Pileus Git - ~andy/linux/blob - fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16
  17 #include <linux/sunrpc/clnt.h>
  18 #include <linux/nfs_fs.h>
  19 #include <linux/nfs_mount.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/backing-dev.h>
  22
  23 #include <asm/uaccess.h>
  24
  25 #include "delegation.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 #define MIN_POOL_WRITE          (32)
  32 #define MIN_POOL_COMMIT         (4)
  33
  34 /*
  35  * Local function declarations
  36  */
  37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  38                                             struct page *,
  39                                             unsigned int, unsigned int);
  40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static void nfs_redirty_request(struct nfs_page *req);
  43 static const struct rpc_call_ops nfs_write_partial_ops;
  44 static const struct rpc_call_ops nfs_write_full_ops;
  45 static const struct rpc_call_ops nfs_commit_ops;
  46
  47 static struct kmem_cache *nfs_wdata_cachep;
  48 static mempool_t *nfs_wdata_mempool;
  49 static mempool_t *nfs_commit_mempool;
  50
  51 struct nfs_write_data *nfs_commitdata_alloc(void)
  52 {
  53         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  54
  55         if (p) {
  56                 memset(p, 0, sizeof(*p));
  57                 INIT_LIST_HEAD(&p->pages);
  58         }
  59         return p;
  60 }
  61
  62 void nfs_commit_free(struct nfs_write_data *p)
  63 {
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  70 {
  71         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  72
  73         if (p) {
  74                 memset(p, 0, sizeof(*p));
  75                 INIT_LIST_HEAD(&p->pages);
  76                 p->npages = pagecount;
  77                 if (pagecount <= ARRAY_SIZE(p->page_array))
  78                         p->pagevec = p->page_array;
  79                 else {
  80                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  81                         if (!p->pagevec) {
  82                                 mempool_free(p, nfs_wdata_mempool);
  83                                 p = NULL;
  84                         }
  85                 }
  86         }
  87         return p;
  88 }
  89
  90 static void nfs_writedata_free(struct nfs_write_data *p)
  91 {
  92         if (p && (p->pagevec != &p->page_array[0]))
  93                 kfree(p->pagevec);
  94         mempool_free(p, nfs_wdata_mempool);
  95 }
  96
  97 void nfs_writedata_release(void *data)
  98 {
  99         struct nfs_write_data *wdata = data;
 100
 101         put_nfs_open_context(wdata->args.context);
 102         nfs_writedata_free(wdata);
 103 }
 104
 105 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 106 {
 107         ctx->error = error;
 108         smp_wmb();
 109         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 110 }
 111
 112 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 113 {
 114         struct nfs_page *req = NULL;
 115
 116         if (PagePrivate(page)) {
 117                 req = (struct nfs_page *)page_private(page);
 118                 if (req != NULL)
 119                         kref_get(&req->wb_kref);
 120         }
 121         return req;
 122 }
 123
 124 static struct nfs_page *nfs_page_find_request(struct page *page)
 125 {
 126         struct inode *inode = page->mapping->host;
 127         struct nfs_page *req = NULL;
 128
 129         spin_lock(&inode->i_lock);
 130         req = nfs_page_find_request_locked(page);
 131         spin_unlock(&inode->i_lock);
 132         return req;
 133 }
 134
 135 /* Adjust the file length if we're writing beyond the end */
 136 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 137 {
 138         struct inode *inode = page->mapping->host;
 139         loff_t end, i_size = i_size_read(inode);
 140         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 141
 142         if (i_size > 0 && page->index < end_index)
 143                 return;
 144         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 145         if (i_size >= end)
 146                 return;
 147         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 148         i_size_write(inode, end);
 149 }
 150
 151 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 152 static void nfs_set_pageerror(struct page *page)
 153 {
 154         SetPageError(page);
 155         nfs_zap_mapping(page->mapping->host, page->mapping);
 156 }
 157
 158 /* We can set the PG_uptodate flag if we see that a write request
 159  * covers the full page.
 160  */
 161 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 162 {
 163         if (PageUptodate(page))
 164                 return;
 165         if (base != 0)
 166                 return;
 167         if (count != nfs_page_length(page))
 168                 return;
 169         SetPageUptodate(page);
 170 }
 171
 172 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 173                 unsigned int offset, unsigned int count)
 174 {
 175         struct nfs_page *req;
 176         int ret;
 177
 178         for (;;) {
 179                 req = nfs_update_request(ctx, page, offset, count);
 180                 if (!IS_ERR(req))
 181                         break;
 182                 ret = PTR_ERR(req);
 183                 if (ret != -EBUSY)
 184                         return ret;
 185                 ret = nfs_wb_page(page->mapping->host, page);
 186                 if (ret != 0)
 187                         return ret;
 188         }
 189         /* Update file length */
 190         nfs_grow_file(page, offset, count);
 191         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 192         nfs_clear_page_tag_locked(req);
 193         return 0;
 194 }
 195
 196 static int wb_priority(struct writeback_control *wbc)
 197 {
 198         if (wbc->for_reclaim)
 199                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 200         if (wbc->for_kupdate)
 201                 return FLUSH_LOWPRI;
 202         return 0;
 203 }
 204
 205 /*
 206  * NFS congestion control
 207  */
 208
 209 int nfs_congestion_kb;
 210
 211 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 212 #define NFS_CONGESTION_OFF_THRESH       \
 213         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 214
 215 static int nfs_set_page_writeback(struct page *page)
 216 {
 217         int ret = test_set_page_writeback(page);
 218
 219         if (!ret) {
 220                 struct inode *inode = page->mapping->host;
 221                 struct nfs_server *nfss = NFS_SERVER(inode);
 222
 223                 if (atomic_long_inc_return(&nfss->writeback) >
 224                                 NFS_CONGESTION_ON_THRESH)
 225                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 226         }
 227         return ret;
 228 }
 229
 230 static void nfs_end_page_writeback(struct page *page)
 231 {
 232         struct inode *inode = page->mapping->host;
 233         struct nfs_server *nfss = NFS_SERVER(inode);
 234
 235         end_page_writeback(page);
 236         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 237                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 238 }
 239
 240 /*
 241  * Find an associated nfs write request, and prepare to flush it out
 242  * May return an error if the user signalled nfs_wait_on_request().
