err.no Git - linux-2.6/blob - fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15
  16 #include <linux/sunrpc/clnt.h>
  17 #include <linux/nfs_fs.h>
  18 #include <linux/nfs_mount.h>
  19 #include <linux/nfs_page.h>
  20 #include <linux/backing-dev.h>
  21
  22 #include <asm/uaccess.h>
  23 #include <linux/smp_lock.h>
  24
  25 #include "delegation.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 #define MIN_POOL_WRITE          (32)
  32 #define MIN_POOL_COMMIT         (4)
  33
  34 /*
  35  * Local function declarations
  36  */
  37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  38                                             struct page *,
  39                                             unsigned int, unsigned int);
  40 static void nfs_mark_request_dirty(struct nfs_page *req);
  41 static int nfs_wait_on_write_congestion(struct address_space *, int);
  42 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  43 static const struct rpc_call_ops nfs_write_partial_ops;
  44 static const struct rpc_call_ops nfs_write_full_ops;
  45 static const struct rpc_call_ops nfs_commit_ops;
  46
  47 static struct kmem_cache *nfs_wdata_cachep;
  48 static mempool_t *nfs_wdata_mempool;
  49 static mempool_t *nfs_commit_mempool;
  50
  51 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
  52
  53 struct nfs_write_data *nfs_commit_alloc(void)
  54 {
  55         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  56
  57         if (p) {
  58                 memset(p, 0, sizeof(*p));
  59                 INIT_LIST_HEAD(&p->pages);
  60         }
  61         return p;
  62 }
  63
  64 void nfs_commit_rcu_free(struct rcu_head *head)
  65 {
  66         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  67         if (p && (p->pagevec != &p->page_array[0]))
  68                 kfree(p->pagevec);
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71
  72 void nfs_commit_free(struct nfs_write_data *wdata)
  73 {
  74         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  75 }
  76
  77 struct nfs_write_data *nfs_writedata_alloc(size_t len)
  78 {
  79         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
  80         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  81
  82         if (p) {
  83                 memset(p, 0, sizeof(*p));
  84                 INIT_LIST_HEAD(&p->pages);
  85                 p->npages = pagecount;
  86                 if (pagecount <= ARRAY_SIZE(p->page_array))
  87                         p->pagevec = p->page_array;
  88                 else {
  89                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  90                         if (!p->pagevec) {
  91                                 mempool_free(p, nfs_wdata_mempool);
  92                                 p = NULL;
  93                         }
  94                 }
  95         }
  96         return p;
  97 }
  98
  99 static void nfs_writedata_rcu_free(struct rcu_head *head)
 100 {
 101         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 102         if (p && (p->pagevec != &p->page_array[0]))
 103                 kfree(p->pagevec);
 104         mempool_free(p, nfs_wdata_mempool);
 105 }
 106
 107 static void nfs_writedata_free(struct nfs_write_data *wdata)
 108 {
 109         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 110 }
 111
 112 void nfs_writedata_release(void *wdata)
 113 {
 114         nfs_writedata_free(wdata);
 115 }
 116
 117 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 118 {
 119         struct nfs_page *req = NULL;
 120
 121         if (PagePrivate(page)) {
 122                 req = (struct nfs_page *)page_private(page);
 123                 if (req != NULL)
 124                         atomic_inc(&req->wb_count);
 125         }
 126         return req;
 127 }
 128
 129 static struct nfs_page *nfs_page_find_request(struct page *page)
 130 {
 131         struct nfs_page *req = NULL;
 132         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 133
 134         spin_lock(req_lock);
 135         req = nfs_page_find_request_locked(page);
 136         spin_unlock(req_lock);
 137         return req;
 138 }
 139
 140 /* Adjust the file length if we're writing beyond the end */
 141 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 142 {
 143         struct inode *inode = page->mapping->host;
 144         loff_t end, i_size = i_size_read(inode);
 145         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 146
 147         if (i_size > 0 && page->index < end_index)
 148                 return;
 149         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 150         if (i_size >= end)
 151                 return;
 152         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 153         i_size_write(inode, end);
 154 }
 155
 156 /* We can set the PG_uptodate flag if we see that a write request
 157  * covers the full page.
 158  */
 159 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 160 {
 161         if (PageUptodate(page))
 162                 return;
 163         if (base != 0)
 164                 return;
 165         if (count != nfs_page_length(page))
 166                 return;
 167         if (count != PAGE_CACHE_SIZE)
 168                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 169         SetPageUptodate(page);
 170 }
 171
 172 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 173                 unsigned int offset, unsigned int count)
 174 {
 175         struct nfs_page *req;
 176         int ret;
 177
 178         for (;;) {
 179                 req = nfs_update_request(ctx, page, offset, count);
 180                 if (!IS_ERR(req))
 181                         break;
 182                 ret = PTR_ERR(req);
 183                 if (ret != -EBUSY)
 184                         return ret;
 185                 ret = nfs_wb_page(page->mapping->host, page);
 186                 if (ret != 0)
 187                         return ret;
 188         }
 189         /* Update file length */
 190         nfs_grow_file(page, offset, count);
 191         /* Set the PG_uptodate flag? */
 192         nfs_mark_uptodate(page, offset, count);
 193         nfs_unlock_request(req);
 194         return 0;
 195 }
 196
 197 static int wb_priority(struct writeback_control *wbc)
 198 {
 199         if (wbc->for_reclaim)
 200                 return FLUSH_HIGHPRI;
 201         if (wbc->for_kupdate)
 202                 return FLUSH_LOWPRI;
 203         return 0;
 204 }
 205
 206 /*
 207  * Find an associated nfs write request, and prepare to flush it out
 208  * Returns 1 if there was no write request, or if the request was
 209  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 210  * was not tagged.
 211  * May also return an error if the user signalled nfs_wait_on_request().
