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 #include "nfstrace.h"
  35
  36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  37
  38 #define MIN_POOL_WRITE          (32)
  39 #define MIN_POOL_COMMIT         (4)
  40
  41 /*
  42  * Local function declarations
  43  */
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_commit_ops;
  46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  48 static const struct nfs_rw_ops nfs_rw_write_ops;
  49
  50 static struct kmem_cache *nfs_wdata_cachep;
  51 static mempool_t *nfs_wdata_mempool;
  52 static struct kmem_cache *nfs_cdata_cachep;
  53 static mempool_t *nfs_commit_mempool;
  54
  55 struct nfs_commit_data *nfs_commitdata_alloc(void)
  56 {
  57         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  58
  59         if (p) {
  60                 memset(p, 0, sizeof(*p));
  61                 INIT_LIST_HEAD(&p->pages);
  62         }
  63         return p;
  64 }
  65 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  66
  67 void nfs_commit_free(struct nfs_commit_data *p)
  68 {
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71 EXPORT_SYMBOL_GPL(nfs_commit_free);
  72
  73 static struct nfs_rw_header *nfs_writehdr_alloc(void)
  74 {
  75         struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  76
  77         if (p)
  78                 memset(p, 0, sizeof(*p));
  79         return p;
  80 }
  81
  82 static void nfs_writehdr_free(struct nfs_rw_header *whdr)
  83 {
  84         mempool_free(whdr, nfs_wdata_mempool);
  85 }
  86
  87 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  88 {
  89         ctx->error = error;
  90         smp_wmb();
  91         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
  92 }
  93
  94 static struct nfs_page *
  95 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
  96 {
  97         struct nfs_page *req = NULL;
  98
  99         if (PagePrivate(page))
 100                 req = (struct nfs_page *)page_private(page);
 101         else if (unlikely(PageSwapCache(page))) {
 102                 struct nfs_page *freq, *t;
 103
 104                 /* Linearly search the commit list for the correct req */
 105                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 106                         if (freq->wb_page == page) {
 107                                 req = freq;
 108                                 break;
 109                         }
 110                 }
 111         }
 112
 113         if (req)
 114                 kref_get(&req->wb_kref);
 115
 116         return req;
 117 }
 118
 119 static struct nfs_page *nfs_page_find_request(struct page *page)
 120 {
 121         struct inode *inode = page_file_mapping(page)->host;
 122         struct nfs_page *req = NULL;
 123
 124         spin_lock(&inode->i_lock);
 125         req = nfs_page_find_request_locked(NFS_I(inode), page);
 126         spin_unlock(&inode->i_lock);
 127         return req;
 128 }
 129
 130 /* Adjust the file length if we're writing beyond the end */
 131 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 132 {
 133         struct inode *inode = page_file_mapping(page)->host;
 134         loff_t end, i_size;
 135         pgoff_t end_index;
 136
 137         spin_lock(&inode->i_lock);
 138         i_size = i_size_read(inode);
 139         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 140         if (i_size > 0 && page_file_index(page) < end_index)
 141                 goto out;
 142         end = page_file_offset(page) + ((loff_t)offset+count);
 143         if (i_size >= end)
 144                 goto out;
 145         i_size_write(inode, end);
 146         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 147 out:
 148         spin_unlock(&inode->i_lock);
 149 }
 150
 151 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 152 static void nfs_set_pageerror(struct page *page)
 153 {
 154         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 155 }
 156
 157 /*
 158  * nfs_page_group_search_locked
 159  * @head - head request of page group
 160  * @page_offset - offset into page
 161  *
 162  * Search page group with head @head to find a request that contains the
 163  * page offset @page_offset.
 164  *
 165  * Returns a pointer to the first matching nfs request, or NULL if no
 166  * match is found.
 167  *
 168  * Must be called with the page group lock held
 169  */
 170 static struct nfs_page *
 171 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 172 {
 173         struct nfs_page *req;
 174
 175         WARN_ON_ONCE(head != head->wb_head);
 176         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
 177
 178         req = head;
 179         do {
 180                 if (page_offset >= req->wb_pgbase &&
 181                     page_offset < (req->wb_pgbase + req->wb_bytes))
 182                         return req;
 183
 184                 req = req->wb_this_page;
 185         } while (req != head);
 186
 187         return NULL;
 188 }
 189
 190 /*
 191  * nfs_page_group_covers_page
 192  * @head - head request of page group
 193  *
 194  * Return true if the page group with head @head covers the whole page,
 195  * returns false otherwise
 196  */
 197 static bool nfs_page_group_covers_page(struct nfs_page *req)
 198 {
 199         struct nfs_page *tmp;
 200         unsigned int pos = 0;
 201         unsigned int len = nfs_page_length(req->wb_page);
 202
 203         nfs_page_group_lock(req);
 204
 205         do {
 206                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 207                 if (tmp) {
 208                         /* no way this should happen */
 209                         WARN_ON_ONCE(tmp->wb_pgbase != pos);
 210                         pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
 211                 }
 212         } while (tmp && pos < len);
 213
 214         nfs_page_group_unlock(req);
 215         WARN_ON_ONCE(pos > len);
 216         return pos == len;
 217 }
 218
 219 /* We can set the PG_uptodate flag if we see that a write request
 220  * covers the full page.
 221  */
 222 static void nfs_mark_uptodate(struct nfs_page *req)
 223 {
 224         if (PageUptodate(req->wb_page))
 225                 return;
 226         if (!nfs_page_group_covers_page(req))
 227                 return;
 228         SetPageUptodate(req->wb_page);
 229 }
 230
 231 static int wb_priority(struct writeback_control *wbc)
 232 {
 233         if (wbc->for_reclaim)
 234                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 235         if (wbc->for_kupdate || wbc->for_background)
 236                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 237         return FLUSH_COND_STABLE;
 238 }
 239
 240 /*
 241  * NFS congestion control
 242  */
 243
 244 int nfs_congestion_kb;
 245
 246 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 247 #define NFS_CONGESTION_OFF_THRESH       \
 248         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 249
 250 static void nfs_set_page_writeback(struct page *page)
 251 {
 252         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 253         int ret = test_set_page_writeback(page);
 254
 255         WARN_ON_ONCE(ret != 0);
 256
 257         if (atomic_long_inc_return(&nfss->writeback) >
 258                         NFS_CONGESTION_ON_THRESH) {
 259                 set_bdi_congested(&nfss->backing_dev_info,
 260                                         BLK_RW_ASYNC);
 261         }
 262 }
 263
 264 static void nfs_end_page_writeback(struct nfs_page *req)
 265 {
 266         struct inode *inode = page_file_mapping(req->wb_page)->host;
 267         struct nfs_server *nfss = NFS_SERVER(inode);
 268
 269         if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
 270                 return;
 271
 272         end_page_writeback(req->wb_page);
 273         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 274                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 275 }
 276
 277 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 278 {
 279         struct inode *inode = page_file_mapping(page)->host;
 280         struct nfs_page *req;
 281         int ret;
 282
 283         spin_lock(&inode->i_lock);
 284         for (;;) {
 285                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 286                 if (req == NULL)
 287                         break;
 288                 if (nfs_lock_request(req))
 289                         break;
 290                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 291                  *       then the call to nfs_lock_request() will always
 292                  *       succeed provided that someone hasn't already marked the
 293                  *       request as dirty (in which case we don't care).
