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/mpage.h>
  55 #include <linux/writeback.h>
  56
  57 #include <linux/sunrpc/clnt.h>
  58 #include <linux/nfs_fs.h>
  59 #include <linux/nfs_mount.h>
  60 #include <linux/nfs_page.h>
  61 #include <asm/uaccess.h>
  62 #include <linux/smp_lock.h>
  63
  64 #include "delegation.h"
  65 #include "iostat.h"
  66
  67 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  68
  69 #define MIN_POOL_WRITE          (32)
  70 #define MIN_POOL_COMMIT         (4)
  71
  72 /*
  73  * Local function declarations
  74  */
  75 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  76                                             struct inode *,
  77                                             struct page *,
  78                                             unsigned int, unsigned int);
  79 static int nfs_wait_on_write_congestion(struct address_space *, int);
  80 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
  81 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
  82                            unsigned int npages, 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_free(struct nfs_write_data *p)
 105 {
 106         if (p && (p->pagevec != &p->page_array[0]))
 107                 kfree(p->pagevec);
 108         mempool_free(p, nfs_commit_mempool);
 109 }
 110
 111 struct nfs_write_data *nfs_writedata_alloc(size_t len)
 112 {
 113         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 114         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
 115
 116         if (p) {
 117                 memset(p, 0, sizeof(*p));
 118                 INIT_LIST_HEAD(&p->pages);
 119                 p->npages = pagecount;
 120                 if (pagecount <= ARRAY_SIZE(p->page_array))
 121                         p->pagevec = p->page_array;
 122                 else {
 123                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 124                         if (!p->pagevec) {
 125                                 mempool_free(p, nfs_wdata_mempool);
 126                                 p = NULL;
 127                         }
 128                 }
 129         }
 130         return p;
 131 }
 132
 133 static void nfs_writedata_free(struct nfs_write_data *p)
 134 {
 135         if (p && (p->pagevec != &p->page_array[0]))
 136                 kfree(p->pagevec);
 137         mempool_free(p, nfs_wdata_mempool);
 138 }
 139
 140 void nfs_writedata_release(void *wdata)
 141 {
 142         nfs_writedata_free(wdata);
 143 }
 144
 145 /* Adjust the file length if we're writing beyond the end */
 146 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 147 {
 148         struct inode *inode = page->mapping->host;
 149         loff_t end, i_size = i_size_read(inode);
 150         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 151
 152         if (i_size > 0 && page->index < end_index)
 153                 return;
 154         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 155         if (i_size >= end)
 156                 return;
 157         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 158         i_size_write(inode, end);
 159 }
 160
 161 /* We can set the PG_uptodate flag if we see that a write request
 162  * covers the full page.
 163  */
 164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 165 {
 166         loff_t end_offs;
 167
 168         if (PageUptodate(page))
 169                 return;
 170         if (base != 0)
 171                 return;
 172         if (count == PAGE_CACHE_SIZE) {
 173                 SetPageUptodate(page);
 174                 return;
 175         }
 176
 177         end_offs = i_size_read(page->mapping->host) - 1;
 178         if (end_offs < 0)
 179                 return;
 180         /* Is this the last page? */
 181         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
 182                 return;
 183         /* This is the last page: set PG_uptodate if we cover the entire
 184          * extent of the data, then zero the rest of the page.
 185          */
 186         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
 187                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 188                 SetPageUptodate(page);
 189         }
 190 }
 191
 192 /*
 193  * Write a page synchronously.
 194  * Offset is the data offset within the page.
 195  */
 196 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 197                 struct page *page, unsigned int offset, unsigned int count,
 198                 int how)
 199 {
 200         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 201         int             result, written = 0;
 202         struct nfs_write_data *wdata;
 203
 204         wdata = nfs_writedata_alloc(wsize);
 205         if (!wdata)
 206                 return -ENOMEM;
 207
 208         wdata->flags = how;
 209         wdata->cred = ctx->cred;
 210         wdata->inode = inode;
 211         wdata->args.fh = NFS_FH(inode);
 212         wdata->args.context = ctx;
 213         wdata->args.pages = &page;
 214         wdata->args.stable = NFS_FILE_SYNC;
 215         wdata->args.pgbase = offset;
 216         wdata->args.count = wsize;
 217         wdata->res.fattr = &wdata->fattr;
 218         wdata->res.verf = &wdata->verf;
 219
 220         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 221                 inode->i_sb->s_id,
 222                 (long long)NFS_FILEID(inode),
 223                 count, (long long)(page_offset(page) + offset));
 224
 225         set_page_writeback(page);
 226         nfs_begin_data_update(inode);
 227         do {
 228                 if (count < wsize)
 229                         wdata->args.count = count;
 230                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 231
 232                 result = NFS_PROTO(inode)->write(wdata);
 233
 234                 if (result < 0) {
 235                         /* Must mark the page invalid after I/O error */
 236                         ClearPageUptodate(page);
 237                         goto io_error;
 238                 }
 239                 if (result < wdata->args.count)
 240                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 241                                         wdata->args.count, result);
 242
 243                 wdata->args.offset += result;
 244                 wdata->args.pgbase += result;
 245                 written += result;
 246                 count -= result;
 247                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 248         } while (count);
 249         /* Update file length */
 250         nfs_grow_file(page, offset, written);
 251         /* Set the PG_uptodate flag? */
 252         nfs_mark_uptodate(page, offset, written);
 253
 254         if (PageError(page))
 255                 ClearPageError(page);
 256
 257 io_error:
 258         nfs_end_data_update(inode);
 259         end_page_writeback(page);
 260         nfs_writedata_free(wdata);
 261         return written ? written : result;
 262 }
 263
 264 static int nfs_writepage_async(struct nfs_open_context *ctx,
 265                 struct inode *inode, struct page *page,
 266                 unsigned int offset, unsigned int count)
 267 {
 268         struct nfs_page *req;
 269
 270         req = nfs_update_request(ctx, inode, page, offset, count);
 271         if (IS_ERR(req))
 272                 return PTR_ERR(req);
 273         /* Update file length */
 274         nfs_grow_file(page, offset, count);
 275         /* Set the PG_uptodate flag? */
 276         nfs_mark_uptodate(page, offset, count);
 277         nfs_unlock_request(req);
 278         return 0;
 279 }
 280
 281 static int wb_priority(struct writeback_control *wbc)
 282 {
 283         if (wbc->for_reclaim)
 284                 return FLUSH_HIGHPRI;
 285         if (wbc->for_kupdate)
 286                 return FLUSH_LOWPRI;
 287         return 0;
 288 }
 289
 290 /*
 291  * Write an mmapped page to the server.
