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 void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 114 {
 115         ctx->error = error;
 116         smp_wmb();
 117         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 118 }
 119
 120 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 121 {
 122         struct nfs_page *req = NULL;
 123
 124         if (PagePrivate(page)) {
 125                 req = (struct nfs_page *)page_private(page);
 126                 if (req != NULL)
 127                         kref_get(&req->wb_kref);
 128         }
 129         return req;
 130 }
 131
 132 static struct nfs_page *nfs_page_find_request(struct page *page)
 133 {
 134         struct inode *inode = page->mapping->host;
 135         struct nfs_page *req = NULL;
 136
 137         spin_lock(&inode->i_lock);
 138         req = nfs_page_find_request_locked(page);
 139         spin_unlock(&inode->i_lock);
 140         return req;
 141 }
 142
 143 /* Adjust the file length if we're writing beyond the end */
 144 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 145 {
 146         struct inode *inode = page->mapping->host;
 147         loff_t end, i_size = i_size_read(inode);
 148         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 149
 150         if (i_size > 0 && page->index < end_index)
 151                 return;
 152         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 153         if (i_size >= end)
 154                 return;
 155         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 156         i_size_write(inode, end);
 157 }
 158
 159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 160 static void nfs_set_pageerror(struct page *page)
 161 {
 162         SetPageError(page);
 163         nfs_zap_mapping(page->mapping->host, page->mapping);
 164 }
 165
 166 /* We can set the PG_uptodate flag if we see that a write request
 167  * covers the full page.
 168  */
 169 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 170 {
 171         if (PageUptodate(page))
 172                 return;
 173         if (base != 0)
 174                 return;
 175         if (count != nfs_page_length(page))
 176                 return;
 177         SetPageUptodate(page);
 178 }
 179
 180 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 181                 unsigned int offset, unsigned int count)
 182 {
 183         struct nfs_page *req;
 184         int ret;
 185
 186         for (;;) {
 187                 req = nfs_update_request(ctx, page, offset, count);
 188                 if (!IS_ERR(req))
 189                         break;
 190                 ret = PTR_ERR(req);
 191                 if (ret != -EBUSY)
 192                         return ret;
 193                 ret = nfs_wb_page(page->mapping->host, page);
 194                 if (ret != 0)
 195                         return ret;
 196         }
 197         /* Update file length */
 198         nfs_grow_file(page, offset, count);
 199         nfs_clear_page_tag_locked(req);
 200         return 0;
 201 }
 202
 203 static int wb_priority(struct writeback_control *wbc)
 204 {
 205         if (wbc->for_reclaim)
 206                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 207         if (wbc->for_kupdate)
 208                 return FLUSH_LOWPRI;
 209         return 0;
 210 }
 211
 212 /*
 213  * NFS congestion control
 214  */
 215
 216 int nfs_congestion_kb;
 217
 218 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 219 #define NFS_CONGESTION_OFF_THRESH       \
 220         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 221
 222 static int nfs_set_page_writeback(struct page *page)
 223 {
 224         int ret = test_set_page_writeback(page);
 225
 226         if (!ret) {
 227                 struct inode *inode = page->mapping->host;
 228                 struct nfs_server *nfss = NFS_SERVER(inode);
 229
 230                 if (atomic_long_inc_return(&nfss->writeback) >
 231                                 NFS_CONGESTION_ON_THRESH)
 232                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 233         }
 234         return ret;
 235 }
 236
 237 static void nfs_end_page_writeback(struct page *page)
 238 {
 239         struct inode *inode = page->mapping->host;
 240         struct nfs_server *nfss = NFS_SERVER(inode);
 241
 242         end_page_writeback(page);
 243         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 244                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 245 }
 246
 247 /*
 248  * Find an associated nfs write request, and prepare to flush it out
 249  * May return an error if the user signalled nfs_wait_on_request().
 250  */
 251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 252                                 struct page *page)
 253 {
 254         struct inode *inode = page->mapping->host;
 255         struct nfs_page *req;
 256         int ret;
 257
 258         spin_lock(&inode->i_lock);
 259         for(;;) {
 260                 req = nfs_page_find_request_locked(page);
 261                 if (req == NULL) {
 262                         spin_unlock(&inode->i_lock);
 263                         return 0;
 264                 }
 265                 if (nfs_set_page_tag_locked(req))
 266                         break;
 267                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 268                  *       then the call to nfs_set_page_tag_locked() will always
 269                  *       succeed provided that someone hasn't already marked the
 270                  *       request as dirty (in which case we don't care).
