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_KERNEL);
 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         int err;
 725
 726         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 727
 728         ioc = nfs_io_completion_alloc(GFP_KERNEL);
 729         if (ioc)
 730                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 731
 732         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 733                                 &nfs_async_write_completion_ops);
 734         pgio.pg_io_completion = ioc;
 735         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 736         nfs_pageio_complete(&pgio);
 737         nfs_io_completion_put(ioc);
 738
 739         if (err < 0)
 740                 goto out_err;
 741         err = pgio.pg_error;
 742         if (err < 0)
 743                 goto out_err;
 744         return 0;
 745 out_err:
 746         return err;
 747 }
 748
 749 /*
 750  * Insert a write request into an inode
 751  */
 752 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 753 {
 754         struct address_space *mapping = page_file_mapping(req->wb_page);
 755         struct nfs_inode *nfsi = NFS_I(inode);
 756
 757         WARN_ON_ONCE(req->wb_this_page != req);
 758
 759         /* Lock the request! */
 760         nfs_lock_request(req);
 761
 762         /*
 763          * Swap-space should not get truncated. Hence no need to plug the race
 764          * with invalidate/truncate.
 765          */
 766         spin_lock(&mapping->private_lock);
 767         if (!nfs_have_writebacks(inode) &&
 768             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 769                 inode_inc_iversion_raw(inode);
 770         if (likely(!PageSwapCache(req->wb_page))) {
 771                 set_bit(PG_MAPPED, &req->wb_flags);
 772                 SetPagePrivate(req->wb_page);
 773                 set_page_private(req->wb_page, (unsigned long)req);
 774         }
 775         spin_unlock(&mapping->private_lock);
 776         atomic_long_inc(&nfsi->nrequests);
 777         /* this a head request for a page group - mark it as having an
 778          * extra reference so sub groups can follow suit.
 779          * This flag also informs pgio layer when to bump nrequests when
 780          * adding subrequests. */
 781         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 782         kref_get(&req->wb_kref);
 783 }
 784
 785 /*
 786  * Remove a write request from an inode
 787  */
 788 static void nfs_inode_remove_request(struct nfs_page *req)
 789 {
 790         struct address_space *mapping = page_file_mapping(req->wb_page);
 791         struct inode *inode = mapping->host;
 792         struct nfs_inode *nfsi = NFS_I(inode);
 793         struct nfs_page *head;
 794
 795         atomic_long_dec(&nfsi->nrequests);
 796         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 797                 head = req->wb_head;
 798
 799                 spin_lock(&mapping->private_lock);
 800                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 801                         set_page_private(head->wb_page, 0);
 802                         ClearPagePrivate(head->wb_page);
 803                         clear_bit(PG_MAPPED, &head->wb_flags);
 804                 }
 805                 spin_unlock(&mapping->private_lock);
 806         }
 807
 808         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 809                 nfs_release_request(req);
 810 }
 811
 812 static void
 813 nfs_mark_request_dirty(struct nfs_page *req)
 814 {
 815         if (req->wb_page)
 816                 __set_page_dirty_nobuffers(req->wb_page);
 817 }
 818
 819 /*
 820  * nfs_page_search_commits_for_head_request_locked
 821  *
 822  * Search through commit lists on @inode for the head request for @page.
 823  * Must be called while holding the inode (which is cinfo) lock.
 824  *
 825  * Returns the head request if found, or NULL if not found.
 826  */
 827 static struct nfs_page *
 828 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 829                                                 struct page *page)
 830 {
 831         struct nfs_page *freq, *t;
 832         struct nfs_commit_info cinfo;
 833         struct inode *inode = &nfsi->vfs_inode;
 834
 835         nfs_init_cinfo_from_inode(&cinfo, inode);
 836
 837         /* search through pnfs commit lists */
 838         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 839         if (freq)
 840                 return freq->wb_head;
 841
 842         /* Linearly search the commit list for the correct request */
 843         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 844                 if (freq->wb_page == page)
 845                         return freq->wb_head;
 846         }
 847
 848         return NULL;
 849 }
 850
 851 /**
 852  * nfs_request_add_commit_list_locked - add request to a commit list
 853  * @req: pointer to a struct nfs_page
 854  * @dst: commit list head
 855  * @cinfo: holds list lock and accounting info
 856  *
 857  * This sets the PG_CLEAN bit, updates the cinfo count of
 858  * number of outstanding requests requiring a commit as well as
 859  * the MM page stats.
 860  *
 861  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 862  * nfs_page lock.
 863  */
 864 void
 865 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 866                             struct nfs_commit_info *cinfo)
 867 {
 868         set_bit(PG_CLEAN, &req->wb_flags);
 869         nfs_list_add_request(req, dst);
 870         atomic_long_inc(&cinfo->mds->ncommit);
 871 }
 872 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 873
 874 /**
 875  * nfs_request_add_commit_list - add request to a commit list
 876  * @req: pointer to a struct nfs_page
 877  * @cinfo: holds list lock and accounting info
 878  *
 879  * This sets the PG_CLEAN bit, updates the cinfo count of
 880  * number of outstanding requests requiring a commit as well as
 881  * the MM page stats.
 882  *
 883  * The caller must _not_ hold the cinfo->lock, but must be
 884  * holding the nfs_page lock.
 885  */
 886 void
 887 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 888 {
 889         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 890         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 891         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 892         if (req->wb_page)
 893                 nfs_mark_page_unstable(req->wb_page, cinfo);
 894 }
 895 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 896
 897 /**
 898  * nfs_request_remove_commit_list - Remove request from a commit list
 899  * @req: pointer to a nfs_page
 900  * @cinfo: holds list lock and accounting info
 901  *
 902  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 903  * number of outstanding requests requiring a commit
 904  * It does not update the MM page stats.