 243  */
 244 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 245                                 struct page *page)
 246 {
 247         struct inode *inode = page->mapping->host;
 248         struct nfs_page *req;
 249         int ret;
 250
 251         spin_lock(&inode->i_lock);
 252         for(;;) {
 253                 req = nfs_page_find_request_locked(page);
 254                 if (req == NULL) {
 255                         spin_unlock(&inode->i_lock);
 256                         return 0;
 257                 }
 258                 if (nfs_set_page_tag_locked(req))
 259                         break;
 260                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 261                  *       then the call to nfs_set_page_tag_locked() will always
 262                  *       succeed provided that someone hasn't already marked the
 263                  *       request as dirty (in which case we don't care).
 264                  */
 265                 spin_unlock(&inode->i_lock);
 266                 ret = nfs_wait_on_request(req);
 267                 nfs_release_request(req);
 268                 if (ret != 0)
 269                         return ret;
 270                 spin_lock(&inode->i_lock);
 271         }
 272         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 273                 /* This request is marked for commit */
 274                 spin_unlock(&inode->i_lock);
 275                 nfs_clear_page_tag_locked(req);
 276                 nfs_pageio_complete(pgio);
 277                 return 0;
 278         }
 279         if (nfs_set_page_writeback(page) != 0) {
 280                 spin_unlock(&inode->i_lock);
 281                 BUG();
 282         }
 283         spin_unlock(&inode->i_lock);
 284         if (!nfs_pageio_add_request(pgio, req)) {
 285                 nfs_redirty_request(req);
 286                 return pgio->pg_error;
 287         }
 288         return 0;
 289 }
 290
 291 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 292 {
 293         struct inode *inode = page->mapping->host;
 294
 295         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 296         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 297
 298         nfs_pageio_cond_complete(pgio, page->index);
 299         return nfs_page_async_flush(pgio, page);
 300 }
 301
 302 /*
 303  * Write an mmapped page to the server.
 304  */
 305 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 306 {
 307         struct nfs_pageio_descriptor pgio;
 308         int err;
 309
 310         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 311         err = nfs_do_writepage(page, wbc, &pgio);
 312         nfs_pageio_complete(&pgio);
 313         if (err < 0)
 314                 return err;
 315         if (pgio.pg_error < 0)
 316                 return pgio.pg_error;
 317         return 0;
 318 }
 319
 320 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 321 {
 322         int ret;
 323
 324         ret = nfs_writepage_locked(page, wbc);
 325         unlock_page(page);
 326         return ret;
 327 }
 328
 329 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 330 {
 331         int ret;
 332
 333         ret = nfs_do_writepage(page, wbc, data);
 334         unlock_page(page);
 335         return ret;
 336 }
 337
 338 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 339 {
 340         struct inode *inode = mapping->host;
 341         struct nfs_pageio_descriptor pgio;
 342         int err;
 343
 344         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 345
 346         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 347         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 348         nfs_pageio_complete(&pgio);
 349         if (err < 0)
 350                 return err;
 351         if (pgio.pg_error < 0)
 352                 return pgio.pg_error;
 353         return 0;
 354 }
 355
 356 /*
 357  * Insert a write request into an inode
 358  */
 359 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 360 {
 361         struct nfs_inode *nfsi = NFS_I(inode);
 362         int error;
 363
 364         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 365         BUG_ON(error);
 366         if (!nfsi->npages) {
 367                 igrab(inode);
 368                 if (nfs_have_delegation(inode, FMODE_WRITE))
 369                         nfsi->change_attr++;
 370         }
 371         SetPagePrivate(req->wb_page);
 372         set_page_private(req->wb_page, (unsigned long)req);
 373         nfsi->npages++;
 374         kref_get(&req->wb_kref);
 375         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 376                                 NFS_PAGE_TAG_LOCKED);
 377 }
 378
 379 /*
 380  * Remove a write request from an inode
 381  */
 382 static void nfs_inode_remove_request(struct nfs_page *req)
 383 {
 384         struct inode *inode = req->wb_context->path.dentry->d_inode;
 385         struct nfs_inode *nfsi = NFS_I(inode);
 386
 387         BUG_ON (!NFS_WBACK_BUSY(req));
 388
 389         spin_lock(&inode->i_lock);
 390         set_page_private(req->wb_page, 0);
 391         ClearPagePrivate(req->wb_page);
 392         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 393         nfsi->npages--;
 394         if (!nfsi->npages) {
 395                 spin_unlock(&inode->i_lock);
 396                 iput(inode);
 397         } else
 398                 spin_unlock(&inode->i_lock);
 399         nfs_clear_request(req);
 400         nfs_release_request(req);
 401 }
 402
 403 static void
 404 nfs_mark_request_dirty(struct nfs_page *req)
 405 {
 406         __set_page_dirty_nobuffers(req->wb_page);
 407 }
 408
 409 /*
 410  * Check if a request is dirty
 411  */
 412 static inline int
 413 nfs_dirty_request(struct nfs_page *req)
 414 {
 415         struct page *page = req->wb_page;
 416
 417         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 418                 return 0;
 419         return !PageWriteback(page);
 420 }
 421
 422 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 423 /*
 424  * Add a request to the inode's commit list.
