fs/nfs/write.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/write.c
   4  *
   5  * Write file data over NFS.
   6  *
   7  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/mm.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/file.h>
  15 #include <linux/writeback.h>
  16 #include <linux/swap.h>
  17 #include <linux/migrate.h>
  18
  19 #include <linux/sunrpc/clnt.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/export.h>
  25 #include <linux/freezer.h>
  26 #include <linux/wait.h>
  27 #include <linux/iversion.h>
  28
  29 #include <linux/uaccess.h>
  30 #include <linux/sched/mm.h>
  31
  32 #include "delegation.h"
  33 #include "internal.h"
  34 #include "iostat.h"
  35 #include "nfs4_fs.h"
  36 #include "fscache.h"
  37 #include "pnfs.h"
  38
  39 #include "nfstrace.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  42
  43 #define MIN_POOL_WRITE          (32)
  44 #define MIN_POOL_COMMIT         (4)
  45
  46 struct nfs_io_completion {
  47         void (*complete)(void *data);
  48         void *data;
  49         struct kref refcount;
  50 };
  51
  52 /*
  53  * Local function declarations
  54  */
  55 static void nfs_redirty_request(struct nfs_page *req);
  56 static const struct rpc_call_ops nfs_commit_ops;
  57 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  58 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  59 static const struct nfs_rw_ops nfs_rw_write_ops;
  60 static void nfs_inode_remove_request(struct nfs_page *req);
  61 static void nfs_clear_request_commit(struct nfs_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(void)
  74 {
  75         struct nfs_commit_data *p;
  76
  77         p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  78         if (!p) {
  79                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  80                 if (!p)
  81                         return NULL;
  82                 memset(p, 0, sizeof(*p));
  83         }
  84         INIT_LIST_HEAD(&p->pages);
  85         return p;
  86 }
  87 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  88
  89 void nfs_commit_free(struct nfs_commit_data *p)
  90 {
  91         mempool_free(p, nfs_commit_mempool);
  92 }
  93 EXPORT_SYMBOL_GPL(nfs_commit_free);
  94
  95 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  96 {
  97         struct nfs_pgio_header *p;
  98
  99         p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
 100         if (!p) {
 101                 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 102                 if (!p)
 103                         return NULL;
 104                 memset(p, 0, sizeof(*p));
 105         }
 106         p->rw_mode = FMODE_WRITE;
 107         return p;
 108 }
 109
 110 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 111 {
 112         mempool_free(hdr, nfs_wdata_mempool);
 113 }
 114
 115 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 116 {
 117         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 118 }
 119
 120 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 121                 void (*complete)(void *), void *data)
 122 {
 123         ioc->complete = complete;
 124         ioc->data = data;
 125         kref_init(&ioc->refcount);
 126 }
 127
 128 static void nfs_io_completion_release(struct kref *kref)
 129 {
 130         struct nfs_io_completion *ioc = container_of(kref,
 131                         struct nfs_io_completion, refcount);
 132         ioc->complete(ioc->data);
 133         kfree(ioc);
 134 }
 135
 136 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 137 {
 138         if (ioc != NULL)
 139                 kref_get(&ioc->refcount);
 140 }
 141
 142 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 143 {
 144         if (ioc != NULL)
 145                 kref_put(&ioc->refcount, nfs_io_completion_release);
 146 }
 147
 148 static void
 149 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 150 {
 151         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 152                 kref_get(&req->wb_kref);
 153                 atomic_long_inc(&NFS_I(inode)->nrequests);
 154         }
 155 }
 156
 157 static int
 158 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 159 {
 160         int ret;
 161
 162         if (!test_bit(PG_REMOVE, &req->wb_flags))
 163                 return 0;
 164         ret = nfs_page_group_lock(req);
 165         if (ret)
 166                 return ret;
 167         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 168                 nfs_page_set_inode_ref(req, inode);
 169         nfs_page_group_unlock(req);
 170         return 0;
 171 }
 172
 173 static struct nfs_page *
 174 nfs_page_private_request(struct page *page)
 175 {
 176         if (!PagePrivate(page))
 177                 return NULL;
 178         return (struct nfs_page *)page_private(page);
 179 }
 180
 181 /*
 182  * nfs_page_find_head_request_locked - find head request associated with @page
 183  *
 184  * must be called while holding the inode lock.
 185  *
 186  * returns matching head request with reference held, or NULL if not found.
 187  */
 188 static struct nfs_page *
 189 nfs_page_find_private_request(struct page *page)
 190 {
 191         struct address_space *mapping = page_file_mapping(page);
 192         struct nfs_page *req;
 193
 194         if (!PagePrivate(page))
 195                 return NULL;
 196         spin_lock(&mapping->private_lock);
 197         req = nfs_page_private_request(page);
 198         if (req) {
 199                 WARN_ON_ONCE(req->wb_head != req);
 200                 kref_get(&req->wb_kref);
 201         }
 202         spin_unlock(&mapping->private_lock);
 203         return req;
 204 }
 205
 206 static struct nfs_page *
 207 nfs_page_find_swap_request(struct page *page)
 208 {
 209         struct inode *inode = page_file_mapping(page)->host;
 210         struct nfs_inode *nfsi = NFS_I(inode);
 211         struct nfs_page *req = NULL;
 212         if (!PageSwapCache(page))
 213                 return NULL;
 214         mutex_lock(&nfsi->commit_mutex);
 215         if (PageSwapCache(page)) {
 216                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 217                         page);
 218                 if (req) {
 219                         WARN_ON_ONCE(req->wb_head != req);
 220                         kref_get(&req->wb_kref);
 221                 }
 222         }
 223         mutex_unlock(&nfsi->commit_mutex);
 224         return req;
 225 }
 226
 227 /*
 228  * nfs_page_find_head_request - find head request associated with @page
 229  *
 230  * returns matching head request with reference held, or NULL if not found.
