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(unsigned int pagecount)
  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                 if (pagecount <= ARRAY_SIZE(p->page_array))
 101                         p->pagevec = p->page_array;
 102                 else {
 103                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 104                         if (!p->pagevec) {
 105                                 mempool_free(p, nfs_commit_mempool);
 106                                 p = NULL;
 107                         }
 108                 }
 109         }
 110         return p;
 111 }
 112
 113 void nfs_commit_free(struct nfs_write_data *p)
 114 {
 115         if (p && (p->pagevec != &p->page_array[0]))
 116                 kfree(p->pagevec);
 117         mempool_free(p, nfs_commit_mempool);
 118 }
 119
 120 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
 121 {
 122         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
 123
 124         if (p) {
 125                 memset(p, 0, sizeof(*p));
 126                 INIT_LIST_HEAD(&p->pages);
 127                 if (pagecount <= ARRAY_SIZE(p->page_array))
 128                         p->pagevec = p->page_array;
 129                 else {
 130                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
 131                         if (!p->pagevec) {
 132                                 mempool_free(p, nfs_wdata_mempool);
 133                                 p = NULL;
 134                         }
 135                 }
 136         }
 137         return p;
 138 }
 139
 140 static void nfs_writedata_free(struct nfs_write_data *p)
 141 {
 142         if (p && (p->pagevec != &p->page_array[0]))
 143                 kfree(p->pagevec);
 144         mempool_free(p, nfs_wdata_mempool);
 145 }
 146
 147 void nfs_writedata_release(void *wdata)
 148 {
 149         nfs_writedata_free(wdata);
 150 }
 151
 152 /* Adjust the file length if we're writing beyond the end */
 153 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 154 {
 155         struct inode *inode = page->mapping->host;
 156         loff_t end, i_size = i_size_read(inode);
 157         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 158
 159         if (i_size > 0 && page->index < end_index)
 160                 return;
 161         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 162         if (i_size >= end)
 163                 return;
 164         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 165         i_size_write(inode, end);
 166 }
 167
 168 /* We can set the PG_uptodate flag if we see that a write request
 169  * covers the full page.
 170  */
 171 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 172 {
 173         loff_t end_offs;
 174
 175         if (PageUptodate(page))
 176                 return;
 177         if (base != 0)
 178                 return;
 179         if (count == PAGE_CACHE_SIZE) {
 180                 SetPageUptodate(page);
 181                 return;
 182         }
 183
 184         end_offs = i_size_read(page->mapping->host) - 1;
 185         if (end_offs < 0)
 186                 return;
 187         /* Is this the last page? */
 188         if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT))
 189                 return;
 190         /* This is the last page: set PG_uptodate if we cover the entire
 191          * extent of the data, then zero the rest of the page.
 192          */
 193         if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) {
 194                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 195                 SetPageUptodate(page);
 196         }
 197 }
 198
 199 /*
 200  * Write a page synchronously.
 201  * Offset is the data offset within the page.
 202  */
 203 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 204                 struct page *page, unsigned int offset, unsigned int count,
 205                 int how)
 206 {
 207         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 208         int             result, written = 0;
 209         struct nfs_write_data *wdata;
 210
 211         wdata = nfs_writedata_alloc(1);
 212         if (!wdata)
 213                 return -ENOMEM;
 214
 215         wdata->flags = how;
 216         wdata->cred = ctx->cred;
 217         wdata->inode = inode;
 218         wdata->args.fh = NFS_FH(inode);
 219         wdata->args.context = ctx;
 220         wdata->args.pages = &page;
 221         wdata->args.stable = NFS_FILE_SYNC;
 222         wdata->args.pgbase = offset;
 223         wdata->args.count = wsize;
 224         wdata->res.fattr = &wdata->fattr;
 225         wdata->res.verf = &wdata->verf;
 226
 227         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 228                 inode->i_sb->s_id,
 229                 (long long)NFS_FILEID(inode),
 230                 count, (long long)(page_offset(page) + offset));
 231
 232         set_page_writeback(page);
 233         nfs_begin_data_update(inode);
 234         do {
 235                 if (count < wsize)
 236                         wdata->args.count = count;
