fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  35
  36 #define MIN_POOL_WRITE          (32)
  37 #define MIN_POOL_COMMIT         (4)
  38
  39 /*
  40  * Local function declarations
  41  */
  42 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  43                                   struct inode *inode, int ioflags);
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_write_partial_ops;
  46 static const struct rpc_call_ops nfs_write_full_ops;
  47 static const struct rpc_call_ops nfs_commit_ops;
  48
  49 static struct kmem_cache *nfs_wdata_cachep;
  50 static mempool_t *nfs_wdata_mempool;
  51 static mempool_t *nfs_commit_mempool;
  52
  53 struct nfs_write_data *nfs_commitdata_alloc(void)
  54 {
  55         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  56
  57         if (p) {
  58                 memset(p, 0, sizeof(*p));
  59                 INIT_LIST_HEAD(&p->pages);
  60         }
  61         return p;
  62 }
  63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  64
  65 void nfs_commit_free(struct nfs_write_data *p)
  66 {
  67         if (p && (p->pagevec != &p->page_array[0]))
  68                 kfree(p->pagevec);
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71 EXPORT_SYMBOL_GPL(nfs_commit_free);
  72
  73 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  74 {
  75         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  76
  77         if (p) {
  78                 memset(p, 0, sizeof(*p));
  79                 INIT_LIST_HEAD(&p->pages);
  80                 p->npages = pagecount;
  81                 if (pagecount <= ARRAY_SIZE(p->page_array))
  82                         p->pagevec = p->page_array;
  83                 else {
  84                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  85                         if (!p->pagevec) {
  86                                 mempool_free(p, nfs_wdata_mempool);
  87                                 p = NULL;
  88                         }
  89                 }
  90         }
  91         return p;
  92 }
  93
  94 void nfs_writedata_free(struct nfs_write_data *p)
  95 {
  96         if (p && (p->pagevec != &p->page_array[0]))
  97                 kfree(p->pagevec);
  98         mempool_free(p, nfs_wdata_mempool);
  99 }
 100
 101 void nfs_writedata_release(struct nfs_write_data *wdata)
 102 {
 103         put_nfs_open_context(wdata->args.context);
 104         nfs_writedata_free(wdata);
 105 }
 106
 107 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 108 {
 109         ctx->error = error;
 110         smp_wmb();
 111         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 112 }
 113
 114 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 115 {
 116         struct nfs_page *req = NULL;
 117
 118         if (PagePrivate(page)) {
 119                 req = (struct nfs_page *)page_private(page);
 120                 if (req != NULL)
 121                         kref_get(&req->wb_kref);
 122         }
 123         return req;
 124 }
 125
 126 static struct nfs_page *nfs_page_find_request(struct page *page)
 127 {
 128         struct inode *inode = page->mapping->host;
 129         struct nfs_page *req = NULL;
 130
 131         spin_lock(&inode->i_lock);
 132         req = nfs_page_find_request_locked(page);
 133         spin_unlock(&inode->i_lock);
 134         return req;
 135 }
 136
 137 /* Adjust the file length if we're writing beyond the end */
 138 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 139 {
 140         struct inode *inode = page->mapping->host;
 141         loff_t end, i_size;
 142         pgoff_t end_index;
 143
 144         spin_lock(&inode->i_lock);
 145         i_size = i_size_read(inode);
 146         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 147         if (i_size > 0 && page->index < end_index)
 148                 goto out;
 149         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 150         if (i_size >= end)
 151                 goto out;
 152         i_size_write(inode, end);
 153         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 154 out:
 155         spin_unlock(&inode->i_lock);
 156 }
 157
 158 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 159 static void nfs_set_pageerror(struct page *page)
 160 {
 161         SetPageError(page);
 162         nfs_zap_mapping(page->mapping->host, page->mapping);
 163 }
 164
 165 /* We can set the PG_uptodate flag if we see that a write request
 166  * covers the full page.
 167  */
 168 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 169 {
 170         if (PageUptodate(page))
 171                 return;
 172         if (base != 0)
 173                 return;
 174         if (count != nfs_page_length(page))
 175                 return;
 176         SetPageUptodate(page);
 177 }
 178
 179 static int wb_priority(struct writeback_control *wbc)
 180 {
 181         if (wbc->for_reclaim)
 182                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 183         if (wbc->for_kupdate || wbc->for_background)
 184                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 185         return FLUSH_COND_STABLE;
 186 }
 187
 188 /*
 189  * NFS congestion control
 190  */
 191
 192 int nfs_congestion_kb;
 193
 194 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 195 #define NFS_CONGESTION_OFF_THRESH       \
 196         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 197
 198 static int nfs_set_page_writeback(struct page *page)
 199 {
 200         int ret = test_set_page_writeback(page);
 201
 202         if (!ret) {
 203                 struct inode *inode = page->mapping->host;
 204                 struct nfs_server *nfss = NFS_SERVER(inode);
 205
 206                 page_cache_get(page);
 207                 if (atomic_long_inc_return(&nfss->writeback) >
 208                                 NFS_CONGESTION_ON_THRESH) {
 209                         set_bdi_congested(&nfss->backing_dev_info,
 210                                                 BLK_RW_ASYNC);
 211                 }
 212         }
 213         return ret;
 214 }
 215
 216 static void nfs_end_page_writeback(struct page *page)
 217 {
 218         struct inode *inode = page->mapping->host;
 219         struct nfs_server *nfss = NFS_SERVER(inode);
 220
 221         end_page_writeback(page);
 222         page_cache_release(page);
 223         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 224                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 225 }
 226
 227 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 228 {
 229         struct inode *inode = page->mapping->host;
 230         struct nfs_page *req;
 231         int ret;
 232
 233         spin_lock(&inode->i_lock);
 234         for (;;) {
 235                 req = nfs_page_find_request_locked(page);
 236                 if (req == NULL)
 237                         break;
 238                 if (nfs_lock_request_dontget(req))
 239                         break;
 240                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 241                  *       then the call to nfs_lock_request_dontget() will always
 242                  *       succeed provided that someone hasn't already marked the
 243                  *       request as dirty (in which case we don't care).
 244                  */
 245                 spin_unlock(&inode->i_lock);
 246                 if (!nonblock)
 247                         ret = nfs_wait_on_request(req);
 248                 else
 249                         ret = -EAGAIN;
 250                 nfs_release_request(req);
 251                 if (ret != 0)
 252                         return ERR_PTR(ret);
 253                 spin_lock(&inode->i_lock);
 254         }
 255         spin_unlock(&inode->i_lock);
 256         return req;
 257 }
 258
 259 /*
 260  * Find an associated nfs write request, and prepare to flush it out
 261  * May return an error if the user signalled nfs_wait_on_request().
