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
  17 #include <linux/sunrpc/clnt.h>
  18 #include <linux/nfs_fs.h>
  19 #include <linux/nfs_mount.h>
  20 #include <linux/nfs_page.h>
  21 #include <linux/backing-dev.h>
  22
  23 #include <asm/uaccess.h>
  24
  25 #include "delegation.h"
  26 #include "internal.h"
  27 #include "iostat.h"
  28 #include "nfs4_fs.h"
  29
  30 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  31
  32 #define MIN_POOL_WRITE          (32)
  33 #define MIN_POOL_COMMIT         (4)
  34
  35 /*
  36  * Local function declarations
  37  */
  38 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  39                                   struct inode *inode, int ioflags);
  40 static void nfs_redirty_request(struct nfs_page *req);
  41 static const struct rpc_call_ops nfs_write_partial_ops;
  42 static const struct rpc_call_ops nfs_write_full_ops;
  43 static const struct rpc_call_ops nfs_commit_ops;
  44
  45 static struct kmem_cache *nfs_wdata_cachep;
  46 static mempool_t *nfs_wdata_mempool;
  47 static mempool_t *nfs_commit_mempool;
  48
  49 struct nfs_write_data *nfs_commitdata_alloc(void)
  50 {
  51         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  52
  53         if (p) {
  54                 memset(p, 0, sizeof(*p));
  55                 INIT_LIST_HEAD(&p->pages);
  56                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
  57         }
  58         return p;
  59 }
  60
  61 void nfs_commit_free(struct nfs_write_data *p)
  62 {
  63         if (p && (p->pagevec != &p->page_array[0]))
  64                 kfree(p->pagevec);
  65         mempool_free(p, nfs_commit_mempool);
  66 }
  67
  68 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  69 {
  70         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  71
  72         if (p) {
  73                 memset(p, 0, sizeof(*p));
  74                 INIT_LIST_HEAD(&p->pages);
  75                 p->npages = pagecount;
  76                 p->res.seq_res.sr_slotid = NFS4_MAX_SLOT_TABLE;
  77                 if (pagecount <= ARRAY_SIZE(p->page_array))
  78                         p->pagevec = p->page_array;
  79                 else {
  80                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  81                         if (!p->pagevec) {
  82                                 mempool_free(p, nfs_wdata_mempool);
  83                                 p = NULL;
  84                         }
  85                 }
  86         }
  87         return p;
  88 }
  89
  90 static void nfs_writedata_free(struct nfs_write_data *p)
  91 {
  92         if (p && (p->pagevec != &p->page_array[0]))
  93                 kfree(p->pagevec);
  94         mempool_free(p, nfs_wdata_mempool);
  95 }
  96
  97 void nfs_writedata_release(void *data)
  98 {
  99         struct nfs_write_data *wdata = data;
 100
 101         put_nfs_open_context(wdata->args.context);
 102         nfs_writedata_free(wdata);
 103 }
 104
 105 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 106 {
 107         ctx->error = error;
 108         smp_wmb();
 109         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 110 }
 111
 112 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 113 {
 114         struct nfs_page *req = NULL;
 115
 116         if (PagePrivate(page)) {
 117                 req = (struct nfs_page *)page_private(page);
 118                 if (req != NULL)
 119                         kref_get(&req->wb_kref);
 120         }
 121         return req;
 122 }
 123
 124 static struct nfs_page *nfs_page_find_request(struct page *page)
 125 {
 126         struct inode *inode = page->mapping->host;
 127         struct nfs_page *req = NULL;
 128
 129         spin_lock(&inode->i_lock);
 130         req = nfs_page_find_request_locked(page);
 131         spin_unlock(&inode->i_lock);
 132         return req;
 133 }
 134
 135 /* Adjust the file length if we're writing beyond the end */
 136 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 137 {
 138         struct inode *inode = page->mapping->host;
 139         loff_t end, i_size;
 140         pgoff_t end_index;
 141
 142         spin_lock(&inode->i_lock);
 143         i_size = i_size_read(inode);
 144         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 145         if (i_size > 0 && page->index < end_index)
 146                 goto out;
 147         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 148         if (i_size >= end)
 149                 goto out;
 150         i_size_write(inode, end);
 151         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 152 out:
 153         spin_unlock(&inode->i_lock);
 154 }
 155
 156 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 157 static void nfs_set_pageerror(struct page *page)
 158 {
 159         SetPageError(page);
 160         nfs_zap_mapping(page->mapping->host, page->mapping);
 161 }
 162
 163 /* We can set the PG_uptodate flag if we see that a write request
 164  * covers the full page.
 165  */
 166 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 167 {
 168         if (PageUptodate(page))
 169                 return;
 170         if (base != 0)
 171                 return;
 172         if (count != nfs_page_length(page))
 173                 return;
 174         SetPageUptodate(page);
 175 }
 176
 177 static int wb_priority(struct writeback_control *wbc)
 178 {
 179         if (wbc->for_reclaim)
 180                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 181         if (wbc->for_kupdate)
 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                 if (atomic_long_inc_return(&nfss->writeback) >
 205                                 NFS_CONGESTION_ON_THRESH) {
 206                         set_bdi_congested(&nfss->backing_dev_info,
 207                                                 BLK_RW_ASYNC);
 208                 }
 209         }
 210         return ret;
 211 }
 212
 213 static void nfs_end_page_writeback(struct page *page)
 214 {
 215         struct inode *inode = page->mapping->host;
 216         struct nfs_server *nfss = NFS_SERVER(inode);
 217
 218         end_page_writeback(page);
 219         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 220                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 221 }
 222
 223 /*
 224  * Find an associated nfs write request, and prepare to flush it out
 225  * May return an error if the user signalled nfs_wait_on_request().
 226  */
 227 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 228                                 struct page *page)
 229 {
 230         struct inode *inode = page->mapping->host;
 231         struct nfs_page *req;
 232         int ret;
 233
 234         spin_lock(&inode->i_lock);
 235         for(;;) {
 236                 req = nfs_page_find_request_locked(page);
 237                 if (req == NULL) {
 238                         spin_unlock(&inode->i_lock);
 239                         return 0;
 240                 }
 241                 if (nfs_set_page_tag_locked(req))
 242                         break;
 243                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 244                  *       then the call to nfs_set_page_tag_locked() will always
 245                  *       succeed provided that someone hasn't already marked the
 246                  *       request as dirty (in which case we don't care).
