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 <[email protected]>
   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_page *req);
  62 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  63                                       struct inode *inode);
  64 static struct nfs_page *
  65 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  66                                                 struct page *page);
  67
  68 static struct kmem_cache *nfs_wdata_cachep;
  69 static mempool_t *nfs_wdata_mempool;
  70 static struct kmem_cache *nfs_cdata_cachep;
  71 static mempool_t *nfs_commit_mempool;
  72
  73 struct nfs_commit_data *nfs_commitdata_alloc(bool never_fail)
  74 {
  75         struct nfs_commit_data *p;
  76
  77         if (never_fail)
  78                 p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  79         else {
  80                 /* It is OK to do some reclaim, not no safe to wait
  81                  * for anything to be returned to the pool.
  82                  * mempool_alloc() cannot handle that particular combination,
  83                  * so we need two separate attempts.
  84                  */
  85                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  86                 if (!p)
  87                         p = kmem_cache_alloc(nfs_cdata_cachep, GFP_NOIO |
  88                                              __GFP_NOWARN | __GFP_NORETRY);
  89                 if (!p)
  90                         return NULL;
  91         }
  92
  93         memset(p, 0, sizeof(*p));
  94         INIT_LIST_HEAD(&p->pages);
  95         return p;
  96 }
  97 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  98
  99 void nfs_commit_free(struct nfs_commit_data *p)
 100 {
 101         mempool_free(p, nfs_commit_mempool);
 102 }
 103 EXPORT_SYMBOL_GPL(nfs_commit_free);
 104
 105 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
 106 {
 107         struct nfs_pgio_header *p = mempool_alloc(nfs_wdata_mempool, GFP_KERNEL);
 108
 109         memset(p, 0, sizeof(*p));
 110         p->rw_mode = FMODE_WRITE;
 111         return p;
 112 }
 113
 114 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 115 {
 116         mempool_free(hdr, nfs_wdata_mempool);
 117 }
 118
 119 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 120 {
 121         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 122 }
 123
 124 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 125                 void (*complete)(void *), void *data)
 126 {
 127         ioc->complete = complete;
 128         ioc->data = data;
 129         kref_init(&ioc->refcount);
 130 }
 131
 132 static void nfs_io_completion_release(struct kref *kref)
 133 {
 134         struct nfs_io_completion *ioc = container_of(kref,
 135                         struct nfs_io_completion, refcount);
 136         ioc->complete(ioc->data);
 137         kfree(ioc);
 138 }
 139
 140 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 141 {
 142         if (ioc != NULL)
 143                 kref_get(&ioc->refcount);
 144 }
 145
 146 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 147 {
 148         if (ioc != NULL)
 149                 kref_put(&ioc->refcount, nfs_io_completion_release);
 150 }
 151
 152 static void
 153 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 154 {
 155         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 156                 kref_get(&req->wb_kref);
 157                 atomic_long_inc(&NFS_I(inode)->nrequests);
 158         }
 159 }
 160
 161 static int
 162 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 163 {
 164         int ret;
 165
 166         if (!test_bit(PG_REMOVE, &req->wb_flags))
 167                 return 0;
 168         ret = nfs_page_group_lock(req);
 169         if (ret)
 170                 return ret;
 171         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 172                 nfs_page_set_inode_ref(req, inode);
 173         nfs_page_group_unlock(req);
 174         return 0;
 175 }
 176
 177 static struct nfs_page *
 178 nfs_page_private_request(struct page *page)
 179 {
 180         if (!PagePrivate(page))
 181                 return NULL;
 182         return (struct nfs_page *)page_private(page);
 183 }
 184
 185 /*
 186  * nfs_page_find_head_request_locked - find head request associated with @page
 187  *
 188  * must be called while holding the inode lock.
 189  *
 190  * returns matching head request with reference held, or NULL if not found.
 191  */
 192 static struct nfs_page *
 193 nfs_page_find_private_request(struct page *page)
 194 {
 195         struct address_space *mapping = page_file_mapping(page);
 196         struct nfs_page *req;
 197
 198         if (!PagePrivate(page))
 199                 return NULL;
 200         spin_lock(&mapping->private_lock);
 201         req = nfs_page_private_request(page);
 202         if (req) {
 203                 WARN_ON_ONCE(req->wb_head != req);
 204                 kref_get(&req->wb_kref);
 205         }
 206         spin_unlock(&mapping->private_lock);
 207         return req;
 208 }
 209
 210 static struct nfs_page *
 211 nfs_page_find_swap_request(struct page *page)
 212 {
 213         struct inode *inode = page_file_mapping(page)->host;
 214         struct nfs_inode *nfsi = NFS_I(inode);
 215         struct nfs_page *req = NULL;
 216         if (!PageSwapCache(page))
 217                 return NULL;
 218         mutex_lock(&nfsi->commit_mutex);
 219         if (PageSwapCache(page)) {
 220                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 221                         page);
 222                 if (req) {
 223                         WARN_ON_ONCE(req->wb_head != req);
 224                         kref_get(&req->wb_kref);
 225                 }
 226         }
 227         mutex_unlock(&nfsi->commit_mutex);
 228         return req;
 229 }
 230
 231 /*
 232  * nfs_page_find_head_request - find head request associated with @page
 233  *
 234  * returns matching head request with reference held, or NULL if not found.
 235  */
 236 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 237 {
 238         struct nfs_page *req;
 239
 240         req = nfs_page_find_private_request(page);
 241         if (!req)
 242                 req = nfs_page_find_swap_request(page);
 243         return req;
 244 }
 245
 246 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
 247 {
 248         struct inode *inode = page_file_mapping(page)->host;
 249         struct nfs_page *req, *head;
 250         int ret;
 251
 252         for (;;) {
 253                 req = nfs_page_find_head_request(page);
 254                 if (!req)
 255                         return req;
 256                 head = nfs_page_group_lock_head(req);
 257                 if (head != req)
 258                         nfs_release_request(req);
 259                 if (IS_ERR(head))
 260                         return head;
 261                 ret = nfs_cancel_remove_inode(head, inode);
 262                 if (ret < 0) {
 263                         nfs_unlock_and_release_request(head);
 264                         return ERR_PTR(ret);
 265                 }
 266                 /* Ensure that nobody removed the request before we locked it */
 267                 if (head == nfs_page_private_request(page))
 268                         break;
 269                 if (PageSwapCache(page))
 270                         break;
 271                 nfs_unlock_and_release_request(head);
 272         }
 273         return head;
 274 }
 275
 276 /* Adjust the file length if we're writing beyond the end */
 277 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 278 {
 279         struct inode *inode = page_file_mapping(page)->host;
 280         loff_t end, i_size;
 281         pgoff_t end_index;
 282
 283         spin_lock(&inode->i_lock);
 284         i_size = i_size_read(inode);
 285         end_index = (i_size - 1) >> PAGE_SHIFT;
 286         if (i_size > 0 && page_index(page) < end_index)
 287                 goto out;
 288         end = page_file_offset(page) + ((loff_t)offset+count);
 289         if (i_size >= end)
 290                 goto out;
 291         trace_nfs_size_grow(inode, end);
 292         i_size_write(inode, end);
 293         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 294         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 295 out:
 296         spin_unlock(&inode->i_lock);
 297 }
 298
 299 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 300 static void nfs_set_pageerror(struct address_space *mapping)
 301 {
 302         struct inode *inode = mapping->host;
 303
 304         nfs_zap_mapping(mapping->host, mapping);
 305         /* Force file size revalidation */
 306         spin_lock(&inode->i_lock);
 307         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 308                                              NFS_INO_REVAL_PAGECACHE |
 309                                              NFS_INO_INVALID_SIZE);
 310         spin_unlock(&inode->i_lock);
 311 }
 312
 313 static void nfs_mapping_set_error(struct page *page, int error)
 314 {
 315         struct address_space *mapping = page_file_mapping(page);
 316
 317         SetPageError(page);
 318         mapping_set_error(mapping, error);
 319         nfs_set_pageerror(mapping);
 320 }
 321
 322 /*
 323  * nfs_page_group_search_locked
 324  * @head - head request of page group
 325  * @page_offset - offset into page
 326  *
 327  * Search page group with head @head to find a request that contains the
 328  * page offset @page_offset.
 329  *
 330  * Returns a pointer to the first matching nfs request, or NULL if no
 331  * match is found.
