SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #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 || wbc->for_background)
 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         nfsi->ncommit++;
 442         spin_unlock(&inode->i_lock);
 443         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 444         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 445         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 446 }
 447
 448 static int
 449 nfs_clear_request_commit(struct nfs_page *req)
 450 {
 451         struct page *page = req->wb_page;
 452
 453         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 454                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 455                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 456                 return 1;
 457         }
 458         return 0;
 459 }
 460
 461 static inline
 462 int nfs_write_need_commit(struct nfs_write_data *data)
 463 {
 464         return data->verf.committed != NFS_FILE_SYNC;
 465 }
 466
 467 static inline
 468 int nfs_reschedule_unstable_write(struct nfs_page *req)
 469 {
 470         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 471                 nfs_mark_request_commit(req);
 472                 return 1;
 473         }
 474         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 475                 nfs_mark_request_dirty(req);
 476                 return 1;
 477         }
 478         return 0;
 479 }
 480 #else
 481 static inline void
 482 nfs_mark_request_commit(struct nfs_page *req)
 483 {
 484 }
 485
 486 static inline int
 487 nfs_clear_request_commit(struct nfs_page *req)
 488 {
 489         return 0;
 490 }
 491
 492 static inline
 493 int nfs_write_need_commit(struct nfs_write_data *data)
 494 {
 495         return 0;
 496 }
 497
 498 static inline
 499 int nfs_reschedule_unstable_write(struct nfs_page *req)
 500 {
 501         return 0;
 502 }
 503 #endif
 504
 505 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 506 static int
 507 nfs_need_commit(struct nfs_inode *nfsi)
 508 {
 509         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 510 }
 511
 512 /*
 513  * nfs_scan_commit - Scan an inode for commit requests
 514  * @inode: NFS inode to scan
 515  * @dst: destination list
 516  * @idx_start: lower bound of page->index to scan.
 517  * @npages: idx_start + npages sets the upper bound to scan.
 518  *
 519  * Moves requests from the inode's 'commit' request list.
 520  * The requests are *not* checked to ensure that they form a contiguous set.
 521  */
 522 static int
 523 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 524 {
 525         struct nfs_inode *nfsi = NFS_I(inode);
 526         int ret;
 527
 528         if (!nfs_need_commit(nfsi))
 529                 return 0;
 530
 531         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 532         if (ret > 0)
 533                 nfsi->ncommit -= ret;
 534         if (nfs_need_commit(NFS_I(inode)))
 535                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 536         return ret;
 537 }
 538 #else
 539 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 540 {
 541         return 0;
 542 }
 543
 544 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 545 {
 546         return 0;
 547 }
 548 #endif
 549
 550 /*
 551  * Search for an existing write request, and attempt to update
 552  * it to reflect a new dirty region on a given page.
 553  *
 554  * If the attempt fails, then the existing request is flushed out
 555  * to disk.
 556  */
 557 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 558                 struct page *page,
 559                 unsigned int offset,
 560                 unsigned int bytes)
 561 {
 562         struct nfs_page *req;
 563         unsigned int rqend;
 564         unsigned int end;
 565         int error;
 566
 567         if (!PagePrivate(page))
 568                 return NULL;
 569
 570         end = offset + bytes;
 571         spin_lock(&inode->i_lock);
 572
 573         for (;;) {
 574                 req = nfs_page_find_request_locked(page);
 575                 if (req == NULL)
 576                         goto out_unlock;
 577
 578                 rqend = req->wb_offset + req->wb_bytes;
 579                 /*
 580                  * Tell the caller to flush out the request if
 581                  * the offsets are non-contiguous.
 582                  * Note: nfs_flush_incompatible() will already
 583                  * have flushed out requests having wrong owners.
 584                  */
 585                 if (offset > rqend
 586                     || end < req->wb_offset)
 587                         goto out_flushme;
 588
 589                 if (nfs_set_page_tag_locked(req))
 590                         break;
 591
 592                 /* The request is locked, so wait and then retry */
 593                 spin_unlock(&inode->i_lock);
 594                 error = nfs_wait_on_request(req);
 595                 nfs_release_request(req);
 596                 if (error != 0)
 597                         goto out_err;
 598                 spin_lock(&inode->i_lock);
 599         }
 600
 601         if (nfs_clear_request_commit(req) &&
 602                         radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 603                                 req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL)
 604                 NFS_I(inode)->ncommit--;
 605
 606         /* Okay, the request matches. Update the region */
 607         if (offset < req->wb_offset) {
 608                 req->wb_offset = offset;
 609                 req->wb_pgbase = offset;
 610         }
 611         if (end > rqend)
 612                 req->wb_bytes = end - req->wb_offset;
 613         else
 614                 req->wb_bytes = rqend - req->wb_offset;
 615 out_unlock:
 616         spin_unlock(&inode->i_lock);
 617         return req;
 618 out_flushme:
 619         spin_unlock(&inode->i_lock);
 620         nfs_release_request(req);
 621         error = nfs_wb_page(inode, page);
 622 out_err:
 623         return ERR_PTR(error);
 624 }
 625
 626 /*
 627  * Try to update an existing write request, or create one if there is none.
