fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23
  24 #include <asm/uaccess.h>
  25
  26 #include "delegation.h"
  27 #include "internal.h"
  28 #include "iostat.h"
  29 #include "nfs4_fs.h"
  30 #include "fscache.h"
  31 #include "pnfs.h"
  32
  33 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  34
  35 #define MIN_POOL_WRITE          (32)
  36 #define MIN_POOL_COMMIT         (4)
  37
  38 /*
  39  * Local function declarations
  40  */
  41 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *desc,
  42                                   struct inode *inode, int ioflags);
  43 static void nfs_redirty_request(struct nfs_page *req);
  44 static const struct rpc_call_ops nfs_write_partial_ops;
  45 static const struct rpc_call_ops nfs_write_full_ops;
  46 static const struct rpc_call_ops nfs_commit_ops;
  47
  48 static struct kmem_cache *nfs_wdata_cachep;
  49 static mempool_t *nfs_wdata_mempool;
  50 static mempool_t *nfs_commit_mempool;
  51
  52 struct nfs_write_data *nfs_commitdata_alloc(void)
  53 {
  54         struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOFS);
  55
  56         if (p) {
  57                 memset(p, 0, sizeof(*p));
  58                 INIT_LIST_HEAD(&p->pages);
  59         }
  60         return p;
  61 }
  62 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  63
  64 void nfs_commit_free(struct nfs_write_data *p)
  65 {
  66         if (p && (p->pagevec != &p->page_array[0]))
  67                 kfree(p->pagevec);
  68         mempool_free(p, nfs_commit_mempool);
  69 }
  70 EXPORT_SYMBOL_GPL(nfs_commit_free);
  71
  72 struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
  73 {
  74         struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, GFP_NOFS);
  75
  76         if (p) {
  77                 memset(p, 0, sizeof(*p));
  78                 INIT_LIST_HEAD(&p->pages);
  79                 p->npages = pagecount;
  80                 if (pagecount <= ARRAY_SIZE(p->page_array))
  81                         p->pagevec = p->page_array;
  82                 else {
  83                         p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_NOFS);
  84                         if (!p->pagevec) {
  85                                 mempool_free(p, nfs_wdata_mempool);
  86                                 p = NULL;
  87                         }
  88                 }
  89         }
  90         return p;
  91 }
  92
  93 void nfs_writedata_free(struct nfs_write_data *p)
  94 {
  95         if (p && (p->pagevec != &p->page_array[0]))
  96                 kfree(p->pagevec);
  97         mempool_free(p, nfs_wdata_mempool);
  98 }
  99
 100 void nfs_writedata_release(struct nfs_write_data *wdata)
 101 {
 102         put_lseg(wdata->lseg);
 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_set_page_tag_locked(req))
 239                         break;
 240                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 241                  *       then the call to nfs_set_page_tag_locked() 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 int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 378 {
 379         struct nfs_inode *nfsi = NFS_I(inode);
 380         int error;
 381
 382         error = radix_tree_preload(GFP_NOFS);
 383         if (error != 0)
 384                 goto out;
 385
 386         /* Lock the request! */
 387         nfs_lock_request_dontget(req);
 388
 389         spin_lock(&inode->i_lock);
 390         error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req);
 391         BUG_ON(error);
 392         if (!nfsi->npages && nfs_have_delegation(inode, FMODE_WRITE))
 393                 nfsi->change_attr++;
 394         set_bit(PG_MAPPED, &req->wb_flags);
 395         SetPagePrivate(req->wb_page);
 396         set_page_private(req->wb_page, (unsigned long)req);
 397         nfsi->npages++;
 398         kref_get(&req->wb_kref);
 399         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index,
 400                                 NFS_PAGE_TAG_LOCKED);
 401         spin_unlock(&inode->i_lock);
 402         radix_tree_preload_end();
 403 out:
 404         return error;
 405 }
 406
 407 /*
 408  * Remove a write request from an inode
 409  */
 410 static void nfs_inode_remove_request(struct nfs_page *req)
 411 {
 412         struct inode *inode = req->wb_context->path.dentry->d_inode;
 413         struct nfs_inode *nfsi = NFS_I(inode);
 414
 415         BUG_ON (!NFS_WBACK_BUSY(req));
 416
 417         spin_lock(&inode->i_lock);
 418         set_page_private(req->wb_page, 0);
 419         ClearPagePrivate(req->wb_page);
 420         clear_bit(PG_MAPPED, &req->wb_flags);
 421         radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
 422         nfsi->npages--;
 423         spin_unlock(&inode->i_lock);
 424         nfs_release_request(req);
 425 }
 426
 427 static void
 428 nfs_mark_request_dirty(struct nfs_page *req)
 429 {
 430         __set_page_dirty_nobuffers(req->wb_page);
 431         __mark_inode_dirty(req->wb_page->mapping->host, I_DIRTY_DATASYNC);
 432 }
 433
 434 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 435 /*
 436  * Add a request to the inode's commit list.
