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/config.h>
  50 #include <linux/types.h>
  51 #include <linux/slab.h>
  52 #include <linux/mm.h>
  53 #include <linux/pagemap.h>
  54 #include <linux/file.h>
  55 #include <linux/mpage.h>
  56 #include <linux/writeback.h>
  57
  58 #include <linux/sunrpc/clnt.h>
  59 #include <linux/nfs_fs.h>
  60 #include <linux/nfs_mount.h>
  61 #include <linux/nfs_page.h>
  62 #include <asm/uaccess.h>
  63 #include <linux/smp_lock.h>
  64
  65 #include "delegation.h"
  66
  67 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  68
  69 #define MIN_POOL_WRITE          (32)
  70 #define MIN_POOL_COMMIT         (4)
  71
  72 /*
  73  * Local function declarations
  74  */
  75 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  76                                             struct inode *,
  77                                             struct page *,
  78                                             unsigned int, unsigned int);
  79 static void nfs_writeback_done_partial(struct nfs_write_data *, int);
  80 static void nfs_writeback_done_full(struct nfs_write_data *, int);
  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 int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
  84                            unsigned int npages, int how);
  85
  86 static kmem_cache_t *nfs_wdata_cachep;
  87 mempool_t *nfs_wdata_mempool;
  88 static mempool_t *nfs_commit_mempool;
  89
  90 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  91
  92 static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
  93 {
  94         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
  95
  96         if (p) {
  97                 memset(p, 0, sizeof(*p));
  98                 INIT_LIST_HEAD(&p->pages);
  99                 if (pagecount < NFS_PAGEVEC_SIZE)
 100                         p->pagevec = &p->page_array[0];
 101                 else {
 102                         size_t size = ++pagecount * sizeof(struct page *);
 103                         p->pagevec = kmalloc(size, GFP_NOFS);
 104                         if (p->pagevec) {
 105                                 memset(p->pagevec, 0, size);
 106                         } else {
 107                                 mempool_free(p, nfs_commit_mempool);
 108                                 p = NULL;
 109                         }
 110                 }
 111         }
 112         return p;
 113 }
 114
 115 static inline void nfs_commit_free(struct nfs_write_data *p)
 116 {
 117         if (p && (p->pagevec != &p->page_array[0]))
 118                 kfree(p->pagevec);
 119         mempool_free(p, nfs_commit_mempool);
 120 }
 121
 122 void nfs_writedata_release(void *wdata)
 123 {
 124         nfs_writedata_free(wdata);
 125 }
 126
 127 /* Adjust the file length if we're writing beyond the end */
 128 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 129 {
 130         struct inode *inode = page->mapping->host;
 131         loff_t end, i_size = i_size_read(inode);
 132         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 133
 134         if (i_size > 0 && page->index < end_index)
 135                 return;
 136         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 137         if (i_size >= end)
 138                 return;
 139         i_size_write(inode, end);
 140 }
 141
 142 /* We can set the PG_uptodate flag if we see that a write request
 143  * covers the full page.
 144  */
 145 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 146 {
 147         loff_t end_offs;
 148
 149         if (PageUptodate(page))
 150                 return;
 151         if (base != 0)
 152                 return;
 153         if (count == PAGE_CACHE_SIZE) {
 154                 SetPageUptodate(page);
 155                 return;
 156         }
 157
 158         end_offs = i_size_read(page->mapping->host) - 1;
 159         if (end_offs < 0)
 160                 return;
 161         /* Is this the last page? */
 162         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
 163                 return;
 164         /* This is the last page: set PG_uptodate if we cover the entire
 165          * extent of the data, then zero the rest of the page.
 166          */
 167         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
 168                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 169                 SetPageUptodate(page);
 170         }
 171 }
 172
 173 /*
 174  * Write a page synchronously.
 175  * Offset is the data offset within the page.
