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                         kref_get(&req->wb_kref);
 121         }
 122         return req;
 123 }
 124
 125 static struct nfs_page *nfs_page_find_request(struct page *page)
 126 {
 127         struct inode *inode = page->mapping->host;
 128         struct nfs_page *req = NULL;
 129
 130         spin_lock(&inode->i_lock);
 131         req = nfs_page_find_request_locked(page);
 132         spin_unlock(&inode->i_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         nfs_unlock_request(req);
 195         return 0;
 196 }
 197
 198 static int wb_priority(struct writeback_control *wbc)
 199 {
 200         if (wbc->for_reclaim)
 201                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 202         if (wbc->for_kupdate)
 203                 return FLUSH_LOWPRI;
 204         return 0;
 205 }
 206
 207 /*
 208  * NFS congestion control
 209  */
 210
 211 int nfs_congestion_kb;
 212
 213 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 214 #define NFS_CONGESTION_OFF_THRESH       \
 215         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 216
 217 static int nfs_set_page_writeback(struct page *page)
 218 {
 219         int ret = test_set_page_writeback(page);
 220
 221         if (!ret) {
 222                 struct inode *inode = page->mapping->host;
 223                 struct nfs_server *nfss = NFS_SERVER(inode);
 224
 225                 if (atomic_long_inc_return(&nfss->writeback) >
 226                                 NFS_CONGESTION_ON_THRESH)
 227                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 228         }
 229         return ret;
 230 }
 231
 232 static void nfs_end_page_writeback(struct page *page)
 233 {
 234         struct inode *inode = page->mapping->host;
 235         struct nfs_server *nfss = NFS_SERVER(inode);
 236
 237         end_page_writeback(page);
 238         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 239                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 240                 congestion_end(WRITE);
 241         }
 242 }
 243
 244 /*
 245  * Find an associated nfs write request, and prepare to flush it out
 246  * Returns 1 if there was no write request, or if the request was
 247  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 248  * was not tagged.
 249  * May also return an error if the user signalled nfs_wait_on_request().
 250  */
 251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 252                                 struct page *page)
 253 {
 254         struct inode *inode = page->mapping->host;
 255         struct nfs_inode *nfsi = NFS_I(inode);
 256         struct nfs_page *req;
 257         int ret;
 258
 259         spin_lock(&inode->i_lock);
 260         for(;;) {
 261                 req = nfs_page_find_request_locked(page);
 262                 if (req == NULL) {
 263                         spin_unlock(&inode->i_lock);
 264                         return 1;
 265                 }
 266                 if (nfs_lock_request_dontget(req))
 267                         break;
 268                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 269                  *       then the call to nfs_lock_request_dontget() will always
 270                  *       succeed provided that someone hasn't already marked the
 271                  *       request as dirty (in which case we don't care).
 272                  */
 273                 spin_unlock(&inode->i_lock);
 274                 ret = nfs_wait_on_request(req);
 275                 nfs_release_request(req);
 276                 if (ret != 0)
 277                         return ret;
 278                 spin_lock(&inode->i_lock);
 279         }
 280         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 281                 /* This request is marked for commit */
 282                 spin_unlock(&inode->i_lock);
 283                 nfs_unlock_request(req);
 284                 nfs_pageio_complete(pgio);
 285                 return 1;
 286         }
 287         if (nfs_set_page_writeback(page) != 0) {
 288                 spin_unlock(&inode->i_lock);
 289                 BUG();
 290         }
 291         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 292                         NFS_PAGE_TAG_LOCKED);
 293         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 294         spin_unlock(&inode->i_lock);
 295         nfs_pageio_add_request(pgio, req);
 296         return ret;
 297 }
 298
 299 /*
 300  * Write an mmapped page to the server.
