fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #include "nfstrace.h"
  35
  36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  37
  38 #define MIN_POOL_WRITE          (32)
  39 #define MIN_POOL_COMMIT         (4)
  40
  41 /*
  42  * Local function declarations
  43  */
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_commit_ops;
  46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  48 static const struct nfs_rw_ops nfs_rw_write_ops;
  49 static void nfs_clear_request_commit(struct nfs_page *req);
  50 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  51                                       struct inode *inode);
  52
  53 static struct kmem_cache *nfs_wdata_cachep;
  54 static mempool_t *nfs_wdata_mempool;
  55 static struct kmem_cache *nfs_cdata_cachep;
  56 static mempool_t *nfs_commit_mempool;
  57
  58 struct nfs_commit_data *nfs_commitdata_alloc(void)
  59 {
  60         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  61
  62         if (p) {
  63                 memset(p, 0, sizeof(*p));
  64                 INIT_LIST_HEAD(&p->pages);
  65         }
  66         return p;
  67 }
  68 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  69
  70 void nfs_commit_free(struct nfs_commit_data *p)
  71 {
  72         mempool_free(p, nfs_commit_mempool);
  73 }
  74 EXPORT_SYMBOL_GPL(nfs_commit_free);
  75
  76 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  77 {
  78         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  79
  80         if (p)
  81                 memset(p, 0, sizeof(*p));
  82         return p;
  83 }
  84
  85 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
  86 {
  87         mempool_free(hdr, nfs_wdata_mempool);
  88 }
  89
  90 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  91 {
  92         ctx->error = error;
  93         smp_wmb();
  94         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
  95 }
  96
  97 /*
  98  * nfs_page_search_commits_for_head_request_locked
  99  *
 100  * Search through commit lists on @inode for the head request for @page.
 101  * Must be called while holding the inode (which is cinfo) lock.
 102  *
 103  * Returns the head request if found, or NULL if not found.
 104  */
 105 static struct nfs_page *
 106 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 107                                                 struct page *page)
 108 {
 109         struct nfs_page *freq, *t;
 110         struct nfs_commit_info cinfo;
 111         struct inode *inode = &nfsi->vfs_inode;
 112
 113         nfs_init_cinfo_from_inode(&cinfo, inode);
 114
 115         /* search through pnfs commit lists */
 116         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 117         if (freq)
 118                 return freq->wb_head;
 119
 120         /* Linearly search the commit list for the correct request */
 121         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 122                 if (freq->wb_page == page)
 123                         return freq->wb_head;
 124         }
 125
 126         return NULL;
 127 }
 128
 129 /*
 130  * nfs_page_find_head_request_locked - find head request associated with @page
 131  *
 132  * must be called while holding the inode lock.
 133  *
 134  * returns matching head request with reference held, or NULL if not found.
 135  */
 136 static struct nfs_page *
 137 nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
 138 {
 139         struct nfs_page *req = NULL;
 140
 141         if (PagePrivate(page))
 142                 req = (struct nfs_page *)page_private(page);
 143         else if (unlikely(PageSwapCache(page)))
 144                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 145                         page);
 146
 147         if (req) {
 148                 WARN_ON_ONCE(req->wb_head != req);
 149                 kref_get(&req->wb_kref);
 150         }
 151
 152         return req;
 153 }
 154
 155 /*
 156  * nfs_page_find_head_request - find head request associated with @page
 157  *
 158  * returns matching head request with reference held, or NULL if not found.
 159  */
 160 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 161 {
 162         struct inode *inode = page_file_mapping(page)->host;
 163         struct nfs_page *req = NULL;
 164
 165         spin_lock(&inode->i_lock);
 166         req = nfs_page_find_head_request_locked(NFS_I(inode), page);
 167         spin_unlock(&inode->i_lock);
 168         return req;
 169 }
 170
 171 /* Adjust the file length if we're writing beyond the end */
 172 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 173 {
 174         struct inode *inode = page_file_mapping(page)->host;
 175         loff_t end, i_size;
 176         pgoff_t end_index;
 177
 178         spin_lock(&inode->i_lock);
 179         i_size = i_size_read(inode);
 180         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 181         if (i_size > 0 && page_file_index(page) < end_index)
 182                 goto out;
 183         end = page_file_offset(page) + ((loff_t)offset+count);
 184         if (i_size >= end)
 185                 goto out;
 186         i_size_write(inode, end);
 187         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 188 out:
 189         spin_unlock(&inode->i_lock);
 190 }
 191
 192 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 193 static void nfs_set_pageerror(struct page *page)
 194 {
 195         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 196 }
 197
 198 /*
 199  * nfs_page_group_search_locked
 200  * @head - head request of page group
 201  * @page_offset - offset into page
 202  *
 203  * Search page group with head @head to find a request that contains the
 204  * page offset @page_offset.
 205  *
 206  * Returns a pointer to the first matching nfs request, or NULL if no
 207  * match is found.
 208  *
 209  * Must be called with the page group lock held
 210  */
 211 static struct nfs_page *
 212 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 213 {
 214         struct nfs_page *req;
 215
 216         WARN_ON_ONCE(head != head->wb_head);
 217         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
 218
 219         req = head;
 220         do {
 221                 if (page_offset >= req->wb_pgbase &&
 222                     page_offset < (req->wb_pgbase + req->wb_bytes))
 223                         return req;
 224
 225                 req = req->wb_this_page;
 226         } while (req != head);
 227
 228         return NULL;
 229 }
 230
 231 /*
 232  * nfs_page_group_covers_page
 233  * @head - head request of page group
 234  *
 235  * Return true if the page group with head @head covers the whole page,
 236  * returns false otherwise
 237  */
 238 static bool nfs_page_group_covers_page(struct nfs_page *req)
 239 {
 240         struct nfs_page *tmp;
 241         unsigned int pos = 0;
 242         unsigned int len = nfs_page_length(req->wb_page);
 243
 244         nfs_page_group_lock(req, false);
 245
 246         do {
 247                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 248                 if (tmp) {
 249                         /* no way this should happen */
 250                         WARN_ON_ONCE(tmp->wb_pgbase != pos);
 251                         pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
 252                 }
 253         } while (tmp && pos < len);
 254
 255         nfs_page_group_unlock(req);
 256         WARN_ON_ONCE(pos > len);
 257         return pos == len;
 258 }
 259
 260 /* We can set the PG_uptodate flag if we see that a write request
 261  * covers the full page.
