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 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23
  24 #include <asm/uaccess.h>
  25
  26 #include "delegation.h"
  27 #include "internal.h"
  28 #include "iostat.h"
  29 #include "nfs4_fs.h"
  30 #include "fscache.h"
  31
  32 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  33
  34 #define MIN_POOL_WRITE          (32)
  35 #define MIN_POOL_COMMIT         (4)
  36
  37 /*
  38  * Local function declarations
  39  */
  40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static void nfs_redirty_request(struct nfs_page *req);
  43 static const struct rpc_call_ops nfs_write_partial_ops;
  44 static const struct rpc_call_ops nfs_write_full_ops;
  45 static const struct rpc_call_ops nfs_commit_ops;
  46
  47 static struct kmem_cache *nfs_wdata_cachep;
  48 static mempool_t *nfs_wdata_mempool;
  49 static mempool_t *nfs_commit_mempool;
  50
  51 struct nfs_write_data *nfs_commitdata_alloc(void)
  52 {
  53         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  54
  55         if (p) {
  56                 memset(p, 0, sizeof(*p));
  57                 INIT_LIST_HEAD(&p->pages);
  58                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
  59         }
  60         return p;
  61 }
  62
  63 void nfs_commit_free(struct nfs_write_data *p)
  64 {
  65         if (p && (p->pagevec != &p->page_array[0]))
  66                 kfree(p->pagevec);
  67         mempool_free(p, nfs_commit_mempool);
  68 }
  69
  70 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  71 {
  72         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  73
  74         if (p) {
  75                 memset(p, 0, sizeof(*p));
  76                 INIT_LIST_HEAD(&p->pages);
  77                 p->npages = pagecount;
  78                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
  79                 if (pagecount <= ARRAY_SIZE(p->page_array))
  80                         p->pagevec = p->page_array;
  81                 else {
  82                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  83                         if (!p->pagevec) {
  84                                 mempool_free(p, nfs_wdata_mempool);
  85                                 p = NULL;
  86                         }
  87                 }
  88         }
  89         return p;
  90 }
  91
  92 void nfs_writedata_free(struct nfs_write_data *p)
  93 {
  94         if (p && (p->pagevec != &p->page_array[0]))
  95                 kfree(p->pagevec);
  96         mempool_free(p, nfs_wdata_mempool);
  97 }
  98
  99 static void nfs_writedata_release(struct nfs_write_data *wdata)
 100 {
 101         put_nfs_open_context(wdata->args.context);
 102         nfs_writedata_free(wdata);
 103 }
 104
 105 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 106 {
 107         ctx->error = error;
 108         smp_wmb();
 109         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 110 }
 111
 112 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 113 {
 114         struct nfs_page *req = NULL;
 115
 116         if (PagePrivate(page)) {
 117                 req = (struct nfs_page *)page_private(page);
 118                 if (req != NULL)
 119                         kref_get(&req->wb_kref);
 120         }
 121         return req;
 122 }
 123
 124 static struct nfs_page *nfs_page_find_request(struct page *page)
 125 {
 126         struct inode *inode = page->mapping->host;
 127         struct nfs_page *req = NULL;
 128
 129         spin_lock(&inode->i_lock);
 130         req = nfs_page_find_request_locked(page);
 131         spin_unlock(&inode->i_lock);
 132         return req;
 133 }
 134
 135 /* Adjust the file length if we're writing beyond the end */
 136 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 137 {
 138         struct inode *inode = page->mapping->host;
 139         loff_t end, i_size;
 140         pgoff_t end_index;
 141
 142         spin_lock(&inode->i_lock);
 143         i_size = i_size_read(inode);
 144         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 145         if (i_size > 0 && page->index < end_index)
 146                 goto out;
 147         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 148         if (i_size >= end)
 149                 goto out;
 150         i_size_write(inode, end);
 151         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 152 out:
 153         spin_unlock(&inode->i_lock);
 154 }
 155
 156 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 157 static void nfs_set_pageerror(struct page *page)
 158 {
 159         SetPageError(page);
 160         nfs_zap_mapping(page->mapping->host, page->mapping);
 161 }
 162
 163 /* We can set the PG_uptodate flag if we see that a write request
 164  * covers the full page.
 165  */
 166 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 167 {
 168         if (PageUptodate(page))
 169                 return;
 170         if (base != 0)
 171                 return;
 172         if (count != nfs_page_length(page))
 173                 return;
 174         SetPageUptodate(page);
 175 }
 176
 177 static int wb_priority(struct writeback_control *wbc)
 178 {
 179         if (wbc->for_reclaim)
 180                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 181         if (wbc->for_kupdate)
 182                 return FLUSH_LOWPRI;
 183         return 0;
 184 }
 185
 186 /*
 187  * NFS congestion control
 188  */
 189
 190 int nfs_congestion_kb;
 191
 192 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 193 #define NFS_CONGESTION_OFF_THRESH       \
 194         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 195
 196 static int nfs_set_page_writeback(struct page *page)
 197 {
 198         int ret = test_set_page_writeback(page);
 199
 200         if (!ret) {
 201                 struct inode *inode = page->mapping->host;
 202                 struct nfs_server *nfss = NFS_SERVER(inode);
 203
 204                 if (atomic_long_inc_return(&nfss->writeback) >
 205                                 NFS_CONGESTION_ON_THRESH) {
 206                         set_bdi_congested(&nfss->backing_dev_info,
 207                                                 BLK_RW_ASYNC);
 208                 }
 209         }
 210         return ret;
 211 }
 212
 213 static void nfs_end_page_writeback(struct page *page)
 214 {
 215         struct inode *inode = page->mapping->host;
 216         struct nfs_server *nfss = NFS_SERVER(inode);
 217
 218         end_page_writeback(page);
 219         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 220                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 221 }
 222
 223 static struct nfs_page *nfs_find_and_lock_request(struct page *page)
 224 {
 225         struct inode *inode = page->mapping->host;
 226         struct nfs_page *req;
 227         int ret;
 228
 229         spin_lock(&inode->i_lock);
 230         for (;;) {
 231                 req = nfs_page_find_request_locked(page);
 232                 if (req == NULL)
 233                         break;
 234                 if (nfs_set_page_tag_locked(req))
 235                         break;
 236                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 237                  *       then the call to nfs_set_page_tag_locked() will always
 238                  *       succeed provided that someone hasn't already marked the
 239                  *       request as dirty (in which case we don't care).
