fs/nfs/write.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/write.c
   4  *
   5  * Write file data over NFS.
   6  *
   7  * Copyright (C) 1996, 1997, Olaf Kirch <[email protected]>
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/mm.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/file.h>
  15 #include <linux/writeback.h>
  16 #include <linux/swap.h>
  17 #include <linux/migrate.h>
  18
  19 #include <linux/sunrpc/clnt.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/export.h>
  25 #include <linux/freezer.h>
  26 #include <linux/wait.h>
  27 #include <linux/iversion.h>
  28
  29 #include <linux/uaccess.h>
  30 #include <linux/sched/mm.h>
  31
  32 #include "delegation.h"
  33 #include "internal.h"
  34 #include "iostat.h"
  35 #include "nfs4_fs.h"
  36 #include "fscache.h"
  37 #include "pnfs.h"
  38
  39 #include "nfstrace.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  42
  43 #define MIN_POOL_WRITE          (32)
  44 #define MIN_POOL_COMMIT         (4)
  45
  46 struct nfs_io_completion {
  47         void (*complete)(void *data);
  48         void *data;
  49         struct kref refcount;
  50 };
  51
  52 /*
  53  * Local function declarations
  54  */
  55 static void nfs_redirty_request(struct nfs_page *req);
  56 static const struct rpc_call_ops nfs_commit_ops;
  57 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  58 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  59 static const struct nfs_rw_ops nfs_rw_write_ops;
  60 static void nfs_clear_request_commit(struct nfs_page *req);
  61 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  62                                       struct inode *inode);
  63 static struct nfs_page *
  64 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  65                                                 struct page *page);
  66
  67 static struct kmem_cache *nfs_wdata_cachep;
  68 static mempool_t *nfs_wdata_mempool;
  69 static struct kmem_cache *nfs_cdata_cachep;
  70 static mempool_t *nfs_commit_mempool;
  71
  72 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
  73 {
  74         struct nfs_commit_data *p;
  75
  76         if (never_fail)
  77                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  78         else {
  79                 /* It is OK to do some reclaim, not no safe to wait
  80                  * for anything to be returned to the pool.
  81                  * mempool_alloc() cannot handle that particular combination,
  82                  * so we need two separate attempts.
  83                  */
  84                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  85                 if (!p)
  86                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
  87                                              __GFP_NOWARN | __GFP_NORETRY);
  88                 if (!p)
  89                         return NULL;
  90         }
  91
  92         memset(p, 0, sizeof(*p));
  93         INIT_LIST_HEAD(&p->pages);
  94         return p;
  95 }
  96 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  97
  98 void nfs_commit_free(struct nfs_commit_data *p)
  99 {
 100         mempool_free(p, nfs_commit_mempool);
 101 }
 102 EXPORT_SYMBOL_GPL(nfs_commit_free);
 103
 104 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
 105 {
 106         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
 107
 108         memset(p, 0, sizeof(*p));
 109         p->rw_mode = FMODE_WRITE;
 110         return p;
 111 }
 112
 113 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 114 {
 115         mempool_free(hdr, nfs_wdata_mempool);
 116 }
 117
 118 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 119 {
 120         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 121 }
 122
 123 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 124                 void (*complete)(void *), void *data)
 125 {
 126         ioc->complete = complete;
 127         ioc->data = data;
 128         kref_init(&ioc->refcount);
 129 }
 130
 131 static void nfs_io_completion_release(struct kref *kref)
 132 {
 133         struct nfs_io_completion *ioc = container_of(kref,
 134                         struct nfs_io_completion, refcount);
 135         ioc->complete(ioc->data);
 136         kfree(ioc);
 137 }
 138
 139 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 140 {
 141         if (ioc != NULL)
 142                 kref_get(&ioc->refcount);
 143 }
 144
 145 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 146 {
 147         if (ioc != NULL)
 148                 kref_put(&ioc->refcount, nfs_io_completion_release);
 149 }
 150
 151 static struct nfs_page *
 152 nfs_page_private_request(struct page *page)
 153 {
 154         if (!PagePrivate(page))
 155                 return NULL;
 156         return (struct nfs_page *)page_private(page);
 157 }
 158
 159 /*
 160  * nfs_page_find_head_request_locked - find head request associated with @page
 161  *
 162  * must be called while holding the inode lock.
 163  *
 164  * returns matching head request with reference held, or NULL if not found.
 165  */
 166 static struct nfs_page *
 167 nfs_page_find_private_request(struct page *page)
 168 {
 169         struct address_space *mapping = page_file_mapping(page);
 170         struct nfs_page *req;
 171
 172         if (!PagePrivate(page))
 173                 return NULL;
 174         spin_lock(&mapping->private_lock);
 175         req = nfs_page_private_request(page);
 176         if (req) {
 177                 WARN_ON_ONCE(req->wb_head != req);
 178                 kref_get(&req->wb_kref);
 179         }
 180         spin_unlock(&mapping->private_lock);
 181         return req;
 182 }
 183
 184 static struct nfs_page *
 185 nfs_page_find_swap_request(struct page *page)
 186 {
 187         struct inode *inode = page_file_mapping(page)->host;
 188         struct nfs_inode *nfsi = NFS_I(inode);
 189         struct nfs_page *req = NULL;
 190         if (!PageSwapCache(page))
 191                 return NULL;
 192         mutex_lock(&nfsi->commit_mutex);
 193         if (PageSwapCache(page)) {
 194                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 195                         page);
 196                 if (req) {
 197                         WARN_ON_ONCE(req->wb_head != req);
 198                         kref_get(&req->wb_kref);
 199                 }
 200         }
 201         mutex_unlock(&nfsi->commit_mutex);
 202         return req;
 203 }
 204
 205 /*
 206  * nfs_page_find_head_request - find head request associated with @page
 207  *
 208  * returns matching head request with reference held, or NULL if not found.
 209  */
 210 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 211 {
 212         struct nfs_page *req;
 213
 214         req = nfs_page_find_private_request(page);
 215         if (!req)
 216                 req = nfs_page_find_swap_request(page);
 217         return req;
 218 }
 219
 220 /* Adjust the file length if we're writing beyond the end */
 221 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 222 {
 223         struct inode *inode = page_file_mapping(page)->host;
 224         loff_t end, i_size;
 225         pgoff_t end_index;
 226
 227         spin_lock(&inode->i_lock);
 228         i_size = i_size_read(inode);
 229         end_index = (i_size - 1) >> PAGE_SHIFT;
 230         if (i_size > 0 && page_index(page) < end_index)
 231                 goto out;
 232         end = page_file_offset(page) + ((loff_t)offset+count);
 233         if (i_size >= end)
 234                 goto out;
 235         i_size_write(inode, end);
 236         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 237         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 238 out:
 239         spin_unlock(&inode->i_lock);
 240 }
 241
 242 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 243 static void nfs_set_pageerror(struct address_space *mapping)
 244 {
 245         nfs_zap_mapping(mapping->host, mapping);
 246 }
 247
 248 static void nfs_mapping_set_error(struct page *page, int error)
 249 {
 250         SetPageError(page);
 251         mapping_set_error(page_file_mapping(page), error);
 252 }
 253
 254 /*
 255  * nfs_page_group_search_locked
 256  * @head - head request of page group
 257  * @page_offset - offset into page
 258  *
 259  * Search page group with head @head to find a request that contains the
 260  * page offset @page_offset.
 261  *
 262  * Returns a pointer to the first matching nfs request, or NULL if no
 263  * match is found.
 264  *
 265  * Must be called with the page group lock held
 266  */
 267 static struct nfs_page *
 268 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 269 {
 270         struct nfs_page *req;
 271
 272         req = head;
 273         do {
 274                 if (page_offset >= req->wb_pgbase &&
 275                     page_offset < (req->wb_pgbase + req->wb_bytes))
 276                         return req;
 277
 278                 req = req->wb_this_page;
 279         } while (req != head);
 280
 281         return NULL;
 282 }
 283
 284 /*
 285  * nfs_page_group_covers_page
 286  * @head - head request of page group
 287  *
 288  * Return true if the page group with head @head covers the whole page,
 289  * returns false otherwise
 290  */
 291 static bool nfs_page_group_covers_page(struct nfs_page *req)
 292 {
 293         struct nfs_page *tmp;
 294         unsigned int pos = 0;
 295         unsigned int len = nfs_page_length(req->wb_page);
 296
 297         nfs_page_group_lock(req);
 298
 299         for (;;) {
 300                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 301                 if (!tmp)
 302                         break;
 303                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 304         }
 305
 306         nfs_page_group_unlock(req);
 307         return pos >= len;
 308 }
 309
 310 /* We can set the PG_uptodate flag if we see that a write request
 311  * covers the full page.
