1 /* 2 * linux/fs/nfs/write.c 3 * 4 * Write file data over NFS. 5 * 6 * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de> 7 */ 8 9 #include <linux/types.h> 10 #include <linux/slab.h> 11 #include <linux/mm.h> 12 #include <linux/pagemap.h> 13 #include <linux/file.h> 14 #include <linux/writeback.h> 15 #include <linux/swap.h> 16 #include <linux/migrate.h> 17 18 #include <linux/sunrpc/clnt.h> 19 #include <linux/nfs_fs.h> 20 #include <linux/nfs_mount.h> 21 #include <linux/nfs_page.h> 22 #include <linux/backing-dev.h> 23 #include <linux/export.h> 24 25 #include <asm/uaccess.h> 26 27 #include "delegation.h" 28 #include "internal.h" 29 #include "iostat.h" 30 #include "nfs4_fs.h" 31 #include "fscache.h" 32 #include "pnfs.h" 33 34 #include "nfstrace.h" 35 36 #define NFSDBG_FACILITY NFSDBG_PAGECACHE 37 38 #define MIN_POOL_WRITE (32) 39 #define MIN_POOL_COMMIT (4) 40 41 /* 42 * Local function declarations 43 */ 44 static void nfs_redirty_request(struct nfs_page *req); 45 static const struct rpc_call_ops nfs_commit_ops; 46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops; 47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops; 48 static const struct nfs_rw_ops nfs_rw_write_ops; 49 static void nfs_clear_request_commit(struct nfs_page *req); 50 51 static struct kmem_cache *nfs_wdata_cachep; 52 static mempool_t *nfs_wdata_mempool; 53 static struct kmem_cache *nfs_cdata_cachep; 54 static mempool_t *nfs_commit_mempool; 55 56 struct nfs_commit_data *nfs_commitdata_alloc(void) 57 { 58 struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO); 59 60 if (p) { 61 memset(p, 0, sizeof(*p)); 62 INIT_LIST_HEAD(&p->pages); 63 } 64 return p; 65 } 66 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc); 67 68 void nfs_commit_free(struct nfs_commit_data *p) 69 { 70 mempool_free(p, nfs_commit_mempool); 71 } 72 EXPORT_SYMBOL_GPL(nfs_commit_free); 73 74 static struct nfs_rw_header *nfs_writehdr_alloc(void) 75 { 76 struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO); 77 78 if (p) 79 memset(p, 0, sizeof(*p)); 80 return p; 81 } 82 83 static void nfs_writehdr_free(struct nfs_rw_header *whdr) 84 { 85 mempool_free(whdr, nfs_wdata_mempool); 86 } 87 88 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) 89 { 90 ctx->error = error; 91 smp_wmb(); 92 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); 93 } 94 95 /* 96 * nfs_page_find_head_request_locked - find head request associated with @page 97 * 98 * must be called while holding the inode lock. 99 * 100 * returns matching head request with reference held, or NULL if not found. 101 */ 102 static struct nfs_page * 103 nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page) 104 { 105 struct nfs_page *req = NULL; 106 107 if (PagePrivate(page)) 108 req = (struct nfs_page *)page_private(page); 109 else if (unlikely(PageSwapCache(page))) { 110 struct nfs_page *freq, *t; 111 112 /* Linearly search the commit list for the correct req */ 113 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) { 114 if (freq->wb_page == page) { 115 req = freq->wb_head; 116 break; 117 } 118 } 119 } 120 121 if (req) { 122 WARN_ON_ONCE(req->wb_head != req); 123 124 kref_get(&req->wb_kref); 125 } 126 127 return req; 128 } 129 130 /* 131 * nfs_page_find_head_request - find head request associated with @page 132 * 133 * returns matching head request with reference held, or NULL if not found. 134 */ 135 static struct nfs_page *nfs_page_find_head_request(struct page *page) 136 { 137 struct inode *inode = page_file_mapping(page)->host; 138 struct nfs_page *req = NULL; 139 140 spin_lock(&inode->i_lock); 141 req = nfs_page_find_head_request_locked(NFS_I(inode), page); 142 spin_unlock(&inode->i_lock); 143 return req; 144 } 145 146 /* Adjust the file length if we're writing beyond the end */ 147 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) 148 { 149 struct inode *inode = page_file_mapping(page)->host; 150 loff_t end, i_size; 151 pgoff_t end_index; 152 153 spin_lock(&inode->i_lock); 154 i_size = i_size_read(inode); 155 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; 156 if (i_size > 0 && page_file_index(page) < end_index) 157 goto out; 158 end = page_file_offset(page) + ((loff_t)offset+count); 159 if (i_size >= end) 160 goto out; 161 i_size_write(inode, end); 162 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); 163 out: 164 spin_unlock(&inode->i_lock); 165 } 166 167 /* A writeback failed: mark the page as bad, and invalidate the page cache */ 168 static void nfs_set_pageerror(struct page *page) 169 { 170 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page)); 171 } 172 173 /* 174 * nfs_page_group_search_locked 175 * @head - head request of page group 176 * @page_offset - offset into page 177 * 178 * Search page group with head @head to find a request that contains the 179 * page offset @page_offset. 180 * 181 * Returns a pointer to the first matching nfs request, or NULL if no 182 * match is found. 183 * 184 * Must be called with the page group lock held 185 */ 186 static struct nfs_page * 187 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset) 188 { 189 struct nfs_page *req; 190 191 WARN_ON_ONCE(head != head->wb_head); 192 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags)); 193 194 req = head; 195 do { 196 if (page_offset >= req->wb_pgbase && 197 page_offset < (req->wb_pgbase + req->wb_bytes)) 198 return req; 199 200 req = req->wb_this_page; 201 } while (req != head); 202 203 return NULL; 204 } 205 206 /* 207 * nfs_page_group_covers_page 208 * @head - head request of page group 209 * 210 * Return true if the page group with head @head covers the whole page, 211 * returns false otherwise 212 */ 213 static bool nfs_page_group_covers_page(struct nfs_page *req) 214 { 215 struct nfs_page *tmp; 216 unsigned int pos = 0; 217 unsigned int len = nfs_page_length(req->wb_page); 218 219 nfs_page_group_lock(req); 220 221 do { 222 tmp = nfs_page_group_search_locked(req->wb_head, pos); 223 if (tmp) { 224 /* no way this should happen */ 225 WARN_ON_ONCE(tmp->wb_pgbase != pos); 226 pos += tmp->wb_bytes - (pos - tmp->wb_pgbase); 227 } 228 } while (tmp && pos < len); 229 230 nfs_page_group_unlock(req); 231 WARN_ON_ONCE(pos > len); 232 return pos == len; 233 } 234 235 /* We can set the PG_uptodate flag if we see that a write request 236 * covers the full page. 237 */ 238 static void nfs_mark_uptodate(struct nfs_page *req) 239 { 240 if (PageUptodate(req->wb_page)) 241 return; 242 if (!nfs_page_group_covers_page(req)) 243 return; 244 SetPageUptodate(req->wb_page); 245 } 246 247 static int wb_priority(struct writeback_control *wbc) 248 { 249 if (wbc->for_reclaim) 250 return FLUSH_HIGHPRI | FLUSH_STABLE; 251 if (wbc->for_kupdate || wbc->for_background) 252 return FLUSH_LOWPRI | FLUSH_COND_STABLE; 253 return FLUSH_COND_STABLE; 254 } 255 256 /* 257 * NFS congestion control 258 */ 259 260 int nfs_congestion_kb; 261 262 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) 263 #define NFS_CONGESTION_OFF_THRESH \ 264 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) 265 266 static void nfs_set_page_writeback(struct page *page) 267 { 268 struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host); 