 240                  */
 241                 spin_unlock(&inode->i_lock);
 242                 ret = nfs_wait_on_request(req);
 243                 nfs_release_request(req);
 244                 if (ret != 0)
 245                         return ERR_PTR(ret);
 246                 spin_lock(&inode->i_lock);
 247         }
 248         spin_unlock(&inode->i_lock);
 249         return req;
 250 }
 251
 252 /*
 253  * Find an associated nfs write request, and prepare to flush it out
 254  * May return an error if the user signalled nfs_wait_on_request().
 255  */
 256 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 257                                 struct page *page)
 258 {
 259         struct nfs_page *req;
 260         int ret = 0;
 261
 262         req = nfs_find_and_lock_request(page);
 263         if (!req)
 264                 goto out;
 265         ret = PTR_ERR(req);
 266         if (IS_ERR(req))
 267                 goto out;
 268
 269         ret = nfs_set_page_writeback(page);
 270         BUG_ON(ret != 0);
 271         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
 272
 273         if (!nfs_pageio_add_request(pgio, req)) {
 274                 nfs_redirty_request(req);
 275                 ret = pgio->pg_error;
 276         }
 277 out:
 278         return ret;
 279 }
 280
 281 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 282 {
 283         struct inode *inode = page->mapping->host;
 284
 285         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 286         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 287
 288         nfs_pageio_cond_complete(pgio, page->index);
 289         return nfs_page_async_flush(pgio, page);
 290 }
 291
 292 /*
 293  * Write an mmapped page to the server.
 294  */
 295 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 296 {
 297         struct nfs_pageio_descriptor pgio;
 298         int err;
 299
 300         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 301         err = nfs_do_writepage(page, wbc, &pgio);
 302         nfs_pageio_complete(&pgio);
 303         if (err < 0)
 304                 return err;
 305         if (pgio.pg_error < 0)
 306                 return pgio.pg_error;
 307         return 0;
 308 }
 309
 310 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 311 {
 312         int ret;
 313
 314         ret = nfs_writepage_locked(page, wbc);
 315         unlock_page(page);
 316         return ret;
 317 }
 318
 319 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 320 {
 321         int ret;
 322
 323         ret = nfs_do_writepage(page, wbc, data);
 324         unlock_page(page);
 325         return ret;
 326 }
 327
 328 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 329 {
 330         struct inode *inode = mapping->host;
 331         unsigned long *bitlock = &NFS_I(inode)->flags;
 332         struct nfs_pageio_descriptor pgio;
 333         int err;
 334
 335         /* Stop dirtying of new pages while we sync */
 336         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 337                         nfs_wait_bit_killable, TASK_KILLABLE);
 338         if (err)
 339                 goto out_err;
 340
 341         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 342
 343         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 344         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 345         nfs_pageio_complete(&pgio);
 346
 347         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 348         smp_mb__after_clear_bit();
 349         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 350
 351         if (err < 0)
 352                 goto out_err;
 353         err = pgio.pg_error;
 354         if (err < 0)
 355                 goto out_err;
 356         return 0;
 357 out_err:
 358         return err;
 359 }
 360
 361 /*
 362  * Insert a write request into an inode
 363  */
 364 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 365 {
 366         struct nfs_inode *nfsi = NFS_I(inode);
 367         int error;
 368
 369         error = radix_tree_preload(GFP_NOFS);
 370         if (error != 0)
 371                 goto out;
 372
 373         /* Lock the request! */
 374         nfs_lock_request_dontget(req);
 375
 376         spin_lock(&inode->i_lock);
 377         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 378         BUG_ON(error);
 379         if (!nfsi->npages) {
 380                 igrab(inode);
 381                 if (nfs_have_delegation(inode, FMODE_WRITE))
 382                         nfsi->change_attr++;
 383         }
 384         SetPagePrivate(req->wb_page);
 385         set_page_private(req->wb_page, (unsigned long)req);
 386         nfsi->npages++;
 387         kref_get(&req->wb_kref);
 388         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 389                                 NFS_PAGE_TAG_LOCKED);
 390         spin_unlock(&inode->i_lock);
 391         radix_tree_preload_end();
 392 out:
 393         return error;
 394 }
 395
 396 /*
 397  * Remove a write request from an inode
 398  */
 399 static void nfs_inode_remove_request(struct nfs_page *req)
 400 {
 401         struct inode *inode = req->wb_context->path.dentry->d_inode;
 402         struct nfs_inode *nfsi = NFS_I(inode);
 403
 404         BUG_ON (!NFS_WBACK_BUSY(req));
 405
 406         spin_lock(&inode->i_lock);
 407         set_page_private(req->wb_page, 0);
 408         ClearPagePrivate(req->wb_page);
 409         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 410         nfsi->npages--;
 411         if (!nfsi->npages) {
 412                 spin_unlock(&inode->i_lock);
 413                 iput(inode);
 414         } else
 415                 spin_unlock(&inode->i_lock);
 416         nfs_clear_request(req);
 417         nfs_release_request(req);
 418 }
 419
 420 static void
 421 nfs_mark_request_dirty(struct nfs_page *req)
 422 {
 423         __set_page_dirty_nobuffers(req->wb_page);
 424 }
 425
 426 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 427 /*
 428  * Add a request to the inode's commit list.
 429  */
 430 static void
 431 nfs_mark_request_commit(struct nfs_page *req)
 432 {
 433         struct inode *inode = req->wb_context->path.dentry->d_inode;
 434         struct nfs_inode *nfsi = NFS_I(inode);
 435
 436         spin_lock(&inode->i_lock);
 437         set_bit(PG_CLEAN, &(req)->wb_flags);
 438         radix_tree_tag_set(&nfsi->nfs_page_tree,
 439                         req->wb_index,
 440                         NFS_PAGE_TAG_COMMIT);
 441         spin_unlock(&inode->i_lock);
 442         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 443         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 444         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 445 }
 446
 447 static int
 448 nfs_clear_request_commit(struct nfs_page *req)
 449 {
 450         struct page *page = req->wb_page;
 451
 452         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 453                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 454                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 455                 return 1;
 456         }
 457         return 0;
 458 }
 459
 460 static inline
 461 int nfs_write_need_commit(struct nfs_write_data *data)
 462 {
 463         return data->verf.committed != NFS_FILE_SYNC;
 464 }
 465
 466 static inline
 467 int nfs_reschedule_unstable_write(struct nfs_page *req)
 468 {
 469         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 470                 nfs_mark_request_commit(req);
 471                 return 1;
 472         }
 473         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 474                 nfs_mark_request_dirty(req);
 475                 return 1;
 476         }
 477         return 0;
 478 }
 479 #else
 480 static inline void
 481 nfs_mark_request_commit(struct nfs_page *req)
 482 {
 483 }
 484
 485 static inline int
 486 nfs_clear_request_commit(struct nfs_page *req)
 487 {
 488         return 0;
 489 }
 490
 491 static inline
 492 int nfs_write_need_commit(struct nfs_write_data *data)
 493 {
 494         return 0;
 495 }
 496
 497 static inline
 498 int nfs_reschedule_unstable_write(struct nfs_page *req)
 499 {
 500         return 0;
 501 }
 502 #endif
 503
 504 /*
 505  * Wait for a request to complete.
