fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16
  17 #include <linux/sunrpc/clnt.h>
  18 #include <linux/nfs_fs.h>
  19 #include <linux/nfs_mount.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/backing-dev.h>
  22
  23 #include <asm/uaccess.h>
  24
  25 #include "delegation.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 #define MIN_POOL_WRITE          (32)
  32 #define MIN_POOL_COMMIT         (4)
  33
  34 /*
  35  * Local function declarations
  36  */
  37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  38                                             struct page *,
  39                                             unsigned int, unsigned int);
  40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static const struct rpc_call_ops nfs_write_partial_ops;
  43 static const struct rpc_call_ops nfs_write_full_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45
  46 static struct kmem_cache *nfs_wdata_cachep;
  47 static mempool_t *nfs_wdata_mempool;
  48 static mempool_t *nfs_commit_mempool;
  49
  50 struct nfs_write_data *nfs_commit_alloc(void)
  51 {
  52         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  53
  54         if (p) {
  55                 memset(p, 0, sizeof(*p));
  56                 INIT_LIST_HEAD(&p->pages);
  57         }
  58         return p;
  59 }
  60
  61 static void nfs_commit_rcu_free(struct rcu_head *head)
  62 {
  63         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 void nfs_commit_free(struct nfs_write_data *wdata)
  70 {
  71         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  72 }
  73
  74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  75 {
  76         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  77
  78         if (p) {
  79                 memset(p, 0, sizeof(*p));
  80                 INIT_LIST_HEAD(&p->pages);
  81                 p->npages = pagecount;
  82                 if (pagecount <= ARRAY_SIZE(p->page_array))
  83                         p->pagevec = p->page_array;
  84                 else {
  85                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  86                         if (!p->pagevec) {
  87                                 mempool_free(p, nfs_wdata_mempool);
  88                                 p = NULL;
  89                         }
  90                 }
  91         }
  92         return p;
  93 }
  94
  95 static void nfs_writedata_rcu_free(struct rcu_head *head)
  96 {
  97         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  98         if (p && (p->pagevec != &p->page_array[0]))
  99                 kfree(p->pagevec);
 100         mempool_free(p, nfs_wdata_mempool);
 101 }
 102
 103 static void nfs_writedata_free(struct nfs_write_data *wdata)
 104 {
 105         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 106 }
 107
 108 void nfs_writedata_release(void *data)
 109 {
 110         struct nfs_write_data *wdata = data;
 111
 112         put_nfs_open_context(wdata->args.context);
 113         nfs_writedata_free(wdata);
 114 }
 115
 116 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 117 {
 118         ctx->error = error;
 119         smp_wmb();
 120         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 121 }
 122
 123 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 124 {
 125         struct nfs_page *req = NULL;
 126
 127         if (PagePrivate(page)) {
 128                 req = (struct nfs_page *)page_private(page);
 129                 if (req != NULL)
 130                         kref_get(&req->wb_kref);
 131         }
 132         return req;
 133 }
 134
 135 static struct nfs_page *nfs_page_find_request(struct page *page)
 136 {
 137         struct inode *inode = page->mapping->host;
 138         struct nfs_page *req = NULL;
 139
 140         spin_lock(&inode->i_lock);
 141         req = nfs_page_find_request_locked(page);
 142         spin_unlock(&inode->i_lock);
 143         return req;
 144 }
 145
 146 /* Adjust the file length if we're writing beyond the end */
 147 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 148 {
 149         struct inode *inode = page->mapping->host;
 150         loff_t end, i_size = i_size_read(inode);
 151         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 152
 153         if (i_size > 0 && page->index < end_index)
 154                 return;
 155         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 156         if (i_size >= end)
 157                 return;
 158         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 159         i_size_write(inode, end);
 160 }
 161
 162 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 163 static void nfs_set_pageerror(struct page *page)
 164 {
 165         SetPageError(page);
 166         nfs_zap_mapping(page->mapping->host, page->mapping);
 167 }
 168
 169 /* We can set the PG_uptodate flag if we see that a write request
 170  * covers the full page.
