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         nfs_fscache_invalidate(inode, 0);
 298 }
 299
 300 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 301 static void nfs_set_pageerror(struct address_space *mapping)
 302 {
 303         struct inode *inode = mapping->host;
 304
 305         nfs_zap_mapping(mapping->host, mapping);
 306         /* Force file size revalidation */
 307         spin_lock(&inode->i_lock);
 308         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 309                                              NFS_INO_REVAL_PAGECACHE |
 310                                              NFS_INO_INVALID_SIZE);
 311         spin_unlock(&inode->i_lock);
 312 }
 313
 314 static void nfs_mapping_set_error(struct page *page, int error)
 315 {
 316         struct address_space *mapping = page_file_mapping(page);
 317
 318         SetPageError(page);
 319         mapping_set_error(mapping, error);
 320         nfs_set_pageerror(mapping);
 321 }
 322
 323 /*
 324  * nfs_page_group_search_locked
 325  * @head - head request of page group
 326  * @page_offset - offset into page
 327  *
 328  * Search page group with head @head to find a request that contains the
 329  * page offset @page_offset.
 330  *
 331  * Returns a pointer to the first matching nfs request, or NULL if no
 332  * match is found.
 333  *
 334  * Must be called with the page group lock held
 335  */
 336 static struct nfs_page *
 337 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 338 {
 339         struct nfs_page *req;
 340
 341         req = head;
 342         do {
 343                 if (page_offset >= req->wb_pgbase &&
 344                     page_offset < (req->wb_pgbase + req->wb_bytes))
 345                         return req;
 346
 347                 req = req->wb_this_page;
 348         } while (req != head);
 349
 350         return NULL;
 351 }
 352
 353 /*
 354  * nfs_page_group_covers_page
 355  * @head - head request of page group
 356  *
 357  * Return true if the page group with head @head covers the whole page,
 358  * returns false otherwise
 359  */
 360 static bool nfs_page_group_covers_page(struct nfs_page *req)
 361 {
 362         struct nfs_page *tmp;
 363         unsigned int pos = 0;
 364         unsigned int len = nfs_page_length(req->wb_page);
 365
 366         nfs_page_group_lock(req);
 367
 368         for (;;) {
 369                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 370                 if (!tmp)
 371                         break;
 372                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 373         }
 374
 375         nfs_page_group_unlock(req);
 376         return pos >= len;
 377 }
 378
 379 /* We can set the PG_uptodate flag if we see that a write request
 380  * covers the full page.
 381  */
 382 static void nfs_mark_uptodate(struct nfs_page *req)
 383 {
 384         if (PageUptodate(req->wb_page))
 385                 return;
 386         if (!nfs_page_group_covers_page(req))
 387                 return;
 388         SetPageUptodate(req->wb_page);
 389 }
 390
 391 static int wb_priority(struct writeback_control *wbc)
 392 {
 393         int ret = 0;
 394
 395         if (wbc->sync_mode == WB_SYNC_ALL)
 396                 ret = FLUSH_COND_STABLE;
 397         return ret;
 398 }
 399
 400 /*
 401  * NFS congestion control
 402  */
 403
 404 int nfs_congestion_kb;
 405
 406 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 407 #define NFS_CONGESTION_OFF_THRESH       \
 408         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 409
 410 static void nfs_set_page_writeback(struct page *page)
 411 {
 412         struct inode *inode = page_file_mapping(page)->host;
 413         struct nfs_server *nfss = NFS_SERVER(inode);
 414         int ret = test_set_page_writeback(page);
 415
 416         WARN_ON_ONCE(ret != 0);
 417
 418         if (atomic_long_inc_return(&nfss->writeback) >
 419                         NFS_CONGESTION_ON_THRESH)
 420                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 421 }
 422
 423 static void nfs_end_page_writeback(struct nfs_page *req)
 424 {
 425         struct inode *inode = page_file_mapping(req->wb_page)->host;
 426         struct nfs_server *nfss = NFS_SERVER(inode);
 427         bool is_done;
 428
 429         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 430         nfs_unlock_request(req);
 431         if (!is_done)
 432                 return;
 433
 434         end_page_writeback(req->wb_page);
 435         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 436                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 437 }
 438
 439 /*
 440  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 441  *
 442  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 443  * @old_head - the old head of the list
 444  *
 445  * All subrequests must be locked and removed from all lists, so at this point
 446  * they are only "active" in this function, and possibly in nfs_wait_on_request
 447  * with a reference held by some other context.
 448  */
 449 static void
 450 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 451                                  struct nfs_page *old_head,
 452                                  struct inode *inode)
 453 {
 454         while (destroy_list) {
 455                 struct nfs_page *subreq = destroy_list;
 456
 457                 destroy_list = (subreq->wb_this_page == old_head) ?
 458                                    NULL : subreq->wb_this_page;
 459
 460                 /* Note: lock subreq in order to change subreq->wb_head */
 461                 nfs_page_set_headlock(subreq);
 462                 WARN_ON_ONCE(old_head != subreq->wb_head);
 463
 464                 /* make sure old group is not used */
 465                 subreq->wb_this_page = subreq;
 466                 subreq->wb_head = subreq;
 467
 468                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 469
 470                 /* Note: races with nfs_page_group_destroy() */
 471                 if (!kref_read(&subreq->wb_kref)) {
 472                         /* Check if we raced with nfs_page_group_destroy() */
 473                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 474                                 nfs_page_clear_headlock(subreq);
 475                                 nfs_free_request(subreq);
 476                         } else
 477                                 nfs_page_clear_headlock(subreq);
 478                         continue;
 479                 }
 480                 nfs_page_clear_headlock(subreq);
 481
 482                 nfs_release_request(old_head);
 483
 484                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 485                         nfs_release_request(subreq);
 486                         atomic_long_dec(&NFS_I(inode)->nrequests);
 487                 }
 488
 489                 /* subreq is now totally disconnected from page group or any
 490                  * write / commit lists. last chance to wake any waiters */
 491                 nfs_unlock_and_release_request(subreq);
 492         }
 493 }
 494
 495 /*
 496  * nfs_join_page_group - destroy subrequests of the head req
 497  * @head: the page used to lookup the "page group" of nfs_page structures
 498  * @inode: Inode to which the request belongs.
 499  *
 500  * This function joins all sub requests to the head request by first
 501  * locking all requests in the group, cancelling any pending operations
 502  * and finally updating the head request to cover the whole range covered by
 503  * the (former) group.  All subrequests are removed from any write or commit
 504  * lists, unlinked from the group and destroyed.
 505  */
 506 void
 507 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
 508 {
 509         struct nfs_page *subreq;
 510         struct nfs_page *destroy_list = NULL;
 511         unsigned int pgbase, off, bytes;
 512
 513         pgbase = head->wb_pgbase;
 514         bytes = head->wb_bytes;
 515         off = head->wb_offset;
 516         for (subreq = head->wb_this_page; subreq != head;
 517                         subreq = subreq->wb_this_page) {
 518                 /* Subrequests should always form a contiguous range */
 519                 if (pgbase > subreq->wb_pgbase) {
 520                         off -= pgbase - subreq->wb_pgbase;
 521                         bytes += pgbase - subreq->wb_pgbase;
 522                         pgbase = subreq->wb_pgbase;
 523                 }
 524                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 525                                 - pgbase, bytes);
 526         }
 527
 528         /* Set the head request's range to cover the former page group */
 529         head->wb_pgbase = pgbase;
 530         head->wb_bytes = bytes;
 531         head->wb_offset = off;
 532
 533         /* Now that all requests are locked, make sure they aren't on any list.