 212  */
 213 static int nfs_page_mark_flush(struct page *page)
 214 {
 215         struct nfs_page *req;
 216         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 217         int ret;
 218
 219         spin_lock(req_lock);
 220         for(;;) {
 221                 req = nfs_page_find_request_locked(page);
 222                 if (req == NULL) {
 223                         spin_unlock(req_lock);
 224                         return 1;
 225                 }
 226                 if (nfs_lock_request_dontget(req))
 227                         break;
 228                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 229                  *       then the call to nfs_lock_request_dontget() will always
 230                  *       succeed provided that someone hasn't already marked the
 231                  *       request as dirty (in which case we don't care).
 232                  */
 233                 spin_unlock(req_lock);
 234                 ret = nfs_wait_on_request(req);
 235                 nfs_release_request(req);
 236                 if (ret != 0)
 237                         return ret;
 238                 spin_lock(req_lock);
 239         }
 240         spin_unlock(req_lock);
 241         if (test_and_set_bit(PG_FLUSHING, &req->wb_flags) == 0) {
 242                 nfs_mark_request_dirty(req);
 243                 set_page_writeback(page);
 244         }
 245         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 246         nfs_unlock_request(req);
 247         return ret;
 248 }
 249
 250 /*
 251  * Write an mmapped page to the server.
 252  */
 253 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 254 {
 255         struct nfs_open_context *ctx;
 256         struct inode *inode = page->mapping->host;
 257         unsigned offset;
 258         int err;
 259
 260         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 261         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 262
 263         err = nfs_page_mark_flush(page);
 264         if (err <= 0)
 265                 goto out;
 266         err = 0;
 267         offset = nfs_page_length(page);
 268         if (!offset)
 269                 goto out;
 270
 271         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 272         if (ctx == NULL) {
 273                 err = -EBADF;
 274                 goto out;
 275         }
 276         err = nfs_writepage_setup(ctx, page, 0, offset);
 277         put_nfs_open_context(ctx);
 278         if (err != 0)
 279                 goto out;
 280         err = nfs_page_mark_flush(page);
 281         if (err > 0)
 282                 err = 0;
 283 out:
 284         if (!wbc->for_writepages)
 285                 nfs_flush_mapping(page->mapping, wbc, FLUSH_STABLE|wb_priority(wbc));
 286         return err;
 287 }
 288
 289 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 290 {
 291         int err;
 292
 293         err = nfs_writepage_locked(page, wbc);
 294         unlock_page(page);
 295         return err;
 296 }
 297
 298 /*
 299  * Note: causes nfs_update_request() to block on the assumption
 300  *       that the writeback is generated due to memory pressure.
 301  */
 302 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 303 {
 304         struct backing_dev_info *bdi = mapping->backing_dev_info;
 305         struct inode *inode = mapping->host;
 306         int err;
 307
 308         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 309
 310         err = generic_writepages(mapping, wbc);
 311         if (err)
 312                 return err;
 313         while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) {
 314                 if (wbc->nonblocking)
 315                         return 0;
 316                 nfs_wait_on_write_congestion(mapping, 0);
 317         }
 318         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 319         if (err < 0)
 320                 goto out;
 321         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 322         err = 0;
 323 out:
 324         clear_bit(BDI_write_congested, &bdi->state);
 325         wake_up_all(&nfs_write_congestion);
 326         congestion_end(WRITE);
 327         return err;
 328 }
 329
 330 /*
 331  * Insert a write request into an inode
 332  */
 333 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 334 {
 335         struct nfs_inode *nfsi = NFS_I(inode);
 336         int error;
 337
 338         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 339         BUG_ON(error == -EEXIST);
 340         if (error)
 341                 return error;
 342         if (!nfsi->npages) {
 343                 igrab(inode);
 344                 nfs_begin_data_update(inode);
 345                 if (nfs_have_delegation(inode, FMODE_WRITE))
 346                         nfsi->change_attr++;
 347         }
 348         SetPagePrivate(req->wb_page);
 349         set_page_private(req->wb_page, (unsigned long)req);
 350         nfsi->npages++;
 351         atomic_inc(&req->wb_count);
 352         return 0;
 353 }
 354
 355 /*
 356  * Insert a write request into an inode
 357  */
 358 static void nfs_inode_remove_request(struct nfs_page *req)
 359 {
 360         struct inode *inode = req->wb_context->dentry->d_inode;
 361         struct nfs_inode *nfsi = NFS_I(inode);
 362
 363         BUG_ON (!NFS_WBACK_BUSY(req));
 364
 365         spin_lock(&nfsi->req_lock);
 366         set_page_private(req->wb_page, 0);
 367         ClearPagePrivate(req->wb_page);
 368         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 369         nfsi->npages--;
 370         if (!nfsi->npages) {
 371                 spin_unlock(&nfsi->req_lock);
 372                 nfs_end_data_update(inode);
 373                 iput(inode);
 374         } else
 375                 spin_unlock(&nfsi->req_lock);
 376         nfs_clear_request(req);
 377         nfs_release_request(req);
 378 }
 379
 380 /*
 381  * Add a request to the inode's dirty list.
 382  */
 383 static void
 384 nfs_mark_request_dirty(struct nfs_page *req)
 385 {
 386         struct inode *inode = req->wb_context->dentry->d_inode;
 387         struct nfs_inode *nfsi = NFS_I(inode);
 388
 389         spin_lock(&nfsi->req_lock);
 390         radix_tree_tag_set(&nfsi->nfs_page_tree,
 391                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 392         nfs_list_add_request(req, &nfsi->dirty);
 393         nfsi->ndirty++;
 394         spin_unlock(&nfsi->req_lock);
 395         __mark_inode_dirty(inode, I_DIRTY_PAGES);
 396 }
 397
 398 static void
 399 nfs_redirty_request(struct nfs_page *req)
 400 {
 401         clear_bit(PG_FLUSHING, &req->wb_flags);
 402         __set_page_dirty_nobuffers(req->wb_page);
 403 }
 404
 405 /*
 406  * Check if a request is dirty
 407  */
 408 static inline int
 409 nfs_dirty_request(struct nfs_page *req)
 410 {
 411         return test_bit(PG_FLUSHING, &req->wb_flags) == 0;
 412 }
 413
 414 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 415 /*
 416  * Add a request to the inode's commit list.
