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(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(req, error);
 596         nfs_mapping_set_error(req->wb_page, error);
 597         nfs_inode_remove_request(req);
 598         nfs_end_page_writeback(req);
 599         nfs_release_request(req);
 600 }
 601
 602 /*
 603  * Find an associated nfs write request, and prepare to flush it out
 604  * May return an error if the user signalled nfs_wait_on_request().
 605  */
 606 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 607                                 struct page *page)
 608 {
 609         struct nfs_page *req;
 610         int ret = 0;
 611
 612         req = nfs_lock_and_join_requests(page);
 613         if (!req)
 614                 goto out;
 615         ret = PTR_ERR(req);
 616         if (IS_ERR(req))
 617                 goto out;
 618
 619         nfs_set_page_writeback(page);
 620         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 621
 622         /* If there is a fatal error that covers this write, just exit */
 623         ret = pgio->pg_error;
 624         if (nfs_error_is_fatal_on_server(ret))
 625                 goto out_launder;
 626
 627         ret = 0;
 628         if (!nfs_pageio_add_request(pgio, req)) {
 629                 ret = pgio->pg_error;
 630                 /*
 631                  * Remove the problematic req upon fatal errors on the server
 632                  */
 633                 if (nfs_error_is_fatal(ret)) {
 634                         if (nfs_error_is_fatal_on_server(ret))
 635                                 goto out_launder;
 636                 } else
 637                         ret = -EAGAIN;
 638                 nfs_redirty_request(req);
 639                 pgio->pg_error = 0;
 640         } else
 641                 nfs_add_stats(page_file_mapping(page)->host,
 642                                 NFSIOS_WRITEPAGES, 1);
 643 out:
 644         return ret;
 645 out_launder:
 646         nfs_write_error(req, ret);
 647         return 0;
 648 }
 649
 650 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc,
 651                             struct nfs_pageio_descriptor *pgio)
 652 {
 653         int ret;
 654
 655         nfs_pageio_cond_complete(pgio, page_index(page));
 656         ret = nfs_page_async_flush(pgio, page);
 657         if (ret == -EAGAIN) {
 658                 redirty_page_for_writepage(wbc, page);
 659                 ret = AOP_WRITEPAGE_ACTIVATE;
 660         }
 661         return ret;
 662 }
 663
 664 /*
 665  * Write an mmapped page to the server.
 666  */
 667 static int nfs_writepage_locked(struct page *page,
 668                                 struct writeback_control *wbc)
 669 {
 670         struct nfs_pageio_descriptor pgio;
 671         struct inode *inode = page_file_mapping(page)->host;
 672         int err;
 673
 674         if (wbc->sync_mode == WB_SYNC_NONE &&
 675             NFS_SERVER(inode)->write_congested)
 676                 return AOP_WRITEPAGE_ACTIVATE;
 677
 678         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 679         nfs_pageio_init_write(&pgio, inode, 0,
 680                                 false, &nfs_async_write_completion_ops);
 681         err = nfs_do_writepage(page, wbc, &pgio);
 682         pgio.pg_error = 0;
 683         nfs_pageio_complete(&pgio);
 684         if (err < 0)
 685                 return err;
 686         if (nfs_error_is_fatal(pgio.pg_error))
 687                 return pgio.pg_error;
 688         return 0;
 689 }
 690
 691 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 692 {
 693         int ret;
 694
 695         ret = nfs_writepage_locked(page, wbc);
 696         if (ret != AOP_WRITEPAGE_ACTIVATE)
 697                 unlock_page(page);
 698         return ret;
 699 }
 700
 701 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 702 {
 703         int ret;
 704
 705         ret = nfs_do_writepage(page, wbc, data);
 706         if (ret != AOP_WRITEPAGE_ACTIVATE)
 707                 unlock_page(page);
 708         return ret;
 709 }
 710
 711 static void nfs_io_completion_commit(void *inode)
 712 {
 713         nfs_commit_inode(inode, 0);
 714 }
 715
 716 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 717 {
 718         struct inode *inode = mapping->host;
 719         struct nfs_pageio_descriptor pgio;
 720         struct nfs_io_completion *ioc = NULL;
 721         unsigned int mntflags = NFS_SERVER(inode)->flags;
 722         int priority = 0;
 723         int err;
 724
 725         if (wbc->sync_mode == WB_SYNC_NONE &&
 726             NFS_SERVER(inode)->write_congested)
 727                 return 0;
 728
 729         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 730
 731         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 732             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 733                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 734                 if (ioc)
 735                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 736                                                inode);
 737                 priority = wb_priority(wbc);
 738         }
 739
 740         nfs_pageio_init_write(&pgio, inode, priority, false,
 741                                 &nfs_async_write_completion_ops);
 742         pgio.pg_io_completion = ioc;
 743         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 744         pgio.pg_error = 0;
 745         nfs_pageio_complete(&pgio);
 746         nfs_io_completion_put(ioc);
 747
 748         if (err < 0)
 749                 goto out_err;
 750         err = pgio.pg_error;
 751         if (nfs_error_is_fatal(err))
 752                 goto out_err;
 753         return 0;
 754 out_err:
 755         return err;
 756 }
 757
 758 /*
 759  * Insert a write request into an inode
 760  */
 761 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 762 {
 763         struct address_space *mapping = page_file_mapping(req->wb_page);
 764         struct nfs_inode *nfsi = NFS_I(inode);
 765
 766         WARN_ON_ONCE(req->wb_this_page != req);
 767
 768         /* Lock the request! */
 769         nfs_lock_request(req);
 770
 771         /*
 772          * Swap-space should not get truncated. Hence no need to plug the race
 773          * with invalidate/truncate.
 774          */
 775         spin_lock(&mapping->private_lock);
 776         if (likely(!PageSwapCache(req->wb_page))) {
 777                 set_bit(PG_MAPPED, &req->wb_flags);
 778                 SetPagePrivate(req->wb_page);
 779                 set_page_private(req->wb_page, (unsigned long)req);
 780         }
 781         spin_unlock(&mapping->private_lock);
 782         atomic_long_inc(&nfsi->nrequests);
 783         /* this a head request for a page group - mark it as having an
 784          * extra reference so sub groups can follow suit.
