git.oblomov.eu Git - linux-2.6/blob - 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 long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how);
  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 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(size_t len)
  75 {
  76         unsigned int pagecount = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
  77         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  78
  79         if (p) {
  80                 memset(p, 0, sizeof(*p));
  81                 INIT_LIST_HEAD(&p->pages);
  82                 p->npages = pagecount;
  83                 if (pagecount <= ARRAY_SIZE(p->page_array))
  84                         p->pagevec = p->page_array;
  85                 else {
  86                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  87                         if (!p->pagevec) {
  88                                 mempool_free(p, nfs_wdata_mempool);
  89                                 p = NULL;
  90                         }
  91                 }
  92         }
  93         return p;
  94 }
  95
  96 static void nfs_writedata_rcu_free(struct rcu_head *head)
  97 {
  98         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  99         if (p && (p->pagevec != &p->page_array[0]))
 100                 kfree(p->pagevec);
 101         mempool_free(p, nfs_wdata_mempool);
 102 }
 103
 104 static void nfs_writedata_free(struct nfs_write_data *wdata)
 105 {
 106         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 107 }
 108
 109 void nfs_writedata_release(void *wdata)
 110 {
 111         nfs_writedata_free(wdata);
 112 }
 113
 114 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 115 {
 116         struct nfs_page *req = NULL;
 117
 118         if (PagePrivate(page)) {
 119                 req = (struct nfs_page *)page_private(page);
 120                 if (req != NULL)
 121                         atomic_inc(&req->wb_count);
 122         }
 123         return req;
 124 }
 125
 126 static struct nfs_page *nfs_page_find_request(struct page *page)
 127 {
 128         struct nfs_page *req = NULL;
 129         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 130
 131         spin_lock(req_lock);
 132         req = nfs_page_find_request_locked(page);
 133         spin_unlock(req_lock);
 134         return req;
 135 }
 136
 137 /* Adjust the file length if we're writing beyond the end */
 138 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 139 {
 140         struct inode *inode = page->mapping->host;
 141         loff_t end, i_size = i_size_read(inode);
 142         unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 143
 144         if (i_size > 0 && page->index < end_index)
 145                 return;
 146         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 147         if (i_size >= end)
 148                 return;
 149         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 150         i_size_write(inode, end);
 151 }
 152
 153 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 154 static void nfs_set_pageerror(struct page *page)
 155 {
 156         SetPageError(page);
 157         nfs_zap_mapping(page->mapping->host, page->mapping);
 158 }
 159
 160 /* We can set the PG_uptodate flag if we see that a write request
 161  * covers the full page.
 162  */
 163 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 164 {
 165         if (PageUptodate(page))
 166                 return;
 167         if (base != 0)
 168                 return;
 169         if (count != nfs_page_length(page))
 170                 return;
 171         if (count != PAGE_CACHE_SIZE)
 172                 memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count);
 173         SetPageUptodate(page);
 174 }
 175
 176 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 177                 unsigned int offset, unsigned int count)
 178 {
 179         struct nfs_page *req;
 180         int ret;
 181
 182         for (;;) {
 183                 req = nfs_update_request(ctx, page, offset, count);
 184                 if (!IS_ERR(req))
 185                         break;
 186                 ret = PTR_ERR(req);
 187                 if (ret != -EBUSY)
 188                         return ret;
 189                 ret = nfs_wb_page(page->mapping->host, page);
 190                 if (ret != 0)
 191                         return ret;
 192         }
 193         /* Update file length */
 194         nfs_grow_file(page, offset, count);
 195         /* Set the PG_uptodate flag? */
 196         nfs_mark_uptodate(page, offset, count);
 197         nfs_unlock_request(req);
 198         return 0;
 199 }
 200
 201 static int wb_priority(struct writeback_control *wbc)
 202 {
 203         if (wbc->for_reclaim)
 204                 return FLUSH_HIGHPRI;
 205         if (wbc->for_kupdate)
 206                 return FLUSH_LOWPRI;
 207         return 0;
 208 }
 209
 210 /*
 211  * NFS congestion control
 212  */
 213
 214 int nfs_congestion_kb;
 215
 216 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 217 #define NFS_CONGESTION_OFF_THRESH       \
 218         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 219
 220 static int nfs_set_page_writeback(struct page *page)
 221 {
 222         int ret = test_set_page_writeback(page);
 223
 224         if (!ret) {
 225                 struct inode *inode = page->mapping->host;
 226                 struct nfs_server *nfss = NFS_SERVER(inode);
 227
 228                 if (atomic_inc_return(&nfss->writeback) >
 229                                 NFS_CONGESTION_ON_THRESH)
 230                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 231         }
 232         return ret;
 233 }
 234
 235 static void nfs_end_page_writeback(struct page *page)
 236 {
 237         struct inode *inode = page->mapping->host;
 238         struct nfs_server *nfss = NFS_SERVER(inode);
 239
 240         end_page_writeback(page);
 241         if (atomic_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 242                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 243                 congestion_end(WRITE);
 244         }
 245 }
 246
 247 /*
 248  * Find an associated nfs write request, and prepare to flush it out
 249  * Returns 1 if there was no write request, or if the request was
 250  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 251  * was not tagged.
 252  * May also return an error if the user signalled nfs_wait_on_request().