 292  */
 293 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 294 {
 295         struct nfs_open_context *ctx;
 296         struct inode *inode = page->mapping->host;
 297         unsigned long end_index;
 298         unsigned offset = PAGE_CACHE_SIZE;
 299         loff_t i_size = i_size_read(inode);
 300         int inode_referenced = 0;
 301         int priority = wb_priority(wbc);
 302         int err;
 303
 304         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 305         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 306
 307         /*
 308          * Note: We need to ensure that we have a reference to the inode
 309          *       if we are to do asynchronous writes. If not, waiting
 310          *       in nfs_wait_on_request() may deadlock with clear_inode().
 311          *
 312          *       If igrab() fails here, then it is in any case safe to
 313          *       call nfs_wb_page(), since there will be no pending writes.
 314          */
 315         if (igrab(inode) != 0)
 316                 inode_referenced = 1;
 317         end_index = i_size >> PAGE_CACHE_SHIFT;
 318
 319         /* Ensure we've flushed out any previous writes */
 320         nfs_wb_page_priority(inode, page, priority);
 321
 322         /* easy case */
 323         if (page->index < end_index)
 324                 goto do_it;
 325         /* things got complicated... */
 326         offset = i_size & (PAGE_CACHE_SIZE-1);
 327
 328         /* OK, are we completely out? */
 329         err = 0; /* potential race with truncate - ignore */
 330         if (page->index >= end_index+1 || !offset)
 331                 goto out;
 332 do_it:
 333         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 334         if (ctx == NULL) {
 335                 err = -EBADF;
 336                 goto out;
 337         }
 338         lock_kernel();
 339         if (!IS_SYNC(inode) && inode_referenced) {
 340                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
 341                 if (!wbc->for_writepages)
 342                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 343         } else {
 344                 err = nfs_writepage_sync(ctx, inode, page, 0,
 345                                                 offset, priority);
 346                 if (err >= 0) {
 347                         if (err != offset)
 348                                 redirty_page_for_writepage(wbc, page);
 349                         err = 0;
 350                 }
 351         }
 352         unlock_kernel();
 353         put_nfs_open_context(ctx);
 354 out:
 355         unlock_page(page);
 356         if (inode_referenced)
 357                 iput(inode);
 358         return err;
 359 }
 360
 361 /*
 362  * Note: causes nfs_update_request() to block on the assumption
 363  *       that the writeback is generated due to memory pressure.
 364  */
 365 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 366 {
 367         struct backing_dev_info *bdi = mapping->backing_dev_info;
 368         struct inode *inode = mapping->host;
 369         int err;
 370
 371         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 372
 373         err = generic_writepages(mapping, wbc);
 374         if (err)
 375                 return err;
 376         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 377                 if (wbc->nonblocking)
 378                         return 0;
 379                 nfs_wait_on_write_congestion(mapping, 0);
 380         }
 381         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 382         if (err < 0)
 383                 goto out;
 384         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 385         wbc->nr_to_write -= err;
 386         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 387                 err = nfs_wait_on_requests(inode, 0, 0);
 388                 if (err < 0)
 389                         goto out;
 390         }
 391         err = nfs_commit_inode(inode, wb_priority(wbc));
 392         if (err > 0) {
 393                 wbc->nr_to_write -= err;
 394                 err = 0;
 395         }
 396 out:
 397         clear_bit(BDI_write_congested, &bdi->state);
 398         wake_up_all(&nfs_write_congestion);
 399         writeback_congestion_end();
 400         return err;
 401 }
 402
 403 /*
 404  * Insert a write request into an inode
 405  */
 406 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 407 {
 408         struct nfs_inode *nfsi = NFS_I(inode);
 409         int error;
 410
 411         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 412         BUG_ON(error == -EEXIST);
 413         if (error)
 414                 return error;
 415         if (!nfsi->npages) {
 416                 igrab(inode);
 417                 nfs_begin_data_update(inode);
 418                 if (nfs_have_delegation(inode, FMODE_WRITE))
 419                         nfsi->change_attr++;
 420         }
 421         SetPagePrivate(req->wb_page);
 422         nfsi->npages++;
 423         atomic_inc(&req->wb_count);
 424         return 0;
 425 }
 426
 427 /*
 428  * Insert a write request into an inode
 429  */
 430 static void nfs_inode_remove_request(struct nfs_page *req)
 431 {
 432         struct inode *inode = req->wb_context->dentry->d_inode;
 433         struct nfs_inode *nfsi = NFS_I(inode);
 434
 435         BUG_ON (!NFS_WBACK_BUSY(req));
 436
 437         spin_lock(&nfsi->req_lock);
 438         ClearPagePrivate(req->wb_page);
 439         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 440         nfsi->npages--;
 441         if (!nfsi->npages) {
 442                 spin_unlock(&nfsi->req_lock);
 443                 nfs_end_data_update(inode);
 444                 iput(inode);
 445         } else
 446                 spin_unlock(&nfsi->req_lock);
 447         nfs_clear_request(req);
 448         nfs_release_request(req);
 449 }
 450
 451 /*
 452  * Find a request
 453  */
 454 static inline struct nfs_page *
 455 _nfs_find_request(struct inode *inode, unsigned long index)
 456 {
 457         struct nfs_inode *nfsi = NFS_I(inode);
 458         struct nfs_page *req;
 459
 460         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 461         if (req)
 462                 atomic_inc(&req->wb_count);
 463         return req;
 464 }
 465
 466 static struct nfs_page *
 467 nfs_find_request(struct inode *inode, unsigned long index)
 468 {
 469         struct nfs_page         *req;
 470         struct nfs_inode        *nfsi = NFS_I(inode);
 471
 472         spin_lock(&nfsi->req_lock);
 473         req = _nfs_find_request(inode, index);
 474         spin_unlock(&nfsi->req_lock);
 475         return req;
 476 }
 477
 478 /*
 479  * Add a request to the inode's dirty list.