 262  */
 263 static void nfs_mark_uptodate(struct nfs_page *req)
 264 {
 265         if (PageUptodate(req->wb_page))
 266                 return;
 267         if (!nfs_page_group_covers_page(req))
 268                 return;
 269         SetPageUptodate(req->wb_page);
 270 }
 271
 272 static int wb_priority(struct writeback_control *wbc)
 273 {
 274         if (wbc->for_reclaim)
 275                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 276         if (wbc->for_kupdate || wbc->for_background)
 277                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 278         return FLUSH_COND_STABLE;
 279 }
 280
 281 /*
 282  * NFS congestion control
 283  */
 284
 285 int nfs_congestion_kb;
 286
 287 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 288 #define NFS_CONGESTION_OFF_THRESH       \
 289         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 290
 291 static void nfs_set_page_writeback(struct page *page)
 292 {
 293         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 294         int ret = test_set_page_writeback(page);
 295
 296         WARN_ON_ONCE(ret != 0);
 297
 298         if (atomic_long_inc_return(&nfss->writeback) >
 299                         NFS_CONGESTION_ON_THRESH) {
 300                 set_bdi_congested(&nfss->backing_dev_info,
 301                                         BLK_RW_ASYNC);
 302         }
 303 }
 304
 305 static void nfs_end_page_writeback(struct nfs_page *req)
 306 {
 307         struct inode *inode = page_file_mapping(req->wb_page)->host;
 308         struct nfs_server *nfss = NFS_SERVER(inode);
 309
 310         if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
 311                 return;
 312
 313         end_page_writeback(req->wb_page);
 314         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 315                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 316 }
 317
 318
 319 /* nfs_page_group_clear_bits
 320  *   @req - an nfs request
 321  * clears all page group related bits from @req
 322  */
 323 static void
 324 nfs_page_group_clear_bits(struct nfs_page *req)
 325 {
 326         clear_bit(PG_TEARDOWN, &req->wb_flags);
 327         clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
 328         clear_bit(PG_UPTODATE, &req->wb_flags);
 329         clear_bit(PG_WB_END, &req->wb_flags);
 330         clear_bit(PG_REMOVE, &req->wb_flags);
 331 }
 332
 333
 334 /*
 335  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 336  *
 337  * this is a helper function for nfs_lock_and_join_requests
 338  *
 339  * @inode - inode associated with request page group, must be holding inode lock
 340  * @head  - head request of page group, must be holding head lock
 341  * @req   - request that couldn't lock and needs to wait on the req bit lock
 342  * @nonblock - if true, don't actually wait
 343  *
 344  * NOTE: this must be called holding page_group bit lock and inode spin lock
 345  *       and BOTH will be released before returning.
 346  *
 347  * returns 0 on success, < 0 on error.
 348  */
 349 static int
 350 nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
 351                           struct nfs_page *req, bool nonblock)
 352         __releases(&inode->i_lock)
 353 {
 354         struct nfs_page *tmp;
 355         int ret;
 356
 357         /* relinquish all the locks successfully grabbed this run */
 358         for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
 359                 nfs_unlock_request(tmp);
 360
 361         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 362
 363         /* grab a ref on the request that will be waited on */
 364         kref_get(&req->wb_kref);
 365
 366         nfs_page_group_unlock(head);
 367         spin_unlock(&inode->i_lock);
 368
 369         /* release ref from nfs_page_find_head_request_locked */
 370         nfs_release_request(head);
 371
 372         if (!nonblock)
 373                 ret = nfs_wait_on_request(req);
 374         else
 375                 ret = -EAGAIN;
 376         nfs_release_request(req);
 377
 378         return ret;
 379 }
 380
 381 /*
 382  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 383  *
 384  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 385  * @old_head - the old head of the list
 386  *
 387  * All subrequests must be locked and removed from all lists, so at this point
 388  * they are only "active" in this function, and possibly in nfs_wait_on_request
 389  * with a reference held by some other context.
 390  */
 391 static void
 392 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 393                                  struct nfs_page *old_head)
 394 {
 395         while (destroy_list) {
 396                 struct nfs_page *subreq = destroy_list;
 397
 398                 destroy_list = (subreq->wb_this_page == old_head) ?
 399                                    NULL : subreq->wb_this_page;
 400
 401                 WARN_ON_ONCE(old_head != subreq->wb_head);
 402
 403                 /* make sure old group is not used */
 404                 subreq->wb_head = subreq;
 405                 subreq->wb_this_page = subreq;
 406
 407                 /* subreq is now totally disconnected from page group or any
 408                  * write / commit lists. last chance to wake any waiters */
 409                 nfs_unlock_request(subreq);
 410
 411                 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 412                         /* release ref on old head request */
 413                         nfs_release_request(old_head);
 414
 415                         nfs_page_group_clear_bits(subreq);
 416
 417                         /* release the PG_INODE_REF reference */
 418                         if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
 419                                 nfs_release_request(subreq);
 420                         else
 421                                 WARN_ON_ONCE(1);
 422                 } else {
 423                         WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
 424                         /* zombie requests have already released the last
 425                          * reference and were waiting on the rest of the
 426                          * group to complete. Since it's no longer part of a
 427                          * group, simply free the request */
 428                         nfs_page_group_clear_bits(subreq);
 429                         nfs_free_request(subreq);
 430                 }
 431         }
 432 }
 433
 434 /*
 435  * nfs_lock_and_join_requests - join all subreqs to the head req and return
 436  *                              a locked reference, cancelling any pending
 437  *                              operations for this page.
 438  *
 439  * @page - the page used to lookup the "page group" of nfs_page structures
 440  * @nonblock - if true, don't block waiting for request locks
 441  *
 442  * This function joins all sub requests to the head request by first
 443  * locking all requests in the group, cancelling any pending operations
 444  * and finally updating the head request to cover the whole range covered by
 445  * the (former) group.  All subrequests are removed from any write or commit
 446  * lists, unlinked from the group and destroyed.
 447  *
 448  * Returns a locked, referenced pointer to the head request - which after
 449  * this call is guaranteed to be the only request associated with the page.
 450  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 451  * error was encountered.
 452  */
 453 static struct nfs_page *
 454 nfs_lock_and_join_requests(struct page *page, bool nonblock)
 455 {
 456         struct inode *inode = page_file_mapping(page)->host;
 457         struct nfs_page *head, *subreq;
 458         struct nfs_page *destroy_list = NULL;
 459         unsigned int total_bytes;
 460         int ret;
 461
 462 try_again:
 463         total_bytes = 0;
 464
 465         WARN_ON_ONCE(destroy_list);
 466
 467         spin_lock(&inode->i_lock);
 468
 469         /*
 470          * A reference is taken only on the head request which acts as a
 471          * reference to the whole page group - the group will not be destroyed
 472          * until the head reference is released.
 473          */
 474         head = nfs_page_find_head_request_locked(NFS_I(inode), page);
 475
 476         if (!head) {
 477                 spin_unlock(&inode->i_lock);
 478                 return NULL;
 479         }
 480
 481         /* holding inode lock, so always make a non-blocking call to try the
 482          * page group lock */
 483         ret = nfs_page_group_lock(head, true);
 484         if (ret < 0) {
 485                 spin_unlock(&inode->i_lock);
 486
 487                 if (!nonblock && ret == -EAGAIN) {
 488                         nfs_page_group_lock_wait(head);
 489                         nfs_release_request(head);
 490                         goto try_again;
 491                 }
 492
 493                 nfs_release_request(head);
 494                 return ERR_PTR(ret);
 495         }
 496
 497         /* lock each request in the page group */
 498         subreq = head;
 499         do {
 500                 /*
 501                  * Subrequests are always contiguous, non overlapping
 502                  * and in order. If not, it's a programming error.
 503                  */
 504                 WARN_ON_ONCE(subreq->wb_offset !=
 505                      (head->wb_offset + total_bytes));
 506
 507                 /* keep track of how many bytes this group covers */
 508                 total_bytes += subreq->wb_bytes;
 509
 510                 if (!nfs_lock_request(subreq)) {
 511                         /* releases page group bit lock and
 512                          * inode spin lock and all references */
 513                         ret = nfs_unroll_locks_and_wait(inode, head,
 514                                 subreq, nonblock);
 515
 516                         if (ret == 0)
 517                                 goto try_again;
 518
 519                         return ERR_PTR(ret);
 520                 }
 521
 522                 subreq = subreq->wb_this_page;
 523         } while (subreq != head);
 524
 525         /* Now that all requests are locked, make sure they aren't on any list.