 785          * This flag also informs pgio layer when to bump nrequests when
 786          * adding subrequests. */
 787         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 788         kref_get(&req->wb_kref);
 789 }
 790
 791 /*
 792  * Remove a write request from an inode
 793  */
 794 static void nfs_inode_remove_request(struct nfs_page *req)
 795 {
 796         struct address_space *mapping = page_file_mapping(req->wb_page);
 797         struct inode *inode = mapping->host;
 798         struct nfs_inode *nfsi = NFS_I(inode);
 799         struct nfs_page *head;
 800
 801         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 802                 head = req->wb_head;
 803
 804                 spin_lock(&mapping->private_lock);
 805                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 806                         set_page_private(head->wb_page, 0);
 807                         ClearPagePrivate(head->wb_page);
 808                         clear_bit(PG_MAPPED, &head->wb_flags);
 809                 }
 810                 spin_unlock(&mapping->private_lock);
 811         }
 812
 813         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 814                 nfs_release_request(req);
 815                 atomic_long_dec(&nfsi->nrequests);
 816         }
 817 }
 818
 819 static void
 820 nfs_mark_request_dirty(struct nfs_page *req)
 821 {
 822         if (req->wb_page)
 823                 __set_page_dirty_nobuffers(req->wb_page);
 824 }
 825
 826 /*
 827  * nfs_page_search_commits_for_head_request_locked
 828  *
 829  * Search through commit lists on @inode for the head request for @page.
 830  * Must be called while holding the inode (which is cinfo) lock.
 831  *
 832  * Returns the head request if found, or NULL if not found.
 833  */
 834 static struct nfs_page *
 835 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 836                                                 struct page *page)
 837 {
 838         struct nfs_page *freq, *t;
 839         struct nfs_commit_info cinfo;
 840         struct inode *inode = &nfsi->vfs_inode;
 841
 842         nfs_init_cinfo_from_inode(&cinfo, inode);
 843
 844         /* search through pnfs commit lists */
 845         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 846         if (freq)
 847                 return freq->wb_head;
 848
 849         /* Linearly search the commit list for the correct request */
 850         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 851                 if (freq->wb_page == page)
 852                         return freq->wb_head;
 853         }
 854
 855         return NULL;
 856 }
 857
 858 /**
 859  * nfs_request_add_commit_list_locked - add request to a commit list
 860  * @req: pointer to a struct nfs_page
 861  * @dst: commit list head
 862  * @cinfo: holds list lock and accounting info
 863  *
 864  * This sets the PG_CLEAN bit, updates the cinfo count of
 865  * number of outstanding requests requiring a commit as well as
 866  * the MM page stats.
 867  *
 868  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 869  * nfs_page lock.
 870  */
 871 void
 872 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 873                             struct nfs_commit_info *cinfo)
 874 {
 875         set_bit(PG_CLEAN, &req->wb_flags);
 876         nfs_list_add_request(req, dst);
 877         atomic_long_inc(&cinfo->mds->ncommit);
 878 }
 879 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 880
 881 /**
 882  * nfs_request_add_commit_list - add request to a commit list
 883  * @req: pointer to a struct nfs_page
 884  * @cinfo: holds list lock and accounting info
 885  *
 886  * This sets the PG_CLEAN bit, updates the cinfo count of
 887  * number of outstanding requests requiring a commit as well as
 888  * the MM page stats.
 889  *
 890  * The caller must _not_ hold the cinfo->lock, but must be
 891  * holding the nfs_page lock.
 892  */
 893 void
 894 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 895 {
 896         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 897         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 898         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 899         if (req->wb_page)
 900                 nfs_mark_page_unstable(req->wb_page, cinfo);
 901 }
 902 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 903
 904 /**
 905  * nfs_request_remove_commit_list - Remove request from a commit list
 906  * @req: pointer to a nfs_page
 907  * @cinfo: holds list lock and accounting info
 908  *
 909  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 910  * number of outstanding requests requiring a commit
 911  * It does not update the MM page stats.
 912  *
 913  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 914  */
 915 void
 916 nfs_request_remove_commit_list(struct nfs_page *req,
 917                                struct nfs_commit_info *cinfo)
 918 {
 919         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 920                 return;
 921         nfs_list_remove_request(req);
 922         atomic_long_dec(&cinfo->mds->ncommit);
 923 }
 924 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 925
 926 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 927                                       struct inode *inode)
 928 {
 929         cinfo->inode = inode;
 930         cinfo->mds = &NFS_I(inode)->commit_info;
 931         cinfo->ds = pnfs_get_ds_info(inode);
 932         cinfo->dreq = NULL;
 933         cinfo->completion_ops = &nfs_commit_completion_ops;
 934 }
 935
 936 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 937                     struct inode *inode,
 938                     struct nfs_direct_req *dreq)
 939 {
 940         if (dreq)
 941                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 942         else
 943                 nfs_init_cinfo_from_inode(cinfo, inode);
 944 }
 945 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 946
 947 /*
 948  * Add a request to the inode's commit list.
