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                 ret = nfs_wait_on_request(req);
 277                 nfs_release_request(req);
 278                 if (ret != 0)
 279                         return ret;
 280                 spin_lock(req_lock);
 281         }
 282         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 283                 /* This request is marked for commit */
 284                 spin_unlock(req_lock);
 285                 nfs_unlock_request(req);
 286                 nfs_pageio_complete(pgio);
 287                 return 1;
 288         }
 289         if (nfs_set_page_writeback(page) != 0) {
 290                 spin_unlock(req_lock);
 291                 BUG();
 292         }
 293         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 294                         NFS_PAGE_TAG_WRITEBACK);
 295         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 296         spin_unlock(req_lock);
 297         nfs_pageio_add_request(pgio, req);
 298         return ret;
 299 }
 300
 301 /*
 302  * Write an mmapped page to the server.
 303  */
 304 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 305 {
 306         struct nfs_pageio_descriptor mypgio, *pgio;
 307         struct nfs_open_context *ctx;
 308         struct inode *inode = page->mapping->host;
 309         unsigned offset;
 310         int err;
 311
 312         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 313         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 314
 315         if (wbc->for_writepages)
 316                 pgio = wbc->fs_private;
 317         else {
 318                 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
 319                 pgio = &mypgio;
 320         }
 321
 322         nfs_pageio_cond_complete(pgio, page->index);
 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         nfs_pageio_cond_complete(pgio, page->index);
 333
 334         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 335         if (ctx == NULL) {
 336                 err = -EBADF;
 337                 goto out;
 338         }
 339         err = nfs_writepage_setup(ctx, page, 0, offset);
 340         put_nfs_open_context(ctx);
 341         if (err != 0)
 342                 goto out;
 343         err = nfs_page_async_flush(pgio, page);
 344         if (err > 0)
 345                 err = 0;
 346 out:
 347         if (!wbc->for_writepages)
 348                 nfs_pageio_complete(pgio);
 349         return err;
 350 }
 351
 352 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 353 {
 354         int err;
 355
 356         err = nfs_writepage_locked(page, wbc);
 357         unlock_page(page);
 358         return err;
 359 }
 360
 361 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 362 {
 363         struct inode *inode = mapping->host;
 364         struct nfs_pageio_descriptor pgio;
 365         int err;
 366
 367         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 368
 369         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 370         wbc->fs_private = &pgio;
 371         err = generic_writepages(mapping, wbc);
 372         nfs_pageio_complete(&pgio);
 373         if (err)
 374                 return err;
 375         if (pgio.pg_error)
 376                 return pgio.pg_error;
 377         return 0;
 378 }
 379
 380 /*
 381  * Insert a write request into an inode
 382  */
 383 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 384 {
 385         struct nfs_inode *nfsi = NFS_I(inode);
 386         int error;
 387
 388         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 389         BUG_ON(error == -EEXIST);
 390         if (error)
 391                 return error;
 392         if (!nfsi->npages) {
 393                 igrab(inode);
 394                 nfs_begin_data_update(inode);
 395                 if (nfs_have_delegation(inode, FMODE_WRITE))
 396                         nfsi->change_attr++;
 397         }
 398         SetPagePrivate(req->wb_page);
 399         set_page_private(req->wb_page, (unsigned long)req);
 400         if (PageDirty(req->wb_page))
 401                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
 402         nfsi->npages++;
 403         atomic_inc(&req->wb_count);
 404         return 0;
 405 }
 406
 407 /*
 408  * Remove a write request from an inode
 409  */
 410 static void nfs_inode_remove_request(struct nfs_page *req)
 411 {
 412         struct inode *inode = req->wb_context->dentry->d_inode;
 413         struct nfs_inode *nfsi = NFS_I(inode);
 414
 415         BUG_ON (!NFS_WBACK_BUSY(req));
 416
 417         spin_lock(&nfsi->req_lock);
 418         set_page_private(req->wb_page, 0);
 419         ClearPagePrivate(req->wb_page);
 420         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 421         if (test_and_clear_bit(PG_NEED_FLUSH, &req->wb_flags))
 422                 __set_page_dirty_nobuffers(req->wb_page);
 423         nfsi->npages--;
 424         if (!nfsi->npages) {
 425                 spin_unlock(&nfsi->req_lock);
 426                 nfs_end_data_update(inode);
 427                 iput(inode);
 428         } else
 429                 spin_unlock(&nfsi->req_lock);
 430         nfs_clear_request(req);
 431         nfs_release_request(req);
 432 }
 433
 434 static void
 435 nfs_redirty_request(struct nfs_page *req)
 436 {
 437         __set_page_dirty_nobuffers(req->wb_page);
 438 }
 439
 440 /*
 441  * Check if a request is dirty
 442  */
 443 static inline int
 444 nfs_dirty_request(struct nfs_page *req)
 445 {
 446         struct page *page = req->wb_page;
 447
 448         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 449                 return 0;
 450         return !PageWriteback(req->wb_page);
 451 }
 452
 453 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 454 /*
 455  * Add a request to the inode's commit list.