 526          * Commit list removal accounting is done after locks are dropped */
 527         subreq = head;
 528         do {
 529                 nfs_clear_request_commit(subreq);
 530                 subreq = subreq->wb_this_page;
 531         } while (subreq != head);
 532
 533         /* unlink subrequests from head, destroy them later */
 534         if (head->wb_this_page != head) {
 535                 /* destroy list will be terminated by head */
 536                 destroy_list = head->wb_this_page;
 537                 head->wb_this_page = head;
 538
 539                 /* change head request to cover whole range that
 540                  * the former page group covered */
 541                 head->wb_bytes = total_bytes;
 542         }
 543
 544         /*
 545          * prepare head request to be added to new pgio descriptor
 546          */
 547         nfs_page_group_clear_bits(head);
 548
 549         /*
 550          * some part of the group was still on the inode list - otherwise
 551          * the group wouldn't be involved in async write.
 552          * grab a reference for the head request, iff it needs one.
 553          */
 554         if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
 555                 kref_get(&head->wb_kref);
 556
 557         nfs_page_group_unlock(head);
 558
 559         /* drop lock to clean uprequests on destroy list */
 560         spin_unlock(&inode->i_lock);
 561
 562         nfs_destroy_unlinked_subrequests(destroy_list, head);
 563
 564         /* still holds ref on head from nfs_page_find_head_request_locked
 565          * and still has lock on head from lock loop */
 566         return head;
 567 }
 568
 569 /*
 570  * Find an associated nfs write request, and prepare to flush it out
 571  * May return an error if the user signalled nfs_wait_on_request().
 572  */
 573 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 574                                 struct page *page, bool nonblock)
 575 {
 576         struct nfs_page *req;
 577         int ret = 0;
 578
 579         req = nfs_lock_and_join_requests(page, nonblock);
 580         if (!req)
 581                 goto out;
 582         ret = PTR_ERR(req);
 583         if (IS_ERR(req))
 584                 goto out;
 585
 586         nfs_set_page_writeback(page);
 587         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 588
 589         ret = 0;
 590         if (!nfs_pageio_add_request(pgio, req)) {
 591                 nfs_redirty_request(req);
 592                 ret = pgio->pg_error;
 593         }
 594 out:
 595         return ret;
 596 }
 597
 598 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 599 {
 600         struct inode *inode = page_file_mapping(page)->host;
 601         int ret;
 602
 603         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 604         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 605
 606         nfs_pageio_cond_complete(pgio, page_file_index(page));
 607         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 608         if (ret == -EAGAIN) {
 609                 redirty_page_for_writepage(wbc, page);
 610                 ret = 0;
 611         }
 612         return ret;
 613 }
 614
 615 /*
 616  * Write an mmapped page to the server.
 617  */
 618 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 619 {
 620         struct nfs_pageio_descriptor pgio;
 621         int err;
 622
 623         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 624                                 false, &nfs_async_write_completion_ops);
 625         err = nfs_do_writepage(page, wbc, &pgio);
 626         nfs_pageio_complete(&pgio);
 627         if (err < 0)
 628                 return err;
 629         if (pgio.pg_error < 0)
 630                 return pgio.pg_error;
 631         return 0;
 632 }
 633
 634 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 635 {
 636         int ret;
 637
 638         ret = nfs_writepage_locked(page, wbc);
 639         unlock_page(page);
 640         return ret;
 641 }
 642
 643 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 644 {
 645         int ret;
 646
 647         ret = nfs_do_writepage(page, wbc, data);
 648         unlock_page(page);
 649         return ret;
 650 }
 651
 652 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 653 {
 654         struct inode *inode = mapping->host;
 655         unsigned long *bitlock = &NFS_I(inode)->flags;
 656         struct nfs_pageio_descriptor pgio;
 657         int err;
 658
 659         /* Stop dirtying of new pages while we sync */
 660         err = wait_on_bit_lock_action(bitlock, NFS_INO_FLUSHING,
 661                         nfs_wait_bit_killable, TASK_KILLABLE);
 662         if (err)
 663                 goto out_err;
 664
 665         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 666
 667         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 668                                 &nfs_async_write_completion_ops);
 669         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 670         nfs_pageio_complete(&pgio);
 671
 672         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 673         smp_mb__after_atomic();
 674         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 675
 676         if (err < 0)
 677                 goto out_err;
 678         err = pgio.pg_error;
 679         if (err < 0)
 680                 goto out_err;
 681         return 0;
 682 out_err:
 683         return err;
 684 }
 685
 686 /*
 687  * Insert a write request into an inode
 688  */
 689 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 690 {
 691         struct nfs_inode *nfsi = NFS_I(inode);
 692
 693         WARN_ON_ONCE(req->wb_this_page != req);
 694
 695         /* Lock the request! */
 696         nfs_lock_request(req);
 697
 698         spin_lock(&inode->i_lock);
 699         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 700                 inode->i_version++;
 701         /*
 702          * Swap-space should not get truncated. Hence no need to plug the race
 703          * with invalidate/truncate.
 704          */
 705         if (likely(!PageSwapCache(req->wb_page))) {
 706                 set_bit(PG_MAPPED, &req->wb_flags);
 707                 SetPagePrivate(req->wb_page);
 708                 set_page_private(req->wb_page, (unsigned long)req);
 709         }
 710         nfsi->npages++;
 711         /* this a head request for a page group - mark it as having an
 712          * extra reference so sub groups can follow suit */
 713         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 714         kref_get(&req->wb_kref);
 715         spin_unlock(&inode->i_lock);
 716 }
 717
 718 /*
 719  * Remove a write request from an inode
 720  */
 721 static void nfs_inode_remove_request(struct nfs_page *req)
 722 {
 723         struct inode *inode = req->wb_context->dentry->d_inode;
 724         struct nfs_inode *nfsi = NFS_I(inode);
 725         struct nfs_page *head;
 726
 727         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 728                 head = req->wb_head;
 729
 730                 spin_lock(&inode->i_lock);
 731                 if (likely(!PageSwapCache(head->wb_page))) {
 732                         set_page_private(head->wb_page, 0);
 733                         ClearPagePrivate(head->wb_page);
 734                         clear_bit(PG_MAPPED, &head->wb_flags);
 735                 }
 736                 nfsi->npages--;
 737                 spin_unlock(&inode->i_lock);
 738         }
 739
 740         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 741                 nfs_release_request(req);
 742         else
 743                 WARN_ON_ONCE(1);
 744 }
 745
 746 static void
 747 nfs_mark_request_dirty(struct nfs_page *req)
 748 {
 749         __set_page_dirty_nobuffers(req->wb_page);
 750 }
 751
 752 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 753 /**
 754  * nfs_request_add_commit_list - add request to a commit list
 755  * @req: pointer to a struct nfs_page
 756  * @dst: commit list head
 757  * @cinfo: holds list lock and accounting info
 758  *
 759  * This sets the PG_CLEAN bit, updates the cinfo count of
 760  * number of outstanding requests requiring a commit as well as
 761  * the MM page stats.
 762  *
 763  * The caller must _not_ hold the cinfo->lock, but must be
 764  * holding the nfs_page lock.