 237                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 238
 239                 result = NFS_PROTO(inode)->write(wdata);
 240
 241                 if (result < 0) {
 242                         /* Must mark the page invalid after I/O error */
 243                         ClearPageUptodate(page);
 244                         goto io_error;
 245                 }
 246                 if (result < wdata->args.count)
 247                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 248                                         wdata->args.count, result);
 249
 250                 wdata->args.offset += result;
 251                 wdata->args.pgbase += result;
 252                 written += result;
 253                 count -= result;
 254                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 255         } while (count);
 256         /* Update file length */
 257         nfs_grow_file(page, offset, written);
 258         /* Set the PG_uptodate flag? */
 259         nfs_mark_uptodate(page, offset, written);
 260
 261         if (PageError(page))
 262                 ClearPageError(page);
 263
 264 io_error:
 265         nfs_end_data_update(inode);
 266         end_page_writeback(page);
 267         nfs_writedata_free(wdata);
 268         return written ? written : result;
 269 }
 270
 271 static int nfs_writepage_async(struct nfs_open_context *ctx,
 272                 struct inode *inode, struct page *page,
 273                 unsigned int offset, unsigned int count)
 274 {
 275         struct nfs_page *req;
 276
 277         req = nfs_update_request(ctx, inode, page, offset, count);
 278         if (IS_ERR(req))
 279                 return PTR_ERR(req);
 280         /* Update file length */
 281         nfs_grow_file(page, offset, count);
 282         /* Set the PG_uptodate flag? */
 283         nfs_mark_uptodate(page, offset, count);
 284         nfs_unlock_request(req);
 285         return 0;
 286 }
 287
 288 static int wb_priority(struct writeback_control *wbc)
 289 {
 290         if (wbc->for_reclaim)
 291                 return FLUSH_HIGHPRI;
 292         if (wbc->for_kupdate)
 293                 return FLUSH_LOWPRI;
 294         return 0;
 295 }
 296
 297 /*
 298  * Write an mmapped page to the server.
 299  */
 300 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 301 {
 302         struct nfs_open_context *ctx;
 303         struct inode *inode = page->mapping->host;
 304         unsigned long end_index;
 305         unsigned offset = PAGE_CACHE_SIZE;
 306         loff_t i_size = i_size_read(inode);
 307         int inode_referenced = 0;
 308         int priority = wb_priority(wbc);
 309         int err;
 310
 311         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 312         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 313
 314         /*
 315          * Note: We need to ensure that we have a reference to the inode
 316          *       if we are to do asynchronous writes. If not, waiting
 317          *       in nfs_wait_on_request() may deadlock with clear_inode().
 318          *
 319          *       If igrab() fails here, then it is in any case safe to
 320          *       call nfs_wb_page(), since there will be no pending writes.
 321          */
 322         if (igrab(inode) != 0)
 323                 inode_referenced = 1;
 324         end_index = i_size >> PAGE_CACHE_SHIFT;
 325
 326         /* Ensure we've flushed out any previous writes */
 327         nfs_wb_page_priority(inode, page, priority);
 328
 329         /* easy case */
 330         if (page->index < end_index)
 331                 goto do_it;
 332         /* things got complicated... */
 333         offset = i_size & (PAGE_CACHE_SIZE-1);
 334
 335         /* OK, are we completely out? */
 336         err = 0; /* potential race with truncate - ignore */
 337         if (page->index >= end_index+1 || !offset)
 338                 goto out;
 339 do_it:
 340         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 341         if (ctx == NULL) {
 342                 err = -EBADF;
 343                 goto out;
 344         }
 345         lock_kernel();
 346         if (!IS_SYNC(inode) && inode_referenced) {
 347                 err = nfs_writepage_async(ctx, inode, page, 0, offset);
 348                 if (!wbc->for_writepages)
 349                         nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 350         } else {
 351                 err = nfs_writepage_sync(ctx, inode, page, 0,
 352                                                 offset, priority);
 353                 if (err >= 0) {
 354                         if (err != offset)
 355                                 redirty_page_for_writepage(wbc, page);
 356                         err = 0;
 357                 }
 358         }
 359         unlock_kernel();
 360         put_nfs_open_context(ctx);
 361 out:
 362         unlock_page(page);
 363         if (inode_referenced)
 364                 iput(inode);
 365         return err;
 366 }
 367
 368 /*
 369  * Note: causes nfs_update_request() to block on the assumption
 370  *       that the writeback is generated due to memory pressure.