 417  */
 418 static void
 419 nfs_mark_request_commit(struct nfs_page *req)
 420 {
 421         struct inode *inode = req->wb_context->dentry->d_inode;
 422         struct nfs_inode *nfsi = NFS_I(inode);
 423
 424         spin_lock(&nfsi->req_lock);
 425         nfs_list_add_request(req, &nfsi->commit);
 426         nfsi->ncommit++;
 427         spin_unlock(&nfsi->req_lock);
 428         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 429         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 430 }
 431 #endif
 432
 433 /*
 434  * Wait for a request to complete.
 435  *
 436  * Interruptible by signals only if mounted with intr flag.
 437  */
 438 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 439 {
 440         struct nfs_inode *nfsi = NFS_I(inode);
 441         struct nfs_page *req;
 442         unsigned long           idx_end, next;
 443         unsigned int            res = 0;
 444         int                     error;
 445
 446         if (npages == 0)
 447                 idx_end = ~0;
 448         else
 449                 idx_end = idx_start + npages - 1;
 450
 451         next = idx_start;
 452         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 453                 if (req->wb_index > idx_end)
 454                         break;
 455
 456                 next = req->wb_index + 1;
 457                 BUG_ON(!NFS_WBACK_BUSY(req));
 458
 459                 atomic_inc(&req->wb_count);
 460                 spin_unlock(&nfsi->req_lock);
 461                 error = nfs_wait_on_request(req);
 462                 nfs_release_request(req);
 463                 spin_lock(&nfsi->req_lock);
 464                 if (error < 0)
 465                         return error;
 466                 res++;
 467         }
 468         return res;
 469 }
 470
 471 static void nfs_cancel_dirty_list(struct list_head *head)
 472 {
 473         struct nfs_page *req;
 474         while(!list_empty(head)) {
 475                 req = nfs_list_entry(head->next);
 476                 nfs_list_remove_request(req);
 477                 nfs_inode_remove_request(req);
 478                 nfs_clear_page_writeback(req);
 479         }
 480 }
 481
 482 static void nfs_cancel_commit_list(struct list_head *head)
 483 {
 484         struct nfs_page *req;
 485
 486         while(!list_empty(head)) {
 487                 req = nfs_list_entry(head->next);
 488                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 489                 nfs_list_remove_request(req);
 490                 nfs_inode_remove_request(req);
 491                 nfs_unlock_request(req);
 492         }
 493 }
 494
 495 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 496 /*
 497  * nfs_scan_commit - Scan an inode for commit requests
 498  * @inode: NFS inode to scan
 499  * @dst: destination list
 500  * @idx_start: lower bound of page->index to scan.
 501  * @npages: idx_start + npages sets the upper bound to scan.
 502  *
 503  * Moves requests from the inode's 'commit' request list.
 504  * The requests are *not* checked to ensure that they form a contiguous set.
 505  */
 506 static int
 507 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 508 {
 509         struct nfs_inode *nfsi = NFS_I(inode);
 510         int res = 0;
 511
 512         if (nfsi->ncommit != 0) {
 513                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 514                 nfsi->ncommit -= res;
 515                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 516                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 517         }
 518         return res;
 519 }
 520 #else
 521 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 522 {
 523         return 0;
 524 }
 525 #endif
 526
 527 static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr)
 528 {
 529         struct backing_dev_info *bdi = mapping->backing_dev_info;
 530         DEFINE_WAIT(wait);
 531         int ret = 0;
 532
 533         might_sleep();
 534
 535         if (!bdi_write_congested(bdi))
 536                 return 0;
 537
 538         nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT);
 539
 540         if (intr) {
 541                 struct rpc_clnt *clnt = NFS_CLIENT(mapping->host);
 542                 sigset_t oldset;
 543
 544                 rpc_clnt_sigmask(clnt, &oldset);
 545                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE);
 546                 if (bdi_write_congested(bdi)) {
 547                         if (signalled())
 548                                 ret = -ERESTARTSYS;
 549                         else
 550                                 schedule();
 551                 }
 552                 rpc_clnt_sigunmask(clnt, &oldset);
 553         } else {
 554                 prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE);
 555                 if (bdi_write_congested(bdi))
 556                         schedule();
 557         }
 558         finish_wait(&nfs_write_congestion, &wait);
 559         return ret;
 560 }
 561
 562
 563 /*
 564  * Try to update any existing write request, or create one if there is none.
 565  * In order to match, the request's credentials must match those of
 566  * the calling process.
 567  *
 568  * Note: Should always be called with the Page Lock held!
 569  */
 570 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 571                 struct page *page, unsigned int offset, unsigned int bytes)
 572 {
 573         struct inode *inode = page->mapping->host;
 574         struct nfs_inode *nfsi = NFS_I(inode);
 575         struct nfs_page         *req, *new = NULL;
 576         unsigned long           rqend, end;
 577
 578         end = offset + bytes;
 579
 580         if (nfs_wait_on_write_congestion(page->mapping, NFS_SERVER(inode)->flags & NFS_MOUNT_INTR))
 581                 return ERR_PTR(-ERESTARTSYS);
 582         for (;;) {
 583                 /* Loop over all inode entries and see if we find
 584                  * A request for the page we wish to update
 585                  */
 586                 spin_lock(&nfsi->req_lock);
 587                 req = nfs_page_find_request_locked(page);
 588                 if (req) {
 589                         if (!nfs_lock_request_dontget(req)) {
 590                                 int error;
 591
 592                                 spin_unlock(&nfsi->req_lock);
 593                                 error = nfs_wait_on_request(req);
 594                                 nfs_release_request(req);
 595                                 if (error < 0) {
 596                                         if (new)
 597                                                 nfs_release_request(new);
 598                                         return ERR_PTR(error);
 599                                 }
 600                                 continue;
 601                         }
 602                         spin_unlock(&nfsi->req_lock);
 603                         if (new)
 604                                 nfs_release_request(new);
 605                         break;
 606                 }
 607
 608                 if (new) {
 609                         int error;
 610                         nfs_lock_request_dontget(new);
 611                         error = nfs_inode_add_request(inode, new);
 612                         if (error) {
 613                                 spin_unlock(&nfsi->req_lock);
 614                                 nfs_unlock_request(new);
 615                                 return ERR_PTR(error);
 616                         }
 617                         spin_unlock(&nfsi->req_lock);
 618                         return new;
 619                 }
 620                 spin_unlock(&nfsi->req_lock);
 621
 622                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 623                 if (IS_ERR(new))
 624                         return new;
 625         }
 626
 627         /* We have a request for our page.
