Linux kernel mirror (for testing) git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel os linux
at v2.6.26-rc9 1594 lines 41 kB view raw
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 17#include <linux/sunrpc/clnt.h> 18#include <linux/nfs_fs.h> 19#include <linux/nfs_mount.h> 20#include <linux/nfs_page.h> 21#include <linux/backing-dev.h> 22 23#include <asm/uaccess.h> 24 25#include "delegation.h" 26#include "internal.h" 27#include "iostat.h" 28 29#define NFSDBG_FACILITY NFSDBG_PAGECACHE 30 31#define MIN_POOL_WRITE (32) 32#define MIN_POOL_COMMIT (4) 33 34/* 35 * Local function declarations 36 */ 37static struct nfs_page * nfs_update_request(struct nfs_open_context*, 38 struct page *, 39 unsigned int, unsigned int); 40static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc, 41 struct inode *inode, int ioflags); 42static void nfs_redirty_request(struct nfs_page *req); 43static const struct rpc_call_ops nfs_write_partial_ops; 44static const struct rpc_call_ops nfs_write_full_ops; 45static const struct rpc_call_ops nfs_commit_ops; 46 47static struct kmem_cache *nfs_wdata_cachep; 48static mempool_t *nfs_wdata_mempool; 49static mempool_t *nfs_commit_mempool; 50 51struct nfs_write_data *nfs_commitdata_alloc(void) 52{ 53 struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS); 54 55 if (p) { 56 memset(p, 0, sizeof(*p)); 57 INIT_LIST_HEAD(&p->pages); 58 } 59 return p; 60} 61 62void nfs_commit_free(struct nfs_write_data *p) 63{ 64 if (p && (p->pagevec != &p->page_array[0])) 65 kfree(p->pagevec); 66 mempool_free(p, nfs_commit_mempool); 67} 68 69struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount) 70{ 71 struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS); 72 73 if (p) { 74 memset(p, 0, sizeof(*p)); 75 INIT_LIST_HEAD(&p->pages); 76 p->npages = pagecount; 77 if (pagecount <= ARRAY_SIZE(p->page_array)) 78 p->pagevec = p->page_array; 79 else { 80 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS); 81 if (!p->pagevec) { 82 mempool_free(p, nfs_wdata_mempool); 83 p = NULL; 84 } 85 } 86 } 87 return p; 88} 89 90static void nfs_writedata_free(struct nfs_write_data *p) 91{ 92 if (p && (p->pagevec != &p->page_array[0])) 93 kfree(p->pagevec); 94 mempool_free(p, nfs_wdata_mempool); 95} 96 97void nfs_writedata_release(void *data) 98{ 99 struct nfs_write_data *wdata = data; 100 101 put_nfs_open_context(wdata->args.context); 102 nfs_writedata_free(wdata); 103} 104 105static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error) 106{ 107 ctx->error = error; 108 smp_wmb(); 109 set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags); 110} 111 112static struct nfs_page *nfs_page_find_request_locked(struct page *page) 113{ 114 struct nfs_page *req = NULL; 115 116 if (PagePrivate(page)) { 117 req = (struct nfs_page *)page_private(page); 118 if (req != NULL) 119 kref_get(&req->wb_kref); 120 } 121 return req; 122} 123 124static struct nfs_page *nfs_page_find_request(struct page *page) 125{ 126 struct inode *inode = page->mapping->host; 127 struct nfs_page *req = NULL; 128 129 spin_lock(&inode->i_lock); 130 req = nfs_page_find_request_locked(page); 131 spin_unlock(&inode->i_lock); 132 return req; 133} 134 135/* Adjust the file length if we're writing beyond the end */ 136static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) 137{ 138 struct inode *inode = page->mapping->host; 139 loff_t end, i_size = i_size_read(inode); 140 pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; 141 142 if (i_size > 0 && page->index < end_index) 143 return; 144 end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count); 145 if (i_size >= end) 146 return; 147 nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); 148 i_size_write(inode, end); 149} 150 151/* A writeback failed: mark the page as bad, and invalidate the page cache */ 152static void nfs_set_pageerror(struct page *page) 153{ 154 SetPageError(page); 155 nfs_zap_mapping(page->mapping->host, page->mapping); 156} 157 158/* We can set the PG_uptodate flag if we see that a write request 159 * covers the full page. 160 */ 161static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count) 162{ 163 if (PageUptodate(page)) 164 return; 165 if (base != 0) 166 return; 167 if (count != nfs_page_length(page)) 168 return; 169 SetPageUptodate(page); 170} 171 172static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page, 173 unsigned int offset, unsigned int count) 174{ 175 struct nfs_page *req; 176 int ret; 177 178 for (;;) { 179 req = nfs_update_request(ctx, page, offset, count); 180 if (!IS_ERR(req)) 181 break; 182 ret = PTR_ERR(req); 183 if (ret != -EBUSY) 184 return ret; 185 ret = nfs_wb_page(page->mapping->host, page); 186 if (ret != 0) 187 return ret; 188 } 189 /* Update file length */ 190 nfs_grow_file(page, offset, count); 191 nfs_clear_page_tag_locked(req); 192 return 0; 193} 194 195static int wb_priority(struct writeback_control *wbc) 196{ 197 if (wbc->for_reclaim) 198 return FLUSH_HIGHPRI | FLUSH_STABLE; 199 if (wbc->for_kupdate) 200 return FLUSH_LOWPRI; 201 return 0; 202} 203 204/* 205 * NFS congestion control 206 */ 207 208int nfs_congestion_kb; 209 210#define NFS_CONGESTION_ON_THRESH (nfs_congestion_kb >> (PAGE_SHIFT-10)) 211#define NFS_CONGESTION_OFF_THRESH \ 212 (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2)) 213 214static int nfs_set_page_writeback(struct page *page) 215{ 216 int ret = test_set_page_writeback(page); 217 218 if (!ret) { 219 struct inode *inode = page->mapping->host; 220 struct nfs_server *nfss = NFS_SERVER(inode); 221 222 if (atomic_long_inc_return(&nfss->writeback) > 223 NFS_CONGESTION_ON_THRESH) 224 set_bdi_congested(&nfss->backing_dev_info, WRITE); 225 } 226 return ret; 227} 228 229static void nfs_end_page_writeback(struct page *page) 230{ 231 struct inode *inode = page->mapping->host; 232 struct nfs_server *nfss = NFS_SERVER(inode); 233 234 end_page_writeback(page); 235 if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) 236 clear_bdi_congested(&nfss->backing_dev_info, WRITE); 237} 238 239/* 240 * Find an associated nfs write request, and prepare to flush it out 241 * May return an error if the user signalled nfs_wait_on_request(). 