 371  */
 372 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 373 {
 374         struct backing_dev_info *bdi = mapping->backing_dev_info;
 375         struct inode *inode = mapping->host;
 376         int err;
 377
 378         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 379
 380         err = generic_writepages(mapping, wbc);
 381         if (err)
 382                 return err;
 383         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 384                 if (wbc->nonblocking)
 385                         return 0;
 386                 nfs_wait_on_write_congestion(mapping, 0);
 387         }
 388         err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc));
 389         if (err < 0)
 390                 goto out;
 391         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 392         wbc->nr_to_write -= err;
 393         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 394                 err = nfs_wait_on_requests(inode, 0, 0);
 395                 if (err < 0)
 396                         goto out;
 397         }
 398         err = nfs_commit_inode(inode, wb_priority(wbc));
 399         if (err > 0) {
 400                 wbc->nr_to_write -= err;
 401                 err = 0;
 402         }
 403 out:
 404         clear_bit(BDI_write_congested, &bdi->state);
 405         wake_up_all(&nfs_write_congestion);
 406         return err;
 407 }
 408
 409 /*
 410  * Insert a write request into an inode
 411  */
 412 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 413 {
 414         struct nfs_inode *nfsi = NFS_I(inode);
 415         int error;
 416
 417         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 418         BUG_ON(error == -EEXIST);
 419         if (error)
 420                 return error;
 421         if (!nfsi->npages) {
 422                 igrab(inode);
 423                 nfs_begin_data_update(inode);
 424                 if (nfs_have_delegation(inode, FMODE_WRITE))
 425                         nfsi->change_attr++;
 426         }
 427         SetPagePrivate(req->wb_page);
 428         nfsi->npages++;
 429         atomic_inc(&req->wb_count);
 430         return 0;
 431 }
 432
 433 /*
 434  * Insert a write request into an inode
 435  */
 436 static void nfs_inode_remove_request(struct nfs_page *req)
 437 {
 438         struct inode *inode = req->wb_context->dentry->d_inode;
 439         struct nfs_inode *nfsi = NFS_I(inode);
 440
 441         BUG_ON (!NFS_WBACK_BUSY(req));
 442
 443         spin_lock(&nfsi->req_lock);
 444         ClearPagePrivate(req->wb_page);
 445         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 446         nfsi->npages--;
 447         if (!nfsi->npages) {
 448                 spin_unlock(&nfsi->req_lock);
 449                 nfs_end_data_update(inode);
 450                 iput(inode);
 451         } else
 452                 spin_unlock(&nfsi->req_lock);
 453         nfs_clear_request(req);
 454         nfs_release_request(req);
 455 }
 456
 457 /*
 458  * Find a request
 459  */
 460 static inline struct nfs_page *
 461 _nfs_find_request(struct inode *inode, unsigned long index)
 462 {
 463         struct nfs_inode *nfsi = NFS_I(inode);
 464         struct nfs_page *req;
 465
 466         req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index);
 467         if (req)
 468                 atomic_inc(&req->wb_count);
 469         return req;
 470 }
 471
 472 static struct nfs_page *
 473 nfs_find_request(struct inode *inode, unsigned long index)
 474 {
 475         struct nfs_page         *req;
 476         struct nfs_inode        *nfsi = NFS_I(inode);
 477
 478         spin_lock(&nfsi->req_lock);
 479         req = _nfs_find_request(inode, index);
 480         spin_unlock(&nfsi->req_lock);
 481         return req;
 482 }
 483
 484 /*
 485  * Add a request to the inode's dirty list.
 486  */
 487 static void
 488 nfs_mark_request_dirty(struct nfs_page *req)
 489 {
 490         struct inode *inode = req->wb_context->dentry->d_inode;
 491         struct nfs_inode *nfsi = NFS_I(inode);
 492
 493         spin_lock(&nfsi->req_lock);
 494         radix_tree_tag_set(&nfsi->nfs_page_tree,
 495                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 496         nfs_list_add_request(req, &nfsi->dirty);
 497         nfsi->ndirty++;
 498         spin_unlock(&nfsi->req_lock);
 499         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 500         mark_inode_dirty(inode);
 501 }
 502
 503 /*
 504  * Check if a request is dirty
 505  */
 506 static inline int
 507 nfs_dirty_request(struct nfs_page *req)
 508 {
 509         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 510         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 511 }
 512
 513 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 514 /*
 515  * Add a request to the inode's commit list.
 516  */
 517 static void
 518 nfs_mark_request_commit(struct nfs_page *req)
 519 {
 520         struct inode *inode = req->wb_context->dentry->d_inode;
 521         struct nfs_inode *nfsi = NFS_I(inode);
 522
 523         spin_lock(&nfsi->req_lock);
 524         nfs_list_add_request(req, &nfsi->commit);
 525         nfsi->ncommit++;
 526         spin_unlock(&nfsi->req_lock);
 527         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 528         mark_inode_dirty(inode);
 529 }
 530 #endif
 531
 532 /*
 533  * Wait for a request to complete.
 534  *
 535  * Interruptible by signals only if mounted with intr flag.
