fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Writing file data over NFS.
   5  *
   6  * We do it like this: When a (user) process wishes to write data to an
   7  * NFS file, a write request is allocated that contains the RPC task data
   8  * plus some info on the page to be written, and added to the inode's
   9  * write chain. If the process writes past the end of the page, an async
  10  * RPC call to write the page is scheduled immediately; otherwise, the call
  11  * is delayed for a few seconds.
  12  *
  13  * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
  14  *
  15  * Write requests are kept on the inode's writeback list. Each entry in
  16  * that list references the page (portion) to be written. When the
  17  * cache timeout has expired, the RPC task is woken up, and tries to
  18  * lock the page. As soon as it manages to do so, the request is moved
  19  * from the writeback list to the writelock list.
  20  *
  21  * Note: we must make sure never to confuse the inode passed in the
  22  * write_page request with the one in page->inode. As far as I understand
  23  * it, these are different when doing a swap-out.
  24  *
  25  * To understand everything that goes on here and in the NFS read code,
  26  * one should be aware that a page is locked in exactly one of the following
  27  * cases:
  28  *
  29  *  -   A write request is in progress.
  30  *  -   A user process is in generic_file_write/nfs_update_page
  31  *  -   A user process is in generic_file_read
  32  *
  33  * Also note that because of the way pages are invalidated in
  34  * nfs_revalidate_inode, the following assertions hold:
  35  *
  36  *  -   If a page is dirty, there will be no read requests (a page will
  37  *      not be re-read unless invalidated by nfs_revalidate_inode).
  38  *  -   If the page is not uptodate, there will be no pending write
  39  *      requests, and no process will be in nfs_update_page.
  40  *
  41  * FIXME: Interaction with the vmscan routines is not optimal yet.
  42  * Either vmscan must be made nfs-savvy, or we need a different page
  43  * reclaim concept that supports something like FS-independent
  44  * buffer_heads with a b_ops-> field.
  45  *
  46  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
  47  */
  48
  49 #include <linux/types.h>
  50 #include <linux/slab.h>
  51 #include <linux/mm.h>
  52 #include <linux/pagemap.h>
  53 #include <linux/file.h>
  54 #include <linux/writeback.h>
  55
  56 #include <linux/sunrpc/clnt.h>
  57 #include <linux/nfs_fs.h>
  58 #include <linux/nfs_mount.h>
  59 #include <linux/nfs_page.h>
  60 #include <linux/backing-dev.h>
  61
  62 #include <asm/uaccess.h>
  63 #include <linux/smp_lock.h>
  64
  65 #include "delegation.h"
  66 #include "internal.h"
  67 #include "iostat.h"
  68
  69 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  70
  71 #define MIN_POOL_WRITE          (32)
  72 #define MIN_POOL_COMMIT         (4)
  73
  74 /*
  75  * Local function declarations
  76  */
  77 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  78                                             struct page *,
  79                                             unsigned int, unsigned int);
  80 static void nfs_mark_request_dirty(struct nfs_page *req);
  81 static int nfs_wait_on_write_congestion(struct address_space *, int);
  82 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
  83 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  84 static const struct rpc_call_ops nfs_write_partial_ops;
  85 static const struct rpc_call_ops nfs_write_full_ops;
  86 static const struct rpc_call_ops nfs_commit_ops;
  87
  88 static struct kmem_cache *nfs_wdata_cachep;
  89 static mempool_t *nfs_wdata_mempool;
  90 static mempool_t *nfs_commit_mempool;
  91
  92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  93
  94 struct nfs_write_data *nfs_commit_alloc(void)
  95 {
  96         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  97
  98         if (p) {
  99                 memset(p, 0, sizeof(*p));
 100                 INIT_LIST_HEAD(&p->pages);
 101         }
 102         return p;
 103 }
 104
 105 void nfs_commit_rcu_free(struct rcu_head *head)
 106 {
 107         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 108         if (p && (p->pagevec != &p->page_array[0]))
 109                 kfree(p->pagevec);
 110         mempool_free(p, nfs_commit_mempool);
 111 }
 112
 113 void nfs_commit_free(struct nfs_write_data *wdata)
 114 {
 115         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
 116 }
 117
 118 struct nfs_write_data *nfs_writedata_alloc(size_t len)
 119 {
 120         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 121         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
 122
 123         if (p) {
 124                 memset(p, 0, sizeof(*p));
 125                 INIT_LIST_HEAD(&p->pages);
 126                 p->npages = pagecount;
 127                 if (pagecount <= ARRAY_SIZE(p->page_array))
 128                         p->pagevec = p->page_array;
 129                 else {
 130                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 131                         if (!p->pagevec) {
 132                                 mempool_free(p, nfs_wdata_mempool);
 133                                 p = NULL;
 134                         }
 135                 }
 136         }
 137         return p;
 138 }
 139
 140 static void nfs_writedata_rcu_free(struct rcu_head *head)
 141 {
 142         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 143         if (p && (p->pagevec != &p->page_array[0]))
 144                 kfree(p->pagevec);
 145         mempool_free(p, nfs_wdata_mempool);
 146 }
 147
 148 static void nfs_writedata_free(struct nfs_write_data *wdata)
 149 {
 150         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 151 }
 152
 153 void nfs_writedata_release(void *wdata)
 154 {
 155         nfs_writedata_free(wdata);
 156 }
 157
 158 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 159 {
 160         struct nfs_page *req = NULL;
 161
 162         if (PagePrivate(page)) {
 163                 req = (struct nfs_page *)page_private(page);
 164                 if (req != NULL)
 165                         atomic_inc(&req->wb_count);
 166         }
 167         return req;
 168 }
 169
 170 static struct nfs_page *nfs_page_find_request(struct page *page)
 171 {
 172         struct nfs_page *req = NULL;
 173         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 174
 175         spin_lock(req_lock);
 176         req = nfs_page_find_request_locked(page);
 177         spin_unlock(req_lock);
 178         return req;
 179 }
 180
 181 /* Adjust the file length if we're writing beyond the end */
 182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         loff_t end, i_size = i_size_read(inode);
 186         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 187
 188         if (i_size > 0 && page->index < end_index)
 189                 return;
 190         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 191         if (i_size >= end)
 192                 return;
 193         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 194         i_size_write(inode, end);
 195 }
 196
 197 /* We can set the PG_uptodate flag if we see that a write request
 198  * covers the full page.
