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 #include <linux/smp_lock.h>
  25
  26 #include "delegation.h"
  27 #include "internal.h"
  28 #include "iostat.h"
  29
  30 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  31
  32 #define MIN_POOL_WRITE          (32)
  33 #define MIN_POOL_COMMIT         (4)
  34
  35 /*
  36  * Local function declarations
  37  */
  38 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  39                                             struct page *,
  40                                             unsigned int, unsigned int);
  41 static void nfs_mark_request_dirty(struct nfs_page *req);
  42 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  43 static const struct rpc_call_ops nfs_write_partial_ops;
  44 static const struct rpc_call_ops nfs_write_full_ops;
  45 static const struct rpc_call_ops nfs_commit_ops;
  46
  47 static struct kmem_cache *nfs_wdata_cachep;
  48 static mempool_t *nfs_wdata_mempool;
  49 static mempool_t *nfs_commit_mempool;
  50
  51 struct nfs_write_data *nfs_commit_alloc(void)
  52 {
  53         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  54
  55         if (p) {
  56                 memset(p, 0, sizeof(*p));
  57                 INIT_LIST_HEAD(&p->pages);
  58         }
  59         return p;
  60 }
  61
  62 void nfs_commit_rcu_free(struct rcu_head *head)
  63 {
  64         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  65         if (p && (p->pagevec != &p->page_array[0]))
  66                 kfree(p->pagevec);
  67         mempool_free(p, nfs_commit_mempool);
  68 }
  69
  70 void nfs_commit_free(struct nfs_write_data *wdata)
  71 {
  72         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  73 }
  74
  75 struct nfs_write_data *nfs_writedata_alloc(size_t len)
  76 {
  77         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
  78         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  79
  80         if (p) {
  81                 memset(p, 0, sizeof(*p));
  82                 INIT_LIST_HEAD(&p->pages);
  83                 p->npages = pagecount;
  84                 if (pagecount <= ARRAY_SIZE(p->page_array))
  85                         p->pagevec = p->page_array;
  86                 else {
  87                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  88                         if (!p->pagevec) {
  89                                 mempool_free(p, nfs_wdata_mempool);
  90                                 p = NULL;
  91                         }
  92                 }
  93         }
  94         return p;
  95 }
  96
  97 static void nfs_writedata_rcu_free(struct rcu_head *head)
  98 {
  99         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
 100         if (p && (p->pagevec != &p->page_array[0]))
 101                 kfree(p->pagevec);
 102         mempool_free(p, nfs_wdata_mempool);
 103 }
 104
 105 static void nfs_writedata_free(struct nfs_write_data *wdata)
 106 {
 107         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 108 }
 109
 110 void nfs_writedata_release(void *wdata)
 111 {
 112         nfs_writedata_free(wdata);
 113 }
 114
 115 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 116 {
 117         struct nfs_page *req = NULL;
 118
 119         if (PagePrivate(page)) {
 120                 req = (struct nfs_page *)page_private(page);
 121                 if (req != NULL)
 122                         atomic_inc(&req->wb_count);
 123         }
 124         return req;
 125 }
 126
 127 static struct nfs_page *nfs_page_find_request(struct page *page)
 128 {
 129         struct nfs_page *req = NULL;
 130         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 131
 132         spin_lock(req_lock);
 133         req = nfs_page_find_request_locked(page);
 134         spin_unlock(req_lock);
 135         return req;
 136 }
 137
 138 /* Adjust the file length if we're writing beyond the end */
 139 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 140 {
 141         struct inode *inode = page->mapping->host;
 142         loff_t end, i_size = i_size_read(inode);
 143         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 144
 145         if (i_size > 0 && page->index < end_index)
 146                 return;
 147         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 148         if (i_size >= end)
 149                 return;
 150         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 151         i_size_write(inode, end);
 152 }
 153
 154 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 155 static void nfs_set_pageerror(struct page *page)
 156 {
 157         SetPageError(page);
 158         nfs_zap_mapping(page->mapping->host, page->mapping);
 159 }
 160
 161 /* We can set the PG_uptodate flag if we see that a write request
 162  * covers the full page.
 163  */
 164 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 165 {
 166         if (PageUptodate(page))
 167                 return;
 168         if (base != 0)
 169                 return;
 170         if (count != nfs_page_length(page))
 171                 return;
 172         if (count != PAGE_CACHE_SIZE)
 173                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 174         SetPageUptodate(page);
 175 }
 176
 177 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 178                 unsigned int offset, unsigned int count)
 179 {
 180         struct nfs_page *req;
 181         int ret;
 182
 183         for (;;) {
 184                 req = nfs_update_request(ctx, page, offset, count);
 185                 if (!IS_ERR(req))
 186                         break;
 187                 ret = PTR_ERR(req);
 188                 if (ret != -EBUSY)
 189                         return ret;
 190                 ret = nfs_wb_page(page->mapping->host, page);
 191                 if (ret != 0)
 192                         return ret;
 193         }
 194         /* Update file length */
 195         nfs_grow_file(page, offset, count);
 196         /* Set the PG_uptodate flag? */
 197         nfs_mark_uptodate(page, offset, count);
 198         nfs_unlock_request(req);
 199         return 0;
 200 }
 201
 202 static int wb_priority(struct writeback_control *wbc)
 203 {
 204         if (wbc->for_reclaim)
 205                 return FLUSH_HIGHPRI;
 206         if (wbc->for_kupdate)
 207                 return FLUSH_LOWPRI;
 208         return 0;
 209 }
 210
 211 /*
 212  * NFS congestion control
 213  */
 214
 215 int nfs_congestion_kb;
 216
 217 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 218 #define NFS_CONGESTION_OFF_THRESH       \
 219         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 220
 221 static void nfs_set_page_writeback(struct page *page)
 222 {
 223         if (!test_set_page_writeback(page)) {
 224                 struct inode *inode = page->mapping->host;
 225                 struct nfs_server *nfss = NFS_SERVER(inode);
 226
 227                 if (atomic_inc_return(&nfss->writeback) >
 228                                 NFS_CONGESTION_ON_THRESH)
 229                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 230         }
 231 }
 232
 233 static void nfs_end_page_writeback(struct page *page)
 234 {
 235         struct inode *inode = page->mapping->host;
 236         struct nfs_server *nfss = NFS_SERVER(inode);
 237
 238         end_page_writeback(page);
 239         if (atomic_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 240                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 241                 congestion_end(WRITE);
 242         }
 243 }
 244
 245 /*
 246  * Find an associated nfs write request, and prepare to flush it out
 247  * Returns 1 if there was no write request, or if the request was
 248  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 249  * was not tagged.