 534          * Commit list removal accounting is done after locks are dropped */
 535         subreq = head;
 536         do {
 537                 nfs_clear_request_commit(subreq);
 538                 subreq = subreq->wb_this_page;
 539         } while (subreq != head);
 540
 541         /* unlink subrequests from head, destroy them later */
 542         if (head->wb_this_page != head) {
 543                 /* destroy list will be terminated by head */
 544                 destroy_list = head->wb_this_page;
 545                 head->wb_this_page = head;
 546         }
 547
 548         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 549 }
 550
 551 /*
 552  * nfs_lock_and_join_requests - join all subreqs to the head req
 553  * @page: the page used to lookup the "page group" of nfs_page structures
 554  *
 555  * This function joins all sub requests to the head request by first
 556  * locking all requests in the group, cancelling any pending operations
 557  * and finally updating the head request to cover the whole range covered by
 558  * the (former) group.  All subrequests are removed from any write or commit
 559  * lists, unlinked from the group and destroyed.
 560  *
 561  * Returns a locked, referenced pointer to the head request - which after
 562  * this call is guaranteed to be the only request associated with the page.
 563  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 564  * error was encountered.
 565  */
 566 static struct nfs_page *
 567 nfs_lock_and_join_requests(struct page *page)
 568 {
 569         struct inode *inode = page_file_mapping(page)->host;
 570         struct nfs_page *head;
 571         int ret;
 572
 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, 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 nfs_pageio_descriptor *pgio,
 608                                 struct page *page)
 609 {
 610         struct nfs_page *req;
 611         int ret = 0;
 612
 613         req = nfs_lock_and_join_requests(page);
 614         if (!req)
 615                 goto out;
 616         ret = PTR_ERR(req);
 617         if (IS_ERR(req))
 618                 goto out;
 619
 620         nfs_set_page_writeback(page);
 621         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 622
 623         /* If there is a fatal error that covers this write, just exit */
 624         ret = pgio->pg_error;
 625         if (nfs_error_is_fatal_on_server(ret))
 626                 goto out_launder;
 627
 628         ret = 0;
 629         if (!nfs_pageio_add_request(pgio, req)) {
 630                 ret = pgio->pg_error;
 631                 /*
 632                  * Remove the problematic req upon fatal errors on the server
 633                  */
 634                 if (nfs_error_is_fatal(ret)) {
 635                         if (nfs_error_is_fatal_on_server(ret))
 636                                 goto out_launder;
 637                 } else
 638                         ret = -EAGAIN;
 639                 nfs_redirty_request(req);
 640                 pgio->pg_error = 0;
 641         } else
 642                 nfs_add_stats(page_file_mapping(page)->host,
 643                                 NFSIOS_WRITEPAGES, 1);
 644 out:
 645         return ret;
 646 out_launder:
 647         nfs_write_error(req, ret);
 648         return 0;
 649 }
 650
 651 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 652                             struct nfs_pageio_descriptor *pgio)
 653 {
 654         int ret;
 655
 656         nfs_pageio_cond_complete(pgio, page_index(page));
 657         ret = nfs_page_async_flush(pgio, page);
 658         if (ret == -EAGAIN) {
 659                 redirty_page_for_writepage(wbc, page);
 660                 ret = AOP_WRITEPAGE_ACTIVATE;
 661         }
 662         return ret;
 663 }
 664
 665 /*
 666  * Write an mmapped page to the server.
 667  */
 668 static int nfs_writepage_locked(struct page *page,
 669                                 struct writeback_control *wbc)
 670 {
 671         struct nfs_pageio_descriptor pgio;
 672         struct inode *inode = page_file_mapping(page)->host;
 673         int err;
 674
 675         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 676         nfs_pageio_init_write(&pgio, inode, 0,
 677                                 false, &nfs_async_write_completion_ops);
 678         err = nfs_do_writepage(page, wbc, &pgio);
 679         pgio.pg_error = 0;
 680         nfs_pageio_complete(&pgio);
 681         if (err < 0)
 682                 return err;
 683         if (nfs_error_is_fatal(pgio.pg_error))
 684                 return pgio.pg_error;
 685         return 0;
 686 }
 687
 688 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 689 {
 690         int ret;
 691
 692         ret = nfs_writepage_locked(page, wbc);
 693         if (ret != AOP_WRITEPAGE_ACTIVATE)
 694                 unlock_page(page);
 695         return ret;
 696 }
 697
 698 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 699 {
 700         int ret;
 701
 702         ret = nfs_do_writepage(page, wbc, data);
 703         if (ret != AOP_WRITEPAGE_ACTIVATE)
 704                 unlock_page(page);
 705         return ret;
 706 }
 707
 708 static void nfs_io_completion_commit(void *inode)
 709 {
 710         nfs_commit_inode(inode, 0);
 711 }
 712
 713 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 714 {
 715         struct inode *inode = mapping->host;
 716         struct nfs_pageio_descriptor pgio;
 717         struct nfs_io_completion *ioc = NULL;
 718         unsigned int mntflags = NFS_SERVER(inode)->flags;
 719         int priority = 0;
 720         int err;
 721
 722         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 723
 724         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 725             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 726                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 727                 if (ioc)
 728                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 729                                                inode);
 730                 priority = wb_priority(wbc);
 731         }
 732
 733         nfs_pageio_init_write(&pgio, inode, priority, false,
 734                                 &nfs_async_write_completion_ops);
 735         pgio.pg_io_completion = ioc;
 736         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 737         pgio.pg_error = 0;
 738         nfs_pageio_complete(&pgio);
 739         nfs_io_completion_put(ioc);
 740
 741         if (err < 0)
 742                 goto out_err;
 743         err = pgio.pg_error;
 744         if (nfs_error_is_fatal(err))
 745                 goto out_err;
 746         return 0;
 747 out_err:
 748         return err;
 749 }
 750
 751 /*
 752  * Insert a write request into an inode
 753  */
 754 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 755 {
 756         struct address_space *mapping = page_file_mapping(req->wb_page);
 757         struct nfs_inode *nfsi = NFS_I(inode);
 758
 759         WARN_ON_ONCE(req->wb_this_page != req);
 760
 761         /* Lock the request! */
 762         nfs_lock_request(req);
 763
 764         /*
 765          * Swap-space should not get truncated. Hence no need to plug the race
 766          * with invalidate/truncate.
 767          */
 768         spin_lock(&mapping->private_lock);
 769         if (likely(!PageSwapCache(req->wb_page))) {
 770                 set_bit(PG_MAPPED, &req->wb_flags);
 771                 SetPagePrivate(req->wb_page);
 772                 set_page_private(req->wb_page, (unsigned long)req);
 773         }
 774         spin_unlock(&mapping->private_lock);
 775         atomic_long_inc(&nfsi->nrequests);
 776         /* this a head request for a page group - mark it as having an
 777          * extra reference so sub groups can follow suit.