 312  */
 313 static void nfs_mark_uptodate(struct nfs_page *req)
 314 {
 315         if (PageUptodate(req->wb_page))
 316                 return;
 317         if (!nfs_page_group_covers_page(req))
 318                 return;
 319         SetPageUptodate(req->wb_page);
 320 }
 321
 322 static int wb_priority(struct writeback_control *wbc)
 323 {
 324         int ret = 0;
 325
 326         if (wbc->sync_mode == WB_SYNC_ALL)
 327                 ret = FLUSH_COND_STABLE;
 328         return ret;
 329 }
 330
 331 /*
 332  * NFS congestion control
 333  */
 334
 335 int nfs_congestion_kb;
 336
 337 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 338 #define NFS_CONGESTION_OFF_THRESH       \
 339         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 340
 341 static void nfs_set_page_writeback(struct page *page)
 342 {
 343         struct inode *inode = page_file_mapping(page)->host;
 344         struct nfs_server *nfss = NFS_SERVER(inode);
 345         int ret = test_set_page_writeback(page);
 346
 347         WARN_ON_ONCE(ret != 0);
 348
 349         if (atomic_long_inc_return(&nfss->writeback) >
 350                         NFS_CONGESTION_ON_THRESH)
 351                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 352 }
 353
 354 static void nfs_end_page_writeback(struct nfs_page *req)
 355 {
 356         struct inode *inode = page_file_mapping(req->wb_page)->host;
 357         struct nfs_server *nfss = NFS_SERVER(inode);
 358         bool is_done;
 359
 360         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 361         nfs_unlock_request(req);
 362         if (!is_done)
 363                 return;
 364
 365         end_page_writeback(req->wb_page);
 366         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 367                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 368 }
 369
 370 /*
 371  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 372  *
 373  * this is a helper function for nfs_lock_and_join_requests
 374  *
 375  * @inode - inode associated with request page group, must be holding inode lock
 376  * @head  - head request of page group, must be holding head lock
 377  * @req   - request that couldn't lock and needs to wait on the req bit lock
 378  *
 379  * NOTE: this must be called holding page_group bit lock
 380  *       which will be released before returning.
 381  *
 382  * returns 0 on success, < 0 on error.
 383  */
 384 static void
 385 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
 386                           struct nfs_page *req)
 387 {
 388         struct nfs_page *tmp;
 389
 390         /* relinquish all the locks successfully grabbed this run */
 391         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 392                 if (!kref_read(&tmp->wb_kref))
 393                         continue;
 394                 nfs_unlock_and_release_request(tmp);
 395         }
 396 }
 397
 398 /*
 399  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 400  *
 401  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 402  * @old_head - the old head of the list
 403  *
 404  * All subrequests must be locked and removed from all lists, so at this point
 405  * they are only "active" in this function, and possibly in nfs_wait_on_request
 406  * with a reference held by some other context.
 407  */
 408 static void
 409 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 410                                  struct nfs_page *old_head,
 411                                  struct inode *inode)
 412 {
 413         while (destroy_list) {
 414                 struct nfs_page *subreq = destroy_list;
 415
 416                 destroy_list = (subreq->wb_this_page == old_head) ?
 417                                    NULL : subreq->wb_this_page;
 418
 419                 WARN_ON_ONCE(old_head != subreq->wb_head);
 420
 421                 /* make sure old group is not used */
 422                 subreq->wb_this_page = subreq;
 423
 424                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 425
 426                 /* Note: races with nfs_page_group_destroy() */
 427                 if (!kref_read(&subreq->wb_kref)) {
 428                         /* Check if we raced with nfs_page_group_destroy() */
 429                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
 430                                 nfs_free_request(subreq);
 431                         continue;
 432                 }
 433
 434                 subreq->wb_head = subreq;
 435
 436                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 437                         nfs_release_request(subreq);
 438                         atomic_long_dec(&NFS_I(inode)->nrequests);
 439                 }
 440
 441                 /* subreq is now totally disconnected from page group or any
 442                  * write / commit lists. last chance to wake any waiters */
 443                 nfs_unlock_and_release_request(subreq);
 444         }
 445 }
 446
 447 /*
 448  * nfs_lock_and_join_requests - join all subreqs to the head req and return
 449  *                              a locked reference, cancelling any pending
 450  *                              operations for this page.
 451  *
 452  * @page - the page used to lookup the "page group" of nfs_page structures
 453  *
 454  * This function joins all sub requests to the head request by first
 455  * locking all requests in the group, cancelling any pending operations
 456  * and finally updating the head request to cover the whole range covered by
 457  * the (former) group.  All subrequests are removed from any write or commit
 458  * lists, unlinked from the group and destroyed.
 459  *
 460  * Returns a locked, referenced pointer to the head request - which after
 461  * this call is guaranteed to be the only request associated with the page.
 462  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 463  * error was encountered.
 464  */
 465 static struct nfs_page *
 466 nfs_lock_and_join_requests(struct page *page)
 467 {
 468         struct inode *inode = page_file_mapping(page)->host;
 469         struct nfs_page *head, *subreq;
 470         struct nfs_page *destroy_list = NULL;
 471         unsigned int total_bytes;
 472         int ret;
 473
 474 try_again:
 475         /*
 476          * A reference is taken only on the head request which acts as a
 477          * reference to the whole page group - the group will not be destroyed
 478          * until the head reference is released.
 479          */
 480         head = nfs_page_find_head_request(page);
 481         if (!head)
 482                 return NULL;
 483
 484         /* lock the page head first in order to avoid an ABBA inefficiency */
 485         if (!nfs_lock_request(head)) {
 486                 ret = nfs_wait_on_request(head);
 487                 nfs_release_request(head);
 488                 if (ret < 0)
 489                         return ERR_PTR(ret);
 490                 goto try_again;
 491         }
 492
 493         /* Ensure that nobody removed the request before we locked it */
 494         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
 495                 nfs_unlock_and_release_request(head);
 496                 goto try_again;
 497         }
 498
 499         ret = nfs_page_group_lock(head);
 500         if (ret < 0)
 501                 goto release_request;
 502
 503         /* lock each request in the page group */
 504         total_bytes = head->wb_bytes;
 505         for (subreq = head->wb_this_page; subreq != head;
 506                         subreq = subreq->wb_this_page) {
 507
 508                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
 509                         if (subreq->wb_offset == head->wb_offset + total_bytes)
 510                                 total_bytes += subreq->wb_bytes;
 511                         continue;
 512                 }
 513
 514                 while (!nfs_lock_request(subreq)) {
 515                         /*
 516                          * Unlock page to allow nfs_page_group_sync_on_bit()
 517                          * to succeed
 518                          */
 519                         nfs_page_group_unlock(head);
 520                         ret = nfs_wait_on_request(subreq);
 521                         if (!ret)
 522                                 ret = nfs_page_group_lock(head);
 523                         if (ret < 0) {
 524                                 nfs_unroll_locks(inode, head, subreq);
 525                                 nfs_release_request(subreq);
 526                                 goto release_request;
 527                         }
 528                 }
 529                 /*
 530                  * Subrequests are always contiguous, non overlapping
 531                  * and in order - but may be repeated (mirrored writes).
 532                  */
 533                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
 534                         /* keep track of how many bytes this group covers */
 535                         total_bytes += subreq->wb_bytes;
 536                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
 537                             ((subreq->wb_offset + subreq->wb_bytes) >
 538                              (head->wb_offset + total_bytes)))) {
 539                         nfs_page_group_unlock(head);
 540                         nfs_unroll_locks(inode, head, subreq);
 541                         nfs_unlock_and_release_request(subreq);
 542                         ret = -EIO;
 543                         goto release_request;
 544                 }
 545         }
 546
 547         /* Now that all requests are locked, make sure they aren't on any list.