 301  */
 302 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 303 {
 304         struct nfs_pageio_descriptor mypgio, *pgio;
 305         struct nfs_open_context *ctx;
 306         struct inode *inode = page->mapping->host;
 307         unsigned offset;
 308         int err;
 309
 310         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 311         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 312
 313         if (wbc->for_writepages)
 314                 pgio = wbc->fs_private;
 315         else {
 316                 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
 317                 pgio = &mypgio;
 318         }
 319
 320         nfs_pageio_cond_complete(pgio, page->index);
 321
 322         err = nfs_page_async_flush(pgio, page);
 323         if (err <= 0)
 324                 goto out;
 325         err = 0;
 326         offset = nfs_page_length(page);
 327         if (!offset)
 328                 goto out;
 329
 330         nfs_pageio_cond_complete(pgio, page->index);
 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         kref_get(&req->wb_kref);
 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->path.dentry->d_inode;
 411         struct nfs_inode *nfsi = NFS_I(inode);
 412
 413         BUG_ON (!NFS_WBACK_BUSY(req));
 414
 415         spin_lock(&inode->i_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(&inode->i_lock);
 424                 nfs_end_data_update(inode);
 425                 iput(inode);
 426         } else
 427                 spin_unlock(&inode->i_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->path.dentry->d_inode;
 459         struct nfs_inode *nfsi = NFS_I(inode);
 460
 461         spin_lock(&inode->i_lock);
 462         nfsi->ncommit++;
 463         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 464         radix_tree_tag_set(&nfsi->nfs_page_tree,
 465                         req->wb_index,
 466                         NFS_PAGE_TAG_COMMIT);
 467         spin_unlock(&inode->i_lock);
 468         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 469         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 470 }
 471
 472 static inline
 473 int nfs_write_need_commit(struct nfs_write_data *data)
 474 {
 475         return data->verf.committed != NFS_FILE_SYNC;
 476 }
 477
 478 static inline
 479 int nfs_reschedule_unstable_write(struct nfs_page *req)
 480 {
 481         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 482                 nfs_mark_request_commit(req);
 483                 return 1;
 484         }
 485         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 486                 nfs_redirty_request(req);
 487                 return 1;
 488         }
 489         return 0;
 490 }
 491 #else
 492 static inline void
 493 nfs_mark_request_commit(struct nfs_page *req)
 494 {
 495 }
 496
 497 static inline
 498 int nfs_write_need_commit(struct nfs_write_data *data)
 499 {
 500         return 0;
 501 }
 502
 503 static inline
 504 int nfs_reschedule_unstable_write(struct nfs_page *req)
 505 {
 506         return 0;
 507 }
 508 #endif
 509
 510 /*
 511  * Wait for a request to complete.
 512  *
 513  * Interruptible by signals only if mounted with intr flag.
 514  */
 515 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 516 {
 517         struct nfs_inode *nfsi = NFS_I(inode);
 518         struct nfs_page *req;
 519         pgoff_t idx_end, next;
 520         unsigned int            res = 0;
 521         int                     error;
 522
 523         if (npages == 0)
 524                 idx_end = ~0;
 525         else
 526                 idx_end = idx_start + npages - 1;
 527
 528         next = idx_start;
 529         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 530                 if (req->wb_index > idx_end)
 531                         break;
 532
 533                 next = req->wb_index + 1;
 534                 BUG_ON(!NFS_WBACK_BUSY(req));
 535
 536                 kref_get(&req->wb_kref);
 537                 spin_unlock(&inode->i_lock);
 538                 error = nfs_wait_on_request(req);
 539                 nfs_release_request(req);
 540                 spin_lock(&inode->i_lock);
 541                 if (error < 0)
 542                         return error;
 543                 res++;
 544         }
 545         return res;
 546 }
 547
 548 static void nfs_cancel_commit_list(struct list_head *head)
 549 {
 550         struct nfs_page *req;
 551
 552         while(!list_empty(head)) {
 553                 req = nfs_list_entry(head->next);
 554                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 555                 nfs_list_remove_request(req);
 556                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 557                 nfs_inode_remove_request(req);
 558                 nfs_unlock_request(req);
 559         }
 560 }
 561
 562 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 563 /*
 564  * nfs_scan_commit - Scan an inode for commit requests
 565  * @inode: NFS inode to scan
 566  * @dst: destination list
 567  * @idx_start: lower bound of page->index to scan.
 568  * @npages: idx_start + npages sets the upper bound to scan.
 569  *
 570  * Moves requests from the inode's 'commit' request list.
 571  * The requests are *not* checked to ensure that they form a contiguous set.
 572  */
 573 static int
 574 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 575 {
 576         struct nfs_inode *nfsi = NFS_I(inode);
 577         int res = 0;
 578
 579         if (nfsi->ncommit != 0) {
 580                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 581                                 NFS_PAGE_TAG_COMMIT);
 582                 nfsi->ncommit -= res;
 583         }
 584         return res;
 585 }
 586 #else
 587 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 588 {
 589         return 0;
 590 }
 591 #endif
 592
 593 /*
 594  * Try to update any existing write request, or create one if there is none.
