SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 kmem_cache_t *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, SLAB_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, SLAB_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         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 444         mark_inode_dirty(inode);
 445 }
 446
 447 static void
 448 nfs_redirty_request(struct nfs_page *req)
 449 {
 450         clear_bit(PG_FLUSHING, &req->wb_flags);
 451         __set_page_dirty_nobuffers(req->wb_page);
 452 }
 453
 454 /*
 455  * Check if a request is dirty
 456  */
 457 static inline int
 458 nfs_dirty_request(struct nfs_page *req)
 459 {
 460         return test_bit(PG_FLUSHING, &req->wb_flags) == 0;
 461 }
 462
 463 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 464 /*
 465  * Add a request to the inode's commit list.
 466  */
 467 static void
 468 nfs_mark_request_commit(struct nfs_page *req)
 469 {
 470         struct inode *inode = req->wb_context->dentry->d_inode;
 471         struct nfs_inode *nfsi = NFS_I(inode);
 472
 473         spin_lock(&nfsi->req_lock);
 474         nfs_list_add_request(req, &nfsi->commit);
 475         nfsi->ncommit++;
 476         spin_unlock(&nfsi->req_lock);
 477         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 478         mark_inode_dirty(inode);
 479 }
 480 #endif
 481
 482 /*
 483  * Wait for a request to complete.
 484  *
 485  * Interruptible by signals only if mounted with intr flag.
 486  */
 487 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 488 {
 489         struct nfs_inode *nfsi = NFS_I(inode);
 490         struct nfs_page *req;
 491         unsigned long           idx_end, next;
 492         unsigned int            res = 0;
 493         int                     error;
 494
 495         if (npages == 0)
 496                 idx_end = ~0;
 497         else
 498                 idx_end = idx_start + npages - 1;
 499
 500         next = idx_start;
 501         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 502                 if (req->wb_index > idx_end)
 503                         break;
 504
 505                 next = req->wb_index + 1;
 506                 BUG_ON(!NFS_WBACK_BUSY(req));
 507
 508                 atomic_inc(&req->wb_count);
 509                 spin_unlock(&nfsi->req_lock);
 510                 error = nfs_wait_on_request(req);
 511                 nfs_release_request(req);
 512                 spin_lock(&nfsi->req_lock);
 513                 if (error < 0)
 514                         return error;
 515                 res++;
 516         }
 517         return res;
 518 }
 519
 520 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 521 {
 522         struct nfs_inode *nfsi = NFS_I(inode);
 523         int ret;
 524
 525         spin_lock(&nfsi->req_lock);
 526         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 527         spin_unlock(&nfsi->req_lock);
 528         return ret;
 529 }
 530
 531 static void nfs_cancel_dirty_list(struct list_head *head)
 532 {
 533         struct nfs_page *req;
 534         while(!list_empty(head)) {
 535                 req = nfs_list_entry(head->next);
 536                 nfs_list_remove_request(req);
 537                 nfs_inode_remove_request(req);
 538                 nfs_clear_page_writeback(req);
 539         }
 540 }
 541
 542 static void nfs_cancel_commit_list(struct list_head *head)
 543 {
 544         struct nfs_page *req;
 545
 546         while(!list_empty(head)) {
 547                 req = nfs_list_entry(head->next);
 548                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 549                 nfs_list_remove_request(req);
 550                 nfs_inode_remove_request(req);
 551                 nfs_unlock_request(req);
 552         }
 553 }
 554
 555 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 556 /*
 557  * nfs_scan_commit - Scan an inode for commit requests
 558  * @inode: NFS inode to scan
 559  * @dst: destination list
 560  * @idx_start: lower bound of page->index to scan.
 561  * @npages: idx_start + npages sets the upper bound to scan.
 562  *
 563  * Moves requests from the inode's 'commit' request list.
 564  * The requests are *not* checked to ensure that they form a contiguous set.