 262  */
 263 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 264                                 struct page *page, bool nonblock)
 265 {
 266         struct nfs_page *req;
 267         int ret = 0;
 268
 269         req = nfs_find_and_lock_request(page, nonblock);
 270         if (!req)
 271                 goto out;
 272         ret = PTR_ERR(req);
 273         if (IS_ERR(req))
 274                 goto out;
 275
 276         ret = nfs_set_page_writeback(page);
 277         BUG_ON(ret != 0);
 278         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
 279
 280         if (!nfs_pageio_add_request(pgio, req)) {
 281                 nfs_redirty_request(req);
 282                 ret = pgio->pg_error;
 283         }
 284 out:
 285         return ret;
 286 }
 287
 288 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 289 {
 290         struct inode *inode = page->mapping->host;
 291         int ret;
 292
 293         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 294         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 295
 296         nfs_pageio_cond_complete(pgio, page->index);
 297         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 298         if (ret == -EAGAIN) {
 299                 redirty_page_for_writepage(wbc, page);
 300                 ret = 0;
 301         }
 302         return ret;
 303 }
 304
 305 /*
 306  * Write an mmapped page to the server.
 307  */
 308 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 309 {
 310         struct nfs_pageio_descriptor pgio;
 311         int err;
 312
 313         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc));
 314         err = nfs_do_writepage(page, wbc, &pgio);
 315         nfs_pageio_complete(&pgio);
 316         if (err < 0)
 317                 return err;
 318         if (pgio.pg_error < 0)
 319                 return pgio.pg_error;
 320         return 0;
 321 }
 322
 323 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 324 {
 325         int ret;
 326
 327         ret = nfs_writepage_locked(page, wbc);
 328         unlock_page(page);
 329         return ret;
 330 }
 331
 332 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 333 {
 334         int ret;
 335
 336         ret = nfs_do_writepage(page, wbc, data);
 337         unlock_page(page);
 338         return ret;
 339 }
 340
 341 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 342 {
 343         struct inode *inode = mapping->host;
 344         unsigned long *bitlock = &NFS_I(inode)->flags;
 345         struct nfs_pageio_descriptor pgio;
 346         int err;
 347
 348         /* Stop dirtying of new pages while we sync */
 349         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 350                         nfs_wait_bit_killable, TASK_KILLABLE);
 351         if (err)
 352                 goto out_err;
 353
 354         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 355
 356         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc));
 357         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 358         nfs_pageio_complete(&pgio);
 359
 360         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 361         smp_mb__after_clear_bit();
 362         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 363
 364         if (err < 0)
 365                 goto out_err;
 366         err = pgio.pg_error;
 367         if (err < 0)
 368                 goto out_err;
 369         return 0;
 370 out_err:
 371         return err;
 372 }
 373
 374 /*
 375  * Insert a write request into an inode
 376  */
 377 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 378 {
 379         struct nfs_inode *nfsi = NFS_I(inode);
 380
 381         /* Lock the request! */
 382         nfs_lock_request_dontget(req);
 383
 384         spin_lock(&inode->i_lock);
 385         if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
 386                 inode->i_version++;
 387         set_bit(PG_MAPPED, &req->wb_flags);
 388         SetPagePrivate(req->wb_page);
 389         set_page_private(req->wb_page, (unsigned long)req);
 390         nfsi->npages++;
 391         kref_get(&req->wb_kref);
 392         spin_unlock(&inode->i_lock);
 393 }
 394
 395 /*
 396  * Remove a write request from an inode
 397  */
 398 static void nfs_inode_remove_request(struct nfs_page *req)
 399 {
 400         struct inode *inode = req->wb_context->dentry->d_inode;
 401         struct nfs_inode *nfsi = NFS_I(inode);
 402
 403         BUG_ON (!NFS_WBACK_BUSY(req));
 404
 405         spin_lock(&inode->i_lock);
 406         set_page_private(req->wb_page, 0);
 407         ClearPagePrivate(req->wb_page);
 408         clear_bit(PG_MAPPED, &req->wb_flags);
 409         nfsi->npages--;
 410         spin_unlock(&inode->i_lock);
 411         nfs_release_request(req);
 412 }
 413
 414 static void
 415 nfs_mark_request_dirty(struct nfs_page *req)
 416 {
 417         __set_page_dirty_nobuffers(req->wb_page);
 418 }
 419
 420 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 421 /**
 422  * nfs_request_add_commit_list - add request to a commit list
 423  * @req: pointer to a struct nfs_page
 424  * @head: commit list head
 425  *
 426  * This sets the PG_CLEAN bit, updates the inode global count of
 427  * number of outstanding requests requiring a commit as well as
 428  * the MM page stats.
 429  *
 430  * The caller must _not_ hold the inode->i_lock, but must be
 431  * holding the nfs_page lock.
 432  */
 433 void
 434 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *head)
 435 {
 436         struct inode *inode = req->wb_context->dentry->d_inode;
 437
 438         set_bit(PG_CLEAN, &(req)->wb_flags);
 439         spin_lock(&inode->i_lock);
 440         nfs_list_add_request(req, head);
 441         NFS_I(inode)->ncommit++;
 442         spin_unlock(&inode->i_lock);
 443         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 444         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 445         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 446 }
 447 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 448
 449 /**
 450  * nfs_request_remove_commit_list - Remove request from a commit list
 451  * @req: pointer to a nfs_page
 452  *
 453  * This clears the PG_CLEAN bit, and updates the inode global count of
 454  * number of outstanding requests requiring a commit
 455  * It does not update the MM page stats.
 456  *
 457  * The caller _must_ hold the inode->i_lock and the nfs_page lock.
 458  */
 459 void
 460 nfs_request_remove_commit_list(struct nfs_page *req)
 461 {
 462         struct inode *inode = req->wb_context->dentry->d_inode;
 463
 464         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 465                 return;
 466         nfs_list_remove_request(req);
 467         NFS_I(inode)->ncommit--;
 468 }
 469 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 470
 471
 472 /*
 473  * Add a request to the inode's commit list.
