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