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 #include <linux/freezer.h> 25 #include <linux/wait.h> 26 27 #include <linux/uaccess.h> 28 29 #include "delegation.h" 30 #include "internal.h" 31 #include "iostat.h" 32 #include "nfs4_fs.h" 33 #include "fscache.h" 34 #include "pnfs.h" 35 36 #include "nfstrace.h" 37 38 #define NFSDBG_FACILITY NFSDBG_PAGECACHE 39 40 #define MIN_POOL_WRITE (32) 41 #define MIN_POOL_COMMIT (4) 42 43 struct nfs_io_completion { 44 void (*complete)(void *data); 45 void *data; 46 struct kref refcount; 47 }; 48 49 /* 50 * Local function declarations 51 */ 52 static void nfs_redirty_request(struct nfs_page *req); 53 static const struct rpc_call_ops nfs_commit_ops; 54 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops; 55 static const struct nfs_commit_completion_ops nfs_commit_completion_ops; 56 static const struct nfs_rw_ops nfs_rw_write_ops; 57 static void nfs_clear_request_commit(struct nfs_page *req); 58 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, 59 struct inode *inode); 60 static struct nfs_page * 61 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, 62 struct page *page); 63 64 static struct kmem_cache *nfs_wdata_cachep; 65 static mempool_t *nfs_wdata_mempool; 66 static struct kmem_cache *nfs_cdata_cachep; 67 static mempool_t *nfs_commit_mempool; 68 69 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail) 70 { 71 struct nfs_commit_data *p; 72 73 if (never_fail) 74 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO); 75 else { 76 /* It is OK to do some reclaim, not no safe to wait 77 * for anything to be returned to the pool. 78 * mempool_alloc() cannot handle that particular combination, 79 * so we need two separate attempts. 80 */ 81 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT); 82 if (!p) 83 p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO | 84 __GFP_NOWARN | __GFP_NORETRY); 85 if (!p) 86 return NULL; 87 } 88 89 memset(p, 0, sizeof(*p)); 90 INIT_LIST_HEAD(&p->pages); 91 return p; 92 } 93 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc); 94 95 void nfs_commit_free(struct nfs_commit_data *p) 96 { 97 mempool_free(p, nfs_commit_mempool); 98 } 99 EXPORT_SYMBOL_GPL(nfs_commit_free); 100 101 static struct nfs_pgio_header *nfs_writehdr_alloc(void) 102 { 103 struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO); 104 105 memset(p, 0, sizeof(*p)); 106 p->rw_mode = FMODE_WRITE; 107 return p; 108 } 109 110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr) 111 { 112 mempool_free(hdr, nfs_wdata_mempool); 113 } 114 115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags) 116 { 117 return kmalloc(sizeof(struct nfs_io_completion), gfp_flags); 118 } 119 120 static void nfs_io_completion_init(struct nfs_io_completion *ioc, 121 void (*complete)(void *), void *data) 122 { 123 ioc->complete = complete; 124 ioc->data = data; 125 kref_init(&ioc->refcount); 126 } 127 128 static void nfs_io_completion_release(struct kref *kref) 129 { 130 struct nfs_io_completion *ioc = container_of(kref, 131 struct nfs_io_completion, refcount); 132 ioc->complete(ioc->data); 133 kfree(ioc); 134 } 135 136 static void nfs_io_completion_get(struct nfs_io_completion *ioc) 137 { 138 if (ioc != NULL) 139 kref_get(&ioc->refcount); 140 } 141 142 static void nfs_io_completion_put(struct nfs_io_completion *ioc) 143 { 144 if (ioc != NULL) 145 kref_put(&ioc->refcount, nfs_io_completion_release); 146 } 147 148 static struct nfs_page * 149 nfs_page_private_request(struct page *page) 150 { 151 if (!PagePrivate(page)) 152 return NULL; 153 return (struct nfs_page *)page_private(page); 154 } 155 156 /* 157 * nfs_page_find_head_request_locked - find head request associated with @page 158 * 159 * must be called while holding the inode lock. 160 * 161 * returns matching head request with reference held, or NULL if not found. 162 */ 163 static struct nfs_page * 164 nfs_page_find_private_request(struct page *page) 165 { 166 struct address_space *mapping = page_file_mapping(page); 167 struct nfs_page *req; 168 169 if (!PagePrivate(page)) 170 return NULL; 171 spin_lock(&mapping->private_lock); 172 req = nfs_page_private_request(page); 173 if (req) { 174 WARN_ON_ONCE(req->wb_head != req); 175 kref_get(&req->wb_kref); 176 } 177 spin_unlock(&mapping->private_lock); 178 return req; 179 } 180 181 static struct nfs_page * 182 nfs_page_find_swap_request(struct page *page) 183 { 184 struct inode *inode = page_file_mapping(page)->host; 185 struct nfs_inode *nfsi = NFS_I(inode); 186 struct nfs_page *req = NULL; 187 if (!PageSwapCache(page)) 188 return NULL; 189 mutex_lock(&nfsi->commit_mutex); 190 if (PageSwapCache(page)) { 191 req = nfs_page_search_commits_for_head_request_locked(nfsi, 192 page); 193 if (req) { 194 WARN_ON_ONCE(req->wb_head != req); 195 kref_get(&req->wb_kref); 196 } 197 } 198 mutex_unlock(&nfsi->commit_mutex); 199 return req; 200 } 201 202 /* 203 * nfs_page_find_head_request - find head request associated with @page 204 * 205 * returns matching head request with reference held, or NULL if not found. 206 */ 207 static struct nfs_page *nfs_page_find_head_request(struct page *page) 208 { 209 struct nfs_page *req; 210 211 req = nfs_page_find_private_request(page); 212 if (!req) 213 req = nfs_page_find_swap_request(page); 214 return req; 215 } 216 217 /* Adjust the file length if we're writing beyond the end */ 218 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) 219 { 220 struct inode *inode = page_file_mapping(page)->host; 221 loff_t end, i_size; 222 pgoff_t end_index; 223 224 spin_lock(&inode->i_lock); 225 i_size = i_size_read(inode); 226 end_index = (i_size - 1) >> PAGE_SHIFT; 227 if (i_size > 0 && page_index(page) < end_index) 228 goto out; 229 end = page_file_offset(page) + ((loff_t)offset+count); 230 if (i_size >= end) 231 goto out; 232 i_size_write(inode, end); 233 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); 234 out: 235 spin_unlock(&inode->i_lock); 236 } 237 238 /* A writeback failed: mark the page as bad, and invalidate the page cache */ 239 static void nfs_set_pageerror(struct page *page) 240 { 241 nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page)); 242 } 243 244 /* 245 * nfs_page_group_search_locked 246 * @head - head request of page group 247 * @page_offset - offset into page 248 * 249 * Search page group with head @head to find a request that contains the 250 * page offset @page_offset. 251 * 252 * Returns a pointer to the first matching nfs request, or NULL if no 253 * match is found. 254 * 255 * Must be called with the page group lock held 256 */ 257 static struct nfs_page * 258 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset) 259 { 260 struct nfs_page *req; 261 262 req = head; 263 do { 264 if (page_offset >= req->wb_pgbase && 265 page_offset < (req->wb_pgbase + req->wb_bytes)) 266 return req; 267 268 req = req->wb_this_page; 269 } while (req != head); 270 271 return NULL; 272 } 273 274 /* 275 * nfs_page_group_covers_page 276 * @head - head request of page group 277 * 278 * Return true if the page group with head @head covers the whole page, 279 * returns false otherwise 280 */ 281 static bool nfs_page_group_covers_page(struct nfs_page *req) 282 { 283 struct nfs_page *tmp; 284 unsigned int pos = 0; 285 unsigned int len = nfs_page_length(req->wb_page); 286 287 nfs_page_group_lock(req); 288 289 for (;;) { 290 tmp = nfs_page_group_search_locked(req->wb_head, pos); 291 if (!tmp) 292 break; 293 pos = tmp->wb_pgbase + tmp->wb_bytes; 294 } 295 296 nfs_page_group_unlock(req); 297 return pos >= len; 298 } 299 300 /* We can set the PG_uptodate flag if we see that a write request 301 * covers the full page. 302 */ 303 static void nfs_mark_uptodate(struct nfs_page *req) 304 { 305 if (PageUptodate(req->wb_page)) 306 return; 307 if (!nfs_page_group_covers_page(req)) 308 return; 309 SetPageUptodate(req->wb_page); 310 } 311 312 static int wb_priority(struct writeback_control *wbc) 313 { 314 int ret = 0; 315 316 if (wbc->sync_mode == WB_SYNC_ALL) 317 ret = FLUSH_COND_STABLE; 318 return ret; 319 } 320 321 /* 322 * NFS congestion control 323 */ 324 325 int nfs_congestion_kb; 326 327 #define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) 328 #define NFS_CONGESTION_OFF_THRESH \ 329 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) 330 331 static void nfs_set_page_writeback(struct page *page) 332 { 333 struct inode *inode = page_file_mapping(page)->host; 334 struct nfs_server *nfss = NFS_SERVER(inode); 335 int ret = test_set_page_writeback(page); 336 337 WARN_ON_ONCE(ret != 0); 338 339 if (atomic_long_inc_return(&nfss->writeback) > 340 NFS_CONGESTION_ON_THRESH) 341 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC); 342 } 343 344 static void nfs_end_page_writeback(struct nfs_page *req) 345 { 346 struct inode *inode = page_file_mapping(req->wb_page)->host; 347 struct nfs_server *nfss = NFS_SERVER(inode); 348 bool is_done; 349 350 is_done = nfs_page_group_sync_on_bit(req, PG_WB_END); 351 nfs_unlock_request(req); 352 if (!is_done) 353 return; 354 355 end_page_writeback(req->wb_page); 356 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) 357 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC); 358 } 359 360 /* 361 * nfs_unroll_locks_and_wait - unlock all newly locked reqs and wait on @req 362 * 363 * this is a helper function for nfs_lock_and_join_requests 364 * 365 * @inode - inode associated with request page group, must be holding inode lock 366 * @head - head request of page group, must be holding head lock 367 * @req - request that couldn't lock and needs to wait on the req bit lock 368 * 369 * NOTE: this must be called holding page_group bit lock 370 * which will be released before returning. 371 * 372 * returns 0 on success, < 0 on error. 373 */ 374 static void 375 nfs_unroll_locks(struct inode *inode, struct nfs_page *head, 376 struct nfs_page *req) 377 { 378 struct nfs_page *tmp; 379 380 /* relinquish all the locks successfully grabbed this run */ 381 for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) { 382 if (!kref_read(&tmp->wb_kref)) 383 continue; 384 nfs_unlock_and_release_request(tmp); 385 } 386 } 387 388 /* 389 * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests 390 * 391 * @destroy_list - request list (using wb_this_page) terminated by @old_head 392 * @old_head - the old head of the list 393 * 394 * All subrequests must be locked and removed from all lists, so at this point 395 * they are only "active" in this function, and possibly in nfs_wait_on_request 396 * with a reference held by some other context. 397 */ 398 static void 399 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list, 400 struct nfs_page *old_head, 401 struct inode *inode) 402 { 403 while (destroy_list) { 404 struct nfs_page *subreq = destroy_list; 405 406 destroy_list = (subreq->wb_this_page == old_head) ? 407 NULL : subreq->wb_this_page; 408 409 WARN_ON_ONCE(old_head != subreq->wb_head); 410 411 /* make sure old group is not used */ 412 subreq->wb_this_page = subreq; 413 414 clear_bit(PG_REMOVE, &subreq->wb_flags); 415 416 /* Note: races with nfs_page_group_destroy() */ 417 if (!kref_read(&subreq->wb_kref)) { 418 /* Check if we raced with nfs_page_group_destroy() */ 419 if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) 420 nfs_free_request(subreq); 421 continue; 422 } 423 424 subreq->wb_head = subreq; 425 426 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) { 427 nfs_release_request(subreq); 428 atomic_long_dec(&NFS_I(inode)->nrequests); 429 } 430 431 /* subreq is now totally disconnected from page group or any 432 * write / commit lists. last chance to wake any waiters */ 433 nfs_unlock_and_release_request(subreq); 434 } 435 } 436 437 /* 438 * nfs_lock_and_join_requests - join all subreqs to the head req and return 439 * a locked reference, cancelling any pending 440 * operations for this page. 441 * 442 * @page - the page used to lookup the "page group" of nfs_page structures 443 * 444 * This function joins all sub requests to the head request by first 445 * locking all requests in the group, cancelling any pending operations 446 * and finally updating the head request to cover the whole range covered by 447 * the (former) group. All subrequests are removed from any write or commit 448 * lists, unlinked from the group and destroyed. 449 * 450 * Returns a locked, referenced pointer to the head request - which after 451 * this call is guaranteed to be the only request associated with the page. 452 * Returns NULL if no requests are found for @page, or a ERR_PTR if an 453 * error was encountered. 454 */ 455 static struct nfs_page * 456 nfs_lock_and_join_requests(struct page *page) 457 { 458 struct inode *inode = page_file_mapping(page)->host; 459 struct nfs_page *head, *subreq; 460 struct nfs_page *destroy_list = NULL; 461 unsigned int total_bytes; 462 int ret; 463 464 try_again: 465 /* 466 * A reference is taken only on the head request which acts as a 467 * reference to the whole page group - the group will not be destroyed 468 * until the head reference is released. 469 */ 470 head = nfs_page_find_head_request(page); 471 if (!head) 472 return NULL; 473 474 /* lock the page head first in order to avoid an ABBA inefficiency */ 475 if (!nfs_lock_request(head)) { 476 ret = nfs_wait_on_request(head); 477 nfs_release_request(head); 478 if (ret < 0) 479 return ERR_PTR(ret); 480 goto try_again; 481 } 482 483 /* Ensure that nobody removed the request before we locked it */ 484 if (head != nfs_page_private_request(page) && !PageSwapCache(page)) { 485 nfs_unlock_and_release_request(head); 486 goto try_again; 487 } 488 489 ret = nfs_page_group_lock(head); 490 if (ret < 0) 491 goto release_request; 492 493 /* lock each request in the page group */ 494 total_bytes = head->wb_bytes; 495 for (subreq = head->wb_this_page; subreq != head; 496 subreq = subreq->wb_this_page) { 497 498 if (!kref_get_unless_zero(&subreq->wb_kref)) { 499 if (subreq->wb_offset == head->wb_offset + total_bytes) 500 total_bytes += subreq->wb_bytes; 501 continue; 502 } 503 504 while (!nfs_lock_request(subreq)) { 505 /* 506 * Unlock page to allow nfs_page_group_sync_on_bit() 507 * to succeed 508 */ 509 nfs_page_group_unlock(head); 510 ret = nfs_wait_on_request(subreq); 511 if (!ret) 512 ret = nfs_page_group_lock(head); 513 if (ret < 0) { 514 nfs_unroll_locks(inode, head, subreq); 515 nfs_release_request(subreq); 516 goto release_request; 517 } 518 } 519 /* 520 * Subrequests are always contiguous, non overlapping 521 * and in order - but may be repeated (mirrored writes). 522 */ 523 if (subreq->wb_offset == (head->wb_offset + total_bytes)) { 524 /* keep track of how many bytes this group covers */ 525 total_bytes += subreq->wb_bytes; 526 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset || 527 ((subreq->wb_offset + subreq->wb_bytes) > 528 (head->wb_offset + total_bytes)))) { 529 nfs_page_group_unlock(head); 530 nfs_unroll_locks(inode, head, subreq); 531 nfs_unlock_and_release_request(subreq); 532 ret = -EIO; 533 goto release_request; 534 } 535 } 536 537 /* Now that all requests are locked, make sure they aren't on any list. 538 * Commit list removal accounting is done after locks are dropped */ 539 subreq = head; 540 do { 541 nfs_clear_request_commit(subreq); 542 subreq = subreq->wb_this_page; 543 } while (subreq != head); 544 545 /* unlink subrequests from head, destroy them later */ 546 if (head->wb_this_page != head) { 547 /* destroy list will be terminated by head */ 548 destroy_list = head->wb_this_page; 549 head->wb_this_page = head; 550 551 /* change head request to cover whole range that 552 * the former page group covered */ 553 head->wb_bytes = total_bytes; 554 } 555 556 /* Postpone destruction of this request */ 557 if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) { 558 set_bit(PG_INODE_REF, &head->wb_flags); 559 kref_get(&head->wb_kref); 560 atomic_long_inc(&NFS_I(inode)->nrequests); 561 } 562 563 nfs_page_group_unlock(head); 564 565 nfs_destroy_unlinked_subrequests(destroy_list, head, inode); 566 567 /* Did we lose a race with nfs_inode_remove_request()? */ 568 if (!