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/config.h>
  50 #include <linux/types.h>
  51 #include <linux/slab.h>
  52 #include <linux/mm.h>
  53 #include <linux/pagemap.h>
  54 #include <linux/file.h>
  55 #include <linux/mpage.h>
  56 #include <linux/writeback.h>
  57
  58 #include <linux/sunrpc/clnt.h>
  59 #include <linux/nfs_fs.h>
  60 #include <linux/nfs_mount.h>
  61 #include <linux/nfs_page.h>
  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_writeback_done(struct rpc_task *, struct nfs_write_data *);
  81 static int nfs_wait_on_write_congestion(struct address_space *, int);
  82 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
  83 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
  84                            unsigned int npages, int how);
  85 static const struct rpc_call_ops nfs_write_partial_ops;
  86 static const struct rpc_call_ops nfs_write_full_ops;
  87 static const struct rpc_call_ops nfs_commit_ops;
  88
  89 static kmem_cache_t *nfs_wdata_cachep;
  90 mempool_t *nfs_wdata_mempool;
  91 static mempool_t *nfs_commit_mempool;
  92
  93 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  94
  95 static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
  96 {
  97         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
  98
  99         if (p) {
 100                 memset(p, 0, sizeof(*p));
 101                 INIT_LIST_HEAD(&p->pages);
 102                 if (pagecount < NFS_PAGEVEC_SIZE)
 103                         p->pagevec = &p->page_array[0];
 104                 else {
 105                         size_t size = ++pagecount * sizeof(struct page *);
 106                         p->pagevec = kzalloc(size, GFP_NOFS);
 107                         if (!p->pagevec) {
 108                                 mempool_free(p, nfs_commit_mempool);
 109                                 p = NULL;
 110                         }
 111                 }
 112         }
 113         return p;
 114 }
 115
 116 static inline void nfs_commit_free(struct nfs_write_data *p)
 117 {
 118         if (p && (p->pagevec != &p->page_array[0]))
 119                 kfree(p->pagevec);
 120         mempool_free(p, nfs_commit_mempool);
 121 }
 122
 123 void nfs_writedata_release(void *wdata)
 124 {
 125         nfs_writedata_free(wdata);
 126 }
 127
 128 /* Adjust the file length if we're writing beyond the end */
 129 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 130 {
 131         struct inode *inode = page->mapping->host;
 132         loff_t end, i_size = i_size_read(inode);
 133         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 134
 135         if (i_size > 0 && page->index < end_index)
 136                 return;
 137         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 138         if (i_size >= end)
 139                 return;
 140         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 141         i_size_write(inode, end);
 142 }
 143
 144 /* We can set the PG_uptodate flag if we see that a write request
 145  * covers the full page.
 146  */
 147 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 148 {
 149         loff_t end_offs;
 150
 151         if (PageUptodate(page))
 152                 return;
 153         if (base != 0)
 154                 return;
 155         if (count == PAGE_CACHE_SIZE) {
 156                 SetPageUptodate(page);
 157                 return;
 158         }
 159
 160         end_offs = i_size_read(page->mapping->host) - 1;
 161         if (end_offs < 0)
 162                 return;
 163         /* Is this the last page? */
 164         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
 165                 return;
 166         /* This is the last page: set PG_uptodate if we cover the entire
 167          * extent of the data, then zero the rest of the page.
 168          */
 169         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
 170                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 171                 SetPageUptodate(page);
 172         }
 173 }
 174
 175 /*
 176  * Write a page synchronously.
 177  * Offset is the data offset within the page.
 178  */
 179 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 180                 struct page *page, unsigned int offset, unsigned int count,
 181                 int how)
 182 {
 183         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 184         int             result, written = 0;
 185         struct nfs_write_data *wdata;
 186
 187         wdata = nfs_writedata_alloc(1);
 188         if (!wdata)
 189                 return -ENOMEM;
 190
 191         wdata->flags = how;
 192         wdata->cred = ctx->cred;
 193         wdata->inode = inode;
 194         wdata->args.fh = NFS_FH(inode);
 195         wdata->args.context = ctx;
 196         wdata->args.pages = &page;
 197         wdata->args.stable = NFS_FILE_SYNC;
 198         wdata->args.pgbase = offset;
 199         wdata->args.count = wsize;
 200         wdata->res.fattr = &wdata->fattr;
 201         wdata->res.verf = &wdata->verf;
 202
 203         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 204                 inode->i_sb->s_id,
 205                 (long long)NFS_FILEID(inode),
 206                 count, (long long)(page_offset(page) + offset));
 207
 208         set_page_writeback(page);
 209         nfs_begin_data_update(inode);
 210         do {
 211                 if (count < wsize)
 212                         wdata->args.count = count;
 213                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 214
 215                 result = NFS_PROTO(inode)->write(wdata);
 216
 217                 if (result < 0) {
 218                         /* Must mark the page invalid after I/O error */
 219                         ClearPageUptodate(page);
 220                         goto io_error;
 221                 }
 222                 if (result < wdata->args.count)
 223                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 224                                         wdata->args.count, result);
 225
 226                 wdata->args.offset += result;
 227                 wdata->args.pgbase += result;
 228                 written += result;
 229                 count -= result;
 230                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 231         } while (count);
 232         /* Update file length */
 233         nfs_grow_file(page, offset, written);
 234         /* Set the PG_uptodate flag? */
 235         nfs_mark_uptodate(page, offset, written);
 236
 237         if (PageError(page))
 238                 ClearPageError(page);
 239
 240 io_error:
 241         nfs_end_data_update(inode);
 242         end_page_writeback(page);
 243         nfs_writedata_free(wdata);
 244         return written ? written : result;
 245 }
 246
 247 static int nfs_writepage_async(struct nfs_open_context *ctx,
 248                 struct inode *inode, struct page *page,
 249                 unsigned int offset, unsigned int count)
 250 {
 251         struct nfs_page *req;
 252
 253         req = nfs_update_request(ctx, inode, page, offset, count);
 254         if (IS_ERR(req))
 255                 return PTR_ERR(req);
 256         /* Update file length */
 257         nfs_grow_file(page, offset, count);
 258         /* Set the PG_uptodate flag? */
 259         nfs_mark_uptodate(page, offset, count);
 260         nfs_unlock_request(req);
 261         return 0;
 262 }
 263
 264 static int wb_priority(struct writeback_control *wbc)
 265 {
 266         if (wbc->for_reclaim)
 267                 return FLUSH_HIGHPRI;
 268         if (wbc->for_kupdate)
 269                 return FLUSH_LOWPRI;
 270         return 0;
 271 }
 272
 273 /*
 274  * Write an mmapped page to the server.