 176  */
 177 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 178                 struct page *page, unsigned int offset, unsigned int count,
 179                 int how)
 180 {
 181         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 182         int             result, written = 0;
 183         struct nfs_write_data *wdata;
 184
 185         wdata = nfs_writedata_alloc(1);
 186         if (!wdata)
 187                 return -ENOMEM;
 188
 189         wdata->flags = how;
 190         wdata->cred = ctx->cred;
 191         wdata->inode = inode;
 192         wdata->args.fh = NFS_FH(inode);
 193         wdata->args.context = ctx;
 194         wdata->args.pages = &page;
 195         wdata->args.stable = NFS_FILE_SYNC;
 196         wdata->args.pgbase = offset;
 197         wdata->args.count = wsize;
 198         wdata->res.fattr = &wdata->fattr;
 199         wdata->res.verf = &wdata->verf;
 200
 201         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 202                 inode->i_sb->s_id,
 203                 (long long)NFS_FILEID(inode),
 204                 count, (long long)(page_offset(page) + offset));
 205
 206         set_page_writeback(page);
 207         nfs_begin_data_update(inode);
 208         do {
 209                 if (count < wsize)
 210                         wdata->args.count = count;
 211                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 212
 213                 result = NFS_PROTO(inode)->write(wdata);
 214
 215                 if (result < 0) {
 216                         /* Must mark the page invalid after I/O error */
 217                         ClearPageUptodate(page);
 218                         goto io_error;
 219                 }
 220                 if (result < wdata->args.count)
 221                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 222                                         wdata->args.count, result);
 223
 224                 wdata->args.offset += result;
 225                 wdata->args.pgbase += result;
 226                 written += result;
 227                 count -= result;
 228         } while (count);
 229         /* Update file length */
 230         nfs_grow_file(page, offset, written);
 231         /* Set the PG_uptodate flag? */
 232         nfs_mark_uptodate(page, offset, written);
 233
 234         if (PageError(page))
 235                 ClearPageError(page);
 236
 237 io_error:
 238         nfs_end_data_update(inode);
 239         end_page_writeback(page);
 240         nfs_writedata_free(wdata);
 241         return written ? written : result;
 242 }
 243
 244 static int nfs_writepage_async(struct nfs_open_context *ctx,
 245                 struct inode *inode, struct page *page,
 246                 unsigned int offset, unsigned int count)
 247 {
 248         struct nfs_page *req;
 249
 250         req = nfs_update_request(ctx, inode, page, offset, count);
 251         if (IS_ERR(req))
 252                 return PTR_ERR(req);
 253         /* Update file length */
 254         nfs_grow_file(page, offset, count);
 255         /* Set the PG_uptodate flag? */
 256         nfs_mark_uptodate(page, offset, count);
 257         nfs_unlock_request(req);
 258         return 0;
 259 }
 260
 261 static int wb_priority(struct writeback_control *wbc)
 262 {
 263         if (wbc->for_reclaim)
 264                 return FLUSH_HIGHPRI;
 265         if (wbc->for_kupdate)
 266                 return FLUSH_LOWPRI;
 267         return 0;
 268 }
 269
 270 /*
 271  * Write an mmapped page to the server.
 272  */
 273 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 274 {
 275         struct nfs_open_context *ctx;
 276         struct inode *inode = page->mapping->host;
 277         unsigned long end_index;
 278         unsigned offset = PAGE_CACHE_SIZE;
 279         loff_t i_size = i_size_read(inode);
 280         int inode_referenced = 0;
 281         int priority = wb_priority(wbc);
 282         int err;
 283
 284         /*
 285          * Note: We need to ensure that we have a reference to the inode
 286          *       if we are to do asynchronous writes. If not, waiting
 287          *       in nfs_wait_on_request() may deadlock with clear_inode().
 288          *
 289          *       If igrab() fails here, then it is in any case safe to
 290          *       call nfs_wb_page(), since there will be no pending writes.
 291          */
 292         if (igrab(inode) != 0)
 293                 inode_referenced = 1;
 294         end_index = i_size >> PAGE_CACHE_SHIFT;
 295
 296         /* Ensure we've flushed out any previous writes */
 297         nfs_wb_page_priority(inode, page, priority);
 298
 299         /* easy case */
 300         if (page->index < end_index)
 301                 goto do_it;
 302         /* things got complicated... */
 303         offset = i_size & (PAGE_CACHE_SIZE-1);
 304
 305         /* OK, are we completely out? */
 306         err = 0; /* potential race with truncate - ignore */
 307         if (page->index >= end_index+1 || !offset)
 308                 goto out;
 309 do_it:
 310         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 311         if (ctx == NULL) {
 312                 err = -EBADF;
 313                 goto out;
 314         }
 315         lock_kernel();
 316         if (!IS_SYNC(inode) && inode_referenced) {
 317                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
 318                 if (!wbc->for_writepages)
 319                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 320         } else {
 321                 err = nfs_writepage_sync(ctx, inode, page, 0,
 322                                                 offset, priority);
 323                 if (err >= 0) {
 324                         if (err != offset)
 325                                 redirty_page_for_writepage(wbc, page);
 326                         err = 0;
 327                 }
 328         }
 329         unlock_kernel();
 330         put_nfs_open_context(ctx);
 331 out:
 332         unlock_page(page);
 333         if (inode_referenced)
 334                 iput(inode);
 335         return err;
 336 }
 337
 338 /*
 339  * Note: causes nfs_update_request() to block on the assumption
 340  *       that the writeback is generated due to memory pressure.