 332  *
 333  * Must be called with the page group lock held
 334  */
 335 static struct nfs_page *
 336 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 337 {
 338         struct nfs_page *req;
 339
 340         req = head;
 341         do {
 342                 if (page_offset >= req->wb_pgbase &&
 343                     page_offset < (req->wb_pgbase + req->wb_bytes))
 344                         return req;
 345
 346                 req = req->wb_this_page;
 347         } while (req != head);
 348
 349         return NULL;
 350 }
 351
 352 /*
 353  * nfs_page_group_covers_page
 354  * @head - head request of page group
 355  *
 356  * Return true if the page group with head @head covers the whole page,
 357  * returns false otherwise
 358  */
 359 static bool nfs_page_group_covers_page(struct nfs_page *req)
 360 {
 361         struct nfs_page *tmp;
 362         unsigned int pos = 0;
 363         unsigned int len = nfs_page_length(req->wb_page);
 364
 365         nfs_page_group_lock(req);
 366
 367         for (;;) {
 368                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 369                 if (!tmp)
 370                         break;
 371                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 372         }
 373
 374         nfs_page_group_unlock(req);
 375         return pos >= len;
 376 }
 377
 378 /* We can set the PG_uptodate flag if we see that a write request
 379  * covers the full page.
 380  */
 381 static void nfs_mark_uptodate(struct nfs_page *req)
 382 {
 383         if (PageUptodate(req->wb_page))
 384                 return;
 385         if (!nfs_page_group_covers_page(req))
 386                 return;
 387         SetPageUptodate(req->wb_page);
 388 }
 389
 390 static int wb_priority(struct writeback_control *wbc)
 391 {
 392         int ret = 0;
 393
 394         if (wbc->sync_mode == WB_SYNC_ALL)
 395                 ret = FLUSH_COND_STABLE;
 396         return ret;
 397 }
 398
 399 /*
 400  * NFS congestion control
 401  */
 402
 403 int nfs_congestion_kb;
 404
 405 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 406 #define NFS_CONGESTION_OFF_THRESH       \
 407         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 408
 409 static void nfs_set_page_writeback(struct page *page)
 410 {
 411         struct inode *inode = page_file_mapping(page)->host;
 412         struct nfs_server *nfss = NFS_SERVER(inode);
 413         int ret = test_set_page_writeback(page);
 414
 415         WARN_ON_ONCE(ret != 0);
 416
 417         if (atomic_long_inc_return(&nfss->writeback) >
 418                         NFS_CONGESTION_ON_THRESH)
 419                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 420 }
 421
 422 static void nfs_end_page_writeback(struct nfs_page *req)
 423 {
 424         struct inode *inode = page_file_mapping(req->wb_page)->host;
 425         struct nfs_server *nfss = NFS_SERVER(inode);
 426         bool is_done;
 427
 428         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 429         nfs_unlock_request(req);
 430         if (!is_done)
 431                 return;
 432
 433         end_page_writeback(req->wb_page);
 434         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 435                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 436 }
 437
 438 /*
 439  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 440  *
 441  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 442  * @old_head - the old head of the list
 443  *
 444  * All subrequests must be locked and removed from all lists, so at this point
 445  * they are only "active" in this function, and possibly in nfs_wait_on_request
 446  * with a reference held by some other context.
 447  */
 448 static void
 449 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 450                                  struct nfs_page *old_head,
 451                                  struct inode *inode)
 452 {
 453         while (destroy_list) {
 454                 struct nfs_page *subreq = destroy_list;
 455
 456                 destroy_list = (subreq->wb_this_page == old_head) ?
 457                                    NULL : subreq->wb_this_page;
 458
 459                 /* Note: lock subreq in order to change subreq->wb_head */
 460                 nfs_page_set_headlock(subreq);
 461                 WARN_ON_ONCE(old_head != subreq->wb_head);
 462
 463                 /* make sure old group is not used */
 464                 subreq->wb_this_page = subreq;
 465                 subreq->wb_head = subreq;
 466
 467                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 468
 469                 /* Note: races with nfs_page_group_destroy() */
 470                 if (!kref_read(&subreq->wb_kref)) {
 471                         /* Check if we raced with nfs_page_group_destroy() */
 472                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 473                                 nfs_page_clear_headlock(subreq);
 474                                 nfs_free_request(subreq);
 475                         } else
 476                                 nfs_page_clear_headlock(subreq);
 477                         continue;
 478                 }
 479                 nfs_page_clear_headlock(subreq);
 480
 481                 nfs_release_request(old_head);
 482
 483                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 484                         nfs_release_request(subreq);
 485                         atomic_long_dec(&NFS_I(inode)->nrequests);
 486                 }
 487
 488                 /* subreq is now totally disconnected from page group or any
 489                  * write / commit lists. last chance to wake any waiters */
 490                 nfs_unlock_and_release_request(subreq);
 491         }
 492 }
 493
 494 /*
 495  * nfs_join_page_group - destroy subrequests of the head req
 496  * @head: the page used to lookup the "page group" of nfs_page structures
 497  * @inode: Inode to which the request belongs.
 498  *
 499  * This function joins all sub requests to the head request by first
 500  * locking all requests in the group, cancelling any pending operations
 501  * and finally updating the head request to cover the whole range covered by
 502  * the (former) group.  All subrequests are removed from any write or commit
 503  * lists, unlinked from the group and destroyed.
 504  */
 505 void
 506 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
 507 {
 508         struct nfs_page *subreq;
 509         struct nfs_page *destroy_list = NULL;
 510         unsigned int pgbase, off, bytes;
 511
 512         pgbase = head->wb_pgbase;
 513         bytes = head->wb_bytes;
 514         off = head->wb_offset;
 515         for (subreq = head->wb_this_page; subreq != head;
 516                         subreq = subreq->wb_this_page) {
 517                 /* Subrequests should always form a contiguous range */
 518                 if (pgbase > subreq->wb_pgbase) {
 519                         off -= pgbase - subreq->wb_pgbase;
 520                         bytes += pgbase - subreq->wb_pgbase;
 521                         pgbase = subreq->wb_pgbase;
 522                 }
 523                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 524                                 - pgbase, bytes);
 525         }
 526
 527         /* Set the head request's range to cover the former page group */
 528         head->wb_pgbase = pgbase;
 529         head->wb_bytes = bytes;
 530         head->wb_offset = off;
 531
 532         /* Now that all requests are locked, make sure they aren't on any list.
 533          * Commit list removal accounting is done after locks are dropped */
 534         subreq = head;
 535         do {
 536                 nfs_clear_request_commit(subreq);
 537                 subreq = subreq->wb_this_page;
 538         } while (subreq != head);
 539
 540         /* unlink subrequests from head, destroy them later */
 541         if (head->wb_this_page != head) {
 542                 /* destroy list will be terminated by head */
 543                 destroy_list = head->wb_this_page;
 544                 head->wb_this_page = head;
 545         }
 546
 547         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 548 }
 549
 550 /*
 551  * nfs_lock_and_join_requests - join all subreqs to the head req
 552  * @page: the page used to lookup the "page group" of nfs_page structures
 553  *
 554  * This function joins all sub requests to the head request by first
 555  * locking all requests in the group, cancelling any pending operations
 556  * and finally updating the head request to cover the whole range covered by
 557  * the (former) group.  All subrequests are removed from any write or commit
 558  * lists, unlinked from the group and destroyed.
 559  *
 560  * Returns a locked, referenced pointer to the head request - which after
 561  * this call is guaranteed to be the only request associated with the page.
 562  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 563  * error was encountered.
 564  */
 565 static struct nfs_page *
 566 nfs_lock_and_join_requests(struct page *page)
 567 {
 568         struct inode *inode = page_file_mapping(page)->host;
 569         struct nfs_page *head;
 570         int ret;
 571
 572         /*
 573          * A reference is taken only on the head request which acts as a
 574          * reference to the whole page group - the group will not be destroyed
 575          * until the head reference is released.
 576          */
 577         head = nfs_find_and_lock_page_request(page);
 578         if (IS_ERR_OR_NULL(head))
 579                 return head;
 580
 581         /* lock each request in the page group */
 582         ret = nfs_page_group_lock_subrequests(head);
 583         if (ret < 0) {
 584                 nfs_unlock_and_release_request(head);
 585                 return ERR_PTR(ret);
 586         }
 587
 588         nfs_join_page_group(head, inode);
 589
 590         return head;
 591 }
 592
 593 static void nfs_write_error(struct nfs_page *req, int error)
 594 {
 595         trace_nfs_write_error(req, error);
 596         nfs_mapping_set_error(req->wb_page, error);
 597         nfs_inode_remove_request(req);
 598         nfs_end_page_writeback(req);
 599         nfs_release_request(req);
 600 }
 601
 602 /*
 603  * Find an associated nfs write request, and prepare to flush it out
 604  * May return an error if the user signalled nfs_wait_on_request().
