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 int 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                 wbc->nr_to_write -= err;
 405                 err = 0;
 406         }
 407 out:
 408         clear_bit(BDI_write_congested, &bdi->state);
 409         wake_up_all(&nfs_write_congestion);
 410         congestion_end(WRITE);
 411         return err;
 412 }
 413
 414 /*
 415  * Insert a write request into an inode
 416  */
 417 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 418 {
 419         struct nfs_inode *nfsi = NFS_I(inode);
 420         int error;
 421
 422         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 423         BUG_ON(error == -EEXIST);
 424         if (error)
 425                 return error;
 426         if (!nfsi->npages) {
 427                 igrab(inode);
 428                 nfs_begin_data_update(inode);
 429                 if (nfs_have_delegation(inode, FMODE_WRITE))
 430                         nfsi->change_attr++;
 431         }
 432         SetPagePrivate(req->wb_page);
 433         nfsi->npages++;
 434         atomic_inc(&req->wb_count);
 435         return 0;
 436 }
 437
 438 /*
 439  * Insert a write request into an inode
 440  */
 441 static void nfs_inode_remove_request(struct nfs_page *req)
 442 {
 443         struct inode *inode = req->wb_context->dentry->d_inode;
 444         struct nfs_inode *nfsi = NFS_I(inode);
 445
 446         BUG_ON (!NFS_WBACK_BUSY(req));
 447
 448         spin_lock(&nfsi->req_lock);
 449         ClearPagePrivate(req->wb_page);
 450         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 451         nfsi->npages--;
 452         if (!nfsi->npages) {
 453                 spin_unlock(&nfsi->req_lock);
 454                 nfs_end_data_update(inode);
 455                 iput(inode);
 456         } else
 457                 spin_unlock(&nfsi->req_lock);
 458         nfs_clear_request(req);
 459         nfs_release_request(req);
 460 }
 461
 462 /*
 463  * Find a request
 464  */
 465 static inline struct nfs_page *
 466 _nfs_find_request(struct inode *inode, unsigned long index)
 467 {
 468         struct nfs_inode *nfsi = NFS_I(inode);
 469         struct nfs_page *req;
 470
 471         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 472         if (req)
 473                 atomic_inc(&req->wb_count);
 474         return req;
 475 }
 476
 477 static struct nfs_page *
 478 nfs_find_request(struct inode *inode, unsigned long index)
 479 {
 480         struct nfs_page         *req;
 481         struct nfs_inode        *nfsi = NFS_I(inode);
 482
 483         spin_lock(&nfsi->req_lock);
 484         req = _nfs_find_request(inode, index);
 485         spin_unlock(&nfsi->req_lock);
 486         return req;
 487 }
 488
 489 /*
 490  * Add a request to the inode's dirty list.
 491  */
 492 static void
 493 nfs_mark_request_dirty(struct nfs_page *req)
 494 {
 495         struct inode *inode = req->wb_context->dentry->d_inode;
 496         struct nfs_inode *nfsi = NFS_I(inode);
 497
 498         spin_lock(&nfsi->req_lock);
 499         radix_tree_tag_set(&nfsi->nfs_page_tree,
 500                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 501         nfs_list_add_request(req, &nfsi->dirty);
 502         nfsi->ndirty++;
 503         spin_unlock(&nfsi->req_lock);
 504         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 505         mark_inode_dirty(inode);
 506 }
 507
 508 /*
 509  * Check if a request is dirty
 510  */
 511 static inline int
 512 nfs_dirty_request(struct nfs_page *req)
 513 {
 514         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 515         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 516 }
 517
 518 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 519 /*
 520  * Add a request to the inode's commit list.
 521  */
 522 static void
 523 nfs_mark_request_commit(struct nfs_page *req)
 524 {
 525         struct inode *inode = req->wb_context->dentry->d_inode;
 526         struct nfs_inode *nfsi = NFS_I(inode);
 527
 528         spin_lock(&nfsi->req_lock);
 529         nfs_list_add_request(req, &nfsi->commit);
 530         nfsi->ncommit++;
 531         spin_unlock(&nfsi->req_lock);
 532         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 533         mark_inode_dirty(inode);
 534 }
 535 #endif
 536
 537 /*
 538  * Wait for a request to complete.
 539  *
 540  * Interruptible by signals only if mounted with intr flag.
