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         struct inode *inode = mapping->host;
 247
 248         nfs_zap_mapping(mapping->host, mapping);
 249         /* Force file size revalidation */
 250         spin_lock(&inode->i_lock);
 251         NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
 252                                         NFS_INO_REVAL_PAGECACHE |
 253                                         NFS_INO_INVALID_SIZE;
 254         spin_unlock(&inode->i_lock);
 255 }
 256
 257 static void nfs_mapping_set_error(struct page *page, int error)
 258 {
 259         SetPageError(page);
 260         mapping_set_error(page_file_mapping(page), error);
 261 }
 262
 263 /*
 264  * nfs_page_group_search_locked
 265  * @head - head request of page group
 266  * @page_offset - offset into page
 267  *
 268  * Search page group with head @head to find a request that contains the
 269  * page offset @page_offset.
 270  *
 271  * Returns a pointer to the first matching nfs request, or NULL if no
 272  * match is found.
 273  *
 274  * Must be called with the page group lock held
 275  */
 276 static struct nfs_page *
 277 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 278 {
 279         struct nfs_page *req;
 280
 281         req = head;
 282         do {
 283                 if (page_offset >= req->wb_pgbase &&
 284                     page_offset < (req->wb_pgbase + req->wb_bytes))
 285                         return req;
 286
 287                 req = req->wb_this_page;
 288         } while (req != head);
 289
 290         return NULL;
 291 }
 292
 293 /*
 294  * nfs_page_group_covers_page
 295  * @head - head request of page group
 296  *
 297  * Return true if the page group with head @head covers the whole page,
 298  * returns false otherwise
 299  */
 300 static bool nfs_page_group_covers_page(struct nfs_page *req)
 301 {
 302         struct nfs_page *tmp;
 303         unsigned int pos = 0;
 304         unsigned int len = nfs_page_length(req->wb_page);
 305
 306         nfs_page_group_lock(req);
 307
 308         for (;;) {
 309                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 310                 if (!tmp)
 311                         break;
 312                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 313         }
 314
 315         nfs_page_group_unlock(req);
 316         return pos >= len;
 317 }
 318
 319 /* We can set the PG_uptodate flag if we see that a write request
 320  * covers the full page.
 321  */
 322 static void nfs_mark_uptodate(struct nfs_page *req)
 323 {
 324         if (PageUptodate(req->wb_page))
 325                 return;
 326         if (!nfs_page_group_covers_page(req))
 327                 return;
 328         SetPageUptodate(req->wb_page);
 329 }
 330
 331 static int wb_priority(struct writeback_control *wbc)
 332 {
 333         int ret = 0;
 334
 335         if (wbc->sync_mode == WB_SYNC_ALL)
 336                 ret = FLUSH_COND_STABLE;
 337         return ret;
 338 }
 339
 340 /*
 341  * NFS congestion control
 342  */
 343
 344 int nfs_congestion_kb;
 345
 346 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 347 #define NFS_CONGESTION_OFF_THRESH       \
 348         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 349
 350 static void nfs_set_page_writeback(struct page *page)
 351 {
 352         struct inode *inode = page_file_mapping(page)->host;
 353         struct nfs_server *nfss = NFS_SERVER(inode);
 354         int ret = test_set_page_writeback(page);
 355
 356         WARN_ON_ONCE(ret != 0);
 357
 358         if (atomic_long_inc_return(&nfss->writeback) >
 359                         NFS_CONGESTION_ON_THRESH)
 360                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 361 }
 362
 363 static void nfs_end_page_writeback(struct nfs_page *req)
 364 {
 365         struct inode *inode = page_file_mapping(req->wb_page)->host;
 366         struct nfs_server *nfss = NFS_SERVER(inode);
 367         bool is_done;
 368
 369         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 370         nfs_unlock_request(req);
 371         if (!is_done)
 372                 return;
 373
 374         end_page_writeback(req->wb_page);
 375         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 376                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 377 }
 378
 379 /*
 380  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 381  *
 382  * this is a helper function for nfs_lock_and_join_requests
 383  *
 384  * @inode - inode associated with request page group, must be holding inode lock
 385  * @head  - head request of page group, must be holding head lock
 386  * @req   - request that couldn't lock and needs to wait on the req bit lock
 387  *
 388  * NOTE: this must be called holding page_group bit lock
 389  *       which will be released before returning.
 390  *
 391  * returns 0 on success, < 0 on error.
 392  */
 393 static void
 394 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
 395                           struct nfs_page *req)
 396 {
 397         struct nfs_page *tmp;
 398
 399         /* relinquish all the locks successfully grabbed this run */
 400         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 401                 if (!kref_read(&tmp->wb_kref))
 402                         continue;
 403                 nfs_unlock_and_release_request(tmp);
 404         }
 405 }
 406
 407 /*
 408  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 409  *
 410  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 411  * @old_head - the old head of the list
 412  *
 413  * All subrequests must be locked and removed from all lists, so at this point
 414  * they are only "active" in this function, and possibly in nfs_wait_on_request
 415  * with a reference held by some other context.
 416  */
 417 static void
 418 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 419                                  struct nfs_page *old_head,
 420                                  struct inode *inode)
 421 {
 422         while (destroy_list) {
 423                 struct nfs_page *subreq = destroy_list;
 424
 425                 destroy_list = (subreq->wb_this_page == old_head) ?
 426                                    NULL : subreq->wb_this_page;
 427
 428                 WARN_ON_ONCE(old_head != subreq->wb_head);
 429
 430                 /* make sure old group is not used */
 431                 subreq->wb_this_page = subreq;
 432
 433                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 434
 435                 /* Note: races with nfs_page_group_destroy() */
 436                 if (!kref_read(&subreq->wb_kref)) {
 437                         /* Check if we raced with nfs_page_group_destroy() */
 438                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
 439                                 nfs_free_request(subreq);
 440                         continue;
 441                 }
 442
 443                 subreq->wb_head = subreq;
 444
 445                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 446                         nfs_release_request(subreq);
 447                         atomic_long_dec(&NFS_I(inode)->nrequests);
 448                 }
 449
 450                 /* subreq is now totally disconnected from page group or any
 451                  * write / commit lists. last chance to wake any waiters */
 452                 nfs_unlock_and_release_request(subreq);
 453         }
 454 }
 455
 456 /*
 457  * nfs_lock_and_join_requests - join all subreqs to the head req and return
 458  *                              a locked reference, cancelling any pending
 459  *                              operations for this page.
 460  *
 461  * @page - the page used to lookup the "page group" of nfs_page structures
 462  *
 463  * This function joins all sub requests to the head request by first
 464  * locking all requests in the group, cancelling any pending operations
 465  * and finally updating the head request to cover the whole range covered by
 466  * the (former) group.  All subrequests are removed from any write or commit
 467  * lists, unlinked from the group and destroyed.
 468  *
 469  * Returns a locked, referenced pointer to the head request - which after
 470  * this call is guaranteed to be the only request associated with the page.
 471  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 472  * error was encountered.
 473  */
 474 static struct nfs_page *
 475 nfs_lock_and_join_requests(struct page *page)
 476 {
 477         struct inode *inode = page_file_mapping(page)->host;
 478         struct nfs_page *head, *subreq;
 479         struct nfs_page *destroy_list = NULL;
 480         unsigned int total_bytes;
 481         int ret;
 482
 483 try_again:
 484         /*
 485          * A reference is taken only on the head request which acts as a
 486          * reference to the whole page group - the group will not be destroyed
 487          * until the head reference is released.
