fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Writing file data over NFS.
   5  *
   6  * We do it like this: When a (user) process wishes to write data to an
   7  * NFS file, a write request is allocated that contains the RPC task data
   8  * plus some info on the page to be written, and added to the inode's
   9  * write chain. If the process writes past the end of the page, an async
  10  * RPC call to write the page is scheduled immediately; otherwise, the call
  11  * is delayed for a few seconds.
  12  *
  13  * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE.
  14  *
  15  * Write requests are kept on the inode's writeback list. Each entry in
  16  * that list references the page (portion) to be written. When the
  17  * cache timeout has expired, the RPC task is woken up, and tries to
  18  * lock the page. As soon as it manages to do so, the request is moved
  19  * from the writeback list to the writelock list.
  20  *
  21  * Note: we must make sure never to confuse the inode passed in the
  22  * write_page request with the one in page->inode. As far as I understand
  23  * it, these are different when doing a swap-out.
  24  *
  25  * To understand everything that goes on here and in the NFS read code,
  26  * one should be aware that a page is locked in exactly one of the following
  27  * cases:
  28  *
  29  *  -   A write request is in progress.
  30  *  -   A user process is in generic_file_write/nfs_update_page
  31  *  -   A user process is in generic_file_read
  32  *
  33  * Also note that because of the way pages are invalidated in
  34  * nfs_revalidate_inode, the following assertions hold:
  35  *
  36  *  -   If a page is dirty, there will be no read requests (a page will
  37  *      not be re-read unless invalidated by nfs_revalidate_inode).
  38  *  -   If the page is not uptodate, there will be no pending write
  39  *      requests, and no process will be in nfs_update_page.
  40  *
  41  * FIXME: Interaction with the vmscan routines is not optimal yet.
  42  * Either vmscan must be made nfs-savvy, or we need a different page
  43  * reclaim concept that supports something like FS-independent
  44  * buffer_heads with a b_ops-> field.
  45  *
  46  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
  47  */
  48
  49 #include <linux/types.h>
  50 #include <linux/slab.h>
  51 #include <linux/mm.h>
  52 #include <linux/pagemap.h>
  53 #include <linux/file.h>
  54 #include <linux/writeback.h>
  55
  56 #include <linux/sunrpc/clnt.h>
  57 #include <linux/nfs_fs.h>
  58 #include <linux/nfs_mount.h>
  59 #include <linux/nfs_page.h>
  60 #include <linux/backing-dev.h>
  61
  62 #include <asm/uaccess.h>
  63 #include <linux/smp_lock.h>
  64
  65 #include "delegation.h"
  66 #include "internal.h"
  67 #include "iostat.h"
  68
  69 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  70
  71 #define MIN_POOL_WRITE          (32)
  72 #define MIN_POOL_COMMIT         (4)
  73
  74 /*
  75  * Local function declarations
  76  */
  77 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  78                                             struct page *,
  79                                             unsigned int, unsigned int);
  80 static int nfs_wait_on_write_congestion(struct address_space *, int);
  81 static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int);
  82 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  83 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how);
  84 static const struct rpc_call_ops nfs_write_partial_ops;
  85 static const struct rpc_call_ops nfs_write_full_ops;
  86 static const struct rpc_call_ops nfs_commit_ops;
  87
  88 static kmem_cache_t *nfs_wdata_cachep;
  89 static mempool_t *nfs_wdata_mempool;
  90 static mempool_t *nfs_commit_mempool;
  91
  92 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  93
  94 struct nfs_write_data *nfs_commit_alloc(void)
  95 {
  96         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
  97
  98         if (p) {
  99                 memset(p, 0, sizeof(*p));
 100                 INIT_LIST_HEAD(&p->pages);
 101         }
 102         return p;
 103 }
 104
 105 void nfs_commit_rcu_free(struct rcu_head *head)
 106 {
 107         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 108         if (p && (p->pagevec != &p->page_array[0]))
 109                 kfree(p->pagevec);
 110         mempool_free(p, nfs_commit_mempool);
 111 }
 112
 113 void nfs_commit_free(struct nfs_write_data *wdata)
 114 {
 115         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
 116 }
 117
 118 struct nfs_write_data *nfs_writedata_alloc(size_t len)
 119 {
 120         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 121         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
 122
 123         if (p) {
 124                 memset(p, 0, sizeof(*p));
 125                 INIT_LIST_HEAD(&p->pages);
 126                 p->npages = pagecount;
 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_rcu_free(struct rcu_head *head)
 141 {
 142         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 143         if (p && (p->pagevec != &p->page_array[0]))
 144                 kfree(p->pagevec);
 145         mempool_free(p, nfs_wdata_mempool);
 146 }
 147
 148 static void nfs_writedata_free(struct nfs_write_data *wdata)
 149 {
 150         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 151 }
 152
 153 void nfs_writedata_release(void *wdata)
 154 {
 155         nfs_writedata_free(wdata);
 156 }
 157
 158 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 159 {
 160         struct nfs_page *req = NULL;
 161
 162         if (PagePrivate(page)) {
 163                 req = (struct nfs_page *)page_private(page);
 164                 if (req != NULL)
 165                         atomic_inc(&req->wb_count);
 166         }
 167         return req;
 168 }
 169
 170 static struct nfs_page *nfs_page_find_request(struct page *page)
 171 {
 172         struct nfs_page *req = NULL;
 173         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 174
 175         spin_lock(req_lock);
 176         req = nfs_page_find_request_locked(page);
 177         spin_unlock(req_lock);
 178         return req;
 179 }
 180
 181 /* Adjust the file length if we're writing beyond the end */
 182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         loff_t end, i_size = i_size_read(inode);
 186         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 187
 188         if (i_size > 0 && page->index < end_index)
 189                 return;
 190         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 191         if (i_size >= end)
 192                 return;
 193         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 194         i_size_write(inode, end);
 195 }
 196
 197 /* We can set the PG_uptodate flag if we see that a write request
 198  * covers the full page.
