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(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                 return AOP_WRITEPAGE_ACTIVATE;
 672
 673         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 674         nfs_pageio_init_write(&pgio, inode, 0,
 675                                 false, &nfs_async_write_completion_ops);
 676         err = nfs_do_writepage(page, wbc, &pgio);
 677         pgio.pg_error = 0;
 678         nfs_pageio_complete(&pgio);
 679         return err;
 680 }
 681
 682 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 683 {
 684         int ret;
 685
 686         ret = nfs_writepage_locked(page, wbc);
 687         if (ret != AOP_WRITEPAGE_ACTIVATE)
 688                 unlock_page(page);
 689         return ret;
 690 }
 691
 692 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 693 {
 694         int ret;
 695
 696         ret = nfs_do_writepage(page, wbc, data);
 697         if (ret != AOP_WRITEPAGE_ACTIVATE)
 698                 unlock_page(page);
 699         return ret;
 700 }
 701
 702 static void nfs_io_completion_commit(void *inode)
 703 {
 704         nfs_commit_inode(inode, 0);
 705 }
 706
 707 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 708 {
 709         struct inode *inode = mapping->host;
 710         struct nfs_pageio_descriptor pgio;
 711         struct nfs_io_completion *ioc = NULL;
 712         unsigned int mntflags = NFS_SERVER(inode)->flags;
 713         int priority = 0;
 714         int err;
 715
 716         if (wbc->sync_mode == WB_SYNC_NONE &&
 717             NFS_SERVER(inode)->write_congested)
 718                 return 0;
 719
 720         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 721
 722         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 723             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 724                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 725                 if (ioc)
 726                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 727                                                inode);
 728                 priority = wb_priority(wbc);
 729         }
 730
 731         do {
 732                 nfs_pageio_init_write(&pgio, inode, priority, false,
 733                                       &nfs_async_write_completion_ops);
 734                 pgio.pg_io_completion = ioc;
 735                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 736                                         &pgio);
 737                 pgio.pg_error = 0;
 738                 nfs_pageio_complete(&pgio);
 739         } while (err < 0 && !nfs_error_is_fatal(err));
 740         nfs_io_completion_put(ioc);
 741
 742         if (err < 0)
 743                 goto out_err;
 744         return 0;
 745 out_err:
 746         return err;
 747 }
 748
 749 /*
 750  * Insert a write request into an inode
 751  */
 752 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 753 {
 754         struct address_space *mapping = page_file_mapping(req->wb_page);
 755         struct nfs_inode *nfsi = NFS_I(inode);
 756
 757         WARN_ON_ONCE(req->wb_this_page != req);
 758
 759         /* Lock the request! */
 760         nfs_lock_request(req);
 761
 762         /*
 763          * Swap-space should not get truncated. Hence no need to plug the race
 764          * with invalidate/truncate.
 765          */
 766         spin_lock(&mapping->private_lock);
 767         if (likely(!PageSwapCache(req->wb_page))) {
 768                 set_bit(PG_MAPPED, &req->wb_flags);
 769                 SetPagePrivate(req->wb_page);
 770                 set_page_private(req->wb_page, (unsigned long)req);
 771         }
 772         spin_unlock(&mapping->private_lock);
 773         atomic_long_inc(&nfsi->nrequests);
 774         /* this a head request for a page group - mark it as having an
 775          * extra reference so sub groups can follow suit.
 776          * This flag also informs pgio layer when to bump nrequests when
 777          * adding subrequests. */
 778         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 779         kref_get(&req->wb_kref);
 780 }
 781
 782 /*
 783  * Remove a write request from an inode
 784  */
 785 static void nfs_inode_remove_request(struct nfs_page *req)
 786 {
 787         struct address_space *mapping = page_file_mapping(req->wb_page);
 788         struct inode *inode = mapping->host;
 789         struct nfs_inode *nfsi = NFS_I(inode);
 790         struct nfs_page *head;
 791
 792         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 793                 head = req->wb_head;
 794
 795                 spin_lock(&mapping->private_lock);
 796                 if (likely(head->wb_page && !PageSwapCache(head->wb_page))) {
 797                         set_page_private(head->wb_page, 0);
 798                         ClearPagePrivate(head->wb_page);
 799                         clear_bit(PG_MAPPED, &head->wb_flags);
 800                 }
 801                 spin_unlock(&mapping->private_lock);
 802         }
 803
 804         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 805                 nfs_release_request(req);
 806                 atomic_long_dec(&nfsi->nrequests);
 807         }
 808 }
 809
 810 static void
 811 nfs_mark_request_dirty(struct nfs_page *req)
 812 {
 813         if (req->wb_page)
 814                 __set_page_dirty_nobuffers(req->wb_page);
 815 }
 816
 817 /*
 818  * nfs_page_search_commits_for_head_request_locked
 819  *
 820  * Search through commit lists on @inode for the head request for @page.