 250  * May also return an error if the user signalled nfs_wait_on_request().
 251  */
 252 static int nfs_page_mark_flush(struct page *page)
 253 {
 254         struct nfs_page *req;
 255         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 256         int ret;
 257
 258         spin_lock(req_lock);
 259         for(;;) {
 260                 req = nfs_page_find_request_locked(page);
 261                 if (req == NULL) {
 262                         spin_unlock(req_lock);
 263                         return 1;
 264                 }
 265                 if (nfs_lock_request_dontget(req))
 266                         break;
 267                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 268                  *       then the call to nfs_lock_request_dontget() 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(req_lock);
 273                 ret = nfs_wait_on_request(req);
 274                 nfs_release_request(req);
 275                 if (ret != 0)
 276                         return ret;
 277                 spin_lock(req_lock);
 278         }
 279         spin_unlock(req_lock);
 280         if (test_and_set_bit(PG_FLUSHING, &req->wb_flags) == 0) {
 281                 nfs_mark_request_dirty(req);
 282                 nfs_set_page_writeback(page);
 283         }
 284         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 285         nfs_unlock_request(req);
 286         return ret;
 287 }
 288
 289 /*
 290  * Write an mmapped page to the server.
 291  */
 292 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 293 {
 294         struct nfs_open_context *ctx;
 295         struct inode *inode = page->mapping->host;
 296         unsigned offset;
 297         int err;
 298
 299         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 300         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 301
 302         err = nfs_page_mark_flush(page);
 303         if (err <= 0)
 304                 goto out;
 305         err = 0;
 306         offset = nfs_page_length(page);
 307         if (!offset)
 308                 goto out;
 309
 310         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 311         if (ctx == NULL) {
 312                 err = -EBADF;
 313                 goto out;
 314         }
 315         err = nfs_writepage_setup(ctx, page, 0, offset);
 316         put_nfs_open_context(ctx);
 317         if (err != 0)
 318                 goto out;
 319         err = nfs_page_mark_flush(page);
 320         if (err > 0)
 321                 err = 0;
 322 out:
 323         if (!wbc->for_writepages)
 324                 nfs_flush_mapping(page->mapping, wbc, FLUSH_STABLE|wb_priority(wbc));
 325         return err;
 326 }
 327
 328 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 329 {
 330         int err;
 331
 332         err = nfs_writepage_locked(page, wbc);
 333         unlock_page(page);
 334         return err;
 335 }
 336
 337 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 338 {
 339         struct inode *inode = mapping->host;
 340         int err;
 341
 342         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 343
 344         err = generic_writepages(mapping, wbc);
 345         if (err)
 346                 return err;
 347         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 348         if (err < 0)
 349                 goto out;
 350         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 351         err = 0;
 352 out:
 353         return err;
 354 }
 355
 356 /*
 357  * Insert a write request into an inode
 358  */
 359 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 360 {
 361         struct nfs_inode *nfsi = NFS_I(inode);
 362         int error;
 363
 364         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 365         BUG_ON(error == -EEXIST);
 366         if (error)
 367                 return error;
 368         if (!nfsi->npages) {
 369                 igrab(inode);
 370                 nfs_begin_data_update(inode);
 371                 if (nfs_have_delegation(inode, FMODE_WRITE))
 372                         nfsi->change_attr++;
 373         }
 374         SetPagePrivate(req->wb_page);
 375         set_page_private(req->wb_page, (unsigned long)req);
 376         nfsi->npages++;
 377         atomic_inc(&req->wb_count);
 378         return 0;
 379 }
 380
 381 /*
 382  * Remove a write request from an inode
 383  */
 384 static void nfs_inode_remove_request(struct nfs_page *req)
 385 {
 386         struct inode *inode = req->wb_context->dentry->d_inode;
 387         struct nfs_inode *nfsi = NFS_I(inode);
 388
 389         BUG_ON (!NFS_WBACK_BUSY(req));
 390
 391         spin_lock(&nfsi->req_lock);
 392         set_page_private(req->wb_page, 0);
 393         ClearPagePrivate(req->wb_page);
 394         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 395         nfsi->npages--;
 396         if (!nfsi->npages) {
 397                 spin_unlock(&nfsi->req_lock);
 398                 nfs_end_data_update(inode);
 399                 iput(inode);
 400         } else
 401                 spin_unlock(&nfsi->req_lock);
 402         nfs_clear_request(req);
 403         nfs_release_request(req);
 404 }
 405
 406 /*
 407  * Add a request to the inode's dirty list.