 199  */
 200 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 201 {
 202         if (PageUptodate(page))
 203                 return;
 204         if (base != 0)
 205                 return;
 206         if (count != nfs_page_length(page))
 207                 return;
 208         if (count != PAGE_CACHE_SIZE)
 209                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 210         SetPageUptodate(page);
 211 }
 212
 213 /*
 214  * Write a page synchronously.
 215  * Offset is the data offset within the page.
 216  */
 217 static int nfs_writepage_sync(struct nfs_open_context *ctx, struct page *page,
 218                 unsigned int offset, unsigned int count, int how)
 219 {
 220         struct inode *inode = page->mapping->host;
 221         unsigned int    wsize = NFS_SERVER(inode)->wsize;
 222         int             result, written = 0;
 223         struct nfs_write_data *wdata;
 224
 225         wdata = nfs_writedata_alloc(wsize);
 226         if (!wdata)
 227                 return -ENOMEM;
 228
 229         wdata->flags = how;
 230         wdata->cred = ctx->cred;
 231         wdata->inode = inode;
 232         wdata->args.fh = NFS_FH(inode);
 233         wdata->args.context = ctx;
 234         wdata->args.pages = &page;
 235         wdata->args.stable = NFS_FILE_SYNC;
 236         wdata->args.pgbase = offset;
 237         wdata->args.count = wsize;
 238         wdata->res.fattr = &wdata->fattr;
 239         wdata->res.verf = &wdata->verf;
 240
 241         dprintk("NFS:      nfs_writepage_sync(%s/%Ld %d@%Ld)\n",
 242                 inode->i_sb->s_id,
 243                 (long long)NFS_FILEID(inode),
 244                 count, (long long)(page_offset(page) + offset));
 245
 246         set_page_writeback(page);
 247         nfs_begin_data_update(inode);
 248         do {
 249                 if (count < wsize)
 250                         wdata->args.count = count;
 251                 wdata->args.offset = page_offset(page) + wdata->args.pgbase;
 252
 253                 result = NFS_PROTO(inode)->write(wdata);
 254
 255                 if (result < 0) {
 256                         /* Must mark the page invalid after I/O error */
 257                         ClearPageUptodate(page);
 258                         goto io_error;
 259                 }
 260                 if (result < wdata->args.count)
 261                         printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n",
 262                                         wdata->args.count, result);
 263
 264                 wdata->args.offset += result;
 265                 wdata->args.pgbase += result;
 266                 written += result;
 267                 count -= result;
 268                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 269         } while (count);
 270         /* Update file length */
 271         nfs_grow_file(page, offset, written);
 272         /* Set the PG_uptodate flag? */
 273         nfs_mark_uptodate(page, offset, written);
 274
 275         if (PageError(page))
 276                 ClearPageError(page);
 277
 278 io_error:
 279         nfs_end_data_update(inode);
 280         end_page_writeback(page);
 281         nfs_writedata_release(wdata);
 282         return written ? written : result;
 283 }
 284
 285 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 286                 unsigned int offset, unsigned int count)
 287 {
 288         struct nfs_page *req;
 289         int ret;
 290
 291         for (;;) {
 292                 req = nfs_update_request(ctx, page, offset, count);
 293                 if (!IS_ERR(req))
 294                         break;
 295                 ret = PTR_ERR(req);
 296                 if (ret != -EBUSY)
 297                         return ret;
 298                 ret = nfs_wb_page(page->mapping->host, page);
 299                 if (ret != 0)
 300                         return ret;
 301         }
 302         /* Update file length */
 303         nfs_grow_file(page, offset, count);
 304         /* Set the PG_uptodate flag? */
 305         nfs_mark_uptodate(page, offset, count);
 306         nfs_unlock_request(req);
 307         return 0;
 308 }
 309
 310 static int wb_priority(struct writeback_control *wbc)
 311 {
 312         if (wbc->for_reclaim)
 313                 return FLUSH_HIGHPRI;
 314         if (wbc->for_kupdate)
 315                 return FLUSH_LOWPRI;
 316         return 0;
 317 }
 318
 319 /*
 320  * Write an mmapped page to the server.
 321  */
 322 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 323 {
 324         struct nfs_open_context *ctx;
 325         struct inode *inode = page->mapping->host;
 326         unsigned offset;
 327         int err;
 328
 329         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 330         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 331
 332         /* Ensure we've flushed out any previous writes */
 333         nfs_wb_page_priority(inode, page, wb_priority(wbc));
 334
 335         err = 0;
 336         offset = nfs_page_length(page);
 337         if (!offset)
 338                 goto out;
 339
 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)) {
 347                 err = nfs_writepage_setup(ctx, page, 0, offset);
 348                 if (!wbc->for_writepages)
 349                         nfs_flush_mapping(page->mapping, wbc, wb_priority(wbc));
 350         } else {
 351                 err = nfs_writepage_sync(ctx, page, 0, offset, wb_priority(wbc));
 352                 if (err >= 0) {
 353                         if (err != offset)
 354                                 redirty_page_for_writepage(wbc, page);
 355                         err = 0;
 356                 }
 357         }
 358         unlock_kernel();
 359         put_nfs_open_context(ctx);
 360 out:
 361         unlock_page(page);
 362         return err;
 363 }
 364
 365 /*
 366  * Note: causes nfs_update_request() to block on the assumption
 367  *       that the writeback is generated due to memory pressure.