 778          * This flag also informs pgio layer when to bump nrequests when
 779          * adding subrequests. */
 780         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 781         kref_get(&req->wb_kref);
 782 }
 783
 784 /*
 785  * Remove a write request from an inode
 786  */
 787 static void nfs_inode_remove_request(struct nfs_page *req)
 788 {
 789         struct address_space *mapping = page_file_mapping(req->wb_page);
 790         struct inode *inode = mapping->host;
 791         struct nfs_inode *nfsi = NFS_I(inode);
 792         struct nfs_page *head;
 793
 794         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 795                 head = req->wb_head;
 796
 797                 spin_lock(&mapping->private_lock);
 798                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 799                         set_page_private(head->wb_page, 0);
 800                         ClearPagePrivate(head->wb_page);
 801                         clear_bit(PG_MAPPED, &head->wb_flags);
 802                 }
 803                 spin_unlock(&mapping->private_lock);
 804         }
 805
 806         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 807                 nfs_release_request(req);
 808                 atomic_long_dec(&nfsi->nrequests);
 809         }
 810 }
 811
 812 static void
 813 nfs_mark_request_dirty(struct nfs_page *req)
 814 {
 815         if (req->wb_page)
 816                 __set_page_dirty_nobuffers(req->wb_page);
 817 }
 818
 819 /*
 820  * nfs_page_search_commits_for_head_request_locked
 821  *
 822  * Search through commit lists on @inode for the head request for @page.
 823  * Must be called while holding the inode (which is cinfo) lock.
 824  *
 825  * Returns the head request if found, or NULL if not found.
 826  */
 827 static struct nfs_page *
 828 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 829                                                 struct page *page)
 830 {
 831         struct nfs_page *freq, *t;
 832         struct nfs_commit_info cinfo;
 833         struct inode *inode = &nfsi->vfs_inode;
 834
 835         nfs_init_cinfo_from_inode(&cinfo, inode);
 836
 837         /* search through pnfs commit lists */
 838         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 839         if (freq)
 840                 return freq->wb_head;
 841
 842         /* Linearly search the commit list for the correct request */
 843         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 844                 if (freq->wb_page == page)
 845                         return freq->wb_head;
 846         }
 847
 848         return NULL;
 849 }
 850
 851 /**
 852  * nfs_request_add_commit_list_locked - add request to a commit list
 853  * @req: pointer to a struct nfs_page
 854  * @dst: commit list head
 855  * @cinfo: holds list lock and accounting info
 856  *
 857  * This sets the PG_CLEAN bit, updates the cinfo count of
 858  * number of outstanding requests requiring a commit as well as
 859  * the MM page stats.
 860  *
 861  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 862  * nfs_page lock.
 863  */
 864 void
 865 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 866                             struct nfs_commit_info *cinfo)
 867 {
 868         set_bit(PG_CLEAN, &req->wb_flags);
 869         nfs_list_add_request(req, dst);
 870         atomic_long_inc(&cinfo->mds->ncommit);
 871 }
 872 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 873
 874 /**
 875  * nfs_request_add_commit_list - add request to a commit list
 876  * @req: pointer to a struct nfs_page
 877  * @cinfo: holds list lock and accounting info
 878  *
 879  * This sets the PG_CLEAN bit, updates the cinfo count of
 880  * number of outstanding requests requiring a commit as well as
 881  * the MM page stats.
 882  *
 883  * The caller must _not_ hold the cinfo->lock, but must be
 884  * holding the nfs_page lock.
 885  */
 886 void
 887 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 888 {
 889         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 890         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 891         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 892         if (req->wb_page)
 893                 nfs_mark_page_unstable(req->wb_page, cinfo);
 894 }
 895 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 896
 897 /**
 898  * nfs_request_remove_commit_list - Remove request from a commit list
 899  * @req: pointer to a nfs_page
 900  * @cinfo: holds list lock and accounting info
 901  *
 902  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 903  * number of outstanding requests requiring a commit
 904  * It does not update the MM page stats.
 905  *
 906  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 907  */
 908 void
 909 nfs_request_remove_commit_list(struct nfs_page *req,
 910                                struct nfs_commit_info *cinfo)
 911 {
 912         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 913                 return;
 914         nfs_list_remove_request(req);
 915         atomic_long_dec(&cinfo->mds->ncommit);
 916 }
 917 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 918
 919 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 920                                       struct inode *inode)
 921 {
 922         cinfo->inode = inode;
 923         cinfo->mds = &NFS_I(inode)->commit_info;
 924         cinfo->ds = pnfs_get_ds_info(inode);
 925         cinfo->dreq = NULL;
 926         cinfo->completion_ops = &nfs_commit_completion_ops;
 927 }
 928
 929 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 930                     struct inode *inode,
 931                     struct nfs_direct_req *dreq)
 932 {
 933         if (dreq)
 934                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 935         else
 936                 nfs_init_cinfo_from_inode(cinfo, inode);
 937 }
 938 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 939
 940 /*
 941  * Add a request to the inode's commit list.
 942  */
 943 void
 944 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 945                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 946 {
 947         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 948                 return;
 949         nfs_request_add_commit_list(req, cinfo);
 950 }
 951
 952 static void
 953 nfs_clear_page_commit(struct page *page)
 954 {
 955         dec_node_page_state(page, NR_WRITEBACK);
 956         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 957                     WB_WRITEBACK);
 958 }
 959
 960 /* Called holding the request lock on @req */
 961 static void
 962 nfs_clear_request_commit(struct nfs_page *req)
 963 {
 964         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 965                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 966                 struct inode *inode = d_inode(ctx->dentry);
 967                 struct nfs_commit_info cinfo;
 968
 969                 nfs_init_cinfo_from_inode(&cinfo, inode);
 970                 mutex_lock(&NFS_I(inode)->commit_mutex);
 971                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 972                         nfs_request_remove_commit_list(req, &cinfo);
 973                 }
 974                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 975                 nfs_clear_page_commit(req->wb_page);
 976         }
 977 }
 978
 979 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 980 {
 981         if (hdr->verf.committed == NFS_DATA_SYNC)
 982                 return hdr->lseg == NULL;
 983         return hdr->verf.committed != NFS_FILE_SYNC;
 984 }
 985
 986 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 987 {
 988         nfs_io_completion_get(hdr->io_completion);
 989 }
 990
 991 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 992 {
 993         struct nfs_commit_info cinfo;
 994         unsigned long bytes = 0;
 995
 996         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 997                 goto out;
 998         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 999         while (!list_empty(&hdr->pages)) {
1000                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1001
1002                 bytes += req->wb_bytes;
1003                 nfs_list_remove_request(req);
1004                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1005                     (hdr->good_bytes < bytes)) {
1006                         trace_nfs_comp_error(req, hdr->error);
1007                         nfs_mapping_set_error(req->wb_page, hdr->error);
1008                         goto remove_req;
1009                 }
1010                 if (nfs_write_need_commit(hdr)) {
1011                         /* Reset wb_nio, since the write was successful. */
1012                         req->wb_nio = 0;
1013                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1014                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1015                                 hdr->pgio_mirror_idx);
1016                         goto next;
1017                 }
1018 remove_req:
1019                 nfs_inode_remove_request(req);
1020 next:
1021                 nfs_end_page_writeback(req);
1022                 nfs_release_request(req);
1023         }
1024 out:
1025         nfs_io_completion_put(hdr->io_completion);
1026         hdr->release(hdr);
1027 }
1028
1029 unsigned long
1030 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1031 {
1032         return atomic_long_read(&cinfo->mds->ncommit);
1033 }
1034
1035 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1036 int
1037 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1038                      struct nfs_commit_info *cinfo, int max)
1039 {
1040         struct nfs_page *req, *tmp;
1041         int ret = 0;
1042
1043         list_for_each_entry_safe(req, tmp, src, wb_list) {
1044                 kref_get(&req->wb_kref);
1045                 if (!nfs_lock_request(req)) {
1046                         nfs_release_request(req);
1047                         continue;
1048                 }
1049                 nfs_request_remove_commit_list(req, cinfo);
1050                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1051                 nfs_list_add_request(req, dst);
1052                 ret++;
1053                 if ((ret == max) && !cinfo->dreq)
1054                         break;
1055                 cond_resched();
1056         }
1057         return ret;
1058 }
1059 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1060
1061 /*
1062  * nfs_scan_commit - Scan an inode for commit requests
1063  * @inode: NFS inode to scan
1064  * @dst: mds destination list
1065  * @cinfo: mds and ds lists of reqs ready to commit
1066  *
1067  * Moves requests from the inode's 'commit' request list.