 271                  */
 272                 spin_unlock(&inode->i_lock);
 273                 ret = nfs_wait_on_request(req);
 274                 nfs_release_request(req);
 275                 if (ret != 0)
 276                         return ret;
 277                 spin_lock(&inode->i_lock);
 278         }
 279         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 280                 /* This request is marked for commit */
 281                 spin_unlock(&inode->i_lock);
 282                 nfs_clear_page_tag_locked(req);
 283                 nfs_pageio_complete(pgio);
 284                 return 0;
 285         }
 286         if (nfs_set_page_writeback(page) != 0) {
 287                 spin_unlock(&inode->i_lock);
 288                 BUG();
 289         }
 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                 if (nfs_have_delegation(inode, FMODE_WRITE))
 375                         nfsi->change_attr++;
 376         }
 377         SetPagePrivate(req->wb_page);
 378         set_page_private(req->wb_page, (unsigned long)req);
 379         nfsi->npages++;
 380         kref_get(&req->wb_kref);
 381         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 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                 iput(inode);
 403         } else
 404                 spin_unlock(&inode->i_lock);
 405         nfs_clear_request(req);
 406         nfs_release_request(req);
 407 }
 408
 409 static void
 410 nfs_redirty_request(struct nfs_page *req)
 411 {
 412         __set_page_dirty_nobuffers(req->wb_page);
 413 }
 414
 415 /*
 416  * Check if a request is dirty
 417  */
 418 static inline int
 419 nfs_dirty_request(struct nfs_page *req)
 420 {
 421         struct page *page = req->wb_page;
 422
 423         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 424                 return 0;
 425         return !PageWriteback(req->wb_page);
 426 }
 427
 428 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 429 /*
 430  * Add a request to the inode's commit list.
 431  */
 432 static void
 433 nfs_mark_request_commit(struct nfs_page *req)
 434 {
 435         struct inode *inode = req->wb_context->path.dentry->d_inode;
 436         struct nfs_inode *nfsi = NFS_I(inode);
 437
 438         spin_lock(&inode->i_lock);
 439         nfsi->ncommit++;
 440         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 441         radix_tree_tag_set(&nfsi->nfs_page_tree,
 442                         req->wb_index,
 443                         NFS_PAGE_TAG_COMMIT);
 444         spin_unlock(&inode->i_lock);
 445         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 446         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 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                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
 534                                 BDI_RECLAIMABLE);
 535                 nfs_list_remove_request(req);
 536                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 537                 nfs_inode_remove_request(req);
 538                 nfs_unlock_request(req);
 539         }
 540 }
 541
 542 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 543 /*
 544  * nfs_scan_commit - Scan an inode for commit requests
 545  * @inode: NFS inode to scan
 546  * @dst: destination list
 547  * @idx_start: lower bound of page->index to scan.
 548  * @npages: idx_start + npages sets the upper bound to scan.
 549  *
 550  * Moves requests from the inode's 'commit' request list.
 551  * The requests are *not* checked to ensure that they form a contiguous set.
 552  */
 553 static int
 554 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 555 {
 556         struct nfs_inode *nfsi = NFS_I(inode);
 557         int res = 0;
 558
 559         if (nfsi->ncommit != 0) {
 560                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 561                                 NFS_PAGE_TAG_COMMIT);
 562                 nfsi->ncommit -= res;
 563         }
 564         return res;
 565 }
 566 #else
 567 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 568 {
 569         return 0;
 570 }
 571 #endif
 572
 573 /*
 574  * Try to update any existing write request, or create one if there is none.
 575  * In order to match, the request's credentials must match those of
 576  * the calling process.
 577  *
 578  * Note: Should always be called with the Page Lock held!
 579  */
 580 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 581                 struct page *page, unsigned int offset, unsigned int bytes)
 582 {
 583         struct address_space *mapping = page->mapping;
 584         struct inode *inode = mapping->host;
 585         struct nfs_page         *req, *new = NULL;
 586         pgoff_t         rqend, end;
 587
 588         end = offset + bytes;
 589
 590         for (;;) {
 591                 /* Loop over all inode entries and see if we find
 592                  * A request for the page we wish to update
 593                  */
 594                 spin_lock(&inode->i_lock);
 595                 req = nfs_page_find_request_locked(page);
 596                 if (req) {
 597                         if (!nfs_set_page_tag_locked(req)) {
 598                                 int error;
 599
 600                                 spin_unlock(&inode->i_lock);
 601                                 error = nfs_wait_on_request(req);
 602                                 nfs_release_request(req);
 603                                 if (error < 0) {
 604                                         if (new)
 605                                                 nfs_release_request(new);
 606                                         return ERR_PTR(error);
 607                                 }
 608                                 continue;
 609                         }
 610                         spin_unlock(&inode->i_lock);
 611                         if (new)
 612                                 nfs_release_request(new);
 613                         break;
 614                 }
 615
 616                 if (new) {
 617                         int error;
 618                         nfs_lock_request_dontget(new);
 619                         error = nfs_inode_add_request(inode, new);
 620                         if (error) {
 621                                 spin_unlock(&inode->i_lock);
 622                                 nfs_unlock_request(new);
 623                                 return ERR_PTR(error);
 624                         }
 625                         spin_unlock(&inode->i_lock);
 626                         req = new;
 627                         goto zero_page;
 628                 }
 629                 spin_unlock(&inode->i_lock);
 630
 631                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 632                 if (IS_ERR(new))
 633                         return new;
 634         }
 635
 636         /* We have a request for our page.
