fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Writing file data over NFS.
   5  *
   6  * We do it like this: When a (user) process wishes to write data to an
   7  * NFS file, a write request is allocated that contains the RPC task data
   8  * plus some info on the page to be written, and added to the inode's
   9  * write chain. If the process writes past the end of the page, an async
  10  * RPC call to write the page is scheduled immediately; otherwise, the call
  11  * is delayed for a few seconds.
  12  *
  13  * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
  14  *
  15  * Write requests are kept on the inode's writeback list. Each entry in
  16  * that list references the page (portion) to be written. When the
  17  * cache timeout has expired, the RPC task is woken up, and tries to
  18  * lock the page. As soon as it manages to do so, the request is moved
  19  * from the writeback list to the writelock list.
  20  *
  21  * Note: we must make sure never to confuse the inode passed in the
  22  * write_page request with the one in page->inode. As far as I understand
  23  * it, these are different when doing a swap-out.
  24  *
  25  * To understand everything that goes on here and in the NFS read code,
  26  * one should be aware that a page is locked in exactly one of the following
  27  * cases:
  28  *
  29  *  -   A write request is in progress.
  30  *  -   A user process is in generic_file_write/nfs_update_page
  31  *  -   A user process is in generic_file_read
  32  *
  33  * Also note that because of the way pages are invalidated in
  34  * nfs_revalidate_inode, the following assertions hold:
  35  *
  36  *  -   If a page is dirty, there will be no read requests (a page will
  37  *      not be re-read unless invalidated by nfs_revalidate_inode).
  38  *  -   If the page is not uptodate, there will be no pending write
  39  *      requests, and no process will be in nfs_update_page.
  40  *
  41  * FIXME: Interaction with the vmscan routines is not optimal yet.
  42  * Either vmscan must be made nfs-savvy, or we need a different page
  43  * reclaim concept that supports something like FS-independent
  44  * buffer_heads with a b_ops-> field.
  45  *
  46  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
  47  */
  48
  49 #include <linux/types.h>
  50 #include <linux/slab.h>
  51 #include <linux/mm.h>
  52 #include <linux/pagemap.h>
  53 #include <linux/file.h>
  54 #include <linux/writeback.h>
  55
  56 #include <linux/sunrpc/clnt.h>
  57 #include <linux/nfs_fs.h>
  58 #include <linux/nfs_mount.h>
  59 #include <linux/nfs_page.h>
  60 #include <linux/backing-dev.h>
  61
  62 #include <asm/uaccess.h>
  63 #include <linux/smp_lock.h>
  64
  65 #include "delegation.h"
  66 #include "internal.h"
  67 #include "iostat.h"
  68
  69 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  70
  71 #define MIN_POOL_WRITE          (32)
  72 #define MIN_POOL_COMMIT         (4)
  73
  74 /*
  75  * Local function declarations
  76  */
  77 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  78                                             struct page *,
  79                                             unsigned int, unsigned int);
  80 static int nfs_wait_on_write_congestion(struct address_space *, int);
  81 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
  82 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  83 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how);
  84 static const struct rpc_call_ops nfs_write_partial_ops;
  85 static const struct rpc_call_ops nfs_write_full_ops;
  86 static const struct rpc_call_ops nfs_commit_ops;
  87
  88 static kmem_cache_t *nfs_wdata_cachep;
  89 static mempool_t *nfs_wdata_mempool;
  90 static mempool_t *nfs_commit_mempool;
  91
  92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  93
  94 struct nfs_write_data *nfs_commit_alloc(void)
  95 {
  96         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
  97
  98         if (p) {
  99                 memset(p, 0, sizeof(*p));
 100                 INIT_LIST_HEAD(&p->pages);
 101         }
 102         return p;
 103 }
 104
 105 void nfs_commit_rcu_free(struct rcu_head *head)
 106 {
 107         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 108         if (p && (p->pagevec != &p->page_array[0]))
 109                 kfree(p->pagevec);
 110         mempool_free(p, nfs_commit_mempool);
 111 }
 112
 113 void nfs_commit_free(struct nfs_write_data *wdata)
 114 {
 115         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
 116 }
 117
 118 struct nfs_write_data *nfs_writedata_alloc(size_t len)
 119 {
 120         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 121         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
 122
 123         if (p) {
 124                 memset(p, 0, sizeof(*p));
 125                 INIT_LIST_HEAD(&p->pages);
 126                 p->npages = pagecount;
 127                 if (pagecount <= ARRAY_SIZE(p->page_array))
 128                         p->pagevec = p->page_array;
 129                 else {
 130                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 131                         if (!p->pagevec) {
 132                                 mempool_free(p, nfs_wdata_mempool);
 133                                 p = NULL;
 134                         }
 135                 }
 136         }
 137         return p;
 138 }
 139
 140 static void nfs_writedata_rcu_free(struct rcu_head *head)
 141 {
 142         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 143         if (p && (p->pagevec != &p->page_array[0]))
 144                 kfree(p->pagevec);
 145         mempool_free(p, nfs_wdata_mempool);
 146 }
 147
 148 static void nfs_writedata_free(struct nfs_write_data *wdata)
 149 {
 150         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 151 }
 152
 153 void nfs_writedata_release(void *wdata)
 154 {
 155         nfs_writedata_free(wdata);
 156 }
 157
 158 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 159 {
 160         struct nfs_page *req = NULL;
 161
 162         if (PagePrivate(page)) {
 163                 req = (struct nfs_page *)page_private(page);
 164                 if (req != NULL)
 165                         atomic_inc(&req->wb_count);
 166         }
 167         return req;
 168 }
 169
 170 static struct nfs_page *nfs_page_find_request(struct page *page)
 171 {
 172         struct nfs_page *req = NULL;
 173         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 174
 175         spin_lock(req_lock);
 176         req = nfs_page_find_request_locked(page);
 177         spin_unlock(req_lock);
 178         return req;
 179 }
 180
 181 /* Adjust the file length if we're writing beyond the end */
 182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         loff_t end, i_size = i_size_read(inode);
 186         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 187
 188         if (i_size > 0 && page->index < end_index)
 189                 return;
 190         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 191         if (i_size >= end)
 192                 return;
 193         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 194         i_size_write(inode, end);
 195 }
 196
 197 /* We can set the PG_uptodate flag if we see that a write request
 198  * covers the full page.