 480  */
 481 static void
 482 nfs_mark_request_dirty(struct nfs_page *req)
 483 {
 484         struct inode *inode = req->wb_context->dentry->d_inode;
 485         struct nfs_inode *nfsi = NFS_I(inode);
 486
 487         spin_lock(&nfsi->req_lock);
 488         radix_tree_tag_set(&nfsi->nfs_page_tree,
 489                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 490         nfs_list_add_request(req, &nfsi->dirty);
 491         nfsi->ndirty++;
 492         spin_unlock(&nfsi->req_lock);
 493         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 494         mark_inode_dirty(inode);
 495 }
 496
 497 /*
 498  * Check if a request is dirty
 499  */
 500 static inline int
 501 nfs_dirty_request(struct nfs_page *req)
 502 {
 503         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 504         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 505 }
 506
 507 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 508 /*
 509  * Add a request to the inode's commit list.
 510  */
 511 static void
 512 nfs_mark_request_commit(struct nfs_page *req)
 513 {
 514         struct inode *inode = req->wb_context->dentry->d_inode;
 515         struct nfs_inode *nfsi = NFS_I(inode);
 516
 517         spin_lock(&nfsi->req_lock);
 518         nfs_list_add_request(req, &nfsi->commit);
 519         nfsi->ncommit++;
 520         spin_unlock(&nfsi->req_lock);
 521         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 522         mark_inode_dirty(inode);
 523 }
 524 #endif
 525
 526 /*
 527  * Wait for a request to complete.
 528  *
 529  * Interruptible by signals only if mounted with intr flag.
 530  */
 531 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 532 {
 533         struct nfs_inode *nfsi = NFS_I(inode);
 534         struct nfs_page *req;
 535         unsigned long           idx_end, next;
 536         unsigned int            res = 0;
 537         int                     error;
 538
 539         if (npages == 0)
 540                 idx_end = ~0;
 541         else
 542                 idx_end = idx_start + npages - 1;
 543
 544         next = idx_start;
 545         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 546                 if (req->wb_index > idx_end)
 547                         break;
 548
 549                 next = req->wb_index + 1;
 550                 BUG_ON(!NFS_WBACK_BUSY(req));
 551
 552                 atomic_inc(&req->wb_count);
 553                 spin_unlock(&nfsi->req_lock);
 554                 error = nfs_wait_on_request(req);
 555                 nfs_release_request(req);
 556                 spin_lock(&nfsi->req_lock);
 557                 if (error < 0)
 558                         return error;
 559                 res++;
 560         }
 561         return res;
 562 }
 563
 564 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 565 {
 566         struct nfs_inode *nfsi = NFS_I(inode);
 567         int ret;
 568
 569         spin_lock(&nfsi->req_lock);
 570         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 571         spin_unlock(&nfsi->req_lock);
 572         return ret;
 573 }
 574
 575 static void nfs_cancel_dirty_list(struct list_head *head)
 576 {
 577         struct nfs_page *req;
 578         while(!list_empty(head)) {
 579                 req = nfs_list_entry(head->next);
 580                 nfs_list_remove_request(req);
 581                 nfs_inode_remove_request(req);
 582                 nfs_clear_page_writeback(req);
 583         }
 584 }
 585
 586 static void nfs_cancel_commit_list(struct list_head *head)
 587 {
 588         struct nfs_page *req;
 589
 590         while(!list_empty(head)) {
 591                 req = nfs_list_entry(head->next);
 592                 nfs_list_remove_request(req);
 593                 nfs_inode_remove_request(req);
 594                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 595                 nfs_clear_page_writeback(req);
 596         }
 597 }
 598
 599 /*
 600  * nfs_scan_dirty - Scan an inode for dirty requests
 601  * @inode: NFS inode to scan
 602  * @dst: destination list
 603  * @idx_start: lower bound of page->index to scan.
 604  * @npages: idx_start + npages sets the upper bound to scan.
 605  *
 606  * Moves requests from the inode's dirty page list.
 607  * The requests are *not* checked to ensure that they form a contiguous set.
 608  */
 609 static int
 610 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 611 {
 612         struct nfs_inode *nfsi = NFS_I(inode);
 613         int res = 0;
 614
 615         if (nfsi->ndirty != 0) {
 616                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
 617                 nfsi->ndirty -= res;
 618                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
 619                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
 620         }
 621         return res;
 622 }
 623
 624 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 625 /*
 626  * nfs_scan_commit - Scan an inode for commit requests
 627  * @inode: NFS inode to scan
 628  * @dst: destination list
 629  * @idx_start: lower bound of page->index to scan.
 630  * @npages: idx_start + npages sets the upper bound to scan.
 631  *
 632  * Moves requests from the inode's 'commit' request list.
 633  * The requests are *not* checked to ensure that they form a contiguous set.