 548          * Commit list removal accounting is done after locks are dropped */
 549         subreq = head;
 550         do {
 551                 nfs_clear_request_commit(subreq);
 552                 subreq = subreq->wb_this_page;
 553         } while (subreq != head);
 554
 555         /* unlink subrequests from head, destroy them later */
 556         if (head->wb_this_page != head) {
 557                 /* destroy list will be terminated by head */
 558                 destroy_list = head->wb_this_page;
 559                 head->wb_this_page = head;
 560
 561                 /* change head request to cover whole range that
 562                  * the former page group covered */
 563                 head->wb_bytes = total_bytes;
 564         }
 565
 566         /* Postpone destruction of this request */
 567         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
 568                 set_bit(PG_INODE_REF, &head->wb_flags);
 569                 kref_get(&head->wb_kref);
 570                 atomic_long_inc(&NFS_I(inode)->nrequests);
 571         }
 572
 573         nfs_page_group_unlock(head);
 574
 575         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 576
 577         /* Did we lose a race with nfs_inode_remove_request()? */
 578         if (!(PagePrivate(page) || PageSwapCache(page))) {
 579                 nfs_unlock_and_release_request(head);
 580                 return NULL;
 581         }
 582
 583         /* still holds ref on head from nfs_page_find_head_request
 584          * and still has lock on head from lock loop */
 585         return head;
 586
 587 release_request:
 588         nfs_unlock_and_release_request(head);
 589         return ERR_PTR(ret);
 590 }
 591
 592 static void nfs_write_error(struct nfs_page *req, int error)
 593 {
 594         nfs_mapping_set_error(req->wb_page, error);
 595         nfs_end_page_writeback(req);
 596         nfs_release_request(req);
 597 }
 598
 599 static bool
 600 nfs_error_is_fatal_on_server(int err)
 601 {
 602         switch (err) {
 603         case 0:
 604         case -ERESTARTSYS:
 605         case -EINTR:
 606                 return false;
 607         }
 608         return nfs_error_is_fatal(err);
 609 }
 610
 611 /*
 612  * Find an associated nfs write request, and prepare to flush it out
 613  * May return an error if the user signalled nfs_wait_on_request().
 614  */
 615 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 616                                 struct page *page)
 617 {
 618         struct address_space *mapping;
 619         struct nfs_page *req;
 620         int ret = 0;
 621
 622         req = nfs_lock_and_join_requests(page);
 623         if (!req)
 624                 goto out;
 625         ret = PTR_ERR(req);
 626         if (IS_ERR(req))
 627                 goto out;
 628
 629         nfs_set_page_writeback(page);
 630         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 631
 632         /* If there is a fatal error that covers this write, just exit */
 633         ret = 0;
 634         mapping = page_file_mapping(page);
 635         if (test_bit(AS_ENOSPC, &mapping->flags) ||
 636             test_bit(AS_EIO, &mapping->flags))
 637                 goto out_launder;
 638
 639         if (!nfs_pageio_add_request(pgio, req)) {
 640                 ret = pgio->pg_error;
 641                 /*
 642                  * Remove the problematic req upon fatal errors on the server
 643                  */
 644                 if (nfs_error_is_fatal(ret)) {
 645                         if (nfs_error_is_fatal_on_server(ret))
 646                                 goto out_launder;
 647                 } else
 648                         ret = -EAGAIN;
 649                 nfs_redirty_request(req);
 650         } else
 651                 nfs_add_stats(page_file_mapping(page)->host,
 652                                 NFSIOS_WRITEPAGES, 1);
 653 out:
 654         return ret;
 655 out_launder:
 656         nfs_write_error(req, ret);
 657         return 0;
 658 }
 659
 660 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 661                             struct nfs_pageio_descriptor *pgio)
 662 {
 663         int ret;
 664
 665         nfs_pageio_cond_complete(pgio, page_index(page));
 666         ret = nfs_page_async_flush(pgio, page);
 667         if (ret == -EAGAIN) {
 668                 redirty_page_for_writepage(wbc, page);
 669                 ret = 0;
 670         }
 671         return ret;
 672 }
 673
 674 /*
 675  * Write an mmapped page to the server.
 676  */
 677 static int nfs_writepage_locked(struct page *page,
 678                                 struct writeback_control *wbc)
 679 {
 680         struct nfs_pageio_descriptor pgio;
 681         struct inode *inode = page_file_mapping(page)->host;
 682         int err;
 683
 684         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 685         nfs_pageio_init_write(&pgio, inode, 0,
 686                                 false, &nfs_async_write_completion_ops);
 687         err = nfs_do_writepage(page, wbc, &pgio);
 688         nfs_pageio_complete(&pgio);
 689         if (err < 0)
 690                 return err;
 691         if (pgio.pg_error < 0)
 692                 return pgio.pg_error;
 693         return 0;
 694 }
 695
 696 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 697 {
 698         int ret;
 699
 700         ret = nfs_writepage_locked(page, wbc);
 701         unlock_page(page);
 702         return ret;
 703 }
 704
 705 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 706 {
 707         int ret;
 708
 709         ret = nfs_do_writepage(page, wbc, data);
 710         unlock_page(page);
 711         return ret;
 712 }
 713
 714 static void nfs_io_completion_commit(void *inode)
 715 {
 716         nfs_commit_inode(inode, 0);
 717 }
 718
 719 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 720 {
 721         struct inode *inode = mapping->host;
 722         struct nfs_pageio_descriptor pgio;
 723         struct nfs_io_completion *ioc;
 724         unsigned int pflags = memalloc_nofs_save();
 725         int err;
 726
 727         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 728
 729         ioc = nfs_io_completion_alloc(GFP_NOFS);
 730         if (ioc)
 731                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 732
 733         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 734                                 &nfs_async_write_completion_ops);
 735         pgio.pg_io_completion = ioc;
 736         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 737         nfs_pageio_complete(&pgio);
 738         nfs_io_completion_put(ioc);
 739
 740         memalloc_nofs_restore(pflags);
 741
 742         if (err < 0)
 743                 goto out_err;
 744         err = pgio.pg_error;
 745         if (err < 0)
 746                 goto out_err;
 747         return 0;
 748 out_err:
 749         return err;
 750 }
 751
 752 /*
 753  * Insert a write request into an inode
 754  */
 755 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 756 {
 757         struct address_space *mapping = page_file_mapping(req->wb_page);
 758         struct nfs_inode *nfsi = NFS_I(inode);
 759
 760         WARN_ON_ONCE(req->wb_this_page != req);
 761
 762         /* Lock the request! */
 763         nfs_lock_request(req);
 764
 765         /*
 766          * Swap-space should not get truncated. Hence no need to plug the race
 767          * with invalidate/truncate.
 768          */
 769         spin_lock(&mapping->private_lock);
 770         if (!nfs_have_writebacks(inode) &&
 771             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 772                 inode_inc_iversion_raw(inode);
 773         if (likely(!PageSwapCache(req->wb_page))) {
 774                 set_bit(PG_MAPPED, &req->wb_flags);
 775                 SetPagePrivate(req->wb_page);
 776                 set_page_private(req->wb_page, (unsigned long)req);
 777         }
 778         spin_unlock(&mapping->private_lock);
 779         atomic_long_inc(&nfsi->nrequests);
 780         /* this a head request for a page group - mark it as having an
 781          * extra reference so sub groups can follow suit.
 782          * This flag also informs pgio layer when to bump nrequests when
 783          * adding subrequests. */
 784         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 785         kref_get(&req->wb_kref);
 786 }
 787
 788 /*
 789  * Remove a write request from an inode
 790  */
 791 static void nfs_inode_remove_request(struct nfs_page *req)
 792 {
 793         struct address_space *mapping = page_file_mapping(req->wb_page);
 794         struct inode *inode = mapping->host;
 795         struct nfs_inode *nfsi = NFS_I(inode);
 796         struct nfs_page *head;
 797
 798         atomic_long_dec(&nfsi->nrequests);
 799         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 800                 head = req->wb_head;
 801
 802                 spin_lock(&mapping->private_lock);
 803                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 804                         set_page_private(head->wb_page, 0);
 805                         ClearPagePrivate(head->wb_page);
 806                         clear_bit(PG_MAPPED, &head->wb_flags);
 807                 }
 808                 spin_unlock(&mapping->private_lock);
 809         }
 810
 811         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 812                 nfs_release_request(req);
 813 }
 814
 815 static void
 816 nfs_mark_request_dirty(struct nfs_page *req)
 817 {
 818         if (req->wb_page)
 819                 __set_page_dirty_nobuffers(req->wb_page);
 820 }
 821
 822 /*
 823  * nfs_page_search_commits_for_head_request_locked
 824  *
 825  * Search through commit lists on @inode for the head request for @page.
 826  * Must be called while holding the inode (which is cinfo) lock.
 827  *
 828  * Returns the head request if found, or NULL if not found.
