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         if (count != PAGE_CACHE_SIZE)
 178                 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
 179         SetPageUptodate(page);
 180 }
 181
 182 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 183                 unsigned int offset, unsigned int count)
 184 {
 185         struct nfs_page *req;
 186         int ret;
 187
 188         for (;;) {
 189                 req = nfs_update_request(ctx, page, offset, count);
 190                 if (!IS_ERR(req))
 191                         break;
 192                 ret = PTR_ERR(req);
 193                 if (ret != -EBUSY)
 194                         return ret;
 195                 ret = nfs_wb_page(page->mapping->host, page);
 196                 if (ret != 0)
 197                         return ret;
 198         }
 199         /* Update file length */
 200         nfs_grow_file(page, offset, count);
 201         nfs_unlock_request(req);
 202         return 0;
 203 }
 204
 205 static int wb_priority(struct writeback_control *wbc)
 206 {
 207         if (wbc->for_reclaim)
 208                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 209         if (wbc->for_kupdate)
 210                 return FLUSH_LOWPRI;
 211         return 0;
 212 }
 213
 214 /*
 215  * NFS congestion control
 216  */
 217
 218 int nfs_congestion_kb;
 219
 220 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 221 #define NFS_CONGESTION_OFF_THRESH       \
 222         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 223
 224 static int nfs_set_page_writeback(struct page *page)
 225 {
 226         int ret = test_set_page_writeback(page);
 227
 228         if (!ret) {
 229                 struct inode *inode = page->mapping->host;
 230                 struct nfs_server *nfss = NFS_SERVER(inode);
 231
 232                 if (atomic_long_inc_return(&nfss->writeback) >
 233                                 NFS_CONGESTION_ON_THRESH)
 234                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 235         }
 236         return ret;
 237 }
 238
 239 static void nfs_end_page_writeback(struct page *page)
 240 {
 241         struct inode *inode = page->mapping->host;
 242         struct nfs_server *nfss = NFS_SERVER(inode);
 243
 244         end_page_writeback(page);
 245         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 246                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 247                 congestion_end(WRITE);
 248         }
 249 }
 250
 251 /*
 252  * Find an associated nfs write request, and prepare to flush it out
 253  * May return an error if the user signalled nfs_wait_on_request().
 254  */
 255 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 256                                 struct page *page)
 257 {
 258         struct inode *inode = page->mapping->host;
 259         struct nfs_inode *nfsi = NFS_I(inode);
 260         struct nfs_page *req;
 261         int ret;
 262
 263         spin_lock(&inode->i_lock);
 264         for(;;) {
 265                 req = nfs_page_find_request_locked(page);
 266                 if (req == NULL) {
 267                         spin_unlock(&inode->i_lock);
 268                         return 0;
 269                 }
 270                 if (nfs_lock_request_dontget(req))
 271                         break;
 272                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 273                  *       then the call to nfs_lock_request_dontget() will always
 274                  *       succeed provided that someone hasn't already marked the
 275                  *       request as dirty (in which case we don't care).
 276                  */
 277                 spin_unlock(&inode->i_lock);
 278                 ret = nfs_wait_on_request(req);
 279                 nfs_release_request(req);
 280                 if (ret != 0)
 281                         return ret;
 282                 spin_lock(&inode->i_lock);
 283         }
 284         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 285                 /* This request is marked for commit */
 286                 spin_unlock(&inode->i_lock);
 287                 nfs_unlock_request(req);
 288                 nfs_pageio_complete(pgio);
 289                 return 0;
 290         }
 291         if (nfs_set_page_writeback(page) != 0) {
 292                 spin_unlock(&inode->i_lock);
 293                 BUG();
 294         }
 295         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 296                         NFS_PAGE_TAG_LOCKED);
 297         spin_unlock(&inode->i_lock);
 298         nfs_pageio_add_request(pgio, req);
 299         return 0;
 300 }
 301
 302 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 303 {
 304         struct inode *inode = page->mapping->host;
 305
 306         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 307         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 308
 309         nfs_pageio_cond_complete(pgio, page->index);
 310         return nfs_page_async_flush(pgio, page);
 311 }
 312
 313 /*
 314  * Write an mmapped page to the server.
