err.no 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, 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_end_page_writeback(req->wb_page);
 895         nfs_clear_page_writeback(req);
 896         return -ENOMEM;
 897 }
 898
 899 /*
 900  * Create an RPC task for the given write request and kick it.
 901  * The page must have been locked by the caller.
 902  *
 903  * It may happen that the page we're passed is not marked dirty.
 904  * This is the case if nfs_updatepage detects a conflicting request
 905  * that has been written but not committed.
 906  */
 907 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 908 {
 909         struct nfs_page         *req;
 910         struct page             **pages;
 911         struct nfs_write_data   *data;
 912         unsigned int            count;
 913
 914         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 915         if (!data)
 916                 goto out_bad;
 917
 918         pages = data->pagevec;
 919         count = 0;
 920         while (!list_empty(head)) {
 921                 req = nfs_list_entry(head->next);
 922                 nfs_list_remove_request(req);
 923                 nfs_list_add_request(req, &data->pages);
 924                 ClearPageError(req->wb_page);
 925                 *pages++ = req->wb_page;
 926                 count += req->wb_bytes;
 927         }
 928         req = nfs_list_entry(data->pages.next);
 929
 930         /* Set up the argument struct */
 931         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 932
 933         nfs_execute_write(data);
 934         return 0;
 935  out_bad:
 936         while (!list_empty(head)) {
 937                 struct nfs_page *req = nfs_list_entry(head->next);
 938                 nfs_list_remove_request(req);
 939                 nfs_redirty_request(req);
 940                 nfs_end_page_writeback(req->wb_page);
 941                 nfs_clear_page_writeback(req);
 942         }
 943         return -ENOMEM;
 944 }
 945
 946 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 947 {
 948         LIST_HEAD(one_request);
 949         int (*flush_one)(struct inode *, struct list_head *, int);
 950         struct nfs_page *req;
 951         int wpages = NFS_SERVER(inode)->wpages;
 952         int wsize = NFS_SERVER(inode)->wsize;
 953         int error;
 954
 955         flush_one = nfs_flush_one;
 956         if (wsize < PAGE_CACHE_SIZE)
 957                 flush_one = nfs_flush_multi;
 958         /* For single writes, FLUSH_STABLE is more efficient */
 959         if (npages <= wpages && npages == NFS_I(inode)->npages
 960                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 961                 how |= FLUSH_STABLE;
 962
 963         do {
 964                 nfs_coalesce_requests(head, &one_request, wpages);
 965                 req = nfs_list_entry(one_request.next);
 966                 error = flush_one(inode, &one_request, how);
 967                 if (error < 0)
 968                         goto out_err;
 969         } while (!list_empty(head));
 970         return 0;
 971 out_err:
 972         while (!list_empty(head)) {
 973                 req = nfs_list_entry(head->next);
 974                 nfs_list_remove_request(req);
 975                 nfs_redirty_request(req);
 976                 nfs_end_page_writeback(req->wb_page);
 977                 nfs_clear_page_writeback(req);
 978         }
 979         return error;
 980 }
 981
 982 /*
 983  * Handle a write reply that flushed part of a page.
 984  */
 985 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 986 {
 987         struct nfs_write_data   *data = calldata;
 988         struct nfs_page         *req = data->req;
 989         struct page             *page = req->wb_page;
 990
 991         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 992                 req->wb_context->dentry->d_inode->i_sb->s_id,
 993                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 994                 req->wb_bytes,
 995                 (long long)req_offset(req));
 996
 997         if (nfs_writeback_done(task, data) != 0)
 998                 return;
 999
1000         if (task->tk_status < 0) {
1001                 nfs_set_pageerror(page);
1002                 req->wb_context->error = task->tk_status;
1003                 dprintk(", error = %d\n", task->tk_status);
1004                 goto out;
1005         }
1006
1007         if (nfs_write_need_commit(data)) {
1008                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
1009
1010                 spin_lock(req_lock);
1011                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1012                         /* Do nothing we need to resend the writes */
1013                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1014                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1015                         dprintk(" defer commit\n");
1016                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1017                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1018                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1019                         dprintk(" server reboot detected\n");
1020                 }
1021                 spin_unlock(req_lock);
1022         } else
1023                 dprintk(" OK\n");
1024
1025 out:
1026         if (atomic_dec_and_test(&req->wb_complete))
1027                 nfs_writepage_release(req);
1028 }
1029
1030 static const struct rpc_call_ops nfs_write_partial_ops = {
1031         .rpc_call_done = nfs_writeback_done_partial,
1032         .rpc_release = nfs_writedata_release,
1033 };
1034
1035 /*
1036  * Handle a write reply that flushes a whole page.
