fs/nfs/write.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/write.c
   4  *
   5  * Write file data over NFS.
   6  *
   7  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/mm.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/file.h>
  15 #include <linux/writeback.h>
  16 #include <linux/swap.h>
  17 #include <linux/migrate.h>
  18
  19 #include <linux/sunrpc/clnt.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/export.h>
  25 #include <linux/freezer.h>
  26 #include <linux/wait.h>
  27 #include <linux/iversion.h>
  28
  29 #include <linux/uaccess.h>
  30 #include <linux/sched/mm.h>
  31
  32 #include "delegation.h"
  33 #include "internal.h"
  34 #include "iostat.h"
  35 #include "nfs4_fs.h"
  36 #include "fscache.h"
  37 #include "pnfs.h"
  38
  39 #include "nfstrace.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  42
  43 #define MIN_POOL_WRITE          (32)
  44 #define MIN_POOL_COMMIT         (4)
  45
  46 struct nfs_io_completion {
  47         void (*complete)(void *data);
  48         void *data;
  49         struct kref refcount;
  50 };
  51
  52 /*
  53  * Local function declarations
  54  */
  55 static void nfs_redirty_request(struct nfs_page *req);
  56 static const struct rpc_call_ops nfs_commit_ops;
  57 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  58 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  59 static const struct nfs_rw_ops nfs_rw_write_ops;
  60 static void nfs_inode_remove_request(struct nfs_page *req);
  61 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
  62                                      struct nfs_page *req);
  63 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  64                                       struct inode *inode);
  65 static struct nfs_page *
  66 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  67                                                 struct page *page);
  68
  69 static struct kmem_cache *nfs_wdata_cachep;
  70 static mempool_t *nfs_wdata_mempool;
  71 static struct kmem_cache *nfs_cdata_cachep;
  72 static mempool_t *nfs_commit_mempool;
  73
  74 struct nfs_commit_data *nfs_commitdata_alloc(void)
  75 {
  76         struct nfs_commit_data *p;
  77
  78         p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  79         if (!p) {
  80                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  81                 if (!p)
  82                         return NULL;
  83                 memset(p, 0, sizeof(*p));
  84         }
  85         INIT_LIST_HEAD(&p->pages);
  86         return p;
  87 }
  88 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  89
  90 void nfs_commit_free(struct nfs_commit_data *p)
  91 {
  92         mempool_free(p, nfs_commit_mempool);
  93 }
  94 EXPORT_SYMBOL_GPL(nfs_commit_free);
  95
  96 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  97 {
  98         struct nfs_pgio_header *p;
  99
 100         p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
 101         if (!p) {
 102                 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 103                 if (!p)
 104                         return NULL;
 105                 memset(p, 0, sizeof(*p));
 106         }
 107         p->rw_mode = FMODE_WRITE;
 108         return p;
 109 }
 110
 111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 112 {
 113         mempool_free(hdr, nfs_wdata_mempool);
 114 }
 115
 116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 117 {
 118         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 119 }
 120
 121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 122                 void (*complete)(void *), void *data)
 123 {
 124         ioc->complete = complete;
 125         ioc->data = data;
 126         kref_init(&ioc->refcount);
 127 }
 128
 129 static void nfs_io_completion_release(struct kref *kref)
 130 {
 131         struct nfs_io_completion *ioc = container_of(kref,
 132                         struct nfs_io_completion, refcount);
 133         ioc->complete(ioc->data);
 134         kfree(ioc);
 135 }
 136
 137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 138 {
 139         if (ioc != NULL)
 140                 kref_get(&ioc->refcount);
 141 }
 142
 143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 144 {
 145         if (ioc != NULL)
 146                 kref_put(&ioc->refcount, nfs_io_completion_release);
 147 }
 148
 149 static void
 150 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 151 {
 152         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 153                 kref_get(&req->wb_kref);
 154                 atomic_long_inc(&NFS_I(inode)->nrequests);
 155         }
 156 }
 157
 158 static int
 159 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 160 {
 161         int ret;
 162
 163         if (!test_bit(PG_REMOVE, &req->wb_flags))
 164                 return 0;
 165         ret = nfs_page_group_lock(req);
 166         if (ret)
 167                 return ret;
 168         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 169                 nfs_page_set_inode_ref(req, inode);
 170         nfs_page_group_unlock(req);
 171         return 0;
 172 }
 173
 174 static struct nfs_page *
 175 nfs_page_private_request(struct page *page)
 176 {
 177         if (!PagePrivate(page))
 178                 return NULL;
 179         return (struct nfs_page *)page_private(page);
 180 }
 181
 182 /*
 183  * nfs_page_find_head_request_locked - find head request associated with @page
 184  *
 185  * must be called while holding the inode lock.
 186  *
 187  * returns matching head request with reference held, or NULL if not found.
 188  */
 189 static struct nfs_page *
 190 nfs_page_find_private_request(struct page *page)
 191 {
 192         struct address_space *mapping = page_file_mapping(page);
 193         struct nfs_page *req;
 194
 195         if (!PagePrivate(page))
 196                 return NULL;
 197         spin_lock(&mapping->private_lock);
 198         req = nfs_page_private_request(page);
 199         if (req) {
 200                 WARN_ON_ONCE(req->wb_head != req);
 201                 kref_get(&req->wb_kref);
 202         }
 203         spin_unlock(&mapping->private_lock);
 204         return req;
 205 }
 206
 207 static struct nfs_page *
 208 nfs_page_find_swap_request(struct page *page)
 209 {
 210         struct inode *inode = page_file_mapping(page)->host;
 211         struct nfs_inode *nfsi = NFS_I(inode);
 212         struct nfs_page *req = NULL;
 213         if (!PageSwapCache(page))
 214                 return NULL;
 215         mutex_lock(&nfsi->commit_mutex);
 216         if (PageSwapCache(page)) {
 217                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 218                         page);
 219                 if (req) {
 220                         WARN_ON_ONCE(req->wb_head != req);
 221                         kref_get(&req->wb_kref);
 222                 }
 223         }
 224         mutex_unlock(&nfsi->commit_mutex);
 225         return req;
 226 }
 227
 228 /*
 229  * nfs_page_find_head_request - find head request associated with @page
 230  *
 231  * returns matching head request with reference held, or NULL if not found.
 232  */
 233 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 234 {
 235         struct nfs_page *req;
 236
 237         req = nfs_page_find_private_request(page);
 238         if (!req)
 239                 req = nfs_page_find_swap_request(page);
 240         return req;
 241 }
 242
 243 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
 244 {
 245         struct inode *inode = page_file_mapping(page)->host;
 246         struct nfs_page *req, *head;
 247         int ret;
 248
 249         for (;;) {
 250                 req = nfs_page_find_head_request(page);
 251                 if (!req)
 252                         return req;
 253                 head = nfs_page_group_lock_head(req);
 254                 if (head != req)
 255                         nfs_release_request(req);
 256                 if (IS_ERR(head))
 257                         return head;
 258                 ret = nfs_cancel_remove_inode(head, inode);
 259                 if (ret < 0) {
 260                         nfs_unlock_and_release_request(head);
 261                         return ERR_PTR(ret);
 262                 }
 263                 /* Ensure that nobody removed the request before we locked it */
 264                 if (head == nfs_page_private_request(page))
 265                         break;
 266                 if (PageSwapCache(page))
 267                         break;
 268                 nfs_unlock_and_release_request(head);
 269         }
 270         return head;
 271 }
 272
 273 /* Adjust the file length if we're writing beyond the end */
 274 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 275 {
 276         struct inode *inode = page_file_mapping(page)->host;
 277         loff_t end, i_size;
 278         pgoff_t end_index;
 279
 280         spin_lock(&inode->i_lock);
 281         i_size = i_size_read(inode);
 282         end_index = (i_size - 1) >> PAGE_SHIFT;
 283         if (i_size > 0 && page_index(page) < end_index)
 284                 goto out;
 285         end = page_file_offset(page) + ((loff_t)offset+count);
 286         if (i_size >= end)
 287                 goto out;
 288         i_size_write(inode, end);
 289         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 290         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 291 out:
 292         spin_unlock(&inode->i_lock);
 293 }
 294
 295 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 296 static void nfs_set_pageerror(struct address_space *mapping)
 297 {
 298         struct inode *inode = mapping->host;
 299
 300         nfs_zap_mapping(mapping->host, mapping);
 301         /* Force file size revalidation */
 302         spin_lock(&inode->i_lock);
 303         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 304                                              NFS_INO_REVAL_PAGECACHE |
 305                                              NFS_INO_INVALID_SIZE);
 306         spin_unlock(&inode->i_lock);
 307 }
 308
 309 static void nfs_mapping_set_error(struct page *page, int error)
 310 {
 311         struct address_space *mapping = page_file_mapping(page);
 312
 313         SetPageError(page);
 314         filemap_set_wb_err(mapping, error);
 315         if (mapping->host)
 316                 errseq_set(&mapping->host->i_sb->s_wb_err,
 317                            error == -ENOSPC ? -ENOSPC : -EIO);
 318         nfs_set_pageerror(mapping);
 319 }
 320
 321 /*
 322  * nfs_page_group_search_locked
 323  * @head - head request of page group
 324  * @page_offset - offset into page
 325  *
 326  * Search page group with head @head to find a request that contains the
 327  * page offset @page_offset.