 474  */
 475 static void
 476 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 477 {
 478         struct inode *inode = req->wb_context->dentry->d_inode;
 479
 480         if (pnfs_mark_request_commit(req, lseg))
 481                 return;
 482         nfs_request_add_commit_list(req, &NFS_I(inode)->commit_list);
 483 }
 484
 485 static void
 486 nfs_clear_page_commit(struct page *page)
 487 {
 488         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 489         dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 490 }
 491
 492 static void
 493 nfs_clear_request_commit(struct nfs_page *req)
 494 {
 495         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 496                 struct inode *inode = req->wb_context->dentry->d_inode;
 497
 498                 if (!pnfs_clear_request_commit(req)) {
 499                         spin_lock(&inode->i_lock);
 500                         nfs_request_remove_commit_list(req);
 501                         spin_unlock(&inode->i_lock);
 502                 }
 503                 nfs_clear_page_commit(req->wb_page);
 504         }
 505 }
 506
 507 static inline
 508 int nfs_write_need_commit(struct nfs_write_data *data)
 509 {
 510         if (data->verf.committed == NFS_DATA_SYNC)
 511                 return data->lseg == NULL;
 512         else
 513                 return data->verf.committed != NFS_FILE_SYNC;
 514 }
 515
 516 static inline
 517 int nfs_reschedule_unstable_write(struct nfs_page *req,
 518                                   struct nfs_write_data *data)
 519 {
 520         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 521                 nfs_mark_request_commit(req, data->lseg);
 522                 return 1;
 523         }
 524         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 525                 nfs_mark_request_dirty(req);
 526                 return 1;
 527         }
 528         return 0;
 529 }
 530 #else
 531 static void
 532 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 533 {
 534 }
 535
 536 static void
 537 nfs_clear_request_commit(struct nfs_page *req)
 538 {
 539 }
 540
 541 static inline
 542 int nfs_write_need_commit(struct nfs_write_data *data)
 543 {
 544         return 0;
 545 }
 546
 547 static inline
 548 int nfs_reschedule_unstable_write(struct nfs_page *req,
 549                                   struct nfs_write_data *data)
 550 {
 551         return 0;
 552 }
 553 #endif
 554
 555 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 556 static int
 557 nfs_need_commit(struct nfs_inode *nfsi)
 558 {
 559         return nfsi->ncommit > 0;
 560 }
 561
 562 /* i_lock held by caller */
 563 static int
 564 nfs_scan_commit_list(struct list_head *src, struct list_head *dst, int max,
 565                 spinlock_t *lock)
 566 {
 567         struct nfs_page *req, *tmp;
 568         int ret = 0;
 569
 570         list_for_each_entry_safe(req, tmp, src, wb_list) {
 571                 if (!nfs_lock_request(req))
 572                         continue;
 573                 if (cond_resched_lock(lock))
 574                         list_safe_reset_next(req, tmp, wb_list);
 575                 nfs_request_remove_commit_list(req);
 576                 nfs_list_add_request(req, dst);
 577                 ret++;
 578                 if (ret == max)
 579                         break;
 580         }
 581         return ret;
 582 }
 583
 584 /*
 585  * nfs_scan_commit - Scan an inode for commit requests
 586  * @inode: NFS inode to scan
 587  * @dst: destination list
 588  *
 589  * Moves requests from the inode's 'commit' request list.
 590  * The requests are *not* checked to ensure that they form a contiguous set.
 591  */
 592 static int
 593 nfs_scan_commit(struct inode *inode, struct list_head *dst)
 594 {
 595         struct nfs_inode *nfsi = NFS_I(inode);
 596         int ret = 0;
 597
 598         spin_lock(&inode->i_lock);
 599         if (nfsi->ncommit > 0) {
 600                 const int max = INT_MAX;
 601
 602                 ret = nfs_scan_commit_list(&nfsi->commit_list, dst, max,
 603                                 &inode->i_lock);
 604                 ret += pnfs_scan_commit_lists(inode, max - ret,
 605                                 &inode->i_lock);
 606         }
 607         spin_unlock(&inode->i_lock);
 608         return ret;
 609 }
 610
 611 #else
 612 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 613 {
 614         return 0;
 615 }
 616
 617 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst)
 618 {
 619         return 0;
 620 }
 621 #endif
 622
 623 /*
 624  * Search for an existing write request, and attempt to update
 625  * it to reflect a new dirty region on a given page.
 626  *
 627  * If the attempt fails, then the existing request is flushed out
 628  * to disk.
 629  */
 630 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 631                 struct page *page,
 632                 unsigned int offset,
 633                 unsigned int bytes)
 634 {
 635         struct nfs_page *req;
 636         unsigned int rqend;
 637         unsigned int end;
 638         int error;
 639
 640         if (!PagePrivate(page))
 641                 return NULL;
 642
 643         end = offset + bytes;
 644         spin_lock(&inode->i_lock);
 645
 646         for (;;) {
 647                 req = nfs_page_find_request_locked(page);
 648                 if (req == NULL)
 649                         goto out_unlock;
 650
 651                 rqend = req->wb_offset + req->wb_bytes;
 652                 /*
 653                  * Tell the caller to flush out the request if
 654                  * the offsets are non-contiguous.
 655                  * Note: nfs_flush_incompatible() will already
 656                  * have flushed out requests having wrong owners.
 657                  */
 658                 if (offset > rqend
 659                     || end < req->wb_offset)
 660                         goto out_flushme;
 661
 662                 if (nfs_lock_request_dontget(req))
 663                         break;
 664
 665                 /* The request is locked, so wait and then retry */
 666                 spin_unlock(&inode->i_lock);
 667                 error = nfs_wait_on_request(req);
 668                 nfs_release_request(req);
 669                 if (error != 0)
 670                         goto out_err;
 671                 spin_lock(&inode->i_lock);
 672         }
 673
 674         /* Okay, the request matches. Update the region */
 675         if (offset < req->wb_offset) {
 676                 req->wb_offset = offset;
 677                 req->wb_pgbase = offset;
 678         }
 679         if (end > rqend)
 680                 req->wb_bytes = end - req->wb_offset;
 681         else
 682                 req->wb_bytes = rqend - req->wb_offset;
 683 out_unlock:
 684         spin_unlock(&inode->i_lock);
 685         if (req)
 686                 nfs_clear_request_commit(req);
 687         return req;
 688 out_flushme:
 689         spin_unlock(&inode->i_lock);
 690         nfs_release_request(req);
 691         error = nfs_wb_page(inode, page);
 692 out_err:
 693         return ERR_PTR(error);
 694 }
 695
 696 /*
 697  * Try to update an existing write request, or create one if there is none.
 698  *
 699  * Note: Should always be called with the Page Lock held to prevent races
 700  * if we have to add a new request. Also assumes that the caller has
 701  * already called nfs_flush_incompatible() if necessary.
 702  */
 703 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 704                 struct page *page, unsigned int offset, unsigned int bytes)
 705 {
 706         struct inode *inode = page->mapping->host;
 707         struct nfs_page *req;
 708
 709         req = nfs_try_to_update_request(inode, page, offset, bytes);
 710         if (req != NULL)
 711                 goto out;
 712         req = nfs_create_request(ctx, inode, page, offset, bytes);
 713         if (IS_ERR(req))
 714                 goto out;
 715         nfs_inode_add_request(inode, req);
 716 out:
 717         return req;
 718 }
 719
 720 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 721                 unsigned int offset, unsigned int count)
 722 {
 723         struct nfs_page *req;
 724
 725         req = nfs_setup_write_request(ctx, page, offset, count);
 726         if (IS_ERR(req))
 727                 return PTR_ERR(req);
 728         /* Update file length */
 729         nfs_grow_file(page, offset, count);
 730         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 731         nfs_mark_request_dirty(req);
 732         nfs_unlock_request(req);
 733         return 0;
 734 }
 735
 736 int nfs_flush_incompatible(struct file *file, struct page *page)
 737 {
 738         struct nfs_open_context *ctx = nfs_file_open_context(file);
 739         struct nfs_page *req;
 740         int do_flush, status;
 741         /*
 742          * Look for a request corresponding to this page. If there
 743          * is one, and it belongs to another file, we flush it out
 744          * before we try to copy anything into the page. Do this
 745          * due to the lack of an ACCESS-type call in NFSv2.