 253  */
 254 static int nfs_page_mark_flush(struct page *page)
 255 {
 256         struct nfs_page *req;
 257         struct nfs_inode *nfsi = NFS_I(page->mapping->host);
 258         spinlock_t *req_lock = &nfsi->req_lock;
 259         int ret;
 260
 261         spin_lock(req_lock);
 262         for(;;) {
 263                 req = nfs_page_find_request_locked(page);
 264                 if (req == NULL) {
 265                         spin_unlock(req_lock);
 266                         return 1;
 267                 }
 268                 if (nfs_lock_request_dontget(req))
 269                         break;
 270                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 271                  *       then the call to nfs_lock_request_dontget() will always
 272                  *       succeed provided that someone hasn't already marked the
 273                  *       request as dirty (in which case we don't care).
 274                  */
 275                 spin_unlock(req_lock);
 276                 ret = nfs_wait_on_request(req);
 277                 nfs_release_request(req);
 278                 if (ret != 0)
 279                         return ret;
 280                 spin_lock(req_lock);
 281         }
 282         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 283                 /* This request is marked for commit */
 284                 spin_unlock(req_lock);
 285                 nfs_unlock_request(req);
 286                 return 1;
 287         }
 288         if (nfs_set_page_writeback(page) == 0) {
 289                 nfs_list_remove_request(req);
 290                 /* add the request to the inode's dirty list. */
 291                 radix_tree_tag_set(&nfsi->nfs_page_tree,
 292                                 req->wb_index, NFS_PAGE_TAG_DIRTY);
 293                 nfs_list_add_request(req, &nfsi->dirty);
 294                 nfsi->ndirty++;
 295                 spin_unlock(req_lock);
 296                 __mark_inode_dirty(page->mapping->host, I_DIRTY_PAGES);
 297         } else
 298                 spin_unlock(req_lock);
 299         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 300         nfs_unlock_request(req);
 301         return ret;
 302 }
 303
 304 /*
 305  * Write an mmapped page to the server.
 306  */
 307 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 308 {
 309         struct nfs_open_context *ctx;
 310         struct inode *inode = page->mapping->host;
 311         unsigned offset;
 312         int err;
 313
 314         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 315         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 316
 317         err = nfs_page_mark_flush(page);
 318         if (err <= 0)
 319                 goto out;
 320         err = 0;
 321         offset = nfs_page_length(page);
 322         if (!offset)
 323                 goto out;
 324
 325         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 326         if (ctx == NULL) {
 327                 err = -EBADF;
 328                 goto out;
 329         }
 330         err = nfs_writepage_setup(ctx, page, 0, offset);
 331         put_nfs_open_context(ctx);
 332         if (err != 0)
 333                 goto out;
 334         err = nfs_page_mark_flush(page);
 335         if (err > 0)
 336                 err = 0;
 337 out:
 338         if (!wbc->for_writepages)
 339                 nfs_flush_mapping(page->mapping, wbc, FLUSH_STABLE|wb_priority(wbc));
 340         return err;
 341 }
 342
 343 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 344 {
 345         int err;
 346
 347         err = nfs_writepage_locked(page, wbc);
 348         unlock_page(page);
 349         return err;
 350 }
 351
 352 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 353 {
 354         struct inode *inode = mapping->host;
 355         int err;
 356
 357         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 358
 359         err = generic_writepages(mapping, wbc);
 360         if (err)
 361                 return err;
 362         err = nfs_flush_mapping(mapping, wbc, wb_priority(wbc));
 363         if (err < 0)
 364                 goto out;
 365         nfs_add_stats(inode, NFSIOS_WRITEPAGES, err);
 366         err = 0;
 367 out:
 368         return err;
 369 }
 370
 371 /*
 372  * Insert a write request into an inode
 373  */
 374 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 375 {
 376         struct nfs_inode *nfsi = NFS_I(inode);
 377         int error;
 378
 379         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 380         BUG_ON(error == -EEXIST);
 381         if (error)
 382                 return error;
 383         if (!nfsi->npages) {
 384                 igrab(inode);
 385                 nfs_begin_data_update(inode);
 386                 if (nfs_have_delegation(inode, FMODE_WRITE))
 387                         nfsi->change_attr++;
 388         }
 389         SetPagePrivate(req->wb_page);
 390         set_page_private(req->wb_page, (unsigned long)req);
 391         if (PageDirty(req->wb_page))
 392                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
 393         nfsi->npages++;
 394         atomic_inc(&req->wb_count);
 395         return 0;
 396 }
 397
 398 /*
 399  * Remove a write request from an inode
 400  */
 401 static void nfs_inode_remove_request(struct nfs_page *req)
 402 {
 403         struct inode *inode = req->wb_context->dentry->d_inode;
 404         struct nfs_inode *nfsi = NFS_I(inode);
 405
 406         BUG_ON (!NFS_WBACK_BUSY(req));
 407
 408         spin_lock(&nfsi->req_lock);
 409         set_page_private(req->wb_page, 0);
 410         ClearPagePrivate(req->wb_page);
 411         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 412         if (test_and_clear_bit(PG_NEED_FLUSH, &req->wb_flags))
 413                 __set_page_dirty_nobuffers(req->wb_page);
 414         nfsi->npages--;
 415         if (!nfsi->npages) {
 416                 spin_unlock(&nfsi->req_lock);
 417                 nfs_end_data_update(inode);
 418                 iput(inode);
 419         } else
 420                 spin_unlock(&nfsi->req_lock);
 421         nfs_clear_request(req);
 422         nfs_release_request(req);
 423 }
 424
 425 static void
 426 nfs_redirty_request(struct nfs_page *req)
 427 {
 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         struct page *page = req->wb_page;
 438
 439         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 440                 return 0;
 441         return !PageWriteback(req->wb_page);
 442 }
 443
 444 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 445 /*
 446  * Add a request to the inode's commit list.