 605  */
 606 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 607                                 struct page *page)
 608 {
 609         struct nfs_page *req;
 610         int ret = 0;
 611
 612         req = nfs_lock_and_join_requests(page);
 613         if (!req)
 614                 goto out;
 615         ret = PTR_ERR(req);
 616         if (IS_ERR(req))
 617                 goto out;
 618
 619         nfs_set_page_writeback(page);
 620         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 621
 622         /* If there is a fatal error that covers this write, just exit */
 623         ret = pgio->pg_error;
 624         if (nfs_error_is_fatal_on_server(ret))
 625                 goto out_launder;
 626
 627         ret = 0;
 628         if (!nfs_pageio_add_request(pgio, req)) {
 629                 ret = pgio->pg_error;
 630                 /*
 631                  * Remove the problematic req upon fatal errors on the server
 632                  */
 633                 if (nfs_error_is_fatal(ret)) {
 634                         if (nfs_error_is_fatal_on_server(ret))
 635                                 goto out_launder;
 636                 } else
 637                         ret = -EAGAIN;
 638                 nfs_redirty_request(req);
 639                 pgio->pg_error = 0;
 640         } else
 641                 nfs_add_stats(page_file_mapping(page)->host,
 642                                 NFSIOS_WRITEPAGES, 1);
 643 out:
 644         return ret;
 645 out_launder:
 646         nfs_write_error(req, ret);
 647         return 0;
 648 }
 649
 650 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 651                             struct nfs_pageio_descriptor *pgio)
 652 {
 653         int ret;
 654
 655         nfs_pageio_cond_complete(pgio, page_index(page));
 656         ret = nfs_page_async_flush(pgio, page);
 657         if (ret == -EAGAIN) {
 658                 redirty_page_for_writepage(wbc, page);
 659                 ret = AOP_WRITEPAGE_ACTIVATE;
 660         }
 661         return ret;
 662 }
 663
 664 /*
 665  * Write an mmapped page to the server.
 666  */
 667 static int nfs_writepage_locked(struct page *page,
 668                                 struct writeback_control *wbc)
 669 {
 670         struct nfs_pageio_descriptor pgio;
 671         struct inode *inode = page_file_mapping(page)->host;
 672         int err;
 673
 674         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 675         nfs_pageio_init_write(&pgio, inode, 0,
 676                                 false, &nfs_async_write_completion_ops);
 677         err = nfs_do_writepage(page, wbc, &pgio);
 678         pgio.pg_error = 0;
 679         nfs_pageio_complete(&pgio);
 680         if (err < 0)
 681                 return err;
 682         if (nfs_error_is_fatal(pgio.pg_error))
 683                 return pgio.pg_error;
 684         return 0;
 685 }
 686
 687 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 688 {
 689         int ret;
 690
 691         ret = nfs_writepage_locked(page, wbc);
 692         if (ret != AOP_WRITEPAGE_ACTIVATE)
 693                 unlock_page(page);
 694         return ret;
 695 }
 696
 697 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 698 {
 699         int ret;
 700
 701         ret = nfs_do_writepage(page, wbc, data);
 702         if (ret != AOP_WRITEPAGE_ACTIVATE)
 703                 unlock_page(page);
 704         return ret;
 705 }
 706
 707 static void nfs_io_completion_commit(void *inode)
 708 {
 709         nfs_commit_inode(inode, 0);
 710 }
 711
 712 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 713 {
 714         struct inode *inode = mapping->host;
 715         struct nfs_pageio_descriptor pgio;
 716         struct nfs_io_completion *ioc = NULL;
 717         unsigned int mntflags = NFS_SERVER(inode)->flags;
 718         int priority = 0;
 719         int err;
 720
 721         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 722
 723         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 724             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 725                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 726                 if (ioc)
 727                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 728                                                inode);
 729                 priority = wb_priority(wbc);
 730         }
 731
 732         nfs_pageio_init_write(&pgio, inode, priority, false,
 733                                 &nfs_async_write_completion_ops);
 734         pgio.pg_io_completion = ioc;
 735         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 736         pgio.pg_error = 0;
 737         nfs_pageio_complete(&pgio);
 738         nfs_io_completion_put(ioc);
 739
 740         if (err < 0)
 741                 goto out_err;
 742         err = pgio.pg_error;
 743         if (nfs_error_is_fatal(err))
 744                 goto out_err;
 745         return 0;
 746 out_err:
 747         return err;
 748 }
 749
 750 /*
 751  * Insert a write request into an inode
 752  */
 753 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 754 {
 755         struct address_space *mapping = page_file_mapping(req->wb_page);
 756         struct nfs_inode *nfsi = NFS_I(inode);
 757
 758         WARN_ON_ONCE(req->wb_this_page != req);
 759
 760         /* Lock the request! */
 761         nfs_lock_request(req);
 762
 763         /*
 764          * Swap-space should not get truncated. Hence no need to plug the race
 765          * with invalidate/truncate.
 766          */
 767         spin_lock(&mapping->private_lock);
 768         if (likely(!PageSwapCache(req->wb_page))) {
 769                 set_bit(PG_MAPPED, &req->wb_flags);
 770                 SetPagePrivate(req->wb_page);
 771                 set_page_private(req->wb_page, (unsigned long)req);
 772         }
 773         spin_unlock(&mapping->private_lock);
 774         atomic_long_inc(&nfsi->nrequests);
 775         /* this a head request for a page group - mark it as having an
 776          * extra reference so sub groups can follow suit.
 777          * This flag also informs pgio layer when to bump nrequests when
 778          * adding subrequests. */
 779         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 780         kref_get(&req->wb_kref);
 781 }
 782
 783 /*
 784  * Remove a write request from an inode
 785  */
 786 static void nfs_inode_remove_request(struct nfs_page *req)
 787 {
 788         struct address_space *mapping = page_file_mapping(req->wb_page);
 789         struct inode *inode = mapping->host;
 790         struct nfs_inode *nfsi = NFS_I(inode);
 791         struct nfs_page *head;
 792
 793         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 794                 head = req->wb_head;
 795
 796                 spin_lock(&mapping->private_lock);
 797                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 798                         set_page_private(head->wb_page, 0);
 799                         ClearPagePrivate(head->wb_page);
 800                         clear_bit(PG_MAPPED, &head->wb_flags);
 801                 }
 802                 spin_unlock(&mapping->private_lock);
 803         }
 804
 805         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 806                 nfs_release_request(req);
 807                 atomic_long_dec(&nfsi->nrequests);
 808         }
 809 }
 810
 811 static void
 812 nfs_mark_request_dirty(struct nfs_page *req)
 813 {
 814         if (req->wb_page)
 815                 __set_page_dirty_nobuffers(req->wb_page);
 816 }
 817
 818 /*
 819  * nfs_page_search_commits_for_head_request_locked
 820  *
 821  * Search through commit lists on @inode for the head request for @page.
 822  * Must be called while holding the inode (which is cinfo) lock.
 823  *
 824  * Returns the head request if found, or NULL if not found.
 825  */
 826 static struct nfs_page *
 827 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 828                                                 struct page *page)
 829 {
 830         struct nfs_page *freq, *t;
 831         struct nfs_commit_info cinfo;
 832         struct inode *inode = &nfsi->vfs_inode;
 833
 834         nfs_init_cinfo_from_inode(&cinfo, inode);
 835
 836         /* search through pnfs commit lists */
 837         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 838         if (freq)
 839                 return freq->wb_head;
 840
 841         /* Linearly search the commit list for the correct request */
 842         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 843                 if (freq->wb_page == page)
 844                         return freq->wb_head;
 845         }
 846
 847         return NULL;
 848 }
 849
 850 /**
 851  * nfs_request_add_commit_list_locked - add request to a commit list
 852  * @req: pointer to a struct nfs_page
 853  * @dst: commit list head
 854  * @cinfo: holds list lock and accounting info
 855  *
 856  * This sets the PG_CLEAN bit, updates the cinfo count of
 857  * number of outstanding requests requiring a commit as well as
 858  * the MM page stats.