 199  */
 200 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 201 {
 202         if (PageUptodate(page))
 203                 return;
 204         if (base != 0)
 205                 return;
 206         if (count != nfs_page_length(page))
 207                 return;
 208         if (count != PAGE_CACHE_SIZE)
 209                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 210         SetPageUptodate(page);
 211 }
 212
 213 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 214                 unsigned int offset, unsigned int count)
 215 {
 216         struct nfs_page *req;
 217         int ret;
 218
 219         for (;;) {
 220                 req = nfs_update_request(ctx, page, offset, count);
 221                 if (!IS_ERR(req))
 222                         break;
 223                 ret = PTR_ERR(req);
 224                 if (ret != -EBUSY)
 225                         return ret;
 226                 ret = nfs_wb_page(page->mapping->host, page);
 227                 if (ret != 0)
 228                         return ret;
 229         }
 230         /* Update file length */
 231         nfs_grow_file(page, offset, count);
 232         /* Set the PG_uptodate flag? */
 233         nfs_mark_uptodate(page, offset, count);
 234         nfs_unlock_request(req);
 235         return 0;
 236 }
 237
 238 static int wb_priority(struct writeback_control *wbc)
 239 {
 240         if (wbc->for_reclaim)
 241                 return FLUSH_HIGHPRI;
 242         if (wbc->for_kupdate)
 243                 return FLUSH_LOWPRI;
 244         return 0;
 245 }
 246
 247 /*
 248  * Write an mmapped page to the server.
 249  */
 250 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 251 {
 252         struct nfs_open_context *ctx;
 253         struct inode *inode = page->mapping->host;
 254         struct nfs_page *req;
 255         unsigned offset;
 256         int err = 0;
 257
 258         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 259         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 260
 261         req = nfs_page_find_request(page);
 262         if (req != NULL) {
 263                 int flushme = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 264                 nfs_release_request(req);
 265                 if (!flushme)
 266                         goto out;
 267                 /* Ensure we've flushed out the invalid write */
 268                 nfs_wb_page_priority(inode, page, wb_priority(wbc));
 269         }
 270
 271         offset = nfs_page_length(page);
 272         if (!offset)
 273                 goto out;
 274
 275         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 276         if (ctx == NULL) {
 277                 err = -EBADF;
 278                 goto out;
 279         }
 280         err = nfs_writepage_setup(ctx, page, 0, offset);
 281         put_nfs_open_context(ctx);
 282
 283 out:
 284         if (!wbc->for_writepages)
 285                 nfs_flush_mapping(page->mapping, wbc, wb_priority(wbc));
 286         unlock_page(page);
 287         return err;
 288 }
 289
 290 /*
 291  * Note: causes nfs_update_request() to block on the assumption
 292  *       that the writeback is generated due to memory pressure.
 293  */
 294 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 295 {
 296         struct backing_dev_info *bdi = mapping->backing_dev_info;
 297         struct inode *inode = mapping->host;
 298         int err;
 299
 300         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 301
 302         err = generic_writepages(mapping, wbc);
 303         if (err)
 304                 return err;
 305         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 306                 if (wbc->nonblocking)
 307                         return 0;
 308                 nfs_wait_on_write_congestion(mapping, 0);
 309         }
 310         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 311         if (err < 0)
 312                 goto out;
 313         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 314         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 315                 err = nfs_wait_on_requests(inode, 0, 0);
 316                 if (err < 0)
 317                         goto out;
 318         }
 319         err = nfs_commit_inode(inode, wb_priority(wbc));
 320         if (err > 0)
 321                 err = 0;
 322 out:
 323         clear_bit(BDI_write_congested, &bdi->state);
 324         wake_up_all(&nfs_write_congestion);
 325         congestion_end(WRITE);
 326         return err;
 327 }
 328
 329 /*
 330  * Insert a write request into an inode
 331  */
 332 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 333 {
 334         struct nfs_inode *nfsi = NFS_I(inode);
 335         int error;
 336
 337         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 338         BUG_ON(error == -EEXIST);
 339         if (error)
 340                 return error;
 341         if (!nfsi->npages) {
 342                 igrab(inode);
 343                 nfs_begin_data_update(inode);
 344                 if (nfs_have_delegation(inode, FMODE_WRITE))
 345                         nfsi->change_attr++;
 346         }
 347         SetPagePrivate(req->wb_page);
 348         set_page_private(req->wb_page, (unsigned long)req);
 349         nfsi->npages++;
 350         atomic_inc(&req->wb_count);
 351         return 0;
 352 }
 353
 354 /*
 355  * Insert a write request into an inode
 356  */
 357 static void nfs_inode_remove_request(struct nfs_page *req)
 358 {
 359         struct inode *inode = req->wb_context->dentry->d_inode;
 360         struct nfs_inode *nfsi = NFS_I(inode);
 361
 362         BUG_ON (!NFS_WBACK_BUSY(req));
 363
 364         spin_lock(&nfsi->req_lock);
 365         set_page_private(req->wb_page, 0);
 366         ClearPagePrivate(req->wb_page);
 367         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 368         nfsi->npages--;
 369         if (!nfsi->npages) {
 370                 spin_unlock(&nfsi->req_lock);
 371                 nfs_end_data_update(inode);
 372                 iput(inode);
 373         } else
 374                 spin_unlock(&nfsi->req_lock);
 375         nfs_clear_request(req);
 376         nfs_release_request(req);
 377 }
 378
 379 /*
 380  * Add a request to the inode's dirty list.