242 */ 243static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio, 244 struct page *page) 245{ 246 struct inode *inode = page->mapping->host; 247 struct nfs_page *req; 248 int ret; 249 250 spin_lock(&inode->i_lock); 251 for(;;) { 252 req = nfs_page_find_request_locked(page); 253 if (req == NULL) { 254 spin_unlock(&inode->i_lock); 255 return 0; 256 } 257 if (nfs_set_page_tag_locked(req)) 258 break; 259 /* Note: If we hold the page lock, as is the case in nfs_writepage, 260 * then the call to nfs_set_page_tag_locked() will always 261 * succeed provided that someone hasn't already marked the 262 * request as dirty (in which case we don't care). 263 */ 264 spin_unlock(&inode->i_lock); 265 ret = nfs_wait_on_request(req); 266 nfs_release_request(req); 267 if (ret != 0) 268 return ret; 269 spin_lock(&inode->i_lock); 270 } 271 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) { 272 /* This request is marked for commit */ 273 spin_unlock(&inode->i_lock); 274 nfs_clear_page_tag_locked(req); 275 nfs_pageio_complete(pgio); 276 return 0; 277 } 278 if (nfs_set_page_writeback(page) != 0) { 279 spin_unlock(&inode->i_lock); 280 BUG(); 281 } 282 spin_unlock(&inode->i_lock); 283 if (!nfs_pageio_add_request(pgio, req)) { 284 nfs_redirty_request(req); 285 return pgio->pg_error; 286 } 287 return 0; 288} 289 290static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio) 291{ 292 struct inode *inode = page->mapping->host; 293 294 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); 295 nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1); 296 297 nfs_pageio_cond_complete(pgio, page->index); 298 return nfs_page_async_flush(pgio, page); 299} 300 301/* 302 * Write an mmapped page to the server. 303 */ 304static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc) 305{ 306 struct nfs_pageio_descriptor pgio; 307 int err; 308 309 nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc)); 310 err = nfs_do_writepage(page, wbc, &pgio); 311 nfs_pageio_complete(&pgio); 312 if (err < 0) 313 return err; 314 if (pgio.pg_error < 0) 315 return pgio.pg_error; 316 return 0; 317} 318 319int nfs_writepage(struct page *page, struct writeback_control *wbc) 320{ 321 int ret; 322 323 ret = nfs_writepage_locked(page, wbc); 324 unlock_page(page); 325 return ret; 326} 327 328static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data) 329{ 330 int ret; 331 332 ret = nfs_do_writepage(page, wbc, data); 333 unlock_page(page); 334 return ret; 335} 336 337int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) 338{ 339 struct inode *inode = mapping->host; 340 struct nfs_pageio_descriptor pgio; 341 int err; 342 343 nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); 344 345 nfs_pageio_init_write(&pgio, inode, wb_priority(wbc)); 346 err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio); 347 nfs_pageio_complete(&pgio); 348 if (err < 0) 349 return err; 350 if (pgio.pg_error < 0) 351 return pgio.pg_error; 352 return 0; 353} 354 355/* 356 * Insert a write request into an inode 357 */ 358static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req) 359{ 360 struct nfs_inode *nfsi = NFS_I(inode); 361 int error; 362 363 error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req); 364 BUG_ON(error); 365 if (!nfsi->npages) { 366 igrab(inode); 367 if (nfs_have_delegation(inode, FMODE_WRITE)) 368 nfsi->change_attr++; 369 } 370 SetPagePrivate(req->wb_page); 371 set_page_private(req->wb_page, (unsigned long)req); 372 nfsi->npages++; 373 kref_get(&req->wb_kref); 374 radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, 375 NFS_PAGE_TAG_LOCKED); 376} 377 378/* 379 * Remove a write request from an inode 380 */ 381static void nfs_inode_remove_request(struct nfs_page *req) 382{ 383 struct inode *inode = req->wb_context->path.dentry->d_inode; 384 struct nfs_inode *nfsi = NFS_I(inode); 385 386 BUG_ON (!NFS_WBACK_BUSY(req)); 387 388 spin_lock(&inode->i_lock); 389 set_page_private(req->wb_page, 0); 390 ClearPagePrivate(req->wb_page); 391 radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index); 392 nfsi->npages--; 393 if (!nfsi->npages) { 394 spin_unlock(&inode->i_lock); 395 iput(inode); 396 } else 397 spin_unlock(&inode->i_lock); 398 nfs_clear_request(req); 399 nfs_release_request(req); 400} 401 402static void 403nfs_mark_request_dirty(struct nfs_page *req) 404{ 405 __set_page_dirty_nobuffers(req->wb_page); 406} 407 408/* 409 * Check if a request is dirty 410 */ 411static inline int 412nfs_dirty_request(struct nfs_page *req) 413{ 414 struct page *page = req->wb_page; 415 416 if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags)) 417 return 0; 418 return !PageWriteback(page); 419} 420 421#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) 422/* 423 * Add a request to the inode's commit list. 424 */ 425static void 426nfs_mark_request_commit(struct nfs_page *req) 427{ 428 struct inode *inode = req->wb_context->path.dentry->d_inode; 429 struct nfs_inode *nfsi = NFS_I(inode); 430 