1037  *
1038  * FIXME: There is an inherent race with invalidate_inode_pages and
1039  *        writebacks since the page->count is kept > 1 for as long
1040  *        as the page has a write request pending.
1041  */
1042 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1043 {
1044         struct nfs_write_data   *data = calldata;
1045         struct nfs_page         *req;
1046         struct page             *page;
1047
1048         if (nfs_writeback_done(task, data) != 0)
1049                 return;
1050
1051         /* Update attributes as result of writeback. */
1052         while (!list_empty(&data->pages)) {
1053                 req = nfs_list_entry(data->pages.next);
1054                 nfs_list_remove_request(req);
1055                 page = req->wb_page;
1056
1057                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1058                         req->wb_context->dentry->d_inode->i_sb->s_id,
1059                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1060                         req->wb_bytes,
1061                         (long long)req_offset(req));
1062
1063                 if (task->tk_status < 0) {
1064                         nfs_set_pageerror(page);
1065                         req->wb_context->error = task->tk_status;
1066                         dprintk(", error = %d\n", task->tk_status);
1067                         goto remove_request;
1068                 }
1069
1070                 if (nfs_write_need_commit(data)) {
1071                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1072                         nfs_mark_request_commit(req);
1073                         nfs_end_page_writeback(page);
1074                         dprintk(" marked for commit\n");
1075                         goto next;
1076                 }
1077                 dprintk(" OK\n");
1078 remove_request:
1079                 nfs_end_page_writeback(page);
1080                 nfs_inode_remove_request(req);
1081         next:
1082                 nfs_clear_page_writeback(req);
1083         }
1084 }
1085
1086 static const struct rpc_call_ops nfs_write_full_ops = {
1087         .rpc_call_done = nfs_writeback_done_full,
1088         .rpc_release = nfs_writedata_release,
1089 };
1090
1091
1092 /*
1093  * This function is called when the WRITE call is complete.
1094  */
1095 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1096 {
1097         struct nfs_writeargs    *argp = &data->args;
1098         struct nfs_writeres     *resp = &data->res;
1099         int status;
1100
1101         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1102                 task->tk_pid, task->tk_status);
1103
1104         /*
1105          * ->write_done will attempt to use post-op attributes to detect
1106          * conflicting writes by other clients.  A strict interpretation
1107          * of close-to-open would allow us to continue caching even if
1108          * another writer had changed the file, but some applications
1109          * depend on tighter cache coherency when writing.
1110          */
1111         status = NFS_PROTO(data->inode)->write_done(task, data);
1112         if (status != 0)
1113                 return status;
1114         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1115
1116 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1117         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1118                 /* We tried a write call, but the server did not
1119                  * commit data to stable storage even though we
1120                  * requested it.
1121                  * Note: There is a known bug in Tru64 < 5.0 in which
1122                  *       the server reports NFS_DATA_SYNC, but performs
1123                  *       NFS_FILE_SYNC. We therefore implement this checking
1124                  *       as a dprintk() in order to avoid filling syslog.
1125                  */
1126                 static unsigned long    complain;
1127
1128                 if (time_before(complain, jiffies)) {
1129                         dprintk("NFS: faulty NFS server %s:"
1130                                 " (committed = %d) != (stable = %d)\n",
1131                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1132                                 resp->verf->committed, argp->stable);
1133                         complain = jiffies + 300 * HZ;
1134                 }
1135         }
1136 #endif
1137         /* Is this a short write? */
1138         if (task->tk_status >= 0 && resp->count < argp->count) {
1139                 static unsigned long    complain;
1140
1141                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1142
1143                 /* Has the server at least made some progress? */
1144                 if (resp->count != 0) {
1145                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1146                         if (resp->verf->committed != NFS_UNSTABLE) {
1147                                 /* Resend from where the server left off */
1148                                 argp->offset += resp->count;
1149                                 argp->pgbase += resp->count;
1150                                 argp->count -= resp->count;
1151                         } else {
1152                                 /* Resend as a stable write in order to avoid
1153                                  * headaches in the case of a server crash.