 231  */
 232 static struct nfs_page *nfs_page_find_head_request(struct page *page)
 233 {
 234         struct nfs_page *req;
 235
 236         req = nfs_page_find_private_request(page);
 237         if (!req)
 238                 req = nfs_page_find_swap_request(page);
 239         return req;
 240 }
 241
 242 static struct nfs_page *nfs_find_and_lock_page_request(struct page *page)
 243 {
 244         struct inode *inode = page_file_mapping(page)->host;
 245         struct nfs_page *req, *head;
 246         int ret;
 247
 248         for (;;) {
 249                 req = nfs_page_find_head_request(page);
 250                 if (!req)
 251                         return req;
 252                 head = nfs_page_group_lock_head(req);
 253                 if (head != req)
 254                         nfs_release_request(req);
 255                 if (IS_ERR(head))
 256                         return head;
 257                 ret = nfs_cancel_remove_inode(head, inode);
 258                 if (ret < 0) {
 259                         nfs_unlock_and_release_request(head);
 260                         return ERR_PTR(ret);
 261                 }
 262                 /* Ensure that nobody removed the request before we locked it */
 263                 if (head == nfs_page_private_request(page))
 264                         break;
 265                 if (PageSwapCache(page))
 266                         break;
 267                 nfs_unlock_and_release_request(head);
 268         }
 269         return head;
 270 }
 271
 272 /* Adjust the file length if we're writing beyond the end */
 273 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 274 {
 275         struct inode *inode = page_file_mapping(page)->host;
 276         loff_t end, i_size;
 277         pgoff_t end_index;
 278
 279         spin_lock(&inode->i_lock);
 280         i_size = i_size_read(inode);
 281         end_index = (i_size - 1) >> PAGE_SHIFT;
 282         if (i_size > 0 && page_index(page) < end_index)
 283                 goto out;
 284         end = page_file_offset(page) + ((loff_t)offset+count);
 285         if (i_size >= end)
 286                 goto out;
 287         trace_nfs_size_grow(inode, end);
 288         i_size_write(inode, end);
 289         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 290         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 291 out:
 292         spin_unlock(&inode->i_lock);
 293         nfs_fscache_invalidate(inode, 0);
 294 }
 295
 296 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 297 static void nfs_set_pageerror(struct address_space *mapping)
 298 {
 299         struct inode *inode = mapping->host;
 300
 301         nfs_zap_mapping(mapping->host, mapping);
 302         /* Force file size revalidation */
 303         spin_lock(&inode->i_lock);
 304         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 305                                              NFS_INO_INVALID_CHANGE |
 306                                              NFS_INO_INVALID_SIZE);
 307         spin_unlock(&inode->i_lock);
 308 }
 309
 310 static void nfs_mapping_set_error(struct page *page, int error)
 311 {
 312         struct address_space *mapping = page_file_mapping(page);
 313
 314         SetPageError(page);
 315         filemap_set_wb_err(mapping, error);
 316         if (mapping->host)
 317                 errseq_set(&mapping->host->i_sb->s_wb_err,
 318                            error == -ENOSPC ? -ENOSPC : -EIO);
 319         nfs_set_pageerror(mapping);
 320 }
 321
 322 /*
 323  * nfs_page_group_search_locked
 324  * @head - head request of page group
 325  * @page_offset - offset into page
 326  *
 327  * Search page group with head @head to find a request that contains the
 328  * page offset @page_offset.
 329  *
 330  * Returns a pointer to the first matching nfs request, or NULL if no
 331  * match is found.
 332  *
 333  * Must be called with the page group lock held
 334  */
 335 static struct nfs_page *
 336 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 337 {
 338         struct nfs_page *req;
 339
 340         req = head;
 341         do {
 342                 if (page_offset >= req->wb_pgbase &&
 343                     page_offset < (req->wb_pgbase + req->wb_bytes))
 344                         return req;
 345
 346                 req = req->wb_this_page;
 347         } while (req != head);
 348
 349         return NULL;
 350 }
 351
 352 /*
 353  * nfs_page_group_covers_page
 354  * @head - head request of page group
 355  *
 356  * Return true if the page group with head @head covers the whole page,
 357  * returns false otherwise
 358  */
 359 static bool nfs_page_group_covers_page(struct nfs_page *req)
 360 {
 361         struct nfs_page *tmp;
 362         unsigned int pos = 0;
 363         unsigned int len = nfs_page_length(req->wb_page);
 364
 365         nfs_page_group_lock(req);
 366
 367         for (;;) {
 368                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 369                 if (!tmp)
 370                         break;
 371                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 372         }
 373
 374         nfs_page_group_unlock(req);
 375         return pos >= len;
 376 }
 377
 378 /* We can set the PG_uptodate flag if we see that a write request
 379  * covers the full page.