 746          * Also do the same if we find a request from an existing
 747          * dropped page.
 748          */
 749         do {
 750                 req = nfs_page_find_request(page);
 751                 if (req == NULL)
 752                         return 0;
 753                 do_flush = req->wb_page != page || req->wb_context != ctx ||
 754                         req->wb_lock_context->lockowner != current->files ||
 755                         req->wb_lock_context->pid != current->tgid;
 756                 nfs_release_request(req);
 757                 if (!do_flush)
 758                         return 0;
 759                 status = nfs_wb_page(page->mapping->host, page);
 760         } while (status == 0);
 761         return status;
 762 }
 763
 764 /*
 765  * If the page cache is marked as unsafe or invalid, then we can't rely on
 766  * the PageUptodate() flag. In this case, we will need to turn off
 767  * write optimisations that depend on the page contents being correct.
 768  */
 769 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 770 {
 771         return PageUptodate(page) &&
 772                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 773 }
 774
 775 /*
 776  * Update and possibly write a cached page of an NFS file.
 777  *
 778  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 779  * things with a page scheduled for an RPC call (e.g. invalidate it).
 780  */
 781 int nfs_updatepage(struct file *file, struct page *page,
 782                 unsigned int offset, unsigned int count)
 783 {
 784         struct nfs_open_context *ctx = nfs_file_open_context(file);
 785         struct inode    *inode = page->mapping->host;
 786         int             status = 0;
 787
 788         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 789
 790         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 791                 file->f_path.dentry->d_parent->d_name.name,
 792                 file->f_path.dentry->d_name.name, count,
 793                 (long long)(page_offset(page) + offset));
 794
 795         /* If we're not using byte range locks, and we know the page
 796          * is up to date, it may be more efficient to extend the write
 797          * to cover the entire page in order to avoid fragmentation
 798          * inefficiencies.
 799          */
 800         if (nfs_write_pageuptodate(page, inode) &&
 801                         inode->i_flock == NULL &&
 802                         !(file->f_flags & O_DSYNC)) {
 803                 count = max(count + offset, nfs_page_length(page));
 804                 offset = 0;
 805         }
 806
 807         status = nfs_writepage_setup(ctx, page, offset, count);
 808         if (status < 0)
 809                 nfs_set_pageerror(page);
 810         else
 811                 __set_page_dirty_nobuffers(page);
 812
 813         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 814                         status, (long long)i_size_read(inode));
 815         return status;
 816 }
 817
 818 static void nfs_writepage_release(struct nfs_page *req,
 819                                   struct nfs_write_data *data)
 820 {
 821         struct page *page = req->wb_page;
 822
 823         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
 824                 nfs_inode_remove_request(req);
 825         nfs_unlock_request(req);
 826         nfs_end_page_writeback(page);
 827 }
 828
 829 static int flush_task_priority(int how)
 830 {
 831         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 832                 case FLUSH_HIGHPRI:
 833                         return RPC_PRIORITY_HIGH;
 834                 case FLUSH_LOWPRI:
 835                         return RPC_PRIORITY_LOW;
 836         }
 837         return RPC_PRIORITY_NORMAL;
 838 }
 839
 840 int nfs_initiate_write(struct nfs_write_data *data,
 841                        struct rpc_clnt *clnt,
 842                        const struct rpc_call_ops *call_ops,
 843                        int how)
 844 {
 845         struct inode *inode = data->inode;
 846         int priority = flush_task_priority(how);
 847         struct rpc_task *task;
 848         struct rpc_message msg = {
 849                 .rpc_argp = &data->args,
 850                 .rpc_resp = &data->res,
 851                 .rpc_cred = data->cred,
 852         };
 853         struct rpc_task_setup task_setup_data = {
 854                 .rpc_client = clnt,
 855                 .task = &data->task,
 856                 .rpc_message = &msg,
 857                 .callback_ops = call_ops,
 858                 .callback_data = data,
 859                 .workqueue = nfsiod_workqueue,
 860                 .flags = RPC_TASK_ASYNC,
 861                 .priority = priority,
 862         };
 863         int ret = 0;
 864
 865         /* Set up the initial task struct.  */
 866         NFS_PROTO(inode)->write_setup(data, &msg);
 867
 868         dprintk("NFS: %5u initiated write call "
 869                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 870                 data->task.tk_pid,
 871                 inode->i_sb->s_id,
 872                 (long long)NFS_FILEID(inode),
 873                 data->args.count,
 874                 (unsigned long long)data->args.offset);
 875
 876         task = rpc_run_task(&task_setup_data);
 877         if (IS_ERR(task)) {
 878                 ret = PTR_ERR(task);
 879                 goto out;
 880         }
 881         if (how & FLUSH_SYNC) {
 882                 ret = rpc_wait_for_completion_task(task);
 883                 if (ret == 0)
 884                         ret = task->tk_status;
 885         }
 886         rpc_put_task(task);
 887 out:
 888         return ret;
 889 }
 890 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 891
 892 /*
 893  * Set up the argument/result storage required for the RPC call.