 447  */
 448 static void
 449 nfs_mark_request_commit(struct nfs_page *req)
 450 {
 451         struct inode *inode = req->wb_context->dentry->d_inode;
 452         struct nfs_inode *nfsi = NFS_I(inode);
 453
 454         spin_lock(&nfsi->req_lock);
 455         nfs_list_add_request(req, &nfsi->commit);
 456         nfsi->ncommit++;
 457         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 458         spin_unlock(&nfsi->req_lock);
 459         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 460         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 461 }
 462
 463 static inline
 464 int nfs_write_need_commit(struct nfs_write_data *data)
 465 {
 466         return data->verf.committed != NFS_FILE_SYNC;
 467 }
 468
 469 static inline
 470 int nfs_reschedule_unstable_write(struct nfs_page *req)
 471 {
 472         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 473                 nfs_mark_request_commit(req);
 474                 return 1;
 475         }
 476         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 477                 nfs_redirty_request(req);
 478                 return 1;
 479         }
 480         return 0;
 481 }
 482 #else
 483 static inline void
 484 nfs_mark_request_commit(struct nfs_page *req)
 485 {
 486 }
 487
 488 static inline
 489 int nfs_write_need_commit(struct nfs_write_data *data)
 490 {
 491         return 0;
 492 }
 493
 494 static inline
 495 int nfs_reschedule_unstable_write(struct nfs_page *req)
 496 {
 497         return 0;
 498 }
 499 #endif
 500
 501 /*
 502  * Wait for a request to complete.
 503  *
 504  * Interruptible by signals only if mounted with intr flag.
 505  */
 506 static int nfs_wait_on_requests_locked(struct inode *inode, unsigned long idx_start, unsigned int npages)
 507 {
 508         struct nfs_inode *nfsi = NFS_I(inode);
 509         struct nfs_page *req;
 510         unsigned long           idx_end, next;
 511         unsigned int            res = 0;
 512         int                     error;
 513
 514         if (npages == 0)
 515                 idx_end = ~0;
 516         else
 517                 idx_end = idx_start + npages - 1;
 518
 519         next = idx_start;
 520         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) {
 521                 if (req->wb_index > idx_end)
 522                         break;
 523
 524                 next = req->wb_index + 1;
 525                 BUG_ON(!NFS_WBACK_BUSY(req));
 526
 527                 atomic_inc(&req->wb_count);
 528                 spin_unlock(&nfsi->req_lock);
 529                 error = nfs_wait_on_request(req);
 530                 nfs_release_request(req);
 531                 spin_lock(&nfsi->req_lock);
 532                 if (error < 0)
 533                         return error;
 534                 res++;
 535         }
 536         return res;
 537 }
 538
 539 static void nfs_cancel_dirty_list(struct list_head *head)
 540 {
 541         struct nfs_page *req;
 542         while(!list_empty(head)) {
 543                 req = nfs_list_entry(head->next);
 544                 nfs_list_remove_request(req);
 545                 nfs_end_page_writeback(req->wb_page);
 546                 nfs_inode_remove_request(req);
 547                 nfs_clear_page_writeback(req);
 548         }
 549 }
 550
 551 static void nfs_cancel_commit_list(struct list_head *head)
 552 {
 553         struct nfs_page *req;
 554
 555         while(!list_empty(head)) {
 556                 req = nfs_list_entry(head->next);
 557                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 558                 nfs_list_remove_request(req);
 559                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 560                 nfs_inode_remove_request(req);
 561                 nfs_unlock_request(req);
 562         }
 563 }
 564
 565 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 566 /*
 567  * nfs_scan_commit - Scan an inode for commit requests
 568  * @inode: NFS inode to scan
 569  * @dst: destination list
 570  * @idx_start: lower bound of page->index to scan.
 571  * @npages: idx_start + npages sets the upper bound to scan.
 572  *
 573  * Moves requests from the inode's 'commit' request list.
 574  * The requests are *not* checked to ensure that they form a contiguous set.
 575  */
 576 static int
 577 nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 578 {
 579         struct nfs_inode *nfsi = NFS_I(inode);
 580         int res = 0;
 581
 582         if (nfsi->ncommit != 0) {
 583                 res = nfs_scan_list(nfsi, &nfsi->commit, dst, idx_start, npages);
 584                 nfsi->ncommit -= res;
 585                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 586                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 587         }
 588         return res;
 589 }
 590 #else
 591 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages)
 592 {
 593         return 0;
 594 }
 595 #endif
 596
 597 /*
 598  * Try to update any existing write request, or create one if there is none.
 599  * In order to match, the request's credentials must match those of
 600  * the calling process.
 601  *
 602  * Note: Should always be called with the Page Lock held!