 294                  */
 295                 spin_unlock(&inode->i_lock);
 296                 if (!nonblock)
 297                         ret = nfs_wait_on_request(req);
 298                 else
 299                         ret = -EAGAIN;
 300                 nfs_release_request(req);
 301                 if (ret != 0)
 302                         return ERR_PTR(ret);
 303                 spin_lock(&inode->i_lock);
 304         }
 305         spin_unlock(&inode->i_lock);
 306         return req;
 307 }
 308
 309 /*
 310  * Find an associated nfs write request, and prepare to flush it out
 311  * May return an error if the user signalled nfs_wait_on_request().
 312  */
 313 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 314                                 struct page *page, bool nonblock)
 315 {
 316         struct nfs_page *req;
 317         int ret = 0;
 318
 319         req = nfs_find_and_lock_request(page, nonblock);
 320         if (!req)
 321                 goto out;
 322         ret = PTR_ERR(req);
 323         if (IS_ERR(req))
 324                 goto out;
 325
 326         nfs_set_page_writeback(page);
 327         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 328
 329         ret = 0;
 330         if (!nfs_pageio_add_request(pgio, req)) {
 331                 nfs_redirty_request(req);
 332                 ret = pgio->pg_error;
 333         }
 334 out:
 335         return ret;
 336 }
 337
 338 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 339 {
 340         struct inode *inode = page_file_mapping(page)->host;
 341         int ret;
 342
 343         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 344         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 345
 346         nfs_pageio_cond_complete(pgio, page_file_index(page));
 347         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 348         if (ret == -EAGAIN) {
 349                 redirty_page_for_writepage(wbc, page);
 350                 ret = 0;
 351         }
 352         return ret;
 353 }
 354
 355 /*
 356  * Write an mmapped page to the server.
 357  */
 358 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 359 {
 360         struct nfs_pageio_descriptor pgio;
 361         int err;
 362
 363         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 364                                 false, &nfs_async_write_completion_ops);
 365         err = nfs_do_writepage(page, wbc, &pgio);
 366         nfs_pageio_complete(&pgio);
 367         if (err < 0)
 368                 return err;
 369         if (pgio.pg_error < 0)
 370                 return pgio.pg_error;
 371         return 0;
 372 }
 373
 374 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 375 {
 376         int ret;
 377
 378         ret = nfs_writepage_locked(page, wbc);
 379         unlock_page(page);
 380         return ret;
 381 }
 382
 383 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 384 {
 385         int ret;
 386
 387         ret = nfs_do_writepage(page, wbc, data);
 388         unlock_page(page);
 389         return ret;
 390 }
 391
 392 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 393 {
 394         struct inode *inode = mapping->host;
 395         unsigned long *bitlock = &NFS_I(inode)->flags;
 396         struct nfs_pageio_descriptor pgio;
 397         int err;
 398
 399         /* Stop dirtying of new pages while we sync */
 400         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 401                         nfs_wait_bit_killable, TASK_KILLABLE);
 402         if (err)
 403                 goto out_err;
 404
 405         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 406
 407         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 408                                 &nfs_async_write_completion_ops);
 409         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 410         nfs_pageio_complete(&pgio);
 411
 412         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 413         smp_mb__after_atomic();
 414         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 415
 416         if (err < 0)
 417                 goto out_err;
 418         err = pgio.pg_error;
 419         if (err < 0)
 420                 goto out_err;
 421         return 0;
 422 out_err:
 423         return err;
 424 }
 425
 426 /*
 427  * Insert a write request into an inode
 428  */
 429 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 430 {
 431         struct nfs_inode *nfsi = NFS_I(inode);
 432
 433         WARN_ON_ONCE(req->wb_this_page != req);
 434
 435         /* Lock the request! */
 436         nfs_lock_request(req);
 437
 438         spin_lock(&inode->i_lock);
 439         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 440                 inode->i_version++;
 441         /*
 442          * Swap-space should not get truncated. Hence no need to plug the race
 443          * with invalidate/truncate.
 444          */
 445         if (likely(!PageSwapCache(req->wb_page))) {
 446                 set_bit(PG_MAPPED, &req->wb_flags);
 447                 SetPagePrivate(req->wb_page);
 448                 set_page_private(req->wb_page, (unsigned long)req);
 449         }
 450         nfsi->npages++;
 451         /* this a head request for a page group - mark it as having an
 452          * extra reference so sub groups can follow suit */
 453         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 454         kref_get(&req->wb_kref);
 455         spin_unlock(&inode->i_lock);
 456 }
 457
 458 /*
 459  * Remove a write request from an inode
 460  */
 461 static void nfs_inode_remove_request(struct nfs_page *req)
 462 {
 463         struct inode *inode = req->wb_context->dentry->d_inode;
 464         struct nfs_inode *nfsi = NFS_I(inode);
 465         struct nfs_page *head;
 466
 467         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 468                 head = req->wb_head;
 469
 470                 spin_lock(&inode->i_lock);
 471                 if (likely(!PageSwapCache(head->wb_page))) {
 472                         set_page_private(head->wb_page, 0);
 473                         ClearPagePrivate(head->wb_page);
 474                         clear_bit(PG_MAPPED, &head->wb_flags);
 475                 }
 476                 nfsi->npages--;
 477                 spin_unlock(&inode->i_lock);
 478         }
 479
 480         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags))
 481                 nfs_release_request(req);
 482 }
 483
 484 static void
 485 nfs_mark_request_dirty(struct nfs_page *req)
 486 {
 487         __set_page_dirty_nobuffers(req->wb_page);
 488 }
 489
 490 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 491 /**
 492  * nfs_request_add_commit_list - add request to a commit list
 493  * @req: pointer to a struct nfs_page
 494  * @dst: commit list head
 495  * @cinfo: holds list lock and accounting info
 496  *
 497  * This sets the PG_CLEAN bit, updates the cinfo count of
 498  * number of outstanding requests requiring a commit as well as
 499  * the MM page stats.