 328  *
 329  * Returns a pointer to the first matching nfs request, or NULL if no
 330  * match is found.
 331  *
 332  * Must be called with the page group lock held
 333  */
 334 static struct nfs_page *
 335 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 336 {
 337         struct nfs_page *req;
 338
 339         req = head;
 340         do {
 341                 if (page_offset >= req->wb_pgbase &&
 342                     page_offset < (req->wb_pgbase + req->wb_bytes))
 343                         return req;
 344
 345                 req = req->wb_this_page;
 346         } while (req != head);
 347
 348         return NULL;
 349 }
 350
 351 /*
 352  * nfs_page_group_covers_page
 353  * @head - head request of page group
 354  *
 355  * Return true if the page group with head @head covers the whole page,
 356  * returns false otherwise
 357  */
 358 static bool nfs_page_group_covers_page(struct nfs_page *req)
 359 {
 360         struct nfs_page *tmp;
 361         unsigned int pos = 0;
 362         unsigned int len = nfs_page_length(req->wb_page);
 363
 364         nfs_page_group_lock(req);
 365
 366         for (;;) {
 367                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 368                 if (!tmp)
 369                         break;
 370                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 371         }
 372
 373         nfs_page_group_unlock(req);
 374         return pos >= len;
 375 }
 376
 377 /* We can set the PG_uptodate flag if we see that a write request
 378  * covers the full page.
 379  */
 380 static void nfs_mark_uptodate(struct nfs_page *req)
 381 {
 382         if (PageUptodate(req->wb_page))
 383                 return;
 384         if (!nfs_page_group_covers_page(req))
 385                 return;
 386         SetPageUptodate(req->wb_page);
 387 }
 388
 389 static int wb_priority(struct writeback_control *wbc)
 390 {
 391         int ret = 0;
 392
 393         if (wbc->sync_mode == WB_SYNC_ALL)
 394                 ret = FLUSH_COND_STABLE;
 395         return ret;
 396 }
 397
 398 /*
 399  * NFS congestion control
 400  */
 401
 402 int nfs_congestion_kb;
 403
 404 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 405 #define NFS_CONGESTION_OFF_THRESH       \
 406         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 407
 408 static void nfs_set_page_writeback(struct page *page)
 409 {
 410         struct inode *inode = page_file_mapping(page)->host;
 411         struct nfs_server *nfss = NFS_SERVER(inode);
 412         int ret = test_set_page_writeback(page);
 413
 414         WARN_ON_ONCE(ret != 0);
 415
 416         if (atomic_long_inc_return(&nfss->writeback) >
 417                         NFS_CONGESTION_ON_THRESH)
 418                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 419 }
 420
 421 static void nfs_end_page_writeback(struct nfs_page *req)
 422 {
 423         struct inode *inode = page_file_mapping(req->wb_page)->host;
 424         struct nfs_server *nfss = NFS_SERVER(inode);
 425         bool is_done;
 426
 427         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 428         nfs_unlock_request(req);
 429         if (!is_done)
 430                 return;
 431
 432         end_page_writeback(req->wb_page);
 433         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 434                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 435 }
 436
 437 /*
 438  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 439  *
 440  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 441  * @old_head - the old head of the list
 442  *
 443  * All subrequests must be locked and removed from all lists, so at this point
 444  * they are only "active" in this function, and possibly in nfs_wait_on_request
 445  * with a reference held by some other context.
 446  */
 447 static void
 448 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 449                                  struct nfs_page *old_head,
 450                                  struct inode *inode)
 451 {
 452         while (destroy_list) {
 453                 struct nfs_page *subreq = destroy_list;
 454
 455                 destroy_list = (subreq->wb_this_page == old_head) ?
 456                                    NULL : subreq->wb_this_page;
 457
 458                 /* Note: lock subreq in order to change subreq->wb_head */
 459                 nfs_page_set_headlock(subreq);
 460                 WARN_ON_ONCE(old_head != subreq->wb_head);
 461
 462                 /* make sure old group is not used */
 463                 subreq->wb_this_page = subreq;
 464                 subreq->wb_head = subreq;
 465
 466                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 467
 468                 /* Note: races with nfs_page_group_destroy() */
 469                 if (!kref_read(&subreq->wb_kref)) {
 470                         /* Check if we raced with nfs_page_group_destroy() */
 471                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 472                                 nfs_page_clear_headlock(subreq);
 473                                 nfs_free_request(subreq);
 474                         } else
 475                                 nfs_page_clear_headlock(subreq);
 476                         continue;
 477                 }
 478                 nfs_page_clear_headlock(subreq);
 479
 480                 nfs_release_request(old_head);
 481
 482                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 483                         nfs_release_request(subreq);
 484                         atomic_long_dec(&NFS_I(inode)->nrequests);
 485                 }
 486
 487                 /* subreq is now totally disconnected from page group or any
 488                  * write / commit lists. last chance to wake any waiters */
 489                 nfs_unlock_and_release_request(subreq);
 490         }
 491 }
 492
 493 /*
 494  * nfs_join_page_group - destroy subrequests of the head req
 495  * @head: the page used to lookup the "page group" of nfs_page structures
 496  * @inode: Inode to which the request belongs.
 497  *
 498  * This function joins all sub requests to the head request by first
 499  * locking all requests in the group, cancelling any pending operations
 500  * and finally updating the head request to cover the whole range covered by
 501  * the (former) group.  All subrequests are removed from any write or commit
 502  * lists, unlinked from the group and destroyed.
 503  */
 504 void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
 505                          struct inode *inode)
 506 {
 507         struct nfs_page *subreq;
 508         struct nfs_page *destroy_list = NULL;
 509         unsigned int pgbase, off, bytes;
 510
 511         pgbase = head->wb_pgbase;
 512         bytes = head->wb_bytes;
 513         off = head->wb_offset;
 514         for (subreq = head->wb_this_page; subreq != head;
 515                         subreq = subreq->wb_this_page) {
 516                 /* Subrequests should always form a contiguous range */
 517                 if (pgbase > subreq->wb_pgbase) {
 518                         off -= pgbase - subreq->wb_pgbase;
 519                         bytes += pgbase - subreq->wb_pgbase;
 520                         pgbase = subreq->wb_pgbase;
 521                 }
 522                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 523                                 - pgbase, bytes);
 524         }
 525
 526         /* Set the head request's range to cover the former page group */
 527         head->wb_pgbase = pgbase;
 528         head->wb_bytes = bytes;
 529         head->wb_offset = off;
 530
 531         /* Now that all requests are locked, make sure they aren't on any list.
 532          * Commit list removal accounting is done after locks are dropped */
 533         subreq = head;
 534         do {
 535                 nfs_clear_request_commit(cinfo, subreq);
 536                 subreq = subreq->wb_this_page;
 537         } while (subreq != head);
 538
 539         /* unlink subrequests from head, destroy them later */
 540         if (head->wb_this_page != head) {
 541                 /* destroy list will be terminated by head */
 542                 destroy_list = head->wb_this_page;
 543                 head->wb_this_page = head;
 544         }
 545
 546         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 547 }
 548
 549 /*
 550  * nfs_lock_and_join_requests - join all subreqs to the head req
 551  * @page: the page used to lookup the "page group" of nfs_page structures
 552  *
 553  * This function joins all sub requests to the head request by first
 554  * locking all requests in the group, cancelling any pending operations
 555  * and finally updating the head request to cover the whole range covered by
 556  * the (former) group.  All subrequests are removed from any write or commit
 557  * lists, unlinked from the group and destroyed.
 558  *
 559  * Returns a locked, referenced pointer to the head request - which after
 560  * this call is guaranteed to be the only request associated with the page.
 561  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 562  * error was encountered.