 315  */
 316 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 317 {
 318         struct nfs_pageio_descriptor pgio;
 319         int err;
 320
 321         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 322         err = nfs_do_writepage(page, wbc, &pgio);
 323         nfs_pageio_complete(&pgio);
 324         if (err < 0)
 325                 return err;
 326         if (pgio.pg_error < 0)
 327                 return pgio.pg_error;
 328         return 0;
 329 }
 330
 331 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 332 {
 333         int ret;
 334
 335         ret = nfs_writepage_locked(page, wbc);
 336         unlock_page(page);
 337         return ret;
 338 }
 339
 340 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 341 {
 342         int ret;
 343
 344         ret = nfs_do_writepage(page, wbc, data);
 345         unlock_page(page);
 346         return ret;
 347 }
 348
 349 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 350 {
 351         struct inode *inode = mapping->host;
 352         struct nfs_pageio_descriptor pgio;
 353         int err;
 354
 355         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 356
 357         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 358         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 359         nfs_pageio_complete(&pgio);
 360         if (err < 0)
 361                 return err;
 362         if (pgio.pg_error < 0)
 363                 return pgio.pg_error;
 364         return 0;
 365 }
 366
 367 /*
 368  * Insert a write request into an inode
 369  */
 370 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 371 {
 372         struct nfs_inode *nfsi = NFS_I(inode);
 373         int error;
 374
 375         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 376         BUG_ON(error == -EEXIST);
 377         if (error)
 378                 return error;
 379         if (!nfsi->npages) {
 380                 igrab(inode);
 381                 nfs_begin_data_update(inode);
 382                 if (nfs_have_delegation(inode, FMODE_WRITE))
 383                         nfsi->change_attr++;
 384         }
 385         SetPagePrivate(req->wb_page);
 386         set_page_private(req->wb_page, (unsigned long)req);
 387         nfsi->npages++;
 388         kref_get(&req->wb_kref);
 389         return 0;
 390 }
 391
 392 /*
 393  * Remove a write request from an inode
 394  */
 395 static void nfs_inode_remove_request(struct nfs_page *req)
 396 {
 397         struct inode *inode = req->wb_context->path.dentry->d_inode;
 398         struct nfs_inode *nfsi = NFS_I(inode);
 399
 400         BUG_ON (!NFS_WBACK_BUSY(req));
 401
 402         spin_lock(&inode->i_lock);
 403         set_page_private(req->wb_page, 0);
 404         ClearPagePrivate(req->wb_page);
 405         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 406         nfsi->npages--;
 407         if (!nfsi->npages) {
 408                 spin_unlock(&inode->i_lock);
 409                 nfs_end_data_update(inode);
 410                 iput(inode);
 411         } else
 412                 spin_unlock(&inode->i_lock);
 413         nfs_clear_request(req);
 414         nfs_release_request(req);
 415 }
 416
 417 static void
 418 nfs_redirty_request(struct nfs_page *req)
 419 {
 420         __set_page_dirty_nobuffers(req->wb_page);
 421 }
 422
 423 /*
 424  * Check if a request is dirty
 425  */
 426 static inline int
 427 nfs_dirty_request(struct nfs_page *req)
 428 {
 429         struct page *page = req->wb_page;
 430
 431         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 432                 return 0;
 433         return !PageWriteback(req->wb_page);
 434 }
 435
 436 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 437 /*
 438  * Add a request to the inode's commit list.
 439  */
 440 static void
 441 nfs_mark_request_commit(struct nfs_page *req)
 442 {
 443         struct inode *inode = req->wb_context->path.dentry->d_inode;
 444         struct nfs_inode *nfsi = NFS_I(inode);
 445
 446         spin_lock(&inode->i_lock);
 447         nfsi->ncommit++;
 448         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 449         radix_tree_tag_set(&nfsi->nfs_page_tree,
 450                         req->wb_index,
 451                         NFS_PAGE_TAG_COMMIT);
 452         spin_unlock(&inode->i_lock);
 453         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 454         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 455 }
 456
 457 static inline
 458 int nfs_write_need_commit(struct nfs_write_data *data)
 459 {
 460         return data->verf.committed != NFS_FILE_SYNC;
 461 }
 462
 463 static inline
 464 int nfs_reschedule_unstable_write(struct nfs_page *req)
 465 {
 466         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 467                 nfs_mark_request_commit(req);
 468                 return 1;
 469         }
 470         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 471                 nfs_redirty_request(req);
 472                 return 1;
 473         }
 474         return 0;
 475 }
 476 #else
 477 static inline void
 478 nfs_mark_request_commit(struct nfs_page *req)
 479 {
 480 }
 481
 482 static inline
 483 int nfs_write_need_commit(struct nfs_write_data *data)
 484 {
 485         return 0;
 486 }
 487
 488 static inline
 489 int nfs_reschedule_unstable_write(struct nfs_page *req)
 490 {
 491         return 0;
 492 }
 493 #endif
 494
 495 /*
 496  * Wait for a request to complete.
 497  *
 498  * Interruptible by signals only if mounted with intr flag.
 499  */
 500 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 501 {
 502         struct nfs_inode *nfsi = NFS_I(inode);
 503         struct nfs_page *req;
 504         pgoff_t idx_end, next;
 505         unsigned int            res = 0;
 506         int                     error;
 507
 508         if (npages == 0)
 509                 idx_end = ~0;
 510         else
 511                 idx_end = idx_start + npages - 1;
 512
 513         next = idx_start;
 514         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 515                 if (req->wb_index > idx_end)
 516                         break;
 517
 518                 next = req->wb_index + 1;
 519                 BUG_ON(!NFS_WBACK_BUSY(req));
 520
 521                 kref_get(&req->wb_kref);
 522                 spin_unlock(&inode->i_lock);
 523                 error = nfs_wait_on_request(req);
 524                 nfs_release_request(req);
 525                 spin_lock(&inode->i_lock);
 526                 if (error < 0)
 527                         return error;
 528                 res++;
 529         }
 530         return res;
 531 }
 532
 533 static void nfs_cancel_commit_list(struct list_head *head)
 534 {
 535         struct nfs_page *req;
 536
 537         while(!list_empty(head)) {
 538                 req = nfs_list_entry(head->next);
 539                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 540                 nfs_list_remove_request(req);
 541                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 542                 nfs_inode_remove_request(req);
 543                 nfs_unlock_request(req);
 544         }
 545 }
 546
 547 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 548 /*
 549  * nfs_scan_commit - Scan an inode for commit requests
 550  * @inode: NFS inode to scan
 551  * @dst: destination list
 552  * @idx_start: lower bound of page->index to scan.