431 spin_lock(&inode->i_lock); 432 nfsi->ncommit++; 433 set_bit(PG_NEED_COMMIT, &(req)->wb_flags); 434 radix_tree_tag_set(&nfsi->nfs_page_tree, 435 req->wb_index, 436 NFS_PAGE_TAG_COMMIT); 437 spin_unlock(&inode->i_lock); 438 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 439 inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE); 440 __mark_inode_dirty(inode, I_DIRTY_DATASYNC); 441} 442 443static inline 444int nfs_write_need_commit(struct nfs_write_data *data) 445{ 446 return data->verf.committed != NFS_FILE_SYNC; 447} 448 449static inline 450int nfs_reschedule_unstable_write(struct nfs_page *req) 451{ 452 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) { 453 nfs_mark_request_commit(req); 454 return 1; 455 } 456 if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) { 457 nfs_mark_request_dirty(req); 458 return 1; 459 } 460 return 0; 461} 462#else 463static inline void 464nfs_mark_request_commit(struct nfs_page *req) 465{ 466} 467 468static inline 469int nfs_write_need_commit(struct nfs_write_data *data) 470{ 471 return 0; 472} 473 474static inline 475int nfs_reschedule_unstable_write(struct nfs_page *req) 476{ 477 return 0; 478} 479#endif 480 481/* 482 * Wait for a request to complete. 483 * 484 * Interruptible by fatal signals only. 485 */ 486static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages) 487{ 488 struct nfs_inode *nfsi = NFS_I(inode); 489 struct nfs_page *req; 490 pgoff_t idx_end, next; 491 unsigned int res = 0; 492 int error; 493 494 if (npages == 0) 495 idx_end = ~0; 496 else 497 idx_end = idx_start + npages - 1; 498 499 next = idx_start; 500 while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) { 501 if (req->wb_index > idx_end) 502 break; 503 504 next = req->wb_index + 1; 505 BUG_ON(!NFS_WBACK_BUSY(req)); 506 507 kref_get(&req->wb_kref); 508 spin_unlock(&inode->i_lock); 509 error = nfs_wait_on_request(req); 510 nfs_release_request(req); 511 spin_lock(&inode->i_lock); 512 if (error < 0) 513 return error; 514 res++; 515 } 516 return res; 517} 518 519static void nfs_cancel_commit_list(struct list_head *head) 520{ 521 struct nfs_page *req; 522 523 while(!list_empty(head)) { 524 req = nfs_list_entry(head->next); 525 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 526 dec_bdi_stat(req->wb_page->mapping->backing_dev_info, 527 BDI_RECLAIMABLE); 528 nfs_list_remove_request(req); 529 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags); 530 nfs_inode_remove_request(req); 531 nfs_unlock_request(req); 532 } 533} 534 535#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) 536/* 537 * nfs_scan_commit - Scan an inode for commit requests 538 * @inode: NFS inode to scan 539 * @dst: destination list 540 * @idx_start: lower bound of page->index to scan. 541 * @npages: idx_start + npages sets the upper bound to scan. 542 * 543 * Moves requests from the inode's 'commit' request list. 544 * The requests are *not* checked to ensure that they form a contiguous set. 545 */ 546static int 547nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages) 548{ 549 struct nfs_inode *nfsi = NFS_I(inode); 550 int res = 0; 551 552 if (nfsi->ncommit != 0) { 553 res = nfs_scan_list(nfsi, dst, idx_start, npages, 554 NFS_PAGE_TAG_COMMIT); 555 nfsi->ncommit -= res; 556 } 557 return res; 558} 559#else 560static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages) 561{ 562 return 0; 563} 564#endif 565 566/* 567 * Try to update any existing write request, or create one if there is none. 568 * In order to match, the request's credentials must match those of 569 * the calling process. 570 * 571 * Note: Should always be called with the Page Lock held! 572 */ 573static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx, 574 struct page *page, unsigned int offset, unsigned int bytes) 575{ 576 struct address_space *mapping = page->mapping; 577 struct inode *inode = mapping->host; 578 struct nfs_page *req, *new = NULL; 579 pgoff_t rqend, end; 580 581 end = offset + bytes; 582 583 for (;;) { 584 /* Loop over all inode entries and see if we find 585 * A request for the page we wish to update 586 */ 587 if (new) { 588 if (radix_tree_preload(GFP_NOFS)) { 589 nfs_release_request(new); 590 return ERR_PTR(-ENOMEM); 591 } 592 } 593 594 spin_lock(&inode->i_lock); 595 req = nfs_page_find_request_locked(page); 596 if (req) { 597 if (!nfs_set_page_tag_locked(req)) { 598 int error; 599 600 spin_unlock(&inode->i_lock); 601 error = nfs_wait_on_request(req); 602 nfs_release_request(req); 603 if (error < 0) { 604 if (new) { 605 radix_tree_preload_end(); 606 nfs_release_request(new); 607 } 608 return ERR_PTR(error); 609 } 610 continue; 611 } 612 spin_unlock(&inode->i_lock); 613 if (new) { 614 radix_tree_preload_end(); 615 nfs_release_request(new); 616 } 617 break; 618 } 619 620 if (new) { 621 nfs_lock_request_dontget(new); 622 nfs_inode_add_request(inode, new); 623 spin_unlock(&inode->i_lock); 624 radix_tree_preload_end(); 625 req = new; 626 goto zero_page; 627 } 628 spin_unlock(&inode->i_lock); 629 630 new = nfs_create_request(ctx, inode, page, offset, bytes); 631 if (IS_ERR(new)) 632 return new; 633 } 634 635 /* We have a request for our page. 636 * If the creds don't match, or the 637 * page addresses don't match, 638 * tell the caller to wait on the conflicting 639 * request. 