 829  */
 830 static struct nfs_page *
 831 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 832                                                 struct page *page)
 833 {
 834         struct nfs_page *freq, *t;
 835         struct nfs_commit_info cinfo;
 836         struct inode *inode = &nfsi->vfs_inode;
 837
 838         nfs_init_cinfo_from_inode(&cinfo, inode);
 839
 840         /* search through pnfs commit lists */
 841         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 842         if (freq)
 843                 return freq->wb_head;
 844
 845         /* Linearly search the commit list for the correct request */
 846         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 847                 if (freq->wb_page == page)
 848                         return freq->wb_head;
 849         }
 850
 851         return NULL;
 852 }
 853
 854 /**
 855  * nfs_request_add_commit_list_locked - add request to a commit list
 856  * @req: pointer to a struct nfs_page
 857  * @dst: commit list head
 858  * @cinfo: holds list lock and accounting info
 859  *
 860  * This sets the PG_CLEAN bit, updates the cinfo count of
 861  * number of outstanding requests requiring a commit as well as
 862  * the MM page stats.
 863  *
 864  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 865  * nfs_page lock.
 866  */
 867 void
 868 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 869                             struct nfs_commit_info *cinfo)
 870 {
 871         set_bit(PG_CLEAN, &req->wb_flags);
 872         nfs_list_add_request(req, dst);
 873         atomic_long_inc(&cinfo->mds->ncommit);
 874 }
 875 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 876
 877 /**
 878  * nfs_request_add_commit_list - add request to a commit list
 879  * @req: pointer to a struct nfs_page
 880  * @cinfo: holds list lock and accounting info
 881  *
 882  * This sets the PG_CLEAN bit, updates the cinfo count of
 883  * number of outstanding requests requiring a commit as well as
 884  * the MM page stats.
 885  *
 886  * The caller must _not_ hold the cinfo->lock, but must be
 887  * holding the nfs_page lock.
 888  */
 889 void
 890 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 891 {
 892         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 893         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 894         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 895         if (req->wb_page)
 896                 nfs_mark_page_unstable(req->wb_page, cinfo);
 897 }
 898 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 899
 900 /**
 901  * nfs_request_remove_commit_list - Remove request from a commit list
 902  * @req: pointer to a nfs_page
 903  * @cinfo: holds list lock and accounting info
 904  *
 905  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 906  * number of outstanding requests requiring a commit
 907  * It does not update the MM page stats.
 908  *
 909  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 910  */
 911 void
 912 nfs_request_remove_commit_list(struct nfs_page *req,
 913                                struct nfs_commit_info *cinfo)
 914 {
 915         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 916                 return;
 917         nfs_list_remove_request(req);
 918         atomic_long_dec(&cinfo->mds->ncommit);
 919 }
 920 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 921
 922 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 923                                       struct inode *inode)
 924 {
 925         cinfo->inode = inode;
 926         cinfo->mds = &NFS_I(inode)->commit_info;
 927         cinfo->ds = pnfs_get_ds_info(inode);
 928         cinfo->dreq = NULL;
 929         cinfo->completion_ops = &nfs_commit_completion_ops;
 930 }
 931
 932 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 933                     struct inode *inode,
 934                     struct nfs_direct_req *dreq)
 935 {
 936         if (dreq)
 937                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 938         else
 939                 nfs_init_cinfo_from_inode(cinfo, inode);
 940 }
 941 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 942
 943 /*
 944  * Add a request to the inode's commit list.
 945  */
 946 void
 947 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 948                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 949 {
 950         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 951                 return;
 952         nfs_request_add_commit_list(req, cinfo);
 953 }
 954
 955 static void
 956 nfs_clear_page_commit(struct page *page)
 957 {
 958         dec_node_page_state(page, NR_UNSTABLE_NFS);
 959         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 960                     WB_RECLAIMABLE);
 961 }
 962
 963 /* Called holding the request lock on @req */
 964 static void
 965 nfs_clear_request_commit(struct nfs_page *req)
 966 {
 967         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 968                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 969                 struct inode *inode = d_inode(ctx->dentry);
 970                 struct nfs_commit_info cinfo;
 971
 972                 nfs_init_cinfo_from_inode(&cinfo, inode);
 973                 mutex_lock(&NFS_I(inode)->commit_mutex);
 974                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 975                         nfs_request_remove_commit_list(req, &cinfo);
 976                 }
 977                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 978                 nfs_clear_page_commit(req->wb_page);
 979         }
 980 }
 981
 982 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 983 {
 984         if (hdr->verf.committed == NFS_DATA_SYNC)
 985                 return hdr->lseg == NULL;
 986         return hdr->verf.committed != NFS_FILE_SYNC;
 987 }
 988
 989 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 990 {
 991         nfs_io_completion_get(hdr->io_completion);
 992 }
 993
 994 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 995 {
 996         struct nfs_commit_info cinfo;
 997         unsigned long bytes = 0;
 998
 999         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1000                 goto out;
1001         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1002         while (!list_empty(&hdr->pages)) {
1003                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1004
1005                 bytes += req->wb_bytes;
1006                 nfs_list_remove_request(req);
1007                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1008                     (hdr->good_bytes < bytes)) {
1009                         nfs_set_pageerror(page_file_mapping(req->wb_page));
1010                         nfs_mapping_set_error(req->wb_page, hdr->error);
1011                         goto remove_req;
1012                 }
1013                 if (nfs_write_need_commit(hdr)) {
1014                         /* Reset wb_nio, since the write was successful. */
1015                         req->wb_nio = 0;
1016                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1017                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1018                                 hdr->pgio_mirror_idx);
1019                         goto next;
1020                 }
1021 remove_req:
1022                 nfs_inode_remove_request(req);
1023 next:
1024                 nfs_end_page_writeback(req);
1025                 nfs_release_request(req);
1026         }
1027 out:
1028         nfs_io_completion_put(hdr->io_completion);
1029         hdr->release(hdr);
1030 }
1031
1032 unsigned long
1033 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1034 {
1035         return atomic_long_read(&cinfo->mds->ncommit);
1036 }
1037
1038 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1039 int
1040 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1041                      struct nfs_commit_info *cinfo, int max)
1042 {
1043         struct nfs_page *req, *tmp;
1044         int ret = 0;
1045
1046 restart:
1047         list_for_each_entry_safe(req, tmp, src, wb_list) {
1048                 kref_get(&req->wb_kref);
1049                 if (!nfs_lock_request(req)) {
1050                         int status;
1051
1052                         /* Prevent deadlock with nfs_lock_and_join_requests */
1053                         if (!list_empty(dst)) {
1054                                 nfs_release_request(req);
1055                                 continue;
1056                         }
1057                         /* Ensure we make progress to prevent livelock */
1058                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1059                         status = nfs_wait_on_request(req);
1060                         nfs_release_request(req);
1061                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1062                         if (status < 0)
1063                                 break;
1064                         goto restart;
1065                 }
1066                 nfs_request_remove_commit_list(req, cinfo);
1067                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1068                 nfs_list_add_request(req, dst);
1069                 ret++;
1070                 if ((ret == max) && !cinfo->dreq)
1071                         break;
1072                 cond_resched();
1073         }
1074         return ret;
1075 }
1076 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1077
1078 /*
1079  * nfs_scan_commit - Scan an inode for commit requests
1080  * @inode: NFS inode to scan
1081  * @dst: mds destination list
1082  * @cinfo: mds and ds lists of reqs ready to commit
1083  *
1084  * Moves requests from the inode's 'commit' request list.
1085  * The requests are *not* checked to ensure that they form a contiguous set.
1086  */
1087 int
1088 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1089                 struct nfs_commit_info *cinfo)
1090 {
1091         int ret = 0;
1092
1093         if (!atomic_long_read(&cinfo->mds->ncommit))
1094                 return 0;
1095         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1096         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1097                 const int max = INT_MAX;
1098
1099                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1100                                            cinfo, max);
1101                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1102         }
1103         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1104         return ret;
1105 }
1106
1107 /*
1108  * Search for an existing write request, and attempt to update
1109  * it to reflect a new dirty region on a given page.
1110  *
1111  * If the attempt fails, then the existing request is flushed out
1112  * to disk.
1113  */
1114 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1115                 struct page *page,
1116                 unsigned int offset,
1117                 unsigned int bytes)
1118 {
1119         struct nfs_page *req;
1120         unsigned int rqend;
1121         unsigned int end;
1122         int error;
1123
1124         end = offset + bytes;
1125
1126         req = nfs_lock_and_join_requests(page);
1127         if (IS_ERR_OR_NULL(req))
1128                 return req;
1129
1130         rqend = req->wb_offset + req->wb_bytes;
1131         /*
1132          * Tell the caller to flush out the request if
1133          * the offsets are non-contiguous.