 634  */
 635 static int
 636 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 637 {
 638         struct nfs_inode *nfsi = NFS_I(inode);
 639         int res = 0;
 640
 641         if (nfsi->ncommit != 0) {
 642                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 643                 nfsi->ncommit -= res;
 644                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 645                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 646         }
 647         return res;
 648 }
 649 #else
 650 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 651 {
 652         return 0;
 653 }
 654 #endif
 655
 656 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 657 {
 658         struct backing_dev_info *bdi = mapping->backing_dev_info;
 659         DEFINE_WAIT(wait);
 660         int ret = 0;
 661
 662         might_sleep();
 663
 664         if (!bdi_write_congested(bdi))
 665                 return 0;
 666
 667         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 668
 669         if (intr) {
 670                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 671                 sigset_t oldset;
 672
 673                 rpc_clnt_sigmask(clnt, &oldset);
 674                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 675                 if (bdi_write_congested(bdi)) {
 676                         if (signalled())
 677                                 ret = -ERESTARTSYS;
 678                         else
 679                                 schedule();
 680                 }
 681                 rpc_clnt_sigunmask(clnt, &oldset);
 682         } else {
 683                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 684                 if (bdi_write_congested(bdi))
 685                         schedule();
 686         }
 687         finish_wait(&nfs_write_congestion, &wait);
 688         return ret;
 689 }
 690
 691
 692 /*
 693  * Try to update any existing write request, or create one if there is none.
 694  * In order to match, the request's credentials must match those of
 695  * the calling process.
 696  *
 697  * Note: Should always be called with the Page Lock held!
 698  */
 699 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 700                 struct inode *inode, struct page *page,
 701                 unsigned int offset, unsigned int bytes)
 702 {
 703         struct nfs_server *server = NFS_SERVER(inode);
 704         struct nfs_inode *nfsi = NFS_I(inode);
 705         struct nfs_page         *req, *new = NULL;
 706         unsigned long           rqend, end;
 707
 708         end = offset + bytes;
 709
 710         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 711                 return ERR_PTR(-ERESTARTSYS);
 712         for (;;) {
 713                 /* Loop over all inode entries and see if we find
 714                  * A request for the page we wish to update
 715                  */
 716                 spin_lock(&nfsi->req_lock);
 717                 req = _nfs_find_request(inode, page->index);
 718                 if (req) {
 719                         if (!nfs_lock_request_dontget(req)) {
 720                                 int error;
 721                                 spin_unlock(&nfsi->req_lock);
 722                                 error = nfs_wait_on_request(req);
 723                                 nfs_release_request(req);
 724                                 if (error < 0) {
 725                                         if (new)
 726                                                 nfs_release_request(new);
 727                                         return ERR_PTR(error);
 728                                 }
 729                                 continue;
 730                         }
 731                         spin_unlock(&nfsi->req_lock);
 732                         if (new)
 733                                 nfs_release_request(new);
 734                         break;
 735                 }
 736
 737                 if (new) {
 738                         int error;
 739                         nfs_lock_request_dontget(new);
 740                         error = nfs_inode_add_request(inode, new);
 741                         if (error) {
 742                                 spin_unlock(&nfsi->req_lock);
 743                                 nfs_unlock_request(new);
 744                                 return ERR_PTR(error);
 745                         }
 746                         spin_unlock(&nfsi->req_lock);
 747                         nfs_mark_request_dirty(new);
 748                         return new;
 749                 }
 750                 spin_unlock(&nfsi->req_lock);
 751
 752                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 753                 if (IS_ERR(new))
 754                         return new;
 755         }
 756
 757         /* We have a request for our page.
 758          * If the creds don't match, or the
 759          * page addresses don't match,
 760          * tell the caller to wait on the conflicting
 761          * request.
 762          */
 763         rqend = req->wb_offset + req->wb_bytes;
 764         if (req->wb_context != ctx
 765             || req->wb_page != page
 766             || !nfs_dirty_request(req)
 767             || offset > rqend || end < req->wb_offset) {
 768                 nfs_unlock_request(req);
 769                 return ERR_PTR(-EBUSY);
 770         }
 771
 772         /* Okay, the request matches. Update the region */
 773         if (offset < req->wb_offset) {
 774                 req->wb_offset = offset;
 775                 req->wb_pgbase = offset;
 776                 req->wb_bytes = rqend - req->wb_offset;
 777         }
 778
 779         if (end > rqend)
 780                 req->wb_bytes = end - req->wb_offset;
 781
 782         return req;
 783 }
 784
 785 int nfs_flush_incompatible(struct file *file, struct page *page)
 786 {
 787         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 788         struct inode    *inode = page->mapping->host;
 789         struct nfs_page *req;
 790         int             status = 0;
 791         /*
 792          * Look for a request corresponding to this page. If there
 793          * is one, and it belongs to another file, we flush it out
 794          * before we try to copy anything into the page. Do this
 795          * due to the lack of an ACCESS-type call in NFSv2.
 796          * Also do the same if we find a request from an existing
 797          * dropped page.
 798          */
 799         req = nfs_find_request(inode, page->index);
 800         if (req) {
 801                 if (req->wb_page != page || ctx != req->wb_context)
 802                         status = nfs_wb_page(inode, page);
 803                 nfs_release_request(req);
 804         }
 805         return (status < 0) ? status : 0;
 806 }
 807
 808 /*
 809  * Update and possibly write a cached page of an NFS file.
 810  *
 811  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 812  * things with a page scheduled for an RPC call (e.g. invalidate it).
 813  */
 814 int nfs_updatepage(struct file *file, struct page *page,
 815                 unsigned int offset, unsigned int count)
 816 {
 817         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 818         struct inode    *inode = page->mapping->host;
 819         struct nfs_page *req;
 820         int             status = 0;
 821
 822         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 823
 824         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 825                 file->f_dentry->d_parent->d_name.name,
 826                 file->f_dentry->d_name.name, count,
 827                 (long long)(page_offset(page) +offset));
 828
 829         if (IS_SYNC(inode)) {
 830                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 831                 if (status > 0) {
 832                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 833                                 SetPageUptodate(page);
 834                         return 0;
 835                 }
 836                 return status;
 837         }
 838
 839         /* If we're not using byte range locks, and we know the page
 840          * is entirely in cache, it may be more efficient to avoid
 841          * fragmenting write requests.