 595  * In order to match, the request's credentials must match those of
 596  * the calling process.
 597  *
 598  * Note: Should always be called with the Page Lock held!
 599  */
 600 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 601                 struct page *page, unsigned int offset, unsigned int bytes)
 602 {
 603         struct address_space *mapping = page->mapping;
 604         struct inode *inode = mapping->host;
 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(&inode->i_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(&inode->i_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(&inode->i_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(&inode->i_lock);
 642                                 nfs_unlock_request(new);
 643                                 return ERR_PTR(error);
 644                         }
 645                         spin_unlock(&inode->i_lock);
 646                         return new;
 647                 }
 648                 spin_unlock(&inode->i_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)) {
 753                 nfs_end_page_writeback(req->wb_page);
 754                 nfs_inode_remove_request(req);
 755         } else if (!nfs_reschedule_unstable_write(req)) {
 756                 /* Set the PG_uptodate flag */
 757                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 758                 nfs_end_page_writeback(req->wb_page);
 759                 nfs_inode_remove_request(req);
 760         } else
 761                 nfs_end_page_writeback(req->wb_page);
 762         nfs_clear_page_tag_locked(req);
 763 }
 764
 765 static inline int flush_task_priority(int how)
 766 {
 767         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 768                 case FLUSH_HIGHPRI:
 769                         return RPC_PRIORITY_HIGH;
 770                 case FLUSH_LOWPRI:
 771                         return RPC_PRIORITY_LOW;
 772         }
 773         return RPC_PRIORITY_NORMAL;
 774 }
 775
 776 /*
 777  * Set up the argument/result storage required for the RPC call.
 778  */
 779 static void nfs_write_rpcsetup(struct nfs_page *req,
 780                 struct nfs_write_data *data,
 781                 const struct rpc_call_ops *call_ops,
 782                 unsigned int count, unsigned int offset,
 783                 int how)
 784 {
 785         struct inode            *inode;
 786         int flags;
 787
 788         /* Set up the RPC argument and reply structs
 789          * NB: take care not to mess about with data->commit et al. */
 790
 791         data->req = req;
 792         data->inode = inode = req->wb_context->path.dentry->d_inode;
 793         data->cred = req->wb_context->cred;
 794
 795         data->args.fh     = NFS_FH(inode);
 796         data->args.offset = req_offset(req) + offset;
 797         data->args.pgbase = req->wb_pgbase + offset;
 798         data->args.pages  = data->pagevec;
 799         data->args.count  = count;
 800         data->args.context = req->wb_context;
 801
 802         data->res.fattr   = &data->fattr;
 803         data->res.count   = count;
 804         data->res.verf    = &data->verf;
 805         nfs_fattr_init(&data->fattr);
 806
 807         /* Set up the initial task struct.  */
 808         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 809         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 810         NFS_PROTO(inode)->write_setup(data, how);
 811
 812         data->task.tk_priority = flush_task_priority(how);
 813         data->task.tk_cookie = (unsigned long)inode;
 814
 815         dprintk("NFS: %5u initiated write call "
 816                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 817                 data->task.tk_pid,
 818                 inode->i_sb->s_id,
 819                 (long long)NFS_FILEID(inode),
 820                 count,
 821                 (unsigned long long)data->args.offset);
 822 }
 823
 824 static void nfs_execute_write(struct nfs_write_data *data)
 825 {
 826         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 827         sigset_t oldset;
 828
 829         rpc_clnt_sigmask(clnt, &oldset);
 830         rpc_execute(&data->task);
 831         rpc_clnt_sigunmask(clnt, &oldset);
 832 }
 833
 834 /*
 835  * Generate multiple small requests to write out a single
 836  * contiguous dirty area on one page.