 563  */
 564 static struct nfs_page *
 565 nfs_lock_and_join_requests(struct page *page)
 566 {
 567         struct inode *inode = page_file_mapping(page)->host;
 568         struct nfs_page *head;
 569         struct nfs_commit_info cinfo;
 570         int ret;
 571
 572         nfs_init_cinfo_from_inode(&cinfo, inode);
 573         /*
 574          * A reference is taken only on the head request which acts as a
 575          * reference to the whole page group - the group will not be destroyed
 576          * until the head reference is released.
 577          */
 578         head = nfs_find_and_lock_page_request(page);
 579         if (IS_ERR_OR_NULL(head))
 580                 return head;
 581
 582         /* lock each request in the page group */
 583         ret = nfs_page_group_lock_subrequests(head);
 584         if (ret < 0) {
 585                 nfs_unlock_and_release_request(head);
 586                 return ERR_PTR(ret);
 587         }
 588
 589         nfs_join_page_group(head, &cinfo, inode);
 590
 591         return head;
 592 }
 593
 594 static void nfs_write_error(struct nfs_page *req, int error)
 595 {
 596         trace_nfs_write_error(req, error);
 597         nfs_mapping_set_error(req->wb_page, error);
 598         nfs_inode_remove_request(req);
 599         nfs_end_page_writeback(req);
 600         nfs_release_request(req);
 601 }
 602
 603 /*
 604  * Find an associated nfs write request, and prepare to flush it out
 605  * May return an error if the user signalled nfs_wait_on_request().
 606  */
 607 static int nfs_page_async_flush(struct page *page,
 608                                 struct writeback_control *wbc,
 609                                 struct nfs_pageio_descriptor *pgio)
 610 {
 611         struct nfs_page *req;
 612         int ret = 0;
 613
 614         req = nfs_lock_and_join_requests(page);
 615         if (!req)
 616                 goto out;
 617         ret = PTR_ERR(req);
 618         if (IS_ERR(req))
 619                 goto out;
 620
 621         nfs_set_page_writeback(page);
 622         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 623
 624         /* If there is a fatal error that covers this write, just exit */
 625         ret = pgio->pg_error;
 626         if (nfs_error_is_fatal_on_server(ret))
 627                 goto out_launder;
 628
 629         ret = 0;
 630         if (!nfs_pageio_add_request(pgio, req)) {
 631                 ret = pgio->pg_error;
 632                 /*
 633                  * Remove the problematic req upon fatal errors on the server
 634                  */
 635                 if (nfs_error_is_fatal_on_server(ret))
 636                         goto out_launder;
 637                 if (wbc->sync_mode == WB_SYNC_NONE)
 638                         ret = AOP_WRITEPAGE_ACTIVATE;
 639                 redirty_page_for_writepage(wbc, page);
 640                 nfs_redirty_request(req);
 641                 pgio->pg_error = 0;
 642         } else
 643                 nfs_add_stats(page_file_mapping(page)->host,
 644                                 NFSIOS_WRITEPAGES, 1);
 645 out:
 646         return ret;
 647 out_launder:
 648         nfs_write_error(req, ret);
 649         return 0;
 650 }
 651
 652 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 653                             struct nfs_pageio_descriptor *pgio)
 654 {
 655         nfs_pageio_cond_complete(pgio, page_index(page));
 656         return nfs_page_async_flush(page, wbc, pgio);
 657 }
 658
 659 /*
 660  * Write an mmapped page to the server.
 661  */
 662 static int nfs_writepage_locked(struct page *page,
 663                                 struct writeback_control *wbc)
 664 {
 665         struct nfs_pageio_descriptor pgio;
 666         struct inode *inode = page_file_mapping(page)->host;
 667         int err;
 668
 669         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 670         nfs_pageio_init_write(&pgio, inode, 0,
 671                                 false, &nfs_async_write_completion_ops);
 672         err = nfs_do_writepage(page, wbc, &pgio);
 673         pgio.pg_error = 0;
 674         nfs_pageio_complete(&pgio);
 675         return err;
 676 }
 677
 678 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 679 {
 680         int ret;
 681
 682         ret = nfs_writepage_locked(page, wbc);
 683         if (ret != AOP_WRITEPAGE_ACTIVATE)
 684                 unlock_page(page);
 685         return ret;
 686 }
 687
 688 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 689 {
 690         int ret;
 691
 692         ret = nfs_do_writepage(page, wbc, data);
 693         if (ret != AOP_WRITEPAGE_ACTIVATE)
 694                 unlock_page(page);
 695         return ret;
 696 }
 697
 698 static void nfs_io_completion_commit(void *inode)
 699 {
 700         nfs_commit_inode(inode, 0);
 701 }
 702
 703 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 704 {
 705         struct inode *inode = mapping->host;
 706         struct nfs_pageio_descriptor pgio;
 707         struct nfs_io_completion *ioc = NULL;
 708         unsigned int mntflags = NFS_SERVER(inode)->flags;
 709         int priority = 0;
 710         int err;
 711
 712         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 713
 714         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 715             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 716                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 717                 if (ioc)
 718                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 719                                                inode);
 720                 priority = wb_priority(wbc);
 721         }
 722
 723         do {
 724                 nfs_pageio_init_write(&pgio, inode, priority, false,
 725                                       &nfs_async_write_completion_ops);
 726                 pgio.pg_io_completion = ioc;
 727                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 728                                         &pgio);
 729                 pgio.pg_error = 0;
 730                 nfs_pageio_complete(&pgio);
 731         } while (err < 0 && !nfs_error_is_fatal(err));
 732         nfs_io_completion_put(ioc);
 733
 734         if (err < 0)
 735                 goto out_err;
 736         return 0;
 737 out_err:
 738         return err;
 739 }
 740
 741 /*
 742  * Insert a write request into an inode
 743  */
 744 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 745 {
 746         struct address_space *mapping = page_file_mapping(req->wb_page);
 747         struct nfs_inode *nfsi = NFS_I(inode);
 748
 749         WARN_ON_ONCE(req->wb_this_page != req);
 750
 751         /* Lock the request! */
 752         nfs_lock_request(req);
 753
 754         /*
 755          * Swap-space should not get truncated. Hence no need to plug the race
 756          * with invalidate/truncate.
 757          */
 758         spin_lock(&mapping->private_lock);
 759         if (likely(!PageSwapCache(req->wb_page))) {
 760                 set_bit(PG_MAPPED, &req->wb_flags);
 761                 SetPagePrivate(req->wb_page);
 762                 set_page_private(req->wb_page, (unsigned long)req);
 763         }
 764         spin_unlock(&mapping->private_lock);
 765         atomic_long_inc(&nfsi->nrequests);
 766         /* this a head request for a page group - mark it as having an
 767          * extra reference so sub groups can follow suit.
 768          * This flag also informs pgio layer when to bump nrequests when
 769          * adding subrequests. */
 770         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 771         kref_get(&req->wb_kref);
 772 }
 773
 774 /*
 775  * Remove a write request from an inode
 776  */
 777 static void nfs_inode_remove_request(struct nfs_page *req)
 778 {
 779         struct address_space *mapping = page_file_mapping(req->wb_page);
 780         struct inode *inode = mapping->host;
 781         struct nfs_inode *nfsi = NFS_I(inode);
 782         struct nfs_page *head;
 783
 784         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 785                 head = req->wb_head;
 786
 787                 spin_lock(&mapping->private_lock);
 788                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 789                         set_page_private(head->wb_page, 0);
 790                         ClearPagePrivate(head->wb_page);
 791                         clear_bit(PG_MAPPED, &head->wb_flags);
 792                 }
 793                 spin_unlock(&mapping->private_lock);
 794         }
 795
 796         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 797                 nfs_release_request(req);
 798                 atomic_long_dec(&nfsi->nrequests);
 799         }
 800 }
 801
 802 static void
 803 nfs_mark_request_dirty(struct nfs_page *req)
 804 {
 805         if (req->wb_page)
 806                 __set_page_dirty_nobuffers(req->wb_page);
 807 }
 808
 809 /*
 810  * nfs_page_search_commits_for_head_request_locked
 811  *
 812  * Search through commit lists on @inode for the head request for @page.
 813  * Must be called while holding the inode (which is cinfo) lock.
 814  *
 815  * Returns the head request if found, or NULL if not found.