 637          * If the creds don't match, or the
 638          * page addresses don't match,
 639          * tell the caller to wait on the conflicting
 640          * request.
 641          */
 642         rqend = req->wb_offset + req->wb_bytes;
 643         if (req->wb_context != ctx
 644             || req->wb_page != page
 645             || !nfs_dirty_request(req)
 646             || offset > rqend || end < req->wb_offset) {
 647                 nfs_clear_page_tag_locked(req);
 648                 return ERR_PTR(-EBUSY);
 649         }
 650
 651         /* Okay, the request matches. Update the region */
 652         if (offset < req->wb_offset) {
 653                 req->wb_offset = offset;
 654                 req->wb_pgbase = offset;
 655                 req->wb_bytes = max(end, rqend) - req->wb_offset;
 656                 goto zero_page;
 657         }
 658
 659         if (end > rqend)
 660                 req->wb_bytes = end - req->wb_offset;
 661
 662         return req;
 663 zero_page:
 664         /* If this page might potentially be marked as up to date,
 665          * then we need to zero any uninitalised data. */
 666         if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE
 667                         && !PageUptodate(req->wb_page))
 668                 zero_user_page(req->wb_page, req->wb_bytes,
 669                                 PAGE_CACHE_SIZE - req->wb_bytes,
 670                                 KM_USER0);
 671         return req;
 672 }
 673
 674 int nfs_flush_incompatible(struct file *file, struct page *page)
 675 {
 676         struct nfs_open_context *ctx = nfs_file_open_context(file);
 677         struct nfs_page *req;
 678         int do_flush, status;
 679         /*
 680          * Look for a request corresponding to this page. If there
 681          * is one, and it belongs to another file, we flush it out
 682          * before we try to copy anything into the page. Do this
 683          * due to the lack of an ACCESS-type call in NFSv2.
 684          * Also do the same if we find a request from an existing
 685          * dropped page.
 686          */
 687         do {
 688                 req = nfs_page_find_request(page);
 689                 if (req == NULL)
 690                         return 0;
 691                 do_flush = req->wb_page != page || req->wb_context != ctx
 692                         || !nfs_dirty_request(req);
 693                 nfs_release_request(req);
 694                 if (!do_flush)
 695                         return 0;
 696                 status = nfs_wb_page(page->mapping->host, page);
 697         } while (status == 0);
 698         return status;
 699 }
 700
 701 /*
 702  * Update and possibly write a cached page of an NFS file.
 703  *
 704  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 705  * things with a page scheduled for an RPC call (e.g. invalidate it).
 706  */
 707 int nfs_updatepage(struct file *file, struct page *page,
 708                 unsigned int offset, unsigned int count)
 709 {
 710         struct nfs_open_context *ctx = nfs_file_open_context(file);
 711         struct inode    *inode = page->mapping->host;
 712         int             status = 0;
 713
 714         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 715
 716         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 717                 file->f_path.dentry->d_parent->d_name.name,
 718                 file->f_path.dentry->d_name.name, count,
 719                 (long long)(page_offset(page) +offset));
 720
 721         /* If we're not using byte range locks, and we know the page
 722          * is entirely in cache, it may be more efficient to avoid
 723          * fragmenting write requests.
 724          */
 725         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 726                 count = max(count + offset, nfs_page_length(page));
 727                 offset = 0;
 728         }
 729
 730         status = nfs_writepage_setup(ctx, page, offset, count);
 731         __set_page_dirty_nobuffers(page);
 732
 733         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 734                         status, (long long)i_size_read(inode));
 735         if (status < 0)
 736                 nfs_set_pageerror(page);
 737         return status;
 738 }
 739
 740 static void nfs_writepage_release(struct nfs_page *req)
 741 {
 742
 743         if (PageError(req->wb_page)) {
 744                 nfs_end_page_writeback(req->wb_page);
 745                 nfs_inode_remove_request(req);
 746         } else if (!nfs_reschedule_unstable_write(req)) {
 747                 /* Set the PG_uptodate flag */
 748                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 749                 nfs_end_page_writeback(req->wb_page);
 750                 nfs_inode_remove_request(req);
 751         } else
 752                 nfs_end_page_writeback(req->wb_page);
 753         nfs_clear_page_tag_locked(req);
 754 }
 755
 756 static int flush_task_priority(int how)
 757 {
 758         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 759                 case FLUSH_HIGHPRI:
 760                         return RPC_PRIORITY_HIGH;
 761                 case FLUSH_LOWPRI:
 762                         return RPC_PRIORITY_LOW;
 763         }
 764         return RPC_PRIORITY_NORMAL;
 765 }
 766
 767 static void nfs_execute_write(struct nfs_write_data *data)
 768 {
 769         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 770         sigset_t oldset;
 771
 772         rpc_clnt_sigmask(clnt, &oldset);
 773         rpc_execute(&data->task);
 774         rpc_clnt_sigunmask(clnt, &oldset);
 775 }
 776
 777 /*
 778  * Set up the argument/result storage required for the RPC call.
