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         struct address_space *mapping = page_file_mapping(page);
 260
 261         SetPageError(page);
 262         mapping_set_error(mapping, error);
 263         nfs_set_pageerror(mapping);
 264 }
 265
 266 /*
 267  * nfs_page_group_search_locked
 268  * @head - head request of page group
 269  * @page_offset - offset into page
 270  *
 271  * Search page group with head @head to find a request that contains the
 272  * page offset @page_offset.
 273  *
 274  * Returns a pointer to the first matching nfs request, or NULL if no
 275  * match is found.
 276  *
 277  * Must be called with the page group lock held
 278  */
 279 static struct nfs_page *
 280 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 281 {
 282         struct nfs_page *req;
 283
 284         req = head;
 285         do {
 286                 if (page_offset >= req->wb_pgbase &&
 287                     page_offset < (req->wb_pgbase + req->wb_bytes))
 288                         return req;
 289
 290                 req = req->wb_this_page;
 291         } while (req != head);
 292
 293         return NULL;
 294 }
 295
 296 /*
 297  * nfs_page_group_covers_page
 298  * @head - head request of page group
 299  *
 300  * Return true if the page group with head @head covers the whole page,
 301  * returns false otherwise
 302  */
 303 static bool nfs_page_group_covers_page(struct nfs_page *req)
 304 {
 305         struct nfs_page *tmp;
 306         unsigned int pos = 0;
 307         unsigned int len = nfs_page_length(req->wb_page);
 308
 309         nfs_page_group_lock(req);
 310
 311         for (;;) {
 312                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 313                 if (!tmp)
 314                         break;
 315                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 316         }
 317
 318         nfs_page_group_unlock(req);
 319         return pos >= len;
 320 }
 321
 322 /* We can set the PG_uptodate flag if we see that a write request
 323  * covers the full page.
 324  */
 325 static void nfs_mark_uptodate(struct nfs_page *req)
 326 {
 327         if (PageUptodate(req->wb_page))
 328                 return;
 329         if (!nfs_page_group_covers_page(req))
 330                 return;
 331         SetPageUptodate(req->wb_page);
 332 }
 333
 334 static int wb_priority(struct writeback_control *wbc)
 335 {
 336         int ret = 0;
 337
 338         if (wbc->sync_mode == WB_SYNC_ALL)
 339                 ret = FLUSH_COND_STABLE;
 340         return ret;
 341 }
 342
 343 /*
 344  * NFS congestion control
 345  */
 346
 347 int nfs_congestion_kb;
 348
 349 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 350 #define NFS_CONGESTION_OFF_THRESH       \
 351         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 352
 353 static void nfs_set_page_writeback(struct page *page)
 354 {
 355         struct inode *inode = page_file_mapping(page)->host;
 356         struct nfs_server *nfss = NFS_SERVER(inode);
 357         int ret = test_set_page_writeback(page);
 358
 359         WARN_ON_ONCE(ret != 0);
 360
 361         if (atomic_long_inc_return(&nfss->writeback) >
 362                         NFS_CONGESTION_ON_THRESH)
 363                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 364 }
 365
 366 static void nfs_end_page_writeback(struct nfs_page *req)
 367 {
 368         struct inode *inode = page_file_mapping(req->wb_page)->host;
 369         struct nfs_server *nfss = NFS_SERVER(inode);
 370         bool is_done;
 371
 372         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 373         nfs_unlock_request(req);
 374         if (!is_done)
 375                 return;
 376
 377         end_page_writeback(req->wb_page);
 378         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 379                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 380 }
 381
 382 /*
 383  * nfs_unroll_locks_and_wait -  unlock all newly locked reqs and wait on @req
 384  *
 385  * this is a helper function for nfs_lock_and_join_requests
 386  *
 387  * @inode - inode associated with request page group, must be holding inode lock
 388  * @head  - head request of page group, must be holding head lock
 389  * @req   - request that couldn't lock and needs to wait on the req bit lock
 390  *
 391  * NOTE: this must be called holding page_group bit lock
 392  *       which will be released before returning.
 393  *
 394  * returns 0 on success, < 0 on error.
 395  */
 396 static void
 397 nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
 398                           struct nfs_page *req)
 399 {
 400         struct nfs_page *tmp;
 401
 402         /* relinquish all the locks successfully grabbed this run */
 403         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 404                 if (!kref_read(&tmp->wb_kref))
 405                         continue;
 406                 nfs_unlock_and_release_request(tmp);
 407         }
 408 }
 409
 410 /*
 411  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 412  *
 413  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 414  * @old_head - the old head of the list
 415  *
 416  * All subrequests must be locked and removed from all lists, so at this point
 417  * they are only "active" in this function, and possibly in nfs_wait_on_request
 418  * with a reference held by some other context.
 419  */
 420 static void
 421 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 422                                  struct nfs_page *old_head,
 423                                  struct inode *inode)
 424 {
 425         while (destroy_list) {
 426                 struct nfs_page *subreq = destroy_list;
 427
 428                 destroy_list = (subreq->wb_this_page == old_head) ?
 429                                    NULL : subreq->wb_this_page;
 430
 431                 WARN_ON_ONCE(old_head != subreq->wb_head);
 432
 433                 /* make sure old group is not used */
 434                 subreq->wb_this_page = subreq;
 435
 436                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 437
 438                 /* Note: races with nfs_page_group_destroy() */
 439                 if (!kref_read(&subreq->wb_kref)) {
 440                         /* Check if we raced with nfs_page_group_destroy() */
 441                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags))
 442                                 nfs_free_request(subreq);
 443                         continue;
 444                 }
 445
 446                 subreq->wb_head = subreq;
 447
 448                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 449                         nfs_release_request(subreq);
 450                         atomic_long_dec(&NFS_I(inode)->nrequests);
 451                 }
 452
 453                 /* subreq is now totally disconnected from page group or any
 454                  * write / commit lists. last chance to wake any waiters */
 455                 nfs_unlock_and_release_request(subreq);
 456         }
 457 }
 458
 459 /*
 460  * nfs_lock_and_join_requests - join all subreqs to the head req and return
 461  *                              a locked reference, cancelling any pending
 462  *                              operations for this page.
 463  *
 464  * @page - the page used to lookup the "page group" of nfs_page structures
 465  *
 466  * This function joins all sub requests to the head request by first
 467  * locking all requests in the group, cancelling any pending operations
 468  * and finally updating the head request to cover the whole range covered by
 469  * the (former) group.  All subrequests are removed from any write or commit
 470  * lists, unlinked from the group and destroyed.
 471  *
 472  * Returns a locked, referenced pointer to the head request - which after
 473  * this call is guaranteed to be the only request associated with the page.
 474  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 475  * error was encountered.
 476  */
 477 static struct nfs_page *
 478 nfs_lock_and_join_requests(struct page *page)
 479 {
 480         struct inode *inode = page_file_mapping(page)->host;
 481         struct nfs_page *head, *subreq;
 482         struct nfs_page *destroy_list = NULL;
 483         unsigned int total_bytes;
 484         int ret;
 485
 486 try_again:
 487         /*
 488          * A reference is taken only on the head request which acts as a
 489          * reference to the whole page group - the group will not be destroyed
 490          * until the head reference is released.