 199  */
 200 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 201 {
 202         if (PageUptodate(page))
 203                 return;
 204         if (base != 0)
 205                 return;
 206         if (count != nfs_page_length(page))
 207                 return;
 208         if (count != PAGE_CACHE_SIZE)
 209                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 210         SetPageUptodate(page);
 211 }
 212
 213 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 214                 unsigned int offset, unsigned int count)
 215 {
 216         struct nfs_page *req;
 217         int ret;
 218
 219         for (;;) {
 220                 req = nfs_update_request(ctx, page, offset, count);
 221                 if (!IS_ERR(req))
 222                         break;
 223                 ret = PTR_ERR(req);
 224                 if (ret != -EBUSY)
 225                         return ret;
 226                 ret = nfs_wb_page(page->mapping->host, page);
 227                 if (ret != 0)
 228                         return ret;
 229         }
 230         /* Update file length */
 231         nfs_grow_file(page, offset, count);
 232         /* Set the PG_uptodate flag? */
 233         nfs_mark_uptodate(page, offset, count);
 234         nfs_unlock_request(req);
 235         return 0;
 236 }
 237
 238 static int wb_priority(struct writeback_control *wbc)
 239 {
 240         if (wbc->for_reclaim)
 241                 return FLUSH_HIGHPRI;
 242         if (wbc->for_kupdate)
 243                 return FLUSH_LOWPRI;
 244         return 0;
 245 }
 246
 247 /*
 248  * Find an associated nfs write request, and prepare to flush it out
 249  * Returns 1 if there was no write request, or if the request was
 250  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 251  * was not tagged.
 252  * May also return an error if the user signalled nfs_wait_on_request().
 253  */
 254 static int nfs_page_mark_flush(struct page *page)
 255 {
 256         struct nfs_page *req;
 257         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 258         int ret;
 259
 260         spin_lock(req_lock);
 261         for(;;) {
 262                 req = nfs_page_find_request_locked(page);
 263                 if (req == NULL) {
 264                         spin_unlock(req_lock);
 265                         return 1;
 266                 }
 267                 if (nfs_lock_request_dontget(req))
 268                         break;
 269                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 270                  *       then the call to nfs_lock_request_dontget() will always
 271                  *       succeed provided that someone hasn't already marked the
 272                  *       request as dirty (in which case we don't care).
 273                  */
 274                 spin_unlock(req_lock);
 275                 ret = nfs_wait_on_request(req);
 276                 nfs_release_request(req);
 277                 if (ret != 0)
 278                         return ret;
 279                 spin_lock(req_lock);
 280         }
 281         spin_unlock(req_lock);
 282         if (test_and_set_bit(PG_FLUSHING, &req->wb_flags) == 0) {
 283                 nfs_mark_request_dirty(req);
 284                 set_page_writeback(page);
 285         }
 286         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 287         nfs_unlock_request(req);
 288         return ret;
 289 }
 290
 291 /*
 292  * Write an mmapped page to the server.
 293  */
 294 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 295 {
 296         struct nfs_open_context *ctx;
 297         struct inode *inode = page->mapping->host;
 298         unsigned offset;
 299         int err;
 300
 301         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 302         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 303
 304         err = nfs_page_mark_flush(page);
 305         if (err <= 0)
 306                 goto out;
 307         err = 0;
 308         offset = nfs_page_length(page);
 309         if (!offset)
 310                 goto out;
 311
 312         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 313         if (ctx == NULL) {
 314                 err = -EBADF;
 315                 goto out;
 316         }
 317         err = nfs_writepage_setup(ctx, page, 0, offset);
 318         put_nfs_open_context(ctx);
 319         if (err != 0)
 320                 goto out;
 321         err = nfs_page_mark_flush(page);
 322         if (err > 0)
 323                 err = 0;
 324 out:
 325         if (!wbc->for_writepages)
 326                 nfs_flush_mapping(page->mapping, wbc, wb_priority(wbc));
 327         return err;
 328 }
 329
 330 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 331 {
 332         int err;
 333
 334         err = nfs_writepage_locked(page, wbc);
 335         unlock_page(page);
 336         return err;
 337 }
 338
 339 /*
 340  * Note: causes nfs_update_request() to block on the assumption
 341  *       that the writeback is generated due to memory pressure.
 342  */
 343 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 344 {
 345         struct backing_dev_info *bdi = mapping->backing_dev_info;
 346         struct inode *inode = mapping->host;
 347         int err;
 348
 349         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 350
 351         err = generic_writepages(mapping, wbc);
 352         if (err)
 353                 return err;
 354         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 355                 if (wbc->nonblocking)
 356                         return 0;
 357                 nfs_wait_on_write_congestion(mapping, 0);
 358         }
 359         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 360         if (err < 0)
 361                 goto out;
 362         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 363         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 364                 err = nfs_wait_on_requests(inode, 0, 0);
 365                 if (err < 0)
 366                         goto out;
 367         }
 368         err = nfs_commit_inode(inode, wb_priority(wbc));
 369         if (err > 0)
 370                 err = 0;
 371 out:
 372         clear_bit(BDI_write_congested, &bdi->state);
 373         wake_up_all(&nfs_write_congestion);
 374         congestion_end(WRITE);
 375         return err;
 376 }
 377
 378 /*
 379  * Insert a write request into an inode
 380  */
 381 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 382 {
 383         struct nfs_inode *nfsi = NFS_I(inode);
 384         int error;
 385
 386         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 387         BUG_ON(error == -EEXIST);
 388         if (error)
 389                 return error;
 390         if (!nfsi->npages) {
 391                 igrab(inode);
 392                 nfs_begin_data_update(inode);
 393                 if (nfs_have_delegation(inode, FMODE_WRITE))
 394                         nfsi->change_attr++;
 395         }
 396         SetPagePrivate(req->wb_page);
 397         set_page_private(req->wb_page, (unsigned long)req);
 398         nfsi->npages++;
 399         atomic_inc(&req->wb_count);
 400         return 0;
 401 }
 402
 403 /*
 404  * Insert a write request into an inode
 405  */
 406 static void nfs_inode_remove_request(struct nfs_page *req)
 407 {
 408         struct inode *inode = req->wb_context->dentry->d_inode;
 409         struct nfs_inode *nfsi = NFS_I(inode);
 410
 411         BUG_ON (!NFS_WBACK_BUSY(req));
 412
 413         spin_lock(&nfsi->req_lock);
 414         set_page_private(req->wb_page, 0);
 415         ClearPagePrivate(req->wb_page);
 416         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 417         nfsi->npages--;
 418         if (!nfsi->npages) {
 419                 spin_unlock(&nfsi->req_lock);
 420                 nfs_end_data_update(inode);
 421                 iput(inode);
 422         } else
 423                 spin_unlock(&nfsi->req_lock);
 424         nfs_clear_request(req);
 425         nfs_release_request(req);
 426 }
 427
 428 /*
 429  * Add a request to the inode's dirty list.