 816  */
 817 static struct nfs_page *
 818 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 819                                                 struct page *page)
 820 {
 821         struct nfs_page *freq, *t;
 822         struct nfs_commit_info cinfo;
 823         struct inode *inode = &nfsi->vfs_inode;
 824
 825         nfs_init_cinfo_from_inode(&cinfo, inode);
 826
 827         /* search through pnfs commit lists */
 828         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 829         if (freq)
 830                 return freq->wb_head;
 831
 832         /* Linearly search the commit list for the correct request */
 833         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 834                 if (freq->wb_page == page)
 835                         return freq->wb_head;
 836         }
 837
 838         return NULL;
 839 }
 840
 841 /**
 842  * nfs_request_add_commit_list_locked - add request to a commit list
 843  * @req: pointer to a struct nfs_page
 844  * @dst: commit list head
 845  * @cinfo: holds list lock and accounting info
 846  *
 847  * This sets the PG_CLEAN bit, updates the cinfo count of
 848  * number of outstanding requests requiring a commit as well as
 849  * the MM page stats.
 850  *
 851  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 852  * nfs_page lock.
 853  */
 854 void
 855 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 856                             struct nfs_commit_info *cinfo)
 857 {
 858         set_bit(PG_CLEAN, &req->wb_flags);
 859         nfs_list_add_request(req, dst);
 860         atomic_long_inc(&cinfo->mds->ncommit);
 861 }
 862 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 863
 864 /**
 865  * nfs_request_add_commit_list - add request to a commit list
 866  * @req: pointer to a struct nfs_page
 867  * @cinfo: holds list lock and accounting info
 868  *
 869  * This sets the PG_CLEAN bit, updates the cinfo count of
 870  * number of outstanding requests requiring a commit as well as
 871  * the MM page stats.
 872  *
 873  * The caller must _not_ hold the cinfo->lock, but must be
 874  * holding the nfs_page lock.
 875  */
 876 void
 877 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 878 {
 879         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 880         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 881         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 882         if (req->wb_page)
 883                 nfs_mark_page_unstable(req->wb_page, cinfo);
 884 }
 885 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 886
 887 /**
 888  * nfs_request_remove_commit_list - Remove request from a commit list
 889  * @req: pointer to a nfs_page
 890  * @cinfo: holds list lock and accounting info
 891  *
 892  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 893  * number of outstanding requests requiring a commit
 894  * It does not update the MM page stats.
 895  *
 896  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 897  */
 898 void
 899 nfs_request_remove_commit_list(struct nfs_page *req,
 900                                struct nfs_commit_info *cinfo)
 901 {
 902         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 903                 return;
 904         nfs_list_remove_request(req);
 905         atomic_long_dec(&cinfo->mds->ncommit);
 906 }
 907 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 908
 909 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 910                                       struct inode *inode)
 911 {
 912         cinfo->inode = inode;
 913         cinfo->mds = &NFS_I(inode)->commit_info;
 914         cinfo->ds = pnfs_get_ds_info(inode);
 915         cinfo->dreq = NULL;
 916         cinfo->completion_ops = &nfs_commit_completion_ops;
 917 }
 918
 919 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 920                     struct inode *inode,
 921                     struct nfs_direct_req *dreq)
 922 {
 923         if (dreq)
 924                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 925         else
 926                 nfs_init_cinfo_from_inode(cinfo, inode);
 927 }
 928 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 929
 930 /*
 931  * Add a request to the inode's commit list.
 932  */
 933 void
 934 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 935                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 936 {
 937         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 938                 return;
 939         nfs_request_add_commit_list(req, cinfo);
 940 }
 941
 942 static void
 943 nfs_clear_page_commit(struct page *page)
 944 {
 945         dec_node_page_state(page, NR_WRITEBACK);
 946         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 947                     WB_WRITEBACK);
 948 }
 949
 950 /* Called holding the request lock on @req */
 951 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 952                                      struct nfs_page *req)
 953 {
 954         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 955                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 956                 struct inode *inode = d_inode(ctx->dentry);
 957
 958                 mutex_lock(&NFS_I(inode)->commit_mutex);
 959                 if (!pnfs_clear_request_commit(req, cinfo)) {
 960                         nfs_request_remove_commit_list(req, cinfo);
 961                 }
 962                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 963                 nfs_clear_page_commit(req->wb_page);
 964         }
 965 }
 966
 967 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 968 {
 969         if (hdr->verf.committed == NFS_DATA_SYNC)
 970                 return hdr->lseg == NULL;
 971         return hdr->verf.committed != NFS_FILE_SYNC;
 972 }
 973
 974 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 975 {
 976         nfs_io_completion_get(hdr->io_completion);
 977 }
 978
 979 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 980 {
 981         struct nfs_commit_info cinfo;
 982         unsigned long bytes = 0;
 983
 984         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 985                 goto out;
 986         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 987         while (!list_empty(&hdr->pages)) {
 988                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 989
 990                 bytes += req->wb_bytes;
 991                 nfs_list_remove_request(req);
 992                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 993                     (hdr->good_bytes < bytes)) {
 994                         trace_nfs_comp_error(req, hdr->error);
 995                         nfs_mapping_set_error(req->wb_page, hdr->error);
 996                         goto remove_req;
 997                 }
 998                 if (nfs_write_need_commit(hdr)) {
 999                         /* Reset wb_nio, since the write was successful. */
1000                         req->wb_nio = 0;
1001                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1002                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1003                                 hdr->pgio_mirror_idx);
1004                         goto next;
1005                 }
1006 remove_req:
1007                 nfs_inode_remove_request(req);
1008 next:
1009                 nfs_end_page_writeback(req);
1010                 nfs_release_request(req);
1011         }
1012 out:
1013         nfs_io_completion_put(hdr->io_completion);
1014         hdr->release(hdr);
1015 }
1016
1017 unsigned long
1018 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1019 {
1020         return atomic_long_read(&cinfo->mds->ncommit);
1021 }
1022
1023 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1024 int
1025 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1026                      struct nfs_commit_info *cinfo, int max)
1027 {
1028         struct nfs_page *req, *tmp;
1029         int ret = 0;
1030
1031         list_for_each_entry_safe(req, tmp, src, wb_list) {
1032                 kref_get(&req->wb_kref);
1033                 if (!nfs_lock_request(req)) {
1034                         nfs_release_request(req);
1035                         continue;
1036                 }
1037                 nfs_request_remove_commit_list(req, cinfo);
1038                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1039                 nfs_list_add_request(req, dst);
1040                 ret++;
1041                 if ((ret == max) && !cinfo->dreq)
1042                         break;
1043                 cond_resched();
1044         }
1045         return ret;
1046 }
1047 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1048
1049 /*
1050  * nfs_scan_commit - Scan an inode for commit requests
1051  * @inode: NFS inode to scan
1052  * @dst: mds destination list
1053  * @cinfo: mds and ds lists of reqs ready to commit
1054  *
1055  * Moves requests from the inode's 'commit' request list.
1056  * The requests are *not* checked to ensure that they form a contiguous set.
1057  */
1058 int
1059 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1060                 struct nfs_commit_info *cinfo)
1061 {
1062         int ret = 0;
1063
1064         if (!atomic_long_read(&cinfo->mds->ncommit))
1065                 return 0;
1066         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1067         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1068                 const int max = INT_MAX;
1069
1070                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1071                                            cinfo, max);
1072                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1073         }
1074         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1075         return ret;
1076 }
1077
1078 /*
1079  * Search for an existing write request, and attempt to update
1080  * it to reflect a new dirty region on a given page.
1081  *
1082  * If the attempt fails, then the existing request is flushed out
1083  * to disk.
1084  */
1085 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1086                 struct page *page,
1087                 unsigned int offset,
1088                 unsigned int bytes)
1089 {
1090         struct nfs_page *req;
1091         unsigned int rqend;
1092         unsigned int end;
1093         int error;
1094
1095         end = offset + bytes;
1096
1097         req = nfs_lock_and_join_requests(page);
1098         if (IS_ERR_OR_NULL(req))
1099                 return req;
1100
1101         rqend = req->wb_offset + req->wb_bytes;
1102         /*
1103          * Tell the caller to flush out the request if
1104          * the offsets are non-contiguous.
1105          * Note: nfs_flush_incompatible() will already
1106          * have flushed out requests having wrong owners.
1107          */
1108         if (offset > rqend || end < req->wb_offset)
1109                 goto out_flushme;
1110
1111         /* Okay, the request matches. Update the region */
1112         if (offset < req->wb_offset) {
1113                 req->wb_offset = offset;
1114                 req->wb_pgbase = offset;
1115         }
1116         if (end > rqend)
1117                 req->wb_bytes = end - req->wb_offset;
1118         else
1119                 req->wb_bytes = rqend - req->wb_offset;
1120         req->wb_nio = 0;
1121         return req;
1122 out_flushme:
1123         /*
1124          * Note: we mark the request dirty here because
1125          * nfs_lock_and_join_requests() cannot preserve
1126          * commit flags, so we have to replay the write.
