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 const struct rpc_call_ops nfs_write_partial_ops;
  43 static const struct rpc_call_ops nfs_write_full_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45
  46 static struct kmem_cache *nfs_wdata_cachep;
  47 static mempool_t *nfs_wdata_mempool;
  48 static mempool_t *nfs_commit_mempool;
  49
  50 struct nfs_write_data *nfs_commit_alloc(void)
  51 {
  52         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  53
  54         if (p) {
  55                 memset(p, 0, sizeof(*p));
  56                 INIT_LIST_HEAD(&p->pages);
  57         }
  58         return p;
  59 }
  60
  61 static void nfs_commit_rcu_free(struct rcu_head *head)
  62 {
  63         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 void nfs_commit_free(struct nfs_write_data *wdata)
  70 {
  71         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  72 }
  73
  74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  75 {
  76         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  77
  78         if (p) {
  79                 memset(p, 0, sizeof(*p));
  80                 INIT_LIST_HEAD(&p->pages);
  81                 p->npages = pagecount;
  82                 if (pagecount <= ARRAY_SIZE(p->page_array))
  83                         p->pagevec = p->page_array;
  84                 else {
  85                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  86                         if (!p->pagevec) {
  87                                 mempool_free(p, nfs_wdata_mempool);
  88                                 p = NULL;
  89                         }
  90                 }
  91         }
  92         return p;
  93 }
  94
  95 static void nfs_writedata_rcu_free(struct rcu_head *head)
  96 {
  97         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  98         if (p && (p->pagevec != &p->page_array[0]))
  99                 kfree(p->pagevec);
 100         mempool_free(p, nfs_wdata_mempool);
 101 }
 102
 103 static void nfs_writedata_free(struct nfs_write_data *wdata)
 104 {
 105         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 106 }
 107
 108 void nfs_writedata_release(void *wdata)
 109 {
 110         nfs_writedata_free(wdata);
 111 }
 112
 113 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 114 {
 115         struct nfs_page *req = NULL;
 116
 117         if (PagePrivate(page)) {
 118                 req = (struct nfs_page *)page_private(page);
 119                 if (req != NULL)
 120                         kref_get(&req->wb_kref);
 121         }
 122         return req;
 123 }
 124
 125 static struct nfs_page *nfs_page_find_request(struct page *page)
 126 {
 127         struct inode *inode = page->mapping->host;
 128         struct nfs_page *req = NULL;
 129
 130         spin_lock(&inode->i_lock);
 131         req = nfs_page_find_request_locked(page);
 132         spin_unlock(&inode->i_lock);
 133         return req;
 134 }
 135
 136 /* Adjust the file length if we're writing beyond the end */
 137 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 138 {
 139         struct inode *inode = page->mapping->host;
 140         loff_t end, i_size = i_size_read(inode);
 141         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 142
 143         if (i_size > 0 && page->index < end_index)
 144                 return;
 145         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 146         if (i_size >= end)
 147                 return;
 148         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 149         i_size_write(inode, end);
 150 }
 151
 152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 153 static void nfs_set_pageerror(struct page *page)
 154 {
 155         SetPageError(page);
 156         nfs_zap_mapping(page->mapping->host, page->mapping);
 157 }
 158
 159 /* We can set the PG_uptodate flag if we see that a write request
 160  * covers the full page.
 161  */
 162 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 163 {
 164         if (PageUptodate(page))
 165                 return;
 166         if (base != 0)
 167                 return;
 168         if (count != nfs_page_length(page))
 169                 return;
 170         if (count != PAGE_CACHE_SIZE)
 171                 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
 172         SetPageUptodate(page);
 173 }
 174
 175 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 176                 unsigned int offset, unsigned int count)
 177 {
 178         struct nfs_page *req;
 179         int ret;
 180
 181         for (;;) {
 182                 req = nfs_update_request(ctx, page, offset, count);
 183                 if (!IS_ERR(req))
 184                         break;
 185                 ret = PTR_ERR(req);
 186                 if (ret != -EBUSY)
 187                         return ret;
 188                 ret = nfs_wb_page(page->mapping->host, page);
 189                 if (ret != 0)
 190                         return ret;
 191         }
 192         /* Update file length */
 193         nfs_grow_file(page, offset, count);
 194         nfs_unlock_request(req);
 195         return 0;
 196 }
 197
 198 static int wb_priority(struct writeback_control *wbc)
 199 {
 200         if (wbc->for_reclaim)
 201                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 202         if (wbc->for_kupdate)
 203                 return FLUSH_LOWPRI;
 204         return 0;
 205 }
 206
 207 /*
 208  * NFS congestion control
 209  */
 210
 211 int nfs_congestion_kb;
 212
 213 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 214 #define NFS_CONGESTION_OFF_THRESH       \
 215         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 216
 217 static int nfs_set_page_writeback(struct page *page)
 218 {
 219         int ret = test_set_page_writeback(page);
 220
 221         if (!ret) {
 222                 struct inode *inode = page->mapping->host;
 223                 struct nfs_server *nfss = NFS_SERVER(inode);
 224
 225                 if (atomic_long_inc_return(&nfss->writeback) >
 226                                 NFS_CONGESTION_ON_THRESH)
 227                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 228         }
 229         return ret;
 230 }
 231
 232 static void nfs_end_page_writeback(struct page *page)
 233 {
 234         struct inode *inode = page->mapping->host;
 235         struct nfs_server *nfss = NFS_SERVER(inode);
 236
 237         end_page_writeback(page);
 238         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 239                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 240                 congestion_end(WRITE);
 241         }
 242 }
 243
 244 /*
 245  * Find an associated nfs write request, and prepare to flush it out
 246  * May return an error if the user signalled nfs_wait_on_request().