 437  */
 438 static void
 439 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 440 {
 441         struct inode *inode = req->wb_context->path.dentry->d_inode;
 442         struct nfs_inode *nfsi = NFS_I(inode);
 443
 444         spin_lock(&inode->i_lock);
 445         set_bit(PG_CLEAN, &(req)->wb_flags);
 446         radix_tree_tag_set(&nfsi->nfs_page_tree,
 447                         req->wb_index,
 448                         NFS_PAGE_TAG_COMMIT);
 449         nfsi->ncommit++;
 450         spin_unlock(&inode->i_lock);
 451         pnfs_mark_request_commit(req, lseg);
 452         inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 453         inc_bdi_stat(req->wb_page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 454         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 455 }
 456
 457 static int
 458 nfs_clear_request_commit(struct nfs_page *req)
 459 {
 460         struct page *page = req->wb_page;
 461
 462         if (test_and_clear_bit(PG_CLEAN, &(req)->wb_flags)) {
 463                 dec_zone_page_state(page, NR_UNSTABLE_NFS);
 464                 dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 465                 return 1;
 466         }
 467         return 0;
 468 }
 469
 470 static inline
 471 int nfs_write_need_commit(struct nfs_write_data *data)
 472 {
 473         if (data->verf.committed == NFS_DATA_SYNC)
 474                 return data->lseg == NULL;
 475         else
 476                 return data->verf.committed != NFS_FILE_SYNC;
 477 }
 478
 479 static inline
 480 int nfs_reschedule_unstable_write(struct nfs_page *req,
 481                                   struct nfs_write_data *data)
 482 {
 483         if (test_and_clear_bit(PG_NEED_COMMIT, &req->wb_flags)) {
 484                 nfs_mark_request_commit(req, data->lseg);
 485                 return 1;
 486         }
 487         if (test_and_clear_bit(PG_NEED_RESCHED, &req->wb_flags)) {
 488                 nfs_mark_request_dirty(req);
 489                 return 1;
 490         }
 491         return 0;
 492 }
 493 #else
 494 static inline void
 495 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg)
 496 {
 497 }
 498
 499 static inline int
 500 nfs_clear_request_commit(struct nfs_page *req)
 501 {
 502         return 0;
 503 }
 504
 505 static inline
 506 int nfs_write_need_commit(struct nfs_write_data *data)
 507 {
 508         return 0;
 509 }
 510
 511 static inline
 512 int nfs_reschedule_unstable_write(struct nfs_page *req,
 513                                   struct nfs_write_data *data)
 514 {
 515         return 0;
 516 }
 517 #endif
 518
 519 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 520 static int
 521 nfs_need_commit(struct nfs_inode *nfsi)
 522 {
 523         return radix_tree_tagged(&nfsi->nfs_page_tree, NFS_PAGE_TAG_COMMIT);
 524 }
 525
 526 /*
 527  * nfs_scan_commit - Scan an inode for commit requests
 528  * @inode: NFS inode to scan
 529  * @dst: destination list
 530  * @idx_start: lower bound of page->index to scan.
 531  * @npages: idx_start + npages sets the upper bound to scan.
 532  *
 533  * Moves requests from the inode's 'commit' request list.
 534  * The requests are *not* checked to ensure that they form a contiguous set.
 535  */
 536 static int
 537 nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 538 {
 539         struct nfs_inode *nfsi = NFS_I(inode);
 540         int ret;
 541
 542         if (!nfs_need_commit(nfsi))
 543                 return 0;
 544
 545         spin_lock(&inode->i_lock);
 546         ret = nfs_scan_list(nfsi, dst, idx_start, npages, NFS_PAGE_TAG_COMMIT);
 547         if (ret > 0)
 548                 nfsi->ncommit -= ret;
 549         spin_unlock(&inode->i_lock);
 550
 551         if (nfs_need_commit(NFS_I(inode)))
 552                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 553
 554         return ret;
 555 }
 556 #else
 557 static inline int nfs_need_commit(struct nfs_inode *nfsi)
 558 {
 559         return 0;
 560 }
 561
 562 static inline int nfs_scan_commit(struct inode *inode, struct list_head *dst, pgoff_t idx_start, unsigned int npages)
 563 {
 564         return 0;
 565 }
 566 #endif
 567
 568 /*
 569  * Search for an existing write request, and attempt to update
 570  * it to reflect a new dirty region on a given page.
 571  *
 572  * If the attempt fails, then the existing request is flushed out
 573  * to disk.
 574  */
 575 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 576                 struct page *page,
 577                 unsigned int offset,
 578                 unsigned int bytes)
 579 {
 580         struct nfs_page *req;
 581         unsigned int rqend;
 582         unsigned int end;
 583         int error;
 584
 585         if (!PagePrivate(page))
 586                 return NULL;
 587
 588         end = offset + bytes;
 589         spin_lock(&inode->i_lock);
 590
 591         for (;;) {
 592                 req = nfs_page_find_request_locked(page);
 593                 if (req == NULL)
 594                         goto out_unlock;
 595
 596                 rqend = req->wb_offset + req->wb_bytes;
 597                 /*
 598                  * Tell the caller to flush out the request if
 599                  * the offsets are non-contiguous.
 600                  * Note: nfs_flush_incompatible() will already
 601                  * have flushed out requests having wrong owners.
 602                  */
 603                 if (offset > rqend
 604                     || end < req->wb_offset)
 605                         goto out_flushme;
 606
 607                 if (nfs_set_page_tag_locked(req))
 608                         break;
 609
 610                 /* The request is locked, so wait and then retry */
 611                 spin_unlock(&inode->i_lock);
 612                 error = nfs_wait_on_request(req);
 613                 nfs_release_request(req);
 614                 if (error != 0)
 615                         goto out_err;
 616                 spin_lock(&inode->i_lock);
 617         }
 618
 619         if (nfs_clear_request_commit(req) &&
 620             radix_tree_tag_clear(&NFS_I(inode)->nfs_page_tree,
 621                                  req->wb_index, NFS_PAGE_TAG_COMMIT) != NULL) {
 622                 NFS_I(inode)->ncommit--;
 623                 pnfs_clear_request_commit(req);
 624         }
 625
 626         /* Okay, the request matches. Update the region */
 627         if (offset < req->wb_offset) {
 628                 req->wb_offset = offset;
 629                 req->wb_pgbase = offset;
 630         }
 631         if (end > rqend)
 632                 req->wb_bytes = end - req->wb_offset;
 633         else
 634                 req->wb_bytes = rqend - req->wb_offset;
 635 out_unlock:
 636         spin_unlock(&inode->i_lock);
 637         return req;
 638 out_flushme:
 639         spin_unlock(&inode->i_lock);
 640         nfs_release_request(req);
 641         error = nfs_wb_page(inode, page);
 642 out_err:
 643         return ERR_PTR(error);
 644 }
 645
 646 /*
 647  * Try to update an existing write request, or create one if there is none.
 648  *
 649  * Note: Should always be called with the Page Lock held to prevent races
 650  * if we have to add a new request. Also assumes that the caller has
 651  * already called nfs_flush_incompatible() if necessary.
