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 #include <linux/freezer.h>
  25 #include <linux/wait.h>
  26
  27 #include <linux/uaccess.h>
  28
  29 #include "delegation.h"
  30 #include "internal.h"
  31 #include "iostat.h"
  32 #include "nfs4_fs.h"
  33 #include "fscache.h"
  34 #include "pnfs.h"
  35
  36 #include "nfstrace.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  39
  40 #define MIN_POOL_WRITE          (32)
  41 #define MIN_POOL_COMMIT         (4)
  42
  43 struct nfs_io_completion {
  44         void (*complete)(void *data);
  45         void *data;
  46         struct kref refcount;
  47 };
  48
  49 /*
  50  * Local function declarations
  51  */
  52 static void nfs_redirty_request(struct nfs_page *req);
  53 static const struct rpc_call_ops nfs_commit_ops;
  54 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  55 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  56 static const struct nfs_rw_ops nfs_rw_write_ops;
  57 static void nfs_clear_request_commit(struct nfs_page *req);
  58 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  59                                       struct inode *inode);
  60 static struct nfs_page *
  61 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  62                                                 struct page *page);
  63
  64 static struct kmem_cache *nfs_wdata_cachep;
  65 static mempool_t *nfs_wdata_mempool;
  66 static struct kmem_cache *nfs_cdata_cachep;
  67 static mempool_t *nfs_commit_mempool;
  68
  69 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
  70 {
  71         struct nfs_commit_data *p;
  72
  73         if (never_fail)
  74                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  75         else {
  76                 /* It is OK to do some reclaim, not no safe to wait
  77                  * for anything to be returned to the pool.
  78                  * mempool_alloc() cannot handle that particular combination,
  79                  * so we need two separate attempts.
  80                  */
  81                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  82                 if (!p)
  83                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
  84                                              __GFP_NOWARN | __GFP_NORETRY);
  85                 if (!p)
  86                         return NULL;
  87         }
  88
  89         memset(p, 0, sizeof(*p));
  90         INIT_LIST_HEAD(&p->pages);
  91         return p;
  92 }
  93 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  94
  95 void nfs_commit_free(struct nfs_commit_data *p)
  96 {
  97         mempool_free(p, nfs_commit_mempool);
  98 }
  99 EXPORT_SYMBOL_GPL(nfs_commit_free);
 100
 101 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
 102 {
 103         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
 104
 105         if (p) {
 106                 memset(p, 0, sizeof(*p));
 107                 p->rw_mode = FMODE_WRITE;
 108         }
 109         return p;
 110 }
 111
 112 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 113 {
 114         mempool_free(hdr, nfs_wdata_mempool);
 115 }
 116
 117 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 118 {
 119         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 120 }
 121
 122 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 123                 void (*complete)(void *), void *data)
 124 {
 125         ioc->complete = complete;
 126         ioc->data = data;
 127         kref_init(&ioc->refcount);
 128 }
 129
 130 static void nfs_io_completion_release(struct kref *kref)
 131 {
 132         struct nfs_io_completion *ioc = container_of(kref,
 133                         struct nfs_io_completion, refcount);
 134         ioc->complete(ioc->data);
 135         kfree(ioc);
 136 }
 137
 138 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 139 {
 140         if (ioc != NULL)
 141                 kref_get(&ioc->refcount);
 142 }
 143
 144 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 145 {
 146         if (ioc != NULL)
 147                 kref_put(&ioc->refcount, nfs_io_completion_release);
 148 }
 149
 150 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 151 {
 152         ctx->error = error;
 153         smp_wmb();
 154         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 155 }
 156
 157 /*
 158  * nfs_page_find_head_request_locked - find head request associated with @page
 159  *
 160  * must be called while holding the inode lock.
 161  *
 162  * returns matching head request with reference held, or NULL if not found.
 163  */
 164 static struct nfs_page *
 165 nfs_page_find_head_request_locked(struct nfs_inode *nfsi, struct page *page)
 166 {
 167         struct nfs_page *req = NULL;
 168
 169         if (PagePrivate(page))
 170                 req = (struct nfs_page *)page_private(page);
 171         else if (unlikely(PageSwapCache(page)))
 172                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 173                         page);
 174
 175         if (req) {
 176                 WARN_ON_ONCE(req->wb_head != req);
 177                 kref_get(&req->wb_kref);
 178         }
 179
 180         return req;
 181 }
 182
 183 /*
 184  * nfs_page_find_head_request - find head request associated with @page
 185  *
 186  * returns matching head request with reference held, or NULL if not found.
 187  */
 188 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 189 {
 190         struct inode *inode = page_file_mapping(page)->host;
 191         struct nfs_page *req = NULL;
 192
 193         spin_lock(&inode->i_lock);
 194         req = nfs_page_find_head_request_locked(NFS_I(inode), page);
 195         spin_unlock(&inode->i_lock);
 196         return req;
 197 }
 198
 199 /* Adjust the file length if we're writing beyond the end */
 200 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 201 {
 202         struct inode *inode = page_file_mapping(page)->host;
 203         loff_t end, i_size;
 204         pgoff_t end_index;
 205
 206         spin_lock(&inode->i_lock);
 207         i_size = i_size_read(inode);
 208         end_index = (i_size - 1) >> PAGE_SHIFT;
 209         if (i_size > 0 && page_index(page) < end_index)
 210                 goto out;
 211         end = page_file_offset(page) + ((loff_t)offset+count);
 212         if (i_size >= end)
 213                 goto out;
 214         i_size_write(inode, end);
 215         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 216 out:
 217         spin_unlock(&inode->i_lock);
 218 }
 219
 220 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 221 static void nfs_set_pageerror(struct page *page)
 222 {
 223         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 224 }
 225
 226 /*
 227  * nfs_page_group_search_locked
 228  * @head - head request of page group
 229  * @page_offset - offset into page
 230  *
 231  * Search page group with head @head to find a request that contains the
 232  * page offset @page_offset.
 233  *
 234  * Returns a pointer to the first matching nfs request, or NULL if no
 235  * match is found.
 236  *
 237  * Must be called with the page group lock held
 238  */
 239 static struct nfs_page *
 240 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 241 {
 242         struct nfs_page *req;
 243
 244         WARN_ON_ONCE(head != head->wb_head);
 245         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
 246
 247         req = head;
 248         do {
 249                 if (page_offset >= req->wb_pgbase &&
 250                     page_offset < (req->wb_pgbase + req->wb_bytes))
 251                         return req;
 252
 253                 req = req->wb_this_page;
 254         } while (req != head);
 255
 256         return NULL;
 257 }
 258
 259 /*
 260  * nfs_page_group_covers_page
 261  * @head - head request of page group
 262  *
 263  * Return true if the page group with head @head covers the whole page,
 264  * returns false otherwise
 265  */
 266 static bool nfs_page_group_covers_page(struct nfs_page *req)
 267 {
 268         struct nfs_page *tmp;
 269         unsigned int pos = 0;
 270         unsigned int len = nfs_page_length(req->wb_page);
 271
 272         nfs_page_group_lock(req, false);
 273
 274         do {
 275                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 276                 if (tmp) {
 277                         /* no way this should happen */
 278                         WARN_ON_ONCE(tmp->wb_pgbase != pos);
 279                         pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
 280                 }
 281         } while (tmp && pos < len);
 282
 283         nfs_page_group_unlock(req);
 284         WARN_ON_ONCE(pos > len);
 285         return pos == len;
 286 }
 287
 288 /* We can set the PG_uptodate flag if we see that a write request
 289  * covers the full page.