 765  */
 766 void
 767 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 768                             struct nfs_commit_info *cinfo)
 769 {
 770         set_bit(PG_CLEAN, &(req)->wb_flags);
 771         spin_lock(cinfo->lock);
 772         nfs_list_add_request(req, dst);
 773         cinfo->mds->ncommit++;
 774         spin_unlock(cinfo->lock);
 775         if (!cinfo->dreq) {
 776                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 777                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 778                              BDI_RECLAIMABLE);
 779                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 780                                    I_DIRTY_DATASYNC);
 781         }
 782 }
 783 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 784
 785 /**
 786  * nfs_request_remove_commit_list - Remove request from a commit list
 787  * @req: pointer to a nfs_page
 788  * @cinfo: holds list lock and accounting info
 789  *
 790  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 791  * number of outstanding requests requiring a commit
 792  * It does not update the MM page stats.
 793  *
 794  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 795  */
 796 void
 797 nfs_request_remove_commit_list(struct nfs_page *req,
 798                                struct nfs_commit_info *cinfo)
 799 {
 800         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 801                 return;
 802         nfs_list_remove_request(req);
 803         cinfo->mds->ncommit--;
 804 }
 805 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 806
 807 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 808                                       struct inode *inode)
 809 {
 810         cinfo->lock = &inode->i_lock;
 811         cinfo->mds = &NFS_I(inode)->commit_info;
 812         cinfo->ds = pnfs_get_ds_info(inode);
 813         cinfo->dreq = NULL;
 814         cinfo->completion_ops = &nfs_commit_completion_ops;
 815 }
 816
 817 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 818                     struct inode *inode,
 819                     struct nfs_direct_req *dreq)
 820 {
 821         if (dreq)
 822                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 823         else
 824                 nfs_init_cinfo_from_inode(cinfo, inode);
 825 }
 826 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 827
 828 /*
 829  * Add a request to the inode's commit list.
 830  */
 831 void
 832 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 833                         struct nfs_commit_info *cinfo)
 834 {
 835         if (pnfs_mark_request_commit(req, lseg, cinfo))
 836                 return;
 837         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 838 }
 839
 840 static void
 841 nfs_clear_page_commit(struct page *page)
 842 {
 843         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 844         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 845 }
 846
 847 /* Called holding inode (/cinfo) lock */
 848 static void
 849 nfs_clear_request_commit(struct nfs_page *req)
 850 {
 851         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 852                 struct inode *inode = req->wb_context->dentry->d_inode;
 853                 struct nfs_commit_info cinfo;
 854
 855                 nfs_init_cinfo_from_inode(&cinfo, inode);
 856                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 857                         nfs_request_remove_commit_list(req, &cinfo);
 858                 }
 859                 nfs_clear_page_commit(req->wb_page);
 860         }
 861 }
 862
 863 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 864 {
 865         if (hdr->verf.committed == NFS_DATA_SYNC)
 866                 return hdr->lseg == NULL;
 867         return hdr->verf.committed != NFS_FILE_SYNC;
 868 }
 869
 870 #else
 871 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 872                                       struct inode *inode)
 873 {
 874 }
 875
 876 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 877                     struct inode *inode,
 878                     struct nfs_direct_req *dreq)
 879 {
 880 }
 881
 882 void
 883 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 884                         struct nfs_commit_info *cinfo)
 885 {
 886 }
 887
 888 static void
 889 nfs_clear_request_commit(struct nfs_page *req)
 890 {
 891 }
 892
 893 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 894 {
 895         return 0;
 896 }
 897
 898 #endif
 899
 900 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 901 {
 902         struct nfs_commit_info cinfo;
 903         unsigned long bytes = 0;
 904
 905         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 906                 goto out;
 907         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 908         while (!list_empty(&hdr->pages)) {
 909                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 910
 911                 bytes += req->wb_bytes;
 912                 nfs_list_remove_request(req);
 913                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 914                     (hdr->good_bytes < bytes)) {
 915                         nfs_set_pageerror(req->wb_page);
 916                         nfs_context_set_write_error(req->wb_context, hdr->error);
 917                         goto remove_req;
 918                 }
 919                 if (nfs_write_need_commit(hdr)) {
 920                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 921                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 922                         goto next;
 923                 }
 924 remove_req:
 925                 nfs_inode_remove_request(req);
 926 next:
 927                 nfs_unlock_request(req);
 928                 nfs_end_page_writeback(req);
 929                 nfs_release_request(req);
 930         }
 931 out:
 932         hdr->release(hdr);
 933 }
 934
 935 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 936 unsigned long
 937 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 938 {
 939         return cinfo->mds->ncommit;
 940 }
 941
 942 /* cinfo->lock held by caller */
 943 int
 944 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 945                      struct nfs_commit_info *cinfo, int max)
 946 {
 947         struct nfs_page *req, *tmp;
 948         int ret = 0;
 949
 950         list_for_each_entry_safe(req, tmp, src, wb_list) {
 951                 if (!nfs_lock_request(req))
 952                         continue;
 953                 kref_get(&req->wb_kref);
 954                 if (cond_resched_lock(cinfo->lock))
 955                         list_safe_reset_next(req, tmp, wb_list);
 956                 nfs_request_remove_commit_list(req, cinfo);
 957                 nfs_list_add_request(req, dst);
 958                 ret++;
 959                 if ((ret == max) && !cinfo->dreq)
 960                         break;
 961         }
 962         return ret;
 963 }
 964
 965 /*
 966  * nfs_scan_commit - Scan an inode for commit requests
 967  * @inode: NFS inode to scan
 968  * @dst: mds destination list
 969  * @cinfo: mds and ds lists of reqs ready to commit
 970  *
 971  * Moves requests from the inode's 'commit' request list.
 972  * The requests are *not* checked to ensure that they form a contiguous set.
 973  */
 974 int
 975 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 976                 struct nfs_commit_info *cinfo)
 977 {
 978         int ret = 0;
 979
 980         spin_lock(cinfo->lock);
 981         if (cinfo->mds->ncommit > 0) {
 982                 const int max = INT_MAX;
 983
 984                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 985                                            cinfo, max);
 986                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 987         }
 988         spin_unlock(cinfo->lock);
 989         return ret;
 990 }
 991
 992 #else
 993 unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 994 {
 995         return 0;
 996 }
 997
 998 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 999                     struct nfs_commit_info *cinfo)
1000 {
1001         return 0;
1002 }
1003 #endif
1004
1005 /*
1006  * Search for an existing write request, and attempt to update
1007  * it to reflect a new dirty region on a given page.
1008  *
1009  * If the attempt fails, then the existing request is flushed out
1010  * to disk.
1011  */
1012 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1013                 struct page *page,
1014                 unsigned int offset,
1015                 unsigned int bytes)
1016 {
1017         struct nfs_page *req;
1018         unsigned int rqend;
1019         unsigned int end;
1020         int error;
1021
1022         if (!PagePrivate(page))
1023                 return NULL;
1024
1025         end = offset + bytes;
1026         spin_lock(&inode->i_lock);
1027
1028         for (;;) {
1029                 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1030                 if (req == NULL)
1031                         goto out_unlock;
1032
1033                 /* should be handled by nfs_flush_incompatible */
1034                 WARN_ON_ONCE(req->wb_head != req);
1035                 WARN_ON_ONCE(req->wb_this_page != req);
1036
1037                 rqend = req->wb_offset + req->wb_bytes;
1038                 /*
1039                  * Tell the caller to flush out the request if
1040                  * the offsets are non-contiguous.
1041                  * Note: nfs_flush_incompatible() will already
1042                  * have flushed out requests having wrong owners.