269 int ret = test_set_page_writeback(page); 270 271 WARN_ON_ONCE(ret != 0); 272 273 if (atomic_long_inc_return(&nfss->writeback) > 274 NFS_CONGESTION_ON_THRESH) { 275 set_bdi_congested(&nfss->backing_dev_info, 276 BLK_RW_ASYNC); 277 } 278 } 279 280 static void nfs_end_page_writeback(struct nfs_page *req) 281 { 282 struct inode *inode = page_file_mapping(req->wb_page)->host; 283 struct nfs_server *nfss = NFS_SERVER(inode); 284 285 if (!nfs_page_group_sync_on_bit(req, PG_WB_END)) 286 return; 287 288 end_page_writeback(req->wb_page); 289 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) 290 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC); 291 } 292 293 294 /* nfs_page_group_clear_bits 295 * @req - an nfs request 296 * clears all page group related bits from @req 297 */ 298 static void 299 nfs_page_group_clear_bits(struct nfs_page *req) 300 { 301 clear_bit(PG_TEARDOWN, &req->wb_flags); 302 clear_bit(PG_UNLOCKPAGE, &req->wb_flags); 303 clear_bit(PG_UPTODATE, &req->wb_flags); 304 clear_bit(PG_WB_END, &req->wb_flags); 305 clear_bit(PG_REMOVE, &req->wb_flags); 306 } 307 308 309 /* 310 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req 311 * 312 * this is a helper function for nfs_lock_and_join_requests 313 * 314 * @inode - inode associated with request page group, must be holding inode lock 315 * @head - head request of page group, must be holding head lock 316 * @req - request that couldn't lock and needs to wait on the req bit lock 317 * @nonblock - if true, don't actually wait 318 * 319 * NOTE: this must be called holding page_group bit lock and inode spin lock 320 * and BOTH will be released before returning. 321 * 322 * returns 0 on success, < 0 on error. 323 */ 324 static int 325 nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head, 326 struct nfs_page *req, bool nonblock) 327 __releases(&inode->i_lock) 328 { 329 struct nfs_page *tmp; 330 int ret; 331 332 /* relinquish all the locks successfully grabbed this run */ 333 for (tmp = head ; tmp != req; tmp = tmp->wb_this_page) 334 nfs_unlock_request(tmp); 335 336 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags)); 337 338 /* grab a ref on the request that will be waited on */ 339 kref_get(&req->wb_kref); 340 341 nfs_page_group_unlock(head); 342 spin_unlock(&inode->i_lock); 343 344 /* release ref from nfs_page_find_head_request_locked */ 345 nfs_release_request(head); 346 347 if (!nonblock) 348 ret = nfs_wait_on_request(req); 349 else 350 ret = -EAGAIN; 351 nfs_release_request(req); 352 353 return ret; 354 } 355 356 /* 357 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests 358 * 359 * @destroy_list - request list (using wb_this_page) terminated by @old_head 360 * @old_head - the old head of the list 361 * 362 * All subrequests must be locked and removed from all lists, so at this point 363 * they are only "active" in this function, and possibly in nfs_wait_on_request 364 * with a reference held by some other context. 365 */ 366 static void 367 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list, 368 struct nfs_page *old_head) 369 { 370 while (destroy_list) { 371 struct nfs_page *subreq = destroy_list; 372 373 destroy_list = (subreq->wb_this_page == old_head) ? 374 NULL : subreq->wb_this_page; 375 376 WARN_ON_ONCE(old_head != subreq->wb_head); 377 378 /* make sure old group is not used */ 379 subreq->wb_head = subreq; 380 subreq->wb_this_page = subreq; 381 382 nfs_clear_request_commit(subreq); 383 384 /* subreq is now totally disconnected from page group or any 385 * write / commit lists. last chance to wake any waiters */ 386 nfs_unlock_request(subreq); 387 388 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) { 389 /* release ref on old head request */ 390 nfs_release_request(old_head); 391 392 nfs_page_group_clear_bits(subreq); 393 394 /* release the PG_INODE_REF reference */ 395 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) 396 nfs_release_request(subreq); 397 else 398 WARN_ON_ONCE(1); 399 } else { 400 WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags)); 401 /* zombie requests have already released the last 402 * reference and were waiting on the rest of the 403 * group to complete. Since it's no longer part of a 404 * group, simply free the request */ 405 nfs_page_group_clear_bits(subreq); 406 nfs_free_request(subreq); 407 } 408 } 409 } 410 411 /* 412 * nfs_lock_and_join_requests - join all subreqs to the head req and return 413 * a locked reference, cancelling any pending 414 * operations for this page. 415 * 416 * @page - the page used to lookup the "page group" of nfs_page structures 417 * @nonblock - if true, don't block waiting for request locks 418 * 419 * This function joins all sub requests to the head request by first 420 * locking all requests in the group, cancelling any pending operations 421 * and finally updating the head request to cover the whole range covered by 422 * the (former) group. All subrequests are removed from any write or commit 423 * lists, unlinked from the group and destroyed. 424 * 425 * Returns a locked, referenced pointer to the head request - which after 426 * this call is guaranteed to be the only request associated with the page. 427 * Returns NULL if no requests are found for @page, or a ERR_PTR if an 428 * error was encountered. 429 */ 430 static struct nfs_page * 431 nfs_lock_and_join_requests(struct page *page, bool nonblock) 432 { 433 struct inode *inode = page_file_mapping(page)->host; 434 struct nfs_page *head, *subreq; 435 struct nfs_page *destroy_list = NULL; 436 unsigned int total_bytes; 437 int ret; 438 439 try_again: 440 total_bytes = 0; 441 442 WARN_ON_ONCE(destroy_list); 443 444 spin_lock(&inode->i_lock); 445 446 /* 447 * A reference is taken only on the head request which acts as a 448 * reference to the whole page group - the group will not be destroyed 449 * until the head reference is released. 450 */ 451 head = nfs_page_find_head_request_locked(NFS_I(inode), page); 452 453 if (!head) { 454 spin_unlock(&inode->i_lock); 455 return NULL; 456 } 457 458 /* lock each request in the page group */ 459 nfs_page_group_lock(head); 460 subreq = head; 461 do { 462 /* 463 * Subrequests are always contiguous, non overlapping 464 * and in order. If not, it's a programming error. 465 */ 466 WARN_ON_ONCE(subreq->wb_offset != 467 (head->wb_offset + total_bytes)); 468 469 /* keep track of how many bytes this group covers */ 470 total_bytes += subreq->wb_bytes; 471 472 if (!nfs_lock_request(subreq)) { 473 /* releases page group bit lock and 474 * inode spin lock and all references */ 475 ret = nfs_unroll_locks_and_wait(inode, head, 476 subreq, nonblock); 477 478 if (ret == 0) 479 goto try_again; 480 481 return ERR_PTR(ret); 482 } 483 484 subreq = subreq->wb_this_page; 485 } while (subreq != head); 486 487 /* Now that all requests are locked, make sure they aren't on any list. 488 * Commit list removal accounting is done after locks are dropped */ 489 subreq = head; 490 do { 491 nfs_list_remove_request(subreq); 492 subreq = subreq->wb_this_page; 493 } while (subreq != head); 494 495 /* unlink subrequests from head, destroy them later */ 496 if (head->wb_this_page != head) { 497 /* destroy list will be terminated by head */ 498 destroy_list = head->wb_this_page; 499 head->wb_this_page = head; 500 501 /* change head request to cover whole range that 502 * the former page group covered */ 503 head->wb_bytes = total_bytes; 504 } 505 506 /* 507 * prepare head request to be added to new pgio descriptor 508 */ 509 nfs_page_group_clear_bits(head); 510 511 /* 512 * some part of the group was still on the inode list - otherwise 513 * the group wouldn't be involved in async write. 