 491          */
 492         head = nfs_page_find_head_request(page);
 493         if (!head)
 494                 return NULL;
 495
 496         /* lock the page head first in order to avoid an ABBA inefficiency */
 497         if (!nfs_lock_request(head)) {
 498                 ret = nfs_wait_on_request(head);
 499                 nfs_release_request(head);
 500                 if (ret < 0)
 501                         return ERR_PTR(ret);
 502                 goto try_again;
 503         }
 504
 505         /* Ensure that nobody removed the request before we locked it */
 506         if (head != nfs_page_private_request(page) && !PageSwapCache(page)) {
 507                 nfs_unlock_and_release_request(head);
 508                 goto try_again;
 509         }
 510
 511         ret = nfs_page_group_lock(head);
 512         if (ret < 0)
 513                 goto release_request;
 514
 515         /* lock each request in the page group */
 516         total_bytes = head->wb_bytes;
 517         for (subreq = head->wb_this_page; subreq != head;
 518                         subreq = subreq->wb_this_page) {
 519
 520                 if (!kref_get_unless_zero(&subreq->wb_kref)) {
 521                         if (subreq->wb_offset == head->wb_offset + total_bytes)
 522                                 total_bytes += subreq->wb_bytes;
 523                         continue;
 524                 }
 525
 526                 while (!nfs_lock_request(subreq)) {
 527                         /*
 528                          * Unlock page to allow nfs_page_group_sync_on_bit()
 529                          * to succeed
 530                          */
 531                         nfs_page_group_unlock(head);
 532                         ret = nfs_wait_on_request(subreq);
 533                         if (!ret)
 534                                 ret = nfs_page_group_lock(head);
 535                         if (ret < 0) {
 536                                 nfs_unroll_locks(inode, head, subreq);
 537                                 nfs_release_request(subreq);
 538                                 goto release_request;
 539                         }
 540                 }
 541                 /*
 542                  * Subrequests are always contiguous, non overlapping
 543                  * and in order - but may be repeated (mirrored writes).
 544                  */
 545                 if (subreq->wb_offset == (head->wb_offset + total_bytes)) {
 546                         /* keep track of how many bytes this group covers */
 547                         total_bytes += subreq->wb_bytes;
 548                 } else if (WARN_ON_ONCE(subreq->wb_offset < head->wb_offset ||
 549                             ((subreq->wb_offset + subreq->wb_bytes) >
 550                              (head->wb_offset + total_bytes)))) {
 551                         nfs_page_group_unlock(head);
 552                         nfs_unroll_locks(inode, head, subreq);
 553                         nfs_unlock_and_release_request(subreq);
 554                         ret = -EIO;
 555                         goto release_request;
 556                 }
 557         }
 558
 559         /* Now that all requests are locked, make sure they aren't on any list.
 560          * Commit list removal accounting is done after locks are dropped */
 561         subreq = head;
 562         do {
 563                 nfs_clear_request_commit(subreq);
 564                 subreq = subreq->wb_this_page;
 565         } while (subreq != head);
 566
 567         /* unlink subrequests from head, destroy them later */
 568         if (head->wb_this_page != head) {
 569                 /* destroy list will be terminated by head */
 570                 destroy_list = head->wb_this_page;
 571                 head->wb_this_page = head;
 572
 573                 /* change head request to cover whole range that
 574                  * the former page group covered */
 575                 head->wb_bytes = total_bytes;
 576         }
 577
 578         /* Postpone destruction of this request */
 579         if (test_and_clear_bit(PG_REMOVE, &head->wb_flags)) {
 580                 set_bit(PG_INODE_REF, &head->wb_flags);
 581                 kref_get(&head->wb_kref);
 582                 atomic_long_inc(&NFS_I(inode)->nrequests);
 583         }
 584
 585         nfs_page_group_unlock(head);
 586
 587         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 588
 589         /* Did we lose a race with nfs_inode_remove_request()? */
 590         if (!(PagePrivate(page) || PageSwapCache(page))) {
 591                 nfs_unlock_and_release_request(head);
 592                 return NULL;
 593         }
 594
 595         /* still holds ref on head from nfs_page_find_head_request
 596          * and still has lock on head from lock loop */
 597         return head;
 598
 599 release_request:
 600         nfs_unlock_and_release_request(head);
 601         return ERR_PTR(ret);
 602 }
 603
 604 static void nfs_write_error(struct nfs_page *req, int error)
 605 {
 606         trace_nfs_write_error(req, error);
 607         nfs_mapping_set_error(req->wb_page, error);
 608         nfs_inode_remove_request(req);
 609         nfs_end_page_writeback(req);
 610         nfs_release_request(req);
 611 }
 612
 613 /*
 614  * Find an associated nfs write request, and prepare to flush it out
 615  * May return an error if the user signalled nfs_wait_on_request().
 616  */
 617 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 618                                 struct page *page)
 619 {
 620         struct nfs_page *req;
 621         int ret = 0;
 622
 623         req = nfs_lock_and_join_requests(page);
 624         if (!req)
 625                 goto out;
 626         ret = PTR_ERR(req);
 627         if (IS_ERR(req))
 628                 goto out;
 629
 630         nfs_set_page_writeback(page);
 631         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 632
 633         /* If there is a fatal error that covers this write, just exit */
 634         ret = pgio->pg_error;
 635         if (nfs_error_is_fatal_on_server(ret))
 636                 goto out_launder;
 637
 638         ret = 0;
 639         if (!nfs_pageio_add_request(pgio, req)) {
 640                 ret = pgio->pg_error;
 641                 /*
 642                  * Remove the problematic req upon fatal errors on the server
 643                  */
 644                 if (nfs_error_is_fatal(ret)) {
 645                         if (nfs_error_is_fatal_on_server(ret))
 646                                 goto out_launder;
 647                 } else
 648                         ret = -EAGAIN;
 649                 nfs_redirty_request(req);
 650                 pgio->pg_error = 0;
 651         } else
 652                 nfs_add_stats(page_file_mapping(page)->host,
 653                                 NFSIOS_WRITEPAGES, 1);
 654 out:
 655         return ret;
 656 out_launder:
 657         nfs_write_error(req, ret);
 658         return 0;
 659 }
 660
 661 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 662                             struct nfs_pageio_descriptor *pgio)
 663 {
 664         int ret;
 665
 666         nfs_pageio_cond_complete(pgio, page_index(page));
 667         ret = nfs_page_async_flush(pgio, page);
 668         if (ret == -EAGAIN) {
 669                 redirty_page_for_writepage(wbc, page);
 670                 ret = AOP_WRITEPAGE_ACTIVATE;
 671         }
 672         return ret;
 673 }
 674
 675 /*
 676  * Write an mmapped page to the server.
