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 void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  38                                   struct inode *inode, int ioflags);
  39 static void nfs_redirty_request(struct nfs_page *req);
  40 static const struct rpc_call_ops nfs_write_partial_ops;
  41 static const struct rpc_call_ops nfs_write_full_ops;
  42 static const struct rpc_call_ops nfs_commit_ops;
  43
  44 static struct kmem_cache *nfs_wdata_cachep;
  45 static mempool_t *nfs_wdata_mempool;
  46 static mempool_t *nfs_commit_mempool;
  47
  48 struct nfs_write_data *nfs_commitdata_alloc(void)
  49 {
  50         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  51
  52         if (p) {
  53                 memset(p, 0, sizeof(*p));
  54                 INIT_LIST_HEAD(&p->pages);
  55         }
  56         return p;
  57 }
  58
  59 void nfs_commit_free(struct nfs_write_data *p)
  60 {
  61         if (p && (p->pagevec != &p->page_array[0]))
  62                 kfree(p->pagevec);
  63         mempool_free(p, nfs_commit_mempool);
  64 }
  65
  66 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  67 {
  68         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  69
  70         if (p) {
  71                 memset(p, 0, sizeof(*p));
  72                 INIT_LIST_HEAD(&p->pages);
  73                 p->npages = pagecount;
  74                 if (pagecount <= ARRAY_SIZE(p->page_array))
  75                         p->pagevec = p->page_array;
  76                 else {
  77                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  78                         if (!p->pagevec) {
  79                                 mempool_free(p, nfs_wdata_mempool);
  80                                 p = NULL;
  81                         }
  82                 }
  83         }
  84         return p;
  85 }
  86
  87 static void nfs_writedata_free(struct nfs_write_data *p)
  88 {
  89         if (p && (p->pagevec != &p->page_array[0]))
  90                 kfree(p->pagevec);
  91         mempool_free(p, nfs_wdata_mempool);
  92 }
  93
  94 void nfs_writedata_release(void *data)
  95 {
  96         struct nfs_write_data *wdata = data;
  97
  98         put_nfs_open_context(wdata->args.context);
  99         nfs_writedata_free(wdata);
 100 }
 101
 102 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 103 {
 104         ctx->error = error;
 105         smp_wmb();
 106         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 107 }
 108
 109 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 110 {
 111         struct nfs_page *req = NULL;
 112
 113         if (PagePrivate(page)) {
 114                 req = (struct nfs_page *)page_private(page);
 115                 if (req != NULL)
 116                         kref_get(&req->wb_kref);
 117         }
 118         return req;
 119 }
 120
 121 static struct nfs_page *nfs_page_find_request(struct page *page)
 122 {
 123         struct inode *inode = page->mapping->host;
 124         struct nfs_page *req = NULL;
 125
 126         spin_lock(&inode->i_lock);
 127         req = nfs_page_find_request_locked(page);
 128         spin_unlock(&inode->i_lock);
 129         return req;
 130 }
 131
 132 /* Adjust the file length if we're writing beyond the end */
 133 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 134 {
 135         struct inode *inode = page->mapping->host;
 136         loff_t end, i_size;
 137         pgoff_t end_index;
 138
 139         spin_lock(&inode->i_lock);
 140         i_size = i_size_read(inode);
 141         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 142         if (i_size > 0 && page->index < end_index)
 143                 goto out;
 144         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 145         if (i_size >= end)
 146                 goto out;
 147         i_size_write(inode, end);
 148         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 149 out:
 150         spin_unlock(&inode->i_lock);
 151 }
 152
 153 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 154 static void nfs_set_pageerror(struct page *page)
 155 {
 156         SetPageError(page);
 157         nfs_zap_mapping(page->mapping->host, page->mapping);
 158 }
 159
 160 /* We can set the PG_uptodate flag if we see that a write request
 161  * covers the full page.
 162  */
 163 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 164 {
 165         if (PageUptodate(page))
 166                 return;
 167         if (base != 0)
 168                 return;
 169         if (count != nfs_page_length(page))
 170                 return;
 171         SetPageUptodate(page);
 172 }
 173
 174 static int wb_priority(struct writeback_control *wbc)
 175 {
 176         if (wbc->for_reclaim)
 177                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 178         if (wbc->for_kupdate)
 179                 return FLUSH_LOWPRI;
 180         return 0;
 181 }
 182
 183 /*
 184  * NFS congestion control
 185  */
 186
 187 int nfs_congestion_kb;
 188
 189 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 190 #define NFS_CONGESTION_OFF_THRESH       \
 191         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 192
 193 static int nfs_set_page_writeback(struct page *page)
 194 {
 195         int ret = test_set_page_writeback(page);
 196
 197         if (!ret) {
 198                 struct inode *inode = page->mapping->host;
 199                 struct nfs_server *nfss = NFS_SERVER(inode);
 200
 201                 if (atomic_long_inc_return(&nfss->writeback) >
 202                                 NFS_CONGESTION_ON_THRESH)
 203                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 204         }
 205         return ret;
 206 }
 207
 208 static void nfs_end_page_writeback(struct page *page)
 209 {
 210         struct inode *inode = page->mapping->host;
 211         struct nfs_server *nfss = NFS_SERVER(inode);
 212
 213         end_page_writeback(page);
 214         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 215                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 216 }
 217
 218 /*
 219  * Find an associated nfs write request, and prepare to flush it out
 220  * May return an error if the user signalled nfs_wait_on_request().
 221  */
 222 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 223                                 struct page *page)
 224 {
 225         struct inode *inode = page->mapping->host;
 226         struct nfs_page *req;
 227         int ret;
 228
 229         spin_lock(&inode->i_lock);
 230         for(;;) {
 231                 req = nfs_page_find_request_locked(page);
 232                 if (req == NULL) {
 233                         spin_unlock(&inode->i_lock);
 234                         return 0;
 235                 }
 236                 if (nfs_set_page_tag_locked(req))
 237                         break;
 238                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 239                  *       then the call to nfs_set_page_tag_locked() will always
 240                  *       succeed provided that someone hasn't already marked the
 241                  *       request as dirty (in which case we don't care).
