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 static int nfs_wait_on_write_congestion(struct address_space *mapping)
 598 {
 599         struct inode *inode = mapping->host;
 600         struct backing_dev_info *bdi = mapping->backing_dev_info;
 601         int ret = 0;
 602
 603         might_sleep();
 604
 605         if (!bdi_write_congested(bdi))
 606                 return 0;
 607
 608         nfs_inc_stats(inode, NFSIOS_CONGESTIONWAIT);
 609
 610         do {
 611                 struct rpc_clnt *clnt = NFS_CLIENT(inode);
 612                 sigset_t oldset;
 613
 614                 rpc_clnt_sigmask(clnt, &oldset);
 615                 ret = congestion_wait_interruptible(WRITE, HZ/10);
 616                 rpc_clnt_sigunmask(clnt, &oldset);
 617                 if (ret == -ERESTARTSYS)
 618                         break;
 619                 ret = 0;
 620         } while (bdi_write_congested(bdi));
 621
 622         return ret;
 623 }
 624
 625 /*
 626  * Try to update any existing write request, or create one if there is none.
 627  * In order to match, the request's credentials must match those of
 628  * the calling process.
 629  *
 630  * Note: Should always be called with the Page Lock held!
 631  */
 632 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 633                 struct page *page, unsigned int offset, unsigned int bytes)
 634 {
 635         struct address_space *mapping = page->mapping;
 636         struct inode *inode = mapping->host;
 637         struct nfs_inode *nfsi = NFS_I(inode);
 638         struct nfs_page         *req, *new = NULL;
 639         unsigned long           rqend, end;
 640
 641         end = offset + bytes;
 642
 643         if (nfs_wait_on_write_congestion(mapping))
 644                 return ERR_PTR(-ERESTARTSYS);
 645         for (;;) {
 646                 /* Loop over all inode entries and see if we find
 647                  * A request for the page we wish to update
 648                  */
 649                 spin_lock(&nfsi->req_lock);
 650                 req = nfs_page_find_request_locked(page);
 651                 if (req) {
 652                         if (!nfs_lock_request_dontget(req)) {
 653                                 int error;
 654
 655                                 spin_unlock(&nfsi->req_lock);
 656                                 error = nfs_wait_on_request(req);
 657                                 nfs_release_request(req);
 658                                 if (error < 0) {
 659                                         if (new)
 660                                                 nfs_release_request(new);
 661                                         return ERR_PTR(error);
 662                                 }
 663                                 continue;
 664                         }
 665                         spin_unlock(&nfsi->req_lock);
 666                         if (new)
 667                                 nfs_release_request(new);
 668                         break;
 669                 }
 670
 671                 if (new) {
 672                         int error;
 673                         nfs_lock_request_dontget(new);
 674                         error = nfs_inode_add_request(inode, new);
 675                         if (error) {
 676                                 spin_unlock(&nfsi->req_lock);
 677                                 nfs_unlock_request(new);
 678                                 return ERR_PTR(error);
 679                         }
 680                         spin_unlock(&nfsi->req_lock);
 681                         return new;
 682                 }
 683                 spin_unlock(&nfsi->req_lock);
 684
 685                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 686                 if (IS_ERR(new))
 687                         return new;
 688         }
 689
 690         /* We have a request for our page.
 691          * If the creds don't match, or the
 692          * page addresses don't match,
 693          * tell the caller to wait on the conflicting
 694          * request.
 695          */
 696         rqend = req->wb_offset + req->wb_bytes;
 697         if (req->wb_context != ctx
 698             || req->wb_page != page
 699             || !nfs_dirty_request(req)
 700             || offset > rqend || end < req->wb_offset) {
 701                 nfs_unlock_request(req);
 702                 return ERR_PTR(-EBUSY);
 703         }
 704
 705         /* Okay, the request matches. Update the region */
 706         if (offset < req->wb_offset) {
 707                 req->wb_offset = offset;
 708                 req->wb_pgbase = offset;
 709                 req->wb_bytes = rqend - req->wb_offset;
 710         }
 711
 712         if (end > rqend)
 713                 req->wb_bytes = end - req->wb_offset;
 714
 715         return req;
 716 }
 717
 718 int nfs_flush_incompatible(struct file *file, struct page *page)
 719 {
 720         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 721         struct nfs_page *req;
 722         int do_flush, status;
 723         /*
 724          * Look for a request corresponding to this page. If there
 725          * is one, and it belongs to another file, we flush it out
 726          * before we try to copy anything into the page. Do this
 727          * due to the lack of an ACCESS-type call in NFSv2.