 565  */
 566 static int
 567 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 568 {
 569         struct nfs_inode *nfsi = NFS_I(inode);
 570         int res = 0;
 571
 572         if (nfsi->ncommit != 0) {
 573                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 574                 nfsi->ncommit -= res;
 575                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 576                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 577         }
 578         return res;
 579 }
 580 #else
 581 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 582 {
 583         return 0;
 584 }
 585 #endif
 586
 587 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 588 {
 589         struct backing_dev_info *bdi = mapping->backing_dev_info;
 590         DEFINE_WAIT(wait);
 591         int ret = 0;
 592
 593         might_sleep();
 594
 595         if (!bdi_write_congested(bdi))
 596                 return 0;
 597
 598         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 599
 600         if (intr) {
 601                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 602                 sigset_t oldset;
 603
 604                 rpc_clnt_sigmask(clnt, &oldset);
 605                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 606                 if (bdi_write_congested(bdi)) {
 607                         if (signalled())
 608                                 ret = -ERESTARTSYS;
 609                         else
 610                                 schedule();
 611                 }
 612                 rpc_clnt_sigunmask(clnt, &oldset);
 613         } else {
 614                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 615                 if (bdi_write_congested(bdi))
 616                         schedule();
 617         }
 618         finish_wait(&nfs_write_congestion, &wait);
 619         return ret;
 620 }
 621
 622
 623 /*
 624  * Try to update any existing write request, or create one if there is none.
 625  * In order to match, the request's credentials must match those of
 626  * the calling process.
 627  *
 628  * Note: Should always be called with the Page Lock held!
 629  */
 630 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 631                 struct page *page, unsigned int offset, unsigned int bytes)
 632 {
 633         struct inode *inode = page->mapping->host;
 634         struct nfs_inode *nfsi = NFS_I(inode);
 635         struct nfs_page         *req, *new = NULL;
 636         unsigned long           rqend, end;
 637
 638         end = offset + bytes;
 639
 640         if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
 641                 return ERR_PTR(-ERESTARTSYS);
 642         for (;;) {
 643                 /* Loop over all inode entries and see if we find
 644                  * A request for the page we wish to update
 645                  */
 646                 spin_lock(&nfsi->req_lock);
 647                 req = nfs_page_find_request_locked(page);
 648                 if (req) {
 649                         if (!nfs_lock_request_dontget(req)) {
 650                                 int error;
 651
 652                                 spin_unlock(&nfsi->req_lock);
 653                                 error = nfs_wait_on_request(req);
 654                                 nfs_release_request(req);
 655                                 if (error < 0) {
 656                                         if (new)
 657                                                 nfs_release_request(new);
 658                                         return ERR_PTR(error);
 659                                 }
 660                                 continue;
 661                         }
 662                         spin_unlock(&nfsi->req_lock);
 663                         if (new)
 664                                 nfs_release_request(new);
 665                         break;
 666                 }
 667
 668                 if (new) {
 669                         int error;
 670                         nfs_lock_request_dontget(new);
 671                         error = nfs_inode_add_request(inode, new);
 672                         if (error) {
 673                                 spin_unlock(&nfsi->req_lock);
 674                                 nfs_unlock_request(new);
 675                                 return ERR_PTR(error);
 676                         }
 677                         spin_unlock(&nfsi->req_lock);
 678                         return new;
 679                 }
 680                 spin_unlock(&nfsi->req_lock);
 681
 682                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 683                 if (IS_ERR(new))
 684                         return new;
 685         }
 686
 687         /* We have a request for our page.
 688          * If the creds don't match, or the
 689          * page addresses don't match,
 690          * tell the caller to wait on the conflicting
 691          * request.
 692          */
 693         rqend = req->wb_offset + req->wb_bytes;
 694         if (req->wb_context != ctx
 695             || req->wb_page != page
 696             || !nfs_dirty_request(req)
 697             || offset > rqend || end < req->wb_offset) {
 698                 nfs_unlock_request(req);
 699                 return ERR_PTR(-EBUSY);
 700         }
 701
 702         /* Okay, the request matches. Update the region */
 703         if (offset < req->wb_offset) {
 704                 req->wb_offset = offset;
 705                 req->wb_pgbase = offset;
 706                 req->wb_bytes = rqend - req->wb_offset;
 707         }
 708
 709         if (end > rqend)
 710                 req->wb_bytes = end - req->wb_offset;
 711
 712         return req;
 713 }
 714
 715 int nfs_flush_incompatible(struct file *file, struct page *page)
 716 {
 717         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 718         struct nfs_page *req;
 719         int do_flush, status;
 720         /*
 721          * Look for a request corresponding to this page. If there
 722          * is one, and it belongs to another file, we flush it out
 723          * before we try to copy anything into the page. Do this
 724          * due to the lack of an ACCESS-type call in NFSv2.
 725          * Also do the same if we find a request from an existing
 726          * dropped page.