 553  * @npages: idx_start + npages sets the upper bound to scan.
 554  *
 555  * Moves requests from the inode's 'commit' request list.
 556  * The requests are *not* checked to ensure that they form a contiguous set.
 557  */
 558 static int
 559 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 560 {
 561         struct nfs_inode *nfsi = NFS_I(inode);
 562         int res = 0;
 563
 564         if (nfsi->ncommit != 0) {
 565                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 566                                 NFS_PAGE_TAG_COMMIT);
 567                 nfsi->ncommit -= res;
 568         }
 569         return res;
 570 }
 571 #else
 572 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 573 {
 574         return 0;
 575 }
 576 #endif
 577
 578 /*
 579  * Try to update any existing write request, or create one if there is none.
 580  * In order to match, the request's credentials must match those of
 581  * the calling process.
 582  *
 583  * Note: Should always be called with the Page Lock held!
 584  */
 585 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 586                 struct page *page, unsigned int offset, unsigned int bytes)
 587 {
 588         struct address_space *mapping = page->mapping;
 589         struct inode *inode = mapping->host;
 590         struct nfs_page         *req, *new = NULL;
 591         pgoff_t         rqend, end;
 592
 593         end = offset + bytes;
 594
 595         for (;;) {
 596                 /* Loop over all inode entries and see if we find
 597                  * A request for the page we wish to update
 598                  */
 599                 spin_lock(&inode->i_lock);
 600                 req = nfs_page_find_request_locked(page);
 601                 if (req) {
 602                         if (!nfs_lock_request_dontget(req)) {
 603                                 int error;
 604
 605                                 spin_unlock(&inode->i_lock);
 606                                 error = nfs_wait_on_request(req);
 607                                 nfs_release_request(req);
 608                                 if (error < 0) {
 609                                         if (new)
 610                                                 nfs_release_request(new);
 611                                         return ERR_PTR(error);
 612                                 }
 613                                 continue;
 614                         }
 615                         spin_unlock(&inode->i_lock);
 616                         if (new)
 617                                 nfs_release_request(new);
 618                         break;
 619                 }
 620
 621                 if (new) {
 622                         int error;
 623                         nfs_lock_request_dontget(new);
 624                         error = nfs_inode_add_request(inode, new);
 625                         if (error) {
 626                                 spin_unlock(&inode->i_lock);
 627                                 nfs_unlock_request(new);
 628                                 return ERR_PTR(error);
 629                         }
 630                         spin_unlock(&inode->i_lock);
 631                         return new;
 632                 }
 633                 spin_unlock(&inode->i_lock);
 634
 635                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 636                 if (IS_ERR(new))
 637                         return new;
 638         }
 639
 640         /* We have a request for our page.
 641          * If the creds don't match, or the
 642          * page addresses don't match,
 643          * tell the caller to wait on the conflicting
 644          * request.
 645          */
 646         rqend = req->wb_offset + req->wb_bytes;
 647         if (req->wb_context != ctx
 648             || req->wb_page != page
 649             || !nfs_dirty_request(req)
 650             || offset > rqend || end < req->wb_offset) {
 651                 nfs_unlock_request(req);
 652                 return ERR_PTR(-EBUSY);
 653         }
 654
 655         /* Okay, the request matches. Update the region */
 656         if (offset < req->wb_offset) {
 657                 req->wb_offset = offset;
 658                 req->wb_pgbase = offset;
 659                 req->wb_bytes = rqend - req->wb_offset;
 660         }
 661
 662         if (end > rqend)
 663                 req->wb_bytes = end - req->wb_offset;
 664
 665         return req;
 666 }
 667
 668 int nfs_flush_incompatible(struct file *file, struct page *page)
 669 {
 670         struct nfs_open_context *ctx = nfs_file_open_context(file);
 671         struct nfs_page *req;
 672         int do_flush, status;
 673         /*
 674          * Look for a request corresponding to this page. If there
 675          * is one, and it belongs to another file, we flush it out
 676          * before we try to copy anything into the page. Do this
 677          * due to the lack of an ACCESS-type call in NFSv2.
 678          * Also do the same if we find a request from an existing
 679          * dropped page.
 680          */
 681         do {
 682                 req = nfs_page_find_request(page);
 683                 if (req == NULL)
 684                         return 0;
 685                 do_flush = req->wb_page != page || req->wb_context != ctx
 686                         || !nfs_dirty_request(req);
 687                 nfs_release_request(req);
 688                 if (!do_flush)
 689                         return 0;
 690                 status = nfs_wb_page(page->mapping->host, page);
 691         } while (status == 0);
 692         return status;
 693 }
 694
 695 /*
 696  * Update and possibly write a cached page of an NFS file.