 652  */
 653 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 654                 struct page *page, unsigned int offset, unsigned int bytes)
 655 {
 656         struct inode *inode = page->mapping->host;
 657         struct nfs_page *req;
 658         int error;
 659
 660         req = nfs_try_to_update_request(inode, page, offset, bytes);
 661         if (req != NULL)
 662                 goto out;
 663         req = nfs_create_request(ctx, inode, page, offset, bytes);
 664         if (IS_ERR(req))
 665                 goto out;
 666         error = nfs_inode_add_request(inode, req);
 667         if (error != 0) {
 668                 nfs_release_request(req);
 669                 req = ERR_PTR(error);
 670         }
 671 out:
 672         return req;
 673 }
 674
 675 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 676                 unsigned int offset, unsigned int count)
 677 {
 678         struct nfs_page *req;
 679
 680         req = nfs_setup_write_request(ctx, page, offset, count);
 681         if (IS_ERR(req))
 682                 return PTR_ERR(req);
 683         /* Update file length */
 684         nfs_grow_file(page, offset, count);
 685         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 686         nfs_mark_request_dirty(req);
 687         nfs_clear_page_tag_locked(req);
 688         return 0;
 689 }
 690
 691 int nfs_flush_incompatible(struct file *file, struct page *page)
 692 {
 693         struct nfs_open_context *ctx = nfs_file_open_context(file);
 694         struct nfs_page *req;
 695         int do_flush, status;
 696         /*
 697          * Look for a request corresponding to this page. If there
 698          * is one, and it belongs to another file, we flush it out
 699          * before we try to copy anything into the page. Do this
 700          * due to the lack of an ACCESS-type call in NFSv2.
 701          * Also do the same if we find a request from an existing
 702          * dropped page.
 703          */
 704         do {
 705                 req = nfs_page_find_request(page);
 706                 if (req == NULL)
 707                         return 0;
 708                 do_flush = req->wb_page != page || req->wb_context != ctx ||
 709                         req->wb_lock_context->lockowner != current->files ||
 710                         req->wb_lock_context->pid != current->tgid;
 711                 nfs_release_request(req);
 712                 if (!do_flush)
 713                         return 0;
 714                 status = nfs_wb_page(page->mapping->host, page);
 715         } while (status == 0);
 716         return status;
 717 }
 718
 719 /*
 720  * If the page cache is marked as unsafe or invalid, then we can't rely on
 721  * the PageUptodate() flag. In this case, we will need to turn off
 722  * write optimisations that depend on the page contents being correct.
 723  */
 724 static int nfs_write_pageuptodate(struct page *page, struct inode *inode)
 725 {
 726         return PageUptodate(page) &&
 727                 !(NFS_I(inode)->cache_validity & (NFS_INO_REVAL_PAGECACHE|NFS_INO_INVALID_DATA));
 728 }
 729
 730 /*
 731  * Update and possibly write a cached page of an NFS file.
 732  *
 733  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 734  * things with a page scheduled for an RPC call (e.g. invalidate it).
 735  */
 736 int nfs_updatepage(struct file *file, struct page *page,
 737                 unsigned int offset, unsigned int count)
 738 {
 739         struct nfs_open_context *ctx = nfs_file_open_context(file);
 740         struct inode    *inode = page->mapping->host;
 741         int             status = 0;
 742
 743         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 744
 745         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 746                 file->f_path.dentry->d_parent->d_name.name,
 747                 file->f_path.dentry->d_name.name, count,
 748                 (long long)(page_offset(page) + offset));
 749
 750         /* If we're not using byte range locks, and we know the page
 751          * is up to date, it may be more efficient to extend the write
 752          * to cover the entire page in order to avoid fragmentation
 753          * inefficiencies.
 754          */
 755         if (nfs_write_pageuptodate(page, inode) &&
 756                         inode->i_flock == NULL &&
 757                         !(file->f_flags & O_DSYNC)) {
 758                 count = max(count + offset, nfs_page_length(page));
 759                 offset = 0;
 760         }
 761
 762         status = nfs_writepage_setup(ctx, page, offset, count);
 763         if (status < 0)
 764                 nfs_set_pageerror(page);
 765
 766         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 767                         status, (long long)i_size_read(inode));
 768         return status;
 769 }
 770
 771 static void nfs_writepage_release(struct nfs_page *req,
 772                                   struct nfs_write_data *data)
 773 {
 774         struct page *page = req->wb_page;
 775
 776         if (PageError(req->wb_page) || !nfs_reschedule_unstable_write(req, data))
 777                 nfs_inode_remove_request(req);
 778         nfs_clear_page_tag_locked(req);
 779         nfs_end_page_writeback(page);
 780 }
 781
 782 static int flush_task_priority(int how)
 783 {
 784         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 785                 case FLUSH_HIGHPRI:
 786                         return RPC_PRIORITY_HIGH;
 787                 case FLUSH_LOWPRI:
 788                         return RPC_PRIORITY_LOW;
 789         }
 790         return RPC_PRIORITY_NORMAL;
 791 }
 792
 793 int nfs_initiate_write(struct nfs_write_data *data,
 794                        struct rpc_clnt *clnt,
 795                        const struct rpc_call_ops *call_ops,
 796                        int how)
 797 {
 798         struct inode *inode = data->inode;
 799         int priority = flush_task_priority(how);
 800         struct rpc_task *task;
 801         struct rpc_message msg = {
 802                 .rpc_argp = &data->args,
 803                 .rpc_resp = &data->res,
 804                 .rpc_cred = data->cred,
 805         };
 806         struct rpc_task_setup task_setup_data = {
 807                 .rpc_client = clnt,
 808                 .task = &data->task,
 809                 .rpc_message = &msg,
 810                 .callback_ops = call_ops,
 811                 .callback_data = data,
 812                 .workqueue = nfsiod_workqueue,
 813                 .flags = RPC_TASK_ASYNC,
 814                 .priority = priority,
 815         };
 816         int ret = 0;
 817
 818         /* Set up the initial task struct.  */
 819         NFS_PROTO(inode)->write_setup(data, &msg);
 820
 821         dprintk("NFS: %5u initiated write call "
 822                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 823                 data->task.tk_pid,
 824                 inode->i_sb->s_id,
 825                 (long long)NFS_FILEID(inode),
 826                 data->args.count,
 827                 (unsigned long long)data->args.offset);
 828
 829         task = rpc_run_task(&task_setup_data);
 830         if (IS_ERR(task)) {
 831                 ret = PTR_ERR(task);
 832                 goto out;
 833         }
 834         if (how & FLUSH_SYNC) {
 835                 ret = rpc_wait_for_completion_task(task);
 836                 if (ret == 0)
 837                         ret = task->tk_status;
 838         }
 839         rpc_put_task(task);
 840 out:
 841         return ret;
 842 }
 843 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 844
 845 /*
 846  * Set up the argument/result storage required for the RPC call.
