git.oblomov.eu 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 #include <linux/swap.h>
  16
  17 #include <linux/sunrpc/clnt.h>
  18 #include <linux/nfs_fs.h>
  19 #include <linux/nfs_mount.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/backing-dev.h>
  22
  23 #include <asm/uaccess.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_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static const struct rpc_call_ops nfs_write_partial_ops;
  43 static const struct rpc_call_ops nfs_write_full_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45
  46 static struct kmem_cache *nfs_wdata_cachep;
  47 static mempool_t *nfs_wdata_mempool;
  48 static mempool_t *nfs_commit_mempool;
  49
  50 struct nfs_write_data *nfs_commit_alloc(void)
  51 {
  52         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  53
  54         if (p) {
  55                 memset(p, 0, sizeof(*p));
  56                 INIT_LIST_HEAD(&p->pages);
  57         }
  58         return p;
  59 }
  60
  61 static void nfs_commit_rcu_free(struct rcu_head *head)
  62 {
  63         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 void nfs_commit_free(struct nfs_write_data *wdata)
  70 {
  71         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  72 }
  73
  74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  75 {
  76         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  77
  78         if (p) {
  79                 memset(p, 0, sizeof(*p));
  80                 INIT_LIST_HEAD(&p->pages);
  81                 p->npages = pagecount;
  82                 if (pagecount <= ARRAY_SIZE(p->page_array))
  83                         p->pagevec = p->page_array;
  84                 else {
  85                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  86                         if (!p->pagevec) {
  87                                 mempool_free(p, nfs_wdata_mempool);
  88                                 p = NULL;
  89                         }
  90                 }
  91         }
  92         return p;
  93 }
  94
  95 static void nfs_writedata_rcu_free(struct rcu_head *head)
  96 {
  97         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  98         if (p && (p->pagevec != &p->page_array[0]))
  99                 kfree(p->pagevec);
 100         mempool_free(p, nfs_wdata_mempool);
 101 }
 102
 103 static void nfs_writedata_free(struct nfs_write_data *wdata)
 104 {
 105         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 106 }
 107
 108 void nfs_writedata_release(void *wdata)
 109 {
 110         nfs_writedata_free(wdata);
 111 }
 112
 113 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 114 {
 115         struct nfs_page *req = NULL;
 116
 117         if (PagePrivate(page)) {
 118                 req = (struct nfs_page *)page_private(page);
 119                 if (req != NULL)
 120                         atomic_inc(&req->wb_count);
 121         }
 122         return req;
 123 }
 124
 125 static struct nfs_page *nfs_page_find_request(struct page *page)
 126 {
 127         struct nfs_page *req = NULL;
 128         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 129
 130         spin_lock(req_lock);
 131         req = nfs_page_find_request_locked(page);
 132         spin_unlock(req_lock);
 133         return req;
 134 }
 135
 136 /* Adjust the file length if we're writing beyond the end */
 137 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 138 {
 139         struct inode *inode = page->mapping->host;
 140         loff_t end, i_size = i_size_read(inode);
 141         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 142
 143         if (i_size > 0 && page->index < end_index)
 144                 return;
 145         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 146         if (i_size >= end)
 147                 return;
 148         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 149         i_size_write(inode, end);
 150 }
 151
 152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 153 static void nfs_set_pageerror(struct page *page)
 154 {
 155         SetPageError(page);
 156         nfs_zap_mapping(page->mapping->host, page->mapping);
 157 }
 158
 159 /* We can set the PG_uptodate flag if we see that a write request
 160  * covers the full page.
 161  */
 162 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 163 {
 164         if (PageUptodate(page))
 165                 return;
 166         if (base != 0)
 167                 return;
 168         if (count != nfs_page_length(page))
 169                 return;
 170         if (count != PAGE_CACHE_SIZE)
 171                 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
 172         SetPageUptodate(page);
 173 }
 174
 175 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 176                 unsigned int offset, unsigned int count)
 177 {
 178         struct nfs_page *req;
 179         int ret;
 180
 181         for (;;) {
 182                 req = nfs_update_request(ctx, page, offset, count);
 183                 if (!IS_ERR(req))
 184                         break;
 185                 ret = PTR_ERR(req);
 186                 if (ret != -EBUSY)
 187                         return ret;
 188                 ret = nfs_wb_page(page->mapping->host, page);
 189                 if (ret != 0)
 190                         return ret;
 191         }
 192         /* Update file length */
 193         nfs_grow_file(page, offset, count);
 194         /* Set the PG_uptodate flag? */
 195         nfs_mark_uptodate(page, offset, count);
 196         nfs_unlock_request(req);
 197         return 0;
 198 }
 199
 200 static int wb_priority(struct writeback_control *wbc)
 201 {
 202         if (wbc->for_reclaim)
 203                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 204         if (wbc->for_kupdate)
 205                 return FLUSH_LOWPRI;
 206         return 0;
 207 }
 208
 209 /*
 210  * NFS congestion control
 211  */
 212
 213 int nfs_congestion_kb;
 214
 215 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 216 #define NFS_CONGESTION_OFF_THRESH       \
 217         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 218
 219 static int nfs_set_page_writeback(struct page *page)
 220 {
 221         int ret = test_set_page_writeback(page);
 222
 223         if (!ret) {
 224                 struct inode *inode = page->mapping->host;
 225                 struct nfs_server *nfss = NFS_SERVER(inode);
 226
 227                 if (atomic_long_inc_return(&nfss->writeback) >
 228                                 NFS_CONGESTION_ON_THRESH)
 229                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 230         }
 231         return ret;
 232 }
 233
 234 static void nfs_end_page_writeback(struct page *page)
 235 {
 236         struct inode *inode = page->mapping->host;
 237         struct nfs_server *nfss = NFS_SERVER(inode);
 238
 239         end_page_writeback(page);
 240         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 241                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 242                 congestion_end(WRITE);
 243         }
 244 }
 245
 246 /*
 247  * Find an associated nfs write request, and prepare to flush it out
 248  * Returns 1 if there was no write request, or if the request was
 249  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 250  * was not tagged.