 905  *
 906  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 907  */
 908 void
 909 nfs_request_remove_commit_list(struct nfs_page *req,
 910                                struct nfs_commit_info *cinfo)
 911 {
 912         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 913                 return;
 914         nfs_list_remove_request(req);
 915         atomic_long_dec(&cinfo->mds->ncommit);
 916 }
 917 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 918
 919 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 920                                       struct inode *inode)
 921 {
 922         cinfo->inode = inode;
 923         cinfo->mds = &NFS_I(inode)->commit_info;
 924         cinfo->ds = pnfs_get_ds_info(inode);
 925         cinfo->dreq = NULL;
 926         cinfo->completion_ops = &nfs_commit_completion_ops;
 927 }
 928
 929 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 930                     struct inode *inode,
 931                     struct nfs_direct_req *dreq)
 932 {
 933         if (dreq)
 934                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 935         else
 936                 nfs_init_cinfo_from_inode(cinfo, inode);
 937 }
 938 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 939
 940 /*
 941  * Add a request to the inode's commit list.
 942  */
 943 void
 944 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 945                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 946 {
 947         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 948                 return;
 949         nfs_request_add_commit_list(req, cinfo);
 950 }
 951
 952 static void
 953 nfs_clear_page_commit(struct page *page)
 954 {
 955         dec_node_page_state(page, NR_UNSTABLE_NFS);
 956         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 957                     WB_RECLAIMABLE);
 958 }
 959
 960 /* Called holding the request lock on @req */
 961 static void
 962 nfs_clear_request_commit(struct nfs_page *req)
 963 {
 964         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 965                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 966                 struct inode *inode = d_inode(ctx->dentry);
 967                 struct nfs_commit_info cinfo;
 968
 969                 nfs_init_cinfo_from_inode(&cinfo, inode);
 970                 mutex_lock(&NFS_I(inode)->commit_mutex);
 971                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 972                         nfs_request_remove_commit_list(req, &cinfo);
 973                 }
 974                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 975                 nfs_clear_page_commit(req->wb_page);
 976         }
 977 }
 978
 979 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 980 {
 981         if (hdr->verf.committed == NFS_DATA_SYNC)
 982                 return hdr->lseg == NULL;
 983         return hdr->verf.committed != NFS_FILE_SYNC;
 984 }
 985
 986 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 987 {
 988         nfs_io_completion_get(hdr->io_completion);
 989 }
 990
 991 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 992 {
 993         struct nfs_commit_info cinfo;
 994         unsigned long bytes = 0;
 995
 996         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 997                 goto out;
 998         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 999         while (!list_empty(&hdr->pages)) {
1000                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1001
1002                 bytes += req->wb_bytes;
1003                 nfs_list_remove_request(req);
1004                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1005                     (hdr->good_bytes < bytes)) {
1006                         nfs_set_pageerror(page_file_mapping(req->wb_page));
1007                         nfs_mapping_set_error(req->wb_page, hdr->error);
1008                         goto remove_req;
1009                 }
1010                 if (nfs_write_need_commit(hdr)) {
1011                         /* Reset wb_nio, since the write was successful. */
1012                         req->wb_nio = 0;
1013                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1014                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1015                                 hdr->pgio_mirror_idx);
1016                         goto next;
1017                 }
1018 remove_req:
1019                 nfs_inode_remove_request(req);
1020 next:
1021                 nfs_end_page_writeback(req);
1022                 nfs_release_request(req);
1023         }
1024 out:
1025         nfs_io_completion_put(hdr->io_completion);
1026         hdr->release(hdr);
1027 }
1028
1029 unsigned long
1030 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1031 {
1032         return atomic_long_read(&cinfo->mds->ncommit);
1033 }
1034
1035 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1036 int
1037 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1038                      struct nfs_commit_info *cinfo, int max)
1039 {
1040         struct nfs_page *req, *tmp;
1041         int ret = 0;
1042
1043 restart:
1044         list_for_each_entry_safe(req, tmp, src, wb_list) {
1045                 kref_get(&req->wb_kref);
1046                 if (!nfs_lock_request(req)) {
1047                         int status;
1048
1049                         /* Prevent deadlock with nfs_lock_and_join_requests */
1050                         if (!list_empty(dst)) {
1051                                 nfs_release_request(req);
1052                                 continue;
1053                         }
1054                         /* Ensure we make progress to prevent livelock */
1055                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1056                         status = nfs_wait_on_request(req);
1057                         nfs_release_request(req);
1058                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1059                         if (status < 0)
1060                                 break;
1061                         goto restart;
1062                 }
1063                 nfs_request_remove_commit_list(req, cinfo);
1064                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1065                 nfs_list_add_request(req, dst);
1066                 ret++;
1067                 if ((ret == max) && !cinfo->dreq)
1068                         break;
1069                 cond_resched();
1070         }
1071         return ret;
1072 }
1073 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1074
1075 /*
1076  * nfs_scan_commit - Scan an inode for commit requests
1077  * @inode: NFS inode to scan
1078  * @dst: mds destination list
1079  * @cinfo: mds and ds lists of reqs ready to commit
1080  *
1081  * Moves requests from the inode's 'commit' request list.
1082  * The requests are *not* checked to ensure that they form a contiguous set.
1083  */
1084 int
1085 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1086                 struct nfs_commit_info *cinfo)
1087 {
1088         int ret = 0;
1089
1090         if (!atomic_long_read(&cinfo->mds->ncommit))
1091                 return 0;
1092         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1093         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1094                 const int max = INT_MAX;
1095
1096                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1097                                            cinfo, max);
1098                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1099         }
1100         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1101         return ret;
1102 }
1103
1104 /*
1105  * Search for an existing write request, and attempt to update
1106  * it to reflect a new dirty region on a given page.