 430  */
 431 static void
 432 nfs_mark_request_dirty(struct nfs_page *req)
 433 {
 434         struct inode *inode = req->wb_context->dentry->d_inode;
 435         struct nfs_inode *nfsi = NFS_I(inode);
 436
 437         spin_lock(&nfsi->req_lock);
 438         radix_tree_tag_set(&nfsi->nfs_page_tree,
 439                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 440         nfs_list_add_request(req, &nfsi->dirty);
 441         nfsi->ndirty++;
 442         spin_unlock(&nfsi->req_lock);
 443         __mark_inode_dirty(inode, I_DIRTY_PAGES);
 444 }
 445
 446 static void
 447 nfs_redirty_request(struct nfs_page *req)
 448 {
 449         clear_bit(PG_FLUSHING, &req->wb_flags);
 450         __set_page_dirty_nobuffers(req->wb_page);
 451 }
 452
 453 /*
 454  * Check if a request is dirty
 455  */
 456 static inline int
 457 nfs_dirty_request(struct nfs_page *req)
 458 {
 459         return test_bit(PG_FLUSHING, &req->wb_flags) == 0;
 460 }
 461
 462 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 463 /*
 464  * Add a request to the inode's commit list.
 465  */
 466 static void
 467 nfs_mark_request_commit(struct nfs_page *req)
 468 {
 469         struct inode *inode = req->wb_context->dentry->d_inode;
 470         struct nfs_inode *nfsi = NFS_I(inode);
 471
 472         spin_lock(&nfsi->req_lock);
 473         nfs_list_add_request(req, &nfsi->commit);
 474         nfsi->ncommit++;
 475         spin_unlock(&nfsi->req_lock);
 476         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 477         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 478 }
 479 #endif
 480
 481 /*
 482  * Wait for a request to complete.
 483  *
 484  * Interruptible by signals only if mounted with intr flag.
 485  */
 486 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 487 {
 488         struct nfs_inode *nfsi = NFS_I(inode);
 489         struct nfs_page *req;
 490         unsigned long           idx_end, next;
 491         unsigned int            res = 0;
 492         int                     error;
 493
 494         if (npages == 0)
 495                 idx_end = ~0;
 496         else
 497                 idx_end = idx_start + npages - 1;
 498
 499         next = idx_start;
 500         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 501                 if (req->wb_index > idx_end)
 502                         break;
 503
 504                 next = req->wb_index + 1;
 505                 BUG_ON(!NFS_WBACK_BUSY(req));
 506
 507                 atomic_inc(&req->wb_count);
 508                 spin_unlock(&nfsi->req_lock);
 509                 error = nfs_wait_on_request(req);
 510                 nfs_release_request(req);
 511                 spin_lock(&nfsi->req_lock);
 512                 if (error < 0)
 513                         return error;
 514                 res++;
 515         }
 516         return res;
 517 }
 518
 519 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 520 {
 521         struct nfs_inode *nfsi = NFS_I(inode);
 522         int ret;
 523
 524         spin_lock(&nfsi->req_lock);
 525         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 526         spin_unlock(&nfsi->req_lock);
 527         return ret;
 528 }
 529
 530 static void nfs_cancel_dirty_list(struct list_head *head)
 531 {
 532         struct nfs_page *req;
 533         while(!list_empty(head)) {
 534                 req = nfs_list_entry(head->next);
 535                 nfs_list_remove_request(req);
 536                 nfs_inode_remove_request(req);
 537                 nfs_clear_page_writeback(req);
 538         }
 539 }
 540
 541 static void nfs_cancel_commit_list(struct list_head *head)
 542 {
 543         struct nfs_page *req;
 544
 545         while(!list_empty(head)) {
 546                 req = nfs_list_entry(head->next);
 547                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 548                 nfs_list_remove_request(req);
 549                 nfs_inode_remove_request(req);
 550                 nfs_unlock_request(req);
 551         }
 552 }
 553
 554 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 555 /*
 556  * nfs_scan_commit - Scan an inode for commit requests
 557  * @inode: NFS inode to scan
 558  * @dst: destination list
 559  * @idx_start: lower bound of page->index to scan.
 560  * @npages: idx_start + npages sets the upper bound to scan.
 561  *
 562  * Moves requests from the inode's 'commit' request list.
 563  * The requests are *not* checked to ensure that they form a contiguous set.
 564  */
 565 static int
 566 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 567 {
 568         struct nfs_inode *nfsi = NFS_I(inode);
 569         int res = 0;
 570
 571         if (nfsi->ncommit != 0) {
 572                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 573                 nfsi->ncommit -= res;
 574                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 575                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 576         }
 577         return res;
 578 }
 579 #else
 580 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 581 {
 582         return 0;
 583 }
 584 #endif
 585
 586 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 587 {
 588         struct backing_dev_info *bdi = mapping->backing_dev_info;
 589         DEFINE_WAIT(wait);
 590         int ret = 0;
 591
 592         might_sleep();
 593
 594         if (!bdi_write_congested(bdi))
 595                 return 0;
 596
 597         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 598
 599         if (intr) {
 600                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 601                 sigset_t oldset;
 602
 603                 rpc_clnt_sigmask(clnt, &oldset);
 604                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 605                 if (bdi_write_congested(bdi)) {
 606                         if (signalled())
 607                                 ret = -ERESTARTSYS;
 608                         else
 609                                 schedule();
 610                 }
 611                 rpc_clnt_sigunmask(clnt, &oldset);
 612         } else {
 613                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 614                 if (bdi_write_congested(bdi))
 615                         schedule();
 616         }
 617         finish_wait(&nfs_write_congestion, &wait);
 618         return ret;
 619 }
 620
 621
 622 /*
 623  * Try to update any existing write request, or create one if there is none.