 251  * May also return an error if the user signalled nfs_wait_on_request().
 252  */
 253 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 254                                 struct page *page)
 255 {
 256         struct nfs_page *req;
 257         struct nfs_inode *nfsi = NFS_I(page->mapping->host);
 258         spinlock_t *req_lock = &nfsi->req_lock;
 259         int ret;
 260
 261         spin_lock(req_lock);
 262         for(;;) {
 263                 req = nfs_page_find_request_locked(page);
 264                 if (req == NULL) {
 265                         spin_unlock(req_lock);
 266                         return 1;
 267                 }
 268                 if (nfs_lock_request_dontget(req))
 269                         break;
 270                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 271                  *       then the call to nfs_lock_request_dontget() will always
 272                  *       succeed provided that someone hasn't already marked the
 273                  *       request as dirty (in which case we don't care).
 274                  */
 275                 spin_unlock(req_lock);
 276                 /* Prevent deadlock! */
 277                 nfs_pageio_complete(pgio);
 278                 ret = nfs_wait_on_request(req);
 279                 nfs_release_request(req);
 280                 if (ret != 0)
 281                         return ret;
 282                 spin_lock(req_lock);
 283         }
 284         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 285                 /* This request is marked for commit */
 286                 spin_unlock(req_lock);
 287                 nfs_unlock_request(req);
 288                 nfs_pageio_complete(pgio);
 289                 return 1;
 290         }
 291         if (nfs_set_page_writeback(page) != 0) {
 292                 spin_unlock(req_lock);
 293                 BUG();
 294         }
 295         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 296                         NFS_PAGE_TAG_WRITEBACK);
 297         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 298         spin_unlock(req_lock);
 299         nfs_pageio_add_request(pgio, req);
 300         return ret;
 301 }
 302
 303 /*
 304  * Write an mmapped page to the server.
 305  */
 306 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 307 {
 308         struct nfs_pageio_descriptor mypgio, *pgio;
 309         struct nfs_open_context *ctx;
 310         struct inode *inode = page->mapping->host;
 311         unsigned offset;
 312         int err;
 313
 314         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 315         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 316
 317         if (wbc->for_writepages)
 318                 pgio = wbc->fs_private;
 319         else {
 320                 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
 321                 pgio = &mypgio;
 322         }
 323
 324         err = nfs_page_async_flush(pgio, page);
 325         if (err <= 0)
 326                 goto out;
 327         err = 0;
 328         offset = nfs_page_length(page);
 329         if (!offset)
 330                 goto out;
 331
 332         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 333         if (ctx == NULL) {
 334                 err = -EBADF;
 335                 goto out;
 336         }
 337         err = nfs_writepage_setup(ctx, page, 0, offset);
 338         put_nfs_open_context(ctx);
 339         if (err != 0)
 340                 goto out;
 341         err = nfs_page_async_flush(pgio, page);
 342         if (err > 0)
 343                 err = 0;
 344 out:
 345         if (!wbc->for_writepages)
 346                 nfs_pageio_complete(pgio);
 347         return err;
 348 }
 349
 350 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 351 {
 352         int err;
 353
 354         err = nfs_writepage_locked(page, wbc);
 355         unlock_page(page);
 356         return err;
 357 }
 358
 359 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 360 {
 361         struct inode *inode = mapping->host;
 362         struct nfs_pageio_descriptor pgio;
 363         int err;
 364
 365         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 366
 367         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 368         wbc->fs_private = &pgio;
 369         err = generic_writepages(mapping, wbc);
 370         nfs_pageio_complete(&pgio);
 371         if (err)
 372                 return err;
 373         if (pgio.pg_error)
 374                 return pgio.pg_error;
 375         return 0;
 376 }
 377
 378 /*
 379  * Insert a write request into an inode
 380  */
 381 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 382 {
 383         struct nfs_inode *nfsi = NFS_I(inode);
 384         int error;
 385
 386         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 387         BUG_ON(error == -EEXIST);
 388         if (error)
 389                 return error;
 390         if (!nfsi->npages) {
 391                 igrab(inode);
 392                 nfs_begin_data_update(inode);
 393                 if (nfs_have_delegation(inode, FMODE_WRITE))
 394                         nfsi->change_attr++;
 395         }
 396         SetPagePrivate(req->wb_page);
 397         set_page_private(req->wb_page, (unsigned long)req);
 398         if (PageDirty(req->wb_page))
 399                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
 400         nfsi->npages++;
 401         atomic_inc(&req->wb_count);
 402         return 0;
 403 }
 404
 405 /*
 406  * Remove a write request from an inode
 407  */
 408 static void nfs_inode_remove_request(struct nfs_page *req)
 409 {
 410         struct inode *inode = req->wb_context->dentry->d_inode;
 411         struct nfs_inode *nfsi = NFS_I(inode);
 412
 413         BUG_ON (!NFS_WBACK_BUSY(req));
 414
 415         spin_lock(&nfsi->req_lock);
 416         set_page_private(req->wb_page, 0);
 417         ClearPagePrivate(req->wb_page);
 418         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 419         if (test_and_clear_bit(PG_NEED_FLUSH, &req->wb_flags))
 420                 __set_page_dirty_nobuffers(req->wb_page);
 421         nfsi->npages--;
 422         if (!nfsi->npages) {
 423                 spin_unlock(&nfsi->req_lock);
 424                 nfs_end_data_update(inode);
 425                 iput(inode);
 426         } else
 427                 spin_unlock(&nfsi->req_lock);
 428         nfs_clear_request(req);
 429         nfs_release_request(req);
 430 }
 431
 432 static void
 433 nfs_redirty_request(struct nfs_page *req)
 434 {
 435         __set_page_dirty_nobuffers(req->wb_page);
 436 }
 437
 438 /*
 439  * Check if a request is dirty
 440  */
 441 static inline int
 442 nfs_dirty_request(struct nfs_page *req)
 443 {
 444         struct page *page = req->wb_page;
 445
 446         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 447                 return 0;
 448         return !PageWriteback(req->wb_page);
 449 }
 450
 451 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 452 /*
 453  * Add a request to the inode's commit list.