 247                  */
 248                 spin_unlock(&inode->i_lock);
 249                 ret = nfs_wait_on_request(req);
 250                 nfs_release_request(req);
 251                 if (ret != 0)
 252                         return ret;
 253                 spin_lock(&inode->i_lock);
 254         }
 255         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 256                 spin_unlock(&inode->i_lock);
 257                 BUG();
 258         }
 259         if (nfs_set_page_writeback(page) != 0) {
 260                 spin_unlock(&inode->i_lock);
 261                 BUG();
 262         }
 263         spin_unlock(&inode->i_lock);
 264         if (!nfs_pageio_add_request(pgio, req)) {
 265                 nfs_redirty_request(req);
 266                 return pgio->pg_error;
 267         }
 268         return 0;
 269 }
 270
 271 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 272 {
 273         struct inode *inode = page->mapping->host;
 274
 275         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 276         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 277
 278         nfs_pageio_cond_complete(pgio, page->index);
 279         return nfs_page_async_flush(pgio, page);
 280 }
 281
 282 /*
 283  * Write an mmapped page to the server.
 284  */
 285 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 286 {
 287         struct nfs_pageio_descriptor pgio;
 288         int err;
 289
 290         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 291         err = nfs_do_writepage(page, wbc, &pgio);
 292         nfs_pageio_complete(&pgio);
 293         if (err < 0)
 294                 return err;
 295         if (pgio.pg_error < 0)
 296                 return pgio.pg_error;
 297         return 0;
 298 }
 299
 300 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 301 {
 302         int ret;
 303
 304         ret = nfs_writepage_locked(page, wbc);
 305         unlock_page(page);
 306         return ret;
 307 }
 308
 309 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 310 {
 311         int ret;
 312
 313         ret = nfs_do_writepage(page, wbc, data);
 314         unlock_page(page);
 315         return ret;
 316 }
 317
 318 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 319 {
 320         struct inode *inode = mapping->host;
 321         unsigned long *bitlock = &NFS_I(inode)->flags;
 322         struct nfs_pageio_descriptor pgio;
 323         int err;
 324
 325         /* Stop dirtying of new pages while we sync */
 326         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 327                         nfs_wait_bit_killable, TASK_KILLABLE);
 328         if (err)
 329                 goto out_err;
 330
 331         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 332
 333         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 334         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 335         nfs_pageio_complete(&pgio);
 336
 337         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 338         smp_mb__after_clear_bit();
 339         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 340
 341         if (err < 0)
 342                 goto out_err;
 343         err = pgio.pg_error;
 344         if (err < 0)
 345                 goto out_err;
 346         return 0;
 347 out_err:
 348         return err;
 349 }
 350
 351 /*
 352  * Insert a write request into an inode
 353  */
 354 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 355 {
 356         struct nfs_inode *nfsi = NFS_I(inode);
 357         int error;
 358
 359         error = radix_tree_preload(GFP_NOFS);
 360         if (error != 0)
 361                 goto out;
 362
 363         /* Lock the request! */
 364         nfs_lock_request_dontget(req);
 365
 366         spin_lock(&inode->i_lock);
 367         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 368         BUG_ON(error);
 369         if (!nfsi->npages) {
 370                 igrab(inode);
 371                 if (nfs_have_delegation(inode, FMODE_WRITE))
 372                         nfsi->change_attr++;
 373         }
 374         SetPagePrivate(req->wb_page);
 375         set_page_private(req->wb_page, (unsigned long)req);
 376         nfsi->npages++;
 377         kref_get(&req->wb_kref);
 378         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 379                                 NFS_PAGE_TAG_LOCKED);
 380         spin_unlock(&inode->i_lock);
 381         radix_tree_preload_end();
 382 out:
 383         return error;
 384 }
 385
 386 /*
 387  * Remove a write request from an inode
 388  */
 389 static void nfs_inode_remove_request(struct nfs_page *req)
 390 {
 391         struct inode *inode = req->wb_context->path.dentry->d_inode;
 392         struct nfs_inode *nfsi = NFS_I(inode);
 393
 394         BUG_ON (!NFS_WBACK_BUSY(req));
 395
 396         spin_lock(&inode->i_lock);
 397         set_page_private(req->wb_page, 0);
 398         ClearPagePrivate(req->wb_page);
 399         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 400         nfsi->npages--;
 401         if (!nfsi->npages) {
 402                 spin_unlock(&inode->i_lock);
 403                 iput(inode);
 404         } else
 405                 spin_unlock(&inode->i_lock);
 406         nfs_clear_request(req);
 407         nfs_release_request(req);
 408 }
 409
 410 static void
 411 nfs_mark_request_dirty(struct nfs_page *req)
 412 {
 413         __set_page_dirty_nobuffers(req->wb_page);
 414 }
 415
 416 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 417 /*
 418  * Add a request to the inode's commit list.
 419  */
 420 static void
 421 nfs_mark_request_commit(struct nfs_page *req)
 422 {
 423         struct inode *inode = req->wb_context->path.dentry->d_inode;
 424         struct nfs_inode *nfsi = NFS_I(inode);
 425
 426         spin_lock(&inode->i_lock);
 427         set_bit(PG_CLEAN, &(req)->wb_flags);
 428         radix_tree_tag_set(&nfsi->nfs_page_tree,
 429                         req->wb_index,
 430                         NFS_PAGE_TAG_COMMIT);
 431         spin_unlock(&inode->i_lock);
 432         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 433         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 434         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 435 }
 436
 437 static int
 438 nfs_clear_request_commit(struct nfs_page *req)
 439 {
 440         struct page *page = req->wb_page;
 441
 442         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 443                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 444                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 445                 return 1;
 446         }
 447         return 0;
 448 }
 449
 450 static inline
 451 int nfs_write_need_commit(struct nfs_write_data *data)
 452 {
 453         return data->verf.committed != NFS_FILE_SYNC;
 454 }
 455
 456 static inline
 457 int nfs_reschedule_unstable_write(struct nfs_page *req)
 458 {
 459         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 460                 nfs_mark_request_commit(req);
 461                 return 1;
 462         }
 463         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 464                 nfs_mark_request_dirty(req);
 465                 return 1;
 466         }
 467         return 0;
 468 }
 469 #else
 470 static inline void
 471 nfs_mark_request_commit(struct nfs_page *req)
 472 {
 473 }
 474
 475 static inline int
 476 nfs_clear_request_commit(struct nfs_page *req)
 477 {
 478         return 0;
 479 }
 480
 481 static inline
 482 int nfs_write_need_commit(struct nfs_write_data *data)
 483 {
 484         return 0;
 485 }
 486
 487 static inline
 488 int nfs_reschedule_unstable_write(struct nfs_page *req)
 489 {
 490         return 0;
 491 }
 492 #endif
 493
 494 /*
 495  * Wait for a request to complete.
