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
  29 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  30
  31 #define MIN_POOL_WRITE          (32)
  32 #define MIN_POOL_COMMIT         (4)
  33
  34 /*
  35  * Local function declarations
  36  */
  37 static struct nfs_page * nfs_update_request(struct nfs_open_context*,
  38                                             struct page *,
  39                                             unsigned int, unsigned int);
  40 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  41                                   struct inode *inode, int ioflags);
  42 static const struct rpc_call_ops nfs_write_partial_ops;
  43 static const struct rpc_call_ops nfs_write_full_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45
  46 static struct kmem_cache *nfs_wdata_cachep;
  47 static mempool_t *nfs_wdata_mempool;
  48 static mempool_t *nfs_commit_mempool;
  49
  50 struct nfs_write_data *nfs_commit_alloc(void)
  51 {
  52         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  53
  54         if (p) {
  55                 memset(p, 0, sizeof(*p));
  56                 INIT_LIST_HEAD(&p->pages);
  57         }
  58         return p;
  59 }
  60
  61 static void nfs_commit_rcu_free(struct rcu_head *head)
  62 {
  63         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  64         if (p && (p->pagevec != &p->page_array[0]))
  65                 kfree(p->pagevec);
  66         mempool_free(p, nfs_commit_mempool);
  67 }
  68
  69 void nfs_commit_free(struct nfs_write_data *wdata)
  70 {
  71         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_commit_rcu_free);
  72 }
  73
  74 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  75 {
  76         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  77
  78         if (p) {
  79                 memset(p, 0, sizeof(*p));
  80                 INIT_LIST_HEAD(&p->pages);
  81                 p->npages = pagecount;
  82                 if (pagecount <= ARRAY_SIZE(p->page_array))
  83                         p->pagevec = p->page_array;
  84                 else {
  85                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  86                         if (!p->pagevec) {
  87                                 mempool_free(p, nfs_wdata_mempool);
  88                                 p = NULL;
  89                         }
  90                 }
  91         }
  92         return p;
  93 }
  94
  95 static void nfs_writedata_rcu_free(struct rcu_head *head)
  96 {
  97         struct nfs_write_data *p = container_of(head, struct nfs_write_data, task.u.tk_rcu);
  98         if (p && (p->pagevec != &p->page_array[0]))
  99                 kfree(p->pagevec);
 100         mempool_free(p, nfs_wdata_mempool);
 101 }
 102
 103 static void nfs_writedata_free(struct nfs_write_data *wdata)
 104 {
 105         call_rcu_bh(&wdata->task.u.tk_rcu, nfs_writedata_rcu_free);
 106 }
 107
 108 void nfs_writedata_release(void *wdata)
 109 {
 110         nfs_writedata_free(wdata);
 111 }
 112
 113 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 114 {
 115         ctx->error = error;
 116         smp_wmb();
 117         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 118 }
 119
 120 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 121 {
 122         struct nfs_page *req = NULL;
 123
 124         if (PagePrivate(page)) {
 125                 req = (struct nfs_page *)page_private(page);
 126                 if (req != NULL)
 127                         kref_get(&req->wb_kref);
 128         }
 129         return req;
 130 }
 131
 132 static struct nfs_page *nfs_page_find_request(struct page *page)
 133 {
 134         struct inode *inode = page->mapping->host;
 135         struct nfs_page *req = NULL;
 136
 137         spin_lock(&inode->i_lock);
 138         req = nfs_page_find_request_locked(page);
 139         spin_unlock(&inode->i_lock);
 140         return req;
 141 }
 142
 143 /* Adjust the file length if we're writing beyond the end */
 144 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 145 {
 146         struct inode *inode = page->mapping->host;
 147         loff_t end, i_size = i_size_read(inode);
 148         pgoff_t end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 149
 150         if (i_size > 0 && page->index < end_index)
 151                 return;
 152         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 153         if (i_size >= end)
 154                 return;
 155         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 156         i_size_write(inode, end);
 157 }
 158
 159 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 160 static void nfs_set_pageerror(struct page *page)
 161 {
 162         SetPageError(page);
 163         nfs_zap_mapping(page->mapping->host, page->mapping);
 164 }
 165
 166 /* We can set the PG_uptodate flag if we see that a write request
 167  * covers the full page.
 168  */
 169 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 170 {
 171         if (PageUptodate(page))
 172                 return;
 173         if (base != 0)
 174                 return;
 175         if (count != nfs_page_length(page))
 176                 return;
 177         SetPageUptodate(page);
 178 }
 179
 180 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 181                 unsigned int offset, unsigned int count)
 182 {
 183         struct nfs_page *req;
 184         int ret;
 185
 186         for (;;) {
 187                 req = nfs_update_request(ctx, page, offset, count);
 188                 if (!IS_ERR(req))
 189                         break;
 190                 ret = PTR_ERR(req);
 191                 if (ret != -EBUSY)
 192                         return ret;
 193                 ret = nfs_wb_page(page->mapping->host, page);
 194                 if (ret != 0)
 195                         return ret;
 196         }
 197         /* Update file length */
 198         nfs_grow_file(page, offset, count);
 199         nfs_clear_page_tag_locked(req);
 200         return 0;
 201 }
 202
 203 static int wb_priority(struct writeback_control *wbc)
 204 {
 205         if (wbc->for_reclaim)
 206                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 207         if (wbc->for_kupdate)
 208                 return FLUSH_LOWPRI;
 209         return 0;
 210 }
 211
 212 /*
 213  * NFS congestion control
 214  */
 215
 216 int nfs_congestion_kb;
 217
 218 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 219 #define NFS_CONGESTION_OFF_THRESH       \
 220         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 221
 222 static int nfs_set_page_writeback(struct page *page)
 223 {
 224         int ret = test_set_page_writeback(page);
 225
 226         if (!ret) {
 227                 struct inode *inode = page->mapping->host;
 228                 struct nfs_server *nfss = NFS_SERVER(inode);
 229
 230                 if (atomic_long_inc_return(&nfss->writeback) >
 231                                 NFS_CONGESTION_ON_THRESH)
 232                         set_bdi_congested(&nfss->backing_dev_info, WRITE);
 233         }
 234         return ret;
 235 }
 236
 237 static void nfs_end_page_writeback(struct page *page)
 238 {
 239         struct inode *inode = page->mapping->host;
 240         struct nfs_server *nfss = NFS_SERVER(inode);
 241
 242         end_page_writeback(page);
 243         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 244                 clear_bdi_congested(&nfss->backing_dev_info, WRITE);
 245 }
 246
 247 /*
 248  * Find an associated nfs write request, and prepare to flush it out
 249  * May return an error if the user signalled nfs_wait_on_request().
