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