 456  */
 457 static void
 458 nfs_mark_request_commit(struct nfs_page *req)
 459 {
 460         struct inode *inode = req->wb_context->dentry->d_inode;
 461         struct nfs_inode *nfsi = NFS_I(inode);
 462
 463         spin_lock(&nfsi->req_lock);
 464         nfs_list_add_request(req, &nfsi->commit);
 465         nfsi->ncommit++;
 466         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 467         spin_unlock(&nfsi->req_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_WRITEBACK)) {
 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                 atomic_inc(&req->wb_count);
 537                 spin_unlock(&nfsi->req_lock);
 538                 error = nfs_wait_on_request(req);
 539                 nfs_release_request(req);
 540                 spin_lock(&nfsi->req_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, &nfsi->commit, dst, idx_start, npages);
 581                 nfsi->ncommit -= res;
 582                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 583                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 584         }
 585         return res;
 586 }
 587 #else
 588 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 589 {
 590         return 0;
 591 }
 592 #endif
 593
 594 /*
 595  * Try to update any existing write request, or create one if there is none.
 596  * In order to match, the request's credentials must match those of
 597  * the calling process.
 598  *
 599  * Note: Should always be called with the Page Lock held!
 600  */
 601 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 602                 struct page *page, unsigned int offset, unsigned int bytes)
 603 {
 604         struct address_space *mapping = page->mapping;
 605         struct inode *inode = mapping->host;
 606         struct nfs_inode *nfsi = NFS_I(inode);
 607         struct nfs_page         *req, *new = NULL;
 608         pgoff_t         rqend, end;
 609
 610         end = offset + bytes;
 611
 612         for (;;) {
 613                 /* Loop over all inode entries and see if we find
 614                  * A request for the page we wish to update
 615                  */
 616                 spin_lock(&nfsi->req_lock);
 617                 req = nfs_page_find_request_locked(page);
 618                 if (req) {
 619                         if (!nfs_lock_request_dontget(req)) {
 620                                 int error;
 621
 622                                 spin_unlock(&nfsi->req_lock);
 623                                 error = nfs_wait_on_request(req);
 624                                 nfs_release_request(req);
 625                                 if (error < 0) {
 626                                         if (new)
 627                                                 nfs_release_request(new);
 628                                         return ERR_PTR(error);
 629                                 }
 630                                 continue;
 631                         }
 632                         spin_unlock(&nfsi->req_lock);
 633                         if (new)
 634                                 nfs_release_request(new);
 635                         break;
 636                 }
 637
 638                 if (new) {
 639                         int error;
 640                         nfs_lock_request_dontget(new);
 641                         error = nfs_inode_add_request(inode, new);
 642                         if (error) {
 643                                 spin_unlock(&nfsi->req_lock);
 644                                 nfs_unlock_request(new);
 645                                 return ERR_PTR(error);
 646                         }
 647                         spin_unlock(&nfsi->req_lock);
 648                         return new;
 649                 }
 650                 spin_unlock(&nfsi->req_lock);
 651
 652                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 653                 if (IS_ERR(new))
 654                         return new;
 655         }
 656
 657         /* We have a request for our page.
 658          * If the creds don't match, or the
 659          * page addresses don't match,
 660          * tell the caller to wait on the conflicting
 661          * request.