 496  *
 497  * Interruptible by fatal signals only.
 498  */
 499 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 500 {
 501         struct nfs_inode *nfsi = NFS_I(inode);
 502         struct nfs_page *req;
 503         pgoff_t idx_end, next;
 504         unsigned int            res = 0;
 505         int                     error;
 506
 507         if (npages == 0)
 508                 idx_end = ~0;
 509         else
 510                 idx_end = idx_start + npages - 1;
 511
 512         next = idx_start;
 513         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 514                 if (req->wb_index > idx_end)
 515                         break;
 516
 517                 next = req->wb_index + 1;
 518                 BUG_ON(!NFS_WBACK_BUSY(req));
 519
 520                 kref_get(&req->wb_kref);
 521                 spin_unlock(&inode->i_lock);
 522                 error = nfs_wait_on_request(req);
 523                 nfs_release_request(req);
 524                 spin_lock(&inode->i_lock);
 525                 if (error < 0)
 526                         return error;
 527                 res++;
 528         }
 529         return res;
 530 }
 531
 532 static void nfs_cancel_commit_list(struct list_head *head)
 533 {
 534         struct nfs_page *req;
 535
 536         while(!list_empty(head)) {
 537                 req = nfs_list_entry(head->next);
 538                 nfs_list_remove_request(req);
 539                 nfs_clear_request_commit(req);
 540                 nfs_inode_remove_request(req);
 541                 nfs_unlock_request(req);
 542         }
 543 }
 544
 545 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 546 static int
 547 nfs_need_commit(struct nfs_inode *nfsi)
 548 {
 549         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 550 }
 551
 552 /*
 553  * nfs_scan_commit - Scan an inode for commit requests
 554  * @inode: NFS inode to scan
 555  * @dst: destination list
 556  * @idx_start: lower bound of page->index to scan.
 557  * @npages: idx_start + npages sets the upper bound to scan.
 558  *
 559  * Moves requests from the inode's 'commit' request list.
 560  * The requests are *not* checked to ensure that they form a contiguous set.
 561  */
 562 static int
 563 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 564 {
 565         struct nfs_inode *nfsi = NFS_I(inode);
 566
 567         if (!nfs_need_commit(nfsi))
 568                 return 0;
 569
 570         return nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 571 }
 572 #else
 573 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 574 {
 575         return 0;
 576 }
 577
 578 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 579 {
 580         return 0;
 581 }
 582 #endif
 583
 584 /*
 585  * Search for an existing write request, and attempt to update
 586  * it to reflect a new dirty region on a given page.
 587  *
 588  * If the attempt fails, then the existing request is flushed out
 589  * to disk.
 590  */
 591 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 592                 struct page *page,
 593                 unsigned int offset,
 594                 unsigned int bytes)
 595 {
 596         struct nfs_page *req;
 597         unsigned int rqend;
 598         unsigned int end;
 599         int error;
 600
 601         if (!PagePrivate(page))
 602                 return NULL;
 603
 604         end = offset + bytes;
 605         spin_lock(&inode->i_lock);
 606
 607         for (;;) {
 608                 req = nfs_page_find_request_locked(page);
 609                 if (req == NULL)
 610                         goto out_unlock;
 611
 612                 rqend = req->wb_offset + req->wb_bytes;
 613                 /*
 614                  * Tell the caller to flush out the request if
 615                  * the offsets are non-contiguous.
 616                  * Note: nfs_flush_incompatible() will already
 617                  * have flushed out requests having wrong owners.
 618                  */
 619                 if (offset > rqend
 620                     || end < req->wb_offset)
 621                         goto out_flushme;
 622
 623                 if (nfs_set_page_tag_locked(req))
 624                         break;
 625
 626                 /* The request is locked, so wait and then retry */
 627                 spin_unlock(&inode->i_lock);
 628                 error = nfs_wait_on_request(req);
 629                 nfs_release_request(req);
 630                 if (error != 0)
 631                         goto out_err;
 632                 spin_lock(&inode->i_lock);
 633         }
 634
 635         if (nfs_clear_request_commit(req))
 636                 radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 637                                 req->wb_index, NFS_PAGE_TAG_COMMIT);
 638
 639         /* Okay, the request matches. Update the region */
 640         if (offset < req->wb_offset) {
 641                 req->wb_offset = offset;
 642                 req->wb_pgbase = offset;
 643         }
 644         if (end > rqend)
 645                 req->wb_bytes = end - req->wb_offset;
 646         else
 647                 req->wb_bytes = rqend - req->wb_offset;
 648 out_unlock:
 649         spin_unlock(&inode->i_lock);
 650         return req;
 651 out_flushme:
 652         spin_unlock(&inode->i_lock);
 653         nfs_release_request(req);
 654         error = nfs_wb_page(inode, page);
 655 out_err:
 656         return ERR_PTR(error);
 657 }
 658
 659 /*
 660  * Try to update an existing write request, or create one if there is none.
 661  *
 662  * Note: Should always be called with the Page Lock held to prevent races
 663  * if we have to add a new request. Also assumes that the caller has
 664  * already called nfs_flush_incompatible() if necessary.