 677  */
 678 static int nfs_writepage_locked(struct page *page,
 679                                 struct writeback_control *wbc)
 680 {
 681         struct nfs_pageio_descriptor pgio;
 682         struct inode *inode = page_file_mapping(page)->host;
 683         int err;
 684
 685         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 686         nfs_pageio_init_write(&pgio, inode, 0,
 687                                 false, &nfs_async_write_completion_ops);
 688         err = nfs_do_writepage(page, wbc, &pgio);
 689         pgio.pg_error = 0;
 690         nfs_pageio_complete(&pgio);
 691         if (err < 0)
 692                 return err;
 693         if (nfs_error_is_fatal(pgio.pg_error))
 694                 return pgio.pg_error;
 695         return 0;
 696 }
 697
 698 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 699 {
 700         int ret;
 701
 702         ret = nfs_writepage_locked(page, wbc);
 703         if (ret != AOP_WRITEPAGE_ACTIVATE)
 704                 unlock_page(page);
 705         return ret;
 706 }
 707
 708 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 709 {
 710         int ret;
 711
 712         ret = nfs_do_writepage(page, wbc, data);
 713         if (ret != AOP_WRITEPAGE_ACTIVATE)
 714                 unlock_page(page);
 715         return ret;
 716 }
 717
 718 static void nfs_io_completion_commit(void *inode)
 719 {
 720         nfs_commit_inode(inode, 0);
 721 }
 722
 723 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 724 {
 725         struct inode *inode = mapping->host;
 726         struct nfs_pageio_descriptor pgio;
 727         struct nfs_io_completion *ioc;
 728         int err;
 729
 730         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 731
 732         ioc = nfs_io_completion_alloc(GFP_KERNEL);
 733         if (ioc)
 734                 nfs_io_completion_init(ioc, nfs_io_completion_commit, inode);
 735
 736         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 737                                 &nfs_async_write_completion_ops);
 738         pgio.pg_io_completion = ioc;
 739         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 740         pgio.pg_error = 0;
 741         nfs_pageio_complete(&pgio);
 742         nfs_io_completion_put(ioc);
 743
 744         if (err < 0)
 745                 goto out_err;
 746         err = pgio.pg_error;
 747         if (nfs_error_is_fatal(err))
 748                 goto out_err;
 749         return 0;
 750 out_err:
 751         return err;
 752 }
 753
 754 /*
 755  * Insert a write request into an inode
 756  */
 757 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 758 {
 759         struct address_space *mapping = page_file_mapping(req->wb_page);
 760         struct nfs_inode *nfsi = NFS_I(inode);
 761
 762         WARN_ON_ONCE(req->wb_this_page != req);
 763
 764         /* Lock the request! */
 765         nfs_lock_request(req);
 766
 767         /*
 768          * Swap-space should not get truncated. Hence no need to plug the race
 769          * with invalidate/truncate.
 770          */
 771         spin_lock(&mapping->private_lock);
 772         if (!nfs_have_writebacks(inode) &&
 773             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 774                 inode_inc_iversion_raw(inode);
 775         if (likely(!PageSwapCache(req->wb_page))) {
 776                 set_bit(PG_MAPPED, &req->wb_flags);
 777                 SetPagePrivate(req->wb_page);
 778                 set_page_private(req->wb_page, (unsigned long)req);
 779         }
 780         spin_unlock(&mapping->private_lock);
 781         atomic_long_inc(&nfsi->nrequests);
 782         /* this a head request for a page group - mark it as having an
 783          * extra reference so sub groups can follow suit.
 784          * This flag also informs pgio layer when to bump nrequests when
 785          * adding subrequests. */
 786         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 787         kref_get(&req->wb_kref);
 788 }
 789
 790 /*
 791  * Remove a write request from an inode
 792  */
 793 static void nfs_inode_remove_request(struct nfs_page *req)
 794 {
 795         struct address_space *mapping = page_file_mapping(req->wb_page);
 796         struct inode *inode = mapping->host;
 797         struct nfs_inode *nfsi = NFS_I(inode);
 798         struct nfs_page *head;
 799
 800         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 801                 head = req->wb_head;
 802
 803                 spin_lock(&mapping->private_lock);
 804                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 805                         set_page_private(head->wb_page, 0);
 806                         ClearPagePrivate(head->wb_page);
 807                         clear_bit(PG_MAPPED, &head->wb_flags);
 808                 }
 809                 spin_unlock(&mapping->private_lock);
 810         }
 811
 812         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 813                 nfs_release_request(req);
 814                 atomic_long_dec(&nfsi->nrequests);
 815         }
 816 }
 817
 818 static void
 819 nfs_mark_request_dirty(struct nfs_page *req)
 820 {
 821         if (req->wb_page)
 822                 __set_page_dirty_nobuffers(req->wb_page);
 823 }
 824
 825 /*
 826  * nfs_page_search_commits_for_head_request_locked
 827  *
 828  * Search through commit lists on @inode for the head request for @page.
 829  * Must be called while holding the inode (which is cinfo) lock.
 830  *
 831  * Returns the head request if found, or NULL if not found.
 832  */
 833 static struct nfs_page *
 834 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 835                                                 struct page *page)
 836 {
 837         struct nfs_page *freq, *t;
 838         struct nfs_commit_info cinfo;
 839         struct inode *inode = &nfsi->vfs_inode;
 840
 841         nfs_init_cinfo_from_inode(&cinfo, inode);
 842
 843         /* search through pnfs commit lists */
 844         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 845         if (freq)
 846                 return freq->wb_head;
 847
 848         /* Linearly search the commit list for the correct request */
 849         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 850                 if (freq->wb_page == page)
 851                         return freq->wb_head;
 852         }
 853
 854         return NULL;
 855 }
 856
 857 /**
 858  * nfs_request_add_commit_list_locked - add request to a commit list
 859  * @req: pointer to a struct nfs_page
 860  * @dst: commit list head
 861  * @cinfo: holds list lock and accounting info
 862  *
 863  * This sets the PG_CLEAN bit, updates the cinfo count of
 864  * number of outstanding requests requiring a commit as well as
 865  * the MM page stats.
 866  *
 867  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 868  * nfs_page lock.
 869  */
 870 void
 871 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 872                             struct nfs_commit_info *cinfo)
 873 {
 874         set_bit(PG_CLEAN, &req->wb_flags);
 875         nfs_list_add_request(req, dst);
 876         atomic_long_inc(&cinfo->mds->ncommit);
 877 }
 878 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 879
 880 /**
 881  * nfs_request_add_commit_list - add request to a commit list
 882  * @req: pointer to a struct nfs_page
 883  * @cinfo: holds list lock and accounting info
 884  *
 885  * This sets the PG_CLEAN bit, updates the cinfo count of
 886  * number of outstanding requests requiring a commit as well as
 887  * the MM page stats.
 888  *
 889  * The caller must _not_ hold the cinfo->lock, but must be
 890  * holding the nfs_page lock.
 891  */
 892 void
 893 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 894 {
 895         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 896         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 897         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 898         if (req->wb_page)
 899                 nfs_mark_page_unstable(req->wb_page, cinfo);
 900 }
 901 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 902
 903 /**
 904  * nfs_request_remove_commit_list - Remove request from a commit list
 905  * @req: pointer to a nfs_page
 906  * @cinfo: holds list lock and accounting info
 907  *
 908  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 909  * number of outstanding requests requiring a commit
 910  * It does not update the MM page stats.