 779  */
 780 static void nfs_write_rpcsetup(struct nfs_page *req,
 781                 struct nfs_write_data *data,
 782                 const struct rpc_call_ops *call_ops,
 783                 unsigned int count, unsigned int offset,
 784                 int how)
 785 {
 786         struct inode *inode = req->wb_context->path.dentry->d_inode;
 787         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 788         int priority = flush_task_priority(how);
 789         struct rpc_message msg = {
 790                 .rpc_argp = &data->args,
 791                 .rpc_resp = &data->res,
 792                 .rpc_cred = req->wb_context->cred,
 793         };
 794         struct rpc_task_setup task_setup_data = {
 795                 .rpc_client = NFS_CLIENT(inode),
 796                 .rpc_message = &msg,
 797                 .callback_ops = call_ops,
 798                 .callback_data = data,
 799                 .flags = flags,
 800                 .priority = priority,
 801         };
 802
 803         /* Set up the RPC argument and reply structs
 804          * NB: take care not to mess about with data->commit et al. */
 805
 806         data->req = req;
 807         data->inode = inode = req->wb_context->path.dentry->d_inode;
 808         data->cred = msg.rpc_cred;
 809
 810         data->args.fh     = NFS_FH(inode);
 811         data->args.offset = req_offset(req) + offset;
 812         data->args.pgbase = req->wb_pgbase + offset;
 813         data->args.pages  = data->pagevec;
 814         data->args.count  = count;
 815         data->args.context = req->wb_context;
 816         data->args.stable  = NFS_UNSTABLE;
 817         if (how & FLUSH_STABLE) {
 818                 data->args.stable = NFS_DATA_SYNC;
 819                 if (!NFS_I(inode)->ncommit)
 820                         data->args.stable = NFS_FILE_SYNC;
 821         }
 822
 823         data->res.fattr   = &data->fattr;
 824         data->res.count   = count;
 825         data->res.verf    = &data->verf;
 826         nfs_fattr_init(&data->fattr);
 827
 828         /* Set up the initial task struct.  */
 829         NFS_PROTO(inode)->write_setup(data, &msg);
 830         rpc_init_task(&data->task, &task_setup_data);
 831
 832         dprintk("NFS: %5u initiated write call "
 833                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 834                 data->task.tk_pid,
 835                 inode->i_sb->s_id,
 836                 (long long)NFS_FILEID(inode),
 837                 count,
 838                 (unsigned long long)data->args.offset);
 839
 840         nfs_execute_write(data);
 841 }
 842
 843 /*
 844  * Generate multiple small requests to write out a single
 845  * contiguous dirty area on one page.
 846  */
 847 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 848 {
 849         struct nfs_page *req = nfs_list_entry(head->next);
 850         struct page *page = req->wb_page;
 851         struct nfs_write_data *data;
 852         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 853         unsigned int offset;
 854         int requests = 0;
 855         LIST_HEAD(list);
 856
 857         nfs_list_remove_request(req);
 858
 859         nbytes = count;
 860         do {
 861                 size_t len = min(nbytes, wsize);
 862
 863                 data = nfs_writedata_alloc(1);
 864                 if (!data)
 865                         goto out_bad;
 866                 list_add(&data->pages, &list);
 867                 requests++;
 868                 nbytes -= len;
 869         } while (nbytes != 0);
 870         atomic_set(&req->wb_complete, requests);
 871
 872         ClearPageError(page);
 873         offset = 0;
 874         nbytes = count;
 875         do {
 876                 data = list_entry(list.next, struct nfs_write_data, pages);
 877                 list_del_init(&data->pages);
 878
 879                 data->pagevec[0] = page;
 880
 881                 if (nbytes < wsize)
 882                         wsize = nbytes;
 883                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 884                                    wsize, offset, how);
 885                 offset += wsize;
 886                 nbytes -= wsize;
 887         } while (nbytes != 0);
 888
 889         return 0;
 890
 891 out_bad:
 892         while (!list_empty(&list)) {
 893                 data = list_entry(list.next, struct nfs_write_data, pages);
 894                 list_del(&data->pages);
 895                 nfs_writedata_release(data);
 896         }
 897         nfs_redirty_request(req);
 898         nfs_end_page_writeback(req->wb_page);
 899         nfs_clear_page_tag_locked(req);
 900         return -ENOMEM;
 901 }
 902
 903 /*
 904  * Create an RPC task for the given write request and kick it.
 905  * The page must have been locked by the caller.
 906  *
 907  * It may happen that the page we're passed is not marked dirty.
 908  * This is the case if nfs_updatepage detects a conflicting request
 909  * that has been written but not committed.