 894  */
 895 static void nfs_write_rpcsetup(struct nfs_page *req,
 896                 struct nfs_write_data *data,
 897                 unsigned int count, unsigned int offset,
 898                 int how)
 899 {
 900         struct inode *inode = req->wb_context->dentry->d_inode;
 901
 902         /* Set up the RPC argument and reply structs
 903          * NB: take care not to mess about with data->commit et al. */
 904
 905         data->req = req;
 906         data->inode = inode = req->wb_context->dentry->d_inode;
 907         data->cred = req->wb_context->cred;
 908
 909         data->args.fh     = NFS_FH(inode);
 910         data->args.offset = req_offset(req) + offset;
 911         /* pnfs_set_layoutcommit needs this */
 912         data->mds_offset = data->args.offset;
 913         data->args.pgbase = req->wb_pgbase + offset;
 914         data->args.pages  = data->pagevec;
 915         data->args.count  = count;
 916         data->args.context = get_nfs_open_context(req->wb_context);
 917         data->args.lock_context = req->wb_lock_context;
 918         data->args.stable  = NFS_UNSTABLE;
 919         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 920         case 0:
 921                 break;
 922         case FLUSH_COND_STABLE:
 923                 if (nfs_need_commit(NFS_I(inode)))
 924                         break;
 925         default:
 926                 data->args.stable = NFS_FILE_SYNC;
 927         }
 928
 929         data->res.fattr   = &data->fattr;
 930         data->res.count   = count;
 931         data->res.verf    = &data->verf;
 932         nfs_fattr_init(&data->fattr);
 933 }
 934
 935 static int nfs_do_write(struct nfs_write_data *data,
 936                 const struct rpc_call_ops *call_ops,
 937                 int how)
 938 {
 939         struct inode *inode = data->args.context->dentry->d_inode;
 940
 941         return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
 942 }
 943
 944 static int nfs_do_multiple_writes(struct list_head *head,
 945                 const struct rpc_call_ops *call_ops,
 946                 int how)
 947 {
 948         struct nfs_write_data *data;
 949         int ret = 0;
 950
 951         while (!list_empty(head)) {
 952                 int ret2;
 953
 954                 data = list_entry(head->next, struct nfs_write_data, list);
 955                 list_del_init(&data->list);
 956
 957                 ret2 = nfs_do_write(data, call_ops, how);
 958                  if (ret == 0)
 959                          ret = ret2;
 960         }
 961         return ret;
 962 }
 963
 964 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 965  * call this on each, which will prepare them to be retried on next
 966  * writeback using standard nfs.
 967  */
 968 static void nfs_redirty_request(struct nfs_page *req)
 969 {
 970         struct page *page = req->wb_page;
 971
 972         nfs_mark_request_dirty(req);
 973         nfs_unlock_request(req);
 974         nfs_end_page_writeback(page);
 975 }
 976
 977 /*
 978  * Generate multiple small requests to write out a single
 979  * contiguous dirty area on one page.
 980  */
 981 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
 982 {
 983         struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
 984         struct page *page = req->wb_page;
 985         struct nfs_write_data *data;
 986         size_t wsize = desc->pg_bsize, nbytes;
 987         unsigned int offset;
 988         int requests = 0;
 989         int ret = 0;
 990
 991         nfs_list_remove_request(req);
 992
 993         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 994             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
 995              desc->pg_count > wsize))
 996                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 997
 998
 999         offset = 0;
1000         nbytes = desc->pg_count;
1001         do {
1002                 size_t len = min(nbytes, wsize);
1003
1004                 data = nfs_writedata_alloc(1);
1005                 if (!data)
1006                         goto out_bad;
1007                 data->pagevec[0] = page;
1008                 nfs_write_rpcsetup(req, data, len, offset, desc->pg_ioflags);
1009                 list_add(&data->list, res);
1010                 requests++;
1011                 nbytes -= len;
1012                 offset += len;
1013         } while (nbytes != 0);
1014         atomic_set(&req->wb_complete, requests);
1015         desc->pg_rpc_callops = &nfs_write_partial_ops;
1016         return ret;
1017
1018 out_bad:
1019         while (!list_empty(res)) {
1020                 data = list_entry(res->next, struct nfs_write_data, list);
1021                 list_del(&data->list);
1022                 nfs_writedata_release(data);
1023         }
1024         nfs_redirty_request(req);
1025         return -ENOMEM;
1026 }
1027
1028 /*
1029  * Create an RPC task for the given write request and kick it.
1030  * The page must have been locked by the caller.
1031  *
1032  * It may happen that the page we're passed is not marked dirty.
1033  * This is the case if nfs_updatepage detects a conflicting request
1034  * that has been written but not committed.
1035  */
1036 static int nfs_flush_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
1037 {
1038         struct nfs_page         *req;
1039         struct page             **pages;
1040         struct nfs_write_data   *data;
1041         struct list_head *head = &desc->pg_list;
1042         int ret = 0;
1043
1044         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
1045                                                       desc->pg_count));
1046         if (!data) {
1047                 while (!list_empty(head)) {
1048                         req = nfs_list_entry(head->next);
1049                         nfs_list_remove_request(req);
1050                         nfs_redirty_request(req);
1051                 }
1052                 ret = -ENOMEM;
1053                 goto out;
1054         }
1055         pages = data->pagevec;
1056         while (!list_empty(head)) {
1057                 req = nfs_list_entry(head->next);
1058                 nfs_list_remove_request(req);
1059                 nfs_list_add_request(req, &data->pages);
1060                 *pages++ = req->wb_page;
1061         }
1062         req = nfs_list_entry(data->pages.next);
1063
1064         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1065             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1066                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1067
1068         /* Set up the argument struct */
1069         nfs_write_rpcsetup(req, data, desc->pg_count, 0, desc->pg_ioflags);
1070         list_add(&data->list, res);
1071         desc->pg_rpc_callops = &nfs_write_full_ops;
1072 out:
1073         return ret;
1074 }
1075
1076 int nfs_generic_flush(struct nfs_pageio_descriptor *desc, struct list_head *head)
1077 {
1078         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1079                 return nfs_flush_multi(desc, head);
1080         return nfs_flush_one(desc, head);
1081 }
1082
1083 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1084 {
1085         LIST_HEAD(head);
1086         int ret;
1087
1088         ret = nfs_generic_flush(desc, &head);
1089         if (ret == 0)
1090                 ret = nfs_do_multiple_writes(&head, desc->pg_rpc_callops,
1091                                 desc->pg_ioflags);
1092         return ret;
1093 }
1094
1095 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1096         .pg_test = nfs_generic_pg_test,
1097         .pg_doio = nfs_generic_pg_writepages,
1098 };
1099
1100 void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1101                                   struct inode *inode, int ioflags)
1102 {
1103         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops,
1104                                 NFS_SERVER(inode)->wsize, ioflags);
1105 }
1106
1107 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1108 {
1109         pgio->pg_ops = &nfs_pageio_write_ops;
1110         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1111 }
1112 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1113
1114 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1115                                   struct inode *inode, int ioflags)
1116 {
1117         if (!pnfs_pageio_init_write(pgio, inode, ioflags))
1118                 nfs_pageio_init_write_mds(pgio, inode, ioflags);
1119 }
1120
1121 /*
1122  * Handle a write reply that flushed part of a page.