 842          */
 843         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 844                 loff_t end_offs = i_size_read(inode) - 1;
 845                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 846
 847                 count += offset;
 848                 offset = 0;
 849                 if (unlikely(end_offs < 0)) {
 850                         /* Do nothing */
 851                 } else if (page->index == end_index) {
 852                         unsigned int pglen;
 853                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 854                         if (count < pglen)
 855                                 count = pglen;
 856                 } else if (page->index < end_index)
 857                         count = PAGE_CACHE_SIZE;
 858         }
 859
 860         /*
 861          * Try to find an NFS request corresponding to this page
 862          * and update it.
 863          * If the existing request cannot be updated, we must flush
 864          * it out now.
 865          */
 866         do {
 867                 req = nfs_update_request(ctx, inode, page, offset, count);
 868                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 869                 if (status != -EBUSY)
 870                         break;
 871                 /* Request could not be updated. Flush it out and try again */
 872                 status = nfs_wb_page(inode, page);
 873         } while (status >= 0);
 874         if (status < 0)
 875                 goto done;
 876
 877         status = 0;
 878
 879         /* Update file length */
 880         nfs_grow_file(page, offset, count);
 881         /* Set the PG_uptodate flag? */
 882         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 883         nfs_unlock_request(req);
 884 done:
 885         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 886                         status, (long long)i_size_read(inode));
 887         if (status < 0)
 888                 ClearPageUptodate(page);
 889         return status;
 890 }
 891
 892 static void nfs_writepage_release(struct nfs_page *req)
 893 {
 894         end_page_writeback(req->wb_page);
 895
 896 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 897         if (!PageError(req->wb_page)) {
 898                 if (NFS_NEED_RESCHED(req)) {
 899                         nfs_mark_request_dirty(req);
 900                         goto out;
 901                 } else if (NFS_NEED_COMMIT(req)) {
 902                         nfs_mark_request_commit(req);
 903                         goto out;
 904                 }
 905         }
 906         nfs_inode_remove_request(req);
 907
 908 out:
 909         nfs_clear_commit(req);
 910         nfs_clear_reschedule(req);
 911 #else
 912         nfs_inode_remove_request(req);
 913 #endif
 914         nfs_clear_page_writeback(req);
 915 }
 916
 917 static inline int flush_task_priority(int how)
 918 {
 919         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 920                 case FLUSH_HIGHPRI:
 921                         return RPC_PRIORITY_HIGH;
 922                 case FLUSH_LOWPRI:
 923                         return RPC_PRIORITY_LOW;
 924         }
 925         return RPC_PRIORITY_NORMAL;
 926 }
 927
 928 /*
 929  * Set up the argument/result storage required for the RPC call.
 930  */
 931 static void nfs_write_rpcsetup(struct nfs_page *req,
 932                 struct nfs_write_data *data,
 933                 const struct rpc_call_ops *call_ops,
 934                 unsigned int count, unsigned int offset,
 935                 int how)
 936 {
 937         struct inode            *inode;
 938         int flags;
 939
 940         /* Set up the RPC argument and reply structs
 941          * NB: take care not to mess about with data->commit et al. */
 942
 943         data->req = req;
 944         data->inode = inode = req->wb_context->dentry->d_inode;
 945         data->cred = req->wb_context->cred;
 946
 947         data->args.fh     = NFS_FH(inode);
 948         data->args.offset = req_offset(req) + offset;
 949         data->args.pgbase = req->wb_pgbase + offset;
 950         data->args.pages  = data->pagevec;
 951         data->args.count  = count;
 952         data->args.context = req->wb_context;
 953
 954         data->res.fattr   = &data->fattr;
 955         data->res.count   = count;
 956         data->res.verf    = &data->verf;
 957         nfs_fattr_init(&data->fattr);
 958
 959         /* Set up the initial task struct.  */
 960         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 961         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 962         NFS_PROTO(inode)->write_setup(data, how);
 963
 964         data->task.tk_priority = flush_task_priority(how);
 965         data->task.tk_cookie = (unsigned long)inode;
 966
 967         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 968                 data->task.tk_pid,
 969                 inode->i_sb->s_id,
 970                 (long long)NFS_FILEID(inode),
 971                 count,
 972                 (unsigned long long)data->args.offset);
 973 }
 974
 975 static void nfs_execute_write(struct nfs_write_data *data)
 976 {
 977         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 978         sigset_t oldset;
 979
 980         rpc_clnt_sigmask(clnt, &oldset);
 981         lock_kernel();
 982         rpc_execute(&data->task);
 983         unlock_kernel();
 984         rpc_clnt_sigunmask(clnt, &oldset);
 985 }
 986
 987 /*
 988  * Generate multiple small requests to write out a single
 989  * contiguous dirty area on one page.
 990  */
 991 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 992 {
 993         struct nfs_page *req = nfs_list_entry(head->next);
 994         struct page *page = req->wb_page;
 995         struct nfs_write_data *data;
 996         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 997         unsigned int offset;
 998         int requests = 0;
 999         LIST_HEAD(list);
1000
1001         nfs_list_remove_request(req);
1002
1003         nbytes = req->wb_bytes;
1004         do {
1005                 size_t len = min(nbytes, wsize);
1006
1007                 data = nfs_writedata_alloc(len);
1008                 if (!data)
1009                         goto out_bad;
1010                 list_add(&data->pages, &list);
1011                 requests++;
1012                 nbytes -= len;
1013         } while (nbytes != 0);
1014         atomic_set(&req->wb_complete, requests);
1015
1016         ClearPageError(page);
1017         set_page_writeback(page);
1018         offset = 0;
1019         nbytes = req->wb_bytes;
1020         do {
1021                 data = list_entry(list.next, struct nfs_write_data, pages);
1022                 list_del_init(&data->pages);
1023
1024                 data->pagevec[0] = page;
1025
1026                 if (nbytes > wsize) {
1027                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1028                                         wsize, offset, how);
1029                         offset += wsize;
1030                         nbytes -= wsize;
1031                 } else {
1032                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1033                                         nbytes, offset, how);
1034                         nbytes = 0;
1035                 }
1036                 nfs_execute_write(data);
1037         } while (nbytes != 0);
1038
1039         return 0;
1040
1041 out_bad:
1042         while (!list_empty(&list)) {
1043                 data = list_entry(list.next, struct nfs_write_data, pages);
1044                 list_del(&data->pages);
1045                 nfs_writedata_free(data);
1046         }
1047         nfs_mark_request_dirty(req);
1048         nfs_clear_page_writeback(req);
1049         return -ENOMEM;
1050 }
1051
1052 /*
1053  * Create an RPC task for the given write request and kick it.