 541  */
 542 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 543 {
 544         struct nfs_inode *nfsi = NFS_I(inode);
 545         struct nfs_page *req;
 546         unsigned long           idx_end, next;
 547         unsigned int            res = 0;
 548         int                     error;
 549
 550         if (npages == 0)
 551                 idx_end = ~0;
 552         else
 553                 idx_end = idx_start + npages - 1;
 554
 555         next = idx_start;
 556         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 557                 if (req->wb_index > idx_end)
 558                         break;
 559
 560                 next = req->wb_index + 1;
 561                 BUG_ON(!NFS_WBACK_BUSY(req));
 562
 563                 atomic_inc(&req->wb_count);
 564                 spin_unlock(&nfsi->req_lock);
 565                 error = nfs_wait_on_request(req);
 566                 nfs_release_request(req);
 567                 spin_lock(&nfsi->req_lock);
 568                 if (error < 0)
 569                         return error;
 570                 res++;
 571         }
 572         return res;
 573 }
 574
 575 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 576 {
 577         struct nfs_inode *nfsi = NFS_I(inode);
 578         int ret;
 579
 580         spin_lock(&nfsi->req_lock);
 581         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 582         spin_unlock(&nfsi->req_lock);
 583         return ret;
 584 }
 585
 586 static void nfs_cancel_dirty_list(struct list_head *head)
 587 {
 588         struct nfs_page *req;
 589         while(!list_empty(head)) {
 590                 req = nfs_list_entry(head->next);
 591                 nfs_list_remove_request(req);
 592                 nfs_inode_remove_request(req);
 593                 nfs_clear_page_writeback(req);
 594         }
 595 }
 596
 597 static void nfs_cancel_commit_list(struct list_head *head)
 598 {
 599         struct nfs_page *req;
 600
 601         while(!list_empty(head)) {
 602                 req = nfs_list_entry(head->next);
 603                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 604                 nfs_list_remove_request(req);
 605                 nfs_inode_remove_request(req);
 606                 nfs_unlock_request(req);
 607         }
 608 }
 609
 610 /*
 611  * nfs_scan_dirty - Scan an inode for dirty requests
 612  * @inode: NFS inode to scan
 613  * @dst: destination list
 614  * @idx_start: lower bound of page->index to scan.
 615  * @npages: idx_start + npages sets the upper bound to scan.
 616  *
 617  * Moves requests from the inode's dirty page list.
 618  * The requests are *not* checked to ensure that they form a contiguous set.
 619  */
 620 static int
 621 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 622 {
 623         struct nfs_inode *nfsi = NFS_I(inode);
 624         int res = 0;
 625
 626         if (nfsi->ndirty != 0) {
 627                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
 628                 nfsi->ndirty -= res;
 629                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
 630                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
 631         }
 632         return res;
 633 }
 634
 635 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 636 /*
 637  * nfs_scan_commit - Scan an inode for commit requests
 638  * @inode: NFS inode to scan
 639  * @dst: destination list
 640  * @idx_start: lower bound of page->index to scan.
 641  * @npages: idx_start + npages sets the upper bound to scan.
 642  *
 643  * Moves requests from the inode's 'commit' request list.
 644  * The requests are *not* checked to ensure that they form a contiguous set.
 645  */
 646 static int
 647 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 648 {
 649         struct nfs_inode *nfsi = NFS_I(inode);
 650         int res = 0;
 651
 652         if (nfsi->ncommit != 0) {
 653                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 654                 nfsi->ncommit -= res;
 655                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 656                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 657         }
 658         return res;
 659 }
 660 #else
 661 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 662 {
 663         return 0;
 664 }
 665 #endif
 666
 667 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 668 {
 669         struct backing_dev_info *bdi = mapping->backing_dev_info;
 670         DEFINE_WAIT(wait);
 671         int ret = 0;
 672
 673         might_sleep();
 674
 675         if (!bdi_write_congested(bdi))
 676                 return 0;
 677
 678         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 679
 680         if (intr) {
 681                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 682                 sigset_t oldset;
 683
 684                 rpc_clnt_sigmask(clnt, &oldset);
 685                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 686                 if (bdi_write_congested(bdi)) {
 687                         if (signalled())
 688                                 ret = -ERESTARTSYS;
 689                         else
 690                                 schedule();
 691                 }
 692                 rpc_clnt_sigunmask(clnt, &oldset);
 693         } else {
 694                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 695                 if (bdi_write_congested(bdi))
 696                         schedule();
 697         }
 698         finish_wait(&nfs_write_congestion, &wait);
 699         return ret;
 700 }
 701
 702
 703 /*
 704  * Try to update any existing write request, or create one if there is none.
 705  * In order to match, the request's credentials must match those of
 706  * the calling process.
 707  *
 708  * Note: Should always be called with the Page Lock held!