(PagePrivate(page) || PageSwapCache(page))) { 569 nfs_unlock_and_release_request(head); 570 return NULL; 571 } 572 573 /* still holds ref on head from nfs_page_find_head_request 574 * and still has lock on head from lock loop */ 575 return head; 576 577 release_request: 578 nfs_unlock_and_release_request(head); 579 return ERR_PTR(ret); 580 } 581 582 static void nfs_write_error_remove_page(struct nfs_page *req) 583 { 584 nfs_end_page_writeback(req); 585 generic_error_remove_page(page_file_mapping(req->wb_page), 586 req->wb_page); 587 nfs_release_request(req); 588 } 589 590 static bool 591 nfs_error_is_fatal_on_server(int err) 592 { 593 switch (err) { 594 case 0: 595 case -ERESTARTSYS: 596 case -EINTR: 597 return false; 598 } 599 return nfs_error_is_fatal(err); 600 } 601 602 /* 603 * Find an associated nfs write request, and prepare to flush it out 604 * May return an error if the user signalled nfs_wait_on_request(). 605 */ 606 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio, 607 struct page *page) 608 { 609 struct nfs_page *req; 610 int ret = 0; 611 612 req = nfs_lock_and_join_requests(page); 613 if (!req) 614 goto out; 615 ret = PTR_ERR(req); 616 if (IS_ERR(req)) 617 goto out; 618 619 nfs_set_page_writeback(page); 620 WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags)); 621 622 ret = 0; 623 /* If there is a fatal error that covers this write, just exit */ 624 if (nfs_error_is_fatal_on_server(req->wb_context->error)) 625 goto out_launder; 626 627 if (!nfs_pageio_add_request(pgio, req)) { 628 ret = pgio->pg_error; 629 /* 630 * Remove the problematic req upon fatal errors on the server 631 */ 632 if (nfs_error_is_fatal(ret)) { 633 nfs_context_set_write_error(req->wb_context, ret); 634 if (nfs_error_is_fatal_on_server(ret)) 635 goto out_launder; 636 } 637 nfs_redirty_request(req); 638 ret = -EAGAIN; 639 } else 640 nfs_add_stats(page_file_mapping(page)->host, 641 NFSIOS_WRITEPAGES, 1); 642 out: 643 return ret; 644 out_launder: 645 nfs_write_error_remove_page(req); 646 return ret; 647 } 648 649 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, 650 struct nfs_pageio_descriptor *pgio) 651 { 652 int ret; 653 654 nfs_pageio_cond_complete(pgio, page_index(page)); 655 ret = nfs_page_async_flush(pgio, page); 656 if (ret == -EAGAIN) { 657 redirty_page_for_writepage(wbc, page); 658 ret = 0; 659 } 660 return ret; 661 } 662 663 /* 664 * Write an mmapped page to the server. 665 */ 666 static int nfs_writepage_locked(struct page *page, 667 struct writeback_control *wbc) 668 { 669 struct nfs_pageio_descriptor pgio; 670 struct inode *inode = page_file_mapping(page)->host; 671 int err; 672 673 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); 674 nfs_pageio_init_write(&pgio, inode, 0, 675 false, &nfs_async_write_completion_ops); 676 err = nfs_do_writepage(page, wbc, &pgio); 677 nfs_pageio_complete(&pgio); 678 if (err < 0) 679 return err; 680 if (pgio.pg_error < 0) 681 return pgio.pg_error; 682 return 0; 683 } 684 685 int nfs_writepage(struct page *page, struct writeback_control *wbc) 686 { 687 int ret; 688 689 ret = nfs_writepage_locked(page, wbc); 690 unlock_page(page); 691 return ret; 692 } 693 694 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data) 695 { 696 int ret; 697 698 ret = nfs_do_writepage(page, wbc, data); 699 unlock_page(page); 700 return ret; 701 } 702 703 static void nfs_io_completion_commit(void *inode) 704 { 705 nfs_commit_inode(inode, 0); 706 } 707 708 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) 709 { 710 struct inode *inode = mapping->host; 711 struct nfs_pageio_descriptor pgio; 712 struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS); 713 int err; 714 715 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); 716 717 if (ioc) 718 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode); 719 720 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false, 721 &nfs_async_write_completion_ops); 722 pgio.pg_io_completion = ioc; 723 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio); 724 nfs_pageio_complete(&pgio); 725 nfs_io_completion_put(ioc); 726 727 if (err < 0) 728 goto out_err; 729 err = pgio.pg_error; 730 if (err < 0) 731 goto out_err; 732 return 0; 733 out_err: 734 return err; 735 } 736 737 /* 738 * Insert a write request into an inode 739 */ 740 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req) 741 { 742 struct address_space *mapping = page_file_mapping(req->wb_page); 743 struct nfs_inode *nfsi = NFS_I(inode); 744 745 WARN_ON_ONCE(req->wb_this_page != req); 746 747 /* Lock the request! */ 748 nfs_lock_request(req); 749 750 /* 751 * Swap-space should not get truncated. Hence no need to plug the race 752 * with invalidate/truncate. 753 */ 754 spin_lock(&mapping->private_lock); 755 if (!nfs_have_writebacks(inode) && 756 NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) { 757 spin_lock(&inode->i_lock); 758 inode->i_version++; 759 spin_unlock(&inode->i_lock); 760 } 761 if (likely(!PageSwapCache(req->wb_page))) { 762 set_bit(PG_MAPPED, &req->wb_flags); 763 SetPagePrivate(req->wb_page); 764 set_page_private(req->wb_page, (unsigned long)req); 765 } 766 spin_unlock(&mapping->private_lock); 767 atomic_long_inc(&nfsi->nrequests); 768 /* this a head request for a page group - mark it as having an 769 * extra reference so sub groups can follow suit. 770 * This flag also informs pgio layer when to bump nrequests when 771 * adding subrequests. */ 772 WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags)); 773 kref_get(&req->wb_kref); 774 } 775 776 /* 777 * Remove a write request from an inode 778 */ 779 static void nfs_inode_remove_request(struct nfs_page *req) 780 { 781 struct address_space *mapping = page_file_mapping(req->wb_page); 782 struct inode *inode = mapping->host; 783 struct nfs_inode *nfsi = NFS_I(inode); 784 struct nfs_page *head; 785 786 atomic_long_dec(&nfsi->nrequests); 787 if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) { 788 head = req->wb_head; 789 790 spin_lock(&mapping->private_lock); 791 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) { 792 set_page_private(head->wb_page, 0); 793 ClearPagePrivate(head->wb_page); 794 clear_bit(PG_MAPPED, &head->wb_flags); 795 } 796 spin_unlock(&mapping->private_lock); 797 } 798 799 if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) 800 nfs_release_request(req); 801 } 802 803 static void 804 nfs_mark_request_dirty(struct nfs_page *req) 805 { 806 if (req->wb_page) 807 __set_page_dirty_nobuffers(req->wb_page); 808 } 809 810 /* 811 * nfs_page_search_commits_for_head_request_locked 812 * 813 * Search through commit lists on @inode for the head request for @page. 814 * Must be called while holding the inode (which is cinfo) lock. 815 * 816 * Returns the head request if found, or NULL if not found. 817 */ 818 static struct nfs_page * 819 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi, 820 struct page *page) 821 { 822 struct nfs_page *freq, *t; 823 struct nfs_commit_info cinfo; 824 struct inode *inode = &nfsi->vfs_inode; 825 826 nfs_init_cinfo_from_inode(&cinfo, inode); 827 828 /* search through pnfs commit lists */ 829 freq = pnfs_search_commit_reqs(inode, &cinfo, page); 830 if (freq) 831 return freq->wb_head; 832 833 /* Linearly search the commit list for the correct request */ 834 list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) { 835 if (freq->wb_page == page) 836 return freq->wb_head; 837 } 838 839 return NULL; 840 } 841 842 /** 843 * nfs_request_add_commit_list_locked - add request to a commit list 844 * @req: pointer to a struct nfs_page 845 * @dst: commit list head 846 * @cinfo: holds list lock and accounting info 847 * 848 * This sets the PG_CLEAN bit, updates the cinfo count of 849 * number of outstanding requests requiring a commit as well as 850 * the MM page stats. 