 847  */
 848 static void nfs_write_rpcsetup(struct nfs_page *req,
 849                 struct nfs_write_data *data,
 850                 unsigned int count, unsigned int offset,
 851                 int how)
 852 {
 853         struct inode *inode = req->wb_context->path.dentry->d_inode;
 854
 855         /* Set up the RPC argument and reply structs
 856          * NB: take care not to mess about with data->commit et al. */
 857
 858         data->req = req;
 859         data->inode = inode = req->wb_context->path.dentry->d_inode;
 860         data->cred = req->wb_context->cred;
 861
 862         data->args.fh     = NFS_FH(inode);
 863         data->args.offset = req_offset(req) + offset;
 864         /* pnfs_set_layoutcommit needs this */
 865         data->mds_offset = data->args.offset;
 866         data->args.pgbase = req->wb_pgbase + offset;
 867         data->args.pages  = data->pagevec;
 868         data->args.count  = count;
 869         data->args.context = get_nfs_open_context(req->wb_context);
 870         data->args.lock_context = req->wb_lock_context;
 871         data->args.stable  = NFS_UNSTABLE;
 872         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 873         case 0:
 874                 break;
 875         case FLUSH_COND_STABLE:
 876                 if (nfs_need_commit(NFS_I(inode)))
 877                         break;
 878         default:
 879                 data->args.stable = NFS_FILE_SYNC;
 880         }
 881
 882         data->res.fattr   = &data->fattr;
 883         data->res.count   = count;
 884         data->res.verf    = &data->verf;
 885         nfs_fattr_init(&data->fattr);
 886 }
 887
 888 static int nfs_do_write(struct nfs_write_data *data,
 889                 const struct rpc_call_ops *call_ops,
 890                 int how)
 891 {
 892         struct inode *inode = data->args.context->path.dentry->d_inode;
 893
 894         return nfs_initiate_write(data, NFS_CLIENT(inode), call_ops, how);
 895 }
 896
 897 static int nfs_do_multiple_writes(struct list_head *head,
 898                 const struct rpc_call_ops *call_ops,
 899                 int how)
 900 {
 901         struct nfs_write_data *data;
 902         int ret = 0;
 903
 904         while (!list_empty(head)) {
 905                 int ret2;
 906
 907                 data = list_entry(head->next, struct nfs_write_data, list);
 908                 list_del_init(&data->list);
 909
 910                 ret2 = nfs_do_write(data, call_ops, how);
 911                  if (ret == 0)
 912                          ret = ret2;
 913         }
 914         return ret;
 915 }
 916
 917 /* If a nfs_flush_* function fails, it should remove reqs from @head and
 918  * call this on each, which will prepare them to be retried on next
 919  * writeback using standard nfs.
 920  */
 921 static void nfs_redirty_request(struct nfs_page *req)
 922 {
 923         struct page *page = req->wb_page;
 924
 925         nfs_mark_request_dirty(req);
 926         nfs_clear_page_tag_locked(req);
 927         nfs_end_page_writeback(page);
 928 }
 929
 930 /*
 931  * Generate multiple small requests to write out a single
 932  * contiguous dirty area on one page.
 933  */
 934 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc, struct list_head *res)
 935 {
 936         struct nfs_page *req = nfs_list_entry(desc->pg_list.next);
 937         struct page *page = req->wb_page;
 938         struct nfs_write_data *data;
 939         size_t wsize = desc->pg_bsize, nbytes;
 940         unsigned int offset;
 941         int requests = 0;
 942         int ret = 0;
 943
 944         nfs_list_remove_request(req);
 945
 946         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 947             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit ||
 948              desc->pg_count > wsize))
 949                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 950
 951
 952         offset = 0;
 953         nbytes = desc->pg_count;
 954         do {
 955                 size_t len = min(nbytes, wsize);
 956
 957                 data = nfs_writedata_alloc(1);
 958                 if (!data)
 959                         goto out_bad;
 960                 data->pagevec[0] = page;
 961                 nfs_write_rpcsetup(req, data, wsize, offset, desc->pg_ioflags);
 962                 list_add(&data->list, res);
 963                 requests++;
 964                 nbytes -= len;
 965                 offset += len;
 966         } while (nbytes != 0);
 967         atomic_set(&req->wb_complete, requests);
 968         desc->pg_rpc_callops = &nfs_write_partial_ops;
 969         return ret;
 970
 971 out_bad:
 972         while (!list_empty(res)) {
 973                 data = list_entry(res->next, struct nfs_write_data, list);
 974                 list_del(&data->list);
 975                 nfs_writedata_free(data);
 976         }
 977         nfs_redirty_request(req);
 978         return -ENOMEM;
 979 }
 980
 981 /*
 982  * Create an RPC task for the given write request and kick it.
 983  * The page must have been locked by the caller.
 984  *
 985  * It may happen that the page we're passed is not marked dirty.
 986  * This is the case if nfs_updatepage detects a conflicting request
 987  * that has been written but not committed.