 536  */
 537 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 538 {
 539         struct nfs_inode *nfsi = NFS_I(inode);
 540         struct nfs_page *req;
 541         unsigned long           idx_end, next;
 542         unsigned int            res = 0;
 543         int                     error;
 544
 545         if (npages == 0)
 546                 idx_end = ~0;
 547         else
 548                 idx_end = idx_start + npages - 1;
 549
 550         next = idx_start;
 551         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 552                 if (req->wb_index > idx_end)
 553                         break;
 554
 555                 next = req->wb_index + 1;
 556                 BUG_ON(!NFS_WBACK_BUSY(req));
 557
 558                 atomic_inc(&req->wb_count);
 559                 spin_unlock(&nfsi->req_lock);
 560                 error = nfs_wait_on_request(req);
 561                 nfs_release_request(req);
 562                 spin_lock(&nfsi->req_lock);
 563                 if (error < 0)
 564                         return error;
 565                 res++;
 566         }
 567         return res;
 568 }
 569
 570 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 571 {
 572         struct nfs_inode *nfsi = NFS_I(inode);
 573         int ret;
 574
 575         spin_lock(&nfsi->req_lock);
 576         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 577         spin_unlock(&nfsi->req_lock);
 578         return ret;
 579 }
 580
 581 static void nfs_cancel_dirty_list(struct list_head *head)
 582 {
 583         struct nfs_page *req;
 584         while(!list_empty(head)) {
 585                 req = nfs_list_entry(head->next);
 586                 nfs_list_remove_request(req);
 587                 nfs_inode_remove_request(req);
 588                 nfs_clear_page_writeback(req);
 589         }
 590 }
 591
 592 static void nfs_cancel_commit_list(struct list_head *head)
 593 {
 594         struct nfs_page *req;
 595
 596         while(!list_empty(head)) {
 597                 req = nfs_list_entry(head->next);
 598                 nfs_list_remove_request(req);
 599                 nfs_inode_remove_request(req);
 600                 nfs_clear_page_writeback(req);
 601                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 602         }
 603 }
 604
 605 /*
 606  * nfs_scan_dirty - Scan an inode for dirty requests
 607  * @inode: NFS inode to scan
 608  * @dst: destination list
 609  * @idx_start: lower bound of page->index to scan.
 610  * @npages: idx_start + npages sets the upper bound to scan.
 611  *
 612  * Moves requests from the inode's dirty page list.
 613  * The requests are *not* checked to ensure that they form a contiguous set.
 614  */
 615 static int
 616 nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 617 {
 618         struct nfs_inode *nfsi = NFS_I(inode);
 619         int res = 0;
 620
 621         if (nfsi->ndirty != 0) {
 622                 res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages);
 623                 nfsi->ndirty -= res;
 624                 if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty))
 625                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n");
 626         }
 627         return res;
 628 }
 629
 630 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 631 /*
 632  * nfs_scan_commit - Scan an inode for commit requests
 633  * @inode: NFS inode to scan
 634  * @dst: destination list
 635  * @idx_start: lower bound of page->index to scan.
 636  * @npages: idx_start + npages sets the upper bound to scan.
 637  *
 638  * Moves requests from the inode's 'commit' request list.
 639  * The requests are *not* checked to ensure that they form a contiguous set.
 640  */
 641 static int
 642 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 643 {
 644         struct nfs_inode *nfsi = NFS_I(inode);
 645         int res = 0;
 646
 647         if (nfsi->ncommit != 0) {
 648                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 649                 nfsi->ncommit -= res;
 650                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 651                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 652         }
 653         return res;
 654 }
 655 #else
 656 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 657 {
 658         return 0;
 659 }
 660 #endif
 661
 662 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 663 {
 664         struct backing_dev_info *bdi = mapping->backing_dev_info;
 665         DEFINE_WAIT(wait);
 666         int ret = 0;
 667
 668         might_sleep();
 669
 670         if (!bdi_write_congested(bdi))
 671                 return 0;
 672
 673         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 674
 675         if (intr) {
 676                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 677                 sigset_t oldset;
 678
 679                 rpc_clnt_sigmask(clnt, &oldset);
 680                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 681                 if (bdi_write_congested(bdi)) {
 682                         if (signalled())
 683                                 ret = -ERESTARTSYS;
 684                         else
 685                                 schedule();
 686                 }
 687                 rpc_clnt_sigunmask(clnt, &oldset);
 688         } else {
 689                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 690                 if (bdi_write_congested(bdi))
 691                         schedule();
 692         }
 693         finish_wait(&nfs_write_congestion, &wait);
 694         return ret;
 695 }
 696
 697
 698 /*
 699  * Try to update any existing write request, or create one if there is none.
 700  * In order to match, the request's credentials must match those of
 701  * the calling process.
 702  *
 703  * Note: Should always be called with the Page Lock held!