 425  */
 426 static void
 427 nfs_mark_request_commit(struct nfs_page *req)
 428 {
 429         struct inode *inode = req->wb_context->path.dentry->d_inode;
 430         struct nfs_inode *nfsi = NFS_I(inode);
 431
 432         spin_lock(&inode->i_lock);
 433         nfsi->ncommit++;
 434         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 435         radix_tree_tag_set(&nfsi->nfs_page_tree,
 436                         req->wb_index,
 437                         NFS_PAGE_TAG_COMMIT);
 438         spin_unlock(&inode->i_lock);
 439         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 440         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 441         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 442 }
 443
 444 static inline
 445 int nfs_write_need_commit(struct nfs_write_data *data)
 446 {
 447         return data->verf.committed != NFS_FILE_SYNC;
 448 }
 449
 450 static inline
 451 int nfs_reschedule_unstable_write(struct nfs_page *req)
 452 {
 453         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 454                 nfs_mark_request_commit(req);
 455                 return 1;
 456         }
 457         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 458                 nfs_mark_request_dirty(req);
 459                 return 1;
 460         }
 461         return 0;
 462 }
 463 #else
 464 static inline void
 465 nfs_mark_request_commit(struct nfs_page *req)
 466 {
 467 }
 468
 469 static inline
 470 int nfs_write_need_commit(struct nfs_write_data *data)
 471 {
 472         return 0;
 473 }
 474
 475 static inline
 476 int nfs_reschedule_unstable_write(struct nfs_page *req)
 477 {
 478         return 0;
 479 }
 480 #endif
 481
 482 /*
 483  * Wait for a request to complete.
 484  *
 485  * Interruptible by fatal signals only.
 486  */
 487 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 488 {
 489         struct nfs_inode *nfsi = NFS_I(inode);
 490         struct nfs_page *req;
 491         pgoff_t idx_end, next;
 492         unsigned int            res = 0;
 493         int                     error;
 494
 495         if (npages == 0)
 496                 idx_end = ~0;
 497         else
 498                 idx_end = idx_start + npages - 1;
 499
 500         next = idx_start;
 501         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 502                 if (req->wb_index > idx_end)
 503                         break;
 504
 505                 next = req->wb_index + 1;
 506                 BUG_ON(!NFS_WBACK_BUSY(req));
 507
 508                 kref_get(&req->wb_kref);
 509                 spin_unlock(&inode->i_lock);
 510                 error = nfs_wait_on_request(req);
 511                 nfs_release_request(req);
 512                 spin_lock(&inode->i_lock);
 513                 if (error < 0)
 514                         return error;
 515                 res++;
 516         }
 517         return res;
 518 }
 519
 520 static void nfs_cancel_commit_list(struct list_head *head)
 521 {
 522         struct nfs_page *req;
 523
 524         while(!list_empty(head)) {
 525                 req = nfs_list_entry(head->next);
 526                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 527                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
 528                                 BDI_RECLAIMABLE);
 529                 nfs_list_remove_request(req);
 530                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 531                 nfs_inode_remove_request(req);
 532                 nfs_unlock_request(req);
 533         }
 534 }
 535
 536 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 537 /*
 538  * nfs_scan_commit - Scan an inode for commit requests
 539  * @inode: NFS inode to scan
 540  * @dst: destination list
 541  * @idx_start: lower bound of page->index to scan.
 542  * @npages: idx_start + npages sets the upper bound to scan.
 543  *
 544  * Moves requests from the inode's 'commit' request list.
 545  * The requests are *not* checked to ensure that they form a contiguous set.
 546  */
 547 static int
 548 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 549 {
 550         struct nfs_inode *nfsi = NFS_I(inode);
 551         int res = 0;
 552
 553         if (nfsi->ncommit != 0) {
 554                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 555                                 NFS_PAGE_TAG_COMMIT);
 556                 nfsi->ncommit -= res;
 557         }
 558         return res;
 559 }
 560 #else
 561 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 562 {
 563         return 0;
 564 }
 565 #endif
 566
 567 /*
 568  * Try to update any existing write request, or create one if there is none.
 569  * In order to match, the request's credentials must match those of
 570  * the calling process.
 571  *
 572  * Note: Should always be called with the Page Lock held!
 573  */
 574 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 575                 struct page *page, unsigned int offset, unsigned int bytes)
 576 {
 577         struct address_space *mapping = page->mapping;
 578         struct inode *inode = mapping->host;
 579         struct nfs_page         *req, *new = NULL;
 580         pgoff_t         rqend, end;
 581
 582         end = offset + bytes;
 583
 584         for (;;) {
 585                 /* Loop over all inode entries and see if we find
 586                  * A request for the page we wish to update
 587                  */
 588                 spin_lock(&inode->i_lock);
 589                 req = nfs_page_find_request_locked(page);
 590                 if (req) {
 591                         if (!nfs_set_page_tag_locked(req)) {
 592                                 int error;
 593
 594                                 spin_unlock(&inode->i_lock);
 595                                 error = nfs_wait_on_request(req);
 596                                 nfs_release_request(req);
 597                                 if (error < 0) {
 598                                         if (new) {
 599                                                 radix_tree_preload_end();
 600                                                 nfs_release_request(new);
 601                                         }
 602                                         return ERR_PTR(error);
 603                                 }
 604                                 continue;
 605                         }
 606                         spin_unlock(&inode->i_lock);
 607                         if (new) {
 608                                 radix_tree_preload_end();
 609                                 nfs_release_request(new);
 610                         }
 611                         break;
 612                 }
 613
 614                 if (new) {
 615                         nfs_lock_request_dontget(new);
 616                         nfs_inode_add_request(inode, new);
 617                         spin_unlock(&inode->i_lock);
 618                         radix_tree_preload_end();
 619                         req = new;
 620                         goto out;
 621                 }
 622                 spin_unlock(&inode->i_lock);
 623
 624                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 625                 if (IS_ERR(new))
 626                         return new;
 627                 if (radix_tree_preload(GFP_NOFS)) {
 628                         nfs_release_request(new);
 629                         return ERR_PTR(-ENOMEM);
 630                 }
 631         }
 632
 633         /* We have a request for our page.
 634          * If the creds don't match, or the
 635          * page addresses don't match,
 636          * tell the caller to wait on the conflicting
 637          * request.