 988  */
 989 static int nfs_flush_one(struct nfs_pageio_descriptor *desc, struct list_head *res)
 990 {
 991         struct nfs_page         *req;
 992         struct page             **pages;
 993         struct nfs_write_data   *data;
 994         struct list_head *head = &desc->pg_list;
 995         int ret = 0;
 996
 997         data = nfs_writedata_alloc(nfs_page_array_len(desc->pg_base,
 998                                                       desc->pg_count));
 999         if (!data) {
1000                 while (!list_empty(head)) {
1001                         req = nfs_list_entry(head->next);
1002                         nfs_list_remove_request(req);
1003                         nfs_redirty_request(req);
1004                 }
1005                 ret = -ENOMEM;
1006                 goto out;
1007         }
1008         pages = data->pagevec;
1009         while (!list_empty(head)) {
1010                 req = nfs_list_entry(head->next);
1011                 nfs_list_remove_request(req);
1012                 nfs_list_add_request(req, &data->pages);
1013                 ClearPageError(req->wb_page);
1014                 *pages++ = req->wb_page;
1015         }
1016         req = nfs_list_entry(data->pages.next);
1017
1018         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1019             (desc->pg_moreio || NFS_I(desc->pg_inode)->ncommit))
1020                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1021
1022         /* Set up the argument struct */
1023         nfs_write_rpcsetup(req, data, desc->pg_count, 0, desc->pg_ioflags);
1024         list_add(&data->list, res);
1025         desc->pg_rpc_callops = &nfs_write_full_ops;
1026 out:
1027         return ret;
1028 }
1029
1030 int nfs_generic_flush(struct nfs_pageio_descriptor *desc, struct list_head *head)
1031 {
1032         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1033                 return nfs_flush_multi(desc, head);
1034         return nfs_flush_one(desc, head);
1035 }
1036
1037 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1038 {
1039         LIST_HEAD(head);
1040         int ret;
1041
1042         ret = nfs_generic_flush(desc, &head);
1043         if (ret == 0)
1044                 ret = nfs_do_multiple_writes(&head, desc->pg_rpc_callops,
1045                                 desc->pg_ioflags);
1046         return ret;
1047 }
1048
1049 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1050         .pg_test = nfs_generic_pg_test,
1051         .pg_doio = nfs_generic_pg_writepages,
1052 };
1053
1054 static void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,
1055                                   struct inode *inode, int ioflags)
1056 {
1057         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops,
1058                                 NFS_SERVER(inode)->wsize, ioflags);
1059 }
1060
1061 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1062 {
1063         pgio->pg_ops = &nfs_pageio_write_ops;
1064         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1065 }
1066 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1067
1068 static void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1069                                   struct inode *inode, int ioflags)
1070 {
1071         if (!pnfs_pageio_init_write(pgio, inode, ioflags))
1072                 nfs_pageio_init_write_mds(pgio, inode, ioflags);
1073 }
1074
1075 /*
1076  * Handle a write reply that flushed part of a page.
1077  */
1078 static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata)
1079 {
1080         struct nfs_write_data   *data = calldata;
1081
1082         dprintk("NFS: %5u write(%s/%lld %d@%lld)",
1083                 task->tk_pid,
1084                 data->req->wb_context->path.dentry->d_inode->i_sb->s_id,
1085                 (long long)
1086                   NFS_FILEID(data->req->wb_context->path.dentry->d_inode),
1087                 data->req->wb_bytes, (long long)req_offset(data->req));
1088
1089         nfs_writeback_done(task, data);
1090 }
1091
1092 static void nfs_writeback_release_partial(void *calldata)
1093 {
1094         struct nfs_write_data   *data = calldata;
1095         struct nfs_page         *req = data->req;
1096         struct page             *page = req->wb_page;
1097         int status = data->task.tk_status;
1098
1099         if (status < 0) {
1100                 nfs_set_pageerror(page);
1101                 nfs_context_set_write_error(req->wb_context, status);
1102                 dprintk(", error = %d\n", status);
1103                 goto out;
1104         }
1105
1106         if (nfs_write_need_commit(data)) {
1107                 struct inode *inode = page->mapping->host;
1108
1109                 spin_lock(&inode->i_lock);
1110                 if (test_bit(PG_NEED_RESCHED, &req->wb_flags)) {
1111                         /* Do nothing we need to resend the writes */
1112                 } else if (!test_and_set_bit(PG_NEED_COMMIT, &req->wb_flags)) {
1113                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1114                         dprintk(" defer commit\n");
1115                 } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) {
1116                         set_bit(PG_NEED_RESCHED, &req->wb_flags);
1117                         clear_bit(PG_NEED_COMMIT, &req->wb_flags);
1118                         dprintk(" server reboot detected\n");
1119                 }
1120                 spin_unlock(&inode->i_lock);
1121         } else
1122                 dprintk(" OK\n");
1123
1124 out:
1125         if (atomic_dec_and_test(&req->wb_complete))
1126                 nfs_writepage_release(req, data);
1127         nfs_writedata_release(calldata);
1128 }
1129
1130 #if defined(CONFIG_NFS_V4_1)
1131 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1132 {
1133         struct nfs_write_data *data = calldata;
1134
1135         if (nfs4_setup_sequence(NFS_SERVER(data->inode),
1136                                 &data->args.seq_args,
1137                                 &data->res.seq_res, 1, task))
1138                 return;
1139         rpc_call_start(task);
1140 }
1141 #endif /* CONFIG_NFS_V4_1 */
1142
1143 static const struct rpc_call_ops nfs_write_partial_ops = {
1144 #if defined(CONFIG_NFS_V4_1)
1145         .rpc_call_prepare = nfs_write_prepare,
1146 #endif /* CONFIG_NFS_V4_1 */
1147         .rpc_call_done = nfs_writeback_done_partial,
1148         .rpc_release = nfs_writeback_release_partial,
1149 };
1150
1151 /*
1152  * Handle a write reply that flushes a whole page.
1153  *
1154  * FIXME: There is an inherent race with invalidate_inode_pages and
1155  *        writebacks since the page->count is kept > 1 for as long
1156  *        as the page has a write request pending.
1157  */
1158 static void nfs_writeback_done_full(struct rpc_task *task, void *calldata)
1159 {
1160         struct nfs_write_data   *data = calldata;
1161
1162         nfs_writeback_done(task, data);
1163 }
1164
1165 static void nfs_writeback_release_full(void *calldata)
1166 {
1167         struct nfs_write_data   *data = calldata;
1168         int status = data->task.tk_status;
1169
1170         /* Update attributes as result of writeback. */
1171         while (!list_empty(&data->pages)) {
1172                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1173                 struct page *page = req->wb_page;
1174
1175                 nfs_list_remove_request(req);
1176
1177                 dprintk("NFS: %5u write (%s/%lld %d@%lld)",
1178                         data->task.tk_pid,
1179                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1180                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1181                         req->wb_bytes,
1182                         (long long)req_offset(req));
1183
1184                 if (status < 0) {
1185                         nfs_set_pageerror(page);
1186                         nfs_context_set_write_error(req->wb_context, status);
1187                         dprintk(", error = %d\n", status);
1188                         goto remove_request;
1189                 }
1190
1191                 if (nfs_write_need_commit(data)) {
1192                         memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf));
1193                         nfs_mark_request_commit(req, data->lseg);
1194                         dprintk(" marked for commit\n");
1195                         goto next;
1196                 }
1197                 dprintk(" OK\n");
1198 remove_request:
1199                 nfs_inode_remove_request(req);
1200         next:
1201                 nfs_clear_page_tag_locked(req);
1202                 nfs_end_page_writeback(page);
1203         }
1204         nfs_writedata_release(calldata);
1205 }
1206
1207 static const struct rpc_call_ops nfs_write_full_ops = {
1208 #if defined(CONFIG_NFS_V4_1)
1209         .rpc_call_prepare = nfs_write_prepare,
1210 #endif /* CONFIG_NFS_V4_1 */
1211         .rpc_call_done = nfs_writeback_done_full,
1212         .rpc_release = nfs_writeback_release_full,
1213 };
1214
1215
1216 /*
1217  * This function is called when the WRITE call is complete.