 242                  */
 243                 spin_unlock(&inode->i_lock);
 244                 ret = nfs_wait_on_request(req);
 245                 nfs_release_request(req);
 246                 if (ret != 0)
 247                         return ret;
 248                 spin_lock(&inode->i_lock);
 249         }
 250         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 251                 spin_unlock(&inode->i_lock);
 252                 BUG();
 253         }
 254         if (nfs_set_page_writeback(page) != 0) {
 255                 spin_unlock(&inode->i_lock);
 256                 BUG();
 257         }
 258         spin_unlock(&inode->i_lock);
 259         if (!nfs_pageio_add_request(pgio, req)) {
 260                 nfs_redirty_request(req);
 261                 return pgio->pg_error;
 262         }
 263         return 0;
 264 }
 265
 266 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 267 {
 268         struct inode *inode = page->mapping->host;
 269
 270         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 271         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 272
 273         nfs_pageio_cond_complete(pgio, page->index);
 274         return nfs_page_async_flush(pgio, page);
 275 }
 276
 277 /*
 278  * Write an mmapped page to the server.
 279  */
 280 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 281 {
 282         struct nfs_pageio_descriptor pgio;
 283         int err;
 284
 285         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 286         err = nfs_do_writepage(page, wbc, &pgio);
 287         nfs_pageio_complete(&pgio);
 288         if (err < 0)
 289                 return err;
 290         if (pgio.pg_error < 0)
 291                 return pgio.pg_error;
 292         return 0;
 293 }
 294
 295 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 296 {
 297         int ret;
 298
 299         ret = nfs_writepage_locked(page, wbc);
 300         unlock_page(page);
 301         return ret;
 302 }
 303
 304 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 305 {
 306         int ret;
 307
 308         ret = nfs_do_writepage(page, wbc, data);
 309         unlock_page(page);
 310         return ret;
 311 }
 312
 313 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 314 {
 315         struct inode *inode = mapping->host;
 316         struct nfs_pageio_descriptor pgio;
 317         int err;
 318
 319         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 320
 321         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 322         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 323         nfs_pageio_complete(&pgio);
 324         if (err < 0)
 325                 return err;
 326         if (pgio.pg_error < 0)
 327                 return pgio.pg_error;
 328         return 0;
 329 }
 330
 331 /*
 332  * Insert a write request into an inode
 333  */
 334 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 335 {
 336         struct nfs_inode *nfsi = NFS_I(inode);
 337         int error;
 338
 339         error = radix_tree_preload(GFP_NOFS);
 340         if (error != 0)
 341                 goto out;
 342
 343         /* Lock the request! */
 344         nfs_lock_request_dontget(req);
 345
 346         spin_lock(&inode->i_lock);
 347         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 348         BUG_ON(error);
 349         if (!nfsi->npages) {
 350                 igrab(inode);
 351                 if (nfs_have_delegation(inode, FMODE_WRITE))
 352                         nfsi->change_attr++;
 353         }
 354         SetPagePrivate(req->wb_page);
 355         set_page_private(req->wb_page, (unsigned long)req);
 356         nfsi->npages++;
 357         kref_get(&req->wb_kref);
 358         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 359                                 NFS_PAGE_TAG_LOCKED);
 360         spin_unlock(&inode->i_lock);
 361         radix_tree_preload_end();
 362 out:
 363         return error;
 364 }
 365
 366 /*
 367  * Remove a write request from an inode
 368  */
 369 static void nfs_inode_remove_request(struct nfs_page *req)
 370 {
 371         struct inode *inode = req->wb_context->path.dentry->d_inode;
 372         struct nfs_inode *nfsi = NFS_I(inode);
 373
 374         BUG_ON (!NFS_WBACK_BUSY(req));
 375
 376         spin_lock(&inode->i_lock);
 377         set_page_private(req->wb_page, 0);
 378         ClearPagePrivate(req->wb_page);
 379         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 380         nfsi->npages--;
 381         if (!nfsi->npages) {
 382                 spin_unlock(&inode->i_lock);
 383                 iput(inode);
 384         } else
 385                 spin_unlock(&inode->i_lock);
 386         nfs_clear_request(req);
 387         nfs_release_request(req);
 388 }
 389
 390 static void
 391 nfs_mark_request_dirty(struct nfs_page *req)
 392 {
 393         __set_page_dirty_nobuffers(req->wb_page);
 394 }
 395
 396 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 397 /*
 398  * Add a request to the inode's commit list.
 399  */
 400 static void
 401 nfs_mark_request_commit(struct nfs_page *req)
 402 {
 403         struct inode *inode = req->wb_context->path.dentry->d_inode;
 404         struct nfs_inode *nfsi = NFS_I(inode);
 405
 406         spin_lock(&inode->i_lock);
 407         nfsi->ncommit++;
 408         set_bit(PG_CLEAN, &(req)->wb_flags);
 409         radix_tree_tag_set(&nfsi->nfs_page_tree,
 410                         req->wb_index,
 411                         NFS_PAGE_TAG_COMMIT);
 412         spin_unlock(&inode->i_lock);
 413         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 414         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 415         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 416 }
 417
 418 static int
 419 nfs_clear_request_commit(struct nfs_page *req)
 420 {
 421         struct page *page = req->wb_page;
 422
 423         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 424                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 425                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 426                 return 1;
 427         }
 428         return 0;
 429 }
 430
 431 static inline
 432 int nfs_write_need_commit(struct nfs_write_data *data)
 433 {
 434         return data->verf.committed != NFS_FILE_SYNC;
 435 }
 436
 437 static inline
 438 int nfs_reschedule_unstable_write(struct nfs_page *req)
 439 {
 440         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 441                 nfs_mark_request_commit(req);
 442                 return 1;
 443         }
 444         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 445                 nfs_mark_request_dirty(req);
 446                 return 1;
 447         }
 448         return 0;
 449 }
 450 #else
 451 static inline void
 452 nfs_mark_request_commit(struct nfs_page *req)
 453 {
 454 }
 455
 456 static inline int
 457 nfs_clear_request_commit(struct nfs_page *req)
 458 {
 459         return 0;
 460 }
 461
 462 static inline
 463 int nfs_write_need_commit(struct nfs_write_data *data)
 464 {
 465         return 0;
 466 }
 467
 468 static inline
 469 int nfs_reschedule_unstable_write(struct nfs_page *req)
 470 {
 471         return 0;
 472 }
 473 #endif
 474
 475 /*
 476  * Wait for a request to complete.