 247  */
 248 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 249                                 struct page *page)
 250 {
 251         struct inode *inode = page->mapping->host;
 252         struct nfs_inode *nfsi = NFS_I(inode);
 253         struct nfs_page *req;
 254         int ret;
 255
 256         spin_lock(&inode->i_lock);
 257         for(;;) {
 258                 req = nfs_page_find_request_locked(page);
 259                 if (req == NULL) {
 260                         spin_unlock(&inode->i_lock);
 261                         return 0;
 262                 }
 263                 if (nfs_lock_request_dontget(req))
 264                         break;
 265                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 266                  *       then the call to nfs_lock_request_dontget() will always
 267                  *       succeed provided that someone hasn't already marked the
 268                  *       request as dirty (in which case we don't care).
 269                  */
 270                 spin_unlock(&inode->i_lock);
 271                 ret = nfs_wait_on_request(req);
 272                 nfs_release_request(req);
 273                 if (ret != 0)
 274                         return ret;
 275                 spin_lock(&inode->i_lock);
 276         }
 277         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 278                 /* This request is marked for commit */
 279                 spin_unlock(&inode->i_lock);
 280                 nfs_unlock_request(req);
 281                 nfs_pageio_complete(pgio);
 282                 return 0;
 283         }
 284         if (nfs_set_page_writeback(page) != 0) {
 285                 spin_unlock(&inode->i_lock);
 286                 BUG();
 287         }
 288         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 289                         NFS_PAGE_TAG_LOCKED);
 290         spin_unlock(&inode->i_lock);
 291         nfs_pageio_add_request(pgio, req);
 292         return 0;
 293 }
 294
 295 /*
 296  * Write an mmapped page to the server.
 297  */
 298 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 299 {
 300         struct nfs_pageio_descriptor mypgio, *pgio;
 301         struct inode *inode = page->mapping->host;
 302         int err;
 303
 304         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 305         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 306
 307         if (wbc->for_writepages)
 308                 pgio = wbc->fs_private;
 309         else {
 310                 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
 311                 pgio = &mypgio;
 312         }
 313
 314         nfs_pageio_cond_complete(pgio, page->index);
 315
 316         err = nfs_page_async_flush(pgio, page);
 317
 318         if (!wbc->for_writepages)
 319                 nfs_pageio_complete(pgio);
 320         return err;
 321 }
 322
 323 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 324 {
 325         int err;
 326
 327         err = nfs_writepage_locked(page, wbc);
 328         unlock_page(page);
 329         return err;
 330 }
 331
 332 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 333 {
 334         struct inode *inode = mapping->host;
 335         struct nfs_pageio_descriptor pgio;
 336         int err;
 337
 338         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 339
 340         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 341         wbc->fs_private = &pgio;
 342         err = generic_writepages(mapping, wbc);
 343         nfs_pageio_complete(&pgio);
 344         if (err)
 345                 return err;
 346         if (pgio.pg_error)
 347                 return pgio.pg_error;
 348         return 0;
 349 }
 350
 351 /*
 352  * Insert a write request into an inode
 353  */
 354 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 355 {
 356         struct nfs_inode *nfsi = NFS_I(inode);
 357         int error;
 358
 359         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 360         BUG_ON(error == -EEXIST);
 361         if (error)
 362                 return error;
 363         if (!nfsi->npages) {
 364                 igrab(inode);
 365                 nfs_begin_data_update(inode);
 366                 if (nfs_have_delegation(inode, FMODE_WRITE))
 367                         nfsi->change_attr++;
 368         }
 369         SetPagePrivate(req->wb_page);
 370         set_page_private(req->wb_page, (unsigned long)req);
 371         nfsi->npages++;
 372         kref_get(&req->wb_kref);
 373         return 0;
 374 }
 375
 376 /*
 377  * Remove a write request from an inode
 378  */
 379 static void nfs_inode_remove_request(struct nfs_page *req)
 380 {
 381         struct inode *inode = req->wb_context->path.dentry->d_inode;
 382         struct nfs_inode *nfsi = NFS_I(inode);
 383
 384         BUG_ON (!NFS_WBACK_BUSY(req));
 385
 386         spin_lock(&inode->i_lock);
 387         set_page_private(req->wb_page, 0);
 388         ClearPagePrivate(req->wb_page);
 389         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 390         nfsi->npages--;
 391         if (!nfsi->npages) {
 392                 spin_unlock(&inode->i_lock);
 393                 nfs_end_data_update(inode);
 394                 iput(inode);
 395         } else
 396                 spin_unlock(&inode->i_lock);
 397         nfs_clear_request(req);
 398         nfs_release_request(req);
 399 }
 400
 401 static void
 402 nfs_redirty_request(struct nfs_page *req)
 403 {
 404         __set_page_dirty_nobuffers(req->wb_page);
 405 }
 406
 407 /*
 408  * Check if a request is dirty
 409  */
 410 static inline int
 411 nfs_dirty_request(struct nfs_page *req)
 412 {
 413         struct page *page = req->wb_page;
 414
 415         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 416                 return 0;
 417         return !PageWriteback(req->wb_page);
 418 }
 419
 420 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 421 /*
 422  * Add a request to the inode's commit list.
