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         i_size_write(inode, end);
 292         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 293         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 294 out:
 295         spin_unlock(&inode->i_lock);
 296 }
 297
 298 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 299 static void nfs_set_pageerror(struct address_space *mapping)
 300 {
 301         struct inode *inode = mapping->host;
 302
 303         nfs_zap_mapping(mapping->host, mapping);
 304         /* Force file size revalidation */
 305         spin_lock(&inode->i_lock);
 306         NFS_I(inode)->cache_validity |= NFS_INO_REVAL_FORCED |
 307                                         NFS_INO_REVAL_PAGECACHE |
 308                                         NFS_INO_INVALID_SIZE;
 309         spin_unlock(&inode->i_lock);
 310 }
 311
 312 static void nfs_mapping_set_error(struct page *page, int error)
 313 {
 314         struct address_space *mapping = page_file_mapping(page);
 315
 316         SetPageError(page);
 317         mapping_set_error(mapping, error);
 318         nfs_set_pageerror(mapping);
 319 }
 320
 321 /*
 322  * nfs_page_group_search_locked
 323  * @head - head request of page group
 324  * @page_offset - offset into page
 325  *
 326  * Search page group with head @head to find a request that contains the
 327  * page offset @page_offset.
 328  *
 329  * Returns a pointer to the first matching nfs request, or NULL if no
 330  * match is found.
 331  *
 332  * Must be called with the page group lock held
 333  */
 334 static struct nfs_page *
 335 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 336 {
 337         struct nfs_page *req;
 338
 339         req = head;
 340         do {
 341                 if (page_offset >= req->wb_pgbase &&
 342                     page_offset < (req->wb_pgbase + req->wb_bytes))
 343                         return req;
 344
 345                 req = req->wb_this_page;
 346         } while (req != head);
 347
 348         return NULL;
 349 }
 350
 351 /*
 352  * nfs_page_group_covers_page
 353  * @head - head request of page group
 354  *
 355  * Return true if the page group with head @head covers the whole page,
 356  * returns false otherwise
 357  */
 358 static bool nfs_page_group_covers_page(struct nfs_page *req)
 359 {
 360         struct nfs_page *tmp;
 361         unsigned int pos = 0;
 362         unsigned int len = nfs_page_length(req->wb_page);
 363
 364         nfs_page_group_lock(req);
 365
 366         for (;;) {
 367                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 368                 if (!tmp)
 369                         break;
 370                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 371         }
 372
 373         nfs_page_group_unlock(req);
 374         return pos >= len;
 375 }
 376
 377 /* We can set the PG_uptodate flag if we see that a write request
 378  * covers the full page.
 379  */
 380 static void nfs_mark_uptodate(struct nfs_page *req)
 381 {
 382         if (PageUptodate(req->wb_page))
 383                 return;
 384         if (!nfs_page_group_covers_page(req))
 385                 return;
 386         SetPageUptodate(req->wb_page);
 387 }
 388
 389 static int wb_priority(struct writeback_control *wbc)
 390 {
 391         int ret = 0;
 392
 393         if (wbc->sync_mode == WB_SYNC_ALL)
 394                 ret = FLUSH_COND_STABLE;
 395         return ret;
 396 }
 397
 398 /*
 399  * NFS congestion control
 400  */
 401
 402 int nfs_congestion_kb;
 403
 404 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 405 #define NFS_CONGESTION_OFF_THRESH       \
 406         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 407
 408 static void nfs_set_page_writeback(struct page *page)
 409 {
 410         struct inode *inode = page_file_mapping(page)->host;
 411         struct nfs_server *nfss = NFS_SERVER(inode);
 412         int ret = test_set_page_writeback(page);
 413
 414         WARN_ON_ONCE(ret != 0);
 415
 416         if (atomic_long_inc_return(&nfss->writeback) >
 417                         NFS_CONGESTION_ON_THRESH)
 418                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 419 }
 420
 421 static void nfs_end_page_writeback(struct nfs_page *req)
 422 {
 423         struct inode *inode = page_file_mapping(req->wb_page)->host;
 424         struct nfs_server *nfss = NFS_SERVER(inode);
 425         bool is_done;
 426
 427         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 428         nfs_unlock_request(req);
 429         if (!is_done)
 430                 return;
 431
 432         end_page_writeback(req->wb_page);
 433         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 434                 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
 435 }
 436
 437 /*
 438  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 439  *
 440  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 441  * @old_head - the old head of the list
 442  *
 443  * All subrequests must be locked and removed from all lists, so at this point
 444  * they are only "active" in this function, and possibly in nfs_wait_on_request
 445  * with a reference held by some other context.
 446  */
 447 static void
 448 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 449                                  struct nfs_page *old_head,
 450                                  struct inode *inode)
 451 {
 452         while (destroy_list) {
 453                 struct nfs_page *subreq = destroy_list;
 454
 455                 destroy_list = (subreq->wb_this_page == old_head) ?
 456                                    NULL : subreq->wb_this_page;
 457
 458                 /* Note: lock subreq in order to change subreq->wb_head */
 459                 nfs_page_set_headlock(subreq);
 460                 WARN_ON_ONCE(old_head != subreq->wb_head);
 461
 462                 /* make sure old group is not used */
 463                 subreq->wb_this_page = subreq;
 464                 subreq->wb_head = subreq;
 465
 466                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 467
 468                 /* Note: races with nfs_page_group_destroy() */
 469                 if (!kref_read(&subreq->wb_kref)) {
 470                         /* Check if we raced with nfs_page_group_destroy() */
 471                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 472                                 nfs_page_clear_headlock(subreq);
 473                                 nfs_free_request(subreq);
 474                         } else
 475                                 nfs_page_clear_headlock(subreq);
 476                         continue;
 477                 }
 478                 nfs_page_clear_headlock(subreq);
 479
 480                 nfs_release_request(old_head);
 481
 482                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 483                         nfs_release_request(subreq);
 484                         atomic_long_dec(&NFS_I(inode)->nrequests);
 485                 }
 486
 487                 /* subreq is now totally disconnected from page group or any
 488                  * write / commit lists. last chance to wake any waiters */
 489                 nfs_unlock_and_release_request(subreq);
 490         }
 491 }
 492
 493 /*
 494  * nfs_join_page_group - destroy subrequests of the head req
 495  * @head: the page used to lookup the "page group" of nfs_page structures
 496  * @inode: Inode to which the request belongs.
 497  *
 498  * This function joins all sub requests to the head request by first
 499  * locking all requests in the group, cancelling any pending operations
 500  * and finally updating the head request to cover the whole range covered by
 501  * the (former) group.  All subrequests are removed from any write or commit
 502  * lists, unlinked from the group and destroyed.