 477  *
 478  * Interruptible by fatal signals only.
 479  */
 480 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 481 {
 482         struct nfs_inode *nfsi = NFS_I(inode);
 483         struct nfs_page *req;
 484         pgoff_t idx_end, next;
 485         unsigned int            res = 0;
 486         int                     error;
 487
 488         if (npages == 0)
 489                 idx_end = ~0;
 490         else
 491                 idx_end = idx_start + npages - 1;
 492
 493         next = idx_start;
 494         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 495                 if (req->wb_index > idx_end)
 496                         break;
 497
 498                 next = req->wb_index + 1;
 499                 BUG_ON(!NFS_WBACK_BUSY(req));
 500
 501                 kref_get(&req->wb_kref);
 502                 spin_unlock(&inode->i_lock);
 503                 error = nfs_wait_on_request(req);
 504                 nfs_release_request(req);
 505                 spin_lock(&inode->i_lock);
 506                 if (error < 0)
 507                         return error;
 508                 res++;
 509         }
 510         return res;
 511 }
 512
 513 static void nfs_cancel_commit_list(struct list_head *head)
 514 {
 515         struct nfs_page *req;
 516
 517         while(!list_empty(head)) {
 518                 req = nfs_list_entry(head->next);
 519                 nfs_list_remove_request(req);
 520                 nfs_clear_request_commit(req);
 521                 nfs_inode_remove_request(req);
 522                 nfs_unlock_request(req);
 523         }
 524 }
 525
 526 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 527 /*
 528  * nfs_scan_commit - Scan an inode for commit requests
 529  * @inode: NFS inode to scan
 530  * @dst: destination list
 531  * @idx_start: lower bound of page->index to scan.
 532  * @npages: idx_start + npages sets the upper bound to scan.
 533  *
 534  * Moves requests from the inode's 'commit' request list.
 535  * The requests are *not* checked to ensure that they form a contiguous set.
 536  */
 537 static int
 538 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 539 {
 540         struct nfs_inode *nfsi = NFS_I(inode);
 541         int res = 0;
 542
 543         if (nfsi->ncommit != 0) {
 544                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 545                                 NFS_PAGE_TAG_COMMIT);
 546                 nfsi->ncommit -= res;
 547         }
 548         return res;
 549 }
 550 #else
 551 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 552 {
 553         return 0;
 554 }
 555 #endif
 556
 557 /*
 558  * Search for an existing write request, and attempt to update
 559  * it to reflect a new dirty region on a given page.
 560  *
 561  * If the attempt fails, then the existing request is flushed out
 562  * to disk.
 563  */
 564 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 565                 struct page *page,
 566                 unsigned int offset,
 567                 unsigned int bytes)
 568 {
 569         struct nfs_page *req;
 570         unsigned int rqend;
 571         unsigned int end;
 572         int error;
 573
 574         if (!PagePrivate(page))
 575                 return NULL;
 576
 577         end = offset + bytes;
 578         spin_lock(&inode->i_lock);
 579
 580         for (;;) {
 581                 req = nfs_page_find_request_locked(page);
 582                 if (req == NULL)
 583                         goto out_unlock;
 584
 585                 rqend = req->wb_offset + req->wb_bytes;
 586                 /*
 587                  * Tell the caller to flush out the request if
 588                  * the offsets are non-contiguous.
 589                  * Note: nfs_flush_incompatible() will already
 590                  * have flushed out requests having wrong owners.
 591                  */
 592                 if (offset > rqend
 593                     || end < req->wb_offset)
 594                         goto out_flushme;
 595
 596                 if (nfs_set_page_tag_locked(req))
 597                         break;
 598
 599                 /* The request is locked, so wait and then retry */
 600                 spin_unlock(&inode->i_lock);
 601                 error = nfs_wait_on_request(req);
 602                 nfs_release_request(req);
 603                 if (error != 0)
 604                         goto out_err;
 605                 spin_lock(&inode->i_lock);
 606         }
 607
 608         if (nfs_clear_request_commit(req))
 609                 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 610                                 req->wb_index, NFS_PAGE_TAG_COMMIT);
 611
 612         /* Okay, the request matches. Update the region */
 613         if (offset < req->wb_offset) {
 614                 req->wb_offset = offset;
 615                 req->wb_pgbase = offset;
 616         }
 617         if (end > rqend)
 618                 req->wb_bytes = end - req->wb_offset;
 619         else
 620                 req->wb_bytes = rqend - req->wb_offset;
 621 out_unlock:
 622         spin_unlock(&inode->i_lock);
 623         return req;
 624 out_flushme:
 625         spin_unlock(&inode->i_lock);
 626         nfs_release_request(req);
 627         error = nfs_wb_page(inode, page);
 628 out_err:
 629         return ERR_PTR(error);
 630 }
 631
 632 /*
 633  * Try to update an existing write request, or create one if there is none.
 634  *
 635  * Note: Should always be called with the Page Lock held to prevent races
 636  * if we have to add a new request. Also assumes that the caller has
 637  * already called nfs_flush_incompatible() if necessary.