 290  */
 291 static void nfs_mark_uptodate(struct nfs_page *req)
 292 {
 293         if (PageUptodate(req->wb_page))
 294                 return;
 295         if (!nfs_page_group_covers_page(req))
 296                 return;
 297         SetPageUptodate(req->wb_page);
 298 }
 299
 300 static int wb_priority(struct writeback_control *wbc)
 301 {
 302         int ret = 0;
 303
 304         if (wbc->sync_mode == WB_SYNC_ALL)
 305                 ret = FLUSH_COND_STABLE;
 306         return ret;
 307 }
 308
 309 /*
 310  * NFS congestion control
 311  */
 312
 313 int nfs_congestion_kb;
 314
 315 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 316 #define NFS_CONGESTION_OFF_THRESH       \
 317         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 318
 319 static void nfs_set_page_writeback(struct page *page)
 320 {
 321         struct inode *inode = page_file_mapping(page)->host;
 322         struct nfs_server *nfss = NFS_SERVER(inode);
 323         int ret = test_set_page_writeback(page);
 324
 325         WARN_ON_ONCE(ret != 0);
 326
 327         if (atomic_long_inc_return(&nfss->writeback) >
 328                         NFS_CONGESTION_ON_THRESH)
 329                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 330 }
 331
 332 static void nfs_end_page_writeback(struct nfs_page *req)
 333 {
 334         struct inode *inode = page_file_mapping(req->wb_page)->host;
 335         struct nfs_server *nfss = NFS_SERVER(inode);
 336
 337         if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
 338                 return;
 339
 340         end_page_writeback(req->wb_page);
 341         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 342                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 343 }
 344
 345
 346 /* nfs_page_group_clear_bits
 347  *   @req - an nfs request
 348  * clears all page group related bits from @req
 349  */
 350 static void
 351 nfs_page_group_clear_bits(struct nfs_page *req)
 352 {
 353         clear_bit(PG_TEARDOWN, &req->wb_flags);
 354         clear_bit(PG_UNLOCKPAGE, &req->wb_flags);
 355         clear_bit(PG_UPTODATE, &req->wb_flags);
 356         clear_bit(PG_WB_END, &req->wb_flags);
 357         clear_bit(PG_REMOVE, &req->wb_flags);
 358 }
 359
 360
 361 /*
 362  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 363  *
 364  * this is a helper function for nfs_lock_and_join_requests
 365  *
 366  * @inode - inode associated with request page group, must be holding inode lock
 367  * @head  - head request of page group, must be holding head lock
 368  * @req   - request that couldn't lock and needs to wait on the req bit lock
 369  * @nonblock - if true, don't actually wait
 370  *
 371  * NOTE: this must be called holding page_group bit lock and inode spin lock
 372  *       and BOTH will be released before returning.
 373  *
 374  * returns 0 on success, < 0 on error.
 375  */
 376 static int
 377 nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
 378                           struct nfs_page *req, bool nonblock)
 379         __releases(&inode->i_lock)
 380 {
 381         struct nfs_page *tmp;
 382         int ret;
 383
 384         /* relinquish all the locks successfully grabbed this run */
 385         for (tmp = head ; tmp != req; tmp = tmp->wb_this_page)
 386                 nfs_unlock_request(tmp);
 387
 388         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 389
 390         /* grab a ref on the request that will be waited on */
 391         kref_get(&req->wb_kref);
 392
 393         nfs_page_group_unlock(head);
 394         spin_unlock(&inode->i_lock);
 395
 396         /* release ref from nfs_page_find_head_request_locked */
 397         nfs_release_request(head);
 398
 399         if (!nonblock)
 400                 ret = nfs_wait_on_request(req);
 401         else
 402                 ret = -EAGAIN;
 403         nfs_release_request(req);
 404
 405         return ret;
 406 }
 407
 408 /*
 409  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 410  *
 411  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 412  * @old_head - the old head of the list
 413  *
 414  * All subrequests must be locked and removed from all lists, so at this point
 415  * they are only "active" in this function, and possibly in nfs_wait_on_request
 416  * with a reference held by some other context.
 417  */
 418 static void
 419 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 420                                  struct nfs_page *old_head)
 421 {
 422         while (destroy_list) {
 423                 struct nfs_page *subreq = destroy_list;
 424
 425                 destroy_list = (subreq->wb_this_page == old_head) ?
 426                                    NULL : subreq->wb_this_page;
 427
 428                 WARN_ON_ONCE(old_head != subreq->wb_head);
 429
 430                 /* make sure old group is not used */
 431                 subreq->wb_head = subreq;
 432                 subreq->wb_this_page = subreq;
 433
 434                 /* subreq is now totally disconnected from page group or any
 435                  * write / commit lists. last chance to wake any waiters */
 436                 nfs_unlock_request(subreq);
 437
 438                 if (!test_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 439                         /* release ref on old head request */
 440                         nfs_release_request(old_head);
 441
 442                         nfs_page_group_clear_bits(subreq);
 443
 444                         /* release the PG_INODE_REF reference */
 445                         if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags))
 446                                 nfs_release_request(subreq);
 447                         else
 448                                 WARN_ON_ONCE(1);
 449                 } else {
 450                         WARN_ON_ONCE(test_bit(PG_CLEAN, &subreq->wb_flags));
 451                         /* zombie requests have already released the last
 452                          * reference and were waiting on the rest of the
 453                          * group to complete. Since it's no longer part of a
 454                          * group, simply free the request */
 455                         nfs_page_group_clear_bits(subreq);
 456                         nfs_free_request(subreq);
 457                 }
 458         }
 459 }
 460
 461 /*
 462  * nfs_lock_and_join_requests - join all subreqs to the head req and return
 463  *                              a locked reference, cancelling any pending
 464  *                              operations for this page.
 465  *
 466  * @page - the page used to lookup the "page group" of nfs_page structures
 467  * @nonblock - if true, don't block waiting for request locks
 468  *
 469  * This function joins all sub requests to the head request by first
 470  * locking all requests in the group, cancelling any pending operations
 471  * and finally updating the head request to cover the whole range covered by
 472  * the (former) group.  All subrequests are removed from any write or commit
 473  * lists, unlinked from the group and destroyed.
 474  *
 475  * Returns a locked, referenced pointer to the head request - which after
 476  * this call is guaranteed to be the only request associated with the page.
 477  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 478  * error was encountered.
 479  */
 480 static struct nfs_page *
 481 nfs_lock_and_join_requests(struct page *page, bool nonblock)
 482 {
 483         struct inode *inode = page_file_mapping(page)->host;
 484         struct nfs_page *head, *subreq;
 485         struct nfs_page *destroy_list = NULL;
 486         unsigned int total_bytes;
 487         int ret;
 488
 489 try_again:
 490         total_bytes = 0;
 491
 492         WARN_ON_ONCE(destroy_list);
 493
 494         spin_lock(&inode->i_lock);
 495
 496         /*
 497          * A reference is taken only on the head request which acts as a
 498          * reference to the whole page group - the group will not be destroyed
 499          * until the head reference is released.
 500          */
 501         head = nfs_page_find_head_request_locked(NFS_I(inode), page);
 502
 503         if (!head) {
 504                 spin_unlock(&inode->i_lock);
 505                 return NULL;
 506         }
 507
 508         /* holding inode lock, so always make a non-blocking call to try the
 509          * page group lock */
 510         ret = nfs_page_group_lock(head, true);
 511         if (ret < 0) {
 512                 spin_unlock(&inode->i_lock);
 513
 514                 if (!nonblock && ret == -EAGAIN) {
 515                         nfs_page_group_lock_wait(head);
 516                         nfs_release_request(head);
 517                         goto try_again;
 518                 }
 519
 520                 nfs_release_request(head);
 521                 return ERR_PTR(ret);
 522         }
 523
 524         /* lock each request in the page group */
 525         subreq = head;
 526         do {
 527                 /*
 528                  * Subrequests are always contiguous, non overlapping
 529                  * and in order - but may be repeated (mirrored writes).
 530                  */
 531                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
 532                         /* keep track of how many bytes this group covers */
 533                         total_bytes += subreq->wb_bytes;
 534                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
 535                             ((subreq->wb_offset + subreq->wb_bytes) >
 536                              (head->wb_offset + total_bytes)))) {
 537                         nfs_page_group_unlock(head);
 538                         spin_unlock(&inode->i_lock);
 539                         return ERR_PTR(-EIO);
 540                 }
 541
 542                 if (!nfs_lock_request(subreq)) {
 543                         /* releases page group bit lock and
 544                          * inode spin lock and all references */
 545                         ret = nfs_unroll_locks_and_wait(inode, head,
 546                                 subreq, nonblock);
 547
 548                         if (ret == 0)
 549                                 goto try_again;
 550
 551                         return ERR_PTR(ret);
 552                 }
 553
 554                 subreq = subreq->wb_this_page;
 555         } while (subreq != head);
 556
 557         /* Now that all requests are locked, make sure they aren't on any list.
