fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16
  17 #include <linux/sunrpc/clnt.h>
  18 #include <linux/nfs_fs.h>
  19 #include <linux/nfs_mount.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/backing-dev.h>
  22
  23 #include <asm/uaccess.h>
  24
  25 #include "delegation.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 #define MIN_POOL_WRITE          (32)
  32 #define MIN_POOL_COMMIT         (4)
  33
  34 /*
  35  * Local function declarations
  36  */
  37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  38                                             struct page *,
  39                                             unsigned int, unsigned int);
  40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static const struct rpc_call_ops nfs_write_partial_ops;
  43 static const struct rpc_call_ops nfs_write_full_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45
  46 static struct kmem_cache *nfs_wdata_cachep;
  47 static mempool_t *nfs_wdata_mempool;
  48 static mempool_t *nfs_commit_mempool;
  49
  50 struct nfs_write_data *nfs_commit_alloc(void)
  51 {
  52         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  53
  54         if (p) {
  55                 memset(p, 0, sizeof(*p));
  56                 INIT_LIST_HEAD(&p->pages);
  57         }
  58         return p;
  59 }
  60
  61 static void nfs_commit_rcu_free(struct rcu_head *head)
  62 {
  63         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 void nfs_commit_free(struct nfs_write_data *wdata)
  70 {
  71         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  72 }
  73
  74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  75 {
  76         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  77
  78         if (p) {
  79                 memset(p, 0, sizeof(*p));
  80                 INIT_LIST_HEAD(&p->pages);
  81                 p->npages = pagecount;
  82                 if (pagecount <= ARRAY_SIZE(p->page_array))
  83                         p->pagevec = p->page_array;
  84                 else {
  85                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  86                         if (!p->pagevec) {
  87                                 mempool_free(p, nfs_wdata_mempool);
  88                                 p = NULL;
  89                         }
  90                 }
  91         }
  92         return p;
  93 }
  94
  95 static void nfs_writedata_rcu_free(struct rcu_head *head)
  96 {
  97         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  98         if (p && (p->pagevec != &p->page_array[0]))
  99                 kfree(p->pagevec);
 100         mempool_free(p, nfs_wdata_mempool);
 101 }
 102
 103 static void nfs_writedata_free(struct nfs_write_data *wdata)
 104 {
 105         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 106 }
 107
 108 void nfs_writedata_release(void *wdata)
 109 {
 110         nfs_writedata_free(wdata);
 111 }
 112
 113 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 114 {
 115         struct nfs_page *req = NULL;
 116
 117         if (PagePrivate(page)) {
 118                 req = (struct nfs_page *)page_private(page);
 119                 if (req != NULL)
 120                         kref_get(&req->wb_kref);
 121         }
 122         return req;
 123 }
 124
 125 static struct nfs_page *nfs_page_find_request(struct page *page)
 126 {
 127         struct nfs_page *req = NULL;
 128         spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 129
 130         spin_lock(req_lock);
 131         req = nfs_page_find_request_locked(page);
 132         spin_unlock(req_lock);
 133         return req;
 134 }
 135
 136 /* Adjust the file length if we're writing beyond the end */
 137 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 138 {
 139         struct inode *inode = page->mapping->host;
 140         loff_t end, i_size = i_size_read(inode);
 141         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 142
 143         if (i_size > 0 && page->index < end_index)
 144                 return;
 145         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 146         if (i_size >= end)
 147                 return;
 148         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 149         i_size_write(inode, end);
 150 }
 151
 152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 153 static void nfs_set_pageerror(struct page *page)
 154 {
 155         SetPageError(page);
 156         nfs_zap_mapping(page->mapping->host, page->mapping);
 157 }
 158
 159 /* We can set the PG_uptodate flag if we see that a write request
 160  * covers the full page.
 161  */
 162 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 163 {
 164         if (PageUptodate(page))
 165                 return;
 166         if (base != 0)
 167                 return;
 168         if (count != nfs_page_length(page))
 169                 return;
 170         if (count != PAGE_CACHE_SIZE)
 171                 zero_user_page(page, count, PAGE_CACHE_SIZE - count, KM_USER0);
 172         SetPageUptodate(page);
 173 }
 174
 175 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 176                 unsigned int offset, unsigned int count)
 177 {
 178         struct nfs_page *req;
 179         int ret;
 180
 181         for (;;) {
 182                 req = nfs_update_request(ctx, page, offset, count);
 183                 if (!IS_ERR(req))
 184                         break;
 185                 ret = PTR_ERR(req);
 186                 if (ret != -EBUSY)
 187                         return ret;
 188                 ret = nfs_wb_page(page->mapping->host, page);
 189                 if (ret != 0)
 190                         return ret;
 191         }
 192         /* Update file length */
 193         nfs_grow_file(page, offset, count);
 194         nfs_unlock_request(req);
 195         return 0;
 196 }
 197
 198 static int wb_priority(struct writeback_control *wbc)
 199 {
 200         if (wbc->for_reclaim)
 201                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 202         if (wbc->for_kupdate)
 203                 return FLUSH_LOWPRI;
 204         return 0;
 205 }
 206
 207 /*
 208  * NFS congestion control
 209  */
 210
 211 int nfs_congestion_kb;
 212
 213 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 214 #define NFS_CONGESTION_OFF_THRESH       \
 215         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 216
 217 static int nfs_set_page_writeback(struct page *page)
 218 {
 219         int ret = test_set_page_writeback(page);
 220
 221         if (!ret) {
 222                 struct inode *inode = page->mapping->host;
 223                 struct nfs_server *nfss = NFS_SERVER(inode);
 224
 225                 if (atomic_long_inc_return(&nfss->writeback) >
 226                                 NFS_CONGESTION_ON_THRESH)
 227                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 228         }
 229         return ret;
 230 }
 231
 232 static void nfs_end_page_writeback(struct page *page)
 233 {
 234         struct inode *inode = page->mapping->host;
 235         struct nfs_server *nfss = NFS_SERVER(inode);
 236
 237         end_page_writeback(page);
 238         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH) {
 239                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 240                 congestion_end(WRITE);
 241         }
 242 }
 243
 244 /*
 245  * Find an associated nfs write request, and prepare to flush it out
 246  * Returns 1 if there was no write request, or if the request was
 247  * already tagged by nfs_set_page_dirty.Returns 0 if the request
 248  * was not tagged.