 506  *
 507  * Interruptible by fatal signals only.
 508  */
 509 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 510 {
 511         struct nfs_inode *nfsi = NFS_I(inode);
 512         struct nfs_page *req;
 513         pgoff_t idx_end, next;
 514         unsigned int            res = 0;
 515         int                     error;
 516
 517         if (npages == 0)
 518                 idx_end = ~0;
 519         else
 520                 idx_end = idx_start + npages - 1;
 521
 522         next = idx_start;
 523         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 524                 if (req->wb_index > idx_end)
 525                         break;
 526
 527                 next = req->wb_index + 1;
 528                 BUG_ON(!NFS_WBACK_BUSY(req));
 529
 530                 kref_get(&req->wb_kref);
 531                 spin_unlock(&inode->i_lock);
 532                 error = nfs_wait_on_request(req);
 533                 nfs_release_request(req);
 534                 spin_lock(&inode->i_lock);
 535                 if (error < 0)
 536                         return error;
 537                 res++;
 538         }
 539         return res;
 540 }
 541
 542 static void nfs_cancel_commit_list(struct list_head *head)
 543 {
 544         struct nfs_page *req;
 545
 546         while(!list_empty(head)) {
 547                 req = nfs_list_entry(head->next);
 548                 nfs_list_remove_request(req);
 549                 nfs_clear_request_commit(req);
 550                 nfs_inode_remove_request(req);
 551                 nfs_unlock_request(req);
 552         }
 553 }
 554
 555 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 556 static int
 557 nfs_need_commit(struct nfs_inode *nfsi)
 558 {
 559         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 560 }
 561
 562 /*
 563  * nfs_scan_commit - Scan an inode for commit requests
 564  * @inode: NFS inode to scan
 565  * @dst: destination list
 566  * @idx_start: lower bound of page->index to scan.
 567  * @npages: idx_start + npages sets the upper bound to scan.
 568  *
 569  * Moves requests from the inode's 'commit' request list.
 570  * The requests are *not* checked to ensure that they form a contiguous set.
 571  */
 572 static int
 573 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 574 {
 575         struct nfs_inode *nfsi = NFS_I(inode);
 576
 577         if (!nfs_need_commit(nfsi))
 578                 return 0;
 579
 580         return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 581 }
 582 #else
 583 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 584 {
 585         return 0;
 586 }
 587
 588 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 589 {
 590         return 0;
 591 }
 592 #endif
 593
 594 /*
 595  * Search for an existing write request, and attempt to update
 596  * it to reflect a new dirty region on a given page.
 597  *
 598  * If the attempt fails, then the existing request is flushed out
 599  * to disk.
 600  */
 601 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 602                 struct page *page,
 603                 unsigned int offset,
 604                 unsigned int bytes)
 605 {
 606         struct nfs_page *req;
 607         unsigned int rqend;
 608         unsigned int end;
 609         int error;
 610
 611         if (!PagePrivate(page))
 612                 return NULL;
 613
 614         end = offset + bytes;
 615         spin_lock(&inode->i_lock);
 616
 617         for (;;) {
 618                 req = nfs_page_find_request_locked(page);
 619                 if (req == NULL)
 620                         goto out_unlock;
 621
 622                 rqend = req->wb_offset + req->wb_bytes;
 623                 /*
 624                  * Tell the caller to flush out the request if
 625                  * the offsets are non-contiguous.
 626                  * Note: nfs_flush_incompatible() will already
 627                  * have flushed out requests having wrong owners.
 628                  */
 629                 if (offset > rqend
 630                     || end < req->wb_offset)
 631                         goto out_flushme;
 632
 633                 if (nfs_set_page_tag_locked(req))
 634                         break;
 635
 636                 /* The request is locked, so wait and then retry */
 637                 spin_unlock(&inode->i_lock);
 638                 error = nfs_wait_on_request(req);
 639                 nfs_release_request(req);
 640                 if (error != 0)
 641                         goto out_err;
 642                 spin_lock(&inode->i_lock);
 643         }
 644
 645         if (nfs_clear_request_commit(req))
 646                 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 647                                 req->wb_index, NFS_PAGE_TAG_COMMIT);
 648
 649         /* Okay, the request matches. Update the region */
 650         if (offset < req->wb_offset) {
 651                 req->wb_offset = offset;
 652                 req->wb_pgbase = offset;
 653         }
 654         if (end > rqend)
 655                 req->wb_bytes = end - req->wb_offset;
 656         else
 657                 req->wb_bytes = rqend - req->wb_offset;
 658 out_unlock:
 659         spin_unlock(&inode->i_lock);
 660         return req;
 661 out_flushme:
 662         spin_unlock(&inode->i_lock);
 663         nfs_release_request(req);
 664         error = nfs_wb_page(inode, page);
 665 out_err:
 666         return ERR_PTR(error);
 667 }
 668
 669 /*
 670  * Try to update an existing write request, or create one if there is none.
 671  *
 672  * Note: Should always be called with the Page Lock held to prevent races
 673  * if we have to add a new request. Also assumes that the caller has
 674  * already called nfs_flush_incompatible() if necessary.