1107  *
1108  * If the attempt fails, then the existing request is flushed out
1109  * to disk.
1110  */
1111 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1112                 struct page *page,
1113                 unsigned int offset,
1114                 unsigned int bytes)
1115 {
1116         struct nfs_page *req;
1117         unsigned int rqend;
1118         unsigned int end;
1119         int error;
1120
1121         end = offset + bytes;
1122
1123         req = nfs_lock_and_join_requests(page);
1124         if (IS_ERR_OR_NULL(req))
1125                 return req;
1126
1127         rqend = req->wb_offset + req->wb_bytes;
1128         /*
1129          * Tell the caller to flush out the request if
1130          * the offsets are non-contiguous.
1131          * Note: nfs_flush_incompatible() will already
1132          * have flushed out requests having wrong owners.
1133          */
1134         if (offset > rqend || end < req->wb_offset)
1135                 goto out_flushme;
1136
1137         /* Okay, the request matches. Update the region */
1138         if (offset < req->wb_offset) {
1139                 req->wb_offset = offset;
1140                 req->wb_pgbase = offset;
1141         }
1142         if (end > rqend)
1143                 req->wb_bytes = end - req->wb_offset;
1144         else
1145                 req->wb_bytes = rqend - req->wb_offset;
1146         req->wb_nio = 0;
1147         return req;
1148 out_flushme:
1149         /*
1150          * Note: we mark the request dirty here because
1151          * nfs_lock_and_join_requests() cannot preserve
1152          * commit flags, so we have to replay the write.
1153          */
1154         nfs_mark_request_dirty(req);
1155         nfs_unlock_and_release_request(req);
1156         error = nfs_wb_page(inode, page);
1157         return (error < 0) ? ERR_PTR(error) : NULL;
1158 }
1159
1160 /*
1161  * Try to update an existing write request, or create one if there is none.
1162  *
1163  * Note: Should always be called with the Page Lock held to prevent races
1164  * if we have to add a new request. Also assumes that the caller has
1165  * already called nfs_flush_incompatible() if necessary.
1166  */
1167 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1168                 struct page *page, unsigned int offset, unsigned int bytes)
1169 {
1170         struct inode *inode = page_file_mapping(page)->host;
1171         struct nfs_page *req;
1172
1173         req = nfs_try_to_update_request(inode, page, offset, bytes);
1174         if (req != NULL)
1175                 goto out;
1176         req = nfs_create_request(ctx, page, offset, bytes);
1177         if (IS_ERR(req))
1178                 goto out;
1179         nfs_inode_add_request(inode, req);
1180 out:
1181         return req;
1182 }
1183
1184 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1185                 unsigned int offset, unsigned int count)
1186 {
1187         struct nfs_page *req;
1188
1189         req = nfs_setup_write_request(ctx, page, offset, count);
1190         if (IS_ERR(req))
1191                 return PTR_ERR(req);
1192         /* Update file length */
1193         nfs_grow_file(page, offset, count);
1194         nfs_mark_uptodate(req);
1195         nfs_mark_request_dirty(req);
1196         nfs_unlock_and_release_request(req);
1197         return 0;
1198 }
1199
1200 int nfs_flush_incompatible(struct file *file, struct page *page)
1201 {
1202         struct nfs_open_context *ctx = nfs_file_open_context(file);
1203         struct nfs_lock_context *l_ctx;
1204         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1205         struct nfs_page *req;
1206         int do_flush, status;
1207         /*
1208          * Look for a request corresponding to this page. If there
1209          * is one, and it belongs to another file, we flush it out
1210          * before we try to copy anything into the page. Do this
1211          * due to the lack of an ACCESS-type call in NFSv2.
1212          * Also do the same if we find a request from an existing
1213          * dropped page.
1214          */
1215         do {
1216                 req = nfs_page_find_head_request(page);
1217                 if (req == NULL)
1218                         return 0;
1219                 l_ctx = req->wb_lock_context;
1220                 do_flush = req->wb_page != page ||
1221                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1222                 if (l_ctx && flctx &&
1223                     !(list_empty_careful(&flctx->flc_posix) &&
1224                       list_empty_careful(&flctx->flc_flock))) {
1225                         do_flush |= l_ctx->lockowner != current->files;
1226                 }
1227                 nfs_release_request(req);
1228                 if (!do_flush)
1229                         return 0;
1230                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1231         } while (status == 0);
1232         return status;
1233 }
1234
1235 /*
1236  * Avoid buffered writes when a open context credential's key would
1237  * expire soon.
1238  *
1239  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1240  *
1241  * Return 0 and set a credential flag which triggers the inode to flush
1242  * and performs  NFS_FILE_SYNC writes if the key will expired within
1243  * RPC_KEY_EXPIRE_TIMEO.
1244  */
1245 int
1246 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1247 {
1248         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1249
1250         if (nfs_ctx_key_to_expire(ctx, inode) &&
1251             !ctx->ll_cred)
1252                 /* Already expired! */
1253                 return -EACCES;
1254         return 0;
1255 }
1256
1257 /*
1258  * Test if the open context credential key is marked to expire soon.
1259  */
1260 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1261 {
1262         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1263         struct rpc_cred *cred = ctx->ll_cred;
1264         struct auth_cred acred = {
1265                 .cred = ctx->cred,
1266         };
1267
1268         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1269                 put_rpccred(cred);
1270                 ctx->ll_cred = NULL;
1271                 cred = NULL;
1272         }
1273         if (!cred)
1274                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1275         if (!cred || IS_ERR(cred))
1276                 return true;
1277         ctx->ll_cred = cred;
1278         return !!(cred->cr_ops->crkey_timeout &&
1279                   cred->cr_ops->crkey_timeout(cred));
1280 }
1281
1282 /*
1283  * If the page cache is marked as unsafe or invalid, then we can't rely on
1284  * the PageUptodate() flag. In this case, we will need to turn off
1285  * write optimisations that depend on the page contents being correct.