 249  * May also return an error if the user signalled nfs_wait_on_request().
 250  */
 251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 252                                 struct page *page)
 253 {
 254         struct nfs_page *req;
 255         struct nfs_inode *nfsi = NFS_I(page->mapping->host);
 256         spinlock_t *req_lock = &nfsi->req_lock;
 257         int ret;
 258
 259         spin_lock(req_lock);
 260         for(;;) {
 261                 req = nfs_page_find_request_locked(page);
 262                 if (req == NULL) {
 263                         spin_unlock(req_lock);
 264                         return 1;
 265                 }
 266                 if (nfs_lock_request_dontget(req))
 267                         break;
 268                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 269                  *       then the call to nfs_lock_request_dontget() will always
 270                  *       succeed provided that someone hasn't already marked the
 271                  *       request as dirty (in which case we don't care).
 272                  */
 273                 spin_unlock(req_lock);
 274                 ret = nfs_wait_on_request(req);
 275                 nfs_release_request(req);
 276                 if (ret != 0)
 277                         return ret;
 278                 spin_lock(req_lock);
 279         }
 280         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 281                 /* This request is marked for commit */
 282                 spin_unlock(req_lock);
 283                 nfs_unlock_request(req);
 284                 nfs_pageio_complete(pgio);
 285                 return 1;
 286         }
 287         if (nfs_set_page_writeback(page) != 0) {
 288                 spin_unlock(req_lock);
 289                 BUG();
 290         }
 291         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 292                         NFS_PAGE_TAG_LOCKED);
 293         ret = test_bit(PG_NEED_FLUSH, &req->wb_flags);
 294         spin_unlock(req_lock);
 295         nfs_pageio_add_request(pgio, req);
 296         return ret;
 297 }
 298
 299 /*
 300  * Write an mmapped page to the server.
 301  */
 302 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 303 {
 304         struct nfs_pageio_descriptor mypgio, *pgio;
 305         struct nfs_open_context *ctx;
 306         struct inode *inode = page->mapping->host;
 307         unsigned offset;
 308         int err;
 309
 310         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 311         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 312
 313         if (wbc->for_writepages)
 314                 pgio = wbc->fs_private;
 315         else {
 316                 nfs_pageio_init_write(&mypgio, inode, wb_priority(wbc));
 317                 pgio = &mypgio;
 318         }
 319
 320         nfs_pageio_cond_complete(pgio, page->index);
 321
 322         err = nfs_page_async_flush(pgio, page);
 323         if (err <= 0)
 324                 goto out;
 325         err = 0;
 326         offset = nfs_page_length(page);
 327         if (!offset)
 328                 goto out;
 329
 330         nfs_pageio_cond_complete(pgio, page->index);
 331
 332         ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE);
 333         if (ctx == NULL) {
 334                 err = -EBADF;
 335                 goto out;
 336         }
 337         err = nfs_writepage_setup(ctx, page, 0, offset);
 338         put_nfs_open_context(ctx);
 339         if (err != 0)
 340                 goto out;
 341         err = nfs_page_async_flush(pgio, page);
 342         if (err > 0)
 343                 err = 0;
 344 out:
 345         if (!wbc->for_writepages)
 346                 nfs_pageio_complete(pgio);
 347         return err;
 348 }
 349
 350 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 351 {
 352         int err;
 353
 354         err = nfs_writepage_locked(page, wbc);
 355         unlock_page(page);
 356         return err;
 357 }
 358
 359 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 360 {
 361         struct inode *inode = mapping->host;
 362         struct nfs_pageio_descriptor pgio;
 363         int err;
 364
 365         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 366
 367         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 368         wbc->fs_private = &pgio;
 369         err = generic_writepages(mapping, wbc);
 370         nfs_pageio_complete(&pgio);
 371         if (err)
 372                 return err;
 373         if (pgio.pg_error)
 374                 return pgio.pg_error;
 375         return 0;
 376 }
 377
 378 /*
 379  * Insert a write request into an inode
 380  */
 381 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 382 {
 383         struct nfs_inode *nfsi = NFS_I(inode);
 384         int error;
 385
 386         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 387         BUG_ON(error == -EEXIST);
 388         if (error)
 389                 return error;
 390         if (!nfsi->npages) {
 391                 igrab(inode);
 392                 nfs_begin_data_update(inode);
 393                 if (nfs_have_delegation(inode, FMODE_WRITE))
 394                         nfsi->change_attr++;
 395         }
 396         SetPagePrivate(req->wb_page);
 397         set_page_private(req->wb_page, (unsigned long)req);
 398         if (PageDirty(req->wb_page))
 399                 set_bit(PG_NEED_FLUSH, &req->wb_flags);
 400         nfsi->npages++;
 401         kref_get(&req->wb_kref);
 402         return 0;
 403 }
 404
 405 /*
 406  * Remove a write request from an inode
 407  */
 408 static void nfs_inode_remove_request(struct nfs_page *req)
 409 {
 410         struct inode *inode = req->wb_context->path.dentry->d_inode;
 411         struct nfs_inode *nfsi = NFS_I(inode);
 412
 413         BUG_ON (!NFS_WBACK_BUSY(req));
 414
 415         spin_lock(&nfsi->req_lock);
 416         set_page_private(req->wb_page, 0);
 417         ClearPagePrivate(req->wb_page);
 418         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 419         if (test_and_clear_bit(PG_NEED_FLUSH, &req->wb_flags))
 420                 __set_page_dirty_nobuffers(req->wb_page);
 421         nfsi->npages--;
 422         if (!nfsi->npages) {
 423                 spin_unlock(&nfsi->req_lock);
 424                 nfs_end_data_update(inode);
 425                 iput(inode);
 426         } else
 427                 spin_unlock(&nfsi->req_lock);
 428         nfs_clear_request(req);
 429         nfs_release_request(req);
 430 }
 431
 432 static void
 433 nfs_redirty_request(struct nfs_page *req)
 434 {
 435         __set_page_dirty_nobuffers(req->wb_page);
 436 }
 437
 438 /*
 439  * Check if a request is dirty
 440  */
 441 static inline int
 442 nfs_dirty_request(struct nfs_page *req)
 443 {
 444         struct page *page = req->wb_page;
 445
 446         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 447                 return 0;
 448         return !PageWriteback(req->wb_page);
 449 }
 450
 451 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 452 /*
 453  * Add a request to the inode's commit list.