1154                                  */
1155                                 argp->stable = NFS_FILE_SYNC;
1156                         }
1157                         rpc_restart_call(task);
1158                         return -EAGAIN;
1159                 }
1160                 if (time_before(complain, jiffies)) {
1161                         printk(KERN_WARNING
1162                                "NFS: Server wrote zero bytes, expected %u.\n",
1163                                         argp->count);
1164                         complain = jiffies + 300 * HZ;
1165                 }
1166                 /* Can't do anything about it except throw an error. */
1167                 task->tk_status = -EIO;
1168         }
1169         return 0;
1170 }
1171
1172
1173 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1174 void nfs_commit_release(void *wdata)
1175 {
1176         nfs_commit_free(wdata);
1177 }
1178
1179 /*
1180  * Set up the argument/result storage required for the RPC call.
1181  */
1182 static void nfs_commit_rpcsetup(struct list_head *head,
1183                 struct nfs_write_data *data,
1184                 int how)
1185 {
1186         struct nfs_page         *first;
1187         struct inode            *inode;
1188         int flags;
1189
1190         /* Set up the RPC argument and reply structs
1191          * NB: take care not to mess about with data->commit et al. */
1192
1193         list_splice_init(head, &data->pages);
1194         first = nfs_list_entry(data->pages.next);
1195         inode = first->wb_context->dentry->d_inode;
1196
1197         data->inode       = inode;
1198         data->cred        = first->wb_context->cred;
1199
1200         data->args.fh     = NFS_FH(data->inode);
1201         /* Note: we always request a commit of the entire inode */
1202         data->args.offset = 0;
1203         data->args.count  = 0;
1204         data->res.count   = 0;
1205         data->res.fattr   = &data->fattr;
1206         data->res.verf    = &data->verf;
1207         nfs_fattr_init(&data->fattr);
1208
1209         /* Set up the initial task struct.  */
1210         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1211         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1212         NFS_PROTO(inode)->commit_setup(data, how);
1213
1214         data->task.tk_priority = flush_task_priority(how);
1215         data->task.tk_cookie = (unsigned long)inode;
1216
1217         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1218 }
1219
1220 /*
1221  * Commit dirty pages
1222  */
1223 static int
1224 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1225 {
1226         struct nfs_write_data   *data;
1227         struct nfs_page         *req;
1228
1229         data = nfs_commit_alloc();
1230
1231         if (!data)
1232                 goto out_bad;
1233
1234         /* Set up the argument struct */
1235         nfs_commit_rpcsetup(head, data, how);
1236
1237         nfs_execute_write(data);
1238         return 0;
1239  out_bad:
1240         while (!list_empty(head)) {
1241                 req = nfs_list_entry(head->next);
1242                 nfs_list_remove_request(req);
1243                 nfs_mark_request_commit(req);
1244                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1245                 nfs_clear_page_writeback(req);
1246         }
1247         return -ENOMEM;
1248 }
1249
1250 /*
1251  * COMMIT call returned
1252  */
1253 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1254 {
1255         struct nfs_write_data   *data = calldata;
1256         struct nfs_page         *req;
1257
1258         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1259                                 task->tk_pid, task->tk_status);
1260
1261         /* Call the NFS version-specific code */
1262         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1263                 return;
1264
1265         while (!list_empty(&data->pages)) {
1266                 req = nfs_list_entry(data->pages.next);
1267                 nfs_list_remove_request(req);
1268                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1269                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1270
1271                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1272                         req->wb_context->dentry->d_inode->i_sb->s_id,
1273                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1274                         req->wb_bytes,
1275                         (long long)req_offset(req));
1276                 if (task->tk_status < 0) {
1277                         req->wb_context->error = task->tk_status;
1278                         nfs_inode_remove_request(req);
1279                         dprintk(", error = %d\n", task->tk_status);
1280                         goto next;
1281                 }
1282
1283                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1284                  * returned by the server against all stored verfs. */