 727          */
 728         do {
 729                 req = nfs_page_find_request(page);
 730                 if (req == NULL)
 731                         return 0;
 732                 do_flush = req->wb_page != page || req->wb_context != ctx
 733                         || !nfs_dirty_request(req);
 734                 nfs_release_request(req);
 735                 if (!do_flush)
 736                         return 0;
 737                 status = nfs_wb_page(page->mapping->host, page);
 738         } while (status == 0);
 739         return status;
 740 }
 741
 742 /*
 743  * Update and possibly write a cached page of an NFS file.
 744  *
 745  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 746  * things with a page scheduled for an RPC call (e.g. invalidate it).
 747  */
 748 int nfs_updatepage(struct file *file, struct page *page,
 749                 unsigned int offset, unsigned int count)
 750 {
 751         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 752         struct inode    *inode = page->mapping->host;
 753         int             status = 0;
 754
 755         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 756
 757         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 758                 file->f_dentry->d_parent->d_name.name,
 759                 file->f_dentry->d_name.name, count,
 760                 (long long)(page_offset(page) +offset));
 761
 762         /* If we're not using byte range locks, and we know the page
 763          * is entirely in cache, it may be more efficient to avoid
 764          * fragmenting write requests.
 765          */
 766         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 767                 count = max(count + offset, nfs_page_length(page));
 768                 offset = 0;
 769         }
 770
 771         status = nfs_writepage_setup(ctx, page, offset, count);
 772         __set_page_dirty_nobuffers(page);
 773
 774         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 775                         status, (long long)i_size_read(inode));
 776         if (status < 0)
 777                 ClearPageUptodate(page);
 778         return status;
 779 }
 780
 781 static void nfs_writepage_release(struct nfs_page *req)
 782 {
 783         end_page_writeback(req->wb_page);
 784
 785 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 786         if (!PageError(req->wb_page)) {
 787                 if (NFS_NEED_RESCHED(req)) {
 788                         nfs_redirty_request(req);
 789                         goto out;
 790                 } else if (NFS_NEED_COMMIT(req)) {
 791                         nfs_mark_request_commit(req);
 792                         goto out;
 793                 }
 794         }
 795         nfs_inode_remove_request(req);
 796
 797 out:
 798         nfs_clear_commit(req);
 799         nfs_clear_reschedule(req);
 800 #else
 801         nfs_inode_remove_request(req);
 802 #endif
 803         nfs_clear_page_writeback(req);
 804 }
 805
 806 static inline int flush_task_priority(int how)
 807 {
 808         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 809                 case FLUSH_HIGHPRI:
 810                         return RPC_PRIORITY_HIGH;
 811                 case FLUSH_LOWPRI:
 812                         return RPC_PRIORITY_LOW;
 813         }
 814         return RPC_PRIORITY_NORMAL;
 815 }
 816
 817 /*
 818  * Set up the argument/result storage required for the RPC call.
 819  */
 820 static void nfs_write_rpcsetup(struct nfs_page *req,
 821                 struct nfs_write_data *data,
 822                 const struct rpc_call_ops *call_ops,
 823                 unsigned int count, unsigned int offset,
 824                 int how)
 825 {
 826         struct inode            *inode;
 827         int flags;
 828
 829         /* Set up the RPC argument and reply structs
 830          * NB: take care not to mess about with data->commit et al. */
 831
 832         data->req = req;
 833         data->inode = inode = req->wb_context->dentry->d_inode;
 834         data->cred = req->wb_context->cred;
 835
 836         data->args.fh     = NFS_FH(inode);
 837         data->args.offset = req_offset(req) + offset;
 838         data->args.pgbase = req->wb_pgbase + offset;
 839         data->args.pages  = data->pagevec;
 840         data->args.count  = count;
 841         data->args.context = req->wb_context;
 842
 843         data->res.fattr   = &data->fattr;
 844         data->res.count   = count;
 845         data->res.verf    = &data->verf;
 846         nfs_fattr_init(&data->fattr);
 847
 848         /* Set up the initial task struct.  */
 849         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 850         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 851         NFS_PROTO(inode)->write_setup(data, how);
 852
 853         data->task.tk_priority = flush_task_priority(how);
 854         data->task.tk_cookie = (unsigned long)inode;
 855
 856         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 857                 data->task.tk_pid,
 858                 inode->i_sb->s_id,
 859                 (long long)NFS_FILEID(inode),
 860                 count,
 861                 (unsigned long long)data->args.offset);
 862 }
 863
 864 static void nfs_execute_write(struct nfs_write_data *data)
 865 {
 866         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 867         sigset_t oldset;
 868
 869         rpc_clnt_sigmask(clnt, &oldset);
 870         rpc_execute(&data->task);
 871         rpc_clnt_sigunmask(clnt, &oldset);
 872 }
 873
 874 /*
 875  * Generate multiple small requests to write out a single
 876  * contiguous dirty area on one page.