 250  */
 251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 252                                 struct page *page)
 253 {
 254         struct inode *inode = page->mapping->host;
 255         struct nfs_page *req;
 256         int ret;
 257
 258         spin_lock(&inode->i_lock);
 259         for(;;) {
 260                 req = nfs_page_find_request_locked(page);
 261                 if (req == NULL) {
 262                         spin_unlock(&inode->i_lock);
 263                         return 0;
 264                 }
 265                 if (nfs_set_page_tag_locked(req))
 266                         break;
 267                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 268                  *       then the call to nfs_set_page_tag_locked() will always
 269                  *       succeed provided that someone hasn't already marked the
 270                  *       request as dirty (in which case we don't care).
 271                  */
 272                 spin_unlock(&inode->i_lock);
 273                 ret = nfs_wait_on_request(req);
 274                 nfs_release_request(req);
 275                 if (ret != 0)
 276                         return ret;
 277                 spin_lock(&inode->i_lock);
 278         }
 279         if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 280                 /* This request is marked for commit */
 281                 spin_unlock(&inode->i_lock);
 282                 nfs_clear_page_tag_locked(req);
 283                 nfs_pageio_complete(pgio);
 284                 return 0;
 285         }
 286         if (nfs_set_page_writeback(page) != 0) {
 287                 spin_unlock(&inode->i_lock);
 288                 BUG();
 289         }
 290         spin_unlock(&inode->i_lock);
 291         nfs_pageio_add_request(pgio, req);
 292         return 0;
 293 }
 294
 295 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 296 {
 297         struct inode *inode = page->mapping->host;
 298
 299         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 300         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 301
 302         nfs_pageio_cond_complete(pgio, page->index);
 303         return nfs_page_async_flush(pgio, page);
 304 }
 305
 306 /*
 307  * Write an mmapped page to the server.
 308  */
 309 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 310 {
 311         struct nfs_pageio_descriptor pgio;
 312         int err;
 313
 314         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 315         err = nfs_do_writepage(page, wbc, &pgio);
 316         nfs_pageio_complete(&pgio);
 317         if (err < 0)
 318                 return err;
 319         if (pgio.pg_error < 0)
 320                 return pgio.pg_error;
 321         return 0;
 322 }
 323
 324 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 325 {
 326         int ret;
 327
 328         ret = nfs_writepage_locked(page, wbc);
 329         unlock_page(page);
 330         return ret;
 331 }
 332
 333 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 334 {
 335         int ret;
 336
 337         ret = nfs_do_writepage(page, wbc, data);
 338         unlock_page(page);
 339         return ret;
 340 }
 341
 342 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 343 {
 344         struct inode *inode = mapping->host;
 345         struct nfs_pageio_descriptor pgio;
 346         int err;
 347
 348         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 349
 350         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 351         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 352         nfs_pageio_complete(&pgio);
 353         if (err < 0)
 354                 return err;
 355         if (pgio.pg_error < 0)
 356                 return pgio.pg_error;
 357         return 0;
 358 }
 359
 360 /*
 361  * Insert a write request into an inode
 362  */
 363 static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 364 {
 365         struct nfs_inode *nfsi = NFS_I(inode);
 366         int error;
 367
 368         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 369         BUG_ON(error == -EEXIST);
 370         if (error)
 371                 return error;
 372         if (!nfsi->npages) {
 373                 igrab(inode);
 374                 if (nfs_have_delegation(inode, FMODE_WRITE))
 375                         nfsi->change_attr++;
 376         }
 377         SetPagePrivate(req->wb_page);
 378         set_page_private(req->wb_page, (unsigned long)req);
 379         nfsi->npages++;
 380         kref_get(&req->wb_kref);
 381         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 382         return 0;
 383 }
 384
 385 /*
 386  * Remove a write request from an inode
 387  */
 388 static void nfs_inode_remove_request(struct nfs_page *req)
 389 {
 390         struct inode *inode = req->wb_context->path.dentry->d_inode;
 391         struct nfs_inode *nfsi = NFS_I(inode);
 392
 393         BUG_ON (!NFS_WBACK_BUSY(req));
 394
 395         spin_lock(&inode->i_lock);
 396         set_page_private(req->wb_page, 0);
 397         ClearPagePrivate(req->wb_page);
 398         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 399         nfsi->npages--;
 400         if (!nfsi->npages) {
 401                 spin_unlock(&inode->i_lock);
 402                 iput(inode);
 403         } else
 404                 spin_unlock(&inode->i_lock);
 405         nfs_clear_request(req);
 406         nfs_release_request(req);
 407 }
 408
 409 static void
 410 nfs_redirty_request(struct nfs_page *req)
 411 {
 412         __set_page_dirty_nobuffers(req->wb_page);
 413 }
 414
 415 /*
 416  * Check if a request is dirty
 417  */
 418 static inline int
 419 nfs_dirty_request(struct nfs_page *req)
 420 {
 421         struct page *page = req->wb_page;
 422
 423         if (page == NULL || test_bit(PG_NEED_COMMIT, &req->wb_flags))
 424                 return 0;
 425         return !PageWriteback(req->wb_page);
 426 }
 427
 428 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 429 /*
 430  * Add a request to the inode's commit list.
 431  */
 432 static void
 433 nfs_mark_request_commit(struct nfs_page *req)
 434 {
 435         struct inode *inode = req->wb_context->path.dentry->d_inode;
 436         struct nfs_inode *nfsi = NFS_I(inode);
 437
 438         spin_lock(&inode->i_lock);
 439         nfsi->ncommit++;
 440         set_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 441         radix_tree_tag_set(&nfsi->nfs_page_tree,
 442                         req->wb_index,
 443                         NFS_PAGE_TAG_COMMIT);
 444         spin_unlock(&inode->i_lock);
 445         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 446         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 447         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 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_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_redirty_request(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
 476 int nfs_write_need_commit(struct nfs_write_data *data)
 477 {
 478         return 0;
 479 }
 480
 481 static inline
 482 int nfs_reschedule_unstable_write(struct nfs_page *req)
 483 {
 484         return 0;
 485 }
 486 #endif
 487
 488 /*
 489  * Wait for a request to complete.