 624  * In order to match, the request's credentials must match those of
 625  * the calling process.
 626  *
 627  * Note: Should always be called with the Page Lock held!
 628  */
 629 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 630                 struct page *page, unsigned int offset, unsigned int bytes)
 631 {
 632         struct inode *inode = page->mapping->host;
 633         struct nfs_inode *nfsi = NFS_I(inode);
 634         struct nfs_page         *req, *new = NULL;
 635         unsigned long           rqend, end;
 636
 637         end = offset + bytes;
 638
 639         if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
 640                 return ERR_PTR(-ERESTARTSYS);
 641         for (;;) {
 642                 /* Loop over all inode entries and see if we find
 643                  * A request for the page we wish to update
 644                  */
 645                 spin_lock(&nfsi->req_lock);
 646                 req = nfs_page_find_request_locked(page);
 647                 if (req) {
 648                         if (!nfs_lock_request_dontget(req)) {
 649                                 int error;
 650
 651                                 spin_unlock(&nfsi->req_lock);
 652                                 error = nfs_wait_on_request(req);
 653                                 nfs_release_request(req);
 654                                 if (error < 0) {
 655                                         if (new)
 656                                                 nfs_release_request(new);
 657                                         return ERR_PTR(error);
 658                                 }
 659                                 continue;
 660                         }
 661                         spin_unlock(&nfsi->req_lock);
 662                         if (new)
 663                                 nfs_release_request(new);
 664                         break;
 665                 }
 666
 667                 if (new) {
 668                         int error;
 669                         nfs_lock_request_dontget(new);
 670                         error = nfs_inode_add_request(inode, new);
 671                         if (error) {
 672                                 spin_unlock(&nfsi->req_lock);
 673                                 nfs_unlock_request(new);
 674                                 return ERR_PTR(error);
 675                         }
 676                         spin_unlock(&nfsi->req_lock);
 677                         return new;
 678                 }
 679                 spin_unlock(&nfsi->req_lock);
 680
 681                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 682                 if (IS_ERR(new))
 683                         return new;
 684         }
 685
 686         /* We have a request for our page.
 687          * If the creds don't match, or the
 688          * page addresses don't match,
 689          * tell the caller to wait on the conflicting
 690          * request.
 691          */
 692         rqend = req->wb_offset + req->wb_bytes;
 693         if (req->wb_context != ctx
 694             || req->wb_page != page
 695             || !nfs_dirty_request(req)
 696             || offset > rqend || end < req->wb_offset) {
 697                 nfs_unlock_request(req);
 698                 return ERR_PTR(-EBUSY);
 699         }
 700
 701         /* Okay, the request matches. Update the region */
 702         if (offset < req->wb_offset) {
 703                 req->wb_offset = offset;
 704                 req->wb_pgbase = offset;
 705                 req->wb_bytes = rqend - req->wb_offset;
 706         }
 707
 708         if (end > rqend)
 709                 req->wb_bytes = end - req->wb_offset;
 710
 711         return req;
 712 }
 713
 714 int nfs_flush_incompatible(struct file *file, struct page *page)
 715 {
 716         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 717         struct nfs_page *req;
 718         int do_flush, status;
 719         /*
 720          * Look for a request corresponding to this page. If there
 721          * is one, and it belongs to another file, we flush it out
 722          * before we try to copy anything into the page. Do this
 723          * due to the lack of an ACCESS-type call in NFSv2.
 724          * Also do the same if we find a request from an existing
 725          * dropped page.
 726          */
 727         do {
 728                 req = nfs_page_find_request(page);
 729                 if (req == NULL)
 730                         return 0;
 731                 do_flush = req->wb_page != page || req->wb_context != ctx
 732                         || !nfs_dirty_request(req);
 733                 nfs_release_request(req);
 734                 if (!do_flush)
 735                         return 0;
 736                 status = nfs_wb_page(page->mapping->host, page);
 737         } while (status == 0);
 738         return status;
 739 }
 740
 741 /*
 742  * Update and possibly write a cached page of an NFS file.
 743  *
 744  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 745  * things with a page scheduled for an RPC call (e.g. invalidate it).
 746  */
 747 int nfs_updatepage(struct file *file, struct page *page,
 748                 unsigned int offset, unsigned int count)
 749 {
 750         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 751         struct inode    *inode = page->mapping->host;
 752         int             status = 0;
 753
 754         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 755
 756         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 757                 file->f_path.dentry->d_parent->d_name.name,
 758                 file->f_path.dentry->d_name.name, count,
 759                 (long long)(page_offset(page) +offset));
 760
 761         /* If we're not using byte range locks, and we know the page
 762          * is entirely in cache, it may be more efficient to avoid
 763          * fragmenting write requests.
 764          */
 765         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 766                 count = max(count + offset, nfs_page_length(page));
 767                 offset = 0;
 768         }
 769
 770         status = nfs_writepage_setup(ctx, page, offset, count);
 771         __set_page_dirty_nobuffers(page);
 772
 773         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 774                         status, (long long)i_size_read(inode));
 775         if (status < 0)
 776                 ClearPageUptodate(page);
 777         return status;
 778 }
 779
 780 static void nfs_writepage_release(struct nfs_page *req)
 781 {
 782         end_page_writeback(req->wb_page);
 783
 784 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 785         if (!PageError(req->wb_page)) {
 786                 if (NFS_NEED_RESCHED(req)) {
 787                         nfs_redirty_request(req);
 788                         goto out;
 789                 } else if (NFS_NEED_COMMIT(req)) {
 790                         nfs_mark_request_commit(req);
 791                         goto out;
 792                 }
 793         }
 794         nfs_inode_remove_request(req);
 795
 796 out:
 797         nfs_clear_commit(req);
 798         nfs_clear_reschedule(req);
 799 #else
 800         nfs_inode_remove_request(req);
 801 #endif
 802         nfs_clear_page_writeback(req);
 803 }
 804
 805 static inline int flush_task_priority(int how)
 806 {
 807         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 808                 case FLUSH_HIGHPRI:
 809                         return RPC_PRIORITY_HIGH;
 810                 case FLUSH_LOWPRI:
 811                         return RPC_PRIORITY_LOW;
 812         }
 813         return RPC_PRIORITY_NORMAL;
 814 }
 815
 816 /*
 817  * Set up the argument/result storage required for the RPC call.