 341  */
 342 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 343 {
 344         struct backing_dev_info *bdi = mapping->backing_dev_info;
 345         struct inode *inode = mapping->host;
 346         int err;
 347
 348         err = generic_writepages(mapping, wbc);
 349         if (err)
 350                 return err;
 351         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 352                 if (wbc->nonblocking)
 353                         return 0;
 354                 nfs_wait_on_write_congestion(mapping, 0);
 355         }
 356         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 357         if (err < 0)
 358                 goto out;
 359         wbc->nr_to_write -= err;
 360         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 361                 err = nfs_wait_on_requests(inode, 0, 0);
 362                 if (err < 0)
 363                         goto out;
 364         }
 365         err = nfs_commit_inode(inode, wb_priority(wbc));
 366         if (err > 0) {
 367                 wbc->nr_to_write -= err;
 368                 err = 0;
 369         }
 370 out:
 371         clear_bit(BDI_write_congested, &bdi->state);
 372         wake_up_all(&nfs_write_congestion);
 373         return err;
 374 }
 375
 376 /*
 377  * Insert a write request into an inode
 378  */
 379 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 380 {
 381         struct nfs_inode *nfsi = NFS_I(inode);
 382         int error;
 383
 384         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 385         BUG_ON(error == -EEXIST);
 386         if (error)
 387                 return error;
 388         if (!nfsi->npages) {
 389                 igrab(inode);
 390                 nfs_begin_data_update(inode);
 391                 if (nfs_have_delegation(inode, FMODE_WRITE))
 392                         nfsi->change_attr++;
 393         }
 394         nfsi->npages++;
 395         atomic_inc(&req->wb_count);
 396         return 0;
 397 }
 398
 399 /*
 400  * Insert a write request into an inode
 401  */
 402 static void nfs_inode_remove_request(struct nfs_page *req)
 403 {
 404         struct inode *inode = req->wb_context->dentry->d_inode;
 405         struct nfs_inode *nfsi = NFS_I(inode);
 406
 407         BUG_ON (!NFS_WBACK_BUSY(req));
 408
 409         spin_lock(&nfsi->req_lock);
 410         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 411         nfsi->npages--;
 412         if (!nfsi->npages) {
 413                 spin_unlock(&nfsi->req_lock);
 414                 nfs_end_data_update(inode);
 415                 iput(inode);
 416         } else
 417                 spin_unlock(&nfsi->req_lock);
 418         nfs_clear_request(req);
 419         nfs_release_request(req);
 420 }
 421
 422 /*
 423  * Find a request
 424  */
 425 static inline struct nfs_page *
 426 _nfs_find_request(struct inode *inode, unsigned long index)
 427 {
 428         struct nfs_inode *nfsi = NFS_I(inode);
 429         struct nfs_page *req;
 430
 431         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 432         if (req)
 433                 atomic_inc(&req->wb_count);
 434         return req;
 435 }
 436
 437 static struct nfs_page *
 438 nfs_find_request(struct inode *inode, unsigned long index)
 439 {
 440         struct nfs_page         *req;
 441         struct nfs_inode        *nfsi = NFS_I(inode);
 442
 443         spin_lock(&nfsi->req_lock);
 444         req = _nfs_find_request(inode, index);
 445         spin_unlock(&nfsi->req_lock);
 446         return req;
 447 }
 448
 449 /*
 450  * Add a request to the inode's dirty list.
 451  */
 452 static void
 453 nfs_mark_request_dirty(struct nfs_page *req)
 454 {
 455         struct inode *inode = req->wb_context->dentry->d_inode;
 456         struct nfs_inode *nfsi = NFS_I(inode);
 457
 458         spin_lock(&nfsi->req_lock);
 459         radix_tree_tag_set(&nfsi->nfs_page_tree,
 460                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 461         nfs_list_add_request(req, &nfsi->dirty);
 462         nfsi->ndirty++;
 463         spin_unlock(&nfsi->req_lock);
 464         inc_page_state(nr_dirty);
 465         mark_inode_dirty(inode);
 466 }
 467
 468 /*
 469  * Check if a request is dirty
 470  */
 471 static inline int
 472 nfs_dirty_request(struct nfs_page *req)
 473 {
 474         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 475         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 476 }
 477
 478 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 479 /*
 480  * Add a request to the inode's commit list.
 481  */
 482 static void
 483 nfs_mark_request_commit(struct nfs_page *req)
 484 {
 485         struct inode *inode = req->wb_context->dentry->d_inode;
 486         struct nfs_inode *nfsi = NFS_I(inode);
 487
 488         spin_lock(&nfsi->req_lock);
 489         nfs_list_add_request(req, &nfsi->commit);
 490         nfsi->ncommit++;
 491         spin_unlock(&nfsi->req_lock);
 492         inc_page_state(nr_unstable);
 493         mark_inode_dirty(inode);
 494 }
 495 #endif
 496
 497 /*
 498  * Wait for a request to complete.
 499  *
 500  * Interruptible by signals only if mounted with intr flag.
 501  */
 502 static int
 503 nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 504 {
 505         struct nfs_inode *nfsi = NFS_I(inode);
 506         struct nfs_page *req;
 507         unsigned long           idx_end, next;
 508         unsigned int            res = 0;
 509         int                     error;
 510
 511         if (npages == 0)
 512                 idx_end = ~0;
 513         else
 514                 idx_end = idx_start + npages - 1;
 515
 516         spin_lock(&nfsi->req_lock);
 517         next = idx_start;
 518         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 519                 if (req->wb_index > idx_end)
 520                         break;
 521
 522                 next = req->wb_index + 1;
 523                 BUG_ON(!NFS_WBACK_BUSY(req));
 524
 525                 atomic_inc(&req->wb_count);
 526                 spin_unlock(&nfsi->req_lock);
 527                 error = nfs_wait_on_request(req);
 528                 nfs_release_request(req);
 529                 if (error < 0)
 530                         return error;
 531                 spin_lock(&nfsi->req_lock);
 532                 res++;
 533         }
 534         spin_unlock(&nfsi->req_lock);
 535         return res;
 536 }
 537
 538 /*
 539  * nfs_scan_dirty - Scan an inode for dirty requests
 540  * @inode: NFS inode to scan
 541  * @dst: destination list
 542  * @idx_start: lower bound of page->index to scan.
 543  * @npages: idx_start + npages sets the upper bound to scan.
 544  *
 545  * Moves requests from the inode's dirty page list.