851 * 852 * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the 853 * nfs_page lock. 854 */ 855 void 856 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst, 857 struct nfs_commit_info *cinfo) 858 { 859 set_bit(PG_CLEAN, &req->wb_flags); 860 nfs_list_add_request(req, dst); 861 atomic_long_inc(&cinfo->mds->ncommit); 862 } 863 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked); 864 865 /** 866 * nfs_request_add_commit_list - add request to a commit list 867 * @req: pointer to a struct nfs_page 868 * @dst: commit list head 869 * @cinfo: holds list lock and accounting info 870 * 871 * This sets the PG_CLEAN bit, updates the cinfo count of 872 * number of outstanding requests requiring a commit as well as 873 * the MM page stats. 874 * 875 * The caller must _not_ hold the cinfo->lock, but must be 876 * holding the nfs_page lock. 877 */ 878 void 879 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo) 880 { 881 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 882 nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo); 883 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 884 if (req->wb_page) 885 nfs_mark_page_unstable(req->wb_page, cinfo); 886 } 887 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list); 888 889 /** 890 * nfs_request_remove_commit_list - Remove request from a commit list 891 * @req: pointer to a nfs_page 892 * @cinfo: holds list lock and accounting info 893 * 894 * This clears the PG_CLEAN bit, and updates the cinfo's count of 895 * number of outstanding requests requiring a commit 896 * It does not update the MM page stats. 897 * 898 * The caller _must_ hold the cinfo->lock and the nfs_page lock. 899 */ 900 void 901 nfs_request_remove_commit_list(struct nfs_page *req, 902 struct nfs_commit_info *cinfo) 903 { 904 if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) 905 return; 906 nfs_list_remove_request(req); 907 atomic_long_dec(&cinfo->mds->ncommit); 908 } 909 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list); 910 911 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo, 912 struct inode *inode) 913 { 914 cinfo->inode = inode; 915 cinfo->mds = &NFS_I(inode)->commit_info; 916 cinfo->ds = pnfs_get_ds_info(inode); 917 cinfo->dreq = NULL; 918 cinfo->completion_ops = &nfs_commit_completion_ops; 919 } 920 921 void nfs_init_cinfo(struct nfs_commit_info *cinfo, 922 struct inode *inode, 923 struct nfs_direct_req *dreq) 924 { 925 if (dreq) 926 nfs_init_cinfo_from_dreq(cinfo, dreq); 927 else 928 nfs_init_cinfo_from_inode(cinfo, inode); 929 } 930 EXPORT_SYMBOL_GPL(nfs_init_cinfo); 931 932 /* 933 * Add a request to the inode's commit list. 934 */ 935 void 936 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg, 937 struct nfs_commit_info *cinfo, u32 ds_commit_idx) 938 { 939 if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx)) 940 return; 941 nfs_request_add_commit_list(req, cinfo); 942 } 943 944 static void 945 nfs_clear_page_commit(struct page *page) 946 { 947 dec_node_page_state(page, NR_UNSTABLE_NFS); 948 dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb, 949 WB_RECLAIMABLE); 950 } 951 952 /* Called holding the request lock on @req */ 953 static void 954 nfs_clear_request_commit(struct nfs_page *req) 955 { 956 if (test_bit(PG_CLEAN, &req->wb_flags)) { 957 struct inode *inode = d_inode(req->wb_context->dentry); 958 struct nfs_commit_info cinfo; 959 960 nfs_init_cinfo_from_inode(&cinfo, inode); 961 mutex_lock(&NFS_I(inode)->commit_mutex); 962 if (!pnfs_clear_request_commit(req, &cinfo)) { 963 nfs_request_remove_commit_list(req, &cinfo); 964 } 965 mutex_unlock(&NFS_I(inode)->commit_mutex); 966 nfs_clear_page_commit(req->wb_page); 967 } 968 } 969 970 int nfs_write_need_commit(struct nfs_pgio_header *hdr) 971 { 972 if (hdr->verf.committed == NFS_DATA_SYNC) 973 return hdr->lseg == NULL; 974 return hdr->verf.committed != NFS_FILE_SYNC; 975 } 976 977 static void nfs_async_write_init(struct nfs_pgio_header *hdr) 978 { 979 nfs_io_completion_get(hdr->io_completion); 980 } 981 982 static void nfs_write_completion(struct nfs_pgio_header *hdr) 983 { 984 struct nfs_commit_info cinfo; 985 unsigned long bytes = 0; 986 987 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) 988 goto out; 989 nfs_init_cinfo_from_inode(&cinfo, hdr->inode); 990 while (!list_empty(&hdr->pages)) { 991 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 992 993 bytes += req->wb_bytes; 994 nfs_list_remove_request(req); 995 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && 996 (hdr->good_bytes < bytes)) { 997 nfs_set_pageerror(req->wb_page); 998 nfs_context_set_write_error(req->wb_context, hdr->error); 999 goto remove_req; 1000 } 1001 if (nfs_write_need_commit(hdr)) { 1002 memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf)); 1003 nfs_mark_request_commit(req, hdr->lseg, &cinfo, 1004 hdr->pgio_mirror_idx); 1005 goto next; 1006 } 1007 remove_req: 1008 nfs_inode_remove_request(req); 1009 next: 1010 nfs_end_page_writeback(req); 1011 nfs_release_request(req); 1012 } 1013 out: 1014 nfs_io_completion_put(hdr->io_completion); 1015 hdr->release(hdr); 1016 } 1017 1018 unsigned long 1019 nfs_reqs_to_commit(struct nfs_commit_info *cinfo) 1020 { 1021 return atomic_long_read(&cinfo->mds->ncommit); 1022 } 1023 1024 /* NFS_I(cinfo->inode)->commit_mutex held by caller */ 1025 int 1026 nfs_scan_commit_list(struct list_head *src, struct list_head *dst, 1027 struct nfs_commit_info *cinfo, int max) 1028 { 1029 struct nfs_page *req, *tmp; 1030 int ret = 0; 1031 1032 restart: 1033 list_for_each_entry_safe(req, tmp, src, wb_list) { 1034 kref_get(&req->wb_kref); 1035 if (!nfs_lock_request(req)) { 1036 int status; 1037 1038 /* Prevent deadlock with nfs_lock_and_join_requests */ 1039 if (!list_empty(dst)) { 1040 nfs_release_request(req); 1041 continue; 1042 } 1043 /* Ensure we make progress to prevent livelock */ 1044 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 1045 status = nfs_wait_on_request(req); 1046 nfs_release_request(req); 1047 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 1048 if (status < 0) 1049 break; 1050 goto restart; 1051 } 1052 nfs_request_remove_commit_list(req, cinfo); 1053 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags); 1054 nfs_list_add_request(req, dst); 1055 ret++; 1056 if ((ret == max) && !cinfo->dreq) 1057 break; 1058 cond_resched(); 1059 } 1060 return ret; 1061 } 1062 EXPORT_SYMBOL_GPL(nfs_scan_commit_list); 1063 1064 /* 1065 * nfs_scan_commit - Scan an inode for commit requests 1066 * @inode: NFS inode to scan 1067 * @dst: mds destination list 1068 * @cinfo: mds and ds lists of reqs ready to commit 1069 * 1070 * Moves requests from the inode's 'commit' request list. 1071 * The requests are *not* checked to ensure that they form a contiguous set. 1072 */ 1073 int 1074 nfs_scan_commit(struct inode *inode, struct list_head *dst, 1075 struct nfs_commit_info *cinfo) 1076 { 1077 int ret = 0; 1078 1079 if (!atomic_long_read(&cinfo->mds->ncommit)) 1080 return 0; 1081 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 1082 if (atomic_long_read(&cinfo->mds->ncommit) > 0) { 1083 const int max = INT_MAX; 1084 1085 ret = nfs_scan_commit_list(&cinfo->mds->list, dst, 1086 cinfo, max); 1087 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret); 1088 } 1089 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 1090 return ret; 1091 } 1092 1093 /* 1094 * Search for an existing write request, and attempt to update 1095 * it to reflect a new dirty region on a given page. 1096 * 1097 * If the attempt fails, then the existing request is flushed out 1098 * to disk. 1099 */ 1100 static struct nfs_page *nfs_try_to_update_request(struct inode *inode, 1101 struct page *page, 1102 unsigned int offset, 1103 unsigned int bytes) 1104 { 1105 struct nfs_page *req; 1106 unsigned int rqend; 1107 unsigned int end; 1108 int error; 1109 1110 end = offset + bytes; 1111 1112 req = nfs_lock_and_join_requests(page); 1113 if (IS_ERR_OR_NULL(req)) 1114 return req; 1115 1116 rqend = req->wb_offset + req->wb_bytes; 1117 /* 1118 * Tell the caller to flush out the request if 1119 * the offsets are non-contiguous. 