 603  */
 604 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 605                 struct page *page, unsigned int offset, unsigned int bytes)
 606 {
 607         struct address_space *mapping = page->mapping;
 608         struct inode *inode = mapping->host;
 609         struct nfs_inode *nfsi = NFS_I(inode);
 610         struct nfs_page         *req, *new = NULL;
 611         unsigned long           rqend, end;
 612
 613         end = offset + bytes;
 614
 615         for (;;) {
 616                 /* Loop over all inode entries and see if we find
 617                  * A request for the page we wish to update
 618                  */
 619                 spin_lock(&nfsi->req_lock);
 620                 req = nfs_page_find_request_locked(page);
 621                 if (req) {
 622                         if (!nfs_lock_request_dontget(req)) {
 623                                 int error;
 624
 625                                 spin_unlock(&nfsi->req_lock);
 626                                 error = nfs_wait_on_request(req);
 627                                 nfs_release_request(req);
 628                                 if (error < 0) {
 629                                         if (new)
 630                                                 nfs_release_request(new);
 631                                         return ERR_PTR(error);
 632                                 }
 633                                 continue;
 634                         }
 635                         spin_unlock(&nfsi->req_lock);
 636                         if (new)
 637                                 nfs_release_request(new);
 638                         break;
 639                 }
 640
 641                 if (new) {
 642                         int error;
 643                         nfs_lock_request_dontget(new);
 644                         error = nfs_inode_add_request(inode, new);
 645                         if (error) {
 646                                 spin_unlock(&nfsi->req_lock);
 647                                 nfs_unlock_request(new);
 648                                 return ERR_PTR(error);
 649                         }
 650                         spin_unlock(&nfsi->req_lock);
 651                         return new;
 652                 }
 653                 spin_unlock(&nfsi->req_lock);
 654
 655                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 656                 if (IS_ERR(new))
 657                         return new;
 658         }
 659
 660         /* We have a request for our page.
 661          * If the creds don't match, or the
 662          * page addresses don't match,
 663          * tell the caller to wait on the conflicting
 664          * request.
 665          */
 666         rqend = req->wb_offset + req->wb_bytes;
 667         if (req->wb_context != ctx
 668             || req->wb_page != page
 669             || !nfs_dirty_request(req)
 670             || offset > rqend || end < req->wb_offset) {
 671                 nfs_unlock_request(req);
 672                 return ERR_PTR(-EBUSY);
 673         }
 674
 675         /* Okay, the request matches. Update the region */
 676         if (offset < req->wb_offset) {
 677                 req->wb_offset = offset;
 678                 req->wb_pgbase = offset;
 679                 req->wb_bytes = rqend - req->wb_offset;
 680         }
 681
 682         if (end > rqend)
 683                 req->wb_bytes = end - req->wb_offset;
 684
 685         return req;
 686 }
 687
 688 int nfs_flush_incompatible(struct file *file, struct page *page)
 689 {
 690         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 691         struct nfs_page *req;
 692         int do_flush, status;
 693         /*
 694          * Look for a request corresponding to this page. If there
 695          * is one, and it belongs to another file, we flush it out
 696          * before we try to copy anything into the page. Do this
 697          * due to the lack of an ACCESS-type call in NFSv2.
 698          * Also do the same if we find a request from an existing
 699          * dropped page.
 700          */
 701         do {
 702                 req = nfs_page_find_request(page);
 703                 if (req == NULL)
 704                         return 0;
 705                 do_flush = req->wb_page != page || req->wb_context != ctx
 706                         || !nfs_dirty_request(req);
 707                 nfs_release_request(req);
 708                 if (!do_flush)
 709                         return 0;
 710                 status = nfs_wb_page(page->mapping->host, page);
 711         } while (status == 0);
 712         return status;
 713 }
 714
 715 /*
 716  * Update and possibly write a cached page of an NFS file.
 717  *
 718  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 719  * things with a page scheduled for an RPC call (e.g. invalidate it).
 720  */
 721 int nfs_updatepage(struct file *file, struct page *page,
 722                 unsigned int offset, unsigned int count)
 723 {
 724         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 725         struct inode    *inode = page->mapping->host;
 726         int             status = 0;
 727
 728         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 729
 730         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 731                 file->f_path.dentry->d_parent->d_name.name,
 732                 file->f_path.dentry->d_name.name, count,
 733                 (long long)(page_offset(page) +offset));
 734
 735         /* If we're not using byte range locks, and we know the page
 736          * is entirely in cache, it may be more efficient to avoid
 737          * fragmenting write requests.
 738          */
 739         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 740                 count = max(count + offset, nfs_page_length(page));
 741                 offset = 0;
 742         }
 743
 744         status = nfs_writepage_setup(ctx, page, offset, count);
 745         __set_page_dirty_nobuffers(page);
 746
 747         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 748                         status, (long long)i_size_read(inode));
 749         if (status < 0)
 750                 nfs_set_pageerror(page);
 751         return status;
 752 }
 753
 754 static void nfs_writepage_release(struct nfs_page *req)
 755 {
 756
 757         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 758                 nfs_end_page_writeback(req->wb_page);
 759                 nfs_inode_remove_request(req);
 760         } else
 761                 nfs_end_page_writeback(req->wb_page);
 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, size_t count, 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 = count;
 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 = count;
 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                         wsize = nbytes;
 874                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 875                                    wsize, offset, how);
 876                 offset += wsize;
 877                 nbytes -= wsize;
 878                 nfs_execute_write(data);
 879         } while (nbytes != 0);
 880
 881         return 0;
 882
 883 out_bad:
 884         while (!list_empty(&list)) {
 885                 data = list_entry(list.next, struct nfs_write_data, pages);
 886                 list_del(&data->pages);
 887                 nfs_writedata_release(data);
 888         }
 889         nfs_redirty_request(req);
 890         nfs_end_page_writeback(req->wb_page);
 891         nfs_clear_page_writeback(req);
 892         return -ENOMEM;
 893 }
 894
 895 /*
 896  * Create an RPC task for the given write request and kick it.
