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 void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 114 {
 115         ctx->error = error;
 116         smp_wmb();
 117         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 118 }
 119
 120 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 121 {
 122         struct nfs_page *req = NULL;
 123
 124         if (PagePrivate(page)) {
 125                 req = (struct nfs_page *)page_private(page);
 126                 if (req != NULL)
 127                         kref_get(&req->wb_kref);
 128         }
 129         return req;
 130 }
 131
 132 static struct nfs_page *nfs_page_find_request(struct page *page)
 133 {
 134         struct inode *inode = page->mapping->host;
 135         struct nfs_page *req = NULL;
 136
 137         spin_lock(&inode->i_lock);
 138         req = nfs_page_find_request_locked(page);
 139         spin_unlock(&inode->i_lock);
 140         return req;
 141 }
 142
 143 /* Adjust the file length if we're writing beyond the end */
 144 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 145 {
 146         struct inode *inode = page->mapping->host;
 147         loff_t end, i_size = i_size_read(inode);
 148         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 149
 150         if (i_size > 0 && page->index < end_index)
 151                 return;
 152         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 153         if (i_size >= end)
 154                 return;
 155         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 156         i_size_write(inode, end);
 157 }
 158
 159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 160 static void nfs_set_pageerror(struct page *page)
 161 {
 162         SetPageError(page);
 163         nfs_zap_mapping(page->mapping->host, page->mapping);
 164 }
 165
 166 /* We can set the PG_uptodate flag if we see that a write request
 167  * covers the full page.
 168  */
 169 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 170 {
 171         if (PageUptodate(page))
 172                 return;
 173         if (base != 0)
 174                 return;
 175         if (count != nfs_page_length(page))
 176                 return;
 177         if (count != PAGE_CACHE_SIZE)
 178                 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
 179         SetPageUptodate(page);
 180 }
 181
 182 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 183                 unsigned int offset, unsigned int count)
 184 {
 185         struct nfs_page *req;
 186         int ret;
 187
 188         for (;;) {
 189                 req = nfs_update_request(ctx, page, offset, count);
 190                 if (!IS_ERR(req))
 191                         break;
 192                 ret = PTR_ERR(req);
 193                 if (ret != -EBUSY)
 194                         return ret;
 195                 ret = nfs_wb_page(page->mapping->host, page);
 196                 if (ret != 0)
 197                         return ret;
 198         }
 199         /* Update file length */
 200         nfs_grow_file(page, offset, count);
 201         nfs_unlock_request(req);
 202         return 0;
 203 }
 204
 205 static int wb_priority(struct writeback_control *wbc)
 206 {
 207         if (wbc->for_reclaim)
 208                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 209         if (wbc->for_kupdate)
 210                 return FLUSH_LOWPRI;
 211         return 0;
 212 }
 213
 214 /*
 215  * NFS congestion control
 216  */
 217
 218 int nfs_congestion_kb;
 219
 220 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 221 #define NFS_CONGESTION_OFF_THRESH       \
 222         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 223
 224 static int nfs_set_page_writeback(struct page *page)
 225 {
 226         int ret = test_set_page_writeback(page);
 227
 228         if (!ret) {
 229                 struct inode *inode = page->mapping->host;
 230                 struct nfs_server *nfss = NFS_SERVER(inode);
 231
 232                 if (atomic_long_inc_return(&nfss->writeback) >
 233                                 NFS_CONGESTION_ON_THRESH)
 234                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 235         }
 236         return ret;
 237 }
 238
 239 static void nfs_end_page_writeback(struct page *page)
 240 {
 241         struct inode *inode = page->mapping->host;
 242         struct nfs_server *nfss = NFS_SERVER(inode);
 243
 244         end_page_writeback(page);
 245         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 246                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 247                 congestion_end(WRITE);
 248         }
 249 }
 250
 251 /*
 252  * Find an associated nfs write request, and prepare to flush it out
 253  * May return an error if the user signalled nfs_wait_on_request().
 254  */
 255 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 256                                 struct page *page)
 257 {
 258         struct inode *inode = page->mapping->host;
 259         struct nfs_inode *nfsi = NFS_I(inode);
 260         struct nfs_page *req;
 261         int ret;
 262
 263         spin_lock(&inode->i_lock);
 264         for(;;) {
 265                 req = nfs_page_find_request_locked(page);
 266                 if (req == NULL) {
 267                         spin_unlock(&inode->i_lock);
 268                         return 0;
 269                 }
 270                 if (nfs_lock_request_dontget(req))
 271                         break;
 272                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 273                  *       then the call to nfs_lock_request_dontget() will always
 274                  *       succeed provided that someone hasn't already marked the
 275                  *       request as dirty (in which case we don't care).
 276                  */
 277                 spin_unlock(&inode->i_lock);
 278                 ret = nfs_wait_on_request(req);
 279                 nfs_release_request(req);
 280                 if (ret != 0)
 281                         return ret;
 282                 spin_lock(&inode->i_lock);
 283         }
 284         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 285                 /* This request is marked for commit */
 286                 spin_unlock(&inode->i_lock);
 287                 nfs_unlock_request(req);
 288                 nfs_pageio_complete(pgio);
 289                 return 0;
 290         }
 291         if (nfs_set_page_writeback(page) != 0) {
 292                 spin_unlock(&inode->i_lock);
 293                 BUG();
 294         }
 295         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 296                         NFS_PAGE_TAG_LOCKED);
 297         spin_unlock(&inode->i_lock);
 298         nfs_pageio_add_request(pgio, req);
 299         return 0;
 300 }
 301
 302 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 303 {
 304         struct inode *inode = page->mapping->host;
 305
 306         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 307         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 308
 309         nfs_pageio_cond_complete(pgio, page->index);
 310         return nfs_page_async_flush(pgio, page);
 311 }
 312
 313 /*
 314  * Write an mmapped page to the server.