 638  */
 639 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 640                 struct page *page, unsigned int offset, unsigned int bytes)
 641 {
 642         struct inode *inode = page->mapping->host;
 643         struct nfs_page *req;
 644         int error;
 645
 646         req = nfs_try_to_update_request(inode, page, offset, bytes);
 647         if (req != NULL)
 648                 goto out;
 649         req = nfs_create_request(ctx, inode, page, offset, bytes);
 650         if (IS_ERR(req))
 651                 goto out;
 652         error = nfs_inode_add_request(inode, req);
 653         if (error != 0) {
 654                 nfs_release_request(req);
 655                 req = ERR_PTR(error);
 656         }
 657 out:
 658         return req;
 659 }
 660
 661 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 662                 unsigned int offset, unsigned int count)
 663 {
 664         struct nfs_page *req;
 665
 666         req = nfs_setup_write_request(ctx, page, offset, count);
 667         if (IS_ERR(req))
 668                 return PTR_ERR(req);
 669         /* Update file length */
 670         nfs_grow_file(page, offset, count);
 671         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 672         nfs_clear_page_tag_locked(req);
 673         return 0;
 674 }
 675
 676 int nfs_flush_incompatible(struct file *file, struct page *page)
 677 {
 678         struct nfs_open_context *ctx = nfs_file_open_context(file);
 679         struct nfs_page *req;
 680         int do_flush, status;
 681         /*
 682          * Look for a request corresponding to this page. If there
 683          * is one, and it belongs to another file, we flush it out
 684          * before we try to copy anything into the page. Do this
 685          * due to the lack of an ACCESS-type call in NFSv2.
 686          * Also do the same if we find a request from an existing
 687          * dropped page.
 688          */
 689         do {
 690                 req = nfs_page_find_request(page);
 691                 if (req == NULL)
 692                         return 0;
 693                 do_flush = req->wb_page != page || req->wb_context != ctx;
 694                 nfs_release_request(req);
 695                 if (!do_flush)
 696                         return 0;
 697                 status = nfs_wb_page(page->mapping->host, page);
 698         } while (status == 0);
 699         return status;
 700 }
 701
 702 /*
 703  * If the page cache is marked as unsafe or invalid, then we can't rely on
 704  * the PageUptodate() flag. In this case, we will need to turn off
 705  * write optimisations that depend on the page contents being correct.
 706  */
 707 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 708 {
 709         return PageUptodate(page) &&
 710                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 711 }
 712
 713 /*
 714  * Update and possibly write a cached page of an NFS file.
 715  *
 716  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 717  * things with a page scheduled for an RPC call (e.g. invalidate it).
 718  */
 719 int nfs_updatepage(struct file *file, struct page *page,
 720                 unsigned int offset, unsigned int count)
 721 {
 722         struct nfs_open_context *ctx = nfs_file_open_context(file);
 723         struct inode    *inode = page->mapping->host;
 724         int             status = 0;
 725
 726         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 727
 728         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 729                 file->f_path.dentry->d_parent->d_name.name,
 730                 file->f_path.dentry->d_name.name, count,
 731                 (long long)(page_offset(page) + offset));
 732
 733         /* If we're not using byte range locks, and we know the page
 734          * is up to date, it may be more efficient to extend the write
 735          * to cover the entire page in order to avoid fragmentation
 736          * inefficiencies.
 737          */
 738         if (nfs_write_pageuptodate(page, inode) &&
 739                         inode->i_flock == NULL &&
 740                         !(file->f_flags & O_SYNC)) {
 741                 count = max(count + offset, nfs_page_length(page));
 742                 offset = 0;
 743         }
 744
 745         status = nfs_writepage_setup(ctx, page, offset, count);
 746         if (status < 0)
 747                 nfs_set_pageerror(page);
 748         else
 749                 __set_page_dirty_nobuffers(page);
 750
 751         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 752                         status, (long long)i_size_read(inode));
 753         return status;
 754 }
 755
 756 static void nfs_writepage_release(struct nfs_page *req)
 757 {
 758
 759         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 760                 nfs_end_page_writeback(req->wb_page);
 761                 nfs_inode_remove_request(req);
 762         } else
 763                 nfs_end_page_writeback(req->wb_page);
 764         nfs_clear_page_tag_locked(req);
 765 }
 766
 767 static int flush_task_priority(int how)
 768 {
 769         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 770                 case FLUSH_HIGHPRI:
 771                         return RPC_PRIORITY_HIGH;
 772                 case FLUSH_LOWPRI:
 773                         return RPC_PRIORITY_LOW;
 774         }
 775         return RPC_PRIORITY_NORMAL;
 776 }
 777
 778 /*
 779  * Set up the argument/result storage required for the RPC call.
 780  */
 781 static int nfs_write_rpcsetup(struct nfs_page *req,
 782                 struct nfs_write_data *data,
 783                 const struct rpc_call_ops *call_ops,
 784                 unsigned int count, unsigned int offset,
 785                 int how)
 786 {
 787         struct inode *inode = req->wb_context->path.dentry->d_inode;
 788         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 789         int priority = flush_task_priority(how);
 790         struct rpc_task *task;
 791         struct rpc_message msg = {
 792                 .rpc_argp = &data->args,
 793                 .rpc_resp = &data->res,
 794                 .rpc_cred = req->wb_context->cred,
 795         };
 796         struct rpc_task_setup task_setup_data = {
 797                 .rpc_client = NFS_CLIENT(inode),
 798                 .task = &data->task,
 799                 .rpc_message = &msg,
 800                 .callback_ops = call_ops,
 801                 .callback_data = data,
 802                 .workqueue = nfsiod_workqueue,
 803                 .flags = flags,
 804                 .priority = priority,
 805         };
 806
 807         /* Set up the RPC argument and reply structs
 808          * NB: take care not to mess about with data->commit et al. */
 809
 810         data->req = req;
 811         data->inode = inode = req->wb_context->path.dentry->d_inode;
 812         data->cred = msg.rpc_cred;
 813
 814         data->args.fh     = NFS_FH(inode);
 815         data->args.offset = req_offset(req) + offset;
 816         data->args.pgbase = req->wb_pgbase + offset;
 817         data->args.pages  = data->pagevec;
 818         data->args.count  = count;
 819         data->args.context = get_nfs_open_context(req->wb_context);
 820         data->args.stable  = NFS_UNSTABLE;
 821         if (how & FLUSH_STABLE) {
 822                 data->args.stable = NFS_DATA_SYNC;
 823                 if (!NFS_I(inode)->ncommit)
 824                         data->args.stable = NFS_FILE_SYNC;
 825         }
 826
 827         data->res.fattr   = &data->fattr;
 828         data->res.count   = count;
 829         data->res.verf    = &data->verf;
 830         nfs_fattr_init(&data->fattr);
 831
 832         /* Set up the initial task struct.  */
 833         NFS_PROTO(inode)->write_setup(data, &msg);
 834
 835         dprintk("NFS: %5u initiated write call "
 836                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 837                 data->task.tk_pid,
 838                 inode->i_sb->s_id,
 839                 (long long)NFS_FILEID(inode),
 840                 count,
 841                 (unsigned long long)data->args.offset);
 842
 843         task = rpc_run_task(&task_setup_data);
 844         if (IS_ERR(task))
 845                 return PTR_ERR(task);
 846         rpc_put_task(task);
 847         return 0;
 848 }
 849
 850 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 851  * call this on each, which will prepare them to be retried on next
 852  * writeback using standard nfs.