 662          */
 663         rqend = req->wb_offset + req->wb_bytes;
 664         if (req->wb_context != ctx
 665             || req->wb_page != page
 666             || !nfs_dirty_request(req)
 667             || offset > rqend || end < req->wb_offset) {
 668                 nfs_unlock_request(req);
 669                 return ERR_PTR(-EBUSY);
 670         }
 671
 672         /* Okay, the request matches. Update the region */
 673         if (offset < req->wb_offset) {
 674                 req->wb_offset = offset;
 675                 req->wb_pgbase = offset;
 676                 req->wb_bytes = rqend - req->wb_offset;
 677         }
 678
 679         if (end > rqend)
 680                 req->wb_bytes = end - req->wb_offset;
 681
 682         return req;
 683 }
 684
 685 int nfs_flush_incompatible(struct file *file, struct page *page)
 686 {
 687         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 688         struct nfs_page *req;
 689         int do_flush, status;
 690         /*
 691          * Look for a request corresponding to this page. If there
 692          * is one, and it belongs to another file, we flush it out
 693          * before we try to copy anything into the page. Do this
 694          * due to the lack of an ACCESS-type call in NFSv2.
 695          * Also do the same if we find a request from an existing
 696          * dropped page.
 697          */
 698         do {
 699                 req = nfs_page_find_request(page);
 700                 if (req == NULL)
 701                         return 0;
 702                 do_flush = req->wb_page != page || req->wb_context != ctx
 703                         || !nfs_dirty_request(req);
 704                 nfs_release_request(req);
 705                 if (!do_flush)
 706                         return 0;
 707                 status = nfs_wb_page(page->mapping->host, page);
 708         } while (status == 0);
 709         return status;
 710 }
 711
 712 /*
 713  * Update and possibly write a cached page of an NFS file.
 714  *
 715  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 716  * things with a page scheduled for an RPC call (e.g. invalidate it).
 717  */
 718 int nfs_updatepage(struct file *file, struct page *page,
 719                 unsigned int offset, unsigned int count)
 720 {
 721         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 722         struct inode    *inode = page->mapping->host;
 723         int             status = 0;
 724
 725         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 726
 727         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 728                 file->f_path.dentry->d_parent->d_name.name,
 729                 file->f_path.dentry->d_name.name, count,
 730                 (long long)(page_offset(page) +offset));
 731
 732         /* If we're not using byte range locks, and we know the page
 733          * is entirely in cache, it may be more efficient to avoid
 734          * fragmenting write requests.
 735          */
 736         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 737                 count = max(count + offset, nfs_page_length(page));
 738                 offset = 0;
 739         }
 740
 741         status = nfs_writepage_setup(ctx, page, offset, count);
 742         __set_page_dirty_nobuffers(page);
 743
 744         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 745                         status, (long long)i_size_read(inode));
 746         if (status < 0)
 747                 nfs_set_pageerror(page);
 748         return status;
 749 }
 750
 751 static void nfs_writepage_release(struct nfs_page *req)
 752 {
 753
 754         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 755                 nfs_end_page_writeback(req->wb_page);
 756                 nfs_inode_remove_request(req);
 757         } else
 758                 nfs_end_page_writeback(req->wb_page);
 759         nfs_clear_page_writeback(req);
 760 }
 761
 762 static inline int flush_task_priority(int how)
 763 {
 764         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 765                 case FLUSH_HIGHPRI:
 766                         return RPC_PRIORITY_HIGH;
 767                 case FLUSH_LOWPRI:
 768                         return RPC_PRIORITY_LOW;
 769         }
 770         return RPC_PRIORITY_NORMAL;
 771 }
 772
 773 /*
 774  * Set up the argument/result storage required for the RPC call.
 775  */
 776 static void nfs_write_rpcsetup(struct nfs_page *req,
 777                 struct nfs_write_data *data,
 778                 const struct rpc_call_ops *call_ops,
 779                 unsigned int count, unsigned int offset,
 780                 int how)
 781 {
 782         struct inode            *inode;
 783         int flags;
 784
 785         /* Set up the RPC argument and reply structs
 786          * NB: take care not to mess about with data->commit et al. */
 787
 788         data->req = req;
 789         data->inode = inode = req->wb_context->dentry->d_inode;
 790         data->cred = req->wb_context->cred;
 791
 792         data->args.fh     = NFS_FH(inode);
 793         data->args.offset = req_offset(req) + offset;
 794         data->args.pgbase = req->wb_pgbase + offset;
 795         data->args.pages  = data->pagevec;
 796         data->args.count  = count;
 797         data->args.context = req->wb_context;
 798
 799         data->res.fattr   = &data->fattr;
 800         data->res.count   = count;
 801         data->res.verf    = &data->verf;
 802         nfs_fattr_init(&data->fattr);
 803
 804         /* Set up the initial task struct.  */
 805         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 806         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 807         NFS_PROTO(inode)->write_setup(data, how);
 808
 809         data->task.tk_priority = flush_task_priority(how);
 810         data->task.tk_cookie = (unsigned long)inode;
 811
 812         dprintk("NFS: %5u initiated write call "
 813                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 814                 data->task.tk_pid,
 815                 inode->i_sb->s_id,
 816                 (long long)NFS_FILEID(inode),
 817                 count,
 818                 (unsigned long long)data->args.offset);
 819 }
 820
 821 static void nfs_execute_write(struct nfs_write_data *data)
 822 {
 823         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 824         sigset_t oldset;
 825
 826         rpc_clnt_sigmask(clnt, &oldset);
 827         rpc_execute(&data->task);
 828         rpc_clnt_sigunmask(clnt, &oldset);
 829 }
 830
 831 /*
 832  * Generate multiple small requests to write out a single
 833  * contiguous dirty area on one page.