 380  */
 381 static void nfs_mark_uptodate(struct nfs_page *req)
 382 {
 383         if (PageUptodate(req->wb_page))
 384                 return;
 385         if (!nfs_page_group_covers_page(req))
 386                 return;
 387         SetPageUptodate(req->wb_page);
 388 }
 389
 390 static int wb_priority(struct writeback_control *wbc)
 391 {
 392         int ret = 0;
 393
 394         if (wbc->sync_mode == WB_SYNC_ALL)
 395                 ret = FLUSH_COND_STABLE;
 396         return ret;
 397 }
 398
 399 /*
 400  * NFS congestion control
 401  */
 402
 403 int nfs_congestion_kb;
 404
 405 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 406 #define NFS_CONGESTION_OFF_THRESH       \
 407         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 408
 409 static void nfs_set_page_writeback(struct page *page)
 410 {
 411         struct inode *inode = page_file_mapping(page)->host;
 412         struct nfs_server *nfss = NFS_SERVER(inode);
 413         int ret = test_set_page_writeback(page);
 414
 415         WARN_ON_ONCE(ret != 0);
 416
 417         if (atomic_long_inc_return(&nfss->writeback) >
 418                         NFS_CONGESTION_ON_THRESH)
 419                 nfss->write_congested = 1;
 420 }
 421
 422 static void nfs_end_page_writeback(struct nfs_page *req)
 423 {
 424         struct inode *inode = page_file_mapping(req->wb_page)->host;
 425         struct nfs_server *nfss = NFS_SERVER(inode);
 426         bool is_done;
 427
 428         is_done = nfs_page_group_sync_on_bit(req, PG_WB_END);
 429         nfs_unlock_request(req);
 430         if (!is_done)
 431                 return;
 432
 433         end_page_writeback(req->wb_page);
 434         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 435                 nfss->write_congested = 0;
 436 }
 437
 438 /*
 439  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 440  *
 441  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 442  * @old_head - the old head of the list
 443  *
 444  * All subrequests must be locked and removed from all lists, so at this point
 445  * they are only "active" in this function, and possibly in nfs_wait_on_request
 446  * with a reference held by some other context.
 447  */
 448 static void
 449 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 450                                  struct nfs_page *old_head,
 451                                  struct inode *inode)
 452 {
 453         while (destroy_list) {
 454                 struct nfs_page *subreq = destroy_list;
 455
 456                 destroy_list = (subreq->wb_this_page == old_head) ?
 457                                    NULL : subreq->wb_this_page;
 458
 459                 /* Note: lock subreq in order to change subreq->wb_head */
 460                 nfs_page_set_headlock(subreq);
 461                 WARN_ON_ONCE(old_head != subreq->wb_head);
 462
 463                 /* make sure old group is not used */
 464                 subreq->wb_this_page = subreq;
 465                 subreq->wb_head = subreq;
 466
 467                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 468
 469                 /* Note: races with nfs_page_group_destroy() */
 470                 if (!kref_read(&subreq->wb_kref)) {
 471                         /* Check if we raced with nfs_page_group_destroy() */
 472                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 473                                 nfs_page_clear_headlock(subreq);
 474                                 nfs_free_request(subreq);
 475                         } else
 476                                 nfs_page_clear_headlock(subreq);
 477                         continue;
 478                 }
 479                 nfs_page_clear_headlock(subreq);
 480
 481                 nfs_release_request(old_head);
 482
 483                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 484                         nfs_release_request(subreq);
 485                         atomic_long_dec(&NFS_I(inode)->nrequests);
 486                 }
 487
 488                 /* subreq is now totally disconnected from page group or any
 489                  * write / commit lists. last chance to wake any waiters */
 490                 nfs_unlock_and_release_request(subreq);
 491         }
 492 }
 493
 494 /*
 495  * nfs_join_page_group - destroy subrequests of the head req
 496  * @head: the page used to lookup the "page group" of nfs_page structures
 497  * @inode: Inode to which the request belongs.
 498  *
 499  * This function joins all sub requests to the head request by first
 500  * locking all requests in the group, cancelling any pending operations
 501  * and finally updating the head request to cover the whole range covered by
 502  * the (former) group.  All subrequests are removed from any write or commit
 503  * lists, unlinked from the group and destroyed.
 504  */
 505 void
 506 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
 507 {
 508         struct nfs_page *subreq;
 509         struct nfs_page *destroy_list = NULL;
 510         unsigned int pgbase, off, bytes;
 511
 512         pgbase = head->wb_pgbase;
 513         bytes = head->wb_bytes;
 514         off = head->wb_offset;
 515         for (subreq = head->wb_this_page; subreq != head;
 516                         subreq = subreq->wb_this_page) {
 517                 /* Subrequests should always form a contiguous range */
 518                 if (pgbase > subreq->wb_pgbase) {
 519                         off -= pgbase - subreq->wb_pgbase;
 520                         bytes += pgbase - subreq->wb_pgbase;
 521                         pgbase = subreq->wb_pgbase;
 522                 }
 523                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 524                                 - pgbase, bytes);
 525         }
 526
 527         /* Set the head request's range to cover the former page group */
 528         head->wb_pgbase = pgbase;
 529         head->wb_bytes = bytes;
 530         head->wb_offset = off;
 531
 532         /* Now that all requests are locked, make sure they aren't on any list.
 533          * Commit list removal accounting is done after locks are dropped */
 534         subreq = head;
 535         do {
 536                 nfs_clear_request_commit(subreq);
 537                 subreq = subreq->wb_this_page;
 538         } while (subreq != head);
 539
 540         /* unlink subrequests from head, destroy them later */
 541         if (head->wb_this_page != head) {
 542                 /* destroy list will be terminated by head */
 543                 destroy_list = head->wb_this_page;
 544                 head->wb_this_page = head;
 545         }
 546
 547         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 548 }
 549
 550 /*
 551  * nfs_lock_and_join_requests - join all subreqs to the head req
 552  * @page: the page used to lookup the "page group" of nfs_page structures
 553  *
 554  * This function joins all sub requests to the head request by first
 555  * locking all requests in the group, cancelling any pending operations
 556  * and finally updating the head request to cover the whole range covered by
 557  * the (former) group.  All subrequests are removed from any write or commit
 558  * lists, unlinked from the group and destroyed.
