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, struct pnfs_layout_segment *lseg)
 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         pnfs_mark_request_commit(req, lseg);
 457         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 458         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 459         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 460 }
 461
 462 static int
 463 nfs_clear_request_commit(struct nfs_page *req)
 464 {
 465         struct page *page = req->wb_page;
 466
 467         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 468                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 469                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 470                 return 1;
 471         }
 472         return 0;
 473 }
 474
 475 static inline
 476 int nfs_write_need_commit(struct nfs_write_data *data)
 477 {
 478         if (data->verf.committed == NFS_DATA_SYNC)
 479                 return data->lseg == NULL;
 480         else
 481                 return data->verf.committed != NFS_FILE_SYNC;
 482 }
 483
 484 static inline
 485 int nfs_reschedule_unstable_write(struct nfs_page *req,
 486                                   struct nfs_write_data *data)
 487 {
 488         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 489                 nfs_mark_request_commit(req, data->lseg);
 490                 return 1;
 491         }
 492         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 493                 nfs_mark_request_dirty(req);
 494                 return 1;
 495         }
 496         return 0;
 497 }
 498 #else
 499 static inline void
 500 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 501 {
 502 }
 503
 504 static inline int
 505 nfs_clear_request_commit(struct nfs_page *req)
 506 {
 507         return 0;
 508 }
 509
 510 static inline
 511 int nfs_write_need_commit(struct nfs_write_data *data)
 512 {
 513         return 0;
 514 }
 515
 516 static inline
 517 int nfs_reschedule_unstable_write(struct nfs_page *req,
 518                                   struct nfs_write_data *data)
 519 {
 520         return 0;
 521 }
 522 #endif
 523
 524 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 525 static int
 526 nfs_need_commit(struct nfs_inode *nfsi)
 527 {
 528         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 529 }
 530
 531 /*
 532  * nfs_scan_commit - Scan an inode for commit requests
 533  * @inode: NFS inode to scan
 534  * @dst: destination list
 535  * @idx_start: lower bound of page->index to scan.
 536  * @npages: idx_start + npages sets the upper bound to scan.
 537  *
 538  * Moves requests from the inode's 'commit' request list.
 539  * The requests are *not* checked to ensure that they form a contiguous set.
 540  */
 541 static int
 542 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 543 {
 544         struct nfs_inode *nfsi = NFS_I(inode);
 545         int ret;
 546
 547         if (!nfs_need_commit(nfsi))
 548                 return 0;
 549
 550         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 551         if (ret > 0)
 552                 nfsi->ncommit -= ret;
 553         if (nfs_need_commit(NFS_I(inode)))
 554                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 555         return ret;
 556 }
 557 #else
 558 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 559 {
 560         return 0;
 561 }
 562
 563 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 564 {
 565         return 0;
 566 }
 567 #endif
 568
 569 /*
 570  * Search for an existing write request, and attempt to update
 571  * it to reflect a new dirty region on a given page.
 572  *
 573  * If the attempt fails, then the existing request is flushed out
 574  * to disk.
 575  */
 576 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 577                 struct page *page,
 578                 unsigned int offset,
 579                 unsigned int bytes)
 580 {
 581         struct nfs_page *req;
 582         unsigned int rqend;
 583         unsigned int end;
 584         int error;
 585
 586         if (!PagePrivate(page))
 587                 return NULL;
 588
 589         end = offset + bytes;
 590         spin_lock(&inode->i_lock);
 591
 592         for (;;) {
 593                 req = nfs_page_find_request_locked(page);
 594                 if (req == NULL)
 595                         goto out_unlock;
 596
 597                 rqend = req->wb_offset + req->wb_bytes;
 598                 /*
 599                  * Tell the caller to flush out the request if
 600                  * the offsets are non-contiguous.
 601                  * Note: nfs_flush_incompatible() will already
 602                  * have flushed out requests having wrong owners.
 603                  */
 604                 if (offset > rqend
 605                     || end < req->wb_offset)
 606                         goto out_flushme;
 607
 608                 if (nfs_set_page_tag_locked(req))
 609                         break;
 610
 611                 /* The request is locked, so wait and then retry */
 612                 spin_unlock(&inode->i_lock);
 613                 error = nfs_wait_on_request(req);
 614                 nfs_release_request(req);
 615                 if (error != 0)
 616                         goto out_err;
 617                 spin_lock(&inode->i_lock);
 618         }
 619
 620         if (nfs_clear_request_commit(req) &&
 621             radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 622                                  req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
 623                 NFS_I(inode)->ncommit--;
 624                 pnfs_clear_request_commit(req);
 625         }
 626
 627         /* Okay, the request matches. Update the region */
 628         if (offset < req->wb_offset) {
 629                 req->wb_offset = offset;
 630                 req->wb_pgbase = offset;
 631         }
 632         if (end > rqend)
 633                 req->wb_bytes = end - req->wb_offset;
 634         else
 635                 req->wb_bytes = rqend - req->wb_offset;
 636 out_unlock:
 637         spin_unlock(&inode->i_lock);
 638         return req;
 639 out_flushme:
 640         spin_unlock(&inode->i_lock);
 641         nfs_release_request(req);
 642         error = nfs_wb_page(inode, page);
 643 out_err:
 644         return ERR_PTR(error);
 645 }
 646
 647 /*
 648  * Try to update an existing write request, or create one if there is none.