 381  */
 382 static void
 383 nfs_mark_request_dirty(struct nfs_page *req)
 384 {
 385         struct inode *inode = req->wb_context->dentry->d_inode;
 386         struct nfs_inode *nfsi = NFS_I(inode);
 387
 388         spin_lock(&nfsi->req_lock);
 389         radix_tree_tag_set(&nfsi->nfs_page_tree,
 390                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 391         nfs_list_add_request(req, &nfsi->dirty);
 392         nfsi->ndirty++;
 393         spin_unlock(&nfsi->req_lock);
 394         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 395         mark_inode_dirty(inode);
 396 }
 397
 398 /*
 399  * Check if a request is dirty
 400  */
 401 static inline int
 402 nfs_dirty_request(struct nfs_page *req)
 403 {
 404         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 405         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 406 }
 407
 408 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 409 /*
 410  * Add a request to the inode's commit list.
 411  */
 412 static void
 413 nfs_mark_request_commit(struct nfs_page *req)
 414 {
 415         struct inode *inode = req->wb_context->dentry->d_inode;
 416         struct nfs_inode *nfsi = NFS_I(inode);
 417
 418         spin_lock(&nfsi->req_lock);
 419         nfs_list_add_request(req, &nfsi->commit);
 420         nfsi->ncommit++;
 421         spin_unlock(&nfsi->req_lock);
 422         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 423         mark_inode_dirty(inode);
 424 }
 425 #endif
 426
 427 /*
 428  * Wait for a request to complete.
 429  *
 430  * Interruptible by signals only if mounted with intr flag.
 431  */
 432 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 433 {
 434         struct nfs_inode *nfsi = NFS_I(inode);
 435         struct nfs_page *req;
 436         unsigned long           idx_end, next;
 437         unsigned int            res = 0;
 438         int                     error;
 439
 440         if (npages == 0)
 441                 idx_end = ~0;
 442         else
 443                 idx_end = idx_start + npages - 1;
 444
 445         next = idx_start;
 446         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 447                 if (req->wb_index > idx_end)
 448                         break;
 449
 450                 next = req->wb_index + 1;
 451                 BUG_ON(!NFS_WBACK_BUSY(req));
 452
 453                 atomic_inc(&req->wb_count);
 454                 spin_unlock(&nfsi->req_lock);
 455                 error = nfs_wait_on_request(req);
 456                 nfs_release_request(req);
 457                 spin_lock(&nfsi->req_lock);
 458                 if (error < 0)
 459                         return error;
 460                 res++;
 461         }
 462         return res;
 463 }
 464
 465 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 466 {
 467         struct nfs_inode *nfsi = NFS_I(inode);
 468         int ret;
 469
 470         spin_lock(&nfsi->req_lock);
 471         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 472         spin_unlock(&nfsi->req_lock);
 473         return ret;
 474 }
 475
 476 static void nfs_cancel_dirty_list(struct list_head *head)
 477 {
 478         struct nfs_page *req;
 479         while(!list_empty(head)) {
 480                 req = nfs_list_entry(head->next);
 481                 nfs_list_remove_request(req);
 482                 nfs_inode_remove_request(req);
 483                 nfs_clear_page_writeback(req);
 484         }
 485 }
 486
 487 static void nfs_cancel_commit_list(struct list_head *head)
 488 {
 489         struct nfs_page *req;
 490
 491         while(!list_empty(head)) {
 492                 req = nfs_list_entry(head->next);
 493                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 494                 nfs_list_remove_request(req);
 495                 nfs_inode_remove_request(req);
 496                 nfs_unlock_request(req);
 497         }
 498 }
 499
 500 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 501 /*
 502  * nfs_scan_commit - Scan an inode for commit requests
 503  * @inode: NFS inode to scan
 504  * @dst: destination list
 505  * @idx_start: lower bound of page->index to scan.
 506  * @npages: idx_start + npages sets the upper bound to scan.
 507  *
 508  * Moves requests from the inode's 'commit' request list.
 509  * The requests are *not* checked to ensure that they form a contiguous set.
 510  */
 511 static int
 512 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 513 {
 514         struct nfs_inode *nfsi = NFS_I(inode);
 515         int res = 0;
 516
 517         if (nfsi->ncommit != 0) {
 518                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 519                 nfsi->ncommit -= res;
 520                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 521                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 522         }
 523         return res;
 524 }
 525 #else
 526 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 527 {
 528         return 0;
 529 }
 530 #endif
 531
 532 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 533 {
 534         struct backing_dev_info *bdi = mapping->backing_dev_info;
 535         DEFINE_WAIT(wait);
 536         int ret = 0;
 537
 538         might_sleep();
 539
 540         if (!bdi_write_congested(bdi))
 541                 return 0;
 542
 543         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 544
 545         if (intr) {
 546                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 547                 sigset_t oldset;
 548
 549                 rpc_clnt_sigmask(clnt, &oldset);
 550                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 551                 if (bdi_write_congested(bdi)) {
 552                         if (signalled())
 553                                 ret = -ERESTARTSYS;
 554                         else
 555                                 schedule();
 556                 }
 557                 rpc_clnt_sigunmask(clnt, &oldset);
 558         } else {
 559                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 560                 if (bdi_write_congested(bdi))
 561                         schedule();
 562         }
 563         finish_wait(&nfs_write_congestion, &wait);
 564         return ret;
 565 }
 566
 567
 568 /*
 569  * Try to update any existing write request, or create one if there is none.
 570  * In order to match, the request's credentials must match those of
 571  * the calling process.
 572  *
 573  * Note: Should always be called with the Page Lock held!