 503  */
 504 void
 505 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
 506 {
 507         struct nfs_page *subreq;
 508         struct nfs_page *destroy_list = NULL;
 509         unsigned int pgbase, off, bytes;
 510
 511         pgbase = head->wb_pgbase;
 512         bytes = head->wb_bytes;
 513         off = head->wb_offset;
 514         for (subreq = head->wb_this_page; subreq != head;
 515                         subreq = subreq->wb_this_page) {
 516                 /* Subrequests should always form a contiguous range */
 517                 if (pgbase > subreq->wb_pgbase) {
 518                         off -= pgbase - subreq->wb_pgbase;
 519                         bytes += pgbase - subreq->wb_pgbase;
 520                         pgbase = subreq->wb_pgbase;
 521                 }
 522                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 523                                 - pgbase, bytes);
 524         }
 525
 526         /* Set the head request's range to cover the former page group */
 527         head->wb_pgbase = pgbase;
 528         head->wb_bytes = bytes;
 529         head->wb_offset = off;
 530
 531         /* Now that all requests are locked, make sure they aren't on any list.
 532          * Commit list removal accounting is done after locks are dropped */
 533         subreq = head;
 534         do {
 535                 nfs_clear_request_commit(subreq);
 536                 subreq = subreq->wb_this_page;
 537         } while (subreq != head);
 538
 539         /* unlink subrequests from head, destroy them later */
 540         if (head->wb_this_page != head) {
 541                 /* destroy list will be terminated by head */
 542                 destroy_list = head->wb_this_page;
 543                 head->wb_this_page = head;
 544         }
 545
 546         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 547 }
 548
 549 /*
 550  * nfs_lock_and_join_requests - join all subreqs to the head req
 551  * @page: the page used to lookup the "page group" of nfs_page structures
 552  *
 553  * This function joins all sub requests to the head request by first
 554  * locking all requests in the group, cancelling any pending operations
 555  * and finally updating the head request to cover the whole range covered by
 556  * the (former) group.  All subrequests are removed from any write or commit
 557  * lists, unlinked from the group and destroyed.
 558  *
 559  * Returns a locked, referenced pointer to the head request - which after
 560  * this call is guaranteed to be the only request associated with the page.
 561  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 562  * error was encountered.
 563  */
 564 static struct nfs_page *
 565 nfs_lock_and_join_requests(struct page *page)
 566 {
 567         struct inode *inode = page_file_mapping(page)->host;
 568         struct nfs_page *head;
 569         int ret;
 570
 571         /*
 572          * A reference is taken only on the head request which acts as a
 573          * reference to the whole page group - the group will not be destroyed
 574          * until the head reference is released.
 575          */
 576         head = nfs_find_and_lock_page_request(page);
 577         if (IS_ERR_OR_NULL(head))
 578                 return head;
 579
 580         /* lock each request in the page group */
 581         ret = nfs_page_group_lock_subrequests(head);
 582         if (ret < 0) {
 583                 nfs_unlock_and_release_request(head);
 584                 return ERR_PTR(ret);
 585         }
 586
 587         nfs_join_page_group(head, inode);
 588
 589         return head;
 590 }
 591
 592 static void nfs_write_error(struct nfs_page *req, int error)
 593 {
 594         trace_nfs_write_error(req, error);
 595         nfs_mapping_set_error(req->wb_page, error);
 596         nfs_inode_remove_request(req);
 597         nfs_end_page_writeback(req);
 598         nfs_release_request(req);
 599 }
 600
 601 /*
 602  * Find an associated nfs write request, and prepare to flush it out
 603  * May return an error if the user signalled nfs_wait_on_request().
 604  */
 605 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 606                                 struct page *page)
 607 {
 608         struct nfs_page *req;
 609         int ret = 0;
 610
 611         req = nfs_lock_and_join_requests(page);
 612         if (!req)
 613                 goto out;
 614         ret = PTR_ERR(req);
 615         if (IS_ERR(req))
 616                 goto out;
 617
 618         nfs_set_page_writeback(page);
 619         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 620
 621         /* If there is a fatal error that covers this write, just exit */
 622         ret = pgio->pg_error;
 623         if (nfs_error_is_fatal_on_server(ret))
 624                 goto out_launder;
 625
 626         ret = 0;
 627         if (!nfs_pageio_add_request(pgio, req)) {
 628                 ret = pgio->pg_error;
 629                 /*
 630                  * Remove the problematic req upon fatal errors on the server
 631                  */
 632                 if (nfs_error_is_fatal(ret)) {
 633                         if (nfs_error_is_fatal_on_server(ret))
 634                                 goto out_launder;
 635                 } else
 636                         ret = -EAGAIN;
 637                 nfs_redirty_request(req);
 638                 pgio->pg_error = 0;
 639         } else
 640                 nfs_add_stats(page_file_mapping(page)->host,
 641                                 NFSIOS_WRITEPAGES, 1);
 642 out:
 643         return ret;
 644 out_launder:
 645         nfs_write_error(req, ret);
 646         return 0;
 647 }
 648
 649 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 650                             struct nfs_pageio_descriptor *pgio)
 651 {
 652         int ret;
 653
 654         nfs_pageio_cond_complete(pgio, page_index(page));
 655         ret = nfs_page_async_flush(pgio, page);
 656         if (ret == -EAGAIN) {
 657                 redirty_page_for_writepage(wbc, page);
 658                 ret = AOP_WRITEPAGE_ACTIVATE;
 659         }
 660         return ret;
 661 }
 662
 663 /*
 664  * Write an mmapped page to the server.