 665  */
 666 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 667                 struct page *page, unsigned int offset, unsigned int bytes)
 668 {
 669         struct inode *inode = page->mapping->host;
 670         struct nfs_page *req;
 671         int error;
 672
 673         req = nfs_try_to_update_request(inode, page, offset, bytes);
 674         if (req != NULL)
 675                 goto out;
 676         req = nfs_create_request(ctx, inode, page, offset, bytes);
 677         if (IS_ERR(req))
 678                 goto out;
 679         error = nfs_inode_add_request(inode, req);
 680         if (error != 0) {
 681                 nfs_release_request(req);
 682                 req = ERR_PTR(error);
 683         }
 684 out:
 685         return req;
 686 }
 687
 688 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 689                 unsigned int offset, unsigned int count)
 690 {
 691         struct nfs_page *req;
 692
 693         req = nfs_setup_write_request(ctx, page, offset, count);
 694         if (IS_ERR(req))
 695                 return PTR_ERR(req);
 696         /* Update file length */
 697         nfs_grow_file(page, offset, count);
 698         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 699         nfs_clear_page_tag_locked(req);
 700         return 0;
 701 }
 702
 703 int nfs_flush_incompatible(struct file *file, struct page *page)
 704 {
 705         struct nfs_open_context *ctx = nfs_file_open_context(file);
 706         struct nfs_page *req;
 707         int do_flush, status;
 708         /*
 709          * Look for a request corresponding to this page. If there
 710          * is one, and it belongs to another file, we flush it out
 711          * before we try to copy anything into the page. Do this
 712          * due to the lack of an ACCESS-type call in NFSv2.
 713          * Also do the same if we find a request from an existing
 714          * dropped page.
 715          */
 716         do {
 717                 req = nfs_page_find_request(page);
 718                 if (req == NULL)
 719                         return 0;
 720                 do_flush = req->wb_page != page || req->wb_context != ctx;
 721                 nfs_release_request(req);
 722                 if (!do_flush)
 723                         return 0;
 724                 status = nfs_wb_page(page->mapping->host, page);
 725         } while (status == 0);
 726         return status;
 727 }
 728
 729 /*
 730  * If the page cache is marked as unsafe or invalid, then we can't rely on
 731  * the PageUptodate() flag. In this case, we will need to turn off
 732  * write optimisations that depend on the page contents being correct.
 733  */
 734 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 735 {
 736         return PageUptodate(page) &&
 737                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 738 }
 739
 740 /*
 741  * Update and possibly write a cached page of an NFS file.
 742  *
 743  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 744  * things with a page scheduled for an RPC call (e.g. invalidate it).
 745  */
 746 int nfs_updatepage(struct file *file, struct page *page,
 747                 unsigned int offset, unsigned int count)
 748 {
 749         struct nfs_open_context *ctx = nfs_file_open_context(file);
 750         struct inode    *inode = page->mapping->host;
 751         int             status = 0;
 752
 753         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 754
 755         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 756                 file->f_path.dentry->d_parent->d_name.name,
 757                 file->f_path.dentry->d_name.name, count,
 758                 (long long)(page_offset(page) + offset));
 759
 760         /* If we're not using byte range locks, and we know the page
 761          * is up to date, it may be more efficient to extend the write
 762          * to cover the entire page in order to avoid fragmentation
 763          * inefficiencies.
 764          */
 765         if (nfs_write_pageuptodate(page, inode) &&
 766                         inode->i_flock == NULL &&
 767                         !(file->f_flags & O_SYNC)) {
 768                 count = max(count + offset, nfs_page_length(page));
 769                 offset = 0;
 770         }
 771
 772         status = nfs_writepage_setup(ctx, page, offset, count);
 773         if (status < 0)
 774                 nfs_set_pageerror(page);
 775         else
 776                 __set_page_dirty_nobuffers(page);
 777
 778         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 779                         status, (long long)i_size_read(inode));
 780         return status;
 781 }
 782
 783 static void nfs_writepage_release(struct nfs_page *req)
 784 {
 785
 786         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req)) {
 787                 nfs_end_page_writeback(req->wb_page);
 788                 nfs_inode_remove_request(req);
 789         } else
 790                 nfs_end_page_writeback(req->wb_page);
 791         nfs_clear_page_tag_locked(req);
 792 }
 793
 794 static int flush_task_priority(int how)
 795 {
 796         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 797                 case FLUSH_HIGHPRI:
 798                         return RPC_PRIORITY_HIGH;
 799                 case FLUSH_LOWPRI:
 800                         return RPC_PRIORITY_LOW;
 801         }
 802         return RPC_PRIORITY_NORMAL;
 803 }
 804
 805 /*
 806  * Set up the argument/result storage required for the RPC call.
 807  */
 808 static int nfs_write_rpcsetup(struct nfs_page *req,
 809                 struct nfs_write_data *data,
 810                 const struct rpc_call_ops *call_ops,
 811                 unsigned int count, unsigned int offset,
 812                 int how)
 813 {
 814         struct inode *inode = req->wb_context->path.dentry->d_inode;
 815         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 816         int priority = flush_task_priority(how);
 817         struct rpc_task *task;
 818         struct rpc_message msg = {
 819                 .rpc_argp = &data->args,
 820                 .rpc_resp = &data->res,
 821                 .rpc_cred = req->wb_context->cred,
 822         };
 823         struct rpc_task_setup task_setup_data = {
 824                 .rpc_client = NFS_CLIENT(inode),
 825                 .task = &data->task,
 826                 .rpc_message = &msg,
 827                 .callback_ops = call_ops,
 828                 .callback_data = data,
 829                 .workqueue = nfsiod_workqueue,
 830                 .flags = flags,
 831                 .priority = priority,
 832         };
 833
 834         /* Set up the RPC argument and reply structs
 835          * NB: take care not to mess about with data->commit et al. */
 836
 837         data->req = req;
 838         data->inode = inode = req->wb_context->path.dentry->d_inode;
 839         data->cred = msg.rpc_cred;
 840
 841         data->args.fh     = NFS_FH(inode);
 842         data->args.offset = req_offset(req) + offset;
 843         data->args.pgbase = req->wb_pgbase + offset;
 844         data->args.pages  = data->pagevec;
 845         data->args.count  = count;
 846         data->args.context = get_nfs_open_context(req->wb_context);
 847         data->args.stable  = NFS_UNSTABLE;
 848         if (how & FLUSH_STABLE) {
 849                 data->args.stable = NFS_DATA_SYNC;
 850                 if (!nfs_need_commit(NFS_I(inode)))
 851                         data->args.stable = NFS_FILE_SYNC;
 852         }
 853
 854         data->res.fattr   = &data->fattr;
 855         data->res.count   = count;
 856         data->res.verf    = &data->verf;
 857         nfs_fattr_init(&data->fattr);
 858
 859         /* Set up the initial task struct.  */
 860         NFS_PROTO(inode)->write_setup(data, &msg);
 861
 862         dprintk("NFS: %5u initiated write call "
 863                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 864                 data->task.tk_pid,
 865                 inode->i_sb->s_id,
 866                 (long long)NFS_FILEID(inode),
 867                 count,
 868                 (unsigned long long)data->args.offset);
 869
 870         task = rpc_run_task(&task_setup_data);
 871         if (IS_ERR(task))
 872                 return PTR_ERR(task);
 873         rpc_put_task(task);
 874         return 0;
 875 }
 876
 877 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 878  * call this on each, which will prepare them to be retried on next
 879  * writeback using standard nfs.