 837  */
 838 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 839 {
 840         struct nfs_page *req = nfs_list_entry(head->next);
 841         struct page *page = req->wb_page;
 842         struct nfs_write_data *data;
 843         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 844         unsigned int offset;
 845         int requests = 0;
 846         LIST_HEAD(list);
 847
 848         nfs_list_remove_request(req);
 849
 850         nbytes = count;
 851         do {
 852                 size_t len = min(nbytes, wsize);
 853
 854                 data = nfs_writedata_alloc(1);
 855                 if (!data)
 856                         goto out_bad;
 857                 list_add(&data->pages, &list);
 858                 requests++;
 859                 nbytes -= len;
 860         } while (nbytes != 0);
 861         atomic_set(&req->wb_complete, requests);
 862
 863         ClearPageError(page);
 864         offset = 0;
 865         nbytes = count;
 866         do {
 867                 data = list_entry(list.next, struct nfs_write_data, pages);
 868                 list_del_init(&data->pages);
 869
 870                 data->pagevec[0] = page;
 871
 872                 if (nbytes < wsize)
 873                         wsize = nbytes;
 874                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 875                                    wsize, offset, how);
 876                 offset += wsize;
 877                 nbytes -= wsize;
 878                 nfs_execute_write(data);
 879         } while (nbytes != 0);
 880
 881         return 0;
 882
 883 out_bad:
 884         while (!list_empty(&list)) {
 885                 data = list_entry(list.next, struct nfs_write_data, pages);
 886                 list_del(&data->pages);
 887                 nfs_writedata_release(data);
 888         }
 889         nfs_redirty_request(req);
 890         nfs_end_page_writeback(req->wb_page);
 891         nfs_clear_page_tag_locked(req);
 892         return -ENOMEM;
 893 }
 894
 895 /*
 896  * Create an RPC task for the given write request and kick it.
 897  * The page must have been locked by the caller.
 898  *
 899  * It may happen that the page we're passed is not marked dirty.
 900  * This is the case if nfs_updatepage detects a conflicting request
 901  * that has been written but not committed.
 902  */
 903 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 904 {
 905         struct nfs_page         *req;
 906         struct page             **pages;
 907         struct nfs_write_data   *data;
 908
 909         data = nfs_writedata_alloc(npages);
 910         if (!data)
 911                 goto out_bad;
 912
 913         pages = data->pagevec;
 914         while (!list_empty(head)) {
 915                 req = nfs_list_entry(head->next);
 916                 nfs_list_remove_request(req);
 917                 nfs_list_add_request(req, &data->pages);
 918                 ClearPageError(req->wb_page);
 919                 *pages++ = req->wb_page;
 920         }
 921         req = nfs_list_entry(data->pages.next);
 922
 923         /* Set up the argument struct */
 924         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 925
 926         nfs_execute_write(data);
 927         return 0;
 928  out_bad:
 929         while (!list_empty(head)) {
 930                 req = nfs_list_entry(head->next);
 931                 nfs_list_remove_request(req);
 932                 nfs_redirty_request(req);
 933                 nfs_end_page_writeback(req->wb_page);
 934                 nfs_clear_page_tag_locked(req);
 935         }
 936         return -ENOMEM;
 937 }
 938
 939 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 940                                   struct inode *inode, int ioflags)
 941 {
 942         int wsize = NFS_SERVER(inode)->wsize;
 943
 944         if (wsize < PAGE_CACHE_SIZE)
 945                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 946         else
 947                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 948 }
 949
 950 /*
 951  * Handle a write reply that flushed part of a page.
 952  */
 953 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 954 {
 955         struct nfs_write_data   *data = calldata;
 956         struct nfs_page         *req = data->req;
 957         struct page             *page = req->wb_page;
 958
 959         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 960                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 961                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 962                 req->wb_bytes,
 963                 (long long)req_offset(req));
 964
 965         if (nfs_writeback_done(task, data) != 0)
 966                 return;
 967
 968         if (task->tk_status < 0) {
 969                 nfs_set_pageerror(page);
 970                 req->wb_context->error = task->tk_status;
 971                 dprintk(", error = %d\n", task->tk_status);
 972                 goto out;
 973         }
 974
 975         if (nfs_write_need_commit(data)) {
 976                 struct inode *inode = page->mapping->host;
 977
 978                 spin_lock(&inode->i_lock);
 979                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 980                         /* Do nothing we need to resend the writes */
 981                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 982                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 983                         dprintk(" defer commit\n");
 984                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 985                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 986                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 987                         dprintk(" server reboot detected\n");
 988                 }
 989                 spin_unlock(&inode->i_lock);
 990         } else
 991                 dprintk(" OK\n");
 992
 993 out:
 994         if (atomic_dec_and_test(&req->wb_complete))
 995                 nfs_writepage_release(req);
 996 }
 997
 998 static const struct rpc_call_ops nfs_write_partial_ops = {
 999         .rpc_call_done = nfs_writeback_done_partial,
1000         .rpc_release = nfs_writedata_release,
1001 };
1002
1003 /*
1004  * Handle a write reply that flushes a whole page.