1218  */
1219 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1220 {
1221         struct nfs_writeargs    *argp = &data->args;
1222         struct nfs_writeres     *resp = &data->res;
1223         struct nfs_server       *server = NFS_SERVER(data->inode);
1224         int status;
1225
1226         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1227                 task->tk_pid, task->tk_status);
1228
1229         /*
1230          * ->write_done will attempt to use post-op attributes to detect
1231          * conflicting writes by other clients.  A strict interpretation
1232          * of close-to-open would allow us to continue caching even if
1233          * another writer had changed the file, but some applications
1234          * depend on tighter cache coherency when writing.
1235          */
1236         status = NFS_PROTO(data->inode)->write_done(task, data);
1237         if (status != 0)
1238                 return;
1239         nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1240
1241 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1242         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1243                 /* We tried a write call, but the server did not
1244                  * commit data to stable storage even though we
1245                  * requested it.
1246                  * Note: There is a known bug in Tru64 < 5.0 in which
1247                  *       the server reports NFS_DATA_SYNC, but performs
1248                  *       NFS_FILE_SYNC. We therefore implement this checking
1249                  *       as a dprintk() in order to avoid filling syslog.
1250                  */
1251                 static unsigned long    complain;
1252
1253                 /* Note this will print the MDS for a DS write */
1254                 if (time_before(complain, jiffies)) {
1255                         dprintk("NFS:       faulty NFS server %s:"
1256                                 " (committed = %d) != (stable = %d)\n",
1257                                 server->nfs_client->cl_hostname,
1258                                 resp->verf->committed, argp->stable);
1259                         complain = jiffies + 300 * HZ;
1260                 }
1261         }
1262 #endif
1263         /* Is this a short write? */
1264         if (task->tk_status >= 0 && resp->count < argp->count) {
1265                 static unsigned long    complain;
1266
1267                 nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE);
1268
1269                 /* Has the server at least made some progress? */
1270                 if (resp->count != 0) {
1271                         /* Was this an NFSv2 write or an NFSv3 stable write? */
1272                         if (resp->verf->committed != NFS_UNSTABLE) {
1273                                 /* Resend from where the server left off */
1274                                 data->mds_offset += resp->count;
1275                                 argp->offset += resp->count;
1276                                 argp->pgbase += resp->count;
1277                                 argp->count -= resp->count;
1278                         } else {
1279                                 /* Resend as a stable write in order to avoid
1280                                  * headaches in the case of a server crash.
1281                                  */
1282                                 argp->stable = NFS_FILE_SYNC;
1283                         }
1284                         nfs_restart_rpc(task, server->nfs_client);
1285                         return;
1286                 }
1287                 if (time_before(complain, jiffies)) {
1288                         printk(KERN_WARNING
1289                                "NFS: Server wrote zero bytes, expected %u.\n",
1290                                         argp->count);
1291                         complain = jiffies + 300 * HZ;
1292                 }
1293                 /* Can't do anything about it except throw an error. */
1294                 task->tk_status = -EIO;
1295         }
1296         return;
1297 }
1298
1299
1300 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
1301 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1302 {
1303         int ret;
1304
1305         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1306                 return 1;
1307         if (!may_wait)
1308                 return 0;
1309         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1310                                 NFS_INO_COMMIT,
1311                                 nfs_wait_bit_killable,
1312                                 TASK_KILLABLE);
1313         return (ret < 0) ? ret : 1;
1314 }
1315
1316 void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1317 {
1318         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1319         smp_mb__after_clear_bit();
1320         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1321 }
1322 EXPORT_SYMBOL_GPL(nfs_commit_clear_lock);
1323
1324 void nfs_commitdata_release(void *data)
1325 {
1326         struct nfs_write_data *wdata = data;
1327
1328         put_lseg(wdata->lseg);
1329         put_nfs_open_context(wdata->args.context);
1330         nfs_commit_free(wdata);
1331 }
1332 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1333
1334 int nfs_initiate_commit(struct nfs_write_data *data, struct rpc_clnt *clnt,
1335                         const struct rpc_call_ops *call_ops,
1336                         int how)
1337 {
1338         struct rpc_task *task;
1339         int priority = flush_task_priority(how);
1340         struct rpc_message msg = {
1341                 .rpc_argp = &data->args,
1342                 .rpc_resp = &data->res,
1343                 .rpc_cred = data->cred,
1344         };
1345         struct rpc_task_setup task_setup_data = {
1346                 .task = &data->task,
1347                 .rpc_client = clnt,
1348                 .rpc_message = &msg,
1349                 .callback_ops = call_ops,
1350                 .callback_data = data,
1351                 .workqueue = nfsiod_workqueue,
1352                 .flags = RPC_TASK_ASYNC,
1353                 .priority = priority,
1354         };
1355         /* Set up the initial task struct.  */
1356         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1357
1358         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1359
1360         task = rpc_run_task(&task_setup_data);
1361         if (IS_ERR(task))
1362                 return PTR_ERR(task);
1363         if (how & FLUSH_SYNC)
1364                 rpc_wait_for_completion_task(task);
1365         rpc_put_task(task);
1366         return 0;
1367 }
1368 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1369
1370 /*
1371  * Set up the argument/result storage required for the RPC call.