 697  *
 698  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 699  * things with a page scheduled for an RPC call (e.g. invalidate it).
 700  */
 701 int nfs_updatepage(struct file *file, struct page *page,
 702                 unsigned int offset, unsigned int count)
 703 {
 704         struct nfs_open_context *ctx = nfs_file_open_context(file);
 705         struct inode    *inode = page->mapping->host;
 706         int             status = 0;
 707
 708         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 709
 710         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 711                 file->f_path.dentry->d_parent->d_name.name,
 712                 file->f_path.dentry->d_name.name, count,
 713                 (long long)(page_offset(page) +offset));
 714
 715         /* If we're not using byte range locks, and we know the page
 716          * is entirely in cache, it may be more efficient to avoid
 717          * fragmenting write requests.
 718          */
 719         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 720                 count = max(count + offset, nfs_page_length(page));
 721                 offset = 0;
 722         }
 723
 724         status = nfs_writepage_setup(ctx, page, offset, count);
 725         __set_page_dirty_nobuffers(page);
 726
 727         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 728                         status, (long long)i_size_read(inode));
 729         if (status < 0)
 730                 nfs_set_pageerror(page);
 731         return status;
 732 }
 733
 734 static void nfs_writepage_release(struct nfs_page *req)
 735 {
 736
 737         if (PageError(req->wb_page)) {
 738                 nfs_end_page_writeback(req->wb_page);
 739                 nfs_inode_remove_request(req);
 740         } else if (!nfs_reschedule_unstable_write(req)) {
 741                 /* Set the PG_uptodate flag */
 742                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 743                 nfs_end_page_writeback(req->wb_page);
 744                 nfs_inode_remove_request(req);
 745         } else
 746                 nfs_end_page_writeback(req->wb_page);
 747         nfs_clear_page_tag_locked(req);
 748 }
 749
 750 static inline int flush_task_priority(int how)
 751 {
 752         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 753                 case FLUSH_HIGHPRI:
 754                         return RPC_PRIORITY_HIGH;
 755                 case FLUSH_LOWPRI:
 756                         return RPC_PRIORITY_LOW;
 757         }
 758         return RPC_PRIORITY_NORMAL;
 759 }
 760
 761 /*
 762  * Set up the argument/result storage required for the RPC call.
 763  */
 764 static void nfs_write_rpcsetup(struct nfs_page *req,
 765                 struct nfs_write_data *data,
 766                 const struct rpc_call_ops *call_ops,
 767                 unsigned int count, unsigned int offset,
 768                 int how)
 769 {
 770         struct inode            *inode;
 771         int flags;
 772
 773         /* Set up the RPC argument and reply structs
 774          * NB: take care not to mess about with data->commit et al. */
 775
 776         data->req = req;
 777         data->inode = inode = req->wb_context->path.dentry->d_inode;
 778         data->cred = req->wb_context->cred;
 779
 780         data->args.fh     = NFS_FH(inode);
 781         data->args.offset = req_offset(req) + offset;
 782         data->args.pgbase = req->wb_pgbase + offset;
 783         data->args.pages  = data->pagevec;
 784         data->args.count  = count;
 785         data->args.context = req->wb_context;
 786
 787         data->res.fattr   = &data->fattr;
 788         data->res.count   = count;
 789         data->res.verf    = &data->verf;
 790         nfs_fattr_init(&data->fattr);
 791
 792         /* Set up the initial task struct.  */
 793         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 794         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 795         NFS_PROTO(inode)->write_setup(data, how);
 796
 797         data->task.tk_priority = flush_task_priority(how);
 798         data->task.tk_cookie = (unsigned long)inode;
 799
 800         dprintk("NFS: %5u initiated write call "
 801                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 802                 data->task.tk_pid,
 803                 inode->i_sb->s_id,
 804                 (long long)NFS_FILEID(inode),
 805                 count,
 806                 (unsigned long long)data->args.offset);
 807 }
 808
 809 static void nfs_execute_write(struct nfs_write_data *data)
 810 {
 811         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 812         sigset_t oldset;
 813
 814         rpc_clnt_sigmask(clnt, &oldset);
 815         rpc_execute(&data->task);
 816         rpc_clnt_sigunmask(clnt, &oldset);
 817 }
 818
 819 /*
 820  * Generate multiple small requests to write out a single
 821  * contiguous dirty area on one page.