 408  */
 409 static void
 410 nfs_mark_request_dirty(struct nfs_page *req)
 411 {
 412         struct inode *inode = req->wb_context->dentry->d_inode;
 413         struct nfs_inode *nfsi = NFS_I(inode);
 414
 415         spin_lock(&nfsi->req_lock);
 416         radix_tree_tag_set(&nfsi->nfs_page_tree,
 417                         req->wb_index, NFS_PAGE_TAG_DIRTY);
 418         nfs_list_add_request(req, &nfsi->dirty);
 419         nfsi->ndirty++;
 420         spin_unlock(&nfsi->req_lock);
 421         __mark_inode_dirty(inode, I_DIRTY_PAGES);
 422 }
 423
 424 static void
 425 nfs_redirty_request(struct nfs_page *req)
 426 {
 427         clear_bit(PG_FLUSHING, &req->wb_flags);
 428         __set_page_dirty_nobuffers(req->wb_page);
 429 }
 430
 431 /*
 432  * Check if a request is dirty
 433  */
 434 static inline int
 435 nfs_dirty_request(struct nfs_page *req)
 436 {
 437         return test_bit(PG_FLUSHING, &req->wb_flags) == 0;
 438 }
 439
 440 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 441 /*
 442  * Add a request to the inode's commit list.
 443  */
 444 static void
 445 nfs_mark_request_commit(struct nfs_page *req)
 446 {
 447         struct inode *inode = req->wb_context->dentry->d_inode;
 448         struct nfs_inode *nfsi = NFS_I(inode);
 449
 450         spin_lock(&nfsi->req_lock);
 451         nfs_list_add_request(req, &nfsi->commit);
 452         nfsi->ncommit++;
 453         spin_unlock(&nfsi->req_lock);
 454         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 455         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 456 }
 457 #endif
 458
 459 /*
 460  * Wait for a request to complete.
 461  *
 462  * Interruptible by signals only if mounted with intr flag.
 463  */
 464 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 465 {
 466         struct nfs_inode *nfsi = NFS_I(inode);
 467         struct nfs_page *req;
 468         unsigned long           idx_end, next;
 469         unsigned int            res = 0;
 470         int                     error;
 471
 472         if (npages == 0)
 473                 idx_end = ~0;
 474         else
 475                 idx_end = idx_start + npages - 1;
 476
 477         next = idx_start;
 478         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 479                 if (req->wb_index > idx_end)
 480                         break;
 481
 482                 next = req->wb_index + 1;
 483                 BUG_ON(!NFS_WBACK_BUSY(req));
 484
 485                 atomic_inc(&req->wb_count);
 486                 spin_unlock(&nfsi->req_lock);
 487                 error = nfs_wait_on_request(req);
 488                 nfs_release_request(req);
 489                 spin_lock(&nfsi->req_lock);
 490                 if (error < 0)
 491                         return error;
 492                 res++;
 493         }
 494         return res;
 495 }
 496
 497 static void nfs_cancel_dirty_list(struct list_head *head)
 498 {
 499         struct nfs_page *req;
 500         while(!list_empty(head)) {
 501                 req = nfs_list_entry(head->next);
 502                 nfs_list_remove_request(req);
 503                 nfs_inode_remove_request(req);
 504                 nfs_clear_page_writeback(req);
 505         }
 506 }
 507
 508 static void nfs_cancel_commit_list(struct list_head *head)
 509 {
 510         struct nfs_page *req;
 511
 512         while(!list_empty(head)) {
 513                 req = nfs_list_entry(head->next);
 514                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 515                 nfs_list_remove_request(req);
 516                 nfs_inode_remove_request(req);
 517                 nfs_unlock_request(req);
 518         }
 519 }
 520
 521 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 522 /*
 523  * nfs_scan_commit - Scan an inode for commit requests
 524  * @inode: NFS inode to scan
 525  * @dst: destination list
 526  * @idx_start: lower bound of page->index to scan.
 527  * @npages: idx_start + npages sets the upper bound to scan.
 528  *
 529  * Moves requests from the inode's 'commit' request list.
 530  * The requests are *not* checked to ensure that they form a contiguous set.
 531  */
 532 static int
 533 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 534 {
 535         struct nfs_inode *nfsi = NFS_I(inode);
 536         int res = 0;
 537
 538         if (nfsi->ncommit != 0) {
 539                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 540                 nfsi->ncommit -= res;
 541                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 542                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 543         }
 544         return res;
 545 }
 546 #else
 547 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 548 {
 549         return 0;
 550 }
 551 #endif
 552
 553 static int nfs_wait_on_write_congestion(struct address_space *mapping)
 554 {
 555         struct inode *inode = mapping->host;
 556         struct backing_dev_info *bdi = mapping->backing_dev_info;
 557         int ret = 0;
 558
 559         might_sleep();
 560
 561         if (!bdi_write_congested(bdi))
 562                 return 0;
 563
 564         nfs_inc_stats(inode, NFSIOS_CONGESTIONWAIT);
 565
 566         do {
 567                 struct rpc_clnt *clnt = NFS_CLIENT(inode);
 568                 sigset_t oldset;
 569
 570                 rpc_clnt_sigmask(clnt, &oldset);
 571                 ret = congestion_wait_interruptible(WRITE, HZ/10);
 572                 rpc_clnt_sigunmask(clnt, &oldset);
 573                 if (ret == -ERESTARTSYS)
 574                         break;
 575                 ret = 0;
 576         } while (bdi_write_congested(bdi));
 577
 578         return ret;
 579 }
 580
 581 /*
 582  * Try to update any existing write request, or create one if there is none.
 583  * In order to match, the request's credentials must match those of
 584  * the calling process.
 585  *
 586  * Note: Should always be called with the Page Lock held!