1005  *
1006  * FIXME: There is an inherent race with invalidate_inode_pages and
1007  *        writebacks since the page->count is kept > 1 for as long
1008  *        as the page has a write request pending.
1009  */
1010 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1011 {
1012         struct nfs_write_data   *data = calldata;
1013         struct nfs_page         *req;
1014         struct page             *page;
1015
1016         if (nfs_writeback_done(task, data) != 0)
1017                 return;
1018
1019         /* Update attributes as result of writeback. */
1020         while (!list_empty(&data->pages)) {
1021                 req = nfs_list_entry(data->pages.next);
1022                 nfs_list_remove_request(req);
1023                 page = req->wb_page;
1024
1025                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1026                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1027                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1028                         req->wb_bytes,
1029                         (long long)req_offset(req));
1030
1031                 if (task->tk_status < 0) {
1032                         nfs_set_pageerror(page);
1033                         req->wb_context->error = task->tk_status;
1034                         dprintk(", error = %d\n", task->tk_status);
1035                         goto remove_request;
1036                 }
1037
1038                 if (nfs_write_need_commit(data)) {
1039                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1040                         nfs_mark_request_commit(req);
1041                         nfs_end_page_writeback(page);
1042                         dprintk(" marked for commit\n");
1043                         goto next;
1044                 }
1045                 /* Set the PG_uptodate flag? */
1046                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1047                 dprintk(" OK\n");
1048 remove_request:
1049                 nfs_end_page_writeback(page);
1050                 nfs_inode_remove_request(req);
1051         next:
1052                 nfs_clear_page_tag_locked(req);
1053         }
1054 }
1055
1056 static const struct rpc_call_ops nfs_write_full_ops = {
1057         .rpc_call_done = nfs_writeback_done_full,
1058         .rpc_release = nfs_writedata_release,
1059 };
1060
1061
1062 /*
1063  * This function is called when the WRITE call is complete.
1064  */
1065 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1066 {
1067         struct nfs_writeargs    *argp = &data->args;
1068         struct nfs_writeres     *resp = &data->res;
1069         int status;
1070
1071         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1072                 task->tk_pid, task->tk_status);
1073
1074         /*
1075          * ->write_done will attempt to use post-op attributes to detect
1076          * conflicting writes by other clients.  A strict interpretation
1077          * of close-to-open would allow us to continue caching even if
1078          * another writer had changed the file, but some applications
1079          * depend on tighter cache coherency when writing.
1080          */
1081         status = NFS_PROTO(data->inode)->write_done(task, data);
1082         if (status != 0)
1083                 return status;
1084         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1085
1086 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1087         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1088                 /* We tried a write call, but the server did not
1089                  * commit data to stable storage even though we
1090                  * requested it.
1091                  * Note: There is a known bug in Tru64 < 5.0 in which
1092                  *       the server reports NFS_DATA_SYNC, but performs
1093                  *       NFS_FILE_SYNC. We therefore implement this checking
1094                  *       as a dprintk() in order to avoid filling syslog.
1095                  */
1096                 static unsigned long    complain;
1097
1098                 if (time_before(complain, jiffies)) {
1099                         dprintk("NFS: faulty NFS server %s:"
1100                                 " (committed = %d) != (stable = %d)\n",
1101                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1102                                 resp->verf->committed, argp->stable);
1103                         complain = jiffies + 300 * HZ;
1104                 }
1105         }
1106 #endif
1107         /* Is this a short write? */
1108         if (task->tk_status >= 0 && resp->count < argp->count) {
1109                 static unsigned long    complain;
1110
1111                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1112
1113                 /* Has the server at least made some progress? */
1114                 if (resp->count != 0) {
1115                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1116                         if (resp->verf->committed != NFS_UNSTABLE) {
1117                                 /* Resend from where the server left off */
1118                                 argp->offset += resp->count;
1119                                 argp->pgbase += resp->count;
1120                                 argp->count -= resp->count;
1121                         } else {
1122                                 /* Resend as a stable write in order to avoid
1123                                  * headaches in the case of a server crash.