 821  * Must be called while holding the inode (which is cinfo) lock.
 822  *
 823  * Returns the head request if found, or NULL if not found.
 824  */
 825 static struct nfs_page *
 826 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 827                                                 struct page *page)
 828 {
 829         struct nfs_page *freq, *t;
 830         struct nfs_commit_info cinfo;
 831         struct inode *inode = &nfsi->vfs_inode;
 832
 833         nfs_init_cinfo_from_inode(&cinfo, inode);
 834
 835         /* search through pnfs commit lists */
 836         freq = pnfs_search_commit_reqs(inode, &cinfo, page);
 837         if (freq)
 838                 return freq->wb_head;
 839
 840         /* Linearly search the commit list for the correct request */
 841         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 842                 if (freq->wb_page == page)
 843                         return freq->wb_head;
 844         }
 845
 846         return NULL;
 847 }
 848
 849 /**
 850  * nfs_request_add_commit_list_locked - add request to a commit list
 851  * @req: pointer to a struct nfs_page
 852  * @dst: commit list head
 853  * @cinfo: holds list lock and accounting info
 854  *
 855  * This sets the PG_CLEAN bit, updates the cinfo count of
 856  * number of outstanding requests requiring a commit as well as
 857  * the MM page stats.
 858  *
 859  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 860  * nfs_page lock.
 861  */
 862 void
 863 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 864                             struct nfs_commit_info *cinfo)
 865 {
 866         set_bit(PG_CLEAN, &req->wb_flags);
 867         nfs_list_add_request(req, dst);
 868         atomic_long_inc(&cinfo->mds->ncommit);
 869 }
 870 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 871
 872 /**
 873  * nfs_request_add_commit_list - add request to a commit list
 874  * @req: pointer to a struct nfs_page
 875  * @cinfo: holds list lock and accounting info
 876  *
 877  * This sets the PG_CLEAN bit, updates the cinfo count of
 878  * number of outstanding requests requiring a commit as well as
 879  * the MM page stats.
 880  *
 881  * The caller must _not_ hold the cinfo->lock, but must be
 882  * holding the nfs_page lock.
 883  */
 884 void
 885 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 886 {
 887         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 888         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 889         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 890         if (req->wb_page)
 891                 nfs_mark_page_unstable(req->wb_page, cinfo);
 892 }
 893 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 894
 895 /**
 896  * nfs_request_remove_commit_list - Remove request from a commit list
 897  * @req: pointer to a nfs_page
 898  * @cinfo: holds list lock and accounting info
 899  *
 900  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 901  * number of outstanding requests requiring a commit
 902  * It does not update the MM page stats.
 903  *
 904  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 905  */
 906 void
 907 nfs_request_remove_commit_list(struct nfs_page *req,
 908                                struct nfs_commit_info *cinfo)
 909 {
 910         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 911                 return;
 912         nfs_list_remove_request(req);
 913         atomic_long_dec(&cinfo->mds->ncommit);
 914 }
 915 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 916
 917 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 918                                       struct inode *inode)
 919 {
 920         cinfo->inode = inode;
 921         cinfo->mds = &NFS_I(inode)->commit_info;
 922         cinfo->ds = pnfs_get_ds_info(inode);
 923         cinfo->dreq = NULL;
 924         cinfo->completion_ops = &nfs_commit_completion_ops;
 925 }
 926
 927 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 928                     struct inode *inode,
 929                     struct nfs_direct_req *dreq)
 930 {
 931         if (dreq)
 932                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 933         else
 934                 nfs_init_cinfo_from_inode(cinfo, inode);
 935 }
 936 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 937
 938 /*
 939  * Add a request to the inode's commit list.
