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 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  35
  36 #define MIN_POOL_WRITE          (32)
  37 #define MIN_POOL_COMMIT         (4)
  38
  39 /*
  40  * Local function declarations
  41  */
  42 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  43                                   struct inode *inode, int ioflags);
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_write_partial_ops;
  46 static const struct rpc_call_ops nfs_write_full_ops;
  47 static const struct rpc_call_ops nfs_commit_ops;
  48
  49 static struct kmem_cache *nfs_wdata_cachep;
  50 static mempool_t *nfs_wdata_mempool;
  51 static mempool_t *nfs_commit_mempool;
  52
  53 struct nfs_write_data *nfs_commitdata_alloc(void)
  54 {
  55         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  56
  57         if (p) {
  58                 memset(p, 0, sizeof(*p));
  59                 INIT_LIST_HEAD(&p->pages);
  60         }
  61         return p;
  62 }
  63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  64
  65 void nfs_commit_free(struct nfs_write_data *p)
  66 {
  67         if (p && (p->pagevec != &p->page_array[0]))
  68                 kfree(p->pagevec);
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71 EXPORT_SYMBOL_GPL(nfs_commit_free);
  72
  73 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  74 {
  75         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  76
  77         if (p) {
  78                 memset(p, 0, sizeof(*p));
  79                 INIT_LIST_HEAD(&p->pages);
  80                 p->npages = pagecount;
  81                 if (pagecount <= ARRAY_SIZE(p->page_array))
  82                         p->pagevec = p->page_array;
  83                 else {
  84                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  85                         if (!p->pagevec) {
  86                                 mempool_free(p, nfs_wdata_mempool);
  87                                 p = NULL;
  88                         }
  89                 }
  90         }
  91         return p;
  92 }
  93
  94 void nfs_writedata_free(struct nfs_write_data *p)
  95 {
  96         if (p && (p->pagevec != &p->page_array[0]))
  97                 kfree(p->pagevec);
  98         mempool_free(p, nfs_wdata_mempool);
  99 }
 100
 101 void nfs_writedata_release(struct nfs_write_data *wdata)
 102 {
 103         put_lseg(wdata->lseg);
 104         put_nfs_open_context(wdata->args.context);
 105         nfs_writedata_free(wdata);
 106 }
 107
 108 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 109 {
 110         ctx->error = error;
 111         smp_wmb();
 112         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 113 }
 114
 115 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 116 {
 117         struct nfs_page *req = NULL;
 118
 119         if (PagePrivate(page)) {
 120                 req = (struct nfs_page *)page_private(page);
 121                 if (req != NULL)
 122                         kref_get(&req->wb_kref);
 123         }
 124         return req;
 125 }
 126
 127 static struct nfs_page *nfs_page_find_request(struct page *page)
 128 {
 129         struct inode *inode = page->mapping->host;
 130         struct nfs_page *req = NULL;
 131
 132         spin_lock(&inode->i_lock);
 133         req = nfs_page_find_request_locked(page);
 134         spin_unlock(&inode->i_lock);
 135         return req;
 136 }
 137
 138 /* Adjust the file length if we're writing beyond the end */
 139 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 140 {
 141         struct inode *inode = page->mapping->host;
 142         loff_t end, i_size;
 143         pgoff_t end_index;
 144
 145         spin_lock(&inode->i_lock);
 146         i_size = i_size_read(inode);
 147         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 148         if (i_size > 0 && page->index < end_index)
 149                 goto out;
 150         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 151         if (i_size >= end)
 152                 goto out;
 153         i_size_write(inode, end);
 154         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 155 out:
 156         spin_unlock(&inode->i_lock);
 157 }
 158
 159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 160 static void nfs_set_pageerror(struct page *page)
 161 {
 162         SetPageError(page);
 163         nfs_zap_mapping(page->mapping->host, page->mapping);
 164 }
 165
 166 /* We can set the PG_uptodate flag if we see that a write request
 167  * covers the full page.
 168  */
 169 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 170 {
 171         if (PageUptodate(page))
 172                 return;
 173         if (base != 0)
 174                 return;
 175         if (count != nfs_page_length(page))
 176                 return;
 177         SetPageUptodate(page);
 178 }
 179
 180 static int wb_priority(struct writeback_control *wbc)
 181 {
 182         if (wbc->for_reclaim)
 183                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 184         if (wbc->for_kupdate || wbc->for_background)
 185                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 186         return FLUSH_COND_STABLE;
 187 }
 188
 189 /*
 190  * NFS congestion control
 191  */
 192
 193 int nfs_congestion_kb;
 194
 195 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 196 #define NFS_CONGESTION_OFF_THRESH       \
 197         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 198
 199 static int nfs_set_page_writeback(struct page *page)
 200 {
 201         int ret = test_set_page_writeback(page);
 202
 203         if (!ret) {
 204                 struct inode *inode = page->mapping->host;
 205                 struct nfs_server *nfss = NFS_SERVER(inode);
 206
 207                 page_cache_get(page);
 208                 if (atomic_long_inc_return(&nfss->writeback) >
 209                                 NFS_CONGESTION_ON_THRESH) {
 210                         set_bdi_congested(&nfss->backing_dev_info,
 211                                                 BLK_RW_ASYNC);
 212                 }
 213         }
 214         return ret;
 215 }
 216
 217 static void nfs_end_page_writeback(struct page *page)
 218 {
 219         struct inode *inode = page->mapping->host;
 220         struct nfs_server *nfss = NFS_SERVER(inode);
 221
 222         end_page_writeback(page);
 223         page_cache_release(page);
 224         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 225                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 226 }
 227
 228 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 229 {
 230         struct inode *inode = page->mapping->host;
 231         struct nfs_page *req;
 232         int ret;
 233
 234         spin_lock(&inode->i_lock);
 235         for (;;) {
 236                 req = nfs_page_find_request_locked(page);
 237                 if (req == NULL)
 238                         break;
 239                 if (nfs_set_page_tag_locked(req))
 240                         break;
 241                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 242                  *       then the call to nfs_set_page_tag_locked() will always
 243                  *       succeed provided that someone hasn't already marked the
 244                  *       request as dirty (in which case we don't care).
 245                  */
 246                 spin_unlock(&inode->i_lock);
 247                 if (!nonblock)
 248                         ret = nfs_wait_on_request(req);
 249                 else
 250                         ret = -EAGAIN;
 251                 nfs_release_request(req);
 252                 if (ret != 0)
 253                         return ERR_PTR(ret);
 254                 spin_lock(&inode->i_lock);
 255         }
 256         spin_unlock(&inode->i_lock);
 257         return req;
 258 }
 259
 260 /*
 261  * Find an associated nfs write request, and prepare to flush it out
 262  * May return an error if the user signalled nfs_wait_on_request().