640 */ 641 rqend = req->wb_offset + req->wb_bytes; 642 if (req->wb_context != ctx 643 || req->wb_page != page 644 || !nfs_dirty_request(req) 645 || offset > rqend || end < req->wb_offset) { 646 nfs_clear_page_tag_locked(req); 647 return ERR_PTR(-EBUSY); 648 } 649 650 /* Okay, the request matches. Update the region */ 651 if (offset < req->wb_offset) { 652 req->wb_offset = offset; 653 req->wb_pgbase = offset; 654 req->wb_bytes = max(end, rqend) - req->wb_offset; 655 goto zero_page; 656 } 657 658 if (end > rqend) 659 req->wb_bytes = end - req->wb_offset; 660 661 return req; 662zero_page: 663 /* If this page might potentially be marked as up to date, 664 * then we need to zero any uninitalised data. */ 665 if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE 666 && !PageUptodate(req->wb_page)) 667 zero_user_segment(req->wb_page, req->wb_bytes, PAGE_CACHE_SIZE); 668 return req; 669} 670 671int nfs_flush_incompatible(struct file *file, struct page *page) 672{ 673 struct nfs_open_context *ctx = nfs_file_open_context(file); 674 struct nfs_page *req; 675 int do_flush, status; 676 /* 677 * Look for a request corresponding to this page. If there 678 * is one, and it belongs to another file, we flush it out 679 * before we try to copy anything into the page. Do this 680 * due to the lack of an ACCESS-type call in NFSv2. 681 * Also do the same if we find a request from an existing 682 * dropped page. 683 */ 684 do { 685 req = nfs_page_find_request(page); 686 if (req == NULL) 687 return 0; 688 do_flush = req->wb_page != page || req->wb_context != ctx 689 || !nfs_dirty_request(req); 690 nfs_release_request(req); 691 if (!do_flush) 692 return 0; 693 status = nfs_wb_page(page->mapping->host, page); 694 } while (status == 0); 695 return status; 696} 697 698/* 699 * If the page cache is marked as unsafe or invalid, then we can't rely on 700 * the PageUptodate() flag. In this case, we will need to turn off 701 * write optimisations that depend on the page contents being correct. 702 */ 703static int nfs_write_pageuptodate(struct page *page, struct inode *inode) 704{ 705 return PageUptodate(page) && 706 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA)); 707} 708 709/* 710 * Update and possibly write a cached page of an NFS file. 711 * 712 * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad 713 * things with a page scheduled for an RPC call (e.g. invalidate it). 714 */ 715int nfs_updatepage(struct file *file, struct page *page, 716 unsigned int offset, unsigned int count) 717{ 718 struct nfs_open_context *ctx = nfs_file_open_context(file); 719 struct inode *inode = page->mapping->host; 720 int status = 0; 721 722 nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); 723 724 dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n", 725 file->f_path.dentry->d_parent->d_name.name, 726 file->f_path.dentry->d_name.name, count, 727 (long long)(page_offset(page) +offset)); 728 729 /* If we're not using byte range locks, and we know the page 730 * is up to date, it may be more efficient to extend the write 731 * to cover the entire page in order to avoid fragmentation 732 * inefficiencies. 733 */ 734 if (nfs_write_pageuptodate(page, inode) && 735 inode->i_flock == NULL && 736 !(file->f_flags & O_SYNC)) { 737 count = max(count + offset, nfs_page_length(page)); 738 offset = 0; 739 } 740 741 status = nfs_writepage_setup(ctx, page, offset, count); 742 if (status < 0) 743 nfs_set_pageerror(page); 744 else 745 __set_page_dirty_nobuffers(page); 746 747 dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n", 748 status, (long long)i_size_read(inode)); 749 return status; 750} 751 752static void nfs_writepage_release(struct nfs_page *req) 753{ 754 755 if (PageError(req->wb_page)) { 756 nfs_end_page_writeback(req->wb_page); 757 nfs_inode_remove_request(req); 758 } else if (!nfs_reschedule_unstable_write(req)) { 759 /* Set the PG_uptodate flag */ 760 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes); 761 nfs_end_page_writeback(req->wb_page); 762 nfs_inode_remove_request(req); 763 } else 764 nfs_end_page_writeback(req->wb_page); 765 nfs_clear_page_tag_locked(req); 766} 767 768static int flush_task_priority(int how) 769{ 770 switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { 771 case FLUSH_HIGHPRI: 772 return RPC_PRIORITY_HIGH; 773 case FLUSH_LOWPRI: 774 return RPC_PRIORITY_LOW; 775 } 776 return RPC_PRIORITY_NORMAL; 777} 778 779/* 780 * Set up the argument/result storage required for the RPC call. 781 */ 782static int nfs_write_rpcsetup(struct nfs_page *req, 783 struct nfs_write_data *data, 784 const struct rpc_call_ops *call_ops, 785 unsigned int count, unsigned int offset, 786 int how) 787{ 788 struct inode *inode = req->wb_context->path.dentry->d_inode; 789 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC; 790 int priority = flush_task_priority(how); 791 struct rpc_task *task; 792 struct rpc_message msg = { 793 .rpc_argp = &data->args, 794 .rpc_resp = &data->res, 795 .rpc_cred = req->wb_context->cred, 796 }; 797 struct rpc_task_setup task_setup_data = { 798 .rpc_client = NFS_CLIENT(inode), 799 .task = &data->task, 800 .rpc_message = &msg, 801 .callback_ops = call_ops, 802 .callback_data = data, 803 .workqueue = nfsiod_workqueue, 804 .flags = flags, 805 .priority = priority, 806 }; 807 808 /* Set up the RPC argument and reply structs 809 * NB: take care not to mess about with data->commit et al. */ 810 811 data->req = req; 812 data->inode = inode = req->wb_context->path.dentry->d_inode; 813 data->cred = msg.rpc_cred; 814 815 data->args.fh = NFS_FH(inode); 816 data->args.offset = req_offset(req) + offset; 817 data->args.pgbase = req->wb_pgbase + offset; 818 data->args.pages = data->pagevec; 819 data->args.count = count; 820 data->args.context = get_nfs_open_context(req->wb_context); 821 data->args.stable = NFS_UNSTABLE; 