 628  *
 629  * Note: Should always be called with the Page Lock held to prevent races
 630  * if we have to add a new request. Also assumes that the caller has
 631  * already called nfs_flush_incompatible() if necessary.
 632  */
 633 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 634                 struct page *page, unsigned int offset, unsigned int bytes)
 635 {
 636         struct inode *inode = page->mapping->host;
 637         struct nfs_page *req;
 638         int error;
 639
 640         req = nfs_try_to_update_request(inode, page, offset, bytes);
 641         if (req != NULL)
 642                 goto out;
 643         req = nfs_create_request(ctx, inode, page, offset, bytes);
 644         if (IS_ERR(req))
 645                 goto out;
 646         error = nfs_inode_add_request(inode, req);
 647         if (error != 0) {
 648                 nfs_release_request(req);
 649                 req = ERR_PTR(error);
 650         }
 651 out:
 652         return req;
 653 }
 654
 655 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 656                 unsigned int offset, unsigned int count)
 657 {
 658         struct nfs_page *req;
 659
 660         req = nfs_setup_write_request(ctx, page, offset, count);
 661         if (IS_ERR(req))
 662                 return PTR_ERR(req);
 663         /* Update file length */
 664         nfs_grow_file(page, offset, count);
 665         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 666         nfs_clear_page_tag_locked(req);
 667         return 0;
 668 }
 669
 670 int nfs_flush_incompatible(struct file *file, struct page *page)
 671 {
 672         struct nfs_open_context *ctx = nfs_file_open_context(file);
 673         struct nfs_page *req;
 674         int do_flush, status;
 675         /*
 676          * Look for a request corresponding to this page. If there
 677          * is one, and it belongs to another file, we flush it out
 678          * before we try to copy anything into the page. Do this
 679          * due to the lack of an ACCESS-type call in NFSv2.
 680          * Also do the same if we find a request from an existing
 681          * dropped page.
 682          */
 683         do {
 684                 req = nfs_page_find_request(page);
 685                 if (req == NULL)
 686                         return 0;
 687                 do_flush = req->wb_page != page || req->wb_context != ctx;
 688                 nfs_release_request(req);
 689                 if (!do_flush)
 690                         return 0;
 691                 status = nfs_wb_page(page->mapping->host, page);
 692         } while (status == 0);
 693         return status;
 694 }
 695
 696 /*
 697  * If the page cache is marked as unsafe or invalid, then we can't rely on
 698  * the PageUptodate() flag. In this case, we will need to turn off
 699  * write optimisations that depend on the page contents being correct.
 700  */
 701 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 702 {
 703         return PageUptodate(page) &&
 704                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 705 }
 706
 707 /*
 708  * Update and possibly write a cached page of an NFS file.
 709  *
 710  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 711  * things with a page scheduled for an RPC call (e.g. invalidate it).
 712  */
 713 int nfs_updatepage(struct file *file, struct page *page,
 714                 unsigned int offset, unsigned int count)
 715 {
 716         struct nfs_open_context *ctx = nfs_file_open_context(file);
 717         struct inode    *inode = page->mapping->host;
 718         int             status = 0;
 719
 720         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 721
 722         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 723                 file->f_path.dentry->d_parent->d_name.name,
 724                 file->f_path.dentry->d_name.name, count,
 725                 (long long)(page_offset(page) + offset));
 726
 727         /* If we're not using byte range locks, and we know the page
 728          * is up to date, it may be more efficient to extend the write
 729          * to cover the entire page in order to avoid fragmentation
 730          * inefficiencies.