 574  */
 575 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 576                 struct page *page, unsigned int offset, unsigned int bytes)
 577 {
 578         struct inode *inode = page->mapping->host;
 579         struct nfs_inode *nfsi = NFS_I(inode);
 580         struct nfs_page         *req, *new = NULL;
 581         unsigned long           rqend, end;
 582
 583         end = offset + bytes;
 584
 585         if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
 586                 return ERR_PTR(-ERESTARTSYS);
 587         for (;;) {
 588                 /* Loop over all inode entries and see if we find
 589                  * A request for the page we wish to update
 590                  */
 591                 spin_lock(&nfsi->req_lock);
 592                 req = nfs_page_find_request_locked(page);
 593                 if (req) {
 594                         if (!nfs_lock_request_dontget(req)) {
 595                                 int error;
 596
 597                                 spin_unlock(&nfsi->req_lock);
 598                                 error = nfs_wait_on_request(req);
 599                                 nfs_release_request(req);
 600                                 if (error < 0) {
 601                                         if (new)
 602                                                 nfs_release_request(new);
 603                                         return ERR_PTR(error);
 604                                 }
 605                                 continue;
 606                         }
 607                         spin_unlock(&nfsi->req_lock);
 608                         if (new)
 609                                 nfs_release_request(new);
 610                         break;
 611                 }
 612
 613                 if (new) {
 614                         int error;
 615                         nfs_lock_request_dontget(new);
 616                         error = nfs_inode_add_request(inode, new);
 617                         if (error) {
 618                                 spin_unlock(&nfsi->req_lock);
 619                                 nfs_unlock_request(new);
 620                                 return ERR_PTR(error);
 621                         }
 622                         spin_unlock(&nfsi->req_lock);
 623                         nfs_mark_request_dirty(new);
 624                         return new;
 625                 }
 626                 spin_unlock(&nfsi->req_lock);
 627
 628                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 629                 if (IS_ERR(new))
 630                         return new;
 631         }
 632
 633         /* We have a request for our page.
 634          * If the creds don't match, or the
 635          * page addresses don't match,
 636          * tell the caller to wait on the conflicting
 637          * request.
 638          */
 639         rqend = req->wb_offset + req->wb_bytes;
 640         if (req->wb_context != ctx
 641             || req->wb_page != page
 642             || !nfs_dirty_request(req)
 643             || offset > rqend || end < req->wb_offset) {
 644                 nfs_unlock_request(req);
 645                 return ERR_PTR(-EBUSY);
 646         }
 647
 648         /* Okay, the request matches. Update the region */
 649         if (offset < req->wb_offset) {
 650                 req->wb_offset = offset;
 651                 req->wb_pgbase = offset;
 652                 req->wb_bytes = rqend - req->wb_offset;
 653         }
 654
 655         if (end > rqend)
 656                 req->wb_bytes = end - req->wb_offset;
 657
 658         return req;
 659 }
 660
 661 int nfs_flush_incompatible(struct file *file, struct page *page)
 662 {
 663         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 664         struct nfs_page *req;
 665         int do_flush, status;
 666         /*
 667          * Look for a request corresponding to this page. If there
 668          * is one, and it belongs to another file, we flush it out
 669          * before we try to copy anything into the page. Do this
 670          * due to the lack of an ACCESS-type call in NFSv2.
 671          * Also do the same if we find a request from an existing
 672          * dropped page.
 673          */
 674         do {
 675                 req = nfs_page_find_request(page);
 676                 if (req == NULL)
 677                         return 0;
 678                 do_flush = req->wb_page != page || req->wb_context != ctx
 679                         || test_bit(PG_NEED_FLUSH, &req->wb_flags);
 680                 nfs_release_request(req);
 681                 if (!do_flush)
 682                         return 0;
 683                 status = nfs_wb_page(page->mapping->host, page);
 684         } while (status == 0);
 685         return status;
 686 }
 687
 688 /*
 689  * Update and possibly write a cached page of an NFS file.
 690  *
 691  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 692  * things with a page scheduled for an RPC call (e.g. invalidate it).
 693  */
 694 int nfs_updatepage(struct file *file, struct page *page,
 695                 unsigned int offset, unsigned int count)
 696 {
 697         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 698         struct inode    *inode = page->mapping->host;
 699         int             status = 0;
 700
 701         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 702
 703         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 704                 file->f_dentry->d_parent->d_name.name,
 705                 file->f_dentry->d_name.name, count,
 706                 (long long)(page_offset(page) +offset));
 707
 708         /* If we're not using byte range locks, and we know the page
 709          * is entirely in cache, it may be more efficient to avoid
 710          * fragmenting write requests.
 711          */
 712         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 713                 count = max(count + offset, nfs_page_length(page));
 714                 offset = 0;
 715         }
 716
 717         status = nfs_writepage_setup(ctx, page, offset, count);
 718
 719         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 720                         status, (long long)i_size_read(inode));
 721         if (status < 0)
 722                 ClearPageUptodate(page);
 723         return status;
 724 }
 725
 726 static void nfs_writepage_release(struct nfs_page *req)
 727 {
 728         end_page_writeback(req->wb_page);
 729
 730 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 731         if (!PageError(req->wb_page)) {
 732                 if (NFS_NEED_RESCHED(req)) {
 733                         nfs_mark_request_dirty(req);
 734                         goto out;
 735                 } else if (NFS_NEED_COMMIT(req)) {
 736                         nfs_mark_request_commit(req);
 737                         goto out;
 738                 }
 739         }
 740         nfs_inode_remove_request(req);
 741
 742 out:
 743         nfs_clear_commit(req);
 744         nfs_clear_reschedule(req);
 745 #else
 746         nfs_inode_remove_request(req);
 747 #endif
 748         nfs_clear_page_writeback(req);
 749 }
 750
 751 static inline int flush_task_priority(int how)
 752 {
 753         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 754                 case FLUSH_HIGHPRI:
 755                         return RPC_PRIORITY_HIGH;
 756                 case FLUSH_LOWPRI:
 757                         return RPC_PRIORITY_LOW;
 758         }
 759         return RPC_PRIORITY_NORMAL;
 760 }
 761
 762 /*
 763  * Set up the argument/result storage required for the RPC call.