 171  */
 172 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 173 {
 174         if (PageUptodate(page))
 175                 return;
 176         if (base != 0)
 177                 return;
 178         if (count != nfs_page_length(page))
 179                 return;
 180         SetPageUptodate(page);
 181 }
 182
 183 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 184                 unsigned int offset, unsigned int count)
 185 {
 186         struct nfs_page *req;
 187         int ret;
 188
 189         for (;;) {
 190                 req = nfs_update_request(ctx, page, offset, count);
 191                 if (!IS_ERR(req))
 192                         break;
 193                 ret = PTR_ERR(req);
 194                 if (ret != -EBUSY)
 195                         return ret;
 196                 ret = nfs_wb_page(page->mapping->host, page);
 197                 if (ret != 0)
 198                         return ret;
 199         }
 200         /* Update file length */
 201         nfs_grow_file(page, offset, count);
 202         nfs_clear_page_tag_locked(req);
 203         return 0;
 204 }
 205
 206 static int wb_priority(struct writeback_control *wbc)
 207 {
 208         if (wbc->for_reclaim)
 209                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 210         if (wbc->for_kupdate)
 211                 return FLUSH_LOWPRI;
 212         return 0;
 213 }
 214
 215 /*
 216  * NFS congestion control
 217  */
 218
 219 int nfs_congestion_kb;
 220
 221 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 222 #define NFS_CONGESTION_OFF_THRESH       \
 223         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 224
 225 static int nfs_set_page_writeback(struct page *page)
 226 {
 227         int ret = test_set_page_writeback(page);
 228
 229         if (!ret) {
 230                 struct inode *inode = page->mapping->host;
 231                 struct nfs_server *nfss = NFS_SERVER(inode);
 232
 233                 if (atomic_long_inc_return(&nfss->writeback) >
 234                                 NFS_CONGESTION_ON_THRESH)
 235                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 236         }
 237         return ret;
 238 }
 239
 240 static void nfs_end_page_writeback(struct page *page)
 241 {
 242         struct inode *inode = page->mapping->host;
 243         struct nfs_server *nfss = NFS_SERVER(inode);
 244
 245         end_page_writeback(page);
 246         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 247                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 248 }
 249
 250 /*
 251  * Find an associated nfs write request, and prepare to flush it out
 252  * May return an error if the user signalled nfs_wait_on_request().
 253  */
 254 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 255                                 struct page *page)
 256 {
 257         struct inode *inode = page->mapping->host;
 258         struct nfs_page *req;
 259         int ret;
 260
 261         spin_lock(&inode->i_lock);
 262         for(;;) {
 263                 req = nfs_page_find_request_locked(page);
 264                 if (req == NULL) {
 265                         spin_unlock(&inode->i_lock);
 266                         return 0;
 267                 }
 268                 if (nfs_set_page_tag_locked(req))
 269                         break;
 270                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 271                  *       then the call to nfs_set_page_tag_locked() will always
 272                  *       succeed provided that someone hasn't already marked the
 273                  *       request as dirty (in which case we don't care).
 274                  */
 275                 spin_unlock(&inode->i_lock);
 276                 ret = nfs_wait_on_request(req);
 277                 nfs_release_request(req);
 278                 if (ret != 0)
 279                         return ret;
 280                 spin_lock(&inode->i_lock);
 281         }
 282         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 283                 /* This request is marked for commit */
 284                 spin_unlock(&inode->i_lock);
 285                 nfs_clear_page_tag_locked(req);
 286                 nfs_pageio_complete(pgio);
 287                 return 0;
 288         }
 289         if (nfs_set_page_writeback(page) != 0) {
 290                 spin_unlock(&inode->i_lock);
 291                 BUG();
 292         }
 293         spin_unlock(&inode->i_lock);
 294         nfs_pageio_add_request(pgio, req);
 295         return 0;
 296 }
 297
 298 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 299 {
 300         struct inode *inode = page->mapping->host;
 301
 302         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 303         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 304
 305         nfs_pageio_cond_complete(pgio, page->index);
 306         return nfs_page_async_flush(pgio, page);
 307 }
 308
 309 /*
 310  * Write an mmapped page to the server.