1120 * Note: nfs_flush_incompatible() will already 1121 * have flushed out requests having wrong owners. 1122 */ 1123 if (offset > rqend || end < req->wb_offset) 1124 goto out_flushme; 1125 1126 /* Okay, the request matches. Update the region */ 1127 if (offset < req->wb_offset) { 1128 req->wb_offset = offset; 1129 req->wb_pgbase = offset; 1130 } 1131 if (end > rqend) 1132 req->wb_bytes = end - req->wb_offset; 1133 else 1134 req->wb_bytes = rqend - req->wb_offset; 1135 return req; 1136 out_flushme: 1137 /* 1138 * Note: we mark the request dirty here because 1139 * nfs_lock_and_join_requests() cannot preserve 1140 * commit flags, so we have to replay the write. 1141 */ 1142 nfs_mark_request_dirty(req); 1143 nfs_unlock_and_release_request(req); 1144 error = nfs_wb_page(inode, page); 1145 return (error < 0) ? ERR_PTR(error) : NULL; 1146 } 1147 1148 /* 1149 * Try to update an existing write request, or create one if there is none. 1150 * 1151 * Note: Should always be called with the Page Lock held to prevent races 1152 * if we have to add a new request. Also assumes that the caller has 1153 * already called nfs_flush_incompatible() if necessary. 1154 */ 1155 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx, 1156 struct page *page, unsigned int offset, unsigned int bytes) 1157 { 1158 struct inode *inode = page_file_mapping(page)->host; 1159 struct nfs_page *req; 1160 1161 req = nfs_try_to_update_request(inode, page, offset, bytes); 1162 if (req != NULL) 1163 goto out; 1164 req = nfs_create_request(ctx, page, NULL, offset, bytes); 1165 if (IS_ERR(req)) 1166 goto out; 1167 nfs_inode_add_request(inode, req); 1168 out: 1169 return req; 1170 } 1171 1172 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page, 1173 unsigned int offset, unsigned int count) 1174 { 1175 struct nfs_page *req; 1176 1177 req = nfs_setup_write_request(ctx, page, offset, count); 1178 if (IS_ERR(req)) 1179 return PTR_ERR(req); 1180 /* Update file length */ 1181 nfs_grow_file(page, offset, count); 1182 nfs_mark_uptodate(req); 1183 nfs_mark_request_dirty(req); 1184 nfs_unlock_and_release_request(req); 1185 return 0; 1186 } 1187 1188 int nfs_flush_incompatible(struct file *file, struct page *page) 1189 { 1190 struct nfs_open_context *ctx = nfs_file_open_context(file); 1191 struct nfs_lock_context *l_ctx; 1192 struct file_lock_context *flctx = file_inode(file)->i_flctx; 1193 struct nfs_page *req; 1194 int do_flush, status; 1195 /* 1196 * Look for a request corresponding to this page. If there 1197 * is one, and it belongs to another file, we flush it out 1198 * before we try to copy anything into the page. Do this 1199 * due to the lack of an ACCESS-type call in NFSv2. 1200 * Also do the same if we find a request from an existing 1201 * dropped page. 1202 */ 1203 do { 1204 req = nfs_page_find_head_request(page); 1205 if (req == NULL) 1206 return 0; 1207 l_ctx = req->wb_lock_context; 1208 do_flush = req->wb_page != page || 1209 !nfs_match_open_context(req->wb_context, ctx); 1210 if (l_ctx && flctx && 1211 !(list_empty_careful(&flctx->flc_posix) && 1212 list_empty_careful(&flctx->flc_flock))) { 1213 do_flush |= l_ctx->lockowner != current->files; 1214 } 1215 nfs_release_request(req); 1216 if (!do_flush) 1217 return 0; 1218 status = nfs_wb_page(page_file_mapping(page)->host, page); 1219 } while (status == 0); 1220 return status; 1221 } 1222 1223 /* 1224 * Avoid buffered writes when a open context credential's key would 1225 * expire soon. 1226 * 1227 * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL. 1228 * 1229 * Return 0 and set a credential flag which triggers the inode to flush 1230 * and performs NFS_FILE_SYNC writes if the key will expired within 1231 * RPC_KEY_EXPIRE_TIMEO. 1232 */ 1233 int 1234 nfs_key_timeout_notify(struct file *filp, struct inode *inode) 1235 { 1236 struct nfs_open_context *ctx = nfs_file_open_context(filp); 1237 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; 1238 1239 return rpcauth_key_timeout_notify(auth, ctx->cred); 1240 } 1241 1242 /* 1243 * Test if the open context credential key is marked to expire soon. 1244 */ 1245 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode) 1246 { 1247 struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth; 1248 1249 return rpcauth_cred_key_to_expire(auth, ctx->cred); 1250 } 1251 1252 /* 1253 * If the page cache is marked as unsafe or invalid, then we can't rely on 1254 * the PageUptodate() flag. In this case, we will need to turn off 1255 * write optimisations that depend on the page contents being correct. 1256 */ 1257 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode) 1258 { 1259 struct nfs_inode *nfsi = NFS_I(inode); 1260 1261 if (nfs_have_delegated_attributes(inode)) 1262 goto out; 1263 if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE) 1264 return false; 1265 smp_rmb(); 1266 if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags)) 1267 return false; 1268 out: 1269 if (nfsi->cache_validity & NFS_INO_INVALID_DATA) 1270 return false; 1271 return PageUptodate(page) != 0; 1272 } 1273 1274 static bool 1275 is_whole_file_wrlock(struct file_lock *fl) 1276 { 1277 return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX && 1278 fl->fl_type == F_WRLCK; 1279 } 1280 1281 /* If we know the page is up to date, and we're not using byte range locks (or 1282 * if we have the whole file locked for writing), it may be more efficient to 1283 * extend the write to cover the entire page in order to avoid fragmentation 1284 * inefficiencies. 1285 * 1286 * If the file is opened for synchronous writes then we can just skip the rest 1287 * of the checks. 1288 */ 1289 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode) 1290 { 1291 int ret; 1292 struct file_lock_context *flctx = inode->i_flctx; 1293 struct file_lock *fl; 1294 1295 if (file->f_flags & O_DSYNC) 1296 return 0; 1297 if (!nfs_write_pageuptodate(page, inode)) 1298 return 0; 1299 if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE)) 1300 return 1; 1301 if (!flctx || (list_empty_careful(&flctx->flc_flock) && 1302 list_empty_careful(&flctx->flc_posix))) 1303 return 1; 1304 1305 /* Check to see if there are whole file write locks */ 1306 ret = 0; 1307 spin_lock(&flctx->flc_lock); 1308 if (!list_empty(&flctx->flc_posix)) { 1309 fl = list_first_entry(&flctx->flc_posix, struct file_lock, 1310 fl_list); 1311 if (is_whole_file_wrlock(fl)) 1312 ret = 1; 1313 } else if (!list_empty(&flctx->flc_flock)) { 1314 fl = list_first_entry(&flctx->flc_flock, struct file_lock, 1315 fl_list); 1316 if (fl->fl_type == F_WRLCK) 1317 ret = 1; 1318 } 1319 spin_unlock(&flctx->flc_lock); 1320 return ret; 1321 } 1322 1323 /* 1324 * Update and possibly write a cached page of an NFS file. 1325 * 1326 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad 1327 * things with a page scheduled for an RPC call (e.g. invalidate it). 