1286  */
1287 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1288 {
1289         struct nfs_inode *nfsi = NFS_I(inode);
1290
1291         if (nfs_have_delegated_attributes(inode))
1292                 goto out;
1293         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1294                 return false;
1295         smp_rmb();
1296         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1297                 return false;
1298 out:
1299         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1300                 return false;
1301         return PageUptodate(page) != 0;
1302 }
1303
1304 static bool
1305 is_whole_file_wrlock(struct file_lock *fl)
1306 {
1307         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1308                         fl->fl_type == F_WRLCK;
1309 }
1310
1311 /* If we know the page is up to date, and we're not using byte range locks (or
1312  * if we have the whole file locked for writing), it may be more efficient to
1313  * extend the write to cover the entire page in order to avoid fragmentation
1314  * inefficiencies.
1315  *
1316  * If the file is opened for synchronous writes then we can just skip the rest
1317  * of the checks.
1318  */
1319 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1320 {
1321         int ret;
1322         struct file_lock_context *flctx = inode->i_flctx;
1323         struct file_lock *fl;
1324
1325         if (file->f_flags & O_DSYNC)
1326                 return 0;
1327         if (!nfs_write_pageuptodate(page, inode))
1328                 return 0;
1329         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1330                 return 1;
1331         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1332                        list_empty_careful(&flctx->flc_posix)))
1333                 return 1;
1334
1335         /* Check to see if there are whole file write locks */
1336         ret = 0;
1337         spin_lock(&flctx->flc_lock);
1338         if (!list_empty(&flctx->flc_posix)) {
1339                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1340                                         fl_list);
1341                 if (is_whole_file_wrlock(fl))
1342                         ret = 1;
1343         } else if (!list_empty(&flctx->flc_flock)) {
1344                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1345                                         fl_list);
1346                 if (fl->fl_type == F_WRLCK)
1347                         ret = 1;
1348         }
1349         spin_unlock(&flctx->flc_lock);
1350         return ret;
1351 }
1352
1353 /*
1354  * Update and possibly write a cached page of an NFS file.
1355  *
1356  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1357  * things with a page scheduled for an RPC call (e.g. invalidate it).
1358  */
1359 int nfs_updatepage(struct file *file, struct page *page,
1360                 unsigned int offset, unsigned int count)
1361 {
1362         struct nfs_open_context *ctx = nfs_file_open_context(file);
1363         struct address_space *mapping = page_file_mapping(page);
1364         struct inode    *inode = mapping->host;
1365         int             status = 0;
1366
1367         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1368
1369         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1370                 file, count, (long long)(page_file_offset(page) + offset));
1371
1372         if (!count)
1373                 goto out;
1374
1375         if (nfs_can_extend_write(file, page, inode)) {
1376                 count = max(count + offset, nfs_page_length(page));
1377                 offset = 0;
1378         }
1379
1380         status = nfs_writepage_setup(ctx, page, offset, count);
1381         if (status < 0)
1382                 nfs_set_pageerror(mapping);
1383         else
1384                 __set_page_dirty_nobuffers(page);
1385 out:
1386         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1387                         status, (long long)i_size_read(inode));
1388         return status;
1389 }
1390
1391 static int flush_task_priority(int how)
1392 {
1393         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1394                 case FLUSH_HIGHPRI:
1395                         return RPC_PRIORITY_HIGH;
1396                 case FLUSH_LOWPRI:
1397                         return RPC_PRIORITY_LOW;
1398         }
1399         return RPC_PRIORITY_NORMAL;
1400 }
1401
1402 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1403                                struct rpc_message *msg,
1404                                const struct nfs_rpc_ops *rpc_ops,
1405                                struct rpc_task_setup *task_setup_data, int how)
1406 {
1407         int priority = flush_task_priority(how);
1408
1409         task_setup_data->priority = priority;
1410         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1411         trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1412                                  hdr->args.stable);
1413 }
1414
1415 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1416  * call this on each, which will prepare them to be retried on next
1417  * writeback using standard nfs.
1418  */
1419 static void nfs_redirty_request(struct nfs_page *req)
1420 {
1421         /* Bump the transmission count */
1422         req->wb_nio++;
1423         nfs_mark_request_dirty(req);
1424         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1425         nfs_end_page_writeback(req);
1426         nfs_release_request(req);
1427 }
1428
1429 static void nfs_async_write_error(struct list_head *head, int error)
1430 {
1431         struct nfs_page *req;
1432
1433         while (!list_empty(head)) {
1434                 req = nfs_list_entry(head->next);
1435                 nfs_list_remove_request(req);
1436                 if (nfs_error_is_fatal(error))
1437                         nfs_write_error(req, error);
1438                 else
1439                         nfs_redirty_request(req);
1440         }
1441 }
1442
1443 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1444 {
1445         nfs_async_write_error(&hdr->pages, 0);
1446         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1447                         hdr->args.offset + hdr->args.count - 1);
1448 }
1449
1450 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1451         .init_hdr = nfs_async_write_init,
1452         .error_cleanup = nfs_async_write_error,
1453         .completion = nfs_write_completion,
1454         .reschedule_io = nfs_async_write_reschedule_io,
1455 };
1456
1457 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1458                                struct inode *inode, int ioflags, bool force_mds,
1459                                const struct nfs_pgio_completion_ops *compl_ops)
1460 {
1461         struct nfs_server *server = NFS_SERVER(inode);
1462         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1463
1464 #ifdef CONFIG_NFS_V4_1
1465         if (server->pnfs_curr_ld && !force_mds)
1466                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1467 #endif
1468         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1469                         server->wsize, ioflags);
1470 }
1471 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1472
1473 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1474 {
1475         struct nfs_pgio_mirror *mirror;
1476
1477         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1478                 pgio->pg_ops->pg_cleanup(pgio);
1479
1480         pgio->pg_ops = &nfs_pgio_rw_ops;
1481
1482         nfs_pageio_stop_mirroring(pgio);
1483
1484         mirror = &pgio->pg_mirrors[0];
1485         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1486 }
1487 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1488
1489
1490 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1491 {
1492         struct nfs_commit_data *data = calldata;
1493
1494         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1495 }
1496
1497 /*
1498  * Special version of should_remove_suid() that ignores capabilities.