822 if (how & FLUSH_STABLE) { 823 data->args.stable = NFS_DATA_SYNC; 824 if (!NFS_I(inode)->ncommit) 825 data->args.stable = NFS_FILE_SYNC; 826 } 827 828 data->res.fattr = &data->fattr; 829 data->res.count = count; 830 data->res.verf = &data->verf; 831 nfs_fattr_init(&data->fattr); 832 833 /* Set up the initial task struct. */ 834 NFS_PROTO(inode)->write_setup(data, &msg); 835 836 dprintk("NFS: %5u initiated write call " 837 "(req %s/%Ld, %u bytes @ offset %Lu)\n", 838 data->task.tk_pid, 839 inode->i_sb->s_id, 840 (long long)NFS_FILEID(inode), 841 count, 842 (unsigned long long)data->args.offset); 843 844 task = rpc_run_task(&task_setup_data); 845 if (IS_ERR(task)) 846 return PTR_ERR(task); 847 rpc_put_task(task); 848 return 0; 849} 850 851/* If a nfs_flush_* function fails, it should remove reqs from @head and 852 * call this on each, which will prepare them to be retried on next 853 * writeback using standard nfs. 854 */ 855static void nfs_redirty_request(struct nfs_page *req) 856{ 857 nfs_mark_request_dirty(req); 858 nfs_end_page_writeback(req->wb_page); 859 nfs_clear_page_tag_locked(req); 860} 861 862/* 863 * Generate multiple small requests to write out a single 864 * contiguous dirty area on one page. 865 */ 866static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how) 867{ 868 struct nfs_page *req = nfs_list_entry(head->next); 869 struct page *page = req->wb_page; 870 struct nfs_write_data *data; 871 size_t wsize = NFS_SERVER(inode)->wsize, nbytes; 872 unsigned int offset; 873 int requests = 0; 874 int ret = 0; 875 LIST_HEAD(list); 876 877 nfs_list_remove_request(req); 878 879 nbytes = count; 880 do { 881 size_t len = min(nbytes, wsize); 882 883 data = nfs_writedata_alloc(1); 884 if (!data) 885 goto out_bad; 886 list_add(&data->pages, &list); 887 requests++; 888 nbytes -= len; 889 } while (nbytes != 0); 890 atomic_set(&req->wb_complete, requests); 891 892 ClearPageError(page); 893 offset = 0; 894 nbytes = count; 895 do { 896 int ret2; 897 898 data = list_entry(list.next, struct nfs_write_data, pages); 899 list_del_init(&data->pages); 900 901 data->pagevec[0] = page; 902 903 if (nbytes < wsize) 904 wsize = nbytes; 905 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops, 906 wsize, offset, how); 907 if (ret == 0) 908 ret = ret2; 909 offset += wsize; 910 nbytes -= wsize; 911 } while (nbytes != 0); 912 913 return ret; 914 915out_bad: 916 while (!list_empty(&list)) { 917 data = list_entry(list.next, struct nfs_write_data, pages); 918 list_del(&data->pages); 919 nfs_writedata_release(data); 920 } 921 nfs_redirty_request(req); 922 return -ENOMEM; 923} 924 925/* 926 * Create an RPC task for the given write request and kick it. 927 * The page must have been locked by the caller. 928 * 929 * It may happen that the page we're passed is not marked dirty. 930 * This is the case if nfs_updatepage detects a conflicting request 931 * that has been written but not committed. 932 */ 933static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how) 934{ 935 struct nfs_page *req; 936 struct page **pages; 937 struct nfs_write_data *data; 938 939 data = nfs_writedata_alloc(npages); 940 if (!data) 941 goto out_bad; 942 943 pages = data->pagevec; 944 while (!list_empty(head)) { 945 req = nfs_list_entry(head->next); 946 nfs_list_remove_request(req); 947 nfs_list_add_request(req, &data->pages); 948 ClearPageError(req->wb_page); 949 *pages++ = req->wb_page; 950 } 951 req = nfs_list_entry(data->pages.next); 952 953 /* Set up the argument struct */ 954 return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how); 955 out_bad: 956 while (!list_empty(head)) { 957 req = nfs_list_entry(head->next); 958 nfs_list_remove_request(req); 959 nfs_redirty_request(req); 960 } 961 return -ENOMEM; 962} 963 964static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, 965 struct inode *inode, int ioflags) 966{ 967 size_t wsize = NFS_SERVER(inode)->wsize; 968 969 if (wsize < PAGE_CACHE_SIZE) 970 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags); 971 else 972 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags); 973} 974 975/* 976 * Handle a write reply that flushed part of a page. 977 */ 978static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata) 979{ 980 struct nfs_write_data *data = calldata; 981 struct nfs_page *req = data->req; 982 983 dprintk("NFS: write (%s/%Ld %d@%Ld)", 984 req->wb_context->path.dentry->d_inode->i_sb->s_id, 985 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode), 986 req->wb_bytes, 987 (long long)req_offset(req)); 988 989 nfs_writeback_done(task, data); 990} 991 992static void nfs_writeback_release_partial(void *calldata) 993{ 994 struct nfs_write_data *data = calldata; 995 struct nfs_page *req = data->req; 996 struct page *page = req->wb_page; 997 int status = data->task.tk_status; 998 999 if (status < 0) { 1000 nfs_set_pageerror(page); 1001 nfs_context_set_write_error(req->wb_context, status); 1002 dprintk(", error = %d\n", status); 1003 goto out; 1004 } 1005 1006 if (nfs_write_need_commit(data)) { 1007 struct inode *inode = page->mapping->host; 1008 1009 spin_lock(&inode->i_lock); 1010 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) { 1011 /* Do nothing we need to resend the writes */ 1012 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) { 1013 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); 1014 dprintk(" defer commit\n"); 1015 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) { 1016 set_bit(PG_NEED_RESCHED, &req->wb_flags); 1017 clear_bit(PG_NEED_COMMIT, &req->wb_flags); 1018 dprintk(" server reboot detected\n"); 1019 } 1020 spin_unlock(&inode->i_lock); 1021 } else 1022 dprintk(" OK\n"); 1023 1024out: 1025 if (atomic_dec_and_test(&req->wb_complete)) 1026 nfs_writepage_release(req); 1027 nfs_writedata_release(calldata); 1028} 1029 1030static const struct rpc_call_ops nfs_write_partial_ops = { 1031 .rpc_call_done = nfs_writeback_done_partial, 1032 .rpc_release = nfs_writeback_release_partial, 1033}; 1034 1035/* 1036 * Handle a write reply that flushes a whole page. 1037 * 1038 * FIXME: There is an inherent race with invalidate_inode_pages and 1039 * writebacks since the page->count is kept > 1 for as long 1040 * as the page has a write request pending. 