 315  */
 316 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 317 {
 318         struct nfs_pageio_descriptor pgio;
 319         int err;
 320
 321         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 322         err = nfs_do_writepage(page, wbc, &pgio);
 323         nfs_pageio_complete(&pgio);
 324         if (err < 0)
 325                 return err;
 326         if (pgio.pg_error < 0)
 327                 return pgio.pg_error;
 328         return 0;
 329 }
 330
 331 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 332 {
 333         int ret;
 334
 335         ret = nfs_writepage_locked(page, wbc);
 336         unlock_page(page);
 337         return ret;
 338 }
 339
 340 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 341 {
 342         int ret;
 343
 344         ret = nfs_do_writepage(page, wbc, data);
 345         unlock_page(page);
 346         return ret;
 347 }
 348
 349 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 350 {
 351         struct inode *inode = mapping->host;
 352         struct nfs_pageio_descriptor pgio;
 353         int err;
 354
 355         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 356
 357         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 358         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 359         nfs_pageio_complete(&pgio);
 360         if (err < 0)
 361                 return err;
 362         if (pgio.pg_error < 0)
 363                 return pgio.pg_error;
 364         return 0;
 365 }
 366
 367 /*
 368  * Insert a write request into an inode
 369  */
 370 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 371 {
 372         struct nfs_inode *nfsi = NFS_I(inode);
 373         int error;
 374
 375         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 376         BUG_ON(error == -EEXIST);
 377         if (error)
 378                 return error;
 379         if (!nfsi->npages) {
 380                 igrab(inode);
 381                 if (nfs_have_delegation(inode, FMODE_WRITE))
 382                         nfsi->change_attr++;
 383         }
 384         SetPagePrivate(req->wb_page);
 385         set_page_private(req->wb_page, (unsigned long)req);
 386         nfsi->npages++;
 387         kref_get(&req->wb_kref);
 388         return 0;
 389 }
 390
 391 /*
 392  * Remove a write request from an inode
 393  */
 394 static void nfs_inode_remove_request(struct nfs_page *req)
 395 {
 396         struct inode *inode = req->wb_context->path.dentry->d_inode;
 397         struct nfs_inode *nfsi = NFS_I(inode);
 398
 399         BUG_ON (!NFS_WBACK_BUSY(req));
 400
 401         spin_lock(&inode->i_lock);
 402         set_page_private(req->wb_page, 0);
 403         ClearPagePrivate(req->wb_page);
 404         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 405         nfsi->npages--;
 406         if (!nfsi->npages) {
 407                 spin_unlock(&inode->i_lock);
 408                 iput(inode);
 409         } else
 410                 spin_unlock(&inode->i_lock);
 411         nfs_clear_request(req);
 412         nfs_release_request(req);
 413 }
 414
 415 static void
 416 nfs_redirty_request(struct nfs_page *req)
 417 {
 418         __set_page_dirty_nobuffers(req->wb_page);
 419 }
 420
 421 /*
 422  * Check if a request is dirty
 423  */
 424 static inline int
 425 nfs_dirty_request(struct nfs_page *req)
 426 {
 427         struct page *page = req->wb_page;
 428
 429         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 430                 return 0;
 431         return !PageWriteback(req->wb_page);
 432 }
 433
 434 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 435 /*
 436  * Add a request to the inode's commit list.
 437  */
 438 static void
 439 nfs_mark_request_commit(struct nfs_page *req)
 440 {
 441         struct inode *inode = req->wb_context->path.dentry->d_inode;
 442         struct nfs_inode *nfsi = NFS_I(inode);
 443
 444         spin_lock(&inode->i_lock);
 445         nfsi->ncommit++;
 446         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 447         radix_tree_tag_set(&nfsi->nfs_page_tree,
 448                         req->wb_index,
 449                         NFS_PAGE_TAG_COMMIT);
 450         spin_unlock(&inode->i_lock);
 451         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 452         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 453 }
 454
 455 static inline
 456 int nfs_write_need_commit(struct nfs_write_data *data)
 457 {
 458         return data->verf.committed != NFS_FILE_SYNC;
 459 }
 460
 461 static inline
 462 int nfs_reschedule_unstable_write(struct nfs_page *req)
 463 {
 464         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 465                 nfs_mark_request_commit(req);
 466                 return 1;
 467         }
 468         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 469                 nfs_redirty_request(req);
 470                 return 1;
 471         }
 472         return 0;
 473 }
 474 #else
 475 static inline void
 476 nfs_mark_request_commit(struct nfs_page *req)
 477 {
 478 }
 479
 480 static inline
 481 int nfs_write_need_commit(struct nfs_write_data *data)
 482 {
 483         return 0;
 484 }
 485
 486 static inline
 487 int nfs_reschedule_unstable_write(struct nfs_page *req)
 488 {
 489         return 0;
 490 }
 491 #endif
 492
 493 /*
 494  * Wait for a request to complete.