 853  */
 854 static void nfs_redirty_request(struct nfs_page *req)
 855 {
 856         nfs_mark_request_dirty(req);
 857         nfs_end_page_writeback(req->wb_page);
 858         nfs_clear_page_tag_locked(req);
 859 }
 860
 861 /*
 862  * Generate multiple small requests to write out a single
 863  * contiguous dirty area on one page.
 864  */
 865 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 866 {
 867         struct nfs_page *req = nfs_list_entry(head->next);
 868         struct page *page = req->wb_page;
 869         struct nfs_write_data *data;
 870         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 871         unsigned int offset;
 872         int requests = 0;
 873         int ret = 0;
 874         LIST_HEAD(list);
 875
 876         nfs_list_remove_request(req);
 877
 878         nbytes = count;
 879         do {
 880                 size_t len = min(nbytes, wsize);
 881
 882                 data = nfs_writedata_alloc(1);
 883                 if (!data)
 884                         goto out_bad;
 885                 list_add(&data->pages, &list);
 886                 requests++;
 887                 nbytes -= len;
 888         } while (nbytes != 0);
 889         atomic_set(&req->wb_complete, requests);
 890
 891         ClearPageError(page);
 892         offset = 0;
 893         nbytes = count;
 894         do {
 895                 int ret2;
 896
 897                 data = list_entry(list.next, struct nfs_write_data, pages);
 898                 list_del_init(&data->pages);
 899
 900                 data->pagevec[0] = page;
 901
 902                 if (nbytes < wsize)
 903                         wsize = nbytes;
 904                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 905                                    wsize, offset, how);
 906                 if (ret == 0)
 907                         ret = ret2;
 908                 offset += wsize;
 909                 nbytes -= wsize;
 910         } while (nbytes != 0);
 911
 912         return ret;
 913
 914 out_bad:
 915         while (!list_empty(&list)) {
 916                 data = list_entry(list.next, struct nfs_write_data, pages);
 917                 list_del(&data->pages);
 918                 nfs_writedata_release(data);
 919         }
 920         nfs_redirty_request(req);
 921         return -ENOMEM;
 922 }
 923
 924 /*
 925  * Create an RPC task for the given write request and kick it.
 926  * The page must have been locked by the caller.
 927  *
 928  * It may happen that the page we're passed is not marked dirty.
 929  * This is the case if nfs_updatepage detects a conflicting request
 930  * that has been written but not committed.
 931  */
 932 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 933 {
 934         struct nfs_page         *req;
 935         struct page             **pages;
 936         struct nfs_write_data   *data;
 937
 938         data = nfs_writedata_alloc(npages);
 939         if (!data)
 940                 goto out_bad;
 941
 942         pages = data->pagevec;
 943         while (!list_empty(head)) {
 944                 req = nfs_list_entry(head->next);
 945                 nfs_list_remove_request(req);
 946                 nfs_list_add_request(req, &data->pages);
 947                 ClearPageError(req->wb_page);
 948                 *pages++ = req->wb_page;
 949         }
 950         req = nfs_list_entry(data->pages.next);
 951
 952         /* Set up the argument struct */
 953         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 954  out_bad:
 955         while (!list_empty(head)) {
 956                 req = nfs_list_entry(head->next);
 957                 nfs_list_remove_request(req);
 958                 nfs_redirty_request(req);
 959         }
 960         return -ENOMEM;
 961 }
 962
 963 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 964                                   struct inode *inode, int ioflags)
 965 {
 966         size_t wsize = NFS_SERVER(inode)->wsize;
 967
 968         if (wsize < PAGE_CACHE_SIZE)
 969                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 970         else
 971                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 972 }
 973
 974 /*
 975  * Handle a write reply that flushed part of a page.
 976  */
 977 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 978 {
 979         struct nfs_write_data   *data = calldata;
 980
 981         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
 982                 task->tk_pid,
 983                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
 984                 (long long)
 985                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
 986                 data->req->wb_bytes, (long long)req_offset(data->req));
 987
 988         nfs_writeback_done(task, data);
 989 }
 990
 991 static void nfs_writeback_release_partial(void *calldata)
 992 {
 993         struct nfs_write_data   *data = calldata;
 994         struct nfs_page         *req = data->req;
 995         struct page             *page = req->wb_page;
 996         int status = data->task.tk_status;
 997
 998         if (status < 0) {
 999                 nfs_set_pageerror(page);
1000                 nfs_context_set_write_error(req->wb_context, status);
1001                 dprintk(", error = %d\n", status);
1002                 goto out;
1003         }
1004
1005         if (nfs_write_need_commit(data)) {
1006                 struct inode *inode = page->mapping->host;
1007
1008                 spin_lock(&inode->i_lock);
1009                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1010                         /* Do nothing we need to resend the writes */
1011                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1012                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1013                         dprintk(" defer commit\n");
1014                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1015                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1016                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1017                         dprintk(" server reboot detected\n");
1018                 }
1019                 spin_unlock(&inode->i_lock);
1020         } else
1021                 dprintk(" OK\n");
1022
1023 out:
1024         if (atomic_dec_and_test(&req->wb_complete))
1025                 nfs_writepage_release(req);
1026         nfs_writedata_release(calldata);
1027 }
1028
1029 static const struct rpc_call_ops nfs_write_partial_ops = {
1030         .rpc_call_done = nfs_writeback_done_partial,
1031         .rpc_release = nfs_writeback_release_partial,
1032 };
1033
1034 /*
1035  * Handle a write reply that flushes a whole page.