 628          * If the creds don't match, or the
 629          * page addresses don't match,
 630          * tell the caller to wait on the conflicting
 631          * request.
 632          */
 633         rqend = req->wb_offset + req->wb_bytes;
 634         if (req->wb_context != ctx
 635             || req->wb_page != page
 636             || !nfs_dirty_request(req)
 637             || offset > rqend || end < req->wb_offset) {
 638                 nfs_unlock_request(req);
 639                 return ERR_PTR(-EBUSY);
 640         }
 641
 642         /* Okay, the request matches. Update the region */
 643         if (offset < req->wb_offset) {
 644                 req->wb_offset = offset;
 645                 req->wb_pgbase = offset;
 646                 req->wb_bytes = rqend - req->wb_offset;
 647         }
 648
 649         if (end > rqend)
 650                 req->wb_bytes = end - req->wb_offset;
 651
 652         return req;
 653 }
 654
 655 int nfs_flush_incompatible(struct file *file, struct page *page)
 656 {
 657         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 658         struct nfs_page *req;
 659         int do_flush, status;
 660         /*
 661          * Look for a request corresponding to this page. If there
 662          * is one, and it belongs to another file, we flush it out
 663          * before we try to copy anything into the page. Do this
 664          * due to the lack of an ACCESS-type call in NFSv2.
 665          * Also do the same if we find a request from an existing
 666          * dropped page.
 667          */
 668         do {
 669                 req = nfs_page_find_request(page);
 670                 if (req == NULL)
 671                         return 0;
 672                 do_flush = req->wb_page != page || req->wb_context != ctx
 673                         || !nfs_dirty_request(req);
 674                 nfs_release_request(req);
 675                 if (!do_flush)
 676                         return 0;
 677                 status = nfs_wb_page(page->mapping->host, page);
 678         } while (status == 0);
 679         return status;
 680 }
 681
 682 /*
 683  * Update and possibly write a cached page of an NFS file.
 684  *
 685  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 686  * things with a page scheduled for an RPC call (e.g. invalidate it).
 687  */
 688 int nfs_updatepage(struct file *file, struct page *page,
 689                 unsigned int offset, unsigned int count)
 690 {
 691         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 692         struct inode    *inode = page->mapping->host;
 693         int             status = 0;
 694
 695         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 696
 697         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 698                 file->f_path.dentry->d_parent->d_name.name,
 699                 file->f_path.dentry->d_name.name, count,
 700                 (long long)(page_offset(page) +offset));
 701
 702         /* If we're not using byte range locks, and we know the page
 703          * is entirely in cache, it may be more efficient to avoid
 704          * fragmenting write requests.
 705          */
 706         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 707                 count = max(count + offset, nfs_page_length(page));
 708                 offset = 0;
 709         }
 710
 711         status = nfs_writepage_setup(ctx, page, offset, count);
 712         __set_page_dirty_nobuffers(page);
 713
 714         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 715                         status, (long long)i_size_read(inode));
 716         if (status < 0)
 717                 ClearPageUptodate(page);
 718         return status;
 719 }
 720
 721 static void nfs_writepage_release(struct nfs_page *req)
 722 {
 723         end_page_writeback(req->wb_page);
 724
 725 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 726         if (!PageError(req->wb_page)) {
 727                 if (NFS_NEED_RESCHED(req)) {
 728                         nfs_redirty_request(req);
 729                         goto out;
 730                 } else if (NFS_NEED_COMMIT(req)) {
 731                         nfs_mark_request_commit(req);
 732                         goto out;
 733                 }
 734         }
 735         nfs_inode_remove_request(req);
 736
 737 out:
 738         nfs_clear_commit(req);
 739         nfs_clear_reschedule(req);
 740 #else
 741         nfs_inode_remove_request(req);
 742 #endif
 743         nfs_clear_page_writeback(req);
 744 }
 745
 746 static inline int flush_task_priority(int how)
 747 {
 748         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 749                 case FLUSH_HIGHPRI:
 750                         return RPC_PRIORITY_HIGH;
 751                 case FLUSH_LOWPRI:
 752                         return RPC_PRIORITY_LOW;
 753         }
 754         return RPC_PRIORITY_NORMAL;
 755 }
 756
 757 /*
 758  * Set up the argument/result storage required for the RPC call.
