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;
 183         return 0;
 184 }
 185
 186 /*
 187  * NFS congestion control
 188  */
 189
 190 int nfs_congestion_kb;
 191
 192 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 193 #define NFS_CONGESTION_OFF_THRESH       \
 194         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 195
 196 static int nfs_set_page_writeback(struct page *page)
 197 {
 198         int ret = test_set_page_writeback(page);
 199
 200         if (!ret) {
 201                 struct inode *inode = page->mapping->host;
 202                 struct nfs_server *nfss = NFS_SERVER(inode);
 203
 204                 page_cache_get(page);
 205                 if (atomic_long_inc_return(&nfss->writeback) >
 206                                 NFS_CONGESTION_ON_THRESH) {
 207                         set_bdi_congested(&nfss->backing_dev_info,
 208                                                 BLK_RW_ASYNC);
 209                 }
 210         }
 211         return ret;
 212 }
 213
 214 static void nfs_end_page_writeback(struct page *page)
 215 {
 216         struct inode *inode = page->mapping->host;
 217         struct nfs_server *nfss = NFS_SERVER(inode);
 218
 219         end_page_writeback(page);
 220         page_cache_release(page);
 221         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 222                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 223 }
 224
 225 static struct nfs_page *nfs_find_and_lock_request(struct page *page, 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) {
 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         nbytes = desc->pg_count;
 916         do {
 917                 size_t len = min(nbytes, wsize);
 918
 919                 data = nfs_writedata_alloc(1);
 920                 if (!data)
 921                         goto out_bad;
 922                 list_add(&data->pages, &list);
 923                 requests++;
 924                 nbytes -= len;
 925         } while (nbytes != 0);
 926         atomic_set(&req->wb_complete, requests);
 927
 928         BUG_ON(desc->pg_lseg);
 929         lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
 930         ClearPageError(page);
 931         offset = 0;
 932         nbytes = desc->pg_count;
 933         do {
 934                 int ret2;
 935
 936                 data = list_entry(list.next, struct nfs_write_data, pages);
 937                 list_del_init(&data->pages);
 938
 939                 data->pagevec[0] = page;
 940
 941                 if (nbytes < wsize)
 942                         wsize = nbytes;
 943                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 944                                           wsize, offset, lseg, desc->pg_ioflags);
 945                 if (ret == 0)
 946                         ret = ret2;
 947                 offset += wsize;
 948                 nbytes -= wsize;
 949         } while (nbytes != 0);
 950
 951         put_lseg(lseg);
 952         desc->pg_lseg = NULL;
 953         return ret;
 954
 955 out_bad:
 956         while (!list_empty(&list)) {
 957                 data = list_entry(list.next, struct nfs_write_data, pages);
 958                 list_del(&data->pages);
 959                 nfs_writedata_free(data);
 960         }
 961         nfs_redirty_request(req);
 962         return -ENOMEM;
 963 }
 964
 965 /*
 966  * Create an RPC task for the given write request and kick it.
 967  * The page must have been locked by the caller.
 968  *
 969  * It may happen that the page we're passed is not marked dirty.
 970  * This is the case if nfs_updatepage detects a conflicting request
 971  * that has been written but not committed.
 972  */
 973 static int nfs_flush_one(struct nfs_pageio_descriptor *desc)
 974 {
 975         struct nfs_page         *req;
 976         struct page             **pages;
 977         struct nfs_write_data   *data;
 978         struct list_head *head = &desc->pg_list;
 979         struct pnfs_layout_segment *lseg = desc->pg_lseg;
 980         int ret;
 981
 982         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
 983                                                       desc->pg_count));
 984         if (!data) {
 985                 while (!list_empty(head)) {
 986                         req = nfs_list_entry(head->next);
 987                         nfs_list_remove_request(req);
 988                         nfs_redirty_request(req);
 989                 }
 990                 ret = -ENOMEM;
 991                 goto out;
 992         }
 993         pages = data->pagevec;
 994         while (!list_empty(head)) {
 995                 req = nfs_list_entry(head->next);
 996                 nfs_list_remove_request(req);
 997                 nfs_list_add_request(req, &data->pages);
 998                 ClearPageError(req->wb_page);
 999                 *pages++ = req->wb_page;
1000         }
1001         req = nfs_list_entry(data->pages.next);
1002         if ((!lseg) && list_is_singular(&data->pages))
1003                 lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
1004
1005         /* Set up the argument struct */
1006         ret = nfs_write_rpcsetup(req, data, &nfs_write_full_ops, desc->pg_count, 0, lseg, desc->pg_ioflags);
1007 out:
1008         put_lseg(lseg); /* Cleans any gotten in ->pg_test */
1009         desc->pg_lseg = NULL;
1010         return ret;
1011 }
1012
1013 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1014                                   struct inode *inode, int ioflags)
1015 {
1016         size_t wsize = NFS_SERVER(inode)->wsize;
1017
1018         pnfs_pageio_init_write(pgio, inode);
1019
1020         if (wsize < PAGE_CACHE_SIZE)
1021                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1022         else
1023                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1024 }
1025
1026 /*
1027  * Handle a write reply that flushed part of a page.