 263  */
 264 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 265                                 struct page *page, bool nonblock)
 266 {
 267         struct nfs_page *req;
 268         int ret = 0;
 269
 270         req = nfs_find_and_lock_request(page, nonblock);
 271         if (!req)
 272                 goto out;
 273         ret = PTR_ERR(req);
 274         if (IS_ERR(req))
 275                 goto out;
 276
 277         ret = nfs_set_page_writeback(page);
 278         BUG_ON(ret != 0);
 279         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
 280
 281         if (!nfs_pageio_add_request(pgio, req)) {
 282                 nfs_redirty_request(req);
 283                 ret = pgio->pg_error;
 284         }
 285 out:
 286         return ret;
 287 }
 288
 289 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 290 {
 291         struct inode *inode = page->mapping->host;
 292         int ret;
 293
 294         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 295         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 296
 297         nfs_pageio_cond_complete(pgio, page->index);
 298         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 299         if (ret == -EAGAIN) {
 300                 redirty_page_for_writepage(wbc, page);
 301                 ret = 0;
 302         }
 303         return ret;
 304 }
 305
 306 /*
 307  * Write an mmapped page to the server.
 308  */
 309 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 310 {
 311         struct nfs_pageio_descriptor pgio;
 312         int err;
 313
 314         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 315         err = nfs_do_writepage(page, wbc, &pgio);
 316         nfs_pageio_complete(&pgio);
 317         if (err < 0)
 318                 return err;
 319         if (pgio.pg_error < 0)
 320                 return pgio.pg_error;
 321         return 0;
 322 }
 323
 324 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 325 {
 326         int ret;
 327
 328         ret = nfs_writepage_locked(page, wbc);
 329         unlock_page(page);
 330         return ret;
 331 }
 332
 333 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 334 {
 335         int ret;
 336
 337         ret = nfs_do_writepage(page, wbc, data);
 338         unlock_page(page);
 339         return ret;
 340 }
 341
 342 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 343 {
 344         struct inode *inode = mapping->host;
 345         unsigned long *bitlock = &NFS_I(inode)->flags;
 346         struct nfs_pageio_descriptor pgio;
 347         int err;
 348
 349         /* Stop dirtying of new pages while we sync */
 350         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 351                         nfs_wait_bit_killable, TASK_KILLABLE);
 352         if (err)
 353                 goto out_err;
 354
 355         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 356
 357         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 358         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 359         nfs_pageio_complete(&pgio);
 360
 361         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 362         smp_mb__after_clear_bit();
 363         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 364
 365         if (err < 0)
 366                 goto out_err;
 367         err = pgio.pg_error;
 368         if (err < 0)
 369                 goto out_err;
 370         return 0;
 371 out_err:
 372         return err;
 373 }
 374
 375 /*
 376  * Insert a write request into an inode
 377  */
 378 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 379 {
 380         struct nfs_inode *nfsi = NFS_I(inode);
 381         int error;
 382
 383         error = radix_tree_preload(GFP_NOFS);
 384         if (error != 0)
 385                 goto out;
 386
 387         /* Lock the request! */
 388         nfs_lock_request_dontget(req);
 389
 390         spin_lock(&inode->i_lock);
 391         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 392         BUG_ON(error);
 393         if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
 394                 inode->i_version++;
 395         set_bit(PG_MAPPED, &req->wb_flags);
 396         SetPagePrivate(req->wb_page);
 397         set_page_private(req->wb_page, (unsigned long)req);
 398         nfsi->npages++;
 399         kref_get(&req->wb_kref);
 400         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 401                                 NFS_PAGE_TAG_LOCKED);
 402         spin_unlock(&inode->i_lock);
 403         radix_tree_preload_end();
 404 out:
 405         return error;
 406 }
 407
 408 /*
 409  * Remove a write request from an inode
 410  */
 411 static void nfs_inode_remove_request(struct nfs_page *req)
 412 {
 413         struct inode *inode = req->wb_context->dentry->d_inode;
 414         struct nfs_inode *nfsi = NFS_I(inode);
 415
 416         BUG_ON (!NFS_WBACK_BUSY(req));
 417
 418         spin_lock(&inode->i_lock);
 419         set_page_private(req->wb_page, 0);
 420         ClearPagePrivate(req->wb_page);
 421         clear_bit(PG_MAPPED, &req->wb_flags);
 422         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 423         nfsi->npages--;
 424         spin_unlock(&inode->i_lock);
 425         nfs_release_request(req);
 426 }
 427
 428 static void
 429 nfs_mark_request_dirty(struct nfs_page *req)
 430 {
 431         __set_page_dirty_nobuffers(req->wb_page);
 432 }
 433
 434 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 435 /*
 436  * Add a request to the inode's commit list.
 437  */
 438 static void
 439 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 440 {
 441         struct inode *inode = req->wb_context->dentry->d_inode;
 442         struct nfs_inode *nfsi = NFS_I(inode);
 443
 444         spin_lock(&inode->i_lock);
 445         set_bit(PG_CLEAN, &(req)->wb_flags);
 446         radix_tree_tag_set(&nfsi->nfs_page_tree,
 447                         req->wb_index,
 448                         NFS_PAGE_TAG_COMMIT);
 449         nfsi->ncommit++;
 450         spin_unlock(&inode->i_lock);
 451         pnfs_mark_request_commit(req, lseg);
 452         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 453         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 454         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 455 }
 456
 457 static int
 458 nfs_clear_request_commit(struct nfs_page *req)
 459 {
 460         struct page *page = req->wb_page;
 461
 462         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 463                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 464                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 465                 return 1;
 466         }
 467         return 0;
 468 }
 469
 470 static inline
 471 int nfs_write_need_commit(struct nfs_write_data *data)
 472 {
 473         if (data->verf.committed == NFS_DATA_SYNC)
 474                 return data->lseg == NULL;
 475         else
 476                 return data->verf.committed != NFS_FILE_SYNC;
 477 }
 478
 479 static inline
 480 int nfs_reschedule_unstable_write(struct nfs_page *req,
 481                                   struct nfs_write_data *data)
 482 {
 483         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 484                 nfs_mark_request_commit(req, data->lseg);
 485                 return 1;
 486         }
 487         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 488                 nfs_mark_request_dirty(req);
 489                 return 1;
 490         }
 491         return 0;
 492 }
 493 #else
 494 static inline void
 495 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 496 {
 497 }
 498
 499 static inline int
 500 nfs_clear_request_commit(struct nfs_page *req)
 501 {
 502         return 0;
 503 }
 504
 505 static inline
 506 int nfs_write_need_commit(struct nfs_write_data *data)
 507 {
 508         return 0;
 509 }
 510
 511 static inline
 512 int nfs_reschedule_unstable_write(struct nfs_page *req,
 513                                   struct nfs_write_data *data)
 514 {
 515         return 0;
 516 }
 517 #endif
 518
 519 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 520 static int
 521 nfs_need_commit(struct nfs_inode *nfsi)
 522 {
 523         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 524 }
 525
 526 /*
 527  * nfs_scan_commit - Scan an inode for commit requests
 528  * @inode: NFS inode to scan
 529  * @dst: destination list
 530  * @idx_start: lower bound of page->index to scan.