 764  */
 765 static void nfs_write_rpcsetup(struct nfs_page *req,
 766                 struct nfs_write_data *data,
 767                 const struct rpc_call_ops *call_ops,
 768                 unsigned int count, unsigned int offset,
 769                 int how)
 770 {
 771         struct inode            *inode;
 772         int flags;
 773
 774         /* Set up the RPC argument and reply structs
 775          * NB: take care not to mess about with data->commit et al. */
 776
 777         data->req = req;
 778         data->inode = inode = req->wb_context->dentry->d_inode;
 779         data->cred = req->wb_context->cred;
 780
 781         data->args.fh     = NFS_FH(inode);
 782         data->args.offset = req_offset(req) + offset;
 783         data->args.pgbase = req->wb_pgbase + offset;
 784         data->args.pages  = data->pagevec;
 785         data->args.count  = count;
 786         data->args.context = req->wb_context;
 787
 788         data->res.fattr   = &data->fattr;
 789         data->res.count   = count;
 790         data->res.verf    = &data->verf;
 791         nfs_fattr_init(&data->fattr);
 792
 793         /* Set up the initial task struct.  */
 794         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 795         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 796         NFS_PROTO(inode)->write_setup(data, how);
 797
 798         data->task.tk_priority = flush_task_priority(how);
 799         data->task.tk_cookie = (unsigned long)inode;
 800
 801         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 802                 data->task.tk_pid,
 803                 inode->i_sb->s_id,
 804                 (long long)NFS_FILEID(inode),
 805                 count,
 806                 (unsigned long long)data->args.offset);
 807 }
 808
 809 static void nfs_execute_write(struct nfs_write_data *data)
 810 {
 811         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 812         sigset_t oldset;
 813
 814         rpc_clnt_sigmask(clnt, &oldset);
 815         rpc_execute(&data->task);
 816         rpc_clnt_sigunmask(clnt, &oldset);
 817 }
 818
 819 /*
 820  * Generate multiple small requests to write out a single
 821  * contiguous dirty area on one page.
 822  */
 823 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 824 {
 825         struct nfs_page *req = nfs_list_entry(head->next);
 826         struct page *page = req->wb_page;
 827         struct nfs_write_data *data;
 828         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 829         unsigned int offset;
 830         int requests = 0;
 831         LIST_HEAD(list);
 832
 833         nfs_list_remove_request(req);
 834
 835         nbytes = req->wb_bytes;
 836         do {
 837                 size_t len = min(nbytes, wsize);
 838
 839                 data = nfs_writedata_alloc(len);
 840                 if (!data)
 841                         goto out_bad;
 842                 list_add(&data->pages, &list);
 843                 requests++;
 844                 nbytes -= len;
 845         } while (nbytes != 0);
 846         atomic_set(&req->wb_complete, requests);
 847
 848         ClearPageError(page);
 849         set_page_writeback(page);
 850         offset = 0;
 851         nbytes = req->wb_bytes;
 852         do {
 853                 data = list_entry(list.next, struct nfs_write_data, pages);
 854                 list_del_init(&data->pages);
 855
 856                 data->pagevec[0] = page;
 857
 858                 if (nbytes > wsize) {
 859                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 860                                         wsize, offset, how);
 861                         offset += wsize;
 862                         nbytes -= wsize;
 863                 } else {
 864                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 865                                         nbytes, offset, how);
 866                         nbytes = 0;
 867                 }
 868                 nfs_execute_write(data);
 869         } while (nbytes != 0);
 870
 871         return 0;
 872
 873 out_bad:
 874         while (!list_empty(&list)) {
 875                 data = list_entry(list.next, struct nfs_write_data, pages);
 876                 list_del(&data->pages);
 877                 nfs_writedata_release(data);
 878         }
 879         nfs_mark_request_dirty(req);
 880         nfs_clear_page_writeback(req);
 881         return -ENOMEM;
 882 }
 883
 884 /*
 885  * Create an RPC task for the given write request and kick it.
 886  * The page must have been locked by the caller.
 887  *
 888  * It may happen that the page we're passed is not marked dirty.
 889  * This is the case if nfs_updatepage detects a conflicting request
 890  * that has been written but not committed.
 891  */
 892 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 893 {
 894         struct nfs_page         *req;
 895         struct page             **pages;
 896         struct nfs_write_data   *data;
 897         unsigned int            count;
 898
 899         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 900         if (!data)
 901                 goto out_bad;
 902
 903         pages = data->pagevec;
 904         count = 0;
 905         while (!list_empty(head)) {
 906                 req = nfs_list_entry(head->next);
 907                 nfs_list_remove_request(req);
 908                 nfs_list_add_request(req, &data->pages);
 909                 ClearPageError(req->wb_page);
 910                 set_page_writeback(req->wb_page);
 911                 *pages++ = req->wb_page;
 912                 count += req->wb_bytes;
 913         }
 914         req = nfs_list_entry(data->pages.next);
 915
 916         /* Set up the argument struct */
 917         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 918
 919         nfs_execute_write(data);
 920         return 0;
 921  out_bad:
 922         while (!list_empty(head)) {
 923                 struct nfs_page *req = nfs_list_entry(head->next);
 924                 nfs_list_remove_request(req);
 925                 nfs_mark_request_dirty(req);
 926                 nfs_clear_page_writeback(req);
 927         }
 928         return -ENOMEM;
 929 }
 930
 931 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 932 {
 933         LIST_HEAD(one_request);
 934         int (*flush_one)(struct inode *, struct list_head *, int);
 935         struct nfs_page *req;
 936         int wpages = NFS_SERVER(inode)->wpages;
 937         int wsize = NFS_SERVER(inode)->wsize;
 938         int error;
 939
 940         flush_one = nfs_flush_one;
 941         if (wsize < PAGE_CACHE_SIZE)
 942                 flush_one = nfs_flush_multi;
 943         /* For single writes, FLUSH_STABLE is more efficient */
 944         if (npages <= wpages && npages == NFS_I(inode)->npages
 945                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 946                 how |= FLUSH_STABLE;
 947
 948         do {
 949                 nfs_coalesce_requests(head, &one_request, wpages);
 950                 req = nfs_list_entry(one_request.next);
 951                 error = flush_one(inode, &one_request, how);
 952                 if (error < 0)
 953                         goto out_err;
 954         } while (!list_empty(head));
 955         return 0;
 956 out_err:
 957         while (!list_empty(head)) {
 958                 req = nfs_list_entry(head->next);
 959                 nfs_list_remove_request(req);
 960                 nfs_mark_request_dirty(req);
 961                 nfs_clear_page_writeback(req);
 962         }
 963         return error;
 964 }
 965
 966 /*
 967  * Handle a write reply that flushed part of a page.