 638          */
 639         rqend = req->wb_offset + req->wb_bytes;
 640         if (req->wb_context != ctx
 641             || req->wb_page != page
 642             || !nfs_dirty_request(req)
 643             || offset > rqend || end < req->wb_offset) {
 644                 nfs_clear_page_tag_locked(req);
 645                 return ERR_PTR(-EBUSY);
 646         }
 647
 648         /* Okay, the request matches. Update the region */
 649         if (offset < req->wb_offset) {
 650                 req->wb_offset = offset;
 651                 req->wb_pgbase = offset;
 652                 req->wb_bytes = max(end, rqend) - req->wb_offset;
 653                 goto out;
 654         }
 655
 656         if (end > rqend)
 657                 req->wb_bytes = end - req->wb_offset;
 658
 659 out:
 660         return req;
 661 }
 662
 663 int nfs_flush_incompatible(struct file *file, struct page *page)
 664 {
 665         struct nfs_open_context *ctx = nfs_file_open_context(file);
 666         struct nfs_page *req;
 667         int do_flush, status;
 668         /*
 669          * Look for a request corresponding to this page. If there
 670          * is one, and it belongs to another file, we flush it out
 671          * before we try to copy anything into the page. Do this
 672          * due to the lack of an ACCESS-type call in NFSv2.
 673          * Also do the same if we find a request from an existing
 674          * dropped page.
 675          */
 676         do {
 677                 req = nfs_page_find_request(page);
 678                 if (req == NULL)
 679                         return 0;
 680                 do_flush = req->wb_page != page || req->wb_context != ctx
 681                         || !nfs_dirty_request(req);
 682                 nfs_release_request(req);
 683                 if (!do_flush)
 684                         return 0;
 685                 status = nfs_wb_page(page->mapping->host, page);
 686         } while (status == 0);
 687         return status;
 688 }
 689
 690 /*
 691  * If the page cache is marked as unsafe or invalid, then we can't rely on
 692  * the PageUptodate() flag. In this case, we will need to turn off
 693  * write optimisations that depend on the page contents being correct.
 694  */
 695 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 696 {
 697         return PageUptodate(page) &&
 698                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 699 }
 700
 701 /*
 702  * Update and possibly write a cached page of an NFS file.
 703  *
 704  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 705  * things with a page scheduled for an RPC call (e.g. invalidate it).
 706  */
 707 int nfs_updatepage(struct file *file, struct page *page,
 708                 unsigned int offset, unsigned int count)
 709 {
 710         struct nfs_open_context *ctx = nfs_file_open_context(file);
 711         struct inode    *inode = page->mapping->host;
 712         int             status = 0;
 713
 714         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 715
 716         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 717                 file->f_path.dentry->d_parent->d_name.name,
 718                 file->f_path.dentry->d_name.name, count,
 719                 (long long)(page_offset(page) +offset));
 720
 721         /* If we're not using byte range locks, and we know the page
 722          * is up to date, it may be more efficient to extend the write
 723          * to cover the entire page in order to avoid fragmentation
 724          * inefficiencies.
 725          */
 726         if (nfs_write_pageuptodate(page, inode) &&
 727                         inode->i_flock == NULL &&
 728                         !(file->f_flags & O_SYNC)) {
 729                 count = max(count + offset, nfs_page_length(page));
 730                 offset = 0;
 731         }
 732
 733         status = nfs_writepage_setup(ctx, page, offset, count);
 734         if (status < 0)
 735                 nfs_set_pageerror(page);
 736         else
 737                 __set_page_dirty_nobuffers(page);
 738
 739         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 740                         status, (long long)i_size_read(inode));
 741         return status;
 742 }
 743
 744 static void nfs_writepage_release(struct nfs_page *req)
 745 {
 746
 747         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 748                 nfs_end_page_writeback(req->wb_page);
 749                 nfs_inode_remove_request(req);
 750         } else
 751                 nfs_end_page_writeback(req->wb_page);
 752         nfs_clear_page_tag_locked(req);
 753 }
 754
 755 static int flush_task_priority(int how)
 756 {
 757         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 758                 case FLUSH_HIGHPRI:
 759                         return RPC_PRIORITY_HIGH;
 760                 case FLUSH_LOWPRI:
 761                         return RPC_PRIORITY_LOW;
 762         }
 763         return RPC_PRIORITY_NORMAL;
 764 }
 765
 766 /*
 767  * Set up the argument/result storage required for the RPC call.
 768  */
 769 static int nfs_write_rpcsetup(struct nfs_page *req,
 770                 struct nfs_write_data *data,
 771                 const struct rpc_call_ops *call_ops,
 772                 unsigned int count, unsigned int offset,
 773                 int how)
 774 {
 775         struct inode *inode = req->wb_context->path.dentry->d_inode;
 776         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 777         int priority = flush_task_priority(how);
 778         struct rpc_task *task;
 779         struct rpc_message msg = {
 780                 .rpc_argp = &data->args,
 781                 .rpc_resp = &data->res,
 782                 .rpc_cred = req->wb_context->cred,
 783         };
 784         struct rpc_task_setup task_setup_data = {
 785                 .rpc_client = NFS_CLIENT(inode),
 786                 .task = &data->task,
 787                 .rpc_message = &msg,
 788                 .callback_ops = call_ops,
 789                 .callback_data = data,
 790                 .workqueue = nfsiod_workqueue,
 791                 .flags = flags,
 792                 .priority = priority,
 793         };
 794
 795         /* Set up the RPC argument and reply structs
 796          * NB: take care not to mess about with data->commit et al. */
 797
 798         data->req = req;
 799         data->inode = inode = req->wb_context->path.dentry->d_inode;
 800         data->cred = msg.rpc_cred;
 801
 802         data->args.fh     = NFS_FH(inode);
 803         data->args.offset = req_offset(req) + offset;
 804         data->args.pgbase = req->wb_pgbase + offset;
 805         data->args.pages  = data->pagevec;
 806         data->args.count  = count;
 807         data->args.context = get_nfs_open_context(req->wb_context);
 808         data->args.stable  = NFS_UNSTABLE;
 809         if (how & FLUSH_STABLE) {
 810                 data->args.stable = NFS_DATA_SYNC;
 811                 if (!NFS_I(inode)->ncommit)
 812                         data->args.stable = NFS_FILE_SYNC;
 813         }
 814
 815         data->res.fattr   = &data->fattr;
 816         data->res.count   = count;
 817         data->res.verf    = &data->verf;
 818         nfs_fattr_init(&data->fattr);
 819
 820         /* Set up the initial task struct.  */
 821         NFS_PROTO(inode)->write_setup(data, &msg);
 822
 823         dprintk("NFS: %5u initiated write call "
 824                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 825                 data->task.tk_pid,
 826                 inode->i_sb->s_id,
 827                 (long long)NFS_FILEID(inode),
 828                 count,
 829                 (unsigned long long)data->args.offset);
 830
 831         task = rpc_run_task(&task_setup_data);
 832         if (IS_ERR(task))
 833                 return PTR_ERR(task);
 834         rpc_put_task(task);
 835         return 0;
 836 }
 837
 838 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 839  * call this on each, which will prepare them to be retried on next
 840  * writeback using standard nfs.