1134          * Note: nfs_flush_incompatible() will already
1135          * have flushed out requests having wrong owners.
1136          */
1137         if (offset > rqend || end < req->wb_offset)
1138                 goto out_flushme;
1139
1140         /* Okay, the request matches. Update the region */
1141         if (offset < req->wb_offset) {
1142                 req->wb_offset = offset;
1143                 req->wb_pgbase = offset;
1144         }
1145         if (end > rqend)
1146                 req->wb_bytes = end - req->wb_offset;
1147         else
1148                 req->wb_bytes = rqend - req->wb_offset;
1149         req->wb_nio = 0;
1150         return req;
1151 out_flushme:
1152         /*
1153          * Note: we mark the request dirty here because
1154          * nfs_lock_and_join_requests() cannot preserve
1155          * commit flags, so we have to replay the write.
1156          */
1157         nfs_mark_request_dirty(req);
1158         nfs_unlock_and_release_request(req);
1159         error = nfs_wb_page(inode, page);
1160         return (error < 0) ? ERR_PTR(error) : NULL;
1161 }
1162
1163 /*
1164  * Try to update an existing write request, or create one if there is none.
1165  *
1166  * Note: Should always be called with the Page Lock held to prevent races
1167  * if we have to add a new request. Also assumes that the caller has
1168  * already called nfs_flush_incompatible() if necessary.
1169  */
1170 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1171                 struct page *page, unsigned int offset, unsigned int bytes)
1172 {
1173         struct inode *inode = page_file_mapping(page)->host;
1174         struct nfs_page *req;
1175
1176         req = nfs_try_to_update_request(inode, page, offset, bytes);
1177         if (req != NULL)
1178                 goto out;
1179         req = nfs_create_request(ctx, page, offset, bytes);
1180         if (IS_ERR(req))
1181                 goto out;
1182         nfs_inode_add_request(inode, req);
1183 out:
1184         return req;
1185 }
1186
1187 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1188                 unsigned int offset, unsigned int count)
1189 {
1190         struct nfs_page *req;
1191
1192         req = nfs_setup_write_request(ctx, page, offset, count);
1193         if (IS_ERR(req))
1194                 return PTR_ERR(req);
1195         /* Update file length */
1196         nfs_grow_file(page, offset, count);
1197         nfs_mark_uptodate(req);
1198         nfs_mark_request_dirty(req);
1199         nfs_unlock_and_release_request(req);
1200         return 0;
1201 }
1202
1203 int nfs_flush_incompatible(struct file *file, struct page *page)
1204 {
1205         struct nfs_open_context *ctx = nfs_file_open_context(file);
1206         struct nfs_lock_context *l_ctx;
1207         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1208         struct nfs_page *req;
1209         int do_flush, status;
1210         /*
1211          * Look for a request corresponding to this page. If there
1212          * is one, and it belongs to another file, we flush it out
1213          * before we try to copy anything into the page. Do this
1214          * due to the lack of an ACCESS-type call in NFSv2.
1215          * Also do the same if we find a request from an existing
1216          * dropped page.
1217          */
1218         do {
1219                 req = nfs_page_find_head_request(page);
1220                 if (req == NULL)
1221                         return 0;
1222                 l_ctx = req->wb_lock_context;
1223                 do_flush = req->wb_page != page ||
1224                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1225                 if (l_ctx && flctx &&
1226                     !(list_empty_careful(&flctx->flc_posix) &&
1227                       list_empty_careful(&flctx->flc_flock))) {
1228                         do_flush |= l_ctx->lockowner != current->files;
1229                 }
1230                 nfs_release_request(req);
1231                 if (!do_flush)
1232                         return 0;
1233                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1234         } while (status == 0);
1235         return status;
1236 }
1237
1238 /*
1239  * Avoid buffered writes when a open context credential's key would
1240  * expire soon.
1241  *
1242  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1243  *
1244  * Return 0 and set a credential flag which triggers the inode to flush
1245  * and performs  NFS_FILE_SYNC writes if the key will expired within
1246  * RPC_KEY_EXPIRE_TIMEO.
1247  */
1248 int
1249 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1250 {
1251         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1252
1253         if (nfs_ctx_key_to_expire(ctx, inode) &&
1254             !ctx->ll_cred)
1255                 /* Already expired! */
1256                 return -EACCES;
1257         return 0;
1258 }
1259
1260 /*
1261  * Test if the open context credential key is marked to expire soon.
1262  */
1263 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1264 {
1265         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1266         struct rpc_cred *cred = ctx->ll_cred;
1267         struct auth_cred acred = {
1268                 .cred = ctx->cred,
1269         };
1270
1271         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1272                 put_rpccred(cred);
1273                 ctx->ll_cred = NULL;
1274                 cred = NULL;
1275         }
1276         if (!cred)
1277                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1278         if (!cred || IS_ERR(cred))
1279                 return true;
1280         ctx->ll_cred = cred;
1281         return !!(cred->cr_ops->crkey_timeout &&
1282                   cred->cr_ops->crkey_timeout(cred));
1283 }
1284
1285 /*
1286  * If the page cache is marked as unsafe or invalid, then we can't rely on
1287  * the PageUptodate() flag. In this case, we will need to turn off
1288  * write optimisations that depend on the page contents being correct.
1289  */
1290 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1291 {
1292         struct nfs_inode *nfsi = NFS_I(inode);
1293
1294         if (nfs_have_delegated_attributes(inode))
1295                 goto out;
1296         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1297                 return false;
1298         smp_rmb();
1299         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1300                 return false;
1301 out:
1302         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1303                 return false;
1304         return PageUptodate(page) != 0;
1305 }
1306
1307 static bool
1308 is_whole_file_wrlock(struct file_lock *fl)
1309 {
1310         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1311                         fl->fl_type == F_WRLCK;
1312 }
1313
1314 /* If we know the page is up to date, and we're not using byte range locks (or
1315  * if we have the whole file locked for writing), it may be more efficient to
1316  * extend the write to cover the entire page in order to avoid fragmentation
1317  * inefficiencies.
1318  *
1319  * If the file is opened for synchronous writes then we can just skip the rest
1320  * of the checks.
1321  */
1322 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1323 {
1324         int ret;
1325         struct file_lock_context *flctx = inode->i_flctx;
1326         struct file_lock *fl;
1327
1328         if (file->f_flags & O_DSYNC)
1329                 return 0;
1330         if (!nfs_write_pageuptodate(page, inode))
1331                 return 0;
1332         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1333                 return 1;
1334         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1335                        list_empty_careful(&flctx->flc_posix)))
1336                 return 1;
1337
1338         /* Check to see if there are whole file write locks */
1339         ret = 0;
1340         spin_lock(&flctx->flc_lock);
1341         if (!list_empty(&flctx->flc_posix)) {
1342                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1343                                         fl_list);
1344                 if (is_whole_file_wrlock(fl))
1345                         ret = 1;
1346         } else if (!list_empty(&flctx->flc_flock)) {
1347                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1348                                         fl_list);
1349                 if (fl->fl_type == F_WRLCK)
1350                         ret = 1;
1351         }
1352         spin_unlock(&flctx->flc_lock);
1353         return ret;
1354 }
1355
1356 /*
1357  * Update and possibly write a cached page of an NFS file.
1358  *
1359  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1360  * things with a page scheduled for an RPC call (e.g. invalidate it).