1028  */
1029 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1030 {
1031         struct nfs_write_data   *data = calldata;
1032
1033         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1034                 task->tk_pid,
1035                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1036                 (long long)
1037                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1038                 data->req->wb_bytes, (long long)req_offset(data->req));
1039
1040         nfs_writeback_done(task, data);
1041 }
1042
1043 static void nfs_writeback_release_partial(void *calldata)
1044 {
1045         struct nfs_write_data   *data = calldata;
1046         struct nfs_page         *req = data->req;
1047         struct page             *page = req->wb_page;
1048         int status = data->task.tk_status;
1049
1050         if (status < 0) {
1051                 nfs_set_pageerror(page);
1052                 nfs_context_set_write_error(req->wb_context, status);
1053                 dprintk(", error = %d\n", status);
1054                 goto out;
1055         }
1056
1057         if (nfs_write_need_commit(data)) {
1058                 struct inode *inode = page->mapping->host;
1059
1060                 spin_lock(&inode->i_lock);
1061                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1062                         /* Do nothing we need to resend the writes */
1063                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1064                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1065                         dprintk(" defer commit\n");
1066                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1067                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1068                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1069                         dprintk(" server reboot detected\n");
1070                 }
1071                 spin_unlock(&inode->i_lock);
1072         } else
1073                 dprintk(" OK\n");
1074
1075 out:
1076         if (atomic_dec_and_test(&req->wb_complete))
1077                 nfs_writepage_release(req);
1078         nfs_writedata_release(calldata);
1079 }
1080
1081 #if defined(CONFIG_NFS_V4_1)
1082 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1083 {
1084         struct nfs_write_data *data = calldata;
1085
1086         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1087                                 &data->args.seq_args,
1088                                 &data->res.seq_res, 1, task))
1089                 return;
1090         rpc_call_start(task);
1091 }
1092 #endif /* CONFIG_NFS_V4_1 */
1093
1094 static const struct rpc_call_ops nfs_write_partial_ops = {
1095 #if defined(CONFIG_NFS_V4_1)
1096         .rpc_call_prepare = nfs_write_prepare,
1097 #endif /* CONFIG_NFS_V4_1 */
1098         .rpc_call_done = nfs_writeback_done_partial,
1099         .rpc_release = nfs_writeback_release_partial,
1100 };
1101
1102 /*
1103  * Handle a write reply that flushes a whole page.
1104  *
1105  * FIXME: There is an inherent race with invalidate_inode_pages and
1106  *        writebacks since the page->count is kept > 1 for as long
1107  *        as the page has a write request pending.
