offset += size; 1144 length -= size; 1145 } 1146 if (length >= PAGE_CACHE_SIZE) { 1147 loff_t size = round_down(length, PAGE_CACHE_SIZE); 1148 truncate_pagecache_range(inode, offset, offset + size - 1); 1149 offset += size; 1150 length -= size; 1151 } 1152 if (length) 1153 ceph_zero_partial_page(inode, offset, length); 1154 } 1155 1156 static int ceph_zero_partial_object(struct inode *inode, 1157 loff_t offset, loff_t *length) 1158 { 1159 struct ceph_inode_info *ci = ceph_inode(inode); 1160 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 1161 struct ceph_osd_request *req; 1162 int ret = 0; 1163 loff_t zero = 0; 1164 int op; 1165 1166 if (!length) { 1167 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE; 1168 length = &zero; 1169 } else { 1170 op = CEPH_OSD_OP_ZERO; 1171 } 1172 1173 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 1174 ceph_vino(inode), 1175 offset, length, 1176 0, 1, op, 1177 CEPH_OSD_FLAG_WRITE | 1178 CEPH_OSD_FLAG_ONDISK, 1179 NULL, 0, 0, false); 1180 if (IS_ERR(req)) { 1181 ret = PTR_ERR(req); 1182 goto out; 1183 } 1184 1185 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap, 1186 &inode->i_mtime); 1187 1188 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); 1189 if (!ret) { 1190 ret = ceph_osdc_wait_request(&fsc->client->osdc, req); 1191 if (ret == -ENOENT) 1192 ret = 0; 1193 } 1194 ceph_osdc_put_request(req); 1195 1196 out: 1197 return ret; 1198 } 1199 1200 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length) 1201 { 1202 int ret = 0; 1203 struct ceph_inode_info *ci = ceph_inode(inode); 1204 s32 stripe_unit = ceph_file_layout_su(ci->i_layout); 1205 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout); 1206 s32 object_size = ceph_file_layout_object_size(ci->i_layout); 1207 u64 object_set_size = object_size * stripe_count; 1208 u64 nearly, t; 1209 1210 /* round offset up to next period boundary */ 1211 nearly = offset + object_set_size - 1; 1212 t = nearly; 1213 nearly -= do_div(t, object_set_size); 1214 1215 while (length && offset < nearly) { 1216 loff_t size = length; 1217 ret = ceph_zero_partial_object(inode, offset, &size); 1218 if (ret < 0) 1219 return ret; 1220 offset += size; 1221 length -= size; 1222 } 1223 while (length >= object_set_size) { 1224 int i; 1225 loff_t pos = offset; 1226 for (i = 0; i < stripe_count; ++i) { 1227 ret = ceph_zero_partial_object(inode, pos, NULL); 1228 if (ret < 0) 1229 return ret; 1230 pos += stripe_unit; 1231 } 1232 offset += object_set_size; 1233 length -= object_set_size; 1234 } 1235 while (length) { 1236 loff_t size = length; 1237 ret = ceph_zero_partial_object(inode, offset, &size); 1238 if (ret < 0) 1239 return ret; 1240 offset += size; 1241 length -= size; 1242 } 1243 return ret; 1244 } 1245 1246 static long ceph_fallocate(struct file *file, int mode, 1247 loff_t offset, loff_t length) 1248 { 1249 struct ceph_file_info *fi = file->private_data; 1250 struct inode *inode = file_inode(file); 1251 struct ceph_inode_info *ci = ceph_inode(inode); 1252 struct ceph_osd_client *osdc = 1253 &ceph_inode_to_client(inode)->client->osdc; 1254 int want, got = 0; 1255 int dirty; 1256 int ret = 0; 1257 loff_t endoff = 0; 1258 loff_t size; 1259 1260 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE)) 1261 return -EOPNOTSUPP; 1262 1263 if (!S_ISREG(inode->i_mode)) 1264 return -EOPNOTSUPP; 1265 1266 mutex_lock(&inode->i_mutex); 1267 1268 if (ceph_snap(inode) != CEPH_NOSNAP) { 1269 ret = -EROFS; 1270 goto unlock; 1271 } 1272 1273 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) && 1274 !(mode & FALLOC_FL_PUNCH_HOLE)) { 1275 ret = -ENOSPC; 1276 goto unlock; 1277 } 1278 1279 if (ci->i_inline_version != CEPH_INLINE_NONE) { 1280 ret = ceph_uninline_data(file, NULL); 1281 if (ret < 0) 1282 goto unlock; 1283 } 1284 1285 size = i_size_read(inode); 1286 if (!(mode & FALLOC_FL_KEEP_SIZE)) 1287 endoff = offset + length; 1288 1289 if (fi->fmode & CEPH_FILE_MODE_LAZY) 1290 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; 1291 else 1292 want = CEPH_CAP_FILE_BUFFER; 1293 1294 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL); 1295 if (ret < 0) 1296 goto unlock; 1297 1298 if (mode & FALLOC_FL_PUNCH_HOLE) { 1299 if (offset < size) 1300 ceph_zero_pagecache_range(inode, offset, length); 1301 ret = ceph_zero_objects(inode, offset, length); 1302 } else if (endoff > size) { 1303 truncate_pagecache_range(inode, size, -1); 1304 if (ceph_inode_set_size(inode, endoff)) 1305 ceph_check_caps(ceph_inode(inode), 1306 CHECK_CAPS_AUTHONLY, NULL); 1307 } 1308 1309 if (!ret) { 1310 spin_lock(&ci->i_ceph_lock); 1311 ci->i_inline_version = CEPH_INLINE_NONE; 1312 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1313 spin_unlock(&ci->i_ceph_lock); 1314 if (dirty) 1315 __mark_inode_dirty(inode, dirty); 1316 } 1317 1318 ceph_put_cap_refs(ci, got); 1319 unlock: 1320 mutex_unlock(&inode->i_mutex); 1321 return ret; 1322 } 1323 1324 const struct file_operations ceph_file_fops = { 1325 .open = ceph_open, 1326 .release = ceph_release, 1327 .llseek = ceph_llseek, 1328 .read = new_sync_read, 1329 .write = new_sync_write, 1330 .read_iter = ceph_read_iter, 1331 .write_iter = ceph_write_iter, 1332 .mmap = ceph_mmap, 1333 .fsync = ceph_fsync, 1334 .lock = ceph_lock, 1335 .flock = ceph_flock, 1336 .splice_read = generic_file_splice_read, 1337 .splice_write = iter_file_splice_write, 1338 .unlocked_ioctl = ceph_ioctl, 1339 .compat_ioctl = ceph_ioctl, 1340 .fallocate = ceph_fallocate, 1341 }; 1342 1343