 859  *
 860  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 861  * nfs_page lock.
 862  */
 863 void
 864 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 865                             struct nfs_commit_info *cinfo)
 866 {
 867         set_bit(PG_CLEAN, &req->wb_flags);
 868         nfs_list_add_request(req, dst);
 869         atomic_long_inc(&cinfo->mds->ncommit);
 870 }
 871 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 872
 873 /**
 874  * nfs_request_add_commit_list - add request to a commit list
 875  * @req: pointer to a struct nfs_page
 876  * @cinfo: holds list lock and accounting info
 877  *
 878  * This sets the PG_CLEAN bit, updates the cinfo count of
 879  * number of outstanding requests requiring a commit as well as
 880  * the MM page stats.
 881  *
 882  * The caller must _not_ hold the cinfo->lock, but must be
 883  * holding the nfs_page lock.
 884  */
 885 void
 886 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 887 {
 888         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 889         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 890         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 891         if (req->wb_page)
 892                 nfs_mark_page_unstable(req->wb_page, cinfo);
 893 }
 894 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 895
 896 /**
 897  * nfs_request_remove_commit_list - Remove request from a commit list
 898  * @req: pointer to a nfs_page
 899  * @cinfo: holds list lock and accounting info
 900  *
 901  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 902  * number of outstanding requests requiring a commit
 903  * It does not update the MM page stats.
 904  *
 905  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 906  */
 907 void
 908 nfs_request_remove_commit_list(struct nfs_page *req,
 909                                struct nfs_commit_info *cinfo)
 910 {
 911         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 912                 return;
 913         nfs_list_remove_request(req);
 914         atomic_long_dec(&cinfo->mds->ncommit);
 915 }
 916 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 917
 918 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 919                                       struct inode *inode)
 920 {
 921         cinfo->inode = inode;
 922         cinfo->mds = &NFS_I(inode)->commit_info;
 923         cinfo->ds = pnfs_get_ds_info(inode);
 924         cinfo->dreq = NULL;
 925         cinfo->completion_ops = &nfs_commit_completion_ops;
 926 }
 927
 928 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 929                     struct inode *inode,
 930                     struct nfs_direct_req *dreq)
 931 {
 932         if (dreq)
 933                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 934         else
 935                 nfs_init_cinfo_from_inode(cinfo, inode);
 936 }
 937 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 938
 939 /*
 940  * Add a request to the inode's commit list.
 941  */
 942 void
 943 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 944                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 945 {
 946         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 947                 return;
 948         nfs_request_add_commit_list(req, cinfo);
 949 }
 950
 951 static void
 952 nfs_clear_page_commit(struct page *page)
 953 {
 954         dec_node_page_state(page, NR_WRITEBACK);
 955         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 956                     WB_WRITEBACK);
 957 }
 958
 959 /* Called holding the request lock on @req */
 960 static void
 961 nfs_clear_request_commit(struct nfs_page *req)
 962 {
 963         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 964                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 965                 struct inode *inode = d_inode(ctx->dentry);
 966                 struct nfs_commit_info cinfo;
 967
 968                 nfs_init_cinfo_from_inode(&cinfo, inode);
 969                 mutex_lock(&NFS_I(inode)->commit_mutex);
 970                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 971                         nfs_request_remove_commit_list(req, &cinfo);
 972                 }
 973                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 974                 nfs_clear_page_commit(req->wb_page);
 975         }
 976 }
 977
 978 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 979 {
 980         if (hdr->verf.committed == NFS_DATA_SYNC)
 981                 return hdr->lseg == NULL;
 982         return hdr->verf.committed != NFS_FILE_SYNC;
 983 }
 984
 985 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 986 {
 987         nfs_io_completion_get(hdr->io_completion);
 988 }
 989
 990 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 991 {
 992         struct nfs_commit_info cinfo;
 993         unsigned long bytes = 0;
 994
 995         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 996                 goto out;
 997         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 998         while (!list_empty(&hdr->pages)) {
 999                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1000
1001                 bytes += req->wb_bytes;
1002                 nfs_list_remove_request(req);
1003                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1004                     (hdr->good_bytes < bytes)) {
1005                         trace_nfs_comp_error(req, hdr->error);
1006                         nfs_mapping_set_error(req->wb_page, hdr->error);
1007                         goto remove_req;
1008                 }
1009                 if (nfs_write_need_commit(hdr)) {
1010                         /* Reset wb_nio, since the write was successful. */
1011                         req->wb_nio = 0;
1012                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1013                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1014                                 hdr->pgio_mirror_idx);
1015                         goto next;
1016                 }
1017 remove_req:
1018                 nfs_inode_remove_request(req);
1019 next:
1020                 nfs_end_page_writeback(req);
1021                 nfs_release_request(req);
1022         }
1023 out:
1024         nfs_io_completion_put(hdr->io_completion);
1025         hdr->release(hdr);
1026 }
1027
1028 unsigned long
1029 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1030 {
1031         return atomic_long_read(&cinfo->mds->ncommit);
1032 }
1033
1034 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1035 int
1036 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1037                      struct nfs_commit_info *cinfo, int max)
1038 {
1039         struct nfs_page *req, *tmp;
1040         int ret = 0;
1041
1042         list_for_each_entry_safe(req, tmp, src, wb_list) {
1043                 kref_get(&req->wb_kref);
1044                 if (!nfs_lock_request(req)) {
1045                         nfs_release_request(req);
1046                         continue;
1047                 }
1048                 nfs_request_remove_commit_list(req, cinfo);
1049                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1050                 nfs_list_add_request(req, dst);
1051                 ret++;
1052                 if ((ret == max) && !cinfo->dreq)
1053                         break;
1054                 cond_resched();
1055         }
1056         return ret;
1057 }
1058 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1059
1060 /*
1061  * nfs_scan_commit - Scan an inode for commit requests
1062  * @inode: NFS inode to scan
1063  * @dst: mds destination list
1064  * @cinfo: mds and ds lists of reqs ready to commit
1065  *
1066  * Moves requests from the inode's 'commit' request list.
1067  * The requests are *not* checked to ensure that they form a contiguous set.
1068  */
1069 int
1070 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1071                 struct nfs_commit_info *cinfo)
1072 {
1073         int ret = 0;
1074
1075         if (!atomic_long_read(&cinfo->mds->ncommit))
1076                 return 0;
1077         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1078         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1079                 const int max = INT_MAX;
1080
1081                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1082                                            cinfo, max);
1083                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1084         }
1085         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1086         return ret;
1087 }
1088
1089 /*
1090  * Search for an existing write request, and attempt to update
1091  * it to reflect a new dirty region on a given page.
1092  *
1093  * If the attempt fails, then the existing request is flushed out
1094  * to disk.
1095  */
1096 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1097                 struct page *page,
1098                 unsigned int offset,
1099                 unsigned int bytes)
1100 {
1101         struct nfs_page *req;
1102         unsigned int rqend;
1103         unsigned int end;
1104         int error;
1105
1106         end = offset + bytes;
1107
1108         req = nfs_lock_and_join_requests(page);
1109         if (IS_ERR_OR_NULL(req))
1110                 return req;
1111
1112         rqend = req->wb_offset + req->wb_bytes;
1113         /*
1114          * Tell the caller to flush out the request if
1115          * the offsets are non-contiguous.
1116          * Note: nfs_flush_incompatible() will already
1117          * have flushed out requests having wrong owners.
1118          */
1119         if (offset > rqend || end < req->wb_offset)
1120                 goto out_flushme;
1121
1122         /* Okay, the request matches. Update the region */
1123         if (offset < req->wb_offset) {
1124                 req->wb_offset = offset;
1125                 req->wb_pgbase = offset;
1126         }
1127         if (end > rqend)
1128                 req->wb_bytes = end - req->wb_offset;
1129         else
1130                 req->wb_bytes = rqend - req->wb_offset;
1131         req->wb_nio = 0;
1132         return req;
1133 out_flushme:
1134         /*
1135          * Note: we mark the request dirty here because
1136          * nfs_lock_and_join_requests() cannot preserve
1137          * commit flags, so we have to replay the write.
1138          */
1139         nfs_mark_request_dirty(req);
1140         nfs_unlock_and_release_request(req);
1141         error = nfs_wb_page(inode, page);
1142         return (error < 0) ? ERR_PTR(error) : NULL;
1143 }
1144
1145 /*
1146  * Try to update an existing write request, or create one if there is none.