 877  */
 878 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 879 {
 880         struct nfs_page *req = nfs_list_entry(head->next);
 881         struct page *page = req->wb_page;
 882         struct nfs_write_data *data;
 883         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 884         unsigned int offset;
 885         int requests = 0;
 886         LIST_HEAD(list);
 887
 888         nfs_list_remove_request(req);
 889
 890         nbytes = req->wb_bytes;
 891         do {
 892                 size_t len = min(nbytes, wsize);
 893
 894                 data = nfs_writedata_alloc(len);
 895                 if (!data)
 896                         goto out_bad;
 897                 list_add(&data->pages, &list);
 898                 requests++;
 899                 nbytes -= len;
 900         } while (nbytes != 0);
 901         atomic_set(&req->wb_complete, requests);
 902
 903         ClearPageError(page);
 904         offset = 0;
 905         nbytes = req->wb_bytes;
 906         do {
 907                 data = list_entry(list.next, struct nfs_write_data, pages);
 908                 list_del_init(&data->pages);
 909
 910                 data->pagevec[0] = page;
 911
 912                 if (nbytes > wsize) {
 913                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 914                                         wsize, offset, how);
 915                         offset += wsize;
 916                         nbytes -= wsize;
 917                 } else {
 918                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 919                                         nbytes, offset, how);
 920                         nbytes = 0;
 921                 }
 922                 nfs_execute_write(data);
 923         } while (nbytes != 0);
 924
 925         return 0;
 926
 927 out_bad:
 928         while (!list_empty(&list)) {
 929                 data = list_entry(list.next, struct nfs_write_data, pages);
 930                 list_del(&data->pages);
 931                 nfs_writedata_release(data);
 932         }
 933         nfs_redirty_request(req);
 934         nfs_clear_page_writeback(req);
 935         return -ENOMEM;
 936 }
 937
 938 /*
 939  * Create an RPC task for the given write request and kick it.
 940  * The page must have been locked by the caller.
 941  *
 942  * It may happen that the page we're passed is not marked dirty.
 943  * This is the case if nfs_updatepage detects a conflicting request
 944  * that has been written but not committed.
 945  */
 946 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 947 {
 948         struct nfs_page         *req;
 949         struct page             **pages;
 950         struct nfs_write_data   *data;
 951         unsigned int            count;
 952
 953         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 954         if (!data)
 955                 goto out_bad;
 956
 957         pages = data->pagevec;
 958         count = 0;
 959         while (!list_empty(head)) {
 960                 req = nfs_list_entry(head->next);
 961                 nfs_list_remove_request(req);
 962                 nfs_list_add_request(req, &data->pages);
 963                 ClearPageError(req->wb_page);
 964                 *pages++ = req->wb_page;
 965                 count += req->wb_bytes;
 966         }
 967         req = nfs_list_entry(data->pages.next);
 968
 969         /* Set up the argument struct */
 970         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 971
 972         nfs_execute_write(data);
 973         return 0;
 974  out_bad:
 975         while (!list_empty(head)) {
 976                 struct nfs_page *req = nfs_list_entry(head->next);
 977                 nfs_list_remove_request(req);
 978                 nfs_redirty_request(req);
 979                 nfs_clear_page_writeback(req);
 980         }
 981         return -ENOMEM;
 982 }
 983
 984 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 985 {
 986         LIST_HEAD(one_request);
 987         int (*flush_one)(struct inode *, struct list_head *, int);
 988         struct nfs_page *req;
 989         int wpages = NFS_SERVER(inode)->wpages;
 990         int wsize = NFS_SERVER(inode)->wsize;
 991         int error;
 992
 993         flush_one = nfs_flush_one;
 994         if (wsize < PAGE_CACHE_SIZE)
 995                 flush_one = nfs_flush_multi;
 996         /* For single writes, FLUSH_STABLE is more efficient */
 997         if (npages <= wpages && npages == NFS_I(inode)->npages
 998                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 999                 how |= FLUSH_STABLE;
1000
1001         do {
1002                 nfs_coalesce_requests(head, &one_request, wpages);
1003                 req = nfs_list_entry(one_request.next);
1004                 error = flush_one(inode, &one_request, how);
1005                 if (error < 0)
1006                         goto out_err;
1007         } while (!list_empty(head));
1008         return 0;
1009 out_err:
1010         while (!list_empty(head)) {
1011                 req = nfs_list_entry(head->next);
1012                 nfs_list_remove_request(req);
1013                 nfs_redirty_request(req);
1014                 nfs_clear_page_writeback(req);
1015         }
1016         return error;
1017 }
1018
1019 /*
1020  * Handle a write reply that flushed part of a page.