 488          */
 489         head = nfs_page_find_head_request(page);
 490         if (!head)
 491                 return NULL;
 492
 493         /* lock the page head first in order to avoid an ABBA inefficiency */
 494         if (!nfs_lock_request(head)) {
 495                 ret = nfs_wait_on_request(head);
 496                 nfs_release_request(head);
 497                 if (ret < 0)
 498                         return ERR_PTR(ret);
 499                 goto try_again;
 500         }
 501
 502         /* Ensure that nobody removed the request before we locked it */
 503         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
 504                 nfs_unlock_and_release_request(head);
 505                 goto try_again;
 506         }
 507
 508         ret = nfs_page_group_lock(head);
 509         if (ret < 0)
 510                 goto release_request;
 511
 512         /* lock each request in the page group */
 513         total_bytes = head->wb_bytes;
 514         for (subreq = head->wb_this_page; subreq != head;
 515                         subreq = subreq->wb_this_page) {
 516
 517                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
 518                         if (subreq->wb_offset == head->wb_offset + total_bytes)
 519                                 total_bytes += subreq->wb_bytes;
 520                         continue;
 521                 }
 522
 523                 while (!nfs_lock_request(subreq)) {
 524                         /*
 525                          * Unlock page to allow nfs_page_group_sync_on_bit()
 526                          * to succeed
 527                          */
 528                         nfs_page_group_unlock(head);
 529                         ret = nfs_wait_on_request(subreq);
 530                         if (!ret)
 531                                 ret = nfs_page_group_lock(head);
 532                         if (ret < 0) {
 533                                 nfs_unroll_locks(inode, head, subreq);
 534                                 nfs_release_request(subreq);
 535                                 goto release_request;
 536                         }
 537                 }
 538                 /*
 539                  * Subrequests are always contiguous, non overlapping
 540                  * and in order - but may be repeated (mirrored writes).
 541                  */
 542                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
 543                         /* keep track of how many bytes this group covers */
 544                         total_bytes += subreq->wb_bytes;
 545                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
 546                             ((subreq->wb_offset + subreq->wb_bytes) >
 547                              (head->wb_offset + total_bytes)))) {
 548                         nfs_page_group_unlock(head);
 549                         nfs_unroll_locks(inode, head, subreq);
 550                         nfs_unlock_and_release_request(subreq);
 551                         ret = -EIO;
 552                         goto release_request;
 553                 }
 554         }
 555
 556         /* Now that all requests are locked, make sure they aren't on any list.
 557          * Commit list removal accounting is done after locks are dropped */
 558         subreq = head;
 559         do {
 560                 nfs_clear_request_commit(subreq);
 561                 subreq = subreq->wb_this_page;
 562         } while (subreq != head);
 563
 564         /* unlink subrequests from head, destroy them later */
 565         if (head->wb_this_page != head) {
 566                 /* destroy list will be terminated by head */
 567                 destroy_list = head->wb_this_page;
 568                 head->wb_this_page = head;
 569
 570                 /* change head request to cover whole range that
 571                  * the former page group covered */
 572                 head->wb_bytes = total_bytes;
 573         }
 574
 575         /* Postpone destruction of this request */
 576         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
 577                 set_bit(PG_INODE_REF, &head->wb_flags);
 578                 kref_get(&head->wb_kref);
 579                 atomic_long_inc(&NFS_I(inode)->nrequests);
 580         }
 581
 582         nfs_page_group_unlock(head);
 583
 584         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 585
 586         /* Did we lose a race with nfs_inode_remove_request()? */
 587         if (!(PagePrivate(page) || PageSwapCache(page))) {
 588                 nfs_unlock_and_release_request(head);
 589                 return NULL;
 590         }
 591
 592         /* still holds ref on head from nfs_page_find_head_request
 593          * and still has lock on head from lock loop */
 594         return head;
 595
 596 release_request:
 597         nfs_unlock_and_release_request(head);
 598         return ERR_PTR(ret);
 599 }
 600
 601 static void nfs_write_error(struct nfs_page *req, int error)
 602 {
 603         nfs_set_pageerror(page_file_mapping(req->wb_page));
 604         trace_nfs_write_error(req, error);
 605         nfs_mapping_set_error(req->wb_page, error);
 606         nfs_inode_remove_request(req);
 607         nfs_end_page_writeback(req);
 608         nfs_release_request(req);
 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 nfs_page *req;
 619         int ret = 0;
 620
 621         req = nfs_lock_and_join_requests(page);
 622         if (!req)
 623                 goto out;
 624         ret = PTR_ERR(req);
 625         if (IS_ERR(req))
 626                 goto out;
 627
 628         nfs_set_page_writeback(page);
 629         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 630
 631         /* If there is a fatal error that covers this write, just exit */
 632         ret = pgio->pg_error;
 633         if (nfs_error_is_fatal_on_server(ret))
 634                 goto out_launder;
 635
 636         ret = 0;
 637         if (!nfs_pageio_add_request(pgio, req)) {
 638                 ret = pgio->pg_error;
 639                 /*
 640                  * Remove the problematic req upon fatal errors on the server
 641                  */
 642                 if (nfs_error_is_fatal(ret)) {
 643                         if (nfs_error_is_fatal_on_server(ret))
 644                                 goto out_launder;
 645                 } else
 646                         ret = -EAGAIN;
 647                 nfs_redirty_request(req);
 648                 pgio->pg_error = 0;
 649         } else
 650                 nfs_add_stats(page_file_mapping(page)->host,
 651                                 NFSIOS_WRITEPAGES, 1);
 652 out:
 653         return ret;
 654 out_launder:
 655         nfs_write_error(req, ret);
 656         return 0;
 657 }
 658
 659 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 660                             struct nfs_pageio_descriptor *pgio)
 661 {
 662         int ret;
 663
 664         nfs_pageio_cond_complete(pgio, page_index(page));
 665         ret = nfs_page_async_flush(pgio, page);
 666         if (ret == -EAGAIN) {
 667                 redirty_page_for_writepage(wbc, page);
 668                 ret = AOP_WRITEPAGE_ACTIVATE;
 669         }
 670         return ret;
 671 }
 672
 673 /*
 674  * Write an mmapped page to the server.
 675  */
 676 static int nfs_writepage_locked(struct page *page,
 677                                 struct writeback_control *wbc)
 678 {
 679         struct nfs_pageio_descriptor pgio;
 680         struct inode *inode = page_file_mapping(page)->host;
 681         int err;
 682
 683         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 684         nfs_pageio_init_write(&pgio, inode, 0,
 685                                 false, &nfs_async_write_completion_ops);
 686         err = nfs_do_writepage(page, wbc, &pgio);
 687         pgio.pg_error = 0;
 688         nfs_pageio_complete(&pgio);
 689         if (err < 0)
 690                 return err;
 691         if (nfs_error_is_fatal(pgio.pg_error))
 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         if (ret != AOP_WRITEPAGE_ACTIVATE)
 702                 unlock_page(page);
 703         return ret;
 704 }
 705
 706 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 707 {
 708         int ret;
 709
 710         ret = nfs_do_writepage(page, wbc, data);
 711         if (ret != AOP_WRITEPAGE_ACTIVATE)
 712                 unlock_page(page);
 713         return ret;
 714 }
 715
 716 static void nfs_io_completion_commit(void *inode)
 717 {
 718         nfs_commit_inode(inode, 0);
 719 }
 720
 721 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 722 {
 723         struct inode *inode = mapping->host;
 724         struct nfs_pageio_descriptor pgio;
 725         struct nfs_io_completion *ioc;
 726         int err;
 727
 728         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 729
 730         ioc = nfs_io_completion_alloc(GFP_KERNEL);
 731         if (ioc)
 732                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 733
 734         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 735                                 &nfs_async_write_completion_ops);
 736         pgio.pg_io_completion = ioc;
 737         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 738         pgio.pg_error = 0;
 739         nfs_pageio_complete(&pgio);
 740         nfs_io_completion_put(ioc);
 741
 742         if (err < 0)
 743                 goto out_err;
 744         err = pgio.pg_error;
 745         if (nfs_error_is_fatal(err))
 746                 goto out_err;
 747         return 0;
 748 out_err:
 749         return err;
 750 }
 751
 752 /*
 753  * Insert a write request into an inode
 754  */
 755 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 756 {
 757         struct address_space *mapping = page_file_mapping(req->wb_page);
 758         struct nfs_inode *nfsi = NFS_I(inode);
 759
 760         WARN_ON_ONCE(req->wb_this_page != req);
 761
 762         /* Lock the request! */
 763         nfs_lock_request(req);
 764
 765         /*
 766          * Swap-space should not get truncated. Hence no need to plug the race
 767          * with invalidate/truncate.