 834  */
 835 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 836 {
 837         struct nfs_page *req = nfs_list_entry(head->next);
 838         struct page *page = req->wb_page;
 839         struct nfs_write_data *data;
 840         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 841         unsigned int offset;
 842         int requests = 0;
 843         LIST_HEAD(list);
 844
 845         nfs_list_remove_request(req);
 846
 847         nbytes = count;
 848         do {
 849                 size_t len = min(nbytes, wsize);
 850
 851                 data = nfs_writedata_alloc(1);
 852                 if (!data)
 853                         goto out_bad;
 854                 list_add(&data->pages, &list);
 855                 requests++;
 856                 nbytes -= len;
 857         } while (nbytes != 0);
 858         atomic_set(&req->wb_complete, requests);
 859
 860         ClearPageError(page);
 861         offset = 0;
 862         nbytes = count;
 863         do {
 864                 data = list_entry(list.next, struct nfs_write_data, pages);
 865                 list_del_init(&data->pages);
 866
 867                 data->pagevec[0] = page;
 868
 869                 if (nbytes < wsize)
 870                         wsize = nbytes;
 871                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 872                                    wsize, offset, how);
 873                 offset += wsize;
 874                 nbytes -= wsize;
 875                 nfs_execute_write(data);
 876         } while (nbytes != 0);
 877
 878         return 0;
 879
 880 out_bad:
 881         while (!list_empty(&list)) {
 882                 data = list_entry(list.next, struct nfs_write_data, pages);
 883                 list_del(&data->pages);
 884                 nfs_writedata_release(data);
 885         }
 886         nfs_redirty_request(req);
 887         nfs_end_page_writeback(req->wb_page);
 888         nfs_clear_page_writeback(req);
 889         return -ENOMEM;
 890 }
 891
 892 /*
 893  * Create an RPC task for the given write request and kick it.
 894  * The page must have been locked by the caller.
 895  *
 896  * It may happen that the page we're passed is not marked dirty.
 897  * This is the case if nfs_updatepage detects a conflicting request
 898  * that has been written but not committed.
 899  */
 900 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 901 {
 902         struct nfs_page         *req;
 903         struct page             **pages;
 904         struct nfs_write_data   *data;
 905
 906         data = nfs_writedata_alloc(npages);
 907         if (!data)
 908                 goto out_bad;
 909
 910         pages = data->pagevec;
 911         while (!list_empty(head)) {
 912                 req = nfs_list_entry(head->next);
 913                 nfs_list_remove_request(req);
 914                 nfs_list_add_request(req, &data->pages);
 915                 ClearPageError(req->wb_page);
 916                 *pages++ = req->wb_page;
 917         }
 918         req = nfs_list_entry(data->pages.next);
 919
 920         /* Set up the argument struct */
 921         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 922
 923         nfs_execute_write(data);
 924         return 0;
 925  out_bad:
 926         while (!list_empty(head)) {
 927                 req = nfs_list_entry(head->next);
 928                 nfs_list_remove_request(req);
 929                 nfs_redirty_request(req);
 930                 nfs_end_page_writeback(req->wb_page);
 931                 nfs_clear_page_writeback(req);
 932         }
 933         return -ENOMEM;
 934 }
 935
 936 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 937                                   struct inode *inode, int ioflags)
 938 {
 939         int wsize = NFS_SERVER(inode)->wsize;
 940
 941         if (wsize < PAGE_CACHE_SIZE)
 942                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 943         else
 944                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 945 }
 946
 947 /*
 948  * Handle a write reply that flushed part of a page.