 897  * The page must have been locked by the caller.
 898  *
 899  * It may happen that the page we're passed is not marked dirty.
 900  * This is the case if nfs_updatepage detects a conflicting request
 901  * that has been written but not committed.
 902  */
 903 static int nfs_flush_one(struct inode *inode, struct list_head *head, size_t count, int how)
 904 {
 905         struct nfs_page         *req;
 906         struct page             **pages;
 907         struct nfs_write_data   *data;
 908
 909         data = nfs_writedata_alloc(count);
 910         if (!data)
 911                 goto out_bad;
 912
 913         pages = data->pagevec;
 914         while (!list_empty(head)) {
 915                 req = nfs_list_entry(head->next);
 916                 nfs_list_remove_request(req);
 917                 nfs_list_add_request(req, &data->pages);
 918                 ClearPageError(req->wb_page);
 919                 *pages++ = req->wb_page;
 920         }
 921         req = nfs_list_entry(data->pages.next);
 922
 923         /* Set up the argument struct */
 924         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 925
 926         nfs_execute_write(data);
 927         return 0;
 928  out_bad:
 929         while (!list_empty(head)) {
 930                 struct nfs_page *req = nfs_list_entry(head->next);
 931                 nfs_list_remove_request(req);
 932                 nfs_redirty_request(req);
 933                 nfs_end_page_writeback(req->wb_page);
 934                 nfs_clear_page_writeback(req);
 935         }
 936         return -ENOMEM;
 937 }
 938
 939 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 940 {
 941         struct nfs_pageio_descriptor desc;
 942         int wpages = NFS_SERVER(inode)->wpages;
 943         int wsize = NFS_SERVER(inode)->wsize;
 944
 945         /* For single writes, FLUSH_STABLE is more efficient */
 946         if (npages <= wpages && npages == NFS_I(inode)->npages
 947                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 948                 how |= FLUSH_STABLE;
 949
 950         if (wsize < PAGE_CACHE_SIZE)
 951                 nfs_pageio_init(&desc, inode, nfs_flush_multi, wsize, how);
 952         else
 953                 nfs_pageio_init(&desc, inode, nfs_flush_one, wsize, how);
 954         nfs_pageio_add_list(&desc, head);
 955         nfs_pageio_complete(&desc);
 956         if (desc.pg_error == 0)
 957                 return 0;
 958         while (!list_empty(head)) {
 959                 struct nfs_page *req = nfs_list_entry(head->next);
 960                 nfs_list_remove_request(req);
 961                 nfs_redirty_request(req);
 962                 nfs_end_page_writeback(req->wb_page);
 963                 nfs_clear_page_writeback(req);
 964         }
 965         return desc.pg_error;
 966 }
 967
 968 /*
 969  * Handle a write reply that flushed part of a page.
 970  */
 971 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 972 {
 973         struct nfs_write_data   *data = calldata;
 974         struct nfs_page         *req = data->req;
 975         struct page             *page = req->wb_page;
 976
 977         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 978                 req->wb_context->dentry->d_inode->i_sb->s_id,
 979                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 980                 req->wb_bytes,
 981                 (long long)req_offset(req));
 982
 983         if (nfs_writeback_done(task, data) != 0)
 984                 return;
 985
 986         if (task->tk_status < 0) {
 987                 nfs_set_pageerror(page);
 988                 req->wb_context->error = task->tk_status;
 989                 dprintk(", error = %d\n", task->tk_status);
 990                 goto out;
 991         }
 992
 993         if (nfs_write_need_commit(data)) {
 994                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 995
 996                 spin_lock(req_lock);
 997                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 998                         /* Do nothing we need to resend the writes */
 999                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1000                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1001                         dprintk(" defer commit\n");
1002                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1003                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1004                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1005                         dprintk(" server reboot detected\n");
1006                 }
1007                 spin_unlock(req_lock);
1008         } else
1009                 dprintk(" OK\n");
1010
1011 out:
1012         if (atomic_dec_and_test(&req->wb_complete))
1013                 nfs_writepage_release(req);
1014 }
1015
1016 static const struct rpc_call_ops nfs_write_partial_ops = {
1017         .rpc_call_done = nfs_writeback_done_partial,
1018         .rpc_release = nfs_writedata_release,
1019 };
1020
1021 /*
1022  * Handle a write reply that flushes a whole page.
1023  *
1024  * FIXME: There is an inherent race with invalidate_inode_pages and
1025  *        writebacks since the page->count is kept > 1 for as long
1026  *        as the page has a write request pending.