 531  * @npages: idx_start + npages sets the upper bound to scan.
 532  *
 533  * Moves requests from the inode's 'commit' request list.
 534  * The requests are *not* checked to ensure that they form a contiguous set.
 535  */
 536 static int
 537 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 538 {
 539         struct nfs_inode *nfsi = NFS_I(inode);
 540         int ret;
 541
 542         if (!nfs_need_commit(nfsi))
 543                 return 0;
 544
 545         spin_lock(&inode->i_lock);
 546         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 547         if (ret > 0)
 548                 nfsi->ncommit -= ret;
 549         spin_unlock(&inode->i_lock);
 550
 551         if (nfs_need_commit(NFS_I(inode)))
 552                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 553
 554         return ret;
 555 }
 556 #else
 557 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 558 {
 559         return 0;
 560 }
 561
 562 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 563 {
 564         return 0;
 565 }
 566 #endif
 567
 568 /*
 569  * Search for an existing write request, and attempt to update
 570  * it to reflect a new dirty region on a given page.
 571  *
 572  * If the attempt fails, then the existing request is flushed out
 573  * to disk.
 574  */
 575 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 576                 struct page *page,
 577                 unsigned int offset,
 578                 unsigned int bytes)
 579 {
 580         struct nfs_page *req;
 581         unsigned int rqend;
 582         unsigned int end;
 583         int error;
 584
 585         if (!PagePrivate(page))
 586                 return NULL;
 587
 588         end = offset + bytes;
 589         spin_lock(&inode->i_lock);
 590
 591         for (;;) {
 592                 req = nfs_page_find_request_locked(page);
 593                 if (req == NULL)
 594                         goto out_unlock;
 595
 596                 rqend = req->wb_offset + req->wb_bytes;
 597                 /*
 598                  * Tell the caller to flush out the request if
 599                  * the offsets are non-contiguous.
 600                  * Note: nfs_flush_incompatible() will already
 601                  * have flushed out requests having wrong owners.
 602                  */
 603                 if (offset > rqend
 604                     || end < req->wb_offset)
 605                         goto out_flushme;
 606
 607                 if (nfs_set_page_tag_locked(req))
 608                         break;
 609
 610                 /* The request is locked, so wait and then retry */
 611                 spin_unlock(&inode->i_lock);
 612                 error = nfs_wait_on_request(req);
 613                 nfs_release_request(req);
 614                 if (error != 0)
 615                         goto out_err;
 616                 spin_lock(&inode->i_lock);
 617         }
 618
 619         if (nfs_clear_request_commit(req) &&
 620             radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 621                                  req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
 622                 NFS_I(inode)->ncommit--;
 623                 pnfs_clear_request_commit(req);
 624         }
 625
 626         /* Okay, the request matches. Update the region */
 627         if (offset < req->wb_offset) {
 628                 req->wb_offset = offset;
 629                 req->wb_pgbase = offset;
 630         }
 631         if (end > rqend)
 632                 req->wb_bytes = end - req->wb_offset;
 633         else
 634                 req->wb_bytes = rqend - req->wb_offset;
 635 out_unlock:
 636         spin_unlock(&inode->i_lock);
 637         return req;
 638 out_flushme:
 639         spin_unlock(&inode->i_lock);
 640         nfs_release_request(req);
 641         error = nfs_wb_page(inode, page);
 642 out_err:
 643         return ERR_PTR(error);
 644 }
 645
 646 /*
 647  * Try to update an existing write request, or create one if there is none.
 648  *
 649  * Note: Should always be called with the Page Lock held to prevent races
 650  * if we have to add a new request. Also assumes that the caller has
 651  * already called nfs_flush_incompatible() if necessary.
 652  */
 653 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 654                 struct page *page, unsigned int offset, unsigned int bytes)
 655 {
 656         struct inode *inode = page->mapping->host;
 657         struct nfs_page *req;
 658         int error;
 659
 660         req = nfs_try_to_update_request(inode, page, offset, bytes);
 661         if (req != NULL)
 662                 goto out;
 663         req = nfs_create_request(ctx, inode, page, offset, bytes);
 664         if (IS_ERR(req))
 665                 goto out;
 666         error = nfs_inode_add_request(inode, req);
 667         if (error != 0) {
 668                 nfs_release_request(req);
 669                 req = ERR_PTR(error);
 670         }
 671 out:
 672         return req;
 673 }
 674
 675 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 676                 unsigned int offset, unsigned int count)
 677 {
 678         struct nfs_page *req;
 679
 680         req = nfs_setup_write_request(ctx, page, offset, count);
 681         if (IS_ERR(req))
 682                 return PTR_ERR(req);
 683         /* Update file length */
 684         nfs_grow_file(page, offset, count);
 685         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 686         nfs_mark_request_dirty(req);
 687         nfs_clear_page_tag_locked(req);
 688         return 0;
 689 }
 690
 691 int nfs_flush_incompatible(struct file *file, struct page *page)
 692 {
 693         struct nfs_open_context *ctx = nfs_file_open_context(file);
 694         struct nfs_page *req;
 695         int do_flush, status;
 696         /*
 697          * Look for a request corresponding to this page. If there
 698          * is one, and it belongs to another file, we flush it out
 699          * before we try to copy anything into the page. Do this
 700          * due to the lack of an ACCESS-type call in NFSv2.
 701          * Also do the same if we find a request from an existing
 702          * dropped page.
 703          */
 704         do {
 705                 req = nfs_page_find_request(page);
 706                 if (req == NULL)
 707                         return 0;
 708                 do_flush = req->wb_page != page || req->wb_context != ctx ||
 709                         req->wb_lock_context->lockowner != current->files ||
 710                         req->wb_lock_context->pid != current->tgid;
 711                 nfs_release_request(req);
 712                 if (!do_flush)
 713                         return 0;
 714                 status = nfs_wb_page(page->mapping->host, page);
 715         } while (status == 0);
 716         return status;
 717 }
 718
 719 /*
 720  * If the page cache is marked as unsafe or invalid, then we can't rely on
 721  * the PageUptodate() flag. In this case, we will need to turn off
 722  * write optimisations that depend on the page contents being correct.
 723  */
 724 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 725 {
 726         return PageUptodate(page) &&
 727                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 728 }
 729
 730 /*
 731  * Update and possibly write a cached page of an NFS file.