 500  *
 501  * The caller must _not_ hold the cinfo->lock, but must be
 502  * holding the nfs_page lock.
 503  */
 504 void
 505 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 506                             struct nfs_commit_info *cinfo)
 507 {
 508         set_bit(PG_CLEAN, &(req)->wb_flags);
 509         spin_lock(cinfo->lock);
 510         nfs_list_add_request(req, dst);
 511         cinfo->mds->ncommit++;
 512         spin_unlock(cinfo->lock);
 513         if (!cinfo->dreq) {
 514                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 515                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 516                              BDI_RECLAIMABLE);
 517                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 518                                    I_DIRTY_DATASYNC);
 519         }
 520 }
 521 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 522
 523 /**
 524  * nfs_request_remove_commit_list - Remove request from a commit list
 525  * @req: pointer to a nfs_page
 526  * @cinfo: holds list lock and accounting info
 527  *
 528  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 529  * number of outstanding requests requiring a commit
 530  * It does not update the MM page stats.
 531  *
 532  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 533  */
 534 void
 535 nfs_request_remove_commit_list(struct nfs_page *req,
 536                                struct nfs_commit_info *cinfo)
 537 {
 538         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 539                 return;
 540         nfs_list_remove_request(req);
 541         cinfo->mds->ncommit--;
 542 }
 543 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 544
 545 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 546                                       struct inode *inode)
 547 {
 548         cinfo->lock = &inode->i_lock;
 549         cinfo->mds = &NFS_I(inode)->commit_info;
 550         cinfo->ds = pnfs_get_ds_info(inode);
 551         cinfo->dreq = NULL;
 552         cinfo->completion_ops = &nfs_commit_completion_ops;
 553 }
 554
 555 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 556                     struct inode *inode,
 557                     struct nfs_direct_req *dreq)
 558 {
 559         if (dreq)
 560                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 561         else
 562                 nfs_init_cinfo_from_inode(cinfo, inode);
 563 }
 564 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 565
 566 /*
 567  * Add a request to the inode's commit list.
 568  */
 569 void
 570 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 571                         struct nfs_commit_info *cinfo)
 572 {
 573         if (pnfs_mark_request_commit(req, lseg, cinfo))
 574                 return;
 575         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 576 }
 577
 578 static void
 579 nfs_clear_page_commit(struct page *page)
 580 {
 581         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 582         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 583 }
 584
 585 static void
 586 nfs_clear_request_commit(struct nfs_page *req)
 587 {
 588         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 589                 struct inode *inode = req->wb_context->dentry->d_inode;
 590                 struct nfs_commit_info cinfo;
 591
 592                 nfs_init_cinfo_from_inode(&cinfo, inode);
 593                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 594                         spin_lock(cinfo.lock);
 595                         nfs_request_remove_commit_list(req, &cinfo);
 596                         spin_unlock(cinfo.lock);
 597                 }
 598                 nfs_clear_page_commit(req->wb_page);
 599         }
 600 }
 601
 602 static inline
 603 int nfs_write_need_commit(struct nfs_pgio_data *data)
 604 {
 605         if (data->verf.committed == NFS_DATA_SYNC)
 606                 return data->header->lseg == NULL;
 607         return data->verf.committed != NFS_FILE_SYNC;
 608 }
 609
 610 #else
 611 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 612                                       struct inode *inode)
 613 {
 614 }
 615
 616 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 617                     struct inode *inode,
 618                     struct nfs_direct_req *dreq)
 619 {
 620 }
 621
 622 void
 623 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 624                         struct nfs_commit_info *cinfo)
 625 {
 626 }
 627
 628 static void
 629 nfs_clear_request_commit(struct nfs_page *req)
 630 {
 631 }
 632
 633 static inline
 634 int nfs_write_need_commit(struct nfs_pgio_data *data)
 635 {
 636         return 0;
 637 }
 638
 639 #endif
 640
 641 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 642 {
 643         struct nfs_commit_info cinfo;
 644         unsigned long bytes = 0;
 645         bool do_destroy;
 646
 647         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 648                 goto out;
 649         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 650         while (!list_empty(&hdr->pages)) {
 651                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 652
 653                 bytes += req->wb_bytes;
 654                 nfs_list_remove_request(req);
 655                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 656                     (hdr->good_bytes < bytes)) {
 657                         nfs_set_pageerror(req->wb_page);
 658                         nfs_context_set_write_error(req->wb_context, hdr->error);
 659                         goto remove_req;
 660                 }
 661                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 662                         nfs_mark_request_dirty(req);
 663                         goto next;
 664                 }
 665                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 666                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 667                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 668                         goto next;
 669                 }
 670 remove_req:
 671                 nfs_inode_remove_request(req);
 672 next:
 673                 nfs_unlock_request(req);
 674                 nfs_end_page_writeback(req);
 675                 do_destroy = !test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags);
 676                 nfs_release_request(req);
 677         }
 678 out:
 679         hdr->release(hdr);
 680 }
 681
 682 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 683 unsigned long
 684 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 685 {
 686         return cinfo->mds->ncommit;
 687 }
 688
 689 /* cinfo->lock held by caller */
 690 int
 691 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 692                      struct nfs_commit_info *cinfo, int max)
 693 {
 694         struct nfs_page *req, *tmp;
 695         int ret = 0;
 696
 697         list_for_each_entry_safe(req, tmp, src, wb_list) {
 698                 if (!nfs_lock_request(req))
 699                         continue;
 700                 kref_get(&req->wb_kref);
 701                 if (cond_resched_lock(cinfo->lock))
 702                         list_safe_reset_next(req, tmp, wb_list);
 703                 nfs_request_remove_commit_list(req, cinfo);
 704                 nfs_list_add_request(req, dst);
 705                 ret++;
 706                 if ((ret == max) && !cinfo->dreq)
 707                         break;
 708         }
 709         return ret;
 710 }
 711
 712 /*
 713  * nfs_scan_commit - Scan an inode for commit requests
 714  * @inode: NFS inode to scan
 715  * @dst: mds destination list
 716  * @cinfo: mds and ds lists of reqs ready to commit
 717  *
 718  * Moves requests from the inode's 'commit' request list.
