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 "iostat.h"
  67
  68 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  69
  70 #define MIN_POOL_WRITE          (32)
  71 #define MIN_POOL_COMMIT         (4)
  72
  73 /*
  74  * Local function declarations
  75  */
  76 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  77                                             struct inode *,
  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 const struct rpc_call_ops nfs_write_partial_ops;
  84 static const struct rpc_call_ops nfs_write_full_ops;
  85 static const struct rpc_call_ops nfs_commit_ops;
  86
  87 static kmem_cache_t *nfs_wdata_cachep;
  88 static mempool_t *nfs_wdata_mempool;
  89 static mempool_t *nfs_commit_mempool;
  90
  91 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  92
  93 struct nfs_write_data *nfs_commit_alloc(void)
  94 {
  95         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
  96
  97         if (p) {
  98                 memset(p, 0, sizeof(*p));
  99                 INIT_LIST_HEAD(&p->pages);
 100         }
 101         return p;
 102 }
 103
 104 void nfs_commit_rcu_free(struct rcu_head *head)
 105 {
 106         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 107         if (p && (p->pagevec != &p->page_array[0]))
 108                 kfree(p->pagevec);
 109         mempool_free(p, nfs_commit_mempool);
 110 }
 111
 112 void nfs_commit_free(struct nfs_write_data *wdata)
 113 {
 114         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
 115 }
 116
 117 struct nfs_write_data *nfs_writedata_alloc(size_t len)
 118 {
 119         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 120         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
 121
 122         if (p) {
 123                 memset(p, 0, sizeof(*p));
 124                 INIT_LIST_HEAD(&p->pages);
 125                 p->npages = pagecount;
 126                 if (pagecount <= ARRAY_SIZE(p->page_array))
 127                         p->pagevec = p->page_array;
 128                 else {
 129                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 130                         if (!p->pagevec) {
 131                                 mempool_free(p, nfs_wdata_mempool);
 132                                 p = NULL;
 133                         }
 134                 }
 135         }
 136         return p;
 137 }
 138
 139 static void nfs_writedata_rcu_free(struct rcu_head *head)
 140 {
 141         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 142         if (p && (p->pagevec != &p->page_array[0]))
 143                 kfree(p->pagevec);
 144         mempool_free(p, nfs_wdata_mempool);
 145 }
 146
 147 static void nfs_writedata_free(struct nfs_write_data *wdata)
 148 {
 149         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 150 }
 151
 152 void nfs_writedata_release(void *wdata)
 153 {
 154         nfs_writedata_free(wdata);
 155 }
 156
 157 /* Adjust the file length if we're writing beyond the end */
 158 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 159 {
 160         struct inode *inode = page->mapping->host;
 161         loff_t end, i_size = i_size_read(inode);
 162         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 163
 164         if (i_size > 0 && page->index < end_index)
 165                 return;
 166         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 167         if (i_size >= end)
 168                 return;
 169         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 170         i_size_write(inode, end);
 171 }
 172
 173 /* We can set the PG_uptodate flag if we see that a write request
 174  * covers the full page.
 175  */
 176 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 177 {
 178         loff_t end_offs;
 179
 180         if (PageUptodate(page))
 181                 return;
 182         if (base != 0)
 183                 return;
 184         if (count == PAGE_CACHE_SIZE) {
 185                 SetPageUptodate(page);
 186                 return;
 187         }
 188
 189         end_offs = i_size_read(page->mapping->host) - 1;
 190         if (end_offs < 0)
 191                 return;
 192         /* Is this the last page? */
 193         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
 194                 return;
 195         /* This is the last page: set PG_uptodate if we cover the entire
 196          * extent of the data, then zero the rest of the page.
 197          */
 198         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
 199                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 200                 SetPageUptodate(page);
 201         }
 202 }
 203
 204 /*
 205  * Write a page synchronously.
 206  * Offset is the data offset within the page.
 207  */
 208 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 209                 struct page *page, unsigned int offset, unsigned int count,
 210                 int how)
 211 {
 212         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 213         int             result, written = 0;
 214         struct nfs_write_data *wdata;
 215
 216         wdata = nfs_writedata_alloc(wsize);
 217         if (!wdata)
 218                 return -ENOMEM;
 219
 220         wdata->flags = how;
 221         wdata->cred = ctx->cred;
 222         wdata->inode = inode;
 223         wdata->args.fh = NFS_FH(inode);
 224         wdata->args.context = ctx;
 225         wdata->args.pages = &page;
 226         wdata->args.stable = NFS_FILE_SYNC;
 227         wdata->args.pgbase = offset;
 228         wdata->args.count = wsize;
 229         wdata->res.fattr = &wdata->fattr;
 230         wdata->res.verf = &wdata->verf;
 231
 232         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 233                 inode->i_sb->s_id,
 234                 (long long)NFS_FILEID(inode),
 235                 count, (long long)(page_offset(page) + offset));
 236
 237         set_page_writeback(page);
 238         nfs_begin_data_update(inode);
 239         do {
 240                 if (count < wsize)
 241                         wdata->args.count = count;
 242                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 243
 244                 result = NFS_PROTO(inode)->write(wdata);
 245
 246                 if (result < 0) {
 247                         /* Must mark the page invalid after I/O error */
 248                         ClearPageUptodate(page);
 249                         goto io_error;
 250                 }
 251                 if (result < wdata->args.count)
 252                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 253                                         wdata->args.count, result);
 254
 255                 wdata->args.offset += result;
 256                 wdata->args.pgbase += result;
 257                 written += result;
 258                 count -= result;
 259                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 260         } while (count);
 261         /* Update file length */
 262         nfs_grow_file(page, offset, written);
 263         /* Set the PG_uptodate flag? */
 264         nfs_mark_uptodate(page, offset, written);
 265
 266         if (PageError(page))
 267                 ClearPageError(page);
 268
 269 io_error:
 270         nfs_end_data_update(inode);
 271         end_page_writeback(page);
 272         nfs_writedata_release(wdata);
 273         return written ? written : result;
 274 }
 275
 276 static int nfs_writepage_async(struct nfs_open_context *ctx,
 277                 struct inode *inode, struct page *page,
 278                 unsigned int offset, unsigned int count)
 279 {
 280         struct nfs_page *req;
 281
 282         req = nfs_update_request(ctx, inode, page, offset, count);
 283         if (IS_ERR(req))
 284                 return PTR_ERR(req);
 285         /* Update file length */
 286         nfs_grow_file(page, offset, count);
 287         /* Set the PG_uptodate flag? */
 288         nfs_mark_uptodate(page, offset, count);
 289         nfs_unlock_request(req);
 290         return 0;
 291 }
 292
 293 static int wb_priority(struct writeback_control *wbc)
 294 {
 295         if (wbc->for_reclaim)
 296                 return FLUSH_HIGHPRI;
 297         if (wbc->for_kupdate)
 298                 return FLUSH_LOWPRI;
 299         return 0;
 300 }
 301
 302 /*
 303  * Write an mmapped page to the server.