1043                  */
1044                 if (offset > rqend
1045                     || end < req->wb_offset)
1046                         goto out_flushme;
1047
1048                 if (nfs_lock_request(req))
1049                         break;
1050
1051                 /* The request is locked, so wait and then retry */
1052                 spin_unlock(&inode->i_lock);
1053                 error = nfs_wait_on_request(req);
1054                 nfs_release_request(req);
1055                 if (error != 0)
1056                         goto out_err;
1057                 spin_lock(&inode->i_lock);
1058         }
1059
1060         /* Okay, the request matches. Update the region */
1061         if (offset < req->wb_offset) {
1062                 req->wb_offset = offset;
1063                 req->wb_pgbase = offset;
1064         }
1065         if (end > rqend)
1066                 req->wb_bytes = end - req->wb_offset;
1067         else
1068                 req->wb_bytes = rqend - req->wb_offset;
1069 out_unlock:
1070         if (req)
1071                 nfs_clear_request_commit(req);
1072         spin_unlock(&inode->i_lock);
1073         return req;
1074 out_flushme:
1075         spin_unlock(&inode->i_lock);
1076         nfs_release_request(req);
1077         error = nfs_wb_page(inode, page);
1078 out_err:
1079         return ERR_PTR(error);
1080 }
1081
1082 /*
1083  * Try to update an existing write request, or create one if there is none.
1084  *
1085  * Note: Should always be called with the Page Lock held to prevent races
1086  * if we have to add a new request. Also assumes that the caller has
1087  * already called nfs_flush_incompatible() if necessary.
1088  */
1089 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1090                 struct page *page, unsigned int offset, unsigned int bytes)
1091 {
1092         struct inode *inode = page_file_mapping(page)->host;
1093         struct nfs_page *req;
1094
1095         req = nfs_try_to_update_request(inode, page, offset, bytes);
1096         if (req != NULL)
1097                 goto out;
1098         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1099         if (IS_ERR(req))
1100                 goto out;
1101         nfs_inode_add_request(inode, req);
1102 out:
1103         return req;
1104 }
1105
1106 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1107                 unsigned int offset, unsigned int count)
1108 {
1109         struct nfs_page *req;
1110
1111         req = nfs_setup_write_request(ctx, page, offset, count);
1112         if (IS_ERR(req))
1113                 return PTR_ERR(req);
1114         /* Update file length */
1115         nfs_grow_file(page, offset, count);
1116         nfs_mark_uptodate(req);
1117         nfs_mark_request_dirty(req);
1118         nfs_unlock_and_release_request(req);
1119         return 0;
1120 }
1121
1122 int nfs_flush_incompatible(struct file *file, struct page *page)
1123 {
1124         struct nfs_open_context *ctx = nfs_file_open_context(file);
1125         struct nfs_lock_context *l_ctx;
1126         struct nfs_page *req;
1127         int do_flush, status;
1128         /*
1129          * Look for a request corresponding to this page. If there
1130          * is one, and it belongs to another file, we flush it out
1131          * before we try to copy anything into the page. Do this
1132          * due to the lack of an ACCESS-type call in NFSv2.
1133          * Also do the same if we find a request from an existing
1134          * dropped page.
1135          */
1136         do {
1137                 req = nfs_page_find_head_request(page);
1138                 if (req == NULL)
1139                         return 0;
1140                 l_ctx = req->wb_lock_context;
1141                 do_flush = req->wb_page != page || req->wb_context != ctx;
1142                 /* for now, flush if more than 1 request in page_group */
1143                 do_flush |= req->wb_this_page != req;
1144                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
1145                         do_flush |= l_ctx->lockowner.l_owner != current->files
1146                                 || l_ctx->lockowner.l_pid != current->tgid;
1147                 }
1148                 nfs_release_request(req);
1149                 if (!do_flush)
1150                         return 0;
1151                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1152         } while (status == 0);
1153         return status;
1154 }
1155
1156 /*
1157  * Avoid buffered writes when a open context credential's key would
1158  * expire soon.
1159  *
1160  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1161  *
1162  * Return 0 and set a credential flag which triggers the inode to flush
1163  * and performs  NFS_FILE_SYNC writes if the key will expired within
1164  * RPC_KEY_EXPIRE_TIMEO.
1165  */
1166 int
1167 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1168 {
1169         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1170         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1171
1172         return rpcauth_key_timeout_notify(auth, ctx->cred);
1173 }
1174
1175 /*
1176  * Test if the open context credential key is marked to expire soon.
1177  */
1178 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
1179 {
1180         return rpcauth_cred_key_to_expire(ctx->cred);
1181 }
1182
1183 /*
1184  * If the page cache is marked as unsafe or invalid, then we can't rely on
1185  * the PageUptodate() flag. In this case, we will need to turn off
1186  * write optimisations that depend on the page contents being correct.
1187  */
1188 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1189 {
1190         struct nfs_inode *nfsi = NFS_I(inode);
1191
1192         if (nfs_have_delegated_attributes(inode))
1193                 goto out;
1194         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1195                 return false;
1196         smp_rmb();
1197         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1198                 return false;
1199 out:
1200         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1201                 return false;
1202         return PageUptodate(page) != 0;
1203 }
1204
1205 /* If we know the page is up to date, and we're not using byte range locks (or
1206  * if we have the whole file locked for writing), it may be more efficient to
1207  * extend the write to cover the entire page in order to avoid fragmentation
1208  * inefficiencies.
1209  *
1210  * If the file is opened for synchronous writes then we can just skip the rest
1211  * of the checks.
1212  */
1213 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1214 {
1215         if (file->f_flags & O_DSYNC)
1216                 return 0;
1217         if (!nfs_write_pageuptodate(page, inode))
1218                 return 0;
1219         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1220                 return 1;
1221         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
1222                         inode->i_flock->fl_end == OFFSET_MAX &&
1223                         inode->i_flock->fl_type != F_RDLCK))
1224                 return 1;
1225         return 0;
1226 }
1227
1228 /*
1229  * Update and possibly write a cached page of an NFS file.
1230  *
1231  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1232  * things with a page scheduled for an RPC call (e.g. invalidate it).
1233  */
1234 int nfs_updatepage(struct file *file, struct page *page,
1235                 unsigned int offset, unsigned int count)
1236 {
1237         struct nfs_open_context *ctx = nfs_file_open_context(file);
1238         struct inode    *inode = page_file_mapping(page)->host;
1239         int             status = 0;
1240
1241         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1242
1243         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1244                 file, count, (long long)(page_file_offset(page) + offset));
1245
1246         if (nfs_can_extend_write(file, page, inode)) {
1247                 count = max(count + offset, nfs_page_length(page));
1248                 offset = 0;
1249         }
1250
1251         status = nfs_writepage_setup(ctx, page, offset, count);
1252         if (status < 0)
1253                 nfs_set_pageerror(page);
1254         else
1255                 __set_page_dirty_nobuffers(page);
1256
1257         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1258                         status, (long long)i_size_read(inode));
1259         return status;
1260 }
1261
1262 static int flush_task_priority(int how)
1263 {
1264         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1265                 case FLUSH_HIGHPRI:
1266                         return RPC_PRIORITY_HIGH;
1267                 case FLUSH_LOWPRI:
1268                         return RPC_PRIORITY_LOW;
1269         }
1270         return RPC_PRIORITY_NORMAL;
1271 }
1272
1273 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1274                                struct rpc_message *msg,
1275                                struct rpc_task_setup *task_setup_data, int how)
1276 {
1277         struct inode *inode = hdr->inode;
1278         int priority = flush_task_priority(how);
1279
1280         task_setup_data->priority = priority;
1281         NFS_PROTO(inode)->write_setup(hdr, msg);
1282
1283         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1284                                  &task_setup_data->rpc_client, msg, hdr);
1285 }
1286
1287 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1288  * call this on each, which will prepare them to be retried on next
1289  * writeback using standard nfs.