1036  *
1037  * FIXME: There is an inherent race with invalidate_inode_pages and
1038  *        writebacks since the page->count is kept > 1 for as long
1039  *        as the page has a write request pending.
1040  */
1041 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1042 {
1043         struct nfs_write_data   *data = calldata;
1044
1045         nfs_writeback_done(task, data);
1046 }
1047
1048 static void nfs_writeback_release_full(void *calldata)
1049 {
1050         struct nfs_write_data   *data = calldata;
1051         int status = data->task.tk_status;
1052
1053         /* Update attributes as result of writeback. */
1054         while (!list_empty(&data->pages)) {
1055                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1056                 struct page *page = req->wb_page;
1057
1058                 nfs_list_remove_request(req);
1059
1060                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1061                         data->task.tk_pid,
1062                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1063                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1064                         req->wb_bytes,
1065                         (long long)req_offset(req));
1066
1067                 if (status < 0) {
1068                         nfs_set_pageerror(page);
1069                         nfs_context_set_write_error(req->wb_context, status);
1070                         dprintk(", error = %d\n", status);
1071                         goto remove_request;
1072                 }
1073
1074                 if (nfs_write_need_commit(data)) {
1075                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1076                         nfs_mark_request_commit(req);
1077                         nfs_end_page_writeback(page);
1078                         dprintk(" marked for commit\n");
1079                         goto next;
1080                 }
1081                 dprintk(" OK\n");
1082 remove_request:
1083                 nfs_end_page_writeback(page);
1084                 nfs_inode_remove_request(req);
1085         next:
1086                 nfs_clear_page_tag_locked(req);
1087         }
1088         nfs_writedata_release(calldata);
1089 }
1090
1091 static const struct rpc_call_ops nfs_write_full_ops = {
1092         .rpc_call_done = nfs_writeback_done_full,
1093         .rpc_release = nfs_writeback_release_full,
1094 };
1095
1096
1097 /*
1098  * This function is called when the WRITE call is complete.
1099  */
1100 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1101 {
1102         struct nfs_writeargs    *argp = &data->args;
1103         struct nfs_writeres     *resp = &data->res;
1104         int status;
1105
1106         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1107                 task->tk_pid, task->tk_status);
1108
1109         /*
1110          * ->write_done will attempt to use post-op attributes to detect
1111          * conflicting writes by other clients.  A strict interpretation
1112          * of close-to-open would allow us to continue caching even if
1113          * another writer had changed the file, but some applications
1114          * depend on tighter cache coherency when writing.
1115          */
1116         status = NFS_PROTO(data->inode)->write_done(task, data);
1117         if (status != 0)
1118                 return status;
1119         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1120
1121 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1122         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1123                 /* We tried a write call, but the server did not
1124                  * commit data to stable storage even though we
1125                  * requested it.
1126                  * Note: There is a known bug in Tru64 < 5.0 in which
1127                  *       the server reports NFS_DATA_SYNC, but performs
1128                  *       NFS_FILE_SYNC. We therefore implement this checking
1129                  *       as a dprintk() in order to avoid filling syslog.
1130                  */
1131                 static unsigned long    complain;
1132
1133                 if (time_before(complain, jiffies)) {
1134                         dprintk("NFS:       faulty NFS server %s:"
1135                                 " (committed = %d) != (stable = %d)\n",
1136                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1137                                 resp->verf->committed, argp->stable);
1138                         complain = jiffies + 300 * HZ;
1139                 }
1140         }
1141 #endif
1142         /* Is this a short write? */
1143         if (task->tk_status >= 0 && resp->count < argp->count) {
1144                 static unsigned long    complain;
1145
1146                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1147
1148                 /* Has the server at least made some progress? */
1149                 if (resp->count != 0) {
1150                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1151                         if (resp->verf->committed != NFS_UNSTABLE) {
1152                                 /* Resend from where the server left off */
1153                                 argp->offset += resp->count;
1154                                 argp->pgbase += resp->count;
1155                                 argp->count -= resp->count;
1156                         } else {
1157                                 /* Resend as a stable write in order to avoid
1158                                  * headaches in the case of a server crash.
1159                                  */
1160                                 argp->stable = NFS_FILE_SYNC;
1161                         }
1162                         rpc_restart_call(task);
1163                         return -EAGAIN;
1164                 }
1165                 if (time_before(complain, jiffies)) {
1166                         printk(KERN_WARNING
1167                                "NFS: Server wrote zero bytes, expected %u.\n",
1168                                         argp->count);
1169                         complain = jiffies + 300 * HZ;
1170                 }
1171                 /* Can't do anything about it except throw an error. */
1172                 task->tk_status = -EIO;
1173         }
1174         return 0;
1175 }
1176
1177
1178 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1179 void nfs_commitdata_release(void *data)
1180 {
1181         struct nfs_write_data *wdata = data;
1182
1183         put_nfs_open_context(wdata->args.context);
1184         nfs_commit_free(wdata);
1185 }
1186
1187 /*
1188  * Set up the argument/result storage required for the RPC call.