 665  */
 666 static int nfs_writepage_locked(struct page *page,
 667                                 struct writeback_control *wbc)
 668 {
 669         struct nfs_pageio_descriptor pgio;
 670         struct inode *inode = page_file_mapping(page)->host;
 671         int err;
 672
 673         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 674         nfs_pageio_init_write(&pgio, inode, 0,
 675                                 false, &nfs_async_write_completion_ops);
 676         err = nfs_do_writepage(page, wbc, &pgio);
 677         pgio.pg_error = 0;
 678         nfs_pageio_complete(&pgio);
 679         if (err < 0)
 680                 return err;
 681         if (nfs_error_is_fatal(pgio.pg_error))
 682                 return pgio.pg_error;
 683         return 0;
 684 }
 685
 686 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 687 {
 688         int ret;
 689
 690         ret = nfs_writepage_locked(page, wbc);
 691         if (ret != AOP_WRITEPAGE_ACTIVATE)
 692                 unlock_page(page);
 693         return ret;
 694 }
 695
 696 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 697 {
 698         int ret;
 699
 700         ret = nfs_do_writepage(page, wbc, data);
 701         if (ret != AOP_WRITEPAGE_ACTIVATE)
 702                 unlock_page(page);
 703         return ret;
 704 }
 705
 706 static void nfs_io_completion_commit(void *inode)
 707 {
 708         nfs_commit_inode(inode, 0);
 709 }
 710
 711 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 712 {
 713         struct inode *inode = mapping->host;
 714         struct nfs_pageio_descriptor pgio;
 715         struct nfs_io_completion *ioc = NULL;
 716         unsigned int mntflags = NFS_SERVER(inode)->flags;
 717         int priority = 0;
 718         int err;
 719
 720         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 721
 722         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 723             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 724                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 725                 if (ioc)
 726                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 727                                                inode);
 728                 priority = wb_priority(wbc);
 729         }
 730
 731         nfs_pageio_init_write(&pgio, inode, priority, false,
 732                                 &nfs_async_write_completion_ops);
 733         pgio.pg_io_completion = ioc;
 734         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 735         pgio.pg_error = 0;
 736         nfs_pageio_complete(&pgio);
 737         nfs_io_completion_put(ioc);
 738
 739         if (err < 0)
 740                 goto out_err;
 741         err = pgio.pg_error;
 742         if (nfs_error_is_fatal(err))
 743                 goto out_err;
 744         return 0;
 745 out_err:
 746         return err;
 747 }
 748
 749 /*
 750  * Insert a write request into an inode
 751  */
 752 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 753 {
 754         struct address_space *mapping = page_file_mapping(req->wb_page);
 755         struct nfs_inode *nfsi = NFS_I(inode);
 756
 757         WARN_ON_ONCE(req->wb_this_page != req);
 758
 759         /* Lock the request! */
 760         nfs_lock_request(req);
 761
 762         /*
 763          * Swap-space should not get truncated. Hence no need to plug the race
 764          * with invalidate/truncate.
 765          */
 766         spin_lock(&mapping->private_lock);
 767         if (!nfs_have_writebacks(inode) &&
 768             NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 769                 inode_inc_iversion_raw(inode);
 770         if (likely(!PageSwapCache(req->wb_page))) {
 771                 set_bit(PG_MAPPED, &req->wb_flags);
 772                 SetPagePrivate(req->wb_page);
 773                 set_page_private(req->wb_page, (unsigned long)req);
 774         }
 775         spin_unlock(&mapping->private_lock);
 776         atomic_long_inc(&nfsi->nrequests);
 777         /* this a head request for a page group - mark it as having an
 778          * extra reference so sub groups can follow suit.
 779          * This flag also informs pgio layer when to bump nrequests when
 780          * adding subrequests. */
 781         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 782         kref_get(&req->wb_kref);
 783 }
 784
 785 /*
 786  * Remove a write request from an inode
 787  */
 788 static void nfs_inode_remove_request(struct nfs_page *req)
 789 {
 790         struct address_space *mapping = page_file_mapping(req->wb_page);
 791         struct inode *inode = mapping->host;
 792         struct nfs_inode *nfsi = NFS_I(inode);
 793         struct nfs_page *head;
 794
 795         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 796                 head = req->wb_head;
 797
 798                 spin_lock(&mapping->private_lock);
 799                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 800                         set_page_private(head->wb_page, 0);
 801                         ClearPagePrivate(head->wb_page);
 802                         clear_bit(PG_MAPPED, &head->wb_flags);
 803                 }
 804                 spin_unlock(&mapping->private_lock);
 805         }
 806
 807         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 808                 nfs_release_request(req);
 809                 atomic_long_dec(&nfsi->nrequests);
 810         }
 811 }
 812
 813 static void
 814 nfs_mark_request_dirty(struct nfs_page *req)
 815 {
 816         if (req->wb_page)
 817                 __set_page_dirty_nobuffers(req->wb_page);
 818 }
 819
 820 /*
 821  * nfs_page_search_commits_for_head_request_locked
 822  *
 823  * Search through commit lists on @inode for the head request for @page.
 824  * Must be called while holding the inode (which is cinfo) lock.
 825  *
 826  * Returns the head request if found, or NULL if not found.
 827  */
 828 static struct nfs_page *
 829 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 830                                                 struct page *page)
 831 {
 832         struct nfs_page *freq, *t;
 833         struct nfs_commit_info cinfo;
 834         struct inode *inode = &nfsi->vfs_inode;
 835
 836         nfs_init_cinfo_from_inode(&cinfo, inode);
 837
 838         /* search through pnfs commit lists */
 839         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 840         if (freq)
 841                 return freq->wb_head;
 842
 843         /* Linearly search the commit list for the correct request */
 844         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 845                 if (freq->wb_page == page)
 846                         return freq->wb_head;
 847         }
 848
 849         return NULL;
 850 }
 851
 852 /**
 853  * nfs_request_add_commit_list_locked - add request to a commit list
 854  * @req: pointer to a struct nfs_page
 855  * @dst: commit list head
 856  * @cinfo: holds list lock and accounting info
 857  *
 858  * This sets the PG_CLEAN bit, updates the cinfo count of
 859  * number of outstanding requests requiring a commit as well as
 860  * the MM page stats.
 861  *
 862  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 863  * nfs_page lock.
 864  */
 865 void
 866 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 867                             struct nfs_commit_info *cinfo)
 868 {
 869         set_bit(PG_CLEAN, &req->wb_flags);
 870         nfs_list_add_request(req, dst);
 871         atomic_long_inc(&cinfo->mds->ncommit);
 872 }
 873 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 874
 875 /**
 876  * nfs_request_add_commit_list - add request to a commit list
 877  * @req: pointer to a struct nfs_page
 878  * @cinfo: holds list lock and accounting info
 879  *
 880  * This sets the PG_CLEAN bit, updates the cinfo count of
 881  * number of outstanding requests requiring a commit as well as
 882  * the MM page stats.
 883  *
 884  * The caller must _not_ hold the cinfo->lock, but must be
 885  * holding the nfs_page lock.
 886  */
 887 void
 888 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 889 {
 890         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 891         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 892         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 893         if (req->wb_page)
 894                 nfs_mark_page_unstable(req->wb_page, cinfo);
 895 }
 896 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 897
 898 /**
 899  * nfs_request_remove_commit_list - Remove request from a commit list
 900  * @req: pointer to a nfs_page
 901  * @cinfo: holds list lock and accounting info
 902  *
 903  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 904  * number of outstanding requests requiring a commit
 905  * It does not update the MM page stats.