1372  */
1373 void nfs_init_commit(struct nfs_write_data *data,
1374                             struct list_head *head,
1375                             struct pnfs_layout_segment *lseg)
1376 {
1377         struct nfs_page *first = nfs_list_entry(head->next);
1378         struct inode *inode = first->wb_context->path.dentry->d_inode;
1379
1380         /* Set up the RPC argument and reply structs
1381          * NB: take care not to mess about with data->commit et al. */
1382
1383         list_splice_init(head, &data->pages);
1384
1385         data->inode       = inode;
1386         data->cred        = first->wb_context->cred;
1387         data->lseg        = lseg; /* reference transferred */
1388         data->mds_ops     = &nfs_commit_ops;
1389
1390         data->args.fh     = NFS_FH(data->inode);
1391         /* Note: we always request a commit of the entire inode */
1392         data->args.offset = 0;
1393         data->args.count  = 0;
1394         data->args.context = get_nfs_open_context(first->wb_context);
1395         data->res.count   = 0;
1396         data->res.fattr   = &data->fattr;
1397         data->res.verf    = &data->verf;
1398         nfs_fattr_init(&data->fattr);
1399 }
1400 EXPORT_SYMBOL_GPL(nfs_init_commit);
1401
1402 void nfs_retry_commit(struct list_head *page_list,
1403                       struct pnfs_layout_segment *lseg)
1404 {
1405         struct nfs_page *req;
1406
1407         while (!list_empty(page_list)) {
1408                 req = nfs_list_entry(page_list->next);
1409                 nfs_list_remove_request(req);
1410                 nfs_mark_request_commit(req, lseg);
1411                 dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1412                 dec_bdi_stat(req->wb_page->mapping->backing_dev_info,
1413                              BDI_RECLAIMABLE);
1414                 nfs_clear_page_tag_locked(req);
1415         }
1416 }
1417 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1418
1419 /*
1420  * Commit dirty pages
1421  */
1422 static int
1423 nfs_commit_list(struct inode *inode, struct list_head *head, int how)
1424 {
1425         struct nfs_write_data   *data;
1426
1427         data = nfs_commitdata_alloc();
1428
1429         if (!data)
1430                 goto out_bad;
1431
1432         /* Set up the argument struct */
1433         nfs_init_commit(data, head, NULL);
1434         return nfs_initiate_commit(data, NFS_CLIENT(inode), data->mds_ops, how);
1435  out_bad:
1436         nfs_retry_commit(head, NULL);
1437         nfs_commit_clear_lock(NFS_I(inode));
1438         return -ENOMEM;
1439 }
1440
1441 /*
1442  * COMMIT call returned
1443  */
1444 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1445 {
1446         struct nfs_write_data   *data = calldata;
1447
1448         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1449                                 task->tk_pid, task->tk_status);
1450
1451         /* Call the NFS version-specific code */
1452         NFS_PROTO(data->inode)->commit_done(task, data);
1453 }
1454
1455 void nfs_commit_release_pages(struct nfs_write_data *data)
1456 {
1457         struct nfs_page *req;
1458         int status = data->task.tk_status;
1459
1460         while (!list_empty(&data->pages)) {
1461                 req = nfs_list_entry(data->pages.next);
1462                 nfs_list_remove_request(req);
1463                 nfs_clear_request_commit(req);
1464
1465                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1466                         req->wb_context->path.dentry->d_inode->i_sb->s_id,
1467                         (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
1468                         req->wb_bytes,
1469                         (long long)req_offset(req));
1470                 if (status < 0) {
1471                         nfs_context_set_write_error(req->wb_context, status);
1472                         nfs_inode_remove_request(req);
1473                         dprintk(", error = %d\n", status);
1474                         goto next;
1475                 }
1476
1477                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1478                  * returned by the server against all stored verfs. */
1479                 if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) {
1480                         /* We have a match */
1481                         nfs_inode_remove_request(req);
1482                         dprintk(" OK\n");
1483                         goto next;
1484                 }
1485                 /* We have a mismatch. Write the page again */
1486                 dprintk(" mismatch\n");
1487                 nfs_mark_request_dirty(req);
1488         next:
1489                 nfs_clear_page_tag_locked(req);
1490         }
1491 }
1492 EXPORT_SYMBOL_GPL(nfs_commit_release_pages);
1493
1494 static void nfs_commit_release(void *calldata)
1495 {
1496         struct nfs_write_data *data = calldata;
1497
1498         nfs_commit_release_pages(data);
1499         nfs_commit_clear_lock(NFS_I(data->inode));
1500         nfs_commitdata_release(calldata);
1501 }
1502
1503 static const struct rpc_call_ops nfs_commit_ops = {
1504 #if defined(CONFIG_NFS_V4_1)
1505         .rpc_call_prepare = nfs_write_prepare,
1506 #endif /* CONFIG_NFS_V4_1 */
1507         .rpc_call_done = nfs_commit_done,
1508         .rpc_release = nfs_commit_release,
1509 };
1510
1511 int nfs_commit_inode(struct inode *inode, int how)
1512 {
1513         LIST_HEAD(head);
1514         int may_wait = how & FLUSH_SYNC;
1515         int res;
1516
1517         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1518         if (res <= 0)
1519                 goto out_mark_dirty;
1520         res = nfs_scan_commit(inode, &head, 0, 0);
1521         if (res) {
1522                 int error;
1523
1524                 error = pnfs_commit_list(inode, &head, how);
1525                 if (error == PNFS_NOT_ATTEMPTED)
1526                         error = nfs_commit_list(inode, &head, how);
1527                 if (error < 0)
1528                         return error;
1529                 if (!may_wait)
1530                         goto out_mark_dirty;
1531                 error = wait_on_bit(&NFS_I(inode)->flags,
1532                                 NFS_INO_COMMIT,
1533                                 nfs_wait_bit_killable,
1534                                 TASK_KILLABLE);
1535                 if (error < 0)
1536                         return error;
1537         } else
1538                 nfs_commit_clear_lock(NFS_I(inode));
1539         return res;
1540         /* Note: If we exit without ensuring that the commit is complete,
1541          * we must mark the inode as dirty. Otherwise, future calls to
1542          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1543          * that the data is on the disk.
1544          */
1545 out_mark_dirty:
1546         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1547         return res;
1548 }
1549
1550 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1551 {
1552         struct nfs_inode *nfsi = NFS_I(inode);
1553         int flags = FLUSH_SYNC;
1554         int ret = 0;
1555
1556         if (wbc->sync_mode == WB_SYNC_NONE) {
1557                 /* Don't commit yet if this is a non-blocking flush and there
1558                  * are a lot of outstanding writes for this mapping.