 880  */
 881 static void nfs_redirty_request(struct nfs_page *req)
 882 {
 883         nfs_mark_request_dirty(req);
 884         nfs_end_page_writeback(req->wb_page);
 885         nfs_clear_page_tag_locked(req);
 886 }
 887
 888 /*
 889  * Generate multiple small requests to write out a single
 890  * contiguous dirty area on one page.
 891  */
 892 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 893 {
 894         struct nfs_page *req = nfs_list_entry(head->next);
 895         struct page *page = req->wb_page;
 896         struct nfs_write_data *data;
 897         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 898         unsigned int offset;
 899         int requests = 0;
 900         int ret = 0;
 901         LIST_HEAD(list);
 902
 903         nfs_list_remove_request(req);
 904
 905         nbytes = count;
 906         do {
 907                 size_t len = min(nbytes, wsize);
 908
 909                 data = nfs_writedata_alloc(1);
 910                 if (!data)
 911                         goto out_bad;
 912                 list_add(&data->pages, &list);
 913                 requests++;
 914                 nbytes -= len;
 915         } while (nbytes != 0);
 916         atomic_set(&req->wb_complete, requests);
 917
 918         ClearPageError(page);
 919         offset = 0;
 920         nbytes = count;
 921         do {
 922                 int ret2;
 923
 924                 data = list_entry(list.next, struct nfs_write_data, pages);
 925                 list_del_init(&data->pages);
 926
 927                 data->pagevec[0] = page;
 928
 929                 if (nbytes < wsize)
 930                         wsize = nbytes;
 931                 ret2 = nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 932                                    wsize, offset, how);
 933                 if (ret == 0)
 934                         ret = ret2;
 935                 offset += wsize;
 936                 nbytes -= wsize;
 937         } while (nbytes != 0);
 938
 939         return ret;
 940
 941 out_bad:
 942         while (!list_empty(&list)) {
 943                 data = list_entry(list.next, struct nfs_write_data, pages);
 944                 list_del(&data->pages);
 945                 nfs_writedata_release(data);
 946         }
 947         nfs_redirty_request(req);
 948         return -ENOMEM;
 949 }
 950
 951 /*
 952  * Create an RPC task for the given write request and kick it.
 953  * The page must have been locked by the caller.
 954  *
 955  * It may happen that the page we're passed is not marked dirty.
 956  * This is the case if nfs_updatepage detects a conflicting request
 957  * that has been written but not committed.
 958  */
 959 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 960 {
 961         struct nfs_page         *req;
 962         struct page             **pages;
 963         struct nfs_write_data   *data;
 964
 965         data = nfs_writedata_alloc(npages);
 966         if (!data)
 967                 goto out_bad;
 968
 969         pages = data->pagevec;
 970         while (!list_empty(head)) {
 971                 req = nfs_list_entry(head->next);
 972                 nfs_list_remove_request(req);
 973                 nfs_list_add_request(req, &data->pages);
 974                 ClearPageError(req->wb_page);
 975                 *pages++ = req->wb_page;
 976         }
 977         req = nfs_list_entry(data->pages.next);
 978
 979         /* Set up the argument struct */
 980         return nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 981  out_bad:
 982         while (!list_empty(head)) {
 983                 req = nfs_list_entry(head->next);
 984                 nfs_list_remove_request(req);
 985                 nfs_redirty_request(req);
 986         }
 987         return -ENOMEM;
 988 }
 989
 990 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 991                                   struct inode *inode, int ioflags)
 992 {
 993         size_t wsize = NFS_SERVER(inode)->wsize;
 994
 995         if (wsize < PAGE_CACHE_SIZE)
 996                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 997         else
 998                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 999 }
1000
1001 /*
1002  * Handle a write reply that flushed part of a page.
1003  */
1004 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1005 {
1006         struct nfs_write_data   *data = calldata;
1007
1008         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1009                 task->tk_pid,
1010                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1011                 (long long)
1012                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1013                 data->req->wb_bytes, (long long)req_offset(data->req));
1014
1015         nfs_writeback_done(task, data);
1016 }
1017
1018 static void nfs_writeback_release_partial(void *calldata)
1019 {
1020         struct nfs_write_data   *data = calldata;
1021         struct nfs_page         *req = data->req;
1022         struct page             *page = req->wb_page;
1023         int status = data->task.tk_status;
1024
1025         if (status < 0) {
1026                 nfs_set_pageerror(page);
1027                 nfs_context_set_write_error(req->wb_context, status);
1028                 dprintk(", error = %d\n", status);
1029                 goto out;
1030         }
1031
1032         if (nfs_write_need_commit(data)) {
1033                 struct inode *inode = page->mapping->host;
1034
1035                 spin_lock(&inode->i_lock);
1036                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1037                         /* Do nothing we need to resend the writes */
1038                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1039                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1040                         dprintk(" defer commit\n");
1041                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1042                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1043                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1044                         dprintk(" server reboot detected\n");
1045                 }
1046                 spin_unlock(&inode->i_lock);
1047         } else
1048                 dprintk(" OK\n");
1049
1050 out:
1051         if (atomic_dec_and_test(&req->wb_complete))
1052                 nfs_writepage_release(req);
1053         nfs_writedata_release(calldata);
1054 }
1055
1056 #if defined(CONFIG_NFS_V4_1)
1057 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1058 {
1059         struct nfs_write_data *data = calldata;
1060         struct nfs_client *clp = (NFS_SERVER(data->inode))->nfs_client;
1061
1062         if (nfs4_setup_sequence(clp, &data->args.seq_args,
1063                                 &data->res.seq_res, 1, task))
1064                 return;
1065         rpc_call_start(task);
1066 }
1067 #endif /* CONFIG_NFS_V4_1 */
1068
1069 static const struct rpc_call_ops nfs_write_partial_ops = {
1070 #if defined(CONFIG_NFS_V4_1)
1071         .rpc_call_prepare = nfs_write_prepare,
1072 #endif /* CONFIG_NFS_V4_1 */
1073         .rpc_call_done = nfs_writeback_done_partial,
1074         .rpc_release = nfs_writeback_release_partial,
1075 };
1076
1077 /*
1078  * Handle a write reply that flushes a whole page.