 546  * The requests are *not* checked to ensure that they form a contiguous set.
 547  */
 548 static int
 549 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 550 {
 551         struct nfs_inode *nfsi = NFS_I(inode);
 552         int res = 0;
 553
 554         if (nfsi->ndirty != 0) {
 555                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
 556                 nfsi->ndirty -= res;
 557                 sub_page_state(nr_dirty,res);
 558                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
 559                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
 560         }
 561         return res;
 562 }
 563
 564 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 565 /*
 566  * nfs_scan_commit - Scan an inode for commit requests
 567  * @inode: NFS inode to scan
 568  * @dst: destination list
 569  * @idx_start: lower bound of page->index to scan.
 570  * @npages: idx_start + npages sets the upper bound to scan.
 571  *
 572  * Moves requests from the inode's 'commit' request list.
 573  * The requests are *not* checked to ensure that they form a contiguous set.
 574  */
 575 static int
 576 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 577 {
 578         struct nfs_inode *nfsi = NFS_I(inode);
 579         int res = 0;
 580
 581         if (nfsi->ncommit != 0) {
 582                 res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);
 583                 nfsi->ncommit -= res;
 584                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 585                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 586         }
 587         return res;
 588 }
 589 #endif
 590
 591 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 592 {
 593         struct backing_dev_info *bdi = mapping->backing_dev_info;
 594         DEFINE_WAIT(wait);
 595         int ret = 0;
 596
 597         might_sleep();
 598
 599         if (!bdi_write_congested(bdi))
 600                 return 0;
 601         if (intr) {
 602                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 603                 sigset_t oldset;
 604
 605                 rpc_clnt_sigmask(clnt, &oldset);
 606                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 607                 if (bdi_write_congested(bdi)) {
 608                         if (signalled())
 609                                 ret = -ERESTARTSYS;
 610                         else
 611                                 schedule();
 612                 }
 613                 rpc_clnt_sigunmask(clnt, &oldset);
 614         } else {
 615                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 616                 if (bdi_write_congested(bdi))
 617                         schedule();
 618         }
 619         finish_wait(&nfs_write_congestion, &wait);
 620         return ret;
 621 }
 622
 623
 624 /*
 625  * Try to update any existing write request, or create one if there is none.
 626  * In order to match, the request's credentials must match those of
 627  * the calling process.
 628  *
 629  * Note: Should always be called with the Page Lock held!
 630  */
 631 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 632                 struct inode *inode, struct page *page,
 633                 unsigned int offset, unsigned int bytes)
 634 {
 635         struct nfs_server *server = NFS_SERVER(inode);
 636         struct nfs_inode *nfsi = NFS_I(inode);
 637         struct nfs_page         *req, *new = NULL;
 638         unsigned long           rqend, end;
 639
 640         end = offset + bytes;
 641
 642         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 643                 return ERR_PTR(-ERESTARTSYS);
 644         for (;;) {
 645                 /* Loop over all inode entries and see if we find
 646                  * A request for the page we wish to update
 647                  */
 648                 spin_lock(&nfsi->req_lock);
 649                 req = _nfs_find_request(inode, page->index);
 650                 if (req) {
 651                         if (!nfs_lock_request_dontget(req)) {
 652                                 int error;
 653                                 spin_unlock(&nfsi->req_lock);
 654                                 error = nfs_wait_on_request(req);
 655                                 nfs_release_request(req);
 656                                 if (error < 0)
 657                                         return ERR_PTR(error);
 658                                 continue;
 659                         }
 660                         spin_unlock(&nfsi->req_lock);
 661                         if (new)
 662                                 nfs_release_request(new);
 663                         break;
 664                 }
 665
 666                 if (new) {
 667                         int error;
 668                         nfs_lock_request_dontget(new);
 669                         error = nfs_inode_add_request(inode, new);
 670                         if (error) {
 671                                 spin_unlock(&nfsi->req_lock);
 672                                 nfs_unlock_request(new);
 673                                 return ERR_PTR(error);
 674                         }
 675                         spin_unlock(&nfsi->req_lock);
 676                         nfs_mark_request_dirty(new);
 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 inode    *inode = page->mapping->host;
 718         struct nfs_page *req;
 719         int             status = 0;
 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         req = nfs_find_request(inode, page->index);
 729         if (req) {
 730                 if (req->wb_page != page || ctx != req->wb_context)
 731                         status = nfs_wb_page(inode, page);
 732                 nfs_release_request(req);
 733         }
 734         return (status < 0) ? status : 0;
 735 }
 736
 737 /*
 738  * Update and possibly write a cached page of an NFS file.
 739  *
 740  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 741  * things with a page scheduled for an RPC call (e.g. invalidate it).
 742  */
 743 int nfs_updatepage(struct file *file, struct page *page,
 744                 unsigned int offset, unsigned int count)
 745 {
 746         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 747         struct inode    *inode = page->mapping->host;
 748         struct nfs_page *req;
 749         int             status = 0;
 750
 751         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 752                 file->f_dentry->d_parent->d_name.name,
 753                 file->f_dentry->d_name.name, count,
 754                 (long long)(page_offset(page) +offset));
 755
 756         if (IS_SYNC(inode)) {
 757                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 758                 if (status > 0) {
 759                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 760                                 SetPageUptodate(page);
 761                         return 0;
 762                 }
 763                 return status;
 764         }
 765
 766         /* If we're not using byte range locks, and we know the page
 767          * is entirely in cache, it may be more efficient to avoid
 768          * fragmenting write requests.