 728          * Also do the same if we find a request from an existing
 729          * dropped page.
 730          */
 731         do {
 732                 req = nfs_page_find_request(page);
 733                 if (req == NULL)
 734                         return 0;
 735                 do_flush = req->wb_page != page || req->wb_context != ctx
 736                         || !nfs_dirty_request(req);
 737                 nfs_release_request(req);
 738                 if (!do_flush)
 739                         return 0;
 740                 status = nfs_wb_page(page->mapping->host, page);
 741         } while (status == 0);
 742         return status;
 743 }
 744
 745 /*
 746  * Update and possibly write a cached page of an NFS file.
 747  *
 748  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 749  * things with a page scheduled for an RPC call (e.g. invalidate it).
 750  */
 751 int nfs_updatepage(struct file *file, struct page *page,
 752                 unsigned int offset, unsigned int count)
 753 {
 754         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 755         struct inode    *inode = page->mapping->host;
 756         int             status = 0;
 757
 758         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 759
 760         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 761                 file->f_path.dentry->d_parent->d_name.name,
 762                 file->f_path.dentry->d_name.name, count,
 763                 (long long)(page_offset(page) +offset));
 764
 765         /* If we're not using byte range locks, and we know the page
 766          * is entirely in cache, it may be more efficient to avoid
 767          * fragmenting write requests.
 768          */
 769         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 770                 count = max(count + offset, nfs_page_length(page));
 771                 offset = 0;
 772         }
 773
 774         status = nfs_writepage_setup(ctx, page, offset, count);
 775         __set_page_dirty_nobuffers(page);
 776
 777         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 778                         status, (long long)i_size_read(inode));
 779         if (status < 0)
 780                 nfs_set_pageerror(page);
 781         return status;
 782 }
 783
 784 static void nfs_writepage_release(struct nfs_page *req)
 785 {
 786
 787         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 788                 nfs_end_page_writeback(req->wb_page);
 789                 nfs_inode_remove_request(req);
 790         } else
 791                 nfs_end_page_writeback(req->wb_page);
 792         nfs_clear_page_writeback(req);
 793 }
 794
 795 static inline int flush_task_priority(int how)
 796 {
 797         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 798                 case FLUSH_HIGHPRI:
 799                         return RPC_PRIORITY_HIGH;
 800                 case FLUSH_LOWPRI:
 801                         return RPC_PRIORITY_LOW;
 802         }
 803         return RPC_PRIORITY_NORMAL;
 804 }
 805
 806 /*
 807  * Set up the argument/result storage required for the RPC call.
 808  */
 809 static void nfs_write_rpcsetup(struct nfs_page *req,
 810                 struct nfs_write_data *data,
 811                 const struct rpc_call_ops *call_ops,
 812                 unsigned int count, unsigned int offset,
 813                 int how)
 814 {
 815         struct inode            *inode;
 816         int flags;
 817
 818         /* Set up the RPC argument and reply structs
 819          * NB: take care not to mess about with data->commit et al. */
 820
 821         data->req = req;
 822         data->inode = inode = req->wb_context->dentry->d_inode;
 823         data->cred = req->wb_context->cred;
 824
 825         data->args.fh     = NFS_FH(inode);
 826         data->args.offset = req_offset(req) + offset;
 827         data->args.pgbase = req->wb_pgbase + offset;
 828         data->args.pages  = data->pagevec;
 829         data->args.count  = count;
 830         data->args.context = req->wb_context;
 831
 832         data->res.fattr   = &data->fattr;
 833         data->res.count   = count;
 834         data->res.verf    = &data->verf;
 835         nfs_fattr_init(&data->fattr);
 836
 837         /* Set up the initial task struct.  */
 838         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 839         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 840         NFS_PROTO(inode)->write_setup(data, how);
 841
 842         data->task.tk_priority = flush_task_priority(how);
 843         data->task.tk_cookie = (unsigned long)inode;
 844
 845         dprintk("NFS: %5u initiated write call "
 846                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 847                 data->task.tk_pid,
 848                 inode->i_sb->s_id,
 849                 (long long)NFS_FILEID(inode),
 850                 count,
 851                 (unsigned long long)data->args.offset);
 852 }
 853
 854 static void nfs_execute_write(struct nfs_write_data *data)
 855 {
 856         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 857         sigset_t oldset;
 858
 859         rpc_clnt_sigmask(clnt, &oldset);
 860         rpc_execute(&data->task);
 861         rpc_clnt_sigunmask(clnt, &oldset);
 862 }
 863
 864 /*
 865  * Generate multiple small requests to write out a single
 866  * contiguous dirty area on one page.