1147  *
1148  * Note: Should always be called with the Page Lock held to prevent races
1149  * if we have to add a new request. Also assumes that the caller has
1150  * already called nfs_flush_incompatible() if necessary.
1151  */
1152 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1153                 struct page *page, unsigned int offset, unsigned int bytes)
1154 {
1155         struct inode *inode = page_file_mapping(page)->host;
1156         struct nfs_page *req;
1157
1158         req = nfs_try_to_update_request(inode, page, offset, bytes);
1159         if (req != NULL)
1160                 goto out;
1161         req = nfs_create_request(ctx, page, offset, bytes);
1162         if (IS_ERR(req))
1163                 goto out;
1164         nfs_inode_add_request(inode, req);
1165 out:
1166         return req;
1167 }
1168
1169 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1170                 unsigned int offset, unsigned int count)
1171 {
1172         struct nfs_page *req;
1173
1174         req = nfs_setup_write_request(ctx, page, offset, count);
1175         if (IS_ERR(req))
1176                 return PTR_ERR(req);
1177         /* Update file length */
1178         nfs_grow_file(page, offset, count);
1179         nfs_mark_uptodate(req);
1180         nfs_mark_request_dirty(req);
1181         nfs_unlock_and_release_request(req);
1182         return 0;
1183 }
1184
1185 int nfs_flush_incompatible(struct file *file, struct page *page)
1186 {
1187         struct nfs_open_context *ctx = nfs_file_open_context(file);
1188         struct nfs_lock_context *l_ctx;
1189         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1190         struct nfs_page *req;
1191         int do_flush, status;
1192         /*
1193          * Look for a request corresponding to this page. If there
1194          * is one, and it belongs to another file, we flush it out
1195          * before we try to copy anything into the page. Do this
1196          * due to the lack of an ACCESS-type call in NFSv2.
1197          * Also do the same if we find a request from an existing
1198          * dropped page.
1199          */
1200         do {
1201                 req = nfs_page_find_head_request(page);
1202                 if (req == NULL)
1203                         return 0;
1204                 l_ctx = req->wb_lock_context;
1205                 do_flush = req->wb_page != page ||
1206                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1207                 if (l_ctx && flctx &&
1208                     !(list_empty_careful(&flctx->flc_posix) &&
1209                       list_empty_careful(&flctx->flc_flock))) {
1210                         do_flush |= l_ctx->lockowner != current->files;
1211                 }
1212                 nfs_release_request(req);
1213                 if (!do_flush)
1214                         return 0;
1215                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1216         } while (status == 0);
1217         return status;
1218 }
1219
1220 /*
1221  * Avoid buffered writes when a open context credential's key would
1222  * expire soon.
1223  *
1224  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1225  *
1226  * Return 0 and set a credential flag which triggers the inode to flush
1227  * and performs  NFS_FILE_SYNC writes if the key will expired within
1228  * RPC_KEY_EXPIRE_TIMEO.
1229  */
1230 int
1231 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1232 {
1233         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1234
1235         if (nfs_ctx_key_to_expire(ctx, inode) &&
1236             !rcu_access_pointer(ctx->ll_cred))
1237                 /* Already expired! */
1238                 return -EACCES;
1239         return 0;
1240 }
1241
1242 /*
1243  * Test if the open context credential key is marked to expire soon.
1244  */
1245 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1246 {
1247         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1248         struct rpc_cred *cred, *new, *old = NULL;
1249         struct auth_cred acred = {
1250                 .cred = ctx->cred,
1251         };
1252         bool ret = false;
1253
1254         rcu_read_lock();
1255         cred = rcu_dereference(ctx->ll_cred);
1256         if (cred && !(cred->cr_ops->crkey_timeout &&
1257                       cred->cr_ops->crkey_timeout(cred)))
1258                 goto out;
1259         rcu_read_unlock();
1260
1261         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1262         if (new == cred) {
1263                 put_rpccred(new);
1264                 return true;
1265         }
1266         if (IS_ERR_OR_NULL(new)) {
1267                 new = NULL;
1268                 ret = true;
1269         } else if (new->cr_ops->crkey_timeout &&
1270                    new->cr_ops->crkey_timeout(new))
1271                 ret = true;
1272
1273         rcu_read_lock();
1274         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1275                                              RCU_INITIALIZER(new)), 1);
1276 out:
1277         rcu_read_unlock();
1278         put_rpccred(old);
1279         return ret;
1280 }
1281
1282 /*
1283  * If the page cache is marked as unsafe or invalid, then we can't rely on
1284  * the PageUptodate() flag. In this case, we will need to turn off
1285  * write optimisations that depend on the page contents being correct.
1286  */
1287 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1288                                    unsigned int pagelen)
1289 {
1290         struct nfs_inode *nfsi = NFS_I(inode);
1291
1292         if (nfs_have_delegated_attributes(inode))
1293                 goto out;
1294         if (nfsi->cache_validity &
1295             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1296                 return false;
1297         smp_rmb();
1298         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1299                 return false;
1300 out:
1301         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1302                 return false;
1303         return PageUptodate(page) != 0;
1304 }
1305
1306 static bool
1307 is_whole_file_wrlock(struct file_lock *fl)
1308 {
1309         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1310                         fl->fl_type == F_WRLCK;
1311 }
1312
1313 /* If we know the page is up to date, and we're not using byte range locks (or
1314  * if we have the whole file locked for writing), it may be more efficient to
1315  * extend the write to cover the entire page in order to avoid fragmentation
1316  * inefficiencies.
1317  *
1318  * If the file is opened for synchronous writes then we can just skip the rest
1319  * of the checks.
1320  */
1321 static int nfs_can_extend_write(struct file *file, struct page *page,
1322                                 struct inode *inode, unsigned int pagelen)
1323 {
1324         int ret;
1325         struct file_lock_context *flctx = inode->i_flctx;
1326         struct file_lock *fl;
1327
1328         if (file->f_flags & O_DSYNC)
1329                 return 0;
1330         if (!nfs_write_pageuptodate(page, inode, pagelen))
1331                 return 0;
1332         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1333                 return 1;
1334         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1335                        list_empty_careful(&flctx->flc_posix)))
1336                 return 1;
1337
1338         /* Check to see if there are whole file write locks */
1339         ret = 0;
1340         spin_lock(&flctx->flc_lock);
1341         if (!list_empty(&flctx->flc_posix)) {
1342                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1343                                         fl_list);
1344                 if (is_whole_file_wrlock(fl))
1345                         ret = 1;
1346         } else if (!list_empty(&flctx->flc_flock)) {
1347                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1348                                         fl_list);
1349                 if (fl->fl_type == F_WRLCK)
1350                         ret = 1;
1351         }
1352         spin_unlock(&flctx->flc_lock);
1353         return ret;
1354 }
1355
1356 /*
1357  * Update and possibly write a cached page of an NFS file.
1358  *
1359  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1360  * things with a page scheduled for an RPC call (e.g. invalidate it).
1361  */
1362 int nfs_updatepage(struct file *file, struct page *page,
1363                 unsigned int offset, unsigned int count)
1364 {
1365         struct nfs_open_context *ctx = nfs_file_open_context(file);
1366         struct address_space *mapping = page_file_mapping(page);
1367         struct inode    *inode = mapping->host;
1368         unsigned int    pagelen = nfs_page_length(page);
1369         int             status = 0;
1370
1371         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1372
1373         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1374                 file, count, (long long)(page_file_offset(page) + offset));
1375
1376         if (!count)
1377                 goto out;
1378
1379         if (nfs_can_extend_write(file, page, inode, pagelen)) {
1380                 count = max(count + offset, pagelen);
1381                 offset = 0;
1382         }
1383
1384         status = nfs_writepage_setup(ctx, page, offset, count);
1385         if (status < 0)
1386                 nfs_set_pageerror(mapping);
1387 out:
1388         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1389                         status, (long long)i_size_read(inode));
1390         return status;
1391 }
1392
1393 static int flush_task_priority(int how)
1394 {
1395         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1396                 case FLUSH_HIGHPRI:
1397                         return RPC_PRIORITY_HIGH;
1398                 case FLUSH_LOWPRI:
1399                         return RPC_PRIORITY_LOW;
1400         }
1401         return RPC_PRIORITY_NORMAL;
1402 }
1403
1404 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1405                                struct rpc_message *msg,
1406                                const struct nfs_rpc_ops *rpc_ops,
1407                                struct rpc_task_setup *task_setup_data, int how)
1408 {
1409         int priority = flush_task_priority(how);
1410
1411         task_setup_data->priority = priority;
1412         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1413         trace_nfs_initiate_write(hdr);
1414 }
1415
1416 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1417  * call this on each, which will prepare them to be retried on next
1418  * writeback using standard nfs.