 490  *
 491  * Interruptible by fatal signals only.
 492  */
 493 static int nfs_wait_on_requests_locked(struct inode *inode, pgoff_t idx_start, unsigned int npages)
 494 {
 495         struct nfs_inode *nfsi = NFS_I(inode);
 496         struct nfs_page *req;
 497         pgoff_t idx_end, next;
 498         unsigned int            res = 0;
 499         int                     error;
 500
 501         if (npages == 0)
 502                 idx_end = ~0;
 503         else
 504                 idx_end = idx_start + npages - 1;
 505
 506         next = idx_start;
 507         while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_LOCKED)) {
 508                 if (req->wb_index > idx_end)
 509                         break;
 510
 511                 next = req->wb_index + 1;
 512                 BUG_ON(!NFS_WBACK_BUSY(req));
 513
 514                 kref_get(&req->wb_kref);
 515                 spin_unlock(&inode->i_lock);
 516                 error = nfs_wait_on_request(req);
 517                 nfs_release_request(req);
 518                 spin_lock(&inode->i_lock);
 519                 if (error < 0)
 520                         return error;
 521                 res++;
 522         }
 523         return res;
 524 }
 525
 526 static void nfs_cancel_commit_list(struct list_head *head)
 527 {
 528         struct nfs_page *req;
 529
 530         while(!list_empty(head)) {
 531                 req = nfs_list_entry(head->next);
 532                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 533                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
 534                                 BDI_RECLAIMABLE);
 535                 nfs_list_remove_request(req);
 536                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
 537                 nfs_inode_remove_request(req);
 538                 nfs_unlock_request(req);
 539         }
 540 }
 541
 542 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 543 /*
 544  * nfs_scan_commit - Scan an inode for commit requests
 545  * @inode: NFS inode to scan
 546  * @dst: destination list
 547  * @idx_start: lower bound of page->index to scan.
 548  * @npages: idx_start + npages sets the upper bound to scan.
 549  *
 550  * Moves requests from the inode's 'commit' request list.
 551  * The requests are *not* checked to ensure that they form a contiguous set.
 552  */
 553 static int
 554 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 555 {
 556         struct nfs_inode *nfsi = NFS_I(inode);
 557         int res = 0;
 558
 559         if (nfsi->ncommit != 0) {
 560                 res = nfs_scan_list(nfsi, dst, idx_start, npages,
 561                                 NFS_PAGE_TAG_COMMIT);
 562                 nfsi->ncommit -= res;
 563         }
 564         return res;
 565 }
 566 #else
 567 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 568 {
 569         return 0;
 570 }
 571 #endif
 572
 573 /*
 574  * Try to update any existing write request, or create one if there is none.
 575  * In order to match, the request's credentials must match those of
 576  * the calling process.
 577  *
 578  * Note: Should always be called with the Page Lock held!
 579  */
 580 static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx,
 581                 struct page *page, unsigned int offset, unsigned int bytes)
 582 {
 583         struct address_space *mapping = page->mapping;
 584         struct inode *inode = mapping->host;
 585         struct nfs_page         *req, *new = NULL;
 586         pgoff_t         rqend, end;
 587
 588         end = offset + bytes;
 589
 590         for (;;) {
 591                 /* Loop over all inode entries and see if we find
 592                  * A request for the page we wish to update
 593                  */
 594                 spin_lock(&inode->i_lock);
 595                 req = nfs_page_find_request_locked(page);
 596                 if (req) {
 597                         if (!nfs_set_page_tag_locked(req)) {
 598                                 int error;
 599
 600                                 spin_unlock(&inode->i_lock);
 601                                 error = nfs_wait_on_request(req);
 602                                 nfs_release_request(req);
 603                                 if (error < 0) {
 604                                         if (new)
 605                                                 nfs_release_request(new);
 606                                         return ERR_PTR(error);
 607                                 }
 608                                 continue;
 609                         }
 610                         spin_unlock(&inode->i_lock);
 611                         if (new)
 612                                 nfs_release_request(new);
 613                         break;
 614                 }
 615
 616                 if (new) {
 617                         int error;
 618                         nfs_lock_request_dontget(new);
 619                         error = nfs_inode_add_request(inode, new);
 620                         if (error) {
 621                                 spin_unlock(&inode->i_lock);
 622                                 nfs_unlock_request(new);
 623                                 return ERR_PTR(error);
 624                         }
 625                         spin_unlock(&inode->i_lock);
 626                         req = new;
 627                         goto zero_page;
 628                 }
 629                 spin_unlock(&inode->i_lock);
 630
 631                 new = nfs_create_request(ctx, inode, page, offset, bytes);
 632                 if (IS_ERR(new))
 633                         return new;
 634         }
 635
 636         /* We have a request for our page.
 637          * If the creds don't match, or the
 638          * page addresses don't match,
 639          * tell the caller to wait on the conflicting
 640          * request.
 641          */
 642         rqend = req->wb_offset + req->wb_bytes;
 643         if (req->wb_context != ctx
 644             || req->wb_page != page
 645             || !nfs_dirty_request(req)
 646             || offset > rqend || end < req->wb_offset) {
 647                 nfs_clear_page_tag_locked(req);
 648                 return ERR_PTR(-EBUSY);
 649         }
 650
 651         /* Okay, the request matches. Update the region */
 652         if (offset < req->wb_offset) {
 653                 req->wb_offset = offset;
 654                 req->wb_pgbase = offset;
 655                 req->wb_bytes = max(end, rqend) - req->wb_offset;
 656                 goto zero_page;
 657         }
 658
 659         if (end > rqend)
 660                 req->wb_bytes = end - req->wb_offset;
 661
 662         return req;
 663 zero_page:
 664         /* If this page might potentially be marked as up to date,
 665          * then we need to zero any uninitalised data. */
 666         if (req->wb_pgbase == 0 && req->wb_bytes != PAGE_CACHE_SIZE
 667                         && !PageUptodate(req->wb_page))
 668                 zero_user_segment(req->wb_page, req->wb_bytes, PAGE_CACHE_SIZE);
 669         return req;
 670 }
 671
 672 int nfs_flush_incompatible(struct file *file, struct page *page)
 673 {
 674         struct nfs_open_context *ctx = nfs_file_open_context(file);
 675         struct nfs_page *req;
 676         int do_flush, status;
 677         /*
 678          * Look for a request corresponding to this page. If there
 679          * is one, and it belongs to another file, we flush it out
 680          * before we try to copy anything into the page. Do this
 681          * due to the lack of an ACCESS-type call in NFSv2.