 275  */
 276 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 277 {
 278         struct nfs_open_context *ctx;
 279         struct inode *inode = page->mapping->host;
 280         unsigned long end_index;
 281         unsigned offset = PAGE_CACHE_SIZE;
 282         loff_t i_size = i_size_read(inode);
 283         int inode_referenced = 0;
 284         int priority = wb_priority(wbc);
 285         int err;
 286
 287         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 288         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 289
 290         /*
 291          * Note: We need to ensure that we have a reference to the inode
 292          *       if we are to do asynchronous writes. If not, waiting
 293          *       in nfs_wait_on_request() may deadlock with clear_inode().
 294          *
 295          *       If igrab() fails here, then it is in any case safe to
 296          *       call nfs_wb_page(), since there will be no pending writes.
 297          */
 298         if (igrab(inode) != 0)
 299                 inode_referenced = 1;
 300         end_index = i_size >> PAGE_CACHE_SHIFT;
 301
 302         /* Ensure we've flushed out any previous writes */
 303         nfs_wb_page_priority(inode, page, priority);
 304
 305         /* easy case */
 306         if (page->index < end_index)
 307                 goto do_it;
 308         /* things got complicated... */
 309         offset = i_size & (PAGE_CACHE_SIZE-1);
 310
 311         /* OK, are we completely out? */
 312         err = 0; /* potential race with truncate - ignore */
 313         if (page->index >= end_index+1 || !offset)
 314                 goto out;
 315 do_it:
 316         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 317         if (ctx == NULL) {
 318                 err = -EBADF;
 319                 goto out;
 320         }
 321         lock_kernel();
 322         if (!IS_SYNC(inode) && inode_referenced) {
 323                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
 324                 if (!wbc->for_writepages)
 325                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 326         } else {
 327                 err = nfs_writepage_sync(ctx, inode, page, 0,
 328                                                 offset, priority);
 329                 if (err >= 0) {
 330                         if (err != offset)
 331                                 redirty_page_for_writepage(wbc, page);
 332                         err = 0;
 333                 }
 334         }
 335         unlock_kernel();
 336         put_nfs_open_context(ctx);
 337 out:
 338         unlock_page(page);
 339         if (inode_referenced)
 340                 iput(inode);
 341         return err;
 342 }
 343
 344 /*
 345  * Note: causes nfs_update_request() to block on the assumption
 346  *       that the writeback is generated due to memory pressure.
 347  */
 348 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 349 {
 350         struct backing_dev_info *bdi = mapping->backing_dev_info;
 351         struct inode *inode = mapping->host;
 352         int err;
 353
 354         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 355
 356         err = generic_writepages(mapping, wbc);
 357         if (err)
 358                 return err;
 359         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 360                 if (wbc->nonblocking)
 361                         return 0;
 362                 nfs_wait_on_write_congestion(mapping, 0);
 363         }
 364         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 365         if (err < 0)
 366                 goto out;
 367         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 368         wbc->nr_to_write -= err;
 369         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 370                 err = nfs_wait_on_requests(inode, 0, 0);
 371                 if (err < 0)
 372                         goto out;
 373         }
 374         err = nfs_commit_inode(inode, wb_priority(wbc));
 375         if (err > 0) {
 376                 wbc->nr_to_write -= err;
 377                 err = 0;
 378         }
 379 out:
 380         clear_bit(BDI_write_congested, &bdi->state);
 381         wake_up_all(&nfs_write_congestion);
 382         return err;
 383 }
 384
 385 /*
 386  * Insert a write request into an inode
 387  */
 388 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 389 {
 390         struct nfs_inode *nfsi = NFS_I(inode);
 391         int error;
 392
 393         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 394         BUG_ON(error == -EEXIST);
 395         if (error)
 396                 return error;
 397         if (!nfsi->npages) {
 398                 igrab(inode);
 399                 nfs_begin_data_update(inode);
 400                 if (nfs_have_delegation(inode, FMODE_WRITE))
 401                         nfsi->change_attr++;
 402         }
 403         nfsi->npages++;
 404         atomic_inc(&req->wb_count);
 405         return 0;
 406 }
 407
 408 /*
 409  * Insert a write request into an inode
 410  */
 411 static void nfs_inode_remove_request(struct nfs_page *req)
 412 {
 413         struct inode *inode = req->wb_context->dentry->d_inode;
 414         struct nfs_inode *nfsi = NFS_I(inode);
 415
 416         BUG_ON (!NFS_WBACK_BUSY(req));
 417
 418         spin_lock(&nfsi->req_lock);
 419         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 420         nfsi->npages--;
 421         if (!nfsi->npages) {
 422                 spin_unlock(&nfsi->req_lock);
 423                 nfs_end_data_update(inode);
 424                 iput(inode);
 425         } else
 426                 spin_unlock(&nfsi->req_lock);
 427         nfs_clear_request(req);
 428         nfs_release_request(req);
 429 }
 430
 431 /*
 432  * Find a request
 433  */
 434 static inline struct nfs_page *
 435 _nfs_find_request(struct inode *inode, unsigned long index)
 436 {
 437         struct nfs_inode *nfsi = NFS_I(inode);
 438         struct nfs_page *req;
 439
 440         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 441         if (req)
 442                 atomic_inc(&req->wb_count);
 443         return req;
 444 }
 445
 446 static struct nfs_page *
 447 nfs_find_request(struct inode *inode, unsigned long index)
 448 {
 449         struct nfs_page         *req;
 450         struct nfs_inode        *nfsi = NFS_I(inode);
 451
 452         spin_lock(&nfsi->req_lock);
 453         req = _nfs_find_request(inode, index);
 454         spin_unlock(&nfsi->req_lock);
 455         return req;
 456 }
 457
 458 /*
 459  * Add a request to the inode's dirty list.