1027  */
1028 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1029 {
1030         struct nfs_write_data   *data = calldata;
1031         struct nfs_page         *req;
1032         struct page             *page;
1033
1034         if (nfs_writeback_done(task, data) != 0)
1035                 return;
1036
1037         /* Update attributes as result of writeback. */
1038         while (!list_empty(&data->pages)) {
1039                 req = nfs_list_entry(data->pages.next);
1040                 nfs_list_remove_request(req);
1041                 page = req->wb_page;
1042
1043                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1044                         req->wb_context->dentry->d_inode->i_sb->s_id,
1045                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1046                         req->wb_bytes,
1047                         (long long)req_offset(req));
1048
1049                 if (task->tk_status < 0) {
1050                         nfs_set_pageerror(page);
1051                         req->wb_context->error = task->tk_status;
1052                         dprintk(", error = %d\n", task->tk_status);
1053                         goto remove_request;
1054                 }
1055
1056                 if (nfs_write_need_commit(data)) {
1057                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1058                         nfs_mark_request_commit(req);
1059                         nfs_end_page_writeback(page);
1060                         dprintk(" marked for commit\n");
1061                         goto next;
1062                 }
1063                 dprintk(" OK\n");
1064 remove_request:
1065                 nfs_end_page_writeback(page);
1066                 nfs_inode_remove_request(req);
1067         next:
1068                 nfs_clear_page_writeback(req);
1069         }
1070 }
1071
1072 static const struct rpc_call_ops nfs_write_full_ops = {
1073         .rpc_call_done = nfs_writeback_done_full,
1074         .rpc_release = nfs_writedata_release,
1075 };
1076
1077
1078 /*
1079  * This function is called when the WRITE call is complete.
1080  */
1081 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1082 {
1083         struct nfs_writeargs    *argp = &data->args;
1084         struct nfs_writeres     *resp = &data->res;
1085         int status;
1086
1087         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1088                 task->tk_pid, task->tk_status);
1089
1090         /*
1091          * ->write_done will attempt to use post-op attributes to detect
1092          * conflicting writes by other clients.  A strict interpretation
1093          * of close-to-open would allow us to continue caching even if
1094          * another writer had changed the file, but some applications
1095          * depend on tighter cache coherency when writing.
1096          */
1097         status = NFS_PROTO(data->inode)->write_done(task, data);
1098         if (status != 0)
1099                 return status;
1100         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1101
1102 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1103         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1104                 /* We tried a write call, but the server did not
1105                  * commit data to stable storage even though we
1106                  * requested it.
1107                  * Note: There is a known bug in Tru64 < 5.0 in which
1108                  *       the server reports NFS_DATA_SYNC, but performs
1109                  *       NFS_FILE_SYNC. We therefore implement this checking
1110                  *       as a dprintk() in order to avoid filling syslog.
1111                  */
1112                 static unsigned long    complain;
1113
1114                 if (time_before(complain, jiffies)) {
1115                         dprintk("NFS: faulty NFS server %s:"
1116                                 " (committed = %d) != (stable = %d)\n",
1117                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1118                                 resp->verf->committed, argp->stable);
1119                         complain = jiffies + 300 * HZ;
1120                 }
1121         }
1122 #endif
1123         /* Is this a short write? */
1124         if (task->tk_status >= 0 && resp->count < argp->count) {
1125                 static unsigned long    complain;
1126
1127                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1128
1129                 /* Has the server at least made some progress? */
1130                 if (resp->count != 0) {
1131                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1132                         if (resp->verf->committed != NFS_UNSTABLE) {
1133                                 /* Resend from where the server left off */
1134                                 argp->offset += resp->count;
1135                                 argp->pgbase += resp->count;
1136                                 argp->count -= resp->count;
1137                         } else {
1138                                 /* Resend as a stable write in order to avoid
1139                                  * headaches in the case of a server crash.
1140                                  */
1141                                 argp->stable = NFS_FILE_SYNC;
1142                         }
1143                         rpc_restart_call(task);
1144                         return -EAGAIN;
1145                 }
1146                 if (time_before(complain, jiffies)) {
1147                         printk(KERN_WARNING
1148                                "NFS: Server wrote zero bytes, expected %u.\n",
1149                                         argp->count);
1150                         complain = jiffies + 300 * HZ;
1151                 }
1152                 /* Can't do anything about it except throw an error. */
1153                 task->tk_status = -EIO;
1154         }
1155         return 0;
1156 }
1157
1158
1159 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1160 void nfs_commit_release(void *wdata)
1161 {
1162         nfs_commit_free(wdata);
1163 }
1164
1165 /*
1166  * Set up the argument/result storage required for the RPC call.