 649  *
 650  * Note: Should always be called with the Page Lock held to prevent races
 651  * if we have to add a new request. Also assumes that the caller has
 652  * already called nfs_flush_incompatible() if necessary.
 653  */
 654 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 655                 struct page *page, unsigned int offset, unsigned int bytes)
 656 {
 657         struct inode *inode = page->mapping->host;
 658         struct nfs_page *req;
 659         int error;
 660
 661         req = nfs_try_to_update_request(inode, page, offset, bytes);
 662         if (req != NULL)
 663                 goto out;
 664         req = nfs_create_request(ctx, inode, page, offset, bytes);
 665         if (IS_ERR(req))
 666                 goto out;
 667         error = nfs_inode_add_request(inode, req);
 668         if (error != 0) {
 669                 nfs_release_request(req);
 670                 req = ERR_PTR(error);
 671         }
 672 out:
 673         return req;
 674 }
 675
 676 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 677                 unsigned int offset, unsigned int count)
 678 {
 679         struct nfs_page *req;
 680
 681         req = nfs_setup_write_request(ctx, page, offset, count);
 682         if (IS_ERR(req))
 683                 return PTR_ERR(req);
 684         nfs_mark_request_dirty(req);
 685         /* Update file length */
 686         nfs_grow_file(page, offset, count);
 687         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 688         nfs_mark_request_dirty(req);
 689         nfs_clear_page_tag_locked(req);
 690         return 0;
 691 }
 692
 693 int nfs_flush_incompatible(struct file *file, struct page *page)
 694 {
 695         struct nfs_open_context *ctx = nfs_file_open_context(file);
 696         struct nfs_page *req;
 697         int do_flush, status;
 698         /*
 699          * Look for a request corresponding to this page. If there
 700          * is one, and it belongs to another file, we flush it out
 701          * before we try to copy anything into the page. Do this
 702          * due to the lack of an ACCESS-type call in NFSv2.
 703          * Also do the same if we find a request from an existing
 704          * dropped page.
 705          */
 706         do {
 707                 req = nfs_page_find_request(page);
 708                 if (req == NULL)
 709                         return 0;
 710                 do_flush = req->wb_page != page || req->wb_context != ctx ||
 711                         req->wb_lock_context->lockowner != current->files ||
 712                         req->wb_lock_context->pid != current->tgid;
 713                 nfs_release_request(req);
 714                 if (!do_flush)
 715                         return 0;
 716                 status = nfs_wb_page(page->mapping->host, page);
 717         } while (status == 0);
 718         return status;
 719 }
 720
 721 /*
 722  * If the page cache is marked as unsafe or invalid, then we can't rely on
 723  * the PageUptodate() flag. In this case, we will need to turn off
 724  * write optimisations that depend on the page contents being correct.
 725  */
 726 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 727 {
 728         return PageUptodate(page) &&
 729                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 730 }
 731
 732 /*
 733  * Update and possibly write a cached page of an NFS file.
 734  *
 735  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 736  * things with a page scheduled for an RPC call (e.g. invalidate it).
 737  */
 738 int nfs_updatepage(struct file *file, struct page *page,
 739                 unsigned int offset, unsigned int count)
 740 {
 741         struct nfs_open_context *ctx = nfs_file_open_context(file);
 742         struct inode    *inode = page->mapping->host;
 743         int             status = 0;
 744
 745         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 746
 747         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 748                 file->f_path.dentry->d_parent->d_name.name,
 749                 file->f_path.dentry->d_name.name, count,
 750                 (long long)(page_offset(page) + offset));
 751
 752         /* If we're not using byte range locks, and we know the page
 753          * is up to date, it may be more efficient to extend the write
 754          * to cover the entire page in order to avoid fragmentation
 755          * inefficiencies.