 949  */
 950 void
 951 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 952                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 953 {
 954         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 955                 return;
 956         nfs_request_add_commit_list(req, cinfo);
 957 }
 958
 959 static void
 960 nfs_clear_page_commit(struct page *page)
 961 {
 962         dec_node_page_state(page, NR_WRITEBACK);
 963         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 964                     WB_WRITEBACK);
 965 }
 966
 967 /* Called holding the request lock on @req */
 968 static void
 969 nfs_clear_request_commit(struct nfs_page *req)
 970 {
 971         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 972                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 973                 struct inode *inode = d_inode(ctx->dentry);
 974                 struct nfs_commit_info cinfo;
 975
 976                 nfs_init_cinfo_from_inode(&cinfo, inode);
 977                 mutex_lock(&NFS_I(inode)->commit_mutex);
 978                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 979                         nfs_request_remove_commit_list(req, &cinfo);
 980                 }
 981                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 982                 nfs_clear_page_commit(req->wb_page);
 983         }
 984 }
 985
 986 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 987 {
 988         if (hdr->verf.committed == NFS_DATA_SYNC)
 989                 return hdr->lseg == NULL;
 990         return hdr->verf.committed != NFS_FILE_SYNC;
 991 }
 992
 993 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 994 {
 995         nfs_io_completion_get(hdr->io_completion);
 996 }
 997
 998 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 999 {
1000         struct nfs_commit_info cinfo;
1001         unsigned long bytes = 0;
1002
1003         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1004                 goto out;
1005         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1006         while (!list_empty(&hdr->pages)) {
1007                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1008
1009                 bytes += req->wb_bytes;
1010                 nfs_list_remove_request(req);
1011                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1012                     (hdr->good_bytes < bytes)) {
1013                         trace_nfs_comp_error(req, hdr->error);
1014                         nfs_mapping_set_error(req->wb_page, hdr->error);
1015                         goto remove_req;
1016                 }
1017                 if (nfs_write_need_commit(hdr)) {
1018                         /* Reset wb_nio, since the write was successful. */
1019                         req->wb_nio = 0;
1020                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1021                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1022                                 hdr->pgio_mirror_idx);
1023                         goto next;
1024                 }
1025 remove_req:
1026                 nfs_inode_remove_request(req);
1027 next:
1028                 nfs_end_page_writeback(req);
1029                 nfs_release_request(req);
1030         }
1031 out:
1032         nfs_io_completion_put(hdr->io_completion);
1033         hdr->release(hdr);
1034 }
1035
1036 unsigned long
1037 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1038 {
1039         return atomic_long_read(&cinfo->mds->ncommit);
1040 }
1041
1042 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1043 int
1044 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1045                      struct nfs_commit_info *cinfo, int max)
1046 {
1047         struct nfs_page *req, *tmp;
1048         int ret = 0;
1049
1050         list_for_each_entry_safe(req, tmp, src, wb_list) {
1051                 kref_get(&req->wb_kref);
1052                 if (!nfs_lock_request(req)) {
1053                         nfs_release_request(req);
1054                         continue;
1055                 }
1056                 nfs_request_remove_commit_list(req, cinfo);
1057                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1058                 nfs_list_add_request(req, dst);
1059                 ret++;
1060                 if ((ret == max) && !cinfo->dreq)
1061                         break;
1062                 cond_resched();
1063         }
1064         return ret;
1065 }
1066 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1067
1068 /*
1069  * nfs_scan_commit - Scan an inode for commit requests
1070  * @inode: NFS inode to scan
1071  * @dst: mds destination list
1072  * @cinfo: mds and ds lists of reqs ready to commit
1073  *
1074  * Moves requests from the inode's 'commit' request list.
1075  * The requests are *not* checked to ensure that they form a contiguous set.
1076  */
1077 int
1078 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1079                 struct nfs_commit_info *cinfo)
1080 {
1081         int ret = 0;
1082
1083         if (!atomic_long_read(&cinfo->mds->ncommit))
1084                 return 0;
1085         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1086         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1087                 const int max = INT_MAX;
1088
1089                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1090                                            cinfo, max);
1091                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1092         }
1093         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1094         return ret;
1095 }
1096
1097 /*
1098  * Search for an existing write request, and attempt to update
1099  * it to reflect a new dirty region on a given page.
1100  *
1101  * If the attempt fails, then the existing request is flushed out
1102  * to disk.
1103  */
1104 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1105                 struct page *page,
1106                 unsigned int offset,
1107                 unsigned int bytes)
1108 {
1109         struct nfs_page *req;
1110         unsigned int rqend;
1111         unsigned int end;
1112         int error;
1113
1114         end = offset + bytes;
1115
1116         req = nfs_lock_and_join_requests(page);
1117         if (IS_ERR_OR_NULL(req))
1118                 return req;
1119
1120         rqend = req->wb_offset + req->wb_bytes;
1121         /*
1122          * Tell the caller to flush out the request if
1123          * the offsets are non-contiguous.
1124          * Note: nfs_flush_incompatible() will already
1125          * have flushed out requests having wrong owners.
1126          */
1127         if (offset > rqend || end < req->wb_offset)
1128                 goto out_flushme;
1129
1130         /* Okay, the request matches. Update the region */
1131         if (offset < req->wb_offset) {
1132                 req->wb_offset = offset;
1133                 req->wb_pgbase = offset;
1134         }
1135         if (end > rqend)
1136                 req->wb_bytes = end - req->wb_offset;
1137         else
1138                 req->wb_bytes = rqend - req->wb_offset;
1139         req->wb_nio = 0;
1140         return req;
1141 out_flushme:
1142         /*
1143          * Note: we mark the request dirty here because
1144          * nfs_lock_and_join_requests() cannot preserve
1145          * commit flags, so we have to replay the write.
1146          */
1147         nfs_mark_request_dirty(req);
1148         nfs_unlock_and_release_request(req);
1149         error = nfs_wb_page(inode, page);
1150         return (error < 0) ? ERR_PTR(error) : NULL;
1151 }
1152
1153 /*
1154  * Try to update an existing write request, or create one if there is none.
1155  *
1156  * Note: Should always be called with the Page Lock held to prevent races
1157  * if we have to add a new request. Also assumes that the caller has
1158  * already called nfs_flush_incompatible() if necessary.
1159  */
1160 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1161                 struct page *page, unsigned int offset, unsigned int bytes)
1162 {
1163         struct inode *inode = page_file_mapping(page)->host;
1164         struct nfs_page *req;
1165
1166         req = nfs_try_to_update_request(inode, page, offset, bytes);
1167         if (req != NULL)
1168                 goto out;
1169         req = nfs_create_request(ctx, page, offset, bytes);
1170         if (IS_ERR(req))
1171                 goto out;
1172         nfs_inode_add_request(inode, req);
1173 out:
1174         return req;
1175 }
1176
1177 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1178                 unsigned int offset, unsigned int count)
1179 {
1180         struct nfs_page *req;
1181
1182         req = nfs_setup_write_request(ctx, page, offset, count);
1183         if (IS_ERR(req))
1184                 return PTR_ERR(req);
1185         /* Update file length */
1186         nfs_grow_file(page, offset, count);
1187         nfs_mark_uptodate(req);
1188         nfs_mark_request_dirty(req);
1189         nfs_unlock_and_release_request(req);
1190         return 0;
1191 }
1192
1193 int nfs_flush_incompatible(struct file *file, struct page *page)
1194 {
1195         struct nfs_open_context *ctx = nfs_file_open_context(file);
1196         struct nfs_lock_context *l_ctx;
1197         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1198         struct nfs_page *req;
1199         int do_flush, status;
1200         /*
1201          * Look for a request corresponding to this page. If there
1202          * is one, and it belongs to another file, we flush it out
1203          * before we try to copy anything into the page. Do this
1204          * due to the lack of an ACCESS-type call in NFSv2.
1205          * Also do the same if we find a request from an existing
1206          * dropped page.