 841  */
 842 static void nfs_redirty_request(struct nfs_page *req)
 843 {
 844         nfs_mark_request_dirty(req);
 845         nfs_end_page_writeback(req->wb_page);
 846         nfs_clear_page_tag_locked(req);
 847 }
 848
 849 /*
 850  * Generate multiple small requests to write out a single
 851  * contiguous dirty area on one page.
 852  */
 853 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 854 {
 855         struct nfs_page *req = nfs_list_entry(head->next);
 856         struct page *page = req->wb_page;
 857         struct nfs_write_data *data;
 858         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 859         unsigned int offset;
 860         int requests = 0;
 861         int ret = 0;
 862         LIST_HEAD(list);
 863
 864         nfs_list_remove_request(req);
 865
 866         nbytes = count;
 867         do {
 868                 size_t len = min(nbytes, wsize);
 869
 870                 data = nfs_writedata_alloc(1);
 871                 if (!data)
 872                         goto out_bad;
 873                 list_add(&data->pages, &list);
 874                 requests++;
 875                 nbytes -= len;
 876         } while (nbytes != 0);
 877         atomic_set(&req->wb_complete, requests);
 878
 879         ClearPageError(page);
 880         offset = 0;
 881         nbytes = count;
 882         do {
 883                 int ret2;
 884
 885                 data = list_entry(list.next, struct nfs_write_data, pages);
 886                 list_del_init(&data->pages);
 887
 888                 data->pagevec[0] = page;
 889
 890                 if (nbytes < wsize)
 891                         wsize = nbytes;
 892                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 893                                    wsize, offset, how);
 894                 if (ret == 0)
 895                         ret = ret2;
 896                 offset += wsize;
 897                 nbytes -= wsize;
 898         } while (nbytes != 0);
 899
 900         return ret;
 901
 902 out_bad:
 903         while (!list_empty(&list)) {
 904                 data = list_entry(list.next, struct nfs_write_data, pages);
 905                 list_del(&data->pages);
 906                 nfs_writedata_release(data);
 907         }
 908         nfs_redirty_request(req);
 909         return -ENOMEM;
 910 }
 911
 912 /*
 913  * Create an RPC task for the given write request and kick it.
 914  * The page must have been locked by the caller.
 915  *
 916  * It may happen that the page we're passed is not marked dirty.
 917  * This is the case if nfs_updatepage detects a conflicting request
 918  * that has been written but not committed.
 919  */
 920 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 921 {
 922         struct nfs_page         *req;
 923         struct page             **pages;
 924         struct nfs_write_data   *data;
 925
 926         data = nfs_writedata_alloc(npages);
 927         if (!data)
 928                 goto out_bad;
 929
 930         pages = data->pagevec;
 931         while (!list_empty(head)) {
 932                 req = nfs_list_entry(head->next);
 933                 nfs_list_remove_request(req);
 934                 nfs_list_add_request(req, &data->pages);
 935                 ClearPageError(req->wb_page);
 936                 *pages++ = req->wb_page;
 937         }
 938         req = nfs_list_entry(data->pages.next);
 939
 940         /* Set up the argument struct */
 941         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 942  out_bad:
 943         while (!list_empty(head)) {
 944                 req = nfs_list_entry(head->next);
 945                 nfs_list_remove_request(req);
 946                 nfs_redirty_request(req);
 947         }
 948         return -ENOMEM;
 949 }
 950
 951 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 952                                   struct inode *inode, int ioflags)
 953 {
 954         size_t wsize = NFS_SERVER(inode)->wsize;
 955
 956         if (wsize < PAGE_CACHE_SIZE)
 957                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 958         else
 959                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 960 }
 961
 962 /*
 963  * Handle a write reply that flushed part of a page.
 964  */
 965 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 966 {
 967         struct nfs_write_data   *data = calldata;
 968         struct nfs_page         *req = data->req;
 969
 970         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 971                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 972                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 973                 req->wb_bytes,
 974                 (long long)req_offset(req));
 975
 976         nfs_writeback_done(task, data);
 977 }
 978
 979 static void nfs_writeback_release_partial(void *calldata)
 980 {
 981         struct nfs_write_data   *data = calldata;
 982         struct nfs_page         *req = data->req;
 983         struct page             *page = req->wb_page;
 984         int status = data->task.tk_status;
 985
 986         if (status < 0) {
 987                 nfs_set_pageerror(page);
 988                 nfs_context_set_write_error(req->wb_context, status);
 989                 dprintk(", error = %d\n", status);
 990                 goto out;
 991         }
 992
 993         if (nfs_write_need_commit(data)) {
 994                 struct inode *inode = page->mapping->host;
 995
 996                 spin_lock(&inode->i_lock);
 997                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 998                         /* Do nothing we need to resend the writes */
 999                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1000                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1001                         dprintk(" defer commit\n");
1002                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1003                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1004                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1005                         dprintk(" server reboot detected\n");
1006                 }
1007                 spin_unlock(&inode->i_lock);
1008         } else
1009                 dprintk(" OK\n");
1010
1011 out:
1012         if (atomic_dec_and_test(&req->wb_complete))
1013                 nfs_writepage_release(req);
1014         nfs_writedata_release(calldata);
1015 }
1016
1017 static const struct rpc_call_ops nfs_write_partial_ops = {
1018         .rpc_call_done = nfs_writeback_done_partial,
1019         .rpc_release = nfs_writeback_release_partial,
1020 };
1021
1022 /*
1023  * Handle a write reply that flushes a whole page.