 968  */
 969 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 970 {
 971         struct nfs_write_data   *data = calldata;
 972         struct nfs_page         *req = data->req;
 973         struct page             *page = req->wb_page;
 974
 975         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 976                 req->wb_context->dentry->d_inode->i_sb->s_id,
 977                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 978                 req->wb_bytes,
 979                 (long long)req_offset(req));
 980
 981         if (nfs_writeback_done(task, data) != 0)
 982                 return;
 983
 984         if (task->tk_status < 0) {
 985                 ClearPageUptodate(page);
 986                 SetPageError(page);
 987                 req->wb_context->error = task->tk_status;
 988                 dprintk(", error = %d\n", task->tk_status);
 989         } else {
 990 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 991                 if (data->verf.committed < NFS_FILE_SYNC) {
 992                         if (!NFS_NEED_COMMIT(req)) {
 993                                 nfs_defer_commit(req);
 994                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 995                                 dprintk(" defer commit\n");
 996                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 997                                 nfs_defer_reschedule(req);
 998                                 dprintk(" server reboot detected\n");
 999                         }
1000                 } else
1001 #endif
1002                         dprintk(" OK\n");
1003         }
1004
1005         if (atomic_dec_and_test(&req->wb_complete))
1006                 nfs_writepage_release(req);
1007 }
1008
1009 static const struct rpc_call_ops nfs_write_partial_ops = {
1010         .rpc_call_done = nfs_writeback_done_partial,
1011         .rpc_release = nfs_writedata_release,
1012 };
1013
1014 /*
1015  * Handle a write reply that flushes a whole page.
1016  *
1017  * FIXME: There is an inherent race with invalidate_inode_pages and
1018  *        writebacks since the page->count is kept > 1 for as long
1019  *        as the page has a write request pending.
1020  */
1021 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1022 {
1023         struct nfs_write_data   *data = calldata;
1024         struct nfs_page         *req;
1025         struct page             *page;
1026
1027         if (nfs_writeback_done(task, data) != 0)
1028                 return;
1029
1030         /* Update attributes as result of writeback. */
1031         while (!list_empty(&data->pages)) {
1032                 req = nfs_list_entry(data->pages.next);
1033                 nfs_list_remove_request(req);
1034                 page = req->wb_page;
1035
1036                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1037                         req->wb_context->dentry->d_inode->i_sb->s_id,
1038                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1039                         req->wb_bytes,
1040                         (long long)req_offset(req));
1041
1042                 if (task->tk_status < 0) {
1043                         ClearPageUptodate(page);
1044                         SetPageError(page);
1045                         req->wb_context->error = task->tk_status;
1046                         end_page_writeback(page);
1047                         nfs_inode_remove_request(req);
1048                         dprintk(", error = %d\n", task->tk_status);
1049                         goto next;
1050                 }
1051                 end_page_writeback(page);
1052
1053 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1054                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1055                         nfs_inode_remove_request(req);
1056                         dprintk(" OK\n");
1057                         goto next;
1058                 }
1059                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1060                 nfs_mark_request_commit(req);
1061                 dprintk(" marked for commit\n");
1062 #else
1063                 nfs_inode_remove_request(req);
1064 #endif
1065         next:
1066                 nfs_clear_page_writeback(req);
1067         }
1068 }
1069
1070 static const struct rpc_call_ops nfs_write_full_ops = {
1071         .rpc_call_done = nfs_writeback_done_full,
1072         .rpc_release = nfs_writedata_release,
1073 };
1074
1075
1076 /*
1077  * This function is called when the WRITE call is complete.
1078  */
1079 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1080 {
1081         struct nfs_writeargs    *argp = &data->args;
1082         struct nfs_writeres     *resp = &data->res;
1083         int status;
1084
1085         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1086                 task->tk_pid, task->tk_status);
1087
1088         /*
1089          * ->write_done will attempt to use post-op attributes to detect
1090          * conflicting writes by other clients.  A strict interpretation
1091          * of close-to-open would allow us to continue caching even if
1092          * another writer had changed the file, but some applications
1093          * depend on tighter cache coherency when writing.
1094          */
1095         status = NFS_PROTO(data->inode)->write_done(task, data);
1096         if (status != 0)
1097                 return status;
1098         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1099
1100 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1101         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1102                 /* We tried a write call, but the server did not
1103                  * commit data to stable storage even though we
1104                  * requested it.
1105                  * Note: There is a known bug in Tru64 < 5.0 in which
1106                  *       the server reports NFS_DATA_SYNC, but performs
1107                  *       NFS_FILE_SYNC. We therefore implement this checking
1108                  *       as a dprintk() in order to avoid filling syslog.