 454  */
 455 static void
 456 nfs_mark_request_commit(struct nfs_page *req)
 457 {
 458         struct inode *inode = req->wb_context->dentry->d_inode;
 459         struct nfs_inode *nfsi = NFS_I(inode);
 460
 461         spin_lock(&nfsi->req_lock);
 462         nfs_list_add_request(req, &nfsi->commit);
 463         nfsi->ncommit++;
 464         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 465         spin_unlock(&nfsi->req_lock);
 466         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 467         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 468 }
 469
 470 static inline
 471 int nfs_write_need_commit(struct nfs_write_data *data)
 472 {
 473         return data->verf.committed != NFS_FILE_SYNC;
 474 }
 475
 476 static inline
 477 int nfs_reschedule_unstable_write(struct nfs_page *req)
 478 {
 479         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 480                 nfs_mark_request_commit(req);
 481                 return 1;
 482         }
 483         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 484                 nfs_redirty_request(req);
 485                 return 1;
 486         }
 487         return 0;
 488 }
 489 #else
 490 static inline void
 491 nfs_mark_request_commit(struct nfs_page *req)
 492 {
 493 }
 494
 495 static inline
 496 int nfs_write_need_commit(struct nfs_write_data *data)
 497 {
 498         return 0;
 499 }
 500
 501 static inline
 502 int nfs_reschedule_unstable_write(struct nfs_page *req)
 503 {
 504         return 0;
 505 }
 506 #endif
 507
 508 /*
 509  * Wait for a request to complete.
 510  *
 511  * Interruptible by signals only if mounted with intr flag.
 512  */
 513 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 514 {
 515         struct nfs_inode *nfsi = NFS_I(inode);
 516         struct nfs_page *req;
 517         pgoff_t idx_end, next;
 518         unsigned int            res = 0;
 519         int                     error;
 520
 521         if (npages == 0)
 522                 idx_end = ~0;
 523         else
 524                 idx_end = idx_start + npages - 1;
 525
 526         next = idx_start;
 527         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 528                 if (req->wb_index > idx_end)
 529                         break;
 530
 531                 next = req->wb_index + 1;
 532                 BUG_ON(!NFS_WBACK_BUSY(req));
 533
 534                 atomic_inc(&req->wb_count);
 535                 spin_unlock(&nfsi->req_lock);
 536                 error = nfs_wait_on_request(req);
 537                 nfs_release_request(req);
 538                 spin_lock(&nfsi->req_lock);
 539                 if (error < 0)
 540                         return error;
 541                 res++;
 542         }
 543         return res;
 544 }
 545
 546 static void nfs_cancel_commit_list(struct list_head *head)
 547 {
 548         struct nfs_page *req;
 549
 550         while(!list_empty(head)) {
 551                 req = nfs_list_entry(head->next);
 552                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 553                 nfs_list_remove_request(req);
 554                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 555                 nfs_inode_remove_request(req);
 556                 nfs_unlock_request(req);
 557         }
 558 }
 559
 560 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 561 /*
 562  * nfs_scan_commit - Scan an inode for commit requests
 563  * @inode: NFS inode to scan
 564  * @dst: destination list
 565  * @idx_start: lower bound of page->index to scan.
 566  * @npages: idx_start + npages sets the upper bound to scan.
 567  *
 568  * Moves requests from the inode's 'commit' request list.
 569  * The requests are *not* checked to ensure that they form a contiguous set.
 570  */
 571 static int
 572 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 573 {
 574         struct nfs_inode *nfsi = NFS_I(inode);
 575         int res = 0;
 576
 577         if (nfsi->ncommit != 0) {
 578                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 579                 nfsi->ncommit -= res;
 580                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 581                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 582         }
 583         return res;
 584 }
 585 #else
 586 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 587 {
 588         return 0;
 589 }
 590 #endif
 591
 592 /*
 593  * Try to update any existing write request, or create one if there is none.
 594  * In order to match, the request's credentials must match those of
 595  * the calling process.
 596  *
 597  * Note: Should always be called with the Page Lock held!