1361  */
1362 int nfs_updatepage(struct file *file, struct page *page,
1363                 unsigned int offset, unsigned int count)
1364 {
1365         struct nfs_open_context *ctx = nfs_file_open_context(file);
1366         struct address_space *mapping = page_file_mapping(page);
1367         struct inode    *inode = mapping->host;
1368         int             status = 0;
1369
1370         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1371
1372         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1373                 file, count, (long long)(page_file_offset(page) + offset));
1374
1375         if (!count)
1376                 goto out;
1377
1378         if (nfs_can_extend_write(file, page, inode)) {
1379                 count = max(count + offset, nfs_page_length(page));
1380                 offset = 0;
1381         }
1382
1383         status = nfs_writepage_setup(ctx, page, offset, count);
1384         if (status < 0)
1385                 nfs_set_pageerror(mapping);
1386         else
1387                 __set_page_dirty_nobuffers(page);
1388 out:
1389         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1390                         status, (long long)i_size_read(inode));
1391         return status;
1392 }
1393
1394 static int flush_task_priority(int how)
1395 {
1396         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1397                 case FLUSH_HIGHPRI:
1398                         return RPC_PRIORITY_HIGH;
1399                 case FLUSH_LOWPRI:
1400                         return RPC_PRIORITY_LOW;
1401         }
1402         return RPC_PRIORITY_NORMAL;
1403 }
1404
1405 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1406                                struct rpc_message *msg,
1407                                const struct nfs_rpc_ops *rpc_ops,
1408                                struct rpc_task_setup *task_setup_data, int how)
1409 {
1410         int priority = flush_task_priority(how);
1411
1412         task_setup_data->priority = priority;
1413         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1414         trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1415                                  hdr->args.stable);
1416 }
1417
1418 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1419  * call this on each, which will prepare them to be retried on next
1420  * writeback using standard nfs.
1421  */
1422 static void nfs_redirty_request(struct nfs_page *req)
1423 {
1424         /* Bump the transmission count */
1425         req->wb_nio++;
1426         nfs_mark_request_dirty(req);
1427         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1428         nfs_end_page_writeback(req);
1429         nfs_release_request(req);
1430 }
1431
1432 static void nfs_async_write_error(struct list_head *head, int error)
1433 {
1434         struct nfs_page *req;
1435
1436         while (!list_empty(head)) {
1437                 req = nfs_list_entry(head->next);
1438                 nfs_list_remove_request(req);
1439                 if (nfs_error_is_fatal(error))
1440                         nfs_write_error(req, error);
1441                 else
1442                         nfs_redirty_request(req);
1443         }
1444 }
1445
1446 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1447 {
1448         nfs_async_write_error(&hdr->pages, 0);
1449         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1450                         hdr->args.offset + hdr->args.count - 1);
1451 }
1452
1453 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1454         .init_hdr = nfs_async_write_init,
1455         .error_cleanup = nfs_async_write_error,
1456         .completion = nfs_write_completion,
1457         .reschedule_io = nfs_async_write_reschedule_io,
1458 };
1459
1460 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1461                                struct inode *inode, int ioflags, bool force_mds,
1462                                const struct nfs_pgio_completion_ops *compl_ops)
1463 {
1464         struct nfs_server *server = NFS_SERVER(inode);
1465         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1466
1467 #ifdef CONFIG_NFS_V4_1
1468         if (server->pnfs_curr_ld && !force_mds)
1469                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1470 #endif
1471         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1472                         server->wsize, ioflags);
1473 }
1474 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1475
1476 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1477 {
1478         struct nfs_pgio_mirror *mirror;
1479
1480         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1481                 pgio->pg_ops->pg_cleanup(pgio);
1482
1483         pgio->pg_ops = &nfs_pgio_rw_ops;
1484
1485         nfs_pageio_stop_mirroring(pgio);
1486
1487         mirror = &pgio->pg_mirrors[0];
1488         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1489 }
1490 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1491
1492
1493 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1494 {
1495         struct nfs_commit_data *data = calldata;
1496
1497         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1498 }
1499
1500 /*
1501  * Special version of should_remove_suid() that ignores capabilities.
1502  */
1503 static int nfs_should_remove_suid(const struct inode *inode)
1504 {
1505         umode_t mode = inode->i_mode;
1506         int kill = 0;
1507
1508         /* suid always must be killed */
1509         if (unlikely(mode & S_ISUID))
1510                 kill = ATTR_KILL_SUID;
1511
1512         /*
1513          * sgid without any exec bits is just a mandatory locking mark; leave
1514          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1515          */
1516         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1517                 kill |= ATTR_KILL_SGID;
1518
1519         if (unlikely(kill && S_ISREG(mode)))
1520                 return kill;
1521
1522         return 0;
1523 }
1524
1525 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1526                 struct nfs_fattr *fattr)
1527 {
1528         struct nfs_pgio_args *argp = &hdr->args;
1529         struct nfs_pgio_res *resp = &hdr->res;
1530         u64 size = argp->offset + resp->count;
1531
1532         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1533                 fattr->size = size;
1534         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1535                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1536                 return;
1537         }
1538         if (size != fattr->size)
1539                 return;
1540         /* Set attribute barrier */
1541         nfs_fattr_set_barrier(fattr);
1542         /* ...and update size */
1543         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1544 }
1545
1546 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1547 {
1548         struct nfs_fattr *fattr = &hdr->fattr;
1549         struct inode *inode = hdr->inode;
1550
1551         spin_lock(&inode->i_lock);
1552         nfs_writeback_check_extend(hdr, fattr);
1553         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1554         spin_unlock(&inode->i_lock);
1555 }
1556 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1557
1558 /*
1559  * This function is called when the WRITE call is complete.
1560  */
1561 static int nfs_writeback_done(struct rpc_task *task,
1562                               struct nfs_pgio_header *hdr,
1563                               struct inode *inode)
1564 {
1565         int status;
1566
1567         /*
1568          * ->write_done will attempt to use post-op attributes to detect
1569          * conflicting writes by other clients.  A strict interpretation
1570          * of close-to-open would allow us to continue caching even if
1571          * another writer had changed the file, but some applications
1572          * depend on tighter cache coherency when writing.
1573          */
1574         status = NFS_PROTO(inode)->write_done(task, hdr);
1575         if (status != 0)
1576                 return status;
1577
1578         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1579         trace_nfs_writeback_done(inode, task->tk_status,
1580                                  hdr->args.offset, hdr->res.verf);
1581
1582         if (hdr->res.verf->committed < hdr->args.stable &&
1583             task->tk_status >= 0) {
1584                 /* We tried a write call, but the server did not
1585                  * commit data to stable storage even though we
1586                  * requested it.
1587                  * Note: There is a known bug in Tru64 < 5.0 in which
1588                  *       the server reports NFS_DATA_SYNC, but performs
1589                  *       NFS_FILE_SYNC. We therefore implement this checking
1590                  *       as a dprintk() in order to avoid filling syslog.
1591                  */
1592                 static unsigned long    complain;
1593
1594                 /* Note this will print the MDS for a DS write */
1595                 if (time_before(complain, jiffies)) {
1596                         dprintk("NFS:       faulty NFS server %s:"
1597                                 " (committed = %d) != (stable = %d)\n",
1598                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1599                                 hdr->res.verf->committed, hdr->args.stable);
1600                         complain = jiffies + 300 * HZ;
1601                 }
1602         }
1603
1604         /* Deal with the suid/sgid bit corner case */
1605         if (nfs_should_remove_suid(inode)) {
1606                 spin_lock(&inode->i_lock);
1607                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1608                 spin_unlock(&inode->i_lock);
1609         }
1610         return 0;
1611 }
1612
1613 /*
1614  * This function is called when the WRITE call is complete.
1615  */
1616 static void nfs_writeback_result(struct rpc_task *task,
1617                                  struct nfs_pgio_header *hdr)
1618 {
1619         struct nfs_pgio_args    *argp = &hdr->args;
1620         struct nfs_pgio_res     *resp = &hdr->res;
1621
1622         if (resp->count < argp->count) {
1623                 static unsigned long    complain;
1624
1625                 /* This a short write! */
1626                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1627
1628                 /* Has the server at least made some progress? */
1629                 if (resp->count == 0) {
1630                         if (time_before(complain, jiffies)) {
1631                                 printk(KERN_WARNING
1632                                        "NFS: Server wrote zero bytes, expected %u.\n",
1633                                        argp->count);
1634                                 complain = jiffies + 300 * HZ;
1635                         }
1636                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1637                         task->tk_status = -EIO;
1638                         return;
1639                 }
1640
1641                 /* For non rpc-based layout drivers, retry-through-MDS */
1642                 if (!task->tk_ops) {
1643                         hdr->pnfs_error = -EAGAIN;
1644                         return;
1645                 }
1646
1647                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1648                 if (resp->verf->committed != NFS_UNSTABLE) {
1649                         /* Resend from where the server left off */
1650                         hdr->mds_offset += resp->count;
1651                         argp->offset += resp->count;
1652                         argp->pgbase += resp->count;
1653                         argp->count -= resp->count;
1654                 } else {
1655                         /* Resend as a stable write in order to avoid
1656                          * headaches in the case of a server crash.