 911  *
 912  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 913  */
 914 void
 915 nfs_request_remove_commit_list(struct nfs_page *req,
 916                                struct nfs_commit_info *cinfo)
 917 {
 918         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 919                 return;
 920         nfs_list_remove_request(req);
 921         atomic_long_dec(&cinfo->mds->ncommit);
 922 }
 923 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 924
 925 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 926                                       struct inode *inode)
 927 {
 928         cinfo->inode = inode;
 929         cinfo->mds = &NFS_I(inode)->commit_info;
 930         cinfo->ds = pnfs_get_ds_info(inode);
 931         cinfo->dreq = NULL;
 932         cinfo->completion_ops = &nfs_commit_completion_ops;
 933 }
 934
 935 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 936                     struct inode *inode,
 937                     struct nfs_direct_req *dreq)
 938 {
 939         if (dreq)
 940                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 941         else
 942                 nfs_init_cinfo_from_inode(cinfo, inode);
 943 }
 944 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 945
 946 /*
 947  * Add a request to the inode's commit list.
 948  */
 949 void
 950 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 951                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 952 {
 953         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 954                 return;
 955         nfs_request_add_commit_list(req, cinfo);
 956 }
 957
 958 static void
 959 nfs_clear_page_commit(struct page *page)
 960 {
 961         dec_node_page_state(page, NR_UNSTABLE_NFS);
 962         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 963                     WB_RECLAIMABLE);
 964 }
 965
 966 /* Called holding the request lock on @req */
 967 static void
 968 nfs_clear_request_commit(struct nfs_page *req)
 969 {
 970         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 971                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 972                 struct inode *inode = d_inode(ctx->dentry);
 973                 struct nfs_commit_info cinfo;
 974
 975                 nfs_init_cinfo_from_inode(&cinfo, inode);
 976                 mutex_lock(&NFS_I(inode)->commit_mutex);
 977                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 978                         nfs_request_remove_commit_list(req, &cinfo);
 979                 }
 980                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 981                 nfs_clear_page_commit(req->wb_page);
 982         }
 983 }
 984
 985 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 986 {
 987         if (hdr->verf.committed == NFS_DATA_SYNC)
 988                 return hdr->lseg == NULL;
 989         return hdr->verf.committed != NFS_FILE_SYNC;
 990 }
 991
 992 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 993 {
 994         nfs_io_completion_get(hdr->io_completion);
 995 }
 996
 997 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 998 {
 999         struct nfs_commit_info cinfo;
1000         unsigned long bytes = 0;
1001
1002         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1003                 goto out;
1004         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1005         while (!list_empty(&hdr->pages)) {
1006                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1007
1008                 bytes += req->wb_bytes;
1009                 nfs_list_remove_request(req);
1010                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1011                     (hdr->good_bytes < bytes)) {
1012                         trace_nfs_comp_error(req, hdr->error);
1013                         nfs_mapping_set_error(req->wb_page, hdr->error);
1014                         goto remove_req;
1015                 }
1016                 if (nfs_write_need_commit(hdr)) {
1017                         /* Reset wb_nio, since the write was successful. */
1018                         req->wb_nio = 0;
1019                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1020                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1021                                 hdr->pgio_mirror_idx);
1022                         goto next;
1023                 }
1024 remove_req:
1025                 nfs_inode_remove_request(req);
1026 next:
1027                 nfs_end_page_writeback(req);
1028                 nfs_release_request(req);
1029         }
1030 out:
1031         nfs_io_completion_put(hdr->io_completion);
1032         hdr->release(hdr);
1033 }
1034
1035 unsigned long
1036 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1037 {
1038         return atomic_long_read(&cinfo->mds->ncommit);
1039 }
1040
1041 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1042 int
1043 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1044                      struct nfs_commit_info *cinfo, int max)
1045 {
1046         struct nfs_page *req, *tmp;
1047         int ret = 0;
1048
1049 restart:
1050         list_for_each_entry_safe(req, tmp, src, wb_list) {
1051                 kref_get(&req->wb_kref);
1052                 if (!nfs_lock_request(req)) {
1053                         int status;
1054
1055                         /* Prevent deadlock with nfs_lock_and_join_requests */
1056                         if (!list_empty(dst)) {
1057                                 nfs_release_request(req);
1058                                 continue;
1059                         }
1060                         /* Ensure we make progress to prevent livelock */
1061                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1062                         status = nfs_wait_on_request(req);
1063                         nfs_release_request(req);
1064                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1065                         if (status < 0)
1066                                 break;
1067                         goto restart;
1068                 }
1069                 nfs_request_remove_commit_list(req, cinfo);
1070                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1071                 nfs_list_add_request(req, dst);
1072                 ret++;
1073                 if ((ret == max) && !cinfo->dreq)
1074                         break;
1075                 cond_resched();
1076         }
1077         return ret;
1078 }
1079 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1080
1081 /*
1082  * nfs_scan_commit - Scan an inode for commit requests
1083  * @inode: NFS inode to scan
1084  * @dst: mds destination list
1085  * @cinfo: mds and ds lists of reqs ready to commit
1086  *
1087  * Moves requests from the inode's 'commit' request list.
1088  * The requests are *not* checked to ensure that they form a contiguous set.
1089  */
1090 int
1091 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1092                 struct nfs_commit_info *cinfo)
1093 {
1094         int ret = 0;
1095
1096         if (!atomic_long_read(&cinfo->mds->ncommit))
1097                 return 0;
1098         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1099         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1100                 const int max = INT_MAX;
1101
1102                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1103                                            cinfo, max);
1104                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1105         }
1106         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1107         return ret;
1108 }
1109
1110 /*
1111  * Search for an existing write request, and attempt to update
1112  * it to reflect a new dirty region on a given page.
1113  *
1114  * If the attempt fails, then the existing request is flushed out
1115  * to disk.
1116  */
1117 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1118                 struct page *page,
1119                 unsigned int offset,
1120                 unsigned int bytes)
1121 {
1122         struct nfs_page *req;
1123         unsigned int rqend;
1124         unsigned int end;
1125         int error;
1126
1127         end = offset + bytes;
1128
1129         req = nfs_lock_and_join_requests(page);
1130         if (IS_ERR_OR_NULL(req))
1131                 return req;
1132
1133         rqend = req->wb_offset + req->wb_bytes;
1134         /*
1135          * Tell the caller to flush out the request if
1136          * the offsets are non-contiguous.
1137          * Note: nfs_flush_incompatible() will already
1138          * have flushed out requests having wrong owners.
1139          */
1140         if (offset > rqend || end < req->wb_offset)
1141                 goto out_flushme;
1142
1143         /* Okay, the request matches. Update the region */
1144         if (offset < req->wb_offset) {
1145                 req->wb_offset = offset;
1146                 req->wb_pgbase = offset;
1147         }
1148         if (end > rqend)
1149                 req->wb_bytes = end - req->wb_offset;
1150         else
1151                 req->wb_bytes = rqend - req->wb_offset;
1152         req->wb_nio = 0;
1153         return req;
1154 out_flushme:
1155         /*
1156          * Note: we mark the request dirty here because
1157          * nfs_lock_and_join_requests() cannot preserve
1158          * commit flags, so we have to replay the write.