 867  */
 868 static int nfs_flush_multi(struct inode *inode, struct list_head *head, int how)
 869 {
 870         struct nfs_page *req = nfs_list_entry(head->next);
 871         struct page *page = req->wb_page;
 872         struct nfs_write_data *data;
 873         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 874         unsigned int offset;
 875         int requests = 0;
 876         LIST_HEAD(list);
 877
 878         nfs_list_remove_request(req);
 879
 880         nbytes = req->wb_bytes;
 881         do {
 882                 size_t len = min(nbytes, wsize);
 883
 884                 data = nfs_writedata_alloc(len);
 885                 if (!data)
 886                         goto out_bad;
 887                 list_add(&data->pages, &list);
 888                 requests++;
 889                 nbytes -= len;
 890         } while (nbytes != 0);
 891         atomic_set(&req->wb_complete, requests);
 892
 893         ClearPageError(page);
 894         offset = 0;
 895         nbytes = req->wb_bytes;
 896         do {
 897                 data = list_entry(list.next, struct nfs_write_data, pages);
 898                 list_del_init(&data->pages);
 899
 900                 data->pagevec[0] = page;
 901
 902                 if (nbytes > wsize) {
 903                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 904                                         wsize, offset, how);
 905                         offset += wsize;
 906                         nbytes -= wsize;
 907                 } else {
 908                         nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 909                                         nbytes, offset, how);
 910                         nbytes = 0;
 911                 }
 912                 nfs_execute_write(data);
 913         } while (nbytes != 0);
 914
 915         return 0;
 916
 917 out_bad:
 918         while (!list_empty(&list)) {
 919                 data = list_entry(list.next, struct nfs_write_data, pages);
 920                 list_del(&data->pages);
 921                 nfs_writedata_release(data);
 922         }
 923         nfs_redirty_request(req);
 924         nfs_end_page_writeback(req->wb_page);
 925         nfs_clear_page_writeback(req);
 926         return -ENOMEM;
 927 }
 928
 929 /*
 930  * Create an RPC task for the given write request and kick it.
 931  * The page must have been locked by the caller.
 932  *
 933  * It may happen that the page we're passed is not marked dirty.
 934  * This is the case if nfs_updatepage detects a conflicting request
 935  * that has been written but not committed.