1054  * The page must have been locked by the caller.
1055  *
1056  * It may happen that the page we're passed is not marked dirty.
1057  * This is the case if nfs_updatepage detects a conflicting request
1058  * that has been written but not committed.
1059  */
1060 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1061 {
1062         struct nfs_page         *req;
1063         struct page             **pages;
1064         struct nfs_write_data   *data;
1065         unsigned int            count;
1066
1067         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
1068         if (!data)
1069                 goto out_bad;
1070
1071         pages = data->pagevec;
1072         count = 0;
1073         while (!list_empty(head)) {
1074                 req = nfs_list_entry(head->next);
1075                 nfs_list_remove_request(req);
1076                 nfs_list_add_request(req, &data->pages);
1077                 ClearPageError(req->wb_page);
1078                 set_page_writeback(req->wb_page);
1079                 *pages++ = req->wb_page;
1080                 count += req->wb_bytes;
1081         }
1082         req = nfs_list_entry(data->pages.next);
1083
1084         /* Set up the argument struct */
1085         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1086
1087         nfs_execute_write(data);
1088         return 0;
1089  out_bad:
1090         while (!list_empty(head)) {
1091                 struct nfs_page *req = nfs_list_entry(head->next);
1092                 nfs_list_remove_request(req);
1093                 nfs_mark_request_dirty(req);
1094                 nfs_clear_page_writeback(req);
1095         }
1096         return -ENOMEM;
1097 }
1098
1099 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1100 {
1101         LIST_HEAD(one_request);
1102         int (*flush_one)(struct inode *, struct list_head *, int);
1103         struct nfs_page *req;
1104         int wpages = NFS_SERVER(inode)->wpages;
1105         int wsize = NFS_SERVER(inode)->wsize;
1106         int error;
1107
1108         flush_one = nfs_flush_one;
1109         if (wsize < PAGE_CACHE_SIZE)
1110                 flush_one = nfs_flush_multi;
1111         /* For single writes, FLUSH_STABLE is more efficient */
1112         if (npages <= wpages && npages == NFS_I(inode)->npages
1113                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1114                 how |= FLUSH_STABLE;
1115
1116         do {
1117                 nfs_coalesce_requests(head, &one_request, wpages);
1118                 req = nfs_list_entry(one_request.next);
1119                 error = flush_one(inode, &one_request, how);
1120                 if (error < 0)
1121                         goto out_err;
1122         } while (!list_empty(head));
1123         return 0;
1124 out_err:
1125         while (!list_empty(head)) {
1126                 req = nfs_list_entry(head->next);
1127                 nfs_list_remove_request(req);
1128                 nfs_mark_request_dirty(req);
1129                 nfs_clear_page_writeback(req);
1130         }
1131         return error;
1132 }
1133
1134 /*
1135  * Handle a write reply that flushed part of a page.
1136  */
1137 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1138 {
1139         struct nfs_write_data   *data = calldata;
1140         struct nfs_page         *req = data->req;
1141         struct page             *page = req->wb_page;
1142
1143         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1144                 req->wb_context->dentry->d_inode->i_sb->s_id,
1145                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1146                 req->wb_bytes,
1147                 (long long)req_offset(req));
1148
1149         if (nfs_writeback_done(task, data) != 0)
1150                 return;
1151
1152         if (task->tk_status < 0) {
1153                 ClearPageUptodate(page);
1154                 SetPageError(page);
1155                 req->wb_context->error = task->tk_status;
1156                 dprintk(", error = %d\n", task->tk_status);
1157         } else {
1158 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1159                 if (data->verf.committed < NFS_FILE_SYNC) {
1160                         if (!NFS_NEED_COMMIT(req)) {
1161                                 nfs_defer_commit(req);
1162                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1163                                 dprintk(" defer commit\n");
1164                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1165                                 nfs_defer_reschedule(req);
1166                                 dprintk(" server reboot detected\n");
1167                         }
1168                 } else
1169 #endif
1170                         dprintk(" OK\n");
1171         }
1172
1173         if (atomic_dec_and_test(&req->wb_complete))
1174                 nfs_writepage_release(req);
1175 }
1176
1177 static const struct rpc_call_ops nfs_write_partial_ops = {
1178         .rpc_call_done = nfs_writeback_done_partial,
1179         .rpc_release = nfs_writedata_release,
1180 };
1181
1182 /*
1183  * Handle a write reply that flushes a whole page.
1184  *
1185  * FIXME: There is an inherent race with invalidate_inode_pages and
1186  *        writebacks since the page->count is kept > 1 for as long
1187  *        as the page has a write request pending.