1207          */
1208         do {
1209                 req = nfs_page_find_head_request(page);
1210                 if (req == NULL)
1211                         return 0;
1212                 l_ctx = req->wb_lock_context;
1213                 do_flush = req->wb_page != page ||
1214                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1215                 if (l_ctx && flctx &&
1216                     !(list_empty_careful(&flctx->flc_posix) &&
1217                       list_empty_careful(&flctx->flc_flock))) {
1218                         do_flush |= l_ctx->lockowner != current->files;
1219                 }
1220                 nfs_release_request(req);
1221                 if (!do_flush)
1222                         return 0;
1223                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1224         } while (status == 0);
1225         return status;
1226 }
1227
1228 /*
1229  * Avoid buffered writes when a open context credential's key would
1230  * expire soon.
1231  *
1232  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1233  *
1234  * Return 0 and set a credential flag which triggers the inode to flush
1235  * and performs  NFS_FILE_SYNC writes if the key will expired within
1236  * RPC_KEY_EXPIRE_TIMEO.
1237  */
1238 int
1239 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1240 {
1241         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1242
1243         if (nfs_ctx_key_to_expire(ctx, inode) &&
1244             !rcu_access_pointer(ctx->ll_cred))
1245                 /* Already expired! */
1246                 return -EACCES;
1247         return 0;
1248 }
1249
1250 /*
1251  * Test if the open context credential key is marked to expire soon.
1252  */
1253 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1254 {
1255         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1256         struct rpc_cred *cred, *new, *old = NULL;
1257         struct auth_cred acred = {
1258                 .cred = ctx->cred,
1259         };
1260         bool ret = false;
1261
1262         rcu_read_lock();
1263         cred = rcu_dereference(ctx->ll_cred);
1264         if (cred && !(cred->cr_ops->crkey_timeout &&
1265                       cred->cr_ops->crkey_timeout(cred)))
1266                 goto out;
1267         rcu_read_unlock();
1268
1269         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1270         if (new == cred) {
1271                 put_rpccred(new);
1272                 return true;
1273         }
1274         if (IS_ERR_OR_NULL(new)) {
1275                 new = NULL;
1276                 ret = true;
1277         } else if (new->cr_ops->crkey_timeout &&
1278                    new->cr_ops->crkey_timeout(new))
1279                 ret = true;
1280
1281         rcu_read_lock();
1282         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1283                                              RCU_INITIALIZER(new)), 1);
1284 out:
1285         rcu_read_unlock();
1286         put_rpccred(old);
1287         return ret;
1288 }
1289
1290 /*
1291  * If the page cache is marked as unsafe or invalid, then we can't rely on
1292  * the PageUptodate() flag. In this case, we will need to turn off
1293  * write optimisations that depend on the page contents being correct.
1294  */
1295 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1296                                    unsigned int pagelen)
1297 {
1298         struct nfs_inode *nfsi = NFS_I(inode);
1299
1300         if (nfs_have_delegated_attributes(inode))
1301                 goto out;
1302         if (nfsi->cache_validity &
1303             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1304                 return false;
1305         smp_rmb();
1306         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1307                 return false;
1308 out:
1309         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1310                 return false;
1311         return PageUptodate(page) != 0;
1312 }
1313
1314 static bool
1315 is_whole_file_wrlock(struct file_lock *fl)
1316 {
1317         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1318                         fl->fl_type == F_WRLCK;
1319 }
1320
1321 /* If we know the page is up to date, and we're not using byte range locks (or
1322  * if we have the whole file locked for writing), it may be more efficient to
1323  * extend the write to cover the entire page in order to avoid fragmentation
1324  * inefficiencies.
1325  *
1326  * If the file is opened for synchronous writes then we can just skip the rest
1327  * of the checks.
1328  */
1329 static int nfs_can_extend_write(struct file *file, struct page *page,
1330                                 struct inode *inode, unsigned int pagelen)
1331 {
1332         int ret;
1333         struct file_lock_context *flctx = inode->i_flctx;
1334         struct file_lock *fl;
1335
1336         if (file->f_flags & O_DSYNC)
1337                 return 0;
1338         if (!nfs_write_pageuptodate(page, inode, pagelen))
1339                 return 0;
1340         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1341                 return 1;
1342         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1343                        list_empty_careful(&flctx->flc_posix)))
1344                 return 1;
1345
1346         /* Check to see if there are whole file write locks */
1347         ret = 0;
1348         spin_lock(&flctx->flc_lock);
1349         if (!list_empty(&flctx->flc_posix)) {
1350                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1351                                         fl_list);
1352                 if (is_whole_file_wrlock(fl))
1353                         ret = 1;
1354         } else if (!list_empty(&flctx->flc_flock)) {
1355                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1356                                         fl_list);
1357                 if (fl->fl_type == F_WRLCK)
1358                         ret = 1;
1359         }
1360         spin_unlock(&flctx->flc_lock);
1361         return ret;
1362 }
1363
1364 /*
1365  * Update and possibly write a cached page of an NFS file.
1366  *
1367  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1368  * things with a page scheduled for an RPC call (e.g. invalidate it).
1369  */
1370 int nfs_updatepage(struct file *file, struct page *page,
1371                 unsigned int offset, unsigned int count)
1372 {
1373         struct nfs_open_context *ctx = nfs_file_open_context(file);
1374         struct address_space *mapping = page_file_mapping(page);
1375         struct inode    *inode = mapping->host;
1376         unsigned int    pagelen = nfs_page_length(page);
1377         int             status = 0;
1378
1379         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1380
1381         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1382                 file, count, (long long)(page_file_offset(page) + offset));
1383
1384         if (!count)
1385                 goto out;
1386
1387         if (nfs_can_extend_write(file, page, inode, pagelen)) {
1388                 count = max(count + offset, pagelen);
1389                 offset = 0;
1390         }
1391
1392         status = nfs_writepage_setup(ctx, page, offset, count);
1393         if (status < 0)
1394                 nfs_set_pageerror(mapping);
1395 out:
1396         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1397                         status, (long long)i_size_read(inode));
1398         return status;
1399 }
1400
1401 static int flush_task_priority(int how)
1402 {
1403         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1404                 case FLUSH_HIGHPRI:
1405                         return RPC_PRIORITY_HIGH;
1406                 case FLUSH_LOWPRI:
1407                         return RPC_PRIORITY_LOW;
1408         }
1409         return RPC_PRIORITY_NORMAL;
1410 }
1411
1412 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1413                                struct rpc_message *msg,
1414                                const struct nfs_rpc_ops *rpc_ops,
1415                                struct rpc_task_setup *task_setup_data, int how)
1416 {
1417         int priority = flush_task_priority(how);
1418
1419         if (IS_SWAPFILE(hdr->inode))
1420                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1421         task_setup_data->priority = priority;
1422         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1423         trace_nfs_initiate_write(hdr);
1424 }
1425
1426 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1427  * call this on each, which will prepare them to be retried on next
1428  * writeback using standard nfs.