 682          * Also do the same if we find a request from an existing
 683          * dropped page.
 684          */
 685         do {
 686                 req = nfs_page_find_request(page);
 687                 if (req == NULL)
 688                         return 0;
 689                 do_flush = req->wb_page != page || req->wb_context != ctx
 690                         || !nfs_dirty_request(req);
 691                 nfs_release_request(req);
 692                 if (!do_flush)
 693                         return 0;
 694                 status = nfs_wb_page(page->mapping->host, page);
 695         } while (status == 0);
 696         return status;
 697 }
 698
 699 /*
 700  * If the page cache is marked as unsafe or invalid, then we can't rely on
 701  * the PageUptodate() flag. In this case, we will need to turn off
 702  * write optimisations that depend on the page contents being correct.
 703  */
 704 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 705 {
 706         return PageUptodate(page) &&
 707                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 708 }
 709
 710 /*
 711  * Update and possibly write a cached page of an NFS file.
 712  *
 713  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 714  * things with a page scheduled for an RPC call (e.g. invalidate it).
 715  */
 716 int nfs_updatepage(struct file *file, struct page *page,
 717                 unsigned int offset, unsigned int count)
 718 {
 719         struct nfs_open_context *ctx = nfs_file_open_context(file);
 720         struct inode    *inode = page->mapping->host;
 721         int             status = 0;
 722
 723         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 724
 725         dprintk("NFS:      nfs_updatepage(%s/%s %d@%Ld)\n",
 726                 file->f_path.dentry->d_parent->d_name.name,
 727                 file->f_path.dentry->d_name.name, count,
 728                 (long long)(page_offset(page) +offset));
 729
 730         /* If we're not using byte range locks, and we know the page
 731          * is up to date, it may be more efficient to extend the write
 732          * to cover the entire page in order to avoid fragmentation
 733          * inefficiencies.
 734          */
 735         if (nfs_write_pageuptodate(page, inode) &&
 736                         inode->i_flock == NULL &&
 737                         !(file->f_mode & O_SYNC)) {
 738                 count = max(count + offset, nfs_page_length(page));
 739                 offset = 0;
 740         }
 741
 742         status = nfs_writepage_setup(ctx, page, offset, count);
 743         __set_page_dirty_nobuffers(page);
 744
 745         dprintk("NFS:      nfs_updatepage returns %d (isize %Ld)\n",
 746                         status, (long long)i_size_read(inode));
 747         if (status < 0)
 748                 nfs_set_pageerror(page);
 749         return status;
 750 }
 751
 752 static void nfs_writepage_release(struct nfs_page *req)
 753 {
 754
 755         if (PageError(req->wb_page)) {
 756                 nfs_end_page_writeback(req->wb_page);
 757                 nfs_inode_remove_request(req);
 758         } else if (!nfs_reschedule_unstable_write(req)) {
 759                 /* Set the PG_uptodate flag */
 760                 nfs_mark_uptodate(req->wb_page, req->wb_pgbase, req->wb_bytes);
 761                 nfs_end_page_writeback(req->wb_page);
 762                 nfs_inode_remove_request(req);
 763         } else
 764                 nfs_end_page_writeback(req->wb_page);
 765         nfs_clear_page_tag_locked(req);
 766 }
 767
 768 static int flush_task_priority(int how)
 769 {
 770         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 771                 case FLUSH_HIGHPRI:
 772                         return RPC_PRIORITY_HIGH;
 773                 case FLUSH_LOWPRI:
 774                         return RPC_PRIORITY_LOW;
 775         }
 776         return RPC_PRIORITY_NORMAL;
 777 }
 778
 779 /*
 780  * Set up the argument/result storage required for the RPC call.
 781  */
 782 static void nfs_write_rpcsetup(struct nfs_page *req,
 783                 struct nfs_write_data *data,
 784                 const struct rpc_call_ops *call_ops,
 785                 unsigned int count, unsigned int offset,
 786                 int how)
 787 {
 788         struct inode *inode = req->wb_context->path.dentry->d_inode;
 789         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
 790         int priority = flush_task_priority(how);
 791         struct rpc_task *task;
 792         struct rpc_message msg = {
 793                 .rpc_argp = &data->args,
 794                 .rpc_resp = &data->res,
 795                 .rpc_cred = req->wb_context->cred,
 796         };
 797         struct rpc_task_setup task_setup_data = {
 798                 .rpc_client = NFS_CLIENT(inode),
 799                 .task = &data->task,
 800                 .rpc_message = &msg,
 801                 .callback_ops = call_ops,
 802                 .callback_data = data,
 803                 .flags = flags,
 804                 .priority = priority,
 805         };
 806
 807         /* Set up the RPC argument and reply structs
 808          * NB: take care not to mess about with data->commit et al. */
 809
 810         data->req = req;
 811         data->inode = inode = req->wb_context->path.dentry->d_inode;
 812         data->cred = msg.rpc_cred;
 813
 814         data->args.fh     = NFS_FH(inode);
 815         data->args.offset = req_offset(req) + offset;
 816         data->args.pgbase = req->wb_pgbase + offset;
 817         data->args.pages  = data->pagevec;
 818         data->args.count  = count;
 819         data->args.context = req->wb_context;
 820         data->args.stable  = NFS_UNSTABLE;
 821         if (how & FLUSH_STABLE) {
 822                 data->args.stable = NFS_DATA_SYNC;
 823                 if (!NFS_I(inode)->ncommit)
 824                         data->args.stable = NFS_FILE_SYNC;
 825         }
 826
 827         data->res.fattr   = &data->fattr;
 828         data->res.count   = count;
 829         data->res.verf    = &data->verf;
 830         nfs_fattr_init(&data->fattr);
 831
 832         /* Set up the initial task struct.  */
 833         NFS_PROTO(inode)->write_setup(data, &msg);
 834
 835         dprintk("NFS: %5u initiated write call "
 836                 "(req %s/%Ld, %u bytes @ offset %Lu)\n",
 837                 data->task.tk_pid,
 838                 inode->i_sb->s_id,
 839                 (long long)NFS_FILEID(inode),
 840                 count,
 841                 (unsigned long long)data->args.offset);
 842
 843         task = rpc_run_task(&task_setup_data);
 844         if (!IS_ERR(task))
 845                 rpc_put_task(task);
 846 }
 847
 848 /*
 849  * Generate multiple small requests to write out a single
 850  * contiguous dirty area on one page.