 598  */
 599 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 600                 struct page *page, unsigned int offset, unsigned int bytes)
 601 {
 602         struct address_space *mapping = page->mapping;
 603         struct inode *inode = mapping->host;
 604         struct nfs_inode *nfsi = NFS_I(inode);
 605         struct nfs_page         *req, *new = NULL;
 606         pgoff_t         rqend, end;
 607
 608         end = offset + bytes;
 609
 610         for (;;) {
 611                 /* Loop over all inode entries and see if we find
 612                  * A request for the page we wish to update
 613                  */
 614                 spin_lock(&nfsi->req_lock);
 615                 req = nfs_page_find_request_locked(page);
 616                 if (req) {
 617                         if (!nfs_lock_request_dontget(req)) {
 618                                 int error;
 619
 620                                 spin_unlock(&nfsi->req_lock);
 621                                 error = nfs_wait_on_request(req);
 622                                 nfs_release_request(req);
 623                                 if (error < 0) {
 624                                         if (new)
 625                                                 nfs_release_request(new);
 626                                         return ERR_PTR(error);
 627                                 }
 628                                 continue;
 629                         }
 630                         spin_unlock(&nfsi->req_lock);
 631                         if (new)
 632                                 nfs_release_request(new);
 633                         break;
 634                 }
 635
 636                 if (new) {
 637                         int error;
 638                         nfs_lock_request_dontget(new);
 639                         error = nfs_inode_add_request(inode, new);
 640                         if (error) {
 641                                 spin_unlock(&nfsi->req_lock);
 642                                 nfs_unlock_request(new);
 643                                 return ERR_PTR(error);
 644                         }
 645                         spin_unlock(&nfsi->req_lock);
 646                         return new;
 647                 }
 648                 spin_unlock(&nfsi->req_lock);
 649
 650                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 651                 if (IS_ERR(new))
 652                         return new;
 653         }
 654
 655         /* We have a request for our page.
 656          * If the creds don't match, or the
 657          * page addresses don't match,
 658          * tell the caller to wait on the conflicting
 659          * request.
 660          */
 661         rqend = req->wb_offset + req->wb_bytes;
 662         if (req->wb_context != ctx
 663             || req->wb_page != page
 664             || !nfs_dirty_request(req)
 665             || offset > rqend || end < req->wb_offset) {
 666                 nfs_unlock_request(req);
 667                 return ERR_PTR(-EBUSY);
 668         }
 669
 670         /* Okay, the request matches. Update the region */
 671         if (offset < req->wb_offset) {
 672                 req->wb_offset = offset;
 673                 req->wb_pgbase = offset;
 674                 req->wb_bytes = rqend - req->wb_offset;
 675         }
 676
 677         if (end > rqend)
 678                 req->wb_bytes = end - req->wb_offset;
 679
 680         return req;
 681 }
 682
 683 int nfs_flush_incompatible(struct file *file, struct page *page)
 684 {
 685         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 686         struct nfs_page *req;
 687         int do_flush, status;
 688         /*
 689          * Look for a request corresponding to this page. If there
 690          * is one, and it belongs to another file, we flush it out
 691          * before we try to copy anything into the page. Do this
 692          * due to the lack of an ACCESS-type call in NFSv2.
 693          * Also do the same if we find a request from an existing
 694          * dropped page.
 695          */
 696         do {
 697                 req = nfs_page_find_request(page);
 698                 if (req == NULL)
 699                         return 0;
 700                 do_flush = req->wb_page != page || req->wb_context != ctx
 701                         || !nfs_dirty_request(req);
 702                 nfs_release_request(req);
 703                 if (!do_flush)
 704                         return 0;
 705                 status = nfs_wb_page(page->mapping->host, page);
 706         } while (status == 0);
 707         return status;
 708 }
 709
 710 /*
 711  * Update and possibly write a cached page of an NFS file.
 712  *
 713  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 714  * things with a page scheduled for an RPC call (e.g. invalidate it).
 715  */
 716 int nfs_updatepage(struct file *file, struct page *page,
 717                 unsigned int offset, unsigned int count)
 718 {
 719         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 720         struct inode    *inode = page->mapping->host;
 721         int             status = 0;
 722
 723         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 724
 725         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 726                 file->f_path.dentry->d_parent->d_name.name,
 727                 file->f_path.dentry->d_name.name, count,
 728                 (long long)(page_offset(page) +offset));
 729
 730         /* If we're not using byte range locks, and we know the page
 731          * is entirely in cache, it may be more efficient to avoid
 732          * fragmenting write requests.
 733          */
 734         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 735                 count = max(count + offset, nfs_page_length(page));
 736                 offset = 0;
 737         }
 738
 739         status = nfs_writepage_setup(ctx, page, offset, count);
 740         __set_page_dirty_nobuffers(page);
 741
 742         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 743                         status, (long long)i_size_read(inode));
 744         if (status < 0)
 745                 nfs_set_pageerror(page);
 746         return status;
 747 }
 748
 749 static void nfs_writepage_release(struct nfs_page *req)
 750 {
 751
 752         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 753                 nfs_end_page_writeback(req->wb_page);
 754                 nfs_inode_remove_request(req);
 755         } else
 756                 nfs_end_page_writeback(req->wb_page);
 757         nfs_clear_page_writeback(req);
 758 }
 759
 760 static inline int flush_task_priority(int how)
 761 {
 762         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 763                 case FLUSH_HIGHPRI:
 764                         return RPC_PRIORITY_HIGH;
 765                 case FLUSH_LOWPRI:
 766                         return RPC_PRIORITY_LOW;
 767         }
 768         return RPC_PRIORITY_NORMAL;
 769 }
 770
 771 /*
 772  * Set up the argument/result storage required for the RPC call.
