git.oblomov.eu Git - linux-2.6/blob - fs/nfs/write.c

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