 822  */
 823 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 824 {
 825         struct nfs_page *req = nfs_list_entry(head->next);
 826         struct page *page = req->wb_page;
 827         struct nfs_write_data *data;
 828         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 829         unsigned int offset;
 830         int requests = 0;
 831         LIST_HEAD(list);
 832
 833         nfs_list_remove_request(req);
 834
 835         nbytes = count;
 836         do {
 837                 size_t len = min(nbytes, wsize);
 838
 839                 data = nfs_writedata_alloc(1);
 840                 if (!data)
 841                         goto out_bad;
 842                 list_add(&data->pages, &list);
 843                 requests++;
 844                 nbytes -= len;
 845         } while (nbytes != 0);
 846         atomic_set(&req->wb_complete, requests);
 847
 848         ClearPageError(page);
 849         offset = 0;
 850         nbytes = count;
 851         do {
 852                 data = list_entry(list.next, struct nfs_write_data, pages);
 853                 list_del_init(&data->pages);
 854
 855                 data->pagevec[0] = page;
 856
 857                 if (nbytes < wsize)
 858                         wsize = nbytes;
 859                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 860                                    wsize, offset, how);
 861                 offset += wsize;
 862                 nbytes -= wsize;
 863                 nfs_execute_write(data);
 864         } while (nbytes != 0);
 865
 866         return 0;
 867
 868 out_bad:
 869         while (!list_empty(&list)) {
 870                 data = list_entry(list.next, struct nfs_write_data, pages);
 871                 list_del(&data->pages);
 872                 nfs_writedata_release(data);
 873         }
 874         nfs_redirty_request(req);
 875         nfs_end_page_writeback(req->wb_page);
 876         nfs_clear_page_tag_locked(req);
 877         return -ENOMEM;
 878 }
 879
 880 /*
 881  * Create an RPC task for the given write request and kick it.
 882  * The page must have been locked by the caller.
 883  *
 884  * It may happen that the page we're passed is not marked dirty.
 885  * This is the case if nfs_updatepage detects a conflicting request
 886  * that has been written but not committed.
 887  */
 888 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 889 {
 890         struct nfs_page         *req;
 891         struct page             **pages;
 892         struct nfs_write_data   *data;
 893
 894         data = nfs_writedata_alloc(npages);
 895         if (!data)
 896                 goto out_bad;
 897
 898         pages = data->pagevec;
 899         while (!list_empty(head)) {
 900                 req = nfs_list_entry(head->next);
 901                 nfs_list_remove_request(req);
 902                 nfs_list_add_request(req, &data->pages);
 903                 ClearPageError(req->wb_page);
 904                 *pages++ = req->wb_page;
 905         }
 906         req = nfs_list_entry(data->pages.next);
 907
 908         /* Set up the argument struct */
 909         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 910
 911         nfs_execute_write(data);
 912         return 0;
 913  out_bad:
 914         while (!list_empty(head)) {
 915                 req = nfs_list_entry(head->next);
 916                 nfs_list_remove_request(req);
 917                 nfs_redirty_request(req);
 918                 nfs_end_page_writeback(req->wb_page);
 919                 nfs_clear_page_tag_locked(req);
 920         }
 921         return -ENOMEM;
 922 }
 923
 924 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 925                                   struct inode *inode, int ioflags)
 926 {
 927         int wsize = NFS_SERVER(inode)->wsize;
 928
 929         if (wsize < PAGE_CACHE_SIZE)
 930                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 931         else
 932                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 933 }
 934
 935 /*
 936  * Handle a write reply that flushed part of a page.
 937  */
 938 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 939 {
 940         struct nfs_write_data   *data = calldata;
 941         struct nfs_page         *req = data->req;
 942         struct page             *page = req->wb_page;
 943
 944         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 945                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 946                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 947                 req->wb_bytes,
 948                 (long long)req_offset(req));
 949
 950         if (nfs_writeback_done(task, data) != 0)
 951                 return;
 952
 953         if (task->tk_status < 0) {
 954                 nfs_set_pageerror(page);
 955                 nfs_context_set_write_error(req->wb_context, task->tk_status);
 956                 dprintk(", error = %d\n", task->tk_status);
 957                 goto out;
 958         }
 959
 960         if (nfs_write_need_commit(data)) {
 961                 struct inode *inode = page->mapping->host;
 962
 963                 spin_lock(&inode->i_lock);
 964                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 965                         /* Do nothing we need to resend the writes */
 966                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 967                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 968                         dprintk(" defer commit\n");
 969                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 970                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 971                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 972                         dprintk(" server reboot detected\n");
 973                 }
 974                 spin_unlock(&inode->i_lock);
 975         } else
 976                 dprintk(" OK\n");
 977
 978 out:
 979         if (atomic_dec_and_test(&req->wb_complete))
 980                 nfs_writepage_release(req);
 981 }
 982
 983 static const struct rpc_call_ops nfs_write_partial_ops = {
 984         .rpc_call_done = nfs_writeback_done_partial,
 985         .rpc_release = nfs_writedata_release,
 986 };
 987
 988 /*
 989  * Handle a write reply that flushes a whole page.
 990  *
 991  * FIXME: There is an inherent race with invalidate_inode_pages and
 992  *        writebacks since the page->count is kept > 1 for as long
 993  *        as the page has a write request pending.