 460  */
 461 static void
 462 nfs_mark_request_dirty(struct nfs_page *req)
 463 {
 464         struct inode *inode = req->wb_context->dentry->d_inode;
 465         struct nfs_inode *nfsi = NFS_I(inode);
 466
 467         spin_lock(&nfsi->req_lock);
 468         radix_tree_tag_set(&nfsi->nfs_page_tree,
 469                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 470         nfs_list_add_request(req, &nfsi->dirty);
 471         nfsi->ndirty++;
 472         spin_unlock(&nfsi->req_lock);
 473         inc_page_state(nr_dirty);
 474         mark_inode_dirty(inode);
 475 }
 476
 477 /*
 478  * Check if a request is dirty
 479  */
 480 static inline int
 481 nfs_dirty_request(struct nfs_page *req)
 482 {
 483         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 484         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 485 }
 486
 487 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 488 /*
 489  * Add a request to the inode's commit list.
 490  */
 491 static void
 492 nfs_mark_request_commit(struct nfs_page *req)
 493 {
 494         struct inode *inode = req->wb_context->dentry->d_inode;
 495         struct nfs_inode *nfsi = NFS_I(inode);
 496
 497         spin_lock(&nfsi->req_lock);
 498         nfs_list_add_request(req, &nfsi->commit);
 499         nfsi->ncommit++;
 500         spin_unlock(&nfsi->req_lock);
 501         inc_page_state(nr_unstable);
 502         mark_inode_dirty(inode);
 503 }
 504 #endif
 505
 506 /*
 507  * Wait for a request to complete.
 508  *
 509  * Interruptible by signals only if mounted with intr flag.
 510  */
 511 static int
 512 nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 513 {
 514         struct nfs_inode *nfsi = NFS_I(inode);
 515         struct nfs_page *req;
 516         unsigned long           idx_end, next;
 517         unsigned int            res = 0;
 518         int                     error;
 519
 520         if (npages == 0)
 521                 idx_end = ~0;
 522         else
 523                 idx_end = idx_start + npages - 1;
 524
 525         spin_lock(&nfsi->req_lock);
 526         next = idx_start;
 527         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 528                 if (req->wb_index > idx_end)
 529                         break;
 530
 531                 next = req->wb_index + 1;
 532                 BUG_ON(!NFS_WBACK_BUSY(req));
 533
 534                 atomic_inc(&req->wb_count);
 535                 spin_unlock(&nfsi->req_lock);
 536                 error = nfs_wait_on_request(req);
 537                 nfs_release_request(req);
 538                 if (error < 0)
 539                         return error;
 540                 spin_lock(&nfsi->req_lock);
 541                 res++;
 542         }
 543         spin_unlock(&nfsi->req_lock);
 544         return res;
 545 }
 546
 547 /*
 548  * nfs_scan_dirty - Scan an inode for dirty requests
 549  * @inode: NFS inode to scan
 550  * @dst: destination list
 551  * @idx_start: lower bound of page->index to scan.
 552  * @npages: idx_start + npages sets the upper bound to scan.
 553  *
 554  * Moves requests from the inode's dirty page list.
 555  * The requests are *not* checked to ensure that they form a contiguous set.
 556  */
 557 static int
 558 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 559 {
 560         struct nfs_inode *nfsi = NFS_I(inode);
 561         int res = 0;
 562
 563         if (nfsi->ndirty != 0) {
 564                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
 565                 nfsi->ndirty -= res;
 566                 sub_page_state(nr_dirty,res);
 567                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
 568                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
 569         }
 570         return res;
 571 }
 572
 573 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 574 /*
 575  * nfs_scan_commit - Scan an inode for commit requests
 576  * @inode: NFS inode to scan
 577  * @dst: destination list
 578  * @idx_start: lower bound of page->index to scan.
 579  * @npages: idx_start + npages sets the upper bound to scan.
 580  *
 581  * Moves requests from the inode's 'commit' request list.
 582  * The requests are *not* checked to ensure that they form a contiguous set.