 311  */
 312 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 313 {
 314         struct nfs_pageio_descriptor pgio;
 315         int err;
 316
 317         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 318         err = nfs_do_writepage(page, wbc, &pgio);
 319         nfs_pageio_complete(&pgio);
 320         if (err < 0)
 321                 return err;
 322         if (pgio.pg_error < 0)
 323                 return pgio.pg_error;
 324         return 0;
 325 }
 326
 327 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 328 {
 329         int ret;
 330
 331         ret = nfs_writepage_locked(page, wbc);
 332         unlock_page(page);
 333         return ret;
 334 }
 335
 336 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 337 {
 338         int ret;
 339
 340         ret = nfs_do_writepage(page, wbc, data);
 341         unlock_page(page);
 342         return ret;
 343 }
 344
 345 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 346 {
 347         struct inode *inode = mapping->host;
 348         struct nfs_pageio_descriptor pgio;
 349         int err;
 350
 351         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 352
 353         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 354         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 355         nfs_pageio_complete(&pgio);
 356         if (err < 0)
 357                 return err;
 358         if (pgio.pg_error < 0)
 359                 return pgio.pg_error;
 360         return 0;
 361 }
 362
 363 /*
 364  * Insert a write request into an inode
 365  */
 366 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 367 {
 368         struct nfs_inode *nfsi = NFS_I(inode);
 369         int error;
 370
 371         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 372         BUG_ON(error == -EEXIST);
 373         if (error)
 374                 return error;
 375         if (!nfsi->npages) {
 376                 igrab(inode);
 377                 if (nfs_have_delegation(inode, FMODE_WRITE))
 378                         nfsi->change_attr++;
 379         }
 380         SetPagePrivate(req->wb_page);
 381         set_page_private(req->wb_page, (unsigned long)req);
 382         nfsi->npages++;
 383         kref_get(&req->wb_kref);
 384         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 385         return 0;
 386 }
 387
 388 /*
 389  * Remove a write request from an inode
 390  */
 391 static void nfs_inode_remove_request(struct nfs_page *req)
 392 {
 393         struct inode *inode = req->wb_context->path.dentry->d_inode;
 394         struct nfs_inode *nfsi = NFS_I(inode);
 395
 396         BUG_ON (!NFS_WBACK_BUSY(req));
 397
 398         spin_lock(&inode->i_lock);
 399         set_page_private(req->wb_page, 0);
 400         ClearPagePrivate(req->wb_page);
 401         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 402         nfsi->npages--;
 403         if (!nfsi->npages) {
 404                 spin_unlock(&inode->i_lock);
 405                 iput(inode);
 406         } else
 407                 spin_unlock(&inode->i_lock);
 408         nfs_clear_request(req);
 409         nfs_release_request(req);
 410 }
 411
 412 static void
 413 nfs_redirty_request(struct nfs_page *req)
 414 {
 415         __set_page_dirty_nobuffers(req->wb_page);
 416 }
 417
 418 /*
 419  * Check if a request is dirty
 420  */
 421 static inline int
 422 nfs_dirty_request(struct nfs_page *req)
 423 {
 424         struct page *page = req->wb_page;
 425
 426         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 427                 return 0;
 428         return !PageWriteback(req->wb_page);
 429 }
 430
 431 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 432 /*
 433  * Add a request to the inode's commit list.
 434  */
 435 static void
 436 nfs_mark_request_commit(struct nfs_page *req)
 437 {
 438         struct inode *inode = req->wb_context->path.dentry->d_inode;
 439         struct nfs_inode *nfsi = NFS_I(inode);
 440
 441         spin_lock(&inode->i_lock);
 442         nfsi->ncommit++;
 443         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 444         radix_tree_tag_set(&nfsi->nfs_page_tree,
 445                         req->wb_index,
 446                         NFS_PAGE_TAG_COMMIT);
 447         spin_unlock(&inode->i_lock);
 448         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 449         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 450         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 451 }
 452
 453 static inline
 454 int nfs_write_need_commit(struct nfs_write_data *data)
 455 {
 456         return data->verf.committed != NFS_FILE_SYNC;
 457 }
 458
 459 static inline
 460 int nfs_reschedule_unstable_write(struct nfs_page *req)
 461 {
 462         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 463                 nfs_mark_request_commit(req);
 464                 return 1;
 465         }
 466         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 467                 nfs_redirty_request(req);
 468                 return 1;
 469         }
 470         return 0;
 471 }
 472 #else
 473 static inline void
 474 nfs_mark_request_commit(struct nfs_page *req)
 475 {
 476 }
 477
 478 static inline
 479 int nfs_write_need_commit(struct nfs_write_data *data)
 480 {
 481         return 0;
 482 }
 483
 484 static inline
 485 int nfs_reschedule_unstable_write(struct nfs_page *req)
 486 {
 487         return 0;
 488 }
 489 #endif
 490
 491 /*
 492  * Wait for a request to complete.