 756          */
 757         if (nfs_write_pageuptodate(page, inode) &&
 758                         inode->i_flock == NULL &&
 759                         !(file->f_flags & O_DSYNC)) {
 760                 count = max(count + offset, nfs_page_length(page));
 761                 offset = 0;
 762         }
 763
 764         status = nfs_writepage_setup(ctx, page, offset, count);
 765         if (status < 0)
 766                 nfs_set_pageerror(page);
 767
 768         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 769                         status, (long long)i_size_read(inode));
 770         return status;
 771 }
 772
 773 static void nfs_writepage_release(struct nfs_page *req,
 774                                   struct nfs_write_data *data)
 775 {
 776         struct page *page = req->wb_page;
 777
 778         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
 779                 nfs_inode_remove_request(req);
 780         nfs_clear_page_tag_locked(req);
 781         nfs_end_page_writeback(page);
 782 }
 783
 784 static int flush_task_priority(int how)
 785 {
 786         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 787                 case FLUSH_HIGHPRI:
 788                         return RPC_PRIORITY_HIGH;
 789                 case FLUSH_LOWPRI:
 790                         return RPC_PRIORITY_LOW;
 791         }
 792         return RPC_PRIORITY_NORMAL;
 793 }
 794
 795 int nfs_initiate_write(struct nfs_write_data *data,
 796                        struct rpc_clnt *clnt,
 797                        const struct rpc_call_ops *call_ops,
 798                        int how)
 799 {
 800         struct inode *inode = data->inode;
 801         int priority = flush_task_priority(how);
 802         struct rpc_task *task;
 803         struct rpc_message msg = {
 804                 .rpc_argp = &data->args,
 805                 .rpc_resp = &data->res,
 806                 .rpc_cred = data->cred,
 807         };
 808         struct rpc_task_setup task_setup_data = {
 809                 .rpc_client = clnt,
 810                 .task = &data->task,
 811                 .rpc_message = &msg,
 812                 .callback_ops = call_ops,
 813                 .callback_data = data,
 814                 .workqueue = nfsiod_workqueue,
 815                 .flags = RPC_TASK_ASYNC,
 816                 .priority = priority,
 817         };
 818         int ret = 0;
 819
 820         /* Set up the initial task struct.  */
 821         NFS_PROTO(inode)->write_setup(data, &msg);
 822
 823         dprintk("NFS: %5u initiated write call "
 824                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 825                 data->task.tk_pid,
 826                 inode->i_sb->s_id,
 827                 (long long)NFS_FILEID(inode),
 828                 data->args.count,
 829                 (unsigned long long)data->args.offset);
 830
 831         task = rpc_run_task(&task_setup_data);
 832         if (IS_ERR(task)) {
 833                 ret = PTR_ERR(task);
 834                 goto out;
 835         }
 836         if (how & FLUSH_SYNC) {
 837                 ret = rpc_wait_for_completion_task(task);
 838                 if (ret == 0)
 839                         ret = task->tk_status;
 840         }
 841         rpc_put_task(task);
 842 out:
 843         return ret;
 844 }
 845 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 846
 847 /*
 848  * Set up the argument/result storage required for the RPC call.
 849  */
 850 static int nfs_write_rpcsetup(struct nfs_page *req,
 851                 struct nfs_write_data *data,
 852                 const struct rpc_call_ops *call_ops,
 853                 unsigned int count, unsigned int offset,
 854                 struct pnfs_layout_segment *lseg,
 855                 int how)
 856 {
 857         struct inode *inode = req->wb_context->path.dentry->d_inode;
 858
 859         /* Set up the RPC argument and reply structs
 860          * NB: take care not to mess about with data->commit et al. */
 861
 862         data->req = req;
 863         data->inode = inode = req->wb_context->path.dentry->d_inode;
 864         data->cred = req->wb_context->cred;
 865         data->lseg = get_lseg(lseg);
 866
 867         data->args.fh     = NFS_FH(inode);
 868         data->args.offset = req_offset(req) + offset;
 869         data->args.pgbase = req->wb_pgbase + offset;
 870         data->args.pages  = data->pagevec;
 871         data->args.count  = count;
 872         data->args.context = get_nfs_open_context(req->wb_context);
 873         data->args.lock_context = req->wb_lock_context;
 874         data->args.stable  = NFS_UNSTABLE;
 875         if (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 876                 data->args.stable = NFS_DATA_SYNC;
 877                 if (!nfs_need_commit(NFS_I(inode)))
 878                         data->args.stable = NFS_FILE_SYNC;
 879         }
 880
 881         data->res.fattr   = &data->fattr;
 882         data->res.count   = count;
 883         data->res.verf    = &data->verf;
 884         nfs_fattr_init(&data->fattr);
 885
 886         if (data->lseg &&
 887             (pnfs_try_to_write_data(data, call_ops, how) == PNFS_ATTEMPTED))
 888                 return 0;
 889
 890         return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
 891 }
 892
 893 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 894  * call this on each, which will prepare them to be retried on next
 895  * writeback using standard nfs.
 896  */
 897 static void nfs_redirty_request(struct nfs_page *req)
 898 {
 899         struct page *page = req->wb_page;
 900
 901         nfs_mark_request_dirty(req);
 902         nfs_clear_page_tag_locked(req);
 903         nfs_end_page_writeback(page);
 904 }
 905
 906 /*
 907  * Generate multiple small requests to write out a single
 908  * contiguous dirty area on one page.