1167  */
1168 static void nfs_commit_rpcsetup(struct list_head *head,
1169                 struct nfs_write_data *data,
1170                 int how)
1171 {
1172         struct nfs_page         *first;
1173         struct inode            *inode;
1174         int flags;
1175
1176         /* Set up the RPC argument and reply structs
1177          * NB: take care not to mess about with data->commit et al. */
1178
1179         list_splice_init(head, &data->pages);
1180         first = nfs_list_entry(data->pages.next);
1181         inode = first->wb_context->dentry->d_inode;
1182
1183         data->inode       = inode;
1184         data->cred        = first->wb_context->cred;
1185
1186         data->args.fh     = NFS_FH(data->inode);
1187         /* Note: we always request a commit of the entire inode */
1188         data->args.offset = 0;
1189         data->args.count  = 0;
1190         data->res.count   = 0;
1191         data->res.fattr   = &data->fattr;
1192         data->res.verf    = &data->verf;
1193         nfs_fattr_init(&data->fattr);
1194
1195         /* Set up the initial task struct.  */
1196         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1197         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1198         NFS_PROTO(inode)->commit_setup(data, how);
1199
1200         data->task.tk_priority = flush_task_priority(how);
1201         data->task.tk_cookie = (unsigned long)inode;
1202
1203         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1204 }
1205
1206 /*
1207  * Commit dirty pages
1208  */
1209 static int
1210 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1211 {
1212         struct nfs_write_data   *data;
1213         struct nfs_page         *req;
1214
1215         data = nfs_commit_alloc();
1216
1217         if (!data)
1218                 goto out_bad;
1219
1220         /* Set up the argument struct */
1221         nfs_commit_rpcsetup(head, data, how);
1222
1223         nfs_execute_write(data);
1224         return 0;
1225  out_bad:
1226         while (!list_empty(head)) {
1227                 req = nfs_list_entry(head->next);
1228                 nfs_list_remove_request(req);
1229                 nfs_mark_request_commit(req);
1230                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1231                 nfs_clear_page_writeback(req);
1232         }
1233         return -ENOMEM;
1234 }
1235
1236 /*
1237  * COMMIT call returned
1238  */
1239 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1240 {
1241         struct nfs_write_data   *data = calldata;
1242         struct nfs_page         *req;
1243
1244         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1245                                 task->tk_pid, task->tk_status);
1246
1247         /* Call the NFS version-specific code */
1248         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1249                 return;
1250
1251         while (!list_empty(&data->pages)) {
1252                 req = nfs_list_entry(data->pages.next);
1253                 nfs_list_remove_request(req);
1254                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1255                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1256
1257                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1258                         req->wb_context->dentry->d_inode->i_sb->s_id,
1259                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1260                         req->wb_bytes,
1261                         (long long)req_offset(req));
1262                 if (task->tk_status < 0) {
1263                         req->wb_context->error = task->tk_status;
1264                         nfs_inode_remove_request(req);
1265                         dprintk(", error = %d\n", task->tk_status);
1266                         goto next;
1267                 }
1268
1269                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1270                  * returned by the server against all stored verfs. */
1271                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1272                         /* We have a match */
1273                         nfs_inode_remove_request(req);
1274                         dprintk(" OK\n");
1275                         goto next;
1276                 }
1277                 /* We have a mismatch. Write the page again */
1278                 dprintk(" mismatch\n");
1279                 nfs_redirty_request(req);
1280         next:
1281                 nfs_clear_page_writeback(req);
1282         }
1283 }
1284
1285 static const struct rpc_call_ops nfs_commit_ops = {
1286         .rpc_call_done = nfs_commit_done,
1287         .rpc_release = nfs_commit_release,
1288 };
1289 #else
1290 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1291 {
1292         return 0;
1293 }
1294 #endif
1295
1296 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1297 {
1298         struct nfs_inode *nfsi = NFS_I(mapping->host);
1299         LIST_HEAD(head);
1300         long res;
1301
1302         spin_lock(&nfsi->req_lock);
1303         res = nfs_scan_dirty(mapping, wbc, &head);
1304         spin_unlock(&nfsi->req_lock);
1305         if (res) {
1306                 int error = nfs_flush_list(mapping->host, &head, res, how);
1307                 if (error < 0)
1308                         return error;
1309         }
1310         return res;
1311 }
1312
1313 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1314 int nfs_commit_inode(struct inode *inode, int how)
1315 {
1316         struct nfs_inode *nfsi = NFS_I(inode);
1317         LIST_HEAD(head);
1318         int res;
1319
1320         spin_lock(&nfsi->req_lock);
1321         res = nfs_scan_commit(inode, &head, 0, 0);
1322         spin_unlock(&nfsi->req_lock);
1323         if (res) {
1324                 int error = nfs_commit_list(inode, &head, how);
1325                 if (error < 0)
1326                         return error;
1327         }
1328         return res;
1329 }
1330 #endif
1331
1332 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1333 {
1334         struct inode *inode = mapping->host;
1335         struct nfs_inode *nfsi = NFS_I(inode);
1336         unsigned long idx_start, idx_end;
1337         unsigned int npages = 0;
1338         LIST_HEAD(head);
1339         int nocommit = how & FLUSH_NOCOMMIT;
1340         long pages, ret;
1341
1342         /* FIXME */
1343         if (wbc->range_cyclic)
1344                 idx_start = 0;
1345         else {
1346                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1347                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1348                 if (idx_end > idx_start) {
1349                         unsigned long l_npages = 1 + idx_end - idx_start;
1350                         npages = l_npages;
1351                         if (sizeof(npages) != sizeof(l_npages) &&
1352                                         (unsigned long)npages != l_npages)
1353                                 npages = 0;
1354                 }
1355         }
1356         how &= ~FLUSH_NOCOMMIT;
1357         spin_lock(&nfsi->req_lock);
1358         do {
1359                 wbc->pages_skipped = 0;
1360                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1361                 if (ret != 0)
1362                         continue;
1363                 pages = nfs_scan_dirty(mapping, wbc, &head);
1364                 if (pages != 0) {
1365                         spin_unlock(&nfsi->req_lock);
1366                         if (how & FLUSH_INVALIDATE) {
1367                                 nfs_cancel_dirty_list(&head);
1368                                 ret = pages;
1369                         } else
1370                                 ret = nfs_flush_list(inode, &head, pages, how);
1371                         spin_lock(&nfsi->req_lock);
1372                         continue;
1373                 }
1374                 if (wbc->pages_skipped != 0)
1375                         continue;
1376                 if (nocommit)
1377                         break;
1378                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1379                 if (pages == 0) {
1380                         if (wbc->pages_skipped != 0)
1381                                 continue;
1382                         break;
1383                 }
1384                 if (how & FLUSH_INVALIDATE) {
1385                         spin_unlock(&nfsi->req_lock);
1386                         nfs_cancel_commit_list(&head);
1387                         ret = pages;
1388                         spin_lock(&nfsi->req_lock);
1389                         continue;
1390                 }
1391                 pages += nfs_scan_commit(inode, &head, 0, 0);
1392                 spin_unlock(&nfsi->req_lock);
1393                 ret = nfs_commit_list(inode, &head, how);
1394                 spin_lock(&nfsi->req_lock);
1395         } while (ret >= 0);
1396         spin_unlock(&nfsi->req_lock);
1397         return ret;
1398 }
1399
1400 /*
1401  * flush the inode to disk.