 304  */
 305 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 306 {
 307         struct nfs_open_context *ctx;
 308         struct inode *inode = page->mapping->host;
 309         unsigned long end_index;
 310         unsigned offset = PAGE_CACHE_SIZE;
 311         loff_t i_size = i_size_read(inode);
 312         int inode_referenced = 0;
 313         int priority = wb_priority(wbc);
 314         int err;
 315
 316         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 317         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 318
 319         /*
 320          * Note: We need to ensure that we have a reference to the inode
 321          *       if we are to do asynchronous writes. If not, waiting
 322          *       in nfs_wait_on_request() may deadlock with clear_inode().
 323          *
 324          *       If igrab() fails here, then it is in any case safe to
 325          *       call nfs_wb_page(), since there will be no pending writes.
 326          */
 327         if (igrab(inode) != 0)
 328                 inode_referenced = 1;
 329         end_index = i_size >> PAGE_CACHE_SHIFT;
 330
 331         /* Ensure we've flushed out any previous writes */
 332         nfs_wb_page_priority(inode, page, priority);
 333
 334         /* easy case */
 335         if (page->index < end_index)
 336                 goto do_it;
 337         /* things got complicated... */
 338         offset = i_size & (PAGE_CACHE_SIZE-1);
 339
 340         /* OK, are we completely out? */
 341         err = 0; /* potential race with truncate - ignore */
 342         if (page->index >= end_index+1 || !offset)
 343                 goto out;
 344 do_it:
 345         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 346         if (ctx == NULL) {
 347                 err = -EBADF;
 348                 goto out;
 349         }
 350         lock_kernel();
 351         if (!IS_SYNC(inode) && inode_referenced) {
 352                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
 353                 if (!wbc->for_writepages)
 354                         nfs_flush_mapping(page->mapping, wbc, wb_priority(wbc));
 355         } else {
 356                 err = nfs_writepage_sync(ctx, inode, page, 0,
 357                                                 offset, priority);
 358                 if (err >= 0) {
 359                         if (err != offset)
 360                                 redirty_page_for_writepage(wbc, page);
 361                         err = 0;
 362                 }
 363         }
 364         unlock_kernel();
 365         put_nfs_open_context(ctx);
 366 out:
 367         unlock_page(page);
 368         if (inode_referenced)
 369                 iput(inode);
 370         return err;
 371 }
 372
 373 /*
 374  * Note: causes nfs_update_request() to block on the assumption
 375  *       that the writeback is generated due to memory pressure.
 376  */
 377 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 378 {
 379         struct backing_dev_info *bdi = mapping->backing_dev_info;
 380         struct inode *inode = mapping->host;
 381         int err;
 382
 383         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 384
 385         err = generic_writepages(mapping, wbc);
 386         if (err)
 387                 return err;
 388         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 389                 if (wbc->nonblocking)
 390                         return 0;
 391                 nfs_wait_on_write_congestion(mapping, 0);
 392         }
 393         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 394         if (err < 0)
 395                 goto out;
 396         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 397         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 398                 err = nfs_wait_on_requests(inode, 0, 0);
 399                 if (err < 0)
 400                         goto out;
 401         }
 402         err = nfs_commit_inode(inode, wb_priority(wbc));
 403         if (err > 0)
 404                 err = 0;
 405 out:
 406         clear_bit(BDI_write_congested, &bdi->state);
 407         wake_up_all(&nfs_write_congestion);
 408         congestion_end(WRITE);
 409         return err;
 410 }
 411
 412 /*
 413  * Insert a write request into an inode
 414  */
 415 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 416 {
 417         struct nfs_inode *nfsi = NFS_I(inode);
 418         int error;
 419
 420         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 421         BUG_ON(error == -EEXIST);
 422         if (error)
 423                 return error;
 424         if (!nfsi->npages) {
 425                 igrab(inode);
 426                 nfs_begin_data_update(inode);
 427                 if (nfs_have_delegation(inode, FMODE_WRITE))
 428                         nfsi->change_attr++;
 429         }
 430         SetPagePrivate(req->wb_page);
 431         nfsi->npages++;
 432         atomic_inc(&req->wb_count);
 433         return 0;
 434 }
 435
 436 /*
 437  * Insert a write request into an inode
 438  */
 439 static void nfs_inode_remove_request(struct nfs_page *req)
 440 {
 441         struct inode *inode = req->wb_context->dentry->d_inode;
 442         struct nfs_inode *nfsi = NFS_I(inode);
 443
 444         BUG_ON (!NFS_WBACK_BUSY(req));
 445
 446         spin_lock(&nfsi->req_lock);
 447         ClearPagePrivate(req->wb_page);
 448         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 449         nfsi->npages--;
 450         if (!nfsi->npages) {
 451                 spin_unlock(&nfsi->req_lock);
 452                 nfs_end_data_update(inode);
 453                 iput(inode);
 454         } else
 455                 spin_unlock(&nfsi->req_lock);
 456         nfs_clear_request(req);
 457         nfs_release_request(req);
 458 }
 459
 460 /*
 461  * Find a request
 462  */
 463 static inline struct nfs_page *
 464 _nfs_find_request(struct inode *inode, unsigned long index)
 465 {
 466         struct nfs_inode *nfsi = NFS_I(inode);
 467         struct nfs_page *req;
 468
 469         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 470         if (req)
 471                 atomic_inc(&req->wb_count);
 472         return req;
 473 }
 474
 475 static struct nfs_page *
 476 nfs_find_request(struct inode *inode, unsigned long index)
 477 {
 478         struct nfs_page         *req;
 479         struct nfs_inode        *nfsi = NFS_I(inode);
 480
 481         spin_lock(&nfsi->req_lock);
 482         req = _nfs_find_request(inode, index);
 483         spin_unlock(&nfsi->req_lock);
 484         return req;
 485 }
 486
 487 /*
 488  * Add a request to the inode's dirty list.