 768          */
 769         spin_lock(&mapping->private_lock);
 770         if (!nfs_have_writebacks(inode) &&
 771             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 772                 inode_inc_iversion_raw(inode);
 773         if (likely(!PageSwapCache(req->wb_page))) {
 774                 set_bit(PG_MAPPED, &req->wb_flags);
 775                 SetPagePrivate(req->wb_page);
 776                 set_page_private(req->wb_page, (unsigned long)req);
 777         }
 778         spin_unlock(&mapping->private_lock);
 779         atomic_long_inc(&nfsi->nrequests);
 780         /* this a head request for a page group - mark it as having an
 781          * extra reference so sub groups can follow suit.
 782          * This flag also informs pgio layer when to bump nrequests when
 783          * adding subrequests. */
 784         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 785         kref_get(&req->wb_kref);
 786 }
 787
 788 /*
 789  * Remove a write request from an inode
 790  */
 791 static void nfs_inode_remove_request(struct nfs_page *req)
 792 {
 793         struct address_space *mapping = page_file_mapping(req->wb_page);
 794         struct inode *inode = mapping->host;
 795         struct nfs_inode *nfsi = NFS_I(inode);
 796         struct nfs_page *head;
 797
 798         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 799                 head = req->wb_head;
 800
 801                 spin_lock(&mapping->private_lock);
 802                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 803                         set_page_private(head->wb_page, 0);
 804                         ClearPagePrivate(head->wb_page);
 805                         clear_bit(PG_MAPPED, &head->wb_flags);
 806                 }
 807                 spin_unlock(&mapping->private_lock);
 808         }
 809
 810         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 811                 nfs_release_request(req);
 812                 atomic_long_dec(&nfsi->nrequests);
 813         }
 814 }
 815
 816 static void
 817 nfs_mark_request_dirty(struct nfs_page *req)
 818 {
 819         if (req->wb_page)
 820                 __set_page_dirty_nobuffers(req->wb_page);
 821 }
 822
 823 /*
 824  * nfs_page_search_commits_for_head_request_locked
 825  *
 826  * Search through commit lists on @inode for the head request for @page.
 827  * Must be called while holding the inode (which is cinfo) lock.
 828  *
 829  * Returns the head request if found, or NULL if not found.
 830  */
 831 static struct nfs_page *
 832 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 833                                                 struct page *page)
 834 {
 835         struct nfs_page *freq, *t;
 836         struct nfs_commit_info cinfo;
 837         struct inode *inode = &nfsi->vfs_inode;
 838
 839         nfs_init_cinfo_from_inode(&cinfo, inode);
 840
 841         /* search through pnfs commit lists */
 842         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 843         if (freq)
 844                 return freq->wb_head;
 845
 846         /* Linearly search the commit list for the correct request */
 847         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 848                 if (freq->wb_page == page)
 849                         return freq->wb_head;
 850         }
 851
 852         return NULL;
 853 }
 854
 855 /**
 856  * nfs_request_add_commit_list_locked - add request to a commit list
 857  * @req: pointer to a struct nfs_page
 858  * @dst: commit list head
 859  * @cinfo: holds list lock and accounting info
 860  *
 861  * This sets the PG_CLEAN bit, updates the cinfo count of
 862  * number of outstanding requests requiring a commit as well as
 863  * the MM page stats.
 864  *
 865  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 866  * nfs_page lock.
 867  */
 868 void
 869 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 870                             struct nfs_commit_info *cinfo)
 871 {
 872         set_bit(PG_CLEAN, &req->wb_flags);
 873         nfs_list_add_request(req, dst);
 874         atomic_long_inc(&cinfo->mds->ncommit);
 875 }
 876 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 877
 878 /**
 879  * nfs_request_add_commit_list - add request to a commit list
 880  * @req: pointer to a struct nfs_page
 881  * @cinfo: holds list lock and accounting info
 882  *
 883  * This sets the PG_CLEAN bit, updates the cinfo count of
 884  * number of outstanding requests requiring a commit as well as
 885  * the MM page stats.
 886  *
 887  * The caller must _not_ hold the cinfo->lock, but must be
 888  * holding the nfs_page lock.
 889  */
 890 void
 891 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 892 {
 893         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 894         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 895         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 896         if (req->wb_page)
 897                 nfs_mark_page_unstable(req->wb_page, cinfo);
 898 }
 899 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 900
 901 /**
 902  * nfs_request_remove_commit_list - Remove request from a commit list
 903  * @req: pointer to a nfs_page
 904  * @cinfo: holds list lock and accounting info
 905  *
 906  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 907  * number of outstanding requests requiring a commit
 908  * It does not update the MM page stats.
 909  *
 910  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 911  */
 912 void
 913 nfs_request_remove_commit_list(struct nfs_page *req,
 914                                struct nfs_commit_info *cinfo)
 915 {
 916         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 917                 return;
 918         nfs_list_remove_request(req);
 919         atomic_long_dec(&cinfo->mds->ncommit);
 920 }
 921 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 922
 923 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 924                                       struct inode *inode)
 925 {
 926         cinfo->inode = inode;
 927         cinfo->mds = &NFS_I(inode)->commit_info;
 928         cinfo->ds = pnfs_get_ds_info(inode);
 929         cinfo->dreq = NULL;
 930         cinfo->completion_ops = &nfs_commit_completion_ops;
 931 }
 932
 933 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 934                     struct inode *inode,
 935                     struct nfs_direct_req *dreq)
 936 {
 937         if (dreq)
 938                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 939         else
 940                 nfs_init_cinfo_from_inode(cinfo, inode);
 941 }
 942 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 943
 944 /*
 945  * Add a request to the inode's commit list.