 587  */
 588 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 589                 struct page *page, unsigned int offset, unsigned int bytes)
 590 {
 591         struct address_space *mapping = page->mapping;
 592         struct inode *inode = mapping->host;
 593         struct nfs_inode *nfsi = NFS_I(inode);
 594         struct nfs_page         *req, *new = NULL;
 595         unsigned long           rqend, end;
 596
 597         end = offset + bytes;
 598
 599         if (nfs_wait_on_write_congestion(mapping))
 600                 return ERR_PTR(-ERESTARTSYS);
 601         for (;;) {
 602                 /* Loop over all inode entries and see if we find
 603                  * A request for the page we wish to update
 604                  */
 605                 spin_lock(&nfsi->req_lock);
 606                 req = nfs_page_find_request_locked(page);
 607                 if (req) {
 608                         if (!nfs_lock_request_dontget(req)) {
 609                                 int error;
 610
 611                                 spin_unlock(&nfsi->req_lock);
 612                                 error = nfs_wait_on_request(req);
 613                                 nfs_release_request(req);
 614                                 if (error < 0) {
 615                                         if (new)
 616                                                 nfs_release_request(new);
 617                                         return ERR_PTR(error);
 618                                 }
 619                                 continue;
 620                         }
 621                         spin_unlock(&nfsi->req_lock);
 622                         if (new)
 623                                 nfs_release_request(new);
 624                         break;
 625                 }
 626
 627                 if (new) {
 628                         int error;
 629                         nfs_lock_request_dontget(new);
 630                         error = nfs_inode_add_request(inode, new);
 631                         if (error) {
 632                                 spin_unlock(&nfsi->req_lock);
 633                                 nfs_unlock_request(new);
 634                                 return ERR_PTR(error);
 635                         }
 636                         spin_unlock(&nfsi->req_lock);
 637                         return new;
 638                 }
 639                 spin_unlock(&nfsi->req_lock);
 640
 641                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 642                 if (IS_ERR(new))
 643                         return new;
 644         }
 645
 646         /* We have a request for our page.
 647          * If the creds don't match, or the
 648          * page addresses don't match,
 649          * tell the caller to wait on the conflicting
 650          * request.
 651          */
 652         rqend = req->wb_offset + req->wb_bytes;
 653         if (req->wb_context != ctx
 654             || req->wb_page != page
 655             || !nfs_dirty_request(req)
 656             || offset > rqend || end < req->wb_offset) {
 657                 nfs_unlock_request(req);
 658                 return ERR_PTR(-EBUSY);
 659         }
 660
 661         /* Okay, the request matches. Update the region */
 662         if (offset < req->wb_offset) {
 663                 req->wb_offset = offset;
 664                 req->wb_pgbase = offset;
 665                 req->wb_bytes = rqend - req->wb_offset;
 666         }
 667
 668         if (end > rqend)
 669                 req->wb_bytes = end - req->wb_offset;
 670
 671         return req;
 672 }
 673
 674 int nfs_flush_incompatible(struct file *file, struct page *page)
 675 {
 676         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 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 = (struct nfs_open_context *)file->private_data;
 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         nfs_end_page_writeback(req->wb_page);
 743
 744 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 745         if (!PageError(req->wb_page)) {
 746                 if (NFS_NEED_RESCHED(req)) {
 747                         nfs_redirty_request(req);
 748                         goto out;
 749                 } else if (NFS_NEED_COMMIT(req)) {
 750                         nfs_mark_request_commit(req);
 751                         goto out;
 752                 }
 753         }
 754         nfs_inode_remove_request(req);
 755
 756 out:
 757         nfs_clear_commit(req);
 758         nfs_clear_reschedule(req);
 759 #else
 760         nfs_inode_remove_request(req);
 761 #endif
 762         nfs_clear_page_writeback(req);
 763 }
 764
 765 static inline int flush_task_priority(int how)
 766 {
 767         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 768                 case FLUSH_HIGHPRI:
 769                         return RPC_PRIORITY_HIGH;
 770                 case FLUSH_LOWPRI:
 771                         return RPC_PRIORITY_LOW;
 772         }
 773         return RPC_PRIORITY_NORMAL;
 774 }
 775
 776 /*
 777  * Set up the argument/result storage required for the RPC call.
 778  */
 779 static void nfs_write_rpcsetup(struct nfs_page *req,
 780                 struct nfs_write_data *data,
 781                 const struct rpc_call_ops *call_ops,
 782                 unsigned int count, unsigned int offset,
 783                 int how)
 784 {
 785         struct inode            *inode;
 786         int flags;
 787
 788         /* Set up the RPC argument and reply structs
 789          * NB: take care not to mess about with data->commit et al. */
 790
 791         data->req = req;
 792         data->inode = inode = req->wb_context->dentry->d_inode;
 793         data->cred = req->wb_context->cred;
 794
 795         data->args.fh     = NFS_FH(inode);
 796         data->args.offset = req_offset(req) + offset;
 797         data->args.pgbase = req->wb_pgbase + offset;
 798         data->args.pages  = data->pagevec;
 799         data->args.count  = count;
 800         data->args.context = req->wb_context;
 801
 802         data->res.fattr   = &data->fattr;
 803         data->res.count   = count;
 804         data->res.verf    = &data->verf;
 805         nfs_fattr_init(&data->fattr);
 806
 807         /* Set up the initial task struct.  */
 808         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 809         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 810         NFS_PROTO(inode)->write_setup(data, how);
 811
 812         data->task.tk_priority = flush_task_priority(how);
 813         data->task.tk_cookie = (unsigned long)inode;
 814
 815         dprintk("NFS: %5u initiated write call "
 816                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 817                 data->task.tk_pid,
 818                 inode->i_sb->s_id,
 819                 (long long)NFS_FILEID(inode),
 820                 count,
 821                 (unsigned long long)data->args.offset);
 822 }
 823
 824 static void nfs_execute_write(struct nfs_write_data *data)
 825 {
 826         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 827         sigset_t oldset;
 828
 829         rpc_clnt_sigmask(clnt, &oldset);
 830         rpc_execute(&data->task);
 831         rpc_clnt_sigunmask(clnt, &oldset);
 832 }
 833
 834 /*
 835  * Generate multiple small requests to write out a single
 836  * contiguous dirty area on one page.