 583  */
 584 static int
 585 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 586 {
 587         struct nfs_inode *nfsi = NFS_I(inode);
 588         int res = 0;
 589
 590         if (nfsi->ncommit != 0) {
 591                 res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages);
 592                 nfsi->ncommit -= res;
 593                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 594                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 595         }
 596         return res;
 597 }
 598 #endif
 599
 600 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 601 {
 602         struct backing_dev_info *bdi = mapping->backing_dev_info;
 603         DEFINE_WAIT(wait);
 604         int ret = 0;
 605
 606         might_sleep();
 607
 608         if (!bdi_write_congested(bdi))
 609                 return 0;
 610
 611         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 612
 613         if (intr) {
 614                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 615                 sigset_t oldset;
 616
 617                 rpc_clnt_sigmask(clnt, &oldset);
 618                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 619                 if (bdi_write_congested(bdi)) {
 620                         if (signalled())
 621                                 ret = -ERESTARTSYS;
 622                         else
 623                                 schedule();
 624                 }
 625                 rpc_clnt_sigunmask(clnt, &oldset);
 626         } else {
 627                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 628                 if (bdi_write_congested(bdi))
 629                         schedule();
 630         }
 631         finish_wait(&nfs_write_congestion, &wait);
 632         return ret;
 633 }
 634
 635
 636 /*
 637  * Try to update any existing write request, or create one if there is none.
 638  * In order to match, the request's credentials must match those of
 639  * the calling process.
 640  *
 641  * Note: Should always be called with the Page Lock held!
 642  */
 643 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 644                 struct inode *inode, struct page *page,
 645                 unsigned int offset, unsigned int bytes)
 646 {
 647         struct nfs_server *server = NFS_SERVER(inode);
 648         struct nfs_inode *nfsi = NFS_I(inode);
 649         struct nfs_page         *req, *new = NULL;
 650         unsigned long           rqend, end;
 651
 652         end = offset + bytes;
 653
 654         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 655                 return ERR_PTR(-ERESTARTSYS);
 656         for (;;) {
 657                 /* Loop over all inode entries and see if we find
 658                  * A request for the page we wish to update
 659                  */
 660                 spin_lock(&nfsi->req_lock);
 661                 req = _nfs_find_request(inode, page->index);
 662                 if (req) {
 663                         if (!nfs_lock_request_dontget(req)) {
 664                                 int error;
 665                                 spin_unlock(&nfsi->req_lock);
 666                                 error = nfs_wait_on_request(req);
 667                                 nfs_release_request(req);
 668                                 if (error < 0) {
 669                                         if (new)
 670                                                 nfs_release_request(new);
 671                                         return ERR_PTR(error);
 672                                 }
 673                                 continue;
 674                         }
 675                         spin_unlock(&nfsi->req_lock);
 676                         if (new)
 677                                 nfs_release_request(new);
 678                         break;
 679                 }
 680
 681                 if (new) {
 682                         int error;
 683                         nfs_lock_request_dontget(new);
 684                         error = nfs_inode_add_request(inode, new);
 685                         if (error) {
 686                                 spin_unlock(&nfsi->req_lock);
 687                                 nfs_unlock_request(new);
 688                                 return ERR_PTR(error);
 689                         }
 690                         spin_unlock(&nfsi->req_lock);
 691                         nfs_mark_request_dirty(new);
 692                         return new;
 693                 }
 694                 spin_unlock(&nfsi->req_lock);
 695
 696                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 697                 if (IS_ERR(new))
 698                         return new;
 699         }
 700
 701         /* We have a request for our page.
 702          * If the creds don't match, or the
 703          * page addresses don't match,
 704          * tell the caller to wait on the conflicting
 705          * request.
 706          */
 707         rqend = req->wb_offset + req->wb_bytes;
 708         if (req->wb_context != ctx
 709             || req->wb_page != page
 710             || !nfs_dirty_request(req)
 711             || offset > rqend || end < req->wb_offset) {
 712                 nfs_unlock_request(req);
 713                 return ERR_PTR(-EBUSY);
 714         }
 715
 716         /* Okay, the request matches. Update the region */
 717         if (offset < req->wb_offset) {
 718                 req->wb_offset = offset;
 719                 req->wb_pgbase = offset;
 720                 req->wb_bytes = rqend - req->wb_offset;
 721         }
 722
 723         if (end > rqend)
 724                 req->wb_bytes = end - req->wb_offset;
 725
 726         return req;
 727 }
 728
 729 int nfs_flush_incompatible(struct file *file, struct page *page)
 730 {
 731         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 732         struct inode    *inode = page->mapping->host;
 733         struct nfs_page *req;
 734         int             status = 0;
 735         /*
 736          * Look for a request corresponding to this page. If there
 737          * is one, and it belongs to another file, we flush it out
 738          * before we try to copy anything into the page. Do this
 739          * due to the lack of an ACCESS-type call in NFSv2.
 740          * Also do the same if we find a request from an existing
 741          * dropped page.
 742          */
 743         req = nfs_find_request(inode, page->index);
 744         if (req) {
 745                 if (req->wb_page != page || ctx != req->wb_context)
 746                         status = nfs_wb_page(inode, page);
 747                 nfs_release_request(req);
 748         }
 749         return (status < 0) ? status : 0;
 750 }
 751
 752 /*
 753  * Update and possibly write a cached page of an NFS file.
 754  *
 755  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 756  * things with a page scheduled for an RPC call (e.g. invalidate it).
 757  */
 758 int nfs_updatepage(struct file *file, struct page *page,
 759                 unsigned int offset, unsigned int count)
 760 {
 761         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 762         struct inode    *inode = page->mapping->host;
 763         struct nfs_page *req;
 764         int             status = 0;
 765
 766         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 767
 768         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 769                 file->f_dentry->d_parent->d_name.name,
 770                 file->f_dentry->d_name.name, count,
 771                 (long long)(page_offset(page) +offset));
 772
 773         if (IS_SYNC(inode)) {
 774                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 775                 if (status > 0) {
 776                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 777                                 SetPageUptodate(page);
 778                         return 0;
 779                 }
 780                 return status;
 781         }
 782
 783         /* If we're not using byte range locks, and we know the page
 784          * is entirely in cache, it may be more efficient to avoid
 785          * fragmenting write requests.