 709  */
 710 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 711                 struct inode *inode, struct page *page,
 712                 unsigned int offset, unsigned int bytes)
 713 {
 714         struct nfs_server *server = NFS_SERVER(inode);
 715         struct nfs_inode *nfsi = NFS_I(inode);
 716         struct nfs_page         *req, *new = NULL;
 717         unsigned long           rqend, end;
 718
 719         end = offset + bytes;
 720
 721         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 722                 return ERR_PTR(-ERESTARTSYS);
 723         for (;;) {
 724                 /* Loop over all inode entries and see if we find
 725                  * A request for the page we wish to update
 726                  */
 727                 spin_lock(&nfsi->req_lock);
 728                 req = _nfs_find_request(inode, page->index);
 729                 if (req) {
 730                         if (!nfs_lock_request_dontget(req)) {
 731                                 int error;
 732                                 spin_unlock(&nfsi->req_lock);
 733                                 error = nfs_wait_on_request(req);
 734                                 nfs_release_request(req);
 735                                 if (error < 0) {
 736                                         if (new)
 737                                                 nfs_release_request(new);
 738                                         return ERR_PTR(error);
 739                                 }
 740                                 continue;
 741                         }
 742                         spin_unlock(&nfsi->req_lock);
 743                         if (new)
 744                                 nfs_release_request(new);
 745                         break;
 746                 }
 747
 748                 if (new) {
 749                         int error;
 750                         nfs_lock_request_dontget(new);
 751                         error = nfs_inode_add_request(inode, new);
 752                         if (error) {
 753                                 spin_unlock(&nfsi->req_lock);
 754                                 nfs_unlock_request(new);
 755                                 return ERR_PTR(error);
 756                         }
 757                         spin_unlock(&nfsi->req_lock);
 758                         nfs_mark_request_dirty(new);
 759                         return new;
 760                 }
 761                 spin_unlock(&nfsi->req_lock);
 762
 763                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 764                 if (IS_ERR(new))
 765                         return new;
 766         }
 767
 768         /* We have a request for our page.
 769          * If the creds don't match, or the
 770          * page addresses don't match,
 771          * tell the caller to wait on the conflicting
 772          * request.
 773          */
 774         rqend = req->wb_offset + req->wb_bytes;
 775         if (req->wb_context != ctx
 776             || req->wb_page != page
 777             || !nfs_dirty_request(req)
 778             || offset > rqend || end < req->wb_offset) {
 779                 nfs_unlock_request(req);
 780                 return ERR_PTR(-EBUSY);
 781         }
 782
 783         /* Okay, the request matches. Update the region */
 784         if (offset < req->wb_offset) {
 785                 req->wb_offset = offset;
 786                 req->wb_pgbase = offset;
 787                 req->wb_bytes = rqend - req->wb_offset;
 788         }
 789
 790         if (end > rqend)
 791                 req->wb_bytes = end - req->wb_offset;
 792
 793         return req;
 794 }
 795
 796 int nfs_flush_incompatible(struct file *file, struct page *page)
 797 {
 798         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 799         struct inode    *inode = page->mapping->host;
 800         struct nfs_page *req;
 801         int             status = 0;
 802         /*
 803          * Look for a request corresponding to this page. If there
 804          * is one, and it belongs to another file, we flush it out
 805          * before we try to copy anything into the page. Do this
 806          * due to the lack of an ACCESS-type call in NFSv2.
 807          * Also do the same if we find a request from an existing
 808          * dropped page.
 809          */
 810         req = nfs_find_request(inode, page->index);
 811         if (req) {
 812                 if (req->wb_page != page || ctx != req->wb_context)
 813                         status = nfs_wb_page(inode, page);
 814                 nfs_release_request(req);
 815         }
 816         return (status < 0) ? status : 0;
 817 }
 818
 819 /*
 820  * Update and possibly write a cached page of an NFS file.
 821  *
 822  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 823  * things with a page scheduled for an RPC call (e.g. invalidate it).
 824  */
 825 int nfs_updatepage(struct file *file, struct page *page,
 826                 unsigned int offset, unsigned int count)
 827 {
 828         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 829         struct inode    *inode = page->mapping->host;
 830         struct nfs_page *req;
 831         int             status = 0;
 832
 833         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 834
 835         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 836                 file->f_dentry->d_parent->d_name.name,
 837                 file->f_dentry->d_name.name, count,
 838                 (long long)(page_offset(page) +offset));
 839
 840         if (IS_SYNC(inode)) {
 841                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 842                 if (status > 0) {
 843                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 844                                 SetPageUptodate(page);
 845                         return 0;
 846                 }
 847                 return status;
 848         }
 849
 850         /* If we're not using byte range locks, and we know the page
 851          * is entirely in cache, it may be more efficient to avoid
 852          * fragmenting write requests.