 493  *
 494  * Interruptible by fatal signals only.
 495  */
 496 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 497 {
 498         struct nfs_inode *nfsi = NFS_I(inode);
 499         struct nfs_page *req;
 500         pgoff_t idx_end, next;
 501         unsigned int            res = 0;
 502         int                     error;
 503
 504         if (npages == 0)
 505                 idx_end = ~0;
 506         else
 507                 idx_end = idx_start + npages - 1;
 508
 509         next = idx_start;
 510         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 511                 if (req->wb_index > idx_end)
 512                         break;
 513
 514                 next = req->wb_index + 1;
 515                 BUG_ON(!NFS_WBACK_BUSY(req));
 516
 517                 kref_get(&req->wb_kref);
 518                 spin_unlock(&inode->i_lock);
 519                 error = nfs_wait_on_request(req);
 520                 nfs_release_request(req);
 521                 spin_lock(&inode->i_lock);
 522                 if (error < 0)
 523                         return error;
 524                 res++;
 525         }
 526         return res;
 527 }
 528
 529 static void nfs_cancel_commit_list(struct list_head *head)
 530 {
 531         struct nfs_page *req;
 532
 533         while(!list_empty(head)) {
 534                 req = nfs_list_entry(head->next);
 535                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 536                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
 537                                 BDI_RECLAIMABLE);
 538                 nfs_list_remove_request(req);
 539                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 540                 nfs_inode_remove_request(req);
 541                 nfs_unlock_request(req);
 542         }
 543 }
 544
 545 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 546 /*
 547  * nfs_scan_commit - Scan an inode for commit requests
 548  * @inode: NFS inode to scan
 549  * @dst: destination list
 550  * @idx_start: lower bound of page->index to scan.
 551  * @npages: idx_start + npages sets the upper bound to scan.
 552  *
 553  * Moves requests from the inode's 'commit' request list.
 554  * The requests are *not* checked to ensure that they form a contiguous set.
 555  */
 556 static int
 557 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 558 {
 559         struct nfs_inode *nfsi = NFS_I(inode);
 560         int res = 0;
 561
 562         if (nfsi->ncommit != 0) {
 563                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 564                                 NFS_PAGE_TAG_COMMIT);
 565                 nfsi->ncommit -= res;
 566         }
 567         return res;
 568 }
 569 #else
 570 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 571 {
 572         return 0;
 573 }
 574 #endif
 575
 576 /*
 577  * Try to update any existing write request, or create one if there is none.
 578  * In order to match, the request's credentials must match those of
 579  * the calling process.
 580  *
 581  * Note: Should always be called with the Page Lock held!
 582  */
 583 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 584                 struct page *page, unsigned int offset, unsigned int bytes)
 585 {
 586         struct address_space *mapping = page->mapping;
 587         struct inode *inode = mapping->host;
 588         struct nfs_page         *req, *new = NULL;
 589         pgoff_t         rqend, end;
 590
 591         end = offset + bytes;
 592
 593         for (;;) {
 594                 /* Loop over all inode entries and see if we find
 595                  * A request for the page we wish to update
 596                  */
 597                 spin_lock(&inode->i_lock);
 598                 req = nfs_page_find_request_locked(page);
 599                 if (req) {
 600                         if (!nfs_set_page_tag_locked(req)) {
 601                                 int error;
 602
 603                                 spin_unlock(&inode->i_lock);
 604                                 error = nfs_wait_on_request(req);
 605                                 nfs_release_request(req);
 606                                 if (error < 0) {
 607                                         if (new)
 608                                                 nfs_release_request(new);
 609                                         return ERR_PTR(error);
 610                                 }
 611                                 continue;
 612                         }
 613                         spin_unlock(&inode->i_lock);
 614                         if (new)
 615                                 nfs_release_request(new);
 616                         break;
 617                 }
 618
 619                 if (new) {
 620                         int error;
 621                         nfs_lock_request_dontget(new);
 622                         error = nfs_inode_add_request(inode, new);
 623                         if (error) {
 624                                 spin_unlock(&inode->i_lock);
 625                                 nfs_unlock_request(new);
 626                                 return ERR_PTR(error);
 627                         }
 628                         spin_unlock(&inode->i_lock);
 629                         req = new;
 630                         goto zero_page;
 631                 }
 632                 spin_unlock(&inode->i_lock);
 633
 634                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 635                 if (IS_ERR(new))
 636                         return new;
 637         }
 638
 639         /* We have a request for our page.