 489  */
 490 static void
 491 nfs_mark_request_dirty(struct nfs_page *req)
 492 {
 493         struct inode *inode = req->wb_context->dentry->d_inode;
 494         struct nfs_inode *nfsi = NFS_I(inode);
 495
 496         spin_lock(&nfsi->req_lock);
 497         radix_tree_tag_set(&nfsi->nfs_page_tree,
 498                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 499         nfs_list_add_request(req, &nfsi->dirty);
 500         nfsi->ndirty++;
 501         spin_unlock(&nfsi->req_lock);
 502         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 503         mark_inode_dirty(inode);
 504 }
 505
 506 /*
 507  * Check if a request is dirty
 508  */
 509 static inline int
 510 nfs_dirty_request(struct nfs_page *req)
 511 {
 512         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 513         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 514 }
 515
 516 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 517 /*
 518  * Add a request to the inode's commit list.
 519  */
 520 static void
 521 nfs_mark_request_commit(struct nfs_page *req)
 522 {
 523         struct inode *inode = req->wb_context->dentry->d_inode;
 524         struct nfs_inode *nfsi = NFS_I(inode);
 525
 526         spin_lock(&nfsi->req_lock);
 527         nfs_list_add_request(req, &nfsi->commit);
 528         nfsi->ncommit++;
 529         spin_unlock(&nfsi->req_lock);
 530         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 531         mark_inode_dirty(inode);
 532 }
 533 #endif
 534
 535 /*
 536  * Wait for a request to complete.
 537  *
 538  * Interruptible by signals only if mounted with intr flag.
 539  */
 540 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 541 {
 542         struct nfs_inode *nfsi = NFS_I(inode);
 543         struct nfs_page *req;
 544         unsigned long           idx_end, next;
 545         unsigned int            res = 0;
 546         int                     error;
 547
 548         if (npages == 0)
 549                 idx_end = ~0;
 550         else
 551                 idx_end = idx_start + npages - 1;
 552
 553         next = idx_start;
 554         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 555                 if (req->wb_index > idx_end)
 556                         break;
 557
 558                 next = req->wb_index + 1;
 559                 BUG_ON(!NFS_WBACK_BUSY(req));
 560
 561                 atomic_inc(&req->wb_count);
 562                 spin_unlock(&nfsi->req_lock);
 563                 error = nfs_wait_on_request(req);
 564                 nfs_release_request(req);
 565                 spin_lock(&nfsi->req_lock);
 566                 if (error < 0)
 567                         return error;
 568                 res++;
 569         }
 570         return res;
 571 }
 572
 573 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 574 {
 575         struct nfs_inode *nfsi = NFS_I(inode);
 576         int ret;
 577
 578         spin_lock(&nfsi->req_lock);
 579         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 580         spin_unlock(&nfsi->req_lock);
 581         return ret;
 582 }
 583
 584 static void nfs_cancel_dirty_list(struct list_head *head)
 585 {
 586         struct nfs_page *req;
 587         while(!list_empty(head)) {
 588                 req = nfs_list_entry(head->next);
 589                 nfs_list_remove_request(req);
 590                 nfs_inode_remove_request(req);
 591                 nfs_clear_page_writeback(req);
 592         }
 593 }
 594
 595 static void nfs_cancel_commit_list(struct list_head *head)
 596 {
 597         struct nfs_page *req;
 598
 599         while(!list_empty(head)) {
 600                 req = nfs_list_entry(head->next);
 601                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 602                 nfs_list_remove_request(req);
 603                 nfs_inode_remove_request(req);
 604                 nfs_unlock_request(req);
 605         }
 606 }
 607
 608 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 609 /*
 610  * nfs_scan_commit - Scan an inode for commit requests
 611  * @inode: NFS inode to scan
 612  * @dst: destination list
 613  * @idx_start: lower bound of page->index to scan.
 614  * @npages: idx_start + npages sets the upper bound to scan.
 615  *
 616  * Moves requests from the inode's 'commit' request list.
 617  * The requests are *not* checked to ensure that they form a contiguous set.
 618  */
 619 static int
 620 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 621 {
 622         struct nfs_inode *nfsi = NFS_I(inode);
 623         int res = 0;
 624
 625         if (nfsi->ncommit != 0) {
 626                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 627                 nfsi->ncommit -= res;
 628                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 629                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 630         }
 631         return res;
 632 }
 633 #else
 634 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 635 {
 636         return 0;
 637 }
 638 #endif
 639
 640 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 641 {
 642         struct backing_dev_info *bdi = mapping->backing_dev_info;
 643         DEFINE_WAIT(wait);
 644         int ret = 0;
 645
 646         might_sleep();
 647
 648         if (!bdi_write_congested(bdi))
 649                 return 0;
 650
 651         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 652
 653         if (intr) {
 654                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 655                 sigset_t oldset;
 656
 657                 rpc_clnt_sigmask(clnt, &oldset);
 658                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 659                 if (bdi_write_congested(bdi)) {
 660                         if (signalled())
 661                                 ret = -ERESTARTSYS;
 662                         else
 663                                 schedule();
 664                 }
 665                 rpc_clnt_sigunmask(clnt, &oldset);
 666         } else {
 667                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 668                 if (bdi_write_congested(bdi))
 669                         schedule();
 670         }
 671         finish_wait(&nfs_write_congestion, &wait);
 672         return ret;
 673 }
 674
 675
 676 /*
 677  * Try to update any existing write request, or create one if there is none.