1079  *
1080  * FIXME: There is an inherent race with invalidate_inode_pages and
1081  *        writebacks since the page->count is kept > 1 for as long
1082  *        as the page has a write request pending.
1083  */
1084 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1085 {
1086         struct nfs_write_data   *data = calldata;
1087
1088         nfs_writeback_done(task, data);
1089 }
1090
1091 static void nfs_writeback_release_full(void *calldata)
1092 {
1093         struct nfs_write_data   *data = calldata;
1094         int status = data->task.tk_status;
1095
1096         /* Update attributes as result of writeback. */
1097         while (!list_empty(&data->pages)) {
1098                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1099                 struct page *page = req->wb_page;
1100
1101                 nfs_list_remove_request(req);
1102
1103                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1104                         data->task.tk_pid,
1105                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1106                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1107                         req->wb_bytes,
1108                         (long long)req_offset(req));
1109
1110                 if (status < 0) {
1111                         nfs_set_pageerror(page);
1112                         nfs_context_set_write_error(req->wb_context, status);
1113                         dprintk(", error = %d\n", status);
1114                         goto remove_request;
1115                 }
1116
1117                 if (nfs_write_need_commit(data)) {
1118                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1119                         nfs_mark_request_commit(req);
1120                         nfs_end_page_writeback(page);
1121                         dprintk(" marked for commit\n");
1122                         goto next;
1123                 }
1124                 dprintk(" OK\n");
1125 remove_request:
1126                 nfs_end_page_writeback(page);
1127                 nfs_inode_remove_request(req);
1128         next:
1129                 nfs_clear_page_tag_locked(req);
1130         }
1131         nfs_writedata_release(calldata);
1132 }
1133
1134 static const struct rpc_call_ops nfs_write_full_ops = {
1135 #if defined(CONFIG_NFS_V4_1)
1136         .rpc_call_prepare = nfs_write_prepare,
1137 #endif /* CONFIG_NFS_V4_1 */
1138         .rpc_call_done = nfs_writeback_done_full,
1139         .rpc_release = nfs_writeback_release_full,
1140 };
1141
1142
1143 /*
1144  * This function is called when the WRITE call is complete.
1145  */
1146 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1147 {
1148         struct nfs_writeargs    *argp = &data->args;
1149         struct nfs_writeres     *resp = &data->res;
1150         struct nfs_server       *server = NFS_SERVER(data->inode);
1151         int status;
1152
1153         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1154                 task->tk_pid, task->tk_status);
1155
1156         /*
1157          * ->write_done will attempt to use post-op attributes to detect
1158          * conflicting writes by other clients.  A strict interpretation
1159          * of close-to-open would allow us to continue caching even if
1160          * another writer had changed the file, but some applications
1161          * depend on tighter cache coherency when writing.
1162          */
1163         status = NFS_PROTO(data->inode)->write_done(task, data);
1164         if (status != 0)
1165                 return status;
1166         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1167
1168 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1169         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1170                 /* We tried a write call, but the server did not
1171                  * commit data to stable storage even though we
1172                  * requested it.
1173                  * Note: There is a known bug in Tru64 < 5.0 in which
1174                  *       the server reports NFS_DATA_SYNC, but performs
1175                  *       NFS_FILE_SYNC. We therefore implement this checking
1176                  *       as a dprintk() in order to avoid filling syslog.
1177                  */
1178                 static unsigned long    complain;
1179
1180                 if (time_before(complain, jiffies)) {
1181                         dprintk("NFS:       faulty NFS server %s:"
1182                                 " (committed = %d) != (stable = %d)\n",
1183                                 server->nfs_client->cl_hostname,
1184                                 resp->verf->committed, argp->stable);
1185                         complain = jiffies + 300 * HZ;
1186                 }
1187         }
1188 #endif
1189         /* Is this a short write? */
1190         if (task->tk_status >= 0 && resp->count < argp->count) {
1191                 static unsigned long    complain;
1192
1193                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1194
1195                 /* Has the server at least made some progress? */
1196                 if (resp->count != 0) {
1197                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1198                         if (resp->verf->committed != NFS_UNSTABLE) {
1199                                 /* Resend from where the server left off */
1200                                 argp->offset += resp->count;
1201                                 argp->pgbase += resp->count;
1202                                 argp->count -= resp->count;
1203                         } else {
1204                                 /* Resend as a stable write in order to avoid
1205                                  * headaches in the case of a server crash.
1206                                  */
1207                                 argp->stable = NFS_FILE_SYNC;
1208                         }
1209                         nfs4_restart_rpc(task, server->nfs_client);
1210                         return -EAGAIN;
1211                 }
1212                 if (time_before(complain, jiffies)) {
1213                         printk(KERN_WARNING
1214                                "NFS: Server wrote zero bytes, expected %u.\n",
1215                                         argp->count);
1216                         complain = jiffies + 300 * HZ;
1217                 }
1218                 /* Can't do anything about it except throw an error. */
1219                 task->tk_status = -EIO;
1220         }
1221         nfs4_sequence_free_slot(server->nfs_client, &data->res.seq_res);
1222         return 0;
1223 }
1224
1225
1226 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1227 void nfs_commitdata_release(void *data)
1228 {
1229         struct nfs_write_data *wdata = data;
1230
1231         put_nfs_open_context(wdata->args.context);
1232         nfs_commit_free(wdata);
1233 }
1234
1235 /*
1236  * Set up the argument/result storage required for the RPC call.