 773  */
 774 static void nfs_write_rpcsetup(struct nfs_page *req,
 775                 struct nfs_write_data *data,
 776                 const struct rpc_call_ops *call_ops,
 777                 unsigned int count, unsigned int offset,
 778                 int how)
 779 {
 780         struct inode            *inode;
 781         int flags;
 782
 783         /* Set up the RPC argument and reply structs
 784          * NB: take care not to mess about with data->commit et al. */
 785
 786         data->req = req;
 787         data->inode = inode = req->wb_context->dentry->d_inode;
 788         data->cred = req->wb_context->cred;
 789
 790         data->args.fh     = NFS_FH(inode);
 791         data->args.offset = req_offset(req) + offset;
 792         data->args.pgbase = req->wb_pgbase + offset;
 793         data->args.pages  = data->pagevec;
 794         data->args.count  = count;
 795         data->args.context = req->wb_context;
 796
 797         data->res.fattr   = &data->fattr;
 798         data->res.count   = count;
 799         data->res.verf    = &data->verf;
 800         nfs_fattr_init(&data->fattr);
 801
 802         /* Set up the initial task struct.  */
 803         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 804         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 805         NFS_PROTO(inode)->write_setup(data, how);
 806
 807         data->task.tk_priority = flush_task_priority(how);
 808         data->task.tk_cookie = (unsigned long)inode;
 809
 810         dprintk("NFS: %5u initiated write call "
 811                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 812                 data->task.tk_pid,
 813                 inode->i_sb->s_id,
 814                 (long long)NFS_FILEID(inode),
 815                 count,
 816                 (unsigned long long)data->args.offset);
 817 }
 818
 819 static void nfs_execute_write(struct nfs_write_data *data)
 820 {
 821         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 822         sigset_t oldset;
 823
 824         rpc_clnt_sigmask(clnt, &oldset);
 825         rpc_execute(&data->task);
 826         rpc_clnt_sigunmask(clnt, &oldset);
 827 }
 828
 829 /*
 830  * Generate multiple small requests to write out a single
 831  * contiguous dirty area on one page.
 832  */
 833 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 834 {
 835         struct nfs_page *req = nfs_list_entry(head->next);
 836         struct page *page = req->wb_page;
 837         struct nfs_write_data *data;
 838         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 839         unsigned int offset;
 840         int requests = 0;
 841         LIST_HEAD(list);
 842
 843         nfs_list_remove_request(req);
 844
 845         nbytes = count;
 846         do {
 847                 size_t len = min(nbytes, wsize);
 848
 849                 data = nfs_writedata_alloc(1);
 850                 if (!data)
 851                         goto out_bad;
 852                 list_add(&data->pages, &list);
 853                 requests++;
 854                 nbytes -= len;
 855         } while (nbytes != 0);
 856         atomic_set(&req->wb_complete, requests);
 857
 858         ClearPageError(page);
 859         offset = 0;
 860         nbytes = count;
 861         do {
 862                 data = list_entry(list.next, struct nfs_write_data, pages);
 863                 list_del_init(&data->pages);
 864
 865                 data->pagevec[0] = page;
 866
 867                 if (nbytes < wsize)
 868                         wsize = nbytes;
 869                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 870                                    wsize, offset, how);
 871                 offset += wsize;
 872                 nbytes -= wsize;
 873                 nfs_execute_write(data);
 874         } while (nbytes != 0);
 875
 876         return 0;
 877
 878 out_bad:
 879         while (!list_empty(&list)) {
 880                 data = list_entry(list.next, struct nfs_write_data, pages);
 881                 list_del(&data->pages);
 882                 nfs_writedata_release(data);
 883         }
 884         nfs_redirty_request(req);
 885         nfs_end_page_writeback(req->wb_page);
 886         nfs_clear_page_writeback(req);
 887         return -ENOMEM;
 888 }
 889
 890 /*
 891  * Create an RPC task for the given write request and kick it.
 892  * The page must have been locked by the caller.
 893  *
 894  * It may happen that the page we're passed is not marked dirty.
 895  * This is the case if nfs_updatepage detects a conflicting request
 896  * that has been written but not committed.
 897  */
 898 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 899 {
 900         struct nfs_page         *req;
 901         struct page             **pages;
 902         struct nfs_write_data   *data;
 903
 904         data = nfs_writedata_alloc(npages);
 905         if (!data)
 906                 goto out_bad;
 907
 908         pages = data->pagevec;
 909         while (!list_empty(head)) {
 910                 req = nfs_list_entry(head->next);
 911                 nfs_list_remove_request(req);
 912                 nfs_list_add_request(req, &data->pages);
 913                 ClearPageError(req->wb_page);
 914                 *pages++ = req->wb_page;
 915         }
 916         req = nfs_list_entry(data->pages.next);
 917
 918         /* Set up the argument struct */
 919         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 920
 921         nfs_execute_write(data);
 922         return 0;
 923  out_bad:
 924         while (!list_empty(head)) {
 925                 req = nfs_list_entry(head->next);
 926                 nfs_list_remove_request(req);
 927                 nfs_redirty_request(req);
 928                 nfs_end_page_writeback(req->wb_page);
 929                 nfs_clear_page_writeback(req);
 930         }
 931         return -ENOMEM;
 932 }
 933
 934 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 935                                   struct inode *inode, int ioflags)
 936 {
 937         int wsize = NFS_SERVER(inode)->wsize;
 938
 939         if (wsize < PAGE_CACHE_SIZE)
 940                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 941         else
 942                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 943 }
 944
 945 /*
 946  * Handle a write reply that flushed part of a page.