 678  * In order to match, the request's credentials must match those of
 679  * the calling process.
 680  *
 681  * Note: Should always be called with the Page Lock held!
 682  */
 683 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 684                 struct inode *inode, struct page *page,
 685                 unsigned int offset, unsigned int bytes)
 686 {
 687         struct nfs_server *server = NFS_SERVER(inode);
 688         struct nfs_inode *nfsi = NFS_I(inode);
 689         struct nfs_page         *req, *new = NULL;
 690         unsigned long           rqend, end;
 691
 692         end = offset + bytes;
 693
 694         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 695                 return ERR_PTR(-ERESTARTSYS);
 696         for (;;) {
 697                 /* Loop over all inode entries and see if we find
 698                  * A request for the page we wish to update
 699                  */
 700                 spin_lock(&nfsi->req_lock);
 701                 req = _nfs_find_request(inode, page->index);
 702                 if (req) {
 703                         if (!nfs_lock_request_dontget(req)) {
 704                                 int error;
 705                                 spin_unlock(&nfsi->req_lock);
 706                                 error = nfs_wait_on_request(req);
 707                                 nfs_release_request(req);
 708                                 if (error < 0) {
 709                                         if (new)
 710                                                 nfs_release_request(new);
 711                                         return ERR_PTR(error);
 712                                 }
 713                                 continue;
 714                         }
 715                         spin_unlock(&nfsi->req_lock);
 716                         if (new)
 717                                 nfs_release_request(new);
 718                         break;
 719                 }
 720
 721                 if (new) {
 722                         int error;
 723                         nfs_lock_request_dontget(new);
 724                         error = nfs_inode_add_request(inode, new);
 725                         if (error) {
 726                                 spin_unlock(&nfsi->req_lock);
 727                                 nfs_unlock_request(new);
 728                                 return ERR_PTR(error);
 729                         }
 730                         spin_unlock(&nfsi->req_lock);
 731                         nfs_mark_request_dirty(new);
 732                         return new;
 733                 }
 734                 spin_unlock(&nfsi->req_lock);
 735
 736                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 737                 if (IS_ERR(new))
 738                         return new;
 739         }
 740
 741         /* We have a request for our page.
 742          * If the creds don't match, or the
 743          * page addresses don't match,
 744          * tell the caller to wait on the conflicting
 745          * request.
 746          */
 747         rqend = req->wb_offset + req->wb_bytes;
 748         if (req->wb_context != ctx
 749             || req->wb_page != page
 750             || !nfs_dirty_request(req)
 751             || offset > rqend || end < req->wb_offset) {
 752                 nfs_unlock_request(req);
 753                 return ERR_PTR(-EBUSY);
 754         }
 755
 756         /* Okay, the request matches. Update the region */
 757         if (offset < req->wb_offset) {
 758                 req->wb_offset = offset;
 759                 req->wb_pgbase = offset;
 760                 req->wb_bytes = rqend - req->wb_offset;
 761         }
 762
 763         if (end > rqend)
 764                 req->wb_bytes = end - req->wb_offset;
 765
 766         return req;
 767 }
 768
 769 int nfs_flush_incompatible(struct file *file, struct page *page)
 770 {
 771         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 772         struct inode    *inode = page->mapping->host;
 773         struct nfs_page *req;
 774         int             status = 0;
 775         /*
 776          * Look for a request corresponding to this page. If there
 777          * is one, and it belongs to another file, we flush it out
 778          * before we try to copy anything into the page. Do this
 779          * due to the lack of an ACCESS-type call in NFSv2.
 780          * Also do the same if we find a request from an existing
 781          * dropped page.
 782          */
 783         req = nfs_find_request(inode, page->index);
 784         if (req) {
 785                 if (req->wb_page != page || ctx != req->wb_context)
 786                         status = nfs_wb_page(inode, page);
 787                 nfs_release_request(req);
 788         }
 789         return (status < 0) ? status : 0;
 790 }
 791
 792 /*
 793  * Update and possibly write a cached page of an NFS file.
 794  *
 795  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 796  * things with a page scheduled for an RPC call (e.g. invalidate it).
 797  */
 798 int nfs_updatepage(struct file *file, struct page *page,
 799                 unsigned int offset, unsigned int count)
 800 {
 801         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 802         struct inode    *inode = page->mapping->host;
 803         struct nfs_page *req;
 804         int             status = 0;
 805
 806         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 807
 808         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 809                 file->f_dentry->d_parent->d_name.name,
 810                 file->f_dentry->d_name.name, count,
 811                 (long long)(page_offset(page) +offset));
 812
 813         if (IS_SYNC(inode)) {
 814                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 815                 if (status > 0) {
 816                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 817                                 SetPageUptodate(page);
 818                         return 0;
 819                 }
 820                 return status;
 821         }
 822
 823         /* If we're not using byte range locks, and we know the page
 824          * is entirely in cache, it may be more efficient to avoid
 825          * fragmenting write requests.
 826          */
 827         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 828                 loff_t end_offs = i_size_read(inode) - 1;
 829                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 830
 831                 count += offset;
 832                 offset = 0;
 833                 if (unlikely(end_offs < 0)) {
 834                         /* Do nothing */
 835                 } else if (page->index == end_index) {
 836                         unsigned int pglen;
 837                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 838                         if (count < pglen)
 839                                 count = pglen;
 840                 } else if (page->index < end_index)
 841                         count = PAGE_CACHE_SIZE;
 842         }
 843
 844         /*
 845          * Try to find an NFS request corresponding to this page
 846          * and update it.