1123  */
1124 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1125 {
1126         struct nfs_write_data   *data = calldata;
1127
1128         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1129                 task->tk_pid,
1130                 data->req->wb_context->dentry->d_inode->i_sb->s_id,
1131                 (long long)
1132                   NFS_FILEID(data->req->wb_context->dentry->d_inode),
1133                 data->req->wb_bytes, (long long)req_offset(data->req));
1134
1135         nfs_writeback_done(task, data);
1136 }
1137
1138 static void nfs_writeback_release_partial(void *calldata)
1139 {
1140         struct nfs_write_data   *data = calldata;
1141         struct nfs_page         *req = data->req;
1142         struct page             *page = req->wb_page;
1143         int status = data->task.tk_status;
1144
1145         if (status < 0) {
1146                 nfs_set_pageerror(page);
1147                 nfs_context_set_write_error(req->wb_context, status);
1148                 dprintk(", error = %d\n", status);
1149                 goto out;
1150         }
1151
1152         if (nfs_write_need_commit(data)) {
1153                 struct inode *inode = page->mapping->host;
1154
1155                 spin_lock(&inode->i_lock);
1156                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1157                         /* Do nothing we need to resend the writes */
1158                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1159                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1160                         dprintk(" defer commit\n");
1161                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1162                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1163                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1164                         dprintk(" server reboot detected\n");
1165                 }
1166                 spin_unlock(&inode->i_lock);
1167         } else
1168                 dprintk(" OK\n");
1169
1170 out:
1171         if (atomic_dec_and_test(&req->wb_complete))
1172                 nfs_writepage_release(req, data);
1173         nfs_writedata_release(calldata);
1174 }
1175
1176 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1177 {
1178         struct nfs_write_data *data = calldata;
1179         NFS_PROTO(data->inode)->write_rpc_prepare(task, data);
1180 }
1181
1182 static const struct rpc_call_ops nfs_write_partial_ops = {
1183         .rpc_call_prepare = nfs_write_prepare,
1184         .rpc_call_done = nfs_writeback_done_partial,
1185         .rpc_release = nfs_writeback_release_partial,
1186 };
1187
1188 /*
1189  * Handle a write reply that flushes a whole page.
1190  *
1191  * FIXME: There is an inherent race with invalidate_inode_pages and
1192  *        writebacks since the page->count is kept > 1 for as long
1193  *        as the page has a write request pending.
1194  */
1195 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1196 {
1197         struct nfs_write_data   *data = calldata;
1198
1199         nfs_writeback_done(task, data);
1200 }
1201
1202 static void nfs_writeback_release_full(void *calldata)
1203 {
1204         struct nfs_write_data   *data = calldata;
1205         int status = data->task.tk_status;
1206
1207         /* Update attributes as result of writeback. */
1208         while (!list_empty(&data->pages)) {
1209                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1210                 struct page *page = req->wb_page;
1211
1212                 nfs_list_remove_request(req);
1213
1214                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1215                         data->task.tk_pid,
1216                         req->wb_context->dentry->d_inode->i_sb->s_id,
1217                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1218                         req->wb_bytes,
1219                         (long long)req_offset(req));
1220
1221                 if (status < 0) {
1222                         nfs_set_pageerror(page);
1223                         nfs_context_set_write_error(req->wb_context, status);
1224                         dprintk(", error = %d\n", status);
1225                         goto remove_request;
1226                 }
1227
1228                 if (nfs_write_need_commit(data)) {
1229                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1230                         nfs_mark_request_commit(req, data->lseg);
1231                         dprintk(" marked for commit\n");
1232                         goto next;
1233                 }
1234                 dprintk(" OK\n");
1235 remove_request:
1236                 nfs_inode_remove_request(req);
1237         next:
1238                 nfs_unlock_request(req);
1239                 nfs_end_page_writeback(page);
1240         }
1241         nfs_writedata_release(calldata);
1242 }
1243
1244 static const struct rpc_call_ops nfs_write_full_ops = {
1245         .rpc_call_prepare = nfs_write_prepare,
1246         .rpc_call_done = nfs_writeback_done_full,
1247         .rpc_release = nfs_writeback_release_full,
1248 };
1249
1250
1251 /*
1252  * This function is called when the WRITE call is complete.
1253  */
1254 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1255 {
1256         struct nfs_writeargs    *argp = &data->args;
1257         struct nfs_writeres     *resp = &data->res;
1258         int status;
1259
1260         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1261                 task->tk_pid, task->tk_status);
1262
1263         /*
1264          * ->write_done will attempt to use post-op attributes to detect
1265          * conflicting writes by other clients.  A strict interpretation
1266          * of close-to-open would allow us to continue caching even if
1267          * another writer had changed the file, but some applications
1268          * depend on tighter cache coherency when writing.
1269          */
1270         status = NFS_PROTO(data->inode)->write_done(task, data);
1271         if (status != 0)
1272                 return;
1273         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1274
1275 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1276         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1277                 /* We tried a write call, but the server did not
1278                  * commit data to stable storage even though we
1279                  * requested it.
1280                  * Note: There is a known bug in Tru64 < 5.0 in which
1281                  *       the server reports NFS_DATA_SYNC, but performs
1282                  *       NFS_FILE_SYNC. We therefore implement this checking
1283                  *       as a dprintk() in order to avoid filling syslog.
1284                  */
1285                 static unsigned long    complain;
1286
1287                 /* Note this will print the MDS for a DS write */
1288                 if (time_before(complain, jiffies)) {
1289                         dprintk("NFS:       faulty NFS server %s:"
1290                                 " (committed = %d) != (stable = %d)\n",
1291                                 NFS_SERVER(data->inode)->nfs_client->cl_hostname,
1292                                 resp->verf->committed, argp->stable);
1293                         complain = jiffies + 300 * HZ;
1294                 }
1295         }
1296 #endif
1297         /* Is this a short write? */
1298         if (task->tk_status >= 0 && resp->count < argp->count) {
1299                 static unsigned long    complain;
1300
1301                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1302
1303                 /* Has the server at least made some progress? */
1304                 if (resp->count != 0) {
1305                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1306                         if (resp->verf->committed != NFS_UNSTABLE) {
1307                                 /* Resend from where the server left off */
1308                                 data->mds_offset += resp->count;
1309                                 argp->offset += resp->count;
1310                                 argp->pgbase += resp->count;
1311                                 argp->count -= resp->count;
1312                         } else {
1313                                 /* Resend as a stable write in order to avoid
1314                                  * headaches in the case of a server crash.