 818  */
 819 static void nfs_write_rpcsetup(struct nfs_page *req,
 820                 struct nfs_write_data *data,
 821                 const struct rpc_call_ops *call_ops,
 822                 unsigned int count, unsigned int offset,
 823                 int how)
 824 {
 825         struct inode            *inode;
 826         int flags;
 827
 828         /* Set up the RPC argument and reply structs
 829          * NB: take care not to mess about with data->commit et al. */
 830
 831         data->req = req;
 832         data->inode = inode = req->wb_context->dentry->d_inode;
 833         data->cred = req->wb_context->cred;
 834
 835         data->args.fh     = NFS_FH(inode);
 836         data->args.offset = req_offset(req) + offset;
 837         data->args.pgbase = req->wb_pgbase + offset;
 838         data->args.pages  = data->pagevec;
 839         data->args.count  = count;
 840         data->args.context = req->wb_context;
 841
 842         data->res.fattr   = &data->fattr;
 843         data->res.count   = count;
 844         data->res.verf    = &data->verf;
 845         nfs_fattr_init(&data->fattr);
 846
 847         /* Set up the initial task struct.  */
 848         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 849         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 850         NFS_PROTO(inode)->write_setup(data, how);
 851
 852         data->task.tk_priority = flush_task_priority(how);
 853         data->task.tk_cookie = (unsigned long)inode;
 854
 855         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 856                 data->task.tk_pid,
 857                 inode->i_sb->s_id,
 858                 (long long)NFS_FILEID(inode),
 859                 count,
 860                 (unsigned long long)data->args.offset);
 861 }
 862
 863 static void nfs_execute_write(struct nfs_write_data *data)
 864 {
 865         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 866         sigset_t oldset;
 867
 868         rpc_clnt_sigmask(clnt, &oldset);
 869         rpc_execute(&data->task);
 870         rpc_clnt_sigunmask(clnt, &oldset);
 871 }
 872
 873 /*
 874  * Generate multiple small requests to write out a single
 875  * contiguous dirty area on one page.
 876  */
 877 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 878 {
 879         struct nfs_page *req = nfs_list_entry(head->next);
 880         struct page *page = req->wb_page;
 881         struct nfs_write_data *data;
 882         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 883         unsigned int offset;
 884         int requests = 0;
 885         LIST_HEAD(list);
 886
 887         nfs_list_remove_request(req);
 888
 889         nbytes = req->wb_bytes;
 890         do {
 891                 size_t len = min(nbytes, wsize);
 892
 893                 data = nfs_writedata_alloc(len);
 894                 if (!data)
 895                         goto out_bad;
 896                 list_add(&data->pages, &list);
 897                 requests++;
 898                 nbytes -= len;
 899         } while (nbytes != 0);
 900         atomic_set(&req->wb_complete, requests);
 901
 902         ClearPageError(page);
 903         offset = 0;
 904         nbytes = req->wb_bytes;
 905         do {
 906                 data = list_entry(list.next, struct nfs_write_data, pages);
 907                 list_del_init(&data->pages);
 908
 909                 data->pagevec[0] = page;
 910
 911                 if (nbytes > wsize) {
 912                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 913                                         wsize, offset, how);
 914                         offset += wsize;
 915                         nbytes -= wsize;
 916                 } else {
 917                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 918                                         nbytes, offset, how);
 919                         nbytes = 0;
 920                 }
 921                 nfs_execute_write(data);
 922         } while (nbytes != 0);
 923
 924         return 0;
 925
 926 out_bad:
 927         while (!list_empty(&list)) {
 928                 data = list_entry(list.next, struct nfs_write_data, pages);
 929                 list_del(&data->pages);
 930                 nfs_writedata_release(data);
 931         }
 932         nfs_redirty_request(req);
 933         nfs_clear_page_writeback(req);
 934         return -ENOMEM;
 935 }
 936
 937 /*
 938  * Create an RPC task for the given write request and kick it.
 939  * The page must have been locked by the caller.
 940  *
 941  * It may happen that the page we're passed is not marked dirty.
 942  * This is the case if nfs_updatepage detects a conflicting request
 943  * that has been written but not committed.
 944  */
 945 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 946 {
 947         struct nfs_page         *req;
 948         struct page             **pages;
 949         struct nfs_write_data   *data;
 950         unsigned int            count;
 951
 952         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 953         if (!data)
 954                 goto out_bad;
 955
 956         pages = data->pagevec;
 957         count = 0;
 958         while (!list_empty(head)) {
 959                 req = nfs_list_entry(head->next);
 960                 nfs_list_remove_request(req);
 961                 nfs_list_add_request(req, &data->pages);
 962                 ClearPageError(req->wb_page);
 963                 *pages++ = req->wb_page;
 964                 count += req->wb_bytes;
 965         }
 966         req = nfs_list_entry(data->pages.next);
 967
 968         /* Set up the argument struct */
 969         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 970
 971         nfs_execute_write(data);
 972         return 0;
 973  out_bad:
 974         while (!list_empty(head)) {
 975                 struct nfs_page *req = nfs_list_entry(head->next);
 976                 nfs_list_remove_request(req);
 977                 nfs_redirty_request(req);
 978                 nfs_clear_page_writeback(req);
 979         }
 980         return -ENOMEM;
 981 }
 982
 983 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 984 {
 985         LIST_HEAD(one_request);
 986         int (*flush_one)(struct inode *, struct list_head *, int);
 987         struct nfs_page *req;
 988         int wpages = NFS_SERVER(inode)->wpages;
 989         int wsize = NFS_SERVER(inode)->wsize;
 990         int error;
 991
 992         flush_one = nfs_flush_one;
 993         if (wsize < PAGE_CACHE_SIZE)
 994                 flush_one = nfs_flush_multi;
 995         /* For single writes, FLUSH_STABLE is more efficient */
 996         if (npages <= wpages && npages == NFS_I(inode)->npages
 997                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 998                 how |= FLUSH_STABLE;
 999
1000         do {
1001                 nfs_coalesce_requests(head, &one_request, wpages);
1002                 req = nfs_list_entry(one_request.next);
1003                 error = flush_one(inode, &one_request, how);
1004                 if (error < 0)
1005                         goto out_err;
1006         } while (!list_empty(head));
1007         return 0;
1008 out_err:
1009         while (!list_empty(head)) {
1010                 req = nfs_list_entry(head->next);
1011                 nfs_list_remove_request(req);
1012                 nfs_redirty_request(req);
1013                 nfs_clear_page_writeback(req);
1014         }
1015         return error;
1016 }
1017
1018 /*
1019  * Handle a write reply that flushed part of a page.