 947  */
 948 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 949 {
 950         struct nfs_write_data   *data = calldata;
 951         struct nfs_page         *req = data->req;
 952         struct page             *page = req->wb_page;
 953
 954         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 955                 req->wb_context->dentry->d_inode->i_sb->s_id,
 956                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 957                 req->wb_bytes,
 958                 (long long)req_offset(req));
 959
 960         if (nfs_writeback_done(task, data) != 0)
 961                 return;
 962
 963         if (task->tk_status < 0) {
 964                 nfs_set_pageerror(page);
 965                 req->wb_context->error = task->tk_status;
 966                 dprintk(", error = %d\n", task->tk_status);
 967                 goto out;
 968         }
 969
 970         if (nfs_write_need_commit(data)) {
 971                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 972
 973                 spin_lock(req_lock);
 974                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 975                         /* Do nothing we need to resend the writes */
 976                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 977                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 978                         dprintk(" defer commit\n");
 979                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 980                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 981                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 982                         dprintk(" server reboot detected\n");
 983                 }
 984                 spin_unlock(req_lock);
 985         } else
 986                 dprintk(" OK\n");
 987
 988 out:
 989         if (atomic_dec_and_test(&req->wb_complete))
 990                 nfs_writepage_release(req);
 991 }
 992
 993 static const struct rpc_call_ops nfs_write_partial_ops = {
 994         .rpc_call_done = nfs_writeback_done_partial,
 995         .rpc_release = nfs_writedata_release,
 996 };
 997
 998 /*
 999  * Handle a write reply that flushes a whole page.
1000  *
1001  * FIXME: There is an inherent race with invalidate_inode_pages and
1002  *        writebacks since the page->count is kept > 1 for as long
1003  *        as the page has a write request pending.
1004  */
1005 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1006 {
1007         struct nfs_write_data   *data = calldata;
1008         struct nfs_page         *req;
1009         struct page             *page;
1010
1011         if (nfs_writeback_done(task, data) != 0)
1012                 return;
1013
1014         /* Update attributes as result of writeback. */
1015         while (!list_empty(&data->pages)) {
1016                 req = nfs_list_entry(data->pages.next);
1017                 nfs_list_remove_request(req);
1018                 page = req->wb_page;
1019
1020                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1021                         req->wb_context->dentry->d_inode->i_sb->s_id,
1022                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1023                         req->wb_bytes,
1024                         (long long)req_offset(req));
1025
1026                 if (task->tk_status < 0) {
1027                         nfs_set_pageerror(page);
1028                         req->wb_context->error = task->tk_status;
1029                         dprintk(", error = %d\n", task->tk_status);
1030                         goto remove_request;
1031                 }
1032
1033                 if (nfs_write_need_commit(data)) {
1034                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1035                         nfs_mark_request_commit(req);
1036                         nfs_end_page_writeback(page);
1037                         dprintk(" marked for commit\n");
1038                         goto next;
1039                 }
1040                 dprintk(" OK\n");
1041 remove_request:
1042                 nfs_end_page_writeback(page);
1043                 nfs_inode_remove_request(req);
1044         next:
1045                 nfs_clear_page_writeback(req);
1046         }
1047 }
1048
1049 static const struct rpc_call_ops nfs_write_full_ops = {
1050         .rpc_call_done = nfs_writeback_done_full,
1051         .rpc_release = nfs_writedata_release,
1052 };
1053
1054
1055 /*
1056  * This function is called when the WRITE call is complete.
1057  */
1058 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1059 {
1060         struct nfs_writeargs    *argp = &data->args;
1061         struct nfs_writeres     *resp = &data->res;
1062         int status;
1063
1064         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1065                 task->tk_pid, task->tk_status);
1066
1067         /*
1068          * ->write_done will attempt to use post-op attributes to detect
1069          * conflicting writes by other clients.  A strict interpretation
1070          * of close-to-open would allow us to continue caching even if
1071          * another writer had changed the file, but some applications
1072          * depend on tighter cache coherency when writing.
1073          */
1074         status = NFS_PROTO(data->inode)->write_done(task, data);
1075         if (status != 0)
1076                 return status;
1077         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1078
1079 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1080         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1081                 /* We tried a write call, but the server did not
1082                  * commit data to stable storage even though we
1083                  * requested it.
1084                  * Note: There is a known bug in Tru64 < 5.0 in which
1085                  *       the server reports NFS_DATA_SYNC, but performs
1086                  *       NFS_FILE_SYNC. We therefore implement this checking
1087                  *       as a dprintk() in order to avoid filling syslog.
1088                  */
1089                 static unsigned long    complain;
1090
1091                 if (time_before(complain, jiffies)) {
1092                         dprintk("NFS: faulty NFS server %s:"
1093                                 " (committed = %d) != (stable = %d)\n",
1094                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1095                                 resp->verf->committed, argp->stable);
1096                         complain = jiffies + 300 * HZ;
1097                 }
1098         }
1099 #endif
1100         /* Is this a short write? */
1101         if (task->tk_status >= 0 && resp->count < argp->count) {
1102                 static unsigned long    complain;
1103
1104                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1105
1106                 /* Has the server at least made some progress? */
1107                 if (resp->count != 0) {
1108                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1109                         if (resp->verf->committed != NFS_UNSTABLE) {
1110                                 /* Resend from where the server left off */
1111                                 argp->offset += resp->count;
1112                                 argp->pgbase += resp->count;
1113                                 argp->count -= resp->count;
1114                         } else {
1115                                 /* Resend as a stable write in order to avoid
1116                                  * headaches in the case of a server crash.
1117                                  */
1118                                 argp->stable = NFS_FILE_SYNC;
1119                         }
1120                         rpc_restart_call(task);
1121                         return -EAGAIN;
1122                 }
1123                 if (time_before(complain, jiffies)) {
1124                         printk(KERN_WARNING
1125                                "NFS: Server wrote zero bytes, expected %u.\n",
1126                                         argp->count);
1127                         complain = jiffies + 300 * HZ;
1128                 }
1129                 /* Can't do anything about it except throw an error. */
1130                 task->tk_status = -EIO;
1131         }
1132         return 0;
1133 }
1134
1135
1136 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1137 void nfs_commit_release(void *wdata)
1138 {
1139         nfs_commit_free(wdata);
1140 }
1141
1142 /*
1143  * Set up the argument/result storage required for the RPC call.