1021  */
1022 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1023 {
1024         struct nfs_write_data   *data = calldata;
1025         struct nfs_page         *req = data->req;
1026         struct page             *page = req->wb_page;
1027
1028         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1029                 req->wb_context->dentry->d_inode->i_sb->s_id,
1030                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1031                 req->wb_bytes,
1032                 (long long)req_offset(req));
1033
1034         if (nfs_writeback_done(task, data) != 0)
1035                 return;
1036
1037         if (task->tk_status < 0) {
1038                 ClearPageUptodate(page);
1039                 SetPageError(page);
1040                 req->wb_context->error = task->tk_status;
1041                 dprintk(", error = %d\n", task->tk_status);
1042         } else {
1043 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1044                 if (data->verf.committed < NFS_FILE_SYNC) {
1045                         if (!NFS_NEED_COMMIT(req)) {
1046                                 nfs_defer_commit(req);
1047                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1048                                 dprintk(" defer commit\n");
1049                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1050                                 nfs_defer_reschedule(req);
1051                                 dprintk(" server reboot detected\n");
1052                         }
1053                 } else
1054 #endif
1055                         dprintk(" OK\n");
1056         }
1057
1058         if (atomic_dec_and_test(&req->wb_complete))
1059                 nfs_writepage_release(req);
1060 }
1061
1062 static const struct rpc_call_ops nfs_write_partial_ops = {
1063         .rpc_call_done = nfs_writeback_done_partial,
1064         .rpc_release = nfs_writedata_release,
1065 };
1066
1067 /*
1068  * Handle a write reply that flushes a whole page.
1069  *
1070  * FIXME: There is an inherent race with invalidate_inode_pages and
1071  *        writebacks since the page->count is kept > 1 for as long
1072  *        as the page has a write request pending.
1073  */
1074 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1075 {
1076         struct nfs_write_data   *data = calldata;
1077         struct nfs_page         *req;
1078         struct page             *page;
1079
1080         if (nfs_writeback_done(task, data) != 0)
1081                 return;
1082
1083         /* Update attributes as result of writeback. */
1084         while (!list_empty(&data->pages)) {
1085                 req = nfs_list_entry(data->pages.next);
1086                 nfs_list_remove_request(req);
1087                 page = req->wb_page;
1088
1089                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1090                         req->wb_context->dentry->d_inode->i_sb->s_id,
1091                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1092                         req->wb_bytes,
1093                         (long long)req_offset(req));
1094
1095                 if (task->tk_status < 0) {
1096                         ClearPageUptodate(page);
1097                         SetPageError(page);
1098                         req->wb_context->error = task->tk_status;
1099                         end_page_writeback(page);
1100                         nfs_inode_remove_request(req);
1101                         dprintk(", error = %d\n", task->tk_status);
1102                         goto next;
1103                 }
1104                 end_page_writeback(page);
1105
1106 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1107                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1108                         nfs_inode_remove_request(req);
1109                         dprintk(" OK\n");
1110                         goto next;
1111                 }
1112                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1113                 nfs_mark_request_commit(req);
1114                 dprintk(" marked for commit\n");
1115 #else
1116                 nfs_inode_remove_request(req);
1117 #endif
1118         next:
1119                 nfs_clear_page_writeback(req);
1120         }
1121 }
1122
1123 static const struct rpc_call_ops nfs_write_full_ops = {
1124         .rpc_call_done = nfs_writeback_done_full,
1125         .rpc_release = nfs_writedata_release,
1126 };
1127
1128
1129 /*
1130  * This function is called when the WRITE call is complete.