 559  *
 560  * Returns a locked, referenced pointer to the head request - which after
 561  * this call is guaranteed to be the only request associated with the page.
 562  * Returns NULL if no requests are found for @page, or a ERR_PTR if an
 563  * error was encountered.
 564  */
 565 static struct nfs_page *
 566 nfs_lock_and_join_requests(struct page *page)
 567 {
 568         struct inode *inode = page_file_mapping(page)->host;
 569         struct nfs_page *head;
 570         int ret;
 571
 572         /*
 573          * A reference is taken only on the head request which acts as a
 574          * reference to the whole page group - the group will not be destroyed
 575          * until the head reference is released.
 576          */
 577         head = nfs_find_and_lock_page_request(page);
 578         if (IS_ERR_OR_NULL(head))
 579                 return head;
 580
 581         /* lock each request in the page group */
 582         ret = nfs_page_group_lock_subrequests(head);
 583         if (ret < 0) {
 584                 nfs_unlock_and_release_request(head);
 585                 return ERR_PTR(ret);
 586         }
 587
 588         nfs_join_page_group(head, inode);
 589
 590         return head;
 591 }
 592
 593 static void nfs_write_error(struct nfs_page *req, int error)
 594 {
 595         trace_nfs_write_error(page_file_mapping(req->wb_page)->host, req,
 596                               error);
 597         nfs_mapping_set_error(req->wb_page, error);
 598         nfs_inode_remove_request(req);
 599         nfs_end_page_writeback(req);
 600         nfs_release_request(req);
 601 }
 602
 603 /*
 604  * Find an associated nfs write request, and prepare to flush it out
 605  * May return an error if the user signalled nfs_wait_on_request().
 606  */
 607 static int nfs_page_async_flush(struct page *page,
 608                                 struct writeback_control *wbc,
 609                                 struct nfs_pageio_descriptor *pgio)
 610 {
 611         struct nfs_page *req;
 612         int ret = 0;
 613
 614         req = nfs_lock_and_join_requests(page);
 615         if (!req)
 616                 goto out;
 617         ret = PTR_ERR(req);
 618         if (IS_ERR(req))
 619                 goto out;
 620
 621         nfs_set_page_writeback(page);
 622         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 623
 624         /* If there is a fatal error that covers this write, just exit */
 625         ret = pgio->pg_error;
 626         if (nfs_error_is_fatal_on_server(ret))
 627                 goto out_launder;
 628
 629         ret = 0;
 630         if (!nfs_pageio_add_request(pgio, req)) {
 631                 ret = pgio->pg_error;
 632                 /*
 633                  * Remove the problematic req upon fatal errors on the server
 634                  */
 635                 if (nfs_error_is_fatal_on_server(ret))
 636                         goto out_launder;
 637                 if (wbc->sync_mode == WB_SYNC_NONE)
 638                         ret = AOP_WRITEPAGE_ACTIVATE;
 639                 redirty_page_for_writepage(wbc, page);
 640                 nfs_redirty_request(req);
 641                 pgio->pg_error = 0;
 642         } else
 643                 nfs_add_stats(page_file_mapping(page)->host,
 644                                 NFSIOS_WRITEPAGES, 1);
 645 out:
 646         return ret;
 647 out_launder:
 648         nfs_write_error(req, ret);
 649         return 0;
 650 }
 651
 652 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 653                             struct nfs_pageio_descriptor *pgio)
 654 {
 655         nfs_pageio_cond_complete(pgio, page_index(page));
 656         return nfs_page_async_flush(page, wbc, pgio);
 657 }
 658
 659 /*
 660  * Write an mmapped page to the server.