1429  */
1430 static void nfs_redirty_request(struct nfs_page *req)
1431 {
1432         /* Bump the transmission count */
1433         req->wb_nio++;
1434         nfs_mark_request_dirty(req);
1435         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1436         nfs_end_page_writeback(req);
1437         nfs_release_request(req);
1438 }
1439
1440 static void nfs_async_write_error(struct list_head *head, int error)
1441 {
1442         struct nfs_page *req;
1443
1444         while (!list_empty(head)) {
1445                 req = nfs_list_entry(head->next);
1446                 nfs_list_remove_request(req);
1447                 if (nfs_error_is_fatal(error))
1448                         nfs_write_error(req, error);
1449                 else
1450                         nfs_redirty_request(req);
1451         }
1452 }
1453
1454 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1455 {
1456         nfs_async_write_error(&hdr->pages, 0);
1457         filemap_fdatawrite_range(hdr->inode->i_mapping, hdr->args.offset,
1458                         hdr->args.offset + hdr->args.count - 1);
1459 }
1460
1461 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1462         .init_hdr = nfs_async_write_init,
1463         .error_cleanup = nfs_async_write_error,
1464         .completion = nfs_write_completion,
1465         .reschedule_io = nfs_async_write_reschedule_io,
1466 };
1467
1468 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1469                                struct inode *inode, int ioflags, bool force_mds,
1470                                const struct nfs_pgio_completion_ops *compl_ops)
1471 {
1472         struct nfs_server *server = NFS_SERVER(inode);
1473         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1474
1475 #ifdef CONFIG_NFS_V4_1
1476         if (server->pnfs_curr_ld && !force_mds)
1477                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1478 #endif
1479         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1480                         server->wsize, ioflags);
1481 }
1482 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1483
1484 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1485 {
1486         struct nfs_pgio_mirror *mirror;
1487
1488         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1489                 pgio->pg_ops->pg_cleanup(pgio);
1490
1491         pgio->pg_ops = &nfs_pgio_rw_ops;
1492
1493         nfs_pageio_stop_mirroring(pgio);
1494
1495         mirror = &pgio->pg_mirrors[0];
1496         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1497 }
1498 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1499
1500
1501 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1502 {
1503         struct nfs_commit_data *data = calldata;
1504
1505         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1506 }
1507
1508 /*
1509  * Special version of should_remove_suid() that ignores capabilities.
1510  */
1511 static int nfs_should_remove_suid(const struct inode *inode)
1512 {
1513         umode_t mode = inode->i_mode;
1514         int kill = 0;
1515
1516         /* suid always must be killed */
1517         if (unlikely(mode & S_ISUID))
1518                 kill = ATTR_KILL_SUID;
1519
1520         /*
1521          * sgid without any exec bits is just a mandatory locking mark; leave
1522          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1523          */
1524         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1525                 kill |= ATTR_KILL_SGID;
1526
1527         if (unlikely(kill && S_ISREG(mode)))
1528                 return kill;
1529
1530         return 0;
1531 }
1532
1533 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1534                 struct nfs_fattr *fattr)
1535 {
1536         struct nfs_pgio_args *argp = &hdr->args;
1537         struct nfs_pgio_res *resp = &hdr->res;
1538         u64 size = argp->offset + resp->count;
1539
1540         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1541                 fattr->size = size;
1542         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1543                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1544                 return;
1545         }
1546         if (size != fattr->size)
1547                 return;
1548         /* Set attribute barrier */
1549         nfs_fattr_set_barrier(fattr);
1550         /* ...and update size */
1551         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1552 }
1553
1554 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1555 {
1556         struct nfs_fattr *fattr = &hdr->fattr;
1557         struct inode *inode = hdr->inode;
1558
1559         spin_lock(&inode->i_lock);
1560         nfs_writeback_check_extend(hdr, fattr);
1561         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1562         spin_unlock(&inode->i_lock);
1563 }
1564 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1565
1566 /*
1567  * This function is called when the WRITE call is complete.
1568  */
1569 static int nfs_writeback_done(struct rpc_task *task,
1570                               struct nfs_pgio_header *hdr,
1571                               struct inode *inode)
1572 {
1573         int status;
1574
1575         /*
1576          * ->write_done will attempt to use post-op attributes to detect
1577          * conflicting writes by other clients.  A strict interpretation
1578          * of close-to-open would allow us to continue caching even if
1579          * another writer had changed the file, but some applications
1580          * depend on tighter cache coherency when writing.
1581          */
1582         status = NFS_PROTO(inode)->write_done(task, hdr);
1583         if (status != 0)
1584                 return status;
1585
1586         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1587         trace_nfs_writeback_done(task, hdr);
1588
1589         if (hdr->res.verf->committed < hdr->args.stable &&
1590             task->tk_status >= 0) {
1591                 /* We tried a write call, but the server did not
1592                  * commit data to stable storage even though we
1593                  * requested it.
1594                  * Note: There is a known bug in Tru64 < 5.0 in which
1595                  *       the server reports NFS_DATA_SYNC, but performs
1596                  *       NFS_FILE_SYNC. We therefore implement this checking
1597                  *       as a dprintk() in order to avoid filling syslog.
1598                  */
1599                 static unsigned long    complain;
1600
1601                 /* Note this will print the MDS for a DS write */
1602                 if (time_before(complain, jiffies)) {
1603                         dprintk("NFS:       faulty NFS server %s:"
1604                                 " (committed = %d) != (stable = %d)\n",
1605                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1606                                 hdr->res.verf->committed, hdr->args.stable);
1607                         complain = jiffies + 300 * HZ;
1608                 }
1609         }
1610
1611         /* Deal with the suid/sgid bit corner case */
1612         if (nfs_should_remove_suid(inode)) {
1613                 spin_lock(&inode->i_lock);
1614                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1615                 spin_unlock(&inode->i_lock);
1616         }
1617         return 0;
1618 }
1619
1620 /*
1621  * This function is called when the WRITE call is complete.