1315                                  */
1316                                 argp->stable = NFS_FILE_SYNC;
1317                         }
1318                         rpc_restart_call_prepare(task);
1319                         return;
1320                 }
1321                 if (time_before(complain, jiffies)) {
1322                         printk(KERN_WARNING
1323                                "NFS: Server wrote zero bytes, expected %u.\n",
1324                                         argp->count);
1325                         complain = jiffies + 300 * HZ;
1326                 }
1327                 /* Can't do anything about it except throw an error. */
1328                 task->tk_status = -EIO;
1329         }
1330         return;
1331 }
1332
1333
1334 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1335 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1336 {
1337         int ret;
1338
1339         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1340                 return 1;
1341         if (!may_wait)
1342                 return 0;
1343         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1344                                 NFS_INO_COMMIT,
1345                                 nfs_wait_bit_killable,
1346                                 TASK_KILLABLE);
1347         return (ret < 0) ? ret : 1;
1348 }
1349
1350 void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1351 {
1352         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1353         smp_mb__after_clear_bit();
1354         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1355 }
1356 EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1357
1358 void nfs_commitdata_release(void *data)
1359 {
1360         struct nfs_write_data *wdata = data;
1361
1362         put_nfs_open_context(wdata->args.context);
1363         nfs_commit_free(wdata);
1364 }
1365 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1366
1367 int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1368                         const struct rpc_call_ops *call_ops,
1369                         int how)
1370 {
1371         struct rpc_task *task;
1372         int priority = flush_task_priority(how);
1373         struct rpc_message msg = {
1374                 .rpc_argp = &data->args,
1375                 .rpc_resp = &data->res,
1376                 .rpc_cred = data->cred,
1377         };
1378         struct rpc_task_setup task_setup_data = {
1379                 .task = &data->task,
1380                 .rpc_client = clnt,
1381                 .rpc_message = &msg,
1382                 .callback_ops = call_ops,
1383                 .callback_data = data,
1384                 .workqueue = nfsiod_workqueue,
1385                 .flags = RPC_TASK_ASYNC,
1386                 .priority = priority,
1387         };
1388         /* Set up the initial task struct.  */
1389         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1390
1391         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1392
1393         task = rpc_run_task(&task_setup_data);
1394         if (IS_ERR(task))
1395                 return PTR_ERR(task);
1396         if (how & FLUSH_SYNC)
1397                 rpc_wait_for_completion_task(task);
1398         rpc_put_task(task);
1399         return 0;
1400 }
1401 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1402
1403 /*
1404  * Set up the argument/result storage required for the RPC call.
1405  */
1406 void nfs_init_commit(struct nfs_write_data *data,
1407                             struct list_head *head,
1408                             struct pnfs_layout_segment *lseg)
1409 {
1410         struct nfs_page *first = nfs_list_entry(head->next);
1411         struct inode *inode = first->wb_context->dentry->d_inode;
1412
1413         /* Set up the RPC argument and reply structs
1414          * NB: take care not to mess about with data->commit et al. */
1415
1416         list_splice_init(head, &data->pages);
1417
1418         data->inode       = inode;
1419         data->cred        = first->wb_context->cred;
1420         data->lseg        = lseg; /* reference transferred */
1421         data->mds_ops     = &nfs_commit_ops;
1422
1423         data->args.fh     = NFS_FH(data->inode);
1424         /* Note: we always request a commit of the entire inode */
1425         data->args.offset = 0;
1426         data->args.count  = 0;
1427         data->args.context = get_nfs_open_context(first->wb_context);
1428         data->res.count   = 0;
1429         data->res.fattr   = &data->fattr;
1430         data->res.verf    = &data->verf;
1431         nfs_fattr_init(&data->fattr);
1432 }
1433 EXPORT_SYMBOL_GPL(nfs_init_commit);
1434
1435 void nfs_retry_commit(struct list_head *page_list,
1436                       struct pnfs_layout_segment *lseg)
1437 {
1438         struct nfs_page *req;
1439
1440         while (!list_empty(page_list)) {
1441                 req = nfs_list_entry(page_list->next);
1442                 nfs_list_remove_request(req);
1443                 nfs_mark_request_commit(req, lseg);
1444                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1445                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1446                              BDI_RECLAIMABLE);
1447                 nfs_unlock_request(req);
1448         }
1449 }
1450 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1451
1452 /*
1453  * Commit dirty pages
1454  */
1455 static int
1456 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1457 {
1458         struct nfs_write_data   *data;
1459
1460         data = nfs_commitdata_alloc();
1461
1462         if (!data)
1463                 goto out_bad;
1464
1465         /* Set up the argument struct */
1466         nfs_init_commit(data, head, NULL);
1467         return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1468  out_bad:
1469         nfs_retry_commit(head, NULL);
1470         nfs_commit_clear_lock(NFS_I(inode));
1471         return -ENOMEM;
1472 }
1473
1474 /*
1475  * COMMIT call returned
1476  */
1477 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1478 {
1479         struct nfs_write_data   *data = calldata;
1480
1481         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1482                                 task->tk_pid, task->tk_status);
1483
1484         /* Call the NFS version-specific code */
1485         NFS_PROTO(data->inode)->commit_done(task, data);
1486 }
1487
1488 void nfs_commit_release_pages(struct nfs_write_data *data)
1489 {
1490         struct nfs_page *req;
1491         int status = data->task.tk_status;
1492
1493         while (!list_empty(&data->pages)) {
1494                 req = nfs_list_entry(data->pages.next);
1495                 nfs_list_remove_request(req);
1496                 nfs_clear_page_commit(req->wb_page);
1497
1498                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1499                         req->wb_context->dentry->d_sb->s_id,
1500                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1501                         req->wb_bytes,
1502                         (long long)req_offset(req));
1503                 if (status < 0) {
1504                         nfs_context_set_write_error(req->wb_context, status);
1505                         nfs_inode_remove_request(req);
1506                         dprintk(", error = %d\n", status);
1507                         goto next;
1508                 }
1509
1510                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1511                  * returned by the server against all stored verfs. */
1512                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1513                         /* We have a match */
1514                         nfs_inode_remove_request(req);
1515                         dprintk(" OK\n");
1516                         goto next;
1517                 }
1518                 /* We have a mismatch. Write the page again */
1519                 dprintk(" mismatch\n");
1520                 nfs_mark_request_dirty(req);
1521         next:
1522                 nfs_unlock_request(req);
1523         }
1524 }
1525 EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1526
1527 static void nfs_commit_release(void *calldata)
1528 {
1529         struct nfs_write_data *data = calldata;
1530
1531         nfs_commit_release_pages(data);
1532         nfs_commit_clear_lock(NFS_I(data->inode));
1533         nfs_commitdata_release(calldata);
1534 }
1535
1536 static const struct rpc_call_ops nfs_commit_ops = {
1537         .rpc_call_prepare = nfs_write_prepare,
1538         .rpc_call_done = nfs_commit_done,
1539         .rpc_release = nfs_commit_release,
1540 };
1541
1542 int nfs_commit_inode(struct inode *inode, int how)
1543 {
1544         LIST_HEAD(head);
1545         int may_wait = how & FLUSH_SYNC;
1546         int res;
1547
1548         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1549         if (res <= 0)
1550                 goto out_mark_dirty;
1551         res = nfs_scan_commit(inode, &head);
1552         if (res) {
1553                 int error;
1554
1555                 error = pnfs_commit_list(inode, &head, how);
1556                 if (error == PNFS_NOT_ATTEMPTED)
1557                         error = nfs_commit_list(inode, &head, how);
1558                 if (error < 0)
1559                         return error;
1560                 if (!may_wait)
1561                         goto out_mark_dirty;
1562                 error = wait_on_bit(&NFS_I(inode)->flags,
1563                                 NFS_INO_COMMIT,
1564                                 nfs_wait_bit_killable,
1565                                 TASK_KILLABLE);
1566                 if (error < 0)
1567                         return error;
1568         } else
1569                 nfs_commit_clear_lock(NFS_I(inode));
1570         return res;
1571         /* Note: If we exit without ensuring that the commit is complete,
1572          * we must mark the inode as dirty. Otherwise, future calls to
1573          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1574          * that the data is on the disk.