 675  */
 676 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 677                 struct page *page, unsigned int offset, unsigned int bytes)
 678 {
 679         struct inode *inode = page->mapping->host;
 680         struct nfs_page *req;
 681         int error;
 682
 683         req = nfs_try_to_update_request(inode, page, offset, bytes);
 684         if (req != NULL)
 685                 goto out;
 686         req = nfs_create_request(ctx, inode, page, offset, bytes);
 687         if (IS_ERR(req))
 688                 goto out;
 689         error = nfs_inode_add_request(inode, req);
 690         if (error != 0) {
 691                 nfs_release_request(req);
 692                 req = ERR_PTR(error);
 693         }
 694 out:
 695         return req;
 696 }
 697
 698 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 699                 unsigned int offset, unsigned int count)
 700 {
 701         struct nfs_page *req;
 702
 703         req = nfs_setup_write_request(ctx, page, offset, count);
 704         if (IS_ERR(req))
 705                 return PTR_ERR(req);
 706         /* Update file length */
 707         nfs_grow_file(page, offset, count);
 708         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 709         nfs_clear_page_tag_locked(req);
 710         return 0;
 711 }
 712
 713 int nfs_flush_incompatible(struct file *file, struct page *page)
 714 {
 715         struct nfs_open_context *ctx = nfs_file_open_context(file);
 716         struct nfs_page *req;
 717         int do_flush, status;
 718         /*
 719          * Look for a request corresponding to this page. If there
 720          * is one, and it belongs to another file, we flush it out
 721          * before we try to copy anything into the page. Do this
 722          * due to the lack of an ACCESS-type call in NFSv2.
 723          * Also do the same if we find a request from an existing
 724          * dropped page.
 725          */
 726         do {
 727                 req = nfs_page_find_request(page);
 728                 if (req == NULL)
 729                         return 0;
 730                 do_flush = req->wb_page != page || req->wb_context != ctx;
 731                 nfs_release_request(req);
 732                 if (!do_flush)
 733                         return 0;
 734                 status = nfs_wb_page(page->mapping->host, page);
 735         } while (status == 0);
 736         return status;
 737 }
 738
 739 /*
 740  * If the page cache is marked as unsafe or invalid, then we can't rely on
 741  * the PageUptodate() flag. In this case, we will need to turn off
 742  * write optimisations that depend on the page contents being correct.
 743  */
 744 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 745 {
 746         return PageUptodate(page) &&
 747                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 748 }
 749
 750 /*
 751  * Update and possibly write a cached page of an NFS file.
 752  *
 753  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 754  * things with a page scheduled for an RPC call (e.g. invalidate it).
 755  */
 756 int nfs_updatepage(struct file *file, struct page *page,
 757                 unsigned int offset, unsigned int count)
 758 {
 759         struct nfs_open_context *ctx = nfs_file_open_context(file);
 760         struct inode    *inode = page->mapping->host;
 761         int             status = 0;
 762
 763         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 764
 765         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 766                 file->f_path.dentry->d_parent->d_name.name,
 767                 file->f_path.dentry->d_name.name, count,
 768                 (long long)(page_offset(page) + offset));
 769
 770         /* If we're not using byte range locks, and we know the page
 771          * is up to date, it may be more efficient to extend the write
 772          * to cover the entire page in order to avoid fragmentation
 773          * inefficiencies.
 774          */
 775         if (nfs_write_pageuptodate(page, inode) &&
 776                         inode->i_flock == NULL &&
 777                         !(file->f_flags & O_SYNC)) {
 778                 count = max(count + offset, nfs_page_length(page));
 779                 offset = 0;
 780         }
 781
 782         status = nfs_writepage_setup(ctx, page, offset, count);
 783         if (status < 0)
 784                 nfs_set_pageerror(page);
 785         else
 786                 __set_page_dirty_nobuffers(page);
 787
 788         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 789                         status, (long long)i_size_read(inode));
 790         return status;
 791 }
 792
 793 static void nfs_writepage_release(struct nfs_page *req)
 794 {
 795
 796         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 797                 nfs_end_page_writeback(req->wb_page);
 798                 nfs_inode_remove_request(req);
 799         } else
 800                 nfs_end_page_writeback(req->wb_page);
 801         nfs_clear_page_tag_locked(req);
 802 }
 803
 804 static int flush_task_priority(int how)
 805 {
 806         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 807                 case FLUSH_HIGHPRI:
 808                         return RPC_PRIORITY_HIGH;
 809                 case FLUSH_LOWPRI:
 810                         return RPC_PRIORITY_LOW;
 811         }
 812         return RPC_PRIORITY_NORMAL;
 813 }
 814
 815 /*
 816  * Set up the argument/result storage required for the RPC call.
 817  */
 818 static int nfs_write_rpcsetup(struct nfs_page *req,
 819                 struct nfs_write_data *data,
 820                 const struct rpc_call_ops *call_ops,
 821                 unsigned int count, unsigned int offset,
 822                 int how)
 823 {
 824         struct inode *inode = req->wb_context->path.dentry->d_inode;
 825         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 826         int priority = flush_task_priority(how);
 827         struct rpc_task *task;
 828         struct rpc_message msg = {
 829                 .rpc_argp = &data->args,
 830                 .rpc_resp = &data->res,
 831                 .rpc_cred = req->wb_context->cred,
 832         };
 833         struct rpc_task_setup task_setup_data = {
 834                 .rpc_client = NFS_CLIENT(inode),
 835                 .task = &data->task,
 836                 .rpc_message = &msg,
 837                 .callback_ops = call_ops,
 838                 .callback_data = data,
 839                 .workqueue = nfsiod_workqueue,
 840                 .flags = flags,
 841                 .priority = priority,
 842         };
 843
 844         /* Set up the RPC argument and reply structs
 845          * NB: take care not to mess about with data->commit et al. */
 846
 847         data->req = req;
 848         data->inode = inode = req->wb_context->path.dentry->d_inode;
 849         data->cred = msg.rpc_cred;
 850
 851         data->args.fh     = NFS_FH(inode);
 852         data->args.offset = req_offset(req) + offset;
 853         data->args.pgbase = req->wb_pgbase + offset;
 854         data->args.pages  = data->pagevec;
 855         data->args.count  = count;
 856         data->args.context = get_nfs_open_context(req->wb_context);
 857         data->args.stable  = NFS_UNSTABLE;
 858         if (how & FLUSH_STABLE) {
 859                 data->args.stable = NFS_DATA_SYNC;
 860                 if (!nfs_need_commit(NFS_I(inode)))
 861                         data->args.stable = NFS_FILE_SYNC;
 862         }
 863
 864         data->res.fattr   = &data->fattr;
 865         data->res.count   = count;
 866         data->res.verf    = &data->verf;
 867         nfs_fattr_init(&data->fattr);
 868
 869         /* Set up the initial task struct.  */
 870         NFS_PROTO(inode)->write_setup(data, &msg);
 871
 872         dprintk("NFS: %5u initiated write call "
 873                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 874                 data->task.tk_pid,
 875                 inode->i_sb->s_id,
 876                 (long long)NFS_FILEID(inode),
 877                 count,
 878                 (unsigned long long)data->args.offset);
 879
 880         task = rpc_run_task(&task_setup_data);
 881         if (IS_ERR(task))
 882                 return PTR_ERR(task);
 883         rpc_put_task(task);
 884         return 0;
 885 }
 886
 887 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 888  * call this on each, which will prepare them to be retried on next
 889  * writeback using standard nfs.