1041 */ 1042static void nfs_writeback_done_full(struct rpc_task *task, void *calldata) 1043{ 1044 struct nfs_write_data *data = calldata; 1045 1046 nfs_writeback_done(task, data); 1047} 1048 1049static void nfs_writeback_release_full(void *calldata) 1050{ 1051 struct nfs_write_data *data = calldata; 1052 int status = data->task.tk_status; 1053 1054 /* Update attributes as result of writeback. */ 1055 while (!list_empty(&data->pages)) { 1056 struct nfs_page *req = nfs_list_entry(data->pages.next); 1057 struct page *page = req->wb_page; 1058 1059 nfs_list_remove_request(req); 1060 1061 dprintk("NFS: write (%s/%Ld %d@%Ld)", 1062 req->wb_context->path.dentry->d_inode->i_sb->s_id, 1063 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode), 1064 req->wb_bytes, 1065 (long long)req_offset(req)); 1066 1067 if (status < 0) { 1068 nfs_set_pageerror(page); 1069 nfs_context_set_write_error(req->wb_context, status); 1070 dprintk(", error = %d\n", status); 1071 goto remove_request; 1072 } 1073 1074 if (nfs_write_need_commit(data)) { 1075 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); 1076 nfs_mark_request_commit(req); 1077 nfs_end_page_writeback(page); 1078 dprintk(" marked for commit\n"); 1079 goto next; 1080 } 1081 /* Set the PG_uptodate flag? */ 1082 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes); 1083 dprintk(" OK\n"); 1084remove_request: 1085 nfs_end_page_writeback(page); 1086 nfs_inode_remove_request(req); 1087 next: 1088 nfs_clear_page_tag_locked(req); 1089 } 1090 nfs_writedata_release(calldata); 1091} 1092 1093static const struct rpc_call_ops nfs_write_full_ops = { 1094 .rpc_call_done = nfs_writeback_done_full, 1095 .rpc_release = nfs_writeback_release_full, 1096}; 1097 1098 1099/* 1100 * This function is called when the WRITE call is complete. 1101 */ 1102int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data) 1103{ 1104 struct nfs_writeargs *argp = &data->args; 1105 struct nfs_writeres *resp = &data->res; 1106 int status; 1107 1108 dprintk("NFS: %5u nfs_writeback_done (status %d)\n", 1109 task->tk_pid, task->tk_status); 1110 1111 /* 1112 * ->write_done will attempt to use post-op attributes to detect 1113 * conflicting writes by other clients. A strict interpretation 1114 * of close-to-open would allow us to continue caching even if 1115 * another writer had changed the file, but some applications 1116 * depend on tighter cache coherency when writing. 1117 */ 1118 status = NFS_PROTO(data->inode)->write_done(task, data); 1119 if (status != 0) 1120 return status; 1121 nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count); 1122 1123#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) 1124 if (resp->verf->committed < argp->stable && task->tk_status >= 0) { 1125 /* We tried a write call, but the server did not 1126 * commit data to stable storage even though we 1127 * requested it. 1128 * Note: There is a known bug in Tru64 < 5.0 in which 1129 * the server reports NFS_DATA_SYNC, but performs 1130 * NFS_FILE_SYNC. We therefore implement this checking 1131 * as a dprintk() in order to avoid filling syslog. 1132 */ 1133 static unsigned long complain; 1134 1135 if (time_before(complain, jiffies)) { 1136 dprintk("NFS: faulty NFS server %s:" 1137 " (committed = %d) != (stable = %d)\n", 1138 NFS_SERVER(data->inode)->nfs_client->cl_hostname, 1139 resp->verf->committed, argp->stable); 1140 complain = jiffies + 300 * HZ; 1141 } 1142 } 1143#endif 1144 /* Is this a short write? */ 1145 if (task->tk_status >= 0 && resp->count < argp->count) { 1146 static unsigned long complain; 1147 1148 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE); 1149 1150 /* Has the server at least made some progress? */ 1151 if (resp->count != 0) { 1152 /* Was this an NFSv2 write or an NFSv3 stable write? */ 1153 if (resp->verf->committed != NFS_UNSTABLE) { 1154 /* Resend from where the server left off */ 1155 argp->offset += resp->count; 1156 argp->pgbase += resp->count; 1157 argp->count -= resp->count; 1158 } else { 1159 /* Resend as a stable write in order to avoid 1160 * headaches in the case of a server crash. 1161 */ 1162 argp->stable = NFS_FILE_SYNC; 1163 } 1164 rpc_restart_call(task); 1165 return -EAGAIN; 1166 } 1167 if (time_before(complain, jiffies)) { 1168 printk(KERN_WARNING 1169 "NFS: Server wrote zero bytes, expected %u.