 495  *
 496  * Interruptible by signals only if mounted with intr flag.
 497  */
 498 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 499 {
 500         struct nfs_inode *nfsi = NFS_I(inode);
 501         struct nfs_page *req;
 502         pgoff_t idx_end, next;
 503         unsigned int            res = 0;
 504         int                     error;
 505
 506         if (npages == 0)
 507                 idx_end = ~0;
 508         else
 509                 idx_end = idx_start + npages - 1;
 510
 511         next = idx_start;
 512         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 513                 if (req->wb_index > idx_end)
 514                         break;
 515
 516                 next = req->wb_index + 1;
 517                 BUG_ON(!NFS_WBACK_BUSY(req));
 518
 519                 kref_get(&req->wb_kref);
 520                 spin_unlock(&inode->i_lock);
 521                 error = nfs_wait_on_request(req);
 522                 nfs_release_request(req);
 523                 spin_lock(&inode->i_lock);
 524                 if (error < 0)
 525                         return error;
 526                 res++;
 527         }
 528         return res;
 529 }
 530
 531 static void nfs_cancel_commit_list(struct list_head *head)
 532 {
 533         struct nfs_page *req;
 534
 535         while(!list_empty(head)) {
 536                 req = nfs_list_entry(head->next);
 537                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 538                 nfs_list_remove_request(req);
 539                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 540                 nfs_inode_remove_request(req);
 541                 nfs_unlock_request(req);
 542         }
 543 }
 544
 545 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 546 /*
 547  * nfs_scan_commit - Scan an inode for commit requests
 548  * @inode: NFS inode to scan
 549  * @dst: destination list
 550  * @idx_start: lower bound of page->index to scan.
 551  * @npages: idx_start + npages sets the upper bound to scan.
 552  *
 553  * Moves requests from the inode's 'commit' request list.
 554  * The requests are *not* checked to ensure that they form a contiguous set.
 555  */
 556 static int
 557 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 558 {
 559         struct nfs_inode *nfsi = NFS_I(inode);
 560         int res = 0;
 561
 562         if (nfsi->ncommit != 0) {
 563                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 564                                 NFS_PAGE_TAG_COMMIT);
 565                 nfsi->ncommit -= res;
 566         }
 567         return res;
 568 }
 569 #else
 570 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 571 {
 572         return 0;
 573 }
 574 #endif
 575
 576 /*
 577  * Try to update any existing write request, or create one if there is none.
 578  * In order to match, the request's credentials must match those of
 579  * the calling process.
 580  *
 581  * Note: Should always be called with the Page Lock held!
 582  */
 583 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 584                 struct page *page, unsigned int offset, unsigned int bytes)
 585 {
 586         struct address_space *mapping = page->mapping;
 587         struct inode *inode = mapping->host;
 588         struct nfs_page         *req, *new = NULL;
 589         pgoff_t         rqend, end;
 590
 591         end = offset + bytes;
 592
 593         for (;;) {
 594                 /* Loop over all inode entries and see if we find
 595                  * A request for the page we wish to update
 596                  */
 597                 spin_lock(&inode->i_lock);
 598                 req = nfs_page_find_request_locked(page);
 599                 if (req) {
 600                         if (!nfs_lock_request_dontget(req)) {
 601                                 int error;
 602
 603                                 spin_unlock(&inode->i_lock);
 604                                 error = nfs_wait_on_request(req);
 605                                 nfs_release_request(req);
 606                                 if (error < 0) {
 607                                         if (new)
 608                                                 nfs_release_request(new);
 609                                         return ERR_PTR(error);
 610                                 }
 611                                 continue;
 612                         }
 613                         spin_unlock(&inode->i_lock);
 614                         if (new)
 615                                 nfs_release_request(new);
 616                         break;
 617                 }
 618
 619                 if (new) {
 620                         int error;
 621                         nfs_lock_request_dontget(new);
 622                         error = nfs_inode_add_request(inode, new);
 623                         if (error) {
 624                                 spin_unlock(&inode->i_lock);
 625                                 nfs_unlock_request(new);
 626                                 return ERR_PTR(error);
 627                         }
 628                         spin_unlock(&inode->i_lock);
 629                         return new;
 630                 }
 631                 spin_unlock(&inode->i_lock);
 632
 633                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 634                 if (IS_ERR(new))
 635                         return new;
 636         }
 637
 638         /* We have a request for our page.
 639          * If the creds don't match, or the
 640          * page addresses don't match,
 641          * tell the caller to wait on the conflicting
 642          * request.
 643          */
 644         rqend = req->wb_offset + req->wb_bytes;
 645         if (req->wb_context != ctx
 646             || req->wb_page != page
 647             || !nfs_dirty_request(req)
 648             || offset > rqend || end < req->wb_offset) {
 649                 nfs_unlock_request(req);
 650                 return ERR_PTR(-EBUSY);
 651         }
 652
 653         /* Okay, the request matches. Update the region */
 654         if (offset < req->wb_offset) {
 655                 req->wb_offset = offset;
 656                 req->wb_pgbase = offset;
 657                 req->wb_bytes = rqend - req->wb_offset;
 658         }
 659
 660         if (end > rqend)
 661                 req->wb_bytes = end - req->wb_offset;
 662
 663         return req;
 664 }
 665
 666 int nfs_flush_incompatible(struct file *file, struct page *page)
 667 {
 668         struct nfs_open_context *ctx = nfs_file_open_context(file);
 669         struct nfs_page *req;
 670         int do_flush, status;
 671         /*
 672          * Look for a request corresponding to this page. If there
 673          * is one, and it belongs to another file, we flush it out
 674          * before we try to copy anything into the page. Do this
 675          * due to the lack of an ACCESS-type call in NFSv2.