 890  */
 891 static void nfs_redirty_request(struct nfs_page *req)
 892 {
 893         nfs_mark_request_dirty(req);
 894         nfs_end_page_writeback(req->wb_page);
 895         nfs_clear_page_tag_locked(req);
 896 }
 897
 898 /*
 899  * Generate multiple small requests to write out a single
 900  * contiguous dirty area on one page.
 901  */
 902 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 903 {
 904         struct nfs_page *req = nfs_list_entry(head->next);
 905         struct page *page = req->wb_page;
 906         struct nfs_write_data *data;
 907         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 908         unsigned int offset;
 909         int requests = 0;
 910         int ret = 0;
 911         LIST_HEAD(list);
 912
 913         nfs_list_remove_request(req);
 914
 915         nbytes = count;
 916         do {
 917                 size_t len = min(nbytes, wsize);
 918
 919                 data = nfs_writedata_alloc(1);
 920                 if (!data)
 921                         goto out_bad;
 922                 list_add(&data->pages, &list);
 923                 requests++;
 924                 nbytes -= len;
 925         } while (nbytes != 0);
 926         atomic_set(&req->wb_complete, requests);
 927
 928         ClearPageError(page);
 929         offset = 0;
 930         nbytes = count;
 931         do {
 932                 int ret2;
 933
 934                 data = list_entry(list.next, struct nfs_write_data, pages);
 935                 list_del_init(&data->pages);
 936
 937                 data->pagevec[0] = page;
 938
 939                 if (nbytes < wsize)
 940                         wsize = nbytes;
 941                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 942                                    wsize, offset, how);
 943                 if (ret == 0)
 944                         ret = ret2;
 945                 offset += wsize;
 946                 nbytes -= wsize;
 947         } while (nbytes != 0);
 948
 949         return ret;
 950
 951 out_bad:
 952         while (!list_empty(&list)) {
 953                 data = list_entry(list.next, struct nfs_write_data, pages);
 954                 list_del(&data->pages);
 955                 nfs_writedata_release(data);
 956         }
 957         nfs_redirty_request(req);
 958         return -ENOMEM;
 959 }
 960
 961 /*
 962  * Create an RPC task for the given write request and kick it.
 963  * The page must have been locked by the caller.
 964  *
 965  * It may happen that the page we're passed is not marked dirty.
 966  * This is the case if nfs_updatepage detects a conflicting request
 967  * that has been written but not committed.
 968  */
 969 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 970 {
 971         struct nfs_page         *req;
 972         struct page             **pages;
 973         struct nfs_write_data   *data;
 974
 975         data = nfs_writedata_alloc(npages);
 976         if (!data)
 977                 goto out_bad;
 978
 979         pages = data->pagevec;
 980         while (!list_empty(head)) {
 981                 req = nfs_list_entry(head->next);
 982                 nfs_list_remove_request(req);
 983                 nfs_list_add_request(req, &data->pages);
 984                 ClearPageError(req->wb_page);
 985                 *pages++ = req->wb_page;
 986         }
 987         req = nfs_list_entry(data->pages.next);
 988
 989         /* Set up the argument struct */
 990         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 991  out_bad:
 992         while (!list_empty(head)) {
 993                 req = nfs_list_entry(head->next);
 994                 nfs_list_remove_request(req);
 995                 nfs_redirty_request(req);
 996         }
 997         return -ENOMEM;
 998 }
 999
1000 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1001                                   struct inode *inode, int ioflags)
1002 {
1003         size_t wsize = NFS_SERVER(inode)->wsize;
1004
1005         if (wsize < PAGE_CACHE_SIZE)
1006                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1007         else
1008                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1009 }
1010
1011 /*
1012  * Handle a write reply that flushed part of a page.
1013  */
1014 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1015 {
1016         struct nfs_write_data   *data = calldata;
1017
1018         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1019                 task->tk_pid,
1020                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1021                 (long long)
1022                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1023                 data->req->wb_bytes, (long long)req_offset(data->req));
1024
1025         nfs_writeback_done(task, data);
1026 }
1027
1028 static void nfs_writeback_release_partial(void *calldata)
1029 {
1030         struct nfs_write_data   *data = calldata;
1031         struct nfs_page         *req = data->req;
1032         struct page             *page = req->wb_page;
1033         int status = data->task.tk_status;
1034
1035         if (status < 0) {
1036                 nfs_set_pageerror(page);
1037                 nfs_context_set_write_error(req->wb_context, status);
1038                 dprintk(", error = %d\n", status);
1039                 goto out;
1040         }
1041
1042         if (nfs_write_need_commit(data)) {
1043                 struct inode *inode = page->mapping->host;
1044
1045                 spin_lock(&inode->i_lock);
1046                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1047                         /* Do nothing we need to resend the writes */
1048                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1049                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1050                         dprintk(" defer commit\n");
1051                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1052                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1053                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1054                         dprintk(" server reboot detected\n");
1055                 }
1056                 spin_unlock(&inode->i_lock);
1057         } else
1058                 dprintk(" OK\n");
1059
1060 out:
1061         if (atomic_dec_and_test(&req->wb_complete))
1062                 nfs_writepage_release(req);
1063         nfs_writedata_release(calldata);
1064 }
1065
1066 #if defined(CONFIG_NFS_V4_1)
1067 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1068 {
1069         struct nfs_write_data *data = calldata;
1070         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1071
1072         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1073                                 &data->res.seq_res, 1, task))
1074                 return;
1075         rpc_call_start(task);
1076 }
1077 #endif /* CONFIG_NFS_V4_1 */
1078
1079 static const struct rpc_call_ops nfs_write_partial_ops = {
1080 #if defined(CONFIG_NFS_V4_1)
1081         .rpc_call_prepare = nfs_write_prepare,
1082 #endif /* CONFIG_NFS_V4_1 */
1083         .rpc_call_done = nfs_writeback_done_partial,
1084         .rpc_release = nfs_writeback_release_partial,
1085 };
1086
1087 /*
1088  * Handle a write reply that flushes a whole page.
