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