1024  *
1025  * FIXME: There is an inherent race with invalidate_inode_pages and
1026  *        writebacks since the page->count is kept > 1 for as long
1027  *        as the page has a write request pending.
1028  */
1029 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1030 {
1031         struct nfs_write_data   *data = calldata;
1032
1033         nfs_writeback_done(task, data);
1034 }
1035
1036 static void nfs_writeback_release_full(void *calldata)
1037 {
1038         struct nfs_write_data   *data = calldata;
1039         int status = data->task.tk_status;
1040
1041         /* Update attributes as result of writeback. */
1042         while (!list_empty(&data->pages)) {
1043                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1044                 struct page *page = req->wb_page;
1045
1046                 nfs_list_remove_request(req);
1047
1048                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1049                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1050                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1051                         req->wb_bytes,
1052                         (long long)req_offset(req));
1053
1054                 if (status < 0) {
1055                         nfs_set_pageerror(page);
1056                         nfs_context_set_write_error(req->wb_context, status);
1057                         dprintk(", error = %d\n", status);
1058                         goto remove_request;
1059                 }
1060
1061                 if (nfs_write_need_commit(data)) {
1062                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1063                         nfs_mark_request_commit(req);
1064                         nfs_end_page_writeback(page);
1065                         dprintk(" marked for commit\n");
1066                         goto next;
1067                 }
1068                 dprintk(" OK\n");
1069 remove_request:
1070                 nfs_end_page_writeback(page);
1071                 nfs_inode_remove_request(req);
1072         next:
1073                 nfs_clear_page_tag_locked(req);
1074         }
1075         nfs_writedata_release(calldata);
1076 }
1077
1078 static const struct rpc_call_ops nfs_write_full_ops = {
1079         .rpc_call_done = nfs_writeback_done_full,
1080         .rpc_release = nfs_writeback_release_full,
1081 };
1082
1083
1084 /*
1085  * This function is called when the WRITE call is complete.
1086  */
1087 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1088 {
1089         struct nfs_writeargs    *argp = &data->args;
1090         struct nfs_writeres     *resp = &data->res;
1091         int status;
1092
1093         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1094                 task->tk_pid, task->tk_status);
1095
1096         /*
1097          * ->write_done will attempt to use post-op attributes to detect
1098          * conflicting writes by other clients.  A strict interpretation
1099          * of close-to-open would allow us to continue caching even if
1100          * another writer had changed the file, but some applications
1101          * depend on tighter cache coherency when writing.
1102          */
1103         status = NFS_PROTO(data->inode)->write_done(task, data);
1104         if (status != 0)
1105                 return status;
1106         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1107
1108 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1109         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1110                 /* We tried a write call, but the server did not
1111                  * commit data to stable storage even though we
1112                  * requested it.
1113                  * Note: There is a known bug in Tru64 < 5.0 in which
1114                  *       the server reports NFS_DATA_SYNC, but performs
1115                  *       NFS_FILE_SYNC. We therefore implement this checking
1116                  *       as a dprintk() in order to avoid filling syslog.
1117                  */
1118                 static unsigned long    complain;
1119
1120                 if (time_before(complain, jiffies)) {
1121                         dprintk("NFS: faulty NFS server %s:"
1122                                 " (committed = %d) != (stable = %d)\n",
1123                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1124                                 resp->verf->committed, argp->stable);
1125                         complain = jiffies + 300 * HZ;
1126                 }
1127         }
1128 #endif
1129         /* Is this a short write? */
1130         if (task->tk_status >= 0 && resp->count < argp->count) {
1131                 static unsigned long    complain;
1132
1133                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1134
1135                 /* Has the server at least made some progress? */
1136                 if (resp->count != 0) {
1137                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1138                         if (resp->verf->committed != NFS_UNSTABLE) {
1139                                 /* Resend from where the server left off */
1140                                 argp->offset += resp->count;
1141                                 argp->pgbase += resp->count;
1142                                 argp->count -= resp->count;
1143                         } else {
1144                                 /* Resend as a stable write in order to avoid
1145                                  * headaches in the case of a server crash.
1146                                  */
1147                                 argp->stable = NFS_FILE_SYNC;
1148                         }
1149                         rpc_restart_call(task);
1150                         return -EAGAIN;
1151                 }
1152                 if (time_before(complain, jiffies)) {
1153                         printk(KERN_WARNING
1154                                "NFS: Server wrote zero bytes, expected %u.\n",
1155                                         argp->count);
1156                         complain = jiffies + 300 * HZ;
1157                 }
1158                 /* Can't do anything about it except throw an error. */
1159                 task->tk_status = -EIO;
1160         }
1161         return 0;
1162 }
1163
1164
1165 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1166 void nfs_commitdata_release(void *data)
1167 {
1168         struct nfs_write_data *wdata = data;
1169
1170         put_nfs_open_context(wdata->args.context);
1171         nfs_commit_free(wdata);
1172 }
1173
1174 /*
1175  * Set up the argument/result storage required for the RPC call.