 640          * If the creds don't match, or the
 641          * page addresses don't match,
 642          * tell the caller to wait on the conflicting
 643          * request.
 644          */
 645         rqend = req->wb_offset + req->wb_bytes;
 646         if (req->wb_context != ctx
 647             || req->wb_page != page
 648             || !nfs_dirty_request(req)
 649             || offset > rqend || end < req->wb_offset) {
 650                 nfs_clear_page_tag_locked(req);
 651                 return ERR_PTR(-EBUSY);
 652         }
 653
 654         /* Okay, the request matches. Update the region */
 655         if (offset < req->wb_offset) {
 656                 req->wb_offset = offset;
 657                 req->wb_pgbase = offset;
 658                 req->wb_bytes = max(end, rqend) - req->wb_offset;
 659                 goto zero_page;
 660         }
 661
 662         if (end > rqend)
 663                 req->wb_bytes = end - req->wb_offset;
 664
 665         return req;
 666 zero_page:
 667         /* If this page might potentially be marked as up to date,
 668          * then we need to zero any uninitalised data. */
 669         if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE
 670                         && !PageUptodate(req->wb_page))
 671                 zero_user_segment(req->wb_page, req->wb_bytes, PAGE_CACHE_SIZE);
 672         return req;
 673 }
 674
 675 int nfs_flush_incompatible(struct file *file, struct page *page)
 676 {
 677         struct nfs_open_context *ctx = nfs_file_open_context(file);
 678         struct nfs_page *req;
 679         int do_flush, status;
 680         /*
 681          * Look for a request corresponding to this page. If there
 682          * is one, and it belongs to another file, we flush it out
 683          * before we try to copy anything into the page. Do this
 684          * due to the lack of an ACCESS-type call in NFSv2.
 685          * Also do the same if we find a request from an existing
 686          * dropped page.
 687          */
 688         do {
 689                 req = nfs_page_find_request(page);
 690                 if (req == NULL)
 691                         return 0;
 692                 do_flush = req->wb_page != page || req->wb_context != ctx
 693                         || !nfs_dirty_request(req);
 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@%Ld)\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         __set_page_dirty_nobuffers(page);
 747
 748         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 749                         status, (long long)i_size_read(inode));
 750         if (status < 0)
 751                 nfs_set_pageerror(page);
 752         return status;
 753 }
 754
 755 static void nfs_writepage_release(struct nfs_page *req)
 756 {
 757
 758         if (PageError(req->wb_page)) {
 759                 nfs_end_page_writeback(req->wb_page);
 760                 nfs_inode_remove_request(req);
 761         } else if (!nfs_reschedule_unstable_write(req)) {
 762                 /* Set the PG_uptodate flag */
 763                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 764                 nfs_end_page_writeback(req->wb_page);
 765                 nfs_inode_remove_request(req);
 766         } else
 767                 nfs_end_page_writeback(req->wb_page);
 768         nfs_clear_page_tag_locked(req);
 769 }
 770
 771 static int flush_task_priority(int how)
 772 {
 773         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 774                 case FLUSH_HIGHPRI:
 775                         return RPC_PRIORITY_HIGH;
 776                 case FLUSH_LOWPRI:
 777                         return RPC_PRIORITY_LOW;
 778         }
 779         return RPC_PRIORITY_NORMAL;
 780 }
 781
 782 /*
 783  * Set up the argument/result storage required for the RPC call.