 847          * If the existing request cannot be updated, we must flush
 848          * it out now.
 849          */
 850         do {
 851                 req = nfs_update_request(ctx, inode, page, offset, count);
 852                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 853                 if (status != -EBUSY)
 854                         break;
 855                 /* Request could not be updated. Flush it out and try again */
 856                 status = nfs_wb_page(inode, page);
 857         } while (status >= 0);
 858         if (status < 0)
 859                 goto done;
 860
 861         status = 0;
 862
 863         /* Update file length */
 864         nfs_grow_file(page, offset, count);
 865         /* Set the PG_uptodate flag? */
 866         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 867         nfs_unlock_request(req);
 868 done:
 869         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 870                         status, (long long)i_size_read(inode));
 871         if (status < 0)
 872                 ClearPageUptodate(page);
 873         return status;
 874 }
 875
 876 static void nfs_writepage_release(struct nfs_page *req)
 877 {
 878         end_page_writeback(req->wb_page);
 879
 880 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 881         if (!PageError(req->wb_page)) {
 882                 if (NFS_NEED_RESCHED(req)) {
 883                         nfs_mark_request_dirty(req);
 884                         goto out;
 885                 } else if (NFS_NEED_COMMIT(req)) {
 886                         nfs_mark_request_commit(req);
 887                         goto out;
 888                 }
 889         }
 890         nfs_inode_remove_request(req);
 891
 892 out:
 893         nfs_clear_commit(req);
 894         nfs_clear_reschedule(req);
 895 #else
 896         nfs_inode_remove_request(req);
 897 #endif
 898         nfs_clear_page_writeback(req);
 899 }
 900
 901 static inline int flush_task_priority(int how)
 902 {
 903         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 904                 case FLUSH_HIGHPRI:
 905                         return RPC_PRIORITY_HIGH;
 906                 case FLUSH_LOWPRI:
 907                         return RPC_PRIORITY_LOW;
 908         }
 909         return RPC_PRIORITY_NORMAL;
 910 }
 911
 912 /*
 913  * Set up the argument/result storage required for the RPC call.
 914  */
 915 static void nfs_write_rpcsetup(struct nfs_page *req,
 916                 struct nfs_write_data *data,
 917                 const struct rpc_call_ops *call_ops,
 918                 unsigned int count, unsigned int offset,
 919                 int how)
 920 {
 921         struct inode            *inode;
 922         int flags;
 923
 924         /* Set up the RPC argument and reply structs
 925          * NB: take care not to mess about with data->commit et al. */
 926
 927         data->req = req;
 928         data->inode = inode = req->wb_context->dentry->d_inode;
 929         data->cred = req->wb_context->cred;
 930
 931         data->args.fh     = NFS_FH(inode);
 932         data->args.offset = req_offset(req) + offset;
 933         data->args.pgbase = req->wb_pgbase + offset;
 934         data->args.pages  = data->pagevec;
 935         data->args.count  = count;
 936         data->args.context = req->wb_context;
 937
 938         data->res.fattr   = &data->fattr;
 939         data->res.count   = count;
 940         data->res.verf    = &data->verf;
 941         nfs_fattr_init(&data->fattr);
 942
 943         /* Set up the initial task struct.  */
 944         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 945         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 946         NFS_PROTO(inode)->write_setup(data, how);
 947
 948         data->task.tk_priority = flush_task_priority(how);
 949         data->task.tk_cookie = (unsigned long)inode;
 950
 951         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 952                 data->task.tk_pid,
 953                 inode->i_sb->s_id,
 954                 (long long)NFS_FILEID(inode),
 955                 count,
 956                 (unsigned long long)data->args.offset);
 957 }
 958
 959 static void nfs_execute_write(struct nfs_write_data *data)
 960 {
 961         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 962         sigset_t oldset;
 963
 964         rpc_clnt_sigmask(clnt, &oldset);
 965         rpc_execute(&data->task);
 966         rpc_clnt_sigunmask(clnt, &oldset);
 967 }
 968
 969 /*
 970  * Generate multiple small requests to write out a single
 971  * contiguous dirty area on one page.
 972  */
 973 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 974 {
 975         struct nfs_page *req = nfs_list_entry(head->next);
 976         struct page *page = req->wb_page;
 977         struct nfs_write_data *data;
 978         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 979         unsigned int offset;
 980         int requests = 0;
 981         LIST_HEAD(list);
 982
 983         nfs_list_remove_request(req);
 984
 985         nbytes = req->wb_bytes;
 986         do {
 987                 size_t len = min(nbytes, wsize);
 988
 989                 data = nfs_writedata_alloc(len);
 990                 if (!data)
 991                         goto out_bad;
 992                 list_add(&data->pages, &list);
 993                 requests++;
 994                 nbytes -= len;
 995         } while (nbytes != 0);
 996         atomic_set(&req->wb_complete, requests);
 997
 998         ClearPageError(page);
 999         set_page_writeback(page);
1000         offset = 0;
1001         nbytes = req->wb_bytes;
1002         do {
1003                 data = list_entry(list.next, struct nfs_write_data, pages);
1004                 list_del_init(&data->pages);
1005
1006                 data->pagevec[0] = page;
1007
1008                 if (nbytes > wsize) {
1009                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1010                                         wsize, offset, how);
1011                         offset += wsize;
1012                         nbytes -= wsize;
1013                 } else {
1014                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1015                                         nbytes, offset, how);
1016                         nbytes = 0;
1017                 }
1018                 nfs_execute_write(data);
1019         } while (nbytes != 0);
1020
1021         return 0;
1022
1023 out_bad:
1024         while (!list_empty(&list)) {
1025                 data = list_entry(list.next, struct nfs_write_data, pages);
1026                 list_del(&data->pages);
1027                 nfs_writedata_release(data);
1028         }
1029         nfs_mark_request_dirty(req);
1030         nfs_clear_page_writeback(req);
1031         return -ENOMEM;
1032 }
1033
1034 /*
1035  * Create an RPC task for the given write request and kick it.