 732  *
 733  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 734  * things with a page scheduled for an RPC call (e.g. invalidate it).
 735  */
 736 int nfs_updatepage(struct file *file, struct page *page,
 737                 unsigned int offset, unsigned int count)
 738 {
 739         struct nfs_open_context *ctx = nfs_file_open_context(file);
 740         struct inode    *inode = page->mapping->host;
 741         int             status = 0;
 742
 743         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 744
 745         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 746                 file->f_path.dentry->d_parent->d_name.name,
 747                 file->f_path.dentry->d_name.name, count,
 748                 (long long)(page_offset(page) + offset));
 749
 750         /* If we're not using byte range locks, and we know the page
 751          * is up to date, it may be more efficient to extend the write
 752          * to cover the entire page in order to avoid fragmentation
 753          * inefficiencies.
 754          */
 755         if (nfs_write_pageuptodate(page, inode) &&
 756                         inode->i_flock == NULL &&
 757                         !(file->f_flags & O_DSYNC)) {
 758                 count = max(count + offset, nfs_page_length(page));
 759                 offset = 0;
 760         }
 761
 762         status = nfs_writepage_setup(ctx, page, offset, count);
 763         if (status < 0)
 764                 nfs_set_pageerror(page);
 765         else
 766                 __set_page_dirty_nobuffers(page);
 767
 768         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 769                         status, (long long)i_size_read(inode));
 770         return status;
 771 }
 772
 773 static void nfs_writepage_release(struct nfs_page *req,
 774                                   struct nfs_write_data *data)
 775 {
 776         struct page *page = req->wb_page;
 777
 778         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
 779                 nfs_inode_remove_request(req);
 780         nfs_clear_page_tag_locked(req);
 781         nfs_end_page_writeback(page);
 782 }
 783
 784 static int flush_task_priority(int how)
 785 {
 786         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 787                 case FLUSH_HIGHPRI:
 788                         return RPC_PRIORITY_HIGH;
 789                 case FLUSH_LOWPRI:
 790                         return RPC_PRIORITY_LOW;
 791         }
 792         return RPC_PRIORITY_NORMAL;
 793 }
 794
 795 int nfs_initiate_write(struct nfs_write_data *data,
 796                        struct rpc_clnt *clnt,
 797                        const struct rpc_call_ops *call_ops,
 798                        int how)
 799 {
 800         struct inode *inode = data->inode;
 801         int priority = flush_task_priority(how);
 802         struct rpc_task *task;
 803         struct rpc_message msg = {
 804                 .rpc_argp = &data->args,
 805                 .rpc_resp = &data->res,
 806                 .rpc_cred = data->cred,
 807         };
 808         struct rpc_task_setup task_setup_data = {
 809                 .rpc_client = clnt,
 810                 .task = &data->task,
 811                 .rpc_message = &msg,
 812                 .callback_ops = call_ops,
 813                 .callback_data = data,
 814                 .workqueue = nfsiod_workqueue,
 815                 .flags = RPC_TASK_ASYNC,
 816                 .priority = priority,
 817         };
 818         int ret = 0;
 819
 820         /* Set up the initial task struct.  */
 821         NFS_PROTO(inode)->write_setup(data, &msg);
 822
 823         dprintk("NFS: %5u initiated write call "
 824                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 825                 data->task.tk_pid,
 826                 inode->i_sb->s_id,
 827                 (long long)NFS_FILEID(inode),
 828                 data->args.count,
 829                 (unsigned long long)data->args.offset);
 830
 831         task = rpc_run_task(&task_setup_data);
 832         if (IS_ERR(task)) {
 833                 ret = PTR_ERR(task);
 834                 goto out;
 835         }
 836         if (how & FLUSH_SYNC) {
 837                 ret = rpc_wait_for_completion_task(task);
 838                 if (ret == 0)
 839                         ret = task->tk_status;
 840         }
 841         rpc_put_task(task);
 842 out:
 843         return ret;
 844 }
 845 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 846
 847 /*
 848  * Set up the argument/result storage required for the RPC call.
 849  */
 850 static void nfs_write_rpcsetup(struct nfs_page *req,
 851                 struct nfs_write_data *data,
 852                 unsigned int count, unsigned int offset,
 853                 int how)
 854 {
 855         struct inode *inode = req->wb_context->dentry->d_inode;
 856
 857         /* Set up the RPC argument and reply structs
 858          * NB: take care not to mess about with data->commit et al. */
 859
 860         data->req = req;
 861         data->inode = inode = req->wb_context->dentry->d_inode;
 862         data->cred = req->wb_context->cred;
 863
 864         data->args.fh     = NFS_FH(inode);
 865         data->args.offset = req_offset(req) + offset;
 866         /* pnfs_set_layoutcommit needs this */
 867         data->mds_offset = data->args.offset;
 868         data->args.pgbase = req->wb_pgbase + offset;
 869         data->args.pages  = data->pagevec;
 870         data->args.count  = count;
 871         data->args.context = get_nfs_open_context(req->wb_context);
 872         data->args.lock_context = req->wb_lock_context;
 873         data->args.stable  = NFS_UNSTABLE;
 874         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 875         case 0:
 876                 break;
 877         case FLUSH_COND_STABLE:
 878                 if (nfs_need_commit(NFS_I(inode)))
 879                         break;
 880         default:
 881                 data->args.stable = NFS_FILE_SYNC;
 882         }
 883
 884         data->res.fattr   = &data->fattr;
 885         data->res.count   = count;
 886         data->res.verf    = &data->verf;
 887         nfs_fattr_init(&data->fattr);
 888 }
 889
 890 static int nfs_do_write(struct nfs_write_data *data,
 891                 const struct rpc_call_ops *call_ops,
 892                 int how)
 893 {
 894         struct inode *inode = data->args.context->dentry->d_inode;
 895
 896         return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
 897 }
 898
 899 static int nfs_do_multiple_writes(struct list_head *head,
 900                 const struct rpc_call_ops *call_ops,
 901                 int how)
 902 {
 903         struct nfs_write_data *data;
 904         int ret = 0;
 905
 906         while (!list_empty(head)) {
 907                 int ret2;
 908
 909                 data = list_entry(head->next, struct nfs_write_data, list);
 910                 list_del_init(&data->list);
 911
 912                 ret2 = nfs_do_write(data, call_ops, how);
 913                  if (ret == 0)
 914                          ret = ret2;
 915         }
 916         return ret;
 917 }
 918
 919 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 920  * call this on each, which will prepare them to be retried on next
 921  * writeback using standard nfs.
 922  */
 923 static void nfs_redirty_request(struct nfs_page *req)
 924 {
 925         struct page *page = req->wb_page;
 926
 927         nfs_mark_request_dirty(req);
 928         nfs_clear_page_tag_locked(req);
 929         nfs_end_page_writeback(page);
 930 }
 931
 932 /*
 933  * Generate multiple small requests to write out a single
 934  * contiguous dirty area on one page.