 784  */
 785 static void nfs_write_rpcsetup(struct nfs_page *req,
 786                 struct nfs_write_data *data,
 787                 const struct rpc_call_ops *call_ops,
 788                 unsigned int count, unsigned int offset,
 789                 int how)
 790 {
 791         struct inode *inode = req->wb_context->path.dentry->d_inode;
 792         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 793         int priority = flush_task_priority(how);
 794         struct rpc_task *task;
 795         struct rpc_message msg = {
 796                 .rpc_argp = &data->args,
 797                 .rpc_resp = &data->res,
 798                 .rpc_cred = req->wb_context->cred,
 799         };
 800         struct rpc_task_setup task_setup_data = {
 801                 .rpc_client = NFS_CLIENT(inode),
 802                 .task = &data->task,
 803                 .rpc_message = &msg,
 804                 .callback_ops = call_ops,
 805                 .callback_data = data,
 806                 .flags = flags,
 807                 .priority = priority,
 808         };
 809
 810         /* Set up the RPC argument and reply structs
 811          * NB: take care not to mess about with data->commit et al. */
 812
 813         data->req = req;
 814         data->inode = inode = req->wb_context->path.dentry->d_inode;
 815         data->cred = msg.rpc_cred;
 816
 817         data->args.fh     = NFS_FH(inode);
 818         data->args.offset = req_offset(req) + offset;
 819         data->args.pgbase = req->wb_pgbase + offset;
 820         data->args.pages  = data->pagevec;
 821         data->args.count  = count;
 822         data->args.context = get_nfs_open_context(req->wb_context);
 823         data->args.stable  = NFS_UNSTABLE;
 824         if (how & FLUSH_STABLE) {
 825                 data->args.stable = NFS_DATA_SYNC;
 826                 if (!NFS_I(inode)->ncommit)
 827                         data->args.stable = NFS_FILE_SYNC;
 828         }
 829
 830         data->res.fattr   = &data->fattr;
 831         data->res.count   = count;
 832         data->res.verf    = &data->verf;
 833         nfs_fattr_init(&data->fattr);
 834
 835         /* Set up the initial task struct.  */
 836         NFS_PROTO(inode)->write_setup(data, &msg);
 837
 838         dprintk("NFS: %5u initiated write call "
 839                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 840                 data->task.tk_pid,
 841                 inode->i_sb->s_id,
 842                 (long long)NFS_FILEID(inode),
 843                 count,
 844                 (unsigned long long)data->args.offset);
 845
 846         task = rpc_run_task(&task_setup_data);
 847         if (!IS_ERR(task))
 848                 rpc_put_task(task);
 849 }
 850
 851 /*
 852  * Generate multiple small requests to write out a single
 853  * contiguous dirty area on one page.
 854  */
 855 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 856 {
 857         struct nfs_page *req = nfs_list_entry(head->next);
 858         struct page *page = req->wb_page;
 859         struct nfs_write_data *data;
 860         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 861         unsigned int offset;
 862         int requests = 0;
 863         LIST_HEAD(list);
 864
 865         nfs_list_remove_request(req);
 866
 867         nbytes = count;
 868         do {
 869                 size_t len = min(nbytes, wsize);
 870
 871                 data = nfs_writedata_alloc(1);
 872                 if (!data)
 873                         goto out_bad;
 874                 list_add(&data->pages, &list);
 875                 requests++;
 876                 nbytes -= len;
 877         } while (nbytes != 0);
 878         atomic_set(&req->wb_complete, requests);
 879
 880         ClearPageError(page);
 881         offset = 0;
 882         nbytes = count;
 883         do {
 884                 data = list_entry(list.next, struct nfs_write_data, pages);
 885                 list_del_init(&data->pages);
 886
 887                 data->pagevec[0] = page;
 888
 889                 if (nbytes < wsize)
 890                         wsize = nbytes;
 891                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 892                                    wsize, offset, how);
 893                 offset += wsize;
 894                 nbytes -= wsize;
 895         } while (nbytes != 0);
 896
 897         return 0;
 898
 899 out_bad:
 900         while (!list_empty(&list)) {
 901                 data = list_entry(list.next, struct nfs_write_data, pages);
 902                 list_del(&data->pages);
 903                 nfs_writedata_release(data);
 904         }
 905         nfs_redirty_request(req);
 906         nfs_end_page_writeback(req->wb_page);
 907         nfs_clear_page_tag_locked(req);
 908         return -ENOMEM;
 909 }
 910
 911 /*
 912  * Create an RPC task for the given write request and kick it.
 913  * The page must have been locked by the caller.
 914  *
 915  * It may happen that the page we're passed is not marked dirty.
 916  * This is the case if nfs_updatepage detects a conflicting request
 917  * that has been written but not committed.