1127          */
1128         nfs_mark_request_dirty(req);
1129         nfs_unlock_and_release_request(req);
1130         error = nfs_wb_page(inode, page);
1131         return (error < 0) ? ERR_PTR(error) : NULL;
1132 }
1133
1134 /*
1135  * Try to update an existing write request, or create one if there is none.
1136  *
1137  * Note: Should always be called with the Page Lock held to prevent races
1138  * if we have to add a new request. Also assumes that the caller has
1139  * already called nfs_flush_incompatible() if necessary.
1140  */
1141 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1142                 struct page *page, unsigned int offset, unsigned int bytes)
1143 {
1144         struct inode *inode = page_file_mapping(page)->host;
1145         struct nfs_page *req;
1146
1147         req = nfs_try_to_update_request(inode, page, offset, bytes);
1148         if (req != NULL)
1149                 goto out;
1150         req = nfs_create_request(ctx, page, offset, bytes);
1151         if (IS_ERR(req))
1152                 goto out;
1153         nfs_inode_add_request(inode, req);
1154 out:
1155         return req;
1156 }
1157
1158 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1159                 unsigned int offset, unsigned int count)
1160 {
1161         struct nfs_page *req;
1162
1163         req = nfs_setup_write_request(ctx, page, offset, count);
1164         if (IS_ERR(req))
1165                 return PTR_ERR(req);
1166         /* Update file length */
1167         nfs_grow_file(page, offset, count);
1168         nfs_mark_uptodate(req);
1169         nfs_mark_request_dirty(req);
1170         nfs_unlock_and_release_request(req);
1171         return 0;
1172 }
1173
1174 int nfs_flush_incompatible(struct file *file, struct page *page)
1175 {
1176         struct nfs_open_context *ctx = nfs_file_open_context(file);
1177         struct nfs_lock_context *l_ctx;
1178         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1179         struct nfs_page *req;
1180         int do_flush, status;
1181         /*
1182          * Look for a request corresponding to this page. If there
1183          * is one, and it belongs to another file, we flush it out
1184          * before we try to copy anything into the page. Do this
1185          * due to the lack of an ACCESS-type call in NFSv2.
1186          * Also do the same if we find a request from an existing
1187          * dropped page.
1188          */
1189         do {
1190                 req = nfs_page_find_head_request(page);
1191                 if (req == NULL)
1192                         return 0;
1193                 l_ctx = req->wb_lock_context;
1194                 do_flush = req->wb_page != page ||
1195                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1196                 if (l_ctx && flctx &&
1197                     !(list_empty_careful(&flctx->flc_posix) &&
1198                       list_empty_careful(&flctx->flc_flock))) {
1199                         do_flush |= l_ctx->lockowner != current->files;
1200                 }
1201                 nfs_release_request(req);
1202                 if (!do_flush)
1203                         return 0;
1204                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1205         } while (status == 0);
1206         return status;
1207 }
1208
1209 /*
1210  * Avoid buffered writes when a open context credential's key would
1211  * expire soon.
1212  *
1213  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1214  *
1215  * Return 0 and set a credential flag which triggers the inode to flush
1216  * and performs  NFS_FILE_SYNC writes if the key will expired within
1217  * RPC_KEY_EXPIRE_TIMEO.
1218  */
1219 int
1220 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1221 {
1222         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1223
1224         if (nfs_ctx_key_to_expire(ctx, inode) &&
1225             !ctx->ll_cred)
1226                 /* Already expired! */
1227                 return -EACCES;
1228         return 0;
1229 }
1230
1231 /*
1232  * Test if the open context credential key is marked to expire soon.
1233  */
1234 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1235 {
1236         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1237         struct rpc_cred *cred = ctx->ll_cred;
1238         struct auth_cred acred = {
1239                 .cred = ctx->cred,
1240         };
1241
1242         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1243                 put_rpccred(cred);
1244                 ctx->ll_cred = NULL;
1245                 cred = NULL;
1246         }
1247         if (!cred)
1248                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1249         if (!cred || IS_ERR(cred))
1250                 return true;
1251         ctx->ll_cred = cred;
1252         return !!(cred->cr_ops->crkey_timeout &&
1253                   cred->cr_ops->crkey_timeout(cred));
1254 }
1255
1256 /*
1257  * If the page cache is marked as unsafe or invalid, then we can't rely on
1258  * the PageUptodate() flag. In this case, we will need to turn off
1259  * write optimisations that depend on the page contents being correct.
1260  */
1261 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1262                                    unsigned int pagelen)
1263 {
1264         struct nfs_inode *nfsi = NFS_I(inode);
1265
1266         if (nfs_have_delegated_attributes(inode))
1267                 goto out;
1268         if (nfsi->cache_validity &
1269             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1270                 return false;
1271         smp_rmb();
1272         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1273                 return false;
1274 out:
1275         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1276                 return false;
1277         return PageUptodate(page) != 0;
1278 }
1279
1280 static bool
1281 is_whole_file_wrlock(struct file_lock *fl)
1282 {
1283         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1284                         fl->fl_type == F_WRLCK;
1285 }
1286
1287 /* If we know the page is up to date, and we're not using byte range locks (or
1288  * if we have the whole file locked for writing), it may be more efficient to
1289  * extend the write to cover the entire page in order to avoid fragmentation
1290  * inefficiencies.
1291  *
1292  * If the file is opened for synchronous writes then we can just skip the rest
1293  * of the checks.
1294  */
1295 static int nfs_can_extend_write(struct file *file, struct page *page,
1296                                 struct inode *inode, unsigned int pagelen)
1297 {
1298         int ret;
1299         struct file_lock_context *flctx = inode->i_flctx;
1300         struct file_lock *fl;
1301
1302         if (file->f_flags & O_DSYNC)
1303                 return 0;
1304         if (!nfs_write_pageuptodate(page, inode, pagelen))
1305                 return 0;
1306         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1307                 return 1;
1308         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1309                        list_empty_careful(&flctx->flc_posix)))
1310                 return 1;
1311
1312         /* Check to see if there are whole file write locks */
1313         ret = 0;
1314         spin_lock(&flctx->flc_lock);
1315         if (!list_empty(&flctx->flc_posix)) {
1316                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1317                                         fl_list);
1318                 if (is_whole_file_wrlock(fl))
1319                         ret = 1;
1320         } else if (!list_empty(&flctx->flc_flock)) {
1321                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1322                                         fl_list);
1323                 if (fl->fl_type == F_WRLCK)
1324                         ret = 1;
1325         }
1326         spin_unlock(&flctx->flc_lock);
1327         return ret;
1328 }
1329
1330 /*
1331  * Update and possibly write a cached page of an NFS file.
1332  *
1333  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1334  * things with a page scheduled for an RPC call (e.g. invalidate it).
1335  */
1336 int nfs_updatepage(struct file *file, struct page *page,
1337                 unsigned int offset, unsigned int count)
1338 {
1339         struct nfs_open_context *ctx = nfs_file_open_context(file);
1340         struct address_space *mapping = page_file_mapping(page);
1341         struct inode    *inode = mapping->host;
1342         unsigned int    pagelen = nfs_page_length(page);
1343         int             status = 0;
1344
1345         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1346
1347         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1348                 file, count, (long long)(page_file_offset(page) + offset));
1349
1350         if (!count)
1351                 goto out;
1352
1353         if (nfs_can_extend_write(file, page, inode, pagelen)) {
1354                 count = max(count + offset, pagelen);
1355                 offset = 0;
1356         }
1357
1358         status = nfs_writepage_setup(ctx, page, offset, count);
1359         if (status < 0)
1360                 nfs_set_pageerror(mapping);
1361         else
1362                 __set_page_dirty_nobuffers(page);
1363 out:
1364         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1365                         status, (long long)i_size_read(inode));
1366         return status;
1367 }
1368
1369 static int flush_task_priority(int how)
1370 {
1371         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1372                 case FLUSH_HIGHPRI:
1373                         return RPC_PRIORITY_HIGH;
1374                 case FLUSH_LOWPRI:
1375                         return RPC_PRIORITY_LOW;
1376         }
1377         return RPC_PRIORITY_NORMAL;
1378 }
1379
1380 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1381                                struct rpc_message *msg,
1382                                const struct nfs_rpc_ops *rpc_ops,
1383                                struct rpc_task_setup *task_setup_data, int how)
1384 {
1385         int priority = flush_task_priority(how);
1386
1387         task_setup_data->priority = priority;
1388         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1389         trace_nfs_initiate_write(hdr);
1390 }
1391
1392 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1393  * call this on each, which will prepare them to be retried on next
1394  * writeback using standard nfs.