 940  */
 941 void
 942 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 943                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 944 {
 945         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 946                 return;
 947         nfs_request_add_commit_list(req, cinfo);
 948 }
 949
 950 static void
 951 nfs_clear_page_commit(struct page *page)
 952 {
 953         dec_node_page_state(page, NR_WRITEBACK);
 954         dec_wb_stat(&inode_to_bdi(page_file_mapping(page)->host)->wb,
 955                     WB_WRITEBACK);
 956 }
 957
 958 /* Called holding the request lock on @req */
 959 static void
 960 nfs_clear_request_commit(struct nfs_page *req)
 961 {
 962         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 963                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 964                 struct inode *inode = d_inode(ctx->dentry);
 965                 struct nfs_commit_info cinfo;
 966
 967                 nfs_init_cinfo_from_inode(&cinfo, inode);
 968                 mutex_lock(&NFS_I(inode)->commit_mutex);
 969                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 970                         nfs_request_remove_commit_list(req, &cinfo);
 971                 }
 972                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 973                 nfs_clear_page_commit(req->wb_page);
 974         }
 975 }
 976
 977 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 978 {
 979         if (hdr->verf.committed == NFS_DATA_SYNC)
 980                 return hdr->lseg == NULL;
 981         return hdr->verf.committed != NFS_FILE_SYNC;
 982 }
 983
 984 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 985 {
 986         nfs_io_completion_get(hdr->io_completion);
 987 }
 988
 989 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 990 {
 991         struct nfs_commit_info cinfo;
 992         unsigned long bytes = 0;
 993
 994         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 995                 goto out;
 996         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 997         while (!list_empty(&hdr->pages)) {
 998                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 999
1000                 bytes += req->wb_bytes;
1001                 nfs_list_remove_request(req);
1002                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1003                     (hdr->good_bytes < bytes)) {
1004                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
1005                         nfs_mapping_set_error(req->wb_page, hdr->error);
1006                         goto remove_req;
1007                 }
1008                 if (nfs_write_need_commit(hdr)) {
1009                         /* Reset wb_nio, since the write was successful. */
1010                         req->wb_nio = 0;
1011                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1012                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1013                                 hdr->pgio_mirror_idx);
1014                         goto next;
1015                 }
1016 remove_req:
1017                 nfs_inode_remove_request(req);
1018 next:
1019                 nfs_end_page_writeback(req);
1020                 nfs_release_request(req);
1021         }
1022 out:
1023         nfs_io_completion_put(hdr->io_completion);
1024         hdr->release(hdr);
1025 }
1026
1027 unsigned long
1028 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1029 {
1030         return atomic_long_read(&cinfo->mds->ncommit);
1031 }
1032
1033 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1034 int
1035 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1036                      struct nfs_commit_info *cinfo, int max)
1037 {
1038         struct nfs_page *req, *tmp;
1039         int ret = 0;
1040
1041         list_for_each_entry_safe(req, tmp, src, wb_list) {
1042                 kref_get(&req->wb_kref);
1043                 if (!nfs_lock_request(req)) {
1044                         nfs_release_request(req);
1045                         continue;
1046                 }
1047                 nfs_request_remove_commit_list(req, cinfo);
1048                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1049                 nfs_list_add_request(req, dst);
1050                 ret++;
1051                 if ((ret == max) && !cinfo->dreq)
1052                         break;
1053                 cond_resched();
1054         }
1055         return ret;
1056 }
1057 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1058
1059 /*
1060  * nfs_scan_commit - Scan an inode for commit requests
1061  * @inode: NFS inode to scan
1062  * @dst: mds destination list
1063  * @cinfo: mds and ds lists of reqs ready to commit
1064  *
1065  * Moves requests from the inode's 'commit' request list.
1066  * The requests are *not* checked to ensure that they form a contiguous set.
1067  */
1068 int
1069 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1070                 struct nfs_commit_info *cinfo)
1071 {
1072         int ret = 0;
1073
1074         if (!atomic_long_read(&cinfo->mds->ncommit))
1075                 return 0;
1076         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1077         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1078                 const int max = INT_MAX;
1079
1080                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1081                                            cinfo, max);
1082                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1083         }
1084         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1085         return ret;
1086 }
1087
1088 /*
1089  * Search for an existing write request, and attempt to update
1090  * it to reflect a new dirty region on a given page.
1091  *
1092  * If the attempt fails, then the existing request is flushed out
1093  * to disk.