 910  */
 911 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 912 {
 913         struct nfs_page         *req;
 914         struct page             **pages;
 915         struct nfs_write_data   *data;
 916
 917         data = nfs_writedata_alloc(npages);
 918         if (!data)
 919                 goto out_bad;
 920
 921         pages = data->pagevec;
 922         while (!list_empty(head)) {
 923                 req = nfs_list_entry(head->next);
 924                 nfs_list_remove_request(req);
 925                 nfs_list_add_request(req, &data->pages);
 926                 ClearPageError(req->wb_page);
 927                 *pages++ = req->wb_page;
 928         }
 929         req = nfs_list_entry(data->pages.next);
 930
 931         /* Set up the argument struct */
 932         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 933
 934         return 0;
 935  out_bad:
 936         while (!list_empty(head)) {
 937                 req = nfs_list_entry(head->next);
 938                 nfs_list_remove_request(req);
 939                 nfs_redirty_request(req);
 940                 nfs_end_page_writeback(req->wb_page);
 941                 nfs_clear_page_tag_locked(req);
 942         }
 943         return -ENOMEM;
 944 }
 945
 946 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 947                                   struct inode *inode, int ioflags)
 948 {
 949         int wsize = NFS_SERVER(inode)->wsize;
 950
 951         if (wsize < PAGE_CACHE_SIZE)
 952                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 953         else
 954                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 955 }
 956
 957 /*
 958  * Handle a write reply that flushed part of a page.
 959  */
 960 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 961 {
 962         struct nfs_write_data   *data = calldata;
 963         struct nfs_page         *req = data->req;
 964         struct page             *page = req->wb_page;
 965
 966         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 967                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 968                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 969                 req->wb_bytes,
 970                 (long long)req_offset(req));
 971
 972         if (nfs_writeback_done(task, data) != 0)
 973                 return;
 974
 975         if (task->tk_status < 0) {
 976                 nfs_set_pageerror(page);
 977                 nfs_context_set_write_error(req->wb_context, task->tk_status);
 978                 dprintk(", error = %d\n", task->tk_status);
 979                 goto out;
 980         }
 981
 982         if (nfs_write_need_commit(data)) {
 983                 struct inode *inode = page->mapping->host;
 984
 985                 spin_lock(&inode->i_lock);
 986                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 987                         /* Do nothing we need to resend the writes */
 988                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 989                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 990                         dprintk(" defer commit\n");
 991                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 992                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 993                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 994                         dprintk(" server reboot detected\n");
 995                 }
 996                 spin_unlock(&inode->i_lock);
 997         } else
 998                 dprintk(" OK\n");
 999
1000 out:
1001         if (atomic_dec_and_test(&req->wb_complete))
1002                 nfs_writepage_release(req);
1003 }
1004
1005 static const struct rpc_call_ops nfs_write_partial_ops = {
1006         .rpc_call_done = nfs_writeback_done_partial,
1007         .rpc_release = nfs_writedata_release,
1008 };
1009
1010 /*
1011  * Handle a write reply that flushes a whole page.
1012  *
1013  * FIXME: There is an inherent race with invalidate_inode_pages and
1014  *        writebacks since the page->count is kept > 1 for as long
1015  *        as the page has a write request pending.
1016  */
1017 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1018 {
1019         struct nfs_write_data   *data = calldata;
1020         struct nfs_page         *req;
1021         struct page             *page;
1022
1023         if (nfs_writeback_done(task, data) != 0)
1024                 return;
1025
1026         /* Update attributes as result of writeback. */
1027         while (!list_empty(&data->pages)) {
1028                 req = nfs_list_entry(data->pages.next);
1029                 nfs_list_remove_request(req);
1030                 page = req->wb_page;
1031
1032                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1033                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1034                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1035                         req->wb_bytes,
1036                         (long long)req_offset(req));
1037
1038                 if (task->tk_status < 0) {
1039                         nfs_set_pageerror(page);
1040                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1041                         dprintk(", error = %d\n", task->tk_status);
1042                         goto remove_request;
1043                 }
1044
1045                 if (nfs_write_need_commit(data)) {
1046                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1047                         nfs_mark_request_commit(req);
1048                         nfs_end_page_writeback(page);
1049                         dprintk(" marked for commit\n");
1050                         goto next;
1051                 }
1052                 /* Set the PG_uptodate flag? */
1053                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1054                 dprintk(" OK\n");
1055 remove_request:
1056                 nfs_end_page_writeback(page);
1057                 nfs_inode_remove_request(req);
1058         next:
1059                 nfs_clear_page_tag_locked(req);
1060         }
1061 }
1062
1063 static const struct rpc_call_ops nfs_write_full_ops = {
1064         .rpc_call_done = nfs_writeback_done_full,
1065         .rpc_release = nfs_writedata_release,
1066 };
1067
1068
1069 /*
1070  * This function is called when the WRITE call is complete.