 909  */
 910 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc)
 911 {
 912         struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
 913         struct page *page = req->wb_page;
 914         struct nfs_write_data *data;
 915         size_t wsize = NFS_SERVER(desc->pg_inode)->wsize, nbytes;
 916         unsigned int offset;
 917         int requests = 0;
 918         int ret = 0;
 919         struct pnfs_layout_segment *lseg;
 920         LIST_HEAD(list);
 921
 922         nfs_list_remove_request(req);
 923
 924         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 925             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
 926              desc->pg_count > wsize))
 927                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 928
 929
 930         nbytes = desc->pg_count;
 931         do {
 932                 size_t len = min(nbytes, wsize);
 933
 934                 data = nfs_writedata_alloc(1);
 935                 if (!data)
 936                         goto out_bad;
 937                 list_add(&data->pages, &list);
 938                 requests++;
 939                 nbytes -= len;
 940         } while (nbytes != 0);
 941         atomic_set(&req->wb_complete, requests);
 942
 943         BUG_ON(desc->pg_lseg);
 944         lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
 945         ClearPageError(page);
 946         offset = 0;
 947         nbytes = desc->pg_count;
 948         do {
 949                 int ret2;
 950
 951                 data = list_entry(list.next, struct nfs_write_data, pages);
 952                 list_del_init(&data->pages);
 953
 954                 data->pagevec[0] = page;
 955
 956                 if (nbytes < wsize)
 957                         wsize = nbytes;
 958                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 959                                           wsize, offset, lseg, desc->pg_ioflags);
 960                 if (ret == 0)
 961                         ret = ret2;
 962                 offset += wsize;
 963                 nbytes -= wsize;
 964         } while (nbytes != 0);
 965
 966         put_lseg(lseg);
 967         desc->pg_lseg = NULL;
 968         return ret;
 969
 970 out_bad:
 971         while (!list_empty(&list)) {
 972                 data = list_entry(list.next, struct nfs_write_data, pages);
 973                 list_del(&data->pages);
 974                 nfs_writedata_free(data);
 975         }
 976         nfs_redirty_request(req);
 977         return -ENOMEM;
 978 }
 979
 980 /*
 981  * Create an RPC task for the given write request and kick it.
 982  * The page must have been locked by the caller.
 983  *
 984  * It may happen that the page we're passed is not marked dirty.
 985  * This is the case if nfs_updatepage detects a conflicting request
 986  * that has been written but not committed.
 987  */
 988 static int nfs_flush_one(struct nfs_pageio_descriptor *desc)
 989 {
 990         struct nfs_page         *req;
 991         struct page             **pages;
 992         struct nfs_write_data   *data;
 993         struct list_head *head = &desc->pg_list;
 994         struct pnfs_layout_segment *lseg = desc->pg_lseg;
 995         int ret;
 996
 997         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
 998                                                       desc->pg_count));
 999         if (!data) {
1000                 while (!list_empty(head)) {
1001                         req = nfs_list_entry(head->next);
1002                         nfs_list_remove_request(req);
1003                         nfs_redirty_request(req);
1004                 }
1005                 ret = -ENOMEM;
1006                 goto out;
1007         }
1008         pages = data->pagevec;
1009         while (!list_empty(head)) {
1010                 req = nfs_list_entry(head->next);
1011                 nfs_list_remove_request(req);
1012                 nfs_list_add_request(req, &data->pages);
1013                 ClearPageError(req->wb_page);
1014                 *pages++ = req->wb_page;
1015         }
1016         req = nfs_list_entry(data->pages.next);
1017         if ((!lseg) && list_is_singular(&data->pages))
1018                 lseg = pnfs_update_layout(desc->pg_inode, req->wb_context, IOMODE_RW);
1019
1020         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1021             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1022                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1023
1024         /* Set up the argument struct */
1025         ret = nfs_write_rpcsetup(req, data, &nfs_write_full_ops, desc->pg_count, 0, lseg, desc->pg_ioflags);
1026 out:
1027         put_lseg(lseg); /* Cleans any gotten in ->pg_test */
1028         desc->pg_lseg = NULL;
1029         return ret;
1030 }
1031
1032 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1033                                   struct inode *inode, int ioflags)
1034 {
1035         size_t wsize = NFS_SERVER(inode)->wsize;
1036
1037         pnfs_pageio_init_write(pgio, inode);
1038
1039         if (wsize < PAGE_CACHE_SIZE)
1040                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
1041         else
1042                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
1043 }
1044
1045 /*
1046  * Handle a write reply that flushed part of a page.