1189  */
1190 static int nfs_commit_rpcsetup(struct list_head *head,
1191                 struct nfs_write_data *data,
1192                 int how)
1193 {
1194         struct nfs_page *first = nfs_list_entry(head->next);
1195         struct inode *inode = first->wb_context->path.dentry->d_inode;
1196         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1197         int priority = flush_task_priority(how);
1198         struct rpc_task *task;
1199         struct rpc_message msg = {
1200                 .rpc_argp = &data->args,
1201                 .rpc_resp = &data->res,
1202                 .rpc_cred = first->wb_context->cred,
1203         };
1204         struct rpc_task_setup task_setup_data = {
1205                 .task = &data->task,
1206                 .rpc_client = NFS_CLIENT(inode),
1207                 .rpc_message = &msg,
1208                 .callback_ops = &nfs_commit_ops,
1209                 .callback_data = data,
1210                 .workqueue = nfsiod_workqueue,
1211                 .flags = flags,
1212                 .priority = priority,
1213         };
1214
1215         /* Set up the RPC argument and reply structs
1216          * NB: take care not to mess about with data->commit et al. */
1217
1218         list_splice_init(head, &data->pages);
1219
1220         data->inode       = inode;
1221         data->cred        = msg.rpc_cred;
1222
1223         data->args.fh     = NFS_FH(data->inode);
1224         /* Note: we always request a commit of the entire inode */
1225         data->args.offset = 0;
1226         data->args.count  = 0;
1227         data->args.context = get_nfs_open_context(first->wb_context);
1228         data->res.count   = 0;
1229         data->res.fattr   = &data->fattr;
1230         data->res.verf    = &data->verf;
1231         nfs_fattr_init(&data->fattr);
1232
1233         /* Set up the initial task struct.  */
1234         NFS_PROTO(inode)->commit_setup(data, &msg);
1235
1236         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1237
1238         task = rpc_run_task(&task_setup_data);
1239         if (IS_ERR(task))
1240                 return PTR_ERR(task);
1241         rpc_put_task(task);
1242         return 0;
1243 }
1244
1245 /*
1246  * Commit dirty pages
1247  */
1248 static int
1249 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1250 {
1251         struct nfs_write_data   *data;
1252         struct nfs_page         *req;
1253
1254         data = nfs_commitdata_alloc();
1255
1256         if (!data)
1257                 goto out_bad;
1258
1259         /* Set up the argument struct */
1260         return nfs_commit_rpcsetup(head, data, how);
1261  out_bad:
1262         while (!list_empty(head)) {
1263                 req = nfs_list_entry(head->next);
1264                 nfs_list_remove_request(req);
1265                 nfs_mark_request_commit(req);
1266                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1267                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1268                                 BDI_RECLAIMABLE);
1269                 nfs_clear_page_tag_locked(req);
1270         }
1271         return -ENOMEM;
1272 }
1273
1274 /*
1275  * COMMIT call returned
1276  */
1277 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1278 {
1279         struct nfs_write_data   *data = calldata;
1280
1281         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1282                                 task->tk_pid, task->tk_status);
1283
1284         /* Call the NFS version-specific code */
1285         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1286                 return;
1287 }
1288
1289 static void nfs_commit_release(void *calldata)
1290 {
1291         struct nfs_write_data   *data = calldata;
1292         struct nfs_page         *req;
1293         int status = data->task.tk_status;
1294
1295         while (!list_empty(&data->pages)) {
1296                 req = nfs_list_entry(data->pages.next);
1297                 nfs_list_remove_request(req);
1298                 nfs_clear_request_commit(req);
1299
1300                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1301                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1302                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1303                         req->wb_bytes,
1304                         (long long)req_offset(req));
1305                 if (status < 0) {
1306                         nfs_context_set_write_error(req->wb_context, status);
1307                         nfs_inode_remove_request(req);
1308                         dprintk(", error = %d\n", status);
1309                         goto next;
1310                 }
1311
1312                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1313                  * returned by the server against all stored verfs. */
1314                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1315                         /* We have a match */
1316                         nfs_inode_remove_request(req);
1317                         dprintk(" OK\n");
1318                         goto next;
1319                 }
1320                 /* We have a mismatch. Write the page again */
1321                 dprintk(" mismatch\n");
1322                 nfs_mark_request_dirty(req);
1323         next:
1324                 nfs_clear_page_tag_locked(req);
1325         }
1326         nfs_commitdata_release(calldata);
1327 }
1328
1329 static const struct rpc_call_ops nfs_commit_ops = {
1330         .rpc_call_done = nfs_commit_done,
1331         .rpc_release = nfs_commit_release,
1332 };
1333
1334 int nfs_commit_inode(struct inode *inode, int how)
1335 {
1336         LIST_HEAD(head);
1337         int res;
1338
1339         spin_lock(&inode->i_lock);
1340         res = nfs_scan_commit(inode, &head, 0, 0);
1341         spin_unlock(&inode->i_lock);
1342         if (res) {
1343                 int error = nfs_commit_list(inode, &head, how);
1344                 if (error < 0)
1345                         return error;
1346         }
1347         return res;
1348 }
1349 #else
1350 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1351 {
1352         return 0;
1353 }
1354 #endif
1355
1356 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1357 {
1358         struct inode *inode = mapping->host;
1359         pgoff_t idx_start, idx_end;
1360         unsigned int npages = 0;
1361         LIST_HEAD(head);
1362         int nocommit = how & FLUSH_NOCOMMIT;
1363         long pages, ret;
1364
1365         /* FIXME */
1366         if (wbc->range_cyclic)
1367                 idx_start = 0;
1368         else {
1369                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1370                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1371                 if (idx_end > idx_start) {
1372                         pgoff_t l_npages = 1 + idx_end - idx_start;
1373                         npages = l_npages;
1374                         if (sizeof(npages) != sizeof(l_npages) &&
1375                                         (pgoff_t)npages != l_npages)
1376                                 npages = 0;
1377                 }
1378         }
1379         how &= ~FLUSH_NOCOMMIT;
1380         spin_lock(&inode->i_lock);
1381         do {
1382                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1383                 if (ret != 0)
1384                         continue;
1385                 if (nocommit)
1386                         break;
1387                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1388                 if (pages == 0)
1389                         break;
1390                 if (how & FLUSH_INVALIDATE) {
1391                         spin_unlock(&inode->i_lock);
1392                         nfs_cancel_commit_list(&head);
1393                         ret = pages;
1394                         spin_lock(&inode->i_lock);
1395                         continue;
1396                 }
1397                 pages += nfs_scan_commit(inode, &head, 0, 0);
1398                 spin_unlock(&inode->i_lock);
1399                 ret = nfs_commit_list(inode, &head, how);
1400                 spin_lock(&inode->i_lock);
1401
1402         } while (ret >= 0);
1403         spin_unlock(&inode->i_lock);
1404         return ret;
1405 }
1406
1407 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1408 {
1409         int ret;
1410
1411         ret = nfs_writepages(mapping, wbc);
1412         if (ret < 0)
1413                 goto out;
1414         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1415         if (ret < 0)
1416                 goto out;
1417         return 0;
1418 out:
1419         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1420         return ret;
1421 }
1422
1423 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1424 static int nfs_write_mapping(struct address_space *mapping, int how)
1425 {
1426         struct writeback_control wbc = {
1427                 .bdi = mapping->backing_dev_info,
1428                 .sync_mode = WB_SYNC_NONE,
1429                 .nr_to_write = LONG_MAX,
1430                 .for_writepages = 1,
1431                 .range_cyclic = 1,
1432         };
1433         int ret;
1434
1435         ret = __nfs_write_mapping(mapping, &wbc, how);
1436         if (ret < 0)
1437                 return ret;
1438         wbc.sync_mode = WB_SYNC_ALL;
1439         return __nfs_write_mapping(mapping, &wbc, how);
1440 }
1441
1442 /*
1443  * flush the inode to disk.