 769          */
 770         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 771                 loff_t end_offs = i_size_read(inode) - 1;
 772                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 773
 774                 count += offset;
 775                 offset = 0;
 776                 if (unlikely(end_offs < 0)) {
 777                         /* Do nothing */
 778                 } else if (page->index == end_index) {
 779                         unsigned int pglen;
 780                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 781                         if (count < pglen)
 782                                 count = pglen;
 783                 } else if (page->index < end_index)
 784                         count = PAGE_CACHE_SIZE;
 785         }
 786
 787         /*
 788          * Try to find an NFS request corresponding to this page
 789          * and update it.
 790          * If the existing request cannot be updated, we must flush
 791          * it out now.
 792          */
 793         do {
 794                 req = nfs_update_request(ctx, inode, page, offset, count);
 795                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 796                 if (status != -EBUSY)
 797                         break;
 798                 /* Request could not be updated. Flush it out and try again */
 799                 status = nfs_wb_page(inode, page);
 800         } while (status >= 0);
 801         if (status < 0)
 802                 goto done;
 803
 804         status = 0;
 805
 806         /* Update file length */
 807         nfs_grow_file(page, offset, count);
 808         /* Set the PG_uptodate flag? */
 809         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 810         nfs_unlock_request(req);
 811 done:
 812         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 813                         status, (long long)i_size_read(inode));
 814         if (status < 0)
 815                 ClearPageUptodate(page);
 816         return status;
 817 }
 818
 819 static void nfs_writepage_release(struct nfs_page *req)
 820 {
 821         end_page_writeback(req->wb_page);
 822
 823 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 824         if (!PageError(req->wb_page)) {
 825                 if (NFS_NEED_RESCHED(req)) {
 826                         nfs_mark_request_dirty(req);
 827                         goto out;
 828                 } else if (NFS_NEED_COMMIT(req)) {
 829                         nfs_mark_request_commit(req);
 830                         goto out;
 831                 }
 832         }
 833         nfs_inode_remove_request(req);
 834
 835 out:
 836         nfs_clear_commit(req);
 837         nfs_clear_reschedule(req);
 838 #else
 839         nfs_inode_remove_request(req);
 840 #endif
 841         nfs_clear_page_writeback(req);
 842 }
 843
 844 static inline int flush_task_priority(int how)
 845 {
 846         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 847                 case FLUSH_HIGHPRI:
 848                         return RPC_PRIORITY_HIGH;
 849                 case FLUSH_LOWPRI:
 850                         return RPC_PRIORITY_LOW;
 851         }
 852         return RPC_PRIORITY_NORMAL;
 853 }
 854
 855 /*
 856  * Set up the argument/result storage required for the RPC call.
 857  */
 858 static void nfs_write_rpcsetup(struct nfs_page *req,
 859                 struct nfs_write_data *data,
 860                 unsigned int count, unsigned int offset,
 861                 int how)
 862 {
 863         struct inode            *inode;
 864
 865         /* Set up the RPC argument and reply structs
 866          * NB: take care not to mess about with data->commit et al. */
 867
 868         data->req = req;
 869         data->inode = inode = req->wb_context->dentry->d_inode;
 870         data->cred = req->wb_context->cred;
 871
 872         data->args.fh     = NFS_FH(inode);
 873         data->args.offset = req_offset(req) + offset;
 874         data->args.pgbase = req->wb_pgbase + offset;
 875         data->args.pages  = data->pagevec;
 876         data->args.count  = count;
 877         data->args.context = req->wb_context;
 878
 879         data->res.fattr   = &data->fattr;
 880         data->res.count   = count;
 881         data->res.verf    = &data->verf;
 882         nfs_fattr_init(&data->fattr);
 883
 884         NFS_PROTO(inode)->write_setup(data, how);
 885
 886         data->task.tk_priority = flush_task_priority(how);
 887         data->task.tk_cookie = (unsigned long)inode;
 888
 889         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 890                 data->task.tk_pid,
 891                 inode->i_sb->s_id,
 892                 (long long)NFS_FILEID(inode),
 893                 count,
 894                 (unsigned long long)data->args.offset);
 895 }
 896
 897 static void nfs_execute_write(struct nfs_write_data *data)
 898 {
 899         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 900         sigset_t oldset;
 901
 902         rpc_clnt_sigmask(clnt, &oldset);
 903         lock_kernel();
 904         rpc_execute(&data->task);
 905         unlock_kernel();
 906         rpc_clnt_sigunmask(clnt, &oldset);
 907 }
 908
 909 /*
 910  * Generate multiple small requests to write out a single
 911  * contiguous dirty area on one page.