1047  */
1048 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1049 {
1050         struct nfs_write_data   *data = calldata;
1051
1052         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1053                 task->tk_pid,
1054                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1055                 (long long)
1056                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1057                 data->req->wb_bytes, (long long)req_offset(data->req));
1058
1059         nfs_writeback_done(task, data);
1060 }
1061
1062 static void nfs_writeback_release_partial(void *calldata)
1063 {
1064         struct nfs_write_data   *data = calldata;
1065         struct nfs_page         *req = data->req;
1066         struct page             *page = req->wb_page;
1067         int status = data->task.tk_status;
1068
1069         if (status < 0) {
1070                 nfs_set_pageerror(page);
1071                 nfs_context_set_write_error(req->wb_context, status);
1072                 dprintk(", error = %d\n", status);
1073                 goto out;
1074         }
1075
1076         if (nfs_write_need_commit(data)) {
1077                 struct inode *inode = page->mapping->host;
1078
1079                 spin_lock(&inode->i_lock);
1080                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1081                         /* Do nothing we need to resend the writes */
1082                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1083                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1084                         dprintk(" defer commit\n");
1085                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1086                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1087                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1088                         dprintk(" server reboot detected\n");
1089                 }
1090                 spin_unlock(&inode->i_lock);
1091         } else
1092                 dprintk(" OK\n");
1093
1094 out:
1095         if (atomic_dec_and_test(&req->wb_complete))
1096                 nfs_writepage_release(req, data);
1097         nfs_writedata_release(calldata);
1098 }
1099
1100 #if defined(CONFIG_NFS_V4_1)
1101 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1102 {
1103         struct nfs_write_data *data = calldata;
1104
1105         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1106                                 &data->args.seq_args,
1107                                 &data->res.seq_res, 1, task))
1108                 return;
1109         rpc_call_start(task);
1110 }
1111 #endif /* CONFIG_NFS_V4_1 */
1112
1113 static const struct rpc_call_ops nfs_write_partial_ops = {
1114 #if defined(CONFIG_NFS_V4_1)
1115         .rpc_call_prepare = nfs_write_prepare,
1116 #endif /* CONFIG_NFS_V4_1 */
1117         .rpc_call_done = nfs_writeback_done_partial,
1118         .rpc_release = nfs_writeback_release_partial,
1119 };
1120
1121 /*
1122  * Handle a write reply that flushes a whole page.
1123  *
1124  * FIXME: There is an inherent race with invalidate_inode_pages and
1125  *        writebacks since the page->count is kept > 1 for as long
1126  *        as the page has a write request pending.
1127  */
1128 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1129 {
1130         struct nfs_write_data   *data = calldata;
1131
1132         nfs_writeback_done(task, data);
1133 }
1134
1135 static void nfs_writeback_release_full(void *calldata)
1136 {
1137         struct nfs_write_data   *data = calldata;
1138         int status = data->task.tk_status;
1139
1140         /* Update attributes as result of writeback. */
1141         while (!list_empty(&data->pages)) {
1142                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1143                 struct page *page = req->wb_page;
1144
1145                 nfs_list_remove_request(req);
1146
1147                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1148                         data->task.tk_pid,
1149                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1150                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1151                         req->wb_bytes,
1152                         (long long)req_offset(req));
1153
1154                 if (status < 0) {
1155                         nfs_set_pageerror(page);
1156                         nfs_context_set_write_error(req->wb_context, status);
1157                         dprintk(", error = %d\n", status);
1158                         goto remove_request;
1159                 }
1160
1161                 if (nfs_write_need_commit(data)) {
1162                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1163                         nfs_mark_request_commit(req, data->lseg);
1164                         dprintk(" marked for commit\n");
1165                         goto next;
1166                 }
1167                 dprintk(" OK\n");
1168 remove_request:
1169                 nfs_inode_remove_request(req);
1170         next:
1171                 nfs_clear_page_tag_locked(req);
1172                 nfs_end_page_writeback(page);
1173         }
1174         nfs_writedata_release(calldata);
1175 }
1176
1177 static const struct rpc_call_ops nfs_write_full_ops = {
1178 #if defined(CONFIG_NFS_V4_1)
1179         .rpc_call_prepare = nfs_write_prepare,
1180 #endif /* CONFIG_NFS_V4_1 */
1181         .rpc_call_done = nfs_writeback_done_full,
1182         .rpc_release = nfs_writeback_release_full,
1183 };
1184
1185
1186 /*
1187  * This function is called when the WRITE call is complete.
1188  */
1189 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1190 {
1191         struct nfs_writeargs    *argp = &data->args;
1192         struct nfs_writeres     *resp = &data->res;
1193         struct nfs_server       *server = NFS_SERVER(data->inode);
1194         int status;
1195
1196         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1197                 task->tk_pid, task->tk_status);
1198
1199         /*
1200          * ->write_done will attempt to use post-op attributes to detect
1201          * conflicting writes by other clients.  A strict interpretation
1202          * of close-to-open would allow us to continue caching even if
1203          * another writer had changed the file, but some applications
1204          * depend on tighter cache coherency when writing.
1205          */
1206         status = NFS_PROTO(data->inode)->write_done(task, data);
1207         if (status != 0)
1208                 return;
1209         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1210
1211 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1212         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1213                 /* We tried a write call, but the server did not
1214                  * commit data to stable storage even though we
1215                  * requested it.
1216                  * Note: There is a known bug in Tru64 < 5.0 in which
1217                  *       the server reports NFS_DATA_SYNC, but performs
1218                  *       NFS_FILE_SYNC. We therefore implement this checking
1219                  *       as a dprintk() in order to avoid filling syslog.