 676          * Also do the same if we find a request from an existing
 677          * dropped page.
 678          */
 679         do {
 680                 req = nfs_page_find_request(page);
 681                 if (req == NULL)
 682                         return 0;
 683                 do_flush = req->wb_page != page || req->wb_context != ctx
 684                         || !nfs_dirty_request(req);
 685                 nfs_release_request(req);
 686                 if (!do_flush)
 687                         return 0;
 688                 status = nfs_wb_page(page->mapping->host, page);
 689         } while (status == 0);
 690         return status;
 691 }
 692
 693 /*
 694  * Update and possibly write a cached page of an NFS file.
 695  *
 696  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 697  * things with a page scheduled for an RPC call (e.g. invalidate it).
 698  */
 699 int nfs_updatepage(struct file *file, struct page *page,
 700                 unsigned int offset, unsigned int count)
 701 {
 702         struct nfs_open_context *ctx = nfs_file_open_context(file);
 703         struct inode    *inode = page->mapping->host;
 704         int             status = 0;
 705
 706         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 707
 708         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 709                 file->f_path.dentry->d_parent->d_name.name,
 710                 file->f_path.dentry->d_name.name, count,
 711                 (long long)(page_offset(page) +offset));
 712
 713         /* If we're not using byte range locks, and we know the page
 714          * is entirely in cache, it may be more efficient to avoid
 715          * fragmenting write requests.
 716          */
 717         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 718                 count = max(count + offset, nfs_page_length(page));
 719                 offset = 0;
 720         }
 721
 722         status = nfs_writepage_setup(ctx, page, offset, count);
 723         __set_page_dirty_nobuffers(page);
 724
 725         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 726                         status, (long long)i_size_read(inode));
 727         if (status < 0)
 728                 nfs_set_pageerror(page);
 729         return status;
 730 }
 731
 732 static void nfs_writepage_release(struct nfs_page *req)
 733 {
 734
 735         if (PageError(req->wb_page)) {
 736                 nfs_end_page_writeback(req->wb_page);
 737                 nfs_inode_remove_request(req);
 738         } else if (!nfs_reschedule_unstable_write(req)) {
 739                 /* Set the PG_uptodate flag */
 740                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 741                 nfs_end_page_writeback(req->wb_page);
 742                 nfs_inode_remove_request(req);
 743         } else
 744                 nfs_end_page_writeback(req->wb_page);
 745         nfs_clear_page_tag_locked(req);
 746 }
 747
 748 static inline int flush_task_priority(int how)
 749 {
 750         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 751                 case FLUSH_HIGHPRI:
 752                         return RPC_PRIORITY_HIGH;
 753                 case FLUSH_LOWPRI:
 754                         return RPC_PRIORITY_LOW;
 755         }
 756         return RPC_PRIORITY_NORMAL;
 757 }
 758
 759 /*
 760  * Set up the argument/result storage required for the RPC call.
 761  */
 762 static void nfs_write_rpcsetup(struct nfs_page *req,
 763                 struct nfs_write_data *data,
 764                 const struct rpc_call_ops *call_ops,
 765                 unsigned int count, unsigned int offset,
 766                 int how)
 767 {
 768         struct inode            *inode;
 769         int flags;
 770
 771         /* Set up the RPC argument and reply structs
 772          * NB: take care not to mess about with data->commit et al. */
 773
 774         data->req = req;
 775         data->inode = inode = req->wb_context->path.dentry->d_inode;
 776         data->cred = req->wb_context->cred;
 777
 778         data->args.fh     = NFS_FH(inode);
 779         data->args.offset = req_offset(req) + offset;
 780         data->args.pgbase = req->wb_pgbase + offset;
 781         data->args.pages  = data->pagevec;
 782         data->args.count  = count;
 783         data->args.context = req->wb_context;
 784
 785         data->res.fattr   = &data->fattr;
 786         data->res.count   = count;
 787         data->res.verf    = &data->verf;
 788         nfs_fattr_init(&data->fattr);
 789
 790         /* Set up the initial task struct.  */
 791         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 792         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 793         NFS_PROTO(inode)->write_setup(data, how);
 794
 795         data->task.tk_priority = flush_task_priority(how);
 796         data->task.tk_cookie = (unsigned long)inode;
 797
 798         dprintk("NFS: %5u initiated write call "
 799                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 800                 data->task.tk_pid,
 801                 inode->i_sb->s_id,
 802                 (long long)NFS_FILEID(inode),
 803                 count,
 804                 (unsigned long long)data->args.offset);
 805 }
 806
 807 static void nfs_execute_write(struct nfs_write_data *data)
 808 {
 809         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 810         sigset_t oldset;
 811
 812         rpc_clnt_sigmask(clnt, &oldset);
 813         rpc_execute(&data->task);
 814         rpc_clnt_sigunmask(clnt, &oldset);
 815 }
 816
 817 /*
 818  * Generate multiple small requests to write out a single
 819  * contiguous dirty area on one page.