 912  */
 913 static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
 914 {
 915         struct nfs_page *req = nfs_list_entry(head->next);
 916         struct page *page = req->wb_page;
 917         struct nfs_write_data *data;
 918         unsigned int wsize = NFS_SERVER(inode)->wsize;
 919         unsigned int nbytes, offset;
 920         int requests = 0;
 921         LIST_HEAD(list);
 922
 923         nfs_list_remove_request(req);
 924
 925         nbytes = req->wb_bytes;
 926         for (;;) {
 927                 data = nfs_writedata_alloc(1);
 928                 if (!data)
 929                         goto out_bad;
 930                 list_add(&data->pages, &list);
 931                 requests++;
 932                 if (nbytes <= wsize)
 933                         break;
 934                 nbytes -= wsize;
 935         }
 936         atomic_set(&req->wb_complete, requests);
 937
 938         ClearPageError(page);
 939         set_page_writeback(page);
 940         offset = 0;
 941         nbytes = req->wb_bytes;
 942         do {
 943                 data = list_entry(list.next, struct nfs_write_data, pages);
 944                 list_del_init(&data->pages);
 945
 946                 data->pagevec[0] = page;
 947                 data->complete = nfs_writeback_done_partial;
 948
 949                 if (nbytes > wsize) {
 950                         nfs_write_rpcsetup(req, data, wsize, offset, how);
 951                         offset += wsize;
 952                         nbytes -= wsize;
 953                 } else {
 954                         nfs_write_rpcsetup(req, data, nbytes, offset, how);
 955                         nbytes = 0;
 956                 }
 957                 nfs_execute_write(data);
 958         } while (nbytes != 0);
 959
 960         return 0;
 961
 962 out_bad:
 963         while (!list_empty(&list)) {
 964                 data = list_entry(list.next, struct nfs_write_data, pages);
 965                 list_del(&data->pages);
 966                 nfs_writedata_free(data);
 967         }
 968         nfs_mark_request_dirty(req);
 969         nfs_clear_page_writeback(req);
 970         return -ENOMEM;
 971 }
 972
 973 /*
 974  * Create an RPC task for the given write request and kick it.
 975  * The page must have been locked by the caller.
 976  *
 977  * It may happen that the page we're passed is not marked dirty.
 978  * This is the case if nfs_updatepage detects a conflicting request
 979  * that has been written but not committed.
 980  */
 981 static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
 982 {
 983         struct nfs_page         *req;
 984         struct page             **pages;
 985         struct nfs_write_data   *data;
 986         unsigned int            count;
 987
 988         if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
 989                 return nfs_flush_multi(head, inode, how);
 990
 991         data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
 992         if (!data)
 993                 goto out_bad;
 994
 995         pages = data->pagevec;
 996         count = 0;
 997         while (!list_empty(head)) {
 998                 req = nfs_list_entry(head->next);
 999                 nfs_list_remove_request(req);
1000                 nfs_list_add_request(req, &data->pages);
1001                 ClearPageError(req->wb_page);
1002                 set_page_writeback(req->wb_page);
1003                 *pages++ = req->wb_page;
1004                 count += req->wb_bytes;
1005         }
1006         req = nfs_list_entry(data->pages.next);
1007
1008         data->complete = nfs_writeback_done_full;
1009         /* Set up the argument struct */
1010         nfs_write_rpcsetup(req, data, count, 0, how);
1011
1012         nfs_execute_write(data);
1013         return 0;
1014  out_bad:
1015         while (!list_empty(head)) {
1016                 struct nfs_page *req = nfs_list_entry(head->next);
1017                 nfs_list_remove_request(req);
1018                 nfs_mark_request_dirty(req);
1019                 nfs_clear_page_writeback(req);
1020         }
1021         return -ENOMEM;
1022 }
1023
1024 static int
1025 nfs_flush_list(struct list_head *head, int wpages, int how)
1026 {
1027         LIST_HEAD(one_request);
1028         struct nfs_page         *req;
1029         int                     error = 0;
1030         unsigned int            pages = 0;
1031
1032         while (!list_empty(head)) {
1033                 pages += nfs_coalesce_requests(head, &one_request, wpages);
1034                 req = nfs_list_entry(one_request.next);
1035                 error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how);
1036                 if (error < 0)
1037                         break;
1038         }
1039         if (error >= 0)
1040                 return pages;
1041
1042         while (!list_empty(head)) {
1043                 req = nfs_list_entry(head->next);
1044                 nfs_list_remove_request(req);
1045                 nfs_mark_request_dirty(req);
1046                 nfs_clear_page_writeback(req);
1047         }
1048         return error;
1049 }
1050
1051 /*
1052  * Handle a write reply that flushed part of a page.
1053  */
1054 static void nfs_writeback_done_partial(struct nfs_write_data *data, int status)
1055 {
1056         struct nfs_page         *req = data->req;
1057         struct page             *page = req->wb_page;
1058
1059         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1060                 req->wb_context->dentry->d_inode->i_sb->s_id,
1061                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1062                 req->wb_bytes,
1063                 (long long)req_offset(req));
1064
1065         if (status < 0) {
1066                 ClearPageUptodate(page);
1067                 SetPageError(page);
1068                 req->wb_context->error = status;
1069                 dprintk(", error = %d\n", status);
1070         } else {
1071 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1072                 if (data->verf.committed < NFS_FILE_SYNC) {
1073                         if (!NFS_NEED_COMMIT(req)) {
1074                                 nfs_defer_commit(req);
1075                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1076                                 dprintk(" defer commit\n");
1077                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1078                                 nfs_defer_reschedule(req);
1079                                 dprintk(" server reboot detected\n");
1080                         }
1081                 } else
1082 #endif
1083                         dprintk(" OK\n");
1084         }
1085
1086         if (atomic_dec_and_test(&req->wb_complete))
1087                 nfs_writepage_release(req);
1088 }
1089
1090 /*
1091  * Handle a write reply that flushes a whole page.