 853          */
 854         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 855                 loff_t end_offs = i_size_read(inode) - 1;
 856                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 857
 858                 count += offset;
 859                 offset = 0;
 860                 if (unlikely(end_offs < 0)) {
 861                         /* Do nothing */
 862                 } else if (page->index == end_index) {
 863                         unsigned int pglen;
 864                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 865                         if (count < pglen)
 866                                 count = pglen;
 867                 } else if (page->index < end_index)
 868                         count = PAGE_CACHE_SIZE;
 869         }
 870
 871         /*
 872          * Try to find an NFS request corresponding to this page
 873          * and update it.
 874          * If the existing request cannot be updated, we must flush
 875          * it out now.
 876          */
 877         do {
 878                 req = nfs_update_request(ctx, inode, page, offset, count);
 879                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 880                 if (status != -EBUSY)
 881                         break;
 882                 /* Request could not be updated. Flush it out and try again */
 883                 status = nfs_wb_page(inode, page);
 884         } while (status >= 0);
 885         if (status < 0)
 886                 goto done;
 887
 888         status = 0;
 889
 890         /* Update file length */
 891         nfs_grow_file(page, offset, count);
 892         /* Set the PG_uptodate flag? */
 893         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 894         nfs_unlock_request(req);
 895 done:
 896         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 897                         status, (long long)i_size_read(inode));
 898         if (status < 0)
 899                 ClearPageUptodate(page);
 900         return status;
 901 }
 902
 903 static void nfs_writepage_release(struct nfs_page *req)
 904 {
 905         end_page_writeback(req->wb_page);
 906
 907 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 908         if (!PageError(req->wb_page)) {
 909                 if (NFS_NEED_RESCHED(req)) {
 910                         nfs_mark_request_dirty(req);
 911                         goto out;
 912                 } else if (NFS_NEED_COMMIT(req)) {
 913                         nfs_mark_request_commit(req);
 914                         goto out;
 915                 }
 916         }
 917         nfs_inode_remove_request(req);
 918
 919 out:
 920         nfs_clear_commit(req);
 921         nfs_clear_reschedule(req);
 922 #else
 923         nfs_inode_remove_request(req);
 924 #endif
 925         nfs_clear_page_writeback(req);
 926 }
 927
 928 static inline int flush_task_priority(int how)
 929 {
 930         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 931                 case FLUSH_HIGHPRI:
 932                         return RPC_PRIORITY_HIGH;
 933                 case FLUSH_LOWPRI:
 934                         return RPC_PRIORITY_LOW;
 935         }
 936         return RPC_PRIORITY_NORMAL;
 937 }
 938
 939 /*
 940  * Set up the argument/result storage required for the RPC call.
 941  */
 942 static void nfs_write_rpcsetup(struct nfs_page *req,
 943                 struct nfs_write_data *data,
 944                 const struct rpc_call_ops *call_ops,
 945                 unsigned int count, unsigned int offset,
 946                 int how)
 947 {
 948         struct inode            *inode;
 949         int flags;
 950
 951         /* Set up the RPC argument and reply structs
 952          * NB: take care not to mess about with data->commit et al. */
 953
 954         data->req = req;
 955         data->inode = inode = req->wb_context->dentry->d_inode;
 956         data->cred = req->wb_context->cred;
 957
 958         data->args.fh     = NFS_FH(inode);
 959         data->args.offset = req_offset(req) + offset;
 960         data->args.pgbase = req->wb_pgbase + offset;
 961         data->args.pages  = data->pagevec;
 962         data->args.count  = count;
 963         data->args.context = req->wb_context;
 964
 965         data->res.fattr   = &data->fattr;
 966         data->res.count   = count;
 967         data->res.verf    = &data->verf;
 968         nfs_fattr_init(&data->fattr);
 969
 970         /* Set up the initial task struct.  */
 971         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 972         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 973         NFS_PROTO(inode)->write_setup(data, how);
 974
 975         data->task.tk_priority = flush_task_priority(how);
 976         data->task.tk_cookie = (unsigned long)inode;
 977
 978         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 979                 data->task.tk_pid,
 980                 inode->i_sb->s_id,
 981                 (long long)NFS_FILEID(inode),
 982                 count,
 983                 (unsigned long long)data->args.offset);
 984 }
 985
 986 static void nfs_execute_write(struct nfs_write_data *data)
 987 {
 988         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 989         sigset_t oldset;
 990
 991         rpc_clnt_sigmask(clnt, &oldset);
 992         rpc_execute(&data->task);
 993         rpc_clnt_sigunmask(clnt, &oldset);
 994 }
 995
 996 /*
 997  * Generate multiple small requests to write out a single
 998  * contiguous dirty area on one page.