 423  */
 424 static void
 425 nfs_mark_request_commit(struct nfs_page *req)
 426 {
 427         struct inode *inode = req->wb_context->path.dentry->d_inode;
 428         struct nfs_inode *nfsi = NFS_I(inode);
 429
 430         spin_lock(&inode->i_lock);
 431         nfsi->ncommit++;
 432         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 433         radix_tree_tag_set(&nfsi->nfs_page_tree,
 434                         req->wb_index,
 435                         NFS_PAGE_TAG_COMMIT);
 436         spin_unlock(&inode->i_lock);
 437         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 438         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 439 }
 440
 441 static inline
 442 int nfs_write_need_commit(struct nfs_write_data *data)
 443 {
 444         return data->verf.committed != NFS_FILE_SYNC;
 445 }
 446
 447 static inline
 448 int nfs_reschedule_unstable_write(struct nfs_page *req)
 449 {
 450         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 451                 nfs_mark_request_commit(req);
 452                 return 1;
 453         }
 454         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 455                 nfs_redirty_request(req);
 456                 return 1;
 457         }
 458         return 0;
 459 }
 460 #else
 461 static inline void
 462 nfs_mark_request_commit(struct nfs_page *req)
 463 {
 464 }
 465
 466 static inline
 467 int nfs_write_need_commit(struct nfs_write_data *data)
 468 {
 469         return 0;
 470 }
 471
 472 static inline
 473 int nfs_reschedule_unstable_write(struct nfs_page *req)
 474 {
 475         return 0;
 476 }
 477 #endif
 478
 479 /*
 480  * Wait for a request to complete.
 481  *
 482  * Interruptible by signals only if mounted with intr flag.
 483  */
 484 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 485 {
 486         struct nfs_inode *nfsi = NFS_I(inode);
 487         struct nfs_page *req;
 488         pgoff_t idx_end, next;
 489         unsigned int            res = 0;
 490         int                     error;
 491
 492         if (npages == 0)
 493                 idx_end = ~0;
 494         else
 495                 idx_end = idx_start + npages - 1;
 496
 497         next = idx_start;
 498         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 499                 if (req->wb_index > idx_end)
 500                         break;
 501
 502                 next = req->wb_index + 1;
 503                 BUG_ON(!NFS_WBACK_BUSY(req));
 504
 505                 kref_get(&req->wb_kref);
 506                 spin_unlock(&inode->i_lock);
 507                 error = nfs_wait_on_request(req);
 508                 nfs_release_request(req);
 509                 spin_lock(&inode->i_lock);
 510                 if (error < 0)
 511                         return error;
 512                 res++;
 513         }
 514         return res;
 515 }
 516
 517 static void nfs_cancel_commit_list(struct list_head *head)
 518 {
 519         struct nfs_page *req;
 520
 521         while(!list_empty(head)) {
 522                 req = nfs_list_entry(head->next);
 523                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 524                 nfs_list_remove_request(req);
 525                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 526                 nfs_inode_remove_request(req);
 527                 nfs_unlock_request(req);
 528         }
 529 }
 530
 531 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 532 /*
 533  * nfs_scan_commit - Scan an inode for commit requests
 534  * @inode: NFS inode to scan
 535  * @dst: destination list
 536  * @idx_start: lower bound of page->index to scan.
 537  * @npages: idx_start + npages sets the upper bound to scan.
 538  *
 539  * Moves requests from the inode's 'commit' request list.
 540  * The requests are *not* checked to ensure that they form a contiguous set.
 541  */
 542 static int
 543 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 544 {
 545         struct nfs_inode *nfsi = NFS_I(inode);
 546         int res = 0;
 547
 548         if (nfsi->ncommit != 0) {
 549                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 550                                 NFS_PAGE_TAG_COMMIT);
 551                 nfsi->ncommit -= res;
 552         }
 553         return res;
 554 }
 555 #else
 556 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 557 {
 558         return 0;
 559 }
 560 #endif
 561
 562 /*
 563  * Try to update any existing write request, or create one if there is none.
 564  * In order to match, the request's credentials must match those of
 565  * the calling process.
 566  *
 567  * Note: Should always be called with the Page Lock held!
 568  */
 569 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 570                 struct page *page, unsigned int offset, unsigned int bytes)
 571 {
 572         struct address_space *mapping = page->mapping;
 573         struct inode *inode = mapping->host;
 574         struct nfs_page         *req, *new = NULL;
 575         pgoff_t         rqend, end;
 576
 577         end = offset + bytes;
 578
 579         for (;;) {
 580                 /* Loop over all inode entries and see if we find
 581                  * A request for the page we wish to update
 582                  */
 583                 spin_lock(&inode->i_lock);
 584                 req = nfs_page_find_request_locked(page);
 585                 if (req) {
 586                         if (!nfs_lock_request_dontget(req)) {
 587                                 int error;
 588
 589                                 spin_unlock(&inode->i_lock);
 590                                 error = nfs_wait_on_request(req);
 591                                 nfs_release_request(req);
 592                                 if (error < 0) {
 593                                         if (new)
 594                                                 nfs_release_request(new);
 595                                         return ERR_PTR(error);
 596                                 }
 597                                 continue;
 598                         }
 599                         spin_unlock(&inode->i_lock);
 600                         if (new)
 601                                 nfs_release_request(new);
 602                         break;
 603                 }
 604
 605                 if (new) {
 606                         int error;
 607                         nfs_lock_request_dontget(new);
 608                         error = nfs_inode_add_request(inode, new);
 609                         if (error) {
 610                                 spin_unlock(&inode->i_lock);
 611                                 nfs_unlock_request(new);
 612                                 return ERR_PTR(error);
 613                         }
 614                         spin_unlock(&inode->i_lock);
 615                         return new;
 616                 }
 617                 spin_unlock(&inode->i_lock);
 618
 619                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 620                 if (IS_ERR(new))
 621                         return new;
 622         }
 623
 624         /* We have a request for our page.