1657                          */
1658                         argp->stable = NFS_FILE_SYNC;
1659                 }
1660                 rpc_restart_call_prepare(task);
1661         }
1662 }
1663
1664 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1665 {
1666         return wait_var_event_killable(&cinfo->rpcs_out,
1667                                        !atomic_read(&cinfo->rpcs_out));
1668 }
1669
1670 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1671 {
1672         atomic_inc(&cinfo->rpcs_out);
1673 }
1674
1675 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1676 {
1677         if (atomic_dec_and_test(&cinfo->rpcs_out))
1678                 wake_up_var(&cinfo->rpcs_out);
1679 }
1680
1681 void nfs_commitdata_release(struct nfs_commit_data *data)
1682 {
1683         put_nfs_open_context(data->context);
1684         nfs_commit_free(data);
1685 }
1686 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1687
1688 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1689                         const struct nfs_rpc_ops *nfs_ops,
1690                         const struct rpc_call_ops *call_ops,
1691                         int how, int flags)
1692 {
1693         struct rpc_task *task;
1694         int priority = flush_task_priority(how);
1695         struct rpc_message msg = {
1696                 .rpc_argp = &data->args,
1697                 .rpc_resp = &data->res,
1698                 .rpc_cred = data->cred,
1699         };
1700         struct rpc_task_setup task_setup_data = {
1701                 .task = &data->task,
1702                 .rpc_client = clnt,
1703                 .rpc_message = &msg,
1704                 .callback_ops = call_ops,
1705                 .callback_data = data,
1706                 .workqueue = nfsiod_workqueue,
1707                 .flags = RPC_TASK_ASYNC | flags,
1708                 .priority = priority,
1709         };
1710         /* Set up the initial task struct.  */
1711         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1712         trace_nfs_initiate_commit(data);
1713
1714         dprintk("NFS: initiated commit call\n");
1715
1716         task = rpc_run_task(&task_setup_data);
1717         if (IS_ERR(task))
1718                 return PTR_ERR(task);
1719         if (how & FLUSH_SYNC)
1720                 rpc_wait_for_completion_task(task);
1721         rpc_put_task(task);
1722         return 0;
1723 }
1724 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1725
1726 static loff_t nfs_get_lwb(struct list_head *head)
1727 {
1728         loff_t lwb = 0;
1729         struct nfs_page *req;
1730
1731         list_for_each_entry(req, head, wb_list)
1732                 if (lwb < (req_offset(req) + req->wb_bytes))
1733                         lwb = req_offset(req) + req->wb_bytes;
1734
1735         return lwb;
1736 }
1737
1738 /*
1739  * Set up the argument/result storage required for the RPC call.
1740  */
1741 void nfs_init_commit(struct nfs_commit_data *data,
1742                      struct list_head *head,
1743                      struct pnfs_layout_segment *lseg,
1744                      struct nfs_commit_info *cinfo)
1745 {
1746         struct nfs_page *first = nfs_list_entry(head->next);
1747         struct nfs_open_context *ctx = nfs_req_openctx(first);
1748         struct inode *inode = d_inode(ctx->dentry);
1749
1750         /* Set up the RPC argument and reply structs
1751          * NB: take care not to mess about with data->commit et al. */
1752
1753         list_splice_init(head, &data->pages);
1754
1755         data->inode       = inode;
1756         data->cred        = ctx->cred;
1757         data->lseg        = lseg; /* reference transferred */
1758         /* only set lwb for pnfs commit */
1759         if (lseg)
1760                 data->lwb = nfs_get_lwb(&data->pages);
1761         data->mds_ops     = &nfs_commit_ops;
1762         data->completion_ops = cinfo->completion_ops;
1763         data->dreq        = cinfo->dreq;
1764
1765         data->args.fh     = NFS_FH(data->inode);
1766         /* Note: we always request a commit of the entire inode */
1767         data->args.offset = 0;
1768         data->args.count  = 0;
1769         data->context     = get_nfs_open_context(ctx);
1770         data->res.fattr   = &data->fattr;
1771         data->res.verf    = &data->verf;
1772         nfs_fattr_init(&data->fattr);
1773 }
1774 EXPORT_SYMBOL_GPL(nfs_init_commit);
1775
1776 void nfs_retry_commit(struct list_head *page_list,
1777                       struct pnfs_layout_segment *lseg,
1778                       struct nfs_commit_info *cinfo,
1779                       u32 ds_commit_idx)
1780 {
1781         struct nfs_page *req;
1782
1783         while (!list_empty(page_list)) {
1784                 req = nfs_list_entry(page_list->next);
1785                 nfs_list_remove_request(req);
1786                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1787                 if (!cinfo->dreq)
1788                         nfs_clear_page_commit(req->wb_page);
1789                 nfs_unlock_and_release_request(req);
1790         }
1791 }
1792 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1793
1794 static void
1795 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1796                 struct nfs_page *req)
1797 {
1798         __set_page_dirty_nobuffers(req->wb_page);
1799 }
1800
1801 /*
1802  * Commit dirty pages
1803  */
1804 static int
1805 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1806                 struct nfs_commit_info *cinfo)
1807 {
1808         struct nfs_commit_data  *data;
1809
1810         /* another commit raced with us */
1811         if (list_empty(head))
1812                 return 0;
1813
1814         data = nfs_commitdata_alloc(true);
1815
1816         /* Set up the argument struct */
1817         nfs_init_commit(data, head, NULL, cinfo);
1818         atomic_inc(&cinfo->mds->rpcs_out);
1819         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1820                                    data->mds_ops, how, 0);
1821 }
1822
1823 /*
1824  * COMMIT call returned
1825  */
1826 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1827 {
1828         struct nfs_commit_data  *data = calldata;
1829
1830         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1831                                 task->tk_pid, task->tk_status);
1832
1833         /* Call the NFS version-specific code */
1834         NFS_PROTO(data->inode)->commit_done(task, data);
1835         trace_nfs_commit_done(data);
1836 }
1837
1838 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1839 {
1840         struct nfs_page *req;
1841         int status = data->task.tk_status;
1842         struct nfs_commit_info cinfo;
1843         struct nfs_server *nfss;
1844
1845         while (!list_empty(&data->pages)) {
1846                 req = nfs_list_entry(data->pages.next);
1847                 nfs_list_remove_request(req);
1848                 if (req->wb_page)
1849                         nfs_clear_page_commit(req->wb_page);
1850
1851                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1852                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1853                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1854                         req->wb_bytes,
1855                         (long long)req_offset(req));
1856                 if (status < 0) {
1857                         if (req->wb_page) {
1858                                 nfs_mapping_set_error(req->wb_page, status);
1859                                 nfs_inode_remove_request(req);
1860                         }
1861                         dprintk_cont(", error = %d\n", status);
1862                         goto next;
1863                 }
1864
1865                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1866                  * returned by the server against all stored verfs. */
1867                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1868                         /* We have a match */
1869                         if (req->wb_page)
1870                                 nfs_inode_remove_request(req);
1871                         dprintk_cont(" OK\n");
1872                         goto next;
1873                 }
1874                 /* We have a mismatch. Write the page again */
1875                 dprintk_cont(" mismatch\n");
1876                 nfs_mark_request_dirty(req);
1877                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1878         next:
1879                 nfs_unlock_and_release_request(req);
1880                 /* Latency breaker */
1881                 cond_resched();
1882         }
1883         nfss = NFS_SERVER(data->inode);
1884         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1885                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1886
1887         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1888         nfs_commit_end(cinfo.mds);
1889 }
1890
1891 static void nfs_commit_release(void *calldata)
1892 {
1893         struct nfs_commit_data *data = calldata;
1894
1895         data->completion_ops->completion(data);
1896         nfs_commitdata_release(calldata);
1897 }
1898
1899 static const struct rpc_call_ops nfs_commit_ops = {
1900         .rpc_call_prepare = nfs_commit_prepare,
1901         .rpc_call_done = nfs_commit_done,
1902         .rpc_release = nfs_commit_release,
1903 };
1904
1905 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1906         .completion = nfs_commit_release_pages,
1907         .resched_write = nfs_commit_resched_write,
1908 };
1909
1910 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1911                             int how, struct nfs_commit_info *cinfo)
1912 {
1913         int status;
1914
1915         status = pnfs_commit_list(inode, head, how, cinfo);
1916         if (status == PNFS_NOT_ATTEMPTED)
1917                 status = nfs_commit_list(inode, head, how, cinfo);
1918         return status;
1919 }
1920
1921 static int __nfs_commit_inode(struct inode *inode, int how,
1922                 struct writeback_control *wbc)
1923 {
1924         LIST_HEAD(head);
1925         struct nfs_commit_info cinfo;
1926         int may_wait = how & FLUSH_SYNC;
1927         int ret, nscan;
1928
1929         nfs_init_cinfo_from_inode(&cinfo, inode);
1930         nfs_commit_begin(cinfo.mds);
1931         for (;;) {
1932                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1933                 if (ret <= 0)
1934                         break;
1935                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1936                 if (ret < 0)
1937                         break;
1938                 ret = 0;
1939                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1940                         if (nscan < wbc->nr_to_write)
1941                                 wbc->nr_to_write -= nscan;
1942                         else
1943                                 wbc->nr_to_write = 0;
1944                 }
1945                 if (nscan < INT_MAX)
1946                         break;
1947                 cond_resched();
1948         }
1949         nfs_commit_end(cinfo.mds);
1950         if (ret || !may_wait)
1951                 return ret;
1952         return wait_on_commit(cinfo.mds);
1953 }
1954
1955 int nfs_commit_inode(struct inode *inode, int how)
1956 {
1957         return __nfs_commit_inode(inode, how, NULL);
1958 }
1959 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1960
1961 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1962 {
1963         struct nfs_inode *nfsi = NFS_I(inode);
1964         int flags = FLUSH_SYNC;
1965         int ret = 0;
1966
1967         if (wbc->sync_mode == WB_SYNC_NONE) {
1968                 /* no commits means nothing needs to be done */
1969                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1970                         goto check_requests_outstanding;
1971
1972                 /* Don't commit yet if this is a non-blocking flush and there
1973                  * are a lot of outstanding writes for this mapping.