1237  */
1238 static int nfs_commit_rpcsetup(struct list_head *head,
1239                 struct nfs_write_data *data,
1240                 int how)
1241 {
1242         struct nfs_page *first = nfs_list_entry(head->next);
1243         struct inode *inode = first->wb_context->path.dentry->d_inode;
1244         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1245         int priority = flush_task_priority(how);
1246         struct rpc_task *task;
1247         struct rpc_message msg = {
1248                 .rpc_argp = &data->args,
1249                 .rpc_resp = &data->res,
1250                 .rpc_cred = first->wb_context->cred,
1251         };
1252         struct rpc_task_setup task_setup_data = {
1253                 .task = &data->task,
1254                 .rpc_client = NFS_CLIENT(inode),
1255                 .rpc_message = &msg,
1256                 .callback_ops = &nfs_commit_ops,
1257                 .callback_data = data,
1258                 .workqueue = nfsiod_workqueue,
1259                 .flags = flags,
1260                 .priority = priority,
1261         };
1262
1263         /* Set up the RPC argument and reply structs
1264          * NB: take care not to mess about with data->commit et al. */
1265
1266         list_splice_init(head, &data->pages);
1267
1268         data->inode       = inode;
1269         data->cred        = msg.rpc_cred;
1270
1271         data->args.fh     = NFS_FH(data->inode);
1272         /* Note: we always request a commit of the entire inode */
1273         data->args.offset = 0;
1274         data->args.count  = 0;
1275         data->args.context = get_nfs_open_context(first->wb_context);
1276         data->res.count   = 0;
1277         data->res.fattr   = &data->fattr;
1278         data->res.verf    = &data->verf;
1279         nfs_fattr_init(&data->fattr);
1280
1281         /* Set up the initial task struct.  */
1282         NFS_PROTO(inode)->commit_setup(data, &msg);
1283
1284         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1285
1286         task = rpc_run_task(&task_setup_data);
1287         if (IS_ERR(task))
1288                 return PTR_ERR(task);
1289         rpc_put_task(task);
1290         return 0;
1291 }
1292
1293 /*
1294  * Commit dirty pages
1295  */
1296 static int
1297 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1298 {
1299         struct nfs_write_data   *data;
1300         struct nfs_page         *req;
1301
1302         data = nfs_commitdata_alloc();
1303
1304         if (!data)
1305                 goto out_bad;
1306
1307         /* Set up the argument struct */
1308         return nfs_commit_rpcsetup(head, data, how);
1309  out_bad:
1310         while (!list_empty(head)) {
1311                 req = nfs_list_entry(head->next);
1312                 nfs_list_remove_request(req);
1313                 nfs_mark_request_commit(req);
1314                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1315                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1316                                 BDI_RECLAIMABLE);
1317                 nfs_clear_page_tag_locked(req);
1318         }
1319         return -ENOMEM;
1320 }
1321
1322 /*
1323  * COMMIT call returned
1324  */
1325 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1326 {
1327         struct nfs_write_data   *data = calldata;
1328
1329         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1330                                 task->tk_pid, task->tk_status);
1331
1332         /* Call the NFS version-specific code */
1333         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1334                 return;
1335 }
1336
1337 static void nfs_commit_release(void *calldata)
1338 {
1339         struct nfs_write_data   *data = calldata;
1340         struct nfs_page         *req;
1341         int status = data->task.tk_status;
1342
1343         while (!list_empty(&data->pages)) {
1344                 req = nfs_list_entry(data->pages.next);
1345                 nfs_list_remove_request(req);
1346                 nfs_clear_request_commit(req);
1347
1348                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1349                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1350                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1351                         req->wb_bytes,
1352                         (long long)req_offset(req));
1353                 if (status < 0) {
1354                         nfs_context_set_write_error(req->wb_context, status);
1355                         nfs_inode_remove_request(req);
1356                         dprintk(", error = %d\n", status);
1357                         goto next;
1358                 }
1359
1360                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1361                  * returned by the server against all stored verfs. */
1362                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1363                         /* We have a match */
1364                         nfs_inode_remove_request(req);
1365                         dprintk(" OK\n");
1366                         goto next;
1367                 }
1368                 /* We have a mismatch. Write the page again */
1369                 dprintk(" mismatch\n");
1370                 nfs_mark_request_dirty(req);
1371         next:
1372                 nfs_clear_page_tag_locked(req);
1373         }
1374         nfs_commitdata_release(calldata);
1375 }
1376
1377 static const struct rpc_call_ops nfs_commit_ops = {
1378 #if defined(CONFIG_NFS_V4_1)
1379         .rpc_call_prepare = nfs_write_prepare,
1380 #endif /* CONFIG_NFS_V4_1 */
1381         .rpc_call_done = nfs_commit_done,
1382         .rpc_release = nfs_commit_release,
1383 };
1384
1385 int nfs_commit_inode(struct inode *inode, int how)
1386 {
1387         LIST_HEAD(head);
1388         int res;
1389
1390         spin_lock(&inode->i_lock);
1391         res = nfs_scan_commit(inode, &head, 0, 0);
1392         spin_unlock(&inode->i_lock);
1393         if (res) {
1394                 int error = nfs_commit_list(inode, &head, how);
1395                 if (error < 0)
1396                         return error;
1397         }
1398         return res;
1399 }
1400 #else
1401 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1402 {
1403         return 0;
1404 }
1405 #endif
1406
1407 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1408 {
1409         struct inode *inode = mapping->host;
1410         pgoff_t idx_start, idx_end;
1411         unsigned int npages = 0;
1412         LIST_HEAD(head);
1413         int nocommit = how & FLUSH_NOCOMMIT;
1414         long pages, ret;
1415
1416         /* FIXME */
1417         if (wbc->range_cyclic)
1418                 idx_start = 0;
1419         else {
1420                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1421                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1422                 if (idx_end > idx_start) {
1423                         pgoff_t l_npages = 1 + idx_end - idx_start;
1424                         npages = l_npages;
1425                         if (sizeof(npages) != sizeof(l_npages) &&
1426                                         (pgoff_t)npages != l_npages)
1427                                 npages = 0;
1428                 }
1429         }
1430         how &= ~FLUSH_NOCOMMIT;
1431         spin_lock(&inode->i_lock);
1432         do {