1094  */
1095 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
1096                 struct page *page,
1097                 unsigned int offset,
1098                 unsigned int bytes)
1099 {
1100         struct nfs_page *req;
1101         unsigned int rqend;
1102         unsigned int end;
1103         int error;
1104
1105         end = offset + bytes;
1106
1107         req = nfs_lock_and_join_requests(page);
1108         if (IS_ERR_OR_NULL(req))
1109                 return req;
1110
1111         rqend = req->wb_offset + req->wb_bytes;
1112         /*
1113          * Tell the caller to flush out the request if
1114          * the offsets are non-contiguous.
1115          * Note: nfs_flush_incompatible() will already
1116          * have flushed out requests having wrong owners.
1117          */
1118         if (offset > rqend || end < req->wb_offset)
1119                 goto out_flushme;
1120
1121         /* Okay, the request matches. Update the region */
1122         if (offset < req->wb_offset) {
1123                 req->wb_offset = offset;
1124                 req->wb_pgbase = offset;
1125         }
1126         if (end > rqend)
1127                 req->wb_bytes = end - req->wb_offset;
1128         else
1129                 req->wb_bytes = rqend - req->wb_offset;
1130         req->wb_nio = 0;
1131         return req;
1132 out_flushme:
1133         /*
1134          * Note: we mark the request dirty here because
1135          * nfs_lock_and_join_requests() cannot preserve
1136          * commit flags, so we have to replay the write.
1137          */
1138         nfs_mark_request_dirty(req);
1139         nfs_unlock_and_release_request(req);
1140         error = nfs_wb_page(inode, page);
1141         return (error < 0) ? ERR_PTR(error) : NULL;
1142 }
1143
1144 /*
1145  * Try to update an existing write request, or create one if there is none.
1146  *
1147  * Note: Should always be called with the Page Lock held to prevent races
1148  * if we have to add a new request. Also assumes that the caller has
1149  * already called nfs_flush_incompatible() if necessary.
1150  */
1151 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
1152                 struct page *page, unsigned int offset, unsigned int bytes)
1153 {
1154         struct inode *inode = page_file_mapping(page)->host;
1155         struct nfs_page *req;
1156
1157         req = nfs_try_to_update_request(inode, page, offset, bytes);
1158         if (req != NULL)
1159                 goto out;
1160         req = nfs_create_request(ctx, page, offset, bytes);
1161         if (IS_ERR(req))
1162                 goto out;
1163         nfs_inode_add_request(inode, req);
1164 out:
1165         return req;
1166 }
1167
1168 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
1169                 unsigned int offset, unsigned int count)
1170 {
1171         struct nfs_page *req;
1172
1173         req = nfs_setup_write_request(ctx, page, offset, count);
1174         if (IS_ERR(req))
1175                 return PTR_ERR(req);
1176         /* Update file length */
1177         nfs_grow_file(page, offset, count);
1178         nfs_mark_uptodate(req);
1179         nfs_mark_request_dirty(req);
1180         nfs_unlock_and_release_request(req);
1181         return 0;
1182 }
1183
1184 int nfs_flush_incompatible(struct file *file, struct page *page)
1185 {
1186         struct nfs_open_context *ctx = nfs_file_open_context(file);
1187         struct nfs_lock_context *l_ctx;
1188         struct file_lock_context *flctx = file_inode(file)->i_flctx;
1189         struct nfs_page *req;
1190         int do_flush, status;
1191         /*
1192          * Look for a request corresponding to this page. If there
1193          * is one, and it belongs to another file, we flush it out
1194          * before we try to copy anything into the page. Do this
1195          * due to the lack of an ACCESS-type call in NFSv2.
1196          * Also do the same if we find a request from an existing
1197          * dropped page.
1198          */
1199         do {
1200                 req = nfs_page_find_head_request(page);
1201                 if (req == NULL)
1202                         return 0;
1203                 l_ctx = req->wb_lock_context;
1204                 do_flush = req->wb_page != page ||
1205                         !nfs_match_open_context(nfs_req_openctx(req), ctx);
1206                 if (l_ctx && flctx &&
1207                     !(list_empty_careful(&flctx->flc_posix) &&
1208                       list_empty_careful(&flctx->flc_flock))) {
1209                         do_flush |= l_ctx->lockowner != current->files;
1210                 }
1211                 nfs_release_request(req);
1212                 if (!do_flush)
1213                         return 0;
1214                 status = nfs_wb_page(page_file_mapping(page)->host, page);
1215         } while (status == 0);
1216         return status;
1217 }
1218
1219 /*
1220  * Avoid buffered writes when a open context credential's key would
1221  * expire soon.