1974                  */
1975                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1976                         goto out_mark_dirty;
1977
1978                 /* don't wait for the COMMIT response */
1979                 flags = 0;
1980         }
1981
1982         ret = __nfs_commit_inode(inode, flags, wbc);
1983         if (!ret) {
1984                 if (flags & FLUSH_SYNC)
1985                         return 0;
1986         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1987                 goto out_mark_dirty;
1988
1989 check_requests_outstanding:
1990         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1991                 return ret;
1992 out_mark_dirty:
1993         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1994         return ret;
1995 }
1996 EXPORT_SYMBOL_GPL(nfs_write_inode);
1997
1998 /*
1999  * Wrapper for filemap_write_and_wait_range()
2000  *
2001  * Needed for pNFS in order to ensure data becomes visible to the
2002  * client.
2003  */
2004 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2005                 loff_t lstart, loff_t lend)
2006 {
2007         int ret;
2008
2009         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2010         if (ret == 0)
2011                 ret = pnfs_sync_inode(mapping->host, true);
2012         return ret;
2013 }
2014 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2015
2016 /*
2017  * flush the inode to disk.
2018  */
2019 int nfs_wb_all(struct inode *inode)
2020 {
2021         int ret;
2022
2023         trace_nfs_writeback_inode_enter(inode);
2024
2025         ret = filemap_write_and_wait(inode->i_mapping);
2026         if (ret)
2027                 goto out;
2028         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2029         if (ret < 0)
2030                 goto out;
2031         pnfs_sync_inode(inode, true);
2032         ret = 0;
2033
2034 out:
2035         trace_nfs_writeback_inode_exit(inode, ret);
2036         return ret;
2037 }
2038 EXPORT_SYMBOL_GPL(nfs_wb_all);
2039
2040 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2041 {
2042         struct nfs_page *req;
2043         int ret = 0;
2044
2045         wait_on_page_writeback(page);
2046
2047         /* blocking call to cancel all requests and join to a single (head)
2048          * request */
2049         req = nfs_lock_and_join_requests(page);
2050
2051         if (IS_ERR(req)) {
2052                 ret = PTR_ERR(req);
2053         } else if (req) {
2054                 /* all requests from this page have been cancelled by
2055                  * nfs_lock_and_join_requests, so just remove the head
2056                  * request from the inode / page_private pointer and
2057                  * release it */
2058                 nfs_inode_remove_request(req);
2059                 nfs_unlock_and_release_request(req);
2060         }
2061
2062         return ret;
2063 }
2064
2065 /*
2066  * Write back all requests on one page - we do this before reading it.
2067  */
2068 int nfs_wb_page(struct inode *inode, struct page *page)
2069 {
2070         loff_t range_start = page_file_offset(page);
2071         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2072         struct writeback_control wbc = {
2073                 .sync_mode = WB_SYNC_ALL,
2074                 .nr_to_write = 0,
2075                 .range_start = range_start,
2076                 .range_end = range_end,
2077         };
2078         int ret;
2079
2080         trace_nfs_writeback_page_enter(inode);
2081
2082         for (;;) {
2083                 wait_on_page_writeback(page);
2084                 if (clear_page_dirty_for_io(page)) {
2085                         ret = nfs_writepage_locked(page, &wbc);
2086                         if (ret < 0)
2087                                 goto out_error;
2088                         continue;
2089                 }
2090                 ret = 0;
2091                 if (!PagePrivate(page))
2092                         break;
2093                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2094                 if (ret < 0)
2095                         goto out_error;
2096         }
2097 out_error:
2098         trace_nfs_writeback_page_exit(inode, ret);
2099         return ret;
2100 }
2101
2102 #ifdef CONFIG_MIGRATION
2103 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2104                 struct page *page, enum migrate_mode mode)
2105 {
2106         /*
2107          * If PagePrivate is set, then the page is currently associated with
2108          * an in-progress read or write request. Don't try to migrate it.
2109          *
2110          * FIXME: we could do this in principle, but we'll need a way to ensure
2111          *        that we can safely release the inode reference while holding
2112          *        the page lock.
2113          */
2114         if (PagePrivate(page))
2115                 return -EBUSY;
2116
2117         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2118                 return -EBUSY;
2119
2120         return migrate_page(mapping, newpage, page, mode);
2121 }
2122 #endif
2123
2124 int __init nfs_init_writepagecache(void)
2125 {
2126         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2127                                              sizeof(struct nfs_pgio_header),
2128                                              0, SLAB_HWCACHE_ALIGN,
2129                                              NULL);
2130         if (nfs_wdata_cachep == NULL)
2131                 return -ENOMEM;
2132
2133         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2134                                                      nfs_wdata_cachep);
2135         if (nfs_wdata_mempool == NULL)
2136                 goto out_destroy_write_cache;
2137
2138         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2139                                              sizeof(struct nfs_commit_data),
2140                                              0, SLAB_HWCACHE_ALIGN,
2141                                              NULL);
2142         if (nfs_cdata_cachep == NULL)
2143                 goto out_destroy_write_mempool;
2144
2145         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2146                                                       nfs_cdata_cachep);
2147         if (nfs_commit_mempool == NULL)
2148                 goto out_destroy_commit_cache;
2149
2150         /*
2151          * NFS congestion size, scale with available memory.
2152          *
2153          *  64MB:    8192k
2154          * 128MB:   11585k
2155          * 256MB:   16384k
2156          * 512MB:   23170k
2157          *   1GB:   32768k
2158          *   2GB:   46340k
2159          *   4GB:   65536k
2160          *   8GB:   92681k
2161          *  16GB:  131072k
2162          *
2163          * This allows larger machines to have larger/more transfers.
2164          * Limit the default to 256M
2165          */
2166         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2167         if (nfs_congestion_kb > 256*1024)
2168                 nfs_congestion_kb = 256*1024;
2169
2170         return 0;
2171
2172 out_destroy_commit_cache:
2173         kmem_cache_destroy(nfs_cdata_cachep);
2174 out_destroy_write_mempool:
2175         mempool_destroy(nfs_wdata_mempool);
2176 out_destroy_write_cache:
2177         kmem_cache_destroy(nfs_wdata_cachep);
2178         return -ENOMEM;
2179 }
2180
2181 void nfs_destroy_writepagecache(void)
2182 {
2183         mempool_destroy(nfs_commit_mempool);
2184         kmem_cache_destroy(nfs_cdata_cachep);
2185         mempool_destroy(nfs_wdata_mempool);
2186         kmem_cache_destroy(nfs_wdata_cachep);
2187 }
2188
2189 static const struct nfs_rw_ops nfs_rw_write_ops = {
2190         .rw_alloc_header        = nfs_writehdr_alloc,
2191         .rw_free_header         = nfs_writehdr_free,
2192         .rw_done                = nfs_writeback_done,
2193         .rw_result              = nfs_writeback_result,
2194         .rw_initiate            = nfs_initiate_write,
2195 };