1068  * The requests are *not* checked to ensure that they form a contiguous set.
1069  */
1070 int
1071 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1072                 struct nfs_commit_info *cinfo)
1073 {
1074         int ret = 0;
1075
1076         if (!atomic_long_read(&cinfo->mds->ncommit))
1077                 return 0;
1078         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1079         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1080                 const int max = INT_MAX;
1081
1082                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1083                                            cinfo, max);
1084                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1085         }
1086         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1087         return ret;
1088 }
1089
1090 /*
1091  * Search for an existing write request, and attempt to update
1092  * it to reflect a new dirty region on a given page.
1093  *
1094  * If the attempt fails, then the existing request is flushed out
1095  * to disk.
1096  */
1097 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1098                 struct page *page,
1099                 unsigned int offset,
1100                 unsigned int bytes)
1101 {
1102         struct nfs_page *req;
1103         unsigned int rqend;
1104         unsigned int end;
1105         int error;
1106
1107         end = offset + bytes;
1108
1109         req = nfs_lock_and_join_requests(page);
1110         if (IS_ERR_OR_NULL(req))
1111                 return req;
1112
1113         rqend = req->wb_offset + req->wb_bytes;
1114         /*
1115          * Tell the caller to flush out the request if
1116          * the offsets are non-contiguous.
1117          * Note: nfs_flush_incompatible() will already
1118          * have flushed out requests having wrong owners.
1119          */
1120         if (offset > rqend || end < req->wb_offset)
1121                 goto out_flushme;
1122
1123         /* Okay, the request matches. Update the region */
1124         if (offset < req->wb_offset) {
1125                 req->wb_offset = offset;
1126                 req->wb_pgbase = offset;
1127         }
1128         if (end > rqend)
1129                 req->wb_bytes = end - req->wb_offset;
1130         else
1131                 req->wb_bytes = rqend - req->wb_offset;
1132         req->wb_nio = 0;
1133         return req;
1134 out_flushme:
1135         /*
1136          * Note: we mark the request dirty here because
1137          * nfs_lock_and_join_requests() cannot preserve
1138          * commit flags, so we have to replay the write.
1139          */
1140         nfs_mark_request_dirty(req);
1141         nfs_unlock_and_release_request(req);
1142         error = nfs_wb_page(inode, page);
1143         return (error < 0) ? ERR_PTR(error) : NULL;
1144 }
1145
1146 /*
1147  * Try to update an existing write request, or create one if there is none.
1148  *
1149  * Note: Should always be called with the Page Lock held to prevent races
1150  * if we have to add a new request. Also assumes that the caller has
1151  * already called nfs_flush_incompatible() if necessary.
1152  */
1153 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1154                 struct page *page, unsigned int offset, unsigned int bytes)
1155 {
1156         struct inode *inode = page_file_mapping(page)->host;
1157         struct nfs_page *req;
1158
1159         req = nfs_try_to_update_request(inode, page, offset, bytes);
1160         if (req != NULL)
1161                 goto out;
1162         req = nfs_create_request(ctx, page, offset, bytes);
1163         if (IS_ERR(req))
1164                 goto out;
1165         nfs_inode_add_request(inode, req);
1166 out:
1167         return req;
1168 }
1169
1170 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1171                 unsigned int offset, unsigned int count)
1172 {
1173         struct nfs_page *req;
1174
1175         req = nfs_setup_write_request(ctx, page, offset, count);
1176         if (IS_ERR(req))
1177                 return PTR_ERR(req);
1178         /* Update file length */
1179         nfs_grow_file(page, offset, count);
1180         nfs_mark_uptodate(req);
1181         nfs_mark_request_dirty(req);
1182         nfs_unlock_and_release_request(req);
1183         return 0;
1184 }
1185
1186 int nfs_flush_incompatible(struct file *file, struct page *page)
1187 {
1188         struct nfs_open_context *ctx = nfs_file_open_context(file);
1189         struct nfs_lock_context *l_ctx;
1190         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1191         struct nfs_page *req;
1192         int do_flush, status;
1193         /*
1194          * Look for a request corresponding to this page. If there
1195          * is one, and it belongs to another file, we flush it out
1196          * before we try to copy anything into the page. Do this
1197          * due to the lack of an ACCESS-type call in NFSv2.
1198          * Also do the same if we find a request from an existing
1199          * dropped page.
1200          */
1201         do {
1202                 req = nfs_page_find_head_request(page);
1203                 if (req == NULL)
1204                         return 0;
1205                 l_ctx = req->wb_lock_context;
1206                 do_flush = req->wb_page != page ||
1207                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1208                 if (l_ctx && flctx &&
1209                     !(list_empty_careful(&flctx->flc_posix) &&
1210                       list_empty_careful(&flctx->flc_flock))) {
1211                         do_flush |= l_ctx->lockowner != current->files;
1212                 }
1213                 nfs_release_request(req);
1214                 if (!do_flush)
1215                         return 0;
1216                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1217         } while (status == 0);
1218         return status;
1219 }
1220
1221 /*
1222  * Avoid buffered writes when a open context credential's key would
1223  * expire soon.
1224  *
1225  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1226  *
1227  * Return 0 and set a credential flag which triggers the inode to flush
1228  * and performs  NFS_FILE_SYNC writes if the key will expired within
1229  * RPC_KEY_EXPIRE_TIMEO.
1230  */
1231 int
1232 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1233 {
1234         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1235
1236         if (nfs_ctx_key_to_expire(ctx, inode) &&
1237             !rcu_access_pointer(ctx->ll_cred))
1238                 /* Already expired! */
1239                 return -EACCES;
1240         return 0;
1241 }
1242
1243 /*
1244  * Test if the open context credential key is marked to expire soon.
1245  */
1246 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1247 {
1248         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1249         struct rpc_cred *cred, *new, *old = NULL;
1250         struct auth_cred acred = {
1251                 .cred = ctx->cred,
1252         };
1253         bool ret = false;
1254
1255         rcu_read_lock();
1256         cred = rcu_dereference(ctx->ll_cred);
1257         if (cred && !(cred->cr_ops->crkey_timeout &&
1258                       cred->cr_ops->crkey_timeout(cred)))
1259                 goto out;
1260         rcu_read_unlock();
1261
1262         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1263         if (new == cred) {
1264                 put_rpccred(new);
1265                 return true;
1266         }
1267         if (IS_ERR_OR_NULL(new)) {
1268                 new = NULL;
1269                 ret = true;
1270         } else if (new->cr_ops->crkey_timeout &&
1271                    new->cr_ops->crkey_timeout(new))
1272                 ret = true;
1273
1274         rcu_read_lock();
1275         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1276                                              RCU_INITIALIZER(new)), 1);
1277 out:
1278         rcu_read_unlock();
1279         put_rpccred(old);
1280         return ret;
1281 }
1282
1283 /*
1284  * If the page cache is marked as unsafe or invalid, then we can't rely on
1285  * the PageUptodate() flag. In this case, we will need to turn off
1286  * write optimisations that depend on the page contents being correct.