1285                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1286                         /* We have a match */
1287                         nfs_inode_remove_request(req);
1288                         dprintk(" OK\n");
1289                         goto next;
1290                 }
1291                 /* We have a mismatch. Write the page again */
1292                 dprintk(" mismatch\n");
1293                 nfs_redirty_request(req);
1294         next:
1295                 nfs_clear_page_writeback(req);
1296         }
1297 }
1298
1299 static const struct rpc_call_ops nfs_commit_ops = {
1300         .rpc_call_done = nfs_commit_done,
1301         .rpc_release = nfs_commit_release,
1302 };
1303 #else
1304 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1305 {
1306         return 0;
1307 }
1308 #endif
1309
1310 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1311 {
1312         struct nfs_inode *nfsi = NFS_I(mapping->host);
1313         LIST_HEAD(head);
1314         long res;
1315
1316         spin_lock(&nfsi->req_lock);
1317         res = nfs_scan_dirty(mapping, wbc, &head);
1318         spin_unlock(&nfsi->req_lock);
1319         if (res) {
1320                 int error = nfs_flush_list(mapping->host, &head, res, how);
1321                 if (error < 0)
1322                         return error;
1323         }
1324         return res;
1325 }
1326
1327 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1328 int nfs_commit_inode(struct inode *inode, int how)
1329 {
1330         struct nfs_inode *nfsi = NFS_I(inode);
1331         LIST_HEAD(head);
1332         int res;
1333
1334         spin_lock(&nfsi->req_lock);
1335         res = nfs_scan_commit(inode, &head, 0, 0);
1336         spin_unlock(&nfsi->req_lock);
1337         if (res) {
1338                 int error = nfs_commit_list(inode, &head, how);
1339                 if (error < 0)
1340                         return error;
1341         }
1342         return res;
1343 }
1344 #endif
1345
1346 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1347 {
1348         struct inode *inode = mapping->host;
1349         struct nfs_inode *nfsi = NFS_I(inode);
1350         unsigned long idx_start, idx_end;
1351         unsigned int npages = 0;
1352         LIST_HEAD(head);
1353         int nocommit = how & FLUSH_NOCOMMIT;
1354         long pages, ret;
1355
1356         /* FIXME */
1357         if (wbc->range_cyclic)
1358                 idx_start = 0;
1359         else {
1360                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1361                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1362                 if (idx_end > idx_start) {
1363                         unsigned long l_npages = 1 + idx_end - idx_start;
1364                         npages = l_npages;
1365                         if (sizeof(npages) != sizeof(l_npages) &&
1366                                         (unsigned long)npages != l_npages)
1367                                 npages = 0;
1368                 }
1369         }
1370         how &= ~FLUSH_NOCOMMIT;
1371         spin_lock(&nfsi->req_lock);
1372         do {
1373                 wbc->pages_skipped = 0;
1374                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1375                 if (ret != 0)
1376                         continue;
1377                 pages = nfs_scan_dirty(mapping, wbc, &head);
1378                 if (pages != 0) {
1379                         spin_unlock(&nfsi->req_lock);
1380                         if (how & FLUSH_INVALIDATE) {
1381                                 nfs_cancel_dirty_list(&head);
1382                                 ret = pages;
1383                         } else
1384                                 ret = nfs_flush_list(inode, &head, pages, how);
1385                         spin_lock(&nfsi->req_lock);
1386                         continue;
1387                 }
1388                 if (wbc->pages_skipped != 0)
1389                         continue;
1390                 if (nocommit)
1391                         break;
1392                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1393                 if (pages == 0) {
1394                         if (wbc->pages_skipped != 0)
1395                                 continue;
1396                         break;
1397                 }
1398                 if (how & FLUSH_INVALIDATE) {
1399                         spin_unlock(&nfsi->req_lock);
1400                         nfs_cancel_commit_list(&head);
1401                         ret = pages;
1402                         spin_lock(&nfsi->req_lock);
1403                         continue;
1404                 }
1405                 pages += nfs_scan_commit(inode, &head, 0, 0);
1406                 spin_unlock(&nfsi->req_lock);
1407                 ret = nfs_commit_list(inode, &head, how);
1408                 spin_lock(&nfsi->req_lock);
1409         } while (ret >= 0);
1410         spin_unlock(&nfsi->req_lock);
1411         return ret;
1412 }
1413
1414 /*
1415  * flush the inode to disk.