1092  *
1093  * FIXME: There is an inherent race with invalidate_inode_pages and
1094  *        writebacks since the page->count is kept > 1 for as long
1095  *        as the page has a write request pending.
1096  */
1097 static void nfs_writeback_done_full(struct nfs_write_data *data, int status)
1098 {
1099         struct nfs_page         *req;
1100         struct page             *page;
1101
1102         /* Update attributes as result of writeback. */
1103         while (!list_empty(&data->pages)) {
1104                 req = nfs_list_entry(data->pages.next);
1105                 nfs_list_remove_request(req);
1106                 page = req->wb_page;
1107
1108                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1109                         req->wb_context->dentry->d_inode->i_sb->s_id,
1110                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1111                         req->wb_bytes,
1112                         (long long)req_offset(req));
1113
1114                 if (status < 0) {
1115                         ClearPageUptodate(page);
1116                         SetPageError(page);
1117                         req->wb_context->error = status;
1118                         end_page_writeback(page);
1119                         nfs_inode_remove_request(req);
1120                         dprintk(", error = %d\n", status);
1121                         goto next;
1122                 }
1123                 end_page_writeback(page);
1124
1125 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1126                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1127                         nfs_inode_remove_request(req);
1128                         dprintk(" OK\n");
1129                         goto next;
1130                 }
1131                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1132                 nfs_mark_request_commit(req);
1133                 dprintk(" marked for commit\n");
1134 #else
1135                 nfs_inode_remove_request(req);
1136 #endif
1137         next:
1138                 nfs_clear_page_writeback(req);
1139         }
1140 }
1141
1142 /*
1143  * This function is called when the WRITE call is complete.
1144  */
1145 void nfs_writeback_done(struct rpc_task *task, void *calldata)
1146 {
1147         struct nfs_write_data   *data = calldata;
1148         struct nfs_writeargs    *argp = &data->args;
1149         struct nfs_writeres     *resp = &data->res;
1150
1151         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1152                 task->tk_pid, task->tk_status);
1153
1154 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1155         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1156                 /* We tried a write call, but the server did not
1157                  * commit data to stable storage even though we
1158                  * requested it.
1159                  * Note: There is a known bug in Tru64 < 5.0 in which
1160                  *       the server reports NFS_DATA_SYNC, but performs
1161                  *       NFS_FILE_SYNC. We therefore implement this checking
1162                  *       as a dprintk() in order to avoid filling syslog.
1163                  */
1164                 static unsigned long    complain;
1165
1166                 if (time_before(complain, jiffies)) {
1167                         dprintk("NFS: faulty NFS server %s:"
1168                                 " (committed = %d) != (stable = %d)\n",
1169                                 NFS_SERVER(data->inode)->hostname,
1170                                 resp->verf->committed, argp->stable);
1171                         complain = jiffies + 300 * HZ;
1172                 }
1173         }
1174 #endif
1175         /* Is this a short write? */
1176         if (task->tk_status >= 0 && resp->count < argp->count) {
1177                 static unsigned long    complain;
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;
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
1206         /*
1207          * Process the nfs_page list
1208          */
1209         data->complete(data, task->tk_status);
1210 }
1211
1212
1213 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1214 void nfs_commit_release(void *wdata)
1215 {
1216         nfs_commit_free(wdata);
1217 }
1218
1219 /*
1220  * Set up the argument/result storage required for the RPC call.
1221  */
1222 static void nfs_commit_rpcsetup(struct list_head *head,
1223                 struct nfs_write_data *data, int how)
1224 {
1225         struct nfs_page         *first;
1226         struct inode            *inode;
1227
1228         /* Set up the RPC argument and reply structs
1229          * NB: take care not to mess about with data->commit et al. */
1230
1231         list_splice_init(head, &data->pages);
1232         first = nfs_list_entry(data->pages.next);
1233         inode = first->wb_context->dentry->d_inode;
1234
1235         data->inode       = inode;
1236         data->cred        = first->wb_context->cred;
1237
1238         data->args.fh     = NFS_FH(data->inode);
1239         /* Note: we always request a commit of the entire inode */
1240         data->args.offset = 0;
1241         data->args.count  = 0;
1242         data->res.count   = 0;
1243         data->res.fattr   = &data->fattr;
1244         data->res.verf    = &data->verf;
1245         nfs_fattr_init(&data->fattr);
1246
1247         NFS_PROTO(inode)->commit_setup(data, how);
1248
1249         data->task.tk_priority = flush_task_priority(how);
1250         data->task.tk_cookie = (unsigned long)inode;
1251
1252         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1253 }
1254
1255 /*
1256  * Commit dirty pages
1257  */
1258 static int
1259 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1260 {
1261         struct nfs_write_data   *data;
1262         struct nfs_page         *req;
1263
1264         data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
1265
1266         if (!data)
1267                 goto out_bad;
1268
1269         /* Set up the argument struct */
1270         nfs_commit_rpcsetup(head, data, how);
1271
1272         nfs_execute_write(data);
1273         return 0;
1274  out_bad:
1275         while (!list_empty(head)) {
1276                 req = nfs_list_entry(head->next);