1287  */
1288 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1289                                    unsigned int pagelen)
1290 {
1291         struct nfs_inode *nfsi = NFS_I(inode);
1292
1293         if (nfs_have_delegated_attributes(inode))
1294                 goto out;
1295         if (nfsi->cache_validity &
1296             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1297                 return false;
1298         smp_rmb();
1299         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1300                 return false;
1301 out:
1302         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1303                 return false;
1304         return PageUptodate(page) != 0;
1305 }
1306
1307 static bool
1308 is_whole_file_wrlock(struct file_lock *fl)
1309 {
1310         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1311                         fl->fl_type == F_WRLCK;
1312 }
1313
1314 /* If we know the page is up to date, and we're not using byte range locks (or
1315  * if we have the whole file locked for writing), it may be more efficient to
1316  * extend the write to cover the entire page in order to avoid fragmentation
1317  * inefficiencies.
1318  *
1319  * If the file is opened for synchronous writes then we can just skip the rest
1320  * of the checks.
1321  */
1322 static int nfs_can_extend_write(struct file *file, struct page *page,
1323                                 struct inode *inode, unsigned int pagelen)
1324 {
1325         int ret;
1326         struct file_lock_context *flctx = inode->i_flctx;
1327         struct file_lock *fl;
1328
1329         if (file->f_flags & O_DSYNC)
1330                 return 0;
1331         if (!nfs_write_pageuptodate(page, inode, pagelen))
1332                 return 0;
1333         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1334                 return 1;
1335         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1336                        list_empty_careful(&flctx->flc_posix)))
1337                 return 1;
1338
1339         /* Check to see if there are whole file write locks */
1340         ret = 0;
1341         spin_lock(&flctx->flc_lock);
1342         if (!list_empty(&flctx->flc_posix)) {
1343                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1344                                         fl_list);
1345                 if (is_whole_file_wrlock(fl))
1346                         ret = 1;
1347         } else if (!list_empty(&flctx->flc_flock)) {
1348                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1349                                         fl_list);
1350                 if (fl->fl_type == F_WRLCK)
1351                         ret = 1;
1352         }
1353         spin_unlock(&flctx->flc_lock);
1354         return ret;
1355 }
1356
1357 /*
1358  * Update and possibly write a cached page of an NFS file.
1359  *
1360  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1361  * things with a page scheduled for an RPC call (e.g. invalidate it).
1362  */
1363 int nfs_updatepage(struct file *file, struct page *page,
1364                 unsigned int offset, unsigned int count)
1365 {
1366         struct nfs_open_context *ctx = nfs_file_open_context(file);
1367         struct address_space *mapping = page_file_mapping(page);
1368         struct inode    *inode = mapping->host;
1369         unsigned int    pagelen = nfs_page_length(page);
1370         int             status = 0;
1371
1372         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1373
1374         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1375                 file, count, (long long)(page_file_offset(page) + offset));
1376
1377         if (!count)
1378                 goto out;
1379
1380         if (nfs_can_extend_write(file, page, inode, pagelen)) {
1381                 count = max(count + offset, pagelen);
1382                 offset = 0;
1383         }
1384
1385         status = nfs_writepage_setup(ctx, page, offset, count);
1386         if (status < 0)
1387                 nfs_set_pageerror(mapping);
1388 out:
1389         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1390                         status, (long long)i_size_read(inode));
1391         return status;
1392 }
1393
1394 static int flush_task_priority(int how)
1395 {
1396         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1397                 case FLUSH_HIGHPRI:
1398                         return RPC_PRIORITY_HIGH;
1399                 case FLUSH_LOWPRI:
1400                         return RPC_PRIORITY_LOW;
1401         }
1402         return RPC_PRIORITY_NORMAL;
1403 }
1404
1405 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1406                                struct rpc_message *msg,
1407                                const struct nfs_rpc_ops *rpc_ops,
1408                                struct rpc_task_setup *task_setup_data, int how)
1409 {
1410         int priority = flush_task_priority(how);
1411
1412         task_setup_data->priority = priority;
1413         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1414         trace_nfs_initiate_write(hdr);
1415 }
1416
1417 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1418  * call this on each, which will prepare them to be retried on next
1419  * writeback using standard nfs.
1420  */
1421 static void nfs_redirty_request(struct nfs_page *req)
1422 {
1423         /* Bump the transmission count */
1424         req->wb_nio++;
1425         nfs_mark_request_dirty(req);
1426         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1427         nfs_end_page_writeback(req);
1428         nfs_release_request(req);
1429 }
1430
1431 static void nfs_async_write_error(struct list_head *head, int error)
1432 {
1433         struct nfs_page *req;
1434
1435         while (!list_empty(head)) {
1436                 req = nfs_list_entry(head->next);
1437                 nfs_list_remove_request(req);
1438                 if (nfs_error_is_fatal(error))
1439                         nfs_write_error(req, error);
1440                 else
1441                         nfs_redirty_request(req);
1442         }
1443 }
1444
1445 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1446 {
1447         nfs_async_write_error(&hdr->pages, 0);
1448         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1449                         hdr->args.offset + hdr->args.count - 1);
1450 }
1451
1452 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1453         .init_hdr = nfs_async_write_init,
1454         .error_cleanup = nfs_async_write_error,
1455         .completion = nfs_write_completion,
1456         .reschedule_io = nfs_async_write_reschedule_io,
1457 };
1458
1459 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1460                                struct inode *inode, int ioflags, bool force_mds,
1461                                const struct nfs_pgio_completion_ops *compl_ops)
1462 {
1463         struct nfs_server *server = NFS_SERVER(inode);
1464         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1465
1466 #ifdef CONFIG_NFS_V4_1
1467         if (server->pnfs_curr_ld && !force_mds)
1468                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1469 #endif
1470         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1471                         server->wsize, ioflags);
1472 }
1473 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1474
1475 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1476 {
1477         struct nfs_pgio_mirror *mirror;
1478
1479         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1480                 pgio->pg_ops->pg_cleanup(pgio);
1481
1482         pgio->pg_ops = &nfs_pgio_rw_ops;
1483
1484         nfs_pageio_stop_mirroring(pgio);
1485
1486         mirror = &pgio->pg_mirrors[0];
1487         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1488 }
1489 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1490
1491
1492 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1493 {
1494         struct nfs_commit_data *data = calldata;
1495
1496         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1497 }
1498
1499 /*
1500  * Special version of should_remove_suid() that ignores capabilities.