1036  * The page must have been locked by the caller.
1037  *
1038  * It may happen that the page we're passed is not marked dirty.
1039  * This is the case if nfs_updatepage detects a conflicting request
1040  * that has been written but not committed.
1041  */
1042 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1043 {
1044         struct nfs_page         *req;
1045         struct page             **pages;
1046         struct nfs_write_data   *data;
1047         unsigned int            count;
1048
1049         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
1050         if (!data)
1051                 goto out_bad;
1052
1053         pages = data->pagevec;
1054         count = 0;
1055         while (!list_empty(head)) {
1056                 req = nfs_list_entry(head->next);
1057                 nfs_list_remove_request(req);
1058                 nfs_list_add_request(req, &data->pages);
1059                 ClearPageError(req->wb_page);
1060                 set_page_writeback(req->wb_page);
1061                 *pages++ = req->wb_page;
1062                 count += req->wb_bytes;
1063         }
1064         req = nfs_list_entry(data->pages.next);
1065
1066         /* Set up the argument struct */
1067         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1068
1069         nfs_execute_write(data);
1070         return 0;
1071  out_bad:
1072         while (!list_empty(head)) {
1073                 struct nfs_page *req = nfs_list_entry(head->next);
1074                 nfs_list_remove_request(req);
1075                 nfs_mark_request_dirty(req);
1076                 nfs_clear_page_writeback(req);
1077         }
1078         return -ENOMEM;
1079 }
1080
1081 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1082 {
1083         LIST_HEAD(one_request);
1084         int (*flush_one)(struct inode *, struct list_head *, int);
1085         struct nfs_page *req;
1086         int wpages = NFS_SERVER(inode)->wpages;
1087         int wsize = NFS_SERVER(inode)->wsize;
1088         int error;
1089
1090         flush_one = nfs_flush_one;
1091         if (wsize < PAGE_CACHE_SIZE)
1092                 flush_one = nfs_flush_multi;
1093         /* For single writes, FLUSH_STABLE is more efficient */
1094         if (npages <= wpages && npages == NFS_I(inode)->npages
1095                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1096                 how |= FLUSH_STABLE;
1097
1098         do {
1099                 nfs_coalesce_requests(head, &one_request, wpages);
1100                 req = nfs_list_entry(one_request.next);
1101                 error = flush_one(inode, &one_request, how);
1102                 if (error < 0)
1103                         goto out_err;
1104         } while (!list_empty(head));
1105         return 0;
1106 out_err:
1107         while (!list_empty(head)) {
1108                 req = nfs_list_entry(head->next);
1109                 nfs_list_remove_request(req);
1110                 nfs_mark_request_dirty(req);
1111                 nfs_clear_page_writeback(req);
1112         }
1113         return error;
1114 }
1115
1116 /*
1117  * Handle a write reply that flushed part of a page.
1118  */
1119 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1120 {
1121         struct nfs_write_data   *data = calldata;
1122         struct nfs_page         *req = data->req;
1123         struct page             *page = req->wb_page;
1124
1125         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1126                 req->wb_context->dentry->d_inode->i_sb->s_id,
1127                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1128                 req->wb_bytes,
1129                 (long long)req_offset(req));
1130
1131         if (nfs_writeback_done(task, data) != 0)
1132                 return;
1133
1134         if (task->tk_status < 0) {
1135                 ClearPageUptodate(page);
1136                 SetPageError(page);
1137                 req->wb_context->error = task->tk_status;
1138                 dprintk(", error = %d\n", task->tk_status);
1139         } else {
1140 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1141                 if (data->verf.committed < NFS_FILE_SYNC) {
1142                         if (!NFS_NEED_COMMIT(req)) {
1143                                 nfs_defer_commit(req);
1144                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1145                                 dprintk(" defer commit\n");
1146                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1147                                 nfs_defer_reschedule(req);
1148                                 dprintk(" server reboot detected\n");
1149                         }
1150                 } else
1151 #endif
1152                         dprintk(" OK\n");
1153         }
1154
1155         if (atomic_dec_and_test(&req->wb_complete))
1156                 nfs_writepage_release(req);
1157 }
1158
1159 static const struct rpc_call_ops nfs_write_partial_ops = {
1160         .rpc_call_done = nfs_writeback_done_partial,
1161         .rpc_release = nfs_writedata_release,
1162 };
1163
1164 /*
1165  * Handle a write reply that flushes a whole page.
1166  *
1167  * FIXME: There is an inherent race with invalidate_inode_pages and
1168  *        writebacks since the page->count is kept > 1 for as long
1169  *        as the page has a write request pending.
1170  */
1171 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1172 {
1173         struct nfs_write_data   *data = calldata;
1174         struct nfs_page         *req;
1175         struct page             *page;
1176
1177         if (nfs_writeback_done(task, data) != 0)
1178                 return;
1179
1180         /* Update attributes as result of writeback. */
1181         while (!list_empty(&data->pages)) {
1182                 req = nfs_list_entry(data->pages.next);
1183                 nfs_list_remove_request(req);
1184                 page = req->wb_page;
1185
1186                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1187                         req->wb_context->dentry->d_inode->i_sb->s_id,
1188                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1189                         req->wb_bytes,
1190                         (long long)req_offset(req));
1191
1192                 if (task->tk_status < 0) {
1193                         ClearPageUptodate(page);
1194                         SetPageError(page);
1195                         req->wb_context->error = task->tk_status;
1196                         end_page_writeback(page);
1197                         nfs_inode_remove_request(req);
1198                         dprintk(", error = %d\n", task->tk_status);
1199                         goto next;
1200                 }
1201                 end_page_writeback(page);
1202
1203 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1204                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1205                         nfs_inode_remove_request(req);
1206                         dprintk(" OK\n");
1207                         goto next;
1208                 }
1209                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1210                 nfs_mark_request_commit(req);
1211                 dprintk(" marked for commit\n");
1212 #else
1213                 nfs_inode_remove_request(req);
1214 #endif
1215         next:
1216                 nfs_clear_page_writeback(req);
1217         }
1218 }
1219
1220 static const struct rpc_call_ops nfs_write_full_ops = {
1221         .rpc_call_done = nfs_writeback_done_full,
1222         .rpc_release = nfs_writedata_release,
1223 };
1224
1225
1226 /*
1227  * This function is called when the WRITE call is complete.