1131  */
1132 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1133 {
1134         struct nfs_writeargs    *argp = &data->args;
1135         struct nfs_writeres     *resp = &data->res;
1136         int status;
1137
1138         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1139                 task->tk_pid, task->tk_status);
1140
1141         /*
1142          * ->write_done will attempt to use post-op attributes to detect
1143          * conflicting writes by other clients.  A strict interpretation
1144          * of close-to-open would allow us to continue caching even if
1145          * another writer had changed the file, but some applications
1146          * depend on tighter cache coherency when writing.
1147          */
1148         status = NFS_PROTO(data->inode)->write_done(task, data);
1149         if (status != 0)
1150                 return status;
1151         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1152
1153 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1154         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1155                 /* We tried a write call, but the server did not
1156                  * commit data to stable storage even though we
1157                  * requested it.
1158                  * Note: There is a known bug in Tru64 < 5.0 in which
1159                  *       the server reports NFS_DATA_SYNC, but performs
1160                  *       NFS_FILE_SYNC. We therefore implement this checking
1161                  *       as a dprintk() in order to avoid filling syslog.
1162                  */
1163                 static unsigned long    complain;
1164
1165                 if (time_before(complain, jiffies)) {
1166                         dprintk("NFS: faulty NFS server %s:"
1167                                 " (committed = %d) != (stable = %d)\n",
1168                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1169                                 resp->verf->committed, argp->stable);
1170                         complain = jiffies + 300 * HZ;
1171                 }
1172         }
1173 #endif
1174         /* Is this a short write? */
1175         if (task->tk_status >= 0 && resp->count < argp->count) {
1176                 static unsigned long    complain;
1177
1178                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1179
1180                 /* Has the server at least made some progress? */
1181                 if (resp->count != 0) {
1182                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1183                         if (resp->verf->committed != NFS_UNSTABLE) {
1184                                 /* Resend from where the server left off */
1185                                 argp->offset += resp->count;
1186                                 argp->pgbase += resp->count;
1187                                 argp->count -= resp->count;
1188                         } else {
1189                                 /* Resend as a stable write in order to avoid
1190                                  * headaches in the case of a server crash.
1191                                  */
1192                                 argp->stable = NFS_FILE_SYNC;
1193                         }
1194                         rpc_restart_call(task);
1195                         return -EAGAIN;
1196                 }
1197                 if (time_before(complain, jiffies)) {
1198                         printk(KERN_WARNING
1199                                "NFS: Server wrote zero bytes, expected %u.\n",
1200                                         argp->count);
1201                         complain = jiffies + 300 * HZ;
1202                 }
1203                 /* Can't do anything about it except throw an error. */
1204                 task->tk_status = -EIO;
1205         }
1206         return 0;
1207 }
1208
1209
1210 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1211 void nfs_commit_release(void *wdata)
1212 {
1213         nfs_commit_free(wdata);
1214 }
1215
1216 /*
1217  * Set up the argument/result storage required for the RPC call.
1218  */
1219 static void nfs_commit_rpcsetup(struct list_head *head,
1220                 struct nfs_write_data *data,
1221                 int how)
1222 {
1223         struct nfs_page         *first;
1224         struct inode            *inode;
1225         int flags;
1226
1227         /* Set up the RPC argument and reply structs
1228          * NB: take care not to mess about with data->commit et al. */
1229
1230         list_splice_init(head, &data->pages);
1231         first = nfs_list_entry(data->pages.next);
1232         inode = first->wb_context->dentry->d_inode;
1233
1234         data->inode       = inode;
1235         data->cred        = first->wb_context->cred;
1236
1237         data->args.fh     = NFS_FH(data->inode);
1238         /* Note: we always request a commit of the entire inode */
1239         data->args.offset = 0;
1240         data->args.count  = 0;
1241         data->res.count   = 0;
1242         data->res.fattr   = &data->fattr;
1243         data->res.verf    = &data->verf;
1244         nfs_fattr_init(&data->fattr);
1245
1246         /* Set up the initial task struct.  */
1247         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1248         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1249         NFS_PROTO(inode)->commit_setup(data, how);
1250
1251         data->task.tk_priority = flush_task_priority(how);
1252         data->task.tk_cookie = (unsigned long)inode;
1253