 906  *
 907  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 908  */
 909 void
 910 nfs_request_remove_commit_list(struct nfs_page *req,
 911                                struct nfs_commit_info *cinfo)
 912 {
 913         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 914                 return;
 915         nfs_list_remove_request(req);
 916         atomic_long_dec(&cinfo->mds->ncommit);
 917 }
 918 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 919
 920 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 921                                       struct inode *inode)
 922 {
 923         cinfo->inode = inode;
 924         cinfo->mds = &NFS_I(inode)->commit_info;
 925         cinfo->ds = pnfs_get_ds_info(inode);
 926         cinfo->dreq = NULL;
 927         cinfo->completion_ops = &nfs_commit_completion_ops;
 928 }
 929
 930 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 931                     struct inode *inode,
 932                     struct nfs_direct_req *dreq)
 933 {
 934         if (dreq)
 935                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 936         else
 937                 nfs_init_cinfo_from_inode(cinfo, inode);
 938 }
 939 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 940
 941 /*
 942  * Add a request to the inode's commit list.
 943  */
 944 void
 945 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 946                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 947 {
 948         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 949                 return;
 950         nfs_request_add_commit_list(req, cinfo);
 951 }
 952
 953 static void
 954 nfs_clear_page_commit(struct page *page)
 955 {
 956         dec_node_page_state(page, NR_WRITEBACK);
 957         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 958                     WB_WRITEBACK);
 959 }
 960
 961 /* Called holding the request lock on @req */
 962 static void
 963 nfs_clear_request_commit(struct nfs_page *req)
 964 {
 965         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 966                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 967                 struct inode *inode = d_inode(ctx->dentry);
 968                 struct nfs_commit_info cinfo;
 969
 970                 nfs_init_cinfo_from_inode(&cinfo, inode);
 971                 mutex_lock(&NFS_I(inode)->commit_mutex);
 972                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 973                         nfs_request_remove_commit_list(req, &cinfo);
 974                 }
 975                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 976                 nfs_clear_page_commit(req->wb_page);
 977         }
 978 }
 979
 980 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 981 {
 982         if (hdr->verf.committed == NFS_DATA_SYNC)
 983                 return hdr->lseg == NULL;
 984         return hdr->verf.committed != NFS_FILE_SYNC;
 985 }
 986
 987 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 988 {
 989         nfs_io_completion_get(hdr->io_completion);
 990 }
 991
 992 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 993 {
 994         struct nfs_commit_info cinfo;
 995         unsigned long bytes = 0;
 996
 997         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 998                 goto out;
 999         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1000         while (!list_empty(&hdr->pages)) {
1001                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1002
1003                 bytes += req->wb_bytes;
1004                 nfs_list_remove_request(req);
1005                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1006                     (hdr->good_bytes < bytes)) {
1007                         trace_nfs_comp_error(req, hdr->error);
1008                         nfs_mapping_set_error(req->wb_page, hdr->error);
1009                         goto remove_req;
1010                 }
1011                 if (nfs_write_need_commit(hdr)) {
1012                         /* Reset wb_nio, since the write was successful. */
1013                         req->wb_nio = 0;
1014                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1015                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1016                                 hdr->pgio_mirror_idx);
1017                         goto next;
1018                 }
1019 remove_req:
1020                 nfs_inode_remove_request(req);
1021 next:
1022                 nfs_end_page_writeback(req);
1023                 nfs_release_request(req);
1024         }
1025 out:
1026         nfs_io_completion_put(hdr->io_completion);
1027         hdr->release(hdr);
1028 }
1029
1030 unsigned long
1031 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1032 {
1033         return atomic_long_read(&cinfo->mds->ncommit);
1034 }
1035
1036 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1037 int
1038 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1039                      struct nfs_commit_info *cinfo, int max)
1040 {
1041         struct nfs_page *req, *tmp;
1042         int ret = 0;
1043
1044 restart:
1045         list_for_each_entry_safe(req, tmp, src, wb_list) {
1046                 kref_get(&req->wb_kref);
1047                 if (!nfs_lock_request(req)) {
1048                         int status;
1049
1050                         /* Prevent deadlock with nfs_lock_and_join_requests */
1051                         if (!list_empty(dst)) {
1052                                 nfs_release_request(req);
1053                                 continue;
1054                         }
1055                         /* Ensure we make progress to prevent livelock */
1056                         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1057                         status = nfs_wait_on_request(req);
1058                         nfs_release_request(req);
1059                         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1060                         if (status < 0)
1061                                 break;
1062                         goto restart;
1063                 }
1064                 nfs_request_remove_commit_list(req, cinfo);
1065                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1066                 nfs_list_add_request(req, dst);
1067                 ret++;
1068                 if ((ret == max) && !cinfo->dreq)
1069                         break;
1070                 cond_resched();
1071         }
1072         return ret;
1073 }
1074 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1075
1076 /*
1077  * nfs_scan_commit - Scan an inode for commit requests
1078  * @inode: NFS inode to scan
1079  * @dst: mds destination list
1080  * @cinfo: mds and ds lists of reqs ready to commit
1081  *
1082  * Moves requests from the inode's 'commit' request list.
1083  * The requests are *not* checked to ensure that they form a contiguous set.
1084  */
1085 int
1086 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1087                 struct nfs_commit_info *cinfo)
1088 {
1089         int ret = 0;
1090
1091         if (!atomic_long_read(&cinfo->mds->ncommit))
1092                 return 0;
1093         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1094         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1095                 const int max = INT_MAX;
1096
1097                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1098                                            cinfo, max);
1099                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1100         }
1101         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1102         return ret;
1103 }
1104
1105 /*
1106  * Search for an existing write request, and attempt to update
1107  * it to reflect a new dirty region on a given page.
1108  *
1109  * If the attempt fails, then the existing request is flushed out
1110  * to disk.
1111  */
1112 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1113                 struct page *page,
1114                 unsigned int offset,
1115                 unsigned int bytes)
1116 {
1117         struct nfs_page *req;
1118         unsigned int rqend;
1119         unsigned int end;
1120         int error;
1121
1122         end = offset + bytes;
1123
1124         req = nfs_lock_and_join_requests(page);
1125         if (IS_ERR_OR_NULL(req))
1126                 return req;
1127
1128         rqend = req->wb_offset + req->wb_bytes;
1129         /*
1130          * Tell the caller to flush out the request if
1131          * the offsets are non-contiguous.
1132          * Note: nfs_flush_incompatible() will already
1133          * have flushed out requests having wrong owners.
1134          */
1135         if (offset > rqend || end < req->wb_offset)
1136                 goto out_flushme;
1137
1138         /* Okay, the request matches. Update the region */
1139         if (offset < req->wb_offset) {
1140                 req->wb_offset = offset;
1141                 req->wb_pgbase = offset;
1142         }
1143         if (end > rqend)
1144                 req->wb_bytes = end - req->wb_offset;
1145         else
1146                 req->wb_bytes = rqend - req->wb_offset;
1147         req->wb_nio = 0;
1148         return req;
1149 out_flushme:
1150         /*
1151          * Note: we mark the request dirty here because
1152          * nfs_lock_and_join_requests() cannot preserve
1153          * commit flags, so we have to replay the write.