 936  */
 937 static int nfs_flush_one(struct inode *inode, struct list_head *head, int how)
 938 {
 939         struct nfs_page         *req;
 940         struct page             **pages;
 941         struct nfs_write_data   *data;
 942         unsigned int            count;
 943
 944         data = nfs_writedata_alloc(NFS_SERVER(inode)->wsize);
 945         if (!data)
 946                 goto out_bad;
 947
 948         pages = data->pagevec;
 949         count = 0;
 950         while (!list_empty(head)) {
 951                 req = nfs_list_entry(head->next);
 952                 nfs_list_remove_request(req);
 953                 nfs_list_add_request(req, &data->pages);
 954                 ClearPageError(req->wb_page);
 955                 *pages++ = req->wb_page;
 956                 count += req->wb_bytes;
 957         }
 958         req = nfs_list_entry(data->pages.next);
 959
 960         /* Set up the argument struct */
 961         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 962
 963         nfs_execute_write(data);
 964         return 0;
 965  out_bad:
 966         while (!list_empty(head)) {
 967                 struct nfs_page *req = nfs_list_entry(head->next);
 968                 nfs_list_remove_request(req);
 969                 nfs_redirty_request(req);
 970                 nfs_end_page_writeback(req->wb_page);
 971                 nfs_clear_page_writeback(req);
 972         }
 973         return -ENOMEM;
 974 }
 975
 976 static int nfs_flush_list(struct inode *inode, struct list_head *head, int npages, int how)
 977 {
 978         LIST_HEAD(one_request);
 979         int (*flush_one)(struct inode *, struct list_head *, int);
 980         struct nfs_page *req;
 981         int wpages = NFS_SERVER(inode)->wpages;
 982         int wsize = NFS_SERVER(inode)->wsize;
 983         int error;
 984
 985         flush_one = nfs_flush_one;
 986         if (wsize < PAGE_CACHE_SIZE)
 987                 flush_one = nfs_flush_multi;
 988         /* For single writes, FLUSH_STABLE is more efficient */
 989         if (npages <= wpages && npages == NFS_I(inode)->npages
 990                         && nfs_list_entry(head->next)->wb_bytes <= wsize)
 991                 how |= FLUSH_STABLE;
 992
 993         do {
 994                 nfs_coalesce_requests(head, &one_request, wpages);
 995                 req = nfs_list_entry(one_request.next);
 996                 error = flush_one(inode, &one_request, how);
 997                 if (error < 0)
 998                         goto out_err;
 999         } while (!list_empty(head));
1000         return 0;
1001 out_err:
1002         while (!list_empty(head)) {
1003                 req = nfs_list_entry(head->next);
1004                 nfs_list_remove_request(req);
1005                 nfs_redirty_request(req);
1006                 nfs_end_page_writeback(req->wb_page);
1007                 nfs_clear_page_writeback(req);
1008         }
1009         return error;
1010 }
1011
1012 /*
1013  * Handle a write reply that flushed part of a page.
1014  */
1015 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1016 {
1017         struct nfs_write_data   *data = calldata;
1018         struct nfs_page         *req = data->req;
1019         struct page             *page = req->wb_page;
1020
1021         dprintk("NFS: write (%s/%Ld %d@%Ld)",
1022                 req->wb_context->dentry->d_inode->i_sb->s_id,
1023                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1024                 req->wb_bytes,
1025                 (long long)req_offset(req));
1026
1027         if (nfs_writeback_done(task, data) != 0)
1028                 return;
1029
1030         if (task->tk_status < 0) {
1031                 nfs_set_pageerror(page);
1032                 req->wb_context->error = task->tk_status;
1033                 dprintk(", error = %d\n", task->tk_status);
1034                 goto out;
1035         }
1036
1037         if (nfs_write_need_commit(data)) {
1038                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
1039
1040                 spin_lock(req_lock);
1041                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1042                         /* Do nothing we need to resend the writes */
1043                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1044                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1045                         dprintk(" defer commit\n");
1046                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1047                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1048                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1049                         dprintk(" server reboot detected\n");
1050                 }
1051                 spin_unlock(req_lock);
1052         } else
1053                 dprintk(" OK\n");
1054
1055 out:
1056         if (atomic_dec_and_test(&req->wb_complete))
1057                 nfs_writepage_release(req);
1058 }
1059
1060 static const struct rpc_call_ops nfs_write_partial_ops = {
1061         .rpc_call_done = nfs_writeback_done_partial,
1062         .rpc_release = nfs_writedata_release,
1063 };
1064
1065 /*
1066  * Handle a write reply that flushes a whole page.
1067  *
1068  * FIXME: There is an inherent race with invalidate_inode_pages and
1069  *        writebacks since the page->count is kept > 1 for as long
1070  *        as the page has a write request pending.