1575          */
1576 out_mark_dirty:
1577         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1578         return res;
1579 }
1580
1581 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1582 {
1583         struct nfs_inode *nfsi = NFS_I(inode);
1584         int flags = FLUSH_SYNC;
1585         int ret = 0;
1586
1587         /* no commits means nothing needs to be done */
1588         if (!nfsi->ncommit)
1589                 return ret;
1590
1591         if (wbc->sync_mode == WB_SYNC_NONE) {
1592                 /* Don't commit yet if this is a non-blocking flush and there
1593                  * are a lot of outstanding writes for this mapping.
1594                  */
1595                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1596                         goto out_mark_dirty;
1597
1598                 /* don't wait for the COMMIT response */
1599                 flags = 0;
1600         }
1601
1602         ret = nfs_commit_inode(inode, flags);
1603         if (ret >= 0) {
1604                 if (wbc->sync_mode == WB_SYNC_NONE) {
1605                         if (ret < wbc->nr_to_write)
1606                                 wbc->nr_to_write -= ret;
1607                         else
1608                                 wbc->nr_to_write = 0;
1609                 }
1610                 return 0;
1611         }
1612 out_mark_dirty:
1613         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1614         return ret;
1615 }
1616 #else
1617 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1618 {
1619         return 0;
1620 }
1621 #endif
1622
1623 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1624 {
1625         int ret;
1626
1627         ret = nfs_commit_unstable_pages(inode, wbc);
1628         if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1629                 int status;
1630                 bool sync = true;
1631
1632                 if (wbc->sync_mode == WB_SYNC_NONE)
1633                         sync = false;
1634
1635                 status = pnfs_layoutcommit_inode(inode, sync);
1636                 if (status < 0)
1637                         return status;
1638         }
1639         return ret;
1640 }
1641
1642 /*
1643  * flush the inode to disk.
1644  */
1645 int nfs_wb_all(struct inode *inode)
1646 {
1647         struct writeback_control wbc = {
1648                 .sync_mode = WB_SYNC_ALL,
1649                 .nr_to_write = LONG_MAX,
1650                 .range_start = 0,
1651                 .range_end = LLONG_MAX,
1652         };
1653
1654         return sync_inode(inode, &wbc);
1655 }
1656
1657 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1658 {
1659         struct nfs_page *req;
1660         int ret = 0;
1661
1662         BUG_ON(!PageLocked(page));
1663         for (;;) {
1664                 wait_on_page_writeback(page);
1665                 req = nfs_page_find_request(page);
1666                 if (req == NULL)
1667                         break;
1668                 if (nfs_lock_request_dontget(req)) {
1669                         nfs_clear_request_commit(req);
1670                         nfs_inode_remove_request(req);
1671                         /*
1672                          * In case nfs_inode_remove_request has marked the
1673                          * page as being dirty
1674                          */
1675                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1676                         nfs_unlock_request(req);
1677                         break;
1678                 }
1679                 ret = nfs_wait_on_request(req);
1680                 nfs_release_request(req);
1681                 if (ret < 0)
1682                         break;
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * Write back all requests on one page - we do this before reading it.
1689  */
1690 int nfs_wb_page(struct inode *inode, struct page *page)
1691 {
1692         loff_t range_start = page_offset(page);
1693         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1694         struct writeback_control wbc = {
1695                 .sync_mode = WB_SYNC_ALL,
1696                 .nr_to_write = 0,
1697                 .range_start = range_start,
1698                 .range_end = range_end,
1699         };
1700         int ret;
1701
1702         for (;;) {
1703                 wait_on_page_writeback(page);
1704                 if (clear_page_dirty_for_io(page)) {
1705                         ret = nfs_writepage_locked(page, &wbc);
1706                         if (ret < 0)
1707                                 goto out_error;
1708                         continue;
1709                 }
1710                 if (!PagePrivate(page))
1711                         break;
1712                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1713                 if (ret < 0)
1714                         goto out_error;
1715         }
1716         return 0;
1717 out_error:
1718         return ret;
1719 }
1720
1721 #ifdef CONFIG_MIGRATION
1722 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1723                 struct page *page, enum migrate_mode mode)
1724 {
1725         /*
1726          * If PagePrivate is set, then the page is currently associated with
1727          * an in-progress read or write request. Don't try to migrate it.
1728          *
1729          * FIXME: we could do this in principle, but we'll need a way to ensure
1730          *        that we can safely release the inode reference while holding
1731          *        the page lock.
1732          */
1733         if (PagePrivate(page))
1734                 return -EBUSY;
1735
1736         nfs_fscache_release_page(page, GFP_KERNEL);
1737
1738         return migrate_page(mapping, newpage, page, mode);
1739 }
1740 #endif
1741
1742 int __init nfs_init_writepagecache(void)
1743 {
1744         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1745                                              sizeof(struct nfs_write_data),
1746                                              0, SLAB_HWCACHE_ALIGN,
1747                                              NULL);
1748         if (nfs_wdata_cachep == NULL)
1749                 return -ENOMEM;
1750
1751         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1752                                                      nfs_wdata_cachep);
1753         if (nfs_wdata_mempool == NULL)
1754                 return -ENOMEM;
1755
1756         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1757                                                       nfs_wdata_cachep);
1758         if (nfs_commit_mempool == NULL)
1759                 return -ENOMEM;
1760
1761         /*
1762          * NFS congestion size, scale with available memory.
1763          *
1764          *  64MB:    8192k
1765          * 128MB:   11585k
1766          * 256MB:   16384k
1767          * 512MB:   23170k
1768          *   1GB:   32768k
1769          *   2GB:   46340k
1770          *   4GB:   65536k
1771          *   8GB:   92681k
1772          *  16GB:  131072k
1773          *
1774          * This allows larger machines to have larger/more transfers.
1775          * Limit the default to 256M
1776          */
1777         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1778         if (nfs_congestion_kb > 256*1024)
1779                 nfs_congestion_kb = 256*1024;
1780
1781         return 0;
1782 }
1783
1784 void nfs_destroy_writepagecache(void)
1785 {
1786         mempool_destroy(nfs_commit_mempool);
1787         mempool_destroy(nfs_wdata_mempool);
1788         kmem_cache_destroy(nfs_wdata_cachep);
1789 }
1790