514 * grab a reference for the head request, iff it needs one. 515 */ 516 if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags)) 517 kref_get(&head->wb_kref); 518 519 nfs_page_group_unlock(head); 520 521 /* drop lock to clear_request_commit the head req and clean up 522 * requests on destroy list */ 523 spin_unlock(&inode->i_lock); 524 525 nfs_destroy_unlinked_subrequests(destroy_list, head); 526 527 /* clean up commit list state */ 528 nfs_clear_request_commit(head); 529 530 /* still holds ref on head from nfs_page_find_head_request_locked 531 * and still has lock on head from lock loop */ 532 return head; 533 } 534 535 /* 536 * Find an associated nfs write request, and prepare to flush it out 537 * May return an error if the user signalled nfs_wait_on_request(). 538 */ 539 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio, 540 struct page *page, bool nonblock) 541 { 542 struct nfs_page *req; 543 int ret = 0; 544 545 req = nfs_lock_and_join_requests(page, nonblock); 546 if (!req) 547 goto out; 548 ret = PTR_ERR(req); 549 if (IS_ERR(req)) 550 goto out; 551 552 nfs_set_page_writeback(page); 553 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags)); 554 555 ret = 0; 556 if (!nfs_pageio_add_request(pgio, req)) { 557 nfs_redirty_request(req); 558 ret = pgio->pg_error; 559 } 560 out: 561 return ret; 562 } 563 564 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio) 565 { 566 struct inode *inode = page_file_mapping(page)->host; 567 int ret; 568 569 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); 570 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1); 571 572 nfs_pageio_cond_complete(pgio, page_file_index(page)); 573 ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE); 574 if (ret == -EAGAIN) { 575 redirty_page_for_writepage(wbc, page); 576 ret = 0; 577 } 578 return ret; 579 } 580 581 /* 582 * Write an mmapped page to the server. 583 */ 584 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc) 585 { 586 struct nfs_pageio_descriptor pgio; 587 int err; 588 589 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc), 590 false, &nfs_async_write_completion_ops); 591 err = nfs_do_writepage(page, wbc, &pgio); 592 nfs_pageio_complete(&pgio); 593 if (err < 0) 594 return err; 595 if (pgio.pg_error < 0) 596 return pgio.pg_error; 597 return 0; 598 } 599 600 int nfs_writepage(struct page *page, struct writeback_control *wbc) 601 { 602 int ret; 603 604 ret = nfs_writepage_locked(page, wbc); 605 unlock_page(page); 606 return ret; 607 } 608 609 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data) 610 { 611 int ret; 612 613 ret = nfs_do_writepage(page, wbc, data); 614 unlock_page(page); 615 return ret; 616 } 617 618 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) 619 { 620 struct inode *inode = mapping->host; 621 unsigned long *bitlock = &NFS_I(inode)->flags; 622 struct nfs_pageio_descriptor pgio; 623 int err; 624 625 /* Stop dirtying of new pages while we sync */ 626 err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING, 627 nfs_wait_bit_killable, TASK_KILLABLE); 628 if (err) 629 goto out_err; 630 631 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); 632 633 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false, 634 &nfs_async_write_completion_ops); 635 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio); 636 nfs_pageio_complete(&pgio); 637 638 clear_bit_unlock(NFS_INO_FLUSHING, bitlock); 639 smp_mb__after_atomic(); 640 wake_up_bit(bitlock, NFS_INO_FLUSHING); 641 642 if (err < 0) 643 goto out_err; 644 err = pgio.pg_error; 645 if (err < 0) 646 goto out_err; 647 return 0; 648 out_err: 649 return err; 650 } 651 652 /* 653 * Insert a write request into an inode 654 */ 655 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req) 656 { 657 struct nfs_inode *nfsi = NFS_I(inode); 658 659 WARN_ON_ONCE(req->wb_this_page != req); 660 661 /* Lock the request! */ 662 nfs_lock_request(req); 663 664 spin_lock(&inode->i_lock); 665 if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) 666 inode->i_version++; 667 /* 668 * Swap-space should not get truncated. Hence no need to plug the race 669 * with invalidate/truncate. 670 */ 671 if (likely(!PageSwapCache(req->wb_page))) { 672 set_bit(PG_MAPPED, &req->wb_flags); 673 SetPagePrivate(req->wb_page); 674 set_page_private(req->wb_page, (unsigned long)req); 675 } 676 nfsi->npages++; 677 /* this a head request for a page group - mark it as having an 678 * extra reference so sub groups can follow suit */ 679 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags)); 680 kref_get(&req->wb_kref); 681 spin_unlock(&inode->i_lock); 682 } 683 684 /* 685 * Remove a write request from an inode 686 */ 687 static void nfs_inode_remove_request(struct nfs_page *req) 688 { 689 struct inode *inode = req->wb_context->dentry->d_inode; 690 struct nfs_inode *nfsi = NFS_I(inode); 691 struct nfs_page *head; 692 693 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) { 694 head = req->wb_head; 695 696 spin_lock(&inode->i_lock); 697 if (likely(!PageSwapCache(head->wb_page))) { 698 set_page_private(head->wb_page, 0); 699 ClearPagePrivate(head->wb_page); 700 clear_bit(PG_MAPPED, &head->wb_flags); 701 } 702 nfsi->npages--; 703 spin_unlock(&inode->i_lock); 704 } 705 706 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) 707 nfs_release_request(req); 708 } 709 710 static void 711 nfs_mark_request_dirty(struct nfs_page *req) 712 { 713 __set_page_dirty_nobuffers(req->wb_page); 714 } 715 716 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) 717 /** 718 * nfs_request_add_commit_list - add request to a commit list 719 * @req: pointer to a struct nfs_page 720 * @dst: commit list head 721 * @cinfo: holds list lock and accounting info 722 * 723 * This sets the PG_CLEAN bit, updates the cinfo count of 724 * number of outstanding requests requiring a commit as well as 725 * the MM page stats. 726 * 727 * The caller must _not_ hold the cinfo->lock, but must be 728 * holding the nfs_page lock. 729 */ 730 void 731 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst, 732 struct nfs_commit_info *cinfo) 733 { 734 set_bit(PG_CLEAN, &(req)->wb_flags); 735 spin_lock(cinfo->lock); 736 nfs_list_add_request(req, dst); 737 cinfo->mds->ncommit++; 738 spin_unlock(cinfo->lock); 739 if (!cinfo->dreq) { 740 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 741 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info, 742 BDI_RECLAIMABLE); 743 __mark_inode_dirty(req->wb_context->dentry->d_inode, 744 I_DIRTY_DATASYNC); 745 } 746 } 747 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list); 748 749 /** 750 * nfs_request_remove_commit_list - Remove request from a commit list 751 * @req: pointer to a nfs_page 752 * @cinfo: holds list lock and accounting info 753 * 754 * This clears the PG_CLEAN bit, and updates the cinfo's count of 755 * number of outstanding requests requiring a commit 756 * It does not update the MM page stats. 