 994  */
 995 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
 996 {
 997         struct nfs_write_data   *data = calldata;
 998         struct nfs_page         *req;
 999         struct page             *page;
1000
1001         if (nfs_writeback_done(task, data) != 0)
1002                 return;
1003
1004         /* Update attributes as result of writeback. */
1005         while (!list_empty(&data->pages)) {
1006                 req = nfs_list_entry(data->pages.next);
1007                 nfs_list_remove_request(req);
1008                 page = req->wb_page;
1009
1010                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1011                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1012                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1013                         req->wb_bytes,
1014                         (long long)req_offset(req));
1015
1016                 if (task->tk_status < 0) {
1017                         nfs_set_pageerror(page);
1018                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1019                         dprintk(", error = %d\n", task->tk_status);
1020                         goto remove_request;
1021                 }
1022
1023                 if (nfs_write_need_commit(data)) {
1024                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1025                         nfs_mark_request_commit(req);
1026                         nfs_end_page_writeback(page);
1027                         dprintk(" marked for commit\n");
1028                         goto next;
1029                 }
1030                 /* Set the PG_uptodate flag? */
1031                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1032                 dprintk(" OK\n");
1033 remove_request:
1034                 nfs_end_page_writeback(page);
1035                 nfs_inode_remove_request(req);
1036         next:
1037                 nfs_clear_page_tag_locked(req);
1038         }
1039 }
1040
1041 static const struct rpc_call_ops nfs_write_full_ops = {
1042         .rpc_call_done = nfs_writeback_done_full,
1043         .rpc_release = nfs_writedata_release,
1044 };
1045
1046
1047 /*
1048  * This function is called when the WRITE call is complete.
1049  */
1050 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1051 {
1052         struct nfs_writeargs    *argp = &data->args;
1053         struct nfs_writeres     *resp = &data->res;
1054         int status;
1055
1056         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1057                 task->tk_pid, task->tk_status);
1058
1059         /*
1060          * ->write_done will attempt to use post-op attributes to detect
1061          * conflicting writes by other clients.  A strict interpretation
1062          * of close-to-open would allow us to continue caching even if
1063          * another writer had changed the file, but some applications
1064          * depend on tighter cache coherency when writing.
1065          */
1066         status = NFS_PROTO(data->inode)->write_done(task, data);
1067         if (status != 0)
1068                 return status;
1069         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1070
1071 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1072         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1073                 /* We tried a write call, but the server did not
1074                  * commit data to stable storage even though we
1075                  * requested it.
1076                  * Note: There is a known bug in Tru64 < 5.0 in which
1077                  *       the server reports NFS_DATA_SYNC, but performs
1078                  *       NFS_FILE_SYNC. We therefore implement this checking
1079                  *       as a dprintk() in order to avoid filling syslog.
1080                  */
1081                 static unsigned long    complain;
1082
1083                 if (time_before(complain, jiffies)) {
1084                         dprintk("NFS: faulty NFS server %s:"
1085                                 " (committed = %d) != (stable = %d)\n",
1086                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1087                                 resp->verf->committed, argp->stable);
1088                         complain = jiffies + 300 * HZ;
1089                 }
1090         }
1091 #endif
1092         /* Is this a short write? */
1093         if (task->tk_status >= 0 && resp->count < argp->count) {
1094                 static unsigned long    complain;
1095
1096                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1097
1098                 /* Has the server at least made some progress? */
1099                 if (resp->count != 0) {
1100                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1101                         if (resp->verf->committed != NFS_UNSTABLE) {
1102                                 /* Resend from where the server left off */
1103                                 argp->offset += resp->count;
1104                                 argp->pgbase += resp->count;
1105                                 argp->count -= resp->count;
1106                         } else {
1107                                 /* Resend as a stable write in order to avoid
1108                                  * headaches in the case of a server crash.
1109                                  */
1110                                 argp->stable = NFS_FILE_SYNC;
1111                         }
1112                         rpc_restart_call(task);
1113                         return -EAGAIN;
1114                 }
1115                 if (time_before(complain, jiffies)) {
1116                         printk(KERN_WARNING
1117                                "NFS: Server wrote zero bytes, expected %u.\n",
1118                                         argp->count);
1119                         complain = jiffies + 300 * HZ;
1120                 }
1121                 /* Can't do anything about it except throw an error. */
1122                 task->tk_status = -EIO;
1123         }
1124         return 0;
1125 }
1126
1127
1128 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1129 void nfs_commit_release(void *wdata)
1130 {
1131         nfs_commit_free(wdata);
1132 }
1133
1134 /*
1135  * Set up the argument/result storage required for the RPC call.