1433                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1434                 if (ret != 0)
1435                         continue;
1436                 if (nocommit)
1437                         break;
1438                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1439                 if (pages == 0)
1440                         break;
1441                 if (how & FLUSH_INVALIDATE) {
1442                         spin_unlock(&inode->i_lock);
1443                         nfs_cancel_commit_list(&head);
1444                         ret = pages;
1445                         spin_lock(&inode->i_lock);
1446                         continue;
1447                 }
1448                 pages += nfs_scan_commit(inode, &head, 0, 0);
1449                 spin_unlock(&inode->i_lock);
1450                 ret = nfs_commit_list(inode, &head, how);
1451                 spin_lock(&inode->i_lock);
1452
1453         } while (ret >= 0);
1454         spin_unlock(&inode->i_lock);
1455         return ret;
1456 }
1457
1458 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1459 {
1460         int ret;
1461
1462         ret = nfs_writepages(mapping, wbc);
1463         if (ret < 0)
1464                 goto out;
1465         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1466         if (ret < 0)
1467                 goto out;
1468         return 0;
1469 out:
1470         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1471         return ret;
1472 }
1473
1474 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1475 static int nfs_write_mapping(struct address_space *mapping, int how)
1476 {
1477         struct writeback_control wbc = {
1478                 .bdi = mapping->backing_dev_info,
1479                 .sync_mode = WB_SYNC_ALL,
1480                 .nr_to_write = LONG_MAX,
1481                 .range_start = 0,
1482                 .range_end = LLONG_MAX,
1483                 .for_writepages = 1,
1484         };
1485
1486         return __nfs_write_mapping(mapping, &wbc, how);
1487 }
1488
1489 /*
1490  * flush the inode to disk.
1491  */
1492 int nfs_wb_all(struct inode *inode)
1493 {
1494         return nfs_write_mapping(inode->i_mapping, 0);
1495 }
1496
1497 int nfs_wb_nocommit(struct inode *inode)
1498 {
1499         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1500 }
1501
1502 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1503 {
1504         struct nfs_page *req;
1505         loff_t range_start = page_offset(page);
1506         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1507         struct writeback_control wbc = {
1508                 .bdi = page->mapping->backing_dev_info,
1509                 .sync_mode = WB_SYNC_ALL,
1510                 .nr_to_write = LONG_MAX,
1511                 .range_start = range_start,
1512                 .range_end = range_end,
1513         };
1514         int ret = 0;
1515
1516         BUG_ON(!PageLocked(page));
1517         for (;;) {
1518                 req = nfs_page_find_request(page);
1519                 if (req == NULL)
1520                         goto out;
1521                 if (test_bit(PG_CLEAN, &req->wb_flags)) {
1522                         nfs_release_request(req);
1523                         break;
1524                 }
1525                 if (nfs_lock_request_dontget(req)) {
1526                         nfs_inode_remove_request(req);
1527                         /*
1528                          * In case nfs_inode_remove_request has marked the
1529                          * page as being dirty
1530                          */
1531                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1532                         nfs_unlock_request(req);
1533                         break;
1534                 }
1535                 ret = nfs_wait_on_request(req);
1536                 if (ret < 0)
1537                         goto out;
1538         }
1539         if (!PagePrivate(page))
1540                 return 0;
1541         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1542 out:
1543         return ret;
1544 }
1545
1546 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1547                                 int how)
1548 {
1549         loff_t range_start = page_offset(page);
1550         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1551         struct writeback_control wbc = {
1552                 .bdi = page->mapping->backing_dev_info,
1553                 .sync_mode = WB_SYNC_ALL,
1554                 .nr_to_write = LONG_MAX,
1555                 .range_start = range_start,
1556                 .range_end = range_end,
1557         };
1558         int ret;
1559
1560         do {
1561                 if (clear_page_dirty_for_io(page)) {
1562                         ret = nfs_writepage_locked(page, &wbc);
1563                         if (ret < 0)
1564                                 goto out_error;
1565                 } else if (!PagePrivate(page))
1566                         break;
1567                 ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1568                 if (ret < 0)
1569                         goto out_error;
1570         } while (PagePrivate(page));
1571         return 0;
1572 out_error:
1573         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1574         return ret;
1575 }
1576
1577 /*
1578  * Write back all requests on one page - we do this before reading it.
1579  */
1580 int nfs_wb_page(struct inode *inode, struct page* page)
1581 {
1582         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1583 }
1584
1585 int __init nfs_init_writepagecache(void)
1586 {
1587         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1588                                              sizeof(struct nfs_write_data),
1589                                              0, SLAB_HWCACHE_ALIGN,
1590                                              NULL);
1591         if (nfs_wdata_cachep == NULL)
1592                 return -ENOMEM;
1593
1594         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1595                                                      nfs_wdata_cachep);
1596         if (nfs_wdata_mempool == NULL)
1597                 return -ENOMEM;
1598
1599         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1600                                                       nfs_wdata_cachep);
1601         if (nfs_commit_mempool == NULL)
1602                 return -ENOMEM;
1603
1604         /*
1605          * NFS congestion size, scale with available memory.
1606          *
1607          *  64MB:    8192k
1608          * 128MB:   11585k
1609          * 256MB:   16384k
1610          * 512MB:   23170k
1611          *   1GB:   32768k
1612          *   2GB:   46340k
1613          *   4GB:   65536k
1614          *   8GB:   92681k
1615          *  16GB:  131072k
1616          *
1617          * This allows larger machines to have larger/more transfers.
1618          * Limit the default to 256M
1619          */
1620         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1621         if (nfs_congestion_kb > 256*1024)
1622                 nfs_congestion_kb = 256*1024;
1623
1624         return 0;
1625 }
1626
1627 void nfs_destroy_writepagecache(void)
1628 {
1629         mempool_destroy(nfs_commit_mempool);
1630         mempool_destroy(nfs_wdata_mempool);
1631         kmem_cache_destroy(nfs_wdata_cachep);
1632 }
1633