1188  */
1189 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1190 {
1191         struct nfs_write_data   *data = calldata;
1192         struct nfs_page         *req;
1193         struct page             *page;
1194
1195         if (nfs_writeback_done(task, data) != 0)
1196                 return;
1197
1198         /* Update attributes as result of writeback. */
1199         while (!list_empty(&data->pages)) {
1200                 req = nfs_list_entry(data->pages.next);
1201                 nfs_list_remove_request(req);
1202                 page = req->wb_page;
1203
1204                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1205                         req->wb_context->dentry->d_inode->i_sb->s_id,
1206                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1207                         req->wb_bytes,
1208                         (long long)req_offset(req));
1209
1210                 if (task->tk_status < 0) {
1211                         ClearPageUptodate(page);
1212                         SetPageError(page);
1213                         req->wb_context->error = task->tk_status;
1214                         end_page_writeback(page);
1215                         nfs_inode_remove_request(req);
1216                         dprintk(", error = %d\n", task->tk_status);
1217                         goto next;
1218                 }
1219                 end_page_writeback(page);
1220
1221 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1222                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1223                         nfs_inode_remove_request(req);
1224                         dprintk(" OK\n");
1225                         goto next;
1226                 }
1227                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1228                 nfs_mark_request_commit(req);
1229                 dprintk(" marked for commit\n");
1230 #else
1231                 nfs_inode_remove_request(req);
1232 #endif
1233         next:
1234                 nfs_clear_page_writeback(req);
1235         }
1236 }
1237
1238 static const struct rpc_call_ops nfs_write_full_ops = {
1239         .rpc_call_done = nfs_writeback_done_full,
1240         .rpc_release = nfs_writedata_release,
1241 };
1242
1243
1244 /*
1245  * This function is called when the WRITE call is complete.
1246  */
1247 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1248 {
1249         struct nfs_writeargs    *argp = &data->args;
1250         struct nfs_writeres     *resp = &data->res;
1251         int status;
1252
1253         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1254                 task->tk_pid, task->tk_status);
1255
1256         /*
1257          * ->write_done will attempt to use post-op attributes to detect
1258          * conflicting writes by other clients.  A strict interpretation
1259          * of close-to-open would allow us to continue caching even if
1260          * another writer had changed the file, but some applications
1261          * depend on tighter cache coherency when writing.
1262          */
1263         status = NFS_PROTO(data->inode)->write_done(task, data);
1264         if (status != 0)
1265                 return status;
1266         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1267
1268 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1269         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1270                 /* We tried a write call, but the server did not
1271                  * commit data to stable storage even though we
1272                  * requested it.
1273                  * Note: There is a known bug in Tru64 < 5.0 in which
1274                  *       the server reports NFS_DATA_SYNC, but performs
1275                  *       NFS_FILE_SYNC. We therefore implement this checking
1276                  *       as a dprintk() in order to avoid filling syslog.
1277                  */
1278                 static unsigned long    complain;
1279
1280                 if (time_before(complain, jiffies)) {
1281                         dprintk("NFS: faulty NFS server %s:"
1282                                 " (committed = %d) != (stable = %d)\n",
1283                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1284                                 resp->verf->committed, argp->stable);
1285                         complain = jiffies + 300 * HZ;
1286                 }
1287         }
1288 #endif
1289         /* Is this a short write? */
1290         if (task->tk_status >= 0 && resp->count < argp->count) {
1291                 static unsigned long    complain;
1292
1293                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1294
1295                 /* Has the server at least made some progress? */
1296                 if (resp->count != 0) {
1297                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1298                         if (resp->verf->committed != NFS_UNSTABLE) {
1299                                 /* Resend from where the server left off */
1300                                 argp->offset += resp->count;
1301                                 argp->pgbase += resp->count;
1302                                 argp->count -= resp->count;
1303                         } else {
1304                                 /* Resend as a stable write in order to avoid
1305                                  * headaches in the case of a server crash.
1306                                  */
1307                                 argp->stable = NFS_FILE_SYNC;
1308                         }
1309                         rpc_restart_call(task);
1310                         return -EAGAIN;
1311                 }
1312                 if (time_before(complain, jiffies)) {
1313                         printk(KERN_WARNING
1314                                "NFS: Server wrote zero bytes, expected %u.\n",
1315                                         argp->count);
1316                         complain = jiffies + 300 * HZ;
1317                 }
1318                 /* Can't do anything about it except throw an error. */
1319                 task->tk_status = -EIO;
1320         }
1321         return 0;
1322 }
1323
1324
1325 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1326 void nfs_commit_release(void *wdata)
1327 {
1328         nfs_commit_free(wdata);
1329 }
1330
1331 /*
1332  * Set up the argument/result storage required for the RPC call.