757 * 758 * The caller _must_ hold the cinfo->lock and the nfs_page lock. 759 */ 760 void 761 nfs_request_remove_commit_list(struct nfs_page *req, 762 struct nfs_commit_info *cinfo) 763 { 764 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) 765 return; 766 nfs_list_remove_request(req); 767 cinfo->mds->ncommit--; 768 } 769 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list); 770 771 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, 772 struct inode *inode) 773 { 774 cinfo->lock = &inode->i_lock; 775 cinfo->mds = &NFS_I(inode)->commit_info; 776 cinfo->ds = pnfs_get_ds_info(inode); 777 cinfo->dreq = NULL; 778 cinfo->completion_ops = &nfs_commit_completion_ops; 779 } 780 781 void nfs_init_cinfo(struct nfs_commit_info *cinfo, 782 struct inode *inode, 783 struct nfs_direct_req *dreq) 784 { 785 if (dreq) 786 nfs_init_cinfo_from_dreq(cinfo, dreq); 787 else 788 nfs_init_cinfo_from_inode(cinfo, inode); 789 } 790 EXPORT_SYMBOL_GPL(nfs_init_cinfo); 791 792 /* 793 * Add a request to the inode's commit list. 794 */ 795 void 796 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, 797 struct nfs_commit_info *cinfo) 798 { 799 if (pnfs_mark_request_commit(req, lseg, cinfo)) 800 return; 801 nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo); 802 } 803 804 static void 805 nfs_clear_page_commit(struct page *page) 806 { 807 dec_zone_page_state(page, NR_UNSTABLE_NFS); 808 dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE); 809 } 810 811 static void 812 nfs_clear_request_commit(struct nfs_page *req) 813 { 814 if (test_bit(PG_CLEAN, &req->wb_flags)) { 815 struct inode *inode = req->wb_context->dentry->d_inode; 816 struct nfs_commit_info cinfo; 817 818 nfs_init_cinfo_from_inode(&cinfo, inode); 819 if (!pnfs_clear_request_commit(req, &cinfo)) { 820 spin_lock(cinfo.lock); 821 nfs_request_remove_commit_list(req, &cinfo); 822 spin_unlock(cinfo.lock); 823 } 824 nfs_clear_page_commit(req->wb_page); 825 } 826 } 827 828 static inline 829 int nfs_write_need_commit(struct nfs_pgio_data *data) 830 { 831 if (data->verf.committed == NFS_DATA_SYNC) 832 return data->header->lseg == NULL; 833 return data->verf.committed != NFS_FILE_SYNC; 834 } 835 836 #else 837 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, 838 struct inode *inode) 839 { 840 } 841 842 void nfs_init_cinfo(struct nfs_commit_info *cinfo, 843 struct inode *inode, 844 struct nfs_direct_req *dreq) 845 { 846 } 847 848 void 849 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, 850 struct nfs_commit_info *cinfo) 851 { 852 } 853 854 static void 855 nfs_clear_request_commit(struct nfs_page *req) 856 { 857 } 858 859 static inline 860 int nfs_write_need_commit(struct nfs_pgio_data *data) 861 { 862 return 0; 863 } 864 865 #endif 866 867 static void nfs_write_completion(struct nfs_pgio_header *hdr) 868 { 869 struct nfs_commit_info cinfo; 870 unsigned long bytes = 0; 871 bool do_destroy; 872 873 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) 874 goto out; 875 nfs_init_cinfo_from_inode(&cinfo, hdr->inode); 876 while (!list_empty(&hdr->pages)) { 877 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 878 879 bytes += req->wb_bytes; 880 nfs_list_remove_request(req); 881 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && 882 (hdr->good_bytes < bytes)) { 883 nfs_set_pageerror(req->wb_page); 884 nfs_context_set_write_error(req->wb_context, hdr->error); 885 goto remove_req; 886 } 887 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) { 888 nfs_mark_request_dirty(req); 889 goto next; 890 } 891 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) { 892 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf)); 893 nfs_mark_request_commit(req, hdr->lseg, &cinfo); 894 goto next; 895 } 896 remove_req: 897 nfs_inode_remove_request(req); 898 next: 899 nfs_unlock_request(req); 900 nfs_end_page_writeback(req); 901 do_destroy = !test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags); 902 nfs_release_request(req); 903 } 904 out: 905 hdr->release(hdr); 906 } 907 908 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) 909 unsigned long 910 nfs_reqs_to_commit(struct nfs_commit_info *cinfo) 911 { 912 return cinfo->mds->ncommit; 913 } 914 915 /* cinfo->lock held by caller */ 916 int 917 nfs_scan_commit_list(struct list_head *src, struct list_head *dst, 918 struct nfs_commit_info *cinfo, int max) 919 { 920 struct nfs_page *req, *tmp; 921 int ret = 0; 922 923 list_for_each_entry_safe(req, tmp, src, wb_list) { 924 if (!nfs_lock_request(req)) 925 continue; 926 kref_get(&req->wb_kref); 927 if (cond_resched_lock(cinfo->lock)) 928 list_safe_reset_next(req, tmp, wb_list); 929 nfs_request_remove_commit_list(req, cinfo); 930 nfs_list_add_request(req, dst); 931 ret++; 932 if ((ret == max) && !cinfo->dreq) 933 break; 934 } 935 return ret; 936 } 937 938 /* 939 * nfs_scan_commit - Scan an inode for commit requests 940 * @inode: NFS inode to scan 941 * @dst: mds destination list 942 * @cinfo: mds and ds lists of reqs ready to commit 943 * 944 * Moves requests from the inode's 'commit' request list. 945 * The requests are *not* checked to ensure that they form a contiguous set. 946 */ 947 int 948 nfs_scan_commit(struct inode *inode, struct list_head *dst, 949 struct nfs_commit_info *cinfo) 950 { 951 int ret = 0; 952 953 spin_lock(cinfo->lock); 954 if (cinfo->mds->ncommit > 0) { 955 const int max = INT_MAX; 956 957 ret = nfs_scan_commit_list(&cinfo->mds->list, dst, 958 cinfo, max); 959 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret); 960 } 961 spin_unlock(cinfo->lock); 962 return ret; 963 } 964 965 #else 966 unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo) 967 { 968 return 0; 969 } 970 971 int nfs_scan_commit(struct inode *inode, struct list_head *dst, 972 struct nfs_commit_info *cinfo) 973 { 974 return 0; 975 } 976 #endif 977 978 /* 979 * Search for an existing write request, and attempt to update 980 * it to reflect a new dirty region on a given page. 981 * 982 * If the attempt fails, then the existing request is flushed out 983 * to disk. 984 */ 985 static struct nfs_page *nfs_try_to_update_request(struct inode *inode, 986 struct page *page, 987 unsigned int offset, 988 unsigned int bytes) 989 { 990 struct nfs_page *req; 991 unsigned int rqend; 992 unsigned int end; 993 int error; 994 995 if (!PagePrivate(page)) 996 return NULL; 997 998 end = offset + bytes; 999 spin_lock(&inode->i_lock); 1000 1001 for (;;) { 1002 req = nfs_page_find_head_request_locked(NFS_I(inode), page); 1003 if (req == NULL) 1004 goto out_unlock; 1005 1006 /* should be handled by nfs_flush_incompatible */ 1007 WARN_ON_ONCE(req->wb_head != req); 1008 WARN_ON_ONCE(req->wb_this_page != req); 1009 1010 rqend = req->wb_offset + req->wb_bytes; 1011 /* 1012 * Tell the caller to flush out the request if 1013 * the offsets are non-contiguous. 1014 * Note: nfs_flush_incompatible() will already 1015 * have flushed out requests having wrong owners. 