 661  */
 662 static int nfs_writepage_locked(struct page *page,
 663                                 struct writeback_control *wbc)
 664 {
 665         struct nfs_pageio_descriptor pgio;
 666         struct inode *inode = page_file_mapping(page)->host;
 667         int err;
 668
 669         if (wbc->sync_mode == WB_SYNC_NONE &&
 670             NFS_SERVER(inode)->write_congested) {
 671                 redirty_page_for_writepage(wbc, page);
 672                 return AOP_WRITEPAGE_ACTIVATE;
 673         }
 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         return err;
 682 }
 683
 684 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 685 {
 686         int ret;
 687
 688         ret = nfs_writepage_locked(page, wbc);
 689         if (ret != AOP_WRITEPAGE_ACTIVATE)
 690                 unlock_page(page);
 691         return ret;
 692 }
 693
 694 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 695 {
 696         int ret;
 697
 698         ret = nfs_do_writepage(page, wbc, data);
 699         if (ret != AOP_WRITEPAGE_ACTIVATE)
 700                 unlock_page(page);
 701         return ret;
 702 }
 703
 704 static void nfs_io_completion_commit(void *inode)
 705 {
 706         nfs_commit_inode(inode, 0);
 707 }
 708
 709 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 710 {
 711         struct inode *inode = mapping->host;
 712         struct nfs_pageio_descriptor pgio;
 713         struct nfs_io_completion *ioc = NULL;
 714         unsigned int mntflags = NFS_SERVER(inode)->flags;
 715         int priority = 0;
 716         int err;
 717
 718         if (wbc->sync_mode == WB_SYNC_NONE &&
 719             NFS_SERVER(inode)->write_congested)
 720                 return 0;
 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         do {
 734                 nfs_pageio_init_write(&pgio, inode, priority, false,
 735                                       &nfs_async_write_completion_ops);
 736                 pgio.pg_io_completion = ioc;
 737                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 738                                         &pgio);
 739                 pgio.pg_error = 0;
 740                 nfs_pageio_complete(&pgio);
 741         } while (err < 0 && !nfs_error_is_fatal(err));
 742         nfs_io_completion_put(ioc);
 743
 744         if (err < 0)
 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(hdr->inode, 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         if (IS_SWAPFILE(hdr->inode))
1413                 task_setup_data->flags |= RPC_TASK_SWAPPER;
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         struct nfs_inode *nfsi = NFS_I(page_file_mapping(req->wb_page)->host);
1426
1427         /* Bump the transmission count */
1428         req->wb_nio++;
1429         nfs_mark_request_dirty(req);
1430         atomic_long_inc(&nfsi->redirtied_pages);
1431         nfs_end_page_writeback(req);
1432         nfs_release_request(req);
1433 }
1434
1435 static void nfs_async_write_error(struct list_head *head, int error)
1436 {
1437         struct nfs_page *req;
1438
1439         while (!list_empty(head)) {
1440                 req = nfs_list_entry(head->next);
1441                 nfs_list_remove_request(req);
1442                 if (nfs_error_is_fatal_on_server(error))
1443                         nfs_write_error(req, error);
1444                 else
1445                         nfs_redirty_request(req);
1446         }
1447 }
1448
1449 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1450 {
1451         nfs_async_write_error(&hdr->pages, 0);
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 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1502                 struct nfs_fattr *fattr)
1503 {
1504         struct nfs_pgio_args *argp = &hdr->args;
1505         struct nfs_pgio_res *resp = &hdr->res;
1506         u64 size = argp->offset + resp->count;
1507
1508         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1509                 fattr->size = size;
1510         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1511                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1512                 return;
1513         }
1514         if (size != fattr->size)
1515                 return;
1516         /* Set attribute barrier */
1517         nfs_fattr_set_barrier(fattr);
1518         /* ...and update size */
1519         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1520 }
1521
1522 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1523 {
1524         struct nfs_fattr *fattr = &hdr->fattr;
1525         struct inode *inode = hdr->inode;
1526
1527         spin_lock(&inode->i_lock);
1528         nfs_writeback_check_extend(hdr, fattr);
1529         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1530         spin_unlock(&inode->i_lock);
1531 }
1532 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1533
1534 /*
1535  * This function is called when the WRITE call is complete.
1536  */
1537 static int nfs_writeback_done(struct rpc_task *task,
1538                               struct nfs_pgio_header *hdr,
1539                               struct inode *inode)
1540 {
1541         int status;
1542
1543         /*
1544          * ->write_done will attempt to use post-op attributes to detect
1545          * conflicting writes by other clients.  A strict interpretation
1546          * of close-to-open would allow us to continue caching even if
1547          * another writer had changed the file, but some applications
1548          * depend on tighter cache coherency when writing.
1549          */
1550         status = NFS_PROTO(inode)->write_done(task, hdr);
1551         if (status != 0)
1552                 return status;
1553
1554         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1555         trace_nfs_writeback_done(task, hdr);
1556
1557         if (task->tk_status >= 0) {
1558                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1559
1560                 if (committed == NFS_UNSTABLE) {
1561                         /*
1562                          * We have some uncommitted data on the server at
1563                          * this point, so ensure that we keep track of that
1564                          * fact irrespective of what later writes do.
1565                          */
1566                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1567                 }
1568
1569                 if (committed < hdr->args.stable) {
1570                         /* We tried a write call, but the server did not
1571                          * commit data to stable storage even though we
1572                          * requested it.
1573                          * Note: There is a known bug in Tru64 < 5.0 in which
1574                          *       the server reports NFS_DATA_SYNC, but performs
1575                          *       NFS_FILE_SYNC. We therefore implement this checking
1576                          *       as a dprintk() in order to avoid filling syslog.
1577                          */
1578                         static unsigned long    complain;
1579
1580                         /* Note this will print the MDS for a DS write */
1581                         if (time_before(complain, jiffies)) {
1582                                 dprintk("NFS:       faulty NFS server %s:"
1583                                         " (committed = %d) != (stable = %d)\n",
1584                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1585                                         committed, hdr->args.stable);
1586                                 complain = jiffies + 300 * HZ;
1587                         }
1588                 }
1589         }
1590
1591         /* Deal with the suid/sgid bit corner case */
1592         if (nfs_should_remove_suid(inode)) {
1593                 spin_lock(&inode->i_lock);
1594                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1595                 spin_unlock(&inode->i_lock);
1596         }
1597         return 0;
1598 }
1599
1600 /*
1601  * This function is called when the WRITE call is complete.