 454  */
 455 static void
 456 nfs_mark_request_commit(struct nfs_page *req)
 457 {
 458         struct inode *inode = req->wb_context->path.dentry->d_inode;
 459         struct nfs_inode *nfsi = NFS_I(inode);
 460
 461         spin_lock(&nfsi->req_lock);
 462         nfs_list_add_request(req, &nfsi->commit);
 463         nfsi->ncommit++;
 464         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 465         radix_tree_tag_set(&nfsi->nfs_page_tree,
 466                         req->wb_index,
 467                         NFS_PAGE_TAG_COMMIT);
 468         spin_unlock(&nfsi->req_lock);
 469         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 470         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 471 }
 472
 473 static inline
 474 int nfs_write_need_commit(struct nfs_write_data *data)
 475 {
 476         return data->verf.committed != NFS_FILE_SYNC;
 477 }
 478
 479 static inline
 480 int nfs_reschedule_unstable_write(struct nfs_page *req)
 481 {
 482         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 483                 nfs_mark_request_commit(req);
 484                 return 1;
 485         }
 486         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 487                 nfs_redirty_request(req);
 488                 return 1;
 489         }
 490         return 0;
 491 }
 492 #else
 493 static inline void
 494 nfs_mark_request_commit(struct nfs_page *req)
 495 {
 496 }
 497
 498 static inline
 499 int nfs_write_need_commit(struct nfs_write_data *data)
 500 {
 501         return 0;
 502 }
 503
 504 static inline
 505 int nfs_reschedule_unstable_write(struct nfs_page *req)
 506 {
 507         return 0;
 508 }
 509 #endif
 510
 511 /*
 512  * Wait for a request to complete.
 513  *
 514  * Interruptible by signals only if mounted with intr flag.
 515  */
 516 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 517 {
 518         struct nfs_inode *nfsi = NFS_I(inode);
 519         struct nfs_page *req;
 520         pgoff_t idx_end, next;
 521         unsigned int            res = 0;
 522         int                     error;
 523
 524         if (npages == 0)
 525                 idx_end = ~0;
 526         else
 527                 idx_end = idx_start + npages - 1;
 528
 529         next = idx_start;
 530         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 531                 if (req->wb_index > idx_end)
 532                         break;
 533
 534                 next = req->wb_index + 1;
 535                 BUG_ON(!NFS_WBACK_BUSY(req));
 536
 537                 kref_get(&req->wb_kref);
 538                 spin_unlock(&nfsi->req_lock);
 539                 error = nfs_wait_on_request(req);
 540                 nfs_release_request(req);
 541                 spin_lock(&nfsi->req_lock);
 542                 if (error < 0)
 543                         return error;
 544                 res++;
 545         }
 546         return res;
 547 }
 548
 549 static void nfs_cancel_commit_list(struct list_head *head)
 550 {
 551         struct nfs_page *req;
 552
 553         while(!list_empty(head)) {
 554                 req = nfs_list_entry(head->next);
 555                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 556                 nfs_list_remove_request(req);
 557                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 558                 nfs_inode_remove_request(req);
 559                 nfs_unlock_request(req);
 560         }
 561 }
 562
 563 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 564 /*
 565  * nfs_scan_commit - Scan an inode for commit requests
 566  * @inode: NFS inode to scan
 567  * @dst: destination list
 568  * @idx_start: lower bound of page->index to scan.
 569  * @npages: idx_start + npages sets the upper bound to scan.
 570  *
 571  * Moves requests from the inode's 'commit' request list.
 572  * The requests are *not* checked to ensure that they form a contiguous set.
 573  */
 574 static int
 575 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 576 {
 577         struct nfs_inode *nfsi = NFS_I(inode);
 578         int res = 0;
 579
 580         if (nfsi->ncommit != 0) {
 581                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 582                                 NFS_PAGE_TAG_COMMIT);
 583                 nfsi->ncommit -= res;
 584                 if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit))
 585                         printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n");
 586         }
 587         return res;
 588 }
 589 #else
 590 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 591 {
 592         return 0;
 593 }
 594 #endif
 595
 596 /*
 597  * Try to update any existing write request, or create one if there is none.
 598  * In order to match, the request's credentials must match those of
 599  * the calling process.
 600  *
 601  * Note: Should always be called with the Page Lock held!