1071  */
1072 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1073 {
1074         struct nfs_write_data   *data = calldata;
1075         struct nfs_page         *req;
1076         struct page             *page;
1077
1078         if (nfs_writeback_done(task, data) != 0)
1079                 return;
1080
1081         /* Update attributes as result of writeback. */
1082         while (!list_empty(&data->pages)) {
1083                 req = nfs_list_entry(data->pages.next);
1084                 nfs_list_remove_request(req);
1085                 page = req->wb_page;
1086
1087                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1088                         req->wb_context->dentry->d_inode->i_sb->s_id,
1089                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1090                         req->wb_bytes,
1091                         (long long)req_offset(req));
1092
1093                 if (task->tk_status < 0) {
1094                         nfs_set_pageerror(page);
1095                         req->wb_context->error = task->tk_status;
1096                         dprintk(", error = %d\n", task->tk_status);
1097                         goto remove_request;
1098                 }
1099
1100                 if (nfs_write_need_commit(data)) {
1101                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1102                         nfs_mark_request_commit(req);
1103                         nfs_end_page_writeback(page);
1104                         dprintk(" marked for commit\n");
1105                         goto next;
1106                 }
1107                 dprintk(" OK\n");
1108 remove_request:
1109                 nfs_end_page_writeback(page);
1110                 nfs_inode_remove_request(req);
1111         next:
1112                 nfs_clear_page_writeback(req);
1113         }
1114 }
1115
1116 static const struct rpc_call_ops nfs_write_full_ops = {
1117         .rpc_call_done = nfs_writeback_done_full,
1118         .rpc_release = nfs_writedata_release,
1119 };
1120
1121
1122 /*
1123  * This function is called when the WRITE call is complete.
1124  */
1125 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1126 {
1127         struct nfs_writeargs    *argp = &data->args;
1128         struct nfs_writeres     *resp = &data->res;
1129         int status;
1130
1131         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1132                 task->tk_pid, task->tk_status);
1133
1134         /*
1135          * ->write_done will attempt to use post-op attributes to detect
1136          * conflicting writes by other clients.  A strict interpretation
1137          * of close-to-open would allow us to continue caching even if
1138          * another writer had changed the file, but some applications
1139          * depend on tighter cache coherency when writing.
1140          */
1141         status = NFS_PROTO(data->inode)->write_done(task, data);
1142         if (status != 0)
1143                 return status;
1144         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1145
1146 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1147         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1148                 /* We tried a write call, but the server did not
1149                  * commit data to stable storage even though we
1150                  * requested it.
1151                  * Note: There is a known bug in Tru64 < 5.0 in which
1152                  *       the server reports NFS_DATA_SYNC, but performs
1153                  *       NFS_FILE_SYNC. We therefore implement this checking
1154                  *       as a dprintk() in order to avoid filling syslog.
1155                  */
1156                 static unsigned long    complain;
1157
1158                 if (time_before(complain, jiffies)) {
1159                         dprintk("NFS: faulty NFS server %s:"
1160                                 " (committed = %d) != (stable = %d)\n",
1161                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1162                                 resp->verf->committed, argp->stable);
1163                         complain = jiffies + 300 * HZ;
1164                 }
1165         }
1166 #endif
1167         /* Is this a short write? */
1168         if (task->tk_status >= 0 && resp->count < argp->count) {
1169                 static unsigned long    complain;
1170
1171                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1172
1173                 /* Has the server at least made some progress? */
1174                 if (resp->count != 0) {
1175                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1176                         if (resp->verf->committed != NFS_UNSTABLE) {
1177                                 /* Resend from where the server left off */
1178                                 argp->offset += resp->count;
1179                                 argp->pgbase += resp->count;
1180                                 argp->count -= resp->count;
1181                         } else {
1182                                 /* Resend as a stable write in order to avoid
1183                                  * headaches in the case of a server crash.
1184                                  */
1185                                 argp->stable = NFS_FILE_SYNC;
1186                         }
1187                         rpc_restart_call(task);
1188                         return -EAGAIN;
1189                 }
1190                 if (time_before(complain, jiffies)) {
1191                         printk(KERN_WARNING
1192                                "NFS: Server wrote zero bytes, expected %u.\n",
1193                                         argp->count);
1194                         complain = jiffies + 300 * HZ;
1195                 }
1196                 /* Can't do anything about it except throw an error. */
1197                 task->tk_status = -EIO;
1198         }
1199         return 0;
1200 }
1201
1202
1203 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1204 void nfs_commit_release(void *wdata)
1205 {
1206         nfs_commit_free(wdata);
1207 }
1208
1209 /*
1210  * Set up the argument/result storage required for the RPC call.