1089  *
1090  * FIXME: There is an inherent race with invalidate_inode_pages and
1091  *        writebacks since the page->count is kept > 1 for as long
1092  *        as the page has a write request pending.
1093  */
1094 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1095 {
1096         struct nfs_write_data   *data = calldata;
1097
1098         nfs_writeback_done(task, data);
1099 }
1100
1101 static void nfs_writeback_release_full(void *calldata)
1102 {
1103         struct nfs_write_data   *data = calldata;
1104         int status = data->task.tk_status;
1105
1106         /* Update attributes as result of writeback. */
1107         while (!list_empty(&data->pages)) {
1108                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1109                 struct page *page = req->wb_page;
1110
1111                 nfs_list_remove_request(req);
1112
1113                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1114                         data->task.tk_pid,
1115                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1116                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1117                         req->wb_bytes,
1118                         (long long)req_offset(req));
1119
1120                 if (status < 0) {
1121                         nfs_set_pageerror(page);
1122                         nfs_context_set_write_error(req->wb_context, status);
1123                         dprintk(", error = %d\n", status);
1124                         goto remove_request;
1125                 }
1126
1127                 if (nfs_write_need_commit(data)) {
1128                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1129                         nfs_mark_request_commit(req);
1130                         nfs_end_page_writeback(page);
1131                         dprintk(" marked for commit\n");
1132                         goto next;
1133                 }
1134                 dprintk(" OK\n");
1135 remove_request:
1136                 nfs_end_page_writeback(page);
1137                 nfs_inode_remove_request(req);
1138         next:
1139                 nfs_clear_page_tag_locked(req);
1140         }
1141         nfs_writedata_release(calldata);
1142 }
1143
1144 static const struct rpc_call_ops nfs_write_full_ops = {
1145 #if defined(CONFIG_NFS_V4_1)
1146         .rpc_call_prepare = nfs_write_prepare,
1147 #endif /* CONFIG_NFS_V4_1 */
1148         .rpc_call_done = nfs_writeback_done_full,
1149         .rpc_release = nfs_writeback_release_full,
1150 };
1151
1152
1153 /*
1154  * This function is called when the WRITE call is complete.
1155  */
1156 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1157 {
1158         struct nfs_writeargs    *argp = &data->args;
1159         struct nfs_writeres     *resp = &data->res;
1160         struct nfs_server       *server = NFS_SERVER(data->inode);
1161         int status;
1162
1163         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1164                 task->tk_pid, task->tk_status);
1165
1166         /*
1167          * ->write_done will attempt to use post-op attributes to detect
1168          * conflicting writes by other clients.  A strict interpretation
1169          * of close-to-open would allow us to continue caching even if
1170          * another writer had changed the file, but some applications
1171          * depend on tighter cache coherency when writing.
1172          */
1173         status = NFS_PROTO(data->inode)->write_done(task, data);
1174         if (status != 0)
1175                 return status;
1176         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1177
1178 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1179         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1180                 /* We tried a write call, but the server did not
1181                  * commit data to stable storage even though we
1182                  * requested it.
1183                  * Note: There is a known bug in Tru64 < 5.0 in which
1184                  *       the server reports NFS_DATA_SYNC, but performs
1185                  *       NFS_FILE_SYNC. We therefore implement this checking
1186                  *       as a dprintk() in order to avoid filling syslog.
1187                  */
1188                 static unsigned long    complain;
1189
1190                 if (time_before(complain, jiffies)) {
1191                         dprintk("NFS:       faulty NFS server %s:"
1192                                 " (committed = %d) != (stable = %d)\n",
1193                                 server->nfs_client->cl_hostname,
1194                                 resp->verf->committed, argp->stable);
1195                         complain = jiffies + 300 * HZ;
1196                 }
1197         }
1198 #endif
1199         /* Is this a short write? */
1200         if (task->tk_status >= 0 && resp->count < argp->count) {
1201                 static unsigned long    complain;
1202
1203                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1204
1205                 /* Has the server at least made some progress? */
1206                 if (resp->count != 0) {
1207                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1208                         if (resp->verf->committed != NFS_UNSTABLE) {
1209                                 /* Resend from where the server left off */
1210                                 argp->offset += resp->count;
1211                                 argp->pgbase += resp->count;
1212                                 argp->count -= resp->count;
1213                         } else {
1214                                 /* Resend as a stable write in order to avoid
1215                                  * headaches in the case of a server crash.
1216                                  */
1217                                 argp->stable = NFS_FILE_SYNC;
1218                         }
1219                         nfs4_restart_rpc(task, server->nfs_client);
1220                         return -EAGAIN;
1221                 }
1222                 if (time_before(complain, jiffies)) {
1223                         printk(KERN_WARNING
1224                                "NFS: Server wrote zero bytes, expected %u.\n",
1225                                         argp->count);
1226                         complain = jiffies + 300 * HZ;
1227                 }
1228                 /* Can't do anything about it except throw an error. */
1229                 task->tk_status = -EIO;
1230         }
1231         nfs4_sequence_free_slot(server->nfs_client, &data->res.seq_res);
1232         return 0;
1233 }
1234
1235
1236 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1237 void nfs_commitdata_release(void *data)
1238 {
1239         struct nfs_write_data *wdata = data;
1240
1241         put_nfs_open_context(wdata->args.context);
1242         nfs_commit_free(wdata);
1243 }
1244
1245 /*
1246  * Set up the argument/result storage required for the RPC call.