1124                                  */
1125                                 argp->stable = NFS_FILE_SYNC;
1126                         }
1127                         rpc_restart_call(task);
1128                         return -EAGAIN;
1129                 }
1130                 if (time_before(complain, jiffies)) {
1131                         printk(KERN_WARNING
1132                                "NFS: Server wrote zero bytes, expected %u.\n",
1133                                         argp->count);
1134                         complain = jiffies + 300 * HZ;
1135                 }
1136                 /* Can't do anything about it except throw an error. */
1137                 task->tk_status = -EIO;
1138         }
1139         return 0;
1140 }
1141
1142
1143 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1144 void nfs_commit_release(void *wdata)
1145 {
1146         nfs_commit_free(wdata);
1147 }
1148
1149 /*
1150  * Set up the argument/result storage required for the RPC call.
1151  */
1152 static void nfs_commit_rpcsetup(struct list_head *head,
1153                 struct nfs_write_data *data,
1154                 int how)
1155 {
1156         struct nfs_page         *first;
1157         struct inode            *inode;
1158         int flags;
1159
1160         /* Set up the RPC argument and reply structs
1161          * NB: take care not to mess about with data->commit et al. */
1162
1163         list_splice_init(head, &data->pages);
1164         first = nfs_list_entry(data->pages.next);
1165         inode = first->wb_context->path.dentry->d_inode;
1166
1167         data->inode       = inode;
1168         data->cred        = first->wb_context->cred;
1169
1170         data->args.fh     = NFS_FH(data->inode);
1171         /* Note: we always request a commit of the entire inode */
1172         data->args.offset = 0;
1173         data->args.count  = 0;
1174         data->res.count   = 0;
1175         data->res.fattr   = &data->fattr;
1176         data->res.verf    = &data->verf;
1177         nfs_fattr_init(&data->fattr);
1178
1179         /* Set up the initial task struct.  */
1180         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1181         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1182         NFS_PROTO(inode)->commit_setup(data, how);
1183
1184         data->task.tk_priority = flush_task_priority(how);
1185         data->task.tk_cookie = (unsigned long)inode;
1186
1187         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1188 }
1189
1190 /*
1191  * Commit dirty pages
1192  */
1193 static int
1194 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1195 {
1196         struct nfs_write_data   *data;
1197         struct nfs_page         *req;
1198
1199         data = nfs_commit_alloc();
1200
1201         if (!data)
1202                 goto out_bad;
1203
1204         /* Set up the argument struct */
1205         nfs_commit_rpcsetup(head, data, how);
1206
1207         nfs_execute_write(data);
1208         return 0;
1209  out_bad:
1210         while (!list_empty(head)) {
1211                 req = nfs_list_entry(head->next);
1212                 nfs_list_remove_request(req);
1213                 nfs_mark_request_commit(req);
1214                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1215                 nfs_clear_page_tag_locked(req);
1216         }
1217         return -ENOMEM;
1218 }
1219
1220 /*
1221  * COMMIT call returned
1222  */
1223 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1224 {
1225         struct nfs_write_data   *data = calldata;
1226         struct nfs_page         *req;
1227
1228         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1229                                 task->tk_pid, task->tk_status);
1230
1231         /* Call the NFS version-specific code */
1232         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1233                 return;
1234
1235         while (!list_empty(&data->pages)) {
1236                 req = nfs_list_entry(data->pages.next);
1237                 nfs_list_remove_request(req);
1238                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1239                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1240
1241                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1242                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1243                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1244                         req->wb_bytes,
1245                         (long long)req_offset(req));
1246                 if (task->tk_status < 0) {
1247                         req->wb_context->error = task->tk_status;
1248                         nfs_inode_remove_request(req);
1249                         dprintk(", error = %d\n", task->tk_status);
1250                         goto next;
1251                 }
1252
1253                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1254                  * returned by the server against all stored verfs. */
1255                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1256                         /* We have a match */
1257                         /* Set the PG_uptodate flag */
1258                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1259                                         req->wb_bytes);
1260                         nfs_inode_remove_request(req);