 935  */
 936 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
 937 {
 938         struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
 939         struct page *page = req->wb_page;
 940         struct nfs_write_data *data;
 941         size_t wsize = desc->pg_bsize, nbytes;
 942         unsigned int offset;
 943         int requests = 0;
 944         int ret = 0;
 945
 946         nfs_list_remove_request(req);
 947
 948         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 949             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
 950              desc->pg_count > wsize))
 951                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 952
 953
 954         offset = 0;
 955         nbytes = desc->pg_count;
 956         do {
 957                 size_t len = min(nbytes, wsize);
 958
 959                 data = nfs_writedata_alloc(1);
 960                 if (!data)
 961                         goto out_bad;
 962                 data->pagevec[0] = page;
 963                 nfs_write_rpcsetup(req, data, len, offset, desc->pg_ioflags);
 964                 list_add(&data->list, res);
 965                 requests++;
 966                 nbytes -= len;
 967                 offset += len;
 968         } while (nbytes != 0);
 969         atomic_set(&req->wb_complete, requests);
 970         desc->pg_rpc_callops = &nfs_write_partial_ops;
 971         return ret;
 972
 973 out_bad:
 974         while (!list_empty(res)) {
 975                 data = list_entry(res->next, struct nfs_write_data, list);
 976                 list_del(&data->list);
 977                 nfs_writedata_free(data);
 978         }
 979         nfs_redirty_request(req);
 980         return -ENOMEM;
 981 }
 982
 983 /*
 984  * Create an RPC task for the given write request and kick it.
 985  * The page must have been locked by the caller.
 986  *
 987  * It may happen that the page we're passed is not marked dirty.
 988  * This is the case if nfs_updatepage detects a conflicting request
 989  * that has been written but not committed.
 990  */
 991 static int nfs_flush_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
 992 {
 993         struct nfs_page         *req;
 994         struct page             **pages;
 995         struct nfs_write_data   *data;
 996         struct list_head *head = &desc->pg_list;
 997         int ret = 0;
 998
 999         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
1000                                                       desc->pg_count));
1001         if (!data) {
1002                 while (!list_empty(head)) {
1003                         req = nfs_list_entry(head->next);
1004                         nfs_list_remove_request(req);
1005                         nfs_redirty_request(req);
1006                 }
1007                 ret = -ENOMEM;
1008                 goto out;
1009         }
1010         pages = data->pagevec;
1011         while (!list_empty(head)) {
1012                 req = nfs_list_entry(head->next);
1013                 nfs_list_remove_request(req);
1014                 nfs_list_add_request(req, &data->pages);
1015                 *pages++ = req->wb_page;
1016         }
1017         req = nfs_list_entry(data->pages.next);
1018
1019         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1020             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1021                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1022
1023         /* Set up the argument struct */
1024         nfs_write_rpcsetup(req, data, desc->pg_count, 0, desc->pg_ioflags);
1025         list_add(&data->list, res);
1026         desc->pg_rpc_callops = &nfs_write_full_ops;
1027 out:
1028         return ret;
1029 }
1030
1031 int nfs_generic_flush(struct nfs_pageio_descriptor *desc, struct list_head *head)
1032 {
1033         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1034                 return nfs_flush_multi(desc, head);
1035         return nfs_flush_one(desc, head);
1036 }
1037
1038 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1039 {
1040         LIST_HEAD(head);
1041         int ret;
1042
1043         ret = nfs_generic_flush(desc, &head);
1044         if (ret == 0)
1045                 ret = nfs_do_multiple_writes(&head, desc->pg_rpc_callops,
1046                                 desc->pg_ioflags);
1047         return ret;
1048 }
1049
1050 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1051         .pg_test = nfs_generic_pg_test,
1052         .pg_doio = nfs_generic_pg_writepages,
1053 };
1054
1055 static void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1056                                   struct inode *inode, int ioflags)
1057 {
1058         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops,
1059                                 NFS_SERVER(inode)->wsize, ioflags);
1060 }
1061
1062 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1063 {
1064         pgio->pg_ops = &nfs_pageio_write_ops;
1065         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1066 }
1067 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1068
1069 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1070                                   struct inode *inode, int ioflags)
1071 {
1072         if (!pnfs_pageio_init_write(pgio, inode, ioflags))
1073                 nfs_pageio_init_write_mds(pgio, inode, ioflags);
1074 }
1075
1076 /*
1077  * Handle a write reply that flushed part of a page.
1078  */
1079 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1080 {
1081         struct nfs_write_data   *data = calldata;
1082
1083         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1084                 task->tk_pid,
1085                 data->req->wb_context->dentry->d_inode->i_sb->s_id,
1086                 (long long)
1087                   NFS_FILEID(data->req->wb_context->dentry->d_inode),
1088                 data->req->wb_bytes, (long long)req_offset(data->req));
1089
1090         nfs_writeback_done(task, data);
1091 }
1092
1093 static void nfs_writeback_release_partial(void *calldata)
1094 {
1095         struct nfs_write_data   *data = calldata;
1096         struct nfs_page         *req = data->req;
1097         struct page             *page = req->wb_page;
1098         int status = data->task.tk_status;
1099
1100         if (status < 0) {
1101                 nfs_set_pageerror(page);
1102                 nfs_context_set_write_error(req->wb_context, status);
1103                 dprintk(", error = %d\n", status);
1104                 goto out;
1105         }
1106
1107         if (nfs_write_need_commit(data)) {
1108                 struct inode *inode = page->mapping->host;
1109
1110                 spin_lock(&inode->i_lock);
1111                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1112                         /* Do nothing we need to resend the writes */
1113                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1114                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1115                         dprintk(" defer commit\n");
1116                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1117                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1118                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1119                         dprintk(" server reboot detected\n");
1120                 }
1121                 spin_unlock(&inode->i_lock);
1122         } else
1123                 dprintk(" OK\n");
1124
1125 out:
1126         if (atomic_dec_and_test(&req->wb_complete))
1127                 nfs_writepage_release(req, data);
1128         nfs_writedata_release(calldata);
1129 }
1130
1131 #if defined(CONFIG_NFS_V4_1)
1132 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1133 {
1134         struct nfs_write_data *data = calldata;
1135
1136         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1137                                 &data->args.seq_args,
1138                                 &data->res.seq_res, 1, task))
1139                 return;
1140         rpc_call_start(task);
1141 }
1142 #endif /* CONFIG_NFS_V4_1 */
1143
1144 static const struct rpc_call_ops nfs_write_partial_ops = {
1145 #if defined(CONFIG_NFS_V4_1)
1146         .rpc_call_prepare = nfs_write_prepare,
1147 #endif /* CONFIG_NFS_V4_1 */
1148         .rpc_call_done = nfs_writeback_done_partial,
1149         .rpc_release = nfs_writeback_release_partial,
1150 };
1151
1152 /*
1153  * Handle a write reply that flushes a whole page.