1247  */
1248 static int nfs_commit_rpcsetup(struct list_head *head,
1249                 struct nfs_write_data *data,
1250                 int how)
1251 {
1252         struct nfs_page *first = nfs_list_entry(head->next);
1253         struct inode *inode = first->wb_context->path.dentry->d_inode;
1254         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1255         int priority = flush_task_priority(how);
1256         struct rpc_task *task;
1257         struct rpc_message msg = {
1258                 .rpc_argp = &data->args,
1259                 .rpc_resp = &data->res,
1260                 .rpc_cred = first->wb_context->cred,
1261         };
1262         struct rpc_task_setup task_setup_data = {
1263                 .task = &data->task,
1264                 .rpc_client = NFS_CLIENT(inode),
1265                 .rpc_message = &msg,
1266                 .callback_ops = &nfs_commit_ops,
1267                 .callback_data = data,
1268                 .workqueue = nfsiod_workqueue,
1269                 .flags = flags,
1270                 .priority = priority,
1271         };
1272
1273         /* Set up the RPC argument and reply structs
1274          * NB: take care not to mess about with data->commit et al. */
1275
1276         list_splice_init(head, &data->pages);
1277
1278         data->inode       = inode;
1279         data->cred        = msg.rpc_cred;
1280
1281         data->args.fh     = NFS_FH(data->inode);
1282         /* Note: we always request a commit of the entire inode */
1283         data->args.offset = 0;
1284         data->args.count  = 0;
1285         data->args.context = get_nfs_open_context(first->wb_context);
1286         data->res.count   = 0;
1287         data->res.fattr   = &data->fattr;
1288         data->res.verf    = &data->verf;
1289         nfs_fattr_init(&data->fattr);
1290
1291         /* Set up the initial task struct.  */
1292         NFS_PROTO(inode)->commit_setup(data, &msg);
1293
1294         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1295
1296         task = rpc_run_task(&task_setup_data);
1297         if (IS_ERR(task))
1298                 return PTR_ERR(task);
1299         rpc_put_task(task);
1300         return 0;
1301 }
1302
1303 /*
1304  * Commit dirty pages
1305  */
1306 static int
1307 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1308 {
1309         struct nfs_write_data   *data;
1310         struct nfs_page         *req;
1311
1312         data = nfs_commitdata_alloc();
1313
1314         if (!data)
1315                 goto out_bad;
1316
1317         /* Set up the argument struct */
1318         return nfs_commit_rpcsetup(head, data, how);
1319  out_bad:
1320         while (!list_empty(head)) {
1321                 req = nfs_list_entry(head->next);
1322                 nfs_list_remove_request(req);
1323                 nfs_mark_request_commit(req);
1324                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1325                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1326                                 BDI_RECLAIMABLE);
1327                 nfs_clear_page_tag_locked(req);
1328         }
1329         return -ENOMEM;
1330 }
1331
1332 /*
1333  * COMMIT call returned
1334  */
1335 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1336 {
1337         struct nfs_write_data   *data = calldata;
1338
1339         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1340                                 task->tk_pid, task->tk_status);
1341
1342         /* Call the NFS version-specific code */
1343         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1344                 return;
1345 }
1346
1347 static void nfs_commit_release(void *calldata)
1348 {
1349         struct nfs_write_data   *data = calldata;
1350         struct nfs_page         *req;
1351         int status = data->task.tk_status;
1352
1353         while (!list_empty(&data->pages)) {
1354                 req = nfs_list_entry(data->pages.next);
1355                 nfs_list_remove_request(req);
1356                 nfs_clear_request_commit(req);
1357
1358                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1359                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1360                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1361                         req->wb_bytes,
1362                         (long long)req_offset(req));
1363                 if (status < 0) {
1364                         nfs_context_set_write_error(req->wb_context, status);
1365                         nfs_inode_remove_request(req);
1366                         dprintk(", error = %d\n", status);
1367                         goto next;
1368                 }
1369
1370                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1371                  * returned by the server against all stored verfs. */
1372                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1373                         /* We have a match */
1374                         nfs_inode_remove_request(req);
1375                         dprintk(" OK\n");
1376                         goto next;
1377                 }
1378                 /* We have a mismatch. Write the page again */
1379                 dprintk(" mismatch\n");
1380                 nfs_mark_request_dirty(req);
1381         next:
1382                 nfs_clear_page_tag_locked(req);
1383         }
1384         nfs_commitdata_release(calldata);
1385 }
1386
1387 static const struct rpc_call_ops nfs_commit_ops = {
1388 #if defined(CONFIG_NFS_V4_1)
1389         .rpc_call_prepare = nfs_write_prepare,
1390 #endif /* CONFIG_NFS_V4_1 */
1391         .rpc_call_done = nfs_commit_done,
1392         .rpc_release = nfs_commit_release,
1393 };
1394
1395 int nfs_commit_inode(struct inode *inode, int how)
1396 {
1397         LIST_HEAD(head);
1398         int res;
1399
1400         spin_lock(&inode->i_lock);
1401         res = nfs_scan_commit(inode, &head, 0, 0);
1402         spin_unlock(&inode->i_lock);
1403         if (res) {
1404                 int error = nfs_commit_list(inode, &head, how);
1405                 if (error < 0)
1406                         return error;
1407         }
1408         return res;
1409 }
1410 #else
1411 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1412 {
1413         return 0;
1414 }
1415 #endif
1416
1417 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1418 {
1419         struct inode *inode = mapping->host;
1420         pgoff_t idx_start, idx_end;
1421         unsigned int npages = 0;
1422         LIST_HEAD(head);
1423         int nocommit = how & FLUSH_NOCOMMIT;
1424         long pages, ret;
1425
1426         /* FIXME */
1427         if (wbc->range_cyclic)
1428                 idx_start = 0;
1429         else {
1430                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1431                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1432                 if (idx_end > idx_start) {
1433                         pgoff_t l_npages = 1 + idx_end - idx_start;
1434                         npages = l_npages;
1435                         if (sizeof(npages) != sizeof(l_npages) &&
1436                                         (pgoff_t)npages != l_npages)
1437                                 npages = 0;
1438                 }
1439         }
1440         how &= ~FLUSH_NOCOMMIT;
1441         spin_lock(&inode->i_lock);
1442         do {
1443                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1444                 if (ret != 0)
1445                         continue;
1446                 if (nocommit)
1447                         break;
1448                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1449                 if (pages == 0)
1450                         break;
1451                 if (how & FLUSH_INVALIDATE) {
1452                         spin_unlock(&inode->i_lock);
1453                         nfs_cancel_commit_list(&head);
1454                         ret = pages;
1455                         spin_lock(&inode->i_lock);
1456                         continue;
1457                 }
1458                 pages += nfs_scan_commit(inode, &head, 0, 0);
1459                 spin_unlock(&inode->i_lock);
1460                 ret = nfs_commit_list(inode, &head, how);
1461                 spin_lock(&inode->i_lock);
1462
1463         } while (ret >= 0);
1464         spin_unlock(&inode->i_lock);
1465         return ret;
1466 }
1467
1468 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1469 {
1470         int ret;
1471
1472         ret = nfs_writepages(mapping, wbc);
1473         if (ret < 0)
1474                 goto out;
1475         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1476         if (ret < 0)
1477                 goto out;
1478         return 0;
1479 out:
1480         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1481         return ret;
1482 }
1483
1484 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1485 static int nfs_write_mapping(struct address_space *mapping, int how)
1486 {
1487         struct writeback_control wbc = {
1488                 .bdi = mapping->backing_dev_info,
1489                 .sync_mode = WB_SYNC_ALL,
1490                 .nr_to_write = LONG_MAX,
1491                 .range_start = 0,
1492                 .range_end = LLONG_MAX,
1493                 .for_writepages = 1,
1494         };
1495
1496         return __nfs_write_mapping(mapping, &wbc, how);
1497 }
1498
1499 /*
1500  * flush the inode to disk.