1261                         dprintk(" OK\n");
1262                         goto next;
1263                 }
1264                 /* We have a mismatch. Write the page again */
1265                 dprintk(" mismatch\n");
1266                 nfs_redirty_request(req);
1267         next:
1268                 nfs_clear_page_tag_locked(req);
1269         }
1270 }
1271
1272 static const struct rpc_call_ops nfs_commit_ops = {
1273         .rpc_call_done = nfs_commit_done,
1274         .rpc_release = nfs_commit_release,
1275 };
1276
1277 int nfs_commit_inode(struct inode *inode, int how)
1278 {
1279         LIST_HEAD(head);
1280         int res;
1281
1282         spin_lock(&inode->i_lock);
1283         res = nfs_scan_commit(inode, &head, 0, 0);
1284         spin_unlock(&inode->i_lock);
1285         if (res) {
1286                 int error = nfs_commit_list(inode, &head, how);
1287                 if (error < 0)
1288                         return error;
1289         }
1290         return res;
1291 }
1292 #else
1293 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1294 {
1295         return 0;
1296 }
1297 #endif
1298
1299 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1300 {
1301         struct inode *inode = mapping->host;
1302         pgoff_t idx_start, idx_end;
1303         unsigned int npages = 0;
1304         LIST_HEAD(head);
1305         int nocommit = how & FLUSH_NOCOMMIT;
1306         long pages, ret;
1307
1308         /* FIXME */
1309         if (wbc->range_cyclic)
1310                 idx_start = 0;
1311         else {
1312                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1313                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1314                 if (idx_end > idx_start) {
1315                         pgoff_t l_npages = 1 + idx_end - idx_start;
1316                         npages = l_npages;
1317                         if (sizeof(npages) != sizeof(l_npages) &&
1318                                         (pgoff_t)npages != l_npages)
1319                                 npages = 0;
1320                 }
1321         }
1322         how &= ~FLUSH_NOCOMMIT;
1323         spin_lock(&inode->i_lock);
1324         do {
1325                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1326                 if (ret != 0)
1327                         continue;
1328                 if (nocommit)
1329                         break;
1330                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1331                 if (pages == 0)
1332                         break;
1333                 if (how & FLUSH_INVALIDATE) {
1334                         spin_unlock(&inode->i_lock);
1335                         nfs_cancel_commit_list(&head);
1336                         ret = pages;
1337                         spin_lock(&inode->i_lock);
1338                         continue;
1339                 }
1340                 pages += nfs_scan_commit(inode, &head, 0, 0);
1341                 spin_unlock(&inode->i_lock);
1342                 ret = nfs_commit_list(inode, &head, how);
1343                 spin_lock(&inode->i_lock);
1344
1345         } while (ret >= 0);
1346         spin_unlock(&inode->i_lock);
1347         return ret;
1348 }
1349
1350 /*
1351  * flush the inode to disk.
1352  */
1353 int nfs_wb_all(struct inode *inode)
1354 {
1355         struct address_space *mapping = inode->i_mapping;
1356         struct writeback_control wbc = {
1357                 .bdi = mapping->backing_dev_info,
1358                 .sync_mode = WB_SYNC_ALL,
1359                 .nr_to_write = LONG_MAX,
1360                 .for_writepages = 1,
1361                 .range_cyclic = 1,
1362         };
1363         int ret;
1364
1365         ret = nfs_writepages(mapping, &wbc);
1366         if (ret < 0)
1367                 goto out;
1368         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1369         if (ret >= 0)
1370                 return 0;
1371 out:
1372         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1373         return ret;
1374 }
1375
1376 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1377 {
1378         struct writeback_control wbc = {
1379                 .bdi = mapping->backing_dev_info,
1380                 .sync_mode = WB_SYNC_ALL,
1381                 .nr_to_write = LONG_MAX,
1382                 .range_start = range_start,
1383                 .range_end = range_end,
1384                 .for_writepages = 1,
1385         };
1386         int ret;
1387
1388         ret = nfs_writepages(mapping, &wbc);
1389         if (ret < 0)
1390                 goto out;
1391         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1392         if (ret >= 0)
1393                 return 0;
1394 out:
1395         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1396         return ret;
1397 }
1398
1399 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1400 {
1401         struct nfs_page *req;
1402         loff_t range_start = page_offset(page);
1403         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1404         struct writeback_control wbc = {
1405                 .bdi = page->mapping->backing_dev_info,
1406                 .sync_mode = WB_SYNC_ALL,
1407                 .nr_to_write = LONG_MAX,
1408                 .range_start = range_start,
1409                 .range_end = range_end,