1154  *
1155  * FIXME: There is an inherent race with invalidate_inode_pages and
1156  *        writebacks since the page->count is kept > 1 for as long
1157  *        as the page has a write request pending.
1158  */
1159 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1160 {
1161         struct nfs_write_data   *data = calldata;
1162
1163         nfs_writeback_done(task, data);
1164 }
1165
1166 static void nfs_writeback_release_full(void *calldata)
1167 {
1168         struct nfs_write_data   *data = calldata;
1169         int ret, status = data->task.tk_status;
1170         struct nfs_pageio_descriptor pgio;
1171
1172         if (data->pnfs_error) {
1173                 nfs_pageio_init_write_mds(&pgio, data->inode, FLUSH_STABLE);
1174                 pgio.pg_recoalesce = 1;
1175         }
1176
1177         /* Update attributes as result of writeback. */
1178         while (!list_empty(&data->pages)) {
1179                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1180                 struct page *page = req->wb_page;
1181
1182                 nfs_list_remove_request(req);
1183
1184                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1185                         data->task.tk_pid,
1186                         req->wb_context->dentry->d_inode->i_sb->s_id,
1187                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1188                         req->wb_bytes,
1189                         (long long)req_offset(req));
1190
1191                 if (data->pnfs_error) {
1192                         dprintk(", pnfs error = %d\n", data->pnfs_error);
1193                         goto next;
1194                 }
1195
1196                 if (status < 0) {
1197                         nfs_set_pageerror(page);
1198                         nfs_context_set_write_error(req->wb_context, status);
1199                         dprintk(", error = %d\n", status);
1200                         goto remove_request;
1201                 }
1202
1203                 if (nfs_write_need_commit(data)) {
1204                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1205                         nfs_mark_request_commit(req, data->lseg);
1206                         dprintk(" marked for commit\n");
1207                         goto next;
1208                 }
1209                 dprintk(" OK\n");
1210 remove_request:
1211                 nfs_inode_remove_request(req);
1212         next:
1213                 nfs_clear_page_tag_locked(req);
1214                 nfs_end_page_writeback(page);
1215                 if (data->pnfs_error) {
1216                         lock_page(page);
1217                         nfs_pageio_cond_complete(&pgio, page->index);
1218                         ret = nfs_page_async_flush(&pgio, page, 0);
1219                         if (ret) {
1220                                 nfs_set_pageerror(page);
1221                                 dprintk("rewrite to MDS error = %d\n", ret);
1222                         }
1223                         unlock_page(page);
1224                 }
1225         }
1226         if (data->pnfs_error)
1227                 nfs_pageio_complete(&pgio);
1228         nfs_writedata_release(calldata);
1229 }
1230
1231 static const struct rpc_call_ops nfs_write_full_ops = {
1232 #if defined(CONFIG_NFS_V4_1)
1233         .rpc_call_prepare = nfs_write_prepare,
1234 #endif /* CONFIG_NFS_V4_1 */
1235         .rpc_call_done = nfs_writeback_done_full,
1236         .rpc_release = nfs_writeback_release_full,
1237 };
1238
1239
1240 /*
1241  * This function is called when the WRITE call is complete.
1242  */
1243 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1244 {
1245         struct nfs_writeargs    *argp = &data->args;
1246         struct nfs_writeres     *resp = &data->res;
1247         int status;
1248
1249         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1250                 task->tk_pid, task->tk_status);
1251
1252         /*
1253          * ->write_done will attempt to use post-op attributes to detect
1254          * conflicting writes by other clients.  A strict interpretation
1255          * of close-to-open would allow us to continue caching even if
1256          * another writer had changed the file, but some applications
1257          * depend on tighter cache coherency when writing.
1258          */
1259         status = NFS_PROTO(data->inode)->write_done(task, data);
1260         if (status != 0)
1261                 return;
1262         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1263
1264 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1265         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1266                 /* We tried a write call, but the server did not
1267                  * commit data to stable storage even though we
1268                  * requested it.
1269                  * Note: There is a known bug in Tru64 < 5.0 in which
1270                  *       the server reports NFS_DATA_SYNC, but performs
1271                  *       NFS_FILE_SYNC. We therefore implement this checking
1272                  *       as a dprintk() in order to avoid filling syslog.
1273                  */
1274                 static unsigned long    complain;
1275
1276                 /* Note this will print the MDS for a DS write */
1277                 if (time_before(complain, jiffies)) {
1278                         dprintk("NFS:       faulty NFS server %s:"
1279                                 " (committed = %d) != (stable = %d)\n",
1280                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1281                                 resp->verf->committed, argp->stable);
1282                         complain = jiffies + 300 * HZ;
1283                 }
1284         }
1285 #endif
1286         /* Is this a short write? */
1287         if (task->tk_status >= 0 && resp->count < argp->count) {
1288                 static unsigned long    complain;
1289
1290                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1291
1292                 /* Has the server at least made some progress? */
1293                 if (resp->count != 0) {
1294                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1295                         if (resp->verf->committed != NFS_UNSTABLE) {
1296                                 /* Resend from where the server left off */
1297                                 data->mds_offset += resp->count;
1298                                 argp->offset += resp->count;
1299                                 argp->pgbase += resp->count;
1300                                 argp->count -= resp->count;
1301                         } else {
1302                                 /* Resend as a stable write in order to avoid
1303                                  * headaches in the case of a server crash.