 731          */
 732         if (nfs_write_pageuptodate(page, inode) &&
 733                         inode->i_flock == NULL &&
 734                         !(file->f_flags & O_DSYNC)) {
 735                 count = max(count + offset, nfs_page_length(page));
 736                 offset = 0;
 737         }
 738
 739         status = nfs_writepage_setup(ctx, page, offset, count);
 740         if (status < 0)
 741                 nfs_set_pageerror(page);
 742         else
 743                 __set_page_dirty_nobuffers(page);
 744
 745         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 746                         status, (long long)i_size_read(inode));
 747         return status;
 748 }
 749
 750 static void nfs_writepage_release(struct nfs_page *req)
 751 {
 752
 753         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 754                 nfs_end_page_writeback(req->wb_page);
 755                 nfs_inode_remove_request(req);
 756         } else
 757                 nfs_end_page_writeback(req->wb_page);
 758         nfs_clear_page_tag_locked(req);
 759 }
 760
 761 static int flush_task_priority(int how)
 762 {
 763         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 764                 case FLUSH_HIGHPRI:
 765                         return RPC_PRIORITY_HIGH;
 766                 case FLUSH_LOWPRI:
 767                         return RPC_PRIORITY_LOW;
 768         }
 769         return RPC_PRIORITY_NORMAL;
 770 }
 771
 772 /*
 773  * Set up the argument/result storage required for the RPC call.
 774  */
 775 static int nfs_write_rpcsetup(struct nfs_page *req,
 776                 struct nfs_write_data *data,
 777                 const struct rpc_call_ops *call_ops,
 778                 unsigned int count, unsigned int offset,
 779                 int how)
 780 {
 781         struct inode *inode = req->wb_context->path.dentry->d_inode;
 782         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 783         int priority = flush_task_priority(how);
 784         struct rpc_task *task;
 785         struct rpc_message msg = {
 786                 .rpc_argp = &data->args,
 787                 .rpc_resp = &data->res,
 788                 .rpc_cred = req->wb_context->cred,
 789         };
 790         struct rpc_task_setup task_setup_data = {
 791                 .rpc_client = NFS_CLIENT(inode),
 792                 .task = &data->task,
 793                 .rpc_message = &msg,
 794                 .callback_ops = call_ops,
 795                 .callback_data = data,
 796                 .workqueue = nfsiod_workqueue,
 797                 .flags = flags,
 798                 .priority = priority,
 799         };
 800
 801         /* Set up the RPC argument and reply structs
 802          * NB: take care not to mess about with data->commit et al. */
 803
 804         data->req = req;
 805         data->inode = inode = req->wb_context->path.dentry->d_inode;
 806         data->cred = msg.rpc_cred;
 807
 808         data->args.fh     = NFS_FH(inode);
 809         data->args.offset = req_offset(req) + offset;
 810         data->args.pgbase = req->wb_pgbase + offset;
 811         data->args.pages  = data->pagevec;
 812         data->args.count  = count;
 813         data->args.context = get_nfs_open_context(req->wb_context);
 814         data->args.stable  = NFS_UNSTABLE;
 815         if (how & FLUSH_STABLE) {
 816                 data->args.stable = NFS_DATA_SYNC;
 817                 if (!nfs_need_commit(NFS_I(inode)))
 818                         data->args.stable = NFS_FILE_SYNC;
 819         }
 820
 821         data->res.fattr   = &data->fattr;
 822         data->res.count   = count;
 823         data->res.verf    = &data->verf;
 824         nfs_fattr_init(&data->fattr);
 825
 826         /* Set up the initial task struct.  */
 827         NFS_PROTO(inode)->write_setup(data, &msg);
 828
 829         dprintk("NFS: %5u initiated write call "
 830                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 831                 data->task.tk_pid,
 832                 inode->i_sb->s_id,
 833                 (long long)NFS_FILEID(inode),
 834                 count,
 835                 (unsigned long long)data->args.offset);
 836
 837         task = rpc_run_task(&task_setup_data);
 838         if (IS_ERR(task))
 839                 return PTR_ERR(task);
 840         rpc_put_task(task);
 841         return 0;
 842 }
 843
 844 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 845  * call this on each, which will prepare them to be retried on next
 846  * writeback using standard nfs.
 847  */
 848 static void nfs_redirty_request(struct nfs_page *req)
 849 {
 850         nfs_mark_request_dirty(req);
 851         nfs_end_page_writeback(req->wb_page);
 852         nfs_clear_page_tag_locked(req);
 853 }
 854
 855 /*
 856  * Generate multiple small requests to write out a single
 857  * contiguous dirty area on one page.