1501  */
1502 static int nfs_should_remove_suid(const struct inode *inode)
1503 {
1504         umode_t mode = inode->i_mode;
1505         int kill = 0;
1506
1507         /* suid always must be killed */
1508         if (unlikely(mode & S_ISUID))
1509                 kill = ATTR_KILL_SUID;
1510
1511         /*
1512          * sgid without any exec bits is just a mandatory locking mark; leave
1513          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1514          */
1515         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1516                 kill |= ATTR_KILL_SGID;
1517
1518         if (unlikely(kill && S_ISREG(mode)))
1519                 return kill;
1520
1521         return 0;
1522 }
1523
1524 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1525                 struct nfs_fattr *fattr)
1526 {
1527         struct nfs_pgio_args *argp = &hdr->args;
1528         struct nfs_pgio_res *resp = &hdr->res;
1529         u64 size = argp->offset + resp->count;
1530
1531         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1532                 fattr->size = size;
1533         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1534                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1535                 return;
1536         }
1537         if (size != fattr->size)
1538                 return;
1539         /* Set attribute barrier */
1540         nfs_fattr_set_barrier(fattr);
1541         /* ...and update size */
1542         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1543 }
1544
1545 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1546 {
1547         struct nfs_fattr *fattr = &hdr->fattr;
1548         struct inode *inode = hdr->inode;
1549
1550         spin_lock(&inode->i_lock);
1551         nfs_writeback_check_extend(hdr, fattr);
1552         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1553         spin_unlock(&inode->i_lock);
1554 }
1555 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1556
1557 /*
1558  * This function is called when the WRITE call is complete.
1559  */
1560 static int nfs_writeback_done(struct rpc_task *task,
1561                               struct nfs_pgio_header *hdr,
1562                               struct inode *inode)
1563 {
1564         int status;
1565
1566         /*
1567          * ->write_done will attempt to use post-op attributes to detect
1568          * conflicting writes by other clients.  A strict interpretation
1569          * of close-to-open would allow us to continue caching even if
1570          * another writer had changed the file, but some applications
1571          * depend on tighter cache coherency when writing.
1572          */
1573         status = NFS_PROTO(inode)->write_done(task, hdr);
1574         if (status != 0)
1575                 return status;
1576
1577         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1578         trace_nfs_writeback_done(task, hdr);
1579
1580         if (hdr->res.verf->committed < hdr->args.stable &&
1581             task->tk_status >= 0) {
1582                 /* We tried a write call, but the server did not
1583                  * commit data to stable storage even though we
1584                  * requested it.
1585                  * Note: There is a known bug in Tru64 < 5.0 in which
1586                  *       the server reports NFS_DATA_SYNC, but performs
1587                  *       NFS_FILE_SYNC. We therefore implement this checking
1588                  *       as a dprintk() in order to avoid filling syslog.
1589                  */
1590                 static unsigned long    complain;
1591
1592                 /* Note this will print the MDS for a DS write */
1593                 if (time_before(complain, jiffies)) {
1594                         dprintk("NFS:       faulty NFS server %s:"
1595                                 " (committed = %d) != (stable = %d)\n",
1596                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1597                                 hdr->res.verf->committed, hdr->args.stable);
1598                         complain = jiffies + 300 * HZ;
1599                 }
1600         }
1601
1602         /* Deal with the suid/sgid bit corner case */
1603         if (nfs_should_remove_suid(inode)) {
1604                 spin_lock(&inode->i_lock);
1605                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1606                 spin_unlock(&inode->i_lock);
1607         }
1608         return 0;
1609 }
1610
1611 /*
1612  * This function is called when the WRITE call is complete.
1613  */
1614 static void nfs_writeback_result(struct rpc_task *task,
1615                                  struct nfs_pgio_header *hdr)
1616 {
1617         struct nfs_pgio_args    *argp = &hdr->args;
1618         struct nfs_pgio_res     *resp = &hdr->res;
1619
1620         if (resp->count < argp->count) {
1621                 static unsigned long    complain;
1622
1623                 /* This a short write! */
1624                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1625
1626                 /* Has the server at least made some progress? */
1627                 if (resp->count == 0) {
1628                         if (time_before(complain, jiffies)) {
1629                                 printk(KERN_WARNING
1630                                        "NFS: Server wrote zero bytes, expected %u.\n",
1631                                        argp->count);
1632                                 complain = jiffies + 300 * HZ;
1633                         }
1634                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1635                         task->tk_status = -EIO;
1636                         return;
1637                 }
1638
1639                 /* For non rpc-based layout drivers, retry-through-MDS */
1640                 if (!task->tk_ops) {
1641                         hdr->pnfs_error = -EAGAIN;
1642                         return;
1643                 }
1644
1645                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1646                 if (resp->verf->committed != NFS_UNSTABLE) {
1647                         /* Resend from where the server left off */
1648                         hdr->mds_offset += resp->count;
1649                         argp->offset += resp->count;
1650                         argp->pgbase += resp->count;
1651                         argp->count -= resp->count;
1652                 } else {
1653                         /* Resend as a stable write in order to avoid
1654                          * headaches in the case of a server crash.
1655                          */
1656                         argp->stable = NFS_FILE_SYNC;
1657                 }
1658                 resp->count = 0;
1659                 resp->verf->committed = 0;
1660                 rpc_restart_call_prepare(task);
1661         }
1662 }
1663
1664 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1665 {
1666         return wait_var_event_killable(&cinfo->rpcs_out,
1667                                        !atomic_read(&cinfo->rpcs_out));
1668 }
1669
1670 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1671 {
1672         atomic_inc(&cinfo->rpcs_out);
1673 }
1674
1675 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1676 {
1677         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1678                 wake_up_var(&cinfo->rpcs_out);
1679                 return true;
1680         }
1681         return false;
1682 }
1683
1684 void nfs_commitdata_release(struct nfs_commit_data *data)
1685 {
1686         put_nfs_open_context(data->context);
1687         nfs_commit_free(data);
1688 }
1689 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1690
1691 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1692                         const struct nfs_rpc_ops *nfs_ops,
1693                         const struct rpc_call_ops *call_ops,
1694                         int how, int flags)
1695 {
1696         struct rpc_task *task;
1697         int priority = flush_task_priority(how);
1698         struct rpc_message msg = {
1699                 .rpc_argp = &data->args,
1700                 .rpc_resp = &data->res,
1701                 .rpc_cred = data->cred,
1702         };
1703         struct rpc_task_setup task_setup_data = {
1704                 .task = &data->task,
1705                 .rpc_client = clnt,
1706                 .rpc_message = &msg,
1707                 .callback_ops = call_ops,
1708                 .callback_data = data,
1709                 .workqueue = nfsiod_workqueue,
1710                 .flags = RPC_TASK_ASYNC | flags,
1711                 .priority = priority,
1712         };
1713         /* Set up the initial task struct.  */
1714         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1715         trace_nfs_initiate_commit(data);
1716
1717         dprintk("NFS: initiated commit call\n");
1718
1719         task = rpc_run_task(&task_setup_data);
1720         if (IS_ERR(task))
1721                 return PTR_ERR(task);
1722         if (how & FLUSH_SYNC)
1723                 rpc_wait_for_completion_task(task);
1724         rpc_put_task(task);
1725         return 0;
1726 }
1727 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1728
1729 static loff_t nfs_get_lwb(struct list_head *head)
1730 {
1731         loff_t lwb = 0;
1732         struct nfs_page *req;
1733
1734         list_for_each_entry(req, head, wb_list)
1735                 if (lwb < (req_offset(req) + req->wb_bytes))
1736                         lwb = req_offset(req) + req->wb_bytes;
1737
1738         return lwb;
1739 }
1740
1741 /*
1742  * Set up the argument/result storage required for the RPC call.