1290  */
1291 static void nfs_redirty_request(struct nfs_page *req)
1292 {
1293         nfs_mark_request_dirty(req);
1294         nfs_unlock_request(req);
1295         nfs_end_page_writeback(req);
1296         nfs_release_request(req);
1297 }
1298
1299 static void nfs_async_write_error(struct list_head *head)
1300 {
1301         struct nfs_page *req;
1302
1303         while (!list_empty(head)) {
1304                 req = nfs_list_entry(head->next);
1305                 nfs_list_remove_request(req);
1306                 nfs_redirty_request(req);
1307         }
1308 }
1309
1310 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1311         .error_cleanup = nfs_async_write_error,
1312         .completion = nfs_write_completion,
1313 };
1314
1315 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1316                                struct inode *inode, int ioflags, bool force_mds,
1317                                const struct nfs_pgio_completion_ops *compl_ops)
1318 {
1319         struct nfs_server *server = NFS_SERVER(inode);
1320         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1321
1322 #ifdef CONFIG_NFS_V4_1
1323         if (server->pnfs_curr_ld && !force_mds)
1324                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1325 #endif
1326         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1327                         server->wsize, ioflags);
1328 }
1329 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1330
1331 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1332 {
1333         pgio->pg_ops = &nfs_pgio_rw_ops;
1334         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1335 }
1336 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1337
1338
1339 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1340 {
1341         struct nfs_commit_data *data = calldata;
1342
1343         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1344 }
1345
1346 static void nfs_writeback_release_common(struct nfs_pgio_header *hdr)
1347 {
1348         /* do nothing! */
1349 }
1350
1351 /*
1352  * Special version of should_remove_suid() that ignores capabilities.
1353  */
1354 static int nfs_should_remove_suid(const struct inode *inode)
1355 {
1356         umode_t mode = inode->i_mode;
1357         int kill = 0;
1358
1359         /* suid always must be killed */
1360         if (unlikely(mode & S_ISUID))
1361                 kill = ATTR_KILL_SUID;
1362
1363         /*
1364          * sgid without any exec bits is just a mandatory locking mark; leave
1365          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1366          */
1367         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1368                 kill |= ATTR_KILL_SGID;
1369
1370         if (unlikely(kill && S_ISREG(mode)))
1371                 return kill;
1372
1373         return 0;
1374 }
1375
1376 /*
1377  * This function is called when the WRITE call is complete.
1378  */
1379 static int nfs_writeback_done(struct rpc_task *task,
1380                               struct nfs_pgio_header *hdr,
1381                               struct inode *inode)
1382 {
1383         int status;
1384
1385         /*
1386          * ->write_done will attempt to use post-op attributes to detect
1387          * conflicting writes by other clients.  A strict interpretation
1388          * of close-to-open would allow us to continue caching even if
1389          * another writer had changed the file, but some applications
1390          * depend on tighter cache coherency when writing.
1391          */
1392         status = NFS_PROTO(inode)->write_done(task, hdr);
1393         if (status != 0)
1394                 return status;
1395         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1396
1397 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1398         if (hdr->res.verf->committed < hdr->args.stable &&
1399             task->tk_status >= 0) {
1400                 /* We tried a write call, but the server did not
1401                  * commit data to stable storage even though we
1402                  * requested it.
1403                  * Note: There is a known bug in Tru64 < 5.0 in which
1404                  *       the server reports NFS_DATA_SYNC, but performs
1405                  *       NFS_FILE_SYNC. We therefore implement this checking
1406                  *       as a dprintk() in order to avoid filling syslog.
1407                  */
1408                 static unsigned long    complain;
1409
1410                 /* Note this will print the MDS for a DS write */
1411                 if (time_before(complain, jiffies)) {
1412                         dprintk("NFS:       faulty NFS server %s:"
1413                                 " (committed = %d) != (stable = %d)\n",
1414                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1415                                 hdr->res.verf->committed, hdr->args.stable);
1416                         complain = jiffies + 300 * HZ;
1417                 }
1418         }
1419 #endif
1420
1421         /* Deal with the suid/sgid bit corner case */
1422         if (nfs_should_remove_suid(inode))
1423                 nfs_mark_for_revalidate(inode);
1424         return 0;
1425 }
1426
1427 /*
1428  * This function is called when the WRITE call is complete.
1429  */
1430 static void nfs_writeback_result(struct rpc_task *task,
1431                                  struct nfs_pgio_header *hdr)
1432 {
1433         struct nfs_pgio_args    *argp = &hdr->args;
1434         struct nfs_pgio_res     *resp = &hdr->res;
1435
1436         if (resp->count < argp->count) {
1437                 static unsigned long    complain;
1438
1439                 /* This a short write! */
1440                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1441
1442                 /* Has the server at least made some progress? */
1443                 if (resp->count == 0) {
1444                         if (time_before(complain, jiffies)) {
1445                                 printk(KERN_WARNING
1446                                        "NFS: Server wrote zero bytes, expected %u.\n",
1447                                        argp->count);
1448                                 complain = jiffies + 300 * HZ;
1449                         }
1450                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1451                         task->tk_status = -EIO;
1452                         return;
1453                 }
1454                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1455                 if (resp->verf->committed != NFS_UNSTABLE) {
1456                         /* Resend from where the server left off */
1457                         hdr->mds_offset += resp->count;
1458                         argp->offset += resp->count;
1459                         argp->pgbase += resp->count;
1460                         argp->count -= resp->count;
1461                 } else {
1462                         /* Resend as a stable write in order to avoid
1463                          * headaches in the case of a server crash.
1464                          */
1465                         argp->stable = NFS_FILE_SYNC;
1466                 }
1467                 rpc_restart_call_prepare(task);
1468         }
1469 }
1470
1471
1472 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1473 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1474 {
1475         int ret;
1476
1477         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1478                 return 1;
1479         if (!may_wait)
1480                 return 0;
1481         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1482                                 NFS_INO_COMMIT,
1483                                 nfs_wait_bit_killable,
1484                                 TASK_KILLABLE);
1485         return (ret < 0) ? ret : 1;
1486 }
1487
1488 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1489 {
1490         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1491         smp_mb__after_atomic();
1492         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1493 }
1494
1495 void nfs_commitdata_release(struct nfs_commit_data *data)
1496 {
1497         put_nfs_open_context(data->context);
1498         nfs_commit_free(data);
1499 }
1500 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1501
1502 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1503                         const struct rpc_call_ops *call_ops,
1504                         int how, int flags)
1505 {
1506         struct rpc_task *task;
1507         int priority = flush_task_priority(how);
1508         struct rpc_message msg = {
1509                 .rpc_argp = &data->args,
1510                 .rpc_resp = &data->res,
1511                 .rpc_cred = data->cred,
1512         };
1513         struct rpc_task_setup task_setup_data = {
1514                 .task = &data->task,
1515                 .rpc_client = clnt,
1516                 .rpc_message = &msg,
1517                 .callback_ops = call_ops,
1518                 .callback_data = data,
1519                 .workqueue = nfsiod_workqueue,
1520                 .flags = RPC_TASK_ASYNC | flags,
1521                 .priority = priority,
1522         };
1523         /* Set up the initial task struct.  */
1524         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1525
1526         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1527
1528         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1529                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1530
1531         task = rpc_run_task(&task_setup_data);
1532         if (IS_ERR(task))
1533                 return PTR_ERR(task);
1534         if (how & FLUSH_SYNC)
1535                 rpc_wait_for_completion_task(task);
1536         rpc_put_task(task);
1537         return 0;
1538 }
1539 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1540
1541 /*
1542  * Set up the argument/result storage required for the RPC call.