 858  */
 859 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 860 {
 861         struct nfs_page *req = nfs_list_entry(head->next);
 862         struct page *page = req->wb_page;
 863         struct nfs_write_data *data;
 864         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 865         unsigned int offset;
 866         int requests = 0;
 867         int ret = 0;
 868         LIST_HEAD(list);
 869
 870         nfs_list_remove_request(req);
 871
 872         nbytes = count;
 873         do {
 874                 size_t len = min(nbytes, wsize);
 875
 876                 data = nfs_writedata_alloc(1);
 877                 if (!data)
 878                         goto out_bad;
 879                 list_add(&data->pages, &list);
 880                 requests++;
 881                 nbytes -= len;
 882         } while (nbytes != 0);
 883         atomic_set(&req->wb_complete, requests);
 884
 885         ClearPageError(page);
 886         offset = 0;
 887         nbytes = count;
 888         do {
 889                 int ret2;
 890
 891                 data = list_entry(list.next, struct nfs_write_data, pages);
 892                 list_del_init(&data->pages);
 893
 894                 data->pagevec[0] = page;
 895
 896                 if (nbytes < wsize)
 897                         wsize = nbytes;
 898                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 899                                    wsize, offset, how);
 900                 if (ret == 0)
 901                         ret = ret2;
 902                 offset += wsize;
 903                 nbytes -= wsize;
 904         } while (nbytes != 0);
 905
 906         return ret;
 907
 908 out_bad:
 909         while (!list_empty(&list)) {
 910                 data = list_entry(list.next, struct nfs_write_data, pages);
 911                 list_del(&data->pages);
 912                 nfs_writedata_release(data);
 913         }
 914         nfs_redirty_request(req);
 915         return -ENOMEM;
 916 }
 917
 918 /*
 919  * Create an RPC task for the given write request and kick it.
 920  * The page must have been locked by the caller.
 921  *
 922  * It may happen that the page we're passed is not marked dirty.
 923  * This is the case if nfs_updatepage detects a conflicting request
 924  * that has been written but not committed.
 925  */
 926 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 927 {
 928         struct nfs_page         *req;
 929         struct page             **pages;
 930         struct nfs_write_data   *data;
 931
 932         data = nfs_writedata_alloc(npages);
 933         if (!data)
 934                 goto out_bad;
 935
 936         pages = data->pagevec;
 937         while (!list_empty(head)) {
 938                 req = nfs_list_entry(head->next);
 939                 nfs_list_remove_request(req);
 940                 nfs_list_add_request(req, &data->pages);
 941                 ClearPageError(req->wb_page);
 942                 *pages++ = req->wb_page;
 943         }
 944         req = nfs_list_entry(data->pages.next);
 945
 946         /* Set up the argument struct */
 947         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 948  out_bad:
 949         while (!list_empty(head)) {
 950                 req = nfs_list_entry(head->next);
 951                 nfs_list_remove_request(req);
 952                 nfs_redirty_request(req);
 953         }
 954         return -ENOMEM;
 955 }
 956
 957 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 958                                   struct inode *inode, int ioflags)
 959 {
 960         size_t wsize = NFS_SERVER(inode)->wsize;
 961
 962         if (wsize < PAGE_CACHE_SIZE)
 963                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 964         else
 965                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 966 }
 967
 968 /*
 969  * Handle a write reply that flushed part of a page.
 970  */
 971 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 972 {
 973         struct nfs_write_data   *data = calldata;
 974
 975         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
 976                 task->tk_pid,
 977                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
 978                 (long long)
 979                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
 980                 data->req->wb_bytes, (long long)req_offset(data->req));
 981
 982         nfs_writeback_done(task, data);
 983 }
 984
 985 static void nfs_writeback_release_partial(void *calldata)
 986 {
 987         struct nfs_write_data   *data = calldata;
 988         struct nfs_page         *req = data->req;
 989         struct page             *page = req->wb_page;
 990         int status = data->task.tk_status;
 991
 992         if (status < 0) {
 993                 nfs_set_pageerror(page);
 994                 nfs_context_set_write_error(req->wb_context, status);
 995                 dprintk(", error = %d\n", status);
 996                 goto out;
 997         }
 998
 999         if (nfs_write_need_commit(data)) {
1000                 struct inode *inode = page->mapping->host;
1001
1002                 spin_lock(&inode->i_lock);
1003                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1004                         /* Do nothing we need to resend the writes */
1005                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1006                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1007                         dprintk(" defer commit\n");
1008                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1009                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1010                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1011                         dprintk(" server reboot detected\n");
1012                 }
1013                 spin_unlock(&inode->i_lock);
1014         } else
1015                 dprintk(" OK\n");
1016
1017 out:
1018         if (atomic_dec_and_test(&req->wb_complete))
1019                 nfs_writepage_release(req);
1020         nfs_writedata_release(calldata);
1021 }
1022
1023 #if defined(CONFIG_NFS_V4_1)
1024 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1025 {
1026         struct nfs_write_data *data = calldata;
1027         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1028
1029         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1030                                 &data->res.seq_res, 1, task))
1031                 return;
1032         rpc_call_start(task);
1033 }
1034 #endif /* CONFIG_NFS_V4_1 */
1035
1036 static const struct rpc_call_ops nfs_write_partial_ops = {
1037 #if defined(CONFIG_NFS_V4_1)
1038         .rpc_call_prepare = nfs_write_prepare,
1039 #endif /* CONFIG_NFS_V4_1 */
1040         .rpc_call_done = nfs_writeback_done_partial,
1041         .rpc_release = nfs_writeback_release_partial,
1042 };
1043
1044 /*
1045  * Handle a write reply that flushes a whole page.