1304                                  */
1305                                 argp->stable = NFS_FILE_SYNC;
1306                         }
1307                         rpc_restart_call_prepare(task);
1308                         return;
1309                 }
1310                 if (time_before(complain, jiffies)) {
1311                         printk(KERN_WARNING
1312                                "NFS: Server wrote zero bytes, expected %u.\n",
1313                                         argp->count);
1314                         complain = jiffies + 300 * HZ;
1315                 }
1316                 /* Can't do anything about it except throw an error. */
1317                 task->tk_status = -EIO;
1318         }
1319         return;
1320 }
1321
1322
1323 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1324 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1325 {
1326         int ret;
1327
1328         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1329                 return 1;
1330         if (!may_wait)
1331                 return 0;
1332         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1333                                 NFS_INO_COMMIT,
1334                                 nfs_wait_bit_killable,
1335                                 TASK_KILLABLE);
1336         return (ret < 0) ? ret : 1;
1337 }
1338
1339 void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1340 {
1341         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1342         smp_mb__after_clear_bit();
1343         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1344 }
1345 EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1346
1347 void nfs_commitdata_release(void *data)
1348 {
1349         struct nfs_write_data *wdata = data;
1350
1351         put_lseg(wdata->lseg);
1352         put_nfs_open_context(wdata->args.context);
1353         nfs_commit_free(wdata);
1354 }
1355 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1356
1357 int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1358                         const struct rpc_call_ops *call_ops,
1359                         int how)
1360 {
1361         struct rpc_task *task;
1362         int priority = flush_task_priority(how);
1363         struct rpc_message msg = {
1364                 .rpc_argp = &data->args,
1365                 .rpc_resp = &data->res,
1366                 .rpc_cred = data->cred,
1367         };
1368         struct rpc_task_setup task_setup_data = {
1369                 .task = &data->task,
1370                 .rpc_client = clnt,
1371                 .rpc_message = &msg,
1372                 .callback_ops = call_ops,
1373                 .callback_data = data,
1374                 .workqueue = nfsiod_workqueue,
1375                 .flags = RPC_TASK_ASYNC,
1376                 .priority = priority,
1377         };
1378         /* Set up the initial task struct.  */
1379         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1380
1381         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1382
1383         task = rpc_run_task(&task_setup_data);
1384         if (IS_ERR(task))
1385                 return PTR_ERR(task);
1386         if (how & FLUSH_SYNC)
1387                 rpc_wait_for_completion_task(task);
1388         rpc_put_task(task);
1389         return 0;
1390 }
1391 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1392
1393 /*
1394  * Set up the argument/result storage required for the RPC call.
1395  */
1396 void nfs_init_commit(struct nfs_write_data *data,
1397                             struct list_head *head,
1398                             struct pnfs_layout_segment *lseg)
1399 {
1400         struct nfs_page *first = nfs_list_entry(head->next);
1401         struct inode *inode = first->wb_context->dentry->d_inode;
1402
1403         /* Set up the RPC argument and reply structs
1404          * NB: take care not to mess about with data->commit et al. */
1405
1406         list_splice_init(head, &data->pages);
1407
1408         data->inode       = inode;
1409         data->cred        = first->wb_context->cred;
1410         data->lseg        = lseg; /* reference transferred */
1411         data->mds_ops     = &nfs_commit_ops;
1412
1413         data->args.fh     = NFS_FH(data->inode);
1414         /* Note: we always request a commit of the entire inode */
1415         data->args.offset = 0;
1416         data->args.count  = 0;
1417         data->args.context = get_nfs_open_context(first->wb_context);
1418         data->res.count   = 0;
1419         data->res.fattr   = &data->fattr;
1420         data->res.verf    = &data->verf;
1421         nfs_fattr_init(&data->fattr);
1422 }
1423 EXPORT_SYMBOL_GPL(nfs_init_commit);
1424
1425 void nfs_retry_commit(struct list_head *page_list,
1426                       struct pnfs_layout_segment *lseg)
1427 {
1428         struct nfs_page *req;
1429
1430         while (!list_empty(page_list)) {
1431                 req = nfs_list_entry(page_list->next);
1432                 nfs_list_remove_request(req);
1433                 nfs_mark_request_commit(req, lseg);
1434                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1435                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1436                              BDI_RECLAIMABLE);
1437                 nfs_clear_page_tag_locked(req);
1438         }
1439 }
1440 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1441
1442 /*
1443  * Commit dirty pages
1444  */
1445 static int
1446 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1447 {
1448         struct nfs_write_data   *data;
1449
1450         data = nfs_commitdata_alloc();
1451
1452         if (!data)
1453                 goto out_bad;
1454
1455         /* Set up the argument struct */
1456         nfs_init_commit(data, head, NULL);
1457         return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1458  out_bad:
1459         nfs_retry_commit(head, NULL);
1460         nfs_commit_clear_lock(NFS_I(inode));
1461         return -ENOMEM;
1462 }
1463
1464 /*
1465  * COMMIT call returned
1466  */
1467 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1468 {
1469         struct nfs_write_data   *data = calldata;
1470
1471         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1472                                 task->tk_pid, task->tk_status);
1473
1474         /* Call the NFS version-specific code */
1475         NFS_PROTO(data->inode)->commit_done(task, data);
1476 }
1477
1478 void nfs_commit_release_pages(struct nfs_write_data *data)
1479 {
1480         struct nfs_page *req;
1481         int status = data->task.tk_status;
1482
1483         while (!list_empty(&data->pages)) {
1484                 req = nfs_list_entry(data->pages.next);
1485                 nfs_list_remove_request(req);
1486                 nfs_clear_request_commit(req);
1487
1488                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1489                         req->wb_context->dentry->d_sb->s_id,
1490                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1491                         req->wb_bytes,
1492                         (long long)req_offset(req));
1493                 if (status < 0) {
1494                         nfs_context_set_write_error(req->wb_context, status);
1495                         nfs_inode_remove_request(req);
1496                         dprintk(", error = %d\n", status);
1497                         goto next;
1498                 }
1499
1500                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1501                  * returned by the server against all stored verfs. */
1502                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1503                         /* We have a match */
1504                         nfs_inode_remove_request(req);
1505                         dprintk(" OK\n");
1506                         goto next;
1507                 }
1508                 /* We have a mismatch. Write the page again */
1509                 dprintk(" mismatch\n");
1510                 nfs_mark_request_dirty(req);
1511         next:
1512                 nfs_clear_page_tag_locked(req);
1513         }
1514 }
1515 EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1516
1517 static void nfs_commit_release(void *calldata)
1518 {
1519         struct nfs_write_data *data = calldata;
1520
1521         nfs_commit_release_pages(data);
1522         nfs_commit_clear_lock(NFS_I(data->inode));
1523         nfs_commitdata_release(calldata);
1524 }
1525
1526 static const struct rpc_call_ops nfs_commit_ops = {
1527 #if defined(CONFIG_NFS_V4_1)
1528         .rpc_call_prepare = nfs_write_prepare,
1529 #endif /* CONFIG_NFS_V4_1 */
1530         .rpc_call_done = nfs_commit_done,
1531         .rpc_release = nfs_commit_release,
1532 };
1533
1534 int nfs_commit_inode(struct inode *inode, int how)
1535 {
1536         LIST_HEAD(head);
1537         int may_wait = how & FLUSH_SYNC;
1538         int res;
1539
1540         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1541         if (res <= 0)
1542                 goto out_mark_dirty;
1543         res = nfs_scan_commit(inode, &head, 0, 0);
1544         if (res) {
1545                 int error;
1546
1547                 error = pnfs_commit_list(inode, &head, how);
1548                 if (error == PNFS_NOT_ATTEMPTED)
1549                         error = nfs_commit_list(inode, &head, how);
1550                 if (error < 0)
1551                         return error;
1552                 if (!may_wait)
1553                         goto out_mark_dirty;
1554                 error = wait_on_bit(&NFS_I(inode)->flags,
1555                                 NFS_INO_COMMIT,
1556                                 nfs_wait_bit_killable,
1557                                 TASK_KILLABLE);
1558                 if (error < 0)
1559                         return error;
1560         } else
1561                 nfs_commit_clear_lock(NFS_I(inode));
1562         return res;
1563         /* Note: If we exit without ensuring that the commit is complete,
1564          * we must mark the inode as dirty. Otherwise, future calls to
1565          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1566          * that the data is on the disk.