1499  */
1500 static int nfs_should_remove_suid(const struct inode *inode)
1501 {
1502         umode_t mode = inode->i_mode;
1503         int kill = 0;
1504
1505         /* suid always must be killed */
1506         if (unlikely(mode & S_ISUID))
1507                 kill = ATTR_KILL_SUID;
1508
1509         /*
1510          * sgid without any exec bits is just a mandatory locking mark; leave
1511          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1512          */
1513         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1514                 kill |= ATTR_KILL_SGID;
1515
1516         if (unlikely(kill && S_ISREG(mode)))
1517                 return kill;
1518
1519         return 0;
1520 }
1521
1522 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1523                 struct nfs_fattr *fattr)
1524 {
1525         struct nfs_pgio_args *argp = &hdr->args;
1526         struct nfs_pgio_res *resp = &hdr->res;
1527         u64 size = argp->offset + resp->count;
1528
1529         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1530                 fattr->size = size;
1531         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1532                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1533                 return;
1534         }
1535         if (size != fattr->size)
1536                 return;
1537         /* Set attribute barrier */
1538         nfs_fattr_set_barrier(fattr);
1539         /* ...and update size */
1540         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1541 }
1542
1543 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1544 {
1545         struct nfs_fattr *fattr = &hdr->fattr;
1546         struct inode *inode = hdr->inode;
1547
1548         spin_lock(&inode->i_lock);
1549         nfs_writeback_check_extend(hdr, fattr);
1550         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1551         spin_unlock(&inode->i_lock);
1552 }
1553 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1554
1555 /*
1556  * This function is called when the WRITE call is complete.
1557  */
1558 static int nfs_writeback_done(struct rpc_task *task,
1559                               struct nfs_pgio_header *hdr,
1560                               struct inode *inode)
1561 {
1562         int status;
1563
1564         /*
1565          * ->write_done will attempt to use post-op attributes to detect
1566          * conflicting writes by other clients.  A strict interpretation
1567          * of close-to-open would allow us to continue caching even if
1568          * another writer had changed the file, but some applications
1569          * depend on tighter cache coherency when writing.
1570          */
1571         status = NFS_PROTO(inode)->write_done(task, hdr);
1572         if (status != 0)
1573                 return status;
1574
1575         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1576         trace_nfs_writeback_done(inode, task->tk_status,
1577                                  hdr->args.offset, hdr->res.verf);
1578
1579         if (hdr->res.verf->committed < hdr->args.stable &&
1580             task->tk_status >= 0) {
1581                 /* We tried a write call, but the server did not
1582                  * commit data to stable storage even though we
1583                  * requested it.
1584                  * Note: There is a known bug in Tru64 < 5.0 in which
1585                  *       the server reports NFS_DATA_SYNC, but performs
1586                  *       NFS_FILE_SYNC. We therefore implement this checking
1587                  *       as a dprintk() in order to avoid filling syslog.
1588                  */
1589                 static unsigned long    complain;
1590
1591                 /* Note this will print the MDS for a DS write */
1592                 if (time_before(complain, jiffies)) {
1593                         dprintk("NFS:       faulty NFS server %s:"
1594                                 " (committed = %d) != (stable = %d)\n",
1595                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1596                                 hdr->res.verf->committed, hdr->args.stable);
1597                         complain = jiffies + 300 * HZ;
1598                 }
1599         }
1600
1601         /* Deal with the suid/sgid bit corner case */
1602         if (nfs_should_remove_suid(inode)) {
1603                 spin_lock(&inode->i_lock);
1604                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1605                 spin_unlock(&inode->i_lock);
1606         }
1607         return 0;
1608 }
1609
1610 /*
1611  * This function is called when the WRITE call is complete.
1612  */
1613 static void nfs_writeback_result(struct rpc_task *task,
1614                                  struct nfs_pgio_header *hdr)
1615 {
1616         struct nfs_pgio_args    *argp = &hdr->args;
1617         struct nfs_pgio_res     *resp = &hdr->res;
1618
1619         if (resp->count < argp->count) {
1620                 static unsigned long    complain;
1621
1622                 /* This a short write! */
1623                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1624
1625                 /* Has the server at least made some progress? */
1626                 if (resp->count == 0) {
1627                         if (time_before(complain, jiffies)) {
1628                                 printk(KERN_WARNING
1629                                        "NFS: Server wrote zero bytes, expected %u.\n",
1630                                        argp->count);
1631                                 complain = jiffies + 300 * HZ;
1632                         }
1633                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1634                         task->tk_status = -EIO;
1635                         return;
1636                 }
1637
1638                 /* For non rpc-based layout drivers, retry-through-MDS */
1639                 if (!task->tk_ops) {
1640                         hdr->pnfs_error = -EAGAIN;
1641                         return;
1642                 }
1643
1644                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1645                 if (resp->verf->committed != NFS_UNSTABLE) {
1646                         /* Resend from where the server left off */
1647                         hdr->mds_offset += resp->count;
1648                         argp->offset += resp->count;
1649                         argp->pgbase += resp->count;
1650                         argp->count -= resp->count;
1651                 } else {
1652                         /* Resend as a stable write in order to avoid
1653                          * headaches in the case of a server crash.