1046  *
1047  * FIXME: There is an inherent race with invalidate_inode_pages and
1048  *        writebacks since the page->count is kept > 1 for as long
1049  *        as the page has a write request pending.
1050  */
1051 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1052 {
1053         struct nfs_write_data   *data = calldata;
1054
1055         nfs_writeback_done(task, data);
1056 }
1057
1058 static void nfs_writeback_release_full(void *calldata)
1059 {
1060         struct nfs_write_data   *data = calldata;
1061         int status = data->task.tk_status;
1062
1063         /* Update attributes as result of writeback. */
1064         while (!list_empty(&data->pages)) {
1065                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1066                 struct page *page = req->wb_page;
1067
1068                 nfs_list_remove_request(req);
1069
1070                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1071                         data->task.tk_pid,
1072                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1073                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1074                         req->wb_bytes,
1075                         (long long)req_offset(req));
1076
1077                 if (status < 0) {
1078                         nfs_set_pageerror(page);
1079                         nfs_context_set_write_error(req->wb_context, status);
1080                         dprintk(", error = %d\n", status);
1081                         goto remove_request;
1082                 }
1083
1084                 if (nfs_write_need_commit(data)) {
1085                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1086                         nfs_mark_request_commit(req);
1087                         nfs_end_page_writeback(page);
1088                         dprintk(" marked for commit\n");
1089                         goto next;
1090                 }
1091                 dprintk(" OK\n");
1092 remove_request:
1093                 nfs_end_page_writeback(page);
1094                 nfs_inode_remove_request(req);
1095         next:
1096                 nfs_clear_page_tag_locked(req);
1097         }
1098         nfs_writedata_release(calldata);
1099 }
1100
1101 static const struct rpc_call_ops nfs_write_full_ops = {
1102 #if defined(CONFIG_NFS_V4_1)
1103         .rpc_call_prepare = nfs_write_prepare,
1104 #endif /* CONFIG_NFS_V4_1 */
1105         .rpc_call_done = nfs_writeback_done_full,
1106         .rpc_release = nfs_writeback_release_full,
1107 };
1108
1109
1110 /*
1111  * This function is called when the WRITE call is complete.
1112  */
1113 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1114 {
1115         struct nfs_writeargs    *argp = &data->args;
1116         struct nfs_writeres     *resp = &data->res;
1117         struct nfs_server       *server = NFS_SERVER(data->inode);
1118         int status;
1119
1120         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1121                 task->tk_pid, task->tk_status);
1122
1123         /*
1124          * ->write_done will attempt to use post-op attributes to detect
1125          * conflicting writes by other clients.  A strict interpretation
1126          * of close-to-open would allow us to continue caching even if
1127          * another writer had changed the file, but some applications
1128          * depend on tighter cache coherency when writing.
1129          */
1130         status = NFS_PROTO(data->inode)->write_done(task, data);
1131         if (status != 0)
1132                 return status;
1133         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1134
1135 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1136         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1137                 /* We tried a write call, but the server did not
1138                  * commit data to stable storage even though we
1139                  * requested it.
1140                  * Note: There is a known bug in Tru64 < 5.0 in which
1141                  *       the server reports NFS_DATA_SYNC, but performs
1142                  *       NFS_FILE_SYNC. We therefore implement this checking
1143                  *       as a dprintk() in order to avoid filling syslog.