1419  */
1420 static void nfs_redirty_request(struct nfs_page *req)
1421 {
1422         /* Bump the transmission count */
1423         req->wb_nio++;
1424         nfs_mark_request_dirty(req);
1425         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1426         nfs_end_page_writeback(req);
1427         nfs_release_request(req);
1428 }
1429
1430 static void nfs_async_write_error(struct list_head *head, int error)
1431 {
1432         struct nfs_page *req;
1433
1434         while (!list_empty(head)) {
1435                 req = nfs_list_entry(head->next);
1436                 nfs_list_remove_request(req);
1437                 if (nfs_error_is_fatal(error))
1438                         nfs_write_error(req, error);
1439                 else
1440                         nfs_redirty_request(req);
1441         }
1442 }
1443
1444 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1445 {
1446         nfs_async_write_error(&hdr->pages, 0);
1447         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1448                         hdr->args.offset + hdr->args.count - 1);
1449 }
1450
1451 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1452         .init_hdr = nfs_async_write_init,
1453         .error_cleanup = nfs_async_write_error,
1454         .completion = nfs_write_completion,
1455         .reschedule_io = nfs_async_write_reschedule_io,
1456 };
1457
1458 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1459                                struct inode *inode, int ioflags, bool force_mds,
1460                                const struct nfs_pgio_completion_ops *compl_ops)
1461 {
1462         struct nfs_server *server = NFS_SERVER(inode);
1463         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1464
1465 #ifdef CONFIG_NFS_V4_1
1466         if (server->pnfs_curr_ld && !force_mds)
1467                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1468 #endif
1469         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1470                         server->wsize, ioflags);
1471 }
1472 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1473
1474 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1475 {
1476         struct nfs_pgio_mirror *mirror;
1477
1478         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1479                 pgio->pg_ops->pg_cleanup(pgio);
1480
1481         pgio->pg_ops = &nfs_pgio_rw_ops;
1482
1483         nfs_pageio_stop_mirroring(pgio);
1484
1485         mirror = &pgio->pg_mirrors[0];
1486         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1487 }
1488 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1489
1490
1491 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1492 {
1493         struct nfs_commit_data *data = calldata;
1494
1495         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1496 }
1497
1498 /*
1499  * Special version of should_remove_suid() that ignores capabilities.
1500  */
1501 static int nfs_should_remove_suid(const struct inode *inode)
1502 {
1503         umode_t mode = inode->i_mode;
1504         int kill = 0;
1505
1506         /* suid always must be killed */
1507         if (unlikely(mode & S_ISUID))
1508                 kill = ATTR_KILL_SUID;
1509
1510         /*
1511          * sgid without any exec bits is just a mandatory locking mark; leave
1512          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1513          */
1514         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1515                 kill |= ATTR_KILL_SGID;
1516
1517         if (unlikely(kill && S_ISREG(mode)))
1518                 return kill;
1519
1520         return 0;
1521 }
1522
1523 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1524                 struct nfs_fattr *fattr)
1525 {
1526         struct nfs_pgio_args *argp = &hdr->args;
1527         struct nfs_pgio_res *resp = &hdr->res;
1528         u64 size = argp->offset + resp->count;
1529
1530         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1531                 fattr->size = size;
1532         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1533                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1534                 return;
1535         }
1536         if (size != fattr->size)
1537                 return;
1538         /* Set attribute barrier */
1539         nfs_fattr_set_barrier(fattr);
1540         /* ...and update size */
1541         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1542 }
1543
1544 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1545 {
1546         struct nfs_fattr *fattr = &hdr->fattr;
1547         struct inode *inode = hdr->inode;
1548
1549         spin_lock(&inode->i_lock);
1550         nfs_writeback_check_extend(hdr, fattr);
1551         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1552         spin_unlock(&inode->i_lock);
1553 }
1554 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1555
1556 /*
1557  * This function is called when the WRITE call is complete.
1558  */
1559 static int nfs_writeback_done(struct rpc_task *task,
1560                               struct nfs_pgio_header *hdr,
1561                               struct inode *inode)
1562 {
1563         int status;
1564
1565         /*
1566          * ->write_done will attempt to use post-op attributes to detect
1567          * conflicting writes by other clients.  A strict interpretation
1568          * of close-to-open would allow us to continue caching even if
1569          * another writer had changed the file, but some applications
1570          * depend on tighter cache coherency when writing.
1571          */
1572         status = NFS_PROTO(inode)->write_done(task, hdr);
1573         if (status != 0)
1574                 return status;
1575
1576         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1577         trace_nfs_writeback_done(task, hdr);
1578
1579         if (hdr->res.verf->committed < hdr->args.stable &&
1580             task->tk_status >= 0) {
1581                 /* We tried a write call, but the server did not
1582                  * commit data to stable storage even though we
1583                  * requested it.
1584                  * Note: There is a known bug in Tru64 < 5.0 in which
1585                  *       the server reports NFS_DATA_SYNC, but performs
1586                  *       NFS_FILE_SYNC. We therefore implement this checking
1587                  *       as a dprintk() in order to avoid filling syslog.
1588                  */
1589                 static unsigned long    complain;
1590
1591                 /* Note this will print the MDS for a DS write */
1592                 if (time_before(complain, jiffies)) {
1593                         dprintk("NFS:       faulty NFS server %s:"
1594                                 " (committed = %d) != (stable = %d)\n",
1595                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1596                                 hdr->res.verf->committed, hdr->args.stable);
1597                         complain = jiffies + 300 * HZ;
1598                 }
1599         }
1600
1601         /* Deal with the suid/sgid bit corner case */
1602         if (nfs_should_remove_suid(inode)) {
1603                 spin_lock(&inode->i_lock);
1604                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1605                 spin_unlock(&inode->i_lock);
1606         }
1607         return 0;
1608 }
1609
1610 /*
1611  * This function is called when the WRITE call is complete.
1612  */
1613 static void nfs_writeback_result(struct rpc_task *task,
1614                                  struct nfs_pgio_header *hdr)
1615 {
1616         struct nfs_pgio_args    *argp = &hdr->args;
1617         struct nfs_pgio_res     *resp = &hdr->res;
1618
1619         if (resp->count < argp->count) {
1620                 static unsigned long    complain;
1621
1622                 /* This a short write! */
1623                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1624
1625                 /* Has the server at least made some progress? */
1626                 if (resp->count == 0) {
1627                         if (time_before(complain, jiffies)) {
1628                                 printk(KERN_WARNING
1629                                        "NFS: Server wrote zero bytes, expected %u.\n",
1630                                        argp->count);
1631                                 complain = jiffies + 300 * HZ;
1632                         }
1633                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1634                         task->tk_status = -EIO;
1635                         return;
1636                 }
1637
1638                 /* For non rpc-based layout drivers, retry-through-MDS */
1639                 if (!task->tk_ops) {
1640                         hdr->pnfs_error = -EAGAIN;
1641                         return;
1642                 }
1643
1644                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1645                 if (resp->verf->committed != NFS_UNSTABLE) {
1646                         /* Resend from where the server left off */
1647                         hdr->mds_offset += resp->count;
1648                         argp->offset += resp->count;
1649                         argp->pgbase += resp->count;
1650                         argp->count -= resp->count;
1651                 } else {
1652                         /* Resend as a stable write in order to avoid
1653                          * headaches in the case of a server crash.