 704  */
 705 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 706                 struct inode *inode, struct page *page,
 707                 unsigned int offset, unsigned int bytes)
 708 {
 709         struct nfs_server *server = NFS_SERVER(inode);
 710         struct nfs_inode *nfsi = NFS_I(inode);
 711         struct nfs_page         *req, *new = NULL;
 712         unsigned long           rqend, end;
 713
 714         end = offset + bytes;
 715
 716         if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR))
 717                 return ERR_PTR(-ERESTARTSYS);
 718         for (;;) {
 719                 /* Loop over all inode entries and see if we find
 720                  * A request for the page we wish to update
 721                  */
 722                 spin_lock(&nfsi->req_lock);
 723                 req = _nfs_find_request(inode, page->index);
 724                 if (req) {
 725                         if (!nfs_lock_request_dontget(req)) {
 726                                 int error;
 727                                 spin_unlock(&nfsi->req_lock);
 728                                 error = nfs_wait_on_request(req);
 729                                 nfs_release_request(req);
 730                                 if (error < 0) {
 731                                         if (new)
 732                                                 nfs_release_request(new);
 733                                         return ERR_PTR(error);
 734                                 }
 735                                 continue;
 736                         }
 737                         spin_unlock(&nfsi->req_lock);
 738                         if (new)
 739                                 nfs_release_request(new);
 740                         break;
 741                 }
 742
 743                 if (new) {
 744                         int error;
 745                         nfs_lock_request_dontget(new);
 746                         error = nfs_inode_add_request(inode, new);
 747                         if (error) {
 748                                 spin_unlock(&nfsi->req_lock);
 749                                 nfs_unlock_request(new);
 750                                 return ERR_PTR(error);
 751                         }
 752                         spin_unlock(&nfsi->req_lock);
 753                         nfs_mark_request_dirty(new);
 754                         return new;
 755                 }
 756                 spin_unlock(&nfsi->req_lock);
 757
 758                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 759                 if (IS_ERR(new))
 760                         return new;
 761         }
 762
 763         /* We have a request for our page.
 764          * If the creds don't match, or the
 765          * page addresses don't match,
 766          * tell the caller to wait on the conflicting
 767          * request.
 768          */
 769         rqend = req->wb_offset + req->wb_bytes;
 770         if (req->wb_context != ctx
 771             || req->wb_page != page
 772             || !nfs_dirty_request(req)
 773             || offset > rqend || end < req->wb_offset) {
 774                 nfs_unlock_request(req);
 775                 return ERR_PTR(-EBUSY);
 776         }
 777
 778         /* Okay, the request matches. Update the region */
 779         if (offset < req->wb_offset) {
 780                 req->wb_offset = offset;
 781                 req->wb_pgbase = offset;
 782                 req->wb_bytes = rqend - req->wb_offset;
 783         }
 784
 785         if (end > rqend)
 786                 req->wb_bytes = end - req->wb_offset;
 787
 788         return req;
 789 }
 790
 791 int nfs_flush_incompatible(struct file *file, struct page *page)
 792 {
 793         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 794         struct inode    *inode = page->mapping->host;
 795         struct nfs_page *req;
 796         int             status = 0;
 797         /*
 798          * Look for a request corresponding to this page. If there
 799          * is one, and it belongs to another file, we flush it out
 800          * before we try to copy anything into the page. Do this
 801          * due to the lack of an ACCESS-type call in NFSv2.
 802          * Also do the same if we find a request from an existing
 803          * dropped page.
 804          */
 805         req = nfs_find_request(inode, page->index);
 806         if (req) {
 807                 if (req->wb_page != page || ctx != req->wb_context)
 808                         status = nfs_wb_page(inode, page);
 809                 nfs_release_request(req);
 810         }
 811         return (status < 0) ? status : 0;
 812 }
 813
 814 /*
 815  * Update and possibly write a cached page of an NFS file.
 816  *
 817  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 818  * things with a page scheduled for an RPC call (e.g. invalidate it).
 819  */
 820 int nfs_updatepage(struct file *file, struct page *page,
 821                 unsigned int offset, unsigned int count)
 822 {
 823         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 824         struct inode    *inode = page->mapping->host;
 825         struct nfs_page *req;
 826         int             status = 0;
 827
 828         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 829
 830         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 831                 file->f_dentry->d_parent->d_name.name,
 832                 file->f_dentry->d_name.name, count,
 833                 (long long)(page_offset(page) +offset));
 834
 835         if (IS_SYNC(inode)) {
 836                 status = nfs_writepage_sync(ctx, inode, page, offset, count, 0);
 837                 if (status > 0) {
 838                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 839                                 SetPageUptodate(page);
 840                         return 0;
 841                 }
 842                 return status;
 843         }
 844
 845         /* If we're not using byte range locks, and we know the page
 846          * is entirely in cache, it may be more efficient to avoid
 847          * fragmenting write requests.