 719  * The requests are *not* checked to ensure that they form a contiguous set.
 720  */
 721 int
 722 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 723                 struct nfs_commit_info *cinfo)
 724 {
 725         int ret = 0;
 726
 727         spin_lock(cinfo->lock);
 728         if (cinfo->mds->ncommit > 0) {
 729                 const int max = INT_MAX;
 730
 731                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 732                                            cinfo, max);
 733                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 734         }
 735         spin_unlock(cinfo->lock);
 736         return ret;
 737 }
 738
 739 #else
 740 unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 741 {
 742         return 0;
 743 }
 744
 745 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 746                     struct nfs_commit_info *cinfo)
 747 {
 748         return 0;
 749 }
 750 #endif
 751
 752 /*
 753  * Search for an existing write request, and attempt to update
 754  * it to reflect a new dirty region on a given page.
 755  *
 756  * If the attempt fails, then the existing request is flushed out
 757  * to disk.
 758  */
 759 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 760                 struct page *page,
 761                 unsigned int offset,
 762                 unsigned int bytes)
 763 {
 764         struct nfs_page *req;
 765         unsigned int rqend;
 766         unsigned int end;
 767         int error;
 768
 769         if (!PagePrivate(page))
 770                 return NULL;
 771
 772         end = offset + bytes;
 773         spin_lock(&inode->i_lock);
 774
 775         for (;;) {
 776                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 777                 if (req == NULL)
 778                         goto out_unlock;
 779
 780                 /* should be handled by nfs_flush_incompatible */
 781                 WARN_ON_ONCE(req->wb_head != req);
 782                 WARN_ON_ONCE(req->wb_this_page != req);
 783
 784                 rqend = req->wb_offset + req->wb_bytes;
 785                 /*
 786                  * Tell the caller to flush out the request if
 787                  * the offsets are non-contiguous.
 788                  * Note: nfs_flush_incompatible() will already
 789                  * have flushed out requests having wrong owners.
 790                  */
 791                 if (offset > rqend
 792                     || end < req->wb_offset)
 793                         goto out_flushme;
 794
 795                 if (nfs_lock_request(req))
 796                         break;
 797
 798                 /* The request is locked, so wait and then retry */
 799                 spin_unlock(&inode->i_lock);
 800                 error = nfs_wait_on_request(req);
 801                 nfs_release_request(req);
 802                 if (error != 0)
 803                         goto out_err;
 804                 spin_lock(&inode->i_lock);
 805         }
 806
 807         /* Okay, the request matches. Update the region */
 808         if (offset < req->wb_offset) {
 809                 req->wb_offset = offset;
 810                 req->wb_pgbase = offset;
 811         }
 812         if (end > rqend)
 813                 req->wb_bytes = end - req->wb_offset;
 814         else
 815                 req->wb_bytes = rqend - req->wb_offset;
 816 out_unlock:
 817         spin_unlock(&inode->i_lock);
 818         if (req)
 819                 nfs_clear_request_commit(req);
 820         return req;
 821 out_flushme:
 822         spin_unlock(&inode->i_lock);
 823         nfs_release_request(req);
 824         error = nfs_wb_page(inode, page);
 825 out_err:
 826         return ERR_PTR(error);
 827 }
 828
 829 /*
 830  * Try to update an existing write request, or create one if there is none.
 831  *
 832  * Note: Should always be called with the Page Lock held to prevent races
 833  * if we have to add a new request. Also assumes that the caller has
 834  * already called nfs_flush_incompatible() if necessary.
 835  */
 836 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 837                 struct page *page, unsigned int offset, unsigned int bytes)
 838 {
 839         struct inode *inode = page_file_mapping(page)->host;
 840         struct nfs_page *req;
 841
 842         req = nfs_try_to_update_request(inode, page, offset, bytes);
 843         if (req != NULL)
 844                 goto out;
 845         req = nfs_create_request(ctx, page, NULL, offset, bytes);
 846         if (IS_ERR(req))
 847                 goto out;
 848         nfs_inode_add_request(inode, req);
 849 out:
 850         return req;
 851 }
 852
 853 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 854                 unsigned int offset, unsigned int count)
 855 {
 856         struct nfs_page *req;
 857
 858         req = nfs_setup_write_request(ctx, page, offset, count);
 859         if (IS_ERR(req))
 860                 return PTR_ERR(req);
 861         /* Update file length */
 862         nfs_grow_file(page, offset, count);
 863         nfs_mark_uptodate(req);
 864         nfs_mark_request_dirty(req);
 865         nfs_unlock_and_release_request(req);
 866         return 0;
 867 }
 868
 869 int nfs_flush_incompatible(struct file *file, struct page *page)
 870 {
 871         struct nfs_open_context *ctx = nfs_file_open_context(file);
 872         struct nfs_lock_context *l_ctx;
 873         struct nfs_page *req;
 874         int do_flush, status;
 875         /*
 876          * Look for a request corresponding to this page. If there
 877          * is one, and it belongs to another file, we flush it out
 878          * before we try to copy anything into the page. Do this
 879          * due to the lack of an ACCESS-type call in NFSv2.
 880          * Also do the same if we find a request from an existing
 881          * dropped page.
 882          */
 883         do {
 884                 req = nfs_page_find_request(page);
 885                 if (req == NULL)
 886                         return 0;
 887                 l_ctx = req->wb_lock_context;
 888                 do_flush = req->wb_page != page || req->wb_context != ctx;
 889                 /* for now, flush if more than 1 request in page_group */
 890                 do_flush |= req->wb_this_page != req;
 891                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 892                         do_flush |= l_ctx->lockowner.l_owner != current->files
 893                                 || l_ctx->lockowner.l_pid != current->tgid;
 894                 }
 895                 nfs_release_request(req);
 896                 if (!do_flush)
 897                         return 0;
 898                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 899         } while (status == 0);
 900         return status;
 901 }
 902
 903 /*
 904  * Avoid buffered writes when a open context credential's key would
 905  * expire soon.
 906  *
 907  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 908  *
 909  * Return 0 and set a credential flag which triggers the inode to flush
 910  * and performs  NFS_FILE_SYNC writes if the key will expired within
 911  * RPC_KEY_EXPIRE_TIMEO.
 912  */
 913 int
 914 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 915 {
 916         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 917         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 918
 919         return rpcauth_key_timeout_notify(auth, ctx->cred);
 920 }
 921
 922 /*
 923  * Test if the open context credential key is marked to expire soon.