1016 */ 1017 if (offset > rqend 1018 || end < req->wb_offset) 1019 goto out_flushme; 1020 1021 if (nfs_lock_request(req)) 1022 break; 1023 1024 /* The request is locked, so wait and then retry */ 1025 spin_unlock(&inode->i_lock); 1026 error = nfs_wait_on_request(req); 1027 nfs_release_request(req); 1028 if (error != 0) 1029 goto out_err; 1030 spin_lock(&inode->i_lock); 1031 } 1032 1033 /* Okay, the request matches. Update the region */ 1034 if (offset < req->wb_offset) { 1035 req->wb_offset = offset; 1036 req->wb_pgbase = offset; 1037 } 1038 if (end > rqend) 1039 req->wb_bytes = end - req->wb_offset; 1040 else 1041 req->wb_bytes = rqend - req->wb_offset; 1042 out_unlock: 1043 spin_unlock(&inode->i_lock); 1044 if (req) 1045 nfs_clear_request_commit(req); 1046 return req; 1047 out_flushme: 1048 spin_unlock(&inode->i_lock); 1049 nfs_release_request(req); 1050 error = nfs_wb_page(inode, page); 1051 out_err: 1052 return ERR_PTR(error); 1053 } 1054 1055 /* 1056 * Try to update an existing write request, or create one if there is none. 1057 * 1058 * Note: Should always be called with the Page Lock held to prevent races 1059 * if we have to add a new request. Also assumes that the caller has 1060 * already called nfs_flush_incompatible() if necessary. 1061 */ 1062 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx, 1063 struct page *page, unsigned int offset, unsigned int bytes) 1064 { 1065 struct inode *inode = page_file_mapping(page)->host; 1066 struct nfs_page *req; 1067 1068 req = nfs_try_to_update_request(inode, page, offset, bytes); 1069 if (req != NULL) 1070 goto out; 1071 req = nfs_create_request(ctx, page, NULL, offset, bytes); 1072 if (IS_ERR(req)) 1073 goto out; 1074 nfs_inode_add_request(inode, req); 1075 out: 1076 return req; 1077 } 1078 1079 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page, 1080 unsigned int offset, unsigned int count) 1081 { 1082 struct nfs_page *req; 1083 1084 req = nfs_setup_write_request(ctx, page, offset, count); 1085 if (IS_ERR(req)) 1086 return PTR_ERR(req); 1087 /* Update file length */ 1088 nfs_grow_file(page, offset, count); 1089 nfs_mark_uptodate(req); 1090 nfs_mark_request_dirty(req); 1091 nfs_unlock_and_release_request(req); 1092 return 0; 1093 } 1094 1095 int nfs_flush_incompatible(struct file *file, struct page *page) 1096 { 1097 struct nfs_open_context *ctx = nfs_file_open_context(file); 1098 struct nfs_lock_context *l_ctx; 1099 struct nfs_page *req; 1100 int do_flush, status; 1101 /* 1102 * Look for a request corresponding to this page. If there 1103 * is one, and it belongs to another file, we flush it out 1104 * before we try to copy anything into the page. Do this 1105 * due to the lack of an ACCESS-type call in NFSv2. 1106 * Also do the same if we find a request from an existing 1107 * dropped page. 1108 */ 1109 do { 1110 req = nfs_page_find_head_request(page); 1111 if (req == NULL) 1112 return 0; 1113 l_ctx = req->wb_lock_context; 1114 do_flush = req->wb_page != page || req->wb_context != ctx; 1115 /* for now, flush if more than 1 request in page_group */ 1116 do_flush |= req->wb_this_page != req; 1117 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) { 1118 do_flush |= l_ctx->lockowner.l_owner != current->files 1119 || l_ctx->lockowner.l_pid != current->tgid; 1120 } 1121 nfs_release_request(req); 1122 if (!do_flush) 1123 return 0; 1124 status = nfs_wb_page(page_file_mapping(page)->host, page); 1125 } while (status == 0); 1126 return status; 1127 } 1128 1129 /* 1130 * Avoid buffered writes when a open context credential's key would 1131 * expire soon. 1132 * 1133 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL. 1134 * 1135 * Return 0 and set a credential flag which triggers the inode to flush 1136 * and performs NFS_FILE_SYNC writes if the key will expired within 1137 * RPC_KEY_EXPIRE_TIMEO. 1138 */ 1139 int 1140 nfs_key_timeout_notify(struct file *filp, struct inode *inode) 1141 { 1142 struct nfs_open_context *ctx = nfs_file_open_context(filp); 1143 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; 1144 1145 return rpcauth_key_timeout_notify(auth, ctx->cred); 1146 } 1147 1148 /* 1149 * Test if the open context credential key is marked to expire soon. 1150 */ 1151 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx) 1152 { 1153 return rpcauth_cred_key_to_expire(ctx->cred); 1154 } 1155 1156 /* 1157 * If the page cache is marked as unsafe or invalid, then we can't rely on 1158 * the PageUptodate() flag. In this case, we will need to turn off 1159 * write optimisations that depend on the page contents being correct. 1160 */ 1161 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode) 1162 { 1163 struct nfs_inode *nfsi = NFS_I(inode); 1164 1165 if (nfs_have_delegated_attributes(inode)) 1166 goto out; 1167 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE) 1168 return false; 1169 smp_rmb(); 1170 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags)) 1171 return false; 1172 out: 1173 if (nfsi->cache_validity & NFS_INO_INVALID_DATA) 1174 return false; 1175 return PageUptodate(page) != 0; 1176 } 1177 1178 /* If we know the page is up to date, and we're not using byte range locks (or 1179 * if we have the whole file locked for writing), it may be more efficient to 1180 * extend the write to cover the entire page in order to avoid fragmentation 1181 * inefficiencies. 1182 * 1183 * If the file is opened for synchronous writes then we can just skip the rest 1184 * of the checks. 1185 */ 1186 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode) 1187 { 1188 if (file->f_flags & O_DSYNC) 1189 return 0; 1190 if (!nfs_write_pageuptodate(page, inode)) 1191 return 0; 1192 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) 1193 return 1; 1194 if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 && 1195 inode->i_flock->fl_end == OFFSET_MAX && 1196 inode->i_flock->fl_type != F_RDLCK)) 1197 return 1; 1198 return 0; 1199 } 1200 1201 /* 1202 * Update and possibly write a cached page of an NFS file. 1203 * 1204 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad 1205 * things with a page scheduled for an RPC call (e.g. invalidate it). 