 848          */
 849         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 850                 loff_t end_offs = i_size_read(inode) - 1;
 851                 unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT;
 852
 853                 count += offset;
 854                 offset = 0;
 855                 if (unlikely(end_offs < 0)) {
 856                         /* Do nothing */
 857                 } else if (page->index == end_index) {
 858                         unsigned int pglen;
 859                         pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1;
 860                         if (count < pglen)
 861                                 count = pglen;
 862                 } else if (page->index < end_index)
 863                         count = PAGE_CACHE_SIZE;
 864         }
 865
 866         /*
 867          * Try to find an NFS request corresponding to this page
 868          * and update it.
 869          * If the existing request cannot be updated, we must flush
 870          * it out now.
 871          */
 872         do {
 873                 req = nfs_update_request(ctx, inode, page, offset, count);
 874                 status = (IS_ERR(req)) ? PTR_ERR(req) : 0;
 875                 if (status != -EBUSY)
 876                         break;
 877                 /* Request could not be updated. Flush it out and try again */
 878                 status = nfs_wb_page(inode, page);
 879         } while (status >= 0);
 880         if (status < 0)
 881                 goto done;
 882
 883         status = 0;
 884
 885         /* Update file length */
 886         nfs_grow_file(page, offset, count);
 887         /* Set the PG_uptodate flag? */
 888         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 889         nfs_unlock_request(req);
 890 done:
 891         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 892                         status, (long long)i_size_read(inode));
 893         if (status < 0)
 894                 ClearPageUptodate(page);
 895         return status;
 896 }
 897
 898 static void nfs_writepage_release(struct nfs_page *req)
 899 {
 900         end_page_writeback(req->wb_page);
 901
 902 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 903         if (!PageError(req->wb_page)) {
 904                 if (NFS_NEED_RESCHED(req)) {
 905                         nfs_mark_request_dirty(req);
 906                         goto out;
 907                 } else if (NFS_NEED_COMMIT(req)) {
 908                         nfs_mark_request_commit(req);
 909                         goto out;
 910                 }
 911         }
 912         nfs_inode_remove_request(req);
 913
 914 out:
 915         nfs_clear_commit(req);
 916         nfs_clear_reschedule(req);
 917 #else
 918         nfs_inode_remove_request(req);
 919 #endif
 920         nfs_clear_page_writeback(req);
 921 }
 922
 923 static inline int flush_task_priority(int how)
 924 {
 925         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 926                 case FLUSH_HIGHPRI:
 927                         return RPC_PRIORITY_HIGH;
 928                 case FLUSH_LOWPRI:
 929                         return RPC_PRIORITY_LOW;
 930         }
 931         return RPC_PRIORITY_NORMAL;
 932 }
 933
 934 /*
 935  * Set up the argument/result storage required for the RPC call.
 936  */
 937 static void nfs_write_rpcsetup(struct nfs_page *req,
 938                 struct nfs_write_data *data,
 939                 const struct rpc_call_ops *call_ops,
 940                 unsigned int count, unsigned int offset,
 941                 int how)
 942 {
 943         struct inode            *inode;
 944         int flags;
 945
 946         /* Set up the RPC argument and reply structs
 947          * NB: take care not to mess about with data->commit et al. */
 948
 949         data->req = req;
 950         data->inode = inode = req->wb_context->dentry->d_inode;
 951         data->cred = req->wb_context->cred;
 952
 953         data->args.fh     = NFS_FH(inode);
 954         data->args.offset = req_offset(req) + offset;
 955         data->args.pgbase = req->wb_pgbase + offset;
 956         data->args.pages  = data->pagevec;
 957         data->args.count  = count;
 958         data->args.context = req->wb_context;
 959
 960         data->res.fattr   = &data->fattr;
 961         data->res.count   = count;
 962         data->res.verf    = &data->verf;
 963         nfs_fattr_init(&data->fattr);
 964
 965         /* Set up the initial task struct.  */
 966         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 967         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 968         NFS_PROTO(inode)->write_setup(data, how);
 969
 970         data->task.tk_priority = flush_task_priority(how);
 971         data->task.tk_cookie = (unsigned long)inode;
 972
 973         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 974                 data->task.tk_pid,
 975                 inode->i_sb->s_id,
 976                 (long long)NFS_FILEID(inode),
 977                 count,
 978                 (unsigned long long)data->args.offset);
 979 }
 980
 981 static void nfs_execute_write(struct nfs_write_data *data)
 982 {
 983         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 984         sigset_t oldset;
 985
 986         rpc_clnt_sigmask(clnt, &oldset);
 987         lock_kernel();
 988         rpc_execute(&data->task);
 989         unlock_kernel();
 990         rpc_clnt_sigunmask(clnt, &oldset);
 991 }
 992
 993 /*
 994  * Generate multiple small requests to write out a single
 995  * contiguous dirty area on one page.