 368  */
 369 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 370 {
 371         struct backing_dev_info *bdi = mapping->backing_dev_info;
 372         struct inode *inode = mapping->host;
 373         int err;
 374
 375         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 376
 377         err = generic_writepages(mapping, wbc);
 378         if (err)
 379                 return err;
 380         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 381                 if (wbc->nonblocking)
 382                         return 0;
 383                 nfs_wait_on_write_congestion(mapping, 0);
 384         }
 385         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 386         if (err < 0)
 387                 goto out;
 388         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 389         if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) {
 390                 err = nfs_wait_on_requests(inode, 0, 0);
 391                 if (err < 0)
 392                         goto out;
 393         }
 394         err = nfs_commit_inode(inode, wb_priority(wbc));
 395         if (err > 0)
 396                 err = 0;
 397 out:
 398         clear_bit(BDI_write_congested, &bdi->state);
 399         wake_up_all(&nfs_write_congestion);
 400         congestion_end(WRITE);
 401         return err;
 402 }
 403
 404 /*
 405  * Insert a write request into an inode
 406  */
 407 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 408 {
 409         struct nfs_inode *nfsi = NFS_I(inode);
 410         int error;
 411
 412         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 413         BUG_ON(error == -EEXIST);
 414         if (error)
 415                 return error;
 416         if (!nfsi->npages) {
 417                 igrab(inode);
 418                 nfs_begin_data_update(inode);
 419                 if (nfs_have_delegation(inode, FMODE_WRITE))
 420                         nfsi->change_attr++;
 421         }
 422         SetPagePrivate(req->wb_page);
 423         set_page_private(req->wb_page, (unsigned long)req);
 424         nfsi->npages++;
 425         atomic_inc(&req->wb_count);
 426         return 0;
 427 }
 428
 429 /*
 430  * Insert a write request into an inode
 431  */
 432 static void nfs_inode_remove_request(struct nfs_page *req)
 433 {
 434         struct inode *inode = req->wb_context->dentry->d_inode;
 435         struct nfs_inode *nfsi = NFS_I(inode);
 436
 437         BUG_ON (!NFS_WBACK_BUSY(req));
 438
 439         spin_lock(&nfsi->req_lock);
 440         set_page_private(req->wb_page, 0);
 441         ClearPagePrivate(req->wb_page);
 442         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 443         nfsi->npages--;
 444         if (!nfsi->npages) {
 445                 spin_unlock(&nfsi->req_lock);
 446                 nfs_end_data_update(inode);
 447                 iput(inode);
 448         } else
 449                 spin_unlock(&nfsi->req_lock);
 450         nfs_clear_request(req);
 451         nfs_release_request(req);
 452 }
 453
 454 /*
 455  * Add a request to the inode's dirty list.
 456  */
 457 static void
 458 nfs_mark_request_dirty(struct nfs_page *req)
 459 {
 460         struct inode *inode = req->wb_context->dentry->d_inode;
 461         struct nfs_inode *nfsi = NFS_I(inode);
 462
 463         spin_lock(&nfsi->req_lock);
 464         radix_tree_tag_set(&nfsi->nfs_page_tree,
 465                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 466         nfs_list_add_request(req, &nfsi->dirty);
 467         nfsi->ndirty++;
 468         spin_unlock(&nfsi->req_lock);
 469         inc_zone_page_state(req->wb_page, NR_FILE_DIRTY);
 470         mark_inode_dirty(inode);
 471 }
 472
 473 /*
 474  * Check if a request is dirty
 475  */
 476 static inline int
 477 nfs_dirty_request(struct nfs_page *req)
 478 {
 479         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 480         return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty;
 481 }
 482
 483 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 484 /*
 485  * Add a request to the inode's commit list.
 486  */
 487 static void
 488 nfs_mark_request_commit(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         nfs_list_add_request(req, &nfsi->commit);
 495         nfsi->ncommit++;
 496         spin_unlock(&nfsi->req_lock);
 497         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 498         mark_inode_dirty(inode);
 499 }
 500 #endif
 501
 502 /*
 503  * Wait for a request to complete.
 504  *
 505  * Interruptible by signals only if mounted with intr flag.
 506  */
 507 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 508 {
 509         struct nfs_inode *nfsi = NFS_I(inode);
 510         struct nfs_page *req;
 511         unsigned long           idx_end, next;
 512         unsigned int            res = 0;
 513         int                     error;
 514
 515         if (npages == 0)
 516                 idx_end = ~0;
 517         else
 518                 idx_end = idx_start + npages - 1;
 519
 520         next = idx_start;
 521         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 522                 if (req->wb_index > idx_end)
 523                         break;
 524
 525                 next = req->wb_index + 1;
 526                 BUG_ON(!NFS_WBACK_BUSY(req));
 527
 528                 atomic_inc(&req->wb_count);
 529                 spin_unlock(&nfsi->req_lock);
 530                 error = nfs_wait_on_request(req);
 531                 nfs_release_request(req);
 532                 spin_lock(&nfsi->req_lock);
 533                 if (error < 0)
 534                         return error;
 535                 res++;
 536         }
 537         return res;
 538 }
 539
 540 static int nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages)
 541 {
 542         struct nfs_inode *nfsi = NFS_I(inode);
 543         int ret;
 544
 545         spin_lock(&nfsi->req_lock);
 546         ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
 547         spin_unlock(&nfsi->req_lock);
 548         return ret;
 549 }
 550
 551 static void nfs_cancel_dirty_list(struct list_head *head)
 552 {
 553         struct nfs_page *req;
 554         while(!list_empty(head)) {
 555                 req = nfs_list_entry(head->next);
 556                 nfs_list_remove_request(req);
 557                 nfs_inode_remove_request(req);
 558                 nfs_clear_page_writeback(req);
 559         }
 560 }
 561
 562 static void nfs_cancel_commit_list(struct list_head *head)
 563 {
 564         struct nfs_page *req;
 565
 566         while(!list_empty(head)) {
 567                 req = nfs_list_entry(head->next);
 568                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 569                 nfs_list_remove_request(req);
 570                 nfs_inode_remove_request(req);
 571                 nfs_unlock_request(req);
 572         }
 573 }
 574
 575 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 576 /*
 577  * nfs_scan_commit - Scan an inode for commit requests
 578  * @inode: NFS inode to scan
 579  * @dst: destination list
 580  * @idx_start: lower bound of page->index to scan.
 581  * @npages: idx_start + npages sets the upper bound to scan.
 582  *
 583  * Moves requests from the inode's 'commit' request list.
 584  * The requests are *not* checked to ensure that they form a contiguous set.