1154          */
1155         nfs_mark_request_dirty(req);
1156         nfs_unlock_and_release_request(req);
1157         error = nfs_wb_page(inode, page);
1158         return (error < 0) ? ERR_PTR(error) : NULL;
1159 }
1160
1161 /*
1162  * Try to update an existing write request, or create one if there is none.
1163  *
1164  * Note: Should always be called with the Page Lock held to prevent races
1165  * if we have to add a new request. Also assumes that the caller has
1166  * already called nfs_flush_incompatible() if necessary.
1167  */
1168 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1169                 struct page *page, unsigned int offset, unsigned int bytes)
1170 {
1171         struct inode *inode = page_file_mapping(page)->host;
1172         struct nfs_page *req;
1173
1174         req = nfs_try_to_update_request(inode, page, offset, bytes);
1175         if (req != NULL)
1176                 goto out;
1177         req = nfs_create_request(ctx, page, offset, bytes);
1178         if (IS_ERR(req))
1179                 goto out;
1180         nfs_inode_add_request(inode, req);
1181 out:
1182         return req;
1183 }
1184
1185 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1186                 unsigned int offset, unsigned int count)
1187 {
1188         struct nfs_page *req;
1189
1190         req = nfs_setup_write_request(ctx, page, offset, count);
1191         if (IS_ERR(req))
1192                 return PTR_ERR(req);
1193         /* Update file length */
1194         nfs_grow_file(page, offset, count);
1195         nfs_mark_uptodate(req);
1196         nfs_mark_request_dirty(req);
1197         nfs_unlock_and_release_request(req);
1198         return 0;
1199 }
1200
1201 int nfs_flush_incompatible(struct file *file, struct page *page)
1202 {
1203         struct nfs_open_context *ctx = nfs_file_open_context(file);
1204         struct nfs_lock_context *l_ctx;
1205         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1206         struct nfs_page *req;
1207         int do_flush, status;
1208         /*
1209          * Look for a request corresponding to this page. If there
1210          * is one, and it belongs to another file, we flush it out
1211          * before we try to copy anything into the page. Do this
1212          * due to the lack of an ACCESS-type call in NFSv2.
1213          * Also do the same if we find a request from an existing
1214          * dropped page.
1215          */
1216         do {
1217                 req = nfs_page_find_head_request(page);
1218                 if (req == NULL)
1219                         return 0;
1220                 l_ctx = req->wb_lock_context;
1221                 do_flush = req->wb_page != page ||
1222                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1223                 if (l_ctx && flctx &&
1224                     !(list_empty_careful(&flctx->flc_posix) &&
1225                       list_empty_careful(&flctx->flc_flock))) {
1226                         do_flush |= l_ctx->lockowner != current->files;
1227                 }
1228                 nfs_release_request(req);
1229                 if (!do_flush)
1230                         return 0;
1231                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1232         } while (status == 0);
1233         return status;
1234 }
1235
1236 /*
1237  * Avoid buffered writes when a open context credential's key would
1238  * expire soon.
1239  *
1240  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1241  *
1242  * Return 0 and set a credential flag which triggers the inode to flush
1243  * and performs  NFS_FILE_SYNC writes if the key will expired within
1244  * RPC_KEY_EXPIRE_TIMEO.
1245  */
1246 int
1247 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1248 {
1249         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1250
1251         if (nfs_ctx_key_to_expire(ctx, inode) &&
1252             !ctx->ll_cred)
1253                 /* Already expired! */
1254                 return -EACCES;
1255         return 0;
1256 }
1257
1258 /*
1259  * Test if the open context credential key is marked to expire soon.
1260  */
1261 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1262 {
1263         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1264         struct rpc_cred *cred = ctx->ll_cred;
1265         struct auth_cred acred = {
1266                 .cred = ctx->cred,
1267         };
1268
1269         if (cred && !cred->cr_ops->crmatch(&acred, cred, 0)) {
1270                 put_rpccred(cred);
1271                 ctx->ll_cred = NULL;
1272                 cred = NULL;
1273         }
1274         if (!cred)
1275                 cred = auth->au_ops->lookup_cred(auth, &acred, 0);
1276         if (!cred || IS_ERR(cred))
1277                 return true;
1278         ctx->ll_cred = cred;
1279         return !!(cred->cr_ops->crkey_timeout &&
1280                   cred->cr_ops->crkey_timeout(cred));
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_REVAL_PAGECACHE | 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         else
1389                 __set_page_dirty_nobuffers(page);
1390 out:
1391         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1392                         status, (long long)i_size_read(inode));
1393         return status;
1394 }
1395
1396 static int flush_task_priority(int how)
1397 {
1398         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1399                 case FLUSH_HIGHPRI:
1400                         return RPC_PRIORITY_HIGH;
1401                 case FLUSH_LOWPRI:
1402                         return RPC_PRIORITY_LOW;
1403         }
1404         return RPC_PRIORITY_NORMAL;
1405 }
1406
1407 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1408                                struct rpc_message *msg,
1409                                const struct nfs_rpc_ops *rpc_ops,
1410                                struct rpc_task_setup *task_setup_data, int how)
1411 {
1412         int priority = flush_task_priority(how);
1413
1414         task_setup_data->priority = priority;
1415         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1416         trace_nfs_initiate_write(hdr);
1417 }
1418
1419 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1420  * call this on each, which will prepare them to be retried on next
1421  * writeback using standard nfs.
1422  */
1423 static void nfs_redirty_request(struct nfs_page *req)
1424 {
1425         /* Bump the transmission count */
1426         req->wb_nio++;
1427         nfs_mark_request_dirty(req);
1428         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1429         nfs_end_page_writeback(req);
1430         nfs_release_request(req);
1431 }
1432
1433 static void nfs_async_write_error(struct list_head *head, int error)
1434 {
1435         struct nfs_page *req;
1436
1437         while (!list_empty(head)) {
1438                 req = nfs_list_entry(head->next);
1439                 nfs_list_remove_request(req);
1440                 if (nfs_error_is_fatal(error))
1441                         nfs_write_error(req, error);
1442                 else
1443                         nfs_redirty_request(req);
1444         }
1445 }
1446
1447 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1448 {
1449         nfs_async_write_error(&hdr->pages, 0);
1450         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1451                         hdr->args.offset + hdr->args.count - 1);
1452 }
1453
1454 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1455         .init_hdr = nfs_async_write_init,
1456         .error_cleanup = nfs_async_write_error,
1457         .completion = nfs_write_completion,
1458         .reschedule_io = nfs_async_write_reschedule_io,
1459 };
1460
1461 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1462                                struct inode *inode, int ioflags, bool force_mds,
1463                                const struct nfs_pgio_completion_ops *compl_ops)
1464 {
1465         struct nfs_server *server = NFS_SERVER(inode);
1466         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1467
1468 #ifdef CONFIG_NFS_V4_1
1469         if (server->pnfs_curr_ld && !force_mds)
1470                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1471 #endif
1472         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1473                         server->wsize, ioflags);
1474 }
1475 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1476
1477 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1478 {
1479         struct nfs_pgio_mirror *mirror;
1480
1481         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1482                 pgio->pg_ops->pg_cleanup(pgio);
1483
1484         pgio->pg_ops = &nfs_pgio_rw_ops;
1485
1486         nfs_pageio_stop_mirroring(pgio);
1487
1488         mirror = &pgio->pg_mirrors[0];
1489         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1490 }
1491 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1492
1493
1494 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1495 {
1496         struct nfs_commit_data *data = calldata;
1497
1498         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1499 }
1500
1501 /*
1502  * Special version of should_remove_suid() that ignores capabilities.