1176  */
1177 static int nfs_commit_rpcsetup(struct list_head *head,
1178                 struct nfs_write_data *data,
1179                 int how)
1180 {
1181         struct nfs_page *first = nfs_list_entry(head->next);
1182         struct inode *inode = first->wb_context->path.dentry->d_inode;
1183         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1184         int priority = flush_task_priority(how);
1185         struct rpc_task *task;
1186         struct rpc_message msg = {
1187                 .rpc_argp = &data->args,
1188                 .rpc_resp = &data->res,
1189                 .rpc_cred = first->wb_context->cred,
1190         };
1191         struct rpc_task_setup task_setup_data = {
1192                 .task = &data->task,
1193                 .rpc_client = NFS_CLIENT(inode),
1194                 .rpc_message = &msg,
1195                 .callback_ops = &nfs_commit_ops,
1196                 .callback_data = data,
1197                 .workqueue = nfsiod_workqueue,
1198                 .flags = flags,
1199                 .priority = priority,
1200         };
1201
1202         /* Set up the RPC argument and reply structs
1203          * NB: take care not to mess about with data->commit et al. */
1204
1205         list_splice_init(head, &data->pages);
1206
1207         data->inode       = inode;
1208         data->cred        = msg.rpc_cred;
1209
1210         data->args.fh     = NFS_FH(data->inode);
1211         /* Note: we always request a commit of the entire inode */
1212         data->args.offset = 0;
1213         data->args.count  = 0;
1214         data->args.context = get_nfs_open_context(first->wb_context);
1215         data->res.count   = 0;
1216         data->res.fattr   = &data->fattr;
1217         data->res.verf    = &data->verf;
1218         nfs_fattr_init(&data->fattr);
1219
1220         /* Set up the initial task struct.  */
1221         NFS_PROTO(inode)->commit_setup(data, &msg);
1222
1223         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1224
1225         task = rpc_run_task(&task_setup_data);
1226         if (IS_ERR(task))
1227                 return PTR_ERR(task);
1228         rpc_put_task(task);
1229         return 0;
1230 }
1231
1232 /*
1233  * Commit dirty pages
1234  */
1235 static int
1236 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1237 {
1238         struct nfs_write_data   *data;
1239         struct nfs_page         *req;
1240
1241         data = nfs_commitdata_alloc();
1242
1243         if (!data)
1244                 goto out_bad;
1245
1246         /* Set up the argument struct */
1247         return nfs_commit_rpcsetup(head, data, how);
1248  out_bad:
1249         while (!list_empty(head)) {
1250                 req = nfs_list_entry(head->next);
1251                 nfs_list_remove_request(req);
1252                 nfs_mark_request_commit(req);
1253                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1254                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1255                                 BDI_RECLAIMABLE);
1256                 nfs_clear_page_tag_locked(req);
1257         }
1258         return -ENOMEM;
1259 }
1260
1261 /*
1262  * COMMIT call returned
1263  */
1264 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1265 {
1266         struct nfs_write_data   *data = calldata;
1267
1268         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1269                                 task->tk_pid, task->tk_status);
1270
1271         /* Call the NFS version-specific code */
1272         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1273                 return;
1274 }
1275
1276 static void nfs_commit_release(void *calldata)
1277 {
1278         struct nfs_write_data   *data = calldata;
1279         struct nfs_page         *req;
1280         int status = data->task.tk_status;
1281
1282         while (!list_empty(&data->pages)) {
1283                 req = nfs_list_entry(data->pages.next);
1284                 nfs_list_remove_request(req);
1285                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1286                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1287                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1288                                 BDI_RECLAIMABLE);
1289
1290                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1291                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1292                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1293                         req->wb_bytes,
1294                         (long long)req_offset(req));
1295                 if (status < 0) {
1296                         nfs_context_set_write_error(req->wb_context, status);
1297                         nfs_inode_remove_request(req);
1298                         dprintk(", error = %d\n", status);
1299                         goto next;
1300                 }
1301
1302                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1303                  * returned by the server against all stored verfs. */
1304                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1305                         /* We have a match */
1306                         nfs_inode_remove_request(req);
1307                         dprintk(" OK\n");
1308                         goto next;
1309                 }
1310                 /* We have a mismatch. Write the page again */
1311                 dprintk(" mismatch\n");
1312                 nfs_mark_request_dirty(req);
1313         next:
1314                 nfs_clear_page_tag_locked(req);
1315         }
1316         nfs_commitdata_release(calldata);
1317 }
1318
1319 static const struct rpc_call_ops nfs_commit_ops = {
1320         .rpc_call_done = nfs_commit_done,
1321         .rpc_release = nfs_commit_release,
1322 };
1323
1324 int nfs_commit_inode(struct inode *inode, int how)
1325 {
1326         LIST_HEAD(head);
1327         int res;
1328
1329         spin_lock(&inode->i_lock);
1330         res = nfs_scan_commit(inode, &head, 0, 0);
1331         spin_unlock(&inode->i_lock);
1332         if (res) {
1333                 int error = nfs_commit_list(inode, &head, how);
1334                 if (error < 0)
1335                         return error;
1336         }
1337         return res;
1338 }
1339 #else
1340 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1341 {
1342         return 0;
1343 }
1344 #endif
1345
1346 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1347 {
1348         struct inode *inode = mapping->host;
1349         pgoff_t idx_start, idx_end;
1350         unsigned int npages = 0;
1351         LIST_HEAD(head);
1352         int nocommit = how & FLUSH_NOCOMMIT;
1353         long pages, ret;
1354
1355         /* FIXME */
1356         if (wbc->range_cyclic)
1357                 idx_start = 0;
1358         else {
1359                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1360                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1361                 if (idx_end > idx_start) {
1362                         pgoff_t l_npages = 1 + idx_end - idx_start;
1363                         npages = l_npages;
1364                         if (sizeof(npages) != sizeof(l_npages) &&
1365                                         (pgoff_t)npages != l_npages)
1366                                 npages = 0;
1367                 }
1368         }
1369         how &= ~FLUSH_NOCOMMIT;
1370         spin_lock(&inode->i_lock);
1371         do {
1372                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1373                 if (ret != 0)
1374                         continue;
1375                 if (nocommit)
1376                         break;
1377                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1378                 if (pages == 0)
1379                         break;
1380                 if (how & FLUSH_INVALIDATE) {
1381                         spin_unlock(&inode->i_lock);
1382                         nfs_cancel_commit_list(&head);
1383                         ret = pages;
1384                         spin_lock(&inode->i_lock);
1385                         continue;
1386                 }
1387                 pages += nfs_scan_commit(inode, &head, 0, 0);
1388                 spin_unlock(&inode->i_lock);
1389                 ret = nfs_commit_list(inode, &head, how);
1390                 spin_lock(&inode->i_lock);
1391
1392         } while (ret >= 0);
1393         spin_unlock(&inode->i_lock);
1394         return ret;
1395 }
1396
1397 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1398 {
1399         int ret;
1400
1401         ret = nfs_writepages(mapping, wbc);
1402         if (ret < 0)
1403                 goto out;
1404         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1405         if (ret < 0)
1406                 goto out;
1407         return 0;
1408 out:
1409         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1410         return ret;
1411 }
1412
1413 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1414 static int nfs_write_mapping(struct address_space *mapping, int how)
1415 {
1416         struct writeback_control wbc = {
1417                 .bdi = mapping->backing_dev_info,
1418                 .sync_mode = WB_SYNC_NONE,
1419                 .nr_to_write = LONG_MAX,
1420                 .for_writepages = 1,
1421                 .range_cyclic = 1,
1422         };
1423         int ret;
1424
1425         ret = __nfs_write_mapping(mapping, &wbc, how);
1426         if (ret < 0)
1427                 return ret;
1428         wbc.sync_mode = WB_SYNC_ALL;
1429         return __nfs_write_mapping(mapping, &wbc, how);
1430 }
1431
1432 /*
1433  * flush the inode to disk.