1071  */
1072 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1073 {
1074         struct nfs_writeargs    *argp = &data->args;
1075         struct nfs_writeres     *resp = &data->res;
1076         int status;
1077
1078         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1079                 task->tk_pid, task->tk_status);
1080
1081         /*
1082          * ->write_done will attempt to use post-op attributes to detect
1083          * conflicting writes by other clients.  A strict interpretation
1084          * of close-to-open would allow us to continue caching even if
1085          * another writer had changed the file, but some applications
1086          * depend on tighter cache coherency when writing.
1087          */
1088         status = NFS_PROTO(data->inode)->write_done(task, data);
1089         if (status != 0)
1090                 return status;
1091         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1092
1093 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1094         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1095                 /* We tried a write call, but the server did not
1096                  * commit data to stable storage even though we
1097                  * requested it.
1098                  * Note: There is a known bug in Tru64 < 5.0 in which
1099                  *       the server reports NFS_DATA_SYNC, but performs
1100                  *       NFS_FILE_SYNC. We therefore implement this checking
1101                  *       as a dprintk() in order to avoid filling syslog.
1102                  */
1103                 static unsigned long    complain;
1104
1105                 if (time_before(complain, jiffies)) {
1106                         dprintk("NFS: faulty NFS server %s:"
1107                                 " (committed = %d) != (stable = %d)\n",
1108                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1109                                 resp->verf->committed, argp->stable);
1110                         complain = jiffies + 300 * HZ;
1111                 }
1112         }
1113 #endif
1114         /* Is this a short write? */
1115         if (task->tk_status >= 0 && resp->count < argp->count) {
1116                 static unsigned long    complain;
1117
1118                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1119
1120                 /* Has the server at least made some progress? */
1121                 if (resp->count != 0) {
1122                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1123                         if (resp->verf->committed != NFS_UNSTABLE) {
1124                                 /* Resend from where the server left off */
1125                                 argp->offset += resp->count;
1126                                 argp->pgbase += resp->count;
1127                                 argp->count -= resp->count;
1128                         } else {
1129                                 /* Resend as a stable write in order to avoid
1130                                  * headaches in the case of a server crash.
1131                                  */
1132                                 argp->stable = NFS_FILE_SYNC;
1133                         }
1134                         rpc_restart_call(task);
1135                         return -EAGAIN;
1136                 }
1137                 if (time_before(complain, jiffies)) {
1138                         printk(KERN_WARNING
1139                                "NFS: Server wrote zero bytes, expected %u.\n",
1140                                         argp->count);
1141                         complain = jiffies + 300 * HZ;
1142                 }
1143                 /* Can't do anything about it except throw an error. */
1144                 task->tk_status = -EIO;
1145         }
1146         return 0;
1147 }
1148
1149
1150 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1151 void nfs_commit_release(void *wdata)
1152 {
1153         nfs_commit_free(wdata);
1154 }
1155
1156 /*
1157  * Set up the argument/result storage required for the RPC call.
1158  */
1159 static void nfs_commit_rpcsetup(struct list_head *head,
1160                 struct nfs_write_data *data,
1161                 int how)
1162 {
1163         struct nfs_page *first = nfs_list_entry(head->next);
1164         struct inode *inode = first->wb_context->path.dentry->d_inode;
1165         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1166         int priority = flush_task_priority(how);
1167         struct rpc_message msg = {
1168                 .rpc_argp = &data->args,
1169                 .rpc_resp = &data->res,
1170                 .rpc_cred = first->wb_context->cred,
1171         };
1172         struct rpc_task_setup task_setup_data = {
1173                 .rpc_client = NFS_CLIENT(inode),
1174                 .rpc_message = &msg,
1175                 .callback_ops = &nfs_commit_ops,
1176                 .callback_data = data,
1177                 .flags = flags,
1178                 .priority = priority,
1179         };
1180
1181         /* Set up the RPC argument and reply structs
1182          * NB: take care not to mess about with data->commit et al. */
1183
1184         list_splice_init(head, &data->pages);
1185
1186         data->inode       = inode;
1187         data->cred        = msg.rpc_cred;
1188
1189         data->args.fh     = NFS_FH(data->inode);
1190         /* Note: we always request a commit of the entire inode */
1191         data->args.offset = 0;
1192         data->args.count  = 0;
1193         data->res.count   = 0;
1194         data->res.fattr   = &data->fattr;
1195         data->res.verf    = &data->verf;
1196         nfs_fattr_init(&data->fattr);
1197
1198         /* Set up the initial task struct.  */
1199         NFS_PROTO(inode)->commit_setup(data, &msg);
1200         rpc_init_task(&data->task, &task_setup_data);
1201
1202         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1203
1204         nfs_execute_write(data);
1205 }
1206
1207 /*
1208  * Commit dirty pages
1209  */
1210 static int
1211 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1212 {
1213         struct nfs_write_data   *data;
1214         struct nfs_page         *req;
1215
1216         data = nfs_commit_alloc();
1217
1218         if (!data)
1219                 goto out_bad;
1220
1221         /* Set up the argument struct */
1222         nfs_commit_rpcsetup(head, data, how);
1223
1224         return 0;
1225  out_bad:
1226         while (!list_empty(head)) {
1227                 req = nfs_list_entry(head->next);
1228                 nfs_list_remove_request(req);
1229                 nfs_mark_request_commit(req);