1144                  */
1145                 static unsigned long    complain;
1146
1147                 if (time_before(complain, jiffies)) {
1148                         dprintk("NFS:       faulty NFS server %s:"
1149                                 " (committed = %d) != (stable = %d)\n",
1150                                 server->nfs_client->cl_hostname,
1151                                 resp->verf->committed, argp->stable);
1152                         complain = jiffies + 300 * HZ;
1153                 }
1154         }
1155 #endif
1156         /* Is this a short write? */
1157         if (task->tk_status >= 0 && resp->count < argp->count) {
1158                 static unsigned long    complain;
1159
1160                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1161
1162                 /* Has the server at least made some progress? */
1163                 if (resp->count != 0) {
1164                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1165                         if (resp->verf->committed != NFS_UNSTABLE) {
1166                                 /* Resend from where the server left off */
1167                                 argp->offset += resp->count;
1168                                 argp->pgbase += resp->count;
1169                                 argp->count -= resp->count;
1170                         } else {
1171                                 /* Resend as a stable write in order to avoid
1172                                  * headaches in the case of a server crash.
1173                                  */
1174                                 argp->stable = NFS_FILE_SYNC;
1175                         }
1176                         nfs_restart_rpc(task, server->nfs_client);
1177                         return -EAGAIN;
1178                 }
1179                 if (time_before(complain, jiffies)) {
1180                         printk(KERN_WARNING
1181                                "NFS: Server wrote zero bytes, expected %u.\n",
1182                                         argp->count);
1183                         complain = jiffies + 300 * HZ;
1184                 }
1185                 /* Can't do anything about it except throw an error. */
1186                 task->tk_status = -EIO;
1187         }
1188         return 0;
1189 }
1190
1191
1192 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1193 static void nfs_commitdata_release(void *data)
1194 {
1195         struct nfs_write_data *wdata = data;
1196
1197         put_nfs_open_context(wdata->args.context);
1198         nfs_commit_free(wdata);
1199 }
1200
1201 /*
1202  * Set up the argument/result storage required for the RPC call.
1203  */
1204 static int nfs_commit_rpcsetup(struct list_head *head,
1205                 struct nfs_write_data *data,
1206                 int how)
1207 {
1208         struct nfs_page *first = nfs_list_entry(head->next);
1209         struct inode *inode = first->wb_context->path.dentry->d_inode;
1210         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1211         int priority = flush_task_priority(how);
1212         struct rpc_task *task;
1213         struct rpc_message msg = {
1214                 .rpc_argp = &data->args,
1215                 .rpc_resp = &data->res,
1216                 .rpc_cred = first->wb_context->cred,
1217         };
1218         struct rpc_task_setup task_setup_data = {
1219                 .task = &data->task,
1220                 .rpc_client = NFS_CLIENT(inode),
1221                 .rpc_message = &msg,
1222                 .callback_ops = &nfs_commit_ops,
1223                 .callback_data = data,
1224                 .workqueue = nfsiod_workqueue,
1225                 .flags = flags,
1226                 .priority = priority,
1227         };
1228
1229         /* Set up the RPC argument and reply structs
1230          * NB: take care not to mess about with data->commit et al. */
1231
1232         list_splice_init(head, &data->pages);
1233
1234         data->inode       = inode;
1235         data->cred        = msg.rpc_cred;
1236
1237         data->args.fh     = NFS_FH(data->inode);
1238         /* Note: we always request a commit of the entire inode */
1239         data->args.offset = 0;
1240         data->args.count  = 0;
1241         data->args.context = get_nfs_open_context(first->wb_context);
1242         data->res.count   = 0;
1243         data->res.fattr   = &data->fattr;
1244         data->res.verf    = &data->verf;
1245         nfs_fattr_init(&data->fattr);
1246
1247         /* Set up the initial task struct.  */
1248         NFS_PROTO(inode)->commit_setup(data, &msg);
1249
1250         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1251
1252         task = rpc_run_task(&task_setup_data);
1253         if (IS_ERR(task))
1254                 return PTR_ERR(task);
1255         rpc_put_task(task);
1256         return 0;
1257 }
1258
1259 /*
1260  * Commit dirty pages
1261  */
1262 static int
1263 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1264 {
1265         struct nfs_write_data   *data;
1266         struct nfs_page         *req;
1267
1268         data = nfs_commitdata_alloc();
1269
1270         if (!data)
1271                 goto out_bad;
1272
1273         /* Set up the argument struct */
1274         return nfs_commit_rpcsetup(head, data, how);
1275  out_bad:
1276         while (!list_empty(head)) {
1277                 req = nfs_list_entry(head->next);
1278                 nfs_list_remove_request(req);
1279                 nfs_mark_request_commit(req);
1280                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1281                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1282                                 BDI_RECLAIMABLE);
1283                 nfs_clear_page_tag_locked(req);
1284         }
1285         return -ENOMEM;
1286 }
1287
1288 /*
1289  * COMMIT call returned
1290  */
1291 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1292 {