1109                  */
1110                 static unsigned long    complain;
1111
1112                 if (time_before(complain, jiffies)) {
1113                         dprintk("NFS: faulty NFS server %s:"
1114                                 " (committed = %d) != (stable = %d)\n",
1115                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1116                                 resp->verf->committed, argp->stable);
1117                         complain = jiffies + 300 * HZ;
1118                 }
1119         }
1120 #endif
1121         /* Is this a short write? */
1122         if (task->tk_status >= 0 && resp->count < argp->count) {
1123                 static unsigned long    complain;
1124
1125                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1126
1127                 /* Has the server at least made some progress? */
1128                 if (resp->count != 0) {
1129                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1130                         if (resp->verf->committed != NFS_UNSTABLE) {
1131                                 /* Resend from where the server left off */
1132                                 argp->offset += resp->count;
1133                                 argp->pgbase += resp->count;
1134                                 argp->count -= resp->count;
1135                         } else {
1136                                 /* Resend as a stable write in order to avoid
1137                                  * headaches in the case of a server crash.
1138                                  */
1139                                 argp->stable = NFS_FILE_SYNC;
1140                         }
1141                         rpc_restart_call(task);
1142                         return -EAGAIN;
1143                 }
1144                 if (time_before(complain, jiffies)) {
1145                         printk(KERN_WARNING
1146                                "NFS: Server wrote zero bytes, expected %u.\n",
1147                                         argp->count);
1148                         complain = jiffies + 300 * HZ;
1149                 }
1150                 /* Can't do anything about it except throw an error. */
1151                 task->tk_status = -EIO;
1152         }
1153         return 0;
1154 }
1155
1156
1157 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1158 void nfs_commit_release(void *wdata)
1159 {
1160         nfs_commit_free(wdata);
1161 }
1162
1163 /*
1164  * Set up the argument/result storage required for the RPC call.
1165  */
1166 static void nfs_commit_rpcsetup(struct list_head *head,
1167                 struct nfs_write_data *data,
1168                 int how)
1169 {
1170         struct nfs_page         *first;
1171         struct inode            *inode;
1172         int flags;
1173
1174         /* Set up the RPC argument and reply structs
1175          * NB: take care not to mess about with data->commit et al. */
1176
1177         list_splice_init(head, &data->pages);
1178         first = nfs_list_entry(data->pages.next);
1179         inode = first->wb_context->dentry->d_inode;
1180
1181         data->inode       = inode;
1182         data->cred        = first->wb_context->cred;
1183
1184         data->args.fh     = NFS_FH(data->inode);
1185         /* Note: we always request a commit of the entire inode */
1186         data->args.offset = 0;
1187         data->args.count  = 0;
1188         data->res.count   = 0;
1189         data->res.fattr   = &data->fattr;
1190         data->res.verf    = &data->verf;
1191         nfs_fattr_init(&data->fattr);
1192
1193         /* Set up the initial task struct.  */
1194         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1195         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1196         NFS_PROTO(inode)->commit_setup(data, how);
1197
1198         data->task.tk_priority = flush_task_priority(how);
1199         data->task.tk_cookie = (unsigned long)inode;
1200
1201         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1202 }
1203
1204 /*
1205  * Commit dirty pages
1206  */
1207 static int
1208 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1209 {
1210         struct nfs_write_data   *data;
1211         struct nfs_page         *req;
1212
1213         data = nfs_commit_alloc();
1214
1215         if (!data)
1216                 goto out_bad;
1217
1218         /* Set up the argument struct */
1219         nfs_commit_rpcsetup(head, data, how);
1220
1221         nfs_execute_write(data);
1222         return 0;
1223  out_bad:
1224         while (!list_empty(head)) {
1225                 req = nfs_list_entry(head->next);
1226                 nfs_list_remove_request(req);
1227                 nfs_mark_request_commit(req);
1228                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1229                 nfs_clear_page_writeback(req);
1230         }
1231         return -ENOMEM;
1232 }
1233
1234 /*
1235  * COMMIT call returned
1236  */
1237 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1238 {
1239         struct nfs_write_data   *data = calldata;
1240         struct nfs_page         *req;
1241
1242         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1243                                 task->tk_pid, task->tk_status);
1244
1245         /* Call the NFS version-specific code */
1246         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1247                 return;
1248
1249         while (!list_empty(&data->pages)) {
1250                 req = nfs_list_entry(data->pages.next);
1251                 nfs_list_remove_request(req);
1252                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1253
1254                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1255                         req->wb_context->dentry->d_inode->i_sb->s_id,
1256                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1257                         req->wb_bytes,
1258                         (long long)req_offset(req));
1259                 if (task->tk_status < 0) {
1260                         req->wb_context->error = task->tk_status;
1261                         nfs_inode_remove_request(req);
1262                         dprintk(", error = %d\n", task->tk_status);
1263                         goto next;
1264                 }
1265
1266                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1267                  * returned by the server against all stored verfs. */
1268                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1269                         /* We have a match */
1270                         nfs_inode_remove_request(req);
1271                         dprintk(" OK\n");
1272                         goto next;
1273                 }
1274                 /* We have a mismatch. Write the page again */
1275                 dprintk(" mismatch\n");
1276                 nfs_mark_request_dirty(req);
1277         next:
1278                 nfs_clear_page_writeback(req);
1279         }
1280 }
1281
1282 static const struct rpc_call_ops nfs_commit_ops = {
1283         .rpc_call_done = nfs_commit_done,
1284         .rpc_release = nfs_commit_release,
1285 };
1286 #else
1287 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1288 {
1289         return 0;
1290 }