1159          */
1160         nfs_mark_request_dirty(req);
1161         nfs_unlock_and_release_request(req);
1162         error = nfs_wb_page(inode, page);
1163         return (error < 0) ? ERR_PTR(error) : NULL;
1164 }
1165
1166 /*
1167  * Try to update an existing write request, or create one if there is none.
1168  *
1169  * Note: Should always be called with the Page Lock held to prevent races
1170  * if we have to add a new request. Also assumes that the caller has
1171  * already called nfs_flush_incompatible() if necessary.
1172  */
1173 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1174                 struct page *page, unsigned int offset, unsigned int bytes)
1175 {
1176         struct inode *inode = page_file_mapping(page)->host;
1177         struct nfs_page *req;
1178
1179         req = nfs_try_to_update_request(inode, page, offset, bytes);
1180         if (req != NULL)
1181                 goto out;
1182         req = nfs_create_request(ctx, page, offset, bytes);
1183         if (IS_ERR(req))
1184                 goto out;
1185         nfs_inode_add_request(inode, req);
1186 out:
1187         return req;
1188 }
1189
1190 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1191                 unsigned int offset, unsigned int count)
1192 {
1193         struct nfs_page *req;
1194
1195         req = nfs_setup_write_request(ctx, page, offset, count);
1196         if (IS_ERR(req))
1197                 return PTR_ERR(req);
1198         /* Update file length */
1199         nfs_grow_file(page, offset, count);
1200         nfs_mark_uptodate(req);
1201         nfs_mark_request_dirty(req);
1202         nfs_unlock_and_release_request(req);
1203         return 0;
1204 }
1205
1206 int nfs_flush_incompatible(struct file *file, struct page *page)
1207 {
1208         struct nfs_open_context *ctx = nfs_file_open_context(file);
1209         struct nfs_lock_context *l_ctx;
1210         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1211         struct nfs_page *req;
1212         int do_flush, status;
1213         /*
1214          * Look for a request corresponding to this page. If there
1215          * is one, and it belongs to another file, we flush it out
1216          * before we try to copy anything into the page. Do this
1217          * due to the lack of an ACCESS-type call in NFSv2.
1218          * Also do the same if we find a request from an existing
1219          * dropped page.
1220          */
1221         do {
1222                 req = nfs_page_find_head_request(page);
1223                 if (req == NULL)
1224                         return 0;
1225                 l_ctx = req->wb_lock_context;
1226                 do_flush = req->wb_page != page ||
1227                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1228                 if (l_ctx && flctx &&
1229                     !(list_empty_careful(&flctx->flc_posix) &&
1230                       list_empty_careful(&flctx->flc_flock))) {
1231                         do_flush |= l_ctx->lockowner != current->files;
1232                 }
1233                 nfs_release_request(req);
1234                 if (!do_flush)
1235                         return 0;
1236                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1237         } while (status == 0);
1238         return status;
1239 }
1240
1241 /*
1242  * Avoid buffered writes when a open context credential's key would
1243  * expire soon.
1244  *
1245  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1246  *
1247  * Return 0 and set a credential flag which triggers the inode to flush
1248  * and performs  NFS_FILE_SYNC writes if the key will expired within
1249  * RPC_KEY_EXPIRE_TIMEO.
1250  */
1251 int
1252 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1253 {
1254         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1255
1256         if (nfs_ctx_key_to_expire(ctx, inode) &&
1257             !ctx->ll_cred)
1258                 /* Already expired! */
1259                 return -EACCES;
1260         return 0;
1261 }
1262
1263 /*
1264  * Test if the open context credential key is marked to expire soon.
1265  */
1266 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1267 {
1268         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1269         struct rpc_cred *cred = ctx->ll_cred;
1270         struct auth_cred acred = {
1271                 .cred = ctx->cred,
1272         };
1273
1274         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1275                 put_rpccred(cred);
1276                 ctx->ll_cred = NULL;
1277                 cred = NULL;
1278         }
1279         if (!cred)
1280                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1281         if (!cred || IS_ERR(cred))
1282                 return true;
1283         ctx->ll_cred = cred;
1284         return !!(cred->cr_ops->crkey_timeout &&
1285                   cred->cr_ops->crkey_timeout(cred));
1286 }
1287
1288 /*
1289  * If the page cache is marked as unsafe or invalid, then we can't rely on
1290  * the PageUptodate() flag. In this case, we will need to turn off
1291  * write optimisations that depend on the page contents being correct.
1292  */
1293 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
1294 {
1295         struct nfs_inode *nfsi = NFS_I(inode);
1296
1297         if (nfs_have_delegated_attributes(inode))
1298                 goto out;
1299         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
1300                 return false;
1301         smp_rmb();
1302         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
1303                 return false;
1304 out:
1305         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
1306                 return false;
1307         return PageUptodate(page) != 0;
1308 }
1309
1310 static bool
1311 is_whole_file_wrlock(struct file_lock *fl)
1312 {
1313         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1314                         fl->fl_type == F_WRLCK;
1315 }
1316
1317 /* If we know the page is up to date, and we're not using byte range locks (or
1318  * if we have the whole file locked for writing), it may be more efficient to
1319  * extend the write to cover the entire page in order to avoid fragmentation
1320  * inefficiencies.
1321  *
1322  * If the file is opened for synchronous writes then we can just skip the rest
1323  * of the checks.
1324  */
1325 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
1326 {
1327         int ret;
1328         struct file_lock_context *flctx = inode->i_flctx;
1329         struct file_lock *fl;
1330
1331         if (file->f_flags & O_DSYNC)
1332                 return 0;
1333         if (!nfs_write_pageuptodate(page, inode))
1334                 return 0;
1335         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1336                 return 1;
1337         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1338                        list_empty_careful(&flctx->flc_posix)))
1339                 return 1;
1340
1341         /* Check to see if there are whole file write locks */
1342         ret = 0;
1343         spin_lock(&flctx->flc_lock);
1344         if (!list_empty(&flctx->flc_posix)) {
1345                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1346                                         fl_list);
1347                 if (is_whole_file_wrlock(fl))
1348                         ret = 1;
1349         } else if (!list_empty(&flctx->flc_flock)) {
1350                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1351                                         fl_list);
1352                 if (fl->fl_type == F_WRLCK)
1353                         ret = 1;
1354         }
1355         spin_unlock(&flctx->flc_lock);
1356         return ret;
1357 }
1358
1359 /*
1360  * Update and possibly write a cached page of an NFS file.
1361  *
1362  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1363  * things with a page scheduled for an RPC call (e.g. invalidate it).