 820  */
 821 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 822 {
 823         struct nfs_page *req = nfs_list_entry(head->next);
 824         struct page *page = req->wb_page;
 825         struct nfs_write_data *data;
 826         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 827         unsigned int offset;
 828         int requests = 0;
 829         LIST_HEAD(list);
 830
 831         nfs_list_remove_request(req);
 832
 833         nbytes = count;
 834         do {
 835                 size_t len = min(nbytes, wsize);
 836
 837                 data = nfs_writedata_alloc(1);
 838                 if (!data)
 839                         goto out_bad;
 840                 list_add(&data->pages, &list);
 841                 requests++;
 842                 nbytes -= len;
 843         } while (nbytes != 0);
 844         atomic_set(&req->wb_complete, requests);
 845
 846         ClearPageError(page);
 847         offset = 0;
 848         nbytes = count;
 849         do {
 850                 data = list_entry(list.next, struct nfs_write_data, pages);
 851                 list_del_init(&data->pages);
 852
 853                 data->pagevec[0] = page;
 854
 855                 if (nbytes < wsize)
 856                         wsize = nbytes;
 857                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 858                                    wsize, offset, how);
 859                 offset += wsize;
 860                 nbytes -= wsize;
 861                 nfs_execute_write(data);
 862         } while (nbytes != 0);
 863
 864         return 0;
 865
 866 out_bad:
 867         while (!list_empty(&list)) {
 868                 data = list_entry(list.next, struct nfs_write_data, pages);
 869                 list_del(&data->pages);
 870                 nfs_writedata_release(data);
 871         }
 872         nfs_redirty_request(req);
 873         nfs_end_page_writeback(req->wb_page);
 874         nfs_clear_page_tag_locked(req);
 875         return -ENOMEM;
 876 }
 877
 878 /*
 879  * Create an RPC task for the given write request and kick it.
 880  * The page must have been locked by the caller.
 881  *
 882  * It may happen that the page we're passed is not marked dirty.
 883  * This is the case if nfs_updatepage detects a conflicting request
 884  * that has been written but not committed.
 885  */
 886 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 887 {
 888         struct nfs_page         *req;
 889         struct page             **pages;
 890         struct nfs_write_data   *data;
 891
 892         data = nfs_writedata_alloc(npages);
 893         if (!data)
 894                 goto out_bad;
 895
 896         pages = data->pagevec;
 897         while (!list_empty(head)) {
 898                 req = nfs_list_entry(head->next);
 899                 nfs_list_remove_request(req);
 900                 nfs_list_add_request(req, &data->pages);
 901                 ClearPageError(req->wb_page);
 902                 *pages++ = req->wb_page;
 903         }
 904         req = nfs_list_entry(data->pages.next);
 905
 906         /* Set up the argument struct */
 907         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 908
 909         nfs_execute_write(data);
 910         return 0;
 911  out_bad:
 912         while (!list_empty(head)) {
 913                 req = nfs_list_entry(head->next);
 914                 nfs_list_remove_request(req);
 915                 nfs_redirty_request(req);
 916                 nfs_end_page_writeback(req->wb_page);
 917                 nfs_clear_page_tag_locked(req);
 918         }
 919         return -ENOMEM;
 920 }
 921
 922 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 923                                   struct inode *inode, int ioflags)
 924 {
 925         int wsize = NFS_SERVER(inode)->wsize;
 926
 927         if (wsize < PAGE_CACHE_SIZE)
 928                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 929         else
 930                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 931 }
 932
 933 /*
 934  * Handle a write reply that flushed part of a page.
 935  */
 936 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 937 {
 938         struct nfs_write_data   *data = calldata;
 939         struct nfs_page         *req = data->req;
 940         struct page             *page = req->wb_page;
 941
 942         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 943                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 944                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 945                 req->wb_bytes,
 946                 (long long)req_offset(req));
 947
 948         if (nfs_writeback_done(task, data) != 0)
 949                 return;
 950
 951         if (task->tk_status < 0) {
 952                 nfs_set_pageerror(page);
 953                 nfs_context_set_write_error(req->wb_context, task->tk_status);
 954                 dprintk(", error = %d\n", task->tk_status);
 955                 goto out;
 956         }
 957
 958         if (nfs_write_need_commit(data)) {
 959                 struct inode *inode = page->mapping->host;
 960
 961                 spin_lock(&inode->i_lock);
 962                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 963                         /* Do nothing we need to resend the writes */
 964                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 965                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 966                         dprintk(" defer commit\n");
 967                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 968                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 969                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 970                         dprintk(" server reboot detected\n");
 971                 }
 972                 spin_unlock(&inode->i_lock);
 973         } else
 974                 dprintk(" OK\n");
 975
 976 out:
 977         if (atomic_dec_and_test(&req->wb_complete))
 978                 nfs_writepage_release(req);
 979 }
 980
 981 static const struct rpc_call_ops nfs_write_partial_ops = {
 982         .rpc_call_done = nfs_writeback_done_partial,
 983         .rpc_release = nfs_writedata_release,
 984 };
 985
 986 /*
 987  * Handle a write reply that flushes a whole page.