 625          * If the creds don't match, or the
 626          * page addresses don't match,
 627          * tell the caller to wait on the conflicting
 628          * request.
 629          */
 630         rqend = req->wb_offset + req->wb_bytes;
 631         if (req->wb_context != ctx
 632             || req->wb_page != page
 633             || !nfs_dirty_request(req)
 634             || offset > rqend || end < req->wb_offset) {
 635                 nfs_unlock_request(req);
 636                 return ERR_PTR(-EBUSY);
 637         }
 638
 639         /* Okay, the request matches. Update the region */
 640         if (offset < req->wb_offset) {
 641                 req->wb_offset = offset;
 642                 req->wb_pgbase = offset;
 643                 req->wb_bytes = rqend - req->wb_offset;
 644         }
 645
 646         if (end > rqend)
 647                 req->wb_bytes = end - req->wb_offset;
 648
 649         return req;
 650 }
 651
 652 int nfs_flush_incompatible(struct file *file, struct page *page)
 653 {
 654         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 655         struct nfs_page *req;
 656         int do_flush, status;
 657         /*
 658          * Look for a request corresponding to this page. If there
 659          * is one, and it belongs to another file, we flush it out
 660          * before we try to copy anything into the page. Do this
 661          * due to the lack of an ACCESS-type call in NFSv2.
 662          * Also do the same if we find a request from an existing
 663          * dropped page.
 664          */
 665         do {
 666                 req = nfs_page_find_request(page);
 667                 if (req == NULL)
 668                         return 0;
 669                 do_flush = req->wb_page != page || req->wb_context != ctx
 670                         || !nfs_dirty_request(req);
 671                 nfs_release_request(req);
 672                 if (!do_flush)
 673                         return 0;
 674                 status = nfs_wb_page(page->mapping->host, page);
 675         } while (status == 0);
 676         return status;
 677 }
 678
 679 /*
 680  * Update and possibly write a cached page of an NFS file.
 681  *
 682  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 683  * things with a page scheduled for an RPC call (e.g. invalidate it).
 684  */
 685 int nfs_updatepage(struct file *file, struct page *page,
 686                 unsigned int offset, unsigned int count)
 687 {
 688         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 689         struct inode    *inode = page->mapping->host;
 690         int             status = 0;
 691
 692         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 693
 694         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 695                 file->f_path.dentry->d_parent->d_name.name,
 696                 file->f_path.dentry->d_name.name, count,
 697                 (long long)(page_offset(page) +offset));
 698
 699         /* If we're not using byte range locks, and we know the page
 700          * is entirely in cache, it may be more efficient to avoid
 701          * fragmenting write requests.
 702          */
 703         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 704                 count = max(count + offset, nfs_page_length(page));
 705                 offset = 0;
 706         }
 707
 708         status = nfs_writepage_setup(ctx, page, offset, count);
 709         __set_page_dirty_nobuffers(page);
 710
 711         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 712                         status, (long long)i_size_read(inode));
 713         if (status < 0)
 714                 nfs_set_pageerror(page);
 715         return status;
 716 }
 717
 718 static void nfs_writepage_release(struct nfs_page *req)
 719 {
 720
 721         if (PageError(req->wb_page)) {
 722                 nfs_end_page_writeback(req->wb_page);
 723                 nfs_inode_remove_request(req);
 724         } else if (!nfs_reschedule_unstable_write(req)) {
 725                 /* Set the PG_uptodate flag */
 726                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 727                 nfs_end_page_writeback(req->wb_page);
 728                 nfs_inode_remove_request(req);
 729         } else
 730                 nfs_end_page_writeback(req->wb_page);
 731         nfs_clear_page_tag_locked(req);
 732 }
 733
 734 static inline int flush_task_priority(int how)
 735 {
 736         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 737                 case FLUSH_HIGHPRI:
 738                         return RPC_PRIORITY_HIGH;
 739                 case FLUSH_LOWPRI:
 740                         return RPC_PRIORITY_LOW;
 741         }
 742         return RPC_PRIORITY_NORMAL;
 743 }
 744
 745 /*
 746  * Set up the argument/result storage required for the RPC call.
 747  */
 748 static void nfs_write_rpcsetup(struct nfs_page *req,
 749                 struct nfs_write_data *data,
 750                 const struct rpc_call_ops *call_ops,
 751                 unsigned int count, unsigned int offset,
 752                 int how)
 753 {
 754         struct inode            *inode;
 755         int flags;
 756
 757         /* Set up the RPC argument and reply structs
 758          * NB: take care not to mess about with data->commit et al. */
 759
 760         data->req = req;
 761         data->inode = inode = req->wb_context->path.dentry->d_inode;
 762         data->cred = req->wb_context->cred;
 763
 764         data->args.fh     = NFS_FH(inode);
 765         data->args.offset = req_offset(req) + offset;
 766         data->args.pgbase = req->wb_pgbase + offset;
 767         data->args.pages  = data->pagevec;
 768         data->args.count  = count;
 769         data->args.context = req->wb_context;
 770
 771         data->res.fattr   = &data->fattr;
 772         data->res.count   = count;
 773         data->res.verf    = &data->verf;
 774         nfs_fattr_init(&data->fattr);
 775
 776         /* Set up the initial task struct.  */
 777         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 778         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 779         NFS_PROTO(inode)->write_setup(data, how);
 780
 781         data->task.tk_priority = flush_task_priority(how);
 782         data->task.tk_cookie = (unsigned long)inode;
 783
 784         dprintk("NFS: %5u initiated write call "
 785                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 786                 data->task.tk_pid,
 787                 inode->i_sb->s_id,
 788                 (long long)NFS_FILEID(inode),
 789                 count,
 790                 (unsigned long long)data->args.offset);
 791 }
 792
 793 static void nfs_execute_write(struct nfs_write_data *data)
 794 {
 795         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 796         sigset_t oldset;
 797
 798         rpc_clnt_sigmask(clnt, &oldset);
 799         rpc_execute(&data->task);
 800         rpc_clnt_sigunmask(clnt, &oldset);
 801 }
 802
 803 /*
 804  * Generate multiple small requests to write out a single
 805  * contiguous dirty area on one page.