 949  */
 950 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 951 {
 952         struct nfs_write_data   *data = calldata;
 953         struct nfs_page         *req = data->req;
 954         struct page             *page = req->wb_page;
 955
 956         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 957                 req->wb_context->dentry->d_inode->i_sb->s_id,
 958                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 959                 req->wb_bytes,
 960                 (long long)req_offset(req));
 961
 962         if (nfs_writeback_done(task, data) != 0)
 963                 return;
 964
 965         if (task->tk_status < 0) {
 966                 nfs_set_pageerror(page);
 967                 req->wb_context->error = task->tk_status;
 968                 dprintk(", error = %d\n", task->tk_status);
 969                 goto out;
 970         }
 971
 972         if (nfs_write_need_commit(data)) {
 973                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 974
 975                 spin_lock(req_lock);
 976                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 977                         /* Do nothing we need to resend the writes */
 978                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 979                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 980                         dprintk(" defer commit\n");
 981                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 982                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 983                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 984                         dprintk(" server reboot detected\n");
 985                 }
 986                 spin_unlock(req_lock);
 987         } else
 988                 dprintk(" OK\n");
 989
 990 out:
 991         if (atomic_dec_and_test(&req->wb_complete))
 992                 nfs_writepage_release(req);
 993 }
 994
 995 static const struct rpc_call_ops nfs_write_partial_ops = {
 996         .rpc_call_done = nfs_writeback_done_partial,
 997         .rpc_release = nfs_writedata_release,
 998 };
 999
1000 /*
1001  * Handle a write reply that flushes a whole page.
1002  *
1003  * FIXME: There is an inherent race with invalidate_inode_pages and
1004  *        writebacks since the page->count is kept > 1 for as long
1005  *        as the page has a write request pending.
1006  */
1007 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1008 {
1009         struct nfs_write_data   *data = calldata;
1010         struct nfs_page         *req;
1011         struct page             *page;
1012
1013         if (nfs_writeback_done(task, data) != 0)
1014                 return;
1015
1016         /* Update attributes as result of writeback. */
1017         while (!list_empty(&data->pages)) {
1018                 req = nfs_list_entry(data->pages.next);
1019                 nfs_list_remove_request(req);
1020                 page = req->wb_page;
1021
1022                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1023                         req->wb_context->dentry->d_inode->i_sb->s_id,
1024                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1025                         req->wb_bytes,
1026                         (long long)req_offset(req));
1027
1028                 if (task->tk_status < 0) {
1029                         nfs_set_pageerror(page);
1030                         req->wb_context->error = task->tk_status;
1031                         dprintk(", error = %d\n", task->tk_status);
1032                         goto remove_request;
1033                 }
1034
1035                 if (nfs_write_need_commit(data)) {
1036                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1037                         nfs_mark_request_commit(req);
1038                         nfs_end_page_writeback(page);
1039                         dprintk(" marked for commit\n");
1040                         goto next;
1041                 }
1042                 dprintk(" OK\n");
1043 remove_request:
1044                 nfs_end_page_writeback(page);
1045                 nfs_inode_remove_request(req);
1046         next:
1047                 nfs_clear_page_writeback(req);
1048         }
1049 }
1050
1051 static const struct rpc_call_ops nfs_write_full_ops = {
1052         .rpc_call_done = nfs_writeback_done_full,
1053         .rpc_release = nfs_writedata_release,
1054 };
1055
1056
1057 /*
1058  * This function is called when the WRITE call is complete.
1059  */
1060 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1061 {
1062         struct nfs_writeargs    *argp = &data->args;
1063         struct nfs_writeres     *resp = &data->res;
1064         int status;
1065
1066         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1067                 task->tk_pid, task->tk_status);
1068
1069         /*
1070          * ->write_done will attempt to use post-op attributes to detect
1071          * conflicting writes by other clients.  A strict interpretation
1072          * of close-to-open would allow us to continue caching even if
1073          * another writer had changed the file, but some applications
1074          * depend on tighter cache coherency when writing.
1075          */
1076         status = NFS_PROTO(data->inode)->write_done(task, data);
1077         if (status != 0)
1078                 return status;
1079         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1080
1081 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1082         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1083                 /* We tried a write call, but the server did not
1084                  * commit data to stable storage even though we
1085                  * requested it.