 786          */
 787         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 788                 loff_t end_offs = i_size_read(inode) - 1;
 789                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 790
 791                 count += offset;
 792                 offset = 0;
 793                 if (unlikely(end_offs < 0)) {
 794                         /* Do nothing */
 795                 } else if (page->index == end_index) {
 796                         unsigned int pglen;
 797                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 798                         if (count < pglen)
 799                                 count = pglen;
 800                 } else if (page->index < end_index)
 801                         count = PAGE_CACHE_SIZE;
 802         }
 803
 804         /*
 805          * Try to find an NFS request corresponding to this page
 806          * and update it.
 807          * If the existing request cannot be updated, we must flush
 808          * it out now.
 809          */
 810         do {
 811                 req = nfs_update_request(ctx, inode, page, offset, count);
 812                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 813                 if (status != -EBUSY)
 814                         break;
 815                 /* Request could not be updated. Flush it out and try again */
 816                 status = nfs_wb_page(inode, page);
 817         } while (status >= 0);
 818         if (status < 0)
 819                 goto done;
 820
 821         status = 0;
 822
 823         /* Update file length */
 824         nfs_grow_file(page, offset, count);
 825         /* Set the PG_uptodate flag? */
 826         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 827         nfs_unlock_request(req);
 828 done:
 829         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 830                         status, (long long)i_size_read(inode));
 831         if (status < 0)
 832                 ClearPageUptodate(page);
 833         return status;
 834 }
 835
 836 static void nfs_writepage_release(struct nfs_page *req)
 837 {
 838         end_page_writeback(req->wb_page);
 839
 840 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 841         if (!PageError(req->wb_page)) {
 842                 if (NFS_NEED_RESCHED(req)) {
 843                         nfs_mark_request_dirty(req);
 844                         goto out;
 845                 } else if (NFS_NEED_COMMIT(req)) {
 846                         nfs_mark_request_commit(req);
 847                         goto out;
 848                 }
 849         }
 850         nfs_inode_remove_request(req);
 851
 852 out:
 853         nfs_clear_commit(req);
 854         nfs_clear_reschedule(req);
 855 #else
 856         nfs_inode_remove_request(req);
 857 #endif
 858         nfs_clear_page_writeback(req);
 859 }
 860
 861 static inline int flush_task_priority(int how)
 862 {
 863         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 864                 case FLUSH_HIGHPRI:
 865                         return RPC_PRIORITY_HIGH;
 866                 case FLUSH_LOWPRI:
 867                         return RPC_PRIORITY_LOW;
 868         }
 869         return RPC_PRIORITY_NORMAL;
 870 }
 871
 872 /*
 873  * Set up the argument/result storage required for the RPC call.
 874  */
 875 static void nfs_write_rpcsetup(struct nfs_page *req,
 876                 struct nfs_write_data *data,
 877                 const struct rpc_call_ops *call_ops,
 878                 unsigned int count, unsigned int offset,
 879                 int how)
 880 {
 881         struct inode            *inode;
 882         int flags;
 883
 884         /* Set up the RPC argument and reply structs
 885          * NB: take care not to mess about with data->commit et al. */
 886
 887         data->req = req;
 888         data->inode = inode = req->wb_context->dentry->d_inode;
 889         data->cred = req->wb_context->cred;
 890
 891         data->args.fh     = NFS_FH(inode);
 892         data->args.offset = req_offset(req) + offset;
 893         data->args.pgbase = req->wb_pgbase + offset;
 894         data->args.pages  = data->pagevec;
 895         data->args.count  = count;
 896         data->args.context = req->wb_context;
 897
 898         data->res.fattr   = &data->fattr;
 899         data->res.count   = count;
 900         data->res.verf    = &data->verf;
 901         nfs_fattr_init(&data->fattr);
 902
 903         /* Set up the initial task struct.  */
 904         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 905         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 906         NFS_PROTO(inode)->write_setup(data, how);
 907
 908         data->task.tk_priority = flush_task_priority(how);
 909         data->task.tk_cookie = (unsigned long)inode;
 910
 911         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 912                 data->task.tk_pid,
 913                 inode->i_sb->s_id,
 914                 (long long)NFS_FILEID(inode),
 915                 count,
 916                 (unsigned long long)data->args.offset);
 917 }
 918
 919 static void nfs_execute_write(struct nfs_write_data *data)
 920 {
 921         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 922         sigset_t oldset;
 923
 924         rpc_clnt_sigmask(clnt, &oldset);
 925         lock_kernel();
 926         rpc_execute(&data->task);
 927         unlock_kernel();
 928         rpc_clnt_sigunmask(clnt, &oldset);
 929 }
 930
 931 /*
 932  * Generate multiple small requests to write out a single
 933  * contiguous dirty area on one page.