 924  */
 925 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 926 {
 927         return rpcauth_cred_key_to_expire(ctx->cred);
 928 }
 929
 930 /*
 931  * If the page cache is marked as unsafe or invalid, then we can't rely on
 932  * the PageUptodate() flag. In this case, we will need to turn off
 933  * write optimisations that depend on the page contents being correct.
 934  */
 935 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 936 {
 937         struct nfs_inode *nfsi = NFS_I(inode);
 938
 939         if (nfs_have_delegated_attributes(inode))
 940                 goto out;
 941         if (nfsi->cache_validity & NFS_INO_REVAL_PAGECACHE)
 942                 return false;
 943         smp_rmb();
 944         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 945                 return false;
 946 out:
 947         if (nfsi->cache_validity & NFS_INO_INVALID_DATA)
 948                 return false;
 949         return PageUptodate(page) != 0;
 950 }
 951
 952 /* If we know the page is up to date, and we're not using byte range locks (or
 953  * if we have the whole file locked for writing), it may be more efficient to
 954  * extend the write to cover the entire page in order to avoid fragmentation
 955  * inefficiencies.
 956  *
 957  * If the file is opened for synchronous writes then we can just skip the rest
 958  * of the checks.
 959  */
 960 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 961 {
 962         if (file->f_flags & O_DSYNC)
 963                 return 0;
 964         if (!nfs_write_pageuptodate(page, inode))
 965                 return 0;
 966         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 967                 return 1;
 968         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 969                         inode->i_flock->fl_end == OFFSET_MAX &&
 970                         inode->i_flock->fl_type != F_RDLCK))
 971                 return 1;
 972         return 0;
 973 }
 974
 975 /*
 976  * Update and possibly write a cached page of an NFS file.
 977  *
 978  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 979  * things with a page scheduled for an RPC call (e.g. invalidate it).
 980  */
 981 int nfs_updatepage(struct file *file, struct page *page,
 982                 unsigned int offset, unsigned int count)
 983 {
 984         struct nfs_open_context *ctx = nfs_file_open_context(file);
 985         struct inode    *inode = page_file_mapping(page)->host;
 986         int             status = 0;
 987
 988         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 989
 990         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 991                 file, count, (long long)(page_file_offset(page) + offset));
 992
 993         if (nfs_can_extend_write(file, page, inode)) {
 994                 count = max(count + offset, nfs_page_length(page));
 995                 offset = 0;
 996         }
 997
 998         status = nfs_writepage_setup(ctx, page, offset, count);
 999         if (status < 0)
1000                 nfs_set_pageerror(page);
1001         else
1002                 __set_page_dirty_nobuffers(page);
1003
1004         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1005                         status, (long long)i_size_read(inode));
1006         return status;
1007 }
1008
1009 static int flush_task_priority(int how)
1010 {
1011         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1012                 case FLUSH_HIGHPRI:
1013                         return RPC_PRIORITY_HIGH;
1014                 case FLUSH_LOWPRI:
1015                         return RPC_PRIORITY_LOW;
1016         }
1017         return RPC_PRIORITY_NORMAL;
1018 }
1019
1020 static void nfs_initiate_write(struct nfs_pgio_data *data, struct rpc_message *msg,
1021                                struct rpc_task_setup *task_setup_data, int how)
1022 {
1023         struct inode *inode = data->header->inode;
1024         int priority = flush_task_priority(how);
1025
1026         task_setup_data->priority = priority;
1027         NFS_PROTO(inode)->write_setup(data, msg);
1028
1029         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1030                                  &task_setup_data->rpc_client, msg, data);
1031 }
1032
1033 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1034  * call this on each, which will prepare them to be retried on next
1035  * writeback using standard nfs.
1036  */
1037 static void nfs_redirty_request(struct nfs_page *req)
1038 {
1039         nfs_mark_request_dirty(req);
1040         nfs_unlock_request(req);
1041         nfs_end_page_writeback(req);
1042         nfs_release_request(req);
1043 }
1044
1045 static void nfs_async_write_error(struct list_head *head)
1046 {
1047         struct nfs_page *req;
1048
1049         while (!list_empty(head)) {
1050                 req = nfs_list_entry(head->next);
1051                 nfs_list_remove_request(req);
1052                 nfs_redirty_request(req);
1053         }
1054 }
1055
1056 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1057         .error_cleanup = nfs_async_write_error,
1058         .completion = nfs_write_completion,
1059 };
1060
1061 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1062                                struct inode *inode, int ioflags, bool force_mds,
1063                                const struct nfs_pgio_completion_ops *compl_ops)
1064 {
1065         struct nfs_server *server = NFS_SERVER(inode);
1066         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1067
1068 #ifdef CONFIG_NFS_V4_1
1069         if (server->pnfs_curr_ld && !force_mds)
1070                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1071 #endif
1072         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1073                         server->wsize, ioflags);
1074 }
1075 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1076
1077 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1078 {
1079         pgio->pg_ops = &nfs_pgio_rw_ops;
1080         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1081 }
1082 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1083
1084
1085 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1086 {
1087         struct nfs_commit_data *data = calldata;
1088
1089         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1090 }
1091
1092 static void nfs_writeback_release_common(struct nfs_pgio_data *data)
1093 {
1094         struct nfs_pgio_header *hdr = data->header;
1095         int status = data->task.tk_status;
1096
1097         if ((status >= 0) && nfs_write_need_commit(data)) {
1098                 spin_lock(&hdr->lock);
1099                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1100                         ; /* Do nothing */
1101                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1102                         memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1103                 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1104                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1105                 spin_unlock(&hdr->lock);
1106         }
1107 }
1108
1109 /*
1110  * Special version of should_remove_suid() that ignores capabilities.
1111  */
1112 static int nfs_should_remove_suid(const struct inode *inode)
1113 {
1114         umode_t mode = inode->i_mode;
1115         int kill = 0;
1116
1117         /* suid always must be killed */
1118         if (unlikely(mode & S_ISUID))
1119                 kill = ATTR_KILL_SUID;
1120
1121         /*
1122          * sgid without any exec bits is just a mandatory locking mark; leave
1123          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1124          */
1125         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1126                 kill |= ATTR_KILL_SGID;
1127
1128         if (unlikely(kill && S_ISREG(mode)))
1129                 return kill;
1130
1131         return 0;
1132 }
1133
1134 /*
1135  * This function is called when the WRITE call is complete.