 806  */
 807 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 808 {
 809         struct nfs_page *req = nfs_list_entry(head->next);
 810         struct page *page = req->wb_page;
 811         struct nfs_write_data *data;
 812         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 813         unsigned int offset;
 814         int requests = 0;
 815         LIST_HEAD(list);
 816
 817         nfs_list_remove_request(req);
 818
 819         nbytes = count;
 820         do {
 821                 size_t len = min(nbytes, wsize);
 822
 823                 data = nfs_writedata_alloc(1);
 824                 if (!data)
 825                         goto out_bad;
 826                 list_add(&data->pages, &list);
 827                 requests++;
 828                 nbytes -= len;
 829         } while (nbytes != 0);
 830         atomic_set(&req->wb_complete, requests);
 831
 832         ClearPageError(page);
 833         offset = 0;
 834         nbytes = count;
 835         do {
 836                 data = list_entry(list.next, struct nfs_write_data, pages);
 837                 list_del_init(&data->pages);
 838
 839                 data->pagevec[0] = page;
 840
 841                 if (nbytes < wsize)
 842                         wsize = nbytes;
 843                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 844                                    wsize, offset, how);
 845                 offset += wsize;
 846                 nbytes -= wsize;
 847                 nfs_execute_write(data);
 848         } while (nbytes != 0);
 849
 850         return 0;
 851
 852 out_bad:
 853         while (!list_empty(&list)) {
 854                 data = list_entry(list.next, struct nfs_write_data, pages);
 855                 list_del(&data->pages);
 856                 nfs_writedata_release(data);
 857         }
 858         nfs_redirty_request(req);
 859         nfs_end_page_writeback(req->wb_page);
 860         nfs_clear_page_tag_locked(req);
 861         return -ENOMEM;
 862 }
 863
 864 /*
 865  * Create an RPC task for the given write request and kick it.
 866  * The page must have been locked by the caller.
 867  *
 868  * It may happen that the page we're passed is not marked dirty.
 869  * This is the case if nfs_updatepage detects a conflicting request
 870  * that has been written but not committed.
 871  */
 872 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 873 {
 874         struct nfs_page         *req;
 875         struct page             **pages;
 876         struct nfs_write_data   *data;
 877
 878         data = nfs_writedata_alloc(npages);
 879         if (!data)
 880                 goto out_bad;
 881
 882         pages = data->pagevec;
 883         while (!list_empty(head)) {
 884                 req = nfs_list_entry(head->next);
 885                 nfs_list_remove_request(req);
 886                 nfs_list_add_request(req, &data->pages);
 887                 ClearPageError(req->wb_page);
 888                 *pages++ = req->wb_page;
 889         }
 890         req = nfs_list_entry(data->pages.next);
 891
 892         /* Set up the argument struct */
 893         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 894
 895         nfs_execute_write(data);
 896         return 0;
 897  out_bad:
 898         while (!list_empty(head)) {
 899                 req = nfs_list_entry(head->next);
 900                 nfs_list_remove_request(req);
 901                 nfs_redirty_request(req);
 902                 nfs_end_page_writeback(req->wb_page);
 903                 nfs_clear_page_tag_locked(req);
 904         }
 905         return -ENOMEM;
 906 }
 907
 908 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 909                                   struct inode *inode, int ioflags)
 910 {
 911         int wsize = NFS_SERVER(inode)->wsize;
 912
 913         if (wsize < PAGE_CACHE_SIZE)
 914                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 915         else
 916                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 917 }
 918
 919 /*
 920  * Handle a write reply that flushed part of a page.
 921  */
 922 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 923 {
 924         struct nfs_write_data   *data = calldata;
 925         struct nfs_page         *req = data->req;
 926         struct page             *page = req->wb_page;
 927
 928         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 929                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 930                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 931                 req->wb_bytes,
 932                 (long long)req_offset(req));
 933
 934         if (nfs_writeback_done(task, data) != 0)
 935                 return;
 936
 937         if (task->tk_status < 0) {
 938                 nfs_set_pageerror(page);
 939                 req->wb_context->error = task->tk_status;
 940                 dprintk(", error = %d\n", task->tk_status);
 941                 goto out;
 942         }
 943
 944         if (nfs_write_need_commit(data)) {
 945                 struct inode *inode = page->mapping->host;
 946
 947                 spin_lock(&inode->i_lock);
 948                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 949                         /* Do nothing we need to resend the writes */
 950                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 951                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 952                         dprintk(" defer commit\n");
 953                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 954                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 955                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 956                         dprintk(" server reboot detected\n");
 957                 }
 958                 spin_unlock(&inode->i_lock);
 959         } else
 960                 dprintk(" OK\n");
 961
 962 out:
 963         if (atomic_dec_and_test(&req->wb_complete))
 964                 nfs_writepage_release(req);
 965 }
 966
 967 static const struct rpc_call_ops nfs_write_partial_ops = {
 968         .rpc_call_done = nfs_writeback_done_partial,
 969         .rpc_release = nfs_writedata_release,
 970 };
 971
 972 /*
 973  * Handle a write reply that flushes a whole page.