1543  */
1544 void nfs_init_commit(struct nfs_commit_data *data,
1545                      struct list_head *head,
1546                      struct pnfs_layout_segment *lseg,
1547                      struct nfs_commit_info *cinfo)
1548 {
1549         struct nfs_page *first = nfs_list_entry(head->next);
1550         struct inode *inode = first->wb_context->dentry->d_inode;
1551
1552         /* Set up the RPC argument and reply structs
1553          * NB: take care not to mess about with data->commit et al. */
1554
1555         list_splice_init(head, &data->pages);
1556
1557         data->inode       = inode;
1558         data->cred        = first->wb_context->cred;
1559         data->lseg        = lseg; /* reference transferred */
1560         data->mds_ops     = &nfs_commit_ops;
1561         data->completion_ops = cinfo->completion_ops;
1562         data->dreq        = cinfo->dreq;
1563
1564         data->args.fh     = NFS_FH(data->inode);
1565         /* Note: we always request a commit of the entire inode */
1566         data->args.offset = 0;
1567         data->args.count  = 0;
1568         data->context     = get_nfs_open_context(first->wb_context);
1569         data->res.fattr   = &data->fattr;
1570         data->res.verf    = &data->verf;
1571         nfs_fattr_init(&data->fattr);
1572 }
1573 EXPORT_SYMBOL_GPL(nfs_init_commit);
1574
1575 void nfs_retry_commit(struct list_head *page_list,
1576                       struct pnfs_layout_segment *lseg,
1577                       struct nfs_commit_info *cinfo)
1578 {
1579         struct nfs_page *req;
1580
1581         while (!list_empty(page_list)) {
1582                 req = nfs_list_entry(page_list->next);
1583                 nfs_list_remove_request(req);
1584                 nfs_mark_request_commit(req, lseg, cinfo);
1585                 if (!cinfo->dreq) {
1586                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1587                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1588                                      BDI_RECLAIMABLE);
1589                 }
1590                 nfs_unlock_and_release_request(req);
1591         }
1592 }
1593 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1594
1595 /*
1596  * Commit dirty pages
1597  */
1598 static int
1599 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1600                 struct nfs_commit_info *cinfo)
1601 {
1602         struct nfs_commit_data  *data;
1603
1604         data = nfs_commitdata_alloc();
1605
1606         if (!data)
1607                 goto out_bad;
1608
1609         /* Set up the argument struct */
1610         nfs_init_commit(data, head, NULL, cinfo);
1611         atomic_inc(&cinfo->mds->rpcs_out);
1612         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1613                                    how, 0);
1614  out_bad:
1615         nfs_retry_commit(head, NULL, cinfo);
1616         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1617         return -ENOMEM;
1618 }
1619
1620 /*
1621  * COMMIT call returned
1622  */
1623 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1624 {
1625         struct nfs_commit_data  *data = calldata;
1626
1627         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1628                                 task->tk_pid, task->tk_status);
1629
1630         /* Call the NFS version-specific code */
1631         NFS_PROTO(data->inode)->commit_done(task, data);
1632 }
1633
1634 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1635 {
1636         struct nfs_page *req;
1637         int status = data->task.tk_status;
1638         struct nfs_commit_info cinfo;
1639
1640         while (!list_empty(&data->pages)) {
1641                 req = nfs_list_entry(data->pages.next);
1642                 nfs_list_remove_request(req);
1643                 nfs_clear_page_commit(req->wb_page);
1644
1645                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1646                         req->wb_context->dentry->d_sb->s_id,
1647                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1648                         req->wb_bytes,
1649                         (long long)req_offset(req));
1650                 if (status < 0) {
1651                         nfs_context_set_write_error(req->wb_context, status);
1652                         nfs_inode_remove_request(req);
1653                         dprintk(", error = %d\n", status);
1654                         goto next;
1655                 }
1656
1657                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1658                  * returned by the server against all stored verfs. */
1659                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1660                         /* We have a match */
1661                         nfs_inode_remove_request(req);
1662                         dprintk(" OK\n");
1663                         goto next;
1664                 }
1665                 /* We have a mismatch. Write the page again */
1666                 dprintk(" mismatch\n");
1667                 nfs_mark_request_dirty(req);
1668                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1669         next:
1670                 nfs_unlock_and_release_request(req);
1671         }
1672         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1673         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1674                 nfs_commit_clear_lock(NFS_I(data->inode));
1675 }
1676
1677 static void nfs_commit_release(void *calldata)
1678 {
1679         struct nfs_commit_data *data = calldata;
1680
1681         data->completion_ops->completion(data);
1682         nfs_commitdata_release(calldata);
1683 }
1684
1685 static const struct rpc_call_ops nfs_commit_ops = {
1686         .rpc_call_prepare = nfs_commit_prepare,
1687         .rpc_call_done = nfs_commit_done,
1688         .rpc_release = nfs_commit_release,
1689 };
1690
1691 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1692         .completion = nfs_commit_release_pages,
1693         .error_cleanup = nfs_commit_clear_lock,
1694 };
1695
1696 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1697                             int how, struct nfs_commit_info *cinfo)
1698 {
1699         int status;
1700
1701         status = pnfs_commit_list(inode, head, how, cinfo);
1702         if (status == PNFS_NOT_ATTEMPTED)
1703                 status = nfs_commit_list(inode, head, how, cinfo);
1704         return status;
1705 }
1706
1707 int nfs_commit_inode(struct inode *inode, int how)
1708 {
1709         LIST_HEAD(head);
1710         struct nfs_commit_info cinfo;
1711         int may_wait = how & FLUSH_SYNC;
1712         int res;
1713
1714         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1715         if (res <= 0)
1716                 goto out_mark_dirty;
1717         nfs_init_cinfo_from_inode(&cinfo, inode);
1718         res = nfs_scan_commit(inode, &head, &cinfo);
1719         if (res) {
1720                 int error;
1721
1722                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1723                 if (error < 0)
1724                         return error;
1725                 if (!may_wait)
1726                         goto out_mark_dirty;
1727                 error = wait_on_bit_action(&NFS_I(inode)->flags,
1728                                 NFS_INO_COMMIT,
1729                                 nfs_wait_bit_killable,
1730                                 TASK_KILLABLE);
1731                 if (error < 0)
1732                         return error;
1733         } else
1734                 nfs_commit_clear_lock(NFS_I(inode));
1735         return res;
1736         /* Note: If we exit without ensuring that the commit is complete,
1737          * we must mark the inode as dirty. Otherwise, future calls to
1738          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1739          * that the data is on the disk.