 602  */
 603 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 604                 struct page *page, unsigned int offset, unsigned int bytes)
 605 {
 606         struct address_space *mapping = page->mapping;
 607         struct inode *inode = mapping->host;
 608         struct nfs_inode *nfsi = NFS_I(inode);
 609         struct nfs_page         *req, *new = NULL;
 610         pgoff_t         rqend, end;
 611
 612         end = offset + bytes;
 613
 614         for (;;) {
 615                 /* Loop over all inode entries and see if we find
 616                  * A request for the page we wish to update
 617                  */
 618                 spin_lock(&nfsi->req_lock);
 619                 req = nfs_page_find_request_locked(page);
 620                 if (req) {
 621                         if (!nfs_lock_request_dontget(req)) {
 622                                 int error;
 623
 624                                 spin_unlock(&nfsi->req_lock);
 625                                 error = nfs_wait_on_request(req);
 626                                 nfs_release_request(req);
 627                                 if (error < 0) {
 628                                         if (new)
 629                                                 nfs_release_request(new);
 630                                         return ERR_PTR(error);
 631                                 }
 632                                 continue;
 633                         }
 634                         spin_unlock(&nfsi->req_lock);
 635                         if (new)
 636                                 nfs_release_request(new);
 637                         break;
 638                 }
 639
 640                 if (new) {
 641                         int error;
 642                         nfs_lock_request_dontget(new);
 643                         error = nfs_inode_add_request(inode, new);
 644                         if (error) {
 645                                 spin_unlock(&nfsi->req_lock);
 646                                 nfs_unlock_request(new);
 647                                 return ERR_PTR(error);
 648                         }
 649                         spin_unlock(&nfsi->req_lock);
 650                         return new;
 651                 }
 652                 spin_unlock(&nfsi->req_lock);
 653
 654                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 655                 if (IS_ERR(new))
 656                         return new;
 657         }
 658
 659         /* We have a request for our page.
 660          * If the creds don't match, or the
 661          * page addresses don't match,
 662          * tell the caller to wait on the conflicting
 663          * request.
 664          */
 665         rqend = req->wb_offset + req->wb_bytes;
 666         if (req->wb_context != ctx
 667             || req->wb_page != page
 668             || !nfs_dirty_request(req)
 669             || offset > rqend || end < req->wb_offset) {
 670                 nfs_unlock_request(req);
 671                 return ERR_PTR(-EBUSY);
 672         }
 673
 674         /* Okay, the request matches. Update the region */
 675         if (offset < req->wb_offset) {
 676                 req->wb_offset = offset;
 677                 req->wb_pgbase = offset;
 678                 req->wb_bytes = rqend - req->wb_offset;
 679         }
 680
 681         if (end > rqend)
 682                 req->wb_bytes = end - req->wb_offset;
 683
 684         return req;
 685 }
 686
 687 int nfs_flush_incompatible(struct file *file, struct page *page)
 688 {
 689         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 690         struct nfs_page *req;
 691         int do_flush, status;
 692         /*
 693          * Look for a request corresponding to this page. If there
 694          * is one, and it belongs to another file, we flush it out
 695          * before we try to copy anything into the page. Do this
 696          * due to the lack of an ACCESS-type call in NFSv2.
 697          * Also do the same if we find a request from an existing
 698          * dropped page.
 699          */
 700         do {
 701                 req = nfs_page_find_request(page);
 702                 if (req == NULL)
 703                         return 0;
 704                 do_flush = req->wb_page != page || req->wb_context != ctx
 705                         || !nfs_dirty_request(req);
 706                 nfs_release_request(req);
 707                 if (!do_flush)
 708                         return 0;
 709                 status = nfs_wb_page(page->mapping->host, page);
 710         } while (status == 0);
 711         return status;
 712 }
 713
 714 /*
 715  * Update and possibly write a cached page of an NFS file.
 716  *
 717  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 718  * things with a page scheduled for an RPC call (e.g. invalidate it).
 719  */
 720 int nfs_updatepage(struct file *file, struct page *page,
 721                 unsigned int offset, unsigned int count)
 722 {
 723         struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data;
 724         struct inode    *inode = page->mapping->host;
 725         int             status = 0;
 726
 727         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 728
 729         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 730                 file->f_path.dentry->d_parent->d_name.name,
 731                 file->f_path.dentry->d_name.name, count,
 732                 (long long)(page_offset(page) +offset));
 733
 734         /* If we're not using byte range locks, and we know the page
 735          * is entirely in cache, it may be more efficient to avoid
 736          * fragmenting write requests.
 737          */
 738         if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) {
 739                 count = max(count + offset, nfs_page_length(page));
 740                 offset = 0;
 741         }
 742
 743         status = nfs_writepage_setup(ctx, page, offset, count);
 744         __set_page_dirty_nobuffers(page);
 745
 746         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 747                         status, (long long)i_size_read(inode));
 748         if (status < 0)
 749                 nfs_set_pageerror(page);
 750         return status;
 751 }
 752
 753 static void nfs_writepage_release(struct nfs_page *req)
 754 {
 755
 756         if (PageError(req->wb_page)) {
 757                 nfs_end_page_writeback(req->wb_page);
 758                 nfs_inode_remove_request(req);
 759         } else if (!nfs_reschedule_unstable_write(req)) {
 760                 /* Set the PG_uptodate flag */
 761                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 762                 nfs_end_page_writeback(req->wb_page);
 763                 nfs_inode_remove_request(req);
 764         } else
 765                 nfs_end_page_writeback(req->wb_page);
 766         nfs_clear_page_tag_locked(req);
 767 }
 768
 769 static inline int flush_task_priority(int how)
 770 {
 771         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 772                 case FLUSH_HIGHPRI:
 773                         return RPC_PRIORITY_HIGH;
 774                 case FLUSH_LOWPRI:
 775                         return RPC_PRIORITY_LOW;
 776         }
 777         return RPC_PRIORITY_NORMAL;
 778 }
 779
 780 /*
 781  * Set up the argument/result storage required for the RPC call.