1364  */
1365 int nfs_updatepage(struct file *file, struct page *page,
1366                 unsigned int offset, unsigned int count)
1367 {
1368         struct nfs_open_context *ctx = nfs_file_open_context(file);
1369         struct address_space *mapping = page_file_mapping(page);
1370         struct inode    *inode = mapping->host;
1371         int             status = 0;
1372
1373         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1374
1375         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1376                 file, count, (long long)(page_file_offset(page) + offset));
1377
1378         if (!count)
1379                 goto out;
1380
1381         if (nfs_can_extend_write(file, page, inode)) {
1382                 count = max(count + offset, nfs_page_length(page));
1383                 offset = 0;
1384         }
1385
1386         status = nfs_writepage_setup(ctx, page, offset, count);
1387         if (status < 0)
1388                 nfs_set_pageerror(mapping);
1389         else
1390                 __set_page_dirty_nobuffers(page);
1391 out:
1392         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1393                         status, (long long)i_size_read(inode));
1394         return status;
1395 }
1396
1397 static int flush_task_priority(int how)
1398 {
1399         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1400                 case FLUSH_HIGHPRI:
1401                         return RPC_PRIORITY_HIGH;
1402                 case FLUSH_LOWPRI:
1403                         return RPC_PRIORITY_LOW;
1404         }
1405         return RPC_PRIORITY_NORMAL;
1406 }
1407
1408 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1409                                struct rpc_message *msg,
1410                                const struct nfs_rpc_ops *rpc_ops,
1411                                struct rpc_task_setup *task_setup_data, int how)
1412 {
1413         int priority = flush_task_priority(how);
1414
1415         task_setup_data->priority = priority;
1416         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1417         trace_nfs_initiate_write(hdr);
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(task, hdr);
1582
1583         if (hdr->res.verf->committed < hdr->args.stable &&
1584             task->tk_status >= 0) {
1585                 /* We tried a write call, but the server did not
1586                  * commit data to stable storage even though we
1587                  * requested it.
1588                  * Note: There is a known bug in Tru64 < 5.0 in which
1589                  *       the server reports NFS_DATA_SYNC, but performs
1590                  *       NFS_FILE_SYNC. We therefore implement this checking
1591                  *       as a dprintk() in order to avoid filling syslog.
1592                  */
1593                 static unsigned long    complain;
1594
1595                 /* Note this will print the MDS for a DS write */
1596                 if (time_before(complain, jiffies)) {
1597                         dprintk("NFS:       faulty NFS server %s:"
1598                                 " (committed = %d) != (stable = %d)\n",
1599                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1600                                 hdr->res.verf->committed, hdr->args.stable);
1601                         complain = jiffies + 300 * HZ;
1602                 }
1603         }
1604
1605         /* Deal with the suid/sgid bit corner case */
1606         if (nfs_should_remove_suid(inode)) {
1607                 spin_lock(&inode->i_lock);
1608                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1609                 spin_unlock(&inode->i_lock);
1610         }
1611         return 0;
1612 }
1613
1614 /*
1615  * This function is called when the WRITE call is complete.
1616  */
1617 static void nfs_writeback_result(struct rpc_task *task,
1618                                  struct nfs_pgio_header *hdr)
1619 {
1620         struct nfs_pgio_args    *argp = &hdr->args;
1621         struct nfs_pgio_res     *resp = &hdr->res;
1622
1623         if (resp->count < argp->count) {
1624                 static unsigned long    complain;
1625
1626                 /* This a short write! */
1627                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1628
1629                 /* Has the server at least made some progress? */
1630                 if (resp->count == 0) {
1631                         if (time_before(complain, jiffies)) {
1632                                 printk(KERN_WARNING
1633                                        "NFS: Server wrote zero bytes, expected %u.\n",
1634                                        argp->count);
1635                                 complain = jiffies + 300 * HZ;
1636                         }
1637                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1638                         task->tk_status = -EIO;
1639                         return;
1640                 }
1641
1642                 /* For non rpc-based layout drivers, retry-through-MDS */
1643                 if (!task->tk_ops) {
1644                         hdr->pnfs_error = -EAGAIN;
1645                         return;
1646                 }
1647
1648                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1649                 if (resp->verf->committed != NFS_UNSTABLE) {
1650                         /* Resend from where the server left off */
1651                         hdr->mds_offset += resp->count;
1652                         argp->offset += resp->count;
1653                         argp->pgbase += resp->count;
1654                         argp->count -= resp->count;
1655                 } else {
1656                         /* Resend as a stable write in order to avoid
1657                          * headaches in the case of a server crash.
1658                          */
1659                         argp->stable = NFS_FILE_SYNC;
1660                 }
1661                 resp->count = 0;
1662                 resp->verf->committed = 0;
1663                 rpc_restart_call_prepare(task);
1664         }
1665 }
1666
1667 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1668 {
1669         return wait_var_event_killable(&cinfo->rpcs_out,
1670                                        !atomic_read(&cinfo->rpcs_out));
1671 }
1672
1673 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1674 {
1675         atomic_inc(&cinfo->rpcs_out);
1676 }
1677
1678 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1679 {
1680         if (atomic_dec_and_test(&cinfo->rpcs_out))
1681                 wake_up_var(&cinfo->rpcs_out);
1682 }
1683
1684 void nfs_commitdata_release(struct nfs_commit_data *data)
1685 {
1686         put_nfs_open_context(data->context);
1687         nfs_commit_free(data);
1688 }
1689 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1690
1691 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1692                         const struct nfs_rpc_ops *nfs_ops,
1693                         const struct rpc_call_ops *call_ops,
1694                         int how, int flags)
1695 {
1696         struct rpc_task *task;
1697         int priority = flush_task_priority(how);
1698         struct rpc_message msg = {
1699                 .rpc_argp = &data->args,
1700                 .rpc_resp = &data->res,
1701                 .rpc_cred = data->cred,
1702         };
1703         struct rpc_task_setup task_setup_data = {
1704                 .task = &data->task,
1705                 .rpc_client = clnt,
1706                 .rpc_message = &msg,
1707                 .callback_ops = call_ops,
1708                 .callback_data = data,
1709                 .workqueue = nfsiod_workqueue,
1710                 .flags = RPC_TASK_ASYNC | flags,
1711                 .priority = priority,
1712         };
1713         /* Set up the initial task struct.  */
1714         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1715         trace_nfs_initiate_commit(data);
1716
1717         dprintk("NFS: initiated commit call\n");
1718
1719         task = rpc_run_task(&task_setup_data);
1720         if (IS_ERR(task))
1721                 return PTR_ERR(task);
1722         if (how & FLUSH_SYNC)
1723                 rpc_wait_for_completion_task(task);
1724         rpc_put_task(task);
1725         return 0;
1726 }
1727 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1728
1729 static loff_t nfs_get_lwb(struct list_head *head)
1730 {
1731         loff_t lwb = 0;
1732         struct nfs_page *req;
1733
1734         list_for_each_entry(req, head, wb_list)
1735                 if (lwb < (req_offset(req) + req->wb_bytes))
1736                         lwb = req_offset(req) + req->wb_bytes;
1737
1738         return lwb;
1739 }
1740
1741 /*
1742  * Set up the argument/result storage required for the RPC call.