1402  */
1403 int nfs_wb_all(struct inode *inode)
1404 {
1405         struct address_space *mapping = inode->i_mapping;
1406         struct writeback_control wbc = {
1407                 .bdi = mapping->backing_dev_info,
1408                 .sync_mode = WB_SYNC_ALL,
1409                 .nr_to_write = LONG_MAX,
1410                 .for_writepages = 1,
1411                 .range_cyclic = 1,
1412         };
1413         int ret;
1414
1415         ret = generic_writepages(mapping, &wbc);
1416         if (ret < 0)
1417                 goto out;
1418         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1419         if (ret >= 0)
1420                 return 0;
1421 out:
1422         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1423         return ret;
1424 }
1425
1426 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1427 {
1428         struct writeback_control wbc = {
1429                 .bdi = mapping->backing_dev_info,
1430                 .sync_mode = WB_SYNC_ALL,
1431                 .nr_to_write = LONG_MAX,
1432                 .range_start = range_start,
1433                 .range_end = range_end,
1434                 .for_writepages = 1,
1435         };
1436         int ret;
1437
1438         if (!(how & FLUSH_NOWRITEPAGE)) {
1439                 ret = generic_writepages(mapping, &wbc);
1440                 if (ret < 0)
1441                         goto out;
1442         }
1443         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1444         if (ret >= 0)
1445                 return 0;
1446 out:
1447         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1448         return ret;
1449 }
1450
1451 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1452 {
1453         loff_t range_start = page_offset(page);
1454         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1455         struct writeback_control wbc = {
1456                 .bdi = page->mapping->backing_dev_info,
1457                 .sync_mode = WB_SYNC_ALL,
1458                 .nr_to_write = LONG_MAX,
1459                 .range_start = range_start,
1460                 .range_end = range_end,
1461         };
1462         int ret;
1463
1464         BUG_ON(!PageLocked(page));
1465         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1466                 ret = nfs_writepage_locked(page, &wbc);
1467                 if (ret < 0)
1468                         goto out;
1469         }
1470         if (!PagePrivate(page))
1471                 return 0;
1472         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1473         if (ret >= 0)
1474                 return 0;
1475 out:
1476         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1477         return ret;
1478 }
1479
1480 /*
1481  * Write back all requests on one page - we do this before reading it.
1482  */
1483 int nfs_wb_page(struct inode *inode, struct page* page)
1484 {
1485         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1486 }
1487
1488 int nfs_set_page_dirty(struct page *page)
1489 {
1490         struct address_space *mapping = page->mapping;
1491         struct inode *inode;
1492         spinlock_t *req_lock;
1493         struct nfs_page *req;
1494         int ret;
1495
1496         if (!mapping)
1497                 goto out_raced;
1498         inode = mapping->host;
1499         if (!inode)
1500                 goto out_raced;
1501         req_lock = &NFS_I(inode)->req_lock;
1502         spin_lock(req_lock);
1503         req = nfs_page_find_request_locked(page);
1504         if (req != NULL) {
1505                 /* Mark any existing write requests for flushing */
1506                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1507                 spin_unlock(req_lock);
1508                 nfs_release_request(req);
1509                 return ret;
1510         }
1511         ret = __set_page_dirty_nobuffers(page);
1512         spin_unlock(req_lock);
1513         return ret;
1514 out_raced:
1515         return !TestSetPageDirty(page);
1516 }
1517
1518
1519 int __init nfs_init_writepagecache(void)
1520 {
1521         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1522                                              sizeof(struct nfs_write_data),
1523                                              0, SLAB_HWCACHE_ALIGN,
1524                                              NULL, NULL);
1525         if (nfs_wdata_cachep == NULL)
1526                 return -ENOMEM;
1527
1528         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1529                                                      nfs_wdata_cachep);
1530         if (nfs_wdata_mempool == NULL)
1531                 return -ENOMEM;
1532
1533         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1534                                                       nfs_wdata_cachep);
1535         if (nfs_commit_mempool == NULL)
1536                 return -ENOMEM;
1537
1538         /*
1539          * NFS congestion size, scale with available memory.
1540          *
1541          *  64MB:    8192k
1542          * 128MB:   11585k
1543          * 256MB:   16384k
1544          * 512MB:   23170k
1545          *   1GB:   32768k
1546          *   2GB:   46340k
1547          *   4GB:   65536k
1548          *   8GB:   92681k
1549          *  16GB:  131072k
1550          *
1551          * This allows larger machines to have larger/more transfers.
1552          * Limit the default to 256M
1553          */
1554         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1555         if (nfs_congestion_kb > 256*1024)
1556                 nfs_congestion_kb = 256*1024;
1557
1558         return 0;
1559 }
1560
1561 void nfs_destroy_writepagecache(void)
1562 {
1563         mempool_destroy(nfs_commit_mempool);
1564         mempool_destroy(nfs_wdata_mempool);
1565         kmem_cache_destroy(nfs_wdata_cachep);
1566 }
1567