1740          */
1741 out_mark_dirty:
1742         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1743         return res;
1744 }
1745
1746 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1747 {
1748         struct nfs_inode *nfsi = NFS_I(inode);
1749         int flags = FLUSH_SYNC;
1750         int ret = 0;
1751
1752         /* no commits means nothing needs to be done */
1753         if (!nfsi->commit_info.ncommit)
1754                 return ret;
1755
1756         if (wbc->sync_mode == WB_SYNC_NONE) {
1757                 /* Don't commit yet if this is a non-blocking flush and there
1758                  * are a lot of outstanding writes for this mapping.
1759                  */
1760                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1761                         goto out_mark_dirty;
1762
1763                 /* don't wait for the COMMIT response */
1764                 flags = 0;
1765         }
1766
1767         ret = nfs_commit_inode(inode, flags);
1768         if (ret >= 0) {
1769                 if (wbc->sync_mode == WB_SYNC_NONE) {
1770                         if (ret < wbc->nr_to_write)
1771                                 wbc->nr_to_write -= ret;
1772                         else
1773                                 wbc->nr_to_write = 0;
1774                 }
1775                 return 0;
1776         }
1777 out_mark_dirty:
1778         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1779         return ret;
1780 }
1781 #else
1782 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1783 {
1784         return 0;
1785 }
1786 #endif
1787
1788 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1789 {
1790         return nfs_commit_unstable_pages(inode, wbc);
1791 }
1792 EXPORT_SYMBOL_GPL(nfs_write_inode);
1793
1794 /*
1795  * flush the inode to disk.
1796  */
1797 int nfs_wb_all(struct inode *inode)
1798 {
1799         struct writeback_control wbc = {
1800                 .sync_mode = WB_SYNC_ALL,
1801                 .nr_to_write = LONG_MAX,
1802                 .range_start = 0,
1803                 .range_end = LLONG_MAX,
1804         };
1805         int ret;
1806
1807         trace_nfs_writeback_inode_enter(inode);
1808
1809         ret = sync_inode(inode, &wbc);
1810
1811         trace_nfs_writeback_inode_exit(inode, ret);
1812         return ret;
1813 }
1814 EXPORT_SYMBOL_GPL(nfs_wb_all);
1815
1816 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1817 {
1818         struct nfs_page *req;
1819         int ret = 0;
1820
1821         wait_on_page_writeback(page);
1822
1823         /* blocking call to cancel all requests and join to a single (head)
1824          * request */
1825         req = nfs_lock_and_join_requests(page, false);
1826
1827         if (IS_ERR(req)) {
1828                 ret = PTR_ERR(req);
1829         } else if (req) {
1830                 /* all requests from this page have been cancelled by
1831                  * nfs_lock_and_join_requests, so just remove the head
1832                  * request from the inode / page_private pointer and
1833                  * release it */
1834                 nfs_inode_remove_request(req);
1835                 /*
1836                  * In case nfs_inode_remove_request has marked the
1837                  * page as being dirty
1838                  */
1839                 cancel_dirty_page(page, PAGE_CACHE_SIZE);
1840                 nfs_unlock_and_release_request(req);
1841         }
1842
1843         return ret;
1844 }
1845
1846 /*
1847  * Write back all requests on one page - we do this before reading it.
1848  */
1849 int nfs_wb_page(struct inode *inode, struct page *page)
1850 {
1851         loff_t range_start = page_file_offset(page);
1852         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1853         struct writeback_control wbc = {
1854                 .sync_mode = WB_SYNC_ALL,
1855                 .nr_to_write = 0,
1856                 .range_start = range_start,
1857                 .range_end = range_end,
1858         };
1859         int ret;
1860
1861         trace_nfs_writeback_page_enter(inode);
1862
1863         for (;;) {
1864                 wait_on_page_writeback(page);
1865                 if (clear_page_dirty_for_io(page)) {
1866                         ret = nfs_writepage_locked(page, &wbc);
1867                         if (ret < 0)
1868                                 goto out_error;
1869                         continue;
1870                 }
1871                 ret = 0;
1872                 if (!PagePrivate(page))
1873                         break;
1874                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1875                 if (ret < 0)
1876                         goto out_error;
1877         }
1878 out_error:
1879         trace_nfs_writeback_page_exit(inode, ret);
1880         return ret;
1881 }
1882
1883 #ifdef CONFIG_MIGRATION
1884 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1885                 struct page *page, enum migrate_mode mode)
1886 {
1887         /*
1888          * If PagePrivate is set, then the page is currently associated with
1889          * an in-progress read or write request. Don't try to migrate it.
1890          *
1891          * FIXME: we could do this in principle, but we'll need a way to ensure
1892          *        that we can safely release the inode reference while holding
1893          *        the page lock.
1894          */
1895         if (PagePrivate(page))
1896                 return -EBUSY;
1897
1898         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1899                 return -EBUSY;
1900
1901         return migrate_page(mapping, newpage, page, mode);
1902 }
1903 #endif
1904
1905 int __init nfs_init_writepagecache(void)
1906 {
1907         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1908                                              sizeof(struct nfs_pgio_header),
1909                                              0, SLAB_HWCACHE_ALIGN,
1910                                              NULL);
1911         if (nfs_wdata_cachep == NULL)
1912                 return -ENOMEM;
1913
1914         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1915                                                      nfs_wdata_cachep);
1916         if (nfs_wdata_mempool == NULL)
1917                 goto out_destroy_write_cache;
1918
1919         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1920                                              sizeof(struct nfs_commit_data),
1921                                              0, SLAB_HWCACHE_ALIGN,
1922                                              NULL);
1923         if (nfs_cdata_cachep == NULL)
1924                 goto out_destroy_write_mempool;
1925
1926         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1927                                                       nfs_cdata_cachep);
1928         if (nfs_commit_mempool == NULL)
1929                 goto out_destroy_commit_cache;
1930
1931         /*
1932          * NFS congestion size, scale with available memory.
1933          *
1934          *  64MB:    8192k
1935          * 128MB:   11585k
1936          * 256MB:   16384k
1937          * 512MB:   23170k
1938          *   1GB:   32768k
1939          *   2GB:   46340k
1940          *   4GB:   65536k
1941          *   8GB:   92681k
1942          *  16GB:  131072k
1943          *
1944          * This allows larger machines to have larger/more transfers.
1945          * Limit the default to 256M
1946          */
1947         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1948         if (nfs_congestion_kb > 256*1024)
1949                 nfs_congestion_kb = 256*1024;
1950
1951         return 0;
1952
1953 out_destroy_commit_cache:
1954         kmem_cache_destroy(nfs_cdata_cachep);
1955 out_destroy_write_mempool:
1956         mempool_destroy(nfs_wdata_mempool);
1957 out_destroy_write_cache:
1958         kmem_cache_destroy(nfs_wdata_cachep);
1959         return -ENOMEM;
1960 }
1961
1962 void nfs_destroy_writepagecache(void)
1963 {
1964         mempool_destroy(nfs_commit_mempool);
1965         kmem_cache_destroy(nfs_cdata_cachep);
1966         mempool_destroy(nfs_wdata_mempool);
1967         kmem_cache_destroy(nfs_wdata_cachep);
1968 }
1969
1970 static const struct nfs_rw_ops nfs_rw_write_ops = {
1971         .rw_mode                = FMODE_WRITE,
1972         .rw_alloc_header        = nfs_writehdr_alloc,
1973         .rw_free_header         = nfs_writehdr_free,
1974         .rw_release             = nfs_writeback_release_common,
1975         .rw_done                = nfs_writeback_done,
1976         .rw_result              = nfs_writeback_result,
1977         .rw_initiate            = nfs_initiate_write,
1978 };