1395  */
1396 static void nfs_redirty_request(struct nfs_page *req)
1397 {
1398         struct nfs_inode *nfsi = NFS_I(page_file_mapping(req->wb_page)->host);
1399
1400         /* Bump the transmission count */
1401         req->wb_nio++;
1402         nfs_mark_request_dirty(req);
1403         atomic_long_inc(&nfsi->redirtied_pages);
1404         nfs_end_page_writeback(req);
1405         nfs_release_request(req);
1406 }
1407
1408 static void nfs_async_write_error(struct list_head *head, int error)
1409 {
1410         struct nfs_page *req;
1411
1412         while (!list_empty(head)) {
1413                 req = nfs_list_entry(head->next);
1414                 nfs_list_remove_request(req);
1415                 if (nfs_error_is_fatal_on_server(error))
1416                         nfs_write_error(req, error);
1417                 else
1418                         nfs_redirty_request(req);
1419         }
1420 }
1421
1422 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1423 {
1424         nfs_async_write_error(&hdr->pages, 0);
1425         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1426                         hdr->args.offset + hdr->args.count - 1);
1427 }
1428
1429 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1430         .init_hdr = nfs_async_write_init,
1431         .error_cleanup = nfs_async_write_error,
1432         .completion = nfs_write_completion,
1433         .reschedule_io = nfs_async_write_reschedule_io,
1434 };
1435
1436 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1437                                struct inode *inode, int ioflags, bool force_mds,
1438                                const struct nfs_pgio_completion_ops *compl_ops)
1439 {
1440         struct nfs_server *server = NFS_SERVER(inode);
1441         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1442
1443 #ifdef CONFIG_NFS_V4_1
1444         if (server->pnfs_curr_ld && !force_mds)
1445                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1446 #endif
1447         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1448                         server->wsize, ioflags);
1449 }
1450 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1451
1452 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1453 {
1454         struct nfs_pgio_mirror *mirror;
1455
1456         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1457                 pgio->pg_ops->pg_cleanup(pgio);
1458
1459         pgio->pg_ops = &nfs_pgio_rw_ops;
1460
1461         nfs_pageio_stop_mirroring(pgio);
1462
1463         mirror = &pgio->pg_mirrors[0];
1464         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1465 }
1466 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1467
1468
1469 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1470 {
1471         struct nfs_commit_data *data = calldata;
1472
1473         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1474 }
1475
1476 /*
1477  * Special version of should_remove_suid() that ignores capabilities.
1478  */
1479 static int nfs_should_remove_suid(const struct inode *inode)
1480 {
1481         umode_t mode = inode->i_mode;
1482         int kill = 0;
1483
1484         /* suid always must be killed */
1485         if (unlikely(mode & S_ISUID))
1486                 kill = ATTR_KILL_SUID;
1487
1488         /*
1489          * sgid without any exec bits is just a mandatory locking mark; leave
1490          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1491          */
1492         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1493                 kill |= ATTR_KILL_SGID;
1494
1495         if (unlikely(kill && S_ISREG(mode)))
1496                 return kill;
1497
1498         return 0;
1499 }
1500
1501 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1502                 struct nfs_fattr *fattr)
1503 {
1504         struct nfs_pgio_args *argp = &hdr->args;
1505         struct nfs_pgio_res *resp = &hdr->res;
1506         u64 size = argp->offset + resp->count;
1507
1508         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1509                 fattr->size = size;
1510         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1511                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1512                 return;
1513         }
1514         if (size != fattr->size)
1515                 return;
1516         /* Set attribute barrier */
1517         nfs_fattr_set_barrier(fattr);
1518         /* ...and update size */
1519         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1520 }
1521
1522 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1523 {
1524         struct nfs_fattr *fattr = &hdr->fattr;
1525         struct inode *inode = hdr->inode;
1526
1527         spin_lock(&inode->i_lock);
1528         nfs_writeback_check_extend(hdr, fattr);
1529         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1530         spin_unlock(&inode->i_lock);
1531 }
1532 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1533
1534 /*
1535  * This function is called when the WRITE call is complete.
1536  */
1537 static int nfs_writeback_done(struct rpc_task *task,
1538                               struct nfs_pgio_header *hdr,
1539                               struct inode *inode)
1540 {
1541         int status;
1542
1543         /*
1544          * ->write_done will attempt to use post-op attributes to detect
1545          * conflicting writes by other clients.  A strict interpretation
1546          * of close-to-open would allow us to continue caching even if
1547          * another writer had changed the file, but some applications
1548          * depend on tighter cache coherency when writing.
1549          */
1550         status = NFS_PROTO(inode)->write_done(task, hdr);
1551         if (status != 0)
1552                 return status;
1553
1554         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1555         trace_nfs_writeback_done(task, hdr);
1556
1557         if (hdr->res.verf->committed < hdr->args.stable &&
1558             task->tk_status >= 0) {
1559                 /* We tried a write call, but the server did not
1560                  * commit data to stable storage even though we
1561                  * requested it.
1562                  * Note: There is a known bug in Tru64 < 5.0 in which
1563                  *       the server reports NFS_DATA_SYNC, but performs
1564                  *       NFS_FILE_SYNC. We therefore implement this checking
1565                  *       as a dprintk() in order to avoid filling syslog.
1566                  */
1567                 static unsigned long    complain;
1568
1569                 /* Note this will print the MDS for a DS write */
1570                 if (time_before(complain, jiffies)) {
1571                         dprintk("NFS:       faulty NFS server %s:"
1572                                 " (committed = %d) != (stable = %d)\n",
1573                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1574                                 hdr->res.verf->committed, hdr->args.stable);
1575                         complain = jiffies + 300 * HZ;
1576                 }
1577         }
1578
1579         /* Deal with the suid/sgid bit corner case */
1580         if (nfs_should_remove_suid(inode)) {
1581                 spin_lock(&inode->i_lock);
1582                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1583                 spin_unlock(&inode->i_lock);
1584         }
1585         return 0;
1586 }
1587
1588 /*
1589  * This function is called when the WRITE call is complete.
1590  */
1591 static void nfs_writeback_result(struct rpc_task *task,
1592                                  struct nfs_pgio_header *hdr)
1593 {
1594         struct nfs_pgio_args    *argp = &hdr->args;
1595         struct nfs_pgio_res     *resp = &hdr->res;
1596
1597         if (resp->count < argp->count) {
1598                 static unsigned long    complain;
1599
1600                 /* This a short write! */
1601                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1602
1603                 /* Has the server at least made some progress? */
1604                 if (resp->count == 0) {
1605                         if (time_before(complain, jiffies)) {
1606                                 printk(KERN_WARNING
1607                                        "NFS: Server wrote zero bytes, expected %u.\n",
1608                                        argp->count);
1609                                 complain = jiffies + 300 * HZ;
1610                         }
1611                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1612                         task->tk_status = -EIO;
1613                         return;
1614                 }
1615
1616                 /* For non rpc-based layout drivers, retry-through-MDS */
1617                 if (!task->tk_ops) {
1618                         hdr->pnfs_error = -EAGAIN;
1619                         return;
1620                 }
1621
1622                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1623                 if (resp->verf->committed != NFS_UNSTABLE) {
1624                         /* Resend from where the server left off */
1625                         hdr->mds_offset += resp->count;
1626                         argp->offset += resp->count;
1627                         argp->pgbase += resp->count;
1628                         argp->count -= resp->count;
1629                 } else {
1630                         /* Resend as a stable write in order to avoid
1631                          * headaches in the case of a server crash.