\n", 1170 argp->count); 1171 complain = jiffies + 300 * HZ; 1172 } 1173 /* Can't do anything about it except throw an error. */ 1174 task->tk_status = -EIO; 1175 } 1176 return 0; 1177} 1178 1179 1180#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) 1181void nfs_commitdata_release(void *data) 1182{ 1183 struct nfs_write_data *wdata = data; 1184 1185 put_nfs_open_context(wdata->args.context); 1186 nfs_commit_free(wdata); 1187} 1188 1189/* 1190 * Set up the argument/result storage required for the RPC call. 1191 */ 1192static int nfs_commit_rpcsetup(struct list_head *head, 1193 struct nfs_write_data *data, 1194 int how) 1195{ 1196 struct nfs_page *first = nfs_list_entry(head->next); 1197 struct inode *inode = first->wb_context->path.dentry->d_inode; 1198 int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC; 1199 int priority = flush_task_priority(how); 1200 struct rpc_task *task; 1201 struct rpc_message msg = { 1202 .rpc_argp = &data->args, 1203 .rpc_resp = &data->res, 1204 .rpc_cred = first->wb_context->cred, 1205 }; 1206 struct rpc_task_setup task_setup_data = { 1207 .task = &data->task, 1208 .rpc_client = NFS_CLIENT(inode), 1209 .rpc_message = &msg, 1210 .callback_ops = &nfs_commit_ops, 1211 .callback_data = data, 1212 .workqueue = nfsiod_workqueue, 1213 .flags = flags, 1214 .priority = priority, 1215 }; 1216 1217 /* Set up the RPC argument and reply structs 1218 * NB: take care not to mess about with data->commit et al. */ 1219 1220 list_splice_init(head, &data->pages); 1221 1222 data->inode = inode; 1223 data->cred = msg.rpc_cred; 1224 1225 data->args.fh = NFS_FH(data->inode); 1226 /* Note: we always request a commit of the entire inode */ 1227 data->args.offset = 0; 1228 data->args.count = 0; 1229 data->args.context = get_nfs_open_context(first->wb_context); 1230 data->res.count = 0; 1231 data->res.fattr = &data->fattr; 1232 data->res.verf = &data->verf; 1233 nfs_fattr_init(&data->fattr); 1234 1235 /* Set up the initial task struct. */ 1236 NFS_PROTO(inode)->commit_setup(data, &msg); 1237 1238 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); 1239 1240 task = rpc_run_task(&task_setup_data); 1241 if (IS_ERR(task)) 1242 return PTR_ERR(task); 1243 rpc_put_task(task); 1244 return 0; 1245} 1246 1247/* 1248 * Commit dirty pages 1249 */ 1250static int 1251nfs_commit_list(struct inode *inode, struct list_head *head, int how) 1252{ 1253 struct nfs_write_data *data; 1254 struct nfs_page *req; 1255 1256 data = nfs_commitdata_alloc(); 1257 1258 if (!data) 1259 goto out_bad; 1260 1261 /* Set up the argument struct */ 1262 return nfs_commit_rpcsetup(head, data, how); 1263 out_bad: 1264 while (!list_empty(head)) { 1265 req = nfs_list_entry(head->next); 1266 nfs_list_remove_request(req); 1267 nfs_mark_request_commit(req); 1268 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 1269 dec_bdi_stat(req->wb_page->mapping->backing_dev_info, 1270 BDI_RECLAIMABLE); 1271 nfs_clear_page_tag_locked(req); 1272 } 1273 return -ENOMEM; 1274} 1275 1276/* 1277 * COMMIT call returned 1278 */ 1279static void nfs_commit_done(struct rpc_task *task, void *calldata) 1280{ 1281 struct nfs_write_data *data = calldata; 1282 1283 dprintk("NFS: %5u nfs_commit_done (status %d)\n", 1284 task->tk_pid, task->tk_status); 1285 1286 /* Call the NFS version-specific code */ 1287 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0) 1288 return; 1289} 1290 1291static void nfs_commit_release(void *calldata) 1292{ 1293 struct nfs_write_data *data = calldata; 1294 struct nfs_page *req; 1295 int status = data->task.tk_status; 1296 1297 while (!list_empty(&data->pages)) { 1298 req = nfs_list_entry(data->pages.next); 1299 nfs_list_remove_request(req); 1300 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags); 1301 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS); 1302 dec_bdi_stat(req->wb_page->mapping->backing_dev_info, 1303 BDI_RECLAIMABLE); 1304 1305 dprintk("NFS: commit (%s/%Ld %d@%Ld)", 1306 req->wb_context->path.dentry->d_inode->i_sb->s_id, 1307 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode), 1308 req->wb_bytes, 1309 (long long)req_offset(req)); 1310 if (status < 0) { 1311 nfs_context_set_write_error(req->wb_context, status); 1312 nfs_inode_remove_request(req); 1313 dprintk(", error = %d\n", status); 1314 goto next; 1315 } 1316 1317 /* Okay, COMMIT succeeded, apparently. Check the verifier 1318 * returned by the server against all stored verfs. */ 1319 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) { 1320 /* We have a match */ 1321 /* Set the PG_uptodate flag */ 1322 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, 1323 req->wb_bytes); 1324 nfs_inode_remove_request(req); 1325 dprintk(" OK\n"); 1326 goto next; 1327 } 1328 /* We have a mismatch. Write the page again */ 1329 dprintk(" mismatch\n"); 1330 nfs_mark_request_dirty(req); 1331 next: 1332 nfs_clear_page_tag_locked(req); 1333 } 1334 nfs_commitdata_release(calldata); 1335} 1336 1337static const struct rpc_call_ops nfs_commit_ops = { 1338 .rpc_call_done = nfs_commit_done, 1339 .rpc_release = nfs_commit_release, 1340}; 1341 1342int nfs_commit_inode(struct inode *inode, int how) 1343{ 1344 LIST_HEAD(head); 1345 int res; 1346 1347 spin_lock(&inode->i_lock); 1348 res = nfs_scan_commit(inode, &head, 0, 0); 1349 spin_unlock(&inode->i_lock); 1350 if (res) { 1351 int error = nfs_commit_list(inode, &head, how); 1352 if (error < 0) 1353 return error; 1354 } 1355 return res; 1356} 1357#else 1358static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how) 1359{ 1360 return 0; 1361} 1362#endif 1363 1364long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how) 1365{ 1366 struct inode *inode = mapping->host; 1367 pgoff_t idx_start, idx_end; 1368 unsigned int npages = 0; 1369 LIST_HEAD(head); 1370 int nocommit = how & FLUSH_NOCOMMIT; 1371 long pages, ret; 1372 1373 /* FIXME */ 1374 if (wbc->range_cyclic) 1375 idx_start = 0; 1376 else { 1377 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT; 1378 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT; 1379 if (idx_end > idx_start) { 1380 pgoff_t l_npages = 1 + idx_end - idx_start; 1381 npages = l_npages; 1382 