1206 */ 1207 int nfs_updatepage(struct file *file, struct page *page, 1208 unsigned int offset, unsigned int count) 1209 { 1210 struct nfs_open_context *ctx = nfs_file_open_context(file); 1211 struct inode *inode = page_file_mapping(page)->host; 1212 int status = 0; 1213 1214 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); 1215 1216 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n", 1217 file, count, (long long)(page_file_offset(page) + offset)); 1218 1219 if (nfs_can_extend_write(file, page, inode)) { 1220 count = max(count + offset, nfs_page_length(page)); 1221 offset = 0; 1222 } 1223 1224 status = nfs_writepage_setup(ctx, page, offset, count); 1225 if (status < 0) 1226 nfs_set_pageerror(page); 1227 else 1228 __set_page_dirty_nobuffers(page); 1229 1230 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n", 1231 status, (long long)i_size_read(inode)); 1232 return status; 1233 } 1234 1235 static int flush_task_priority(int how) 1236 { 1237 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { 1238 case FLUSH_HIGHPRI: 1239 return RPC_PRIORITY_HIGH; 1240 case FLUSH_LOWPRI: 1241 return RPC_PRIORITY_LOW; 1242 } 1243 return RPC_PRIORITY_NORMAL; 1244 } 1245 1246 static void nfs_initiate_write(struct nfs_pgio_data *data, struct rpc_message *msg, 1247 struct rpc_task_setup *task_setup_data, int how) 1248 { 1249 struct inode *inode = data->header->inode; 1250 int priority = flush_task_priority(how); 1251 1252 task_setup_data->priority = priority; 1253 NFS_PROTO(inode)->write_setup(data, msg); 1254 1255 nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client, 1256 &task_setup_data->rpc_client, msg, data); 1257 } 1258 1259 /* If a nfs_flush_* function fails, it should remove reqs from @head and 1260 * call this on each, which will prepare them to be retried on next 1261 * writeback using standard nfs. 1262 */ 1263 static void nfs_redirty_request(struct nfs_page *req) 1264 { 1265 nfs_mark_request_dirty(req); 1266 nfs_unlock_request(req); 1267 nfs_end_page_writeback(req); 1268 nfs_release_request(req); 1269 } 1270 1271 static void nfs_async_write_error(struct list_head *head) 1272 { 1273 struct nfs_page *req; 1274 1275 while (!list_empty(head)) { 1276 req = nfs_list_entry(head->next); 1277 nfs_list_remove_request(req); 1278 nfs_redirty_request(req); 1279 } 1280 } 1281 1282 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = { 1283 .error_cleanup = nfs_async_write_error, 1284 .completion = nfs_write_completion, 1285 }; 1286 1287 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, 1288 struct inode *inode, int ioflags, bool force_mds, 1289 const struct nfs_pgio_completion_ops *compl_ops) 1290 { 1291 struct nfs_server *server = NFS_SERVER(inode); 1292 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; 1293 1294 #ifdef CONFIG_NFS_V4_1 1295 if (server->pnfs_curr_ld && !force_mds) 1296 pg_ops = server->pnfs_curr_ld->pg_write_ops; 1297 #endif 1298 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops, 1299 server->wsize, ioflags); 1300 } 1301 EXPORT_SYMBOL_GPL(nfs_pageio_init_write); 1302 1303 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio) 1304 { 1305 pgio->pg_ops = &nfs_pgio_rw_ops; 1306 pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize; 1307 } 1308 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds); 1309 1310 1311 void nfs_commit_prepare(struct rpc_task *task, void *calldata) 1312 { 1313 struct nfs_commit_data *data = calldata; 1314 1315 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data); 1316 } 1317 1318 static void nfs_writeback_release_common(struct nfs_pgio_data *data) 1319 { 1320 struct nfs_pgio_header *hdr = data->header; 1321 int status = data->task.tk_status; 1322 1323 if ((status >= 0) && nfs_write_need_commit(data)) { 1324 spin_lock(&hdr->lock); 1325 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) 1326 ; /* Do nothing */ 1327 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) 1328 memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf)); 1329 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf))) 1330 set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags); 1331 spin_unlock(&hdr->lock); 1332 } 1333 } 1334 1335 /* 1336 * Special version of should_remove_suid() that ignores capabilities. 1337 */ 1338 static int nfs_should_remove_suid(const struct inode *inode) 1339 { 1340 umode_t mode = inode->i_mode; 1341 int kill = 0; 1342 1343 /* suid always must be killed */ 1344 if (unlikely(mode & S_ISUID)) 1345 kill = ATTR_KILL_SUID; 1346 1347 /* 1348 * sgid without any exec bits is just a mandatory locking mark; leave 1349 * it alone. If some exec bits are set, it's a real sgid; kill it. 1350 */ 1351 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP))) 1352 kill |= ATTR_KILL_SGID; 1353 1354 if (unlikely(kill && S_ISREG(mode))) 1355 return kill; 1356 1357 return 0; 1358 } 1359 1360 /* 1361 * This function is called when the WRITE call is complete. 1362 */ 1363 static int nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data, 1364 struct inode *inode) 1365 { 1366 int status; 1367 1368 /* 1369 * ->write_done will attempt to use post-op attributes to detect 1370 * conflicting writes by other clients. A strict interpretation 1371 * of close-to-open would allow us to continue caching even if 1372 * another writer had changed the file, but some applications 1373 * depend on tighter cache coherency when writing. 1374 */ 1375 status = NFS_PROTO(inode)->write_done(task, data); 1376 if (status != 0) 1377 return status; 1378 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, data->res.count); 1379 1380 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) 1381 if (data->res.verf->committed < data->args.stable && task->tk_status >= 0) { 1382 /* We tried a write call, but the server did not 1383 * commit data to stable storage even though we 1384 * requested it. 1385 * Note: There is a known bug in Tru64 < 5.0 in which 1386 * the server reports NFS_DATA_SYNC, but performs 1387 * NFS_FILE_SYNC. We therefore implement this checking 1388 * as a dprintk() in order to avoid filling syslog. 1389 */ 1390 static unsigned long complain; 1391 1392 /* Note this will print the MDS for a DS write */ 1393 if (time_before(complain, jiffies)) { 1394 dprintk("NFS: faulty NFS server %s:" 1395 " (committed = %d) != (stable = %d)\n", 1396 NFS_SERVER(inode)->nfs_client->cl_hostname, 1397 data->res.verf->committed, data->args.stable); 1398 complain = jiffies + 300 * HZ; 1399 } 1400 } 1401 #endif 1402 1403 /* Deal with the suid/sgid bit corner case */ 1404 if (nfs_should_remove_suid(inode)) 1405 nfs_mark_for_revalidate(inode); 1406 return 0; 1407 } 1408 1409 /* 1410 * This function is called when the WRITE call is complete. 1411 */ 1412 static void nfs_writeback_result(struct rpc_task *task, struct nfs_pgio_data *data) 1413 { 1414 struct nfs_pgio_args *argp = &data->args; 1415 struct nfs_pgio_res *resp = &data->res; 1416 1417 if (resp->count < argp->count) { 1418 static unsigned long complain; 1419 1420 /* This a short write! */ 1421 nfs_inc_stats(data->header->inode, NFSIOS_SHORTWRITE); 1422 1423 /* Has the server at least made some progress? */ 1424 if (resp->count == 0) { 1425 if (time_before(complain, jiffies)) { 1426 printk(KERN_WARNING 1427 "NFS: Server wrote zero bytes, expected %u.\n", 1428 argp->count); 1429 complain = jiffies + 300 * HZ; 1430 } 1431 nfs_set_pgio_error(data->header, -EIO, argp->offset); 1432 task->tk_status = -EIO; 1433 return; 1434 } 1435 /* Was this an NFSv2 write or an NFSv3 stable write? */ 1436 if (resp->verf->committed != NFS_UNSTABLE) { 1437 /* Resend from where the server left off */ 1438 data->mds_offset += resp->count; 1439 argp->offset += resp->count; 1440 argp->pgbase += resp->count; 1441 argp->count -= resp->count; 1442 } else { 1443 /* Resend as a stable write in order to avoid 1444 * headaches in the case of a server crash. 