1622  */
1623 static void nfs_writeback_result(struct rpc_task *task,
1624                                  struct nfs_pgio_header *hdr)
1625 {
1626         struct nfs_pgio_args    *argp = &hdr->args;
1627         struct nfs_pgio_res     *resp = &hdr->res;
1628
1629         if (resp->count < argp->count) {
1630                 static unsigned long    complain;
1631
1632                 /* This a short write! */
1633                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1634
1635                 /* Has the server at least made some progress? */
1636                 if (resp->count == 0) {
1637                         if (time_before(complain, jiffies)) {
1638                                 printk(KERN_WARNING
1639                                        "NFS: Server wrote zero bytes, expected %u.\n",
1640                                        argp->count);
1641                                 complain = jiffies + 300 * HZ;
1642                         }
1643                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1644                         task->tk_status = -EIO;
1645                         return;
1646                 }
1647
1648                 /* For non rpc-based layout drivers, retry-through-MDS */
1649                 if (!task->tk_ops) {
1650                         hdr->pnfs_error = -EAGAIN;
1651                         return;
1652                 }
1653
1654                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1655                 if (resp->verf->committed != NFS_UNSTABLE) {
1656                         /* Resend from where the server left off */
1657                         hdr->mds_offset += resp->count;
1658                         argp->offset += resp->count;
1659                         argp->pgbase += resp->count;
1660                         argp->count -= resp->count;
1661                 } else {
1662                         /* Resend as a stable write in order to avoid
1663                          * headaches in the case of a server crash.
1664                          */
1665                         argp->stable = NFS_FILE_SYNC;
1666                 }
1667                 resp->count = 0;
1668                 resp->verf->committed = 0;
1669                 rpc_restart_call_prepare(task);
1670         }
1671 }
1672
1673 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1674 {
1675         return wait_var_event_killable(&cinfo->rpcs_out,
1676                                        !atomic_read(&cinfo->rpcs_out));
1677 }
1678
1679 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1680 {
1681         atomic_inc(&cinfo->rpcs_out);
1682 }
1683
1684 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1685 {
1686         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1687                 wake_up_var(&cinfo->rpcs_out);
1688                 return true;
1689         }
1690         return false;
1691 }
1692
1693 void nfs_commitdata_release(struct nfs_commit_data *data)
1694 {
1695         put_nfs_open_context(data->context);
1696         nfs_commit_free(data);
1697 }
1698 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1699
1700 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1701                         const struct nfs_rpc_ops *nfs_ops,
1702                         const struct rpc_call_ops *call_ops,
1703                         int how, int flags)
1704 {
1705         struct rpc_task *task;
1706         int priority = flush_task_priority(how);
1707         struct rpc_message msg = {
1708                 .rpc_argp = &data->args,
1709                 .rpc_resp = &data->res,
1710                 .rpc_cred = data->cred,
1711         };
1712         struct rpc_task_setup task_setup_data = {
1713                 .task = &data->task,
1714                 .rpc_client = clnt,
1715                 .rpc_message = &msg,
1716                 .callback_ops = call_ops,
1717                 .callback_data = data,
1718                 .workqueue = nfsiod_workqueue,
1719                 .flags = RPC_TASK_ASYNC | flags,
1720                 .priority = priority,
1721         };
1722         /* Set up the initial task struct.  */
1723         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1724         trace_nfs_initiate_commit(data);
1725
1726         dprintk("NFS: initiated commit call\n");
1727
1728         task = rpc_run_task(&task_setup_data);
1729         if (IS_ERR(task))
1730                 return PTR_ERR(task);
1731         if (how & FLUSH_SYNC)
1732                 rpc_wait_for_completion_task(task);
1733         rpc_put_task(task);
1734         return 0;
1735 }
1736 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1737
1738 static loff_t nfs_get_lwb(struct list_head *head)
1739 {
1740         loff_t lwb = 0;
1741         struct nfs_page *req;
1742
1743         list_for_each_entry(req, head, wb_list)
1744                 if (lwb < (req_offset(req) + req->wb_bytes))
1745                         lwb = req_offset(req) + req->wb_bytes;
1746
1747         return lwb;
1748 }
1749
1750 /*
1751  * Set up the argument/result storage required for the RPC call.
1752  */
1753 void nfs_init_commit(struct nfs_commit_data *data,
1754                      struct list_head *head,
1755                      struct pnfs_layout_segment *lseg,
1756                      struct nfs_commit_info *cinfo)
1757 {
1758         struct nfs_page *first;
1759         struct nfs_open_context *ctx;
1760         struct inode *inode;
1761
1762         /* Set up the RPC argument and reply structs
1763          * NB: take care not to mess about with data->commit et al. */
1764
1765         if (head)
1766                 list_splice_init(head, &data->pages);
1767
1768         first = nfs_list_entry(data->pages.next);
1769         ctx = nfs_req_openctx(first);
1770         inode = d_inode(ctx->dentry);
1771
1772         data->inode       = inode;
1773         data->cred        = ctx->cred;
1774         data->lseg        = lseg; /* reference transferred */
1775         /* only set lwb for pnfs commit */
1776         if (lseg)
1777                 data->lwb = nfs_get_lwb(&data->pages);
1778         data->mds_ops     = &nfs_commit_ops;
1779         data->completion_ops = cinfo->completion_ops;
1780         data->dreq        = cinfo->dreq;
1781
1782         data->args.fh     = NFS_FH(data->inode);
1783         /* Note: we always request a commit of the entire inode */
1784         data->args.offset = 0;
1785         data->args.count  = 0;
1786         data->context     = get_nfs_open_context(ctx);
1787         data->res.fattr   = &data->fattr;
1788         data->res.verf    = &data->verf;
1789         nfs_fattr_init(&data->fattr);
1790         nfs_commit_begin(cinfo->mds);
1791 }
1792 EXPORT_SYMBOL_GPL(nfs_init_commit);
1793
1794 void nfs_retry_commit(struct list_head *page_list,
1795                       struct pnfs_layout_segment *lseg,
1796                       struct nfs_commit_info *cinfo,
1797                       u32 ds_commit_idx)
1798 {
1799         struct nfs_page *req;
1800
1801         while (!list_empty(page_list)) {
1802                 req = nfs_list_entry(page_list->next);
1803                 nfs_list_remove_request(req);
1804                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1805                 if (!cinfo->dreq)
1806                         nfs_clear_page_commit(req->wb_page);
1807                 nfs_unlock_and_release_request(req);
1808         }
1809 }
1810 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1811
1812 static void
1813 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1814                 struct nfs_page *req)
1815 {
1816         __set_page_dirty_nobuffers(req->wb_page);
1817 }
1818
1819 /*
1820  * Commit dirty pages
1821  */
1822 static int
1823 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1824                 struct nfs_commit_info *cinfo)
1825 {