1144  */
1145 static void nfs_commit_rpcsetup(struct list_head *head,
1146                 struct nfs_write_data *data,
1147                 int how)
1148 {
1149         struct nfs_page         *first;
1150         struct inode            *inode;
1151         int flags;
1152
1153         /* Set up the RPC argument and reply structs
1154          * NB: take care not to mess about with data->commit et al. */
1155
1156         list_splice_init(head, &data->pages);
1157         first = nfs_list_entry(data->pages.next);
1158         inode = first->wb_context->dentry->d_inode;
1159
1160         data->inode       = inode;
1161         data->cred        = first->wb_context->cred;
1162
1163         data->args.fh     = NFS_FH(data->inode);
1164         /* Note: we always request a commit of the entire inode */
1165         data->args.offset = 0;
1166         data->args.count  = 0;
1167         data->res.count   = 0;
1168         data->res.fattr   = &data->fattr;
1169         data->res.verf    = &data->verf;
1170         nfs_fattr_init(&data->fattr);
1171
1172         /* Set up the initial task struct.  */
1173         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1174         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1175         NFS_PROTO(inode)->commit_setup(data, how);
1176
1177         data->task.tk_priority = flush_task_priority(how);
1178         data->task.tk_cookie = (unsigned long)inode;
1179
1180         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1181 }
1182
1183 /*
1184  * Commit dirty pages
1185  */
1186 static int
1187 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1188 {
1189         struct nfs_write_data   *data;
1190         struct nfs_page         *req;
1191
1192         data = nfs_commit_alloc();
1193
1194         if (!data)
1195                 goto out_bad;
1196
1197         /* Set up the argument struct */
1198         nfs_commit_rpcsetup(head, data, how);
1199
1200         nfs_execute_write(data);
1201         return 0;
1202  out_bad:
1203         while (!list_empty(head)) {
1204                 req = nfs_list_entry(head->next);
1205                 nfs_list_remove_request(req);
1206                 nfs_mark_request_commit(req);
1207                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1208                 nfs_clear_page_writeback(req);
1209         }
1210         return -ENOMEM;
1211 }
1212
1213 /*
1214  * COMMIT call returned
1215  */
1216 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1217 {
1218         struct nfs_write_data   *data = calldata;
1219         struct nfs_page         *req;
1220
1221         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1222                                 task->tk_pid, task->tk_status);
1223
1224         /* Call the NFS version-specific code */
1225         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1226                 return;
1227
1228         while (!list_empty(&data->pages)) {
1229                 req = nfs_list_entry(data->pages.next);
1230                 nfs_list_remove_request(req);
1231                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1232                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1233
1234                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1235                         req->wb_context->dentry->d_inode->i_sb->s_id,
1236                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1237                         req->wb_bytes,
1238                         (long long)req_offset(req));
1239                 if (task->tk_status < 0) {
1240                         req->wb_context->error = task->tk_status;
1241                         nfs_inode_remove_request(req);
1242                         dprintk(", error = %d\n", task->tk_status);
1243                         goto next;
1244                 }
1245
1246                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1247                  * returned by the server against all stored verfs. */
1248                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1249                         /* We have a match */
1250                         nfs_inode_remove_request(req);
1251                         dprintk(" OK\n");
1252                         goto next;
1253                 }
1254                 /* We have a mismatch. Write the page again */
1255                 dprintk(" mismatch\n");
1256                 nfs_redirty_request(req);
1257         next:
1258                 nfs_clear_page_writeback(req);
1259         }
1260 }
1261
1262 static const struct rpc_call_ops nfs_commit_ops = {
1263         .rpc_call_done = nfs_commit_done,
1264         .rpc_release = nfs_commit_release,
1265 };
1266
1267 int nfs_commit_inode(struct inode *inode, int how)
1268 {
1269         struct nfs_inode *nfsi = NFS_I(inode);
1270         LIST_HEAD(head);
1271         int res;
1272
1273         spin_lock(&nfsi->req_lock);
1274         res = nfs_scan_commit(inode, &head, 0, 0);
1275         spin_unlock(&nfsi->req_lock);
1276         if (res) {
1277                 int error = nfs_commit_list(inode, &head, how);
1278                 if (error < 0)
1279                         return error;
1280         }
1281         return res;
1282 }
1283 #else
1284 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1285 {
1286         return 0;
1287 }
1288 #endif
1289
1290 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1291 {
1292         struct inode *inode = mapping->host;
1293         struct nfs_inode *nfsi = NFS_I(inode);
1294         pgoff_t idx_start, idx_end;
1295         unsigned int npages = 0;
1296         LIST_HEAD(head);
1297         int nocommit = how & FLUSH_NOCOMMIT;
1298         long pages, ret;
1299
1300         /* FIXME */
1301         if (wbc->range_cyclic)
1302                 idx_start = 0;
1303         else {
1304                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1305                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1306                 if (idx_end > idx_start) {
1307                         pgoff_t l_npages = 1 + idx_end - idx_start;
1308                         npages = l_npages;
1309                         if (sizeof(npages) != sizeof(l_npages) &&
1310                                         (pgoff_t)npages != l_npages)
1311                                 npages = 0;
1312                 }
1313         }
1314         how &= ~FLUSH_NOCOMMIT;
1315         spin_lock(&nfsi->req_lock);
1316         do {
1317                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1318                 if (ret != 0)
1319                         continue;
1320                 if (nocommit)
1321                         break;
1322                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1323                 if (pages == 0)
1324                         break;
1325                 if (how & FLUSH_INVALIDATE) {
1326                         spin_unlock(&nfsi->req_lock);
1327                         nfs_cancel_commit_list(&head);
1328                         ret = pages;
1329                         spin_lock(&nfsi->req_lock);
1330                         continue;
1331                 }
1332                 pages += nfs_scan_commit(inode, &head, 0, 0);
1333                 spin_unlock(&nfsi->req_lock);
1334                 ret = nfs_commit_list(inode, &head, how);
1335                 spin_lock(&nfsi->req_lock);
1336         } while (ret >= 0);
1337         spin_unlock(&nfsi->req_lock);
1338         return ret;
1339 }
1340
1341 /*
1342  * flush the inode to disk.
