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