 999  */
1000 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
1001 {
1002         struct nfs_page *req = nfs_list_entry(head->next);
1003         struct page *page = req->wb_page;
1004         struct nfs_write_data *data;
1005         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
1006         unsigned int offset;
1007         int requests = 0;
1008         LIST_HEAD(list);
1009
1010         nfs_list_remove_request(req);
1011
1012         nbytes = req->wb_bytes;
1013         do {
1014                 size_t len = min(nbytes, wsize);
1015
1016                 data = nfs_writedata_alloc(len);
1017                 if (!data)
1018                         goto out_bad;
1019                 list_add(&data->pages, &list);
1020                 requests++;
1021                 nbytes -= len;
1022         } while (nbytes != 0);
1023         atomic_set(&req->wb_complete, requests);
1024
1025         ClearPageError(page);
1026         set_page_writeback(page);
1027         offset = 0;
1028         nbytes = req->wb_bytes;
1029         do {
1030                 data = list_entry(list.next, struct nfs_write_data, pages);
1031                 list_del_init(&data->pages);
1032
1033                 data->pagevec[0] = page;
1034
1035                 if (nbytes > wsize) {
1036                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1037                                         wsize, offset, how);
1038                         offset += wsize;
1039                         nbytes -= wsize;
1040                 } else {
1041                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1042                                         nbytes, offset, how);
1043                         nbytes = 0;
1044                 }
1045                 nfs_execute_write(data);
1046         } while (nbytes != 0);
1047
1048         return 0;
1049
1050 out_bad:
1051         while (!list_empty(&list)) {
1052                 data = list_entry(list.next, struct nfs_write_data, pages);
1053                 list_del(&data->pages);
1054                 nfs_writedata_release(data);
1055         }
1056         nfs_mark_request_dirty(req);
1057         nfs_clear_page_writeback(req);
1058         return -ENOMEM;
1059 }
1060
1061 /*
1062  * Create an RPC task for the given write request and kick it.
1063  * The page must have been locked by the caller.
1064  *
1065  * It may happen that the page we're passed is not marked dirty.
1066  * This is the case if nfs_updatepage detects a conflicting request
1067  * that has been written but not committed.
1068  */
1069 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1070 {
1071         struct nfs_page         *req;
1072         struct page             **pages;
1073         struct nfs_write_data   *data;
1074         unsigned int            count;
1075
1076         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
1077         if (!data)
1078                 goto out_bad;
1079
1080         pages = data->pagevec;
1081         count = 0;
1082         while (!list_empty(head)) {
1083                 req = nfs_list_entry(head->next);
1084                 nfs_list_remove_request(req);
1085                 nfs_list_add_request(req, &data->pages);
1086                 ClearPageError(req->wb_page);
1087                 set_page_writeback(req->wb_page);
1088                 *pages++ = req->wb_page;
1089                 count += req->wb_bytes;
1090         }
1091         req = nfs_list_entry(data->pages.next);
1092
1093         /* Set up the argument struct */
1094         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1095
1096         nfs_execute_write(data);
1097         return 0;
1098  out_bad:
1099         while (!list_empty(head)) {
1100                 struct nfs_page *req = nfs_list_entry(head->next);
1101                 nfs_list_remove_request(req);
1102                 nfs_mark_request_dirty(req);
1103                 nfs_clear_page_writeback(req);
1104         }
1105         return -ENOMEM;
1106 }
1107
1108 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1109 {
1110         LIST_HEAD(one_request);
1111         int (*flush_one)(struct inode *, struct list_head *, int);
1112         struct nfs_page *req;
1113         int wpages = NFS_SERVER(inode)->wpages;
1114         int wsize = NFS_SERVER(inode)->wsize;
1115         int error;
1116
1117         flush_one = nfs_flush_one;
1118         if (wsize < PAGE_CACHE_SIZE)
1119                 flush_one = nfs_flush_multi;
1120         /* For single writes, FLUSH_STABLE is more efficient */
1121         if (npages <= wpages && npages == NFS_I(inode)->npages
1122                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1123                 how |= FLUSH_STABLE;
1124
1125         do {
1126                 nfs_coalesce_requests(head, &one_request, wpages);
1127                 req = nfs_list_entry(one_request.next);
1128                 error = flush_one(inode, &one_request, how);
1129                 if (error < 0)
1130                         goto out_err;
1131         } while (!list_empty(head));
1132         return 0;
1133 out_err:
1134         while (!list_empty(head)) {
1135                 req = nfs_list_entry(head->next);
1136                 nfs_list_remove_request(req);
1137                 nfs_mark_request_dirty(req);
1138                 nfs_clear_page_writeback(req);
1139         }
1140         return error;
1141 }
1142
1143 /*
1144  * Handle a write reply that flushed part of a page.
