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