1654                          */
1655                         argp->stable = NFS_FILE_SYNC;
1656                 }
1657                 rpc_restart_call_prepare(task);
1658         }
1659 }
1660
1661 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1662 {
1663         return wait_var_event_killable(&cinfo->rpcs_out,
1664                                        !atomic_read(&cinfo->rpcs_out));
1665 }
1666
1667 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1668 {
1669         atomic_inc(&cinfo->rpcs_out);
1670 }
1671
1672 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1673 {
1674         if (atomic_dec_and_test(&cinfo->rpcs_out))
1675                 wake_up_var(&cinfo->rpcs_out);
1676 }
1677
1678 void nfs_commitdata_release(struct nfs_commit_data *data)
1679 {
1680         put_nfs_open_context(data->context);
1681         nfs_commit_free(data);
1682 }
1683 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1684
1685 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1686                         const struct nfs_rpc_ops *nfs_ops,
1687                         const struct rpc_call_ops *call_ops,
1688                         int how, int flags)
1689 {
1690         struct rpc_task *task;
1691         int priority = flush_task_priority(how);
1692         struct rpc_message msg = {
1693                 .rpc_argp = &data->args,
1694                 .rpc_resp = &data->res,
1695                 .rpc_cred = data->cred,
1696         };
1697         struct rpc_task_setup task_setup_data = {
1698                 .task = &data->task,
1699                 .rpc_client = clnt,
1700                 .rpc_message = &msg,
1701                 .callback_ops = call_ops,
1702                 .callback_data = data,
1703                 .workqueue = nfsiod_workqueue,
1704                 .flags = RPC_TASK_ASYNC | flags,
1705                 .priority = priority,
1706         };
1707         /* Set up the initial task struct.  */
1708         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1709         trace_nfs_initiate_commit(data);
1710
1711         dprintk("NFS: initiated commit call\n");
1712
1713         task = rpc_run_task(&task_setup_data);
1714         if (IS_ERR(task))
1715                 return PTR_ERR(task);
1716         if (how & FLUSH_SYNC)
1717                 rpc_wait_for_completion_task(task);
1718         rpc_put_task(task);
1719         return 0;
1720 }
1721 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1722
1723 static loff_t nfs_get_lwb(struct list_head *head)
1724 {
1725         loff_t lwb = 0;
1726         struct nfs_page *req;
1727
1728         list_for_each_entry(req, head, wb_list)
1729                 if (lwb < (req_offset(req) + req->wb_bytes))
1730                         lwb = req_offset(req) + req->wb_bytes;
1731
1732         return lwb;
1733 }
1734
1735 /*
1736  * Set up the argument/result storage required for the RPC call.
1737  */
1738 void nfs_init_commit(struct nfs_commit_data *data,
1739                      struct list_head *head,
1740                      struct pnfs_layout_segment *lseg,
1741                      struct nfs_commit_info *cinfo)
1742 {
1743         struct nfs_page *first = nfs_list_entry(head->next);
1744         struct nfs_open_context *ctx = nfs_req_openctx(first);
1745         struct inode *inode = d_inode(ctx->dentry);
1746
1747         /* Set up the RPC argument and reply structs
1748          * NB: take care not to mess about with data->commit et al. */
1749
1750         list_splice_init(head, &data->pages);
1751
1752         data->inode       = inode;
1753         data->cred        = ctx->cred;
1754         data->lseg        = lseg; /* reference transferred */
1755         /* only set lwb for pnfs commit */
1756         if (lseg)
1757                 data->lwb = nfs_get_lwb(&data->pages);
1758         data->mds_ops     = &nfs_commit_ops;
1759         data->completion_ops = cinfo->completion_ops;
1760         data->dreq        = cinfo->dreq;
1761
1762         data->args.fh     = NFS_FH(data->inode);
1763         /* Note: we always request a commit of the entire inode */
1764         data->args.offset = 0;
1765         data->args.count  = 0;
1766         data->context     = get_nfs_open_context(ctx);
1767         data->res.fattr   = &data->fattr;
1768         data->res.verf    = &data->verf;
1769         nfs_fattr_init(&data->fattr);
1770 }
1771 EXPORT_SYMBOL_GPL(nfs_init_commit);
1772
1773 void nfs_retry_commit(struct list_head *page_list,
1774                       struct pnfs_layout_segment *lseg,
1775                       struct nfs_commit_info *cinfo,
1776                       u32 ds_commit_idx)
1777 {
1778         struct nfs_page *req;
1779
1780         while (!list_empty(page_list)) {
1781                 req = nfs_list_entry(page_list->next);
1782                 nfs_list_remove_request(req);
1783                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1784                 if (!cinfo->dreq)
1785                         nfs_clear_page_commit(req->wb_page);
1786                 nfs_unlock_and_release_request(req);
1787         }
1788 }
1789 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1790
1791 static void
1792 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1793                 struct nfs_page *req)
1794 {
1795         __set_page_dirty_nobuffers(req->wb_page);
1796 }
1797
1798 /*
1799  * Commit dirty pages
1800  */
1801 static int
1802 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1803                 struct nfs_commit_info *cinfo)
1804 {
1805         struct nfs_commit_data  *data;
1806
1807         /* another commit raced with us */
1808         if (list_empty(head))
1809                 return 0;
1810
1811         data = nfs_commitdata_alloc(true);
1812
1813         /* Set up the argument struct */
1814         nfs_init_commit(data, head, NULL, cinfo);
1815         atomic_inc(&cinfo->mds->rpcs_out);
1816         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1817                                    data->mds_ops, how, 0);
1818 }
1819
1820 /*
1821  * COMMIT call returned
1822  */
1823 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1824 {
1825         struct nfs_commit_data  *data = calldata;
1826
1827         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1828                                 task->tk_pid, task->tk_status);
1829
1830         /* Call the NFS version-specific code */
1831         NFS_PROTO(data->inode)->commit_done(task, data);
1832         trace_nfs_commit_done(data);
1833 }
1834
1835 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1836 {
1837         struct nfs_page *req;
1838         int status = data->task.tk_status;
1839         struct nfs_commit_info cinfo;
1840         struct nfs_server *nfss;
1841