 759  */
 760 static void nfs_write_rpcsetup(struct nfs_page *req,
 761                 struct nfs_write_data *data,
 762                 const struct rpc_call_ops *call_ops,
 763                 unsigned int count, unsigned int offset,
 764                 int how)
 765 {
 766         struct inode            *inode;
 767         int flags;
 768
 769         /* Set up the RPC argument and reply structs
 770          * NB: take care not to mess about with data->commit et al. */
 771
 772         data->req = req;
 773         data->inode = inode = req->wb_context->dentry->d_inode;
 774         data->cred = req->wb_context->cred;
 775
 776         data->args.fh     = NFS_FH(inode);
 777         data->args.offset = req_offset(req) + offset;
 778         data->args.pgbase = req->wb_pgbase + offset;
 779         data->args.pages  = data->pagevec;
 780         data->args.count  = count;
 781         data->args.context = req->wb_context;
 782
 783         data->res.fattr   = &data->fattr;
 784         data->res.count   = count;
 785         data->res.verf    = &data->verf;
 786         nfs_fattr_init(&data->fattr);
 787
 788         /* Set up the initial task struct.  */
 789         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 790         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 791         NFS_PROTO(inode)->write_setup(data, how);
 792
 793         data->task.tk_priority = flush_task_priority(how);
 794         data->task.tk_cookie = (unsigned long)inode;
 795
 796         dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 797                 data->task.tk_pid,
 798                 inode->i_sb->s_id,
 799                 (long long)NFS_FILEID(inode),
 800                 count,
 801                 (unsigned long long)data->args.offset);
 802 }
 803
 804 static void nfs_execute_write(struct nfs_write_data *data)
 805 {
 806         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 807         sigset_t oldset;
 808
 809         rpc_clnt_sigmask(clnt, &oldset);
 810         rpc_execute(&data->task);
 811         rpc_clnt_sigunmask(clnt, &oldset);
 812 }
 813
 814 /*
 815  * Generate multiple small requests to write out a single
 816  * contiguous dirty area on one page.
 817  */
 818 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 819 {
 820         struct nfs_page *req = nfs_list_entry(head->next);
 821         struct page *page = req->wb_page;
 822         struct nfs_write_data *data;
 823         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 824         unsigned int offset;
 825         int requests = 0;
 826         LIST_HEAD(list);
 827
 828         nfs_list_remove_request(req);
 829
 830         nbytes = req->wb_bytes;
 831         do {
 832                 size_t len = min(nbytes, wsize);
 833
 834                 data = nfs_writedata_alloc(len);
 835                 if (!data)
 836                         goto out_bad;
 837                 list_add(&data->pages, &list);
 838                 requests++;
 839                 nbytes -= len;
 840         } while (nbytes != 0);
 841         atomic_set(&req->wb_complete, requests);
 842
 843         ClearPageError(page);
 844         offset = 0;
 845         nbytes = req->wb_bytes;
 846         do {
 847                 data = list_entry(list.next, struct nfs_write_data, pages);
 848                 list_del_init(&data->pages);
 849
 850                 data->pagevec[0] = page;
 851
 852                 if (nbytes > wsize) {
 853                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 854                                         wsize, offset, how);
 855                         offset += wsize;
 856                         nbytes -= wsize;
 857                 } else {
 858                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 859                                         nbytes, offset, how);
 860                         nbytes = 0;
 861                 }
 862                 nfs_execute_write(data);
 863         } while (nbytes != 0);
 864
 865         return 0;
 866
 867 out_bad:
 868         while (!list_empty(&list)) {
 869                 data = list_entry(list.next, struct nfs_write_data, pages);
 870                 list_del(&data->pages);
 871                 nfs_writedata_release(data);
 872         }
 873         nfs_redirty_request(req);
 874         nfs_clear_page_writeback(req);
 875         return -ENOMEM;
 876 }
 877
 878 /*
 879  * Create an RPC task for the given write request and kick it.
 880  * The page must have been locked by the caller.
 881  *
 882  * It may happen that the page we're passed is not marked dirty.
 883  * This is the case if nfs_updatepage detects a conflicting request
 884  * that has been written but not committed.
 885  */
 886 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 887 {
 888         struct nfs_page         *req;
 889         struct page             **pages;
 890         struct nfs_write_data   *data;
 891         unsigned int            count;
 892
 893         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 894         if (!data)
 895                 goto out_bad;
 896
 897         pages = data->pagevec;
 898         count = 0;
 899         while (!list_empty(head)) {
 900                 req = nfs_list_entry(head->next);
 901                 nfs_list_remove_request(req);
 902                 nfs_list_add_request(req, &data->pages);
 903                 ClearPageError(req->wb_page);
 904                 *pages++ = req->wb_page;
 905                 count += req->wb_bytes;
 906         }
 907         req = nfs_list_entry(data->pages.next);
 908
 909         /* Set up the argument struct */
 910         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 911
 912         nfs_execute_write(data);
 913         return 0;
 914  out_bad:
 915         while (!list_empty(head)) {
 916                 struct nfs_page *req = nfs_list_entry(head->next);
 917                 nfs_list_remove_request(req);
 918                 nfs_redirty_request(req);
 919                 nfs_clear_page_writeback(req);
 920         }
 921         return -ENOMEM;
 922 }
 923
 924 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 925 {
 926         LIST_HEAD(one_request);
 927         int (*flush_one)(struct inode *, struct list_head *, int);
 928         struct nfs_page *req;
 929         int wpages = NFS_SERVER(inode)->wpages;
 930         int wsize = NFS_SERVER(inode)->wsize;
 931         int error;
 932
 933         flush_one = nfs_flush_one;
 934         if (wsize < PAGE_CACHE_SIZE)
 935                 flush_one = nfs_flush_multi;
 936         /* For single writes, FLUSH_STABLE is more efficient */
 937         if (npages <= wpages && npages == NFS_I(inode)->npages
 938                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 939                 how |= FLUSH_STABLE;
 940
 941         do {
 942                 nfs_coalesce_requests(head, &one_request, wpages);
 943                 req = nfs_list_entry(one_request.next);
 944                 error = flush_one(inode, &one_request, how);
 945                 if (error < 0)
 946                         goto out_err;
 947         } while (!list_empty(head));
 948         return 0;
 949 out_err:
 950         while (!list_empty(head)) {
 951                 req = nfs_list_entry(head->next);
 952                 nfs_list_remove_request(req);
 953                 nfs_redirty_request(req);
 954                 nfs_clear_page_writeback(req);
 955         }
 956         return error;
 957 }
 958
 959 /*
 960  * Handle a write reply that flushed part of a page.