1743  */
1744 void nfs_init_commit(struct nfs_commit_data *data,
1745                      struct list_head *head,
1746                      struct pnfs_layout_segment *lseg,
1747                      struct nfs_commit_info *cinfo)
1748 {
1749         struct nfs_page *first = nfs_list_entry(head->next);
1750         struct nfs_open_context *ctx = nfs_req_openctx(first);
1751         struct inode *inode = d_inode(ctx->dentry);
1752
1753         /* Set up the RPC argument and reply structs
1754          * NB: take care not to mess about with data->commit et al. */
1755
1756         list_splice_init(head, &data->pages);
1757
1758         data->inode       = inode;
1759         data->cred        = ctx->cred;
1760         data->lseg        = lseg; /* reference transferred */
1761         /* only set lwb for pnfs commit */
1762         if (lseg)
1763                 data->lwb = nfs_get_lwb(&data->pages);
1764         data->mds_ops     = &nfs_commit_ops;
1765         data->completion_ops = cinfo->completion_ops;
1766         data->dreq        = cinfo->dreq;
1767
1768         data->args.fh     = NFS_FH(data->inode);
1769         /* Note: we always request a commit of the entire inode */
1770         data->args.offset = 0;
1771         data->args.count  = 0;
1772         data->context     = get_nfs_open_context(ctx);
1773         data->res.fattr   = &data->fattr;
1774         data->res.verf    = &data->verf;
1775         nfs_fattr_init(&data->fattr);
1776 }
1777 EXPORT_SYMBOL_GPL(nfs_init_commit);
1778
1779 void nfs_retry_commit(struct list_head *page_list,
1780                       struct pnfs_layout_segment *lseg,
1781                       struct nfs_commit_info *cinfo,
1782                       u32 ds_commit_idx)
1783 {
1784         struct nfs_page *req;
1785
1786         while (!list_empty(page_list)) {
1787                 req = nfs_list_entry(page_list->next);
1788                 nfs_list_remove_request(req);
1789                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1790                 if (!cinfo->dreq)
1791                         nfs_clear_page_commit(req->wb_page);
1792                 nfs_unlock_and_release_request(req);
1793         }
1794 }
1795 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1796
1797 static void
1798 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1799                 struct nfs_page *req)
1800 {
1801         __set_page_dirty_nobuffers(req->wb_page);
1802 }
1803
1804 /*
1805  * Commit dirty pages
1806  */
1807 static int
1808 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1809                 struct nfs_commit_info *cinfo)
1810 {
1811         struct nfs_commit_data  *data;
1812
1813         /* another commit raced with us */
1814         if (list_empty(head))
1815                 return 0;
1816
1817         data = nfs_commitdata_alloc(true);
1818
1819         /* Set up the argument struct */
1820         nfs_init_commit(data, head, NULL, cinfo);
1821         atomic_inc(&cinfo->mds->rpcs_out);
1822         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1823                                    data->mds_ops, how, 0);
1824 }
1825
1826 /*
1827  * COMMIT call returned
1828  */
1829 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1830 {
1831         struct nfs_commit_data  *data = calldata;
1832
1833         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1834                                 task->tk_pid, task->tk_status);
1835
1836         /* Call the NFS version-specific code */
1837         NFS_PROTO(data->inode)->commit_done(task, data);
1838         trace_nfs_commit_done(task, data);
1839 }
1840
1841 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1842 {
1843         const struct nfs_writeverf *verf = data->res.verf;
1844         struct nfs_page *req;
1845         int status = data->task.tk_status;
1846         struct nfs_commit_info cinfo;
1847         struct nfs_server *nfss;
1848
1849         while (!list_empty(&data->pages)) {
1850                 req = nfs_list_entry(data->pages.next);
1851                 nfs_list_remove_request(req);
1852                 if (req->wb_page)
1853                         nfs_clear_page_commit(req->wb_page);
1854
1855                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1856                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1857                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1858                         req->wb_bytes,
1859                         (long long)req_offset(req));
1860                 if (status < 0) {
1861                         if (req->wb_page) {
1862                                 trace_nfs_commit_error(req, status);
1863                                 nfs_mapping_set_error(req->wb_page, status);
1864                                 nfs_inode_remove_request(req);
1865                         }
1866                         dprintk_cont(", error = %d\n", status);
1867                         goto next;
1868                 }
1869
1870                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1871                  * returned by the server against all stored verfs. */
1872                 if (verf->committed > NFS_UNSTABLE &&
1873                     !nfs_write_verifier_cmp(&req->wb_verf, &verf->verifier)) {
1874                         /* We have a match */
1875                         if (req->wb_page)
1876                                 nfs_inode_remove_request(req);
1877                         dprintk_cont(" OK\n");
1878                         goto next;
1879                 }
1880                 /* We have a mismatch. Write the page again */
1881                 dprintk_cont(" mismatch\n");
1882                 nfs_mark_request_dirty(req);
1883                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1884         next:
1885                 nfs_unlock_and_release_request(req);
1886                 /* Latency breaker */
1887                 cond_resched();
1888         }
1889         nfss = NFS_SERVER(data->inode);
1890         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1891                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1892
1893         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1894         nfs_commit_end(cinfo.mds);
1895 }
1896
1897 static void nfs_commit_release(void *calldata)
1898 {
1899         struct nfs_commit_data *data = calldata;
1900
1901         data->completion_ops->completion(data);
1902         nfs_commitdata_release(calldata);
1903 }
1904
1905 static const struct rpc_call_ops nfs_commit_ops = {
1906         .rpc_call_prepare = nfs_commit_prepare,
1907         .rpc_call_done = nfs_commit_done,
1908         .rpc_release = nfs_commit_release,
1909 };
1910
1911 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1912         .completion = nfs_commit_release_pages,
1913         .resched_write = nfs_commit_resched_write,
1914 };
1915
1916 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1917                             int how, struct nfs_commit_info *cinfo)
1918 {
1919         int status;
1920
1921         status = pnfs_commit_list(inode, head, how, cinfo);
1922         if (status == PNFS_NOT_ATTEMPTED)
1923                 status = nfs_commit_list(inode, head, how, cinfo);
1924         return status;
1925 }
1926
1927 static int __nfs_commit_inode(struct inode *inode, int how,
1928                 struct writeback_control *wbc)
1929 {
1930         LIST_HEAD(head);
1931         struct nfs_commit_info cinfo;
1932         int may_wait = how & FLUSH_SYNC;
1933         int ret, nscan;
1934
1935         nfs_init_cinfo_from_inode(&cinfo, inode);
1936         nfs_commit_begin(cinfo.mds);
1937         for (;;) {
1938                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1939                 if (ret <= 0)
1940                         break;
1941                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1942                 if (ret < 0)
1943                         break;
1944                 ret = 0;
1945                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1946                         if (nscan < wbc->nr_to_write)
1947                                 wbc->nr_to_write -= nscan;
1948                         else
1949                                 wbc->nr_to_write = 0;
1950                 }
1951                 if (nscan < INT_MAX)
1952                         break;
1953                 cond_resched();
1954         }
1955         nfs_commit_end(cinfo.mds);
1956         if (ret || !may_wait)
1957                 return ret;
1958         return wait_on_commit(cinfo.mds);
1959 }
1960
1961 int nfs_commit_inode(struct inode *inode, int how)
1962 {
1963         return __nfs_commit_inode(inode, how, NULL);
1964 }
1965 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1966
1967 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1968 {
1969         struct nfs_inode *nfsi = NFS_I(inode);
1970         int flags = FLUSH_SYNC;
1971         int ret = 0;
1972
1973         if (wbc->sync_mode == WB_SYNC_NONE) {
1974                 /* no commits means nothing needs to be done */
1975                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1976                         goto check_requests_outstanding;
1977
1978                 /* Don't commit yet if this is a non-blocking flush and there
1979                  * are a lot of outstanding writes for this mapping.