1654                          */
1655                         argp->stable = NFS_FILE_SYNC;
1656                 }
1657                 resp->count = 0;
1658                 resp->verf->committed = 0;
1659                 rpc_restart_call_prepare(task);
1660         }
1661 }
1662
1663 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1664 {
1665         return wait_var_event_killable(&cinfo->rpcs_out,
1666                                        !atomic_read(&cinfo->rpcs_out));
1667 }
1668
1669 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1670 {
1671         atomic_inc(&cinfo->rpcs_out);
1672 }
1673
1674 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1675 {
1676         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1677                 wake_up_var(&cinfo->rpcs_out);
1678                 return true;
1679         }
1680         return false;
1681 }
1682
1683 void nfs_commitdata_release(struct nfs_commit_data *data)
1684 {
1685         put_nfs_open_context(data->context);
1686         nfs_commit_free(data);
1687 }
1688 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1689
1690 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1691                         const struct nfs_rpc_ops *nfs_ops,
1692                         const struct rpc_call_ops *call_ops,
1693                         int how, int flags)
1694 {
1695         struct rpc_task *task;
1696         int priority = flush_task_priority(how);
1697         struct rpc_message msg = {
1698                 .rpc_argp = &data->args,
1699                 .rpc_resp = &data->res,
1700                 .rpc_cred = data->cred,
1701         };
1702         struct rpc_task_setup task_setup_data = {
1703                 .task = &data->task,
1704                 .rpc_client = clnt,
1705                 .rpc_message = &msg,
1706                 .callback_ops = call_ops,
1707                 .callback_data = data,
1708                 .workqueue = nfsiod_workqueue,
1709                 .flags = RPC_TASK_ASYNC | flags,
1710                 .priority = priority,
1711         };
1712         /* Set up the initial task struct.  */
1713         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1714         trace_nfs_initiate_commit(data);
1715
1716         dprintk("NFS: initiated commit call\n");
1717
1718         task = rpc_run_task(&task_setup_data);
1719         if (IS_ERR(task))
1720                 return PTR_ERR(task);
1721         if (how & FLUSH_SYNC)
1722                 rpc_wait_for_completion_task(task);
1723         rpc_put_task(task);
1724         return 0;
1725 }
1726 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1727
1728 static loff_t nfs_get_lwb(struct list_head *head)
1729 {
1730         loff_t lwb = 0;
1731         struct nfs_page *req;
1732
1733         list_for_each_entry(req, head, wb_list)
1734                 if (lwb < (req_offset(req) + req->wb_bytes))
1735                         lwb = req_offset(req) + req->wb_bytes;
1736
1737         return lwb;
1738 }
1739
1740 /*
1741  * Set up the argument/result storage required for the RPC call.
1742  */
1743 void nfs_init_commit(struct nfs_commit_data *data,
1744                      struct list_head *head,
1745                      struct pnfs_layout_segment *lseg,
1746                      struct nfs_commit_info *cinfo)
1747 {
1748         struct nfs_page *first;
1749         struct nfs_open_context *ctx;
1750         struct inode *inode;
1751
1752         /* Set up the RPC argument and reply structs
1753          * NB: take care not to mess about with data->commit et al. */
1754
1755         if (head)
1756                 list_splice_init(head, &data->pages);
1757
1758         first = nfs_list_entry(data->pages.next);
1759         ctx = nfs_req_openctx(first);
1760         inode = d_inode(ctx->dentry);
1761
1762         data->inode       = inode;
1763         data->cred        = ctx->cred;
1764         data->lseg        = lseg; /* reference transferred */
1765         /* only set lwb for pnfs commit */
1766         if (lseg)
1767                 data->lwb = nfs_get_lwb(&data->pages);
1768         data->mds_ops     = &nfs_commit_ops;
1769         data->completion_ops = cinfo->completion_ops;
1770         data->dreq        = cinfo->dreq;
1771
1772         data->args.fh     = NFS_FH(data->inode);
1773         /* Note: we always request a commit of the entire inode */
1774         data->args.offset = 0;
1775         data->args.count  = 0;
1776         data->context     = get_nfs_open_context(ctx);
1777         data->res.fattr   = &data->fattr;
1778         data->res.verf    = &data->verf;
1779         nfs_fattr_init(&data->fattr);
1780         nfs_commit_begin(cinfo->mds);
1781 }
1782 EXPORT_SYMBOL_GPL(nfs_init_commit);
1783
1784 void nfs_retry_commit(struct list_head *page_list,
1785                       struct pnfs_layout_segment *lseg,
1786                       struct nfs_commit_info *cinfo,
1787                       u32 ds_commit_idx)
1788 {
1789         struct nfs_page *req;
1790
1791         while (!list_empty(page_list)) {
1792                 req = nfs_list_entry(page_list->next);
1793                 nfs_list_remove_request(req);
1794                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1795                 if (!cinfo->dreq)
1796                         nfs_clear_page_commit(req->wb_page);
1797                 nfs_unlock_and_release_request(req);
1798         }
1799 }
1800 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1801
1802 static void
1803 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1804                 struct nfs_page *req)
1805 {
1806         __set_page_dirty_nobuffers(req->wb_page);
1807 }
1808
1809 /*
1810  * Commit dirty pages
1811  */
1812 static int
1813 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1814                 struct nfs_commit_info *cinfo)
1815 {
1816         struct nfs_commit_data  *data;
1817         unsigned short task_flags = 0;
1818
1819         /* another commit raced with us */
1820         if (list_empty(head))
1821                 return 0;
1822
1823         data = nfs_commitdata_alloc(true);
1824
1825         /* Set up the argument struct */
1826         nfs_init_commit(data, head, NULL, cinfo);
1827         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1828                 task_flags = RPC_TASK_MOVEABLE;
1829         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1830                                    data->mds_ops, how,
1831                                    RPC_TASK_CRED_NOREF | task_flags);
1832 }
1833
1834 /*
1835  * COMMIT call returned
1836  */
1837 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1838 {
1839         struct nfs_commit_data  *data = calldata;
1840
1841         /* Call the NFS version-specific code */
1842         NFS_PROTO(data->inode)->commit_done(task, data);
1843         trace_nfs_commit_done(task, data);
1844 }
1845
1846 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1847 {
1848         const struct nfs_writeverf *verf = data->res.verf;
1849         struct nfs_page *req;
1850         int status = data->task.tk_status;
1851         struct nfs_commit_info cinfo;
1852         struct nfs_server *nfss;
1853
1854         while (!list_empty(&data->pages)) {
1855                 req = nfs_list_entry(data->pages.next);
1856                 nfs_list_remove_request(req);
1857                 if (req->wb_page)
1858                         nfs_clear_page_commit(req->wb_page);
1859
1860                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1861                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1862                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1863                         req->wb_bytes,
1864                         (long long)req_offset(req));
1865                 if (status < 0) {
1866                         if (req->wb_page) {
1867                                 trace_nfs_commit_error(req, status);
1868                                 nfs_mapping_set_error(req->wb_page, status);
1869                                 nfs_inode_remove_request(req);
1870                         }
1871                         dprintk_cont(", error = %d\n", status);
1872                         goto next;
1873                 }
1874
1875                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1876                  * returned by the server against all stored verfs. */
1877                 if (nfs_write_match_verf(verf, req)) {
1878                         /* We have a match */
1879                         if (req->wb_page)
1880                                 nfs_inode_remove_request(req);
1881                         dprintk_cont(" OK\n");
1882                         goto next;
1883                 }
1884                 /* We have a mismatch. Write the page again */
1885                 dprintk_cont(" mismatch\n");
1886                 nfs_mark_request_dirty(req);
1887                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1888         next:
1889                 nfs_unlock_and_release_request(req);
1890                 /* Latency breaker */
1891                 cond_resched();
1892         }
1893         nfss = NFS_SERVER(data->inode);
1894         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1895                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1896
1897         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1898         nfs_commit_end(cinfo.mds);
1899 }
1900
1901 static void nfs_commit_release(void *calldata)
1902 {
1903         struct nfs_commit_data *data = calldata;
1904
1905         data->completion_ops->completion(data);
1906         nfs_commitdata_release(calldata);
1907 }
1908
1909 static const struct rpc_call_ops nfs_commit_ops = {
1910         .rpc_call_prepare = nfs_commit_prepare,
1911         .rpc_call_done = nfs_commit_done,
1912         .rpc_release = nfs_commit_release,
1913 };
1914
1915 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1916         .completion = nfs_commit_release_pages,
1917         .resched_write = nfs_commit_resched_write,
1918 };
1919
1920 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1921                             int how, struct nfs_commit_info *cinfo)
1922 {
1923         int status;
1924
1925         status = pnfs_commit_list(inode, head, how, cinfo);
1926         if (status == PNFS_NOT_ATTEMPTED)
1927                 status = nfs_commit_list(inode, head, how, cinfo);
1928         return status;
1929 }
1930
1931 static int __nfs_commit_inode(struct inode *inode, int how,
1932                 struct writeback_control *wbc)
1933 {
1934         LIST_HEAD(head);
1935         struct nfs_commit_info cinfo;
1936         int may_wait = how & FLUSH_SYNC;
1937         int ret, nscan;
1938
1939         how &= ~FLUSH_SYNC;
1940         nfs_init_cinfo_from_inode(&cinfo, inode);
1941         nfs_commit_begin(cinfo.mds);
1942         for (;;) {
1943                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1944                 if (ret <= 0)
1945                         break;
1946                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1947                 if (ret < 0)
1948                         break;
1949                 ret = 0;
1950                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1951                         if (nscan < wbc->nr_to_write)
1952                                 wbc->nr_to_write -= nscan;
1953                         else
1954                                 wbc->nr_to_write = 0;
1955                 }
1956                 if (nscan < INT_MAX)
1957                         break;
1958                 cond_resched();
1959         }
1960         nfs_commit_end(cinfo.mds);
1961         if (ret || !may_wait)
1962                 return ret;
1963         return wait_on_commit(cinfo.mds);
1964 }
1965
1966 int nfs_commit_inode(struct inode *inode, int how)
1967 {
1968         return __nfs_commit_inode(inode, how, NULL);
1969 }
1970 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1971
1972 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1973 {
1974         struct nfs_inode *nfsi = NFS_I(inode);
1975         int flags = FLUSH_SYNC;
1976         int ret = 0;
1977
1978         if (wbc->sync_mode == WB_SYNC_NONE) {
1979                 /* no commits means nothing needs to be done */
1980                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1981                         goto check_requests_outstanding;
1982
1983                 /* Don't commit yet if this is a non-blocking flush and there
1984                  * are a lot of outstanding writes for this mapping.