1842         while (!list_empty(&data->pages)) {
1843                 req = nfs_list_entry(data->pages.next);
1844                 nfs_list_remove_request(req);
1845                 if (req->wb_page)
1846                         nfs_clear_page_commit(req->wb_page);
1847
1848                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1849                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1850                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1851                         req->wb_bytes,
1852                         (long long)req_offset(req));
1853                 if (status < 0) {
1854                         if (req->wb_page) {
1855                                 nfs_mapping_set_error(req->wb_page, status);
1856                                 nfs_inode_remove_request(req);
1857                         }
1858                         dprintk_cont(", error = %d\n", status);
1859                         goto next;
1860                 }
1861
1862                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1863                  * returned by the server against all stored verfs. */
1864                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1865                         /* We have a match */
1866                         if (req->wb_page)
1867                                 nfs_inode_remove_request(req);
1868                         dprintk_cont(" OK\n");
1869                         goto next;
1870                 }
1871                 /* We have a mismatch. Write the page again */
1872                 dprintk_cont(" mismatch\n");
1873                 nfs_mark_request_dirty(req);
1874                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1875         next:
1876                 nfs_unlock_and_release_request(req);
1877                 /* Latency breaker */
1878                 cond_resched();
1879         }
1880         nfss = NFS_SERVER(data->inode);
1881         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1882                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1883
1884         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1885         nfs_commit_end(cinfo.mds);
1886 }
1887
1888 static void nfs_commit_release(void *calldata)
1889 {
1890         struct nfs_commit_data *data = calldata;
1891
1892         data->completion_ops->completion(data);
1893         nfs_commitdata_release(calldata);
1894 }
1895
1896 static const struct rpc_call_ops nfs_commit_ops = {
1897         .rpc_call_prepare = nfs_commit_prepare,
1898         .rpc_call_done = nfs_commit_done,
1899         .rpc_release = nfs_commit_release,
1900 };
1901
1902 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1903         .completion = nfs_commit_release_pages,
1904         .resched_write = nfs_commit_resched_write,
1905 };
1906
1907 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1908                             int how, struct nfs_commit_info *cinfo)
1909 {
1910         int status;
1911
1912         status = pnfs_commit_list(inode, head, how, cinfo);
1913         if (status == PNFS_NOT_ATTEMPTED)
1914                 status = nfs_commit_list(inode, head, how, cinfo);
1915         return status;
1916 }
1917
1918 static int __nfs_commit_inode(struct inode *inode, int how,
1919                 struct writeback_control *wbc)
1920 {
1921         LIST_HEAD(head);
1922         struct nfs_commit_info cinfo;
1923         int may_wait = how & FLUSH_SYNC;
1924         int ret, nscan;
1925
1926         nfs_init_cinfo_from_inode(&cinfo, inode);
1927         nfs_commit_begin(cinfo.mds);
1928         for (;;) {
1929                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1930                 if (ret <= 0)
1931                         break;
1932                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1933                 if (ret < 0)
1934                         break;
1935                 ret = 0;
1936                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1937                         if (nscan < wbc->nr_to_write)
1938                                 wbc->nr_to_write -= nscan;
1939                         else
1940                                 wbc->nr_to_write = 0;
1941                 }
1942                 if (nscan < INT_MAX)
1943                         break;
1944                 cond_resched();
1945         }
1946         nfs_commit_end(cinfo.mds);
1947         if (ret || !may_wait)
1948                 return ret;
1949         return wait_on_commit(cinfo.mds);
1950 }
1951
1952 int nfs_commit_inode(struct inode *inode, int how)
1953 {
1954         return __nfs_commit_inode(inode, how, NULL);
1955 }
1956 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1957
1958 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1959 {
1960         struct nfs_inode *nfsi = NFS_I(inode);
1961         int flags = FLUSH_SYNC;
1962         int ret = 0;
1963
1964         if (wbc->sync_mode == WB_SYNC_NONE) {
1965                 /* no commits means nothing needs to be done */
1966                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1967                         goto check_requests_outstanding;
1968
1969                 /* Don't commit yet if this is a non-blocking flush and there
1970                  * are a lot of outstanding writes for this mapping.
1971                  */
1972                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1973                         goto out_mark_dirty;
1974
1975                 /* don't wait for the COMMIT response */
1976                 flags = 0;
1977         }
1978
1979         ret = __nfs_commit_inode(inode, flags, wbc);
1980         if (!ret) {
1981                 if (flags & FLUSH_SYNC)
1982                         return 0;
1983         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1984                 goto out_mark_dirty;
1985
1986 check_requests_outstanding:
1987         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1988                 return ret;
1989 out_mark_dirty:
1990         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1991         return ret;
1992 }
1993 EXPORT_SYMBOL_GPL(nfs_write_inode);
1994
1995 /*
1996  * Wrapper for filemap_write_and_wait_range()
1997  *
1998  * Needed for pNFS in order to ensure data becomes visible to the
1999  * client.
2000  */
2001 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2002                 loff_t lstart, loff_t lend)
2003 {
2004         int ret;
2005
2006         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2007         if (ret == 0)
2008                 ret = pnfs_sync_inode(mapping->host, true);
2009         return ret;
2010 }
2011 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2012
2013 /*
2014  * flush the inode to disk.