 837  */
 838 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 839 {
 840         struct nfs_page *req = nfs_list_entry(head->next);
 841         struct page *page = req->wb_page;
 842         struct nfs_write_data *data;
 843         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 844         unsigned int offset;
 845         int requests = 0;
 846         LIST_HEAD(list);
 847
 848         nfs_list_remove_request(req);
 849
 850         nbytes = req->wb_bytes;
 851         do {
 852                 size_t len = min(nbytes, wsize);
 853
 854                 data = nfs_writedata_alloc(len);
 855                 if (!data)
 856                         goto out_bad;
 857                 list_add(&data->pages, &list);
 858                 requests++;
 859                 nbytes -= len;
 860         } while (nbytes != 0);
 861         atomic_set(&req->wb_complete, requests);
 862
 863         ClearPageError(page);
 864         offset = 0;
 865         nbytes = req->wb_bytes;
 866         do {
 867                 data = list_entry(list.next, struct nfs_write_data, pages);
 868                 list_del_init(&data->pages);
 869
 870                 data->pagevec[0] = page;
 871
 872                 if (nbytes > wsize) {
 873                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 874                                         wsize, offset, how);
 875                         offset += wsize;
 876                         nbytes -= wsize;
 877                 } else {
 878                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 879                                         nbytes, offset, how);
 880                         nbytes = 0;
 881                 }
 882                 nfs_execute_write(data);
 883         } while (nbytes != 0);
 884
 885         return 0;
 886
 887 out_bad:
 888         while (!list_empty(&list)) {
 889                 data = list_entry(list.next, struct nfs_write_data, pages);
 890                 list_del(&data->pages);
 891                 nfs_writedata_release(data);
 892         }
 893         nfs_redirty_request(req);
 894         nfs_clear_page_writeback(req);
 895         return -ENOMEM;
 896 }
 897
 898 /*
 899  * Create an RPC task for the given write request and kick it.
 900  * The page must have been locked by the caller.
 901  *
 902  * It may happen that the page we're passed is not marked dirty.
 903  * This is the case if nfs_updatepage detects a conflicting request
 904  * that has been written but not committed.
 905  */
 906 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 907 {
 908         struct nfs_page         *req;
 909         struct page             **pages;
 910         struct nfs_write_data   *data;
 911         unsigned int            count;
 912
 913         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 914         if (!data)
 915                 goto out_bad;
 916
 917         pages = data->pagevec;
 918         count = 0;
 919         while (!list_empty(head)) {
 920                 req = nfs_list_entry(head->next);
 921                 nfs_list_remove_request(req);
 922                 nfs_list_add_request(req, &data->pages);
 923                 ClearPageError(req->wb_page);
 924                 *pages++ = req->wb_page;
 925                 count += req->wb_bytes;
 926         }
 927         req = nfs_list_entry(data->pages.next);
 928
 929         /* Set up the argument struct */
 930         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 931
 932         nfs_execute_write(data);
 933         return 0;
 934  out_bad:
 935         while (!list_empty(head)) {
 936                 struct nfs_page *req = nfs_list_entry(head->next);
 937                 nfs_list_remove_request(req);
 938                 nfs_redirty_request(req);
 939                 nfs_clear_page_writeback(req);
 940         }
 941         return -ENOMEM;
 942 }
 943
 944 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 945 {
 946         LIST_HEAD(one_request);
 947         int (*flush_one)(struct inode *, struct list_head *, int);
 948         struct nfs_page *req;
 949         int wpages = NFS_SERVER(inode)->wpages;
 950         int wsize = NFS_SERVER(inode)->wsize;
 951         int error;
 952
 953         flush_one = nfs_flush_one;
 954         if (wsize < PAGE_CACHE_SIZE)
 955                 flush_one = nfs_flush_multi;
 956         /* For single writes, FLUSH_STABLE is more efficient */
 957         if (npages <= wpages && npages == NFS_I(inode)->npages
 958                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 959                 how |= FLUSH_STABLE;
 960
 961         do {
 962                 nfs_coalesce_requests(head, &one_request, wpages);
 963                 req = nfs_list_entry(one_request.next);
 964                 error = flush_one(inode, &one_request, how);
 965                 if (error < 0)
 966                         goto out_err;
 967         } while (!list_empty(head));
 968         return 0;
 969 out_err:
 970         while (!list_empty(head)) {
 971                 req = nfs_list_entry(head->next);
 972                 nfs_list_remove_request(req);
 973                 nfs_redirty_request(req);
 974                 nfs_clear_page_writeback(req);
 975         }
 976         return error;
 977 }
 978
 979 /*
 980  * Handle a write reply that flushed part of a page.
 981  */
 982 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 983 {
 984         struct nfs_write_data   *data = calldata;
 985         struct nfs_page         *req = data->req;
 986         struct page             *page = req->wb_page;
 987
 988         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 989                 req->wb_context->dentry->d_inode->i_sb->s_id,
 990                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 991                 req->wb_bytes,
 992                 (long long)req_offset(req));
 993
 994         if (nfs_writeback_done(task, data) != 0)
 995                 return;
 996
 997         if (task->tk_status < 0) {
 998                 nfs_set_pageerror(page);
 999                 req->wb_context->error = task->tk_status;
1000                 dprintk(", error = %d\n", task->tk_status);
1001         } else {
1002 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1003                 if (data->verf.committed < NFS_FILE_SYNC) {
1004                         if (!NFS_NEED_COMMIT(req)) {
1005                                 nfs_defer_commit(req);
1006                                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1007                                 dprintk(" defer commit\n");
1008                         } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1009                                 nfs_defer_reschedule(req);
1010                                 dprintk(" server reboot detected\n");
1011                         }
1012                 } else
1013 #endif
1014                         dprintk(" OK\n");
1015         }
1016
1017         if (atomic_dec_and_test(&req->wb_complete))
1018                 nfs_writepage_release(req);
1019 }
1020
1021 static const struct rpc_call_ops nfs_write_partial_ops = {
1022         .rpc_call_done = nfs_writeback_done_partial,
1023         .rpc_release = nfs_writedata_release,
1024 };
1025
1026 /*
1027  * Handle a write reply that flushes a whole page.