 558          * Commit list removal accounting is done after locks are dropped */
 559         subreq = head;
 560         do {
 561                 nfs_clear_request_commit(subreq);
 562                 subreq = subreq->wb_this_page;
 563         } while (subreq != head);
 564
 565         /* unlink subrequests from head, destroy them later */
 566         if (head->wb_this_page != head) {
 567                 /* destroy list will be terminated by head */
 568                 destroy_list = head->wb_this_page;
 569                 head->wb_this_page = head;
 570
 571                 /* change head request to cover whole range that
 572                  * the former page group covered */
 573                 head->wb_bytes = total_bytes;
 574         }
 575
 576         /*
 577          * prepare head request to be added to new pgio descriptor
 578          */
 579         nfs_page_group_clear_bits(head);
 580
 581         /*
 582          * some part of the group was still on the inode list - otherwise
 583          * the group wouldn't be involved in async write.
 584          * grab a reference for the head request, iff it needs one.
 585          */
 586         if (!test_and_set_bit(PG_INODE_REF, &head->wb_flags))
 587                 kref_get(&head->wb_kref);
 588
 589         nfs_page_group_unlock(head);
 590
 591         /* drop lock to clean uprequests on destroy list */
 592         spin_unlock(&inode->i_lock);
 593
 594         nfs_destroy_unlinked_subrequests(destroy_list, head);
 595
 596         /* still holds ref on head from nfs_page_find_head_request_locked
 597          * and still has lock on head from lock loop */
 598         return head;
 599 }
 600
 601 static void nfs_write_error_remove_page(struct nfs_page *req)
 602 {
 603         nfs_unlock_request(req);
 604         nfs_end_page_writeback(req);
 605         generic_error_remove_page(page_file_mapping(req->wb_page),
 606                                   req->wb_page);
 607         nfs_release_request(req);
 608 }
 609
 610 static bool
 611 nfs_error_is_fatal_on_server(int err)
 612 {
 613         switch (err) {
 614         case 0:
 615         case -ERESTARTSYS:
 616         case -EINTR:
 617                 return false;
 618         }
 619         return nfs_error_is_fatal(err);
 620 }
 621
 622 /*
 623  * Find an associated nfs write request, and prepare to flush it out
 624  * May return an error if the user signalled nfs_wait_on_request().
 625  */
 626 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 627                                 struct page *page, bool nonblock)
 628 {
 629         struct nfs_page *req;
 630         int ret = 0;
 631
 632         req = nfs_lock_and_join_requests(page, nonblock);
 633         if (!req)
 634                 goto out;
 635         ret = PTR_ERR(req);
 636         if (IS_ERR(req))
 637                 goto out;
 638
 639         nfs_set_page_writeback(page);
 640         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 641
 642         ret = 0;
 643         /* If there is a fatal error that covers this write, just exit */
 644         if (nfs_error_is_fatal_on_server(req->wb_context->error))
 645                 goto out_launder;
 646
 647         if (!nfs_pageio_add_request(pgio, req)) {
 648                 ret = pgio->pg_error;
 649                 /*
 650                  * Remove the problematic req upon fatal errors on the server
 651                  */
 652                 if (nfs_error_is_fatal(ret)) {
 653                         nfs_context_set_write_error(req->wb_context, ret);
 654                         if (nfs_error_is_fatal_on_server(ret))
 655                                 goto out_launder;
 656                 }
 657                 nfs_redirty_request(req);
 658                 ret = -EAGAIN;
 659         } else
 660                 nfs_add_stats(page_file_mapping(page)->host,
 661                                 NFSIOS_WRITEPAGES, 1);
 662 out:
 663         return ret;
 664 out_launder:
 665         nfs_write_error_remove_page(req);
 666         return ret;
 667 }
 668
 669 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 670                             struct nfs_pageio_descriptor *pgio)
 671 {
 672         int ret;
 673
 674         nfs_pageio_cond_complete(pgio, page_index(page));
 675         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 676         if (ret == -EAGAIN) {
 677                 redirty_page_for_writepage(wbc, page);
 678                 ret = 0;
 679         }
 680         return ret;
 681 }
 682
 683 /*
 684  * Write an mmapped page to the server.
 685  */
 686 static int nfs_writepage_locked(struct page *page,
 687                                 struct writeback_control *wbc)
 688 {
 689         struct nfs_pageio_descriptor pgio;
 690         struct inode *inode = page_file_mapping(page)->host;
 691         int err;
 692
 693         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 694         nfs_pageio_init_write(&pgio, inode, 0,
 695                                 false, &nfs_async_write_completion_ops);
 696         err = nfs_do_writepage(page, wbc, &pgio);
 697         nfs_pageio_complete(&pgio);
 698         if (err < 0)
 699                 return err;
 700         if (pgio.pg_error < 0)
 701                 return pgio.pg_error;
 702         return 0;
 703 }
 704
 705 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 706 {
 707         int ret;
 708
 709         ret = nfs_writepage_locked(page, wbc);
 710         unlock_page(page);
 711         return ret;
 712 }
 713
 714 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 715 {
 716         int ret;
 717
 718         ret = nfs_do_writepage(page, wbc, data);
 719         unlock_page(page);
 720         return ret;
 721 }
 722
 723 static void nfs_io_completion_commit(void *inode)
 724 {
 725         nfs_commit_inode(inode, 0);
 726 }
 727
 728 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 729 {
 730         struct inode *inode = mapping->host;
 731         struct nfs_pageio_descriptor pgio;
 732         struct nfs_io_completion *ioc = nfs_io_completion_alloc(GFP_NOFS);
 733         int err;
 734
 735         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 736
 737         if (ioc)
 738                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 739
 740         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 741                                 &nfs_async_write_completion_ops);
 742         pgio.pg_io_completion = ioc;
 743         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 744         nfs_pageio_complete(&pgio);
 745         nfs_io_completion_put(ioc);
 746
 747         if (err < 0)
 748                 goto out_err;
 749         err = pgio.pg_error;
 750         if (err < 0)
 751                 goto out_err;
 752         return 0;
 753 out_err:
 754         return err;
 755 }
 756
 757 /*
 758  * Insert a write request into an inode
 759  */
 760 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 761 {
 762         struct nfs_inode *nfsi = NFS_I(inode);
 763
 764         WARN_ON_ONCE(req->wb_this_page != req);
 765
 766         /* Lock the request! */
 767         nfs_lock_request(req);
 768
 769         spin_lock(&inode->i_lock);
 770         if (!nfsi->nrequests &&
 771             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 772                 inode->i_version++;
 773         /*
 774          * Swap-space should not get truncated. Hence no need to plug the race
 775          * with invalidate/truncate.
 776          */
 777         if (likely(!PageSwapCache(req->wb_page))) {
 778                 set_bit(PG_MAPPED, &req->wb_flags);
 779                 SetPagePrivate(req->wb_page);
 780                 set_page_private(req->wb_page, (unsigned long)req);
 781         }
 782         nfsi->nrequests++;
 783         /* this a head request for a page group - mark it as having an
 784          * extra reference so sub groups can follow suit.
 785          * This flag also informs pgio layer when to bump nrequests when
 786          * adding subrequests. */
 787         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 788         kref_get(&req->wb_kref);
 789         spin_unlock(&inode->i_lock);
 790 }
 791
 792 /*
 793  * Remove a write request from an inode
 794  */
 795 static void nfs_inode_remove_request(struct nfs_page *req)
 796 {
 797         struct inode *inode = d_inode(req->wb_context->dentry);
 798         struct nfs_inode *nfsi = NFS_I(inode);
 799         struct nfs_page *head;
 800
 801         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 802                 head = req->wb_head;
 803
 804                 spin_lock(&inode->i_lock);
 805                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 806                         set_page_private(head->wb_page, 0);
 807                         ClearPagePrivate(head->wb_page);
 808                         clear_bit(PG_MAPPED, &head->wb_flags);
 809                 }
 810                 nfsi->nrequests--;
 811                 spin_unlock(&inode->i_lock);
 812         } else {
 813                 spin_lock(&inode->i_lock);
 814                 nfsi->nrequests--;
 815                 spin_unlock(&inode->i_lock);
 816         }
 817
 818         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 819                 nfs_release_request(req);
 820 }
 821
 822 static void
 823 nfs_mark_request_dirty(struct nfs_page *req)
 824 {
 825         if (req->wb_page)
 826                 __set_page_dirty_nobuffers(req->wb_page);
 827 }
 828
 829 /*
 830  * nfs_page_search_commits_for_head_request_locked
 831  *
 832  * Search through commit lists on @inode for the head request for @page.
 833  * Must be called while holding the inode (which is cinfo) lock.
 834  *
 835  * Returns the head request if found, or NULL if not found.