 934  */
 935 static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
 936 {
 937         struct nfs_page *req = nfs_list_entry(head->next);
 938         struct page *page = req->wb_page;
 939         struct nfs_write_data *data;
 940         unsigned int wsize = NFS_SERVER(inode)->wsize;
 941         unsigned int nbytes, offset;
 942         int requests = 0;
 943         LIST_HEAD(list);
 944
 945         nfs_list_remove_request(req);
 946
 947         nbytes = req->wb_bytes;
 948         for (;;) {
 949                 data = nfs_writedata_alloc(1);
 950                 if (!data)
 951                         goto out_bad;
 952                 list_add(&data->pages, &list);
 953                 requests++;
 954                 if (nbytes <= wsize)
 955                         break;
 956                 nbytes -= wsize;
 957         }
 958         atomic_set(&req->wb_complete, requests);
 959
 960         ClearPageError(page);
 961         set_page_writeback(page);
 962         offset = 0;
 963         nbytes = req->wb_bytes;
 964         do {
 965                 data = list_entry(list.next, struct nfs_write_data, pages);
 966                 list_del_init(&data->pages);
 967
 968                 data->pagevec[0] = page;
 969
 970                 if (nbytes > wsize) {
 971                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 972                                         wsize, offset, how);
 973                         offset += wsize;
 974                         nbytes -= wsize;
 975                 } else {
 976                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 977                                         nbytes, offset, how);
 978                         nbytes = 0;
 979                 }
 980                 nfs_execute_write(data);
 981         } while (nbytes != 0);
 982
 983         return 0;
 984
 985 out_bad:
 986         while (!list_empty(&list)) {
 987                 data = list_entry(list.next, struct nfs_write_data, pages);
 988                 list_del(&data->pages);
 989                 nfs_writedata_free(data);
 990         }
 991         nfs_mark_request_dirty(req);
 992         nfs_clear_page_writeback(req);
 993         return -ENOMEM;
 994 }
 995
 996 /*
 997  * Create an RPC task for the given write request and kick it.
 998  * The page must have been locked by the caller.
 999  *
1000  * It may happen that the page we're passed is not marked dirty.
1001  * This is the case if nfs_updatepage detects a conflicting request
1002  * that has been written but not committed.
1003  */
1004 static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
1005 {
1006         struct nfs_page         *req;
1007         struct page             **pages;
1008         struct nfs_write_data   *data;
1009         unsigned int            count;
1010
1011         if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
1012                 return nfs_flush_multi(head, inode, how);
1013
1014         data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
1015         if (!data)
1016                 goto out_bad;
1017
1018         pages = data->pagevec;
1019         count = 0;
1020         while (!list_empty(head)) {
1021                 req = nfs_list_entry(head->next);
1022                 nfs_list_remove_request(req);
1023                 nfs_list_add_request(req, &data->pages);
1024                 ClearPageError(req->wb_page);
1025                 set_page_writeback(req->wb_page);
1026                 *pages++ = req->wb_page;
1027                 count += req->wb_bytes;
1028         }
1029         req = nfs_list_entry(data->pages.next);
1030
1031         /* Set up the argument struct */
1032         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1033
1034         nfs_execute_write(data);
1035         return 0;
1036  out_bad:
1037         while (!list_empty(head)) {
1038                 struct nfs_page *req = nfs_list_entry(head->next);
1039                 nfs_list_remove_request(req);
1040                 nfs_mark_request_dirty(req);
1041                 nfs_clear_page_writeback(req);
1042         }
1043         return -ENOMEM;
1044 }
1045
1046 static int
1047 nfs_flush_list(struct list_head *head, int wpages, int how)
1048 {
1049         LIST_HEAD(one_request);
1050         struct nfs_page         *req;
1051         int                     error = 0;
1052         unsigned int            pages = 0;
1053
1054         while (!list_empty(head)) {
1055                 pages += nfs_coalesce_requests(head, &one_request, wpages);
1056                 req = nfs_list_entry(one_request.next);
1057                 error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how);
1058                 if (error < 0)
1059                         break;
1060         }
1061         if (error >= 0)
1062                 return pages;
1063
1064         while (!list_empty(head)) {
1065                 req = nfs_list_entry(head->next);
1066                 nfs_list_remove_request(req);
1067                 nfs_mark_request_dirty(req);
1068                 nfs_clear_page_writeback(req);
1069         }
1070         return error;
1071 }
1072
1073 /*
1074  * Handle a write reply that flushed part of a page.
1075  */
1076 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1077 {
1078         struct nfs_write_data   *data = calldata;
1079         struct nfs_page         *req = data->req;
1080         struct page             *page = req->wb_page;
1081
1082         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1083                 req->wb_context->dentry->d_inode->i_sb->s_id,
1084                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1085                 req->wb_bytes,
1086                 (long long)req_offset(req));
1087
1088         if (nfs_writeback_done(task, data) != 0)
1089                 return;
1090
1091         if (task->tk_status < 0) {
1092                 ClearPageUptodate(page);
1093                 SetPageError(page);
1094                 req->wb_context->error = task->tk_status;
1095                 dprintk(", error = %d\n", task->tk_status);
1096         } else {
1097 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1098                 if (data->verf.committed < NFS_FILE_SYNC) {
1099                         if (!NFS_NEED_COMMIT(req)) {
1100                                 nfs_defer_commit(req);
1101                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1102                                 dprintk(" defer commit\n");
1103                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1104                                 nfs_defer_reschedule(req);
1105                                 dprintk(" server reboot detected\n");
1106                         }
1107                 } else
1108 #endif
1109                         dprintk(" OK\n");
1110         }
1111
1112         if (atomic_dec_and_test(&req->wb_complete))
1113                 nfs_writepage_release(req);
1114 }
1115
1116 static const struct rpc_call_ops nfs_write_partial_ops = {
1117         .rpc_call_done = nfs_writeback_done_partial,
1118         .rpc_release = nfs_writedata_release,
1119 };
1120
1121 /*
1122  * Handle a write reply that flushes a whole page.
1123  *
1124  * FIXME: There is an inherent race with invalidate_inode_pages and
1125  *        writebacks since the page->count is kept > 1 for as long
1126  *        as the page has a write request pending.