 946  */
 947 void
 948 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 949                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 950 {
 951         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 952                 return;
 953         nfs_request_add_commit_list(req, cinfo);
 954 }
 955
 956 static void
 957 nfs_clear_page_commit(struct page *page)
 958 {
 959         dec_node_page_state(page, NR_UNSTABLE_NFS);
 960         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 961                     WB_RECLAIMABLE);
 962 }
 963
 964 /* Called holding the request lock on @req */
 965 static void
 966 nfs_clear_request_commit(struct nfs_page *req)
 967 {
 968         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 969                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 970                 struct inode *inode = d_inode(ctx->dentry);
 971                 struct nfs_commit_info cinfo;
 972
 973                 nfs_init_cinfo_from_inode(&cinfo, inode);
 974                 mutex_lock(&NFS_I(inode)->commit_mutex);
 975                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 976                         nfs_request_remove_commit_list(req, &cinfo);
 977                 }
 978                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 979                 nfs_clear_page_commit(req->wb_page);
 980         }
 981 }
 982
 983 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 984 {
 985         if (hdr->verf.committed == NFS_DATA_SYNC)
 986                 return hdr->lseg == NULL;
 987         return hdr->verf.committed != NFS_FILE_SYNC;
 988 }
 989
 990 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 991 {
 992         nfs_io_completion_get(hdr->io_completion);
 993 }
 994
 995 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 996 {
 997         struct nfs_commit_info cinfo;
 998         unsigned long bytes = 0;
 999
1000         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1001                 goto out;
1002         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1003         while (!list_empty(&hdr->pages)) {
1004                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1005
1006                 bytes += req->wb_bytes;
1007                 nfs_list_remove_request(req);
1008                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1009                     (hdr->good_bytes < bytes)) {
1010                         nfs_set_pageerror(page_file_mapping(req->wb_page));
1011                         trace_nfs_comp_error(req, hdr->error);
1012                         nfs_mapping_set_error(req->wb_page, hdr->error);
1013                         goto remove_req;
1014                 }
1015                 if (nfs_write_need_commit(hdr)) {
1016                         /* Reset wb_nio, since the write was successful. */
1017                         req->wb_nio = 0;
1018                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1019                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1020                                 hdr->pgio_mirror_idx);
1021                         goto next;
1022                 }
1023 remove_req:
1024                 nfs_inode_remove_request(req);
1025 next:
1026                 nfs_end_page_writeback(req);
1027                 nfs_release_request(req);
1028         }
1029 out:
1030         nfs_io_completion_put(hdr->io_completion);
1031         hdr->release(hdr);
1032 }
1033
1034 unsigned long
1035 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1036 {
1037         return atomic_long_read(&cinfo->mds->ncommit);
1038 }
1039
1040 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1041 int
1042 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1043                      struct nfs_commit_info *cinfo, int max)
1044 {
1045         struct nfs_page *req, *tmp;
1046         int ret = 0;
1047
1048 restart:
1049         list_for_each_entry_safe(req, tmp, src, wb_list) {
1050                 kref_get(&req->wb_kref);
1051                 if (!nfs_lock_request(req)) {
1052                         int status;
1053
1054                         /* Prevent deadlock with nfs_lock_and_join_requests */
1055                         if (!list_empty(dst)) {
1056                                 nfs_release_request(req);
1057                                 continue;
1058                         }
1059                         /* Ensure we make progress to prevent livelock */
1060                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1061                         status = nfs_wait_on_request(req);
1062                         nfs_release_request(req);
1063                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1064                         if (status < 0)
1065                                 break;
1066                         goto restart;
1067                 }
1068                 nfs_request_remove_commit_list(req, cinfo);
1069                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1070                 nfs_list_add_request(req, dst);
1071                 ret++;
1072                 if ((ret == max) && !cinfo->dreq)
1073                         break;
1074                 cond_resched();
1075         }
1076         return ret;
1077 }
1078 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1079
1080 /*
1081  * nfs_scan_commit - Scan an inode for commit requests
1082  * @inode: NFS inode to scan
1083  * @dst: mds destination list
1084  * @cinfo: mds and ds lists of reqs ready to commit
1085  *
1086  * Moves requests from the inode's 'commit' request list.
1087  * The requests are *not* checked to ensure that they form a contiguous set.
1088  */
1089 int
1090 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1091                 struct nfs_commit_info *cinfo)
1092 {
1093         int ret = 0;
1094
1095         if (!atomic_long_read(&cinfo->mds->ncommit))
1096                 return 0;
1097         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1098         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1099                 const int max = INT_MAX;
1100
1101                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1102                                            cinfo, max);
1103                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1104         }
1105         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1106         return ret;
1107 }
1108
1109 /*
1110  * Search for an existing write request, and attempt to update
1111  * it to reflect a new dirty region on a given page.
1112  *
1113  * If the attempt fails, then the existing request is flushed out
1114  * to disk.
1115  */
1116 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1117                 struct page *page,
1118                 unsigned int offset,
1119                 unsigned int bytes)
1120 {
1121         struct nfs_page *req;
1122         unsigned int rqend;
1123         unsigned int end;
1124         int error;
1125
1126         end = offset + bytes;
1127
1128         req = nfs_lock_and_join_requests(page);
1129         if (IS_ERR_OR_NULL(req))
1130                 return req;
1131
1132         rqend = req->wb_offset + req->wb_bytes;
1133         /*
1134          * Tell the caller to flush out the request if
1135          * the offsets are non-contiguous.
1136          * Note: nfs_flush_incompatible() will already
1137          * have flushed out requests having wrong owners.
1138          */
1139         if (offset > rqend || end < req->wb_offset)
1140                 goto out_flushme;
1141
1142         /* Okay, the request matches. Update the region */
1143         if (offset < req->wb_offset) {
1144                 req->wb_offset = offset;
1145                 req->wb_pgbase = offset;
1146         }
1147         if (end > rqend)
1148                 req->wb_bytes = end - req->wb_offset;
1149         else
1150                 req->wb_bytes = rqend - req->wb_offset;
1151         req->wb_nio = 0;
1152         return req;
1153 out_flushme:
1154         /*
1155          * Note: we mark the request dirty here because
1156          * nfs_lock_and_join_requests() cannot preserve
1157          * commit flags, so we have to replay the write.
1158          */
1159         nfs_mark_request_dirty(req);
1160         nfs_unlock_and_release_request(req);
1161         error = nfs_wb_page(inode, page);
1162         return (error < 0) ? ERR_PTR(error) : NULL;
1163 }
1164
1165 /*
1166  * Try to update an existing write request, or create one if there is none.
1167  *
1168  * Note: Should always be called with the Page Lock held to prevent races
1169  * if we have to add a new request. Also assumes that the caller has
1170  * already called nfs_flush_incompatible() if necessary.
1171  */
1172 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1173                 struct page *page, unsigned int offset, unsigned int bytes)
1174 {
1175         struct inode *inode = page_file_mapping(page)->host;
1176         struct nfs_page *req;
1177
1178         req = nfs_try_to_update_request(inode, page, offset, bytes);
1179         if (req != NULL)
1180                 goto out;
1181         req = nfs_create_request(ctx, page, offset, bytes);
1182         if (IS_ERR(req))
1183                 goto out;
1184         nfs_inode_add_request(inode, req);
1185 out:
1186         return req;
1187 }
1188
1189 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1190                 unsigned int offset, unsigned int count)
1191 {
1192         struct nfs_page *req;
1193
1194         req = nfs_setup_write_request(ctx, page, offset, count);
1195         if (IS_ERR(req))
1196                 return PTR_ERR(req);
1197         /* Update file length */
1198         nfs_grow_file(page, offset, count);
1199         nfs_mark_uptodate(req);
1200         nfs_mark_request_dirty(req);
1201         nfs_unlock_and_release_request(req);
1202         return 0;
1203 }
1204
1205 int nfs_flush_incompatible(struct file *file, struct page *page)
1206 {
1207         struct nfs_open_context *ctx = nfs_file_open_context(file);
1208         struct nfs_lock_context *l_ctx;
1209         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1210         struct nfs_page *req;
1211         int do_flush, status;
1212         /*
1213          * Look for a request corresponding to this page. If there
1214          * is one, and it belongs to another file, we flush it out
1215          * before we try to copy anything into the page. Do this
1216          * due to the lack of an ACCESS-type call in NFSv2.
1217          * Also do the same if we find a request from an existing
1218          * dropped page.