1222  *
1223  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1224  *
1225  * Return 0 and set a credential flag which triggers the inode to flush
1226  * and performs  NFS_FILE_SYNC writes if the key will expired within
1227  * RPC_KEY_EXPIRE_TIMEO.
1228  */
1229 int
1230 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1231 {
1232         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1233
1234         if (nfs_ctx_key_to_expire(ctx, inode) &&
1235             !rcu_access_pointer(ctx->ll_cred))
1236                 /* Already expired! */
1237                 return -EACCES;
1238         return 0;
1239 }
1240
1241 /*
1242  * Test if the open context credential key is marked to expire soon.
1243  */
1244 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1245 {
1246         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1247         struct rpc_cred *cred, *new, *old = NULL;
1248         struct auth_cred acred = {
1249                 .cred = ctx->cred,
1250         };
1251         bool ret = false;
1252
1253         rcu_read_lock();
1254         cred = rcu_dereference(ctx->ll_cred);
1255         if (cred && !(cred->cr_ops->crkey_timeout &&
1256                       cred->cr_ops->crkey_timeout(cred)))
1257                 goto out;
1258         rcu_read_unlock();
1259
1260         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1261         if (new == cred) {
1262                 put_rpccred(new);
1263                 return true;
1264         }
1265         if (IS_ERR_OR_NULL(new)) {
1266                 new = NULL;
1267                 ret = true;
1268         } else if (new->cr_ops->crkey_timeout &&
1269                    new->cr_ops->crkey_timeout(new))
1270                 ret = true;
1271
1272         rcu_read_lock();
1273         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1274                                              RCU_INITIALIZER(new)), 1);
1275 out:
1276         rcu_read_unlock();
1277         put_rpccred(old);
1278         return ret;
1279 }
1280
1281 /*
1282  * If the page cache is marked as unsafe or invalid, then we can't rely on
1283  * the PageUptodate() flag. In this case, we will need to turn off
1284  * write optimisations that depend on the page contents being correct.
1285  */
1286 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode,
1287                                    unsigned int pagelen)
1288 {
1289         struct nfs_inode *nfsi = NFS_I(inode);
1290
1291         if (nfs_have_delegated_attributes(inode))
1292                 goto out;
1293         if (nfsi->cache_validity &
1294             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1295                 return false;
1296         smp_rmb();
1297         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1298                 return false;
1299 out:
1300         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1301                 return false;
1302         return PageUptodate(page) != 0;
1303 }
1304
1305 static bool
1306 is_whole_file_wrlock(struct file_lock *fl)
1307 {
1308         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1309                         fl->fl_type == F_WRLCK;
1310 }
1311
1312 /* If we know the page is up to date, and we're not using byte range locks (or
1313  * if we have the whole file locked for writing), it may be more efficient to
1314  * extend the write to cover the entire page in order to avoid fragmentation
1315  * inefficiencies.
1316  *
1317  * If the file is opened for synchronous writes then we can just skip the rest
1318  * of the checks.
1319  */
1320 static int nfs_can_extend_write(struct file *file, struct page *page,
1321                                 struct inode *inode, unsigned int pagelen)
1322 {
1323         int ret;
1324         struct file_lock_context *flctx = inode->i_flctx;
1325         struct file_lock *fl;
1326
1327         if (file->f_flags & O_DSYNC)
1328                 return 0;
1329         if (!nfs_write_pageuptodate(page, inode, pagelen))
1330                 return 0;
1331         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1332                 return 1;
1333         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1334                        list_empty_careful(&flctx->flc_posix)))
1335                 return 1;
1336
1337         /* Check to see if there are whole file write locks */
1338         ret = 0;
1339         spin_lock(&flctx->flc_lock);
1340         if (!list_empty(&flctx->flc_posix)) {
1341                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1342                                         fl_list);
1343                 if (is_whole_file_wrlock(fl))
1344                         ret = 1;
1345         } else if (!list_empty(&flctx->flc_flock)) {
1346                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1347                                         fl_list);
1348                 if (fl->fl_type == F_WRLCK)
1349                         ret = 1;
1350         }
1351         spin_unlock(&flctx->flc_lock);
1352         return ret;
1353 }
1354
1355 /*
1356  * Update and possibly write a cached page of an NFS file.