1501  */
1502 int nfs_wb_all(struct inode *inode)
1503 {
1504         return nfs_write_mapping(inode->i_mapping, 0);
1505 }
1506
1507 int nfs_wb_nocommit(struct inode *inode)
1508 {
1509         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1510 }
1511
1512 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1513 {
1514         struct nfs_page *req;
1515         loff_t range_start = page_offset(page);
1516         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1517         struct writeback_control wbc = {
1518                 .bdi = page->mapping->backing_dev_info,
1519                 .sync_mode = WB_SYNC_ALL,
1520                 .nr_to_write = LONG_MAX,
1521                 .range_start = range_start,
1522                 .range_end = range_end,
1523         };
1524         int ret = 0;
1525
1526         BUG_ON(!PageLocked(page));
1527         for (;;) {
1528                 req = nfs_page_find_request(page);
1529                 if (req == NULL)
1530                         goto out;
1531                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1532                         nfs_release_request(req);
1533                         break;
1534                 }
1535                 if (nfs_lock_request_dontget(req)) {
1536                         nfs_inode_remove_request(req);
1537                         /*
1538                          * In case nfs_inode_remove_request has marked the
1539                          * page as being dirty
1540                          */
1541                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1542                         nfs_unlock_request(req);
1543                         break;
1544                 }
1545                 ret = nfs_wait_on_request(req);
1546                 if (ret < 0)
1547                         goto out;
1548         }
1549         if (!PagePrivate(page))
1550                 return 0;
1551         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1552 out:
1553         return ret;
1554 }
1555
1556 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1557                                 int how)
1558 {
1559         loff_t range_start = page_offset(page);
1560         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1561         struct writeback_control wbc = {
1562                 .bdi = page->mapping->backing_dev_info,
1563                 .sync_mode = WB_SYNC_ALL,
1564                 .nr_to_write = LONG_MAX,
1565                 .range_start = range_start,
1566                 .range_end = range_end,
1567         };
1568         int ret;
1569
1570         do {
1571                 if (clear_page_dirty_for_io(page)) {
1572                         ret = nfs_writepage_locked(page, &wbc);
1573                         if (ret < 0)
1574                                 goto out_error;
1575                 } else if (!PagePrivate(page))
1576                         break;
1577                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1578                 if (ret < 0)
1579                         goto out_error;
1580         } while (PagePrivate(page));
1581         return 0;
1582 out_error:
1583         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1584         return ret;
1585 }
1586
1587 /*
1588  * Write back all requests on one page - we do this before reading it.
1589  */
1590 int nfs_wb_page(struct inode *inode, struct page* page)
1591 {
1592         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1593 }
1594
1595 #ifdef CONFIG_MIGRATION
1596 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1597                 struct page *page)
1598 {
1599         struct nfs_page *req;
1600         int ret;
1601
1602         if (PageFsCache(page))
1603                 nfs_fscache_release_page(page, GFP_KERNEL);
1604
1605         req = nfs_find_and_lock_request(page);
1606         ret = PTR_ERR(req);
1607         if (IS_ERR(req))
1608                 goto out;
1609
1610         ret = migrate_page(mapping, newpage, page);
1611         if (!req)
1612                 goto out;
1613         if (ret)
1614                 goto out_unlock;
1615         page_cache_get(newpage);
1616         req->wb_page = newpage;
1617         SetPagePrivate(newpage);
1618         set_page_private(newpage, page_private(page));
1619         ClearPagePrivate(page);
1620         set_page_private(page, 0);
1621         page_cache_release(page);
1622 out_unlock:
1623         nfs_clear_page_tag_locked(req);
1624         nfs_release_request(req);
1625 out:
1626         return ret;
1627 }
1628 #endif
1629
1630 int __init nfs_init_writepagecache(void)
1631 {
1632         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1633                                              sizeof(struct nfs_write_data),
1634                                              0, SLAB_HWCACHE_ALIGN,
1635                                              NULL);
1636         if (nfs_wdata_cachep == NULL)
1637                 return -ENOMEM;
1638
1639         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1640                                                      nfs_wdata_cachep);
1641         if (nfs_wdata_mempool == NULL)
1642                 return -ENOMEM;
1643
1644         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1645                                                       nfs_wdata_cachep);
1646         if (nfs_commit_mempool == NULL)
1647                 return -ENOMEM;
1648
1649         /*
1650          * NFS congestion size, scale with available memory.
1651          *
1652          *  64MB:    8192k
1653          * 128MB:   11585k
1654          * 256MB:   16384k
1655          * 512MB:   23170k
1656          *   1GB:   32768k
1657          *   2GB:   46340k
1658          *   4GB:   65536k
1659          *   8GB:   92681k
1660          *  16GB:  131072k
1661          *
1662          * This allows larger machines to have larger/more transfers.
1663          * Limit the default to 256M
1664          */
1665         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1666         if (nfs_congestion_kb > 256*1024)
1667                 nfs_congestion_kb = 256*1024;
1668
1669         return 0;
1670 }
1671
1672 void nfs_destroy_writepagecache(void)
1673 {
1674         mempool_destroy(nfs_commit_mempool);
1675         mempool_destroy(nfs_wdata_mempool);
1676         kmem_cache_destroy(nfs_wdata_cachep);
1677 }
1678