1343  */
1344 int nfs_wb_all(struct inode *inode)
1345 {
1346         struct address_space *mapping = inode->i_mapping;
1347         struct writeback_control wbc = {
1348                 .bdi = mapping->backing_dev_info,
1349                 .sync_mode = WB_SYNC_ALL,
1350                 .nr_to_write = LONG_MAX,
1351                 .for_writepages = 1,
1352                 .range_cyclic = 1,
1353         };
1354         int ret;
1355
1356         ret = nfs_writepages(mapping, &wbc);
1357         if (ret < 0)
1358                 goto out;
1359         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1360         if (ret >= 0)
1361                 return 0;
1362 out:
1363         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1364         return ret;
1365 }
1366
1367 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1368 {
1369         struct writeback_control wbc = {
1370                 .bdi = mapping->backing_dev_info,
1371                 .sync_mode = WB_SYNC_ALL,
1372                 .nr_to_write = LONG_MAX,
1373                 .range_start = range_start,
1374                 .range_end = range_end,
1375                 .for_writepages = 1,
1376         };
1377         int ret;
1378
1379         ret = nfs_writepages(mapping, &wbc);
1380         if (ret < 0)
1381                 goto out;
1382         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1383         if (ret >= 0)
1384                 return 0;
1385 out:
1386         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1387         return ret;
1388 }
1389
1390 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1391 {
1392         loff_t range_start = page_offset(page);
1393         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1394         struct writeback_control wbc = {
1395                 .bdi = page->mapping->backing_dev_info,
1396                 .sync_mode = WB_SYNC_ALL,
1397                 .nr_to_write = LONG_MAX,
1398                 .range_start = range_start,
1399                 .range_end = range_end,
1400         };
1401         int ret;
1402
1403         BUG_ON(!PageLocked(page));
1404         if (clear_page_dirty_for_io(page)) {
1405                 ret = nfs_writepage_locked(page, &wbc);
1406                 if (ret < 0)
1407                         goto out;
1408         }
1409         if (!PagePrivate(page))
1410                 return 0;
1411         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1412         if (ret >= 0)
1413                 return 0;
1414 out:
1415         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1416         return ret;
1417 }
1418
1419 /*
1420  * Write back all requests on one page - we do this before reading it.
1421  */
1422 int nfs_wb_page(struct inode *inode, struct page* page)
1423 {
1424         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1425 }
1426
1427 int nfs_set_page_dirty(struct page *page)
1428 {
1429         struct address_space *mapping = page->mapping;
1430         struct inode *inode;
1431         spinlock_t *req_lock;
1432         struct nfs_page *req;
1433         int ret;
1434
1435         if (!mapping)
1436                 goto out_raced;
1437         inode = mapping->host;
1438         if (!inode)
1439                 goto out_raced;
1440         req_lock = &NFS_I(inode)->req_lock;
1441         spin_lock(req_lock);
1442         req = nfs_page_find_request_locked(page);
1443         if (req != NULL) {
1444                 /* Mark any existing write requests for flushing */
1445                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1446                 spin_unlock(req_lock);
1447                 nfs_release_request(req);
1448                 return ret;
1449         }
1450         ret = __set_page_dirty_nobuffers(page);
1451         spin_unlock(req_lock);
1452         return ret;
1453 out_raced:
1454         return !TestSetPageDirty(page);
1455 }
1456
1457
1458 int __init nfs_init_writepagecache(void)
1459 {
1460         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1461                                              sizeof(struct nfs_write_data),
1462                                              0, SLAB_HWCACHE_ALIGN,
1463                                              NULL, NULL);
1464         if (nfs_wdata_cachep == NULL)
1465                 return -ENOMEM;
1466
1467         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1468                                                      nfs_wdata_cachep);
1469         if (nfs_wdata_mempool == NULL)
1470                 return -ENOMEM;
1471
1472         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1473                                                       nfs_wdata_cachep);
1474         if (nfs_commit_mempool == NULL)
1475                 return -ENOMEM;
1476
1477         /*
1478          * NFS congestion size, scale with available memory.
1479          *
1480          *  64MB:    8192k
1481          * 128MB:   11585k
1482          * 256MB:   16384k
1483          * 512MB:   23170k
1484          *   1GB:   32768k
1485          *   2GB:   46340k
1486          *   4GB:   65536k
1487          *   8GB:   92681k
1488          *  16GB:  131072k
1489          *
1490          * This allows larger machines to have larger/more transfers.
1491          * Limit the default to 256M
1492          */
1493         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1494         if (nfs_congestion_kb > 256*1024)
1495                 nfs_congestion_kb = 256*1024;
1496
1497         return 0;
1498 }
1499
1500 void nfs_destroy_writepagecache(void)
1501 {
1502         mempool_destroy(nfs_commit_mempool);
1503         mempool_destroy(nfs_wdata_mempool);
1504         kmem_cache_destroy(nfs_wdata_cachep);
1505 }
1506