 988  *
 989  * FIXME: There is an inherent race with invalidate_inode_pages and
 990  *        writebacks since the page->count is kept > 1 for as long
 991  *        as the page has a write request pending.
 992  */
 993 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
 994 {
 995         struct nfs_write_data   *data = calldata;
 996         struct nfs_page         *req;
 997         struct page             *page;
 998
 999         if (nfs_writeback_done(task, data) != 0)
1000                 return;
1001
1002         /* Update attributes as result of writeback. */
1003         while (!list_empty(&data->pages)) {
1004                 req = nfs_list_entry(data->pages.next);
1005                 nfs_list_remove_request(req);
1006                 page = req->wb_page;
1007
1008                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1009                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1010                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1011                         req->wb_bytes,
1012                         (long long)req_offset(req));
1013
1014                 if (task->tk_status < 0) {
1015                         nfs_set_pageerror(page);
1016                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1017                         dprintk(", error = %d\n", task->tk_status);
1018                         goto remove_request;
1019                 }
1020
1021                 if (nfs_write_need_commit(data)) {
1022                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1023                         nfs_mark_request_commit(req);
1024                         nfs_end_page_writeback(page);
1025                         dprintk(" marked for commit\n");
1026                         goto next;
1027                 }
1028                 /* Set the PG_uptodate flag? */
1029                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1030                 dprintk(" OK\n");
1031 remove_request:
1032                 nfs_end_page_writeback(page);
1033                 nfs_inode_remove_request(req);
1034         next:
1035                 nfs_clear_page_tag_locked(req);
1036         }
1037 }
1038
1039 static const struct rpc_call_ops nfs_write_full_ops = {
1040         .rpc_call_done = nfs_writeback_done_full,
1041         .rpc_release = nfs_writedata_release,
1042 };
1043
1044
1045 /*
1046  * This function is called when the WRITE call is complete.
1047  */
1048 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1049 {
1050         struct nfs_writeargs    *argp = &data->args;
1051         struct nfs_writeres     *resp = &data->res;
1052         int status;
1053
1054         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1055                 task->tk_pid, task->tk_status);
1056
1057         /*
1058          * ->write_done will attempt to use post-op attributes to detect
1059          * conflicting writes by other clients.  A strict interpretation
1060          * of close-to-open would allow us to continue caching even if
1061          * another writer had changed the file, but some applications
1062          * depend on tighter cache coherency when writing.
1063          */
1064         status = NFS_PROTO(data->inode)->write_done(task, data);
1065         if (status != 0)
1066                 return status;
1067         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1068
1069 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1070         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1071                 /* We tried a write call, but the server did not
1072                  * commit data to stable storage even though we
1073                  * requested it.
1074                  * Note: There is a known bug in Tru64 < 5.0 in which
1075                  *       the server reports NFS_DATA_SYNC, but performs
1076                  *       NFS_FILE_SYNC. We therefore implement this checking
1077                  *       as a dprintk() in order to avoid filling syslog.
1078                  */
1079                 static unsigned long    complain;
1080
1081                 if (time_before(complain, jiffies)) {
1082                         dprintk("NFS: faulty NFS server %s:"
1083                                 " (committed = %d) != (stable = %d)\n",
1084                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1085                                 resp->verf->committed, argp->stable);
1086                         complain = jiffies + 300 * HZ;
1087                 }
1088         }
1089 #endif
1090         /* Is this a short write? */
1091         if (task->tk_status >= 0 && resp->count < argp->count) {
1092                 static unsigned long    complain;
1093
1094                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1095
1096                 /* Has the server at least made some progress? */
1097                 if (resp->count != 0) {
1098                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1099                         if (resp->verf->committed != NFS_UNSTABLE) {
1100                                 /* Resend from where the server left off */
1101                                 argp->offset += resp->count;
1102                                 argp->pgbase += resp->count;
1103                                 argp->count -= resp->count;
1104                         } else {
1105                                 /* Resend as a stable write in order to avoid
1106                                  * headaches in the case of a server crash.
1107                                  */
1108                                 argp->stable = NFS_FILE_SYNC;
1109                         }
1110                         rpc_restart_call(task);
1111                         return -EAGAIN;
1112                 }
1113                 if (time_before(complain, jiffies)) {
1114                         printk(KERN_WARNING
1115                                "NFS: Server wrote zero bytes, expected %u.\n",
1116                                         argp->count);
1117                         complain = jiffies + 300 * HZ;
1118                 }
1119                 /* Can't do anything about it except throw an error. */
1120                 task->tk_status = -EIO;
1121         }
1122         return 0;
1123 }
1124
1125
1126 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1127 void nfs_commit_release(void *wdata)
1128 {
1129         nfs_commit_free(wdata);
1130 }
1131
1132 /*
1133  * Set up the argument/result storage required for the RPC call.
