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