1211  */
1212 static void nfs_commit_rpcsetup(struct list_head *head,
1213                 struct nfs_write_data *data,
1214                 int how)
1215 {
1216         struct nfs_page         *first;
1217         struct inode            *inode;
1218         int flags;
1219
1220         /* Set up the RPC argument and reply structs
1221          * NB: take care not to mess about with data->commit et al. */
1222
1223         list_splice_init(head, &data->pages);
1224         first = nfs_list_entry(data->pages.next);
1225         inode = first->wb_context->dentry->d_inode;
1226
1227         data->inode       = inode;
1228         data->cred        = first->wb_context->cred;
1229
1230         data->args.fh     = NFS_FH(data->inode);
1231         /* Note: we always request a commit of the entire inode */
1232         data->args.offset = 0;
1233         data->args.count  = 0;
1234         data->res.count   = 0;
1235         data->res.fattr   = &data->fattr;
1236         data->res.verf    = &data->verf;
1237         nfs_fattr_init(&data->fattr);
1238
1239         /* Set up the initial task struct.  */
1240         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1241         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1242         NFS_PROTO(inode)->commit_setup(data, how);
1243
1244         data->task.tk_priority = flush_task_priority(how);
1245         data->task.tk_cookie = (unsigned long)inode;
1246
1247         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1248 }
1249
1250 /*
1251  * Commit dirty pages
1252  */
1253 static int
1254 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1255 {
1256         struct nfs_write_data   *data;
1257         struct nfs_page         *req;
1258
1259         data = nfs_commit_alloc();
1260
1261         if (!data)
1262                 goto out_bad;
1263
1264         /* Set up the argument struct */
1265         nfs_commit_rpcsetup(head, data, how);
1266
1267         nfs_execute_write(data);
1268         return 0;
1269  out_bad:
1270         while (!list_empty(head)) {
1271                 req = nfs_list_entry(head->next);
1272                 nfs_list_remove_request(req);
1273                 nfs_mark_request_commit(req);
1274                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1275                 nfs_clear_page_writeback(req);
1276         }
1277         return -ENOMEM;
1278 }
1279
1280 /*
1281  * COMMIT call returned
1282  */
1283 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1284 {
1285         struct nfs_write_data   *data = calldata;
1286         struct nfs_page         *req;
1287
1288         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1289                                 task->tk_pid, task->tk_status);
1290
1291         /* Call the NFS version-specific code */
1292         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1293                 return;
1294
1295         while (!list_empty(&data->pages)) {
1296                 req = nfs_list_entry(data->pages.next);
1297                 nfs_list_remove_request(req);
1298                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1299                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1300
1301                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1302                         req->wb_context->dentry->d_inode->i_sb->s_id,
1303                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1304                         req->wb_bytes,
1305                         (long long)req_offset(req));
1306                 if (task->tk_status < 0) {
1307                         req->wb_context->error = task->tk_status;
1308                         nfs_inode_remove_request(req);
1309                         dprintk(", error = %d\n", task->tk_status);
1310                         goto next;
1311                 }
1312
1313                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1314                  * returned by the server against all stored verfs. */
1315                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1316                         /* We have a match */
1317                         nfs_inode_remove_request(req);
1318                         dprintk(" OK\n");
1319                         goto next;
1320                 }
1321                 /* We have a mismatch. Write the page again */
1322                 dprintk(" mismatch\n");
1323                 nfs_redirty_request(req);
1324         next:
1325                 nfs_clear_page_writeback(req);
1326         }