1086                  * Note: There is a known bug in Tru64 < 5.0 in which
1087                  *       the server reports NFS_DATA_SYNC, but performs
1088                  *       NFS_FILE_SYNC. We therefore implement this checking
1089                  *       as a dprintk() in order to avoid filling syslog.
1090                  */
1091                 static unsigned long    complain;
1092
1093                 if (time_before(complain, jiffies)) {
1094                         dprintk("NFS: faulty NFS server %s:"
1095                                 " (committed = %d) != (stable = %d)\n",
1096                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1097                                 resp->verf->committed, argp->stable);
1098                         complain = jiffies + 300 * HZ;
1099                 }
1100         }
1101 #endif
1102         /* Is this a short write? */
1103         if (task->tk_status >= 0 && resp->count < argp->count) {
1104                 static unsigned long    complain;
1105
1106                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1107
1108                 /* Has the server at least made some progress? */
1109                 if (resp->count != 0) {
1110                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1111                         if (resp->verf->committed != NFS_UNSTABLE) {
1112                                 /* Resend from where the server left off */
1113                                 argp->offset += resp->count;
1114                                 argp->pgbase += resp->count;
1115                                 argp->count -= resp->count;
1116                         } else {
1117                                 /* Resend as a stable write in order to avoid
1118                                  * headaches in the case of a server crash.
1119                                  */
1120                                 argp->stable = NFS_FILE_SYNC;
1121                         }
1122                         rpc_restart_call(task);
1123                         return -EAGAIN;
1124                 }
1125                 if (time_before(complain, jiffies)) {
1126                         printk(KERN_WARNING
1127                                "NFS: Server wrote zero bytes, expected %u.\n",
1128                                         argp->count);
1129                         complain = jiffies + 300 * HZ;
1130                 }
1131                 /* Can't do anything about it except throw an error. */
1132                 task->tk_status = -EIO;
1133         }
1134         return 0;
1135 }
1136
1137
1138 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1139 void nfs_commit_release(void *wdata)
1140 {
1141         nfs_commit_free(wdata);
1142 }
1143
1144 /*
1145  * Set up the argument/result storage required for the RPC call.
1146  */
1147 static void nfs_commit_rpcsetup(struct list_head *head,
1148                 struct nfs_write_data *data,
1149                 int how)
1150 {
1151         struct nfs_page         *first;
1152         struct inode            *inode;
1153         int flags;
1154
1155         /* Set up the RPC argument and reply structs
1156          * NB: take care not to mess about with data->commit et al. */
1157
1158         list_splice_init(head, &data->pages);
1159         first = nfs_list_entry(data->pages.next);
1160         inode = first->wb_context->dentry->d_inode;
1161
1162         data->inode       = inode;
1163         data->cred        = first->wb_context->cred;
1164
1165         data->args.fh     = NFS_FH(data->inode);
1166         /* Note: we always request a commit of the entire inode */
1167         data->args.offset = 0;
1168         data->args.count  = 0;
1169         data->res.count   = 0;
1170         data->res.fattr   = &data->fattr;
1171         data->res.verf    = &data->verf;
1172         nfs_fattr_init(&data->fattr);
1173
1174         /* Set up the initial task struct.  */
1175         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1176         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1177         NFS_PROTO(inode)->commit_setup(data, how);
1178
1179         data->task.tk_priority = flush_task_priority(how);
1180         data->task.tk_cookie = (unsigned long)inode;
1181
1182         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1183 }
1184
1185 /*
1186  * Commit dirty pages
1187  */
1188 static int
1189 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1190 {
1191         struct nfs_write_data   *data;
1192         struct nfs_page         *req;
1193
1194         data = nfs_commit_alloc();
1195
1196         if (!data)
1197                 goto out_bad;
1198
1199         /* Set up the argument struct */
1200         nfs_commit_rpcsetup(head, data, how);
1201
1202         nfs_execute_write(data);
1203         return 0;
1204  out_bad:
1205         while (!list_empty(head)) {
1206                 req = nfs_list_entry(head->next);
1207                 nfs_list_remove_request(req);
1208                 nfs_mark_request_commit(req);
1209                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1210                 nfs_clear_page_writeback(req);
1211         }