1145  */
1146 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1147 {
1148         struct nfs_write_data   *data = calldata;
1149         struct nfs_page         *req = data->req;
1150         struct page             *page = req->wb_page;
1151
1152         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1153                 req->wb_context->dentry->d_inode->i_sb->s_id,
1154                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1155                 req->wb_bytes,
1156                 (long long)req_offset(req));
1157
1158         if (nfs_writeback_done(task, data) != 0)
1159                 return;
1160
1161         if (task->tk_status < 0) {
1162                 ClearPageUptodate(page);
1163                 SetPageError(page);
1164                 req->wb_context->error = task->tk_status;
1165                 dprintk(", error = %d\n", task->tk_status);
1166         } else {
1167 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1168                 if (data->verf.committed < NFS_FILE_SYNC) {
1169                         if (!NFS_NEED_COMMIT(req)) {
1170                                 nfs_defer_commit(req);
1171                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1172                                 dprintk(" defer commit\n");
1173                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1174                                 nfs_defer_reschedule(req);
1175                                 dprintk(" server reboot detected\n");
1176                         }
1177                 } else
1178 #endif
1179                         dprintk(" OK\n");
1180         }
1181
1182         if (atomic_dec_and_test(&req->wb_complete))
1183                 nfs_writepage_release(req);
1184 }
1185
1186 static const struct rpc_call_ops nfs_write_partial_ops = {
1187         .rpc_call_done = nfs_writeback_done_partial,
1188         .rpc_release = nfs_writedata_release,
1189 };
1190
1191 /*
1192  * Handle a write reply that flushes a whole page.
1193  *
1194  * FIXME: There is an inherent race with invalidate_inode_pages and
1195  *        writebacks since the page->count is kept > 1 for as long
1196  *        as the page has a write request pending.
1197  */
1198 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1199 {
1200         struct nfs_write_data   *data = calldata;
1201         struct nfs_page         *req;
1202         struct page             *page;
1203
1204         if (nfs_writeback_done(task, data) != 0)
1205                 return;
1206
1207         /* Update attributes as result of writeback. */
1208         while (!list_empty(&data->pages)) {
1209                 req = nfs_list_entry(data->pages.next);
1210                 nfs_list_remove_request(req);
1211                 page = req->wb_page;
1212
1213                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1214                         req->wb_context->dentry->d_inode->i_sb->s_id,
1215                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1216                         req->wb_bytes,
1217                         (long long)req_offset(req));
1218
1219                 if (task->tk_status < 0) {
1220                         ClearPageUptodate(page);
1221                         SetPageError(page);
1222                         req->wb_context->error = task->tk_status;
1223                         end_page_writeback(page);
1224                         nfs_inode_remove_request(req);
1225                         dprintk(", error = %d\n", task->tk_status);
1226                         goto next;
1227                 }
1228                 end_page_writeback(page);
1229
1230 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1231                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1232                         nfs_inode_remove_request(req);
1233                         dprintk(" OK\n");
1234                         goto next;
1235                 }
1236                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1237                 nfs_mark_request_commit(req);
1238                 dprintk(" marked for commit\n");
1239 #else
1240                 nfs_inode_remove_request(req);
1241 #endif
1242         next:
1243                 nfs_clear_page_writeback(req);
1244         }
1245 }
1246
1247 static const struct rpc_call_ops nfs_write_full_ops = {
1248         .rpc_call_done = nfs_writeback_done_full,
1249         .rpc_release = nfs_writedata_release,
1250 };
1251
1252
1253 /*
1254  * This function is called when the WRITE call is complete.
1255  */
1256 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1257 {
1258         struct nfs_writeargs    *argp = &data->args;
1259         struct nfs_writeres     *resp = &data->res;
1260         int status;
1261
1262         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1263                 task->tk_pid, task->tk_status);
1264
1265         /*
1266          * ->write_done will attempt to use post-op attributes to detect
1267          * conflicting writes by other clients.  A strict interpretation
1268          * of close-to-open would allow us to continue caching even if
1269          * another writer had changed the file, but some applications
1270          * depend on tighter cache coherency when writing.
1271          */
1272         status = NFS_PROTO(data->inode)->write_done(task, data);
1273         if (status != 0)
1274                 return status;
1275         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1276
1277 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1278         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1279                 /* We tried a write call, but the server did not
1280                  * commit data to stable storage even though we
1281                  * requested it.
1282                  * Note: There is a known bug in Tru64 < 5.0 in which
1283                  *       the server reports NFS_DATA_SYNC, but performs
1284                  *       NFS_FILE_SYNC. We therefore implement this checking
1285                  *       as a dprintk() in order to avoid filling syslog.