1327 }
1328
1329 static const struct rpc_call_ops nfs_commit_ops = {
1330         .rpc_call_done = nfs_commit_done,
1331         .rpc_release = nfs_commit_release,
1332 };
1333 #else
1334 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1335 {
1336         return 0;
1337 }
1338 #endif
1339
1340 static long nfs_flush_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1341 {
1342         struct nfs_inode *nfsi = NFS_I(mapping->host);
1343         LIST_HEAD(head);
1344         long res;
1345
1346         spin_lock(&nfsi->req_lock);
1347         res = nfs_scan_dirty(mapping, wbc, &head);
1348         spin_unlock(&nfsi->req_lock);
1349         if (res) {
1350                 int error = nfs_flush_list(mapping->host, &head, res, how);
1351                 if (error < 0)
1352                         return error;
1353         }
1354         return res;
1355 }
1356
1357 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1358 int nfs_commit_inode(struct inode *inode, int how)
1359 {
1360         struct nfs_inode *nfsi = NFS_I(inode);
1361         LIST_HEAD(head);
1362         int res;
1363
1364         spin_lock(&nfsi->req_lock);
1365         res = nfs_scan_commit(inode, &head, 0, 0);
1366         spin_unlock(&nfsi->req_lock);
1367         if (res) {
1368                 int error = nfs_commit_list(inode, &head, how);
1369                 if (error < 0)
1370                         return error;
1371         }
1372         return res;
1373 }
1374 #endif
1375
1376 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1377 {
1378         struct inode *inode = mapping->host;
1379         struct nfs_inode *nfsi = NFS_I(inode);
1380         unsigned long idx_start, idx_end;
1381         unsigned int npages = 0;
1382         LIST_HEAD(head);
1383         int nocommit = how & FLUSH_NOCOMMIT;
1384         long pages, ret;
1385
1386         /* FIXME */
1387         if (wbc->range_cyclic)
1388                 idx_start = 0;
1389         else {
1390                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1391                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1392                 if (idx_end > idx_start) {
1393                         unsigned long l_npages = 1 + idx_end - idx_start;
1394                         npages = l_npages;
1395                         if (sizeof(npages) != sizeof(l_npages) &&
1396                                         (unsigned long)npages != l_npages)
1397                                 npages = 0;
1398                 }
1399         }
1400         how &= ~FLUSH_NOCOMMIT;
1401         spin_lock(&nfsi->req_lock);
1402         do {
1403                 wbc->pages_skipped = 0;
1404                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1405                 if (ret != 0)
1406                         continue;
1407                 pages = nfs_scan_dirty(mapping, wbc, &head);
1408                 if (pages != 0) {
1409                         spin_unlock(&nfsi->req_lock);
1410                         if (how & FLUSH_INVALIDATE) {
1411                                 nfs_cancel_dirty_list(&head);
1412                                 ret = pages;
1413                         } else
1414                                 ret = nfs_flush_list(inode, &head, pages, how);
1415                         spin_lock(&nfsi->req_lock);
1416                         continue;
1417                 }
1418                 if (wbc->pages_skipped != 0)
1419                         continue;
1420                 if (nocommit)
1421                         break;
1422                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1423                 if (pages == 0) {
1424                         if (wbc->pages_skipped != 0)
1425                                 continue;
1426                         break;
1427                 }
1428                 if (how & FLUSH_INVALIDATE) {
1429                         spin_unlock(&nfsi->req_lock);
1430                         nfs_cancel_commit_list(&head);
1431                         ret = pages;
1432                         spin_lock(&nfsi->req_lock);
1433                         continue;
1434                 }
1435                 pages += nfs_scan_commit(inode, &head, 0, 0);
1436                 spin_unlock(&nfsi->req_lock);
1437                 ret = nfs_commit_list(inode, &head, how);
1438                 spin_lock(&nfsi->req_lock);
1439         } while (ret >= 0);
1440         spin_unlock(&nfsi->req_lock);
1441         return ret;
1442 }
1443
1444 /*
1445  * flush the inode to disk.