 585  */
 586 static int
 587 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 588 {
 589         struct nfs_inode *nfsi = NFS_I(inode);
 590         int res = 0;
 591
 592         if (nfsi->ncommit != 0) {
 593                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 594                 nfsi->ncommit -= res;
 595                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 596                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 597         }
 598         return res;
 599 }
 600 #else
 601 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 602 {
 603         return 0;
 604 }
 605 #endif
 606
 607 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 608 {
 609         struct backing_dev_info *bdi = mapping->backing_dev_info;
 610         DEFINE_WAIT(wait);
 611         int ret = 0;
 612
 613         might_sleep();
 614
 615         if (!bdi_write_congested(bdi))
 616                 return 0;
 617
 618         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 619
 620         if (intr) {
 621                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 622                 sigset_t oldset;
 623
 624                 rpc_clnt_sigmask(clnt, &oldset);
 625                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 626                 if (bdi_write_congested(bdi)) {
 627                         if (signalled())
 628                                 ret = -ERESTARTSYS;
 629                         else
 630                                 schedule();
 631                 }
 632                 rpc_clnt_sigunmask(clnt, &oldset);
 633         } else {
 634                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 635                 if (bdi_write_congested(bdi))
 636                         schedule();
 637         }
 638         finish_wait(&nfs_write_congestion, &wait);
 639         return ret;
 640 }
 641
 642
 643 /*
 644  * Try to update any existing write request, or create one if there is none.
 645  * In order to match, the request's credentials must match those of
 646  * the calling process.
 647  *
 648  * Note: Should always be called with the Page Lock held!
 649  */
 650 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 651                 struct page *page, unsigned int offset, unsigned int bytes)
 652 {
 653         struct inode *inode = page->mapping->host;
 654         struct nfs_inode *nfsi = NFS_I(inode);
 655         struct nfs_page         *req, *new = NULL;
 656         unsigned long           rqend, end;
 657
 658         end = offset + bytes;
 659
 660         if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
 661                 return ERR_PTR(-ERESTARTSYS);
 662         for (;;) {
 663                 /* Loop over all inode entries and see if we find
 664                  * A request for the page we wish to update
 665                  */
 666                 spin_lock(&nfsi->req_lock);
 667                 req = nfs_page_find_request_locked(page);
 668                 if (req) {
 669                         if (!nfs_lock_request_dontget(req)) {
 670                                 int error;
 671
 672                                 spin_unlock(&nfsi->req_lock);
 673                                 error = nfs_wait_on_request(req);
 674                                 nfs_release_request(req);
 675                                 if (error < 0) {
 676                                         if (new)
 677                                                 nfs_release_request(new);
 678                                         return ERR_PTR(error);
 679                                 }
 680                                 continue;
 681                         }
 682                         spin_unlock(&nfsi->req_lock);
 683                         if (new)
 684                                 nfs_release_request(new);
 685                         break;
 686                 }
 687
 688                 if (new) {
 689                         int error;
 690                         nfs_lock_request_dontget(new);
 691                         error = nfs_inode_add_request(inode, new);
 692                         if (error) {
 693                                 spin_unlock(&nfsi->req_lock);
 694                                 nfs_unlock_request(new);
 695                                 return ERR_PTR(error);
 696                         }
 697                         spin_unlock(&nfsi->req_lock);
 698                         nfs_mark_request_dirty(new);
 699                         return new;
 700                 }
 701                 spin_unlock(&nfsi->req_lock);
 702
 703                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 704                 if (IS_ERR(new))
 705                         return new;
 706         }
 707
 708         /* We have a request for our page.
 709          * If the creds don't match, or the
 710          * page addresses don't match,
 711          * tell the caller to wait on the conflicting
 712          * request.
 713          */
 714         rqend = req->wb_offset + req->wb_bytes;
 715         if (req->wb_context != ctx
 716             || req->wb_page != page
 717             || !nfs_dirty_request(req)
 718             || offset > rqend || end < req->wb_offset) {
 719                 nfs_unlock_request(req);
 720                 return ERR_PTR(-EBUSY);
 721         }
 722
 723         /* Okay, the request matches. Update the region */
 724         if (offset < req->wb_offset) {
 725                 req->wb_offset = offset;
 726                 req->wb_pgbase = offset;
 727                 req->wb_bytes = rqend - req->wb_offset;
 728         }
 729
 730         if (end > rqend)
 731                 req->wb_bytes = end - req->wb_offset;
 732
 733         return req;
 734 }
 735
 736 int nfs_flush_incompatible(struct file *file, struct page *page)
 737 {
 738         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 739         struct nfs_page *req;
 740         int             status = 0;
 741         /*
 742          * Look for a request corresponding to this page. If there
 743          * is one, and it belongs to another file, we flush it out
 744          * before we try to copy anything into the page. Do this
 745          * due to the lack of an ACCESS-type call in NFSv2.
 746          * Also do the same if we find a request from an existing
 747          * dropped page.
 748          */
 749         req = nfs_page_find_request(page);
 750         if (req != NULL) {
 751                 int do_flush = req->wb_page != page || req->wb_context != ctx;
 752
 753                 nfs_release_request(req);
 754                 if (do_flush)
 755                         status = nfs_wb_page(page->mapping->host, page);
 756         }
 757         return (status < 0) ? status : 0;
 758 }
 759
 760 /*
 761  * Update and possibly write a cached page of an NFS file.
 762  *
 763  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 764  * things with a page scheduled for an RPC call (e.g. invalidate it).
 765  */
 766 int nfs_updatepage(struct file *file, struct page *page,
 767                 unsigned int offset, unsigned int count)
 768 {
 769         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 770         struct inode    *inode = page->mapping->host;
 771         int             status = 0;
 772
 773         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 774
 775         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 776                 file->f_dentry->d_parent->d_name.name,
 777                 file->f_dentry->d_name.name, count,
 778                 (long long)(page_offset(page) +offset));
 779
 780         if (IS_SYNC(inode)) {
 781                 status = nfs_writepage_sync(ctx, page, offset, count, 0);
 782                 if (status > 0) {
 783                         if (offset == 0 && status == PAGE_CACHE_SIZE)
 784                                 SetPageUptodate(page);
 785                         return 0;
 786                 }
 787                 return status;
 788         }
 789
 790         /* If we're not using byte range locks, and we know the page
 791          * is entirely in cache, it may be more efficient to avoid
 792          * fragmenting write requests.