 782  */
 783 static void nfs_write_rpcsetup(struct nfs_page *req,
 784                 struct nfs_write_data *data,
 785                 const struct rpc_call_ops *call_ops,
 786                 unsigned int count, unsigned int offset,
 787                 int how)
 788 {
 789         struct inode            *inode;
 790         int flags;
 791
 792         /* Set up the RPC argument and reply structs
 793          * NB: take care not to mess about with data->commit et al. */
 794
 795         data->req = req;
 796         data->inode = inode = req->wb_context->path.dentry->d_inode;
 797         data->cred = req->wb_context->cred;
 798
 799         data->args.fh     = NFS_FH(inode);
 800         data->args.offset = req_offset(req) + offset;
 801         data->args.pgbase = req->wb_pgbase + offset;
 802         data->args.pages  = data->pagevec;
 803         data->args.count  = count;
 804         data->args.context = req->wb_context;
 805
 806         data->res.fattr   = &data->fattr;
 807         data->res.count   = count;
 808         data->res.verf    = &data->verf;
 809         nfs_fattr_init(&data->fattr);
 810
 811         /* Set up the initial task struct.  */
 812         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 813         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data);
 814         NFS_PROTO(inode)->write_setup(data, how);
 815
 816         data->task.tk_priority = flush_task_priority(how);
 817         data->task.tk_cookie = (unsigned long)inode;
 818
 819         dprintk("NFS: %5u initiated write call "
 820                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 821                 data->task.tk_pid,
 822                 inode->i_sb->s_id,
 823                 (long long)NFS_FILEID(inode),
 824                 count,
 825                 (unsigned long long)data->args.offset);
 826 }
 827
 828 static void nfs_execute_write(struct nfs_write_data *data)
 829 {
 830         struct rpc_clnt *clnt = NFS_CLIENT(data->inode);
 831         sigset_t oldset;
 832
 833         rpc_clnt_sigmask(clnt, &oldset);
 834         rpc_execute(&data->task);
 835         rpc_clnt_sigunmask(clnt, &oldset);
 836 }
 837
 838 /*
 839  * Generate multiple small requests to write out a single
 840  * contiguous dirty area on one page.
 841  */
 842 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 843 {
 844         struct nfs_page *req = nfs_list_entry(head->next);
 845         struct page *page = req->wb_page;
 846         struct nfs_write_data *data;
 847         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 848         unsigned int offset;
 849         int requests = 0;
 850         LIST_HEAD(list);
 851
 852         nfs_list_remove_request(req);
 853
 854         nbytes = count;
 855         do {
 856                 size_t len = min(nbytes, wsize);
 857
 858                 data = nfs_writedata_alloc(1);
 859                 if (!data)
 860                         goto out_bad;
 861                 list_add(&data->pages, &list);
 862                 requests++;
 863                 nbytes -= len;
 864         } while (nbytes != 0);
 865         atomic_set(&req->wb_complete, requests);
 866
 867         ClearPageError(page);
 868         offset = 0;
 869         nbytes = count;
 870         do {
 871                 data = list_entry(list.next, struct nfs_write_data, pages);
 872                 list_del_init(&data->pages);
 873
 874                 data->pagevec[0] = page;
 875
 876                 if (nbytes < wsize)
 877                         wsize = nbytes;
 878                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 879                                    wsize, offset, how);
 880                 offset += wsize;
 881                 nbytes -= wsize;
 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_tag_locked(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, unsigned int npages, size_t count, int how)
 908 {
 909         struct nfs_page         *req;
 910         struct page             **pages;
 911         struct nfs_write_data   *data;
 912
 913         data = nfs_writedata_alloc(npages);
 914         if (!data)
 915                 goto out_bad;
 916
 917         pages = data->pagevec;
 918         while (!list_empty(head)) {
 919                 req = nfs_list_entry(head->next);
 920                 nfs_list_remove_request(req);
 921                 nfs_list_add_request(req, &data->pages);
 922                 ClearPageError(req->wb_page);
 923                 *pages++ = req->wb_page;
 924         }
 925         req = nfs_list_entry(data->pages.next);
 926
 927         /* Set up the argument struct */
 928         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 929
 930         nfs_execute_write(data);
 931         return 0;
 932  out_bad:
 933         while (!list_empty(head)) {
 934                 req = nfs_list_entry(head->next);
 935                 nfs_list_remove_request(req);
 936                 nfs_redirty_request(req);
 937                 nfs_end_page_writeback(req->wb_page);
 938                 nfs_clear_page_tag_locked(req);
 939         }
 940         return -ENOMEM;
 941 }
 942
 943 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 944                                   struct inode *inode, int ioflags)
 945 {
 946         int wsize = NFS_SERVER(inode)->wsize;
 947
 948         if (wsize < PAGE_CACHE_SIZE)
 949                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 950         else
 951                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 952 }
 953
 954 /*
 955  * Handle a write reply that flushed part of a page.