1293         struct nfs_write_data   *data = calldata;
1294
1295         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1296                                 task->tk_pid, task->tk_status);
1297
1298         /* Call the NFS version-specific code */
1299         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1300                 return;
1301 }
1302
1303 static void nfs_commit_release(void *calldata)
1304 {
1305         struct nfs_write_data   *data = calldata;
1306         struct nfs_page         *req;
1307         int status = data->task.tk_status;
1308
1309         while (!list_empty(&data->pages)) {
1310                 req = nfs_list_entry(data->pages.next);
1311                 nfs_list_remove_request(req);
1312                 nfs_clear_request_commit(req);
1313
1314                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1315                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1316                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1317                         req->wb_bytes,
1318                         (long long)req_offset(req));
1319                 if (status < 0) {
1320                         nfs_context_set_write_error(req->wb_context, status);
1321                         nfs_inode_remove_request(req);
1322                         dprintk(", error = %d\n", status);
1323                         goto next;
1324                 }
1325
1326                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1327                  * returned by the server against all stored verfs. */
1328                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1329                         /* We have a match */
1330                         nfs_inode_remove_request(req);
1331                         dprintk(" OK\n");
1332                         goto next;
1333                 }
1334                 /* We have a mismatch. Write the page again */
1335                 dprintk(" mismatch\n");
1336                 nfs_mark_request_dirty(req);
1337         next:
1338                 nfs_clear_page_tag_locked(req);
1339         }
1340         nfs_commitdata_release(calldata);
1341 }
1342
1343 static const struct rpc_call_ops nfs_commit_ops = {
1344 #if defined(CONFIG_NFS_V4_1)
1345         .rpc_call_prepare = nfs_write_prepare,
1346 #endif /* CONFIG_NFS_V4_1 */
1347         .rpc_call_done = nfs_commit_done,
1348         .rpc_release = nfs_commit_release,
1349 };
1350
1351 static int nfs_commit_inode(struct inode *inode, int how)
1352 {
1353         LIST_HEAD(head);
1354         int res;
1355
1356         spin_lock(&inode->i_lock);
1357         res = nfs_scan_commit(inode, &head, 0, 0);
1358         spin_unlock(&inode->i_lock);
1359         if (res) {
1360                 int error = nfs_commit_list(inode, &head, how);
1361                 if (error < 0)
1362                         return error;
1363         }
1364         return res;
1365 }
1366
1367 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1368 {
1369         struct nfs_inode *nfsi = NFS_I(inode);
1370         int flags = FLUSH_SYNC;
1371         int ret = 0;
1372
1373         /* Don't commit yet if this is a non-blocking flush and there are
1374          * lots of outstanding writes for this mapping.
1375          */
1376         if (wbc->sync_mode == WB_SYNC_NONE &&
1377             nfsi->ncommit <= (nfsi->npages >> 1))
1378                 goto out_mark_dirty;
1379
1380         if (wbc->nonblocking || wbc->for_background)
1381                 flags = 0;
1382         ret = nfs_commit_inode(inode, flags);
1383         if (ret >= 0) {
1384                 if (wbc->sync_mode == WB_SYNC_NONE) {
1385                         if (ret < wbc->nr_to_write)
1386                                 wbc->nr_to_write -= ret;
1387                         else
1388                                 wbc->nr_to_write = 0;
1389                 }
1390                 return 0;
1391         }
1392 out_mark_dirty:
1393         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1394         return ret;
1395 }
1396 #else
1397 static int nfs_commit_inode(struct inode *inode, int how)
1398 {
1399         return 0;
1400 }
1401
1402 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1403 {
1404         return 0;
1405 }
1406 #endif
1407
1408 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1409 {
1410         return nfs_commit_unstable_pages(inode, wbc);
1411 }
1412
1413 /*
1414  * flush the inode to disk.
1415  */
1416 int nfs_wb_all(struct inode *inode)
1417 {
1418         struct writeback_control wbc = {
1419                 .sync_mode = WB_SYNC_ALL,
1420                 .nr_to_write = LONG_MAX,
1421                 .range_start = 0,
1422                 .range_end = LLONG_MAX,
1423         };
1424
1425         return sync_inode(inode, &wbc);
1426 }
1427
1428 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1429 {
1430         struct nfs_page *req;
1431         int ret = 0;
1432
1433         BUG_ON(!PageLocked(page));
1434         for (;;) {
1435                 req = nfs_page_find_request(page);
1436                 if (req == NULL)
1437                         break;
1438                 if (nfs_lock_request_dontget(req)) {
1439                         nfs_inode_remove_request(req);
1440                         /*
1441                          * In case nfs_inode_remove_request has marked the
1442                          * page as being dirty
1443                          */
1444                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1445                         nfs_unlock_request(req);
1446                         break;
1447                 }
1448                 ret = nfs_wait_on_request(req);
1449                 nfs_release_request(req);
1450                 if (ret < 0)
1451                         break;
1452         }
1453         return ret;
1454 }
1455
1456 /*
1457  * Write back all requests on one page - we do this before reading it.