1020  */
1021 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1022 {
1023         struct nfs_write_data   *data = calldata;
1024         struct nfs_page         *req = data->req;
1025         struct page             *page = req->wb_page;
1026
1027         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1028                 req->wb_context->dentry->d_inode->i_sb->s_id,
1029                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1030                 req->wb_bytes,
1031                 (long long)req_offset(req));
1032
1033         if (nfs_writeback_done(task, data) != 0)
1034                 return;
1035
1036         if (task->tk_status < 0) {
1037                 ClearPageUptodate(page);
1038                 SetPageError(page);
1039                 req->wb_context->error = task->tk_status;
1040                 dprintk(", error = %d\n", task->tk_status);
1041         } else {
1042 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1043                 if (data->verf.committed < NFS_FILE_SYNC) {
1044                         if (!NFS_NEED_COMMIT(req)) {
1045                                 nfs_defer_commit(req);
1046                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1047                                 dprintk(" defer commit\n");
1048                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1049                                 nfs_defer_reschedule(req);
1050                                 dprintk(" server reboot detected\n");
1051                         }
1052                 } else
1053 #endif
1054                         dprintk(" OK\n");
1055         }
1056
1057         if (atomic_dec_and_test(&req->wb_complete))
1058                 nfs_writepage_release(req);
1059 }
1060
1061 static const struct rpc_call_ops nfs_write_partial_ops = {
1062         .rpc_call_done = nfs_writeback_done_partial,
1063         .rpc_release = nfs_writedata_release,
1064 };
1065
1066 /*
1067  * Handle a write reply that flushes a whole page.
1068  *
1069  * FIXME: There is an inherent race with invalidate_inode_pages and
1070  *        writebacks since the page->count is kept > 1 for as long
1071  *        as the page has a write request pending.
1072  */
1073 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1074 {
1075         struct nfs_write_data   *data = calldata;
1076         struct nfs_page         *req;
1077         struct page             *page;
1078
1079         if (nfs_writeback_done(task, data) != 0)
1080                 return;
1081
1082         /* Update attributes as result of writeback. */
1083         while (!list_empty(&data->pages)) {
1084                 req = nfs_list_entry(data->pages.next);
1085                 nfs_list_remove_request(req);
1086                 page = req->wb_page;
1087
1088                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1089                         req->wb_context->dentry->d_inode->i_sb->s_id,
1090                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1091                         req->wb_bytes,
1092                         (long long)req_offset(req));
1093
1094                 if (task->tk_status < 0) {
1095                         ClearPageUptodate(page);
1096                         SetPageError(page);
1097                         req->wb_context->error = task->tk_status;
1098                         end_page_writeback(page);
1099                         nfs_inode_remove_request(req);
1100                         dprintk(", error = %d\n", task->tk_status);
1101                         goto next;
1102                 }
1103                 end_page_writeback(page);
1104
1105 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1106                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1107                         nfs_inode_remove_request(req);
1108                         dprintk(" OK\n");
1109                         goto next;
1110                 }
1111                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1112                 nfs_mark_request_commit(req);
1113                 dprintk(" marked for commit\n");
1114 #else
1115                 nfs_inode_remove_request(req);
1116 #endif
1117         next:
1118                 nfs_clear_page_writeback(req);
1119         }
1120 }
1121
1122 static const struct rpc_call_ops nfs_write_full_ops = {
1123         .rpc_call_done = nfs_writeback_done_full,
1124         .rpc_release = nfs_writedata_release,
1125 };
1126
1127
1128 /*
1129  * This function is called when the WRITE call is complete.
1130  */
1131 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1132 {
1133         struct nfs_writeargs    *argp = &data->args;
1134         struct nfs_writeres     *resp = &data->res;
1135         int status;
1136
1137         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1138                 task->tk_pid, task->tk_status);
1139
1140         /*
1141          * ->write_done will attempt to use post-op attributes to detect
1142          * conflicting writes by other clients.  A strict interpretation
1143          * of close-to-open would allow us to continue caching even if
1144          * another writer had changed the file, but some applications
1145          * depend on tighter cache coherency when writing.
1146          */
1147         status = NFS_PROTO(data->inode)->write_done(task, data);
1148         if (status != 0)
1149                 return status;
1150         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1151
1152 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1153         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1154                 /* We tried a write call, but the server did not
1155                  * commit data to stable storage even though we
1156                  * requested it.
1157                  * Note: There is a known bug in Tru64 < 5.0 in which
1158                  *       the server reports NFS_DATA_SYNC, but performs
1159                  *       NFS_FILE_SYNC. We therefore implement this checking
1160                  *       as a dprintk() in order to avoid filling syslog.
1161                  */
1162                 static unsigned long    complain;
1163
1164                 if (time_before(complain, jiffies)) {
1165                         dprintk("NFS: faulty NFS server %s:"
1166                                 " (committed = %d) != (stable = %d)\n",
1167                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1168                                 resp->verf->committed, argp->stable);
1169                         complain = jiffies + 300 * HZ;
1170                 }
1171         }
1172 #endif
1173         /* Is this a short write? */
1174         if (task->tk_status >= 0 && resp->count < argp->count) {
1175                 static unsigned long    complain;
1176
1177                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1178
1179                 /* Has the server at least made some progress? */
1180                 if (resp->count != 0) {
1181                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1182                         if (resp->verf->committed != NFS_UNSTABLE) {
1183                                 /* Resend from where the server left off */
1184                                 argp->offset += resp->count;
1185                                 argp->pgbase += resp->count;
1186                                 argp->count -= resp->count;
1187                         } else {
1188                                 /* Resend as a stable write in order to avoid
1189                                  * headaches in the case of a server crash.