 793          */
 794         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 795                 count = max(count + offset, nfs_page_length(page));
 796                 offset = 0;
 797         }
 798
 799         status = nfs_writepage_setup(ctx, page, offset, count);
 800
 801         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 802                         status, (long long)i_size_read(inode));
 803         if (status < 0)
 804                 ClearPageUptodate(page);
 805         return status;
 806 }
 807
 808 static void nfs_writepage_release(struct nfs_page *req)
 809 {
 810         end_page_writeback(req->wb_page);
 811
 812 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 813         if (!PageError(req->wb_page)) {
 814                 if (NFS_NEED_RESCHED(req)) {
 815                         nfs_mark_request_dirty(req);
 816                         goto out;
 817                 } else if (NFS_NEED_COMMIT(req)) {
 818                         nfs_mark_request_commit(req);
 819                         goto out;
 820                 }
 821         }
 822         nfs_inode_remove_request(req);
 823
 824 out:
 825         nfs_clear_commit(req);
 826         nfs_clear_reschedule(req);
 827 #else
 828         nfs_inode_remove_request(req);
 829 #endif
 830         nfs_clear_page_writeback(req);
 831 }
 832
 833 static inline int flush_task_priority(int how)
 834 {
 835         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 836                 case FLUSH_HIGHPRI:
 837                         return RPC_PRIORITY_HIGH;
 838                 case FLUSH_LOWPRI:
 839                         return RPC_PRIORITY_LOW;
 840         }
 841         return RPC_PRIORITY_NORMAL;
 842 }
 843
 844 /*
 845  * Set up the argument/result storage required for the RPC call.
 846  */
 847 static void nfs_write_rpcsetup(struct nfs_page *req,
 848                 struct nfs_write_data *data,
 849                 const struct rpc_call_ops *call_ops,
 850                 unsigned int count, unsigned int offset,
 851                 int how)
 852 {
 853         struct inode            *inode;
 854         int flags;
 855
 856         /* Set up the RPC argument and reply structs
 857          * NB: take care not to mess about with data->commit et al. */
 858
 859         data->req = req;
 860         data->inode = inode = req->wb_context->dentry->d_inode;
 861         data->cred = req->wb_context->cred;
 862
 863         data->args.fh     = NFS_FH(inode);
 864         data->args.offset = req_offset(req) + offset;
 865         data->args.pgbase = req->wb_pgbase + offset;
 866         data->args.pages  = data->pagevec;
 867         data->args.count  = count;
 868         data->args.context = req->wb_context;
 869
 870         data->res.fattr   = &data->fattr;
 871         data->res.count   = count;
 872         data->res.verf    = &data->verf;
 873         nfs_fattr_init(&data->fattr);
 874
 875         /* Set up the initial task struct.  */
 876         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 877         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 878         NFS_PROTO(inode)->write_setup(data, how);
 879
 880         data->task.tk_priority = flush_task_priority(how);
 881         data->task.tk_cookie = (unsigned long)inode;
 882
 883         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 884                 data->task.tk_pid,
 885                 inode->i_sb->s_id,
 886                 (long long)NFS_FILEID(inode),
 887                 count,
 888                 (unsigned long long)data->args.offset);
 889 }
 890
 891 static void nfs_execute_write(struct nfs_write_data *data)
 892 {
 893         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 894         sigset_t oldset;
 895
 896         rpc_clnt_sigmask(clnt, &oldset);
 897         rpc_execute(&data->task);
 898         rpc_clnt_sigunmask(clnt, &oldset);
 899 }
 900
 901 /*
 902  * Generate multiple small requests to write out a single
 903  * contiguous dirty area on one page.
 904  */
 905 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 906 {
 907         struct nfs_page *req = nfs_list_entry(head->next);
 908         struct page *page = req->wb_page;
 909         struct nfs_write_data *data;
 910         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 911         unsigned int offset;
 912         int requests = 0;
 913         LIST_HEAD(list);
 914
 915         nfs_list_remove_request(req);
 916
 917         nbytes = req->wb_bytes;
 918         do {
 919                 size_t len = min(nbytes, wsize);
 920
 921                 data = nfs_writedata_alloc(len);
 922                 if (!data)
 923                         goto out_bad;
 924                 list_add(&data->pages, &list);
 925                 requests++;
 926                 nbytes -= len;
 927         } while (nbytes != 0);
 928         atomic_set(&req->wb_complete, requests);
 929
 930         ClearPageError(page);
 931         set_page_writeback(page);
 932         offset = 0;
 933         nbytes = req->wb_bytes;
 934         do {
 935                 data = list_entry(list.next, struct nfs_write_data, pages);
 936                 list_del_init(&data->pages);
 937
 938                 data->pagevec[0] = page;
 939
 940                 if (nbytes > wsize) {
 941                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 942                                         wsize, offset, how);
 943                         offset += wsize;
 944                         nbytes -= wsize;
 945                 } else {
 946                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 947                                         nbytes, offset, how);
 948                         nbytes = 0;
 949                 }
 950                 nfs_execute_write(data);
 951         } while (nbytes != 0);
 952
 953         return 0;
 954
 955 out_bad:
 956         while (!list_empty(&list)) {
 957                 data = list_entry(list.next, struct nfs_write_data, pages);
 958                 list_del(&data->pages);
 959                 nfs_writedata_release(data);
 960         }
 961         nfs_mark_request_dirty(req);
 962         nfs_clear_page_writeback(req);
 963         return -ENOMEM;
 964 }
 965
 966 /*
 967  * Create an RPC task for the given write request and kick it.
 968  * The page must have been locked by the caller.
 969  *
 970  * It may happen that the page we're passed is not marked dirty.
 971  * This is the case if nfs_updatepage detects a conflicting request
 972  * that has been written but not committed.