1136  */
1137 static void nfs_commit_rpcsetup(struct list_head *head,
1138                 struct nfs_write_data *data,
1139                 int how)
1140 {
1141         struct nfs_page         *first;
1142         struct inode            *inode;
1143         int flags;
1144
1145         /* Set up the RPC argument and reply structs
1146          * NB: take care not to mess about with data->commit et al. */
1147
1148         list_splice_init(head, &data->pages);
1149         first = nfs_list_entry(data->pages.next);
1150         inode = first->wb_context->path.dentry->d_inode;
1151
1152         data->inode       = inode;
1153         data->cred        = first->wb_context->cred;
1154
1155         data->args.fh     = NFS_FH(data->inode);
1156         /* Note: we always request a commit of the entire inode */
1157         data->args.offset = 0;
1158         data->args.count  = 0;
1159         data->res.count   = 0;
1160         data->res.fattr   = &data->fattr;
1161         data->res.verf    = &data->verf;
1162         nfs_fattr_init(&data->fattr);
1163
1164         /* Set up the initial task struct.  */
1165         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1166         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1167         NFS_PROTO(inode)->commit_setup(data, how);
1168
1169         data->task.tk_priority = flush_task_priority(how);
1170         data->task.tk_cookie = (unsigned long)inode;
1171
1172         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1173 }
1174
1175 /*
1176  * Commit dirty pages
1177  */
1178 static int
1179 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1180 {
1181         struct nfs_write_data   *data;
1182         struct nfs_page         *req;
1183
1184         data = nfs_commit_alloc();
1185
1186         if (!data)
1187                 goto out_bad;
1188
1189         /* Set up the argument struct */
1190         nfs_commit_rpcsetup(head, data, how);
1191
1192         nfs_execute_write(data);
1193         return 0;
1194  out_bad:
1195         while (!list_empty(head)) {
1196                 req = nfs_list_entry(head->next);
1197                 nfs_list_remove_request(req);
1198                 nfs_mark_request_commit(req);
1199                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1200                 nfs_clear_page_tag_locked(req);
1201         }
1202         return -ENOMEM;
1203 }
1204
1205 /*
1206  * COMMIT call returned
1207  */
1208 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1209 {
1210         struct nfs_write_data   *data = calldata;
1211         struct nfs_page         *req;
1212
1213         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1214                                 task->tk_pid, task->tk_status);
1215
1216         /* Call the NFS version-specific code */
1217         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1218                 return;
1219
1220         while (!list_empty(&data->pages)) {
1221                 req = nfs_list_entry(data->pages.next);
1222                 nfs_list_remove_request(req);
1223                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1224                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1225
1226                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1227                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1228                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1229                         req->wb_bytes,
1230                         (long long)req_offset(req));
1231                 if (task->tk_status < 0) {
1232                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1233                         nfs_inode_remove_request(req);
1234                         dprintk(", error = %d\n", task->tk_status);
1235                         goto next;
1236                 }
1237
1238                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1239                  * returned by the server against all stored verfs. */
1240                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1241                         /* We have a match */
1242                         /* Set the PG_uptodate flag */
1243                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1244                                         req->wb_bytes);
1245                         nfs_inode_remove_request(req);
1246                         dprintk(" OK\n");
1247                         goto next;
1248                 }
1249                 /* We have a mismatch. Write the page again */
1250                 dprintk(" mismatch\n");
1251                 nfs_redirty_request(req);
1252         next:
1253                 nfs_clear_page_tag_locked(req);
1254         }
1255 }
1256
1257 static const struct rpc_call_ops nfs_commit_ops = {
1258         .rpc_call_done = nfs_commit_done,
1259         .rpc_release = nfs_commit_release,
1260 };
1261
1262 int nfs_commit_inode(struct inode *inode, int how)
1263 {
1264         LIST_HEAD(head);
1265         int res;
1266
1267         spin_lock(&inode->i_lock);
1268         res = nfs_scan_commit(inode, &head, 0, 0);
1269         spin_unlock(&inode->i_lock);
1270         if (res) {
1271                 int error = nfs_commit_list(inode, &head, how);
1272                 if (error < 0)
1273                         return error;
1274         }
1275         return res;
1276 }
1277 #else
1278 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1279 {
1280         return 0;
1281 }
1282 #endif
1283
1284 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1285 {
1286         struct inode *inode = mapping->host;
1287         pgoff_t idx_start, idx_end;
1288         unsigned int npages = 0;
1289         LIST_HEAD(head);
1290         int nocommit = how & FLUSH_NOCOMMIT;
1291         long pages, ret;
1292
1293         /* FIXME */
1294         if (wbc->range_cyclic)
1295                 idx_start = 0;
1296         else {
1297                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1298                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1299                 if (idx_end > idx_start) {
1300                         pgoff_t l_npages = 1 + idx_end - idx_start;
1301                         npages = l_npages;
1302                         if (sizeof(npages) != sizeof(l_npages) &&
1303                                         (pgoff_t)npages != l_npages)
1304                                 npages = 0;
1305                 }
1306         }
1307         how &= ~FLUSH_NOCOMMIT;
1308         spin_lock(&inode->i_lock);
1309         do {
1310                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1311                 if (ret != 0)
1312                         continue;
1313                 if (nocommit)
1314                         break;
1315                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1316                 if (pages == 0)
1317                         break;
1318                 if (how & FLUSH_INVALIDATE) {
1319                         spin_unlock(&inode->i_lock);
1320                         nfs_cancel_commit_list(&head);
1321                         ret = pages;
1322                         spin_lock(&inode->i_lock);
1323                         continue;
1324                 }
1325                 pages += nfs_scan_commit(inode, &head, 0, 0);
1326                 spin_unlock(&inode->i_lock);
1327                 ret = nfs_commit_list(inode, &head, how);
1328                 spin_lock(&inode->i_lock);
1329
1330         } while (ret >= 0);
1331         spin_unlock(&inode->i_lock);
1332         return ret;
1333 }
1334
1335 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1336 {
1337         int ret;
1338
1339         ret = nfs_writepages(mapping, wbc);
1340         if (ret < 0)
1341                 goto out;
1342         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1343         if (ret < 0)
1344                 goto out;
1345         return 0;
1346 out:
1347         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1348         return ret;
1349 }
1350
1351 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1352 static int nfs_write_mapping(struct address_space *mapping, int how)
1353 {
1354         struct writeback_control wbc = {
1355                 .bdi = mapping->backing_dev_info,
1356                 .sync_mode = WB_SYNC_NONE,
1357                 .nr_to_write = LONG_MAX,
1358                 .for_writepages = 1,
1359                 .range_cyclic = 1,
1360         };
1361         int ret;
1362
1363         ret = __nfs_write_mapping(mapping, &wbc, how);
1364         if (ret < 0)
1365                 return ret;
1366         wbc.sync_mode = WB_SYNC_ALL;
1367         return __nfs_write_mapping(mapping, &wbc, how);
1368 }
1369
1370 /*
1371  * flush the inode to disk.