1416  */
1417 int nfs_wb_all(struct inode *inode)
1418 {
1419         struct address_space *mapping = inode->i_mapping;
1420         struct writeback_control wbc = {
1421                 .bdi = mapping->backing_dev_info,
1422                 .sync_mode = WB_SYNC_ALL,
1423                 .nr_to_write = LONG_MAX,
1424                 .for_writepages = 1,
1425                 .range_cyclic = 1,
1426         };
1427         int ret;
1428
1429         ret = generic_writepages(mapping, &wbc);
1430         if (ret < 0)
1431                 goto out;
1432         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1433         if (ret >= 0)
1434                 return 0;
1435 out:
1436         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1437         return ret;
1438 }
1439
1440 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1441 {
1442         struct writeback_control wbc = {
1443                 .bdi = mapping->backing_dev_info,
1444                 .sync_mode = WB_SYNC_ALL,
1445                 .nr_to_write = LONG_MAX,
1446                 .range_start = range_start,
1447                 .range_end = range_end,
1448                 .for_writepages = 1,
1449         };
1450         int ret;
1451
1452         if (!(how & FLUSH_NOWRITEPAGE)) {
1453                 ret = generic_writepages(mapping, &wbc);
1454                 if (ret < 0)
1455                         goto out;
1456         }
1457         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1458         if (ret >= 0)
1459                 return 0;
1460 out:
1461         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1462         return ret;
1463 }
1464
1465 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1466 {
1467         loff_t range_start = page_offset(page);
1468         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1469         struct writeback_control wbc = {
1470                 .bdi = page->mapping->backing_dev_info,
1471                 .sync_mode = WB_SYNC_ALL,
1472                 .nr_to_write = LONG_MAX,
1473                 .range_start = range_start,
1474                 .range_end = range_end,
1475         };
1476         int ret;
1477
1478         BUG_ON(!PageLocked(page));
1479         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1480                 ret = nfs_writepage_locked(page, &wbc);
1481                 if (ret < 0)
1482                         goto out;
1483         }
1484         if (!PagePrivate(page))
1485                 return 0;
1486         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1487         if (ret >= 0)
1488                 return 0;
1489 out:
1490         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1491         return ret;
1492 }
1493
1494 /*
1495  * Write back all requests on one page - we do this before reading it.
1496  */
1497 int nfs_wb_page(struct inode *inode, struct page* page)
1498 {
1499         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1500 }
1501
1502 int nfs_set_page_dirty(struct page *page)
1503 {
1504         struct address_space *mapping = page->mapping;
1505         struct inode *inode;
1506         spinlock_t *req_lock;
1507         struct nfs_page *req;
1508         int ret;
1509
1510         if (!mapping)
1511                 goto out_raced;
1512         inode = mapping->host;
1513         if (!inode)
1514                 goto out_raced;
1515         req_lock = &NFS_I(inode)->req_lock;
1516         spin_lock(req_lock);
1517         req = nfs_page_find_request_locked(page);
1518         if (req != NULL) {
1519                 /* Mark any existing write requests for flushing */
1520                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1521                 spin_unlock(req_lock);
1522                 nfs_release_request(req);
1523                 return ret;
1524         }
1525         ret = __set_page_dirty_nobuffers(page);
1526         spin_unlock(req_lock);
1527         return ret;
1528 out_raced:
1529         return !TestSetPageDirty(page);
1530 }
1531
1532
1533 int __init nfs_init_writepagecache(void)
1534 {
1535         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1536                                              sizeof(struct nfs_write_data),
1537                                              0, SLAB_HWCACHE_ALIGN,
1538                                              NULL, NULL);
1539         if (nfs_wdata_cachep == NULL)
1540                 return -ENOMEM;
1541
1542         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1543                                                      nfs_wdata_cachep);
1544         if (nfs_wdata_mempool == NULL)
1545                 return -ENOMEM;
1546
1547         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1548                                                       nfs_wdata_cachep);
1549         if (nfs_commit_mempool == NULL)
1550                 return -ENOMEM;
1551
1552         /*
1553          * NFS congestion size, scale with available memory.
1554          *
1555          *  64MB:    8192k
1556          * 128MB:   11585k
1557          * 256MB:   16384k
1558          * 512MB:   23170k
1559          *   1GB:   32768k
1560          *   2GB:   46340k
1561          *   4GB:   65536k
1562          *   8GB:   92681k
1563          *  16GB:  131072k
1564          *
1565          * This allows larger machines to have larger/more transfers.
1566          * Limit the default to 256M
1567          */
1568         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1569         if (nfs_congestion_kb > 256*1024)
1570                 nfs_congestion_kb = 256*1024;
1571
1572         return 0;
1573 }
1574
1575 void nfs_destroy_writepagecache(void)
1576 {
1577         mempool_destroy(nfs_commit_mempool);
1578         mempool_destroy(nfs_wdata_mempool);
1579         kmem_cache_destroy(nfs_wdata_cachep);
1580 }
1581