1127  */
1128 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1129 {
1130         struct nfs_write_data   *data = calldata;
1131         struct nfs_page         *req;
1132         struct page             *page;
1133
1134         if (nfs_writeback_done(task, data) != 0)
1135                 return;
1136
1137         /* Update attributes as result of writeback. */
1138         while (!list_empty(&data->pages)) {
1139                 req = nfs_list_entry(data->pages.next);
1140                 nfs_list_remove_request(req);
1141                 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 (task->tk_status < 0) {
1150                         ClearPageUptodate(page);
1151                         SetPageError(page);
1152                         req->wb_context->error = task->tk_status;
1153                         end_page_writeback(page);
1154                         nfs_inode_remove_request(req);
1155                         dprintk(", error = %d\n", task->tk_status);
1156                         goto next;
1157                 }
1158                 end_page_writeback(page);
1159
1160 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1161                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1162                         nfs_inode_remove_request(req);
1163                         dprintk(" OK\n");
1164                         goto next;
1165                 }
1166                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1167                 nfs_mark_request_commit(req);
1168                 dprintk(" marked for commit\n");
1169 #else
1170                 nfs_inode_remove_request(req);
1171 #endif
1172         next:
1173                 nfs_clear_page_writeback(req);
1174         }
1175 }
1176
1177 static const struct rpc_call_ops nfs_write_full_ops = {
1178         .rpc_call_done = nfs_writeback_done_full,
1179         .rpc_release = nfs_writedata_release,
1180 };
1181
1182
1183 /*
1184  * This function is called when the WRITE call is complete.
1185  */
1186 static int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1187 {
1188         struct nfs_writeargs    *argp = &data->args;
1189         struct nfs_writeres     *resp = &data->res;
1190         int status;
1191
1192         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1193                 task->tk_pid, task->tk_status);
1194
1195         /* Call the NFS version-specific code */
1196         status = NFS_PROTO(data->inode)->write_done(task, data);
1197         if (status != 0)
1198                 return status;
1199         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1200
1201 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1202         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1203                 /* We tried a write call, but the server did not
1204                  * commit data to stable storage even though we
1205                  * requested it.
1206                  * Note: There is a known bug in Tru64 < 5.0 in which
1207                  *       the server reports NFS_DATA_SYNC, but performs
1208                  *       NFS_FILE_SYNC. We therefore implement this checking
1209                  *       as a dprintk() in order to avoid filling syslog.
1210                  */
1211                 static unsigned long    complain;
1212
1213                 if (time_before(complain, jiffies)) {
1214                         dprintk("NFS: faulty NFS server %s:"
1215                                 " (committed = %d) != (stable = %d)\n",
1216                                 NFS_SERVER(data->inode)->hostname,
1217                                 resp->verf->committed, argp->stable);
1218                         complain = jiffies + 300 * HZ;
1219                 }
1220         }
1221 #endif
1222         /* Is this a short write? */
1223         if (task->tk_status >= 0 && resp->count < argp->count) {
1224                 static unsigned long    complain;
1225
1226                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1227
1228                 /* Has the server at least made some progress? */
1229                 if (resp->count != 0) {
1230                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1231                         if (resp->verf->committed != NFS_UNSTABLE) {
1232                                 /* Resend from where the server left off */
1233                                 argp->offset += resp->count;
1234                                 argp->pgbase += resp->count;
1235                                 argp->count -= resp->count;
1236                         } else {
1237                                 /* Resend as a stable write in order to avoid
1238                                  * headaches in the case of a server crash.
1239                                  */
1240                                 argp->stable = NFS_FILE_SYNC;
1241                         }
1242                         rpc_restart_call(task);
1243                         return -EAGAIN;
1244                 }
1245                 if (time_before(complain, jiffies)) {
1246                         printk(KERN_WARNING
1247                                "NFS: Server wrote zero bytes, expected %u.\n",
1248                                         argp->count);
1249                         complain = jiffies + 300 * HZ;
1250                 }
1251                 /* Can't do anything about it except throw an error. */
1252                 task->tk_status = -EIO;
1253         }
1254         return 0;
1255 }
1256
1257
1258 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1259 void nfs_commit_release(void *wdata)
1260 {
1261         nfs_commit_free(wdata);
1262 }
1263
1264 /*
1265  * Set up the argument/result storage required for the RPC call.