1567          */
1568 out_mark_dirty:
1569         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1570         return res;
1571 }
1572
1573 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1574 {
1575         struct nfs_inode *nfsi = NFS_I(inode);
1576         int flags = FLUSH_SYNC;
1577         int ret = 0;
1578
1579         /* no commits means nothing needs to be done */
1580         if (!nfsi->ncommit)
1581                 return ret;
1582
1583         if (wbc->sync_mode == WB_SYNC_NONE) {
1584                 /* Don't commit yet if this is a non-blocking flush and there
1585                  * are a lot of outstanding writes for this mapping.
1586                  */
1587                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1588                         goto out_mark_dirty;
1589
1590                 /* don't wait for the COMMIT response */
1591                 flags = 0;
1592         }
1593
1594         ret = nfs_commit_inode(inode, flags);
1595         if (ret >= 0) {
1596                 if (wbc->sync_mode == WB_SYNC_NONE) {
1597                         if (ret < wbc->nr_to_write)
1598                                 wbc->nr_to_write -= ret;
1599                         else
1600                                 wbc->nr_to_write = 0;
1601                 }
1602                 return 0;
1603         }
1604 out_mark_dirty:
1605         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1606         return ret;
1607 }
1608 #else
1609 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1610 {
1611         return 0;
1612 }
1613 #endif
1614
1615 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1616 {
1617         int ret;
1618
1619         ret = nfs_commit_unstable_pages(inode, wbc);
1620         if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1621                 int status;
1622                 bool sync = true;
1623
1624                 if (wbc->sync_mode == WB_SYNC_NONE)
1625                         sync = false;
1626
1627                 status = pnfs_layoutcommit_inode(inode, sync);
1628                 if (status < 0)
1629                         return status;
1630         }
1631         return ret;
1632 }
1633
1634 /*
1635  * flush the inode to disk.
1636  */
1637 int nfs_wb_all(struct inode *inode)
1638 {
1639         struct writeback_control wbc = {
1640                 .sync_mode = WB_SYNC_ALL,
1641                 .nr_to_write = LONG_MAX,
1642                 .range_start = 0,
1643                 .range_end = LLONG_MAX,
1644         };
1645
1646         return sync_inode(inode, &wbc);
1647 }
1648
1649 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1650 {
1651         struct nfs_page *req;
1652         int ret = 0;
1653
1654         BUG_ON(!PageLocked(page));
1655         for (;;) {
1656                 wait_on_page_writeback(page);
1657                 req = nfs_page_find_request(page);
1658                 if (req == NULL)
1659                         break;
1660                 if (nfs_lock_request_dontget(req)) {
1661                         nfs_inode_remove_request(req);
1662                         /*
1663                          * In case nfs_inode_remove_request has marked the
1664                          * page as being dirty
1665                          */
1666                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1667                         nfs_unlock_request(req);
1668                         break;
1669                 }
1670                 ret = nfs_wait_on_request(req);
1671                 nfs_release_request(req);
1672                 if (ret < 0)
1673                         break;
1674         }
1675         return ret;
1676 }
1677
1678 /*
1679  * Write back all requests on one page - we do this before reading it.
1680  */
1681 int nfs_wb_page(struct inode *inode, struct page *page)
1682 {
1683         loff_t range_start = page_offset(page);
1684         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1685         struct writeback_control wbc = {
1686                 .sync_mode = WB_SYNC_ALL,
1687                 .nr_to_write = 0,
1688                 .range_start = range_start,
1689                 .range_end = range_end,
1690         };
1691         int ret;
1692
1693         for (;;) {
1694                 wait_on_page_writeback(page);
1695                 if (clear_page_dirty_for_io(page)) {
1696                         ret = nfs_writepage_locked(page, &wbc);
1697                         if (ret < 0)
1698                                 goto out_error;
1699                         continue;
1700                 }
1701                 if (!PagePrivate(page))
1702                         break;
1703                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1704                 if (ret < 0)
1705                         goto out_error;
1706         }
1707         return 0;
1708 out_error:
1709         return ret;
1710 }
1711
1712 #ifdef CONFIG_MIGRATION
1713 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1714                 struct page *page)
1715 {
1716         /*
1717          * If PagePrivate is set, then the page is currently associated with
1718          * an in-progress read or write request. Don't try to migrate it.
1719          *
1720          * FIXME: we could do this in principle, but we'll need a way to ensure
1721          *        that we can safely release the inode reference while holding
1722          *        the page lock.
1723          */
1724         if (PagePrivate(page))
1725                 return -EBUSY;
1726
1727         nfs_fscache_release_page(page, GFP_KERNEL);
1728
1729         return migrate_page(mapping, newpage, page);
1730 }
1731 #endif
1732
1733 int __init nfs_init_writepagecache(void)
1734 {
1735         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1736                                              sizeof(struct nfs_write_data),
1737                                              0, SLAB_HWCACHE_ALIGN,
1738                                              NULL);
1739         if (nfs_wdata_cachep == NULL)
1740                 return -ENOMEM;
1741
1742         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1743                                                      nfs_wdata_cachep);
1744         if (nfs_wdata_mempool == NULL)
1745                 return -ENOMEM;
1746
1747         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1748                                                       nfs_wdata_cachep);
1749         if (nfs_commit_mempool == NULL)
1750                 return -ENOMEM;
1751
1752         /*
1753          * NFS congestion size, scale with available memory.
1754          *
1755          *  64MB:    8192k
1756          * 128MB:   11585k
1757          * 256MB:   16384k
1758          * 512MB:   23170k
1759          *   1GB:   32768k
1760          *   2GB:   46340k
1761          *   4GB:   65536k
1762          *   8GB:   92681k
1763          *  16GB:  131072k
1764          *
1765          * This allows larger machines to have larger/more transfers.
1766          * Limit the default to 256M
1767          */
1768         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1769         if (nfs_congestion_kb > 256*1024)
1770                 nfs_congestion_kb = 256*1024;
1771
1772         return 0;
1773 }
1774
1775 void nfs_destroy_writepagecache(void)
1776 {
1777         mempool_destroy(nfs_commit_mempool);
1778         mempool_destroy(nfs_wdata_mempool);
1779         kmem_cache_destroy(nfs_wdata_cachep);
1780 }
1781