1458  */
1459 int nfs_wb_page(struct inode *inode, struct page *page)
1460 {
1461         loff_t range_start = page_offset(page);
1462         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1463         struct writeback_control wbc = {
1464                 .sync_mode = WB_SYNC_ALL,
1465                 .nr_to_write = 0,
1466                 .range_start = range_start,
1467                 .range_end = range_end,
1468         };
1469         struct nfs_page *req;
1470         int need_commit;
1471         int ret;
1472
1473         while(PagePrivate(page)) {
1474                 if (clear_page_dirty_for_io(page)) {
1475                         ret = nfs_writepage_locked(page, &wbc);
1476                         if (ret < 0)
1477                                 goto out_error;
1478                 }
1479                 req = nfs_find_and_lock_request(page);
1480                 if (!req)
1481                         break;
1482                 if (IS_ERR(req)) {
1483                         ret = PTR_ERR(req);
1484                         goto out_error;
1485                 }
1486                 need_commit = test_bit(PG_CLEAN, &req->wb_flags);
1487                 nfs_clear_page_tag_locked(req);
1488                 if (need_commit) {
1489                         ret = nfs_commit_inode(inode, FLUSH_SYNC);
1490                         if (ret < 0)
1491                                 goto out_error;
1492                 }
1493         }
1494         return 0;
1495 out_error:
1496         return ret;
1497 }
1498
1499 #ifdef CONFIG_MIGRATION
1500 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1501                 struct page *page)
1502 {
1503         struct nfs_page *req;
1504         int ret;
1505
1506         nfs_fscache_release_page(page, GFP_KERNEL);
1507
1508         req = nfs_find_and_lock_request(page);
1509         ret = PTR_ERR(req);
1510         if (IS_ERR(req))
1511                 goto out;
1512
1513         ret = migrate_page(mapping, newpage, page);
1514         if (!req)
1515                 goto out;
1516         if (ret)
1517                 goto out_unlock;
1518         page_cache_get(newpage);
1519         spin_lock(&mapping->host->i_lock);
1520         req->wb_page = newpage;
1521         SetPagePrivate(newpage);
1522         set_page_private(newpage, (unsigned long)req);
1523         ClearPagePrivate(page);
1524         set_page_private(page, 0);
1525         spin_unlock(&mapping->host->i_lock);
1526         page_cache_release(page);
1527 out_unlock:
1528         nfs_clear_page_tag_locked(req);
1529 out:
1530         return ret;
1531 }
1532 #endif
1533
1534 int __init nfs_init_writepagecache(void)
1535 {
1536         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1537                                              sizeof(struct nfs_write_data),
1538                                              0, SLAB_HWCACHE_ALIGN,
1539                                              NULL);
1540         if (nfs_wdata_cachep == NULL)
1541                 return -ENOMEM;
1542
1543         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1544                                                      nfs_wdata_cachep);
1545         if (nfs_wdata_mempool == NULL)
1546                 return -ENOMEM;
1547
1548         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1549                                                       nfs_wdata_cachep);
1550         if (nfs_commit_mempool == NULL)
1551                 return -ENOMEM;
1552
1553         /*
1554          * NFS congestion size, scale with available memory.
1555          *
1556          *  64MB:    8192k
1557          * 128MB:   11585k
1558          * 256MB:   16384k
1559          * 512MB:   23170k
1560          *   1GB:   32768k
1561          *   2GB:   46340k
1562          *   4GB:   65536k
1563          *   8GB:   92681k
1564          *  16GB:  131072k
1565          *
1566          * This allows larger machines to have larger/more transfers.
1567          * Limit the default to 256M
1568          */
1569         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1570         if (nfs_congestion_kb > 256*1024)
1571                 nfs_congestion_kb = 256*1024;
1572
1573         return 0;
1574 }
1575
1576 void nfs_destroy_writepagecache(void)
1577 {
1578         mempool_destroy(nfs_commit_mempool);
1579         mempool_destroy(nfs_wdata_mempool);
1580         kmem_cache_destroy(nfs_wdata_cachep);
1581 }
1582