 956  */
 957 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 958 {
 959         struct nfs_write_data   *data = calldata;
 960         struct nfs_page         *req = data->req;
 961         struct page             *page = req->wb_page;
 962
 963         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 964                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 965                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 966                 req->wb_bytes,
 967                 (long long)req_offset(req));
 968
 969         if (nfs_writeback_done(task, data) != 0)
 970                 return;
 971
 972         if (task->tk_status < 0) {
 973                 nfs_set_pageerror(page);
 974                 req->wb_context->error = task->tk_status;
 975                 dprintk(", error = %d\n", task->tk_status);
 976                 goto out;
 977         }
 978
 979         if (nfs_write_need_commit(data)) {
 980                 spinlock_t *req_lock = &NFS_I(page->mapping->host)->req_lock;
 981
 982                 spin_lock(req_lock);
 983                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 984                         /* Do nothing we need to resend the writes */
 985                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 986                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 987                         dprintk(" defer commit\n");
 988                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 989                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 990                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 991                         dprintk(" server reboot detected\n");
 992                 }
 993                 spin_unlock(req_lock);
 994         } else
 995                 dprintk(" OK\n");
 996
 997 out:
 998         if (atomic_dec_and_test(&req->wb_complete))
 999                 nfs_writepage_release(req);
1000 }
1001
1002 static const struct rpc_call_ops nfs_write_partial_ops = {
1003         .rpc_call_done = nfs_writeback_done_partial,
1004         .rpc_release = nfs_writedata_release,
1005 };
1006
1007 /*
1008  * Handle a write reply that flushes a whole page.
1009  *
1010  * FIXME: There is an inherent race with invalidate_inode_pages and
1011  *        writebacks since the page->count is kept > 1 for as long
1012  *        as the page has a write request pending.
1013  */
1014 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1015 {
1016         struct nfs_write_data   *data = calldata;
1017         struct nfs_page         *req;
1018         struct page             *page;
1019
1020         if (nfs_writeback_done(task, data) != 0)
1021                 return;
1022
1023         /* Update attributes as result of writeback. */
1024         while (!list_empty(&data->pages)) {
1025                 req = nfs_list_entry(data->pages.next);
1026                 nfs_list_remove_request(req);
1027                 page = req->wb_page;
1028
1029                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1030                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1031                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1032                         req->wb_bytes,
1033                         (long long)req_offset(req));
1034
1035                 if (task->tk_status < 0) {
1036                         nfs_set_pageerror(page);
1037                         req->wb_context->error = task->tk_status;
1038                         dprintk(", error = %d\n", task->tk_status);
1039                         goto remove_request;
1040                 }
1041
1042                 if (nfs_write_need_commit(data)) {
1043                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1044                         nfs_mark_request_commit(req);
1045                         nfs_end_page_writeback(page);
1046                         dprintk(" marked for commit\n");
1047                         goto next;
1048                 }
1049                 /* Set the PG_uptodate flag? */
1050                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1051                 dprintk(" OK\n");
1052 remove_request:
1053                 nfs_end_page_writeback(page);
1054                 nfs_inode_remove_request(req);
1055         next:
1056                 nfs_clear_page_tag_locked(req);
1057         }
1058 }
1059
1060 static const struct rpc_call_ops nfs_write_full_ops = {
1061         .rpc_call_done = nfs_writeback_done_full,
1062         .rpc_release = nfs_writedata_release,
1063 };
1064
1065
1066 /*
1067  * This function is called when the WRITE call is complete.
1068  */
1069 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1070 {
1071         struct nfs_writeargs    *argp = &data->args;
1072         struct nfs_writeres     *resp = &data->res;
1073         int status;
1074
1075         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1076                 task->tk_pid, task->tk_status);
1077
1078         /*
1079          * ->write_done will attempt to use post-op attributes to detect
1080          * conflicting writes by other clients.  A strict interpretation
1081          * of close-to-open would allow us to continue caching even if
1082          * another writer had changed the file, but some applications
1083          * depend on tighter cache coherency when writing.
1084          */
1085         status = NFS_PROTO(data->inode)->write_done(task, data);
1086         if (status != 0)
1087                 return status;
1088         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1089
1090 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1091         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1092                 /* We tried a write call, but the server did not
1093                  * commit data to stable storage even though we
1094                  * requested it.
1095                  * Note: There is a known bug in Tru64 < 5.0 in which
1096                  *       the server reports NFS_DATA_SYNC, but performs
1097                  *       NFS_FILE_SYNC. We therefore implement this checking
1098                  *       as a dprintk() in order to avoid filling syslog.
1099                  */
1100                 static unsigned long    complain;
1101
1102                 if (time_before(complain, jiffies)) {
1103                         dprintk("NFS: faulty NFS server %s:"
1104                                 " (committed = %d) != (stable = %d)\n",
1105                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1106                                 resp->verf->committed, argp->stable);
1107                         complain = jiffies + 300 * HZ;
1108                 }
1109         }
1110 #endif
1111         /* Is this a short write? */
1112         if (task->tk_status >= 0 && resp->count < argp->count) {
1113                 static unsigned long    complain;
1114
1115                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1116
1117                 /* Has the server at least made some progress? */
1118                 if (resp->count != 0) {
1119                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1120                         if (resp->verf->committed != NFS_UNSTABLE) {
1121                                 /* Resend from where the server left off */
1122                                 argp->offset += resp->count;
1123                                 argp->pgbase += resp->count;
1124                                 argp->count -= resp->count;
1125                         } else {
1126                                 /* Resend as a stable write in order to avoid
1127                                  * headaches in the case of a server crash.
1128                                  */
1129                                 argp->stable = NFS_FILE_SYNC;
1130                         }
1131                         rpc_restart_call(task);
1132                         return -EAGAIN;
1133                 }
1134                 if (time_before(complain, jiffies)) {
1135                         printk(KERN_WARNING
1136                                "NFS: Server wrote zero bytes, expected %u.\n",
1137                                         argp->count);
1138                         complain = jiffies + 300 * HZ;
1139                 }
1140                 /* Can't do anything about it except throw an error. */
1141                 task->tk_status = -EIO;
1142         }
1143         return 0;
1144 }
1145
1146
1147 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1148 void nfs_commit_release(void *wdata)
1149 {
1150         nfs_commit_free(wdata);
1151 }
1152
1153 /*
1154  * Set up the argument/result storage required for the RPC call.