 974  *
 975  * FIXME: There is an inherent race with invalidate_inode_pages and
 976  *        writebacks since the page->count is kept > 1 for as long
 977  *        as the page has a write request pending.
 978  */
 979 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
 980 {
 981         struct nfs_write_data   *data = calldata;
 982         struct nfs_page         *req;
 983         struct page             *page;
 984
 985         if (nfs_writeback_done(task, data) != 0)
 986                 return;
 987
 988         /* Update attributes as result of writeback. */
 989         while (!list_empty(&data->pages)) {
 990                 req = nfs_list_entry(data->pages.next);
 991                 nfs_list_remove_request(req);
 992                 page = req->wb_page;
 993
 994                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
 995                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
 996                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 997                         req->wb_bytes,
 998                         (long long)req_offset(req));
 999
1000                 if (task->tk_status < 0) {
1001                         nfs_set_pageerror(page);
1002                         req->wb_context->error = task->tk_status;
1003                         dprintk(", error = %d\n", task->tk_status);
1004                         goto remove_request;
1005                 }
1006
1007                 if (nfs_write_need_commit(data)) {
1008                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1009                         nfs_mark_request_commit(req);
1010                         nfs_end_page_writeback(page);
1011                         dprintk(" marked for commit\n");
1012                         goto next;
1013                 }
1014                 /* Set the PG_uptodate flag? */
1015                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1016                 dprintk(" OK\n");
1017 remove_request:
1018                 nfs_end_page_writeback(page);
1019                 nfs_inode_remove_request(req);
1020         next:
1021                 nfs_clear_page_tag_locked(req);
1022         }
1023 }
1024
1025 static const struct rpc_call_ops nfs_write_full_ops = {
1026         .rpc_call_done = nfs_writeback_done_full,
1027         .rpc_release = nfs_writedata_release,
1028 };
1029
1030
1031 /*
1032  * This function is called when the WRITE call is complete.
1033  */
1034 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1035 {
1036         struct nfs_writeargs    *argp = &data->args;
1037         struct nfs_writeres     *resp = &data->res;
1038         int status;
1039
1040         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1041                 task->tk_pid, task->tk_status);
1042
1043         /*
1044          * ->write_done will attempt to use post-op attributes to detect
1045          * conflicting writes by other clients.  A strict interpretation
1046          * of close-to-open would allow us to continue caching even if
1047          * another writer had changed the file, but some applications
1048          * depend on tighter cache coherency when writing.
1049          */
1050         status = NFS_PROTO(data->inode)->write_done(task, data);
1051         if (status != 0)
1052                 return status;
1053         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1054
1055 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1056         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1057                 /* We tried a write call, but the server did not
1058                  * commit data to stable storage even though we
1059                  * requested it.
1060                  * Note: There is a known bug in Tru64 < 5.0 in which
1061                  *       the server reports NFS_DATA_SYNC, but performs
1062                  *       NFS_FILE_SYNC. We therefore implement this checking
1063                  *       as a dprintk() in order to avoid filling syslog.
1064                  */
1065                 static unsigned long    complain;
1066
1067                 if (time_before(complain, jiffies)) {
1068                         dprintk("NFS: faulty NFS server %s:"
1069                                 " (committed = %d) != (stable = %d)\n",
1070                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1071                                 resp->verf->committed, argp->stable);
1072                         complain = jiffies + 300 * HZ;
1073                 }
1074         }
1075 #endif
1076         /* Is this a short write? */
1077         if (task->tk_status >= 0 && resp->count < argp->count) {
1078                 static unsigned long    complain;
1079
1080                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1081
1082                 /* Has the server at least made some progress? */
1083                 if (resp->count != 0) {
1084                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1085                         if (resp->verf->committed != NFS_UNSTABLE) {
1086                                 /* Resend from where the server left off */
1087                                 argp->offset += resp->count;
1088                                 argp->pgbase += resp->count;
1089                                 argp->count -= resp->count;
1090                         } else {
1091                                 /* Resend as a stable write in order to avoid
1092                                  * headaches in the case of a server crash.
1093                                  */
1094                                 argp->stable = NFS_FILE_SYNC;
1095                         }
1096                         rpc_restart_call(task);
1097                         return -EAGAIN;
1098                 }
1099                 if (time_before(complain, jiffies)) {
1100                         printk(KERN_WARNING
1101                                "NFS: Server wrote zero bytes, expected %u.\n",
1102                                         argp->count);
1103                         complain = jiffies + 300 * HZ;
1104                 }
1105                 /* Can't do anything about it except throw an error. */
1106                 task->tk_status = -EIO;
1107         }
1108         return 0;
1109 }
1110
1111
1112 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1113 void nfs_commit_release(void *wdata)
1114 {
1115         nfs_commit_free(wdata);
1116 }
1117
1118 /*
1119  * Set up the argument/result storage required for the RPC call.