1602  */
1603 static void nfs_writeback_result(struct rpc_task *task,
1604                                  struct nfs_pgio_header *hdr)
1605 {
1606         struct nfs_pgio_args    *argp = &hdr->args;
1607         struct nfs_pgio_res     *resp = &hdr->res;
1608
1609         if (resp->count < argp->count) {
1610                 static unsigned long    complain;
1611
1612                 /* This a short write! */
1613                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1614
1615                 /* Has the server at least made some progress? */
1616                 if (resp->count == 0) {
1617                         if (time_before(complain, jiffies)) {
1618                                 printk(KERN_WARNING
1619                                        "NFS: Server wrote zero bytes, expected %u.\n",
1620                                        argp->count);
1621                                 complain = jiffies + 300 * HZ;
1622                         }
1623                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1624                         task->tk_status = -EIO;
1625                         return;
1626                 }
1627
1628                 /* For non rpc-based layout drivers, retry-through-MDS */
1629                 if (!task->tk_ops) {
1630                         hdr->pnfs_error = -EAGAIN;
1631                         return;
1632                 }
1633
1634                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1635                 if (resp->verf->committed != NFS_UNSTABLE) {
1636                         /* Resend from where the server left off */
1637                         hdr->mds_offset += resp->count;
1638                         argp->offset += resp->count;
1639                         argp->pgbase += resp->count;
1640                         argp->count -= resp->count;
1641                 } else {
1642                         /* Resend as a stable write in order to avoid
1643                          * headaches in the case of a server crash.
1644                          */
1645                         argp->stable = NFS_FILE_SYNC;
1646                 }
1647                 resp->count = 0;
1648                 resp->verf->committed = 0;
1649                 rpc_restart_call_prepare(task);
1650         }
1651 }
1652
1653 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1654 {
1655         return wait_var_event_killable(&cinfo->rpcs_out,
1656                                        !atomic_read(&cinfo->rpcs_out));
1657 }
1658
1659 void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1660 {
1661         atomic_inc(&cinfo->rpcs_out);
1662 }
1663
1664 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1665 {
1666         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1667                 wake_up_var(&cinfo->rpcs_out);
1668                 return true;
1669         }
1670         return false;
1671 }
1672
1673 void nfs_commitdata_release(struct nfs_commit_data *data)
1674 {
1675         put_nfs_open_context(data->context);
1676         nfs_commit_free(data);
1677 }
1678 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1679
1680 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1681                         const struct nfs_rpc_ops *nfs_ops,
1682                         const struct rpc_call_ops *call_ops,
1683                         int how, int flags)
1684 {
1685         struct rpc_task *task;
1686         int priority = flush_task_priority(how);
1687         struct rpc_message msg = {
1688                 .rpc_argp = &data->args,
1689                 .rpc_resp = &data->res,
1690                 .rpc_cred = data->cred,
1691         };
1692         struct rpc_task_setup task_setup_data = {
1693                 .task = &data->task,
1694                 .rpc_client = clnt,
1695                 .rpc_message = &msg,
1696                 .callback_ops = call_ops,
1697                 .callback_data = data,
1698                 .workqueue = nfsiod_workqueue,
1699                 .flags = RPC_TASK_ASYNC | flags,
1700                 .priority = priority,
1701         };
1702
1703         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1704                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1705
1706         /* Set up the initial task struct.  */
1707         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1708         trace_nfs_initiate_commit(data);
1709
1710         dprintk("NFS: initiated commit call\n");
1711
1712         task = rpc_run_task(&task_setup_data);
1713         if (IS_ERR(task))
1714                 return PTR_ERR(task);
1715         if (how & FLUSH_SYNC)
1716                 rpc_wait_for_completion_task(task);
1717         rpc_put_task(task);
1718         return 0;
1719 }
1720 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1721
1722 static loff_t nfs_get_lwb(struct list_head *head)
1723 {
1724         loff_t lwb = 0;
1725         struct nfs_page *req;
1726
1727         list_for_each_entry(req, head, wb_list)
1728                 if (lwb < (req_offset(req) + req->wb_bytes))
1729                         lwb = req_offset(req) + req->wb_bytes;
1730
1731         return lwb;
1732 }
1733
1734 /*
1735  * Set up the argument/result storage required for the RPC call.
1736  */
1737 void nfs_init_commit(struct nfs_commit_data *data,
1738                      struct list_head *head,
1739                      struct pnfs_layout_segment *lseg,
1740                      struct nfs_commit_info *cinfo)
1741 {
1742         struct nfs_page *first;
1743         struct nfs_open_context *ctx;
1744         struct inode *inode;
1745
1746         /* Set up the RPC argument and reply structs
1747          * NB: take care not to mess about with data->commit et al. */
1748
1749         if (head)
1750                 list_splice_init(head, &data->pages);
1751
1752         first = nfs_list_entry(data->pages.next);
1753         ctx = nfs_req_openctx(first);
1754         inode = d_inode(ctx->dentry);
1755
1756         data->inode       = inode;
1757         data->cred        = ctx->cred;
1758         data->lseg        = lseg; /* reference transferred */
1759         /* only set lwb for pnfs commit */
1760         if (lseg)
1761                 data->lwb = nfs_get_lwb(&data->pages);
1762         data->mds_ops     = &nfs_commit_ops;
1763         data->completion_ops = cinfo->completion_ops;
1764         data->dreq        = cinfo->dreq;
1765
1766         data->args.fh     = NFS_FH(data->inode);
1767         /* Note: we always request a commit of the entire inode */
1768         data->args.offset = 0;
1769         data->args.count  = 0;
1770         data->context     = get_nfs_open_context(ctx);
1771         data->res.fattr   = &data->fattr;
1772         data->res.verf    = &data->verf;
1773         nfs_fattr_init(&data->fattr);
1774         nfs_commit_begin(cinfo->mds);
1775 }
1776 EXPORT_SYMBOL_GPL(nfs_init_commit);
1777