1980                  */
1981                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1982                         goto out_mark_dirty;
1983
1984                 /* don't wait for the COMMIT response */
1985                 flags = 0;
1986         }
1987
1988         ret = __nfs_commit_inode(inode, flags, wbc);
1989         if (!ret) {
1990                 if (flags & FLUSH_SYNC)
1991                         return 0;
1992         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1993                 goto out_mark_dirty;
1994
1995 check_requests_outstanding:
1996         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1997                 return ret;
1998 out_mark_dirty:
1999         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2000         return ret;
2001 }
2002 EXPORT_SYMBOL_GPL(nfs_write_inode);
2003
2004 /*
2005  * Wrapper for filemap_write_and_wait_range()
2006  *
2007  * Needed for pNFS in order to ensure data becomes visible to the
2008  * client.
2009  */
2010 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2011                 loff_t lstart, loff_t lend)
2012 {
2013         int ret;
2014
2015         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2016         if (ret == 0)
2017                 ret = pnfs_sync_inode(mapping->host, true);
2018         return ret;
2019 }
2020 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2021
2022 /*
2023  * flush the inode to disk.
2024  */
2025 int nfs_wb_all(struct inode *inode)
2026 {
2027         int ret;
2028
2029         trace_nfs_writeback_inode_enter(inode);
2030
2031         ret = filemap_write_and_wait(inode->i_mapping);
2032         if (ret)
2033                 goto out;
2034         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2035         if (ret < 0)
2036                 goto out;
2037         pnfs_sync_inode(inode, true);
2038         ret = 0;
2039
2040 out:
2041         trace_nfs_writeback_inode_exit(inode, ret);
2042         return ret;
2043 }
2044 EXPORT_SYMBOL_GPL(nfs_wb_all);
2045
2046 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2047 {
2048         struct nfs_page *req;
2049         int ret = 0;
2050
2051         wait_on_page_writeback(page);
2052
2053         /* blocking call to cancel all requests and join to a single (head)
2054          * request */
2055         req = nfs_lock_and_join_requests(page);
2056
2057         if (IS_ERR(req)) {
2058                 ret = PTR_ERR(req);
2059         } else if (req) {
2060                 /* all requests from this page have been cancelled by
2061                  * nfs_lock_and_join_requests, so just remove the head
2062                  * request from the inode / page_private pointer and
2063                  * release it */
2064                 nfs_inode_remove_request(req);
2065                 nfs_unlock_and_release_request(req);
2066         }
2067
2068         return ret;
2069 }
2070
2071 /*
2072  * Write back all requests on one page - we do this before reading it.
2073  */
2074 int nfs_wb_page(struct inode *inode, struct page *page)
2075 {
2076         loff_t range_start = page_file_offset(page);
2077         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2078         struct writeback_control wbc = {
2079                 .sync_mode = WB_SYNC_ALL,
2080                 .nr_to_write = 0,
2081                 .range_start = range_start,
2082                 .range_end = range_end,
2083         };
2084         int ret;
2085
2086         trace_nfs_writeback_page_enter(inode);
2087
2088         for (;;) {
2089                 wait_on_page_writeback(page);
2090                 if (clear_page_dirty_for_io(page)) {
2091                         ret = nfs_writepage_locked(page, &wbc);
2092                         if (ret < 0)
2093                                 goto out_error;
2094                         continue;
2095                 }
2096                 ret = 0;
2097                 if (!PagePrivate(page))
2098                         break;
2099                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2100                 if (ret < 0)
2101                         goto out_error;
2102         }
2103 out_error:
2104         trace_nfs_writeback_page_exit(inode, ret);
2105         return ret;
2106 }
2107
2108 #ifdef CONFIG_MIGRATION
2109 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2110                 struct page *page, enum migrate_mode mode)
2111 {
2112         /*
2113          * If PagePrivate is set, then the page is currently associated with
2114          * an in-progress read or write request. Don't try to migrate it.
2115          *
2116          * FIXME: we could do this in principle, but we'll need a way to ensure
2117          *        that we can safely release the inode reference while holding
2118          *        the page lock.
2119          */
2120         if (PagePrivate(page))
2121                 return -EBUSY;
2122
2123         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2124                 return -EBUSY;
2125
2126         return migrate_page(mapping, newpage, page, mode);
2127 }
2128 #endif
2129
2130 int __init nfs_init_writepagecache(void)
2131 {
2132         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2133                                              sizeof(struct nfs_pgio_header),
2134                                              0, SLAB_HWCACHE_ALIGN,
2135                                              NULL);
2136         if (nfs_wdata_cachep == NULL)
2137                 return -ENOMEM;
2138
2139         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2140                                                      nfs_wdata_cachep);
2141         if (nfs_wdata_mempool == NULL)
2142                 goto out_destroy_write_cache;
2143
2144         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2145                                              sizeof(struct nfs_commit_data),
2146                                              0, SLAB_HWCACHE_ALIGN,
2147                                              NULL);
2148         if (nfs_cdata_cachep == NULL)
2149                 goto out_destroy_write_mempool;
2150
2151         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2152                                                       nfs_cdata_cachep);
2153         if (nfs_commit_mempool == NULL)
2154                 goto out_destroy_commit_cache;
2155
2156         /*
2157          * NFS congestion size, scale with available memory.
2158          *
2159          *  64MB:    8192k
2160          * 128MB:   11585k
2161          * 256MB:   16384k
2162          * 512MB:   23170k
2163          *   1GB:   32768k
2164          *   2GB:   46340k
2165          *   4GB:   65536k
2166          *   8GB:   92681k
2167          *  16GB:  131072k
2168          *
2169          * This allows larger machines to have larger/more transfers.
2170          * Limit the default to 256M
2171          */
2172         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2173         if (nfs_congestion_kb > 256*1024)
2174                 nfs_congestion_kb = 256*1024;
2175
2176         return 0;
2177
2178 out_destroy_commit_cache:
2179         kmem_cache_destroy(nfs_cdata_cachep);
2180 out_destroy_write_mempool:
2181         mempool_destroy(nfs_wdata_mempool);
2182 out_destroy_write_cache:
2183         kmem_cache_destroy(nfs_wdata_cachep);
2184         return -ENOMEM;
2185 }
2186
2187 void nfs_destroy_writepagecache(void)
2188 {
2189         mempool_destroy(nfs_commit_mempool);
2190         kmem_cache_destroy(nfs_cdata_cachep);
2191         mempool_destroy(nfs_wdata_mempool);
2192         kmem_cache_destroy(nfs_wdata_cachep);
2193 }
2194
2195 static const struct nfs_rw_ops nfs_rw_write_ops = {
2196         .rw_alloc_header        = nfs_writehdr_alloc,
2197         .rw_free_header         = nfs_writehdr_free,
2198         .rw_done                = nfs_writeback_done,
2199         .rw_result              = nfs_writeback_result,
2200         .rw_initiate            = nfs_initiate_write,
2201 };