1028  *
1029  * FIXME: There is an inherent race with invalidate_inode_pages and
1030  *        writebacks since the page->count is kept > 1 for as long
1031  *        as the page has a write request pending.
1032  */
1033 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1034 {
1035         struct nfs_write_data   *data = calldata;
1036         struct nfs_page         *req;
1037         struct page             *page;
1038
1039         if (nfs_writeback_done(task, data) != 0)
1040                 return;
1041
1042         /* Update attributes as result of writeback. */
1043         while (!list_empty(&data->pages)) {
1044                 req = nfs_list_entry(data->pages.next);
1045                 nfs_list_remove_request(req);
1046                 page = req->wb_page;
1047
1048                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1049                         req->wb_context->dentry->d_inode->i_sb->s_id,
1050                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1051                         req->wb_bytes,
1052                         (long long)req_offset(req));
1053
1054                 if (task->tk_status < 0) {
1055                         nfs_set_pageerror(page);
1056                         req->wb_context->error = task->tk_status;
1057                         nfs_end_page_writeback(page);
1058                         nfs_inode_remove_request(req);
1059                         dprintk(", error = %d\n", task->tk_status);
1060                         goto next;
1061                 }
1062                 nfs_end_page_writeback(page);
1063
1064 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1065                 if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) {
1066                         nfs_inode_remove_request(req);
1067                         dprintk(" OK\n");
1068                         goto next;
1069                 }
1070                 memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1071                 nfs_mark_request_commit(req);
1072                 dprintk(" marked for commit\n");
1073 #else
1074                 nfs_inode_remove_request(req);
1075 #endif
1076         next:
1077                 nfs_clear_page_writeback(req);
1078         }
1079 }
1080
1081 static const struct rpc_call_ops nfs_write_full_ops = {
1082         .rpc_call_done = nfs_writeback_done_full,
1083         .rpc_release = nfs_writedata_release,
1084 };
1085
1086
1087 /*
1088  * This function is called when the WRITE call is complete.
1089  */
1090 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1091 {
1092         struct nfs_writeargs    *argp = &data->args;
1093         struct nfs_writeres     *resp = &data->res;
1094         int status;
1095
1096         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1097                 task->tk_pid, task->tk_status);
1098
1099         /*
1100          * ->write_done will attempt to use post-op attributes to detect
1101          * conflicting writes by other clients.  A strict interpretation
1102          * of close-to-open would allow us to continue caching even if
1103          * another writer had changed the file, but some applications
1104          * depend on tighter cache coherency when writing.
1105          */
1106         status = NFS_PROTO(data->inode)->write_done(task, data);
1107         if (status != 0)
1108                 return status;
1109         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1110
1111 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1112         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1113                 /* We tried a write call, but the server did not
1114                  * commit data to stable storage even though we
1115                  * requested it.
1116                  * Note: There is a known bug in Tru64 < 5.0 in which
1117                  *       the server reports NFS_DATA_SYNC, but performs
1118                  *       NFS_FILE_SYNC. We therefore implement this checking
1119                  *       as a dprintk() in order to avoid filling syslog.
1120                  */
1121                 static unsigned long    complain;
1122
1123                 if (time_before(complain, jiffies)) {
1124                         dprintk("NFS: faulty NFS server %s:"
1125                                 " (committed = %d) != (stable = %d)\n",
1126                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1127                                 resp->verf->committed, argp->stable);
1128                         complain = jiffies + 300 * HZ;
1129                 }
1130         }
1131 #endif
1132         /* Is this a short write? */
1133         if (task->tk_status >= 0 && resp->count < argp->count) {
1134                 static unsigned long    complain;
1135
1136                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1137
1138                 /* Has the server at least made some progress? */
1139                 if (resp->count != 0) {
1140                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1141                         if (resp->verf->committed != NFS_UNSTABLE) {
1142                                 /* Resend from where the server left off */
1143                                 argp->offset += resp->count;
1144                                 argp->pgbase += resp->count;
1145                                 argp->count -= resp->count;
1146                         } else {
1147                                 /* Resend as a stable write in order to avoid
1148                                  * headaches in the case of a server crash.
1149                                  */
1150                                 argp->stable = NFS_FILE_SYNC;
1151                         }
1152                         rpc_restart_call(task);
1153                         return -EAGAIN;
1154                 }
1155                 if (time_before(complain, jiffies)) {
1156                         printk(KERN_WARNING
1157                                "NFS: Server wrote zero bytes, expected %u.\n",
1158                                         argp->count);
1159                         complain = jiffies + 300 * HZ;
1160                 }
1161                 /* Can't do anything about it except throw an error. */
1162                 task->tk_status = -EIO;
1163         }
1164         return 0;
1165 }
1166
1167
1168 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1169 void nfs_commit_release(void *wdata)
1170 {
1171         nfs_commit_free(wdata);
1172 }
1173
1174 /*
1175  * Set up the argument/result storage required for the RPC call.