 851  */
 852 static int nfs_flush_multi(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 853 {
 854         struct nfs_page *req = nfs_list_entry(head->next);
 855         struct page *page = req->wb_page;
 856         struct nfs_write_data *data;
 857         size_t wsize = NFS_SERVER(inode)->wsize, nbytes;
 858         unsigned int offset;
 859         int requests = 0;
 860         LIST_HEAD(list);
 861
 862         nfs_list_remove_request(req);
 863
 864         nbytes = count;
 865         do {
 866                 size_t len = min(nbytes, wsize);
 867
 868                 data = nfs_writedata_alloc(1);
 869                 if (!data)
 870                         goto out_bad;
 871                 list_add(&data->pages, &list);
 872                 requests++;
 873                 nbytes -= len;
 874         } while (nbytes != 0);
 875         atomic_set(&req->wb_complete, requests);
 876
 877         ClearPageError(page);
 878         offset = 0;
 879         nbytes = count;
 880         do {
 881                 data = list_entry(list.next, struct nfs_write_data, pages);
 882                 list_del_init(&data->pages);
 883
 884                 data->pagevec[0] = page;
 885
 886                 if (nbytes < wsize)
 887                         wsize = nbytes;
 888                 nfs_write_rpcsetup(req, data, &nfs_write_partial_ops,
 889                                    wsize, offset, how);
 890                 offset += wsize;
 891                 nbytes -= wsize;
 892         } while (nbytes != 0);
 893
 894         return 0;
 895
 896 out_bad:
 897         while (!list_empty(&list)) {
 898                 data = list_entry(list.next, struct nfs_write_data, pages);
 899                 list_del(&data->pages);
 900                 nfs_writedata_release(data);
 901         }
 902         nfs_redirty_request(req);
 903         nfs_end_page_writeback(req->wb_page);
 904         nfs_clear_page_tag_locked(req);
 905         return -ENOMEM;
 906 }
 907
 908 /*
 909  * Create an RPC task for the given write request and kick it.
 910  * The page must have been locked by the caller.
 911  *
 912  * It may happen that the page we're passed is not marked dirty.
 913  * This is the case if nfs_updatepage detects a conflicting request
 914  * that has been written but not committed.
 915  */
 916 static int nfs_flush_one(struct inode *inode, struct list_head *head, unsigned int npages, size_t count, int how)
 917 {
 918         struct nfs_page         *req;
 919         struct page             **pages;
 920         struct nfs_write_data   *data;
 921
 922         data = nfs_writedata_alloc(npages);
 923         if (!data)
 924                 goto out_bad;
 925
 926         pages = data->pagevec;
 927         while (!list_empty(head)) {
 928                 req = nfs_list_entry(head->next);
 929                 nfs_list_remove_request(req);
 930                 nfs_list_add_request(req, &data->pages);
 931                 ClearPageError(req->wb_page);
 932                 *pages++ = req->wb_page;
 933         }
 934         req = nfs_list_entry(data->pages.next);
 935
 936         /* Set up the argument struct */
 937         nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how);
 938
 939         return 0;
 940  out_bad:
 941         while (!list_empty(head)) {
 942                 req = nfs_list_entry(head->next);
 943                 nfs_list_remove_request(req);
 944                 nfs_redirty_request(req);
 945                 nfs_end_page_writeback(req->wb_page);
 946                 nfs_clear_page_tag_locked(req);
 947         }
 948         return -ENOMEM;
 949 }
 950
 951 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
 952                                   struct inode *inode, int ioflags)
 953 {
 954         size_t wsize = NFS_SERVER(inode)->wsize;
 955
 956         if (wsize < PAGE_CACHE_SIZE)
 957                 nfs_pageio_init(pgio, inode, nfs_flush_multi, wsize, ioflags);
 958         else
 959                 nfs_pageio_init(pgio, inode, nfs_flush_one, wsize, ioflags);
 960 }
 961
 962 /*
 963  * Handle a write reply that flushed part of a page.
 964  */
 965 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
 966 {
 967         struct nfs_write_data   *data = calldata;
 968         struct nfs_page         *req = data->req;
 969         struct page             *page = req->wb_page;
 970
 971         dprintk("NFS: write (%s/%Ld %d@%Ld)",
 972                 req->wb_context->path.dentry->d_inode->i_sb->s_id,
 973                 (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
 974                 req->wb_bytes,
 975                 (long long)req_offset(req));
 976
 977         if (nfs_writeback_done(task, data) != 0)
 978                 return;
 979
 980         if (task->tk_status < 0) {
 981                 nfs_set_pageerror(page);
 982                 nfs_context_set_write_error(req->wb_context, task->tk_status);
 983                 dprintk(", error = %d\n", task->tk_status);
 984                 goto out;
 985         }
 986
 987         if (nfs_write_need_commit(data)) {
 988                 struct inode *inode = page->mapping->host;
 989
 990                 spin_lock(&inode->i_lock);
 991                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 992                         /* Do nothing we need to resend the writes */
 993                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 994                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
 995                         dprintk(" defer commit\n");
 996                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
 997                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
 998                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
 999                         dprintk(" server reboot detected\n");
1000                 }
1001                 spin_unlock(&inode->i_lock);
1002         } else
1003                 dprintk(" OK\n");
1004
1005 out:
1006         if (atomic_dec_and_test(&req->wb_complete))
1007                 nfs_writepage_release(req);
1008 }
1009
1010 static const struct rpc_call_ops nfs_write_partial_ops = {
1011         .rpc_call_done = nfs_writeback_done_partial,
1012         .rpc_release = nfs_writedata_release,
1013 };
1014
1015 /*
1016  * Handle a write reply that flushes a whole page.