 996  */
 997 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 998 {
 999         struct nfs_page *req = nfs_list_entry(head->next);
1000         struct page *page = req->wb_page;
1001         struct nfs_write_data *data;
1002         unsigned int wsize = NFS_SERVER(inode)->wsize;
1003         unsigned int nbytes, offset;
1004         int requests = 0;
1005         LIST_HEAD(list);
1006
1007         nfs_list_remove_request(req);
1008
1009         nbytes = req->wb_bytes;
1010         for (;;) {
1011                 data = nfs_writedata_alloc(1);
1012                 if (!data)
1013                         goto out_bad;
1014                 list_add(&data->pages, &list);
1015                 requests++;
1016                 if (nbytes <= wsize)
1017                         break;
1018                 nbytes -= wsize;
1019         }
1020         atomic_set(&req->wb_complete, requests);
1021
1022         ClearPageError(page);
1023         set_page_writeback(page);
1024         offset = 0;
1025         nbytes = req->wb_bytes;
1026         do {
1027                 data = list_entry(list.next, struct nfs_write_data, pages);
1028                 list_del_init(&data->pages);
1029
1030                 data->pagevec[0] = page;
1031
1032                 if (nbytes > wsize) {
1033                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1034                                         wsize, offset, how);
1035                         offset += wsize;
1036                         nbytes -= wsize;
1037                 } else {
1038                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
1039                                         nbytes, offset, how);
1040                         nbytes = 0;
1041                 }
1042                 nfs_execute_write(data);
1043         } while (nbytes != 0);
1044
1045         return 0;
1046
1047 out_bad:
1048         while (!list_empty(&list)) {
1049                 data = list_entry(list.next, struct nfs_write_data, pages);
1050                 list_del(&data->pages);
1051                 nfs_writedata_free(data);
1052         }
1053         nfs_mark_request_dirty(req);
1054         nfs_clear_page_writeback(req);
1055         return -ENOMEM;
1056 }
1057
1058 /*
1059  * Create an RPC task for the given write request and kick it.
1060  * The page must have been locked by the caller.
1061  *
1062  * It may happen that the page we're passed is not marked dirty.
1063  * This is the case if nfs_updatepage detects a conflicting request
1064  * that has been written but not committed.
1065  */
1066 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
1067 {
1068         struct nfs_page         *req;
1069         struct page             **pages;
1070         struct nfs_write_data   *data;
1071         unsigned int            count;
1072
1073         data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
1074         if (!data)
1075                 goto out_bad;
1076
1077         pages = data->pagevec;
1078         count = 0;
1079         while (!list_empty(head)) {
1080                 req = nfs_list_entry(head->next);
1081                 nfs_list_remove_request(req);
1082                 nfs_list_add_request(req, &data->pages);
1083                 ClearPageError(req->wb_page);
1084                 set_page_writeback(req->wb_page);
1085                 *pages++ = req->wb_page;
1086                 count += req->wb_bytes;
1087         }
1088         req = nfs_list_entry(data->pages.next);
1089
1090         /* Set up the argument struct */
1091         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1092
1093         nfs_execute_write(data);
1094         return 0;
1095  out_bad:
1096         while (!list_empty(head)) {
1097                 struct nfs_page *req = nfs_list_entry(head->next);
1098                 nfs_list_remove_request(req);
1099                 nfs_mark_request_dirty(req);
1100                 nfs_clear_page_writeback(req);
1101         }
1102         return -ENOMEM;
1103 }
1104
1105 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1106 {
1107         LIST_HEAD(one_request);
1108         int (*flush_one)(struct inode *, struct list_head *, int);
1109         struct nfs_page *req;
1110         int wpages = NFS_SERVER(inode)->wpages;
1111         int wsize = NFS_SERVER(inode)->wsize;
1112         int error;
1113
1114         flush_one = nfs_flush_one;
1115         if (wsize < PAGE_CACHE_SIZE)
1116                 flush_one = nfs_flush_multi;
1117         /* For single writes, FLUSH_STABLE is more efficient */
1118         if (npages <= wpages && npages == NFS_I(inode)->npages
1119                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1120                 how |= FLUSH_STABLE;
1121
1122         do {
1123                 nfs_coalesce_requests(head, &one_request, wpages);
1124                 req = nfs_list_entry(one_request.next);
1125                 error = flush_one(inode, &one_request, how);
1126                 if (error < 0)
1127                         goto out_err;
1128         } while (!list_empty(head));
1129         return 0;
1130 out_err:
1131         while (!list_empty(head)) {
1132                 req = nfs_list_entry(head->next);
1133                 nfs_list_remove_request(req);
1134                 nfs_mark_request_dirty(req);
1135                 nfs_clear_page_writeback(req);
1136         }
1137         return error;
1138 }
1139
1140 /*
1141  * Handle a write reply that flushed part of a page.