 961  */
 962 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 963 {
 964         struct nfs_write_data   *data = calldata;
 965         struct nfs_page         *req = data->req;
 966         struct page             *page = req->wb_page;
 967
 968         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 969                 req->wb_context->dentry->d_inode->i_sb->s_id,
 970                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 971                 req->wb_bytes,
 972                 (long long)req_offset(req));
 973
 974         if (nfs_writeback_done(task, data) != 0)
 975                 return;
 976
 977         if (task->tk_status < 0) {
 978                 ClearPageUptodate(page);
 979                 SetPageError(page);
 980                 req->wb_context->error = task->tk_status;
 981                 dprintk(", error = %d\n", task->tk_status);
 982         } else {
 983 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 984                 if (data->verf.committed < NFS_FILE_SYNC) {
 985                         if (!NFS_NEED_COMMIT(req)) {
 986                                 nfs_defer_commit(req);
 987                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 988                                 dprintk(" defer commit\n");
 989                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 990                                 nfs_defer_reschedule(req);
 991                                 dprintk(" server reboot detected\n");
 992                         }
 993                 } else
 994 #endif
 995                         dprintk(" OK\n");
 996         }
 997
 998         if (atomic_dec_and_test(&req->wb_complete))
 999                 nfs_writepage_release(req);
1000 }
1001
1002 static const struct rpc_call_ops nfs_write_partial_ops = {
1003         .rpc_call_done = nfs_writeback_done_partial,
1004         .rpc_release = nfs_writedata_release,
1005 };
1006
1007 /*
1008  * Handle a write reply that flushes a whole page.
1009  *
1010  * FIXME: There is an inherent race with invalidate_inode_pages and
1011  *        writebacks since the page->count is kept > 1 for as long
1012  *        as the page has a write request pending.
1013  */
1014 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1015 {
1016         struct nfs_write_data   *data = calldata;
1017         struct nfs_page         *req;
1018         struct page             *page;
1019
1020         if (nfs_writeback_done(task, data) != 0)
1021                 return;
1022
1023         /* Update attributes as result of writeback. */
1024         while (!list_empty(&data->pages)) {
1025                 req = nfs_list_entry(data->pages.next);
1026                 nfs_list_remove_request(req);
1027                 page = req->wb_page;
1028
1029                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1030                         req->wb_context->dentry->d_inode->i_sb->s_id,
1031                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1032                         req->wb_bytes,
1033                         (long long)req_offset(req));
1034
1035                 if (task->tk_status < 0) {
1036                         ClearPageUptodate(page);
1037                         SetPageError(page);
1038                         req->wb_context->error = task->tk_status;
1039                         end_page_writeback(page);
1040                         nfs_inode_remove_request(req);
1041                         dprintk(", error = %d\n", task->tk_status);
1042                         goto next;
1043                 }
1044                 end_page_writeback(page);
1045
1046 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1047                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1048                         nfs_inode_remove_request(req);
1049                         dprintk(" OK\n");
1050                         goto next;
1051                 }
1052                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1053                 nfs_mark_request_commit(req);
1054                 dprintk(" marked for commit\n");
1055 #else
1056                 nfs_inode_remove_request(req);
1057 #endif
1058         next:
1059                 nfs_clear_page_writeback(req);
1060         }
1061 }
1062
1063 static const struct rpc_call_ops nfs_write_full_ops = {
1064         .rpc_call_done = nfs_writeback_done_full,
1065         .rpc_release = nfs_writedata_release,
1066 };
1067
1068
1069 /*
1070  * This function is called when the WRITE call is complete.
1071  */
1072 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1073 {
1074         struct nfs_writeargs    *argp = &data->args;
1075         struct nfs_writeres     *resp = &data->res;
1076         int status;
1077
1078         dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
1079                 task->tk_pid, task->tk_status);
1080
1081         /*
1082          * ->write_done will attempt to use post-op attributes to detect
1083          * conflicting writes by other clients.  A strict interpretation
1084          * of close-to-open would allow us to continue caching even if
1085          * another writer had changed the file, but some applications
1086          * depend on tighter cache coherency when writing.
1087          */
1088         status = NFS_PROTO(data->inode)->write_done(task, data);
1089         if (status != 0)
1090                 return status;
1091         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1092
1093 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1094         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1095                 /* We tried a write call, but the server did not
1096                  * commit data to stable storage even though we
1097                  * requested it.
1098                  * Note: There is a known bug in Tru64 < 5.0 in which
1099                  *       the server reports NFS_DATA_SYNC, but performs
1100                  *       NFS_FILE_SYNC. We therefore implement this checking
1101                  *       as a dprintk() in order to avoid filling syslog.
1102                  */
1103                 static unsigned long    complain;
1104
1105                 if (time_before(complain, jiffies)) {
1106                         dprintk("NFS: faulty NFS server %s:"
1107                                 " (committed = %d) != (stable = %d)\n",
1108                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1109                                 resp->verf->committed, argp->stable);
1110                         complain = jiffies + 300 * HZ;
1111                 }
1112         }
1113 #endif
1114         /* Is this a short write? */
1115         if (task->tk_status >= 0 && resp->count < argp->count) {
1116                 static unsigned long    complain;
1117
1118                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1119
1120                 /* Has the server at least made some progress? */
1121                 if (resp->count != 0) {
1122                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1123                         if (resp->verf->committed != NFS_UNSTABLE) {
1124                                 /* Resend from where the server left off */
1125                                 argp->offset += resp->count;
1126                                 argp->pgbase += resp->count;
1127                                 argp->count -= resp->count;
1128                         } else {
1129                                 /* Resend as a stable write in order to avoid
1130                                  * headaches in the case of a server crash.