1266  */
1267 static void nfs_commit_rpcsetup(struct list_head *head,
1268                 struct nfs_write_data *data,
1269                 int how)
1270 {
1271         struct nfs_page         *first;
1272         struct inode            *inode;
1273         int flags;
1274
1275         /* Set up the RPC argument and reply structs
1276          * NB: take care not to mess about with data->commit et al. */
1277
1278         list_splice_init(head, &data->pages);
1279         first = nfs_list_entry(data->pages.next);
1280         inode = first->wb_context->dentry->d_inode;
1281
1282         data->inode       = inode;
1283         data->cred        = first->wb_context->cred;
1284
1285         data->args.fh     = NFS_FH(data->inode);
1286         /* Note: we always request a commit of the entire inode */
1287         data->args.offset = 0;
1288         data->args.count  = 0;
1289         data->res.count   = 0;
1290         data->res.fattr   = &data->fattr;
1291         data->res.verf    = &data->verf;
1292         nfs_fattr_init(&data->fattr);
1293
1294         /* Set up the initial task struct.  */
1295         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1296         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1297         NFS_PROTO(inode)->commit_setup(data, how);
1298
1299         data->task.tk_priority = flush_task_priority(how);
1300         data->task.tk_cookie = (unsigned long)inode;
1301
1302         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1303 }
1304
1305 /*
1306  * Commit dirty pages
1307  */
1308 static int
1309 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1310 {
1311         struct nfs_write_data   *data;
1312         struct nfs_page         *req;
1313
1314         data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
1315
1316         if (!data)
1317                 goto out_bad;
1318
1319         /* Set up the argument struct */
1320         nfs_commit_rpcsetup(head, data, how);
1321
1322         nfs_execute_write(data);
1323         return 0;
1324  out_bad:
1325         while (!list_empty(head)) {
1326                 req = nfs_list_entry(head->next);
1327                 nfs_list_remove_request(req);
1328                 nfs_mark_request_commit(req);
1329                 nfs_clear_page_writeback(req);
1330         }
1331         return -ENOMEM;
1332 }
1333
1334 /*
1335  * COMMIT call returned
1336  */
1337 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1338 {
1339         struct nfs_write_data   *data = calldata;
1340         struct nfs_page         *req;
1341         int res = 0;
1342
1343         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1344                                 task->tk_pid, task->tk_status);
1345
1346         /* Call the NFS version-specific code */
1347         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1348                 return;
1349
1350         while (!list_empty(&data->pages)) {
1351                 req = nfs_list_entry(data->pages.next);
1352                 nfs_list_remove_request(req);
1353
1354                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1355                         req->wb_context->dentry->d_inode->i_sb->s_id,
1356                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1357                         req->wb_bytes,
1358                         (long long)req_offset(req));
1359                 if (task->tk_status < 0) {
1360                         req->wb_context->error = task->tk_status;
1361                         nfs_inode_remove_request(req);
1362                         dprintk(", error = %d\n", task->tk_status);
1363                         goto next;
1364                 }
1365
1366                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1367                  * returned by the server against all stored verfs. */
1368                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1369                         /* We have a match */
1370                         nfs_inode_remove_request(req);
1371                         dprintk(" OK\n");
1372                         goto next;
1373                 }
1374                 /* We have a mismatch. Write the page again */
1375                 dprintk(" mismatch\n");
1376                 nfs_mark_request_dirty(req);
1377         next:
1378                 nfs_clear_page_writeback(req);
1379                 res++;
1380         }
1381         sub_page_state(nr_unstable,res);
1382 }
1383
1384 static const struct rpc_call_ops nfs_commit_ops = {
1385         .rpc_call_done = nfs_commit_done,
1386         .rpc_release = nfs_commit_release,
1387 };
1388 #endif
1389
1390 static int nfs_flush_inode(struct inode *inode, unsigned long idx_start,
1391                            unsigned int npages, int how)
1392 {
1393         struct nfs_inode *nfsi = NFS_I(inode);
1394         LIST_HEAD(head);
1395         int                     res,
1396                                 error = 0;
1397
1398         spin_lock(&nfsi->req_lock);
1399         res = nfs_scan_dirty(inode, &head, idx_start, npages);
1400         spin_unlock(&nfsi->req_lock);
1401         if (res) {
1402                 struct nfs_server *server = NFS_SERVER(inode);
1403
1404                 /* For single writes, FLUSH_STABLE is more efficient */
1405                 if (res == nfsi->npages && nfsi->npages <= server->wpages) {
1406                         if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize)
1407                                 how |= FLUSH_STABLE;
1408                 }
1409                 error = nfs_flush_list(&head, server->wpages, how);
1410         }
1411         if (error < 0)
1412                 return error;
1413         return res;
1414 }
1415
1416 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1417 int nfs_commit_inode(struct inode *inode, int how)
1418 {
1419         struct nfs_inode *nfsi = NFS_I(inode);
1420         LIST_HEAD(head);
1421         int                     res,
1422                                 error = 0;
1423
1424         spin_lock(&nfsi->req_lock);
1425         res = nfs_scan_commit(inode, &head, 0, 0);
1426         spin_unlock(&nfsi->req_lock);
1427         if (res) {
1428                 error = nfs_commit_list(inode, &head, how);
1429                 if (error < 0)
1430                         return error;
1431         }
1432         return res;
1433 }
1434 #endif
1435
1436 int nfs_sync_inode(struct inode *inode, unsigned long idx_start,
1437                   unsigned int npages, int how)
1438 {
1439         int nocommit = how & FLUSH_NOCOMMIT;
1440         int wait = how & FLUSH_WAIT;
1441         int error;
1442
1443         how &= ~(FLUSH_WAIT|FLUSH_NOCOMMIT);
1444
1445         do {
1446                 if (wait) {
1447                         error = nfs_wait_on_requests(inode, idx_start, npages);
1448                         if (error != 0)
1449                                 continue;
1450                 }
1451                 error = nfs_flush_inode(inode, idx_start, npages, how);
1452                 if (error != 0)
1453                         continue;
1454                 if (!nocommit)
1455                         error = nfs_commit_inode(inode, how);
1456         } while (error > 0);
1457         return error;
1458 }
1459
1460 int nfs_init_writepagecache(void)
1461 {
1462         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1463                                              sizeof(struct nfs_write_data),
1464                                              0, SLAB_HWCACHE_ALIGN,
1465                                              NULL, NULL);
1466         if (nfs_wdata_cachep == NULL)
1467                 return -ENOMEM;
1468
1469         nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE,
1470                                            mempool_alloc_slab,
1471                                            mempool_free_slab,
1472                                            nfs_wdata_cachep);
1473         if (nfs_wdata_mempool == NULL)
1474                 return -ENOMEM;
1475
1476         nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT,
1477                                            mempool_alloc_slab,
1478                                            mempool_free_slab,
1479                                            nfs_wdata_cachep);
1480         if (nfs_commit_mempool == NULL)
1481                 return -ENOMEM;
1482
1483         return 0;
1484 }
1485
1486 void nfs_destroy_writepagecache(void)
1487 {
1488         mempool_destroy(nfs_commit_mempool);
1489         mempool_destroy(nfs_wdata_mempool);
1490         if (kmem_cache_destroy(nfs_wdata_cachep))
1491                 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1492 }
1493