1277                 nfs_list_remove_request(req);
1278                 nfs_mark_request_commit(req);
1279                 nfs_clear_page_writeback(req);
1280         }
1281         return -ENOMEM;
1282 }
1283
1284 /*
1285  * COMMIT call returned
1286  */
1287 void nfs_commit_done(struct rpc_task *task, void *calldata)
1288 {
1289         struct nfs_write_data   *data = calldata;
1290         struct nfs_page         *req;
1291         int res = 0;
1292
1293         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1294                                 task->tk_pid, task->tk_status);
1295
1296         while (!list_empty(&data->pages)) {
1297                 req = nfs_list_entry(data->pages.next);
1298                 nfs_list_remove_request(req);
1299
1300                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1301                         req->wb_context->dentry->d_inode->i_sb->s_id,
1302                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1303                         req->wb_bytes,
1304                         (long long)req_offset(req));
1305                 if (task->tk_status < 0) {
1306                         req->wb_context->error = task->tk_status;
1307                         nfs_inode_remove_request(req);
1308                         dprintk(", error = %d\n", task->tk_status);
1309                         goto next;
1310                 }
1311
1312                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1313                  * returned by the server against all stored verfs. */
1314                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1315                         /* We have a match */
1316                         nfs_inode_remove_request(req);
1317                         dprintk(" OK\n");
1318                         goto next;
1319                 }
1320                 /* We have a mismatch. Write the page again */
1321                 dprintk(" mismatch\n");
1322                 nfs_mark_request_dirty(req);
1323         next:
1324                 nfs_clear_page_writeback(req);
1325                 res++;
1326         }
1327         sub_page_state(nr_unstable,res);
1328 }
1329 #endif
1330
1331 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1332                            unsigned int npages, int how)
1333 {
1334         struct nfs_inode *nfsi = NFS_I(inode);
1335         LIST_HEAD(head);
1336         int                     res,
1337                                 error = 0;
1338
1339         spin_lock(&nfsi->req_lock);
1340         res = nfs_scan_dirty(inode, &head, idx_start, npages);
1341         spin_unlock(&nfsi->req_lock);
1342         if (res) {
1343                 struct nfs_server *server = NFS_SERVER(inode);
1344
1345                 /* For single writes, FLUSH_STABLE is more efficient */
1346                 if (res == nfsi->npages && nfsi->npages <= server->wpages) {
1347                         if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize)
1348                                 how |= FLUSH_STABLE;
1349                 }
1350                 error = nfs_flush_list(&head, server->wpages, how);
1351         }
1352         if (error < 0)
1353                 return error;
1354         return res;
1355 }
1356
1357 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1358 int nfs_commit_inode(struct inode *inode, int how)
1359 {
1360         struct nfs_inode *nfsi = NFS_I(inode);
1361         LIST_HEAD(head);
1362         int                     res,
1363                                 error = 0;
1364
1365         spin_lock(&nfsi->req_lock);
1366         res = nfs_scan_commit(inode, &head, 0, 0);
1367         spin_unlock(&nfsi->req_lock);
1368         if (res) {
1369                 error = nfs_commit_list(inode, &head, how);
1370                 if (error < 0)
1371                         return error;
1372         }
1373         return res;
1374 }
1375 #endif
1376
1377 int nfs_sync_inode(struct inode *inode, unsigned long idx_start,
1378                   unsigned int npages, int how)
1379 {
1380         int nocommit = how & FLUSH_NOCOMMIT;
1381         int wait = how & FLUSH_WAIT;
1382         int error;
1383
1384         how &= ~(FLUSH_WAIT|FLUSH_NOCOMMIT);
1385
1386         do {
1387                 if (wait) {
1388                         error = nfs_wait_on_requests(inode, idx_start, npages);
1389                         if (error != 0)
1390                                 continue;
1391                 }
1392                 error = nfs_flush_inode(inode, idx_start, npages, how);
1393                 if (error != 0)
1394                         continue;
1395                 if (!nocommit)
1396                         error = nfs_commit_inode(inode, how);
1397         } while (error > 0);
1398         return error;
1399 }
1400
1401 int nfs_init_writepagecache(void)
1402 {
1403         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1404                                              sizeof(struct nfs_write_data),
1405                                              0, SLAB_HWCACHE_ALIGN,
1406                                              NULL, NULL);
1407         if (nfs_wdata_cachep == NULL)
1408                 return -ENOMEM;
1409
1410         nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE,
1411                                            mempool_alloc_slab,
1412                                            mempool_free_slab,
1413                                            nfs_wdata_cachep);
1414         if (nfs_wdata_mempool == NULL)
1415                 return -ENOMEM;
1416
1417         nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT,
1418                                            mempool_alloc_slab,
1419                                            mempool_free_slab,
1420                                            nfs_wdata_cachep);
1421         if (nfs_commit_mempool == NULL)
1422                 return -ENOMEM;
1423
1424         return 0;
1425 }
1426
1427 void nfs_destroy_writepagecache(void)
1428 {
1429         mempool_destroy(nfs_commit_mempool);
1430         mempool_destroy(nfs_wdata_mempool);
1431         if (kmem_cache_destroy(nfs_wdata_cachep))
1432                 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1433 }
1434