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