1826         struct nfs_commit_data  *data;
1827         unsigned short task_flags = 0;
1828
1829         /* another commit raced with us */
1830         if (list_empty(head))
1831                 return 0;
1832
1833         data = nfs_commitdata_alloc();
1834         if (!data) {
1835                 nfs_retry_commit(head, NULL, cinfo, -1);
1836                 return -ENOMEM;
1837         }
1838
1839         /* Set up the argument struct */
1840         nfs_init_commit(data, head, NULL, cinfo);
1841         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1842                 task_flags = RPC_TASK_MOVEABLE;
1843         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1844                                    data->mds_ops, how,
1845                                    RPC_TASK_CRED_NOREF | task_flags);
1846 }
1847
1848 /*
1849  * COMMIT call returned
1850  */
1851 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1852 {
1853         struct nfs_commit_data  *data = calldata;
1854
1855         /* Call the NFS version-specific code */
1856         NFS_PROTO(data->inode)->commit_done(task, data);
1857         trace_nfs_commit_done(task, data);
1858 }
1859
1860 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1861 {
1862         const struct nfs_writeverf *verf = data->res.verf;
1863         struct nfs_page *req;
1864         int status = data->task.tk_status;
1865         struct nfs_commit_info cinfo;
1866         struct nfs_server *nfss;
1867
1868         while (!list_empty(&data->pages)) {
1869                 req = nfs_list_entry(data->pages.next);
1870                 nfs_list_remove_request(req);
1871                 if (req->wb_page)
1872                         nfs_clear_page_commit(req->wb_page);
1873
1874                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1875                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1876                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1877                         req->wb_bytes,
1878                         (long long)req_offset(req));
1879                 if (status < 0) {
1880                         if (req->wb_page) {
1881                                 trace_nfs_commit_error(req, status);
1882                                 nfs_mapping_set_error(req->wb_page, status);
1883                                 nfs_inode_remove_request(req);
1884                         }
1885                         dprintk_cont(", error = %d\n", status);
1886                         goto next;
1887                 }
1888
1889                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1890                  * returned by the server against all stored verfs. */
1891                 if (nfs_write_match_verf(verf, req)) {
1892                         /* We have a match */
1893                         if (req->wb_page)
1894                                 nfs_inode_remove_request(req);
1895                         dprintk_cont(" OK\n");
1896                         goto next;
1897                 }
1898                 /* We have a mismatch. Write the page again */
1899                 dprintk_cont(" mismatch\n");
1900                 nfs_mark_request_dirty(req);
1901                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1902         next:
1903                 nfs_unlock_and_release_request(req);
1904                 /* Latency breaker */
1905                 cond_resched();
1906         }
1907         nfss = NFS_SERVER(data->inode);
1908         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1909                 nfss->write_congested = 0;
1910
1911         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1912         nfs_commit_end(cinfo.mds);
1913 }
1914
1915 static void nfs_commit_release(void *calldata)
1916 {
1917         struct nfs_commit_data *data = calldata;
1918
1919         data->completion_ops->completion(data);
1920         nfs_commitdata_release(calldata);
1921 }
1922
1923 static const struct rpc_call_ops nfs_commit_ops = {
1924         .rpc_call_prepare = nfs_commit_prepare,
1925         .rpc_call_done = nfs_commit_done,
1926         .rpc_release = nfs_commit_release,
1927 };
1928
1929 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1930         .completion = nfs_commit_release_pages,
1931         .resched_write = nfs_commit_resched_write,
1932 };
1933
1934 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1935                             int how, struct nfs_commit_info *cinfo)
1936 {
1937         int status;
1938
1939         status = pnfs_commit_list(inode, head, how, cinfo);
1940         if (status == PNFS_NOT_ATTEMPTED)
1941                 status = nfs_commit_list(inode, head, how, cinfo);
1942         return status;
1943 }
1944
1945 static int __nfs_commit_inode(struct inode *inode, int how,
1946                 struct writeback_control *wbc)
1947 {
1948         LIST_HEAD(head);
1949         struct nfs_commit_info cinfo;
1950         int may_wait = how & FLUSH_SYNC;
1951         int ret, nscan;
1952
1953         how &= ~FLUSH_SYNC;
1954         nfs_init_cinfo_from_inode(&cinfo, inode);
1955         nfs_commit_begin(cinfo.mds);
1956         for (;;) {
1957                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1958                 if (ret <= 0)
1959                         break;
1960                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1961                 if (ret < 0)
1962                         break;
1963                 ret = 0;
1964                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1965                         if (nscan < wbc->nr_to_write)
1966                                 wbc->nr_to_write -= nscan;
1967                         else
1968                                 wbc->nr_to_write = 0;
1969                 }
1970                 if (nscan < INT_MAX)
1971                         break;
1972                 cond_resched();
1973         }
1974         nfs_commit_end(cinfo.mds);
1975         if (ret || !may_wait)
1976                 return ret;
1977         return wait_on_commit(cinfo.mds);
1978 }
1979
1980 int nfs_commit_inode(struct inode *inode, int how)
1981 {
1982         return __nfs_commit_inode(inode, how, NULL);
1983 }
1984 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1985
1986 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1987 {
1988         struct nfs_inode *nfsi = NFS_I(inode);
1989         int flags = FLUSH_SYNC;
1990         int ret = 0;
1991
1992         if (wbc->sync_mode == WB_SYNC_NONE) {
1993                 /* no commits means nothing needs to be done */
1994                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1995                         goto check_requests_outstanding;
1996
1997                 /* Don't commit yet if this is a non-blocking flush and there
1998                  * are a lot of outstanding writes for this mapping.
1999                  */
2000                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
2001                         goto out_mark_dirty;
2002
2003                 /* don't wait for the COMMIT response */
2004                 flags = 0;
2005         }
2006
2007         ret = __nfs_commit_inode(inode, flags, wbc);
2008         if (!ret) {
2009                 if (flags & FLUSH_SYNC)
2010                         return 0;
2011         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
2012                 goto out_mark_dirty;
2013
2014 check_requests_outstanding:
2015         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2016                 return ret;
2017 out_mark_dirty:
2018         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2019         return ret;
2020 }
2021 EXPORT_SYMBOL_GPL(nfs_write_inode);
2022
2023 /*
2024  * Wrapper for filemap_write_and_wait_range()
2025  *
2026  * Needed for pNFS in order to ensure data becomes visible to the
2027  * client.