1286                  */
1287                 static unsigned long    complain;
1288
1289                 if (time_before(complain, jiffies)) {
1290                         dprintk("NFS: faulty NFS server %s:"
1291                                 " (committed = %d) != (stable = %d)\n",
1292                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1293                                 resp->verf->committed, argp->stable);
1294                         complain = jiffies + 300 * HZ;
1295                 }
1296         }
1297 #endif
1298         /* Is this a short write? */
1299         if (task->tk_status >= 0 && resp->count < argp->count) {
1300                 static unsigned long    complain;
1301
1302                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1303
1304                 /* Has the server at least made some progress? */
1305                 if (resp->count != 0) {
1306                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1307                         if (resp->verf->committed != NFS_UNSTABLE) {
1308                                 /* Resend from where the server left off */
1309                                 argp->offset += resp->count;
1310                                 argp->pgbase += resp->count;
1311                                 argp->count -= resp->count;
1312                         } else {
1313                                 /* Resend as a stable write in order to avoid
1314                                  * headaches in the case of a server crash.
1315                                  */
1316                                 argp->stable = NFS_FILE_SYNC;
1317                         }
1318                         rpc_restart_call(task);
1319                         return -EAGAIN;
1320                 }
1321                 if (time_before(complain, jiffies)) {
1322                         printk(KERN_WARNING
1323                                "NFS: Server wrote zero bytes, expected %u.\n",
1324                                         argp->count);
1325                         complain = jiffies + 300 * HZ;
1326                 }
1327                 /* Can't do anything about it except throw an error. */
1328                 task->tk_status = -EIO;
1329         }
1330         return 0;
1331 }
1332
1333
1334 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1335 void nfs_commit_release(void *wdata)
1336 {
1337         nfs_commit_free(wdata);
1338 }
1339
1340 /*
1341  * Set up the argument/result storage required for the RPC call.
1342  */
1343 static void nfs_commit_rpcsetup(struct list_head *head,
1344                 struct nfs_write_data *data,
1345                 int how)
1346 {
1347         struct nfs_page         *first;
1348         struct inode            *inode;
1349         int flags;
1350
1351         /* Set up the RPC argument and reply structs
1352          * NB: take care not to mess about with data->commit et al. */
1353
1354         list_splice_init(head, &data->pages);
1355         first = nfs_list_entry(data->pages.next);
1356         inode = first->wb_context->dentry->d_inode;
1357
1358         data->inode       = inode;
1359         data->cred        = first->wb_context->cred;
1360
1361         data->args.fh     = NFS_FH(data->inode);
1362         /* Note: we always request a commit of the entire inode */
1363         data->args.offset = 0;
1364         data->args.count  = 0;
1365         data->res.count   = 0;
1366         data->res.fattr   = &data->fattr;
1367         data->res.verf    = &data->verf;
1368         nfs_fattr_init(&data->fattr);
1369
1370         /* Set up the initial task struct.  */
1371         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1372         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1373         NFS_PROTO(inode)->commit_setup(data, how);
1374
1375         data->task.tk_priority = flush_task_priority(how);
1376         data->task.tk_cookie = (unsigned long)inode;
1377
1378         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1379 }
1380
1381 /*
1382  * Commit dirty pages
1383  */
1384 static int
1385 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1386 {
1387         struct nfs_write_data   *data;
1388         struct nfs_page         *req;
1389
1390         data = nfs_commit_alloc();
1391
1392         if (!data)
1393                 goto out_bad;
1394
1395         /* Set up the argument struct */
1396         nfs_commit_rpcsetup(head, data, how);
1397
1398         nfs_execute_write(data);
1399         return 0;
1400  out_bad:
1401         while (!list_empty(head)) {
1402                 req = nfs_list_entry(head->next);
1403                 nfs_list_remove_request(req);
1404                 nfs_mark_request_commit(req);
1405                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1406                 nfs_clear_page_writeback(req);
1407         }
1408         return -ENOMEM;
1409 }
1410
1411 /*
1412  * COMMIT call returned
1413  */
1414 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1415 {
1416         struct nfs_write_data   *data = calldata;
1417         struct nfs_page         *req;
1418
1419         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1420                                 task->tk_pid, task->tk_status);
1421
1422         /* Call the NFS version-specific code */
1423         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1424                 return;
1425
1426         while (!list_empty(&data->pages)) {
1427                 req = nfs_list_entry(data->pages.next);
1428                 nfs_list_remove_request(req);
1429                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1430
1431                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1432                         req->wb_context->dentry->d_inode->i_sb->s_id,
1433                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1434                         req->wb_bytes,