1985                  */
1986                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1987                         goto out_mark_dirty;
1988
1989                 /* don't wait for the COMMIT response */
1990                 flags = 0;
1991         }
1992
1993         ret = __nfs_commit_inode(inode, flags, wbc);
1994         if (!ret) {
1995                 if (flags & FLUSH_SYNC)
1996                         return 0;
1997         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1998                 goto out_mark_dirty;
1999
2000 check_requests_outstanding:
2001         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2002                 return ret;
2003 out_mark_dirty:
2004         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2005         return ret;
2006 }
2007 EXPORT_SYMBOL_GPL(nfs_write_inode);
2008
2009 /*
2010  * Wrapper for filemap_write_and_wait_range()
2011  *
2012  * Needed for pNFS in order to ensure data becomes visible to the
2013  * client.
2014  */
2015 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2016                 loff_t lstart, loff_t lend)
2017 {
2018         int ret;
2019
2020         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2021         if (ret == 0)
2022                 ret = pnfs_sync_inode(mapping->host, true);
2023         return ret;
2024 }
2025 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2026
2027 /*
2028  * flush the inode to disk.
2029  */
2030 int nfs_wb_all(struct inode *inode)
2031 {
2032         int ret;
2033
2034         trace_nfs_writeback_inode_enter(inode);
2035
2036         ret = filemap_write_and_wait(inode->i_mapping);
2037         if (ret)
2038                 goto out;
2039         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2040         if (ret < 0)
2041                 goto out;
2042         pnfs_sync_inode(inode, true);
2043         ret = 0;
2044
2045 out:
2046         trace_nfs_writeback_inode_exit(inode, ret);
2047         return ret;
2048 }
2049 EXPORT_SYMBOL_GPL(nfs_wb_all);
2050
2051 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2052 {
2053         struct nfs_page *req;
2054         int ret = 0;
2055
2056         wait_on_page_writeback(page);
2057
2058         /* blocking call to cancel all requests and join to a single (head)
2059          * request */
2060         req = nfs_lock_and_join_requests(page);
2061
2062         if (IS_ERR(req)) {
2063                 ret = PTR_ERR(req);
2064         } else if (req) {
2065                 /* all requests from this page have been cancelled by
2066                  * nfs_lock_and_join_requests, so just remove the head
2067                  * request from the inode / page_private pointer and
2068                  * release it */
2069                 nfs_inode_remove_request(req);
2070                 nfs_unlock_and_release_request(req);
2071         }
2072
2073         return ret;
2074 }
2075
2076 /*
2077  * Write back all requests on one page - we do this before reading it.
2078  */
2079 int nfs_wb_page(struct inode *inode, struct page *page)
2080 {
2081         loff_t range_start = page_file_offset(page);
2082         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2083         struct writeback_control wbc = {
2084                 .sync_mode = WB_SYNC_ALL,
2085                 .nr_to_write = 0,
2086                 .range_start = range_start,
2087                 .range_end = range_end,
2088         };
2089         int ret;
2090
2091         trace_nfs_writeback_page_enter(inode);
2092
2093         for (;;) {
2094                 wait_on_page_writeback(page);
2095                 if (clear_page_dirty_for_io(page)) {
2096                         ret = nfs_writepage_locked(page, &wbc);
2097                         if (ret < 0)
2098                                 goto out_error;
2099                         continue;
2100                 }
2101                 ret = 0;
2102                 if (!PagePrivate(page))
2103                         break;
2104                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2105                 if (ret < 0)
2106                         goto out_error;
2107         }
2108 out_error:
2109         trace_nfs_writeback_page_exit(inode, ret);
2110         return ret;
2111 }
2112
2113 #ifdef CONFIG_MIGRATION
2114 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2115                 struct page *page, enum migrate_mode mode)
2116 {
2117         /*
2118          * If PagePrivate is set, then the page is currently associated with
2119          * an in-progress read or write request. Don't try to migrate it.
2120          *
2121          * FIXME: we could do this in principle, but we'll need a way to ensure
2122          *        that we can safely release the inode reference while holding
2123          *        the page lock.
2124          */
2125         if (PagePrivate(page))
2126                 return -EBUSY;
2127
2128         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2129                 return -EBUSY;
2130
2131         return migrate_page(mapping, newpage, page, mode);
2132 }
2133 #endif
2134
2135 int __init nfs_init_writepagecache(void)
2136 {
2137         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2138                                              sizeof(struct nfs_pgio_header),
2139                                              0, SLAB_HWCACHE_ALIGN,
2140                                              NULL);
2141         if (nfs_wdata_cachep == NULL)
2142                 return -ENOMEM;
2143
2144         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2145                                                      nfs_wdata_cachep);
2146         if (nfs_wdata_mempool == NULL)
2147                 goto out_destroy_write_cache;
2148
2149         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2150                                              sizeof(struct nfs_commit_data),
2151                                              0, SLAB_HWCACHE_ALIGN,
2152                                              NULL);
2153         if (nfs_cdata_cachep == NULL)
2154                 goto out_destroy_write_mempool;
2155
2156         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2157                                                       nfs_cdata_cachep);
2158         if (nfs_commit_mempool == NULL)
2159                 goto out_destroy_commit_cache;
2160
2161         /*
2162          * NFS congestion size, scale with available memory.
2163          *
2164          *  64MB:    8192k
2165          * 128MB:   11585k
2166          * 256MB:   16384k
2167          * 512MB:   23170k
2168          *   1GB:   32768k
2169          *   2GB:   46340k
2170          *   4GB:   65536k
2171          *   8GB:   92681k
2172          *  16GB:  131072k
2173          *
2174          * This allows larger machines to have larger/more transfers.
2175          * Limit the default to 256M
2176          */
2177         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2178         if (nfs_congestion_kb > 256*1024)
2179                 nfs_congestion_kb = 256*1024;
2180
2181         return 0;
2182
2183 out_destroy_commit_cache:
2184         kmem_cache_destroy(nfs_cdata_cachep);
2185 out_destroy_write_mempool:
2186         mempool_destroy(nfs_wdata_mempool);
2187 out_destroy_write_cache:
2188         kmem_cache_destroy(nfs_wdata_cachep);
2189         return -ENOMEM;
2190 }
2191
2192 void nfs_destroy_writepagecache(void)
2193 {
2194         mempool_destroy(nfs_commit_mempool);
2195         kmem_cache_destroy(nfs_cdata_cachep);
2196         mempool_destroy(nfs_wdata_mempool);
2197         kmem_cache_destroy(nfs_wdata_cachep);
2198 }
2199
2200 static const struct nfs_rw_ops nfs_rw_write_ops = {
2201         .rw_alloc_header        = nfs_writehdr_alloc,
2202         .rw_free_header         = nfs_writehdr_free,
2203         .rw_done                = nfs_writeback_done,
2204         .rw_result              = nfs_writeback_result,
2205         .rw_initiate            = nfs_initiate_write,
2206 };