1228  */
1229 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1230 {
1231         struct nfs_writeargs    *argp = &data->args;
1232         struct nfs_writeres     *resp = &data->res;
1233         int status;
1234
1235         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1236                 task->tk_pid, task->tk_status);
1237
1238         /*
1239          * ->write_done will attempt to use post-op attributes to detect
1240          * conflicting writes by other clients.  A strict interpretation
1241          * of close-to-open would allow us to continue caching even if
1242          * another writer had changed the file, but some applications
1243          * depend on tighter cache coherency when writing.
1244          */
1245         status = NFS_PROTO(data->inode)->write_done(task, data);
1246         if (status != 0)
1247                 return status;
1248         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1249
1250 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1251         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1252                 /* We tried a write call, but the server did not
1253                  * commit data to stable storage even though we
1254                  * requested it.
1255                  * Note: There is a known bug in Tru64 < 5.0 in which
1256                  *       the server reports NFS_DATA_SYNC, but performs
1257                  *       NFS_FILE_SYNC. We therefore implement this checking
1258                  *       as a dprintk() in order to avoid filling syslog.
1259                  */
1260                 static unsigned long    complain;
1261
1262                 if (time_before(complain, jiffies)) {
1263                         dprintk("NFS: faulty NFS server %s:"
1264                                 " (committed = %d) != (stable = %d)\n",
1265                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1266                                 resp->verf->committed, argp->stable);
1267                         complain = jiffies + 300 * HZ;
1268                 }
1269         }
1270 #endif
1271         /* Is this a short write? */
1272         if (task->tk_status >= 0 && resp->count < argp->count) {
1273                 static unsigned long    complain;
1274
1275                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1276
1277                 /* Has the server at least made some progress? */
1278                 if (resp->count != 0) {
1279                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1280                         if (resp->verf->committed != NFS_UNSTABLE) {
1281                                 /* Resend from where the server left off */
1282                                 argp->offset += resp->count;
1283                                 argp->pgbase += resp->count;
1284                                 argp->count -= resp->count;
1285                         } else {
1286                                 /* Resend as a stable write in order to avoid
1287                                  * headaches in the case of a server crash.
1288                                  */
1289                                 argp->stable = NFS_FILE_SYNC;
1290                         }
1291                         rpc_restart_call(task);
1292                         return -EAGAIN;
1293                 }
1294                 if (time_before(complain, jiffies)) {
1295                         printk(KERN_WARNING
1296                                "NFS: Server wrote zero bytes, expected %u.\n",
1297                                         argp->count);
1298                         complain = jiffies + 300 * HZ;
1299                 }
1300                 /* Can't do anything about it except throw an error. */
1301                 task->tk_status = -EIO;
1302         }
1303         return 0;
1304 }
1305
1306
1307 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1308 void nfs_commit_release(void *wdata)
1309 {
1310         nfs_commit_free(wdata);
1311 }
1312
1313 /*
1314  * Set up the argument/result storage required for the RPC call.
1315  */
1316 static void nfs_commit_rpcsetup(struct list_head *head,
1317                 struct nfs_write_data *data,
1318                 int how)
1319 {
1320         struct nfs_page         *first;
1321         struct inode            *inode;
1322         int flags;
1323
1324         /* Set up the RPC argument and reply structs
1325          * NB: take care not to mess about with data->commit et al. */
1326
1327         list_splice_init(head, &data->pages);
1328         first = nfs_list_entry(data->pages.next);
1329         inode = first->wb_context->dentry->d_inode;
1330
1331         data->inode       = inode;
1332         data->cred        = first->wb_context->cred;
1333
1334         data->args.fh     = NFS_FH(data->inode);
1335         /* Note: we always request a commit of the entire inode */
1336         data->args.offset = 0;
1337         data->args.count  = 0;
1338         data->res.count   = 0;
1339         data->res.fattr   = &data->fattr;
1340         data->res.verf    = &data->verf;
1341         nfs_fattr_init(&data->fattr);
1342
1343         /* Set up the initial task struct.  */
1344         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1345         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1346         NFS_PROTO(inode)->commit_setup(data, how);
1347
1348         data->task.tk_priority = flush_task_priority(how);
1349         data->task.tk_cookie = (unsigned long)inode;
1350
1351         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1352 }
1353
1354 /*
1355  * Commit dirty pages
1356  */
1357 static int
1358 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1359 {
1360         struct nfs_write_data   *data;
1361         struct nfs_page         *req;
1362
1363         data = nfs_commit_alloc();
1364
1365         if (!data)
1366                 goto out_bad;
1367
1368         /* Set up the argument struct */
1369         nfs_commit_rpcsetup(head, data, how);
1370
1371         nfs_execute_write(data);
1372         return 0;
1373  out_bad:
1374         while (!list_empty(head)) {
1375                 req = nfs_list_entry(head->next);
1376                 nfs_list_remove_request(req);
1377                 nfs_mark_request_commit(req);
1378                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1379                 nfs_clear_page_writeback(req);
1380         }
1381         return -ENOMEM;
1382 }
1383
1384 /*
1385  * COMMIT call returned
1386  */
1387 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1388 {
1389         struct nfs_write_data   *data = calldata;
1390         struct nfs_page         *req;
1391
1392         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1393                                 task->tk_pid, task->tk_status);