1176  */
1177 static void nfs_commit_rpcsetup(struct list_head *head,
1178                 struct nfs_write_data *data,
1179                 int how)
1180 {
1181         struct nfs_page         *first;
1182         struct inode            *inode;
1183         int flags;
1184
1185         /* Set up the RPC argument and reply structs
1186          * NB: take care not to mess about with data->commit et al. */
1187
1188         list_splice_init(head, &data->pages);
1189         first = nfs_list_entry(data->pages.next);
1190         inode = first->wb_context->dentry->d_inode;
1191
1192         data->inode       = inode;
1193         data->cred        = first->wb_context->cred;
1194
1195         data->args.fh     = NFS_FH(data->inode);
1196         /* Note: we always request a commit of the entire inode */
1197         data->args.offset = 0;
1198         data->args.count  = 0;
1199         data->res.count   = 0;
1200         data->res.fattr   = &data->fattr;
1201         data->res.verf    = &data->verf;
1202         nfs_fattr_init(&data->fattr);
1203
1204         /* Set up the initial task struct.  */
1205         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1206         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1207         NFS_PROTO(inode)->commit_setup(data, how);
1208
1209         data->task.tk_priority = flush_task_priority(how);
1210         data->task.tk_cookie = (unsigned long)inode;
1211
1212         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1213 }
1214
1215 /*
1216  * Commit dirty pages
1217  */
1218 static int
1219 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1220 {
1221         struct nfs_write_data   *data;
1222         struct nfs_page         *req;
1223
1224         data = nfs_commit_alloc();
1225
1226         if (!data)
1227                 goto out_bad;
1228
1229         /* Set up the argument struct */
1230         nfs_commit_rpcsetup(head, data, how);
1231
1232         nfs_execute_write(data);
1233         return 0;
1234  out_bad:
1235         while (!list_empty(head)) {
1236                 req = nfs_list_entry(head->next);
1237                 nfs_list_remove_request(req);
1238                 nfs_mark_request_commit(req);
1239                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1240                 nfs_clear_page_writeback(req);
1241         }
1242         return -ENOMEM;
1243 }
1244
1245 /*
1246  * COMMIT call returned
1247  */
1248 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1249 {
1250         struct nfs_write_data   *data = calldata;
1251         struct nfs_page         *req;
1252
1253         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1254                                 task->tk_pid, task->tk_status);
1255
1256         /* Call the NFS version-specific code */
1257         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1258                 return;
1259
1260         while (!list_empty(&data->pages)) {
1261                 req = nfs_list_entry(data->pages.next);
1262                 nfs_list_remove_request(req);
1263                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1264
1265                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1266                         req->wb_context->dentry->d_inode->i_sb->s_id,
1267                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1268                         req->wb_bytes,
1269                         (long long)req_offset(req));
1270                 if (task->tk_status < 0) {
1271                         req->wb_context->error = task->tk_status;
1272                         nfs_inode_remove_request(req);
1273                         dprintk(", error = %d\n", task->tk_status);
1274                         goto next;
1275                 }
1276
1277                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1278                  * returned by the server against all stored verfs. */
1279                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1280                         /* We have a match */
1281                         nfs_inode_remove_request(req);
1282                         dprintk(" OK\n");
1283                         goto next;
1284                 }
1285                 /* We have a mismatch. Write the page again */
1286                 dprintk(" mismatch\n");
1287                 nfs_redirty_request(req);
1288         next:
1289                 nfs_clear_page_writeback(req);
1290         }
1291 }
1292
1293 static const struct rpc_call_ops nfs_commit_ops = {
1294         .rpc_call_done = nfs_commit_done,
1295         .rpc_release = nfs_commit_release,
1296 };
1297 #else
1298 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1299 {
1300         return 0;
1301 }
1302 #endif
1303
1304 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1305 {
1306         struct nfs_inode *nfsi = NFS_I(mapping->host);
1307         LIST_HEAD(head);
1308         long res;
1309
1310         spin_lock(&nfsi->req_lock);
1311         res = nfs_scan_dirty(mapping, wbc, &head);
1312         spin_unlock(&nfsi->req_lock);
1313         if (res) {
1314                 int error = nfs_flush_list(mapping->host, &head, res, how);
1315                 if (error < 0)
1316                         return error;
1317         }
1318         return res;
1319 }
1320
1321 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1322 int nfs_commit_inode(struct inode *inode, int how)
1323 {
1324         struct nfs_inode *nfsi = NFS_I(inode);
1325         LIST_HEAD(head);
1326         int res;
1327
1328         spin_lock(&nfsi->req_lock);
1329         res = nfs_scan_commit(inode, &head, 0, 0);
1330         spin_unlock(&nfsi->req_lock);
1331         if (res) {
1332                 int error = nfs_commit_list(inode, &head, how);
1333                 if (error < 0)
1334                         return error;
1335         }
1336         return res;
1337 }
1338 #endif
1339
1340 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1341 {
1342         struct inode *inode = mapping->host;
1343         struct nfs_inode *nfsi = NFS_I(inode);
1344         unsigned long idx_start, idx_end;
1345         unsigned int npages = 0;
1346         LIST_HEAD(head);
1347         int nocommit = how & FLUSH_NOCOMMIT;
1348         long pages, ret;
1349
1350         /* FIXME */
1351         if (wbc->range_cyclic)
1352                 idx_start = 0;
1353         else {
1354                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1355                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1356                 if (idx_end > idx_start) {
1357                         unsigned long l_npages = 1 + idx_end - idx_start;
1358                         npages = l_npages;
1359                         if (sizeof(npages) != sizeof(l_npages) &&
1360                                         (unsigned long)npages != l_npages)
1361                                 npages = 0;
1362                 }
1363         }
1364         how &= ~FLUSH_NOCOMMIT;
1365         spin_lock(&nfsi->req_lock);
1366         do {
1367                 wbc->pages_skipped = 0;
1368                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1369                 if (ret != 0)
1370                         continue;
1371                 pages = nfs_scan_dirty(mapping, wbc, &head);
1372                 if (pages != 0) {
1373                         spin_unlock(&nfsi->req_lock);
1374                         if (how & FLUSH_INVALIDATE) {
1375                                 nfs_cancel_dirty_list(&head);
1376                                 ret = pages;
1377                         } else
1378                                 ret = nfs_flush_list(inode, &head, pages, how);
1379                         spin_lock(&nfsi->req_lock);
1380                         continue;
1381                 }
1382                 if (wbc->pages_skipped != 0)
1383                         continue;
1384                 if (nocommit)
1385                         break;
1386                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1387                 if (pages == 0) {
1388                         if (wbc->pages_skipped != 0)
1389                                 continue;
1390                         break;
1391                 }
1392                 if (how & FLUSH_INVALIDATE) {
1393                         spin_unlock(&nfsi->req_lock);
1394                         nfs_cancel_commit_list(&head);
1395                         ret = pages;
1396                         spin_lock(&nfsi->req_lock);
1397                         continue;
1398                 }
1399                 pages += nfs_scan_commit(inode, &head, 0, 0);
1400                 spin_unlock(&nfsi->req_lock);
1401                 ret = nfs_commit_list(inode, &head, how);
1402                 spin_lock(&nfsi->req_lock);
1403         } while (ret >= 0);
1404         spin_unlock(&nfsi->req_lock);
1405         return ret;
1406 }
1407
1408 /*
1409  * flush the inode to disk.