1131                                  */
1132                                 argp->stable = NFS_FILE_SYNC;
1133                         }
1134                         rpc_restart_call(task);
1135                         return -EAGAIN;
1136                 }
1137                 if (time_before(complain, jiffies)) {
1138                         printk(KERN_WARNING
1139                                "NFS: Server wrote zero bytes, expected %u.\n",
1140                                         argp->count);
1141                         complain = jiffies + 300 * HZ;
1142                 }
1143                 /* Can't do anything about it except throw an error. */
1144                 task->tk_status = -EIO;
1145         }
1146         return 0;
1147 }
1148
1149
1150 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1151 void nfs_commit_release(void *wdata)
1152 {
1153         nfs_commit_free(wdata);
1154 }
1155
1156 /*
1157  * Set up the argument/result storage required for the RPC call.
1158  */
1159 static void nfs_commit_rpcsetup(struct list_head *head,
1160                 struct nfs_write_data *data,
1161                 int how)
1162 {
1163         struct nfs_page         *first;
1164         struct inode            *inode;
1165         int flags;
1166
1167         /* Set up the RPC argument and reply structs
1168          * NB: take care not to mess about with data->commit et al. */
1169
1170         list_splice_init(head, &data->pages);
1171         first = nfs_list_entry(data->pages.next);
1172         inode = first->wb_context->dentry->d_inode;
1173
1174         data->inode       = inode;
1175         data->cred        = first->wb_context->cred;
1176
1177         data->args.fh     = NFS_FH(data->inode);
1178         /* Note: we always request a commit of the entire inode */
1179         data->args.offset = 0;
1180         data->args.count  = 0;
1181         data->res.count   = 0;
1182         data->res.fattr   = &data->fattr;
1183         data->res.verf    = &data->verf;
1184         nfs_fattr_init(&data->fattr);
1185
1186         /* Set up the initial task struct.  */
1187         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1188         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1189         NFS_PROTO(inode)->commit_setup(data, how);
1190
1191         data->task.tk_priority = flush_task_priority(how);
1192         data->task.tk_cookie = (unsigned long)inode;
1193
1194         dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid);
1195 }
1196
1197 /*
1198  * Commit dirty pages
1199  */
1200 static int
1201 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1202 {
1203         struct nfs_write_data   *data;
1204         struct nfs_page         *req;
1205
1206         data = nfs_commit_alloc();
1207
1208         if (!data)
1209                 goto out_bad;
1210
1211         /* Set up the argument struct */
1212         nfs_commit_rpcsetup(head, data, how);
1213
1214         nfs_execute_write(data);
1215         return 0;
1216  out_bad:
1217         while (!list_empty(head)) {
1218                 req = nfs_list_entry(head->next);
1219                 nfs_list_remove_request(req);
1220                 nfs_mark_request_commit(req);
1221                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1222                 nfs_clear_page_writeback(req);
1223         }
1224         return -ENOMEM;
1225 }
1226
1227 /*
1228  * COMMIT call returned
1229  */
1230 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1231 {
1232         struct nfs_write_data   *data = calldata;
1233         struct nfs_page         *req;
1234
1235         dprintk("NFS: %4d nfs_commit_done (status %d)\n",
1236                                 task->tk_pid, task->tk_status);
1237
1238         /* Call the NFS version-specific code */
1239         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1240                 return;
1241
1242         while (!list_empty(&data->pages)) {
1243                 req = nfs_list_entry(data->pages.next);
1244                 nfs_list_remove_request(req);
1245                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1246
1247                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1248                         req->wb_context->dentry->d_inode->i_sb->s_id,
1249                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1250                         req->wb_bytes,
1251                         (long long)req_offset(req));
1252                 if (task->tk_status < 0) {
1253                         req->wb_context->error = task->tk_status;
1254                         nfs_inode_remove_request(req);
1255                         dprintk(", error = %d\n", task->tk_status);
1256                         goto next;
1257                 }
1258
1259                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1260                  * returned by the server against all stored verfs. */
1261                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1262                         /* We have a match */
1263                         nfs_inode_remove_request(req);
1264                         dprintk(" OK\n");
1265                         goto next;
1266                 }
1267                 /* We have a mismatch. Write the page again */
1268                 dprintk(" mismatch\n");
1269                 nfs_redirty_request(req);
1270         next:
1271                 nfs_clear_page_writeback(req);
1272         }
1273 }
1274
1275 static const struct rpc_call_ops nfs_commit_ops = {
1276         .rpc_call_done = nfs_commit_done,
1277         .rpc_release = nfs_commit_release,
1278 };
1279 #else
1280 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1281 {
1282         return 0;
1283 }
1284 #endif
1285
1286 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1287 {
1288         struct nfs_inode *nfsi = NFS_I(mapping->host);
1289         LIST_HEAD(head);
1290         long res;
1291
1292         spin_lock(&nfsi->req_lock);
1293         res = nfs_scan_dirty(mapping, wbc, &head);
1294         spin_unlock(&nfsi->req_lock);
1295         if (res) {
1296                 int error = nfs_flush_list(mapping->host, &head, res, how);
1297                 if (error < 0)
1298                         return error;
1299         }
1300         return res;
1301 }
1302
1303 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1304 int nfs_commit_inode(struct inode *inode, int how)
1305 {
1306         struct nfs_inode *nfsi = NFS_I(inode);
1307         LIST_HEAD(head);
1308         int res;
1309
1310         spin_lock(&nfsi->req_lock);
1311         res = nfs_scan_commit(inode, &head, 0, 0);
1312         spin_unlock(&nfsi->req_lock);
1313         if (res) {
1314                 int error = nfs_commit_list(inode, &head, how);
1315                 if (error < 0)
1316                         return error;
1317         }
1318         return res;
1319 }
1320 #endif
1321
1322 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1323 {
1324         struct inode *inode = mapping->host;
1325         struct nfs_inode *nfsi = NFS_I(inode);
1326         unsigned long idx_start, idx_end;
1327         unsigned int npages = 0;
1328         LIST_HEAD(head);
1329         int nocommit = how & FLUSH_NOCOMMIT;