1136  */
1137 static int nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data,
1138                               struct inode *inode)
1139 {
1140         int status;
1141
1142         /*
1143          * ->write_done will attempt to use post-op attributes to detect
1144          * conflicting writes by other clients.  A strict interpretation
1145          * of close-to-open would allow us to continue caching even if
1146          * another writer had changed the file, but some applications
1147          * depend on tighter cache coherency when writing.
1148          */
1149         status = NFS_PROTO(inode)->write_done(task, data);
1150         if (status != 0)
1151                 return status;
1152         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, data->res.count);
1153
1154 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1155         if (data->res.verf->committed < data->args.stable && task->tk_status >= 0) {
1156                 /* We tried a write call, but the server did not
1157                  * commit data to stable storage even though we
1158                  * requested it.
1159                  * Note: There is a known bug in Tru64 < 5.0 in which
1160                  *       the server reports NFS_DATA_SYNC, but performs
1161                  *       NFS_FILE_SYNC. We therefore implement this checking
1162                  *       as a dprintk() in order to avoid filling syslog.
1163                  */
1164                 static unsigned long    complain;
1165
1166                 /* Note this will print the MDS for a DS write */
1167                 if (time_before(complain, jiffies)) {
1168                         dprintk("NFS:       faulty NFS server %s:"
1169                                 " (committed = %d) != (stable = %d)\n",
1170                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1171                                 data->res.verf->committed, data->args.stable);
1172                         complain = jiffies + 300 * HZ;
1173                 }
1174         }
1175 #endif
1176
1177         /* Deal with the suid/sgid bit corner case */
1178         if (nfs_should_remove_suid(inode))
1179                 nfs_mark_for_revalidate(inode);
1180         return 0;
1181 }
1182
1183 /*
1184  * This function is called when the WRITE call is complete.
1185  */
1186 static void nfs_writeback_result(struct rpc_task *task, struct nfs_pgio_data *data)
1187 {
1188         struct nfs_pgio_args    *argp = &data->args;
1189         struct nfs_pgio_res     *resp = &data->res;
1190
1191         if (resp->count < argp->count) {
1192                 static unsigned long    complain;
1193
1194                 /* This a short write! */
1195                 nfs_inc_stats(data->header->inode, NFSIOS_SHORTWRITE);
1196
1197                 /* Has the server at least made some progress? */
1198                 if (resp->count == 0) {
1199                         if (time_before(complain, jiffies)) {
1200                                 printk(KERN_WARNING
1201                                        "NFS: Server wrote zero bytes, expected %u.\n",
1202                                        argp->count);
1203                                 complain = jiffies + 300 * HZ;
1204                         }
1205                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1206                         task->tk_status = -EIO;
1207                         return;
1208                 }
1209                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1210                 if (resp->verf->committed != NFS_UNSTABLE) {
1211                         /* Resend from where the server left off */
1212                         data->mds_offset += resp->count;
1213                         argp->offset += resp->count;
1214                         argp->pgbase += resp->count;
1215                         argp->count -= resp->count;
1216                 } else {
1217                         /* Resend as a stable write in order to avoid
1218                          * headaches in the case of a server crash.
1219                          */
1220                         argp->stable = NFS_FILE_SYNC;
1221                 }
1222                 rpc_restart_call_prepare(task);
1223         }
1224 }
1225
1226
1227 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1228 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1229 {
1230         int ret;
1231
1232         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1233                 return 1;
1234         if (!may_wait)
1235                 return 0;
1236         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1237                                 NFS_INO_COMMIT,
1238                                 nfs_wait_bit_killable,
1239                                 TASK_KILLABLE);
1240         return (ret < 0) ? ret : 1;
1241 }
1242
1243 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1244 {
1245         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1246         smp_mb__after_atomic();
1247         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1248 }
1249
1250 void nfs_commitdata_release(struct nfs_commit_data *data)
1251 {
1252         put_nfs_open_context(data->context);
1253         nfs_commit_free(data);
1254 }
1255 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1256
1257 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1258                         const struct rpc_call_ops *call_ops,
1259                         int how, int flags)
1260 {
1261         struct rpc_task *task;
1262         int priority = flush_task_priority(how);
1263         struct rpc_message msg = {
1264                 .rpc_argp = &data->args,
1265                 .rpc_resp = &data->res,
1266                 .rpc_cred = data->cred,
1267         };
1268         struct rpc_task_setup task_setup_data = {
1269                 .task = &data->task,
1270                 .rpc_client = clnt,
1271                 .rpc_message = &msg,
1272                 .callback_ops = call_ops,
1273                 .callback_data = data,
1274                 .workqueue = nfsiod_workqueue,
1275                 .flags = RPC_TASK_ASYNC | flags,
1276                 .priority = priority,
1277         };
1278         /* Set up the initial task struct.  */
1279         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1280
1281         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1282
1283         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1284                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1285
1286         task = rpc_run_task(&task_setup_data);
1287         if (IS_ERR(task))
1288                 return PTR_ERR(task);
1289         if (how & FLUSH_SYNC)
1290                 rpc_wait_for_completion_task(task);
1291         rpc_put_task(task);
1292         return 0;
1293 }
1294 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1295
1296 /*
1297  * Set up the argument/result storage required for the RPC call.
