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