1219          */
1220         do {
1221                 req = nfs_page_find_head_request(page);
1222                 if (req == NULL)
1223                         return 0;
1224                 l_ctx = req->wb_lock_context;
1225                 do_flush = req->wb_page != page ||
1226                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1227                 if (l_ctx && flctx &&
1228                     !(list_empty_careful(&flctx->flc_posix) &&
1229                       list_empty_careful(&flctx->flc_flock))) {
1230                         do_flush |= l_ctx->lockowner != current->files;
1231                 }
1232                 nfs_release_request(req);
1233                 if (!do_flush)
1234                         return 0;
1235                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1236         } while (status == 0);
1237         return status;
1238 }
1239
1240 /*
1241  * Avoid buffered writes when a open context credential's key would
1242  * expire soon.
1243  *
1244  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1245  *
1246  * Return 0 and set a credential flag which triggers the inode to flush
1247  * and performs  NFS_FILE_SYNC writes if the key will expired within
1248  * RPC_KEY_EXPIRE_TIMEO.
1249  */
1250 int
1251 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1252 {
1253         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1254
1255         if (nfs_ctx_key_to_expire(ctx, inode) &&
1256             !ctx->ll_cred)
1257                 /* Already expired! */
1258                 return -EACCES;
1259         return 0;
1260 }
1261
1262 /*
1263  * Test if the open context credential key is marked to expire soon.
1264  */
1265 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1266 {
1267         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1268         struct rpc_cred *cred = ctx->ll_cred;
1269         struct auth_cred acred = {
1270                 .cred = ctx->cred,
1271         };
1272
1273         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1274                 put_rpccred(cred);
1275                 ctx->ll_cred = NULL;
1276                 cred = NULL;
1277         }
1278         if (!cred)
1279                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1280         if (!cred || IS_ERR(cred))
1281                 return true;
1282         ctx->ll_cred = cred;
1283         return !!(cred->cr_ops->crkey_timeout &&
1284                   cred->cr_ops->crkey_timeout(cred));
1285 }
1286
1287 /*
1288  * If the page cache is marked as unsafe or invalid, then we can't rely on
1289  * the PageUptodate() flag. In this case, we will need to turn off
1290  * write optimisations that depend on the page contents being correct.
1291  */
1292 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1293 {
1294         struct nfs_inode *nfsi = NFS_I(inode);
1295
1296         if (nfs_have_delegated_attributes(inode))
1297                 goto out;
1298         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1299                 return false;
1300         smp_rmb();
1301         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1302                 return false;
1303 out:
1304         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1305                 return false;
1306         return PageUptodate(page) != 0;
1307 }
1308
1309 static bool
1310 is_whole_file_wrlock(struct file_lock *fl)
1311 {
1312         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1313                         fl->fl_type == F_WRLCK;
1314 }
1315
1316 /* If we know the page is up to date, and we're not using byte range locks (or
1317  * if we have the whole file locked for writing), it may be more efficient to
1318  * extend the write to cover the entire page in order to avoid fragmentation
1319  * inefficiencies.
1320  *
1321  * If the file is opened for synchronous writes then we can just skip the rest
1322  * of the checks.
1323  */
1324 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1325 {
1326         int ret;
1327         struct file_lock_context *flctx = inode->i_flctx;
1328         struct file_lock *fl;
1329
1330         if (file->f_flags & O_DSYNC)
1331                 return 0;
1332         if (!nfs_write_pageuptodate(page, inode))
1333                 return 0;
1334         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1335                 return 1;
1336         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1337                        list_empty_careful(&flctx->flc_posix)))
1338                 return 1;
1339
1340         /* Check to see if there are whole file write locks */
1341         ret = 0;
1342         spin_lock(&flctx->flc_lock);
1343         if (!list_empty(&flctx->flc_posix)) {
1344                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1345                                         fl_list);
1346                 if (is_whole_file_wrlock(fl))
1347                         ret = 1;
1348         } else if (!list_empty(&flctx->flc_flock)) {
1349                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1350                                         fl_list);
1351                 if (fl->fl_type == F_WRLCK)
1352                         ret = 1;
1353         }
1354         spin_unlock(&flctx->flc_lock);
1355         return ret;
1356 }
1357
1358 /*
1359  * Update and possibly write a cached page of an NFS file.
1360  *
1361  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1362  * things with a page scheduled for an RPC call (e.g. invalidate it).
1363  */
1364 int nfs_updatepage(struct file *file, struct page *page,
1365                 unsigned int offset, unsigned int count)
1366 {
1367         struct nfs_open_context *ctx = nfs_file_open_context(file);
1368         struct address_space *mapping = page_file_mapping(page);
1369         struct inode    *inode = mapping->host;
1370         int             status = 0;
1371
1372         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1373
1374         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1375                 file, count, (long long)(page_file_offset(page) + offset));
1376
1377         if (!count)
1378                 goto out;
1379
1380         if (nfs_can_extend_write(file, page, inode)) {
1381                 count = max(count + offset, nfs_page_length(page));
1382                 offset = 0;
1383         }
1384
1385         status = nfs_writepage_setup(ctx, page, offset, count);
1386         if (status < 0)
1387                 nfs_set_pageerror(mapping);
1388         else
1389                 __set_page_dirty_nobuffers(page);
1390 out:
1391         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1392                         status, (long long)i_size_read(inode));
1393         return status;
1394 }
1395
1396 static int flush_task_priority(int how)
1397 {
1398         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1399                 case FLUSH_HIGHPRI:
1400                         return RPC_PRIORITY_HIGH;
1401                 case FLUSH_LOWPRI:
1402                         return RPC_PRIORITY_LOW;
1403         }
1404         return RPC_PRIORITY_NORMAL;
1405 }
1406
1407 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1408                                struct rpc_message *msg,
1409                                const struct nfs_rpc_ops *rpc_ops,
1410                                struct rpc_task_setup *task_setup_data, int how)
1411 {
1412         int priority = flush_task_priority(how);
1413
1414         task_setup_data->priority = priority;
1415         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1416         trace_nfs_initiate_write(hdr->inode, hdr->io_start, hdr->good_bytes,
1417                                  hdr->args.stable);
1418 }
1419
1420 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1421  * call this on each, which will prepare them to be retried on next
1422  * writeback using standard nfs.
1423  */
1424 static void nfs_redirty_request(struct nfs_page *req)
1425 {
1426         /* Bump the transmission count */
1427         req->wb_nio++;
1428         nfs_mark_request_dirty(req);
1429         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1430         nfs_end_page_writeback(req);
1431         nfs_release_request(req);
1432 }
1433
1434 static void nfs_async_write_error(struct list_head *head, int error)
1435 {
1436         struct nfs_page *req;
1437
1438         while (!list_empty(head)) {
1439                 req = nfs_list_entry(head->next);
1440                 nfs_list_remove_request(req);
1441                 if (nfs_error_is_fatal(error))
1442                         nfs_write_error(req, error);
1443                 else
1444                         nfs_redirty_request(req);
1445         }
1446 }
1447
1448 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1449 {
1450         nfs_async_write_error(&hdr->pages, 0);
1451         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1452                         hdr->args.offset + hdr->args.count - 1);
1453 }
1454
1455 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1456         .init_hdr = nfs_async_write_init,
1457         .error_cleanup = nfs_async_write_error,
1458         .completion = nfs_write_completion,
1459         .reschedule_io = nfs_async_write_reschedule_io,
1460 };
1461
1462 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1463                                struct inode *inode, int ioflags, bool force_mds,
1464                                const struct nfs_pgio_completion_ops *compl_ops)
1465 {
1466         struct nfs_server *server = NFS_SERVER(inode);
1467         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1468
1469 #ifdef CONFIG_NFS_V4_1
1470         if (server->pnfs_curr_ld && !force_mds)
1471                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1472 #endif
1473         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1474                         server->wsize, ioflags);
1475 }
1476 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1477
1478 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1479 {
1480         struct nfs_pgio_mirror *mirror;
1481
1482         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1483                 pgio->pg_ops->pg_cleanup(pgio);
1484
1485         pgio->pg_ops = &nfs_pgio_rw_ops;
1486
1487         nfs_pageio_stop_mirroring(pgio);
1488
1489         mirror = &pgio->pg_mirrors[0];
1490         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1491 }
1492 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1493
1494
1495 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1496 {
1497         struct nfs_commit_data *data = calldata;
1498
1499         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1500 }
1501
1502 /*
1503  * Special version of should_remove_suid() that ignores capabilities.