1155  */
1156 static void nfs_commit_rpcsetup(struct list_head *head,
1157                 struct nfs_write_data *data,
1158                 int how)
1159 {
1160         struct nfs_page         *first;
1161         struct inode            *inode;
1162         int flags;
1163
1164         /* Set up the RPC argument and reply structs
1165          * NB: take care not to mess about with data->commit et al. */
1166
1167         list_splice_init(head, &data->pages);
1168         first = nfs_list_entry(data->pages.next);
1169         inode = first->wb_context->path.dentry->d_inode;
1170
1171         data->inode       = inode;
1172         data->cred        = first->wb_context->cred;
1173
1174         data->args.fh     = NFS_FH(data->inode);
1175         /* Note: we always request a commit of the entire inode */
1176         data->args.offset = 0;
1177         data->args.count  = 0;
1178         data->res.count   = 0;
1179         data->res.fattr   = &data->fattr;
1180         data->res.verf    = &data->verf;
1181         nfs_fattr_init(&data->fattr);
1182
1183         /* Set up the initial task struct.  */
1184         flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1185         rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data);
1186         NFS_PROTO(inode)->commit_setup(data, how);
1187
1188         data->task.tk_priority = flush_task_priority(how);
1189         data->task.tk_cookie = (unsigned long)inode;
1190
1191         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1192 }
1193
1194 /*
1195  * Commit dirty pages
1196  */
1197 static int
1198 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1199 {
1200         struct nfs_write_data   *data;
1201         struct nfs_page         *req;
1202
1203         data = nfs_commit_alloc();
1204
1205         if (!data)
1206                 goto out_bad;
1207
1208         /* Set up the argument struct */
1209         nfs_commit_rpcsetup(head, data, how);
1210
1211         nfs_execute_write(data);
1212         return 0;
1213  out_bad:
1214         while (!list_empty(head)) {
1215                 req = nfs_list_entry(head->next);
1216                 nfs_list_remove_request(req);
1217                 nfs_mark_request_commit(req);
1218                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1219                 nfs_clear_page_tag_locked(req);
1220         }
1221         return -ENOMEM;
1222 }
1223
1224 /*
1225  * COMMIT call returned
1226  */
1227 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1228 {
1229         struct nfs_write_data   *data = calldata;
1230         struct nfs_page         *req;
1231
1232         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1233                                 task->tk_pid, task->tk_status);
1234
1235         /* Call the NFS version-specific code */
1236         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1237                 return;
1238
1239         while (!list_empty(&data->pages)) {
1240                 req = nfs_list_entry(data->pages.next);
1241                 nfs_list_remove_request(req);
1242                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1243                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1244
1245                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1246                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1247                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1248                         req->wb_bytes,
1249                         (long long)req_offset(req));
1250                 if (task->tk_status < 0) {
1251                         req->wb_context->error = task->tk_status;
1252                         nfs_inode_remove_request(req);
1253                         dprintk(", error = %d\n", task->tk_status);
1254                         goto next;
1255                 }
1256
1257                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1258                  * returned by the server against all stored verfs. */
1259                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1260                         /* We have a match */
1261                         /* Set the PG_uptodate flag */
1262                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1263                                         req->wb_bytes);
1264                         nfs_inode_remove_request(req);
1265                         dprintk(" OK\n");
1266                         goto next;
1267                 }
1268                 /* We have a mismatch. Write the page again */
1269                 dprintk(" mismatch\n");
1270                 nfs_redirty_request(req);
1271         next:
1272                 nfs_clear_page_tag_locked(req);
1273         }
1274 }
1275
1276 static const struct rpc_call_ops nfs_commit_ops = {
1277         .rpc_call_done = nfs_commit_done,
1278         .rpc_release = nfs_commit_release,
1279 };
1280
1281 int nfs_commit_inode(struct inode *inode, int how)
1282 {
1283         struct nfs_inode *nfsi = NFS_I(inode);
1284         LIST_HEAD(head);
1285         int res;
1286
1287         spin_lock(&nfsi->req_lock);
1288         res = nfs_scan_commit(inode, &head, 0, 0);
1289         spin_unlock(&nfsi->req_lock);
1290         if (res) {
1291                 int error = nfs_commit_list(inode, &head, how);
1292                 if (error < 0)
1293                         return error;
1294         }
1295         return res;
1296 }
1297 #else
1298 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1299 {
1300         return 0;
1301 }
1302 #endif
1303
1304 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1305 {
1306         struct inode *inode = mapping->host;
1307         struct nfs_inode *nfsi = NFS_I(inode);
1308         pgoff_t idx_start, idx_end;
1309         unsigned int npages = 0;
1310         LIST_HEAD(head);
1311         int nocommit = how & FLUSH_NOCOMMIT;
1312         long pages, ret;
1313
1314         /* FIXME */
1315         if (wbc->range_cyclic)
1316                 idx_start = 0;
1317         else {
1318                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1319                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1320                 if (idx_end > idx_start) {
1321                         pgoff_t l_npages = 1 + idx_end - idx_start;
1322                         npages = l_npages;
1323                         if (sizeof(npages) != sizeof(l_npages) &&
1324                                         (pgoff_t)npages != l_npages)
1325                                 npages = 0;
1326                 }
1327         }
1328         how &= ~FLUSH_NOCOMMIT;
1329         spin_lock(&nfsi->req_lock);
1330         do {
1331                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1332                 if (ret != 0)
1333                         continue;
1334                 if (nocommit)
1335                         break;
1336                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1337                 if (pages == 0)
1338                         break;
1339                 if (how & FLUSH_INVALIDATE) {
1340                         spin_unlock(&nfsi->req_lock);
1341                         nfs_cancel_commit_list(&head);
1342                         ret = pages;
1343                         spin_lock(&nfsi->req_lock);
1344                         continue;
1345                 }
1346                 pages += nfs_scan_commit(inode, &head, 0, 0);
1347                 spin_unlock(&nfsi->req_lock);
1348                 ret = nfs_commit_list(inode, &head, how);
1349                 spin_lock(&nfsi->req_lock);
1350         } while (ret >= 0);
1351         spin_unlock(&nfsi->req_lock);
1352         return ret;
1353 }
1354
1355 /*
1356  * flush the inode to disk.