 836  */
 837 static struct nfs_page *
 838 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 839                                                 struct page *page)
 840 {
 841         struct nfs_page *freq, *t;
 842         struct nfs_commit_info cinfo;
 843         struct inode *inode = &nfsi->vfs_inode;
 844
 845         nfs_init_cinfo_from_inode(&cinfo, inode);
 846
 847         /* search through pnfs commit lists */
 848         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 849         if (freq)
 850                 return freq->wb_head;
 851
 852         /* Linearly search the commit list for the correct request */
 853         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 854                 if (freq->wb_page == page)
 855                         return freq->wb_head;
 856         }
 857
 858         return NULL;
 859 }
 860
 861 /**
 862  * nfs_request_add_commit_list_locked - add request to a commit list
 863  * @req: pointer to a struct nfs_page
 864  * @dst: commit list head
 865  * @cinfo: holds list lock and accounting info
 866  *
 867  * This sets the PG_CLEAN bit, updates the cinfo count of
 868  * number of outstanding requests requiring a commit as well as
 869  * the MM page stats.
 870  *
 871  * The caller must hold cinfo->inode->i_lock, and the nfs_page lock.
 872  */
 873 void
 874 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 875                             struct nfs_commit_info *cinfo)
 876 {
 877         set_bit(PG_CLEAN, &req->wb_flags);
 878         nfs_list_add_request(req, dst);
 879         cinfo->mds->ncommit++;
 880 }
 881 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 882
 883 /**
 884  * nfs_request_add_commit_list - add request to a commit list
 885  * @req: pointer to a struct nfs_page
 886  * @dst: commit list head
 887  * @cinfo: holds list lock and accounting info
 888  *
 889  * This sets the PG_CLEAN bit, updates the cinfo count of
 890  * number of outstanding requests requiring a commit as well as
 891  * the MM page stats.
 892  *
 893  * The caller must _not_ hold the cinfo->lock, but must be
 894  * holding the nfs_page lock.
 895  */
 896 void
 897 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 898 {
 899         spin_lock(&cinfo->inode->i_lock);
 900         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 901         spin_unlock(&cinfo->inode->i_lock);
 902         if (req->wb_page)
 903                 nfs_mark_page_unstable(req->wb_page, cinfo);
 904 }
 905 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 906
 907 /**
 908  * nfs_request_remove_commit_list - Remove request from a commit list
 909  * @req: pointer to a nfs_page
 910  * @cinfo: holds list lock and accounting info
 911  *
 912  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 913  * number of outstanding requests requiring a commit
 914  * It does not update the MM page stats.
 915  *
 916  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 917  */
 918 void
 919 nfs_request_remove_commit_list(struct nfs_page *req,
 920                                struct nfs_commit_info *cinfo)
 921 {
 922         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 923                 return;
 924         nfs_list_remove_request(req);
 925         cinfo->mds->ncommit--;
 926 }
 927 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 928
 929 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 930                                       struct inode *inode)
 931 {
 932         cinfo->inode = inode;
 933         cinfo->mds = &NFS_I(inode)->commit_info;
 934         cinfo->ds = pnfs_get_ds_info(inode);
 935         cinfo->dreq = NULL;
 936         cinfo->completion_ops = &nfs_commit_completion_ops;
 937 }
 938
 939 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 940                     struct inode *inode,
 941                     struct nfs_direct_req *dreq)
 942 {
 943         if (dreq)
 944                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 945         else
 946                 nfs_init_cinfo_from_inode(cinfo, inode);
 947 }
 948 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 949
 950 /*
 951  * Add a request to the inode's commit list.
 952  */
 953 void
 954 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 955                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 956 {
 957         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 958                 return;
 959         nfs_request_add_commit_list(req, cinfo);
 960 }
 961
 962 static void
 963 nfs_clear_page_commit(struct page *page)
 964 {
 965         dec_node_page_state(page, NR_UNSTABLE_NFS);
 966         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 967                     WB_RECLAIMABLE);
 968 }
 969
 970 /* Called holding inode (/cinfo) lock */
 971 static void
 972 nfs_clear_request_commit(struct nfs_page *req)
 973 {
 974         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 975                 struct inode *inode = d_inode(req->wb_context->dentry);
 976                 struct nfs_commit_info cinfo;
 977
 978                 nfs_init_cinfo_from_inode(&cinfo, inode);
 979                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 980                         nfs_request_remove_commit_list(req, &cinfo);
 981                 }
 982                 nfs_clear_page_commit(req->wb_page);
 983         }
 984 }
 985
 986 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 987 {
 988         if (hdr->verf.committed == NFS_DATA_SYNC)
 989                 return hdr->lseg == NULL;
 990         return hdr->verf.committed != NFS_FILE_SYNC;
 991 }
 992
 993 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 994 {
 995         nfs_io_completion_get(hdr->io_completion);
 996 }
 997
 998 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 999 {
1000         struct nfs_commit_info cinfo;
1001         unsigned long bytes = 0;
1002
1003         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1004                 goto out;
1005         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1006         while (!list_empty(&hdr->pages)) {
1007                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1008
1009                 bytes += req->wb_bytes;
1010                 nfs_list_remove_request(req);
1011                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1012                     (hdr->good_bytes < bytes)) {
1013                         nfs_set_pageerror(req->wb_page);
1014                         nfs_context_set_write_error(req->wb_context, hdr->error);
1015                         goto remove_req;
1016                 }
1017                 if (nfs_write_need_commit(hdr)) {
1018                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1019                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1020                                 hdr->pgio_mirror_idx);
1021                         goto next;
1022                 }
1023 remove_req:
1024                 nfs_inode_remove_request(req);
1025 next:
1026                 nfs_unlock_request(req);
1027                 nfs_end_page_writeback(req);
1028                 nfs_release_request(req);
1029         }
1030 out:
1031         nfs_io_completion_put(hdr->io_completion);
1032         hdr->release(hdr);
1033 }
1034
1035 unsigned long
1036 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1037 {
1038         return cinfo->mds->ncommit;
1039 }
1040
1041 /* cinfo->inode->i_lock held by caller */
1042 int
1043 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1044                      struct nfs_commit_info *cinfo, int max)
1045 {
1046         struct nfs_page *req, *tmp;
1047         int ret = 0;
1048
1049         list_for_each_entry_safe(req, tmp, src, wb_list) {
1050                 if (!nfs_lock_request(req))
1051                         continue;
1052                 kref_get(&req->wb_kref);
1053                 if (cond_resched_lock(&cinfo->inode->i_lock))
1054                         list_safe_reset_next(req, tmp, wb_list);
1055                 nfs_request_remove_commit_list(req, cinfo);
1056                 nfs_list_add_request(req, dst);
1057                 ret++;
1058                 if ((ret == max) && !cinfo->dreq)
1059                         break;
1060         }
1061         return ret;
1062 }
1063
1064 /*
1065  * nfs_scan_commit - Scan an inode for commit requests
1066  * @inode: NFS inode to scan
1067  * @dst: mds destination list
1068  * @cinfo: mds and ds lists of reqs ready to commit
1069  *
1070  * Moves requests from the inode's 'commit' request list.
1071  * The requests are *not* checked to ensure that they form a contiguous set.
1072  */
1073 int
1074 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1075                 struct nfs_commit_info *cinfo)
1076 {
1077         int ret = 0;
1078
1079         spin_lock(&cinfo->inode->i_lock);
1080         if (cinfo->mds->ncommit > 0) {
1081                 const int max = INT_MAX;
1082
1083                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1084                                            cinfo, max);
1085                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1086         }
1087         spin_unlock(&cinfo->inode->i_lock);
1088         return ret;
1089 }
1090
1091 /*
1092  * Search for an existing write request, and attempt to update
1093  * it to reflect a new dirty region on a given page.
1094  *
1095  * If the attempt fails, then the existing request is flushed out
1096  * to disk.
1097  */
1098 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1099                 struct page *page,
1100                 unsigned int offset,
1101                 unsigned int bytes)
1102 {
1103         struct nfs_page *req;
1104         unsigned int rqend;
1105         unsigned int end;
1106         int error;
1107
1108         if (!PagePrivate(page))
1109                 return NULL;
1110
1111         end = offset + bytes;
1112         spin_lock(&inode->i_lock);
1113
1114         for (;;) {
1115                 req = nfs_page_find_head_request_locked(NFS_I(inode), page);
1116                 if (req == NULL)
1117                         goto out_unlock;
1118
1119                 /* should be handled by nfs_flush_incompatible */
1120                 WARN_ON_ONCE(req->wb_head != req);
1121                 WARN_ON_ONCE(req->wb_this_page != req);
1122
1123                 rqend = req->wb_offset + req->wb_bytes;
1124                 /*
1125                  * Tell the caller to flush out the request if
1126                  * the offsets are non-contiguous.