1291 #endif
1292
1293 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1294 {
1295         struct nfs_inode *nfsi = NFS_I(mapping->host);
1296         LIST_HEAD(head);
1297         long res;
1298
1299         spin_lock(&nfsi->req_lock);
1300         res = nfs_scan_dirty(mapping, wbc, &head);
1301         spin_unlock(&nfsi->req_lock);
1302         if (res) {
1303                 int error = nfs_flush_list(mapping->host, &head, res, how);
1304                 if (error < 0)
1305                         return error;
1306         }
1307         return res;
1308 }
1309
1310 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1311 int nfs_commit_inode(struct inode *inode, int how)
1312 {
1313         struct nfs_inode *nfsi = NFS_I(inode);
1314         LIST_HEAD(head);
1315         int res;
1316
1317         spin_lock(&nfsi->req_lock);
1318         res = nfs_scan_commit(inode, &head, 0, 0);
1319         spin_unlock(&nfsi->req_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 #endif
1328
1329 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1330 {
1331         struct inode *inode = mapping->host;
1332         struct nfs_inode *nfsi = NFS_I(inode);
1333         unsigned long idx_start, idx_end;
1334         unsigned int npages = 0;
1335         LIST_HEAD(head);
1336         int nocommit = how & FLUSH_NOCOMMIT;
1337         long pages, ret;
1338
1339         /* FIXME */
1340         if (wbc->range_cyclic)
1341                 idx_start = 0;
1342         else {
1343                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1344                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1345                 if (idx_end > idx_start) {
1346                         unsigned long l_npages = 1 + idx_end - idx_start;
1347                         npages = l_npages;
1348                         if (sizeof(npages) != sizeof(l_npages) &&
1349                                         (unsigned long)npages != l_npages)
1350                                 npages = 0;
1351                 }
1352         }
1353         how &= ~FLUSH_NOCOMMIT;
1354         spin_lock(&nfsi->req_lock);
1355         do {
1356                 wbc->pages_skipped = 0;
1357                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1358                 if (ret != 0)
1359                         continue;
1360                 pages = nfs_scan_dirty(mapping, wbc, &head);
1361                 if (pages != 0) {
1362                         spin_unlock(&nfsi->req_lock);
1363                         if (how & FLUSH_INVALIDATE) {
1364                                 nfs_cancel_dirty_list(&head);
1365                                 ret = pages;
1366                         } else
1367                                 ret = nfs_flush_list(inode, &head, pages, how);
1368                         spin_lock(&nfsi->req_lock);
1369                         continue;
1370                 }
1371                 if (wbc->pages_skipped != 0)
1372                         continue;
1373                 if (nocommit)
1374                         break;
1375                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1376                 if (pages == 0) {
1377                         if (wbc->pages_skipped != 0)
1378                                 continue;
1379                         break;
1380                 }
1381                 if (how & FLUSH_INVALIDATE) {
1382                         spin_unlock(&nfsi->req_lock);
1383                         nfs_cancel_commit_list(&head);
1384                         ret = pages;
1385                         spin_lock(&nfsi->req_lock);
1386                         continue;
1387                 }
1388                 pages += nfs_scan_commit(inode, &head, 0, 0);
1389                 spin_unlock(&nfsi->req_lock);
1390                 ret = nfs_commit_list(inode, &head, how);
1391                 spin_lock(&nfsi->req_lock);
1392         } while (ret >= 0);
1393         spin_unlock(&nfsi->req_lock);
1394         return ret;
1395 }
1396
1397 /*
1398  * flush the inode to disk.
1399  */
1400 int nfs_wb_all(struct inode *inode)
1401 {
1402         struct address_space *mapping = inode->i_mapping;
1403         struct writeback_control wbc = {
1404                 .bdi = mapping->backing_dev_info,
1405                 .sync_mode = WB_SYNC_ALL,
1406                 .nr_to_write = LONG_MAX,
1407                 .range_cyclic = 1,
1408         };
1409         int ret;
1410
1411         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1412         if (ret >= 0)
1413                 return 0;
1414         return ret;
1415 }
1416
1417 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1418 {
1419         struct writeback_control wbc = {
1420                 .bdi = mapping->backing_dev_info,
1421                 .sync_mode = WB_SYNC_ALL,
1422                 .nr_to_write = LONG_MAX,
1423                 .range_start = range_start,
1424                 .range_end = range_end,
1425         };
1426         int ret;
1427
1428         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1429         if (ret >= 0)
1430                 return 0;
1431         return ret;
1432 }
1433
1434 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1435 {
1436         loff_t range_start = page_offset(page);
1437         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1438
1439         return nfs_sync_mapping_range(inode->i_mapping, range_start, range_end, how | FLUSH_STABLE);
1440 }
1441
1442 /*
1443  * Write back all requests on one page - we do this before reading it.
1444  */
1445 int nfs_wb_page(struct inode *inode, struct page* page)
1446 {
1447         return nfs_wb_page_priority(inode, page, 0);
1448 }
1449
1450 int nfs_set_page_dirty(struct page *page)
1451 {
1452         struct nfs_page *req;
1453
1454         req = nfs_page_find_request(page);
1455         if (req != NULL) {
1456                 /* Mark any existing write requests for flushing */
1457                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1458                 nfs_release_request(req);
1459         }
1460         return __set_page_dirty_nobuffers(page);
1461 }
1462
1463
1464 int __init nfs_init_writepagecache(void)
1465 {
1466         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1467                                              sizeof(struct nfs_write_data),
1468                                              0, SLAB_HWCACHE_ALIGN,
1469                                              NULL, NULL);
1470         if (nfs_wdata_cachep == NULL)
1471                 return -ENOMEM;
1472
1473         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1474                                                      nfs_wdata_cachep);
1475         if (nfs_wdata_mempool == NULL)
1476                 return -ENOMEM;
1477
1478         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1479                                                       nfs_wdata_cachep);
1480         if (nfs_commit_mempool == NULL)
1481                 return -ENOMEM;
1482
1483         return 0;
1484 }
1485
1486 void nfs_destroy_writepagecache(void)
1487 {
1488         mempool_destroy(nfs_commit_mempool);
1489         mempool_destroy(nfs_wdata_mempool);
1490         kmem_cache_destroy(nfs_wdata_cachep);
1491 }
1492