1504  */
1505 static int nfs_should_remove_suid(const struct inode *inode)
1506 {
1507         umode_t mode = inode->i_mode;
1508         int kill = 0;
1509
1510         /* suid always must be killed */
1511         if (unlikely(mode & S_ISUID))
1512                 kill = ATTR_KILL_SUID;
1513
1514         /*
1515          * sgid without any exec bits is just a mandatory locking mark; leave
1516          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1517          */
1518         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1519                 kill |= ATTR_KILL_SGID;
1520
1521         if (unlikely(kill && S_ISREG(mode)))
1522                 return kill;
1523
1524         return 0;
1525 }
1526
1527 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1528                 struct nfs_fattr *fattr)
1529 {
1530         struct nfs_pgio_args *argp = &hdr->args;
1531         struct nfs_pgio_res *resp = &hdr->res;
1532         u64 size = argp->offset + resp->count;
1533
1534         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1535                 fattr->size = size;
1536         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1537                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1538                 return;
1539         }
1540         if (size != fattr->size)
1541                 return;
1542         /* Set attribute barrier */
1543         nfs_fattr_set_barrier(fattr);
1544         /* ...and update size */
1545         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1546 }
1547
1548 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1549 {
1550         struct nfs_fattr *fattr = &hdr->fattr;
1551         struct inode *inode = hdr->inode;
1552
1553         spin_lock(&inode->i_lock);
1554         nfs_writeback_check_extend(hdr, fattr);
1555         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1556         spin_unlock(&inode->i_lock);
1557 }
1558 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1559
1560 /*
1561  * This function is called when the WRITE call is complete.
1562  */
1563 static int nfs_writeback_done(struct rpc_task *task,
1564                               struct nfs_pgio_header *hdr,
1565                               struct inode *inode)
1566 {
1567         int status;
1568
1569         /*
1570          * ->write_done will attempt to use post-op attributes to detect
1571          * conflicting writes by other clients.  A strict interpretation
1572          * of close-to-open would allow us to continue caching even if
1573          * another writer had changed the file, but some applications
1574          * depend on tighter cache coherency when writing.
1575          */
1576         status = NFS_PROTO(inode)->write_done(task, hdr);
1577         if (status != 0)
1578                 return status;
1579
1580         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1581         trace_nfs_writeback_done(inode, task->tk_status,
1582                                  hdr->args.offset, hdr->res.verf);
1583
1584         if (hdr->res.verf->committed < hdr->args.stable &&
1585             task->tk_status >= 0) {
1586                 /* We tried a write call, but the server did not
1587                  * commit data to stable storage even though we
1588                  * requested it.
1589                  * Note: There is a known bug in Tru64 < 5.0 in which
1590                  *       the server reports NFS_DATA_SYNC, but performs
1591                  *       NFS_FILE_SYNC. We therefore implement this checking
1592                  *       as a dprintk() in order to avoid filling syslog.
1593                  */
1594                 static unsigned long    complain;
1595
1596                 /* Note this will print the MDS for a DS write */
1597                 if (time_before(complain, jiffies)) {
1598                         dprintk("NFS:       faulty NFS server %s:"
1599                                 " (committed = %d) != (stable = %d)\n",
1600                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1601                                 hdr->res.verf->committed, hdr->args.stable);
1602                         complain = jiffies + 300 * HZ;
1603                 }
1604         }
1605
1606         /* Deal with the suid/sgid bit corner case */
1607         if (nfs_should_remove_suid(inode)) {
1608                 spin_lock(&inode->i_lock);
1609                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1610                 spin_unlock(&inode->i_lock);
1611         }
1612         return 0;
1613 }
1614
1615 /*
1616  * This function is called when the WRITE call is complete.
1617  */
1618 static void nfs_writeback_result(struct rpc_task *task,
1619                                  struct nfs_pgio_header *hdr)
1620 {
1621         struct nfs_pgio_args    *argp = &hdr->args;
1622         struct nfs_pgio_res     *resp = &hdr->res;
1623
1624         if (resp->count < argp->count) {
1625                 static unsigned long    complain;
1626
1627                 /* This a short write! */
1628                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1629
1630                 /* Has the server at least made some progress? */
1631                 if (resp->count == 0) {
1632                         if (time_before(complain, jiffies)) {
1633                                 printk(KERN_WARNING
1634                                        "NFS: Server wrote zero bytes, expected %u.\n",
1635                                        argp->count);
1636                                 complain = jiffies + 300 * HZ;
1637                         }
1638                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1639                         task->tk_status = -EIO;
1640                         return;
1641                 }
1642
1643                 /* For non rpc-based layout drivers, retry-through-MDS */
1644                 if (!task->tk_ops) {
1645                         hdr->pnfs_error = -EAGAIN;
1646                         return;
1647                 }
1648
1649                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1650                 if (resp->verf->committed != NFS_UNSTABLE) {
1651                         /* Resend from where the server left off */
1652                         hdr->mds_offset += resp->count;
1653                         argp->offset += resp->count;
1654                         argp->pgbase += resp->count;
1655                         argp->count -= resp->count;
1656                 } else {
1657                         /* Resend as a stable write in order to avoid
1658                          * headaches in the case of a server crash.
1659                          */
1660                         argp->stable = NFS_FILE_SYNC;
1661                 }
1662                 rpc_restart_call_prepare(task);
1663         }
1664 }
1665
1666 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1667 {
1668         return wait_var_event_killable(&cinfo->rpcs_out,
1669                                        !atomic_read(&cinfo->rpcs_out));
1670 }
1671
1672 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1673 {
1674         atomic_inc(&cinfo->rpcs_out);
1675 }
1676
1677 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1678 {
1679         if (atomic_dec_and_test(&cinfo->rpcs_out))
1680                 wake_up_var(&cinfo->rpcs_out);
1681 }
1682
1683 void nfs_commitdata_release(struct nfs_commit_data *data)
1684 {
1685         put_nfs_open_context(data->context);
1686         nfs_commit_free(data);
1687 }
1688 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1689
1690 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1691                         const struct nfs_rpc_ops *nfs_ops,
1692                         const struct rpc_call_ops *call_ops,
1693                         int how, int flags)
1694 {
1695         struct rpc_task *task;
1696         int priority = flush_task_priority(how);
1697         struct rpc_message msg = {
1698                 .rpc_argp = &data->args,
1699                 .rpc_resp = &data->res,
1700                 .rpc_cred = data->cred,
1701         };
1702         struct rpc_task_setup task_setup_data = {
1703                 .task = &data->task,
1704                 .rpc_client = clnt,
1705                 .rpc_message = &msg,
1706                 .callback_ops = call_ops,
1707                 .callback_data = data,
1708                 .workqueue = nfsiod_workqueue,
1709                 .flags = RPC_TASK_ASYNC | flags,
1710                 .priority = priority,
1711         };
1712         /* Set up the initial task struct.  */
1713         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1714         trace_nfs_initiate_commit(data);
1715
1716         dprintk("NFS: initiated commit call\n");
1717
1718         task = rpc_run_task(&task_setup_data);
1719         if (IS_ERR(task))
1720                 return PTR_ERR(task);
1721         if (how & FLUSH_SYNC)
1722                 rpc_wait_for_completion_task(task);
1723         rpc_put_task(task);
1724         return 0;
1725 }
1726 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1727
1728 static loff_t nfs_get_lwb(struct list_head *head)
1729 {
1730         loff_t lwb = 0;
1731         struct nfs_page *req;
1732
1733         list_for_each_entry(req, head, wb_list)
1734                 if (lwb < (req_offset(req) + req->wb_bytes))
1735                         lwb = req_offset(req) + req->wb_bytes;
1736
1737         return lwb;
1738 }
1739
1740 /*
1741  * Set up the argument/result storage required for the RPC call.