1254         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1255 }
1256
1257 /*
1258  * Commit dirty pages
1259  */
1260 static int
1261 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1262 {
1263         struct nfs_write_data   *data;
1264         struct nfs_page         *req;
1265
1266         data = nfs_commit_alloc();
1267
1268         if (!data)
1269                 goto out_bad;
1270
1271         /* Set up the argument struct */
1272         nfs_commit_rpcsetup(head, data, how);
1273
1274         nfs_execute_write(data);
1275         return 0;
1276  out_bad:
1277         while (!list_empty(head)) {
1278                 req = nfs_list_entry(head->next);
1279                 nfs_list_remove_request(req);
1280                 nfs_mark_request_commit(req);
1281                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1282                 nfs_clear_page_writeback(req);
1283         }
1284         return -ENOMEM;
1285 }
1286
1287 /*
1288  * COMMIT call returned
1289  */
1290 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1291 {
1292         struct nfs_write_data   *data = calldata;
1293         struct nfs_page         *req;
1294
1295         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1296                                 task->tk_pid, task->tk_status);
1297
1298         /* Call the NFS version-specific code */
1299         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1300                 return;
1301
1302         while (!list_empty(&data->pages)) {
1303                 req = nfs_list_entry(data->pages.next);
1304                 nfs_list_remove_request(req);
1305                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1306
1307                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1308                         req->wb_context->dentry->d_inode->i_sb->s_id,
1309                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1310                         req->wb_bytes,
1311                         (long long)req_offset(req));
1312                 if (task->tk_status < 0) {
1313                         req->wb_context->error = task->tk_status;
1314                         nfs_inode_remove_request(req);
1315                         dprintk(", error = %d\n", task->tk_status);
1316                         goto next;
1317                 }
1318
1319                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1320                  * returned by the server against all stored verfs. */
1321                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1322                         /* We have a match */
1323                         nfs_inode_remove_request(req);
1324                         dprintk(" OK\n");
1325                         goto next;
1326                 }
1327                 /* We have a mismatch. Write the page again */
1328                 dprintk(" mismatch\n");
1329                 nfs_redirty_request(req);
1330         next:
1331                 nfs_clear_page_writeback(req);
1332         }
1333 }
1334
1335 static const struct rpc_call_ops nfs_commit_ops = {
1336         .rpc_call_done = nfs_commit_done,
1337         .rpc_release = nfs_commit_release,
1338 };
1339 #else
1340 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1341 {
1342         return 0;
1343 }
1344 #endif
1345
1346 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1347 {
1348         struct nfs_inode *nfsi = NFS_I(mapping->host);
1349         LIST_HEAD(head);
1350         long res;
1351
1352         spin_lock(&nfsi->req_lock);
1353         res = nfs_scan_dirty(mapping, wbc, &head);
1354         spin_unlock(&nfsi->req_lock);
1355         if (res) {
1356                 int error = nfs_flush_list(mapping->host, &head, res, how);
1357                 if (error < 0)
1358                         return error;
1359         }
1360         return res;
1361 }
1362
1363 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1364 int nfs_commit_inode(struct inode *inode, int how)
1365 {
1366         struct nfs_inode *nfsi = NFS_I(inode);
1367         LIST_HEAD(head);
1368         int res;
1369
1370         spin_lock(&nfsi->req_lock);
1371         res = nfs_scan_commit(inode, &head, 0, 0);
1372         spin_unlock(&nfsi->req_lock);
1373         if (res) {
1374                 int error = nfs_commit_list(inode, &head, how);
1375                 if (error < 0)
1376                         return error;
1377         }
1378         return res;
1379 }
1380 #endif
1381
1382 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1383 {
1384         struct inode *inode = mapping->host;
1385         struct nfs_inode *nfsi = NFS_I(inode);
1386         unsigned long idx_start, idx_end;
1387         unsigned int npages = 0;
1388         LIST_HEAD(head);
1389         int nocommit = how & FLUSH_NOCOMMIT;
1390         long pages, ret;
1391
1392         /* FIXME */
1393         if (wbc->range_cyclic)
1394                 idx_start = 0;
1395         else {
1396                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1397                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1398                 if (idx_end > idx_start) {
1399                         unsigned long l_npages = 1 + idx_end - idx_start;
1400                         npages = l_npages;
1401                         if (sizeof(npages) != sizeof(l_npages) &&
1402                                         (unsigned long)npages != l_npages)
1403                                 npages = 0;
1404                 }
1405         }