1220                  */
1221                 static unsigned long    complain;
1222
1223                 /* Note this will print the MDS for a DS write */
1224                 if (time_before(complain, jiffies)) {
1225                         dprintk("NFS:       faulty NFS server %s:"
1226                                 " (committed = %d) != (stable = %d)\n",
1227                                 server->nfs_client->cl_hostname,
1228                                 resp->verf->committed, argp->stable);
1229                         complain = jiffies + 300 * HZ;
1230                 }
1231         }
1232 #endif
1233         /* Is this a short write? */
1234         if (task->tk_status >= 0 && resp->count < argp->count) {
1235                 static unsigned long    complain;
1236
1237                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1238
1239                 /* Has the server at least made some progress? */
1240                 if (resp->count != 0) {
1241                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1242                         if (resp->verf->committed != NFS_UNSTABLE) {
1243                                 /* Resend from where the server left off */
1244                                 data->mds_offset += resp->count;
1245                                 argp->offset += resp->count;
1246                                 argp->pgbase += resp->count;
1247                                 argp->count -= resp->count;
1248                         } else {
1249                                 /* Resend as a stable write in order to avoid
1250                                  * headaches in the case of a server crash.
1251                                  */
1252                                 argp->stable = NFS_FILE_SYNC;
1253                         }
1254                         nfs_restart_rpc(task, server->nfs_client);
1255                         return;
1256                 }
1257                 if (time_before(complain, jiffies)) {
1258                         printk(KERN_WARNING
1259                                "NFS: Server wrote zero bytes, expected %u.\n",
1260                                         argp->count);
1261                         complain = jiffies + 300 * HZ;
1262                 }
1263                 /* Can't do anything about it except throw an error. */
1264                 task->tk_status = -EIO;
1265         }
1266         return;
1267 }
1268
1269
1270 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1271 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1272 {
1273         int ret;
1274
1275         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1276                 return 1;
1277         if (!may_wait)
1278                 return 0;
1279         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1280                                 NFS_INO_COMMIT,
1281                                 nfs_wait_bit_killable,
1282                                 TASK_KILLABLE);
1283         return (ret < 0) ? ret : 1;
1284 }
1285
1286 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1287 {
1288         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1289         smp_mb__after_clear_bit();
1290         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1291 }
1292
1293
1294 static void nfs_commitdata_release(void *data)
1295 {
1296         struct nfs_write_data *wdata = data;
1297
1298         put_lseg(wdata->lseg);
1299         put_nfs_open_context(wdata->args.context);
1300         nfs_commit_free(wdata);
1301 }
1302
1303 static int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1304                         const struct rpc_call_ops *call_ops,
1305                         int how)
1306 {
1307         struct rpc_task *task;
1308         int priority = flush_task_priority(how);
1309         struct rpc_message msg = {
1310                 .rpc_argp = &data->args,
1311                 .rpc_resp = &data->res,
1312                 .rpc_cred = data->cred,
1313         };
1314         struct rpc_task_setup task_setup_data = {
1315                 .task = &data->task,
1316                 .rpc_client = clnt,
1317                 .rpc_message = &msg,
1318                 .callback_ops = call_ops,
1319                 .callback_data = data,
1320                 .workqueue = nfsiod_workqueue,
1321                 .flags = RPC_TASK_ASYNC,
1322                 .priority = priority,
1323         };
1324         /* Set up the initial task struct.  */
1325         NFS_PROTO(data->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  * Set up the argument/result storage required for the RPC call.
1340  */
1341 static void nfs_init_commit(struct nfs_write_data *data,
1342                             struct list_head *head,
1343                             struct pnfs_layout_segment *lseg)
1344 {
1345         struct nfs_page *first = nfs_list_entry(head->next);
1346         struct inode *inode = first->wb_context->path.dentry->d_inode;
1347
1348         /* Set up the RPC argument and reply structs
1349          * NB: take care not to mess about with data->commit et al. */
1350
1351         list_splice_init(head, &data->pages);
1352
1353         data->inode       = inode;
1354         data->cred        = first->wb_context->cred;
1355         data->lseg        = lseg; /* reference transferred */
1356         data->mds_ops     = &nfs_commit_ops;
1357
1358         data->args.fh     = NFS_FH(data->inode);
1359         /* Note: we always request a commit of the entire inode */
1360         data->args.offset = 0;
1361         data->args.count  = 0;
1362         data->args.context = get_nfs_open_context(first->wb_context);
1363         data->res.count   = 0;
1364         data->res.fattr   = &data->fattr;
1365         data->res.verf    = &data->verf;
1366         nfs_fattr_init(&data->fattr);
1367 }
1368
1369 static void nfs_retry_commit(struct list_head *page_list,
1370                       struct pnfs_layout_segment *lseg)
1371 {
1372         struct nfs_page *req;
1373
1374         while (!list_empty(page_list)) {
1375                 req = nfs_list_entry(page_list->next);
1376                 nfs_list_remove_request(req);
1377                 nfs_mark_request_commit(req, lseg);
1378                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1379                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1380                              BDI_RECLAIMABLE);
1381                 nfs_clear_page_tag_locked(req);
1382         }
1383 }
1384