1559                  */
1560                 if (nfsi->ncommit <= (nfsi->npages >> 1))
1561                         goto out_mark_dirty;
1562
1563                 /* don't wait for the COMMIT response */
1564                 flags = 0;
1565         }
1566
1567         ret = nfs_commit_inode(inode, flags);
1568         if (ret >= 0) {
1569                 if (wbc->sync_mode == WB_SYNC_NONE) {
1570                         if (ret < wbc->nr_to_write)
1571                                 wbc->nr_to_write -= ret;
1572                         else
1573                                 wbc->nr_to_write = 0;
1574                 }
1575                 return 0;
1576         }
1577 out_mark_dirty:
1578         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1579         return ret;
1580 }
1581 #else
1582 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1583 {
1584         return 0;
1585 }
1586 #endif
1587
1588 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1589 {
1590         int ret;
1591
1592         ret = nfs_commit_unstable_pages(inode, wbc);
1593         if (ret >= 0 && test_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags)) {
1594                 int status;
1595                 bool sync = true;
1596
1597                 if (wbc->sync_mode == WB_SYNC_NONE || wbc->nonblocking ||
1598                     wbc->for_background)
1599                         sync = false;
1600
1601                 status = pnfs_layoutcommit_inode(inode, sync);
1602                 if (status < 0)
1603                         return status;
1604         }
1605         return ret;
1606 }
1607
1608 /*
1609  * flush the inode to disk.
1610  */
1611 int nfs_wb_all(struct inode *inode)
1612 {
1613         struct writeback_control wbc = {
1614                 .sync_mode = WB_SYNC_ALL,
1615                 .nr_to_write = LONG_MAX,
1616                 .range_start = 0,
1617                 .range_end = LLONG_MAX,
1618         };
1619
1620         return sync_inode(inode, &wbc);
1621 }
1622
1623 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1624 {
1625         struct nfs_page *req;
1626         int ret = 0;
1627
1628         BUG_ON(!PageLocked(page));
1629         for (;;) {
1630                 wait_on_page_writeback(page);
1631                 req = nfs_page_find_request(page);
1632                 if (req == NULL)
1633                         break;
1634                 if (nfs_lock_request_dontget(req)) {
1635                         nfs_inode_remove_request(req);
1636                         /*
1637                          * In case nfs_inode_remove_request has marked the
1638                          * page as being dirty
1639                          */
1640                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1641                         nfs_unlock_request(req);
1642                         break;
1643                 }
1644                 ret = nfs_wait_on_request(req);
1645                 nfs_release_request(req);
1646                 if (ret < 0)
1647                         break;
1648         }
1649         return ret;
1650 }
1651
1652 /*
1653  * Write back all requests on one page - we do this before reading it.
1654  */
1655 int nfs_wb_page(struct inode *inode, struct page *page)
1656 {
1657         loff_t range_start = page_offset(page);
1658         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1659         struct writeback_control wbc = {
1660                 .sync_mode = WB_SYNC_ALL,
1661                 .nr_to_write = 0,
1662                 .range_start = range_start,
1663                 .range_end = range_end,
1664         };
1665         int ret;
1666
1667         for (;;) {
1668                 wait_on_page_writeback(page);
1669                 if (clear_page_dirty_for_io(page)) {
1670                         ret = nfs_writepage_locked(page, &wbc);
1671                         if (ret < 0)
1672                                 goto out_error;
1673                         continue;
1674                 }
1675                 if (!PagePrivate(page))
1676                         break;
1677                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1678                 if (ret < 0)
1679                         goto out_error;
1680         }
1681         return 0;
1682 out_error:
1683         return ret;
1684 }
1685
1686 #ifdef CONFIG_MIGRATION
1687 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1688                 struct page *page)
1689 {
1690         struct nfs_page *req;
1691         int ret;
1692
1693         nfs_fscache_release_page(page, GFP_KERNEL);
1694
1695         req = nfs_find_and_lock_request(page, false);
1696         ret = PTR_ERR(req);
1697         if (IS_ERR(req))
1698                 goto out;
1699
1700         ret = migrate_page(mapping, newpage, page);
1701         if (!req)
1702                 goto out;
1703         if (ret)
1704                 goto out_unlock;
1705         page_cache_get(newpage);
1706         spin_lock(&mapping->host->i_lock);
1707         req->wb_page = newpage;
1708         SetPagePrivate(newpage);
1709         set_page_private(newpage, (unsigned long)req);
1710         ClearPagePrivate(page);
1711         set_page_private(page, 0);
1712         spin_unlock(&mapping->host->i_lock);
1713         page_cache_release(page);
1714 out_unlock:
1715         nfs_clear_page_tag_locked(req);
1716 out:
1717         return ret;
1718 }
1719 #endif
1720
1721 int __init nfs_init_writepagecache(void)
1722 {
1723         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1724                                              sizeof(struct nfs_write_data),
1725                                              0, SLAB_HWCACHE_ALIGN,
1726                                              NULL);
1727         if (nfs_wdata_cachep == NULL)
1728                 return -ENOMEM;
1729
1730         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1731                                                      nfs_wdata_cachep);
1732         if (nfs_wdata_mempool == NULL)
1733                 return -ENOMEM;
1734
1735         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1736                                                       nfs_wdata_cachep);
1737         if (nfs_commit_mempool == NULL)
1738                 return -ENOMEM;
1739
1740         /*
1741          * NFS congestion size, scale with available memory.
1742          *
1743          *  64MB:    8192k
1744          * 128MB:   11585k
1745          * 256MB:   16384k
1746          * 512MB:   23170k
1747          *   1GB:   32768k
1748          *   2GB:   46340k
1749          *   4GB:   65536k
1750          *   8GB:   92681k
1751          *  16GB:  131072k
1752          *
1753          * This allows larger machines to have larger/more transfers.
1754          * Limit the default to 256M
1755          */
1756         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1757         if (nfs_congestion_kb > 256*1024)
1758                 nfs_congestion_kb = 256*1024;
1759
1760         return 0;
1761 }
1762
1763 void nfs_destroy_writepagecache(void)
1764 {
1765         mempool_destroy(nfs_commit_mempool);
1766         mempool_destroy(nfs_wdata_mempool);
1767         kmem_cache_destroy(nfs_wdata_cachep);
1768 }
1769