1127                  * Note: nfs_flush_incompatible() will already
1128                  * have flushed out requests having wrong owners.
1129                  */
1130                 if (offset > rqend
1131                     || end < req->wb_offset)
1132                         goto out_flushme;
1133
1134                 if (nfs_lock_request(req))
1135                         break;
1136
1137                 /* The request is locked, so wait and then retry */
1138                 spin_unlock(&inode->i_lock);
1139                 error = nfs_wait_on_request(req);
1140                 nfs_release_request(req);
1141                 if (error != 0)
1142                         goto out_err;
1143                 spin_lock(&inode->i_lock);
1144         }
1145
1146         /* Okay, the request matches. Update the region */
1147         if (offset < req->wb_offset) {
1148                 req->wb_offset = offset;
1149                 req->wb_pgbase = offset;
1150         }
1151         if (end > rqend)
1152                 req->wb_bytes = end - req->wb_offset;
1153         else
1154                 req->wb_bytes = rqend - req->wb_offset;
1155 out_unlock:
1156         if (req)
1157                 nfs_clear_request_commit(req);
1158         spin_unlock(&inode->i_lock);
1159         return req;
1160 out_flushme:
1161         spin_unlock(&inode->i_lock);
1162         nfs_release_request(req);
1163         error = nfs_wb_page(inode, page);
1164 out_err:
1165         return ERR_PTR(error);
1166 }
1167
1168 /*
1169  * Try to update an existing write request, or create one if there is none.
1170  *
1171  * Note: Should always be called with the Page Lock held to prevent races
1172  * if we have to add a new request. Also assumes that the caller has
1173  * already called nfs_flush_incompatible() if necessary.
1174  */
1175 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1176                 struct page *page, unsigned int offset, unsigned int bytes)
1177 {
1178         struct inode *inode = page_file_mapping(page)->host;
1179         struct nfs_page *req;
1180
1181         req = nfs_try_to_update_request(inode, page, offset, bytes);
1182         if (req != NULL)
1183                 goto out;
1184         req = nfs_create_request(ctx, page, NULL, offset, bytes);
1185         if (IS_ERR(req))
1186                 goto out;
1187         nfs_inode_add_request(inode, req);
1188 out:
1189         return req;
1190 }
1191
1192 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1193                 unsigned int offset, unsigned int count)
1194 {
1195         struct nfs_page *req;
1196
1197         req = nfs_setup_write_request(ctx, page, offset, count);
1198         if (IS_ERR(req))
1199                 return PTR_ERR(req);
1200         /* Update file length */
1201         nfs_grow_file(page, offset, count);
1202         nfs_mark_uptodate(req);
1203         nfs_mark_request_dirty(req);
1204         nfs_unlock_and_release_request(req);
1205         return 0;
1206 }
1207
1208 int nfs_flush_incompatible(struct file *file, struct page *page)
1209 {
1210         struct nfs_open_context *ctx = nfs_file_open_context(file);
1211         struct nfs_lock_context *l_ctx;
1212         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1213         struct nfs_page *req;
1214         int do_flush, status;
1215         /*
1216          * Look for a request corresponding to this page. If there
1217          * is one, and it belongs to another file, we flush it out
1218          * before we try to copy anything into the page. Do this
1219          * due to the lack of an ACCESS-type call in NFSv2.
1220          * Also do the same if we find a request from an existing
1221          * dropped page.
1222          */
1223         do {
1224                 req = nfs_page_find_head_request(page);
1225                 if (req == NULL)
1226                         return 0;
1227                 l_ctx = req->wb_lock_context;
1228                 do_flush = req->wb_page != page ||
1229                         !nfs_match_open_context(req->wb_context, ctx);
1230                 /* for now, flush if more than 1 request in page_group */
1231                 do_flush |= req->wb_this_page != req;
1232                 if (l_ctx && flctx &&
1233                     !(list_empty_careful(&flctx->flc_posix) &&
1234                       list_empty_careful(&flctx->flc_flock))) {
1235                         do_flush |= l_ctx->lockowner != current->files;
1236                 }
1237                 nfs_release_request(req);
1238                 if (!do_flush)
1239                         return 0;
1240                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1241         } while (status == 0);
1242         return status;
1243 }
1244
1245 /*
1246  * Avoid buffered writes when a open context credential's key would
1247  * expire soon.
1248  *
1249  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1250  *
1251  * Return 0 and set a credential flag which triggers the inode to flush
1252  * and performs  NFS_FILE_SYNC writes if the key will expired within
1253  * RPC_KEY_EXPIRE_TIMEO.
1254  */
1255 int
1256 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1257 {
1258         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1259         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1260
1261         return rpcauth_key_timeout_notify(auth, ctx->cred);
1262 }
1263
1264 /*
1265  * Test if the open context credential key is marked to expire soon.
1266  */
1267 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1268 {
1269         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1270
1271         return rpcauth_cred_key_to_expire(auth, ctx->cred);
1272 }
1273
1274 /*
1275  * If the page cache is marked as unsafe or invalid, then we can't rely on
1276  * the PageUptodate() flag. In this case, we will need to turn off
1277  * write optimisations that depend on the page contents being correct.
1278  */
1279 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1280 {
1281         struct nfs_inode *nfsi = NFS_I(inode);
1282
1283         if (nfs_have_delegated_attributes(inode))
1284                 goto out;
1285         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1286                 return false;
1287         smp_rmb();
1288         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1289                 return false;
1290 out:
1291         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1292                 return false;
1293         return PageUptodate(page) != 0;
1294 }
1295
1296 static bool
1297 is_whole_file_wrlock(struct file_lock *fl)
1298 {
1299         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1300                         fl->fl_type == F_WRLCK;
1301 }
1302
1303 /* If we know the page is up to date, and we're not using byte range locks (or
1304  * if we have the whole file locked for writing), it may be more efficient to
1305  * extend the write to cover the entire page in order to avoid fragmentation
1306  * inefficiencies.
1307  *
1308  * If the file is opened for synchronous writes then we can just skip the rest
1309  * of the checks.
1310  */
1311 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1312 {
1313         int ret;
1314         struct file_lock_context *flctx = inode->i_flctx;
1315         struct file_lock *fl;
1316
1317         if (file->f_flags & O_DSYNC)
1318                 return 0;
1319         if (!nfs_write_pageuptodate(page, inode))
1320                 return 0;
1321         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1322                 return 1;
1323         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1324                        list_empty_careful(&flctx->flc_posix)))
1325                 return 1;
1326
1327         /* Check to see if there are whole file write locks */
1328         ret = 0;
1329         spin_lock(&flctx->flc_lock);
1330         if (!list_empty(&flctx->flc_posix)) {
1331                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1332                                         fl_list);
1333                 if (is_whole_file_wrlock(fl))
1334                         ret = 1;
1335         } else if (!list_empty(&flctx->flc_flock)) {
1336                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1337                                         fl_list);
1338                 if (fl->fl_type == F_WRLCK)
1339                         ret = 1;
1340         }
1341         spin_unlock(&flctx->flc_lock);
1342         return ret;
1343 }
1344
1345 /*
1346  * Update and possibly write a cached page of an NFS file.
1347  *
1348  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1349  * things with a page scheduled for an RPC call (e.g. invalidate it).