1632                          */
1633                         argp->stable = NFS_FILE_SYNC;
1634                 }
1635                 resp->count = 0;
1636                 resp->verf->committed = 0;
1637                 rpc_restart_call_prepare(task);
1638         }
1639 }
1640
1641 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1642 {
1643         return wait_var_event_killable(&cinfo->rpcs_out,
1644                                        !atomic_read(&cinfo->rpcs_out));
1645 }
1646
1647 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1648 {
1649         atomic_inc(&cinfo->rpcs_out);
1650 }
1651
1652 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1653 {
1654         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1655                 wake_up_var(&cinfo->rpcs_out);
1656                 return true;
1657         }
1658         return false;
1659 }
1660
1661 void nfs_commitdata_release(struct nfs_commit_data *data)
1662 {
1663         put_nfs_open_context(data->context);
1664         nfs_commit_free(data);
1665 }
1666 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1667
1668 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1669                         const struct nfs_rpc_ops *nfs_ops,
1670                         const struct rpc_call_ops *call_ops,
1671                         int how, int flags)
1672 {
1673         struct rpc_task *task;
1674         int priority = flush_task_priority(how);
1675         struct rpc_message msg = {
1676                 .rpc_argp = &data->args,
1677                 .rpc_resp = &data->res,
1678                 .rpc_cred = data->cred,
1679         };
1680         struct rpc_task_setup task_setup_data = {
1681                 .task = &data->task,
1682                 .rpc_client = clnt,
1683                 .rpc_message = &msg,
1684                 .callback_ops = call_ops,
1685                 .callback_data = data,
1686                 .workqueue = nfsiod_workqueue,
1687                 .flags = RPC_TASK_ASYNC | flags,
1688                 .priority = priority,
1689         };
1690
1691         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1692                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1693
1694         /* Set up the initial task struct.  */
1695         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1696         trace_nfs_initiate_commit(data);
1697
1698         dprintk("NFS: initiated commit call\n");
1699
1700         task = rpc_run_task(&task_setup_data);
1701         if (IS_ERR(task))
1702                 return PTR_ERR(task);
1703         if (how & FLUSH_SYNC)
1704                 rpc_wait_for_completion_task(task);
1705         rpc_put_task(task);
1706         return 0;
1707 }
1708 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1709
1710 static loff_t nfs_get_lwb(struct list_head *head)
1711 {
1712         loff_t lwb = 0;
1713         struct nfs_page *req;
1714
1715         list_for_each_entry(req, head, wb_list)
1716                 if (lwb < (req_offset(req) + req->wb_bytes))
1717                         lwb = req_offset(req) + req->wb_bytes;
1718
1719         return lwb;
1720 }
1721
1722 /*
1723  * Set up the argument/result storage required for the RPC call.
1724  */
1725 void nfs_init_commit(struct nfs_commit_data *data,
1726                      struct list_head *head,
1727                      struct pnfs_layout_segment *lseg,
1728                      struct nfs_commit_info *cinfo)
1729 {
1730         struct nfs_page *first;
1731         struct nfs_open_context *ctx;
1732         struct inode *inode;
1733
1734         /* Set up the RPC argument and reply structs
1735          * NB: take care not to mess about with data->commit et al. */
1736
1737         if (head)
1738                 list_splice_init(head, &data->pages);
1739
1740         first = nfs_list_entry(data->pages.next);
1741         ctx = nfs_req_openctx(first);
1742         inode = d_inode(ctx->dentry);
1743
1744         data->inode       = inode;
1745         data->cred        = ctx->cred;
1746         data->lseg        = lseg; /* reference transferred */
1747         /* only set lwb for pnfs commit */
1748         if (lseg)
1749                 data->lwb = nfs_get_lwb(&data->pages);
1750         data->mds_ops     = &nfs_commit_ops;
1751         data->completion_ops = cinfo->completion_ops;
1752         data->dreq        = cinfo->dreq;
1753
1754         data->args.fh     = NFS_FH(data->inode);
1755         /* Note: we always request a commit of the entire inode */
1756         data->args.offset = 0;
1757         data->args.count  = 0;
1758         data->context     = get_nfs_open_context(ctx);
1759         data->res.fattr   = &data->fattr;
1760         data->res.verf    = &data->verf;
1761         nfs_fattr_init(&data->fattr);
1762         nfs_commit_begin(cinfo->mds);
1763 }
1764 EXPORT_SYMBOL_GPL(nfs_init_commit);
1765
1766 void nfs_retry_commit(struct list_head *page_list,
1767                       struct pnfs_layout_segment *lseg,
1768                       struct nfs_commit_info *cinfo,
1769                       u32 ds_commit_idx)
1770 {
1771         struct nfs_page *req;
1772
1773         while (!list_empty(page_list)) {
1774                 req = nfs_list_entry(page_list->next);
1775                 nfs_list_remove_request(req);
1776                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1777                 if (!cinfo->dreq)
1778                         nfs_clear_page_commit(req->wb_page);
1779                 nfs_unlock_and_release_request(req);
1780         }
1781 }
1782 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1783
1784 static void
1785 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1786                 struct nfs_page *req)
1787 {
1788         __set_page_dirty_nobuffers(req->wb_page);
1789 }
1790
1791 /*
1792  * Commit dirty pages
1793  */
1794 static int
1795 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1796                 struct nfs_commit_info *cinfo)
1797 {
1798         struct nfs_commit_data  *data;
1799         unsigned short task_flags = 0;
1800
1801         /* another commit raced with us */
1802         if (list_empty(head))
1803                 return 0;
1804
1805         data = nfs_commitdata_alloc();
1806         if (!data) {
1807                 nfs_retry_commit(head, NULL, cinfo, -1);
1808                 return -ENOMEM;
1809         }
1810
1811         /* Set up the argument struct */
1812         nfs_init_commit(data, head, NULL, cinfo);
1813         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1814                 task_flags = RPC_TASK_MOVEABLE;
1815         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1816                                    data->mds_ops, how,
1817                                    RPC_TASK_CRED_NOREF | task_flags);
1818 }
1819
1820 /*
1821  * COMMIT call returned
1822  */
1823 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1824 {
1825         struct nfs_commit_data  *data = calldata;
1826
1827         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1828                                 task->tk_pid, task->tk_status);
1829
1830         /* Call the NFS version-specific code */
1831         NFS_PROTO(data->inode)->commit_done(task, data);
1832         trace_nfs_commit_done(task, data);
1833 }
1834
1835 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1836 {
1837         const struct nfs_writeverf *verf = data->res.verf;
1838         struct nfs_page *req;
1839         int status = data->task.tk_status;
1840         struct nfs_commit_info cinfo;
1841         struct nfs_server *nfss;
1842
1843         while (!list_empty(&data->pages)) {
1844                 req = nfs_list_entry(data->pages.next);
1845                 nfs_list_remove_request(req);
1846                 if (req->wb_page)
1847                         nfs_clear_page_commit(req->wb_page);
1848
1849                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1850                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1851                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1852                         req->wb_bytes,
1853                         (long long)req_offset(req));
1854                 if (status < 0) {
1855                         if (req->wb_page) {
1856                                 trace_nfs_commit_error(req, status);
1857                                 nfs_mapping_set_error(req->wb_page, status);
1858                                 nfs_inode_remove_request(req);
1859                         }
1860                         dprintk_cont(", error = %d\n", status);
1861                         goto next;
1862                 }
1863
1864                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1865                  * returned by the server against all stored verfs. */
1866                 if (nfs_write_match_verf(verf, req)) {
1867                         /* We have a match */
1868                         if (req->wb_page)
1869                                 nfs_inode_remove_request(req);
1870                         dprintk_cont(" OK\n");
1871                         goto next;
1872                 }
1873                 /* We have a mismatch. Write the page again */
1874                 dprintk_cont(" mismatch\n");
1875                 nfs_mark_request_dirty(req);
1876                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1877         next:
1878                 nfs_unlock_and_release_request(req);
1879                 /* Latency breaker */
1880                 cond_resched();
1881         }
1882         nfss = NFS_SERVER(data->inode);
1883         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1884                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1885
1886         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1887         nfs_commit_end(cinfo.mds);
1888 }
1889
1890 static void nfs_commit_release(void *calldata)
1891 {
1892         struct nfs_commit_data *data = calldata;
1893
1894         data->completion_ops->completion(data);
1895         nfs_commitdata_release(calldata);
1896 }
1897
1898 static const struct rpc_call_ops nfs_commit_ops = {
1899         .rpc_call_prepare = nfs_commit_prepare,
1900         .rpc_call_done = nfs_commit_done,
1901         .rpc_release = nfs_commit_release,
1902 };
1903
1904 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1905         .completion = nfs_commit_release_pages,
1906         .resched_write = nfs_commit_resched_write,
1907 };
1908
1909 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1910                             int how, struct nfs_commit_info *cinfo)
1911 {
1912         int status;
1913
1914         status = pnfs_commit_list(inode, head, how, cinfo);
1915         if (status == PNFS_NOT_ATTEMPTED)
1916                 status = nfs_commit_list(inode, head, how, cinfo);
1917         return status;
1918 }
1919
1920 static int __nfs_commit_inode(struct inode *inode, int how,
1921                 struct writeback_control *wbc)
1922 {
1923         LIST_HEAD(head);
1924         struct nfs_commit_info cinfo;
1925         int may_wait = how & FLUSH_SYNC;
1926         int ret, nscan;
1927
1928         how &= ~FLUSH_SYNC;
1929         nfs_init_cinfo_from_inode(&cinfo, inode);
1930         nfs_commit_begin(cinfo.mds);
1931         for (;;) {
1932                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1933                 if (ret <= 0)
1934                         break;
1935                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1936                 if (ret < 0)
1937                         break;
1938                 ret = 0;
1939                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1940                         if (nscan < wbc->nr_to_write)
1941                                 wbc->nr_to_write -= nscan;
1942                         else
1943                                 wbc->nr_to_write = 0;
1944                 }
1945                 if (nscan < INT_MAX)
1946                         break;
1947                 cond_resched();
1948         }
1949         nfs_commit_end(cinfo.mds);
1950         if (ret || !may_wait)
1951                 return ret;
1952         return wait_on_commit(cinfo.mds);
1953 }
1954
1955 int nfs_commit_inode(struct inode *inode, int how)
1956 {
1957         return __nfs_commit_inode(inode, how, NULL);
1958 }
1959 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1960
1961 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1962 {
1963         struct nfs_inode *nfsi = NFS_I(inode);
1964         int flags = FLUSH_SYNC;
1965         int ret = 0;
1966
1967         if (wbc->sync_mode == WB_SYNC_NONE) {
1968                 /* no commits means nothing needs to be done */
1969                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1970                         goto check_requests_outstanding;
1971
1972                 /* Don't commit yet if this is a non-blocking flush and there
1973                  * are a lot of outstanding writes for this mapping.