1742  */
1743 void nfs_init_commit(struct nfs_commit_data *data,
1744                      struct list_head *head,
1745                      struct pnfs_layout_segment *lseg,
1746                      struct nfs_commit_info *cinfo)
1747 {
1748         struct nfs_page *first = nfs_list_entry(head->next);
1749         struct nfs_open_context *ctx = nfs_req_openctx(first);
1750         struct inode *inode = d_inode(ctx->dentry);
1751
1752         /* Set up the RPC argument and reply structs
1753          * NB: take care not to mess about with data->commit et al. */
1754
1755         list_splice_init(head, &data->pages);
1756
1757         data->inode       = inode;
1758         data->cred        = ctx->cred;
1759         data->lseg        = lseg; /* reference transferred */
1760         /* only set lwb for pnfs commit */
1761         if (lseg)
1762                 data->lwb = nfs_get_lwb(&data->pages);
1763         data->mds_ops     = &nfs_commit_ops;
1764         data->completion_ops = cinfo->completion_ops;
1765         data->dreq        = cinfo->dreq;
1766
1767         data->args.fh     = NFS_FH(data->inode);
1768         /* Note: we always request a commit of the entire inode */
1769         data->args.offset = 0;
1770         data->args.count  = 0;
1771         data->context     = get_nfs_open_context(ctx);
1772         data->res.fattr   = &data->fattr;
1773         data->res.verf    = &data->verf;
1774         nfs_fattr_init(&data->fattr);
1775 }
1776 EXPORT_SYMBOL_GPL(nfs_init_commit);
1777
1778 void nfs_retry_commit(struct list_head *page_list,
1779                       struct pnfs_layout_segment *lseg,
1780                       struct nfs_commit_info *cinfo,
1781                       u32 ds_commit_idx)
1782 {
1783         struct nfs_page *req;
1784
1785         while (!list_empty(page_list)) {
1786                 req = nfs_list_entry(page_list->next);
1787                 nfs_list_remove_request(req);
1788                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1789                 if (!cinfo->dreq)
1790                         nfs_clear_page_commit(req->wb_page);
1791                 nfs_unlock_and_release_request(req);
1792         }
1793 }
1794 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1795
1796 static void
1797 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1798                 struct nfs_page *req)
1799 {
1800         __set_page_dirty_nobuffers(req->wb_page);
1801 }
1802
1803 /*
1804  * Commit dirty pages
1805  */
1806 static int
1807 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1808                 struct nfs_commit_info *cinfo)
1809 {
1810         struct nfs_commit_data  *data;
1811
1812         /* another commit raced with us */
1813         if (list_empty(head))
1814                 return 0;
1815
1816         data = nfs_commitdata_alloc(true);
1817
1818         /* Set up the argument struct */
1819         nfs_init_commit(data, head, NULL, cinfo);
1820         atomic_inc(&cinfo->mds->rpcs_out);
1821         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1822                                    data->mds_ops, how, 0);
1823 }
1824
1825 /*
1826  * COMMIT call returned
1827  */
1828 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1829 {
1830         struct nfs_commit_data  *data = calldata;
1831
1832         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1833                                 task->tk_pid, task->tk_status);
1834
1835         /* Call the NFS version-specific code */
1836         NFS_PROTO(data->inode)->commit_done(task, data);
1837         trace_nfs_commit_done(data);
1838 }
1839
1840 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1841 {
1842         struct nfs_page *req;
1843         int status = data->task.tk_status;
1844         struct nfs_commit_info cinfo;
1845         struct nfs_server *nfss;
1846
1847         while (!list_empty(&data->pages)) {
1848                 req = nfs_list_entry(data->pages.next);
1849                 nfs_list_remove_request(req);
1850                 if (req->wb_page)
1851                         nfs_clear_page_commit(req->wb_page);
1852
1853                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1854                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1855                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1856                         req->wb_bytes,
1857                         (long long)req_offset(req));
1858                 if (status < 0) {
1859                         if (req->wb_page) {
1860                                 trace_nfs_commit_error(req, status);
1861                                 nfs_mapping_set_error(req->wb_page, status);
1862                                 nfs_inode_remove_request(req);
1863                         }
1864                         dprintk_cont(", error = %d\n", status);
1865                         goto next;
1866                 }
1867
1868                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1869                  * returned by the server against all stored verfs. */
1870                 if (!nfs_write_verifier_cmp(&req->wb_verf, &data->verf.verifier)) {
1871                         /* We have a match */
1872                         if (req->wb_page)
1873                                 nfs_inode_remove_request(req);
1874                         dprintk_cont(" OK\n");
1875                         goto next;
1876                 }
1877                 /* We have a mismatch. Write the page again */
1878                 dprintk_cont(" mismatch\n");
1879                 nfs_mark_request_dirty(req);
1880                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1881         next:
1882                 nfs_unlock_and_release_request(req);
1883                 /* Latency breaker */
1884                 cond_resched();
1885         }
1886         nfss = NFS_SERVER(data->inode);
1887         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1888                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1889
1890         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1891         nfs_commit_end(cinfo.mds);
1892 }
1893
1894 static void nfs_commit_release(void *calldata)
1895 {
1896         struct nfs_commit_data *data = calldata;
1897
1898         data->completion_ops->completion(data);
1899         nfs_commitdata_release(calldata);
1900 }
1901
1902 static const struct rpc_call_ops nfs_commit_ops = {
1903         .rpc_call_prepare = nfs_commit_prepare,
1904         .rpc_call_done = nfs_commit_done,
1905         .rpc_release = nfs_commit_release,
1906 };
1907
1908 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1909         .completion = nfs_commit_release_pages,
1910         .resched_write = nfs_commit_resched_write,
1911 };
1912
1913 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1914                             int how, struct nfs_commit_info *cinfo)
1915 {
1916         int status;
1917
1918         status = pnfs_commit_list(inode, head, how, cinfo);
1919         if (status == PNFS_NOT_ATTEMPTED)
1920                 status = nfs_commit_list(inode, head, how, cinfo);
1921         return status;
1922 }
1923
1924 static int __nfs_commit_inode(struct inode *inode, int how,
1925                 struct writeback_control *wbc)
1926 {
1927         LIST_HEAD(head);
1928         struct nfs_commit_info cinfo;
1929         int may_wait = how & FLUSH_SYNC;
1930         int ret, nscan;
1931
1932         nfs_init_cinfo_from_inode(&cinfo, inode);
1933         nfs_commit_begin(cinfo.mds);
1934         for (;;) {
1935                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1936                 if (ret <= 0)
1937                         break;
1938                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1939                 if (ret < 0)
1940                         break;
1941                 ret = 0;
1942                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1943                         if (nscan < wbc->nr_to_write)
1944                                 wbc->nr_to_write -= nscan;
1945                         else
1946                                 wbc->nr_to_write = 0;
1947                 }
1948                 if (nscan < INT_MAX)
1949                         break;
1950                 cond_resched();
1951         }
1952         nfs_commit_end(cinfo.mds);
1953         if (ret || !may_wait)
1954                 return ret;
1955         return wait_on_commit(cinfo.mds);
1956 }
1957
1958 int nfs_commit_inode(struct inode *inode, int how)
1959 {
1960         return __nfs_commit_inode(inode, how, NULL);
1961 }
1962 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1963
1964 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1965 {
1966         struct nfs_inode *nfsi = NFS_I(inode);
1967         int flags = FLUSH_SYNC;
1968         int ret = 0;
1969
1970         if (wbc->sync_mode == WB_SYNC_NONE) {
1971                 /* no commits means nothing needs to be done */
1972                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1973                         goto check_requests_outstanding;
1974
1975                 /* Don't commit yet if this is a non-blocking flush and there
1976                  * are a lot of outstanding writes for this mapping.