1108  */
1109 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1110 {
1111         struct nfs_write_data   *data = calldata;
1112
1113         nfs_writeback_done(task, data);
1114 }
1115
1116 static void nfs_writeback_release_full(void *calldata)
1117 {
1118         struct nfs_write_data   *data = calldata;
1119         int status = data->task.tk_status;
1120
1121         /* Update attributes as result of writeback. */
1122         while (!list_empty(&data->pages)) {
1123                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1124                 struct page *page = req->wb_page;
1125
1126                 nfs_list_remove_request(req);
1127
1128                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1129                         data->task.tk_pid,
1130                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1131                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1132                         req->wb_bytes,
1133                         (long long)req_offset(req));
1134
1135                 if (status < 0) {
1136                         nfs_set_pageerror(page);
1137                         nfs_context_set_write_error(req->wb_context, status);
1138                         dprintk(", error = %d\n", status);
1139                         goto remove_request;
1140                 }
1141
1142                 if (nfs_write_need_commit(data)) {
1143                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1144                         nfs_mark_request_commit(req);
1145                         dprintk(" marked for commit\n");
1146                         goto next;
1147                 }
1148                 dprintk(" OK\n");
1149 remove_request:
1150                 nfs_inode_remove_request(req);
1151         next:
1152                 nfs_clear_page_tag_locked(req);
1153                 nfs_end_page_writeback(page);
1154         }
1155         nfs_writedata_release(calldata);
1156 }
1157
1158 static const struct rpc_call_ops nfs_write_full_ops = {
1159 #if defined(CONFIG_NFS_V4_1)
1160         .rpc_call_prepare = nfs_write_prepare,
1161 #endif /* CONFIG_NFS_V4_1 */
1162         .rpc_call_done = nfs_writeback_done_full,
1163         .rpc_release = nfs_writeback_release_full,
1164 };
1165
1166
1167 /*
1168  * This function is called when the WRITE call is complete.
1169  */
1170 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1171 {
1172         struct nfs_writeargs    *argp = &data->args;
1173         struct nfs_writeres     *resp = &data->res;
1174         struct nfs_server       *server = NFS_SERVER(data->inode);
1175         int status;
1176
1177         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1178                 task->tk_pid, task->tk_status);
1179
1180         /*
1181          * ->write_done will attempt to use post-op attributes to detect
1182          * conflicting writes by other clients.  A strict interpretation
1183          * of close-to-open would allow us to continue caching even if
1184          * another writer had changed the file, but some applications
1185          * depend on tighter cache coherency when writing.
1186          */
1187         status = NFS_PROTO(data->inode)->write_done(task, data);
1188         if (status != 0)
1189                 return;
1190         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1191
1192 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1193         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1194                 /* We tried a write call, but the server did not
1195                  * commit data to stable storage even though we
1196                  * requested it.
1197                  * Note: There is a known bug in Tru64 < 5.0 in which
1198                  *       the server reports NFS_DATA_SYNC, but performs
1199                  *       NFS_FILE_SYNC. We therefore implement this checking
1200                  *       as a dprintk() in order to avoid filling syslog.
1201                  */
1202                 static unsigned long    complain;
1203
1204                 /* Note this will print the MDS for a DS write */
1205                 if (time_before(complain, jiffies)) {
1206                         dprintk("NFS:       faulty NFS server %s:"
1207                                 " (committed = %d) != (stable = %d)\n",
1208                                 server->nfs_client->cl_hostname,
1209                                 resp->verf->committed, argp->stable);
1210                         complain = jiffies + 300 * HZ;
1211                 }
1212         }
1213 #endif
1214         /* Is this a short write? */
1215         if (task->tk_status >= 0 && resp->count < argp->count) {
1216                 static unsigned long    complain;
1217
1218                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1219
1220                 /* Has the server at least made some progress? */
1221                 if (resp->count != 0) {
1222                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1223                         if (resp->verf->committed != NFS_UNSTABLE) {
1224                                 /* Resend from where the server left off */
1225                                 data->mds_offset += resp->count;
1226                                 argp->offset += resp->count;
1227                                 argp->pgbase += resp->count;
1228                                 argp->count -= resp->count;
1229                         } else {
1230                                 /* Resend as a stable write in order to avoid
1231                                  * headaches in the case of a server crash.
1232                                  */
1233                                 argp->stable = NFS_FILE_SYNC;
1234                         }
1235                         nfs_restart_rpc(task, server->nfs_client);
1236                         return;
1237                 }
1238                 if (time_before(complain, jiffies)) {
1239                         printk(KERN_WARNING
1240                                "NFS: Server wrote zero bytes, expected %u.\n",
1241                                         argp->count);
1242                         complain = jiffies + 300 * HZ;
1243                 }
1244                 /* Can't do anything about it except throw an error. */
1245                 task->tk_status = -EIO;
1246         }
1247         return;
1248 }
1249
1250
1251 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1252 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1253 {
1254         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1255                 return 1;
1256         if (may_wait && !out_of_line_wait_on_bit_lock(&nfsi->flags,
1257                                 NFS_INO_COMMIT, nfs_wait_bit_killable,
1258                                 TASK_KILLABLE))
1259                 return 1;
1260         return 0;
1261 }
1262
1263 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1264 {
1265         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1266         smp_mb__after_clear_bit();
1267         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1268 }
1269
1270
1271 static void nfs_commitdata_release(void *data)
1272 {
1273         struct nfs_write_data *wdata = data;
1274
1275         put_nfs_open_context(wdata->args.context);
1276         nfs_commit_free(wdata);
1277 }
1278
1279 /*
1280  * Set up the argument/result storage required for the RPC call.