1445 */ 1446 argp->stable = NFS_FILE_SYNC; 1447 } 1448 rpc_restart_call_prepare(task); 1449 } 1450 } 1451 1452 1453 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4) 1454 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait) 1455 { 1456 int ret; 1457 1458 if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags)) 1459 return 1; 1460 if (!may_wait) 1461 return 0; 1462 ret = out_of_line_wait_on_bit_lock(&nfsi->flags, 1463 NFS_INO_COMMIT, 1464 nfs_wait_bit_killable, 1465 TASK_KILLABLE); 1466 return (ret < 0) ? ret : 1; 1467 } 1468 1469 static void nfs_commit_clear_lock(struct nfs_inode *nfsi) 1470 { 1471 clear_bit(NFS_INO_COMMIT, &nfsi->flags); 1472 smp_mb__after_atomic(); 1473 wake_up_bit(&nfsi->flags, NFS_INO_COMMIT); 1474 } 1475 1476 void nfs_commitdata_release(struct nfs_commit_data *data) 1477 { 1478 put_nfs_open_context(data->context); 1479 nfs_commit_free(data); 1480 } 1481 EXPORT_SYMBOL_GPL(nfs_commitdata_release); 1482 1483 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data, 1484 const struct rpc_call_ops *call_ops, 1485 int how, int flags) 1486 { 1487 struct rpc_task *task; 1488 int priority = flush_task_priority(how); 1489 struct rpc_message msg = { 1490 .rpc_argp = &data->args, 1491 .rpc_resp = &data->res, 1492 .rpc_cred = data->cred, 1493 }; 1494 struct rpc_task_setup task_setup_data = { 1495 .task = &data->task, 1496 .rpc_client = clnt, 1497 .rpc_message = &msg, 1498 .callback_ops = call_ops, 1499 .callback_data = data, 1500 .workqueue = nfsiod_workqueue, 1501 .flags = RPC_TASK_ASYNC | flags, 1502 .priority = priority, 1503 }; 1504 /* Set up the initial task struct. */ 1505 NFS_PROTO(data->inode)->commit_setup(data, &msg); 1506 1507 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); 1508 1509 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client, 1510 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg); 1511 1512 task = rpc_run_task(&task_setup_data); 1513 if (IS_ERR(task)) 1514 return PTR_ERR(task); 1515 if (how & FLUSH_SYNC) 1516 rpc_wait_for_completion_task(task); 1517 rpc_put_task(task); 1518 return 0; 1519 } 1520 EXPORT_SYMBOL_GPL(nfs_initiate_commit); 1521 1522 /* 1523 * Set up the argument/result storage required for the RPC call. 1524 */ 1525 void nfs_init_commit(struct nfs_commit_data *data, 1526 struct list_head *head, 1527 struct pnfs_layout_segment *lseg, 1528 struct nfs_commit_info *cinfo) 1529 { 1530 struct nfs_page *first = nfs_list_entry(head->next); 1531 struct inode *inode = first->wb_context->dentry->d_inode; 1532 1533 /* Set up the RPC argument and reply structs 1534 * NB: take care not to mess about with data->commit et al. */ 1535 1536 list_splice_init(head, &data->pages); 1537 1538 data->inode = inode; 1539 data->cred = first->wb_context->cred; 1540 data->lseg = lseg; /* reference transferred */ 1541 data->mds_ops = &nfs_commit_ops; 1542 data->completion_ops = cinfo->completion_ops; 1543 data->dreq = cinfo->dreq; 1544 1545 data->args.fh = NFS_FH(data->inode); 1546 /* Note: we always request a commit of the entire inode */ 1547 data->args.offset = 0; 1548 data->args.count = 0; 1549 data->context = get_nfs_open_context(first->wb_context); 1550 data->res.fattr = &data->fattr; 1551 data->res.verf = &data->verf; 1552 nfs_fattr_init(&data->fattr); 1553 } 1554 EXPORT_SYMBOL_GPL(nfs_init_commit); 1555 1556 void nfs_retry_commit(struct list_head *page_list, 1557 struct pnfs_layout_segment *lseg, 1558 struct nfs_commit_info *cinfo) 1559 { 1560 struct nfs_page *req; 1561 1562 while (!list_empty(page_list)) { 1563 req = nfs_list_entry(page_list->next); 1564 nfs_list_remove_request(req); 1565 nfs_mark_request_commit(req, lseg, cinfo); 1566 if (!cinfo->dreq) { 1567 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 1568 dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info, 1569 BDI_RECLAIMABLE); 1570 } 1571 nfs_unlock_and_release_request(req); 1572 } 1573 } 1574 EXPORT_SYMBOL_GPL(nfs_retry_commit); 1575 1576 /* 1577 * Commit dirty pages 1578 */ 1579 static int 1580 nfs_commit_list(struct inode *inode, struct list_head *head, int how, 1581 struct nfs_commit_info *cinfo) 1582 { 1583 struct nfs_commit_data *data; 1584 1585 data = nfs_commitdata_alloc(); 1586 1587 if (!data) 1588 goto out_bad; 1589 1590 /* Set up the argument struct */ 1591 nfs_init_commit(data, head, NULL, cinfo); 1592 atomic_inc(&cinfo->mds->rpcs_out); 1593 return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops, 1594 how, 0); 1595 out_bad: 1596 nfs_retry_commit(head, NULL, cinfo); 1597 cinfo->completion_ops->error_cleanup(NFS_I(inode)); 1598 return -ENOMEM; 1599 } 1600 1601 /* 1602 * COMMIT call returned 1603 */ 1604 static void nfs_commit_done(struct rpc_task *task, void *calldata) 1605 { 1606 struct nfs_commit_data *data = calldata; 1607 1608 dprintk("NFS: %5u nfs_commit_done (status %d)\n", 1609 task->tk_pid, task->tk_status); 1610 1611 /* Call the NFS version-specific code */ 1612 NFS_PROTO(data->inode)->commit_done(task, data); 1613 } 1614 1615 static void nfs_commit_release_pages(struct nfs_commit_data *data) 1616 { 1617 struct nfs_page *req; 1618 int status = data->task.tk_status; 1619 struct nfs_commit_info cinfo; 1620 1621 while (!list_empty(&data->pages)) { 1622 req = nfs_list_entry(data->pages.next); 1623 nfs_list_remove_request(req); 1624 nfs_clear_page_commit(req->wb_page); 1625 1626 dprintk("NFS: commit (%s/%llu %d@%lld)", 1627 req->wb_context->dentry->d_sb->s_id, 1628 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode), 1629 req->wb_bytes, 1630 (long long)req_offset(req)); 1631 if (status < 0) { 1632 nfs_context_set_write_error(req->wb_context, status); 1633 nfs_inode_remove_request(req); 1634 dprintk(", error = %d\n", status); 1635 goto next; 1636 } 1637 1638 /* Okay, COMMIT succeeded, apparently. Check the verifier 1639 * returned by the server against all stored verfs. */ 1640 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) { 1641 /* We have a match */ 1642 nfs_inode_remove_request(req); 1643 dprintk(" OK\n"); 1644 goto next; 1645 } 1646 /* We have a mismatch. Write the page again */ 1647 dprintk(" mismatch\n"); 1648 nfs_mark_request_dirty(req); 1649 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags); 1650 next: 1651 nfs_unlock_and_release_request(req); 1652 } 1653 nfs_init_cinfo(&cinfo, data->inode, data->dreq); 1654 if (atomic_dec_and_test(&cinfo.mds->rpcs_out)) 1655 nfs_commit_clear_lock(NFS_I(data->inode)); 1656 } 1657 1658 static void nfs_commit_release(void *calldata) 1659 { 1660 struct nfs_commit_data *data = calldata; 1661 1662 data->completion_ops->completion(data); 1663 nfs_commitdata_release(calldata); 1664 } 1665 1666 static const struct rpc_call_ops nfs_commit_ops = { 1667 .rpc_call_prepare = nfs_commit_prepare, 1668 .rpc_call_done = nfs_commit_done, 1669 .rpc_release = nfs_commit_release, 1670 }; 1671 1672 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = { 1673 .completion = nfs_commit_release_pages, 1674 .error_cleanup = nfs_commit_clear_lock, 1675 }; 1676 1677 int nfs_generic_commit_list(struct inode *inode, struct list_head *head, 1678 int how, struct nfs_commit_info *cinfo) 1679 { 1680 int status; 1681 1682 status = pnfs_commit_list(inode, head, how, cinfo); 1683 if (status == PNFS_NOT_ATTEMPTED) 1684 status = nfs_commit_list(inode, head, how, cinfo); 1685 return status; 1686 } 1687 1688 int nfs_commit_inode(struct inode *inode, int how) 1689 { 1690 LIST_HEAD(head); 1691 struct nfs_commit_info cinfo; 1692 int may_wait = how & FLUSH_SYNC; 1693 int res; 1694 1695 res = nfs_commit_set_lock(NFS_I(inode), may_wait); 1696 if (res <= 0) 1697 goto out_mark_dirty; 1698 nfs_init_cinfo_from_inode(&cinfo, inode); 1699 res = nfs_scan_commit(inode, &head, &cinfo); 1700 if (res) { 1701 int error; 1702 1703 error = nfs_generic_commit_list(inode, &head, how, &cinfo); 1704 if (error < 0) 1705 return error; 1706 if (!may_wait) 1707 goto out_mark_dirty; 1708 error = wait_on_bit(&NFS_I(inode)->flags, 1709 NFS_INO_COMMIT, 1710 nfs_wait_bit_killable, 1711 TASK_KILLABLE); 1712 if (error < 0) 1713 return error; 1714 } else 1715 nfs_commit_clear_lock(NFS_I(inode)); 1716 return res; 1717 /* Note: If we exit without ensuring that the commit is complete, 1718 * we must mark the inode as dirty. Otherwise, future calls to 1719 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure 1720 * that the data is on the disk. 1721 */ 1722 out_mark_dirty: 1723 __mark_inode_dirty(inode, I_DIRTY_DATASYNC); 1724 return res; 1725 } 1726 1727 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc) 1728 { 1729 struct nfs_inode *nfsi = NFS_I(inode); 1730 int flags = FLUSH_SYNC; 1731 int ret = 0; 1732 1733 /* no commits means nothing needs to be done */ 1734 if (!nfsi->commit_info.ncommit) 1735 return ret; 1736 1737 if (wbc->sync_mode == WB_SYNC_NONE) { 1738 /* Don't commit yet if this is a non-blocking flush and there 1739 * are a lot of outstanding writes for this mapping. 1740 */ 1741 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1)) 1742 goto out_mark_dirty; 1743 1744 /* don't wait for the COMMIT response */ 1745 flags = 0; 1746 } 1747 1748 ret = nfs_commit_inode(inode, flags); 1749 if (ret >= 0) { 1750 if (wbc->sync_mode == WB_SYNC_NONE) { 1751 if (ret < wbc->nr_to_write) 1752 wbc->nr_to_write -= ret; 1753 else 1754 wbc->nr_to_write = 0; 1755 } 1756 return 0; 1757 } 1758 out_mark_dirty: 1759 __mark_inode_dirty(inode, I_DIRTY_DATASYNC); 1760 return ret; 1761 } 1762 #else 1763 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc) 1764 { 1765 return 0; 1766 } 1767 #endif 1768 1769 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc) 1770 { 1771 return nfs_commit_unstable_pages(inode, wbc); 1772 } 1773 EXPORT_SYMBOL_GPL(nfs_write_inode); 1774 1775 /* 1776 * flush the inode to disk. 1777 */ 1778 int nfs_wb_all(struct inode *inode) 1779 { 1780 struct writeback_control wbc = { 1781 .sync_mode = WB_SYNC_ALL, 1782 .nr_to_write = LONG_MAX, 1783 .range_start = 0, 1784 .range_end = LLONG_MAX, 1785 }; 1786 int ret; 1787 1788 trace_nfs_writeback_inode_enter(inode); 1789 1790 ret = sync_inode(inode, &wbc); 1791 1792 trace_nfs_writeback_inode_exit(inode, ret); 1793 return ret; 1794 } 1795 EXPORT_SYMBOL_GPL(nfs_wb_all); 1796 1797 int nfs_wb_page_cancel(struct inode *inode, struct page *page) 1798 { 1799 struct nfs_page *req; 1800 int ret = 0; 1801 1802 for (;;) { 1803 wait_on_page_writeback(page); 1804 req = nfs_page_find_head_request(page); 1805 if (req == NULL) 1806 break; 1807 if (nfs_lock_request(req)) { 1808 nfs_clear_request_commit(req); 1809 nfs_inode_remove_request(req); 1810 /* 1811 * In case nfs_inode_remove_request has marked the 1812 * page as being dirty 1813 */ 1814 cancel_dirty_page(page, PAGE_CACHE_SIZE); 1815 nfs_unlock_and_release_request(req); 1816 break; 1817 } 1818 ret = nfs_wait_on_request(req); 1819 nfs_release_request(req); 1820 if (ret < 0) 1821 break; 1822 } 1823 return ret; 1824 } 1825 1826 /* 1827 * Write back all requests on one page - we do this before reading it. 1828 */ 1829 int nfs_wb_page(struct inode *inode, struct page *page) 1830 { 1831 loff_t range_start = page_file_offset(page); 1832 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1); 1833 struct writeback_control wbc = { 1834 .sync_mode = WB_SYNC_ALL, 1835 .nr_to_write = 0, 1836 .range_start = range_start, 1837 .range_end = range_end, 1838 }; 1839 int ret; 1840 1841 trace_nfs_writeback_page_enter(inode); 1842 1843 for (;;) { 1844 wait_on_page_writeback(page); 1845 if (clear_page_dirty_for_io(page)) { 1846 ret = nfs_writepage_locked(page, &wbc); 1847 if (ret < 0) 1848 goto out_error; 1849 continue; 1850 } 1851 ret = 0; 1852 if (!PagePrivate(page)) 1853 break; 1854 ret = nfs_commit_inode(inode, FLUSH_SYNC); 1855 if (ret < 0) 1856 goto out_error; 1857 } 1858 out_error: 1859 trace_nfs_writeback_page_exit(inode, ret); 1860 return ret; 1861 } 1862 1863 #ifdef CONFIG_MIGRATION 1864 int nfs_migrate_page(struct address_space *mapping, struct page *newpage, 1865 struct page *page, enum migrate_mode mode) 1866 { 1867 /* 1868 * If PagePrivate is set, then the page is currently associated with 1869 * an in-progress read or write request. Don't try to migrate it. 1870 * 1871 * FIXME: we could do this in principle, but we'll need a way to ensure 1872 * that we can safely release the inode reference while holding 1873 * the page lock. 1874 */ 1875 if (PagePrivate(page)) 1876 return -EBUSY; 1877 1878 if (!nfs_fscache_release_page(page, GFP_KERNEL)) 1879 return -EBUSY; 1880 1881 return migrate_page(mapping, newpage, page, mode); 1882 } 1883 #endif 1884 1885 int __init nfs_init_writepagecache(void) 1886 { 1887 nfs_wdata_cachep = kmem_cache_create("nfs_write_data", 1888 sizeof(struct nfs_rw_header), 1889 0, SLAB_HWCACHE_ALIGN, 1890 NULL); 1891 if (nfs_wdata_cachep == NULL) 1892 return -ENOMEM; 1893 1894 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, 1895 nfs_wdata_cachep); 1896 if (nfs_wdata_mempool == NULL) 1897 goto out_destroy_write_cache; 1898 1899 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data", 1900 sizeof(struct nfs_commit_data), 1901 0, SLAB_HWCACHE_ALIGN, 1902 NULL); 1903 if (nfs_cdata_cachep == NULL) 1904 goto out_destroy_write_mempool; 1905 1906 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, 1907 nfs_cdata_cachep); 1908 if (nfs_commit_mempool == NULL) 1909 goto out_destroy_commit_cache; 1910 1911 /* 1912 * NFS congestion size, scale with available memory. 1913 * 1914 * 64MB: 8192k 1915 * 128MB: 11585k 1916 * 256MB: 16384k 1917 * 512MB: 23170k 1918 * 1GB: 32768k 1919 * 2GB: 46340k 1920 * 4GB: 65536k 1921 * 8GB: 92681k 1922 * 16GB: 131072k 1923 * 1924 * This allows larger machines to have larger/more transfers. 1925 * Limit the default to 256M 1926 */ 1927 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10); 1928 if (nfs_congestion_kb > 256*1024) 1929 nfs_congestion_kb = 256*1024; 1930 1931 return 0; 1932 1933 out_destroy_commit_cache: 1934 kmem_cache_destroy(nfs_cdata_cachep); 1935 out_destroy_write_mempool: 1936 mempool_destroy(nfs_wdata_mempool); 1937 out_destroy_write_cache: 1938 kmem_cache_destroy(nfs_wdata_cachep); 1939 return -ENOMEM; 1940 } 1941 1942 void nfs_destroy_writepagecache(void) 1943 { 1944 mempool_destroy(nfs_commit_mempool); 1945 kmem_cache_destroy(nfs_cdata_cachep); 1946 mempool_destroy(nfs_wdata_mempool); 1947 kmem_cache_destroy(nfs_wdata_cachep); 1948 } 1949 1950 static const struct nfs_rw_ops nfs_rw_write_ops = { 1951 .rw_mode = FMODE_WRITE, 1952 .rw_alloc_header = nfs_writehdr_alloc, 1953 .rw_free_header = nfs_writehdr_free, 1954 .rw_release = nfs_writeback_release_common, 1955 .rw_done = nfs_writeback_done, 1956 .rw_result = nfs_writeback_result, 1957 .rw_initiate = nfs_initiate_write, 1958 }; 1959