1350  */
1351 int nfs_updatepage(struct file *file, struct page *page,
1352                 unsigned int offset, unsigned int count)
1353 {
1354         struct nfs_open_context *ctx = nfs_file_open_context(file);
1355         struct inode    *inode = page_file_mapping(page)->host;
1356         int             status = 0;
1357
1358         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1359
1360         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1361                 file, count, (long long)(page_file_offset(page) + offset));
1362
1363         if (!count)
1364                 goto out;
1365
1366         if (nfs_can_extend_write(file, page, inode)) {
1367                 count = max(count + offset, nfs_page_length(page));
1368                 offset = 0;
1369         }
1370
1371         status = nfs_writepage_setup(ctx, page, offset, count);
1372         if (status < 0)
1373                 nfs_set_pageerror(page);
1374         else
1375                 __set_page_dirty_nobuffers(page);
1376 out:
1377         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1378                         status, (long long)i_size_read(inode));
1379         return status;
1380 }
1381
1382 static int flush_task_priority(int how)
1383 {
1384         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1385                 case FLUSH_HIGHPRI:
1386                         return RPC_PRIORITY_HIGH;
1387                 case FLUSH_LOWPRI:
1388                         return RPC_PRIORITY_LOW;
1389         }
1390         return RPC_PRIORITY_NORMAL;
1391 }
1392
1393 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1394                                struct rpc_message *msg,
1395                                const struct nfs_rpc_ops *rpc_ops,
1396                                struct rpc_task_setup *task_setup_data, int how)
1397 {
1398         int priority = flush_task_priority(how);
1399
1400         task_setup_data->priority = priority;
1401         rpc_ops->write_setup(hdr, msg);
1402
1403         nfs4_state_protect_write(NFS_SERVER(hdr->inode)->nfs_client,
1404                                  &task_setup_data->rpc_client, msg, hdr);
1405 }
1406
1407 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1408  * call this on each, which will prepare them to be retried on next
1409  * writeback using standard nfs.
1410  */
1411 static void nfs_redirty_request(struct nfs_page *req)
1412 {
1413         nfs_mark_request_dirty(req);
1414         set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1415         nfs_unlock_request(req);
1416         nfs_end_page_writeback(req);
1417         nfs_release_request(req);
1418 }
1419
1420 static void nfs_async_write_error(struct list_head *head)
1421 {
1422         struct nfs_page *req;
1423
1424         while (!list_empty(head)) {
1425                 req = nfs_list_entry(head->next);
1426                 nfs_list_remove_request(req);
1427                 nfs_redirty_request(req);
1428         }
1429 }
1430
1431 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1432 {
1433         nfs_async_write_error(&hdr->pages);
1434 }
1435
1436 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1437         .init_hdr = nfs_async_write_init,
1438         .error_cleanup = nfs_async_write_error,
1439         .completion = nfs_write_completion,
1440         .reschedule_io = nfs_async_write_reschedule_io,
1441 };
1442
1443 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1444                                struct inode *inode, int ioflags, bool force_mds,
1445                                const struct nfs_pgio_completion_ops *compl_ops)
1446 {
1447         struct nfs_server *server = NFS_SERVER(inode);
1448         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1449
1450 #ifdef CONFIG_NFS_V4_1
1451         if (server->pnfs_curr_ld && !force_mds)
1452                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1453 #endif
1454         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1455                         server->wsize, ioflags, GFP_NOIO);
1456 }
1457 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1458
1459 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1460 {
1461         struct nfs_pgio_mirror *mirror;
1462
1463         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1464                 pgio->pg_ops->pg_cleanup(pgio);
1465
1466         pgio->pg_ops = &nfs_pgio_rw_ops;
1467
1468         nfs_pageio_stop_mirroring(pgio);
1469
1470         mirror = &pgio->pg_mirrors[0];
1471         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1472 }
1473 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1474
1475
1476 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1477 {
1478         struct nfs_commit_data *data = calldata;
1479
1480         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1481 }
1482
1483 /*
1484  * Special version of should_remove_suid() that ignores capabilities.
1485  */
1486 static int nfs_should_remove_suid(const struct inode *inode)
1487 {
1488         umode_t mode = inode->i_mode;
1489         int kill = 0;
1490
1491         /* suid always must be killed */
1492         if (unlikely(mode & S_ISUID))
1493                 kill = ATTR_KILL_SUID;
1494
1495         /*
1496          * sgid without any exec bits is just a mandatory locking mark; leave
1497          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1498          */
1499         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1500                 kill |= ATTR_KILL_SGID;
1501
1502         if (unlikely(kill && S_ISREG(mode)))
1503                 return kill;
1504
1505         return 0;
1506 }
1507
1508 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1509                 struct nfs_fattr *fattr)
1510 {
1511         struct nfs_pgio_args *argp = &hdr->args;
1512         struct nfs_pgio_res *resp = &hdr->res;
1513         u64 size = argp->offset + resp->count;
1514
1515         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1516                 fattr->size = size;
1517         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1518                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1519                 return;
1520         }
1521         if (size != fattr->size)
1522                 return;
1523         /* Set attribute barrier */
1524         nfs_fattr_set_barrier(fattr);
1525         /* ...and update size */
1526         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1527 }
1528
1529 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1530 {
1531         struct nfs_fattr *fattr = &hdr->fattr;
1532         struct inode *inode = hdr->inode;
1533
1534         spin_lock(&inode->i_lock);
1535         nfs_writeback_check_extend(hdr, fattr);
1536         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1537         spin_unlock(&inode->i_lock);
1538 }
1539 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1540
1541 /*
1542  * This function is called when the WRITE call is complete.
1543  */
1544 static int nfs_writeback_done(struct rpc_task *task,
1545                               struct nfs_pgio_header *hdr,
1546                               struct inode *inode)
1547 {
1548         int status;
1549
1550         /*
1551          * ->write_done will attempt to use post-op attributes to detect
1552          * conflicting writes by other clients.  A strict interpretation
1553          * of close-to-open would allow us to continue caching even if
1554          * another writer had changed the file, but some applications
1555          * depend on tighter cache coherency when writing.
1556          */
1557         status = NFS_PROTO(inode)->write_done(task, hdr);
1558         if (status != 0)
1559                 return status;
1560         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1561
1562         if (hdr->res.verf->committed < hdr->args.stable &&
1563             task->tk_status >= 0) {
1564                 /* We tried a write call, but the server did not
1565                  * commit data to stable storage even though we
1566                  * requested it.
1567                  * Note: There is a known bug in Tru64 < 5.0 in which
1568                  *       the server reports NFS_DATA_SYNC, but performs
1569                  *       NFS_FILE_SYNC. We therefore implement this checking
1570                  *       as a dprintk() in order to avoid filling syslog.
1571                  */
1572                 static unsigned long    complain;
1573
1574                 /* Note this will print the MDS for a DS write */
1575                 if (time_before(complain, jiffies)) {
1576                         dprintk("NFS:       faulty NFS server %s:"
1577                                 " (committed = %d) != (stable = %d)\n",
1578                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1579                                 hdr->res.verf->committed, hdr->args.stable);
1580                         complain = jiffies + 300 * HZ;
1581                 }
1582         }
1583
1584         /* Deal with the suid/sgid bit corner case */
1585         if (nfs_should_remove_suid(inode))
1586                 nfs_mark_for_revalidate(inode);
1587         return 0;
1588 }
1589
1590 /*
1591  * This function is called when the WRITE call is complete.
1592  */
1593 static void nfs_writeback_result(struct rpc_task *task,
1594                                  struct nfs_pgio_header *hdr)
1595 {
1596         struct nfs_pgio_args    *argp = &hdr->args;
1597         struct nfs_pgio_res     *resp = &hdr->res;
1598
1599         if (resp->count < argp->count) {
1600                 static unsigned long    complain;
1601
1602                 /* This a short write! */
1603                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1604
1605                 /* Has the server at least made some progress? */
1606                 if (resp->count == 0) {
1607                         if (time_before(complain, jiffies)) {
1608                                 printk(KERN_WARNING
1609                                        "NFS: Server wrote zero bytes, expected %u.\n",
1610                                        argp->count);
1611                                 complain = jiffies + 300 * HZ;
1612                         }
1613                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1614                         task->tk_status = -EIO;
1615                         return;
1616                 }
1617
1618                 /* For non rpc-based layout drivers, retry-through-MDS */
1619                 if (!task->tk_ops) {
1620                         hdr->pnfs_error = -EAGAIN;
1621                         return;
1622                 }
1623
1624                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1625                 if (resp->verf->committed != NFS_UNSTABLE) {
1626                         /* Resend from where the server left off */
1627                         hdr->mds_offset += resp->count;
1628                         argp->offset += resp->count;
1629                         argp->pgbase += resp->count;
1630                         argp->count -= resp->count;
1631                 } else {
1632                         /* Resend as a stable write in order to avoid
1633                          * headaches in the case of a server crash.
