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