1120  */
1121 static void nfs_commit_rpcsetup(struct list_head *head,
1122                 struct nfs_write_data *data,
1123                 int how)
1124 {
1125         struct nfs_page         *first;
1126         struct inode            *inode;
1127         int flags;
1128
1129         /* Set up the RPC argument and reply structs
1130          * NB: take care not to mess about with data->commit et al. */
1131
1132         list_splice_init(head, &data->pages);
1133         first = nfs_list_entry(data->pages.next);
1134         inode = first->wb_context->path.dentry->d_inode;
1135
1136         data->inode       = inode;
1137         data->cred        = first->wb_context->cred;
1138
1139         data->args.fh     = NFS_FH(data->inode);
1140         /* Note: we always request a commit of the entire inode */
1141         data->args.offset = 0;
1142         data->args.count  = 0;
1143         data->res.count   = 0;
1144         data->res.fattr   = &data->fattr;
1145         data->res.verf    = &data->verf;
1146         nfs_fattr_init(&data->fattr);
1147
1148         /* Set up the initial task struct.  */
1149         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1150         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1151         NFS_PROTO(inode)->commit_setup(data, how);
1152
1153         data->task.tk_priority = flush_task_priority(how);
1154         data->task.tk_cookie = (unsigned long)inode;
1155
1156         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1157 }
1158
1159 /*
1160  * Commit dirty pages
1161  */
1162 static int
1163 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1164 {
1165         struct nfs_write_data   *data;
1166         struct nfs_page         *req;
1167
1168         data = nfs_commit_alloc();
1169
1170         if (!data)
1171                 goto out_bad;
1172
1173         /* Set up the argument struct */
1174         nfs_commit_rpcsetup(head, data, how);
1175
1176         nfs_execute_write(data);
1177         return 0;
1178  out_bad:
1179         while (!list_empty(head)) {
1180                 req = nfs_list_entry(head->next);
1181                 nfs_list_remove_request(req);
1182                 nfs_mark_request_commit(req);
1183                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1184                 nfs_clear_page_tag_locked(req);
1185         }
1186         return -ENOMEM;
1187 }
1188
1189 /*
1190  * COMMIT call returned
1191  */
1192 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1193 {
1194         struct nfs_write_data   *data = calldata;
1195         struct nfs_page         *req;
1196
1197         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1198                                 task->tk_pid, task->tk_status);
1199
1200         /* Call the NFS version-specific code */
1201         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1202                 return;
1203
1204         while (!list_empty(&data->pages)) {
1205                 req = nfs_list_entry(data->pages.next);
1206                 nfs_list_remove_request(req);
1207                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1208                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1209
1210                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1211                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1212                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1213                         req->wb_bytes,
1214                         (long long)req_offset(req));
1215                 if (task->tk_status < 0) {
1216                         req->wb_context->error = task->tk_status;
1217                         nfs_inode_remove_request(req);
1218                         dprintk(", error = %d\n", task->tk_status);
1219                         goto next;
1220                 }
1221
1222                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1223                  * returned by the server against all stored verfs. */
1224                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1225                         /* We have a match */
1226                         /* Set the PG_uptodate flag */
1227                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1228                                         req->wb_bytes);
1229                         nfs_inode_remove_request(req);
1230                         dprintk(" OK\n");
1231                         goto next;
1232                 }
1233                 /* We have a mismatch. Write the page again */
1234                 dprintk(" mismatch\n");
1235                 nfs_redirty_request(req);
1236         next:
1237                 nfs_clear_page_tag_locked(req);
1238         }
1239 }
1240
1241 static const struct rpc_call_ops nfs_commit_ops = {
1242         .rpc_call_done = nfs_commit_done,
1243         .rpc_release = nfs_commit_release,
1244 };
1245
1246 int nfs_commit_inode(struct inode *inode, int how)
1247 {
1248         LIST_HEAD(head);
1249         int res;
1250
1251         spin_lock(&inode->i_lock);
1252         res = nfs_scan_commit(inode, &head, 0, 0);
1253         spin_unlock(&inode->i_lock);
1254         if (res) {
1255                 int error = nfs_commit_list(inode, &head, how);
1256                 if (error < 0)
1257                         return error;
1258         }
1259         return res;
1260 }
1261 #else
1262 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1263 {
1264         return 0;
1265 }
1266 #endif
1267
1268 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1269 {
1270         struct inode *inode = mapping->host;
1271         pgoff_t idx_start, idx_end;
1272         unsigned int npages = 0;
1273         LIST_HEAD(head);
1274         int nocommit = how & FLUSH_NOCOMMIT;
1275         long pages, ret;
1276
1277         /* FIXME */
1278         if (wbc->range_cyclic)
1279                 idx_start = 0;
1280         else {
1281                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1282                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1283                 if (idx_end > idx_start) {
1284                         pgoff_t l_npages = 1 + idx_end - idx_start;
1285                         npages = l_npages;
1286                         if (sizeof(npages) != sizeof(l_npages) &&
1287                                         (pgoff_t)npages != l_npages)
1288                                 npages = 0;
1289                 }
1290         }
1291         how &= ~FLUSH_NOCOMMIT;
1292         spin_lock(&inode->i_lock);
1293         do {
1294                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1295                 if (ret != 0)
1296                         continue;
1297                 if (nocommit)
1298                         break;
1299                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1300                 if (pages == 0)
1301                         break;
1302                 if (how & FLUSH_INVALIDATE) {
1303                         spin_unlock(&inode->i_lock);
1304                         nfs_cancel_commit_list(&head);
1305                         ret = pages;
1306                         spin_lock(&inode->i_lock);
1307                         continue;
1308                 }
1309                 pages += nfs_scan_commit(inode, &head, 0, 0);
1310                 spin_unlock(&inode->i_lock);
1311                 ret = nfs_commit_list(inode, &head, how);
1312                 spin_lock(&inode->i_lock);
1313
1314         } while (ret >= 0);
1315         spin_unlock(&inode->i_lock);
1316         return ret;
1317 }
1318
1319 /*
1320  * flush the inode to disk.