1503  */
1504 static int nfs_should_remove_suid(const struct inode *inode)
1505 {
1506         umode_t mode = inode->i_mode;
1507         int kill = 0;
1508
1509         /* suid always must be killed */
1510         if (unlikely(mode & S_ISUID))
1511                 kill = ATTR_KILL_SUID;
1512
1513         /*
1514          * sgid without any exec bits is just a mandatory locking mark; leave
1515          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1516          */
1517         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1518                 kill |= ATTR_KILL_SGID;
1519
1520         if (unlikely(kill && S_ISREG(mode)))
1521                 return kill;
1522
1523         return 0;
1524 }
1525
1526 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1527                 struct nfs_fattr *fattr)
1528 {
1529         struct nfs_pgio_args *argp = &hdr->args;
1530         struct nfs_pgio_res *resp = &hdr->res;
1531         u64 size = argp->offset + resp->count;
1532
1533         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1534                 fattr->size = size;
1535         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1536                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1537                 return;
1538         }
1539         if (size != fattr->size)
1540                 return;
1541         /* Set attribute barrier */
1542         nfs_fattr_set_barrier(fattr);
1543         /* ...and update size */
1544         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1545 }
1546
1547 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1548 {
1549         struct nfs_fattr *fattr = &hdr->fattr;
1550         struct inode *inode = hdr->inode;
1551
1552         spin_lock(&inode->i_lock);
1553         nfs_writeback_check_extend(hdr, fattr);
1554         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1555         spin_unlock(&inode->i_lock);
1556 }
1557 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1558
1559 /*
1560  * This function is called when the WRITE call is complete.
1561  */
1562 static int nfs_writeback_done(struct rpc_task *task,
1563                               struct nfs_pgio_header *hdr,
1564                               struct inode *inode)
1565 {
1566         int status;
1567
1568         /*
1569          * ->write_done will attempt to use post-op attributes to detect
1570          * conflicting writes by other clients.  A strict interpretation
1571          * of close-to-open would allow us to continue caching even if
1572          * another writer had changed the file, but some applications
1573          * depend on tighter cache coherency when writing.
1574          */
1575         status = NFS_PROTO(inode)->write_done(task, hdr);
1576         if (status != 0)
1577                 return status;
1578
1579         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1580         trace_nfs_writeback_done(task, hdr);
1581
1582         if (hdr->res.verf->committed < hdr->args.stable &&
1583             task->tk_status >= 0) {
1584                 /* We tried a write call, but the server did not
1585                  * commit data to stable storage even though we
1586                  * requested it.
1587                  * Note: There is a known bug in Tru64 < 5.0 in which
1588                  *       the server reports NFS_DATA_SYNC, but performs
1589                  *       NFS_FILE_SYNC. We therefore implement this checking
1590                  *       as a dprintk() in order to avoid filling syslog.
1591                  */
1592                 static unsigned long    complain;
1593
1594                 /* Note this will print the MDS for a DS write */
1595                 if (time_before(complain, jiffies)) {
1596                         dprintk("NFS:       faulty NFS server %s:"
1597                                 " (committed = %d) != (stable = %d)\n",
1598                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1599                                 hdr->res.verf->committed, hdr->args.stable);
1600                         complain = jiffies + 300 * HZ;
1601                 }
1602         }
1603
1604         /* Deal with the suid/sgid bit corner case */
1605         if (nfs_should_remove_suid(inode)) {
1606                 spin_lock(&inode->i_lock);
1607                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_OTHER;
1608                 spin_unlock(&inode->i_lock);
1609         }
1610         return 0;
1611 }
1612
1613 /*
1614  * This function is called when the WRITE call is complete.
1615  */
1616 static void nfs_writeback_result(struct rpc_task *task,
1617                                  struct nfs_pgio_header *hdr)
1618 {
1619         struct nfs_pgio_args    *argp = &hdr->args;
1620         struct nfs_pgio_res     *resp = &hdr->res;
1621
1622         if (resp->count < argp->count) {
1623                 static unsigned long    complain;
1624
1625                 /* This a short write! */
1626                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1627
1628                 /* Has the server at least made some progress? */
1629                 if (resp->count == 0) {
1630                         if (time_before(complain, jiffies)) {
1631                                 printk(KERN_WARNING
1632                                        "NFS: Server wrote zero bytes, expected %u.\n",
1633                                        argp->count);
1634                                 complain = jiffies + 300 * HZ;
1635                         }
1636                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1637                         task->tk_status = -EIO;
1638                         return;
1639                 }
1640
1641                 /* For non rpc-based layout drivers, retry-through-MDS */
1642                 if (!task->tk_ops) {
1643                         hdr->pnfs_error = -EAGAIN;
1644                         return;
1645                 }
1646
1647                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1648                 if (resp->verf->committed != NFS_UNSTABLE) {
1649                         /* Resend from where the server left off */
1650                         hdr->mds_offset += resp->count;
1651                         argp->offset += resp->count;
1652                         argp->pgbase += resp->count;
1653                         argp->count -= resp->count;
1654                 } else {
1655                         /* Resend as a stable write in order to avoid
1656                          * headaches in the case of a server crash.