 973  */
 974 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 975 {
 976         struct nfs_page         *req;
 977         struct page             **pages;
 978         struct nfs_write_data   *data;
 979         unsigned int            count;
 980
 981         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 982         if (!data)
 983                 goto out_bad;
 984
 985         pages = data->pagevec;
 986         count = 0;
 987         while (!list_empty(head)) {
 988                 req = nfs_list_entry(head->next);
 989                 nfs_list_remove_request(req);
 990                 nfs_list_add_request(req, &data->pages);
 991                 ClearPageError(req->wb_page);
 992                 set_page_writeback(req->wb_page);
 993                 *pages++ = req->wb_page;
 994                 count += req->wb_bytes;
 995         }
 996         req = nfs_list_entry(data->pages.next);
 997
 998         /* Set up the argument struct */
 999         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
1000
1001         nfs_execute_write(data);
1002         return 0;
1003  out_bad:
1004         while (!list_empty(head)) {
1005                 struct nfs_page *req = nfs_list_entry(head->next);
1006                 nfs_list_remove_request(req);
1007                 nfs_mark_request_dirty(req);
1008                 nfs_clear_page_writeback(req);
1009         }
1010         return -ENOMEM;
1011 }
1012
1013 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
1014 {
1015         LIST_HEAD(one_request);
1016         int (*flush_one)(struct inode *, struct list_head *, int);
1017         struct nfs_page *req;
1018         int wpages = NFS_SERVER(inode)->wpages;
1019         int wsize = NFS_SERVER(inode)->wsize;
1020         int error;
1021
1022         flush_one = nfs_flush_one;
1023         if (wsize < PAGE_CACHE_SIZE)
1024                 flush_one = nfs_flush_multi;
1025         /* For single writes, FLUSH_STABLE is more efficient */
1026         if (npages <= wpages && npages == NFS_I(inode)->npages
1027                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
1028                 how |= FLUSH_STABLE;
1029
1030         do {
1031                 nfs_coalesce_requests(head, &one_request, wpages);
1032                 req = nfs_list_entry(one_request.next);
1033                 error = flush_one(inode, &one_request, how);
1034                 if (error < 0)
1035                         goto out_err;
1036         } while (!list_empty(head));
1037         return 0;
1038 out_err:
1039         while (!list_empty(head)) {
1040                 req = nfs_list_entry(head->next);
1041                 nfs_list_remove_request(req);
1042                 nfs_mark_request_dirty(req);
1043                 nfs_clear_page_writeback(req);
1044         }
1045         return error;
1046 }
1047
1048 /*
1049  * Handle a write reply that flushed part of a page.
1050  */
1051 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1052 {
1053         struct nfs_write_data   *data = calldata;
1054         struct nfs_page         *req = data->req;
1055         struct page             *page = req->wb_page;
1056
1057         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1058                 req->wb_context->dentry->d_inode->i_sb->s_id,
1059                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1060                 req->wb_bytes,
1061                 (long long)req_offset(req));
1062
1063         if (nfs_writeback_done(task, data) != 0)
1064                 return;
1065
1066         if (task->tk_status < 0) {
1067                 ClearPageUptodate(page);
1068                 SetPageError(page);
1069                 req->wb_context->error = task->tk_status;
1070                 dprintk(", error = %d\n", task->tk_status);
1071         } else {
1072 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1073                 if (data->verf.committed < NFS_FILE_SYNC) {
1074                         if (!NFS_NEED_COMMIT(req)) {
1075                                 nfs_defer_commit(req);
1076                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1077                                 dprintk(" defer commit\n");
1078                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1079                                 nfs_defer_reschedule(req);
1080                                 dprintk(" server reboot detected\n");
1081                         }
1082                 } else
1083 #endif
1084                         dprintk(" OK\n");
1085         }
1086
1087         if (atomic_dec_and_test(&req->wb_complete))
1088                 nfs_writepage_release(req);
1089 }
1090
1091 static const struct rpc_call_ops nfs_write_partial_ops = {
1092         .rpc_call_done = nfs_writeback_done_partial,
1093         .rpc_release = nfs_writedata_release,
1094 };
1095
1096 /*
1097  * Handle a write reply that flushes a whole page.
1098  *
1099  * FIXME: There is an inherent race with invalidate_inode_pages and
1100  *        writebacks since the page->count is kept > 1 for as long
1101  *        as the page has a write request pending.
1102  */
1103 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1104 {
1105         struct nfs_write_data   *data = calldata;
1106         struct nfs_page         *req;
1107         struct page             *page;
1108
1109         if (nfs_writeback_done(task, data) != 0)
1110                 return;
1111
1112         /* Update attributes as result of writeback. */
1113         while (!list_empty(&data->pages)) {
1114                 req = nfs_list_entry(data->pages.next);
1115                 nfs_list_remove_request(req);
1116                 page = req->wb_page;
1117
1118                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1119                         req->wb_context->dentry->d_inode->i_sb->s_id,
1120                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1121                         req->wb_bytes,
1122                         (long long)req_offset(req));
1123
1124                 if (task->tk_status < 0) {
1125                         ClearPageUptodate(page);
1126                         SetPageError(page);
1127                         req->wb_context->error = task->tk_status;
1128                         end_page_writeback(page);
1129                         nfs_inode_remove_request(req);
1130                         dprintk(", error = %d\n", task->tk_status);
1131                         goto next;
1132                 }
1133                 end_page_writeback(page);
1134
1135 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1136                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1137                         nfs_inode_remove_request(req);
1138                         dprintk(" OK\n");
1139                         goto next;
1140                 }
1141                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1142                 nfs_mark_request_commit(req);
1143                 dprintk(" marked for commit\n");
1144 #else
1145                 nfs_inode_remove_request(req);
1146 #endif
1147         next:
1148                 nfs_clear_page_writeback(req);
1149         }
1150 }
1151
1152 static const struct rpc_call_ops nfs_write_full_ops = {
1153         .rpc_call_done = nfs_writeback_done_full,
1154         .rpc_release = nfs_writedata_release,
1155 };
1156
1157
1158 /*
1159  * This function is called when the WRITE call is complete.
1160  */
1161 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1162 {
1163         struct nfs_writeargs    *argp = &data->args;
1164         struct nfs_writeres     *resp = &data->res;
1165         int status;
1166
1167         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1168                 task->tk_pid, task->tk_status);
1169
1170         /*
1171          * ->write_done will attempt to use post-op attributes to detect
1172          * conflicting writes by other clients.  A strict interpretation
1173          * of close-to-open would allow us to continue caching even if
1174          * another writer had changed the file, but some applications
1175          * depend on tighter cache coherency when writing.