1134  */
1135 static void nfs_commit_rpcsetup(struct list_head *head,
1136                 struct nfs_write_data *data,
1137                 int how)
1138 {
1139         struct nfs_page         *first;
1140         struct inode            *inode;
1141         int flags;
1142
1143         /* Set up the RPC argument and reply structs
1144          * NB: take care not to mess about with data->commit et al. */
1145
1146         list_splice_init(head, &data->pages);
1147         first = nfs_list_entry(data->pages.next);
1148         inode = first->wb_context->path.dentry->d_inode;
1149
1150         data->inode       = inode;
1151         data->cred        = first->wb_context->cred;
1152
1153         data->args.fh     = NFS_FH(data->inode);
1154         /* Note: we always request a commit of the entire inode */
1155         data->args.offset = 0;
1156         data->args.count  = 0;
1157         data->res.count   = 0;
1158         data->res.fattr   = &data->fattr;
1159         data->res.verf    = &data->verf;
1160         nfs_fattr_init(&data->fattr);
1161
1162         /* Set up the initial task struct.  */
1163         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1164         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1165         NFS_PROTO(inode)->commit_setup(data, how);
1166
1167         data->task.tk_priority = flush_task_priority(how);
1168         data->task.tk_cookie = (unsigned long)inode;
1169
1170         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1171 }
1172
1173 /*
1174  * Commit dirty pages
1175  */
1176 static int
1177 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1178 {
1179         struct nfs_write_data   *data;
1180         struct nfs_page         *req;
1181
1182         data = nfs_commit_alloc();
1183
1184         if (!data)
1185                 goto out_bad;
1186
1187         /* Set up the argument struct */
1188         nfs_commit_rpcsetup(head, data, how);
1189
1190         nfs_execute_write(data);
1191         return 0;
1192  out_bad:
1193         while (!list_empty(head)) {
1194                 req = nfs_list_entry(head->next);
1195                 nfs_list_remove_request(req);
1196                 nfs_mark_request_commit(req);
1197                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1198                 nfs_clear_page_tag_locked(req);
1199         }
1200         return -ENOMEM;
1201 }
1202
1203 /*
1204  * COMMIT call returned
1205  */
1206 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1207 {
1208         struct nfs_write_data   *data = calldata;
1209         struct nfs_page         *req;
1210
1211         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1212                                 task->tk_pid, task->tk_status);
1213
1214         /* Call the NFS version-specific code */
1215         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1216                 return;
1217
1218         while (!list_empty(&data->pages)) {
1219                 req = nfs_list_entry(data->pages.next);
1220                 nfs_list_remove_request(req);
1221                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1222                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1223
1224                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1225                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1226                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1227                         req->wb_bytes,
1228                         (long long)req_offset(req));
1229                 if (task->tk_status < 0) {
1230                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1231                         nfs_inode_remove_request(req);
1232                         dprintk(", error = %d\n", task->tk_status);
1233                         goto next;
1234                 }
1235
1236                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1237                  * returned by the server against all stored verfs. */
1238                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1239                         /* We have a match */
1240                         /* Set the PG_uptodate flag */
1241                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1242                                         req->wb_bytes);
1243                         nfs_inode_remove_request(req);
1244                         dprintk(" OK\n");
1245                         goto next;
1246                 }
1247                 /* We have a mismatch. Write the page again */
1248                 dprintk(" mismatch\n");
1249                 nfs_redirty_request(req);
1250         next:
1251                 nfs_clear_page_tag_locked(req);
1252         }
1253 }
1254
1255 static const struct rpc_call_ops nfs_commit_ops = {
1256         .rpc_call_done = nfs_commit_done,
1257         .rpc_release = nfs_commit_release,
1258 };
1259
1260 int nfs_commit_inode(struct inode *inode, int how)
1261 {
1262         LIST_HEAD(head);
1263         int res;
1264
1265         spin_lock(&inode->i_lock);
1266         res = nfs_scan_commit(inode, &head, 0, 0);
1267         spin_unlock(&inode->i_lock);
1268         if (res) {
1269                 int error = nfs_commit_list(inode, &head, how);
1270                 if (error < 0)
1271                         return error;
1272         }
1273         return res;
1274 }
1275 #else
1276 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1277 {
1278         return 0;
1279 }
1280 #endif
1281
1282 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1283 {
1284         struct inode *inode = mapping->host;
1285         pgoff_t idx_start, idx_end;
1286         unsigned int npages = 0;
1287         LIST_HEAD(head);
1288         int nocommit = how & FLUSH_NOCOMMIT;
1289         long pages, ret;
1290
1291         /* FIXME */
1292         if (wbc->range_cyclic)
1293                 idx_start = 0;
1294         else {
1295                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1296                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1297                 if (idx_end > idx_start) {
1298                         pgoff_t l_npages = 1 + idx_end - idx_start;
1299                         npages = l_npages;
1300                         if (sizeof(npages) != sizeof(l_npages) &&
1301                                         (pgoff_t)npages != l_npages)
1302                                 npages = 0;
1303                 }
1304         }
1305         how &= ~FLUSH_NOCOMMIT;
1306         spin_lock(&inode->i_lock);
1307         do {
1308                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1309                 if (ret != 0)
1310                         continue;
1311                 if (nocommit)
1312                         break;
1313                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1314                 if (pages == 0)
1315                         break;
1316                 if (how & FLUSH_INVALIDATE) {
1317                         spin_unlock(&inode->i_lock);
1318                         nfs_cancel_commit_list(&head);
1319                         ret = pages;
1320                         spin_lock(&inode->i_lock);
1321                         continue;
1322                 }
1323                 pages += nfs_scan_commit(inode, &head, 0, 0);
1324                 spin_unlock(&inode->i_lock);
1325                 ret = nfs_commit_list(inode, &head, how);
1326                 spin_lock(&inode->i_lock);
1327
1328         } while (ret >= 0);
1329         spin_unlock(&inode->i_lock);
1330         return ret;
1331 }
1332
1333 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1334 {
1335         int ret;
1336
1337         ret = nfs_writepages(mapping, wbc);
1338         if (ret < 0)
1339                 goto out;
1340         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1341         if (ret < 0)
1342                 goto out;
1343         return 0;
1344 out:
1345         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1346         return ret;
1347 }
1348
1349 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1350 static int nfs_write_mapping(struct address_space *mapping, int how)
1351 {
1352         struct writeback_control wbc = {
1353                 .bdi = mapping->backing_dev_info,
1354                 .sync_mode = WB_SYNC_NONE,
1355                 .nr_to_write = LONG_MAX,
1356                 .for_writepages = 1,
1357                 .range_cyclic = 1,
1358         };
1359         int ret;
1360
1361         ret = __nfs_write_mapping(mapping, &wbc, how);
1362         if (ret < 0)
1363                 return ret;
1364         wbc.sync_mode = WB_SYNC_ALL;
1365         return __nfs_write_mapping(mapping, &wbc, how);
1366 }
1367
1368 /*
1369  * flush the inode to disk.