1281  */
1282 static int nfs_commit_rpcsetup(struct list_head *head,
1283                 struct nfs_write_data *data,
1284                 int how)
1285 {
1286         struct nfs_page *first = nfs_list_entry(head->next);
1287         struct inode *inode = first->wb_context->path.dentry->d_inode;
1288         int priority = flush_task_priority(how);
1289         struct rpc_task *task;
1290         struct rpc_message msg = {
1291                 .rpc_argp = &data->args,
1292                 .rpc_resp = &data->res,
1293                 .rpc_cred = first->wb_context->cred,
1294         };
1295         struct rpc_task_setup task_setup_data = {
1296                 .task = &data->task,
1297                 .rpc_client = NFS_CLIENT(inode),
1298                 .rpc_message = &msg,
1299                 .callback_ops = &nfs_commit_ops,
1300                 .callback_data = data,
1301                 .workqueue = nfsiod_workqueue,
1302                 .flags = RPC_TASK_ASYNC,
1303                 .priority = priority,
1304         };
1305
1306         /* Set up the RPC argument and reply structs
1307          * NB: take care not to mess about with data->commit et al. */
1308
1309         list_splice_init(head, &data->pages);
1310
1311         data->inode       = inode;
1312         data->cred        = msg.rpc_cred;
1313
1314         data->args.fh     = NFS_FH(data->inode);
1315         /* Note: we always request a commit of the entire inode */
1316         data->args.offset = 0;
1317         data->args.count  = 0;
1318         data->args.context = get_nfs_open_context(first->wb_context);
1319         data->res.count   = 0;
1320         data->res.fattr   = &data->fattr;
1321         data->res.verf    = &data->verf;
1322         nfs_fattr_init(&data->fattr);
1323
1324         /* Set up the initial task struct.  */
1325         NFS_PROTO(inode)->commit_setup(data, &msg);
1326
1327         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1328
1329         task = rpc_run_task(&task_setup_data);
1330         if (IS_ERR(task))
1331                 return PTR_ERR(task);
1332         if (how & FLUSH_SYNC)
1333                 rpc_wait_for_completion_task(task);
1334         rpc_put_task(task);
1335         return 0;
1336 }
1337
1338 /*
1339  * Commit dirty pages
1340  */
1341 static int
1342 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1343 {
1344         struct nfs_write_data   *data;
1345         struct nfs_page         *req;
1346
1347         data = nfs_commitdata_alloc();
1348
1349         if (!data)
1350                 goto out_bad;
1351
1352         /* Set up the argument struct */
1353         return nfs_commit_rpcsetup(head, data, how);
1354  out_bad:
1355         while (!list_empty(head)) {
1356                 req = nfs_list_entry(head->next);
1357                 nfs_list_remove_request(req);
1358                 nfs_mark_request_commit(req);
1359                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1360                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1361                                 BDI_RECLAIMABLE);
1362                 nfs_clear_page_tag_locked(req);
1363         }
1364         nfs_commit_clear_lock(NFS_I(inode));
1365         return -ENOMEM;
1366 }
1367
1368 /*
1369  * COMMIT call returned
1370  */
1371 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1372 {
1373         struct nfs_write_data   *data = calldata;
1374
1375         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1376                                 task->tk_pid, task->tk_status);
1377
1378         /* Call the NFS version-specific code */
1379         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1380                 return;
1381 }
1382
1383 static void nfs_commit_release(void *calldata)
1384 {
1385         struct nfs_write_data   *data = calldata;
1386         struct nfs_page         *req;
1387         int status = data->task.tk_status;
1388
1389         while (!list_empty(&data->pages)) {
1390                 req = nfs_list_entry(data->pages.next);
1391                 nfs_list_remove_request(req);
1392                 nfs_clear_request_commit(req);
1393
1394                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1395                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1396                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1397                         req->wb_bytes,