1142  */
1143 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1144 {
1145         struct nfs_write_data   *data = calldata;
1146         struct nfs_page         *req = data->req;
1147         struct page             *page = req->wb_page;
1148
1149         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1150                 req->wb_context->dentry->d_inode->i_sb->s_id,
1151                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1152                 req->wb_bytes,
1153                 (long long)req_offset(req));
1154
1155         if (nfs_writeback_done(task, data) != 0)
1156                 return;
1157
1158         if (task->tk_status < 0) {
1159                 ClearPageUptodate(page);
1160                 SetPageError(page);
1161                 req->wb_context->error = task->tk_status;
1162                 dprintk(", error = %d\n", task->tk_status);
1163         } else {
1164 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1165                 if (data->verf.committed < NFS_FILE_SYNC) {
1166                         if (!NFS_NEED_COMMIT(req)) {
1167                                 nfs_defer_commit(req);
1168                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1169                                 dprintk(" defer commit\n");
1170                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1171                                 nfs_defer_reschedule(req);
1172                                 dprintk(" server reboot detected\n");
1173                         }
1174                 } else
1175 #endif
1176                         dprintk(" OK\n");
1177         }
1178
1179         if (atomic_dec_and_test(&req->wb_complete))
1180                 nfs_writepage_release(req);
1181 }
1182
1183 static const struct rpc_call_ops nfs_write_partial_ops = {
1184         .rpc_call_done = nfs_writeback_done_partial,
1185         .rpc_release = nfs_writedata_release,
1186 };
1187
1188 /*
1189  * Handle a write reply that flushes a whole page.
1190  *
1191  * FIXME: There is an inherent race with invalidate_inode_pages and
1192  *        writebacks since the page->count is kept > 1 for as long
1193  *        as the page has a write request pending.
1194  */
1195 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1196 {
1197         struct nfs_write_data   *data = calldata;
1198         struct nfs_page         *req;
1199         struct page             *page;
1200
1201         if (nfs_writeback_done(task, data) != 0)
1202                 return;
1203
1204         /* Update attributes as result of writeback. */
1205         while (!list_empty(&data->pages)) {
1206                 req = nfs_list_entry(data->pages.next);
1207                 nfs_list_remove_request(req);
1208                 page = req->wb_page;
1209
1210                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1211                         req->wb_context->dentry->d_inode->i_sb->s_id,
1212                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1213                         req->wb_bytes,
1214                         (long long)req_offset(req));
1215
1216                 if (task->tk_status < 0) {
1217                         ClearPageUptodate(page);
1218                         SetPageError(page);
1219                         req->wb_context->error = task->tk_status;
1220                         end_page_writeback(page);
1221                         nfs_inode_remove_request(req);
1222                         dprintk(", error = %d\n", task->tk_status);
1223                         goto next;
1224                 }
1225                 end_page_writeback(page);
1226
1227 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1228                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1229                         nfs_inode_remove_request(req);
1230                         dprintk(" OK\n");
1231                         goto next;
1232                 }
1233                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1234                 nfs_mark_request_commit(req);
1235                 dprintk(" marked for commit\n");
1236 #else
1237                 nfs_inode_remove_request(req);
1238 #endif
1239         next:
1240                 nfs_clear_page_writeback(req);
1241         }
1242 }
1243
1244 static const struct rpc_call_ops nfs_write_full_ops = {
1245         .rpc_call_done = nfs_writeback_done_full,
1246         .rpc_release = nfs_writedata_release,
1247 };
1248
1249
1250 /*
1251  * This function is called when the WRITE call is complete.
1252  */
1253 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1254 {
1255         struct nfs_writeargs    *argp = &data->args;
1256         struct nfs_writeres     *resp = &data->res;
1257         int status;
1258
1259         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1260                 task->tk_pid, task->tk_status);
1261
1262         /* Call the NFS version-specific code */
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)->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(NFS_SERVER(inode)->wpages);
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                 nfs_clear_page_writeback(req);
1397                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
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         if (kmem_cache_destroy(nfs_wdata_cachep))
1569                 printk(KERN_INFO "nfs_write_data: not all structures were freed\n");
1570 }
1571