1328 */ 1329 int nfs_updatepage(struct file *file, struct page *page, 1330 unsigned int offset, unsigned int count) 1331 { 1332 struct nfs_open_context *ctx = nfs_file_open_context(file); 1333 struct inode *inode = page_file_mapping(page)->host; 1334 int status = 0; 1335 1336 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); 1337 1338 dprintk("NFS: nfs_updatepage(%pD2 %d@%lld)\n", 1339 file, count, (long long)(page_file_offset(page) + offset)); 1340 1341 if (!count) 1342 goto out; 1343 1344 if (nfs_can_extend_write(file, page, inode)) { 1345 count = max(count + offset, nfs_page_length(page)); 1346 offset = 0; 1347 } 1348 1349 status = nfs_writepage_setup(ctx, page, offset, count); 1350 if (status < 0) 1351 nfs_set_pageerror(page); 1352 else 1353 __set_page_dirty_nobuffers(page); 1354 out: 1355 dprintk("NFS: nfs_updatepage returns %d (isize %lld)\n", 1356 status, (long long)i_size_read(inode)); 1357 return status; 1358 } 1359 1360 static int flush_task_priority(int how) 1361 { 1362 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { 1363 case FLUSH_HIGHPRI: 1364 return RPC_PRIORITY_HIGH; 1365 case FLUSH_LOWPRI: 1366 return RPC_PRIORITY_LOW; 1367 } 1368 return RPC_PRIORITY_NORMAL; 1369 } 1370 1371 static void nfs_initiate_write(struct nfs_pgio_header *hdr, 1372 struct rpc_message *msg, 1373 const struct nfs_rpc_ops *rpc_ops, 1374 struct rpc_task_setup *task_setup_data, int how) 1375 { 1376 int priority = flush_task_priority(how); 1377 1378 task_setup_data->priority = priority; 1379 rpc_ops->write_setup(hdr, msg); 1380 trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes, 1381 hdr->args.stable); 1382 1383 nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client, 1384 &task_setup_data->rpc_client, msg, hdr); 1385 } 1386 1387 /* If a nfs_flush_* function fails, it should remove reqs from @head and 1388 * call this on each, which will prepare them to be retried on next 1389 * writeback using standard nfs. 1390 */ 1391 static void nfs_redirty_request(struct nfs_page *req) 1392 { 1393 nfs_mark_request_dirty(req); 1394 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags); 1395 nfs_end_page_writeback(req); 1396 nfs_release_request(req); 1397 } 1398 1399 static void nfs_async_write_error(struct list_head *head) 1400 { 1401 struct nfs_page *req; 1402 1403 while (!list_empty(head)) { 1404 req = nfs_list_entry(head->next); 1405 nfs_list_remove_request(req); 1406 nfs_redirty_request(req); 1407 } 1408 } 1409 1410 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr) 1411 { 1412 nfs_async_write_error(&hdr->pages); 1413 } 1414 1415 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = { 1416 .init_hdr = nfs_async_write_init, 1417 .error_cleanup = nfs_async_write_error, 1418 .completion = nfs_write_completion, 1419 .reschedule_io = nfs_async_write_reschedule_io, 1420 }; 1421 1422 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, 1423 struct inode *inode, int ioflags, bool force_mds, 1424 const struct nfs_pgio_completion_ops *compl_ops) 1425 { 1426 struct nfs_server *server = NFS_SERVER(inode); 1427 const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops; 1428 1429 #ifdef CONFIG_NFS_V4_1 1430 if (server->pnfs_curr_ld && !force_mds) 1431 pg_ops = server->pnfs_curr_ld->pg_write_ops; 1432 #endif 1433 nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops, 1434 server->wsize, ioflags); 1435 } 1436 EXPORT_SYMBOL_GPL(nfs_pageio_init_write); 1437 1438 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio) 1439 { 1440 struct nfs_pgio_mirror *mirror; 1441 1442 if (pgio->pg_ops && pgio->pg_ops->pg_cleanup) 1443 pgio->pg_ops->pg_cleanup(pgio); 1444 1445 pgio->pg_ops = &nfs_pgio_rw_ops; 1446 1447 nfs_pageio_stop_mirroring(pgio); 1448 1449 mirror = &pgio->pg_mirrors[0]; 1450 mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize; 1451 } 1452 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds); 1453 1454 1455 void nfs_commit_prepare(struct rpc_task *task, void *calldata) 1456 { 1457 struct nfs_commit_data *data = calldata; 1458 1459 NFS_PROTO(data->inode)->commit_rpc_prepare(task, data); 1460 } 1461 1462 /* 1463 * Special version of should_remove_suid() that ignores capabilities. 1464 */ 1465 static int nfs_should_remove_suid(const struct inode *inode) 1466 { 1467 umode_t mode = inode->i_mode; 1468 int kill = 0; 1469 1470 /* suid always must be killed */ 1471 if (unlikely(mode & S_ISUID)) 1472 kill = ATTR_KILL_SUID; 1473 1474 /* 1475 * sgid without any exec bits is just a mandatory locking mark; leave 1476 * it alone. If some exec bits are set, it's a real sgid; kill it. 1477 */ 1478 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP))) 1479 kill |= ATTR_KILL_SGID; 1480 1481 if (unlikely(kill && S_ISREG(mode))) 1482 return kill; 1483 1484 return 0; 1485 } 1486 1487 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr, 1488 struct nfs_fattr *fattr) 1489 { 1490 struct nfs_pgio_args *argp = &hdr->args; 1491 struct nfs_pgio_res *resp = &hdr->res; 1492 u64 size = argp->offset + resp->count; 1493 1494 if (!(fattr->valid & NFS_ATTR_FATTR_SIZE)) 1495 fattr->size = size; 1496 if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) { 1497 fattr->valid &= ~NFS_ATTR_FATTR_SIZE; 1498 return; 1499 } 1500 if (size != fattr->size) 1501 return; 1502 /* Set attribute barrier */ 1503 nfs_fattr_set_barrier(fattr); 1504 /* ...and update size */ 1505 fattr->valid |= NFS_ATTR_FATTR_SIZE; 1506 } 1507 1508 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr) 1509 { 1510 struct nfs_fattr *fattr = &hdr->fattr; 1511 struct inode *inode = hdr->inode; 1512 1513 spin_lock(&inode->i_lock); 1514 nfs_writeback_check_extend(hdr, fattr); 1515 nfs_post_op_update_inode_force_wcc_locked(inode, fattr); 1516 spin_unlock(&inode->i_lock); 1517 } 1518 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode); 1519 1520 /* 1521 * This function is called when the WRITE call is complete. 1522 */ 1523 static int nfs_writeback_done(struct rpc_task *task, 1524 struct nfs_pgio_header *hdr, 1525 struct inode *inode) 1526 { 1527 int status; 1528 1529 /* 1530 * ->write_done will attempt to use post-op attributes to detect 1531 * conflicting writes by other clients. A strict interpretation 1532 * of close-to-open would allow us to continue caching even if 1533 * another writer had changed the file, but some applications 1534 * depend on tighter cache coherency when writing. 1535 */ 1536 status = NFS_PROTO(inode)->write_done(task, hdr); 1537 if (status != 0) 1538 return status; 1539 1540 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count); 1541 trace_nfs_writeback_done(inode, task->tk_status, 1542 hdr->args.offset, hdr->res.verf); 1543 1544 if (hdr->res.verf->committed < hdr->args.stable && 1545 task->tk_status >= 0) { 1546 /* We tried a write call, but the server did not 1547 * commit data to stable storage even though we 1548 * requested it. 1549 * Note: There is a known bug in Tru64 < 5.0 in which 1550 * the server reports NFS_DATA_SYNC, but performs 1551 * NFS_FILE_SYNC. We therefore implement this checking 1552 * as a dprintk() in order to avoid filling syslog. 1553 */ 1554 static unsigned long complain; 1555 1556 /* Note this will print the MDS for a DS write */ 1557 if (time_before(complain, jiffies)) { 1558 dprintk("NFS: faulty NFS server %s:" 1559 " (committed = %d) != (stable = %d)\n", 1560 NFS_SERVER(inode)->nfs_client->cl_hostname, 1561 hdr->res.verf->committed, hdr->args.stable); 1562 complain = jiffies + 300 * HZ; 1563 } 1564 } 1565 1566 /* Deal with the suid/sgid bit corner case */ 1567 if (nfs_should_remove_suid(inode)) 1568 nfs_mark_for_revalidate(inode); 1569 return 0; 1570 } 1571 1572 /* 1573 * This function is called when the WRITE call is complete. 1574 */ 1575 static void nfs_writeback_result(struct rpc_task *task, 1576 struct nfs_pgio_header *hdr) 1577 { 1578 struct nfs_pgio_args *argp = &hdr->args; 1579 struct nfs_pgio_res *resp = &hdr->res; 1580 1581 if (resp->count < argp->count) { 1582 static unsigned long complain; 1583 1584 /* This a short write! */ 1585 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE); 1586 1587 /* Has the server at least made some progress? */ 1588 if (resp->count == 0) { 1589 if (time_before(complain, jiffies)) { 1590 printk(KERN_WARNING 1591 "NFS: Server wrote zero bytes, expected %u.