1398                         (long long)req_offset(req));
1399                 if (status < 0) {
1400                         nfs_context_set_write_error(req->wb_context, status);
1401                         nfs_inode_remove_request(req);
1402                         dprintk(", error = %d\n", status);
1403                         goto next;
1404                 }
1405
1406                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1407                  * returned by the server against all stored verfs. */
1408                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1409                         /* We have a match */
1410                         nfs_inode_remove_request(req);
1411                         dprintk(" OK\n");
1412                         goto next;
1413                 }
1414                 /* We have a mismatch. Write the page again */
1415                 dprintk(" mismatch\n");
1416                 nfs_mark_request_dirty(req);
1417         next:
1418                 nfs_clear_page_tag_locked(req);
1419         }
1420         nfs_commit_clear_lock(NFS_I(data->inode));
1421         nfs_commitdata_release(calldata);
1422 }
1423
1424 static const struct rpc_call_ops nfs_commit_ops = {
1425 #if defined(CONFIG_NFS_V4_1)
1426         .rpc_call_prepare = nfs_write_prepare,
1427 #endif /* CONFIG_NFS_V4_1 */
1428         .rpc_call_done = nfs_commit_done,
1429         .rpc_release = nfs_commit_release,
1430 };
1431
1432 int nfs_commit_inode(struct inode *inode, int how)
1433 {
1434         LIST_HEAD(head);
1435         int may_wait = how & FLUSH_SYNC;
1436         int res = 0;
1437
1438         if (!nfs_commit_set_lock(NFS_I(inode), may_wait))
1439                 goto out_mark_dirty;
1440         spin_lock(&inode->i_lock);
1441         res = nfs_scan_commit(inode, &head, 0, 0);
1442         spin_unlock(&inode->i_lock);
1443         if (res) {
1444                 int error = nfs_commit_list(inode, &head, how);
1445                 if (error < 0)
1446                         return error;
1447                 if (may_wait)
1448                         wait_on_bit(&NFS_I(inode)->flags, NFS_INO_COMMIT,
1449                                         nfs_wait_bit_killable,
1450                                         TASK_KILLABLE);
1451                 else
1452                         goto out_mark_dirty;
1453         } else
1454                 nfs_commit_clear_lock(NFS_I(inode));
1455         return res;
1456         /* Note: If we exit without ensuring that the commit is complete,
1457          * we must mark the inode as dirty. Otherwise, future calls to
1458          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1459          * that the data is on the disk.
1460          */
1461 out_mark_dirty:
1462         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1463         return res;
1464 }
1465
1466 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1467 {
1468         struct nfs_inode *nfsi = NFS_I(inode);
1469         int flags = FLUSH_SYNC;
1470         int ret = 0;
1471
1472         if (wbc->sync_mode == WB_SYNC_NONE) {
1473                 /* Don't commit yet if this is a non-blocking flush and there
1474                  * are a lot of outstanding writes for this mapping.
1475                  */
1476                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1477                         goto out_mark_dirty;
1478
1479                 /* don't wait for the COMMIT response */
1480                 flags = 0;
1481         }
1482
1483         ret = nfs_commit_inode(inode, flags);
1484         if (ret >= 0) {
1485                 if (wbc->sync_mode == WB_SYNC_NONE) {
1486                         if (ret < wbc->nr_to_write)
1487                                 wbc->nr_to_write -= ret;
1488                         else
1489                                 wbc->nr_to_write = 0;
1490                 }
1491                 return 0;
1492         }
1493 out_mark_dirty:
1494         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1495         return ret;
1496 }
1497 #else
1498 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1499 {
1500         return 0;
1501 }
1502 #endif
1503
1504 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1505 {
1506         return nfs_commit_unstable_pages(inode, wbc);
1507 }
1508
1509 /*
1510  * flush the inode to disk.