2015  */
2016 int nfs_wb_all(struct inode *inode)
2017 {
2018         int ret;
2019
2020         trace_nfs_writeback_inode_enter(inode);
2021
2022         ret = filemap_write_and_wait(inode->i_mapping);
2023         if (ret)
2024                 goto out;
2025         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2026         if (ret < 0)
2027                 goto out;
2028         pnfs_sync_inode(inode, true);
2029         ret = 0;
2030
2031 out:
2032         trace_nfs_writeback_inode_exit(inode, ret);
2033         return ret;
2034 }
2035 EXPORT_SYMBOL_GPL(nfs_wb_all);
2036
2037 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2038 {
2039         struct nfs_page *req;
2040         int ret = 0;
2041
2042         wait_on_page_writeback(page);
2043
2044         /* blocking call to cancel all requests and join to a single (head)
2045          * request */
2046         req = nfs_lock_and_join_requests(page);
2047
2048         if (IS_ERR(req)) {
2049                 ret = PTR_ERR(req);
2050         } else if (req) {
2051                 /* all requests from this page have been cancelled by
2052                  * nfs_lock_and_join_requests, so just remove the head
2053                  * request from the inode / page_private pointer and
2054                  * release it */
2055                 nfs_inode_remove_request(req);
2056                 nfs_unlock_and_release_request(req);
2057         }
2058
2059         return ret;
2060 }
2061
2062 /*
2063  * Write back all requests on one page - we do this before reading it.
2064  */
2065 int nfs_wb_page(struct inode *inode, struct page *page)
2066 {
2067         loff_t range_start = page_file_offset(page);
2068         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2069         struct writeback_control wbc = {
2070                 .sync_mode = WB_SYNC_ALL,
2071                 .nr_to_write = 0,
2072                 .range_start = range_start,
2073                 .range_end = range_end,
2074         };
2075         int ret;
2076
2077         trace_nfs_writeback_page_enter(inode);
2078
2079         for (;;) {
2080                 wait_on_page_writeback(page);
2081                 if (clear_page_dirty_for_io(page)) {
2082                         ret = nfs_writepage_locked(page, &wbc);
2083                         if (ret < 0)
2084                                 goto out_error;
2085                         continue;
2086                 }
2087                 ret = 0;
2088                 if (!PagePrivate(page))
2089                         break;
2090                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2091                 if (ret < 0)
2092                         goto out_error;
2093         }
2094 out_error:
2095         trace_nfs_writeback_page_exit(inode, ret);
2096         return ret;
2097 }
2098
2099 #ifdef CONFIG_MIGRATION
2100 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2101                 struct page *page, enum migrate_mode mode)
2102 {
2103         /*
2104          * If PagePrivate is set, then the page is currently associated with
2105          * an in-progress read or write request. Don't try to migrate it.
2106          *
2107          * FIXME: we could do this in principle, but we'll need a way to ensure
2108          *        that we can safely release the inode reference while holding
2109          *        the page lock.
2110          */
2111         if (PagePrivate(page))
2112                 return -EBUSY;
2113
2114         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2115                 return -EBUSY;
2116
2117         return migrate_page(mapping, newpage, page, mode);
2118 }
2119 #endif
2120
2121 int __init nfs_init_writepagecache(void)
2122 {
2123         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2124                                              sizeof(struct nfs_pgio_header),
2125                                              0, SLAB_HWCACHE_ALIGN,
2126                                              NULL);
2127         if (nfs_wdata_cachep == NULL)
2128                 return -ENOMEM;
2129
2130         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2131                                                      nfs_wdata_cachep);
2132         if (nfs_wdata_mempool == NULL)
2133                 goto out_destroy_write_cache;
2134
2135         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2136                                              sizeof(struct nfs_commit_data),
2137                                              0, SLAB_HWCACHE_ALIGN,
2138                                              NULL);
2139         if (nfs_cdata_cachep == NULL)
2140                 goto out_destroy_write_mempool;
2141
2142         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2143                                                       nfs_cdata_cachep);
2144         if (nfs_commit_mempool == NULL)
2145                 goto out_destroy_commit_cache;
2146
2147         /*
2148          * NFS congestion size, scale with available memory.
2149          *
2150          *  64MB:    8192k
2151          * 128MB:   11585k
2152          * 256MB:   16384k
2153          * 512MB:   23170k
2154          *   1GB:   32768k
2155          *   2GB:   46340k
2156          *   4GB:   65536k
2157          *   8GB:   92681k
2158          *  16GB:  131072k
2159          *
2160          * This allows larger machines to have larger/more transfers.
2161          * Limit the default to 256M
2162          */
2163         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2164         if (nfs_congestion_kb > 256*1024)
2165                 nfs_congestion_kb = 256*1024;
2166
2167         return 0;
2168
2169 out_destroy_commit_cache:
2170         kmem_cache_destroy(nfs_cdata_cachep);
2171 out_destroy_write_mempool:
2172         mempool_destroy(nfs_wdata_mempool);
2173 out_destroy_write_cache:
2174         kmem_cache_destroy(nfs_wdata_cachep);
2175         return -ENOMEM;
2176 }
2177
2178 void nfs_destroy_writepagecache(void)
2179 {
2180         mempool_destroy(nfs_commit_mempool);
2181         kmem_cache_destroy(nfs_cdata_cachep);
2182         mempool_destroy(nfs_wdata_mempool);
2183         kmem_cache_destroy(nfs_wdata_cachep);
2184 }
2185
2186 static const struct nfs_rw_ops nfs_rw_write_ops = {
2187         .rw_alloc_header        = nfs_writehdr_alloc,
2188         .rw_free_header         = nfs_writehdr_free,
2189         .rw_done                = nfs_writeback_done,
2190         .rw_result              = nfs_writeback_result,
2191         .rw_initiate            = nfs_initiate_write,
2192 };