1372  */
1373 int nfs_wb_all(struct inode *inode)
1374 {
1375         return nfs_write_mapping(inode->i_mapping, 0);
1376 }
1377
1378 int nfs_wb_nocommit(struct inode *inode)
1379 {
1380         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1381 }
1382
1383 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1384 {
1385         struct nfs_page *req;
1386         loff_t range_start = page_offset(page);
1387         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1388         struct writeback_control wbc = {
1389                 .bdi = page->mapping->backing_dev_info,
1390                 .sync_mode = WB_SYNC_ALL,
1391                 .nr_to_write = LONG_MAX,
1392                 .range_start = range_start,
1393                 .range_end = range_end,
1394         };
1395         int ret = 0;
1396
1397         BUG_ON(!PageLocked(page));
1398         for (;;) {
1399                 req = nfs_page_find_request(page);
1400                 if (req == NULL)
1401                         goto out;
1402                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1403                         nfs_release_request(req);
1404                         break;
1405                 }
1406                 if (nfs_lock_request_dontget(req)) {
1407                         nfs_inode_remove_request(req);
1408                         /*
1409                          * In case nfs_inode_remove_request has marked the
1410                          * page as being dirty
1411                          */
1412                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1413                         nfs_unlock_request(req);
1414                         break;
1415                 }
1416                 ret = nfs_wait_on_request(req);
1417                 if (ret < 0)
1418                         goto out;
1419         }
1420         if (!PagePrivate(page))
1421                 return 0;
1422         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1423 out:
1424         return ret;
1425 }
1426
1427 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1428 {
1429         loff_t range_start = page_offset(page);
1430         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1431         struct writeback_control wbc = {
1432                 .bdi = page->mapping->backing_dev_info,
1433                 .sync_mode = WB_SYNC_ALL,
1434                 .nr_to_write = LONG_MAX,
1435                 .range_start = range_start,
1436                 .range_end = range_end,
1437         };
1438         int ret;
1439
1440         BUG_ON(!PageLocked(page));
1441         if (clear_page_dirty_for_io(page)) {
1442                 ret = nfs_writepage_locked(page, &wbc);
1443                 if (ret < 0)
1444                         goto out;
1445         }
1446         if (!PagePrivate(page))
1447                 return 0;
1448         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1449         if (ret >= 0)
1450                 return 0;
1451 out:
1452         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1453         return ret;
1454 }
1455
1456 /*
1457  * Write back all requests on one page - we do this before reading it.
1458  */
1459 int nfs_wb_page(struct inode *inode, struct page* page)
1460 {
1461         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1462 }
1463
1464 int __init nfs_init_writepagecache(void)
1465 {
1466         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1467                                              sizeof(struct nfs_write_data),
1468                                              0, SLAB_HWCACHE_ALIGN,
1469                                              NULL);
1470         if (nfs_wdata_cachep == NULL)
1471                 return -ENOMEM;
1472
1473         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1474                                                      nfs_wdata_cachep);
1475         if (nfs_wdata_mempool == NULL)
1476                 return -ENOMEM;
1477
1478         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1479                                                       nfs_wdata_cachep);
1480         if (nfs_commit_mempool == NULL)
1481                 return -ENOMEM;
1482
1483         /*
1484          * NFS congestion size, scale with available memory.
1485          *
1486          *  64MB:    8192k
1487          * 128MB:   11585k
1488          * 256MB:   16384k
1489          * 512MB:   23170k
1490          *   1GB:   32768k
1491          *   2GB:   46340k
1492          *   4GB:   65536k
1493          *   8GB:   92681k
1494          *  16GB:  131072k
1495          *
1496          * This allows larger machines to have larger/more transfers.
1497          * Limit the default to 256M
1498          */
1499         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1500         if (nfs_congestion_kb > 256*1024)
1501                 nfs_congestion_kb = 256*1024;
1502
1503         return 0;
1504 }
1505
1506 void nfs_destroy_writepagecache(void)
1507 {
1508         mempool_destroy(nfs_commit_mempool);
1509         mempool_destroy(nfs_wdata_mempool);
1510         kmem_cache_destroy(nfs_wdata_cachep);
1511 }
1512