1190                                  */
1191                                 argp->stable = NFS_FILE_SYNC;
1192                         }
1193                         rpc_restart_call(task);
1194                         return -EAGAIN;
1195                 }
1196                 if (time_before(complain, jiffies)) {
1197                         printk(KERN_WARNING
1198                                "NFS: Server wrote zero bytes, expected %u.\n",
1199                                         argp->count);
1200                         complain = jiffies + 300 * HZ;
1201                 }
1202                 /* Can't do anything about it except throw an error. */
1203                 task->tk_status = -EIO;
1204         }
1205         return 0;
1206 }
1207
1208
1209 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1210 void nfs_commit_release(void *wdata)
1211 {
1212         nfs_commit_free(wdata);
1213 }
1214
1215 /*
1216  * Set up the argument/result storage required for the RPC call.
1217  */
1218 static void nfs_commit_rpcsetup(struct list_head *head,
1219                 struct nfs_write_data *data,
1220                 int how)
1221 {
1222         struct nfs_page         *first;
1223         struct inode            *inode;
1224         int flags;
1225
1226         /* Set up the RPC argument and reply structs
1227          * NB: take care not to mess about with data->commit et al. */
1228
1229         list_splice_init(head, &data->pages);
1230         first = nfs_list_entry(data->pages.next);
1231         inode = first->wb_context->dentry->d_inode;
1232
1233         data->inode       = inode;
1234         data->cred        = first->wb_context->cred;
1235
1236         data->args.fh     = NFS_FH(data->inode);
1237         /* Note: we always request a commit of the entire inode */
1238         data->args.offset = 0;
1239         data->args.count  = 0;
1240         data->res.count   = 0;
1241         data->res.fattr   = &data->fattr;
1242         data->res.verf    = &data->verf;
1243         nfs_fattr_init(&data->fattr);
1244
1245         /* Set up the initial task struct.  */
1246         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1247         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1248         NFS_PROTO(inode)->commit_setup(data, how);
1249
1250         data->task.tk_priority = flush_task_priority(how);
1251         data->task.tk_cookie = (unsigned long)inode;
1252
1253         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1254 }
1255
1256 /*
1257  * Commit dirty pages
1258  */
1259 static int
1260 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1261 {
1262         struct nfs_write_data   *data;
1263         struct nfs_page         *req;
1264
1265         data = nfs_commit_alloc();
1266
1267         if (!data)
1268                 goto out_bad;
1269
1270         /* Set up the argument struct */
1271         nfs_commit_rpcsetup(head, data, how);
1272
1273         nfs_execute_write(data);
1274         return 0;
1275  out_bad:
1276         while (!list_empty(head)) {
1277                 req = nfs_list_entry(head->next);
1278                 nfs_list_remove_request(req);
1279                 nfs_mark_request_commit(req);
1280                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1281                 nfs_clear_page_writeback(req);
1282         }
1283         return -ENOMEM;
1284 }
1285
1286 /*
1287  * COMMIT call returned
1288  */
1289 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1290 {
1291         struct nfs_write_data   *data = calldata;
1292         struct nfs_page         *req;
1293
1294         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1295                                 task->tk_pid, task->tk_status);
1296
1297         /* Call the NFS version-specific code */
1298         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1299                 return;
1300
1301         while (!list_empty(&data->pages)) {
1302                 req = nfs_list_entry(data->pages.next);
1303                 nfs_list_remove_request(req);
1304                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1305
1306                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1307                         req->wb_context->dentry->d_inode->i_sb->s_id,
1308                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1309                         req->wb_bytes,
1310                         (long long)req_offset(req));
1311                 if (task->tk_status < 0) {
1312                         req->wb_context->error = task->tk_status;
1313                         nfs_inode_remove_request(req);
1314                         dprintk(", error = %d\n", task->tk_status);
1315                         goto next;
1316                 }
1317
1318                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1319                  * returned by the server against all stored verfs. */
1320                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1321                         /* We have a match */
1322                         nfs_inode_remove_request(req);
1323                         dprintk(" OK\n");
1324                         goto next;
1325                 }
1326                 /* We have a mismatch. Write the page again */
1327                 dprintk(" mismatch\n");
1328                 nfs_redirty_request(req);
1329         next:
1330                 nfs_clear_page_writeback(req);
1331         }
1332 }
1333
1334 static const struct rpc_call_ops nfs_commit_ops = {
1335         .rpc_call_done = nfs_commit_done,
1336         .rpc_release = nfs_commit_release,
1337 };
1338 #else
1339 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1340 {
1341         return 0;
1342 }
1343 #endif
1344
1345 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1346 {
1347         struct nfs_inode *nfsi = NFS_I(mapping->host);
1348         LIST_HEAD(head);
1349         long res;
1350
1351         spin_lock(&nfsi->req_lock);
1352         res = nfs_scan_dirty(mapping, wbc, &head);
1353         spin_unlock(&nfsi->req_lock);
1354         if (res) {
1355                 int error = nfs_flush_list(mapping->host, &head, res, how);
1356                 if (error < 0)
1357                         return error;
1358         }
1359         return res;
1360 }
1361
1362 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1363 int nfs_commit_inode(struct inode *inode, int how)
1364 {
1365         struct nfs_inode *nfsi = NFS_I(inode);
1366         LIST_HEAD(head);
1367         int res;
1368
1369         spin_lock(&nfsi->req_lock);
1370         res = nfs_scan_commit(inode, &head, 0, 0);
1371         spin_unlock(&nfsi->req_lock);
1372         if (res) {
1373                 int error = nfs_commit_list(inode, &head, how);
1374                 if (error < 0)
1375                         return error;
1376         }
1377         return res;
1378 }
1379 #endif
1380
1381 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1382 {
1383         struct inode *inode = mapping->host;
1384         struct nfs_inode *nfsi = NFS_I(inode);
1385         unsigned long idx_start, idx_end;
1386         unsigned int npages = 0;
1387         LIST_HEAD(head);
1388         int nocommit = how & FLUSH_NOCOMMIT;