1778 void nfs_retry_commit(struct list_head *page_list,
1779                       struct pnfs_layout_segment *lseg,
1780                       struct nfs_commit_info *cinfo,
1781                       u32 ds_commit_idx)
1782 {
1783         struct nfs_page *req;
1784
1785         while (!list_empty(page_list)) {
1786                 req = nfs_list_entry(page_list->next);
1787                 nfs_list_remove_request(req);
1788                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1789                 if (!cinfo->dreq)
1790                         nfs_clear_page_commit(req->wb_page);
1791                 nfs_unlock_and_release_request(req);
1792         }
1793 }
1794 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1795
1796 static void
1797 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1798                 struct nfs_page *req)
1799 {
1800         __set_page_dirty_nobuffers(req->wb_page);
1801 }
1802
1803 /*
1804  * Commit dirty pages
1805  */
1806 static int
1807 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1808                 struct nfs_commit_info *cinfo)
1809 {
1810         struct nfs_commit_data  *data;
1811         unsigned short task_flags = 0;
1812
1813         /* another commit raced with us */
1814         if (list_empty(head))
1815                 return 0;
1816
1817         data = nfs_commitdata_alloc();
1818         if (!data) {
1819                 nfs_retry_commit(head, NULL, cinfo, -1);
1820                 return -ENOMEM;
1821         }
1822
1823         /* Set up the argument struct */
1824         nfs_init_commit(data, head, NULL, cinfo);
1825         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1826                 task_flags = RPC_TASK_MOVEABLE;
1827         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1828                                    data->mds_ops, how,
1829                                    RPC_TASK_CRED_NOREF | task_flags);
1830 }
1831
1832 /*
1833  * COMMIT call returned
1834  */
1835 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1836 {
1837         struct nfs_commit_data  *data = calldata;
1838
1839         /* Call the NFS version-specific code */
1840         NFS_PROTO(data->inode)->commit_done(task, data);
1841         trace_nfs_commit_done(task, data);
1842 }
1843
1844 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1845 {
1846         const struct nfs_writeverf *verf = data->res.verf;
1847         struct nfs_page *req;
1848         int status = data->task.tk_status;
1849         struct nfs_commit_info cinfo;
1850         struct nfs_server *nfss;
1851
1852         while (!list_empty(&data->pages)) {
1853                 req = nfs_list_entry(data->pages.next);
1854                 nfs_list_remove_request(req);
1855                 if (req->wb_page)
1856                         nfs_clear_page_commit(req->wb_page);
1857
1858                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1859                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1860                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1861                         req->wb_bytes,
1862                         (long long)req_offset(req));
1863                 if (status < 0) {
1864                         if (req->wb_page) {
1865                                 trace_nfs_commit_error(data->inode, req,
1866                                                        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                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
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                 nfss->write_congested = 0;
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         how &= ~FLUSH_SYNC;
1939         nfs_init_cinfo_from_inode(&cinfo, inode);
1940         nfs_commit_begin(cinfo.mds);
1941         for (;;) {
1942                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1943                 if (ret <= 0)
1944                         break;
1945                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1946                 if (ret < 0)
1947                         break;
1948                 ret = 0;
1949                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1950                         if (nscan < wbc->nr_to_write)
1951                                 wbc->nr_to_write -= nscan;
1952                         else
1953                                 wbc->nr_to_write = 0;
1954                 }
1955                 if (nscan < INT_MAX)
1956                         break;
1957                 cond_resched();
1958         }
1959         nfs_commit_end(cinfo.mds);
1960         if (ret || !may_wait)
1961                 return ret;
1962         return wait_on_commit(cinfo.mds);
1963 }
1964
1965 int nfs_commit_inode(struct inode *inode, int how)
1966 {
1967         return __nfs_commit_inode(inode, how, NULL);
1968 }
1969 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1970
1971 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1972 {
1973         struct nfs_inode *nfsi = NFS_I(inode);
1974         int flags = FLUSH_SYNC;
1975         int ret = 0;
1976
1977         if (wbc->sync_mode == WB_SYNC_NONE) {
1978                 /* no commits means nothing needs to be done */
1979                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1980                         goto check_requests_outstanding;
1981
1982                 /* Don't commit yet if this is a non-blocking flush and there
1983                  * are a lot of outstanding writes for this mapping.
1984                  */
1985                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1986                         goto out_mark_dirty;
1987
1988                 /* don't wait for the COMMIT response */
1989                 flags = 0;
1990         }
1991
1992         ret = __nfs_commit_inode(inode, flags, wbc);
1993         if (!ret) {
1994                 if (flags & FLUSH_SYNC)
1995                         return 0;
1996         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1997                 goto out_mark_dirty;
1998
1999 check_requests_outstanding:
2000         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2001                 return ret;
2002 out_mark_dirty:
2003         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2004         return ret;
2005 }
2006 EXPORT_SYMBOL_GPL(nfs_write_inode);
2007
2008 /*
2009  * Wrapper for filemap_write_and_wait_range()
2010  *
2011  * Needed for pNFS in order to ensure data becomes visible to the
2012  * client.
2013  */
2014 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2015                 loff_t lstart, loff_t lend)
2016 {
2017         int ret;
2018
2019         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2020         if (ret == 0)
2021                 ret = pnfs_sync_inode(mapping->host, true);
2022         return ret;
2023 }
2024 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2025
2026 /*
2027  * flush the inode to disk.