\n", 1592 argp->count); 1593 complain = jiffies + 300 * HZ; 1594 } 1595 nfs_set_pgio_error(hdr, -EIO, argp->offset); 1596 task->tk_status = -EIO; 1597 return; 1598 } 1599 1600 /* For non rpc-based layout drivers, retry-through-MDS */ 1601 if (!task->tk_ops) { 1602 hdr->pnfs_error = -EAGAIN; 1603 return; 1604 } 1605 1606 /* Was this an NFSv2 write or an NFSv3 stable write? */ 1607 if (resp->verf->committed != NFS_UNSTABLE) { 1608 /* Resend from where the server left off */ 1609 hdr->mds_offset += resp->count; 1610 argp->offset += resp->count; 1611 argp->pgbase += resp->count; 1612 argp->count -= resp->count; 1613 } else { 1614 /* Resend as a stable write in order to avoid 1615 * headaches in the case of a server crash. 1616 */ 1617 argp->stable = NFS_FILE_SYNC; 1618 } 1619 rpc_restart_call_prepare(task); 1620 } 1621 } 1622 1623 static int wait_on_commit(struct nfs_mds_commit_info *cinfo) 1624 { 1625 return wait_on_atomic_t(&cinfo->rpcs_out, 1626 nfs_wait_atomic_killable, TASK_KILLABLE); 1627 } 1628 1629 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo) 1630 { 1631 atomic_inc(&cinfo->rpcs_out); 1632 } 1633 1634 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo) 1635 { 1636 if (atomic_dec_and_test(&cinfo->rpcs_out)) 1637 wake_up_atomic_t(&cinfo->rpcs_out); 1638 } 1639 1640 void nfs_commitdata_release(struct nfs_commit_data *data) 1641 { 1642 put_nfs_open_context(data->context); 1643 nfs_commit_free(data); 1644 } 1645 EXPORT_SYMBOL_GPL(nfs_commitdata_release); 1646 1647 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data, 1648 const struct nfs_rpc_ops *nfs_ops, 1649 const struct rpc_call_ops *call_ops, 1650 int how, int flags) 1651 { 1652 struct rpc_task *task; 1653 int priority = flush_task_priority(how); 1654 struct rpc_message msg = { 1655 .rpc_argp = &data->args, 1656 .rpc_resp = &data->res, 1657 .rpc_cred = data->cred, 1658 }; 1659 struct rpc_task_setup task_setup_data = { 1660 .task = &data->task, 1661 .rpc_client = clnt, 1662 .rpc_message = &msg, 1663 .callback_ops = call_ops, 1664 .callback_data = data, 1665 .workqueue = nfsiod_workqueue, 1666 .flags = RPC_TASK_ASYNC | flags, 1667 .priority = priority, 1668 }; 1669 /* Set up the initial task struct. */ 1670 nfs_ops->commit_setup(data, &msg); 1671 trace_nfs_initiate_commit(data); 1672 1673 dprintk("NFS: initiated commit call\n"); 1674 1675 nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client, 1676 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg); 1677 1678 task = rpc_run_task(&task_setup_data); 1679 if (IS_ERR(task)) 1680 return PTR_ERR(task); 1681 if (how & FLUSH_SYNC) 1682 rpc_wait_for_completion_task(task); 1683 rpc_put_task(task); 1684 return 0; 1685 } 1686 EXPORT_SYMBOL_GPL(nfs_initiate_commit); 1687 1688 static loff_t nfs_get_lwb(struct list_head *head) 1689 { 1690 loff_t lwb = 0; 1691 struct nfs_page *req; 1692 1693 list_for_each_entry(req, head, wb_list) 1694 if (lwb < (req_offset(req) + req->wb_bytes)) 1695 lwb = req_offset(req) + req->wb_bytes; 1696 1697 return lwb; 1698 } 1699 1700 /* 1701 * Set up the argument/result storage required for the RPC call. 1702 */ 1703 void nfs_init_commit(struct nfs_commit_data *data, 1704 struct list_head *head, 1705 struct pnfs_layout_segment *lseg, 1706 struct nfs_commit_info *cinfo) 1707 { 1708 struct nfs_page *first = nfs_list_entry(head->next); 1709 struct inode *inode = d_inode(first->wb_context->dentry); 1710 1711 /* Set up the RPC argument and reply structs 1712 * NB: take care not to mess about with data->commit et al. */ 1713 1714 list_splice_init(head, &data->pages); 1715 1716 data->inode = inode; 1717 data->cred = first->wb_context->cred; 1718 data->lseg = lseg; /* reference transferred */ 1719 /* only set lwb for pnfs commit */ 1720 if (lseg) 1721 data->lwb = nfs_get_lwb(&data->pages); 1722 data->mds_ops = &nfs_commit_ops; 1723 data->completion_ops = cinfo->completion_ops; 1724 data->dreq = cinfo->dreq; 1725 1726 data->args.fh = NFS_FH(data->inode); 1727 /* Note: we always request a commit of the entire inode */ 1728 data->args.offset = 0; 1729 data->args.count = 0; 1730 data->context = get_nfs_open_context(first->wb_context); 1731 data->res.fattr = &data->fattr; 1732 data->res.verf = &data->verf; 1733 nfs_fattr_init(&data->fattr); 1734 } 1735 EXPORT_SYMBOL_GPL(nfs_init_commit); 1736 1737 void nfs_retry_commit(struct list_head *page_list, 1738 struct pnfs_layout_segment *lseg, 1739 struct nfs_commit_info *cinfo, 1740 u32 ds_commit_idx) 1741 { 1742 struct nfs_page *req; 1743 1744 while (!list_empty(page_list)) { 1745 req = nfs_list_entry(page_list->next); 1746 nfs_list_remove_request(req); 1747 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx); 1748 if (!cinfo->dreq) 1749 nfs_clear_page_commit(req->wb_page); 1750 nfs_unlock_and_release_request(req); 1751 } 1752 } 1753 EXPORT_SYMBOL_GPL(nfs_retry_commit); 1754 1755 static void 1756 nfs_commit_resched_write(struct nfs_commit_info *cinfo, 1757 struct nfs_page *req) 1758 { 1759 __set_page_dirty_nobuffers(req->wb_page); 1760 } 1761 1762 /* 1763 * Commit dirty pages 1764 */ 1765 static int 1766 nfs_commit_list(struct inode *inode, struct list_head *head, int how, 1767 struct nfs_commit_info *cinfo) 1768 { 1769 struct nfs_commit_data *data; 1770 1771 /* another commit raced with us */ 1772 if (list_empty(head)) 1773 return 0; 1774 1775 data = nfs_commitdata_alloc(true); 1776 1777 /* Set up the argument struct */ 1778 nfs_init_commit(data, head, NULL, cinfo); 1779 atomic_inc(&cinfo->mds->rpcs_out); 1780 return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode), 1781 data->mds_ops, how, 0); 1782 } 1783 1784 /* 1785 * COMMIT call returned 1786 */ 1787 static void nfs_commit_done(struct rpc_task *task, void *calldata) 1788 { 1789 struct nfs_commit_data *data = calldata; 1790 1791 dprintk("NFS: %5u nfs_commit_done (status %d)\n", 1792 task->tk_pid, task->tk_status); 1793 1794 /* Call the NFS version-specific code */ 1795 NFS_PROTO(data->inode)->commit_done(task, data); 1796 trace_nfs_commit_done(data); 1797 } 1798 1799 static void nfs_commit_release_pages(struct nfs_commit_data *data) 1800 { 1801 struct nfs_page *req; 1802 int status = data->task.tk_status; 1803 struct nfs_commit_info cinfo; 1804 struct nfs_server *nfss; 1805 1806 while (!list_empty(&data->pages)) { 1807 req = nfs_list_entry(data->pages.next); 1808 nfs_list_remove_request(req); 1809 if (req->wb_page) 1810 nfs_clear_page_commit(req->wb_page); 1811 1812 dprintk("NFS: commit (%s/%llu %d@%lld)", 1813 req->wb_context->dentry->d_sb->s_id, 1814 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)), 1815 req->wb_bytes, 1816 (long long)req_offset(req)); 1817 if (status < 0) { 1818 nfs_context_set_write_error(req->wb_context, status); 1819 if (req->wb_page) 1820 nfs_inode_remove_request(req); 1821 dprintk_cont(", error = %d\n", status); 1822 goto next; 1823 } 1824 1825 /* Okay, COMMIT succeeded, apparently. Check the verifier 1826 * returned by the server against all stored verfs. */ 1827 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) { 1828 /* We have a match */ 1829 if (req->wb_page) 1830 nfs_inode_remove_request(req); 1831 dprintk_cont(" OK\n"); 1832 goto next; 1833 } 1834 /* We have a mismatch. Write the page again */ 1835 dprintk_cont(" mismatch\n"); 1836 nfs_mark_request_dirty(req); 1837 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags); 1838 next: 1839 nfs_unlock_and_release_request(req); 1840 } 1841 nfss = NFS_SERVER(data->inode); 1842 if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) 1843 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC); 1844 1845 nfs_init_cinfo(&cinfo, data->inode, data->dreq); 1846 nfs_commit_end(cinfo.mds); 1847 } 1848 1849 static void nfs_commit_release(void *calldata) 1850 { 1851 struct nfs_commit_data *data = calldata; 1852 1853 data->completion_ops->completion(data); 1854 nfs_commitdata_release(calldata); 1855 } 1856 1857 static const struct rpc_call_ops nfs_commit_ops = { 1858 .rpc_call_prepare = nfs_commit_prepare, 1859 .rpc_call_done = nfs_commit_done, 1860 .rpc_release = nfs_commit_release, 1861 }; 1862 1863 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = { 1864 .completion = nfs_commit_release_pages, 1865 .resched_write = nfs_commit_resched_write, 1866 }; 1867 1868 int nfs_generic_commit_list(struct inode *inode, struct list_head *head, 1869 int how, struct nfs_commit_info *cinfo) 1870 { 1871 int status; 1872 1873 status = pnfs_commit_list(inode, head, how, cinfo); 1874 if (status == PNFS_NOT_ATTEMPTED) 1875 status = nfs_commit_list(inode, head, how, cinfo); 1876 return status; 1877 } 1878 1879 int nfs_commit_inode(struct inode *inode, int how) 1880 { 1881 LIST_HEAD(head); 1882 struct nfs_commit_info cinfo; 1883 int may_wait = how & FLUSH_SYNC; 1884 int error = 0; 1885 int res; 1886 1887 nfs_init_cinfo_from_inode(&cinfo, inode); 1888 nfs_commit_begin(cinfo.mds); 1889 res = nfs_scan_commit(inode, &head, &cinfo); 1890 if (res) 1891 error = nfs_generic_commit_list(inode, &head, how, &cinfo); 1892 nfs_commit_end(cinfo.mds); 1893 if (res == 0) 1894 return res; 1895 if (error < 0) 1896 goto out_error; 1897 if (!may_wait) 1898 goto out_mark_dirty; 1899 error = wait_on_commit(cinfo.mds); 1900 if (error < 0) 1901 return error; 1902 return res; 1903 out_error: 1904 res = error; 1905 /* Note: If we exit without ensuring that the commit is complete, 1906 * we must mark the inode as dirty. Otherwise, future calls to 1907 * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure 1908 * that the data is on the disk. 1909 */ 1910 out_mark_dirty: 1911 __mark_inode_dirty(inode, I_DIRTY_DATASYNC); 1912 return res; 1913 } 1914 EXPORT_SYMBOL_GPL(nfs_commit_inode); 1915 1916 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc) 1917 { 1918 struct nfs_inode *nfsi = NFS_I(inode); 1919 int flags = FLUSH_SYNC; 1920 int ret = 0; 1921 1922 /* no commits means nothing needs to be done */ 1923 if (!atomic_long_read(&nfsi->commit_info.ncommit)) 1924 return ret; 1925 1926 if (wbc->sync_mode == WB_SYNC_NONE) { 1927 /* Don't commit yet if this is a non-blocking flush and there 1928 * are a lot of outstanding writes for this mapping. 1929 */ 1930 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)) 1931 goto out_mark_dirty; 1932 1933 /* don't wait for the COMMIT response */ 1934 flags = 0; 1935 } 1936 1937 ret = nfs_commit_inode(inode, flags); 1938 if (ret >= 0) { 1939 if (wbc->sync_mode == WB_SYNC_NONE) { 1940 if (ret < wbc->nr_to_write) 1941 wbc->nr_to_write -= ret; 1942 else 1943 wbc->nr_to_write = 0; 1944 } 1945 return 0; 1946 } 1947 out_mark_dirty: 1948 __mark_inode_dirty(inode, I_DIRTY_DATASYNC); 1949 return ret; 1950 } 1951 EXPORT_SYMBOL_GPL(nfs_write_inode); 1952 1953 /* 1954 * Wrapper for filemap_write_and_wait_range() 1955 * 1956 * Needed for pNFS in order to ensure data becomes visible to the 1957 * client. 1958 */ 1959 int nfs_filemap_write_and_wait_range(struct address_space *mapping, 1960 loff_t lstart, loff_t lend) 1961 { 1962 int ret; 1963 1964 ret = filemap_write_and_wait_range(mapping, lstart, lend); 1965 if (ret == 0) 1966 ret = pnfs_sync_inode(mapping->host, true); 1967 return ret; 1968 } 1969 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range); 1970 1971 /* 1972 * flush the inode to disk. 1973 */ 1974 int nfs_wb_all(struct inode *inode) 1975 { 1976 int ret; 1977 1978 trace_nfs_writeback_inode_enter(inode); 1979 1980 ret = filemap_write_and_wait(inode->i_mapping); 1981 if (ret) 1982 goto out; 1983 ret = nfs_commit_inode(inode, FLUSH_SYNC); 1984 if (ret < 0) 1985 goto out; 1986 pnfs_sync_inode(inode, true); 1987 ret = 0; 1988 1989 out: 1990 trace_nfs_writeback_inode_exit(inode, ret); 1991 return ret; 1992 } 1993 EXPORT_SYMBOL_GPL(nfs_wb_all); 1994 1995 int nfs_wb_page_cancel(struct inode *inode, struct page *page) 1996 { 1997 struct nfs_page *req; 1998 int ret = 0; 1999 2000 wait_on_page_writeback(page); 2001 2002 /* blocking call to cancel all requests and join to a single (head) 2003 * request */ 2004 req = nfs_lock_and_join_requests(page); 2005 2006 if (IS_ERR(req)) { 2007 ret = PTR_ERR(req); 2008 } else if (req) { 2009 /* all requests from this page have been cancelled by 2010 * nfs_lock_and_join_requests, so just remove the head 2011 * request from the inode / page_private pointer and 2012 * release it */ 2013 nfs_inode_remove_request(req); 2014 nfs_unlock_and_release_request(req); 2015 } 2016 2017 return ret; 2018 } 2019 2020 /* 2021 * Write back all requests on one page - we do this before reading it. 2022 */ 2023 int nfs_wb_page(struct inode *inode, struct page *page) 2024 { 2025 loff_t range_start = page_file_offset(page); 2026 loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1); 2027 struct writeback_control wbc = { 2028 .sync_mode = WB_SYNC_ALL, 2029 .nr_to_write = 0, 2030 .range_start = range_start, 2031 .range_end = range_end, 2032 }; 2033 int ret; 2034 2035 trace_nfs_writeback_page_enter(inode); 2036 2037 for (;;) { 2038 wait_on_page_writeback(page); 2039 if (clear_page_dirty_for_io(page)) { 2040 ret = nfs_writepage_locked(page, &wbc); 2041 if (ret < 0) 2042 goto out_error; 2043 continue; 2044 } 2045 ret = 0; 2046 if (!PagePrivate(page)) 2047 break; 2048 ret = nfs_commit_inode(inode, FLUSH_SYNC); 2049 if (ret < 0) 2050 goto out_error; 2051 } 2052 out_error: 2053 trace_nfs_writeback_page_exit(inode, ret); 2054 return ret; 2055 } 2056 2057 #ifdef CONFIG_MIGRATION 2058 int nfs_migrate_page(struct address_space *mapping, struct page *newpage, 2059 struct page *page, enum migrate_mode mode) 2060 { 2061 /* 2062 * If PagePrivate is set, then the page is currently associated with 2063 * an in-progress read or write request. Don't try to migrate it. 2064 * 2065 * FIXME: we could do this in principle, but we'll need a way to ensure 2066 * that we can safely release the inode reference while holding 2067 * the page lock. 2068 */ 2069 if (PagePrivate(page)) 2070 return -EBUSY; 2071 2072 if (!nfs_fscache_release_page(page, GFP_KERNEL)) 2073 return -EBUSY; 2074 2075 return migrate_page(mapping, newpage, page, mode); 2076 } 2077 #endif 2078 2079 int __init nfs_init_writepagecache(void) 2080 { 2081 nfs_wdata_cachep = kmem_cache_create("nfs_write_data", 2082 sizeof(struct nfs_pgio_header), 2083 0, SLAB_HWCACHE_ALIGN, 2084 NULL); 2085 if (nfs_wdata_cachep == NULL) 2086 return -ENOMEM; 2087 2088 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, 2089 nfs_wdata_cachep); 2090 if (nfs_wdata_mempool == NULL) 2091 goto out_destroy_write_cache; 2092 2093 nfs_cdata_cachep = kmem_cache_create("nfs_commit_data", 2094 sizeof(struct nfs_commit_data), 2095 0, SLAB_HWCACHE_ALIGN, 2096 NULL); 2097 if (nfs_cdata_cachep == NULL) 2098 goto out_destroy_write_mempool; 2099 2100 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, 2101 nfs_cdata_cachep); 2102 if (nfs_commit_mempool == NULL) 2103 goto out_destroy_commit_cache; 2104 2105 /* 2106 * NFS congestion size, scale with available memory. 2107 * 2108 * 64MB: 8192k 2109 * 128MB: 11585k 2110 * 256MB: 16384k 2111 * 512MB: 23170k 2112 * 1GB: 32768k 2113 * 2GB: 46340k 2114 * 4GB: 65536k 2115 * 8GB: 92681k 2116 * 16GB: 131072k 2117 * 2118 * This allows larger machines to have larger/more transfers. 2119 * Limit the default to 256M 2120 */ 2121 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10); 2122 if (nfs_congestion_kb > 256*1024) 2123 nfs_congestion_kb = 256*1024; 2124 2125 return 0; 2126 2127 out_destroy_commit_cache: 2128 kmem_cache_destroy(nfs_cdata_cachep); 2129 out_destroy_write_mempool: 2130 mempool_destroy(nfs_wdata_mempool); 2131 out_destroy_write_cache: 2132 kmem_cache_destroy(nfs_wdata_cachep); 2133 return -ENOMEM; 2134 } 2135 2136 void nfs_destroy_writepagecache(void) 2137 { 2138 mempool_destroy(nfs_commit_mempool); 2139 kmem_cache_destroy(nfs_cdata_cachep); 2140 mempool_destroy(nfs_wdata_mempool); 2141 kmem_cache_destroy(nfs_wdata_cachep); 2142 } 2143 2144 static const struct nfs_rw_ops nfs_rw_write_ops = { 2145 .rw_alloc_header = nfs_writehdr_alloc, 2146 .rw_free_header = nfs_writehdr_free, 2147 .rw_done = nfs_writeback_done, 2148 .rw_result = nfs_writeback_result, 2149 .rw_initiate = nfs_initiate_write, 2150 }; 2151