1394
1395         /* Call the NFS version-specific code */
1396         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1397                 return;
1398
1399         while (!list_empty(&data->pages)) {
1400                 req = nfs_list_entry(data->pages.next);
1401                 nfs_list_remove_request(req);
1402                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1403
1404                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1405                         req->wb_context->dentry->d_inode->i_sb->s_id,
1406                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1407                         req->wb_bytes,
1408                         (long long)req_offset(req));
1409                 if (task->tk_status < 0) {
1410                         req->wb_context->error = task->tk_status;
1411                         nfs_inode_remove_request(req);
1412                         dprintk(", error = %d\n", task->tk_status);
1413                         goto next;
1414                 }
1415
1416                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1417                  * returned by the server against all stored verfs. */
1418                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1419                         /* We have a match */
1420                         nfs_inode_remove_request(req);
1421                         dprintk(" OK\n");
1422                         goto next;
1423                 }
1424                 /* We have a mismatch. Write the page again */
1425                 dprintk(" mismatch\n");
1426                 nfs_mark_request_dirty(req);
1427         next:
1428                 nfs_clear_page_writeback(req);
1429         }
1430 }
1431
1432 static const struct rpc_call_ops nfs_commit_ops = {
1433         .rpc_call_done = nfs_commit_done,
1434         .rpc_release = nfs_commit_release,
1435 };
1436 #else
1437 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1438 {
1439         return 0;
1440 }
1441 #endif
1442
1443 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1444 {
1445         struct nfs_inode *nfsi = NFS_I(mapping->host);
1446         LIST_HEAD(head);
1447         long res;
1448
1449         spin_lock(&nfsi->req_lock);
1450         res = nfs_scan_dirty(mapping, wbc, &head);
1451         spin_unlock(&nfsi->req_lock);
1452         if (res) {
1453                 int error = nfs_flush_list(mapping->host, &head, res, how);
1454                 if (error < 0)
1455                         return error;
1456         }
1457         return res;
1458 }
1459
1460 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1461 int nfs_commit_inode(struct inode *inode, int how)
1462 {
1463         struct nfs_inode *nfsi = NFS_I(inode);
1464         LIST_HEAD(head);
1465         int res;
1466
1467         spin_lock(&nfsi->req_lock);
1468         res = nfs_scan_commit(inode, &head, 0, 0);
1469         spin_unlock(&nfsi->req_lock);
1470         if (res) {
1471                 int error = nfs_commit_list(inode, &head, how);
1472                 if (error < 0)
1473                         return error;
1474         }
1475         return res;
1476 }
1477 #endif
1478
1479 long nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
1480                 unsigned int npages, int how)
1481 {
1482         struct nfs_inode *nfsi = NFS_I(inode);
1483         struct address_space *mapping = inode->i_mapping;
1484         struct writeback_control wbc = {
1485                 .bdi = mapping->backing_dev_info,
1486                 .sync_mode = WB_SYNC_ALL,
1487                 .nr_to_write = LONG_MAX,
1488                 .range_start = ((loff_t)idx_start) << PAGE_CACHE_SHIFT,
1489                 .range_end = ((loff_t)(idx_start + npages - 1)) << PAGE_CACHE_SHIFT,
1490         };
1491         LIST_HEAD(head);
1492         int nocommit = how & FLUSH_NOCOMMIT;
1493         long pages, ret;
1494
1495         how &= ~FLUSH_NOCOMMIT;
1496         spin_lock(&nfsi->req_lock);
1497         do {
1498                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1499                 if (ret != 0)
1500                         continue;
1501                 pages = nfs_scan_dirty(mapping, &wbc, &head);
1502                 if (pages != 0) {
1503                         spin_unlock(&nfsi->req_lock);
1504                         if (how & FLUSH_INVALIDATE) {
1505                                 nfs_cancel_dirty_list(&head);
1506                                 ret = pages;
1507                         } else
1508                                 ret = nfs_flush_list(inode, &head, pages, how);
1509                         spin_lock(&nfsi->req_lock);
1510                         continue;
1511                 }
1512                 if (nocommit)
1513                         break;
1514                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1515                 if (pages == 0)
1516                         break;
1517                 if (how & FLUSH_INVALIDATE) {
1518                         spin_unlock(&nfsi->req_lock);
1519                         nfs_cancel_commit_list(&head);
1520                         ret = pages;
1521                         spin_lock(&nfsi->req_lock);
1522                         continue;
1523                 }
1524                 pages += nfs_scan_commit(inode, &head, 0, 0);
1525                 spin_unlock(&nfsi->req_lock);
1526                 ret = nfs_commit_list(inode, &head, how);
1527                 spin_lock(&nfsi->req_lock);
1528         } while (ret >= 0);
1529         spin_unlock(&nfsi->req_lock);
1530         return ret;
1531 }
1532
1533 int __init nfs_init_writepagecache(void)
1534 {
1535         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1536                                              sizeof(struct nfs_write_data),
1537                                              0, SLAB_HWCACHE_ALIGN,
1538                                              NULL, NULL);
1539         if (nfs_wdata_cachep == NULL)
1540                 return -ENOMEM;
1541
1542         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1543                                                      nfs_wdata_cachep);
1544         if (nfs_wdata_mempool == NULL)
1545                 return -ENOMEM;
1546
1547         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1548                                                       nfs_wdata_cachep);
1549         if (nfs_commit_mempool == NULL)
1550                 return -ENOMEM;
1551
1552         return 0;
1553 }
1554
1555 void nfs_destroy_writepagecache(void)
1556 {
1557         mempool_destroy(nfs_commit_mempool);
1558         mempool_destroy(nfs_wdata_mempool);
1559         kmem_cache_destroy(nfs_wdata_cachep);
1560 }
1561