1017  *
1018  * FIXME: There is an inherent race with invalidate_inode_pages and
1019  *        writebacks since the page->count is kept > 1 for as long
1020  *        as the page has a write request pending.
1021  */
1022 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1023 {
1024         struct nfs_write_data   *data = calldata;
1025         struct nfs_page         *req;
1026         struct page             *page;
1027
1028         if (nfs_writeback_done(task, data) != 0)
1029                 return;
1030
1031         /* Update attributes as result of writeback. */
1032         while (!list_empty(&data->pages)) {
1033                 req = nfs_list_entry(data->pages.next);
1034                 nfs_list_remove_request(req);
1035                 page = req->wb_page;
1036
1037                 dprintk("NFS: write (%s/%Ld %d@%Ld)",
1038                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1039                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1040                         req->wb_bytes,
1041                         (long long)req_offset(req));
1042
1043                 if (task->tk_status < 0) {
1044                         nfs_set_pageerror(page);
1045                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1046                         dprintk(", error = %d\n", task->tk_status);
1047                         goto remove_request;
1048                 }
1049
1050                 if (nfs_write_need_commit(data)) {
1051                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1052                         nfs_mark_request_commit(req);
1053                         nfs_end_page_writeback(page);
1054                         dprintk(" marked for commit\n");
1055                         goto next;
1056                 }
1057                 /* Set the PG_uptodate flag? */
1058                 nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
1059                 dprintk(" OK\n");
1060 remove_request:
1061                 nfs_end_page_writeback(page);
1062                 nfs_inode_remove_request(req);
1063         next:
1064                 nfs_clear_page_tag_locked(req);
1065         }
1066 }
1067
1068 static const struct rpc_call_ops nfs_write_full_ops = {
1069         .rpc_call_done = nfs_writeback_done_full,
1070         .rpc_release = nfs_writedata_release,
1071 };
1072
1073
1074 /*
1075  * This function is called when the WRITE call is complete.
1076  */
1077 int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1078 {
1079         struct nfs_writeargs    *argp = &data->args;
1080         struct nfs_writeres     *resp = &data->res;
1081         int status;
1082
1083         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1084                 task->tk_pid, task->tk_status);
1085
1086         /*
1087          * ->write_done will attempt to use post-op attributes to detect
1088          * conflicting writes by other clients.  A strict interpretation
1089          * of close-to-open would allow us to continue caching even if
1090          * another writer had changed the file, but some applications
1091          * depend on tighter cache coherency when writing.
1092          */
1093         status = NFS_PROTO(data->inode)->write_done(task, data);
1094         if (status != 0)
1095                 return status;
1096         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1097
1098 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1099         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1100                 /* We tried a write call, but the server did not
1101                  * commit data to stable storage even though we
1102                  * requested it.
1103                  * Note: There is a known bug in Tru64 < 5.0 in which
1104                  *       the server reports NFS_DATA_SYNC, but performs
1105                  *       NFS_FILE_SYNC. We therefore implement this checking
1106                  *       as a dprintk() in order to avoid filling syslog.
1107                  */
1108                 static unsigned long    complain;
1109
1110                 if (time_before(complain, jiffies)) {
1111                         dprintk("NFS: faulty NFS server %s:"
1112                                 " (committed = %d) != (stable = %d)\n",
1113                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1114                                 resp->verf->committed, argp->stable);
1115                         complain = jiffies + 300 * HZ;
1116                 }
1117         }
1118 #endif
1119         /* Is this a short write? */
1120         if (task->tk_status >= 0 && resp->count < argp->count) {
1121                 static unsigned long    complain;
1122
1123                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1124
1125                 /* Has the server at least made some progress? */
1126                 if (resp->count != 0) {
1127                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1128                         if (resp->verf->committed != NFS_UNSTABLE) {
1129                                 /* Resend from where the server left off */
1130                                 argp->offset += resp->count;
1131                                 argp->pgbase += resp->count;
1132                                 argp->count -= resp->count;
1133                         } else {
1134                                 /* Resend as a stable write in order to avoid
1135                                  * headaches in the case of a server crash.
1136                                  */
1137                                 argp->stable = NFS_FILE_SYNC;
1138                         }
1139                         rpc_restart_call(task);
1140                         return -EAGAIN;
1141                 }
1142                 if (time_before(complain, jiffies)) {
1143                         printk(KERN_WARNING
1144                                "NFS: Server wrote zero bytes, expected %u.\n",
1145                                         argp->count);
1146                         complain = jiffies + 300 * HZ;
1147                 }
1148                 /* Can't do anything about it except throw an error. */
1149                 task->tk_status = -EIO;
1150         }
1151         return 0;
1152 }
1153
1154
1155 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1156 void nfs_commit_release(void *wdata)
1157 {
1158         nfs_commit_free(wdata);
1159 }
1160
1161 /*
1162  * Set up the argument/result storage required for the RPC call.