1230                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1231                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1232                                 BDI_RECLAIMABLE);
1233                 nfs_clear_page_tag_locked(req);
1234         }
1235         return -ENOMEM;
1236 }
1237
1238 /*
1239  * COMMIT call returned
1240  */
1241 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1242 {
1243         struct nfs_write_data   *data = calldata;
1244         struct nfs_page         *req;
1245
1246         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1247                                 task->tk_pid, task->tk_status);
1248
1249         /* Call the NFS version-specific code */
1250         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1251                 return;
1252
1253         while (!list_empty(&data->pages)) {
1254                 req = nfs_list_entry(data->pages.next);
1255                 nfs_list_remove_request(req);
1256                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1257                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1258                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1259                                 BDI_RECLAIMABLE);
1260
1261                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1262                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1263                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1264                         req->wb_bytes,
1265                         (long long)req_offset(req));
1266                 if (task->tk_status < 0) {
1267                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1268                         nfs_inode_remove_request(req);
1269                         dprintk(", error = %d\n", task->tk_status);
1270                         goto next;
1271                 }
1272
1273                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1274                  * returned by the server against all stored verfs. */
1275                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1276                         /* We have a match */
1277                         /* Set the PG_uptodate flag */
1278                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1279                                         req->wb_bytes);
1280                         nfs_inode_remove_request(req);
1281                         dprintk(" OK\n");
1282                         goto next;
1283                 }
1284                 /* We have a mismatch. Write the page again */
1285                 dprintk(" mismatch\n");
1286                 nfs_redirty_request(req);
1287         next:
1288                 nfs_clear_page_tag_locked(req);
1289         }
1290 }
1291
1292 static const struct rpc_call_ops nfs_commit_ops = {
1293         .rpc_call_done = nfs_commit_done,
1294         .rpc_release = nfs_commit_release,
1295 };
1296
1297 int nfs_commit_inode(struct inode *inode, int how)
1298 {
1299         LIST_HEAD(head);
1300         int res;
1301
1302         spin_lock(&inode->i_lock);
1303         res = nfs_scan_commit(inode, &head, 0, 0);
1304         spin_unlock(&inode->i_lock);
1305         if (res) {
1306                 int error = nfs_commit_list(inode, &head, how);
1307                 if (error < 0)
1308                         return error;
1309         }
1310         return res;
1311 }
1312 #else
1313 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1314 {
1315         return 0;
1316 }
1317 #endif
1318
1319 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1320 {
1321         struct inode *inode = mapping->host;
1322         pgoff_t idx_start, idx_end;
1323         unsigned int npages = 0;
1324         LIST_HEAD(head);
1325         int nocommit = how & FLUSH_NOCOMMIT;
1326         long pages, ret;
1327
1328         /* FIXME */
1329         if (wbc->range_cyclic)
1330                 idx_start = 0;
1331         else {
1332                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1333                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1334                 if (idx_end > idx_start) {
1335                         pgoff_t l_npages = 1 + idx_end - idx_start;
1336                         npages = l_npages;
1337                         if (sizeof(npages) != sizeof(l_npages) &&
1338                                         (pgoff_t)npages != l_npages)
1339                                 npages = 0;
1340                 }
1341         }
1342         how &= ~FLUSH_NOCOMMIT;
1343         spin_lock(&inode->i_lock);
1344         do {
1345                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1346                 if (ret != 0)
1347                         continue;
1348                 if (nocommit)
1349                         break;
1350                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1351                 if (pages == 0)
1352                         break;
1353                 if (how & FLUSH_INVALIDATE) {
1354                         spin_unlock(&inode->i_lock);
1355                         nfs_cancel_commit_list(&head);
1356                         ret = pages;
1357                         spin_lock(&inode->i_lock);
1358                         continue;
1359                 }
1360                 pages += nfs_scan_commit(inode, &head, 0, 0);
1361                 spin_unlock(&inode->i_lock);
1362                 ret = nfs_commit_list(inode, &head, how);
1363                 spin_lock(&inode->i_lock);
1364
1365         } while (ret >= 0);
1366         spin_unlock(&inode->i_lock);
1367         return ret;
1368 }
1369
1370 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1371 {
1372         int ret;
1373
1374         ret = nfs_writepages(mapping, wbc);
1375         if (ret < 0)
1376                 goto out;
1377         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1378         if (ret < 0)
1379                 goto out;
1380         return 0;
1381 out:
1382         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1383         return ret;
1384 }
1385
1386 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1387 static int nfs_write_mapping(struct address_space *mapping, int how)
1388 {
1389         struct writeback_control wbc = {
1390                 .bdi = mapping->backing_dev_info,
1391                 .sync_mode = WB_SYNC_NONE,
1392                 .nr_to_write = LONG_MAX,
1393                 .for_writepages = 1,
1394                 .range_cyclic = 1,
1395         };
1396         int ret;
1397
1398         ret = __nfs_write_mapping(mapping, &wbc, how);
1399         if (ret < 0)
1400                 return ret;
1401         wbc.sync_mode = WB_SYNC_ALL;
1402         return __nfs_write_mapping(mapping, &wbc, how);
1403 }
1404
1405 /*
1406  * flush the inode to disk.