1385 /*
1386  * Commit dirty pages
1387  */
1388 static int
1389 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1390 {
1391         struct nfs_write_data   *data;
1392
1393         data = nfs_commitdata_alloc();
1394
1395         if (!data)
1396                 goto out_bad;
1397
1398         /* Set up the argument struct */
1399         nfs_init_commit(data, head, NULL);
1400         return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1401  out_bad:
1402         nfs_retry_commit(head, NULL);
1403         nfs_commit_clear_lock(NFS_I(inode));
1404         return -ENOMEM;
1405 }
1406
1407 /*
1408  * COMMIT call returned
1409  */
1410 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1411 {
1412         struct nfs_write_data   *data = calldata;
1413
1414         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1415                                 task->tk_pid, task->tk_status);
1416
1417         /* Call the NFS version-specific code */
1418         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1419                 return;
1420 }
1421
1422 static void nfs_commit_release_pages(struct nfs_write_data *data)
1423 {
1424         struct nfs_page *req;
1425         int status = data->task.tk_status;
1426
1427         while (!list_empty(&data->pages)) {
1428                 req = nfs_list_entry(data->pages.next);
1429                 nfs_list_remove_request(req);
1430                 nfs_clear_request_commit(req);
1431
1432                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1433                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1434                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1435                         req->wb_bytes,
1436                         (long long)req_offset(req));
1437                 if (status < 0) {
1438                         nfs_context_set_write_error(req->wb_context, status);
1439                         nfs_inode_remove_request(req);
1440                         dprintk(", error = %d\n", status);
1441                         goto next;
1442                 }
1443
1444                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1445                  * returned by the server against all stored verfs. */
1446                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1447                         /* We have a match */
1448                         nfs_inode_remove_request(req);
1449                         dprintk(" OK\n");
1450                         goto next;
1451                 }
1452                 /* We have a mismatch. Write the page again */
1453                 dprintk(" mismatch\n");
1454                 nfs_mark_request_dirty(req);
1455         next:
1456                 nfs_clear_page_tag_locked(req);
1457         }
1458 }
1459
1460 static void nfs_commit_release(void *calldata)
1461 {
1462         struct nfs_write_data *data = calldata;
1463
1464         nfs_commit_release_pages(data);
1465         nfs_commit_clear_lock(NFS_I(data->inode));
1466         nfs_commitdata_release(calldata);
1467 }
1468
1469 static const struct rpc_call_ops nfs_commit_ops = {
1470 #if defined(CONFIG_NFS_V4_1)
1471         .rpc_call_prepare = nfs_write_prepare,
1472 #endif /* CONFIG_NFS_V4_1 */
1473         .rpc_call_done = nfs_commit_done,
1474         .rpc_release = nfs_commit_release,
1475 };
1476
1477 int nfs_commit_inode(struct inode *inode, int how)
1478 {
1479         LIST_HEAD(head);
1480         int may_wait = how & FLUSH_SYNC;
1481         int res;
1482
1483         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1484         if (res <= 0)
1485                 goto out_mark_dirty;
1486         spin_lock(&inode->i_lock);
1487         res = nfs_scan_commit(inode, &head, 0, 0);
1488         spin_unlock(&inode->i_lock);
1489         if (res) {
1490                 int error;
1491
1492                 error = pnfs_commit_list(inode, &head, how);
1493                 if (error == PNFS_NOT_ATTEMPTED)
1494                         error = nfs_commit_list(inode, &head, how);
1495                 if (error < 0)
1496                         return error;
1497                 if (!may_wait)
1498                         goto out_mark_dirty;
1499                 error = wait_on_bit(&NFS_I(inode)->flags,
1500                                 NFS_INO_COMMIT,
1501                                 nfs_wait_bit_killable,
1502                                 TASK_KILLABLE);
1503                 if (error < 0)
1504                         return error;
1505         } else
1506                 nfs_commit_clear_lock(NFS_I(inode));
1507         return res;
1508         /* Note: If we exit without ensuring that the commit is complete,
1509          * we must mark the inode as dirty. Otherwise, future calls to
1510          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1511          * that the data is on the disk.
1512          */
1513 out_mark_dirty:
1514         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1515         return res;
1516 }
1517
1518 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1519 {
1520         struct nfs_inode *nfsi = NFS_I(inode);
1521         int flags = FLUSH_SYNC;
1522         int ret = 0;
1523
1524         if (wbc->sync_mode == WB_SYNC_NONE) {
1525                 /* Don't commit yet if this is a non-blocking flush and there
1526                  * are a lot of outstanding writes for this mapping.
1527                  */
1528                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1529                         goto out_mark_dirty;
1530
1531                 /* don't wait for the COMMIT response */
1532                 flags = 0;
1533         }
1534
1535         ret = nfs_commit_inode(inode, flags);
1536         if (ret >= 0) {
1537                 if (wbc->sync_mode == WB_SYNC_NONE) {
1538                         if (ret < wbc->nr_to_write)
1539                                 wbc->nr_to_write -= ret;
1540                         else
1541                                 wbc->nr_to_write = 0;
1542                 }
1543                 return 0;
1544         }
1545 out_mark_dirty:
1546         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1547         return ret;
1548 }
1549 #else
1550 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1551 {
1552         return 0;
1553 }
1554 #endif
1555
1556 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1557 {
1558         return nfs_commit_unstable_pages(inode, wbc);
1559 }
1560
1561 /*
1562  * flush the inode to disk.
