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