1434  */
1435 int nfs_wb_all(struct inode *inode)
1436 {
1437         return nfs_write_mapping(inode->i_mapping, 0);
1438 }
1439
1440 int nfs_wb_nocommit(struct inode *inode)
1441 {
1442         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1443 }
1444
1445 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1446 {
1447         struct nfs_page *req;
1448         loff_t range_start = page_offset(page);
1449         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1450         struct writeback_control wbc = {
1451                 .bdi = page->mapping->backing_dev_info,
1452                 .sync_mode = WB_SYNC_ALL,
1453                 .nr_to_write = LONG_MAX,
1454                 .range_start = range_start,
1455                 .range_end = range_end,
1456         };
1457         int ret = 0;
1458
1459         BUG_ON(!PageLocked(page));
1460         for (;;) {
1461                 req = nfs_page_find_request(page);
1462                 if (req == NULL)
1463                         goto out;
1464                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1465                         nfs_release_request(req);
1466                         break;
1467                 }
1468                 if (nfs_lock_request_dontget(req)) {
1469                         nfs_inode_remove_request(req);
1470                         /*
1471                          * In case nfs_inode_remove_request has marked the
1472                          * page as being dirty
1473                          */
1474                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1475                         nfs_unlock_request(req);
1476                         break;
1477                 }
1478                 ret = nfs_wait_on_request(req);
1479                 if (ret < 0)
1480                         goto out;
1481         }
1482         if (!PagePrivate(page))
1483                 return 0;
1484         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1485 out:
1486         return ret;
1487 }
1488
1489 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1490                                 int how)
1491 {
1492         loff_t range_start = page_offset(page);
1493         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1494         struct writeback_control wbc = {
1495                 .bdi = page->mapping->backing_dev_info,
1496                 .sync_mode = WB_SYNC_ALL,
1497                 .nr_to_write = LONG_MAX,
1498                 .range_start = range_start,
1499                 .range_end = range_end,
1500         };
1501         int ret;
1502
1503         do {
1504                 if (clear_page_dirty_for_io(page)) {
1505                         ret = nfs_writepage_locked(page, &wbc);
1506                         if (ret < 0)
1507                                 goto out_error;
1508                 } else if (!PagePrivate(page))
1509                         break;
1510                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1511                 if (ret < 0)
1512                         goto out_error;
1513         } while (PagePrivate(page));
1514         return 0;
1515 out_error:
1516         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1517         return ret;
1518 }
1519
1520 /*
1521  * Write back all requests on one page - we do this before reading it.
1522  */
1523 int nfs_wb_page(struct inode *inode, struct page* page)
1524 {
1525         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1526 }
1527
1528 int __init nfs_init_writepagecache(void)
1529 {
1530         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1531                                              sizeof(struct nfs_write_data),
1532                                              0, SLAB_HWCACHE_ALIGN,
1533                                              NULL);
1534         if (nfs_wdata_cachep == NULL)
1535                 return -ENOMEM;
1536
1537         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1538                                                      nfs_wdata_cachep);
1539         if (nfs_wdata_mempool == NULL)
1540                 return -ENOMEM;
1541
1542         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1543                                                       nfs_wdata_cachep);
1544         if (nfs_commit_mempool == NULL)
1545                 return -ENOMEM;
1546
1547         /*
1548          * NFS congestion size, scale with available memory.
1549          *
1550          *  64MB:    8192k
1551          * 128MB:   11585k
1552          * 256MB:   16384k
1553          * 512MB:   23170k
1554          *   1GB:   32768k
1555          *   2GB:   46340k
1556          *   4GB:   65536k
1557          *   8GB:   92681k
1558          *  16GB:  131072k
1559          *
1560          * This allows larger machines to have larger/more transfers.
1561          * Limit the default to 256M
1562          */
1563         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1564         if (nfs_congestion_kb > 256*1024)
1565                 nfs_congestion_kb = 256*1024;
1566
1567         return 0;
1568 }
1569
1570 void nfs_destroy_writepagecache(void)
1571 {
1572         mempool_destroy(nfs_commit_mempool);
1573         mempool_destroy(nfs_wdata_mempool);
1574         kmem_cache_destroy(nfs_wdata_cachep);
1575 }
1576