1370  */
1371 int nfs_wb_all(struct inode *inode)
1372 {
1373         return nfs_write_mapping(inode->i_mapping, 0);
1374 }
1375
1376 int nfs_wb_nocommit(struct inode *inode)
1377 {
1378         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1379 }
1380
1381 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1382 {
1383         struct nfs_page *req;
1384         loff_t range_start = page_offset(page);
1385         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1386         struct writeback_control wbc = {
1387                 .bdi = page->mapping->backing_dev_info,
1388                 .sync_mode = WB_SYNC_ALL,
1389                 .nr_to_write = LONG_MAX,
1390                 .range_start = range_start,
1391                 .range_end = range_end,
1392         };
1393         int ret = 0;
1394
1395         BUG_ON(!PageLocked(page));
1396         for (;;) {
1397                 req = nfs_page_find_request(page);
1398                 if (req == NULL)
1399                         goto out;
1400                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1401                         nfs_release_request(req);
1402                         break;
1403                 }
1404                 if (nfs_lock_request_dontget(req)) {
1405                         nfs_inode_remove_request(req);
1406                         /*
1407                          * In case nfs_inode_remove_request has marked the
1408                          * page as being dirty
1409                          */
1410                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1411                         nfs_unlock_request(req);
1412                         break;
1413                 }
1414                 ret = nfs_wait_on_request(req);
1415                 if (ret < 0)
1416                         goto out;
1417         }
1418         if (!PagePrivate(page))
1419                 return 0;
1420         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1421 out:
1422         return ret;
1423 }
1424
1425 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1426 {
1427         loff_t range_start = page_offset(page);
1428         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1429         struct writeback_control wbc = {
1430                 .bdi = page->mapping->backing_dev_info,
1431                 .sync_mode = WB_SYNC_ALL,
1432                 .nr_to_write = LONG_MAX,
1433                 .range_start = range_start,
1434                 .range_end = range_end,
1435         };
1436         int ret;
1437
1438         BUG_ON(!PageLocked(page));
1439         if (clear_page_dirty_for_io(page)) {
1440                 ret = nfs_writepage_locked(page, &wbc);
1441                 if (ret < 0)
1442                         goto out;
1443         }
1444         if (!PagePrivate(page))
1445                 return 0;
1446         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1447         if (ret >= 0)
1448                 return 0;
1449 out:
1450         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1451         return ret;
1452 }
1453
1454 /*
1455  * Write back all requests on one page - we do this before reading it.
1456  */
1457 int nfs_wb_page(struct inode *inode, struct page* page)
1458 {
1459         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1460 }
1461
1462 int __init nfs_init_writepagecache(void)
1463 {
1464         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1465                                              sizeof(struct nfs_write_data),
1466                                              0, SLAB_HWCACHE_ALIGN,
1467                                              NULL);
1468         if (nfs_wdata_cachep == NULL)
1469                 return -ENOMEM;
1470
1471         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1472                                                      nfs_wdata_cachep);
1473         if (nfs_wdata_mempool == NULL)
1474                 return -ENOMEM;
1475
1476         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1477                                                       nfs_wdata_cachep);
1478         if (nfs_commit_mempool == NULL)
1479                 return -ENOMEM;
1480
1481         /*
1482          * NFS congestion size, scale with available memory.
1483          *
1484          *  64MB:    8192k
1485          * 128MB:   11585k
1486          * 256MB:   16384k
1487          * 512MB:   23170k
1488          *   1GB:   32768k
1489          *   2GB:   46340k
1490          *   4GB:   65536k
1491          *   8GB:   92681k
1492          *  16GB:  131072k
1493          *
1494          * This allows larger machines to have larger/more transfers.
1495          * Limit the default to 256M
1496          */
1497         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1498         if (nfs_congestion_kb > 256*1024)
1499                 nfs_congestion_kb = 256*1024;
1500
1501         return 0;
1502 }
1503
1504 void nfs_destroy_writepagecache(void)
1505 {
1506         mempool_destroy(nfs_commit_mempool);
1507         mempool_destroy(nfs_wdata_mempool);
1508         kmem_cache_destroy(nfs_wdata_cachep);
1509 }
1510