1357  *
1358  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1359  * things with a page scheduled for an RPC call (e.g. invalidate it).
1360  */
1361 int nfs_updatepage(struct file *file, struct page *page,
1362                 unsigned int offset, unsigned int count)
1363 {
1364         struct nfs_open_context *ctx = nfs_file_open_context(file);
1365         struct address_space *mapping = page_file_mapping(page);
1366         struct inode    *inode = mapping->host;
1367         unsigned int    pagelen = nfs_page_length(page);
1368         int             status = 0;
1369
1370         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1371
1372         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
1373                 file, count, (long long)(page_file_offset(page) + offset));
1374
1375         if (!count)
1376                 goto out;
1377
1378         if (nfs_can_extend_write(file, page, inode, pagelen)) {
1379                 count = max(count + offset, pagelen);
1380                 offset = 0;
1381         }
1382
1383         status = nfs_writepage_setup(ctx, page, offset, count);
1384         if (status < 0)
1385                 nfs_set_pageerror(mapping);
1386 out:
1387         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
1388                         status, (long long)i_size_read(inode));
1389         return status;
1390 }
1391
1392 static int flush_task_priority(int how)
1393 {
1394         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1395                 case FLUSH_HIGHPRI:
1396                         return RPC_PRIORITY_HIGH;
1397                 case FLUSH_LOWPRI:
1398                         return RPC_PRIORITY_LOW;
1399         }
1400         return RPC_PRIORITY_NORMAL;
1401 }
1402
1403 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1404                                struct rpc_message *msg,
1405                                const struct nfs_rpc_ops *rpc_ops,
1406                                struct rpc_task_setup *task_setup_data, int how)
1407 {
1408         int priority = flush_task_priority(how);
1409
1410         if (IS_SWAPFILE(hdr->inode))
1411                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1412         task_setup_data->priority = priority;
1413         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1414         trace_nfs_initiate_write(hdr);
1415 }
1416
1417 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1418  * call this on each, which will prepare them to be retried on next
1419  * writeback using standard nfs.
1420  */
1421 static void nfs_redirty_request(struct nfs_page *req)
1422 {
1423         /* Bump the transmission count */
1424         req->wb_nio++;
1425         nfs_mark_request_dirty(req);
1426         set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1427         nfs_end_page_writeback(req);
1428         nfs_release_request(req);
1429 }
1430
1431 static void nfs_async_write_error(struct list_head *head, int error)
1432 {
1433         struct nfs_page *req;
1434
1435         while (!list_empty(head)) {
1436                 req = nfs_list_entry(head->next);
1437                 nfs_list_remove_request(req);
1438                 if (nfs_error_is_fatal_on_server(error))
1439                         nfs_write_error(req, error);
1440                 else
1441                         nfs_redirty_request(req);
1442         }
1443 }
1444
1445 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1446 {
1447         nfs_async_write_error(&hdr->pages, 0);
1448 }
1449
1450 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1451         .init_hdr = nfs_async_write_init,
1452         .error_cleanup = nfs_async_write_error,
1453         .completion = nfs_write_completion,
1454         .reschedule_io = nfs_async_write_reschedule_io,
1455 };
1456
1457 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1458                                struct inode *inode, int ioflags, bool force_mds,
1459                                const struct nfs_pgio_completion_ops *compl_ops)
1460 {
1461         struct nfs_server *server = NFS_SERVER(inode);
1462         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1463
1464 #ifdef CONFIG_NFS_V4_1
1465         if (server->pnfs_curr_ld && !force_mds)
1466                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1467 #endif
1468         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1469                         server->wsize, ioflags);
1470 }
1471 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1472
1473 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1474 {
1475         struct nfs_pgio_mirror *mirror;
1476
1477         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1478                 pgio->pg_ops->pg_cleanup(pgio);
1479
1480         pgio->pg_ops = &nfs_pgio_rw_ops;
1481
1482         nfs_pageio_stop_mirroring(pgio);
1483
1484         mirror = &pgio->pg_mirrors[0];
1485         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1486 }
1487 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1488
1489
1490 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1491 {
1492         struct nfs_commit_data *data = calldata;
1493
1494         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1495 }
1496
1497 /*
1498  * Special version of should_remove_suid() that ignores capabilities.