1298  */
1299 void nfs_init_commit(struct nfs_commit_data *data,
1300                      struct list_head *head,
1301                      struct pnfs_layout_segment *lseg,
1302                      struct nfs_commit_info *cinfo)
1303 {
1304         struct nfs_page *first = nfs_list_entry(head->next);
1305         struct inode *inode = first->wb_context->dentry->d_inode;
1306
1307         /* Set up the RPC argument and reply structs
1308          * NB: take care not to mess about with data->commit et al. */
1309
1310         list_splice_init(head, &data->pages);
1311
1312         data->inode       = inode;
1313         data->cred        = first->wb_context->cred;
1314         data->lseg        = lseg; /* reference transferred */
1315         data->mds_ops     = &nfs_commit_ops;
1316         data->completion_ops = cinfo->completion_ops;
1317         data->dreq        = cinfo->dreq;
1318
1319         data->args.fh     = NFS_FH(data->inode);
1320         /* Note: we always request a commit of the entire inode */
1321         data->args.offset = 0;
1322         data->args.count  = 0;
1323         data->context     = get_nfs_open_context(first->wb_context);
1324         data->res.fattr   = &data->fattr;
1325         data->res.verf    = &data->verf;
1326         nfs_fattr_init(&data->fattr);
1327 }
1328 EXPORT_SYMBOL_GPL(nfs_init_commit);
1329
1330 void nfs_retry_commit(struct list_head *page_list,
1331                       struct pnfs_layout_segment *lseg,
1332                       struct nfs_commit_info *cinfo)
1333 {
1334         struct nfs_page *req;
1335
1336         while (!list_empty(page_list)) {
1337                 req = nfs_list_entry(page_list->next);
1338                 nfs_list_remove_request(req);
1339                 nfs_mark_request_commit(req, lseg, cinfo);
1340                 if (!cinfo->dreq) {
1341                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1342                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1343                                      BDI_RECLAIMABLE);
1344                 }
1345                 nfs_unlock_and_release_request(req);
1346         }
1347 }
1348 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1349
1350 /*
1351  * Commit dirty pages
1352  */
1353 static int
1354 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1355                 struct nfs_commit_info *cinfo)
1356 {
1357         struct nfs_commit_data  *data;
1358
1359         data = nfs_commitdata_alloc();
1360
1361         if (!data)
1362                 goto out_bad;
1363
1364         /* Set up the argument struct */
1365         nfs_init_commit(data, head, NULL, cinfo);
1366         atomic_inc(&cinfo->mds->rpcs_out);
1367         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1368                                    how, 0);
1369  out_bad:
1370         nfs_retry_commit(head, NULL, cinfo);
1371         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1372         return -ENOMEM;
1373 }
1374
1375 /*
1376  * COMMIT call returned
1377  */
1378 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1379 {
1380         struct nfs_commit_data  *data = calldata;
1381
1382         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1383                                 task->tk_pid, task->tk_status);
1384
1385         /* Call the NFS version-specific code */
1386         NFS_PROTO(data->inode)->commit_done(task, data);
1387 }
1388
1389 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1390 {
1391         struct nfs_page *req;
1392         int status = data->task.tk_status;
1393         struct nfs_commit_info cinfo;
1394
1395         while (!list_empty(&data->pages)) {
1396                 req = nfs_list_entry(data->pages.next);
1397                 nfs_list_remove_request(req);
1398                 nfs_clear_page_commit(req->wb_page);
1399
1400                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1401                         req->wb_context->dentry->d_sb->s_id,
1402                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1403                         req->wb_bytes,
1404                         (long long)req_offset(req));
1405                 if (status < 0) {
1406                         nfs_context_set_write_error(req->wb_context, status);
1407                         nfs_inode_remove_request(req);
1408                         dprintk(", error = %d\n", status);
1409                         goto next;
1410                 }
1411
1412                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1413                  * returned by the server against all stored verfs. */
1414                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1415                         /* We have a match */
1416                         nfs_inode_remove_request(req);
1417                         dprintk(" OK\n");
1418                         goto next;
1419                 }
1420                 /* We have a mismatch. Write the page again */
1421                 dprintk(" mismatch\n");
1422                 nfs_mark_request_dirty(req);
1423                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1424         next:
1425                 nfs_unlock_and_release_request(req);
1426         }
1427         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1428         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1429                 nfs_commit_clear_lock(NFS_I(data->inode));
1430 }
1431
1432 static void nfs_commit_release(void *calldata)
1433 {
1434         struct nfs_commit_data *data = calldata;
1435
1436         data->completion_ops->completion(data);
1437         nfs_commitdata_release(calldata);
1438 }
1439
1440 static const struct rpc_call_ops nfs_commit_ops = {
1441         .rpc_call_prepare = nfs_commit_prepare,
1442         .rpc_call_done = nfs_commit_done,
1443         .rpc_release = nfs_commit_release,
1444 };
1445
1446 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1447         .completion = nfs_commit_release_pages,
1448         .error_cleanup = nfs_commit_clear_lock,
1449 };
1450
1451 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1452                             int how, struct nfs_commit_info *cinfo)
1453 {
1454         int status;
1455
1456         status = pnfs_commit_list(inode, head, how, cinfo);
1457         if (status == PNFS_NOT_ATTEMPTED)
1458                 status = nfs_commit_list(inode, head, how, cinfo);
1459         return status;
1460 }
1461
1462 int nfs_commit_inode(struct inode *inode, int how)
1463 {
1464         LIST_HEAD(head);
1465         struct nfs_commit_info cinfo;
1466         int may_wait = how & FLUSH_SYNC;
1467         int res;
1468
1469         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1470         if (res <= 0)
1471                 goto out_mark_dirty;
1472         nfs_init_cinfo_from_inode(&cinfo, inode);
1473         res = nfs_scan_commit(inode, &head, &cinfo);
1474         if (res) {
1475                 int error;
1476
1477                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1478                 if (error < 0)
1479                         return error;
1480                 if (!may_wait)
1481                         goto out_mark_dirty;
1482                 error = wait_on_bit(&NFS_I(inode)->flags,
1483                                 NFS_INO_COMMIT,
1484                                 nfs_wait_bit_killable,
1485                                 TASK_KILLABLE);
1486                 if (error < 0)
1487                         return error;
1488         } else
1489                 nfs_commit_clear_lock(NFS_I(inode));
1490         return res;
1491         /* Note: If we exit without ensuring that the commit is complete,
1492          * we must mark the inode as dirty. Otherwise, future calls to
1493          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1494          * that the data is on the disk.
1495          */
1496 out_mark_dirty:
1497         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1498         return res;
1499 }
1500
1501 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1502 {
1503         struct nfs_inode *nfsi = NFS_I(inode);
1504         int flags = FLUSH_SYNC;
1505         int ret = 0;
1506
1507         /* no commits means nothing needs to be done */
1508         if (!nfsi->commit_info.ncommit)
1509                 return ret;
1510
1511         if (wbc->sync_mode == WB_SYNC_NONE) {
1512                 /* Don't commit yet if this is a non-blocking flush and there
1513                  * are a lot of outstanding writes for this mapping.
1514                  */
1515                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1516                         goto out_mark_dirty;
1517
1518                 /* don't wait for the COMMIT response */
1519                 flags = 0;
1520         }
1521
1522         ret = nfs_commit_inode(inode, flags);
1523         if (ret >= 0) {
1524                 if (wbc->sync_mode == WB_SYNC_NONE) {
1525                         if (ret < wbc->nr_to_write)
1526                                 wbc->nr_to_write -= ret;
1527                         else
1528                                 wbc->nr_to_write = 0;
1529                 }
1530                 return 0;
1531         }
1532 out_mark_dirty:
1533         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1534         return ret;
1535 }
1536 #else
1537 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1538 {
1539         return 0;
1540 }
1541 #endif
1542
1543 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1544 {
1545         return nfs_commit_unstable_pages(inode, wbc);
1546 }
1547 EXPORT_SYMBOL_GPL(nfs_write_inode);
1548
1549 /*
1550  * flush the inode to disk.