1163  */
1164 static void nfs_commit_rpcsetup(struct list_head *head,
1165                 struct nfs_write_data *data,
1166                 int how)
1167 {
1168         struct nfs_page *first = nfs_list_entry(head->next);
1169         struct inode *inode = first->wb_context->path.dentry->d_inode;
1170         int flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC;
1171         int priority = flush_task_priority(how);
1172         struct rpc_task *task;
1173         struct rpc_message msg = {
1174                 .rpc_argp = &data->args,
1175                 .rpc_resp = &data->res,
1176                 .rpc_cred = first->wb_context->cred,
1177         };
1178         struct rpc_task_setup task_setup_data = {
1179                 .task = &data->task,
1180                 .rpc_client = NFS_CLIENT(inode),
1181                 .rpc_message = &msg,
1182                 .callback_ops = &nfs_commit_ops,
1183                 .callback_data = data,
1184                 .flags = flags,
1185                 .priority = priority,
1186         };
1187
1188         /* Set up the RPC argument and reply structs
1189          * NB: take care not to mess about with data->commit et al. */
1190
1191         list_splice_init(head, &data->pages);
1192
1193         data->inode       = inode;
1194         data->cred        = msg.rpc_cred;
1195
1196         data->args.fh     = NFS_FH(data->inode);
1197         /* Note: we always request a commit of the entire inode */
1198         data->args.offset = 0;
1199         data->args.count  = 0;
1200         data->res.count   = 0;
1201         data->res.fattr   = &data->fattr;
1202         data->res.verf    = &data->verf;
1203         nfs_fattr_init(&data->fattr);
1204
1205         /* Set up the initial task struct.  */
1206         NFS_PROTO(inode)->commit_setup(data, &msg);
1207
1208         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1209
1210         task = rpc_run_task(&task_setup_data);
1211         if (!IS_ERR(task))
1212                 rpc_put_task(task);
1213 }
1214
1215 /*
1216  * Commit dirty pages
1217  */
1218 static int
1219 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1220 {
1221         struct nfs_write_data   *data;
1222         struct nfs_page         *req;
1223
1224         data = nfs_commit_alloc();
1225
1226         if (!data)
1227                 goto out_bad;
1228
1229         /* Set up the argument struct */
1230         nfs_commit_rpcsetup(head, data, how);
1231
1232         return 0;
1233  out_bad:
1234         while (!list_empty(head)) {
1235                 req = nfs_list_entry(head->next);
1236                 nfs_list_remove_request(req);
1237                 nfs_mark_request_commit(req);
1238                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1239                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1240                                 BDI_RECLAIMABLE);
1241                 nfs_clear_page_tag_locked(req);
1242         }
1243         return -ENOMEM;
1244 }
1245
1246 /*
1247  * COMMIT call returned
1248  */
1249 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1250 {
1251         struct nfs_write_data   *data = calldata;
1252         struct nfs_page         *req;
1253
1254         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1255                                 task->tk_pid, task->tk_status);
1256
1257         /* Call the NFS version-specific code */
1258         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
1259                 return;
1260
1261         while (!list_empty(&data->pages)) {
1262                 req = nfs_list_entry(data->pages.next);
1263                 nfs_list_remove_request(req);
1264                 clear_bit(PG_NEED_COMMIT, &(req)->wb_flags);
1265                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1266                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1267                                 BDI_RECLAIMABLE);
1268
1269                 dprintk("NFS: commit (%s/%Ld %d@%Ld)",
1270                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1271                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1272                         req->wb_bytes,
1273                         (long long)req_offset(req));
1274                 if (task->tk_status < 0) {
1275                         nfs_context_set_write_error(req->wb_context, task->tk_status);
1276                         nfs_inode_remove_request(req);
1277                         dprintk(", error = %d\n", task->tk_status);
1278                         goto next;
1279                 }
1280
1281                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1282                  * returned by the server against all stored verfs. */
1283                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1284                         /* We have a match */
1285                         /* Set the PG_uptodate flag */
1286                         nfs_mark_uptodate(req->wb_page, req->wb_pgbase,
1287                                         req->wb_bytes);
1288                         nfs_inode_remove_request(req);
1289                         dprintk(" OK\n");
1290                         goto next;
1291                 }
1292                 /* We have a mismatch. Write the page again */
1293                 dprintk(" mismatch\n");
1294                 nfs_redirty_request(req);
1295         next:
1296                 nfs_clear_page_tag_locked(req);
1297         }
1298 }
1299
1300 static const struct rpc_call_ops nfs_commit_ops = {
1301         .rpc_call_done = nfs_commit_done,
1302         .rpc_release = nfs_commit_release,
1303 };
1304
1305 int nfs_commit_inode(struct inode *inode, int how)
1306 {
1307         LIST_HEAD(head);
1308         int res;
1309
1310         spin_lock(&inode->i_lock);
1311         res = nfs_scan_commit(inode, &head, 0, 0);
1312         spin_unlock(&inode->i_lock);
1313         if (res) {
1314                 int error = nfs_commit_list(inode, &head, how);
1315                 if (error < 0)
1316                         return error;
1317         }
1318         return res;
1319 }
1320 #else
1321 static inline int nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1322 {
1323         return 0;
1324 }
1325 #endif
1326
1327 long nfs_sync_mapping_wait(struct address_space *mapping, struct writeback_control *wbc, int how)
1328 {
1329         struct inode *inode = mapping->host;
1330         pgoff_t idx_start, idx_end;
1331         unsigned int npages = 0;
1332         LIST_HEAD(head);
1333         int nocommit = how & FLUSH_NOCOMMIT;
1334         long pages, ret;
1335
1336         /* FIXME */
1337         if (wbc->range_cyclic)
1338                 idx_start = 0;
1339         else {
1340                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
1341                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
1342                 if (idx_end > idx_start) {
1343                         pgoff_t l_npages = 1 + idx_end - idx_start;
1344                         npages = l_npages;
1345                         if (sizeof(npages) != sizeof(l_npages) &&
1346                                         (pgoff_t)npages != l_npages)
1347                                 npages = 0;
1348                 }
1349         }
1350         how &= ~FLUSH_NOCOMMIT;
1351         spin_lock(&inode->i_lock);
1352         do {
1353                 ret = nfs_wait_on_requests_locked(inode, idx_start, npages);
1354                 if (ret != 0)
1355                         continue;
1356                 if (nocommit)
1357                         break;
1358                 pages = nfs_scan_commit(inode, &head, idx_start, npages);
1359                 if (pages == 0)
1360                         break;
1361                 if (how & FLUSH_INVALIDATE) {
1362                         spin_unlock(&inode->i_lock);
1363                         nfs_cancel_commit_list(&head);
1364                         ret = pages;
1365                         spin_lock(&inode->i_lock);
1366                         continue;
1367                 }
1368                 pages += nfs_scan_commit(inode, &head, 0, 0);
1369                 spin_unlock(&inode->i_lock);
1370                 ret = nfs_commit_list(inode, &head, how);
1371                 spin_lock(&inode->i_lock);
1372
1373         } while (ret >= 0);
1374         spin_unlock(&inode->i_lock);
1375         return ret;
1376 }
1377
1378 static int __nfs_write_mapping(struct address_space *mapping, struct writeback_control *wbc, int how)
1379 {
1380         int ret;
1381
1382         ret = nfs_writepages(mapping, wbc);
1383         if (ret < 0)
1384                 goto out;
1385         ret = nfs_sync_mapping_wait(mapping, wbc, how);
1386         if (ret < 0)
1387                 goto out;
1388         return 0;
1389 out:
1390         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1391         return ret;
1392 }
1393
1394 /* Two pass sync: first using WB_SYNC_NONE, then WB_SYNC_ALL */
1395 static int nfs_write_mapping(struct address_space *mapping, int how)
1396 {
1397         struct writeback_control wbc = {
1398                 .bdi = mapping->backing_dev_info,
1399                 .sync_mode = WB_SYNC_NONE,
1400                 .nr_to_write = LONG_MAX,
1401                 .for_writepages = 1,
1402                 .range_cyclic = 1,
1403         };
1404         int ret;
1405
1406         ret = __nfs_write_mapping(mapping, &wbc, how);
1407         if (ret < 0)
1408                 return ret;
1409         wbc.sync_mode = WB_SYNC_ALL;
1410         return __nfs_write_mapping(mapping, &wbc, how);
1411 }
1412
1413 /*
1414  * flush the inode to disk.