1743  */
1744 void nfs_init_commit(struct nfs_commit_data *data,
1745                      struct list_head *head,
1746                      struct pnfs_layout_segment *lseg,
1747                      struct nfs_commit_info *cinfo)
1748 {
1749         struct nfs_page *first;
1750         struct nfs_open_context *ctx;
1751         struct inode *inode;
1752
1753         /* Set up the RPC argument and reply structs
1754          * NB: take care not to mess about with data->commit et al. */
1755
1756         if (head)
1757                 list_splice_init(head, &data->pages);
1758
1759         first = nfs_list_entry(data->pages.next);
1760         ctx = nfs_req_openctx(first);
1761         inode = d_inode(ctx->dentry);
1762
1763         data->inode       = inode;
1764         data->cred        = ctx->cred;
1765         data->lseg        = lseg; /* reference transferred */
1766         /* only set lwb for pnfs commit */
1767         if (lseg)
1768                 data->lwb = nfs_get_lwb(&data->pages);
1769         data->mds_ops     = &nfs_commit_ops;
1770         data->completion_ops = cinfo->completion_ops;
1771         data->dreq        = cinfo->dreq;
1772
1773         data->args.fh     = NFS_FH(data->inode);
1774         /* Note: we always request a commit of the entire inode */
1775         data->args.offset = 0;
1776         data->args.count  = 0;
1777         data->context     = get_nfs_open_context(ctx);
1778         data->res.fattr   = &data->fattr;
1779         data->res.verf    = &data->verf;
1780         nfs_fattr_init(&data->fattr);
1781         nfs_commit_begin(cinfo->mds);
1782 }
1783 EXPORT_SYMBOL_GPL(nfs_init_commit);
1784
1785 void nfs_retry_commit(struct list_head *page_list,
1786                       struct pnfs_layout_segment *lseg,
1787                       struct nfs_commit_info *cinfo,
1788                       u32 ds_commit_idx)
1789 {
1790         struct nfs_page *req;
1791
1792         while (!list_empty(page_list)) {
1793                 req = nfs_list_entry(page_list->next);
1794                 nfs_list_remove_request(req);
1795                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1796                 if (!cinfo->dreq)
1797                         nfs_clear_page_commit(req->wb_page);
1798                 nfs_unlock_and_release_request(req);
1799         }
1800 }
1801 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1802
1803 static void
1804 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1805                 struct nfs_page *req)
1806 {
1807         __set_page_dirty_nobuffers(req->wb_page);
1808 }
1809
1810 /*
1811  * Commit dirty pages
1812  */
1813 static int
1814 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1815                 struct nfs_commit_info *cinfo)
1816 {
1817         struct nfs_commit_data  *data;
1818         unsigned short task_flags = 0;
1819
1820         /* another commit raced with us */
1821         if (list_empty(head))
1822                 return 0;
1823
1824         data = nfs_commitdata_alloc(true);
1825
1826         /* Set up the argument struct */
1827         nfs_init_commit(data, head, NULL, cinfo);
1828         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1829                 task_flags = RPC_TASK_MOVEABLE;
1830         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1831                                    data->mds_ops, how,
1832                                    RPC_TASK_CRED_NOREF | task_flags);
1833 }
1834
1835 /*
1836  * COMMIT call returned
1837  */
1838 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1839 {
1840         struct nfs_commit_data  *data = calldata;
1841
1842         /* Call the NFS version-specific code */
1843         NFS_PROTO(data->inode)->commit_done(task, data);
1844         trace_nfs_commit_done(task, data);
1845 }
1846
1847 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1848 {
1849         const struct nfs_writeverf *verf = data->res.verf;
1850         struct nfs_page *req;
1851         int status = data->task.tk_status;
1852         struct nfs_commit_info cinfo;
1853         struct nfs_server *nfss;
1854
1855         while (!list_empty(&data->pages)) {
1856                 req = nfs_list_entry(data->pages.next);
1857                 nfs_list_remove_request(req);
1858                 if (req->wb_page)
1859                         nfs_clear_page_commit(req->wb_page);
1860
1861                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1862                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1863                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1864                         req->wb_bytes,
1865                         (long long)req_offset(req));
1866                 if (status < 0) {
1867                         if (req->wb_page) {
1868                                 trace_nfs_commit_error(req, status);
1869                                 nfs_mapping_set_error(req->wb_page, status);
1870                                 nfs_inode_remove_request(req);
1871                         }
1872                         dprintk_cont(", error = %d\n", status);
1873                         goto next;
1874                 }
1875
1876                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1877                  * returned by the server against all stored verfs. */
1878                 if (nfs_write_match_verf(verf, req)) {
1879                         /* We have a match */
1880                         if (req->wb_page)
1881                                 nfs_inode_remove_request(req);
1882                         dprintk_cont(" OK\n");
1883                         goto next;
1884                 }
1885                 /* We have a mismatch. Write the page again */
1886                 dprintk_cont(" mismatch\n");
1887                 nfs_mark_request_dirty(req);
1888                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1889         next:
1890                 nfs_unlock_and_release_request(req);
1891                 /* Latency breaker */
1892                 cond_resched();
1893         }
1894         nfss = NFS_SERVER(data->inode);
1895         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1896                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1897
1898         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1899         nfs_commit_end(cinfo.mds);
1900 }
1901
1902 static void nfs_commit_release(void *calldata)
1903 {
1904         struct nfs_commit_data *data = calldata;
1905
1906         data->completion_ops->completion(data);
1907         nfs_commitdata_release(calldata);
1908 }
1909
1910 static const struct rpc_call_ops nfs_commit_ops = {
1911         .rpc_call_prepare = nfs_commit_prepare,
1912         .rpc_call_done = nfs_commit_done,
1913         .rpc_release = nfs_commit_release,
1914 };
1915
1916 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1917         .completion = nfs_commit_release_pages,
1918         .resched_write = nfs_commit_resched_write,
1919 };
1920
1921 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1922                             int how, struct nfs_commit_info *cinfo)
1923 {
1924         int status;
1925
1926         status = pnfs_commit_list(inode, head, how, cinfo);
1927         if (status == PNFS_NOT_ATTEMPTED)
1928                 status = nfs_commit_list(inode, head, how, cinfo);
1929         return status;
1930 }
1931
1932 static int __nfs_commit_inode(struct inode *inode, int how,
1933                 struct writeback_control *wbc)
1934 {
1935         LIST_HEAD(head);
1936         struct nfs_commit_info cinfo;
1937         int may_wait = how & FLUSH_SYNC;
1938         int ret, nscan;
1939
1940         how &= ~FLUSH_SYNC;
1941         nfs_init_cinfo_from_inode(&cinfo, inode);
1942         nfs_commit_begin(cinfo.mds);
1943         for (;;) {
1944                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1945                 if (ret <= 0)
1946                         break;
1947                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1948                 if (ret < 0)
1949                         break;
1950                 ret = 0;
1951                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1952                         if (nscan < wbc->nr_to_write)
1953                                 wbc->nr_to_write -= nscan;
1954                         else
1955                                 wbc->nr_to_write = 0;
1956                 }
1957                 if (nscan < INT_MAX)
1958                         break;
1959                 cond_resched();
1960         }
1961         nfs_commit_end(cinfo.mds);
1962         if (ret || !may_wait)
1963                 return ret;
1964         return wait_on_commit(cinfo.mds);
1965 }
1966
1967 int nfs_commit_inode(struct inode *inode, int how)
1968 {
1969         return __nfs_commit_inode(inode, how, NULL);
1970 }
1971 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1972
1973 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1974 {
1975         struct nfs_inode *nfsi = NFS_I(inode);
1976         int flags = FLUSH_SYNC;
1977         int ret = 0;
1978
1979         if (wbc->sync_mode == WB_SYNC_NONE) {
1980                 /* no commits means nothing needs to be done */
1981                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1982                         goto check_requests_outstanding;
1983
1984                 /* Don't commit yet if this is a non-blocking flush and there
1985                  * are a lot of outstanding writes for this mapping.