1410         };
1411         int ret = 0;
1412
1413         BUG_ON(!PageLocked(page));
1414         for (;;) {
1415                 req = nfs_page_find_request(page);
1416                 if (req == NULL)
1417                         goto out;
1418                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1419                         nfs_release_request(req);
1420                         break;
1421                 }
1422                 if (nfs_lock_request_dontget(req)) {
1423                         nfs_inode_remove_request(req);
1424                         /*
1425                          * In case nfs_inode_remove_request has marked the
1426                          * page as being dirty
1427                          */
1428                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1429                         nfs_unlock_request(req);
1430                         break;
1431                 }
1432                 ret = nfs_wait_on_request(req);
1433                 if (ret < 0)
1434                         goto out;
1435         }
1436         if (!PagePrivate(page))
1437                 return 0;
1438         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1439 out:
1440         return ret;
1441 }
1442
1443 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1444 {
1445         loff_t range_start = page_offset(page);
1446         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1447         struct writeback_control wbc = {
1448                 .bdi = page->mapping->backing_dev_info,
1449                 .sync_mode = WB_SYNC_ALL,
1450                 .nr_to_write = LONG_MAX,
1451                 .range_start = range_start,
1452                 .range_end = range_end,
1453         };
1454         int ret;
1455
1456         BUG_ON(!PageLocked(page));
1457         if (clear_page_dirty_for_io(page)) {
1458                 ret = nfs_writepage_locked(page, &wbc);
1459                 if (ret < 0)
1460                         goto out;
1461         }
1462         if (!PagePrivate(page))
1463                 return 0;
1464         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1465         if (ret >= 0)
1466                 return 0;
1467 out:
1468         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1469         return ret;
1470 }
1471
1472 /*
1473  * Write back all requests on one page - we do this before reading it.
1474  */
1475 int nfs_wb_page(struct inode *inode, struct page* page)
1476 {
1477         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1478 }
1479
1480 int nfs_set_page_dirty(struct page *page)
1481 {
1482         struct address_space *mapping = page->mapping;
1483         struct inode *inode;
1484         struct nfs_page *req;
1485         int ret;
1486
1487         if (!mapping)
1488                 goto out_raced;
1489         inode = mapping->host;
1490         if (!inode)
1491                 goto out_raced;
1492         spin_lock(&inode->i_lock);
1493         req = nfs_page_find_request_locked(page);
1494         if (req != NULL) {
1495                 /* Mark any existing write requests for flushing */
1496                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1497                 spin_unlock(&inode->i_lock);
1498                 nfs_release_request(req);
1499                 return ret;
1500         }
1501         ret = __set_page_dirty_nobuffers(page);
1502         spin_unlock(&inode->i_lock);
1503         return ret;
1504 out_raced:
1505         return !TestSetPageDirty(page);
1506 }
1507
1508
1509 int __init nfs_init_writepagecache(void)
1510 {
1511         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1512                                              sizeof(struct nfs_write_data),
1513                                              0, SLAB_HWCACHE_ALIGN,
1514                                              NULL);
1515         if (nfs_wdata_cachep == NULL)
1516                 return -ENOMEM;
1517
1518         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1519                                                      nfs_wdata_cachep);
1520         if (nfs_wdata_mempool == NULL)
1521                 return -ENOMEM;
1522
1523         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1524                                                       nfs_wdata_cachep);
1525         if (nfs_commit_mempool == NULL)
1526                 return -ENOMEM;
1527
1528         /*
1529          * NFS congestion size, scale with available memory.
1530          *
1531          *  64MB:    8192k
1532          * 128MB:   11585k
1533          * 256MB:   16384k
1534          * 512MB:   23170k
1535          *   1GB:   32768k
1536          *   2GB:   46340k
1537          *   4GB:   65536k
1538          *   8GB:   92681k
1539          *  16GB:  131072k
1540          *
1541          * This allows larger machines to have larger/more transfers.
1542          * Limit the default to 256M
1543          */
1544         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1545         if (nfs_congestion_kb > 256*1024)
1546                 nfs_congestion_kb = 256*1024;
1547
1548         return 0;
1549 }
1550
1551 void nfs_destroy_writepagecache(void)
1552 {
1553         mempool_destroy(nfs_commit_mempool);
1554         mempool_destroy(nfs_wdata_mempool);
1555         kmem_cache_destroy(nfs_wdata_cachep);
1556 }
1557