1511  */
1512 int nfs_wb_all(struct inode *inode)
1513 {
1514         struct writeback_control wbc = {
1515                 .sync_mode = WB_SYNC_ALL,
1516                 .nr_to_write = LONG_MAX,
1517                 .range_start = 0,
1518                 .range_end = LLONG_MAX,
1519         };
1520
1521         return sync_inode(inode, &wbc);
1522 }
1523
1524 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1525 {
1526         struct nfs_page *req;
1527         int ret = 0;
1528
1529         BUG_ON(!PageLocked(page));
1530         for (;;) {
1531                 wait_on_page_writeback(page);
1532                 req = nfs_page_find_request(page);
1533                 if (req == NULL)
1534                         break;
1535                 if (nfs_lock_request_dontget(req)) {
1536                         nfs_inode_remove_request(req);
1537                         /*
1538                          * In case nfs_inode_remove_request has marked the
1539                          * page as being dirty
1540                          */
1541                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1542                         nfs_unlock_request(req);
1543                         break;
1544                 }
1545                 ret = nfs_wait_on_request(req);
1546                 nfs_release_request(req);
1547                 if (ret < 0)
1548                         break;
1549         }
1550         return ret;
1551 }
1552
1553 /*
1554  * Write back all requests on one page - we do this before reading it.
1555  */
1556 int nfs_wb_page(struct inode *inode, struct page *page)
1557 {
1558         loff_t range_start = page_offset(page);
1559         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1560         struct writeback_control wbc = {
1561                 .sync_mode = WB_SYNC_ALL,
1562                 .nr_to_write = 0,
1563                 .range_start = range_start,
1564                 .range_end = range_end,
1565         };
1566         int ret;
1567
1568         for (;;) {
1569                 wait_on_page_writeback(page);
1570                 if (clear_page_dirty_for_io(page)) {
1571                         ret = nfs_writepage_locked(page, &wbc);
1572                         if (ret < 0)
1573                                 goto out_error;
1574                         continue;
1575                 }
1576                 if (!PagePrivate(page))
1577                         break;
1578                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1579                 if (ret < 0)
1580                         goto out_error;
1581         }
1582         return 0;
1583 out_error:
1584         return ret;
1585 }
1586
1587 #ifdef CONFIG_MIGRATION
1588 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1589                 struct page *page)
1590 {
1591         struct nfs_page *req;
1592         int ret;
1593
1594         nfs_fscache_release_page(page, GFP_KERNEL);
1595
1596         req = nfs_find_and_lock_request(page, false);
1597         ret = PTR_ERR(req);
1598         if (IS_ERR(req))
1599                 goto out;
1600
1601         ret = migrate_page(mapping, newpage, page);
1602         if (!req)
1603                 goto out;
1604         if (ret)
1605                 goto out_unlock;
1606         page_cache_get(newpage);
1607         spin_lock(&mapping->host->i_lock);
1608         req->wb_page = newpage;
1609         SetPagePrivate(newpage);
1610         set_page_private(newpage, (unsigned long)req);
1611         ClearPagePrivate(page);
1612         set_page_private(page, 0);
1613         spin_unlock(&mapping->host->i_lock);
1614         page_cache_release(page);
1615 out_unlock:
1616         nfs_clear_page_tag_locked(req);
1617 out:
1618         return ret;
1619 }
1620 #endif
1621
1622 int __init nfs_init_writepagecache(void)
1623 {
1624         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1625                                              sizeof(struct nfs_write_data),
1626                                              0, SLAB_HWCACHE_ALIGN,
1627                                              NULL);
1628         if (nfs_wdata_cachep == NULL)
1629                 return -ENOMEM;
1630
1631         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1632                                                      nfs_wdata_cachep);
1633         if (nfs_wdata_mempool == NULL)
1634                 return -ENOMEM;
1635
1636         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1637                                                       nfs_wdata_cachep);
1638         if (nfs_commit_mempool == NULL)
1639                 return -ENOMEM;
1640
1641         /*
1642          * NFS congestion size, scale with available memory.
1643          *
1644          *  64MB:    8192k
1645          * 128MB:   11585k
1646          * 256MB:   16384k
1647          * 512MB:   23170k
1648          *   1GB:   32768k
1649          *   2GB:   46340k
1650          *   4GB:   65536k
1651          *   8GB:   92681k
1652          *  16GB:  131072k
1653          *
1654          * This allows larger machines to have larger/more transfers.
1655          * Limit the default to 256M
1656          */
1657         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1658         if (nfs_congestion_kb > 256*1024)
1659                 nfs_congestion_kb = 256*1024;
1660
1661         return 0;
1662 }
1663
1664 void nfs_destroy_writepagecache(void)
1665 {
1666         mempool_destroy(nfs_commit_mempool);
1667         mempool_destroy(nfs_wdata_mempool);
1668         kmem_cache_destroy(nfs_wdata_cachep);
1669 }
1670