1407  */
1408 int nfs_wb_all(struct inode *inode)
1409 {
1410         return nfs_write_mapping(inode->i_mapping, 0);
1411 }
1412
1413 int nfs_wb_nocommit(struct inode *inode)
1414 {
1415         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1416 }
1417
1418 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1419 {
1420         struct nfs_page *req;
1421         loff_t range_start = page_offset(page);
1422         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1423         struct writeback_control wbc = {
1424                 .bdi = page->mapping->backing_dev_info,
1425                 .sync_mode = WB_SYNC_ALL,
1426                 .nr_to_write = LONG_MAX,
1427                 .range_start = range_start,
1428                 .range_end = range_end,
1429         };
1430         int ret = 0;
1431
1432         BUG_ON(!PageLocked(page));
1433         for (;;) {
1434                 req = nfs_page_find_request(page);
1435                 if (req == NULL)
1436                         goto out;
1437                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1438                         nfs_release_request(req);
1439                         break;
1440                 }
1441                 if (nfs_lock_request_dontget(req)) {
1442                         nfs_inode_remove_request(req);
1443                         /*
1444                          * In case nfs_inode_remove_request has marked the
1445                          * page as being dirty
1446                          */
1447                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1448                         nfs_unlock_request(req);
1449                         break;
1450                 }
1451                 ret = nfs_wait_on_request(req);
1452                 if (ret < 0)
1453                         goto out;
1454         }
1455         if (!PagePrivate(page))
1456                 return 0;
1457         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1458 out:
1459         return ret;
1460 }
1461
1462 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1463                                 int how)
1464 {
1465         loff_t range_start = page_offset(page);
1466         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1467         struct writeback_control wbc = {
1468                 .bdi = page->mapping->backing_dev_info,
1469                 .sync_mode = WB_SYNC_ALL,
1470                 .nr_to_write = LONG_MAX,
1471                 .range_start = range_start,
1472                 .range_end = range_end,
1473         };
1474         int ret;
1475
1476         BUG_ON(!PageLocked(page));
1477         if (clear_page_dirty_for_io(page)) {
1478                 ret = nfs_writepage_locked(page, &wbc);
1479                 if (ret < 0)
1480                         goto out;
1481         }
1482         if (!PagePrivate(page))
1483                 return 0;
1484         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1485         if (ret >= 0)
1486                 return 0;
1487 out:
1488         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1489         return ret;
1490 }
1491
1492 /*
1493  * Write back all requests on one page - we do this before reading it.
1494  */
1495 int nfs_wb_page(struct inode *inode, struct page* page)
1496 {
1497         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1498 }
1499
1500 int __init nfs_init_writepagecache(void)
1501 {
1502         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1503                                              sizeof(struct nfs_write_data),
1504                                              0, SLAB_HWCACHE_ALIGN,
1505                                              NULL);
1506         if (nfs_wdata_cachep == NULL)
1507                 return -ENOMEM;
1508
1509         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1510                                                      nfs_wdata_cachep);
1511         if (nfs_wdata_mempool == NULL)
1512                 return -ENOMEM;
1513
1514         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1515                                                       nfs_wdata_cachep);
1516         if (nfs_commit_mempool == NULL)
1517                 return -ENOMEM;
1518
1519         /*
1520          * NFS congestion size, scale with available memory.
1521          *
1522          *  64MB:    8192k
1523          * 128MB:   11585k
1524          * 256MB:   16384k
1525          * 512MB:   23170k
1526          *   1GB:   32768k
1527          *   2GB:   46340k
1528          *   4GB:   65536k
1529          *   8GB:   92681k
1530          *  16GB:  131072k
1531          *
1532          * This allows larger machines to have larger/more transfers.
1533          * Limit the default to 256M
1534          */
1535         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1536         if (nfs_congestion_kb > 256*1024)
1537                 nfs_congestion_kb = 256*1024;
1538
1539         return 0;
1540 }
1541
1542 void nfs_destroy_writepagecache(void)
1543 {
1544         mempool_destroy(nfs_commit_mempool);
1545         mempool_destroy(nfs_wdata_mempool);
1546         kmem_cache_destroy(nfs_wdata_cachep);
1547 }
1548