1435                         (long long)req_offset(req));
1436                 if (task->tk_status < 0) {
1437                         req->wb_context->error = task->tk_status;
1438                         nfs_inode_remove_request(req);
1439                         dprintk(", error = %d\n", task->tk_status);
1440                         goto next;
1441                 }
1442
1443                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1444                  * returned by the server against all stored verfs. */
1445                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1446                         /* We have a match */
1447                         nfs_inode_remove_request(req);
1448                         dprintk(" OK\n");
1449                         goto next;
1450                 }
1451                 /* We have a mismatch. Write the page again */
1452                 dprintk(" mismatch\n");
1453                 nfs_mark_request_dirty(req);
1454         next:
1455                 nfs_clear_page_writeback(req);
1456         }
1457 }
1458
1459 static const struct rpc_call_ops nfs_commit_ops = {
1460         .rpc_call_done = nfs_commit_done,
1461         .rpc_release = nfs_commit_release,
1462 };
1463 #else
1464 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1465 {
1466         return 0;
1467 }
1468 #endif
1469
1470 static int nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1471 {
1472         struct nfs_inode *nfsi = NFS_I(mapping->host);
1473         LIST_HEAD(head);
1474         pgoff_t index = wbc->range_start >> PAGE_CACHE_SHIFT;
1475         unsigned long npages = 1 + (wbc->range_end >> PAGE_CACHE_SHIFT) - index;
1476         int res;
1477
1478         spin_lock(&nfsi->req_lock);
1479         res = nfs_scan_dirty(mapping->host, &head, index, npages);
1480         spin_unlock(&nfsi->req_lock);
1481         if (res) {
1482                 int error = nfs_flush_list(mapping->host, &head, res, how);
1483                 if (error < 0)
1484                         return error;
1485                 wbc->nr_to_write -= res;
1486         }
1487         return res;
1488 }
1489
1490 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1491 int nfs_commit_inode(struct inode *inode, int how)
1492 {
1493         struct nfs_inode *nfsi = NFS_I(inode);
1494         LIST_HEAD(head);
1495         int res;
1496
1497         spin_lock(&nfsi->req_lock);
1498         res = nfs_scan_commit(inode, &head, 0, 0);
1499         spin_unlock(&nfsi->req_lock);
1500         if (res) {
1501                 int error = nfs_commit_list(inode, &head, how);
1502                 if (error < 0)
1503                         return error;
1504         }
1505         return res;
1506 }
1507 #endif
1508
1509 int nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
1510                 unsigned int npages, int how)
1511 {
1512         struct nfs_inode *nfsi = NFS_I(inode);
1513         LIST_HEAD(head);
1514         int nocommit = how & FLUSH_NOCOMMIT;
1515         int pages, ret;
1516
1517         how &= ~FLUSH_NOCOMMIT;
1518         spin_lock(&nfsi->req_lock);
1519         do {
1520                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1521                 if (ret != 0)
1522                         continue;
1523                 pages = nfs_scan_dirty(inode, &head, idx_start, npages);
1524                 if (pages != 0) {
1525                         spin_unlock(&nfsi->req_lock);
1526                         if (how & FLUSH_INVALIDATE) {
1527                                 nfs_cancel_dirty_list(&head);
1528                                 ret = pages;
1529                         } else
1530                                 ret = nfs_flush_list(inode, &head, pages, how);
1531                         spin_lock(&nfsi->req_lock);
1532                         continue;
1533                 }
1534                 if (nocommit)
1535                         break;
1536                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1537                 if (pages == 0)
1538                         break;
1539                 if (how & FLUSH_INVALIDATE) {
1540                         spin_unlock(&nfsi->req_lock);
1541                         nfs_cancel_commit_list(&head);
1542                         ret = pages;
1543                         spin_lock(&nfsi->req_lock);
1544                         continue;
1545                 }
1546                 pages += nfs_scan_commit(inode, &head, 0, 0);
1547                 spin_unlock(&nfsi->req_lock);
1548                 ret = nfs_commit_list(inode, &head, how);
1549                 spin_lock(&nfsi->req_lock);
1550         } while (ret >= 0);
1551         spin_unlock(&nfsi->req_lock);
1552         return ret;
1553 }
1554
1555 int __init nfs_init_writepagecache(void)
1556 {
1557         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1558                                              sizeof(struct nfs_write_data),
1559                                              0, SLAB_HWCACHE_ALIGN,
1560                                              NULL, NULL);
1561         if (nfs_wdata_cachep == NULL)
1562                 return -ENOMEM;
1563
1564         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1565                                                      nfs_wdata_cachep);
1566         if (nfs_wdata_mempool == NULL)
1567                 return -ENOMEM;
1568
1569         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1570                                                       nfs_wdata_cachep);
1571         if (nfs_commit_mempool == NULL)
1572                 return -ENOMEM;
1573
1574         return 0;
1575 }
1576
1577 void nfs_destroy_writepagecache(void)
1578 {
1579         mempool_destroy(nfs_commit_mempool);
1580         mempool_destroy(nfs_wdata_mempool);
1581         kmem_cache_destroy(nfs_wdata_cachep);
1582 }
1583