1406         how &= ~FLUSH_NOCOMMIT;
1407         spin_lock(&nfsi->req_lock);
1408         do {
1409                 wbc->pages_skipped = 0;
1410                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1411                 if (ret != 0)
1412                         continue;
1413                 pages = nfs_scan_dirty(mapping, wbc, &head);
1414                 if (pages != 0) {
1415                         spin_unlock(&nfsi->req_lock);
1416                         if (how & FLUSH_INVALIDATE) {
1417                                 nfs_cancel_dirty_list(&head);
1418                                 ret = pages;
1419                         } else
1420                                 ret = nfs_flush_list(inode, &head, pages, how);
1421                         spin_lock(&nfsi->req_lock);
1422                         continue;
1423                 }
1424                 if (wbc->pages_skipped != 0)
1425                         continue;
1426                 if (nocommit)
1427                         break;
1428                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1429                 if (pages == 0) {
1430                         if (wbc->pages_skipped != 0)
1431                                 continue;
1432                         break;
1433                 }
1434                 if (how & FLUSH_INVALIDATE) {
1435                         spin_unlock(&nfsi->req_lock);
1436                         nfs_cancel_commit_list(&head);
1437                         ret = pages;
1438                         spin_lock(&nfsi->req_lock);
1439                         continue;
1440                 }
1441                 pages += nfs_scan_commit(inode, &head, 0, 0);
1442                 spin_unlock(&nfsi->req_lock);
1443                 ret = nfs_commit_list(inode, &head, how);
1444                 spin_lock(&nfsi->req_lock);
1445         } while (ret >= 0);
1446         spin_unlock(&nfsi->req_lock);
1447         return ret;
1448 }
1449
1450 /*
1451  * flush the inode to disk.
1452  */
1453 int nfs_wb_all(struct inode *inode)
1454 {
1455         struct address_space *mapping = inode->i_mapping;
1456         struct writeback_control wbc = {
1457                 .bdi = mapping->backing_dev_info,
1458                 .sync_mode = WB_SYNC_ALL,
1459                 .nr_to_write = LONG_MAX,
1460                 .for_writepages = 1,
1461                 .range_cyclic = 1,
1462         };
1463         int ret;
1464
1465         ret = generic_writepages(mapping, &wbc);
1466         if (ret < 0)
1467                 goto out;
1468         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1469         if (ret >= 0)
1470                 return 0;
1471 out:
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         return ret;
1497 }
1498
1499 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1500 {
1501         loff_t range_start = page_offset(page);
1502         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1503         struct writeback_control wbc = {
1504                 .bdi = page->mapping->backing_dev_info,
1505                 .sync_mode = WB_SYNC_ALL,
1506                 .nr_to_write = LONG_MAX,
1507                 .range_start = range_start,
1508                 .range_end = range_end,
1509         };
1510         int ret;
1511
1512         BUG_ON(!PageLocked(page));
1513         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1514                 ret = nfs_writepage_locked(page, &wbc);
1515                 if (ret < 0)
1516                         goto out;
1517         }
1518         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1519         if (ret >= 0)
1520                 return 0;
1521 out:
1522         return ret;
1523 }
1524
1525 /*
1526  * Write back all requests on one page - we do this before reading it.
1527  */
1528 int nfs_wb_page(struct inode *inode, struct page* page)
1529 {
1530         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1531 }
1532
1533 int nfs_set_page_dirty(struct page *page)
1534 {
1535         struct nfs_page *req;
1536
1537         req = nfs_page_find_request(page);
1538         if (req != NULL) {
1539                 /* Mark any existing write requests for flushing */
1540                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1541                 nfs_release_request(req);
1542         }
1543         return __set_page_dirty_nobuffers(page);
1544 }
1545
1546
1547 int __init nfs_init_writepagecache(void)
1548 {
1549         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1550                                              sizeof(struct nfs_write_data),
1551                                              0, SLAB_HWCACHE_ALIGN,
1552                                              NULL, NULL);
1553         if (nfs_wdata_cachep == NULL)
1554                 return -ENOMEM;
1555
1556         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1557                                                      nfs_wdata_cachep);
1558         if (nfs_wdata_mempool == NULL)
1559                 return -ENOMEM;
1560
1561         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1562                                                       nfs_wdata_cachep);
1563         if (nfs_commit_mempool == NULL)
1564                 return -ENOMEM;
1565
1566         return 0;
1567 }
1568
1569 void nfs_destroy_writepagecache(void)
1570 {
1571         mempool_destroy(nfs_commit_mempool);
1572         mempool_destroy(nfs_wdata_mempool);
1573         kmem_cache_destroy(nfs_wdata_cachep);
1574 }
1575