1444  */
1445 int nfs_wb_all(struct inode *inode)
1446 {
1447         return nfs_write_mapping(inode->i_mapping, 0);
1448 }
1449
1450 int nfs_wb_nocommit(struct inode *inode)
1451 {
1452         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1453 }
1454
1455 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1456 {
1457         struct nfs_page *req;
1458         loff_t range_start = page_offset(page);
1459         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1460         struct writeback_control wbc = {
1461                 .bdi = page->mapping->backing_dev_info,
1462                 .sync_mode = WB_SYNC_ALL,
1463                 .nr_to_write = LONG_MAX,
1464                 .range_start = range_start,
1465                 .range_end = range_end,
1466         };
1467         int ret = 0;
1468
1469         BUG_ON(!PageLocked(page));
1470         for (;;) {
1471                 req = nfs_page_find_request(page);
1472                 if (req == NULL)
1473                         goto out;
1474                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1475                         nfs_release_request(req);
1476                         break;
1477                 }
1478                 if (nfs_lock_request_dontget(req)) {
1479                         nfs_inode_remove_request(req);
1480                         /*
1481                          * In case nfs_inode_remove_request has marked the
1482                          * page as being dirty
1483                          */
1484                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1485                         nfs_unlock_request(req);
1486                         break;
1487                 }
1488                 ret = nfs_wait_on_request(req);
1489                 if (ret < 0)
1490                         goto out;
1491         }
1492         if (!PagePrivate(page))
1493                 return 0;
1494         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1495 out:
1496         return ret;
1497 }
1498
1499 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1500                                 int how)
1501 {
1502         loff_t range_start = page_offset(page);
1503         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1504         struct writeback_control wbc = {
1505                 .bdi = page->mapping->backing_dev_info,
1506                 .sync_mode = WB_SYNC_ALL,
1507                 .nr_to_write = LONG_MAX,
1508                 .range_start = range_start,
1509                 .range_end = range_end,
1510         };
1511         int ret;
1512
1513         do {
1514                 if (clear_page_dirty_for_io(page)) {
1515                         ret = nfs_writepage_locked(page, &wbc);
1516                         if (ret < 0)
1517                                 goto out_error;
1518                 } else if (!PagePrivate(page))
1519                         break;
1520                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1521                 if (ret < 0)
1522                         goto out_error;
1523         } while (PagePrivate(page));
1524         return 0;
1525 out_error:
1526         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1527         return ret;
1528 }
1529
1530 /*
1531  * Write back all requests on one page - we do this before reading it.
1532  */
1533 int nfs_wb_page(struct inode *inode, struct page* page)
1534 {
1535         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1536 }
1537
1538 int __init nfs_init_writepagecache(void)
1539 {
1540         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1541                                              sizeof(struct nfs_write_data),
1542                                              0, SLAB_HWCACHE_ALIGN,
1543                                              NULL);
1544         if (nfs_wdata_cachep == NULL)
1545                 return -ENOMEM;
1546
1547         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1548                                                      nfs_wdata_cachep);
1549         if (nfs_wdata_mempool == NULL)
1550                 return -ENOMEM;
1551
1552         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1553                                                       nfs_wdata_cachep);
1554         if (nfs_commit_mempool == NULL)
1555                 return -ENOMEM;
1556
1557         /*
1558          * NFS congestion size, scale with available memory.
1559          *
1560          *  64MB:    8192k
1561          * 128MB:   11585k
1562          * 256MB:   16384k
1563          * 512MB:   23170k
1564          *   1GB:   32768k
1565          *   2GB:   46340k
1566          *   4GB:   65536k
1567          *   8GB:   92681k
1568          *  16GB:  131072k
1569          *
1570          * This allows larger machines to have larger/more transfers.
1571          * Limit the default to 256M
1572          */
1573         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1574         if (nfs_congestion_kb > 256*1024)
1575                 nfs_congestion_kb = 256*1024;
1576
1577         return 0;
1578 }
1579
1580 void nfs_destroy_writepagecache(void)
1581 {
1582         mempool_destroy(nfs_commit_mempool);
1583         mempool_destroy(nfs_wdata_mempool);
1584         kmem_cache_destroy(nfs_wdata_cachep);
1585 }
1586