1657                          */
1658                         argp->stable = NFS_FILE_SYNC;
1659                 }
1660                 resp->count = 0;
1661                 resp->verf->committed = 0;
1662                 rpc_restart_call_prepare(task);
1663         }
1664 }
1665
1666 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1667 {
1668         return wait_var_event_killable(&cinfo->rpcs_out,
1669                                        !atomic_read(&cinfo->rpcs_out));
1670 }
1671
1672 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1673 {
1674         atomic_inc(&cinfo->rpcs_out);
1675 }
1676
1677 static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1678 {
1679         if (atomic_dec_and_test(&cinfo->rpcs_out))
1680                 wake_up_var(&cinfo->rpcs_out);
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 }
1781 EXPORT_SYMBOL_GPL(nfs_init_commit);
1782
1783 void nfs_retry_commit(struct list_head *page_list,
1784                       struct pnfs_layout_segment *lseg,
1785                       struct nfs_commit_info *cinfo,
1786                       u32 ds_commit_idx)
1787 {
1788         struct nfs_page *req;
1789
1790         while (!list_empty(page_list)) {
1791                 req = nfs_list_entry(page_list->next);
1792                 nfs_list_remove_request(req);
1793                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1794                 if (!cinfo->dreq)
1795                         nfs_clear_page_commit(req->wb_page);
1796                 nfs_unlock_and_release_request(req);
1797         }
1798 }
1799 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1800
1801 static void
1802 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1803                 struct nfs_page *req)
1804 {
1805         __set_page_dirty_nobuffers(req->wb_page);
1806 }
1807
1808 /*
1809  * Commit dirty pages
1810  */
1811 static int
1812 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1813                 struct nfs_commit_info *cinfo)
1814 {
1815         struct nfs_commit_data  *data;
1816
1817         /* another commit raced with us */
1818         if (list_empty(head))
1819                 return 0;
1820
1821         data = nfs_commitdata_alloc(true);
1822
1823         /* Set up the argument struct */
1824         nfs_init_commit(data, head, NULL, cinfo);
1825         atomic_inc(&cinfo->mds->rpcs_out);
1826         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1827                                    data->mds_ops, how, RPC_TASK_CRED_NOREF);
1828 }
1829
1830 /*
1831  * COMMIT call returned
1832  */
1833 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1834 {
1835         struct nfs_commit_data  *data = calldata;
1836
1837         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1838                                 task->tk_pid, task->tk_status);
1839
1840         /* Call the NFS version-specific code */
1841         NFS_PROTO(data->inode)->commit_done(task, data);
1842         trace_nfs_commit_done(task, data);
1843 }
1844
1845 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1846 {
1847         const struct nfs_writeverf *verf = data->res.verf;
1848         struct nfs_page *req;
1849         int status = data->task.tk_status;
1850         struct nfs_commit_info cinfo;
1851         struct nfs_server *nfss;
1852
1853         while (!list_empty(&data->pages)) {
1854                 req = nfs_list_entry(data->pages.next);
1855                 nfs_list_remove_request(req);
1856                 if (req->wb_page)
1857                         nfs_clear_page_commit(req->wb_page);
1858
1859                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1860                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1861                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1862                         req->wb_bytes,
1863                         (long long)req_offset(req));
1864                 if (status < 0) {
1865                         if (req->wb_page) {
1866                                 trace_nfs_commit_error(req, status);
1867                                 nfs_mapping_set_error(req->wb_page, status);
1868                                 nfs_inode_remove_request(req);
1869                         }
1870                         dprintk_cont(", error = %d\n", status);
1871                         goto next;
1872                 }
1873
1874                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1875                  * returned by the server against all stored verfs. */
1876                 if (nfs_write_match_verf(verf, req)) {
1877                         /* We have a match */
1878                         if (req->wb_page)
1879                                 nfs_inode_remove_request(req);
1880                         dprintk_cont(" OK\n");
1881                         goto next;
1882                 }
1883                 /* We have a mismatch. Write the page again */
1884                 dprintk_cont(" mismatch\n");
1885                 nfs_mark_request_dirty(req);
1886                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1887         next:
1888                 nfs_unlock_and_release_request(req);
1889                 /* Latency breaker */
1890                 cond_resched();
1891         }
1892         nfss = NFS_SERVER(data->inode);
1893         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1894                 clear_bdi_congested(inode_to_bdi(data->inode), BLK_RW_ASYNC);
1895
1896         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1897         nfs_commit_end(cinfo.mds);
1898 }
1899
1900 static void nfs_commit_release(void *calldata)
1901 {
1902         struct nfs_commit_data *data = calldata;
1903
1904         data->completion_ops->completion(data);
1905         nfs_commitdata_release(calldata);
1906 }
1907
1908 static const struct rpc_call_ops nfs_commit_ops = {
1909         .rpc_call_prepare = nfs_commit_prepare,
1910         .rpc_call_done = nfs_commit_done,
1911         .rpc_release = nfs_commit_release,
1912 };
1913
1914 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1915         .completion = nfs_commit_release_pages,
1916         .resched_write = nfs_commit_resched_write,
1917 };
1918
1919 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1920                             int how, struct nfs_commit_info *cinfo)
1921 {
1922         int status;
1923
1924         status = pnfs_commit_list(inode, head, how, cinfo);
1925         if (status == PNFS_NOT_ATTEMPTED)
1926                 status = nfs_commit_list(inode, head, how, cinfo);
1927         return status;
1928 }
1929
1930 static int __nfs_commit_inode(struct inode *inode, int how,
1931                 struct writeback_control *wbc)
1932 {
1933         LIST_HEAD(head);
1934         struct nfs_commit_info cinfo;
1935         int may_wait = how & FLUSH_SYNC;
1936         int ret, nscan;
1937
1938         nfs_init_cinfo_from_inode(&cinfo, inode);
1939         nfs_commit_begin(cinfo.mds);
1940         for (;;) {
1941                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1942                 if (ret <= 0)
1943                         break;
1944                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1945                 if (ret < 0)
1946                         break;
1947                 ret = 0;
1948                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1949                         if (nscan < wbc->nr_to_write)
1950                                 wbc->nr_to_write -= nscan;
1951                         else
1952                                 wbc->nr_to_write = 0;
1953                 }
1954                 if (nscan < INT_MAX)
1955                         break;
1956                 cond_resched();
1957         }
1958         nfs_commit_end(cinfo.mds);
1959         if (ret || !may_wait)
1960                 return ret;
1961         return wait_on_commit(cinfo.mds);
1962 }
1963
1964 int nfs_commit_inode(struct inode *inode, int how)
1965 {
1966         return __nfs_commit_inode(inode, how, NULL);
1967 }
1968 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1969
1970 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1971 {
1972         struct nfs_inode *nfsi = NFS_I(inode);
1973         int flags = FLUSH_SYNC;
1974         int ret = 0;
1975
1976         if (wbc->sync_mode == WB_SYNC_NONE) {
1977                 /* no commits means nothing needs to be done */
1978                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1979                         goto check_requests_outstanding;
1980
1981                 /* Don't commit yet if this is a non-blocking flush and there
1982                  * are a lot of outstanding writes for this mapping.