1212         return -ENOMEM;
1213 }
1214
1215 /*
1216  * COMMIT call returned
1217  */
1218 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1219 {
1220         struct nfs_write_data   *data = calldata;
1221         struct nfs_page         *req;
1222
1223         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1224                                 task->tk_pid, task->tk_status);
1225
1226         /* Call the NFS version-specific code */
1227         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1228                 return;
1229
1230         while (!list_empty(&data->pages)) {
1231                 req = nfs_list_entry(data->pages.next);
1232                 nfs_list_remove_request(req);
1233                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1234                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1235
1236                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1237                         req->wb_context->dentry->d_inode->i_sb->s_id,
1238                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1239                         req->wb_bytes,
1240                         (long long)req_offset(req));
1241                 if (task->tk_status < 0) {
1242                         req->wb_context->error = task->tk_status;
1243                         nfs_inode_remove_request(req);
1244                         dprintk(", error = %d\n", task->tk_status);
1245                         goto next;
1246                 }
1247
1248                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1249                  * returned by the server against all stored verfs. */
1250                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1251                         /* We have a match */
1252                         nfs_inode_remove_request(req);
1253                         dprintk(" OK\n");
1254                         goto next;
1255                 }
1256                 /* We have a mismatch. Write the page again */
1257                 dprintk(" mismatch\n");
1258                 nfs_redirty_request(req);
1259         next:
1260                 nfs_clear_page_writeback(req);
1261         }
1262 }
1263
1264 static const struct rpc_call_ops nfs_commit_ops = {
1265         .rpc_call_done = nfs_commit_done,
1266         .rpc_release = nfs_commit_release,
1267 };
1268
1269 int nfs_commit_inode(struct inode *inode, int how)
1270 {
1271         struct nfs_inode *nfsi = NFS_I(inode);
1272         LIST_HEAD(head);
1273         int res;
1274
1275         spin_lock(&nfsi->req_lock);
1276         res = nfs_scan_commit(inode, &head, 0, 0);
1277         spin_unlock(&nfsi->req_lock);
1278         if (res) {
1279                 int error = nfs_commit_list(inode, &head, how);
1280                 if (error < 0)
1281                         return error;
1282         }
1283         return res;
1284 }
1285 #else
1286 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1287 {
1288         return 0;
1289 }
1290 #endif
1291
1292 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1293 {
1294         struct inode *inode = mapping->host;
1295         struct nfs_inode *nfsi = NFS_I(inode);
1296         pgoff_t idx_start, idx_end;
1297         unsigned int npages = 0;
1298         LIST_HEAD(head);
1299         int nocommit = how & FLUSH_NOCOMMIT;
1300         long pages, ret;
1301
1302         /* FIXME */
1303         if (wbc->range_cyclic)
1304                 idx_start = 0;
1305         else {
1306                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1307                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1308                 if (idx_end > idx_start) {
1309                         pgoff_t l_npages = 1 + idx_end - idx_start;
1310                         npages = l_npages;
1311                         if (sizeof(npages) != sizeof(l_npages) &&
1312                                         (pgoff_t)npages != l_npages)
1313                                 npages = 0;
1314                 }
1315         }
1316         how &= ~FLUSH_NOCOMMIT;
1317         spin_lock(&nfsi->req_lock);
1318         do {
1319                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1320                 if (ret != 0)
1321                         continue;
1322                 if (nocommit)
1323                         break;
1324                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1325                 if (pages == 0)
1326                         break;
1327                 if (how & FLUSH_INVALIDATE) {
1328                         spin_unlock(&nfsi->req_lock);
1329                         nfs_cancel_commit_list(&head);
1330                         ret = pages;
1331                         spin_lock(&nfsi->req_lock);
1332                         continue;
1333                 }
1334                 pages += nfs_scan_commit(inode, &head, 0, 0);
1335                 spin_unlock(&nfsi->req_lock);
1336                 ret = nfs_commit_list(inode, &head, how);
1337                 spin_lock(&nfsi->req_lock);
1338         } while (ret >= 0);
1339         spin_unlock(&nfsi->req_lock);
1340         return ret;
1341 }
1342
1343 /*
1344  * flush the inode to disk.