1974                  */
1975                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1976                         goto out_mark_dirty;
1977
1978                 /* don't wait for the COMMIT response */
1979                 flags = 0;
1980         }
1981
1982         ret = __nfs_commit_inode(inode, flags, wbc);
1983         if (!ret) {
1984                 if (flags & FLUSH_SYNC)
1985                         return 0;
1986         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1987                 goto out_mark_dirty;
1988
1989 check_requests_outstanding:
1990         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1991                 return ret;
1992 out_mark_dirty:
1993         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1994         return ret;
1995 }
1996 EXPORT_SYMBOL_GPL(nfs_write_inode);
1997
1998 /*
1999  * Wrapper for filemap_write_and_wait_range()
2000  *
2001  * Needed for pNFS in order to ensure data becomes visible to the
2002  * client.
2003  */
2004 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2005                 loff_t lstart, loff_t lend)
2006 {
2007         int ret;
2008
2009         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2010         if (ret == 0)
2011                 ret = pnfs_sync_inode(mapping->host, true);
2012         return ret;
2013 }
2014 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2015
2016 /*
2017  * flush the inode to disk.
2018  */
2019 int nfs_wb_all(struct inode *inode)
2020 {
2021         int ret;
2022
2023         trace_nfs_writeback_inode_enter(inode);
2024
2025         ret = filemap_write_and_wait(inode->i_mapping);
2026         if (ret)
2027                 goto out;
2028         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2029         if (ret < 0)
2030                 goto out;
2031         pnfs_sync_inode(inode, true);
2032         ret = 0;
2033
2034 out:
2035         trace_nfs_writeback_inode_exit(inode, ret);
2036         return ret;
2037 }
2038 EXPORT_SYMBOL_GPL(nfs_wb_all);
2039
2040 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2041 {
2042         struct nfs_page *req;
2043         int ret = 0;
2044
2045         wait_on_page_writeback(page);
2046
2047         /* blocking call to cancel all requests and join to a single (head)
2048          * request */
2049         req = nfs_lock_and_join_requests(page);
2050
2051         if (IS_ERR(req)) {
2052                 ret = PTR_ERR(req);
2053         } else if (req) {
2054                 /* all requests from this page have been cancelled by
2055                  * nfs_lock_and_join_requests, so just remove the head
2056                  * request from the inode / page_private pointer and
2057                  * release it */
2058                 nfs_inode_remove_request(req);
2059                 nfs_unlock_and_release_request(req);
2060         }
2061
2062         return ret;
2063 }
2064
2065 /*
2066  * Write back all requests on one page - we do this before reading it.
2067  */
2068 int nfs_wb_page(struct inode *inode, struct page *page)
2069 {
2070         loff_t range_start = page_file_offset(page);
2071         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2072         struct writeback_control wbc = {
2073                 .sync_mode = WB_SYNC_ALL,
2074                 .nr_to_write = 0,
2075                 .range_start = range_start,
2076                 .range_end = range_end,
2077         };
2078         int ret;
2079
2080         trace_nfs_writeback_page_enter(inode);
2081
2082         for (;;) {
2083                 wait_on_page_writeback(page);
2084                 if (clear_page_dirty_for_io(page)) {
2085                         ret = nfs_writepage_locked(page, &wbc);
2086                         if (ret < 0)
2087                                 goto out_error;
2088                         continue;
2089                 }
2090                 ret = 0;
2091                 if (!PagePrivate(page))
2092                         break;
2093                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2094                 if (ret < 0)
2095                         goto out_error;
2096         }
2097 out_error:
2098         trace_nfs_writeback_page_exit(inode, ret);
2099         return ret;
2100 }
2101
2102 #ifdef CONFIG_MIGRATION
2103 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2104                 struct page *page, enum migrate_mode mode)
2105 {
2106         /*
2107          * If PagePrivate is set, then the page is currently associated with
2108          * an in-progress read or write request. Don't try to migrate it.
2109          *
2110          * FIXME: we could do this in principle, but we'll need a way to ensure
2111          *        that we can safely release the inode reference while holding
2112          *        the page lock.
2113          */
2114         if (PagePrivate(page))
2115                 return -EBUSY;
2116
2117         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2118                 return -EBUSY;
2119
2120         return migrate_page(mapping, newpage, page, mode);
2121 }
2122 #endif
2123
2124 int __init nfs_init_writepagecache(void)
2125 {
2126         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2127                                              sizeof(struct nfs_pgio_header),
2128                                              0, SLAB_HWCACHE_ALIGN,
2129                                              NULL);
2130         if (nfs_wdata_cachep == NULL)
2131                 return -ENOMEM;
2132
2133         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2134                                                      nfs_wdata_cachep);
2135         if (nfs_wdata_mempool == NULL)
2136                 goto out_destroy_write_cache;
2137
2138         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2139                                              sizeof(struct nfs_commit_data),
2140                                              0, SLAB_HWCACHE_ALIGN,
2141                                              NULL);
2142         if (nfs_cdata_cachep == NULL)
2143                 goto out_destroy_write_mempool;
2144
2145         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2146                                                       nfs_cdata_cachep);
2147         if (nfs_commit_mempool == NULL)
2148                 goto out_destroy_commit_cache;
2149
2150         /*
2151          * NFS congestion size, scale with available memory.
2152          *
2153          *  64MB:    8192k
2154          * 128MB:   11585k
2155          * 256MB:   16384k
2156          * 512MB:   23170k
2157          *   1GB:   32768k
2158          *   2GB:   46340k
2159          *   4GB:   65536k
2160          *   8GB:   92681k
2161          *  16GB:  131072k
2162          *
2163          * This allows larger machines to have larger/more transfers.
2164          * Limit the default to 256M
2165          */
2166         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2167         if (nfs_congestion_kb > 256*1024)
2168                 nfs_congestion_kb = 256*1024;
2169
2170         return 0;
2171
2172 out_destroy_commit_cache:
2173         kmem_cache_destroy(nfs_cdata_cachep);
2174 out_destroy_write_mempool:
2175         mempool_destroy(nfs_wdata_mempool);
2176 out_destroy_write_cache:
2177         kmem_cache_destroy(nfs_wdata_cachep);
2178         return -ENOMEM;
2179 }
2180
2181 void nfs_destroy_writepagecache(void)
2182 {
2183         mempool_destroy(nfs_commit_mempool);
2184         kmem_cache_destroy(nfs_cdata_cachep);
2185         mempool_destroy(nfs_wdata_mempool);
2186         kmem_cache_destroy(nfs_wdata_cachep);
2187 }
2188
2189 static const struct nfs_rw_ops nfs_rw_write_ops = {
2190         .rw_alloc_header        = nfs_writehdr_alloc,
2191         .rw_free_header         = nfs_writehdr_free,
2192         .rw_done                = nfs_writeback_done,
2193         .rw_result              = nfs_writeback_result,
2194         .rw_initiate            = nfs_initiate_write,
2195 };