1357  */
1358 int nfs_wb_all(struct inode *inode)
1359 {
1360         struct address_space *mapping = inode->i_mapping;
1361         struct writeback_control wbc = {
1362                 .bdi = mapping->backing_dev_info,
1363                 .sync_mode = WB_SYNC_ALL,
1364                 .nr_to_write = LONG_MAX,
1365                 .for_writepages = 1,
1366                 .range_cyclic = 1,
1367         };
1368         int ret;
1369
1370         ret = nfs_writepages(mapping, &wbc);
1371         if (ret < 0)
1372                 goto out;
1373         ret = nfs_sync_mapping_wait(mapping, &wbc, 0);
1374         if (ret >= 0)
1375                 return 0;
1376 out:
1377         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1378         return ret;
1379 }
1380
1381 int nfs_sync_mapping_range(struct address_space *mapping, loff_t range_start, loff_t range_end, int how)
1382 {
1383         struct writeback_control wbc = {
1384                 .bdi = mapping->backing_dev_info,
1385                 .sync_mode = WB_SYNC_ALL,
1386                 .nr_to_write = LONG_MAX,
1387                 .range_start = range_start,
1388                 .range_end = range_end,
1389                 .for_writepages = 1,
1390         };
1391         int ret;
1392
1393         ret = nfs_writepages(mapping, &wbc);
1394         if (ret < 0)
1395                 goto out;
1396         ret = nfs_sync_mapping_wait(mapping, &wbc, how);
1397         if (ret >= 0)
1398                 return 0;
1399 out:
1400         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1401         return ret;
1402 }
1403
1404 int nfs_wb_page_priority(struct inode *inode, struct page *page, int how)
1405 {
1406         loff_t range_start = page_offset(page);
1407         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1408         struct writeback_control wbc = {
1409                 .bdi = page->mapping->backing_dev_info,
1410                 .sync_mode = WB_SYNC_ALL,
1411                 .nr_to_write = LONG_MAX,
1412                 .range_start = range_start,
1413                 .range_end = range_end,
1414         };
1415         int ret;
1416
1417         BUG_ON(!PageLocked(page));
1418         if (clear_page_dirty_for_io(page)) {
1419                 ret = nfs_writepage_locked(page, &wbc);
1420                 if (ret < 0)
1421                         goto out;
1422         }
1423         if (!PagePrivate(page))
1424                 return 0;
1425         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1426         if (ret >= 0)
1427                 return 0;
1428 out:
1429         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1430         return ret;
1431 }
1432
1433 /*
1434  * Write back all requests on one page - we do this before reading it.
1435  */
1436 int nfs_wb_page(struct inode *inode, struct page* page)
1437 {
1438         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1439 }
1440
1441 int nfs_set_page_dirty(struct page *page)
1442 {
1443         struct address_space *mapping = page->mapping;
1444         struct inode *inode;
1445         spinlock_t *req_lock;
1446         struct nfs_page *req;
1447         int ret;
1448
1449         if (!mapping)
1450                 goto out_raced;
1451         inode = mapping->host;
1452         if (!inode)
1453                 goto out_raced;
1454         req_lock = &NFS_I(inode)->req_lock;
1455         spin_lock(req_lock);
1456         req = nfs_page_find_request_locked(page);
1457         if (req != NULL) {
1458                 /* Mark any existing write requests for flushing */
1459                 ret = !test_and_set_bit(PG_NEED_FLUSH, &req->wb_flags);
1460                 spin_unlock(req_lock);
1461                 nfs_release_request(req);
1462                 return ret;
1463         }
1464         ret = __set_page_dirty_nobuffers(page);
1465         spin_unlock(req_lock);
1466         return ret;
1467 out_raced:
1468         return !TestSetPageDirty(page);
1469 }
1470
1471
1472 int __init nfs_init_writepagecache(void)
1473 {
1474         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1475                                              sizeof(struct nfs_write_data),
1476                                              0, SLAB_HWCACHE_ALIGN,
1477                                              NULL, NULL);
1478         if (nfs_wdata_cachep == NULL)
1479                 return -ENOMEM;
1480
1481         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1482                                                      nfs_wdata_cachep);
1483         if (nfs_wdata_mempool == NULL)
1484                 return -ENOMEM;
1485
1486         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1487                                                       nfs_wdata_cachep);
1488         if (nfs_commit_mempool == NULL)
1489                 return -ENOMEM;
1490
1491         /*
1492          * NFS congestion size, scale with available memory.
1493          *
1494          *  64MB:    8192k
1495          * 128MB:   11585k
1496          * 256MB:   16384k
1497          * 512MB:   23170k
1498          *   1GB:   32768k
1499          *   2GB:   46340k
1500          *   4GB:   65536k
1501          *   8GB:   92681k
1502          *  16GB:  131072k
1503          *
1504          * This allows larger machines to have larger/more transfers.
1505          * Limit the default to 256M
1506          */
1507         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1508         if (nfs_congestion_kb > 256*1024)
1509                 nfs_congestion_kb = 256*1024;
1510
1511         return 0;
1512 }
1513
1514 void nfs_destroy_writepagecache(void)
1515 {
1516         mempool_destroy(nfs_commit_mempool);
1517         mempool_destroy(nfs_wdata_mempool);
1518         kmem_cache_destroy(nfs_wdata_cachep);
1519 }
1520