1330         long pages, ret;
1331
1332         /* FIXME */
1333         if (wbc->range_cyclic)
1334                 idx_start = 0;
1335         else {
1336                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1337                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1338                 if (idx_end > idx_start) {
1339                         unsigned long l_npages = 1 + idx_end - idx_start;
1340                         npages = l_npages;
1341                         if (sizeof(npages) != sizeof(l_npages) &&
1342                                         (unsigned long)npages != l_npages)
1343                                 npages = 0;
1344                 }
1345         }
1346         how &= ~FLUSH_NOCOMMIT;
1347         spin_lock(&nfsi->req_lock);
1348         do {
1349                 wbc->pages_skipped = 0;
1350                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1351                 if (ret != 0)
1352                         continue;
1353                 pages = nfs_scan_dirty(mapping, wbc, &head);
1354                 if (pages != 0) {
1355                         spin_unlock(&nfsi->req_lock);
1356                         if (how & FLUSH_INVALIDATE) {
1357                                 nfs_cancel_dirty_list(&head);
1358                                 ret = pages;
1359                         } else
1360                                 ret = nfs_flush_list(inode, &head, pages, how);
1361                         spin_lock(&nfsi->req_lock);
1362                         continue;
1363                 }
1364                 if (wbc->pages_skipped != 0)
1365                         continue;
1366                 if (nocommit)
1367                         break;
1368                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1369                 if (pages == 0) {
1370                         if (wbc->pages_skipped != 0)
1371                                 continue;
1372                         break;
1373                 }
1374                 if (how & FLUSH_INVALIDATE) {
1375                         spin_unlock(&nfsi->req_lock);
1376                         nfs_cancel_commit_list(&head);
1377                         ret = pages;
1378                         spin_lock(&nfsi->req_lock);
1379                         continue;
1380                 }
1381                 pages += nfs_scan_commit(inode, &head, 0, 0);
1382                 spin_unlock(&nfsi->req_lock);
1383                 ret = nfs_commit_list(inode, &head, how);
1384                 spin_lock(&nfsi->req_lock);
1385         } while (ret >= 0);
1386         spin_unlock(&nfsi->req_lock);
1387         return ret;
1388 }
1389
1390 /*
1391  * flush the inode to disk.
1392  */
1393 int nfs_wb_all(struct inode *inode)
1394 {
1395         struct address_space *mapping = inode->i_mapping;
1396         struct writeback_control wbc = {
1397                 .bdi = mapping->backing_dev_info,
1398                 .sync_mode = WB_SYNC_ALL,
1399                 .nr_to_write = LONG_MAX,
1400                 .for_writepages = 1,
1401                 .range_cyclic = 1,
1402         };
1403         int ret;
1404
1405         ret = generic_writepages(mapping, &wbc);
1406         if (ret < 0)
1407                 goto out;
1408         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1409         if (ret >= 0)
1410                 return 0;
1411 out:
1412         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1413         return ret;
1414 }
1415
1416 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1417 {
1418         struct writeback_control wbc = {
1419                 .bdi = mapping->backing_dev_info,
1420                 .sync_mode = WB_SYNC_ALL,
1421                 .nr_to_write = LONG_MAX,
1422                 .range_start = range_start,
1423                 .range_end = range_end,
1424                 .for_writepages = 1,
1425         };
1426         int ret;
1427
1428         if (!(how & FLUSH_NOWRITEPAGE)) {
1429                 ret = generic_writepages(mapping, &wbc);
1430                 if (ret < 0)
1431                         goto out;
1432         }
1433         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1434         if (ret >= 0)
1435                 return 0;
1436 out:
1437         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1438         return ret;
1439 }
1440
1441 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1442 {
1443         loff_t range_start = page_offset(page);
1444         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1445         struct writeback_control wbc = {
1446                 .bdi = page->mapping->backing_dev_info,
1447                 .sync_mode = WB_SYNC_ALL,
1448                 .nr_to_write = LONG_MAX,
1449                 .range_start = range_start,
1450                 .range_end = range_end,
1451         };
1452         int ret;
1453
1454         BUG_ON(!PageLocked(page));
1455         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1456                 ret = nfs_writepage_locked(page, &wbc);
1457                 if (ret < 0)
1458                         goto out;
1459         }
1460         if (!PagePrivate(page))
1461                 return 0;
1462         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1463         if (ret >= 0)
1464                 return 0;
1465 out:
1466         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1467         return ret;
1468 }
1469
1470 /*
1471  * Write back all requests on one page - we do this before reading it.
1472  */
1473 int nfs_wb_page(struct inode *inode, struct page* page)
1474 {
1475         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1476 }
1477
1478 int nfs_set_page_dirty(struct page *page)
1479 {
1480         struct nfs_page *req;
1481
1482         req = nfs_page_find_request(page);
1483         if (req != NULL) {
1484                 /* Mark any existing write requests for flushing */
1485                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1486                 nfs_release_request(req);
1487         }
1488         return __set_page_dirty_nobuffers(page);
1489 }
1490
1491
1492 int __init nfs_init_writepagecache(void)
1493 {
1494         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1495                                              sizeof(struct nfs_write_data),
1496                                              0, SLAB_HWCACHE_ALIGN,
1497                                              NULL, NULL);
1498         if (nfs_wdata_cachep == NULL)
1499                 return -ENOMEM;
1500
1501         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1502                                                      nfs_wdata_cachep);
1503         if (nfs_wdata_mempool == NULL)
1504                 return -ENOMEM;
1505
1506         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1507                                                       nfs_wdata_cachep);
1508         if (nfs_commit_mempool == NULL)
1509                 return -ENOMEM;
1510
1511         return 0;
1512 }
1513
1514 void nfs_destroy_writepagecache(void)
1515 {
1516         mempool_destroy(nfs_commit_mempool);
1517         mempool_destroy(nfs_wdata_mempool);
1518         kmem_cache_destroy(nfs_wdata_cachep);
1519 }
1520