 918  */
 919 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 920 {
 921         struct nfs_page         *req;
 922         struct page             **pages;
 923         struct nfs_write_data   *data;
 924
 925         data = nfs_writedata_alloc(npages);
 926         if (!data)
 927                 goto out_bad;
 928
 929         pages = data->pagevec;
 930         while (!list_empty(head)) {
 931                 req = nfs_list_entry(head->next);
 932                 nfs_list_remove_request(req);
 933                 nfs_list_add_request(req, &data->pages);
 934                 ClearPageError(req->wb_page);
 935                 *pages++ = req->wb_page;
 936         }
 937         req = nfs_list_entry(data->pages.next);
 938
 939         /* Set up the argument struct */
 940         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 941
 942         return 0;
 943  out_bad:
 944         while (!list_empty(head)) {
 945                 req = nfs_list_entry(head->next);
 946                 nfs_list_remove_request(req);
 947                 nfs_redirty_request(req);
 948                 nfs_end_page_writeback(req->wb_page);
 949                 nfs_clear_page_tag_locked(req);
 950         }
 951         return -ENOMEM;
 952 }
 953
 954 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 955                                   struct inode *inode, int ioflags)
 956 {
 957         size_t wsize = NFS_SERVER(inode)->wsize;
 958
 959         if (wsize < PAGE_CACHE_SIZE)
 960                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 961         else
 962                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 963 }
 964
 965 /*
 966  * Handle a write reply that flushed part of a page.
 967  */
 968 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 969 {
 970         struct nfs_write_data   *data = calldata;
 971         struct nfs_page         *req = data->req;
 972         struct page             *page = req->wb_page;
 973
 974         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 975                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 976                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 977                 req->wb_bytes,
 978                 (long long)req_offset(req));
 979
 980         if (nfs_writeback_done(task, data) != 0)
 981                 return;
 982
 983         if (task->tk_status < 0) {
 984                 nfs_set_pageerror(page);
 985                 nfs_context_set_write_error(req->wb_context, task->tk_status);
 986                 dprintk(", error = %d\n", task->tk_status);
 987                 goto out;
 988         }
 989
 990         if (nfs_write_need_commit(data)) {
 991                 struct inode *inode = page->mapping->host;
 992
 993                 spin_lock(&inode->i_lock);
 994                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 995                         /* Do nothing we need to resend the writes */
 996                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 997                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 998                         dprintk(" defer commit\n");
 999                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1000                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1001                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1002                         dprintk(" server reboot detected\n");
1003                 }
1004                 spin_unlock(&inode->i_lock);
1005         } else
1006                 dprintk(" OK\n");
1007
1008 out:
1009         if (atomic_dec_and_test(&req->wb_complete))
1010                 nfs_writepage_release(req);
1011 }
1012
1013 static const struct rpc_call_ops nfs_write_partial_ops = {
1014         .rpc_call_done = nfs_writeback_done_partial,
1015         .rpc_release = nfs_writedata_release,
1016 };
1017
1018 /*
1019  * Handle a write reply that flushes a whole page.
1020  *
1021  * FIXME: There is an inherent race with invalidate_inode_pages and
1022  *        writebacks since the page->count is kept > 1 for as long
1023  *        as the page has a write request pending.
1024  */
1025 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1026 {
1027         struct nfs_write_data   *data = calldata;
1028         struct nfs_page         *req;
1029         struct page             *page;
1030
1031         if (nfs_writeback_done(task, data) != 0)
1032                 return;
1033
1034         /* Update attributes as result of writeback. */
1035         while (!list_empty(&data->pages)) {
1036                 req = nfs_list_entry(data->pages.next);
1037                 nfs_list_remove_request(req);
1038                 page = req->wb_page;
1039
1040                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1041                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1042                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1043                         req->wb_bytes,
1044                         (long long)req_offset(req));
1045
1046                 if (task->tk_status < 0) {
1047                         nfs_set_pageerror(page);
1048                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1049                         dprintk(", error = %d\n", task->tk_status);
1050                         goto remove_request;
1051                 }
1052
1053                 if (nfs_write_need_commit(data)) {
1054                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1055                         nfs_mark_request_commit(req);
1056                         nfs_end_page_writeback(page);
1057                         dprintk(" marked for commit\n");
1058                         goto next;
1059                 }
1060                 /* Set the PG_uptodate flag? */
1061                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1062                 dprintk(" OK\n");
1063 remove_request:
1064                 nfs_end_page_writeback(page);
1065                 nfs_inode_remove_request(req);
1066         next:
1067                 nfs_clear_page_tag_locked(req);
1068         }
1069 }
1070
1071 static const struct rpc_call_ops nfs_write_full_ops = {
1072         .rpc_call_done = nfs_writeback_done_full,
1073         .rpc_release = nfs_writedata_release,
1074 };
1075
1076
1077 /*
1078  * This function is called when the WRITE call is complete.