1986                  */
1987                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1988                         goto out_mark_dirty;
1989
1990                 /* don't wait for the COMMIT response */
1991                 flags = 0;
1992         }
1993
1994         ret = __nfs_commit_inode(inode, flags, wbc);
1995         if (!ret) {
1996                 if (flags & FLUSH_SYNC)
1997                         return 0;
1998         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1999                 goto out_mark_dirty;
2000
2001 check_requests_outstanding:
2002         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2003                 return ret;
2004 out_mark_dirty:
2005         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2006         return ret;
2007 }
2008 EXPORT_SYMBOL_GPL(nfs_write_inode);
2009
2010 /*
2011  * Wrapper for filemap_write_and_wait_range()
2012  *
2013  * Needed for pNFS in order to ensure data becomes visible to the
2014  * client.
2015  */
2016 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2017                 loff_t lstart, loff_t lend)
2018 {
2019         int ret;
2020
2021         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2022         if (ret == 0)
2023                 ret = pnfs_sync_inode(mapping->host, true);
2024         return ret;
2025 }
2026 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2027
2028 /*
2029  * flush the inode to disk.
2030  */
2031 int nfs_wb_all(struct inode *inode)
2032 {
2033         int ret;
2034
2035         trace_nfs_writeback_inode_enter(inode);
2036
2037         ret = filemap_write_and_wait(inode->i_mapping);
2038         if (ret)
2039                 goto out;
2040         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2041         if (ret < 0)
2042                 goto out;
2043         pnfs_sync_inode(inode, true);
2044         ret = 0;
2045
2046 out:
2047         trace_nfs_writeback_inode_exit(inode, ret);
2048         return ret;
2049 }
2050 EXPORT_SYMBOL_GPL(nfs_wb_all);
2051
2052 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2053 {
2054         struct nfs_page *req;
2055         int ret = 0;
2056
2057         wait_on_page_writeback(page);
2058
2059         /* blocking call to cancel all requests and join to a single (head)
2060          * request */
2061         req = nfs_lock_and_join_requests(page);
2062
2063         if (IS_ERR(req)) {
2064                 ret = PTR_ERR(req);
2065         } else if (req) {
2066                 /* all requests from this page have been cancelled by
2067                  * nfs_lock_and_join_requests, so just remove the head
2068                  * request from the inode / page_private pointer and
2069                  * release it */
2070                 nfs_inode_remove_request(req);
2071                 nfs_unlock_and_release_request(req);
2072         }
2073
2074         return ret;
2075 }
2076
2077 /*
2078  * Write back all requests on one page - we do this before reading it.
2079  */
2080 int nfs_wb_page(struct inode *inode, struct page *page)
2081 {
2082         loff_t range_start = page_file_offset(page);
2083         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2084         struct writeback_control wbc = {
2085                 .sync_mode = WB_SYNC_ALL,
2086                 .nr_to_write = 0,
2087                 .range_start = range_start,
2088                 .range_end = range_end,
2089         };
2090         int ret;
2091
2092         trace_nfs_writeback_page_enter(inode);
2093
2094         for (;;) {
2095                 wait_on_page_writeback(page);
2096                 if (clear_page_dirty_for_io(page)) {
2097                         ret = nfs_writepage_locked(page, &wbc);
2098                         if (ret < 0)
2099                                 goto out_error;
2100                         continue;
2101                 }
2102                 ret = 0;
2103                 if (!PagePrivate(page))
2104                         break;
2105                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2106                 if (ret < 0)
2107                         goto out_error;
2108         }
2109 out_error:
2110         trace_nfs_writeback_page_exit(inode, ret);
2111         return ret;
2112 }
2113
2114 #ifdef CONFIG_MIGRATION
2115 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2116                 struct page *page, enum migrate_mode mode)
2117 {
2118         /*
2119          * If PagePrivate is set, then the page is currently associated with
2120          * an in-progress read or write request. Don't try to migrate it.
2121          *
2122          * FIXME: we could do this in principle, but we'll need a way to ensure
2123          *        that we can safely release the inode reference while holding
2124          *        the page lock.
2125          */
2126         if (PagePrivate(page))
2127                 return -EBUSY;
2128
2129         if (PageFsCache(page)) {
2130                 if (mode == MIGRATE_ASYNC)
2131                         return -EBUSY;
2132                 wait_on_page_fscache(page);
2133         }
2134
2135         return migrate_page(mapping, newpage, page, mode);
2136 }
2137 #endif
2138
2139 int __init nfs_init_writepagecache(void)
2140 {
2141         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2142                                              sizeof(struct nfs_pgio_header),
2143                                              0, SLAB_HWCACHE_ALIGN,
2144                                              NULL);
2145         if (nfs_wdata_cachep == NULL)
2146                 return -ENOMEM;
2147
2148         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2149                                                      nfs_wdata_cachep);
2150         if (nfs_wdata_mempool == NULL)
2151                 goto out_destroy_write_cache;
2152
2153         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2154                                              sizeof(struct nfs_commit_data),
2155                                              0, SLAB_HWCACHE_ALIGN,
2156                                              NULL);
2157         if (nfs_cdata_cachep == NULL)
2158                 goto out_destroy_write_mempool;
2159
2160         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2161                                                       nfs_cdata_cachep);
2162         if (nfs_commit_mempool == NULL)
2163                 goto out_destroy_commit_cache;
2164
2165         /*
2166          * NFS congestion size, scale with available memory.
2167          *
2168          *  64MB:    8192k
2169          * 128MB:   11585k
2170          * 256MB:   16384k
2171          * 512MB:   23170k
2172          *   1GB:   32768k
2173          *   2GB:   46340k
2174          *   4GB:   65536k
2175          *   8GB:   92681k
2176          *  16GB:  131072k
2177          *
2178          * This allows larger machines to have larger/more transfers.
2179          * Limit the default to 256M
2180          */
2181         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2182         if (nfs_congestion_kb > 256*1024)
2183                 nfs_congestion_kb = 256*1024;
2184
2185         return 0;
2186
2187 out_destroy_commit_cache:
2188         kmem_cache_destroy(nfs_cdata_cachep);
2189 out_destroy_write_mempool:
2190         mempool_destroy(nfs_wdata_mempool);
2191 out_destroy_write_cache:
2192         kmem_cache_destroy(nfs_wdata_cachep);
2193         return -ENOMEM;
2194 }
2195
2196 void nfs_destroy_writepagecache(void)
2197 {
2198         mempool_destroy(nfs_commit_mempool);
2199         kmem_cache_destroy(nfs_cdata_cachep);
2200         mempool_destroy(nfs_wdata_mempool);
2201         kmem_cache_destroy(nfs_wdata_cachep);
2202 }
2203
2204 static const struct nfs_rw_ops nfs_rw_write_ops = {
2205         .rw_alloc_header        = nfs_writehdr_alloc,
2206         .rw_free_header         = nfs_writehdr_free,
2207         .rw_done                = nfs_writeback_done,
2208         .rw_result              = nfs_writeback_result,
2209         .rw_initiate            = nfs_initiate_write,
2210 };