1446  */
1447 int nfs_wb_all(struct inode *inode)
1448 {
1449         struct address_space *mapping = inode->i_mapping;
1450         struct writeback_control wbc = {
1451                 .bdi = mapping->backing_dev_info,
1452                 .sync_mode = WB_SYNC_ALL,
1453                 .nr_to_write = LONG_MAX,
1454                 .for_writepages = 1,
1455                 .range_cyclic = 1,
1456         };
1457         int ret;
1458
1459         ret = generic_writepages(mapping, &wbc);
1460         if (ret < 0)
1461                 goto out;
1462         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1463         if (ret >= 0)
1464                 return 0;
1465 out:
1466         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1467         return ret;
1468 }
1469
1470 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1471 {
1472         struct writeback_control wbc = {
1473                 .bdi = mapping->backing_dev_info,
1474                 .sync_mode = WB_SYNC_ALL,
1475                 .nr_to_write = LONG_MAX,
1476                 .range_start = range_start,
1477                 .range_end = range_end,
1478                 .for_writepages = 1,
1479         };
1480         int ret;
1481
1482         if (!(how & FLUSH_NOWRITEPAGE)) {
1483                 ret = generic_writepages(mapping, &wbc);
1484                 if (ret < 0)
1485                         goto out;
1486         }
1487         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1488         if (ret >= 0)
1489                 return 0;
1490 out:
1491         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1492         return ret;
1493 }
1494
1495 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1496 {
1497         loff_t range_start = page_offset(page);
1498         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1499         struct writeback_control wbc = {
1500                 .bdi = page->mapping->backing_dev_info,
1501                 .sync_mode = WB_SYNC_ALL,
1502                 .nr_to_write = LONG_MAX,
1503                 .range_start = range_start,
1504                 .range_end = range_end,
1505         };
1506         int ret;
1507
1508         BUG_ON(!PageLocked(page));
1509         if (!(how & FLUSH_NOWRITEPAGE) && clear_page_dirty_for_io(page)) {
1510                 ret = nfs_writepage_locked(page, &wbc);
1511                 if (ret < 0)
1512                         goto out;
1513         }
1514         if (!PagePrivate(page))
1515                 return 0;
1516         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1517         if (ret >= 0)
1518                 return 0;
1519 out:
1520         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1521         return ret;
1522 }
1523
1524 /*
1525  * Write back all requests on one page - we do this before reading it.
1526  */
1527 int nfs_wb_page(struct inode *inode, struct page* page)
1528 {
1529         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1530 }
1531
1532 int nfs_set_page_dirty(struct page *page)
1533 {
1534         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
1535         struct nfs_page *req;
1536         int ret;
1537
1538         spin_lock(req_lock);
1539         req = nfs_page_find_request_locked(page);
1540         if (req != NULL) {
1541                 /* Mark any existing write requests for flushing */
1542                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1543                 spin_unlock(req_lock);
1544                 nfs_release_request(req);
1545                 return ret;
1546         }
1547         ret = __set_page_dirty_nobuffers(page);
1548         spin_unlock(req_lock);
1549         return ret;
1550 }
1551
1552
1553 int __init nfs_init_writepagecache(void)
1554 {
1555         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1556                                              sizeof(struct nfs_write_data),
1557                                              0, SLAB_HWCACHE_ALIGN,
1558                                              NULL, NULL);
1559         if (nfs_wdata_cachep == NULL)
1560                 return -ENOMEM;
1561
1562         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1563                                                      nfs_wdata_cachep);
1564         if (nfs_wdata_mempool == NULL)
1565                 return -ENOMEM;
1566
1567         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1568                                                       nfs_wdata_cachep);
1569         if (nfs_commit_mempool == NULL)
1570                 return -ENOMEM;
1571
1572         /*
1573          * NFS congestion size, scale with available memory.
1574          *
1575          *  64MB:    8192k
1576          * 128MB:   11585k
1577          * 256MB:   16384k
1578          * 512MB:   23170k
1579          *   1GB:   32768k
1580          *   2GB:   46340k
1581          *   4GB:   65536k
1582          *   8GB:   92681k
1583          *  16GB:  131072k
1584          *
1585          * This allows larger machines to have larger/more transfers.
1586          * Limit the default to 256M
1587          */
1588         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1589         if (nfs_congestion_kb > 256*1024)
1590                 nfs_congestion_kb = 256*1024;
1591
1592         return 0;
1593 }
1594
1595 void nfs_destroy_writepagecache(void)
1596 {
1597         mempool_destroy(nfs_commit_mempool);
1598         mempool_destroy(nfs_wdata_mempool);
1599         kmem_cache_destroy(nfs_wdata_cachep);
1600 }
1601