1983                  */
1984                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1985                         goto out_mark_dirty;
1986
1987                 /* don't wait for the COMMIT response */
1988                 flags = 0;
1989         }
1990
1991         ret = __nfs_commit_inode(inode, flags, wbc);
1992         if (!ret) {
1993                 if (flags & FLUSH_SYNC)
1994                         return 0;
1995         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1996                 goto out_mark_dirty;
1997
1998 check_requests_outstanding:
1999         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2000                 return ret;
2001 out_mark_dirty:
2002         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2003         return ret;
2004 }
2005 EXPORT_SYMBOL_GPL(nfs_write_inode);
2006
2007 /*
2008  * Wrapper for filemap_write_and_wait_range()
2009  *
2010  * Needed for pNFS in order to ensure data becomes visible to the
2011  * client.
2012  */
2013 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2014                 loff_t lstart, loff_t lend)
2015 {
2016         int ret;
2017
2018         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2019         if (ret == 0)
2020                 ret = pnfs_sync_inode(mapping->host, true);
2021         return ret;
2022 }
2023 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2024
2025 /*
2026  * flush the inode to disk.
2027  */
2028 int nfs_wb_all(struct inode *inode)
2029 {
2030         int ret;
2031
2032         trace_nfs_writeback_inode_enter(inode);
2033
2034         ret = filemap_write_and_wait(inode->i_mapping);
2035         if (ret)
2036                 goto out;
2037         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2038         if (ret < 0)
2039                 goto out;
2040         pnfs_sync_inode(inode, true);
2041         ret = 0;
2042
2043 out:
2044         trace_nfs_writeback_inode_exit(inode, ret);
2045         return ret;
2046 }
2047 EXPORT_SYMBOL_GPL(nfs_wb_all);
2048
2049 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
2050 {
2051         struct nfs_page *req;
2052         int ret = 0;
2053
2054         wait_on_page_writeback(page);
2055
2056         /* blocking call to cancel all requests and join to a single (head)
2057          * request */
2058         req = nfs_lock_and_join_requests(page);
2059
2060         if (IS_ERR(req)) {
2061                 ret = PTR_ERR(req);
2062         } else if (req) {
2063                 /* all requests from this page have been cancelled by
2064                  * nfs_lock_and_join_requests, so just remove the head
2065                  * request from the inode / page_private pointer and
2066                  * release it */
2067                 nfs_inode_remove_request(req);
2068                 nfs_unlock_and_release_request(req);
2069         }
2070
2071         return ret;
2072 }
2073
2074 /*
2075  * Write back all requests on one page - we do this before reading it.
2076  */
2077 int nfs_wb_page(struct inode *inode, struct page *page)
2078 {
2079         loff_t range_start = page_file_offset(page);
2080         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2081         struct writeback_control wbc = {
2082                 .sync_mode = WB_SYNC_ALL,
2083                 .nr_to_write = 0,
2084                 .range_start = range_start,
2085                 .range_end = range_end,
2086         };
2087         int ret;
2088
2089         trace_nfs_writeback_page_enter(inode);
2090
2091         for (;;) {
2092                 wait_on_page_writeback(page);
2093                 if (clear_page_dirty_for_io(page)) {
2094                         ret = nfs_writepage_locked(page, &wbc);
2095                         if (ret < 0)
2096                                 goto out_error;
2097                         continue;
2098                 }
2099                 ret = 0;
2100                 if (!PagePrivate(page))
2101                         break;
2102                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2103                 if (ret < 0)
2104                         goto out_error;
2105         }
2106 out_error:
2107         trace_nfs_writeback_page_exit(inode, ret);
2108         return ret;
2109 }
2110
2111 #ifdef CONFIG_MIGRATION
2112 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2113                 struct page *page, enum migrate_mode mode)
2114 {
2115         /*
2116          * If PagePrivate is set, then the page is currently associated with
2117          * an in-progress read or write request. Don't try to migrate it.
2118          *
2119          * FIXME: we could do this in principle, but we'll need a way to ensure
2120          *        that we can safely release the inode reference while holding
2121          *        the page lock.
2122          */
2123         if (PagePrivate(page))
2124                 return -EBUSY;
2125
2126         if (!nfs_fscache_release_page(page, GFP_KERNEL))
2127                 return -EBUSY;
2128
2129         return migrate_page(mapping, newpage, page, mode);
2130 }
2131 #endif
2132
2133 int __init nfs_init_writepagecache(void)
2134 {
2135         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2136                                              sizeof(struct nfs_pgio_header),
2137                                              0, SLAB_HWCACHE_ALIGN,
2138                                              NULL);
2139         if (nfs_wdata_cachep == NULL)
2140                 return -ENOMEM;
2141
2142         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2143                                                      nfs_wdata_cachep);
2144         if (nfs_wdata_mempool == NULL)
2145                 goto out_destroy_write_cache;
2146
2147         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2148                                              sizeof(struct nfs_commit_data),
2149                                              0, SLAB_HWCACHE_ALIGN,
2150                                              NULL);
2151         if (nfs_cdata_cachep == NULL)
2152                 goto out_destroy_write_mempool;
2153
2154         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2155                                                       nfs_cdata_cachep);
2156         if (nfs_commit_mempool == NULL)
2157                 goto out_destroy_commit_cache;
2158
2159         /*
2160          * NFS congestion size, scale with available memory.
2161          *
2162          *  64MB:    8192k
2163          * 128MB:   11585k
2164          * 256MB:   16384k
2165          * 512MB:   23170k
2166          *   1GB:   32768k
2167          *   2GB:   46340k
2168          *   4GB:   65536k
2169          *   8GB:   92681k
2170          *  16GB:  131072k
2171          *
2172          * This allows larger machines to have larger/more transfers.
2173          * Limit the default to 256M
2174          */
2175         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2176         if (nfs_congestion_kb > 256*1024)
2177                 nfs_congestion_kb = 256*1024;
2178
2179         return 0;
2180
2181 out_destroy_commit_cache:
2182         kmem_cache_destroy(nfs_cdata_cachep);
2183 out_destroy_write_mempool:
2184         mempool_destroy(nfs_wdata_mempool);
2185 out_destroy_write_cache:
2186         kmem_cache_destroy(nfs_wdata_cachep);
2187         return -ENOMEM;
2188 }
2189
2190 void nfs_destroy_writepagecache(void)
2191 {
2192         mempool_destroy(nfs_commit_mempool);
2193         kmem_cache_destroy(nfs_cdata_cachep);
2194         mempool_destroy(nfs_wdata_mempool);
2195         kmem_cache_destroy(nfs_wdata_cachep);
2196 }
2197
2198 static const struct nfs_rw_ops nfs_rw_write_ops = {
2199         .rw_alloc_header        = nfs_writehdr_alloc,
2200         .rw_free_header         = nfs_writehdr_free,
2201         .rw_done                = nfs_writeback_done,
2202         .rw_result              = nfs_writeback_result,
2203         .rw_initiate            = nfs_initiate_write,
2204 };