1345  */
1346 int nfs_wb_all(struct inode *inode)
1347 {
1348         struct address_space *mapping = inode->i_mapping;
1349         struct writeback_control wbc = {
1350                 .bdi = mapping->backing_dev_info,
1351                 .sync_mode = WB_SYNC_ALL,
1352                 .nr_to_write = LONG_MAX,
1353                 .for_writepages = 1,
1354                 .range_cyclic = 1,
1355         };
1356         int ret;
1357
1358         ret = nfs_writepages(mapping, &wbc);
1359         if (ret < 0)
1360                 goto out;
1361         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1362         if (ret >= 0)
1363                 return 0;
1364 out:
1365         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1366         return ret;
1367 }
1368
1369 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1370 {
1371         struct writeback_control wbc = {
1372                 .bdi = mapping->backing_dev_info,
1373                 .sync_mode = WB_SYNC_ALL,
1374                 .nr_to_write = LONG_MAX,
1375                 .range_start = range_start,
1376                 .range_end = range_end,
1377                 .for_writepages = 1,
1378         };
1379         int ret;
1380
1381         ret = nfs_writepages(mapping, &wbc);
1382         if (ret < 0)
1383                 goto out;
1384         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1385         if (ret >= 0)
1386                 return 0;
1387 out:
1388         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1389         return ret;
1390 }
1391
1392 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1393 {
1394         loff_t range_start = page_offset(page);
1395         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1396         struct writeback_control wbc = {
1397                 .bdi = page->mapping->backing_dev_info,
1398                 .sync_mode = WB_SYNC_ALL,
1399                 .nr_to_write = LONG_MAX,
1400                 .range_start = range_start,
1401                 .range_end = range_end,
1402         };
1403         int ret;
1404
1405         BUG_ON(!PageLocked(page));
1406         if (clear_page_dirty_for_io(page)) {
1407                 ret = nfs_writepage_locked(page, &wbc);
1408                 if (ret < 0)
1409                         goto out;
1410         }
1411         if (!PagePrivate(page))
1412                 return 0;
1413         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1414         if (ret >= 0)
1415                 return 0;
1416 out:
1417         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1418         return ret;
1419 }
1420
1421 /*
1422  * Write back all requests on one page - we do this before reading it.
1423  */
1424 int nfs_wb_page(struct inode *inode, struct page* page)
1425 {
1426         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1427 }
1428
1429 int nfs_set_page_dirty(struct page *page)
1430 {
1431         struct address_space *mapping = page->mapping;
1432         struct inode *inode;
1433         spinlock_t *req_lock;
1434         struct nfs_page *req;
1435         int ret;
1436
1437         if (!mapping)
1438                 goto out_raced;
1439         inode = mapping->host;
1440         if (!inode)
1441                 goto out_raced;
1442         req_lock = &NFS_I(inode)->req_lock;
1443         spin_lock(req_lock);
1444         req = nfs_page_find_request_locked(page);
1445         if (req != NULL) {
1446                 /* Mark any existing write requests for flushing */
1447                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1448                 spin_unlock(req_lock);
1449                 nfs_release_request(req);
1450                 return ret;
1451         }
1452         ret = __set_page_dirty_nobuffers(page);
1453         spin_unlock(req_lock);
1454         return ret;
1455 out_raced:
1456         return !TestSetPageDirty(page);
1457 }
1458
1459
1460 int __init nfs_init_writepagecache(void)
1461 {
1462         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1463                                              sizeof(struct nfs_write_data),
1464                                              0, SLAB_HWCACHE_ALIGN,
1465                                              NULL, NULL);
1466         if (nfs_wdata_cachep == NULL)
1467                 return -ENOMEM;
1468
1469         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1470                                                      nfs_wdata_cachep);
1471         if (nfs_wdata_mempool == NULL)
1472                 return -ENOMEM;
1473
1474         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1475                                                       nfs_wdata_cachep);
1476         if (nfs_commit_mempool == NULL)
1477                 return -ENOMEM;
1478
1479         /*
1480          * NFS congestion size, scale with available memory.
1481          *
1482          *  64MB:    8192k
1483          * 128MB:   11585k
1484          * 256MB:   16384k
1485          * 512MB:   23170k
1486          *   1GB:   32768k
1487          *   2GB:   46340k
1488          *   4GB:   65536k
1489          *   8GB:   92681k
1490          *  16GB:  131072k
1491          *
1492          * This allows larger machines to have larger/more transfers.
1493          * Limit the default to 256M
1494          */
1495         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1496         if (nfs_congestion_kb > 256*1024)
1497                 nfs_congestion_kb = 256*1024;
1498
1499         return 0;
1500 }
1501
1502 void nfs_destroy_writepagecache(void)
1503 {
1504         mempool_destroy(nfs_commit_mempool);
1505         mempool_destroy(nfs_wdata_mempool);
1506         kmem_cache_destroy(nfs_wdata_cachep);
1507 }
1508