1176          */
1177         status = NFS_PROTO(data->inode)->write_done(task, data);
1178         if (status != 0)
1179                 return status;
1180         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1181
1182 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1183         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1184                 /* We tried a write call, but the server did not
1185                  * commit data to stable storage even though we
1186                  * requested it.
1187                  * Note: There is a known bug in Tru64 < 5.0 in which
1188                  *       the server reports NFS_DATA_SYNC, but performs
1189                  *       NFS_FILE_SYNC. We therefore implement this checking
1190                  *       as a dprintk() in order to avoid filling syslog.
1191                  */
1192                 static unsigned long    complain;
1193
1194                 if (time_before(complain, jiffies)) {
1195                         dprintk("NFS: faulty NFS server %s:"
1196                                 " (committed = %d) != (stable = %d)\n",
1197                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1198                                 resp->verf->committed, argp->stable);
1199                         complain = jiffies + 300 * HZ;
1200                 }
1201         }
1202 #endif
1203         /* Is this a short write? */
1204         if (task->tk_status >= 0 && resp->count < argp->count) {
1205                 static unsigned long    complain;
1206
1207                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1208
1209                 /* Has the server at least made some progress? */
1210                 if (resp->count != 0) {
1211                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1212                         if (resp->verf->committed != NFS_UNSTABLE) {
1213                                 /* Resend from where the server left off */
1214                                 argp->offset += resp->count;
1215                                 argp->pgbase += resp->count;
1216                                 argp->count -= resp->count;
1217                         } else {
1218                                 /* Resend as a stable write in order to avoid
1219                                  * headaches in the case of a server crash.
1220                                  */
1221                                 argp->stable = NFS_FILE_SYNC;
1222                         }
1223                         rpc_restart_call(task);
1224                         return -EAGAIN;
1225                 }
1226                 if (time_before(complain, jiffies)) {
1227                         printk(KERN_WARNING
1228                                "NFS: Server wrote zero bytes, expected %u.\n",
1229                                         argp->count);
1230                         complain = jiffies + 300 * HZ;
1231                 }
1232                 /* Can't do anything about it except throw an error. */
1233                 task->tk_status = -EIO;
1234         }
1235         return 0;
1236 }
1237
1238
1239 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1240 void nfs_commit_release(void *wdata)
1241 {
1242         nfs_commit_free(wdata);
1243 }
1244
1245 /*
1246  * Set up the argument/result storage required for the RPC call.
1247  */
1248 static void nfs_commit_rpcsetup(struct list_head *head,
1249                 struct nfs_write_data *data,
1250                 int how)
1251 {
1252         struct nfs_page         *first;
1253         struct inode            *inode;
1254         int flags;
1255
1256         /* Set up the RPC argument and reply structs
1257          * NB: take care not to mess about with data->commit et al. */
1258
1259         list_splice_init(head, &data->pages);
1260         first = nfs_list_entry(data->pages.next);
1261         inode = first->wb_context->dentry->d_inode;
1262
1263         data->inode       = inode;
1264         data->cred        = first->wb_context->cred;
1265
1266         data->args.fh     = NFS_FH(data->inode);
1267         /* Note: we always request a commit of the entire inode */
1268         data->args.offset = 0;
1269         data->args.count  = 0;
1270         data->res.count   = 0;
1271         data->res.fattr   = &data->fattr;
1272         data->res.verf    = &data->verf;
1273         nfs_fattr_init(&data->fattr);
1274
1275         /* Set up the initial task struct.  */
1276         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1277         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1278         NFS_PROTO(inode)->commit_setup(data, how);
1279
1280         data->task.tk_priority = flush_task_priority(how);
1281         data->task.tk_cookie = (unsigned long)inode;
1282
1283         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1284 }
1285
1286 /*
1287  * Commit dirty pages
1288  */
1289 static int
1290 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1291 {
1292         struct nfs_write_data   *data;
1293         struct nfs_page         *req;
1294
1295         data = nfs_commit_alloc();
1296
1297         if (!data)
1298                 goto out_bad;
1299
1300         /* Set up the argument struct */
1301         nfs_commit_rpcsetup(head, data, how);
1302
1303         nfs_execute_write(data);
1304         return 0;
1305  out_bad:
1306         while (!list_empty(head)) {
1307                 req = nfs_list_entry(head->next);
1308                 nfs_list_remove_request(req);
1309                 nfs_mark_request_commit(req);
1310                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1311                 nfs_clear_page_writeback(req);
1312         }
1313         return -ENOMEM;
1314 }
1315
1316 /*
1317  * COMMIT call returned
1318  */
1319 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1320 {
1321         struct nfs_write_data   *data = calldata;
1322         struct nfs_page         *req;
1323
1324         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1325                                 task->tk_pid, task->tk_status);
1326
1327         /* Call the NFS version-specific code */
1328         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1329                 return;
1330
1331         while (!list_empty(&data->pages)) {
1332                 req = nfs_list_entry(data->pages.next);
1333                 nfs_list_remove_request(req);
1334                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1335
1336                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1337                         req->wb_context->dentry->d_inode->i_sb->s_id,
1338                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1339                         req->wb_bytes,
1340                         (long long)req_offset(req));
1341                 if (task->tk_status < 0) {
1342                         req->wb_context->error = task->tk_status;
1343                         nfs_inode_remove_request(req);
1344                         dprintk(", error = %d\n", task->tk_status);
1345                         goto next;
1346                 }
1347
1348                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1349                  * returned by the server against all stored verfs. */
1350                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1351                         /* We have a match */
1352                         nfs_inode_remove_request(req);
1353                         dprintk(" OK\n");
1354                         goto next;