2028  */
2029 int nfs_wb_all(struct inode *inode)
2030 {
2031         int ret;
2032
2033         trace_nfs_writeback_inode_enter(inode);
2034
2035         ret = filemap_write_and_wait(inode->i_mapping);
2036         if (ret)
2037                 goto out;
2038         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2039         if (ret < 0)
2040                 goto out;
2041         pnfs_sync_inode(inode, true);
2042         ret = 0;
2043
2044 out:
2045         trace_nfs_writeback_inode_exit(inode, ret);
2046         return ret;
2047 }
2048 EXPORT_SYMBOL_GPL(nfs_wb_all);
2049
2050 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2051 {
2052         struct nfs_page *req;
2053         int ret = 0;
2054
2055         folio_wait_writeback(folio);
2056
2057         /* blocking call to cancel all requests and join to a single (head)
2058          * request */
2059         req = nfs_lock_and_join_requests(&folio->page);
2060
2061         if (IS_ERR(req)) {
2062                 ret = PTR_ERR(req);
2063         } else if (req) {
2064                 /* all requests from this folio have been cancelled by
2065                  * nfs_lock_and_join_requests, so just remove the head
2066                  * request from the inode / page_private pointer and
2067                  * release it */
2068                 nfs_inode_remove_request(req);
2069                 nfs_unlock_and_release_request(req);
2070         }
2071
2072         return ret;
2073 }
2074
2075 /*
2076  * Write back all requests on one page - we do this before reading it.
2077  */
2078 int nfs_wb_page(struct inode *inode, struct page *page)
2079 {
2080         loff_t range_start = page_file_offset(page);
2081         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2082         struct writeback_control wbc = {
2083                 .sync_mode = WB_SYNC_ALL,
2084                 .nr_to_write = 0,
2085                 .range_start = range_start,
2086                 .range_end = range_end,
2087         };
2088         int ret;
2089
2090         trace_nfs_writeback_page_enter(inode);
2091
2092         for (;;) {
2093                 wait_on_page_writeback(page);
2094                 if (clear_page_dirty_for_io(page)) {
2095                         ret = nfs_writepage_locked(page, &wbc);
2096                         if (ret < 0)
2097                                 goto out_error;
2098                         continue;
2099                 }
2100                 ret = 0;
2101                 if (!PagePrivate(page))
2102                         break;
2103                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2104                 if (ret < 0)
2105                         goto out_error;
2106         }
2107 out_error:
2108         trace_nfs_writeback_page_exit(inode, ret);
2109         return ret;
2110 }
2111
2112 #ifdef CONFIG_MIGRATION
2113 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2114                 struct folio *src, enum migrate_mode mode)
2115 {
2116         /*
2117          * If the private flag is set, the folio is currently associated with
2118          * an in-progress read or write request. Don't try to migrate it.
2119          *
2120          * FIXME: we could do this in principle, but we'll need a way to ensure
2121          *        that we can safely release the inode reference while holding
2122          *        the folio lock.
2123          */
2124         if (folio_test_private(src))
2125                 return -EBUSY;
2126
2127         if (folio_test_fscache(src)) {
2128                 if (mode == MIGRATE_ASYNC)
2129                         return -EBUSY;
2130                 folio_wait_fscache(src);
2131         }
2132
2133         return migrate_folio(mapping, dst, src, mode);
2134 }
2135 #endif
2136
2137 int __init nfs_init_writepagecache(void)
2138 {
2139         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2140                                              sizeof(struct nfs_pgio_header),
2141                                              0, SLAB_HWCACHE_ALIGN,
2142                                              NULL);
2143         if (nfs_wdata_cachep == NULL)
2144                 return -ENOMEM;
2145
2146         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2147                                                      nfs_wdata_cachep);
2148         if (nfs_wdata_mempool == NULL)
2149                 goto out_destroy_write_cache;
2150
2151         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2152                                              sizeof(struct nfs_commit_data),
2153                                              0, SLAB_HWCACHE_ALIGN,
2154                                              NULL);
2155         if (nfs_cdata_cachep == NULL)
2156                 goto out_destroy_write_mempool;
2157
2158         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2159                                                       nfs_cdata_cachep);
2160         if (nfs_commit_mempool == NULL)
2161                 goto out_destroy_commit_cache;
2162
2163         /*
2164          * NFS congestion size, scale with available memory.
2165          *
2166          *  64MB:    8192k
2167          * 128MB:   11585k
2168          * 256MB:   16384k
2169          * 512MB:   23170k
2170          *   1GB:   32768k
2171          *   2GB:   46340k
2172          *   4GB:   65536k
2173          *   8GB:   92681k
2174          *  16GB:  131072k
2175          *
2176          * This allows larger machines to have larger/more transfers.
2177          * Limit the default to 256M
2178          */
2179         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2180         if (nfs_congestion_kb > 256*1024)
2181                 nfs_congestion_kb = 256*1024;
2182
2183         return 0;
2184
2185 out_destroy_commit_cache:
2186         kmem_cache_destroy(nfs_cdata_cachep);
2187 out_destroy_write_mempool:
2188         mempool_destroy(nfs_wdata_mempool);
2189 out_destroy_write_cache:
2190         kmem_cache_destroy(nfs_wdata_cachep);
2191         return -ENOMEM;
2192 }
2193
2194 void nfs_destroy_writepagecache(void)
2195 {
2196         mempool_destroy(nfs_commit_mempool);
2197         kmem_cache_destroy(nfs_cdata_cachep);
2198         mempool_destroy(nfs_wdata_mempool);
2199         kmem_cache_destroy(nfs_wdata_cachep);
2200 }
2201
2202 static const struct nfs_rw_ops nfs_rw_write_ops = {
2203         .rw_alloc_header        = nfs_writehdr_alloc,
2204         .rw_free_header         = nfs_writehdr_free,
2205         .rw_done                = nfs_writeback_done,
2206         .rw_result              = nfs_writeback_result,
2207         .rw_initiate            = nfs_initiate_write,
2208 };