1634                          */
1635                         argp->stable = NFS_FILE_SYNC;
1636                 }
1637                 rpc_restart_call_prepare(task);
1638         }
1639 }
1640
1641 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1642 {
1643         return wait_on_atomic_t(&cinfo->rpcs_out,
1644                         nfs_wait_atomic_killable, TASK_KILLABLE);
1645 }
1646
1647 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1648 {
1649         atomic_inc(&cinfo->rpcs_out);
1650 }
1651
1652 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1653 {
1654         if (atomic_dec_and_test(&cinfo->rpcs_out))
1655                 wake_up_atomic_t(&cinfo->rpcs_out);
1656 }
1657
1658 void nfs_commitdata_release(struct nfs_commit_data *data)
1659 {
1660         put_nfs_open_context(data->context);
1661         nfs_commit_free(data);
1662 }
1663 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1664
1665 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1666                         const struct nfs_rpc_ops *nfs_ops,
1667                         const struct rpc_call_ops *call_ops,
1668                         int how, int flags)
1669 {
1670         struct rpc_task *task;
1671         int priority = flush_task_priority(how);
1672         struct rpc_message msg = {
1673                 .rpc_argp = &data->args,
1674                 .rpc_resp = &data->res,
1675                 .rpc_cred = data->cred,
1676         };
1677         struct rpc_task_setup task_setup_data = {
1678                 .task = &data->task,
1679                 .rpc_client = clnt,
1680                 .rpc_message = &msg,
1681                 .callback_ops = call_ops,
1682                 .callback_data = data,
1683                 .workqueue = nfsiod_workqueue,
1684                 .flags = RPC_TASK_ASYNC | flags,
1685                 .priority = priority,
1686         };
1687         /* Set up the initial task struct.  */
1688         nfs_ops->commit_setup(data, &msg);
1689
1690         dprintk("NFS: initiated commit call\n");
1691
1692         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1693                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1694
1695         task = rpc_run_task(&task_setup_data);
1696         if (IS_ERR(task))
1697                 return PTR_ERR(task);
1698         if (how & FLUSH_SYNC)
1699                 rpc_wait_for_completion_task(task);
1700         rpc_put_task(task);
1701         return 0;
1702 }
1703 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1704
1705 static loff_t nfs_get_lwb(struct list_head *head)
1706 {
1707         loff_t lwb = 0;
1708         struct nfs_page *req;
1709
1710         list_for_each_entry(req, head, wb_list)
1711                 if (lwb < (req_offset(req) + req->wb_bytes))
1712                         lwb = req_offset(req) + req->wb_bytes;
1713
1714         return lwb;
1715 }
1716
1717 /*
1718  * Set up the argument/result storage required for the RPC call.
1719  */
1720 void nfs_init_commit(struct nfs_commit_data *data,
1721                      struct list_head *head,
1722                      struct pnfs_layout_segment *lseg,
1723                      struct nfs_commit_info *cinfo)
1724 {
1725         struct nfs_page *first = nfs_list_entry(head->next);
1726         struct inode *inode = d_inode(first->wb_context->dentry);
1727
1728         /* Set up the RPC argument and reply structs
1729          * NB: take care not to mess about with data->commit et al. */
1730
1731         list_splice_init(head, &data->pages);
1732
1733         data->inode       = inode;
1734         data->cred        = first->wb_context->cred;
1735         data->lseg        = lseg; /* reference transferred */
1736         /* only set lwb for pnfs commit */
1737         if (lseg)
1738                 data->lwb = nfs_get_lwb(&data->pages);
1739         data->mds_ops     = &nfs_commit_ops;
1740         data->completion_ops = cinfo->completion_ops;
1741         data->dreq        = cinfo->dreq;
1742
1743         data->args.fh     = NFS_FH(data->inode);
1744         /* Note: we always request a commit of the entire inode */
1745         data->args.offset = 0;
1746         data->args.count  = 0;
1747         data->context     = get_nfs_open_context(first->wb_context);
1748         data->res.fattr   = &data->fattr;
1749         data->res.verf    = &data->verf;
1750         nfs_fattr_init(&data->fattr);
1751 }
1752 EXPORT_SYMBOL_GPL(nfs_init_commit);
1753
1754 void nfs_retry_commit(struct list_head *page_list,
1755                       struct pnfs_layout_segment *lseg,
1756                       struct nfs_commit_info *cinfo,
1757                       u32 ds_commit_idx)
1758 {
1759         struct nfs_page *req;
1760
1761         while (!list_empty(page_list)) {
1762                 req = nfs_list_entry(page_list->next);
1763                 nfs_list_remove_request(req);
1764                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1765                 if (!cinfo->dreq)
1766                         nfs_clear_page_commit(req->wb_page);
1767                 nfs_unlock_and_release_request(req);
1768         }
1769 }
1770 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1771
1772 static void
1773 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1774                 struct nfs_page *req)
1775 {
1776         __set_page_dirty_nobuffers(req->wb_page);
1777 }
1778
1779 /*
1780  * Commit dirty pages
1781  */
1782 static int
1783 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1784                 struct nfs_commit_info *cinfo)
1785 {
1786         struct nfs_commit_data  *data;
1787
1788         /* another commit raced with us */
1789         if (list_empty(head))
1790                 return 0;
1791
1792         data = nfs_commitdata_alloc(true);
1793
1794         /* Set up the argument struct */
1795         nfs_init_commit(data, head, NULL, cinfo);
1796         atomic_inc(&cinfo->mds->rpcs_out);
1797         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1798                                    data->mds_ops, how, 0);
1799 }
1800
1801 /*
1802  * COMMIT call returned
1803  */
1804 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1805 {
1806         struct nfs_commit_data  *data = calldata;
1807
1808         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1809                                 task->tk_pid, task->tk_status);
1810
1811         /* Call the NFS version-specific code */
1812         NFS_PROTO(data->inode)->commit_done(task, data);
1813 }
1814
1815 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1816 {
1817         struct nfs_page *req;
1818         int status = data->task.tk_status;
1819         struct nfs_commit_info cinfo;
1820         struct nfs_server *nfss;
1821
1822         while (!list_empty(&data->pages)) {
1823                 req = nfs_list_entry(data->pages.next);
1824                 nfs_list_remove_request(req);
1825                 if (req->wb_page)
1826                         nfs_clear_page_commit(req->wb_page);
1827
1828                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1829                         req->wb_context->dentry->d_sb->s_id,
1830                         (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
1831                         req->wb_bytes,
1832                         (long long)req_offset(req));
1833                 if (status < 0) {
1834                         nfs_context_set_write_error(req->wb_context, status);
1835                         if (req->wb_page)
1836                                 nfs_inode_remove_request(req);
1837                         dprintk_cont(", error = %d\n", status);
1838                         goto next;
1839                 }
1840
1841                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1842                  * returned by the server against all stored verfs. */
1843                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1844                         /* We have a match */
1845                         if (req->wb_page)
1846                                 nfs_inode_remove_request(req);
1847                         dprintk_cont(" OK\n");
1848                         goto next;
1849                 }
1850                 /* We have a mismatch. Write the page again */
1851                 dprintk_cont(" mismatch\n");
1852                 nfs_mark_request_dirty(req);
1853                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1854         next:
1855                 nfs_unlock_and_release_request(req);
1856         }
1857         nfss = NFS_SERVER(data->inode);
1858         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1859                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1860
1861         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1862         nfs_commit_end(cinfo.mds);
1863 }
1864
1865 static void nfs_commit_release(void *calldata)
1866 {
1867         struct nfs_commit_data *data = calldata;
1868
1869         data->completion_ops->completion(data);
1870         nfs_commitdata_release(calldata);
1871 }
1872
1873 static const struct rpc_call_ops nfs_commit_ops = {
1874         .rpc_call_prepare = nfs_commit_prepare,
1875         .rpc_call_done = nfs_commit_done,
1876         .rpc_release = nfs_commit_release,
1877 };
1878
1879 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1880         .completion = nfs_commit_release_pages,
1881         .resched_write = nfs_commit_resched_write,
1882 };
1883
1884 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1885                             int how, struct nfs_commit_info *cinfo)
1886 {
1887         int status;
1888
1889         status = pnfs_commit_list(inode, head, how, cinfo);
1890         if (status == PNFS_NOT_ATTEMPTED)
1891                 status = nfs_commit_list(inode, head, how, cinfo);
1892         return status;
1893 }
1894
1895 int nfs_commit_inode(struct inode *inode, int how)
1896 {
1897         LIST_HEAD(head);
1898         struct nfs_commit_info cinfo;
1899         int may_wait = how & FLUSH_SYNC;
1900         int error = 0;
1901         int res;
1902
1903         nfs_init_cinfo_from_inode(&cinfo, inode);
1904         nfs_commit_begin(cinfo.mds);
1905         res = nfs_scan_commit(inode, &head, &cinfo);
1906         if (res)
1907                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1908         nfs_commit_end(cinfo.mds);
1909         if (error < 0)
1910                 goto out_error;
1911         if (!may_wait)
1912                 goto out_mark_dirty;
1913         error = wait_on_commit(cinfo.mds);
1914         if (error < 0)
1915                 return error;
1916         return res;
1917 out_error:
1918         res = error;
1919         /* Note: If we exit without ensuring that the commit is complete,
1920          * we must mark the inode as dirty. Otherwise, future calls to
1921          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1922          * that the data is on the disk.