1321  */
1322 int nfs_wb_all(struct inode *inode)
1323 {
1324         struct address_space *mapping = inode->i_mapping;
1325         struct writeback_control wbc = {
1326                 .bdi = mapping->backing_dev_info,
1327                 .sync_mode = WB_SYNC_ALL,
1328                 .nr_to_write = LONG_MAX,
1329                 .for_writepages = 1,
1330                 .range_cyclic = 1,
1331         };
1332         int ret;
1333
1334         ret = nfs_writepages(mapping, &wbc);
1335         if (ret < 0)
1336                 goto out;
1337         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1338         if (ret >= 0)
1339                 return 0;
1340 out:
1341         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1342         return ret;
1343 }
1344
1345 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1346 {
1347         struct writeback_control wbc = {
1348                 .bdi = mapping->backing_dev_info,
1349                 .sync_mode = WB_SYNC_ALL,
1350                 .nr_to_write = LONG_MAX,
1351                 .range_start = range_start,
1352                 .range_end = range_end,
1353                 .for_writepages = 1,
1354         };
1355         int ret;
1356
1357         ret = nfs_writepages(mapping, &wbc);
1358         if (ret < 0)
1359                 goto out;
1360         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1361         if (ret >= 0)
1362                 return 0;
1363 out:
1364         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1365         return ret;
1366 }
1367
1368 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1369 {
1370         struct nfs_page *req;
1371         loff_t range_start = page_offset(page);
1372         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1373         struct writeback_control wbc = {
1374                 .bdi = page->mapping->backing_dev_info,
1375                 .sync_mode = WB_SYNC_ALL,
1376                 .nr_to_write = LONG_MAX,
1377                 .range_start = range_start,
1378                 .range_end = range_end,
1379         };
1380         int ret = 0;
1381
1382         BUG_ON(!PageLocked(page));
1383         for (;;) {
1384                 req = nfs_page_find_request(page);
1385                 if (req == NULL)
1386                         goto out;
1387                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1388                         nfs_release_request(req);
1389                         break;
1390                 }
1391                 if (nfs_lock_request_dontget(req)) {
1392                         nfs_inode_remove_request(req);
1393                         /*
1394                          * In case nfs_inode_remove_request has marked the
1395                          * page as being dirty
1396                          */
1397                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1398                         nfs_unlock_request(req);
1399                         break;
1400                 }
1401                 ret = nfs_wait_on_request(req);
1402                 if (ret < 0)
1403                         goto out;
1404         }
1405         if (!PagePrivate(page))
1406                 return 0;
1407         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1408 out:
1409         return ret;
1410 }
1411
1412 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1413 {
1414         loff_t range_start = page_offset(page);
1415         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1416         struct writeback_control wbc = {
1417                 .bdi = page->mapping->backing_dev_info,
1418                 .sync_mode = WB_SYNC_ALL,
1419                 .nr_to_write = LONG_MAX,
1420                 .range_start = range_start,
1421                 .range_end = range_end,
1422         };
1423         int ret;
1424
1425         BUG_ON(!PageLocked(page));
1426         if (clear_page_dirty_for_io(page)) {
1427                 ret = nfs_writepage_locked(page, &wbc);
1428                 if (ret < 0)
1429                         goto out;
1430         }
1431         if (!PagePrivate(page))
1432                 return 0;
1433         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1434         if (ret >= 0)
1435                 return 0;
1436 out:
1437         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1438         return ret;
1439 }
1440
1441 /*
1442  * Write back all requests on one page - we do this before reading it.
1443  */
1444 int nfs_wb_page(struct inode *inode, struct page* page)
1445 {
1446         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1447 }
1448
1449 int __init nfs_init_writepagecache(void)
1450 {
1451         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1452                                              sizeof(struct nfs_write_data),
1453                                              0, SLAB_HWCACHE_ALIGN,
1454                                              NULL);
1455         if (nfs_wdata_cachep == NULL)
1456                 return -ENOMEM;
1457
1458         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1459                                                      nfs_wdata_cachep);
1460         if (nfs_wdata_mempool == NULL)
1461                 return -ENOMEM;
1462
1463         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1464                                                       nfs_wdata_cachep);
1465         if (nfs_commit_mempool == NULL)
1466                 return -ENOMEM;
1467
1468         /*
1469          * NFS congestion size, scale with available memory.
1470          *
1471          *  64MB:    8192k
1472          * 128MB:   11585k
1473          * 256MB:   16384k
1474          * 512MB:   23170k
1475          *   1GB:   32768k
1476          *   2GB:   46340k
1477          *   4GB:   65536k
1478          *   8GB:   92681k
1479          *  16GB:  131072k
1480          *
1481          * This allows larger machines to have larger/more transfers.
1482          * Limit the default to 256M
1483          */
1484         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1485         if (nfs_congestion_kb > 256*1024)
1486                 nfs_congestion_kb = 256*1024;
1487
1488         return 0;
1489 }
1490
1491 void nfs_destroy_writepagecache(void)
1492 {
1493         mempool_destroy(nfs_commit_mempool);
1494         mempool_destroy(nfs_wdata_mempool);
1495         kmem_cache_destroy(nfs_wdata_cachep);
1496 }
1497