1333  */
1334 static void nfs_commit_rpcsetup(struct list_head *head,
1335                 struct nfs_write_data *data,
1336                 int how)
1337 {
1338         struct nfs_page         *first;
1339         struct inode            *inode;
1340         int flags;
1341
1342         /* Set up the RPC argument and reply structs
1343          * NB: take care not to mess about with data->commit et al. */
1344
1345         list_splice_init(head, &data->pages);
1346         first = nfs_list_entry(data->pages.next);
1347         inode = first->wb_context->dentry->d_inode;
1348
1349         data->inode       = inode;
1350         data->cred        = first->wb_context->cred;
1351
1352         data->args.fh     = NFS_FH(data->inode);
1353         /* Note: we always request a commit of the entire inode */
1354         data->args.offset = 0;
1355         data->args.count  = 0;
1356         data->res.count   = 0;
1357         data->res.fattr   = &data->fattr;
1358         data->res.verf    = &data->verf;
1359         nfs_fattr_init(&data->fattr);
1360
1361         /* Set up the initial task struct.  */
1362         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1363         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1364         NFS_PROTO(inode)->commit_setup(data, how);
1365
1366         data->task.tk_priority = flush_task_priority(how);
1367         data->task.tk_cookie = (unsigned long)inode;
1368
1369         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1370 }
1371
1372 /*
1373  * Commit dirty pages
1374  */
1375 static int
1376 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1377 {
1378         struct nfs_write_data   *data;
1379         struct nfs_page         *req;
1380
1381         data = nfs_commit_alloc();
1382
1383         if (!data)
1384                 goto out_bad;
1385
1386         /* Set up the argument struct */
1387         nfs_commit_rpcsetup(head, data, how);
1388
1389         nfs_execute_write(data);
1390         return 0;
1391  out_bad:
1392         while (!list_empty(head)) {
1393                 req = nfs_list_entry(head->next);
1394                 nfs_list_remove_request(req);
1395                 nfs_mark_request_commit(req);
1396                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1397                 nfs_clear_page_writeback(req);
1398         }
1399         return -ENOMEM;
1400 }
1401
1402 /*
1403  * COMMIT call returned
1404  */
1405 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1406 {
1407         struct nfs_write_data   *data = calldata;
1408         struct nfs_page         *req;
1409
1410         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1411                                 task->tk_pid, task->tk_status);
1412
1413         /* Call the NFS version-specific code */
1414         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1415                 return;
1416
1417         while (!list_empty(&data->pages)) {
1418                 req = nfs_list_entry(data->pages.next);
1419                 nfs_list_remove_request(req);
1420                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1421
1422                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1423                         req->wb_context->dentry->d_inode->i_sb->s_id,
1424                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1425                         req->wb_bytes,
1426                         (long long)req_offset(req));
1427                 if (task->tk_status < 0) {
1428                         req->wb_context->error = task->tk_status;
1429                         nfs_inode_remove_request(req);
1430                         dprintk(", error = %d\n", task->tk_status);
1431                         goto next;
1432                 }
1433
1434                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1435                  * returned by the server against all stored verfs. */
1436                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1437                         /* We have a match */
1438                         nfs_inode_remove_request(req);
1439                         dprintk(" OK\n");
1440                         goto next;
1441                 }
1442                 /* We have a mismatch. Write the page again */
1443                 dprintk(" mismatch\n");
1444                 nfs_mark_request_dirty(req);
1445         next:
1446                 nfs_clear_page_writeback(req);
1447         }
1448 }
1449
1450 static const struct rpc_call_ops nfs_commit_ops = {
1451         .rpc_call_done = nfs_commit_done,
1452         .rpc_release = nfs_commit_release,
1453 };
1454 #else
1455 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1456 {
1457         return 0;
1458 }
1459 #endif
1460
1461 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1462                            unsigned int npages, int how)
1463 {
1464         struct nfs_inode *nfsi = NFS_I(inode);
1465         LIST_HEAD(head);
1466         int res;
1467
1468         spin_lock(&nfsi->req_lock);
1469         res = nfs_scan_dirty(inode, &head, idx_start, npages);
1470         spin_unlock(&nfsi->req_lock);
1471         if (res) {
1472                 int error = nfs_flush_list(inode, &head, res, how);
1473                 if (error < 0)
1474                         return error;
1475         }
1476         return res;
1477 }
1478
1479 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1480 int nfs_commit_inode(struct inode *inode, int how)
1481 {
1482         struct nfs_inode *nfsi = NFS_I(inode);
1483         LIST_HEAD(head);
1484         int res;
1485
1486         spin_lock(&nfsi->req_lock);
1487         res = nfs_scan_commit(inode, &head, 0, 0);
1488         spin_unlock(&nfsi->req_lock);
1489         if (res) {
1490                 int error = nfs_commit_list(inode, &head, how);
1491                 if (error < 0)
1492                         return error;
1493         }
1494         return res;
1495 }
1496 #endif
1497
1498 int nfs_sync_inode_wait(struct inode *inode, unsigned long idx_start,
1499                 unsigned int npages, int how)
1500 {
1501         struct nfs_inode *nfsi = NFS_I(inode);
1502         LIST_HEAD(head);
1503         int nocommit = how & FLUSH_NOCOMMIT;
1504         int pages, ret;
1505
1506         how &= ~FLUSH_NOCOMMIT;
1507         spin_lock(&nfsi->req_lock);
1508         do {
1509                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1510                 if (ret != 0)
1511                         continue;
1512                 pages = nfs_scan_dirty(inode, &head, idx_start, npages);
1513                 if (pages != 0) {
1514                         spin_unlock(&nfsi->req_lock);
1515                         if (how & FLUSH_INVALIDATE)
1516                                 nfs_cancel_dirty_list(&head);
1517                         else
1518                                 ret = nfs_flush_list(inode, &head, pages, how);
1519                         spin_lock(&nfsi->req_lock);
1520                         continue;
1521                 }
1522                 if (nocommit)
1523                         break;
1524                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1525                 if (pages == 0)
1526                         break;
1527                 if (how & FLUSH_INVALIDATE) {
1528                         spin_unlock(&nfsi->req_lock);
1529                         nfs_cancel_commit_list(&head);
1530                         spin_lock(&nfsi->req_lock);
1531                         continue;
1532                 }
1533                 pages += nfs_scan_commit(inode, &head, 0, 0);
1534                 spin_unlock(&nfsi->req_lock);
1535                 ret = nfs_commit_list(inode, &head, how);
1536                 spin_lock(&nfsi->req_lock);
1537         } while (ret >= 0);
1538         spin_unlock(&nfsi->req_lock);
1539         return ret;
1540 }
1541
1542 int __init nfs_init_writepagecache(void)
1543 {
1544         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1545                                              sizeof(struct nfs_write_data),
1546                                              0, SLAB_HWCACHE_ALIGN,
1547                                              NULL, NULL);
1548         if (nfs_wdata_cachep == NULL)
1549                 return -ENOMEM;
1550
1551         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1552                                                      nfs_wdata_cachep);
1553         if (nfs_wdata_mempool == NULL)
1554                 return -ENOMEM;
1555
1556         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1557                                                       nfs_wdata_cachep);
1558         if (nfs_commit_mempool == NULL)
1559                 return -ENOMEM;
1560
1561         return 0;
1562 }
1563
1564 void nfs_destroy_writepagecache(void)
1565 {
1566         mempool_destroy(nfs_commit_mempool);
1567         mempool_destroy(nfs_wdata_mempool);
1568         kmem_cache_destroy(nfs_wdata_cachep);
1569 }
1570