1355                 }
1356                 /* We have a mismatch. Write the page again */
1357                 dprintk(" mismatch\n");
1358                 nfs_mark_request_dirty(req);
1359         next:
1360                 nfs_clear_page_writeback(req);
1361         }
1362 }
1363
1364 static const struct rpc_call_ops nfs_commit_ops = {
1365         .rpc_call_done = nfs_commit_done,
1366         .rpc_release = nfs_commit_release,
1367 };
1368 #else
1369 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1370 {
1371         return 0;
1372 }
1373 #endif
1374
1375 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1376 {
1377         struct nfs_inode *nfsi = NFS_I(mapping->host);
1378         LIST_HEAD(head);
1379         long res;
1380
1381         spin_lock(&nfsi->req_lock);
1382         res = nfs_scan_dirty(mapping, wbc, &head);
1383         spin_unlock(&nfsi->req_lock);
1384         if (res) {
1385                 int error = nfs_flush_list(mapping->host, &head, res, how);
1386                 if (error < 0)
1387                         return error;
1388         }
1389         return res;
1390 }
1391
1392 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1393 int nfs_commit_inode(struct inode *inode, int how)
1394 {
1395         struct nfs_inode *nfsi = NFS_I(inode);
1396         LIST_HEAD(head);
1397         int res;
1398
1399         spin_lock(&nfsi->req_lock);
1400         res = nfs_scan_commit(inode, &head, 0, 0);
1401         spin_unlock(&nfsi->req_lock);
1402         if (res) {
1403                 int error = nfs_commit_list(inode, &head, how);
1404                 if (error < 0)
1405                         return error;
1406         }
1407         return res;
1408 }
1409 #endif
1410
1411 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1412 {
1413         struct inode *inode = mapping->host;
1414         struct nfs_inode *nfsi = NFS_I(inode);
1415         unsigned long idx_start, idx_end;
1416         unsigned int npages = 0;
1417         LIST_HEAD(head);
1418         int nocommit = how & FLUSH_NOCOMMIT;
1419         long pages, ret;
1420
1421         /* FIXME */
1422         if (wbc->range_cyclic)
1423                 idx_start = 0;
1424         else {
1425                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1426                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1427                 if (idx_end > idx_start) {
1428                         unsigned long l_npages = 1 + idx_end - idx_start;
1429                         npages = l_npages;
1430                         if (sizeof(npages) != sizeof(l_npages) &&
1431                                         (unsigned long)npages != l_npages)
1432                                 npages = 0;
1433                 }
1434         }
1435         how &= ~FLUSH_NOCOMMIT;
1436         spin_lock(&nfsi->req_lock);
1437         do {
1438                 wbc->pages_skipped = 0;
1439                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1440                 if (ret != 0)
1441                         continue;
1442                 pages = nfs_scan_dirty(mapping, wbc, &head);
1443                 if (pages != 0) {
1444                         spin_unlock(&nfsi->req_lock);
1445                         if (how & FLUSH_INVALIDATE) {
1446                                 nfs_cancel_dirty_list(&head);
1447                                 ret = pages;
1448                         } else
1449                                 ret = nfs_flush_list(inode, &head, pages, how);
1450                         spin_lock(&nfsi->req_lock);
1451                         continue;
1452                 }
1453                 if (wbc->pages_skipped != 0)
1454                         continue;
1455                 if (nocommit)
1456                         break;
1457                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1458                 if (pages == 0) {
1459                         if (wbc->pages_skipped != 0)
1460                                 continue;
1461                         break;
1462                 }
1463                 if (how & FLUSH_INVALIDATE) {
1464                         spin_unlock(&nfsi->req_lock);
1465                         nfs_cancel_commit_list(&head);
1466                         ret = pages;
1467                         spin_lock(&nfsi->req_lock);
1468                         continue;
1469                 }
1470                 pages += nfs_scan_commit(inode, &head, 0, 0);
1471                 spin_unlock(&nfsi->req_lock);
1472                 ret = nfs_commit_list(inode, &head, how);
1473                 spin_lock(&nfsi->req_lock);
1474         } while (ret >= 0);
1475         spin_unlock(&nfsi->req_lock);
1476         return ret;
1477 }
1478
1479 /*
1480  * flush the inode to disk.
1481  */
1482 int nfs_wb_all(struct inode *inode)
1483 {
1484         struct address_space *mapping = inode->i_mapping;
1485         struct writeback_control wbc = {
1486                 .bdi = mapping->backing_dev_info,
1487                 .sync_mode = WB_SYNC_ALL,
1488                 .nr_to_write = LONG_MAX,
1489                 .range_cyclic = 1,
1490         };
1491         int ret;
1492
1493         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1494         if (ret >= 0)
1495                 return 0;
1496         return ret;
1497 }
1498
1499 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1500 {
1501         struct writeback_control wbc = {
1502                 .bdi = mapping->backing_dev_info,
1503                 .sync_mode = WB_SYNC_ALL,
1504                 .nr_to_write = LONG_MAX,
1505                 .range_start = range_start,
1506                 .range_end = range_end,
1507         };
1508         int ret;
1509
1510         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1511         if (ret >= 0)
1512                 return 0;
1513         return ret;
1514 }
1515
1516 static int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1517 {
1518         loff_t range_start = page_offset(page);
1519         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1520
1521         return nfs_sync_mapping_range(inode->i_mapping, range_start, range_end, how | FLUSH_STABLE);
1522 }
1523
1524 /*
1525  * Write back all requests on one page - we do this before reading it.
1526  */
1527 int nfs_wb_page(struct inode *inode, struct page* page)
1528 {
1529         return nfs_wb_page_priority(inode, page, 0);
1530 }
1531
1532
1533 int __init nfs_init_writepagecache(void)
1534 {
1535         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1536                                              sizeof(struct nfs_write_data),
1537                                              0, SLAB_HWCACHE_ALIGN,
1538                                              NULL, NULL);
1539         if (nfs_wdata_cachep == NULL)
1540                 return -ENOMEM;
1541
1542         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1543                                                      nfs_wdata_cachep);
1544         if (nfs_wdata_mempool == NULL)
1545                 return -ENOMEM;
1546
1547         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1548                                                       nfs_wdata_cachep);
1549         if (nfs_commit_mempool == NULL)
1550                 return -ENOMEM;
1551
1552         return 0;
1553 }
1554
1555 void nfs_destroy_writepagecache(void)
1556 {
1557         mempool_destroy(nfs_commit_mempool);
1558         mempool_destroy(nfs_wdata_mempool);
1559         kmem_cache_destroy(nfs_wdata_cachep);
1560 }
1561