1415  */
1416 int nfs_wb_all(struct inode *inode)
1417 {
1418         return nfs_write_mapping(inode->i_mapping, 0);
1419 }
1420
1421 int nfs_wb_nocommit(struct inode *inode)
1422 {
1423         return nfs_write_mapping(inode->i_mapping, FLUSH_NOCOMMIT);
1424 }
1425
1426 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1427 {
1428         struct nfs_page *req;
1429         loff_t range_start = page_offset(page);
1430         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1431         struct writeback_control wbc = {
1432                 .bdi = page->mapping->backing_dev_info,
1433                 .sync_mode = WB_SYNC_ALL,
1434                 .nr_to_write = LONG_MAX,
1435                 .range_start = range_start,
1436                 .range_end = range_end,
1437         };
1438         int ret = 0;
1439
1440         BUG_ON(!PageLocked(page));
1441         for (;;) {
1442                 req = nfs_page_find_request(page);
1443                 if (req == NULL)
1444                         goto out;
1445                 if (test_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1446                         nfs_release_request(req);
1447                         break;
1448                 }
1449                 if (nfs_lock_request_dontget(req)) {
1450                         nfs_inode_remove_request(req);
1451                         /*
1452                          * In case nfs_inode_remove_request has marked the
1453                          * page as being dirty
1454                          */
1455                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1456                         nfs_unlock_request(req);
1457                         break;
1458                 }
1459                 ret = nfs_wait_on_request(req);
1460                 if (ret < 0)
1461                         goto out;
1462         }
1463         if (!PagePrivate(page))
1464                 return 0;
1465         ret = nfs_sync_mapping_wait(page->mapping, &wbc, FLUSH_INVALIDATE);
1466 out:
1467         return ret;
1468 }
1469
1470 static int nfs_wb_page_priority(struct inode *inode, struct page *page,
1471                                 int how)
1472 {
1473         loff_t range_start = page_offset(page);
1474         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1475         struct writeback_control wbc = {
1476                 .bdi = page->mapping->backing_dev_info,
1477                 .sync_mode = WB_SYNC_ALL,
1478                 .nr_to_write = LONG_MAX,
1479                 .range_start = range_start,
1480                 .range_end = range_end,
1481         };
1482         int ret;
1483
1484         BUG_ON(!PageLocked(page));
1485         if (clear_page_dirty_for_io(page)) {
1486                 ret = nfs_writepage_locked(page, &wbc);
1487                 if (ret < 0)
1488                         goto out;
1489         }
1490         if (!PagePrivate(page))
1491                 return 0;
1492         ret = nfs_sync_mapping_wait(page->mapping, &wbc, how);
1493         if (ret >= 0)
1494                 return 0;
1495 out:
1496         __mark_inode_dirty(inode, I_DIRTY_PAGES);
1497         return ret;
1498 }
1499
1500 /*
1501  * Write back all requests on one page - we do this before reading it.
1502  */
1503 int nfs_wb_page(struct inode *inode, struct page* page)
1504 {
1505         return nfs_wb_page_priority(inode, page, FLUSH_STABLE);
1506 }
1507
1508 int __init nfs_init_writepagecache(void)
1509 {
1510         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1511                                              sizeof(struct nfs_write_data),
1512                                              0, SLAB_HWCACHE_ALIGN,
1513                                              NULL);
1514         if (nfs_wdata_cachep == NULL)
1515                 return -ENOMEM;
1516
1517         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1518                                                      nfs_wdata_cachep);
1519         if (nfs_wdata_mempool == NULL)
1520                 return -ENOMEM;
1521
1522         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1523                                                       nfs_wdata_cachep);
1524         if (nfs_commit_mempool == NULL)
1525                 return -ENOMEM;
1526
1527         /*
1528          * NFS congestion size, scale with available memory.
1529          *
1530          *  64MB:    8192k
1531          * 128MB:   11585k
1532          * 256MB:   16384k
1533          * 512MB:   23170k
1534          *   1GB:   32768k
1535          *   2GB:   46340k
1536          *   4GB:   65536k
1537          *   8GB:   92681k
1538          *  16GB:  131072k
1539          *
1540          * This allows larger machines to have larger/more transfers.
1541          * Limit the default to 256M
1542          */
1543         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1544         if (nfs_congestion_kb > 256*1024)
1545                 nfs_congestion_kb = 256*1024;
1546
1547         return 0;
1548 }
1549
1550 void nfs_destroy_writepagecache(void)
1551 {
1552         mempool_destroy(nfs_commit_mempool);
1553         mempool_destroy(nfs_wdata_mempool);
1554         kmem_cache_destroy(nfs_wdata_cachep);
1555 }
1556