fs/nfs/write.c

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