1563  */
1564 int nfs_wb_all(struct inode *inode)
1565 {
1566         struct writeback_control wbc = {
1567                 .sync_mode = WB_SYNC_ALL,
1568                 .nr_to_write = LONG_MAX,
1569                 .range_start = 0,
1570                 .range_end = LLONG_MAX,
1571         };
1572
1573         return sync_inode(inode, &wbc);
1574 }
1575
1576 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1577 {
1578         struct nfs_page *req;
1579         int ret = 0;
1580
1581         BUG_ON(!PageLocked(page));
1582         for (;;) {
1583                 wait_on_page_writeback(page);
1584                 req = nfs_page_find_request(page);
1585                 if (req == NULL)
1586                         break;
1587                 if (nfs_lock_request_dontget(req)) {
1588                         nfs_inode_remove_request(req);
1589                         /*
1590                          * In case nfs_inode_remove_request has marked the
1591                          * page as being dirty
1592                          */
1593                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1594                         nfs_unlock_request(req);
1595                         break;
1596                 }
1597                 ret = nfs_wait_on_request(req);
1598                 nfs_release_request(req);
1599                 if (ret < 0)
1600                         break;
1601         }
1602         return ret;
1603 }
1604
1605 /*
1606  * Write back all requests on one page - we do this before reading it.
1607  */
1608 int nfs_wb_page(struct inode *inode, struct page *page)
1609 {
1610         loff_t range_start = page_offset(page);
1611         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1612         struct writeback_control wbc = {
1613                 .sync_mode = WB_SYNC_ALL,
1614                 .nr_to_write = 0,
1615                 .range_start = range_start,
1616                 .range_end = range_end,
1617         };
1618         int ret;
1619
1620         for (;;) {
1621                 wait_on_page_writeback(page);
1622                 if (clear_page_dirty_for_io(page)) {
1623                         ret = nfs_writepage_locked(page, &wbc);
1624                         if (ret < 0)
1625                                 goto out_error;
1626                         continue;
1627                 }
1628                 if (!PagePrivate(page))
1629                         break;
1630                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1631                 if (ret < 0)
1632                         goto out_error;
1633         }
1634         return 0;
1635 out_error:
1636         return ret;
1637 }
1638
1639 #ifdef CONFIG_MIGRATION
1640 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1641                 struct page *page)
1642 {
1643         struct nfs_page *req;
1644         int ret;
1645
1646         nfs_fscache_release_page(page, GFP_KERNEL);
1647
1648         req = nfs_find_and_lock_request(page, false);
1649         ret = PTR_ERR(req);
1650         if (IS_ERR(req))
1651                 goto out;
1652
1653         ret = migrate_page(mapping, newpage, page);
1654         if (!req)
1655                 goto out;
1656         if (ret)
1657                 goto out_unlock;
1658         page_cache_get(newpage);
1659         spin_lock(&mapping->host->i_lock);
1660         req->wb_page = newpage;
1661         SetPagePrivate(newpage);
1662         set_page_private(newpage, (unsigned long)req);
1663         ClearPagePrivate(page);
1664         set_page_private(page, 0);
1665         spin_unlock(&mapping->host->i_lock);
1666         page_cache_release(page);
1667 out_unlock:
1668         nfs_clear_page_tag_locked(req);
1669 out:
1670         return ret;
1671 }
1672 #endif
1673
1674 int __init nfs_init_writepagecache(void)
1675 {
1676         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1677                                              sizeof(struct nfs_write_data),
1678                                              0, SLAB_HWCACHE_ALIGN,
1679                                              NULL);
1680         if (nfs_wdata_cachep == NULL)
1681                 return -ENOMEM;
1682
1683         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1684                                                      nfs_wdata_cachep);
1685         if (nfs_wdata_mempool == NULL)
1686                 return -ENOMEM;
1687
1688         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1689                                                       nfs_wdata_cachep);
1690         if (nfs_commit_mempool == NULL)
1691                 return -ENOMEM;
1692
1693         /*
1694          * NFS congestion size, scale with available memory.
1695          *
1696          *  64MB:    8192k
1697          * 128MB:   11585k
1698          * 256MB:   16384k
1699          * 512MB:   23170k
1700          *   1GB:   32768k
1701          *   2GB:   46340k
1702          *   4GB:   65536k
1703          *   8GB:   92681k
1704          *  16GB:  131072k
1705          *
1706          * This allows larger machines to have larger/more transfers.
1707          * Limit the default to 256M
1708          */
1709         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1710         if (nfs_congestion_kb > 256*1024)
1711                 nfs_congestion_kb = 256*1024;
1712
1713         return 0;
1714 }
1715
1716 void nfs_destroy_writepagecache(void)
1717 {
1718         mempool_destroy(nfs_commit_mempool);
1719         mempool_destroy(nfs_wdata_mempool);
1720         kmem_cache_destroy(nfs_wdata_cachep);
1721 }
1722