1977                  */
1978                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1979                         goto out_mark_dirty;
1980
1981                 /* don't wait for the COMMIT response */
1982                 flags = 0;
1983         }
1984
1985         ret = __nfs_commit_inode(inode, flags, wbc);
1986         if (!ret) {
1987                 if (flags & FLUSH_SYNC)
1988                         return 0;
1989         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1990                 goto out_mark_dirty;
1991
1992 check_requests_outstanding:
1993         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1994                 return ret;
1995 out_mark_dirty:
1996         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1997         return ret;
1998 }
1999 EXPORT_SYMBOL_GPL(nfs_write_inode);
2000
2001 /*
2002  * Wrapper for filemap_write_and_wait_range()
2003  *
2004  * Needed for pNFS in order to ensure data becomes visible to the
2005  * client.
2006  */
2007 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2008                 loff_t lstart, loff_t lend)
2009 {
2010         int ret;
2011
2012         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2013         if (ret == 0)
2014                 ret = pnfs_sync_inode(mapping->host, true);
2015         return ret;
2016 }
2017 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2018
2019 /*
2020  * flush the inode to disk.
2021  */
2022 int nfs_wb_all(struct inode *inode)
2023 {
2024         int ret;
2025
2026         trace_nfs_writeback_inode_enter(inode);
2027
2028         ret = filemap_write_and_wait(inode->i_mapping);
2029         if (ret)
2030                 goto out;
2031         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2032         if (ret < 0)
2033                 goto out;
2034         pnfs_sync_inode(inode, true);
2035         ret = 0;
2036
2037 out:
2038         trace_nfs_writeback_inode_exit(inode, ret);
2039         return ret;
2040 }
2041 EXPORT_SYMBOL_GPL(nfs_wb_all);
2042
2043 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2044 {
2045         struct nfs_page *req;
2046         int ret = 0;
2047
2048         wait_on_page_writeback(page);
2049
2050         /* blocking call to cancel all requests and join to a single (head)
2051          * request */
2052         req = nfs_lock_and_join_requests(page);
2053
2054         if (IS_ERR(req)) {
2055                 ret = PTR_ERR(req);
2056         } else if (req) {
2057                 /* all requests from this page have been cancelled by
2058                  * nfs_lock_and_join_requests, so just remove the head
2059                  * request from the inode / page_private pointer and
2060                  * release it */
2061                 nfs_inode_remove_request(req);
2062                 nfs_unlock_and_release_request(req);
2063         }
2064
2065         return ret;
2066 }
2067
2068 /*
2069  * Write back all requests on one page - we do this before reading it.
2070  */
2071 int nfs_wb_page(struct inode *inode, struct page *page)
2072 {
2073         loff_t range_start = page_file_offset(page);
2074         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2075         struct writeback_control wbc = {
2076                 .sync_mode = WB_SYNC_ALL,
2077                 .nr_to_write = 0,
2078                 .range_start = range_start,
2079                 .range_end = range_end,
2080         };
2081         int ret;
2082
2083         trace_nfs_writeback_page_enter(inode);
2084
2085         for (;;) {
2086                 wait_on_page_writeback(page);
2087                 if (clear_page_dirty_for_io(page)) {
2088                         ret = nfs_writepage_locked(page, &wbc);
2089                         if (ret < 0)
2090                                 goto out_error;
2091                         continue;
2092                 }
2093                 ret = 0;
2094                 if (!PagePrivate(page))
2095                         break;
2096                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2097                 if (ret < 0)
2098                         goto out_error;
2099         }
2100 out_error:
2101         trace_nfs_writeback_page_exit(inode, ret);
2102         return ret;
2103 }
2104
2105 #ifdef CONFIG_MIGRATION
2106 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2107                 struct page *page, enum migrate_mode mode)
2108 {
2109         /*
2110          * If PagePrivate is set, then the page is currently associated with
2111          * an in-progress read or write request. Don't try to migrate it.
2112          *
2113          * FIXME: we could do this in principle, but we'll need a way to ensure
2114          *        that we can safely release the inode reference while holding
2115          *        the page lock.
2116          */
2117         if (PagePrivate(page))
2118                 return -EBUSY;
2119
2120         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2121                 return -EBUSY;
2122
2123         return migrate_page(mapping, newpage, page, mode);
2124 }
2125 #endif
2126
2127 int __init nfs_init_writepagecache(void)
2128 {
2129         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2130                                              sizeof(struct nfs_pgio_header),
2131                                              0, SLAB_HWCACHE_ALIGN,
2132                                              NULL);
2133         if (nfs_wdata_cachep == NULL)
2134                 return -ENOMEM;
2135
2136         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2137                                                      nfs_wdata_cachep);
2138         if (nfs_wdata_mempool == NULL)
2139                 goto out_destroy_write_cache;
2140
2141         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2142                                              sizeof(struct nfs_commit_data),
2143                                              0, SLAB_HWCACHE_ALIGN,
2144                                              NULL);
2145         if (nfs_cdata_cachep == NULL)
2146                 goto out_destroy_write_mempool;
2147
2148         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2149                                                       nfs_cdata_cachep);
2150         if (nfs_commit_mempool == NULL)
2151                 goto out_destroy_commit_cache;
2152
2153         /*
2154          * NFS congestion size, scale with available memory.
2155          *
2156          *  64MB:    8192k
2157          * 128MB:   11585k
2158          * 256MB:   16384k
2159          * 512MB:   23170k
2160          *   1GB:   32768k
2161          *   2GB:   46340k
2162          *   4GB:   65536k
2163          *   8GB:   92681k
2164          *  16GB:  131072k
2165          *
2166          * This allows larger machines to have larger/more transfers.
2167          * Limit the default to 256M
2168          */
2169         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2170         if (nfs_congestion_kb > 256*1024)
2171                 nfs_congestion_kb = 256*1024;
2172
2173         return 0;
2174
2175 out_destroy_commit_cache:
2176         kmem_cache_destroy(nfs_cdata_cachep);
2177 out_destroy_write_mempool:
2178         mempool_destroy(nfs_wdata_mempool);
2179 out_destroy_write_cache:
2180         kmem_cache_destroy(nfs_wdata_cachep);
2181         return -ENOMEM;
2182 }
2183
2184 void nfs_destroy_writepagecache(void)
2185 {
2186         mempool_destroy(nfs_commit_mempool);
2187         kmem_cache_destroy(nfs_cdata_cachep);
2188         mempool_destroy(nfs_wdata_mempool);
2189         kmem_cache_destroy(nfs_wdata_cachep);
2190 }
2191
2192 static const struct nfs_rw_ops nfs_rw_write_ops = {
2193         .rw_alloc_header        = nfs_writehdr_alloc,
2194         .rw_free_header         = nfs_writehdr_free,
2195         .rw_done                = nfs_writeback_done,
2196         .rw_result              = nfs_writeback_result,
2197         .rw_initiate            = nfs_initiate_write,
2198 };