1923          */
1924 out_mark_dirty:
1925         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1926         return res;
1927 }
1928 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1929
1930 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1931 {
1932         struct nfs_inode *nfsi = NFS_I(inode);
1933         int flags = FLUSH_SYNC;
1934         int ret = 0;
1935
1936         /* no commits means nothing needs to be done */
1937         if (!nfsi->commit_info.ncommit)
1938                 return ret;
1939
1940         if (wbc->sync_mode == WB_SYNC_NONE) {
1941                 /* Don't commit yet if this is a non-blocking flush and there
1942                  * are a lot of outstanding writes for this mapping.
1943                  */
1944                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1945                         goto out_mark_dirty;
1946
1947                 /* don't wait for the COMMIT response */
1948                 flags = 0;
1949         }
1950
1951         ret = nfs_commit_inode(inode, flags);
1952         if (ret >= 0) {
1953                 if (wbc->sync_mode == WB_SYNC_NONE) {
1954                         if (ret < wbc->nr_to_write)
1955                                 wbc->nr_to_write -= ret;
1956                         else
1957                                 wbc->nr_to_write = 0;
1958                 }
1959                 return 0;
1960         }
1961 out_mark_dirty:
1962         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1963         return ret;
1964 }
1965 EXPORT_SYMBOL_GPL(nfs_write_inode);
1966
1967 /*
1968  * Wrapper for filemap_write_and_wait_range()
1969  *
1970  * Needed for pNFS in order to ensure data becomes visible to the
1971  * client.
1972  */
1973 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1974                 loff_t lstart, loff_t lend)
1975 {
1976         int ret;
1977
1978         ret = filemap_write_and_wait_range(mapping, lstart, lend);
1979         if (ret == 0)
1980                 ret = pnfs_sync_inode(mapping->host, true);
1981         return ret;
1982 }
1983 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
1984
1985 /*
1986  * flush the inode to disk.
1987  */
1988 int nfs_wb_all(struct inode *inode)
1989 {
1990         int ret;
1991
1992         trace_nfs_writeback_inode_enter(inode);
1993
1994         ret = filemap_write_and_wait(inode->i_mapping);
1995         if (ret)
1996                 goto out;
1997         ret = nfs_commit_inode(inode, FLUSH_SYNC);
1998         if (ret < 0)
1999                 goto out;
2000         pnfs_sync_inode(inode, true);
2001         ret = 0;
2002
2003 out:
2004         trace_nfs_writeback_inode_exit(inode, ret);
2005         return ret;
2006 }
2007 EXPORT_SYMBOL_GPL(nfs_wb_all);
2008
2009 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2010 {
2011         struct nfs_page *req;
2012         int ret = 0;
2013
2014         wait_on_page_writeback(page);
2015
2016         /* blocking call to cancel all requests and join to a single (head)
2017          * request */
2018         req = nfs_lock_and_join_requests(page, false);
2019
2020         if (IS_ERR(req)) {
2021                 ret = PTR_ERR(req);
2022         } else if (req) {
2023                 /* all requests from this page have been cancelled by
2024                  * nfs_lock_and_join_requests, so just remove the head
2025                  * request from the inode / page_private pointer and
2026                  * release it */
2027                 nfs_inode_remove_request(req);
2028                 nfs_unlock_and_release_request(req);
2029         }
2030
2031         return ret;
2032 }
2033
2034 /*
2035  * Write back all requests on one page - we do this before reading it.
2036  */
2037 int nfs_wb_page(struct inode *inode, struct page *page)
2038 {
2039         loff_t range_start = page_file_offset(page);
2040         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2041         struct writeback_control wbc = {
2042                 .sync_mode = WB_SYNC_ALL,
2043                 .nr_to_write = 0,
2044                 .range_start = range_start,
2045                 .range_end = range_end,
2046         };
2047         int ret;
2048
2049         trace_nfs_writeback_page_enter(inode);
2050
2051         for (;;) {
2052                 wait_on_page_writeback(page);
2053                 if (clear_page_dirty_for_io(page)) {
2054                         ret = nfs_writepage_locked(page, &wbc);
2055                         if (ret < 0)
2056                                 goto out_error;
2057                         continue;
2058                 }
2059                 ret = 0;
2060                 if (!PagePrivate(page))
2061                         break;
2062                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2063                 if (ret < 0)
2064                         goto out_error;
2065         }
2066 out_error:
2067         trace_nfs_writeback_page_exit(inode, ret);
2068         return ret;
2069 }
2070
2071 #ifdef CONFIG_MIGRATION
2072 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2073                 struct page *page, enum migrate_mode mode)
2074 {
2075         /*
2076          * If PagePrivate is set, then the page is currently associated with
2077          * an in-progress read or write request. Don't try to migrate it.
2078          *
2079          * FIXME: we could do this in principle, but we'll need a way to ensure
2080          *        that we can safely release the inode reference while holding
2081          *        the page lock.
2082          */
2083         if (PagePrivate(page))
2084                 return -EBUSY;
2085
2086         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2087                 return -EBUSY;
2088
2089         return migrate_page(mapping, newpage, page, mode);
2090 }
2091 #endif
2092
2093 int __init nfs_init_writepagecache(void)
2094 {
2095         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2096                                              sizeof(struct nfs_pgio_header),
2097                                              0, SLAB_HWCACHE_ALIGN,
2098                                              NULL);
2099         if (nfs_wdata_cachep == NULL)
2100                 return -ENOMEM;
2101
2102         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2103                                                      nfs_wdata_cachep);
2104         if (nfs_wdata_mempool == NULL)
2105                 goto out_destroy_write_cache;
2106
2107         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2108                                              sizeof(struct nfs_commit_data),
2109                                              0, SLAB_HWCACHE_ALIGN,
2110                                              NULL);
2111         if (nfs_cdata_cachep == NULL)
2112                 goto out_destroy_write_mempool;
2113
2114         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2115                                                       nfs_cdata_cachep);
2116         if (nfs_commit_mempool == NULL)
2117                 goto out_destroy_commit_cache;
2118
2119         /*
2120          * NFS congestion size, scale with available memory.
2121          *
2122          *  64MB:    8192k
2123          * 128MB:   11585k
2124          * 256MB:   16384k
2125          * 512MB:   23170k
2126          *   1GB:   32768k
2127          *   2GB:   46340k
2128          *   4GB:   65536k
2129          *   8GB:   92681k
2130          *  16GB:  131072k
2131          *
2132          * This allows larger machines to have larger/more transfers.
2133          * Limit the default to 256M
2134          */
2135         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
2136         if (nfs_congestion_kb > 256*1024)
2137                 nfs_congestion_kb = 256*1024;
2138
2139         return 0;
2140
2141 out_destroy_commit_cache:
2142         kmem_cache_destroy(nfs_cdata_cachep);
2143 out_destroy_write_mempool:
2144         mempool_destroy(nfs_wdata_mempool);
2145 out_destroy_write_cache:
2146         kmem_cache_destroy(nfs_wdata_cachep);
2147         return -ENOMEM;
2148 }
2149
2150 void nfs_destroy_writepagecache(void)
2151 {
2152         mempool_destroy(nfs_commit_mempool);
2153         kmem_cache_destroy(nfs_cdata_cachep);
2154         mempool_destroy(nfs_wdata_mempool);
2155         kmem_cache_destroy(nfs_wdata_cachep);
2156 }
2157
2158 static const struct nfs_rw_ops nfs_rw_write_ops = {
2159         .rw_alloc_header        = nfs_writehdr_alloc,
2160         .rw_free_header         = nfs_writehdr_free,
2161         .rw_done                = nfs_writeback_done,
2162         .rw_result              = nfs_writeback_result,
2163         .rw_initiate            = nfs_initiate_write,
2164 };