if (sizeof(npages) != sizeof(l_npages) && 1383 (pgoff_t)npages != l_npages) 1384 npages = 0; 1385 } 1386 } 1387 how &= ~FLUSH_NOCOMMIT; 1388 spin_lock(&inode->i_lock); 1389 do { 1390 ret = nfs_wait_on_requests_locked(inode, idx_start, npages); 1391 if (ret != 0) 1392 continue; 1393 if (nocommit) 1394 break; 1395 pages = nfs_scan_commit(inode, &head, idx_start, npages); 1396 if (pages == 0) 1397 break; 1398 if (how & FLUSH_INVALIDATE) { 1399 spin_unlock(&inode->i_lock); 1400 nfs_cancel_commit_list(&head); 1401 ret = pages; 1402 spin_lock(&inode->i_lock); 1403 continue; 1404 } 1405 pages += nfs_scan_commit(inode, &head, 0, 0); 1406 spin_unlock(&inode->i_lock); 1407 ret = nfs_commit_list(inode, &head, how); 1408 spin_lock(&inode->i_lock); 1409 1410 } while (ret >= 0); 1411 spin_unlock(&inode->i_lock); 1412 return ret; 1413} 1414 1415static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how) 1416{ 1417 int ret; 1418 1419 ret = nfs_writepages(mapping, wbc); 1420 if (ret < 0) 1421 goto out; 1422 ret = nfs_sync_mapping_wait(mapping, wbc, how); 1423 if (ret < 0) 1424 goto out; 1425 return 0; 1426out: 1427 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 1428 return ret; 1429} 1430 1431/* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */ 1432static int nfs_write_mapping(struct address_space *mapping, int how) 1433{ 1434 struct writeback_control wbc = { 1435 .bdi = mapping->backing_dev_info, 1436 .sync_mode = WB_SYNC_NONE, 1437 .nr_to_write = LONG_MAX, 1438 .for_writepages = 1, 1439 .range_cyclic = 1, 1440 }; 1441 int ret; 1442 1443 ret = __nfs_write_mapping(mapping, &wbc, how); 1444 if (ret < 0) 1445 return ret; 1446 wbc.sync_mode = WB_SYNC_ALL; 1447 return __nfs_write_mapping(mapping, &wbc, how); 1448} 1449 1450/* 1451 * flush the inode to disk. 1452 */ 1453int nfs_wb_all(struct inode *inode) 1454{ 1455 return nfs_write_mapping(inode->i_mapping, 0); 1456} 1457 1458int nfs_wb_nocommit(struct inode *inode) 1459{ 1460 return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT); 1461} 1462 1463int nfs_wb_page_cancel(struct inode *inode, struct page *page) 1464{ 1465 struct nfs_page *req; 1466 loff_t range_start = page_offset(page); 1467 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1); 1468 struct writeback_control wbc = { 1469 .bdi = page->mapping->backing_dev_info, 1470 .sync_mode = WB_SYNC_ALL, 1471 .nr_to_write = LONG_MAX, 1472 .range_start = range_start, 1473 .range_end = range_end, 1474 }; 1475 int ret = 0; 1476 1477 BUG_ON(!PageLocked(page)); 1478 for (;;) { 1479 req = nfs_page_find_request(page); 1480 if (req == NULL) 1481 goto out; 1482 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) { 1483 nfs_release_request(req); 1484 break; 1485 } 1486 if (nfs_lock_request_dontget(req)) { 1487 nfs_inode_remove_request(req); 1488 /* 1489 * In case nfs_inode_remove_request has marked the 1490 * page as being dirty 1491 */ 1492 cancel_dirty_page(page, PAGE_CACHE_SIZE); 1493 nfs_unlock_request(req); 1494 break; 1495 } 1496 ret = nfs_wait_on_request(req); 1497 if (ret < 0) 1498 goto out; 1499 } 1500 if (!PagePrivate(page)) 1501 return 0; 1502 ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE); 1503out: 1504 return ret; 1505} 1506 1507static int nfs_wb_page_priority(struct inode *inode, struct page *page, 1508 int how) 1509{ 1510 loff_t range_start = page_offset(page); 1511 loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1); 1512 struct writeback_control wbc = { 1513 .bdi = page->mapping->backing_dev_info, 1514 .sync_mode = WB_SYNC_ALL, 1515 .nr_to_write = LONG_MAX, 1516 .range_start = range_start, 1517 .range_end = range_end, 1518 }; 1519 int ret; 1520 1521 do { 1522 if (clear_page_dirty_for_io(page)) { 1523 ret = nfs_writepage_locked(page, &wbc); 1524 if (ret < 0) 1525 goto out_error; 1526 } else if (!PagePrivate(page)) 1527 break; 1528 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how); 1529 if (ret < 0) 1530 goto out_error; 1531 } while (PagePrivate(page)); 1532 return 0; 1533out_error: 1534 __mark_inode_dirty(inode, I_DIRTY_PAGES); 1535 return ret; 1536} 1537 1538/* 1539 * Write back all requests on one page - we do this before reading it. 1540 */ 1541int nfs_wb_page(struct inode *inode, struct page* page) 1542{ 1543 return nfs_wb_page_priority(inode, page, FLUSH_STABLE); 1544} 1545 1546int __init nfs_init_writepagecache(void) 1547{ 1548 nfs_wdata_cachep = kmem_cache_create("nfs_write_data", 1549 sizeof(struct nfs_write_data), 1550 0, SLAB_HWCACHE_ALIGN, 1551 NULL); 1552 if (nfs_wdata_cachep == NULL) 1553 return -ENOMEM; 1554 1555 nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE, 1556 nfs_wdata_cachep); 1557 if (nfs_wdata_mempool == NULL) 1558 return -ENOMEM; 1559 1560 nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT, 1561 nfs_wdata_cachep); 1562 if (nfs_commit_mempool == NULL) 1563 return -ENOMEM; 1564 1565 /* 1566 * NFS congestion size, scale with available memory. 1567 * 1568 * 64MB: 8192k 1569 * 128MB: 11585k 1570 * 256MB: 16384k 1571 * 512MB: 23170k 1572 * 1GB: 32768k 1573 * 2GB: 46340k 1574 * 4GB: 65536k 1575 * 8GB: 92681k 1576 * 16GB: 131072k 1577 * 1578 * This allows larger machines to have larger/more transfers. 1579 * Limit the default to 256M 1580 */ 1581 nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10); 1582 if (nfs_congestion_kb > 256*1024) 1583 nfs_congestion_kb = 256*1024; 1584 1585 return 0; 1586} 1587 1588void nfs_destroy_writepagecache(void) 1589{ 1590 mempool_destroy(nfs_commit_mempool); 1591 mempool_destroy(nfs_wdata_mempool); 1592 kmem_cache_destroy(nfs_wdata_cachep); 1593} 1594