2028  */
2029 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2030                 loff_t lstart, loff_t lend)
2031 {
2032         int ret;
2033
2034         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2035         if (ret == 0)
2036                 ret = pnfs_sync_inode(mapping->host, true);
2037         return ret;
2038 }
2039 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2040
2041 /*
2042  * flush the inode to disk.
2043  */
2044 int nfs_wb_all(struct inode *inode)
2045 {
2046         int ret;
2047
2048         trace_nfs_writeback_inode_enter(inode);
2049
2050         ret = filemap_write_and_wait(inode->i_mapping);
2051         if (ret)
2052                 goto out;
2053         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2054         if (ret < 0)
2055                 goto out;
2056         pnfs_sync_inode(inode, true);
2057         ret = 0;
2058
2059 out:
2060         trace_nfs_writeback_inode_exit(inode, ret);
2061         return ret;
2062 }
2063 EXPORT_SYMBOL_GPL(nfs_wb_all);
2064
2065 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2066 {
2067         struct nfs_page *req;
2068         int ret = 0;
2069
2070         folio_wait_writeback(folio);
2071
2072         /* blocking call to cancel all requests and join to a single (head)
2073          * request */
2074         req = nfs_lock_and_join_requests(&folio->page);
2075
2076         if (IS_ERR(req)) {
2077                 ret = PTR_ERR(req);
2078         } else if (req) {
2079                 /* all requests from this folio have been cancelled by
2080                  * nfs_lock_and_join_requests, so just remove the head
2081                  * request from the inode / page_private pointer and
2082                  * release it */
2083                 nfs_inode_remove_request(req);
2084                 nfs_unlock_and_release_request(req);
2085         }
2086
2087         return ret;
2088 }
2089
2090 /*
2091  * Write back all requests on one page - we do this before reading it.
2092  */
2093 int nfs_wb_page(struct inode *inode, struct page *page)
2094 {
2095         loff_t range_start = page_file_offset(page);
2096         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2097         struct writeback_control wbc = {
2098                 .sync_mode = WB_SYNC_ALL,
2099                 .nr_to_write = 0,
2100                 .range_start = range_start,
2101                 .range_end = range_end,
2102         };
2103         int ret;
2104
2105         trace_nfs_writeback_page_enter(inode);
2106
2107         for (;;) {
2108                 wait_on_page_writeback(page);
2109                 if (clear_page_dirty_for_io(page)) {
2110                         ret = nfs_writepage_locked(page, &wbc);
2111                         if (ret < 0)
2112                                 goto out_error;
2113                         continue;
2114                 }
2115                 ret = 0;
2116                 if (!PagePrivate(page))
2117                         break;
2118                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2119                 if (ret < 0)
2120                         goto out_error;
2121         }
2122 out_error:
2123         trace_nfs_writeback_page_exit(inode, ret);
2124         return ret;
2125 }
2126
2127 #ifdef CONFIG_MIGRATION
2128 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
2129                 struct page *page, enum migrate_mode mode)
2130 {
2131         /*
2132          * If PagePrivate is set, then the page is currently associated with
2133          * an in-progress read or write request. Don't try to migrate it.
2134          *
2135          * FIXME: we could do this in principle, but we'll need a way to ensure
2136          *        that we can safely release the inode reference while holding
2137          *        the page lock.
2138          */
2139         if (PagePrivate(page))
2140                 return -EBUSY;
2141
2142         if (PageFsCache(page)) {
2143                 if (mode == MIGRATE_ASYNC)
2144                         return -EBUSY;
2145                 wait_on_page_fscache(page);
2146         }
2147
2148         return migrate_page(mapping, newpage, page, mode);
2149 }
2150 #endif
2151
2152 int __init nfs_init_writepagecache(void)
2153 {
2154         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2155                                              sizeof(struct nfs_pgio_header),
2156                                              0, SLAB_HWCACHE_ALIGN,
2157                                              NULL);
2158         if (nfs_wdata_cachep == NULL)
2159                 return -ENOMEM;
2160
2161         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2162                                                      nfs_wdata_cachep);
2163         if (nfs_wdata_mempool == NULL)
2164                 goto out_destroy_write_cache;
2165
2166         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2167                                              sizeof(struct nfs_commit_data),
2168                                              0, SLAB_HWCACHE_ALIGN,
2169                                              NULL);
2170         if (nfs_cdata_cachep == NULL)
2171                 goto out_destroy_write_mempool;
2172
2173         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2174                                                       nfs_cdata_cachep);
2175         if (nfs_commit_mempool == NULL)
2176                 goto out_destroy_commit_cache;
2177
2178         /*
2179          * NFS congestion size, scale with available memory.
2180          *
2181          *  64MB:    8192k
2182          * 128MB:   11585k
2183          * 256MB:   16384k
2184          * 512MB:   23170k
2185          *   1GB:   32768k
2186          *   2GB:   46340k
2187          *   4GB:   65536k
2188          *   8GB:   92681k
2189          *  16GB:  131072k
2190          *
2191          * This allows larger machines to have larger/more transfers.
2192          * Limit the default to 256M
2193          */
2194         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2195         if (nfs_congestion_kb > 256*1024)
2196                 nfs_congestion_kb = 256*1024;
2197
2198         return 0;
2199
2200 out_destroy_commit_cache:
2201         kmem_cache_destroy(nfs_cdata_cachep);
2202 out_destroy_write_mempool:
2203         mempool_destroy(nfs_wdata_mempool);
2204 out_destroy_write_cache:
2205         kmem_cache_destroy(nfs_wdata_cachep);
2206         return -ENOMEM;
2207 }
2208
2209 void nfs_destroy_writepagecache(void)
2210 {
2211         mempool_destroy(nfs_commit_mempool);
2212         kmem_cache_destroy(nfs_cdata_cachep);
2213         mempool_destroy(nfs_wdata_mempool);
2214         kmem_cache_destroy(nfs_wdata_cachep);
2215 }
2216
2217 static const struct nfs_rw_ops nfs_rw_write_ops = {
2218         .rw_alloc_header        = nfs_writehdr_alloc,
2219         .rw_free_header         = nfs_writehdr_free,
2220         .rw_done                = nfs_writeback_done,
2221         .rw_result              = nfs_writeback_result,
2222         .rw_initiate            = nfs_initiate_write,
2223 };