1079  */
1080 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1081 {
1082         struct nfs_writeargs    *argp = &data->args;
1083         struct nfs_writeres     *resp = &data->res;
1084         int status;
1085
1086         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1087                 task->tk_pid, task->tk_status);
1088
1089         /*
1090          * ->write_done will attempt to use post-op attributes to detect
1091          * conflicting writes by other clients.  A strict interpretation
1092          * of close-to-open would allow us to continue caching even if
1093          * another writer had changed the file, but some applications
1094          * depend on tighter cache coherency when writing.
1095          */
1096         status = NFS_PROTO(data->inode)->write_done(task, data);
1097         if (status != 0)
1098                 return status;
1099         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1100
1101 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1102         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1103                 /* We tried a write call, but the server did not
1104                  * commit data to stable storage even though we
1105                  * requested it.
1106                  * Note: There is a known bug in Tru64 < 5.0 in which
1107                  *       the server reports NFS_DATA_SYNC, but performs
1108                  *       NFS_FILE_SYNC. We therefore implement this checking
1109                  *       as a dprintk() in order to avoid filling syslog.
1110                  */
1111                 static unsigned long    complain;
1112
1113                 if (time_before(complain, jiffies)) {
1114                         dprintk("NFS: faulty NFS server %s:"
1115                                 " (committed = %d) != (stable = %d)\n",
1116                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1117                                 resp->verf->committed, argp->stable);
1118                         complain = jiffies + 300 * HZ;
1119                 }
1120         }
1121 #endif
1122         /* Is this a short write? */
1123         if (task->tk_status >= 0 && resp->count < argp->count) {
1124                 static unsigned long    complain;
1125
1126                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1127
1128                 /* Has the server at least made some progress? */
1129                 if (resp->count != 0) {
1130                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1131                         if (resp->verf->committed != NFS_UNSTABLE) {
1132                                 /* Resend from where the server left off */
1133                                 argp->offset += resp->count;
1134                                 argp->pgbase += resp->count;
1135                                 argp->count -= resp->count;
1136                         } else {
1137                                 /* Resend as a stable write in order to avoid
1138                                  * headaches in the case of a server crash.
1139                                  */
1140                                 argp->stable = NFS_FILE_SYNC;
1141                         }
1142                         rpc_restart_call(task);
1143                         return -EAGAIN;
1144                 }
1145                 if (time_before(complain, jiffies)) {
1146                         printk(KERN_WARNING
1147                                "NFS: Server wrote zero bytes, expected %u.\n",
1148                                         argp->count);
1149                         complain = jiffies + 300 * HZ;
1150                 }
1151                 /* Can't do anything about it except throw an error. */
1152                 task->tk_status = -EIO;
1153         }
1154         return 0;
1155 }
1156
1157
1158 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1159 void nfs_commit_release(void *data)
1160 {
1161         struct nfs_write_data *wdata = data;
1162
1163         put_nfs_open_context(wdata->args.context);
1164         nfs_commit_free(wdata);
1165 }
1166
1167 /*
1168  * Set up the argument/result storage required for the RPC call.
1169  */
1170 static void nfs_commit_rpcsetup(struct list_head *head,
1171                 struct nfs_write_data *data,
1172                 int how)
1173 {
1174         struct nfs_page *first = nfs_list_entry(head->next);
1175         struct inode *inode = first->wb_context->path.dentry->d_inode;
1176         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1177         int priority = flush_task_priority(how);
1178         struct rpc_task *task;
1179         struct rpc_message msg = {
1180                 .rpc_argp = &data->args,
1181                 .rpc_resp = &data->res,
1182                 .rpc_cred = first->wb_context->cred,
1183         };
1184         struct rpc_task_setup task_setup_data = {
1185                 .task = &data->task,
1186                 .rpc_client = NFS_CLIENT(inode),
1187                 .rpc_message = &msg,
1188                 .callback_ops = &nfs_commit_ops,
1189                 .callback_data = data,
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