1389         long pages, ret;
1390
1391         /* FIXME */
1392         if (wbc->range_cyclic)
1393                 idx_start = 0;
1394         else {
1395                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1396                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1397                 if (idx_end > idx_start) {
1398                         unsigned long l_npages = 1 + idx_end - idx_start;
1399                         npages = l_npages;
1400                         if (sizeof(npages) != sizeof(l_npages) &&
1401                                         (unsigned long)npages != l_npages)
1402                                 npages = 0;
1403                 }
1404         }
1405         how &= ~FLUSH_NOCOMMIT;
1406         spin_lock(&nfsi->req_lock);
1407         do {
1408                 wbc->pages_skipped = 0;
1409                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1410                 if (ret != 0)
1411                         continue;
1412                 pages = nfs_scan_dirty(mapping, wbc, &head);
1413                 if (pages != 0) {
1414                         spin_unlock(&nfsi->req_lock);
1415                         if (how & FLUSH_INVALIDATE) {
1416                                 nfs_cancel_dirty_list(&head);
1417                                 ret = pages;
1418                         } else
1419                                 ret = nfs_flush_list(inode, &head, pages, how);
1420                         spin_lock(&nfsi->req_lock);
1421                         continue;
1422                 }
1423                 if (wbc->pages_skipped != 0)
1424                         continue;
1425                 if (nocommit)
1426                         break;
1427                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1428                 if (pages == 0) {
1429                         if (wbc->pages_skipped != 0)
1430                                 continue;
1431                         break;
1432                 }
1433                 if (how & FLUSH_INVALIDATE) {
1434                         spin_unlock(&nfsi->req_lock);
1435                         nfs_cancel_commit_list(&head);
1436                         ret = pages;
1437                         spin_lock(&nfsi->req_lock);
1438                         continue;
1439                 }
1440                 pages += nfs_scan_commit(inode, &head, 0, 0);
1441                 spin_unlock(&nfsi->req_lock);
1442                 ret = nfs_commit_list(inode, &head, how);
1443                 spin_lock(&nfsi->req_lock);
1444         } while (ret >= 0);
1445         spin_unlock(&nfsi->req_lock);
1446         return ret;
1447 }
1448
1449 /*
1450  * flush the inode to disk.
1451  */
1452 int nfs_wb_all(struct inode *inode)
1453 {
1454         struct address_space *mapping = inode->i_mapping;
1455         struct writeback_control wbc = {
1456                 .bdi = mapping->backing_dev_info,
1457                 .sync_mode = WB_SYNC_ALL,
1458                 .nr_to_write = LONG_MAX,
1459                 .for_writepages = 1,
1460                 .range_cyclic = 1,
1461         };
1462         int ret;
1463
1464         ret = generic_writepages(mapping, &wbc);
1465         if (ret < 0)
1466                 goto out;
1467         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1468         if (ret >= 0)
1469                 return 0;
1470 out:
1471         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1472         return ret;
1473 }
1474
1475 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1476 {
1477         struct writeback_control wbc = {
1478                 .bdi = mapping->backing_dev_info,
1479                 .sync_mode = WB_SYNC_ALL,
1480                 .nr_to_write = LONG_MAX,
1481                 .range_start = range_start,
1482                 .range_end = range_end,
1483                 .for_writepages = 1,
1484         };
1485         int ret;
1486
1487         if (!(how & FLUSH_NOWRITEPAGE)) {
1488                 ret = generic_writepages(mapping, &wbc);
1489                 if (ret < 0)
1490                         goto out;
1491         }
1492         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1493         if (ret >= 0)
1494                 return 0;
1495 out:
1496         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1497         return ret;
1498 }
1499
1500 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1501 {
1502         loff_t range_start = page_offset(page);
1503         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1504         struct writeback_control wbc = {
1505                 .bdi = page->mapping->backing_dev_info,
1506                 .sync_mode = WB_SYNC_ALL,
1507                 .nr_to_write = LONG_MAX,
1508                 .range_start = range_start,
1509                 .range_end = range_end,
1510         };
1511         int ret;
1512
1513         BUG_ON(!PageLocked(page));
1514         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1515                 ret = nfs_writepage_locked(page, &wbc);
1516                 if (ret < 0)
1517                         goto out;
1518         }
1519         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1520         if (ret >= 0)
1521                 return 0;
1522 out:
1523         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1524         return ret;
1525 }
1526
1527 /*
1528  * Write back all requests on one page - we do this before reading it.
1529  */
1530 int nfs_wb_page(struct inode *inode, struct page* page)
1531 {
1532         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1533 }
1534
1535 int nfs_set_page_dirty(struct page *page)
1536 {
1537         struct nfs_page *req;
1538
1539         req = nfs_page_find_request(page);
1540         if (req != NULL) {
1541                 /* Mark any existing write requests for flushing */
1542                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1543                 nfs_release_request(req);
1544         }
1545         return __set_page_dirty_nobuffers(page);
1546 }
1547
1548
1549 int __init nfs_init_writepagecache(void)
1550 {
1551         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1552                                              sizeof(struct nfs_write_data),
1553                                              0, SLAB_HWCACHE_ALIGN,
1554                                              NULL, NULL);
1555         if (nfs_wdata_cachep == NULL)
1556                 return -ENOMEM;
1557
1558         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1559                                                      nfs_wdata_cachep);
1560         if (nfs_wdata_mempool == NULL)
1561                 return -ENOMEM;
1562
1563         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1564                                                       nfs_wdata_cachep);
1565         if (nfs_commit_mempool == NULL)
1566                 return -ENOMEM;
1567
1568         return 0;
1569 }
1570
1571 void nfs_destroy_writepagecache(void)
1572 {
1573         mempool_destroy(nfs_commit_mempool);
1574         mempool_destroy(nfs_wdata_mempool);
1575         kmem_cache_destroy(nfs_wdata_cachep);
1576 }
1577