1551  */
1552 int nfs_wb_all(struct inode *inode)
1553 {
1554         struct writeback_control wbc = {
1555                 .sync_mode = WB_SYNC_ALL,
1556                 .nr_to_write = LONG_MAX,
1557                 .range_start = 0,
1558                 .range_end = LLONG_MAX,
1559         };
1560         int ret;
1561
1562         trace_nfs_writeback_inode_enter(inode);
1563
1564         ret = sync_inode(inode, &wbc);
1565
1566         trace_nfs_writeback_inode_exit(inode, ret);
1567         return ret;
1568 }
1569 EXPORT_SYMBOL_GPL(nfs_wb_all);
1570
1571 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1572 {
1573         struct nfs_page *req;
1574         int ret = 0;
1575
1576         for (;;) {
1577                 wait_on_page_writeback(page);
1578                 req = nfs_page_find_request(page);
1579                 if (req == NULL)
1580                         break;
1581                 if (nfs_lock_request(req)) {
1582                         nfs_clear_request_commit(req);
1583                         nfs_inode_remove_request(req);
1584                         /*
1585                          * In case nfs_inode_remove_request has marked the
1586                          * page as being dirty
1587                          */
1588                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1589                         nfs_unlock_and_release_request(req);
1590                         break;
1591                 }
1592                 ret = nfs_wait_on_request(req);
1593                 nfs_release_request(req);
1594                 if (ret < 0)
1595                         break;
1596         }
1597         return ret;
1598 }
1599
1600 /*
1601  * Write back all requests on one page - we do this before reading it.
1602  */
1603 int nfs_wb_page(struct inode *inode, struct page *page)
1604 {
1605         loff_t range_start = page_file_offset(page);
1606         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1607         struct writeback_control wbc = {
1608                 .sync_mode = WB_SYNC_ALL,
1609                 .nr_to_write = 0,
1610                 .range_start = range_start,
1611                 .range_end = range_end,
1612         };
1613         int ret;
1614
1615         trace_nfs_writeback_page_enter(inode);
1616
1617         for (;;) {
1618                 wait_on_page_writeback(page);
1619                 if (clear_page_dirty_for_io(page)) {
1620                         ret = nfs_writepage_locked(page, &wbc);
1621                         if (ret < 0)
1622                                 goto out_error;
1623                         continue;
1624                 }
1625                 ret = 0;
1626                 if (!PagePrivate(page))
1627                         break;
1628                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1629                 if (ret < 0)
1630                         goto out_error;
1631         }
1632 out_error:
1633         trace_nfs_writeback_page_exit(inode, ret);
1634         return ret;
1635 }
1636
1637 #ifdef CONFIG_MIGRATION
1638 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1639                 struct page *page, enum migrate_mode mode)
1640 {
1641         /*
1642          * If PagePrivate is set, then the page is currently associated with
1643          * an in-progress read or write request. Don't try to migrate it.
1644          *
1645          * FIXME: we could do this in principle, but we'll need a way to ensure
1646          *        that we can safely release the inode reference while holding
1647          *        the page lock.
1648          */
1649         if (PagePrivate(page))
1650                 return -EBUSY;
1651
1652         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1653                 return -EBUSY;
1654
1655         return migrate_page(mapping, newpage, page, mode);
1656 }
1657 #endif
1658
1659 int __init nfs_init_writepagecache(void)
1660 {
1661         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1662                                              sizeof(struct nfs_rw_header),
1663                                              0, SLAB_HWCACHE_ALIGN,
1664                                              NULL);
1665         if (nfs_wdata_cachep == NULL)
1666                 return -ENOMEM;
1667
1668         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1669                                                      nfs_wdata_cachep);
1670         if (nfs_wdata_mempool == NULL)
1671                 goto out_destroy_write_cache;
1672
1673         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1674                                              sizeof(struct nfs_commit_data),
1675                                              0, SLAB_HWCACHE_ALIGN,
1676                                              NULL);
1677         if (nfs_cdata_cachep == NULL)
1678                 goto out_destroy_write_mempool;
1679
1680         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1681                                                       nfs_cdata_cachep);
1682         if (nfs_commit_mempool == NULL)
1683                 goto out_destroy_commit_cache;
1684
1685         /*
1686          * NFS congestion size, scale with available memory.
1687          *
1688          *  64MB:    8192k
1689          * 128MB:   11585k
1690          * 256MB:   16384k
1691          * 512MB:   23170k
1692          *   1GB:   32768k
1693          *   2GB:   46340k
1694          *   4GB:   65536k
1695          *   8GB:   92681k
1696          *  16GB:  131072k
1697          *
1698          * This allows larger machines to have larger/more transfers.
1699          * Limit the default to 256M
1700          */
1701         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1702         if (nfs_congestion_kb > 256*1024)
1703                 nfs_congestion_kb = 256*1024;
1704
1705         return 0;
1706
1707 out_destroy_commit_cache:
1708         kmem_cache_destroy(nfs_cdata_cachep);
1709 out_destroy_write_mempool:
1710         mempool_destroy(nfs_wdata_mempool);
1711 out_destroy_write_cache:
1712         kmem_cache_destroy(nfs_wdata_cachep);
1713         return -ENOMEM;
1714 }
1715
1716 void nfs_destroy_writepagecache(void)
1717 {
1718         mempool_destroy(nfs_commit_mempool);
1719         kmem_cache_destroy(nfs_cdata_cachep);
1720         mempool_destroy(nfs_wdata_mempool);
1721         kmem_cache_destroy(nfs_wdata_cachep);
1722 }
1723
1724 static const struct nfs_rw_ops nfs_rw_write_ops = {
1725         .rw_mode                = FMODE_WRITE,
1726         .rw_alloc_header        = nfs_writehdr_alloc,
1727         .rw_free_header         = nfs_writehdr_free,
1728         .rw_release             = nfs_writeback_release_common,
1729         .rw_done                = nfs_writeback_done,
1730         .rw_result              = nfs_writeback_result,
1731         .rw_initiate            = nfs_initiate_write,
1732 };