1410  */
1411 int nfs_wb_all(struct inode *inode)
1412 {
1413         struct address_space *mapping = inode->i_mapping;
1414         struct writeback_control wbc = {
1415                 .bdi = mapping->backing_dev_info,
1416                 .sync_mode = WB_SYNC_ALL,
1417                 .nr_to_write = LONG_MAX,
1418                 .for_writepages = 1,
1419                 .range_cyclic = 1,
1420         };
1421         int ret;
1422
1423         ret = generic_writepages(mapping, &wbc);
1424         if (ret < 0)
1425                 goto out;
1426         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1427         if (ret >= 0)
1428                 return 0;
1429 out:
1430         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1431         return ret;
1432 }
1433
1434 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1435 {
1436         struct writeback_control wbc = {
1437                 .bdi = mapping->backing_dev_info,
1438                 .sync_mode = WB_SYNC_ALL,
1439                 .nr_to_write = LONG_MAX,
1440                 .range_start = range_start,
1441                 .range_end = range_end,
1442                 .for_writepages = 1,
1443         };
1444         int ret;
1445
1446         if (!(how & FLUSH_NOWRITEPAGE)) {
1447                 ret = generic_writepages(mapping, &wbc);
1448                 if (ret < 0)
1449                         goto out;
1450         }
1451         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1452         if (ret >= 0)
1453                 return 0;
1454 out:
1455         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1456         return ret;
1457 }
1458
1459 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1460 {
1461         loff_t range_start = page_offset(page);
1462         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1463         struct writeback_control wbc = {
1464                 .bdi = page->mapping->backing_dev_info,
1465                 .sync_mode = WB_SYNC_ALL,
1466                 .nr_to_write = LONG_MAX,
1467                 .range_start = range_start,
1468                 .range_end = range_end,
1469         };
1470         int ret;
1471
1472         BUG_ON(!PageLocked(page));
1473         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1474                 ret = nfs_writepage_locked(page, &wbc);
1475                 if (ret < 0)
1476                         goto out;
1477         }
1478         if (!PagePrivate(page))
1479                 return 0;
1480         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1481         if (ret >= 0)
1482                 return 0;
1483 out:
1484         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1485         return ret;
1486 }
1487
1488 /*
1489  * Write back all requests on one page - we do this before reading it.
1490  */
1491 int nfs_wb_page(struct inode *inode, struct page* page)
1492 {
1493         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1494 }
1495
1496 int nfs_set_page_dirty(struct page *page)
1497 {
1498         struct nfs_page *req;
1499
1500         req = nfs_page_find_request(page);
1501         if (req != NULL) {
1502                 /* Mark any existing write requests for flushing */
1503                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
1504                 nfs_release_request(req);
1505         }
1506         return __set_page_dirty_nobuffers(page);
1507 }
1508
1509
1510 int __init nfs_init_writepagecache(void)
1511 {
1512         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1513                                              sizeof(struct nfs_write_data),
1514                                              0, SLAB_HWCACHE_ALIGN,
1515                                              NULL, NULL);
1516         if (nfs_wdata_cachep == NULL)
1517                 return -ENOMEM;
1518
1519         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1520                                                      nfs_wdata_cachep);
1521         if (nfs_wdata_mempool == NULL)
1522                 return -ENOMEM;
1523
1524         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1525                                                       nfs_wdata_cachep);
1526         if (nfs_commit_mempool == NULL)
1527                 return -ENOMEM;
1528
1529         /*
1530          * NFS congestion size, scale with available memory.
1531          *
1532          *  64MB:    8192k
1533          * 128MB:   11585k
1534          * 256MB:   16384k
1535          * 512MB:   23170k
1536          *   1GB:   32768k
1537          *   2GB:   46340k
1538          *   4GB:   65536k
1539          *   8GB:   92681k
1540          *  16GB:  131072k
1541          *
1542          * This allows larger machines to have larger/more transfers.
1543          * Limit the default to 256M
1544          */
1545         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1546         if (nfs_congestion_kb > 256*1024)
1547                 nfs_congestion_kb = 256*1024;
1548
1549         return 0;
1550 }
1551
1552 void nfs_destroy_writepagecache(void)
1553 {
1554         mempool_destroy(nfs_commit_mempool);
1555         mempool_destroy(nfs_wdata_mempool);
1556         kmem_cache_destroy(nfs_wdata_cachep);
1557 }
1558