1499  */
1500 static int nfs_should_remove_suid(const struct inode *inode)
1501 {
1502         umode_t mode = inode->i_mode;
1503         int kill = 0;
1504
1505         /* suid always must be killed */
1506         if (unlikely(mode & S_ISUID))
1507                 kill = ATTR_KILL_SUID;
1508
1509         /*
1510          * sgid without any exec bits is just a mandatory locking mark; leave
1511          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1512          */
1513         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1514                 kill |= ATTR_KILL_SGID;
1515
1516         if (unlikely(kill && S_ISREG(mode)))
1517                 return kill;
1518
1519         return 0;
1520 }
1521
1522 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1523                 struct nfs_fattr *fattr)
1524 {
1525         struct nfs_pgio_args *argp = &hdr->args;
1526         struct nfs_pgio_res *resp = &hdr->res;
1527         u64 size = argp->offset + resp->count;
1528
1529         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1530                 fattr->size = size;
1531         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1532                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1533                 return;
1534         }
1535         if (size != fattr->size)
1536                 return;
1537         /* Set attribute barrier */
1538         nfs_fattr_set_barrier(fattr);
1539         /* ...and update size */
1540         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1541 }
1542
1543 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1544 {
1545         struct nfs_fattr *fattr = &hdr->fattr;
1546         struct inode *inode = hdr->inode;
1547
1548         spin_lock(&inode->i_lock);
1549         nfs_writeback_check_extend(hdr, fattr);
1550         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1551         spin_unlock(&inode->i_lock);
1552 }
1553 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1554
1555 /*
1556  * This function is called when the WRITE call is complete.
1557  */
1558 static int nfs_writeback_done(struct rpc_task *task,
1559                               struct nfs_pgio_header *hdr,
1560                               struct inode *inode)
1561 {
1562         int status;
1563
1564         /*
1565          * ->write_done will attempt to use post-op attributes to detect
1566          * conflicting writes by other clients.  A strict interpretation
1567          * of close-to-open would allow us to continue caching even if
1568          * another writer had changed the file, but some applications
1569          * depend on tighter cache coherency when writing.
1570          */
1571         status = NFS_PROTO(inode)->write_done(task, hdr);
1572         if (status != 0)
1573                 return status;
1574
1575         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1576         trace_nfs_writeback_done(task, hdr);
1577
1578         if (task->tk_status >= 0) {
1579                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1580
1581                 if (committed == NFS_UNSTABLE) {
1582                         /*
1583                          * We have some uncommitted data on the server at
1584                          * this point, so ensure that we keep track of that
1585                          * fact irrespective of what later writes do.
1586                          */
1587                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1588                 }
1589
1590                 if (committed < hdr->args.stable) {
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                                         committed, hdr->args.stable);
1607                                 complain = jiffies + 300 * HZ;
1608                         }
1609                 }
1610         }
1611
1612         /* Deal with the suid/sgid bit corner case */
1613         if (nfs_should_remove_suid(inode)) {
1614                 spin_lock(&inode->i_lock);
1615                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1616                 spin_unlock(&inode->i_lock);
1617         }
1618         return 0;
1619 }
1620
1621 /*
1622  * This function is called when the WRITE call is complete.
1623  */
1624 static void nfs_writeback_result(struct rpc_task *task,
1625                                  struct nfs_pgio_header *hdr)
1626 {
1627         struct nfs_pgio_args    *argp = &hdr->args;
1628         struct nfs_pgio_res     *resp = &hdr->res;
1629
1630         if (resp->count < argp->count) {
1631                 static unsigned long    complain;
1632
1633                 /* This a short write! */
1634                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1635
1636                 /* Has the server at least made some progress? */
1637                 if (resp->count == 0) {
1638                         if (time_before(complain, jiffies)) {
1639                                 printk(KERN_WARNING
1640                                        "NFS: Server wrote zero bytes, expected %u.\n",
1641                                        argp->count);
1642                                 complain = jiffies + 300 * HZ;
1643                         }
1644                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1645                         task->tk_status = -EIO;
1646                         return;
1647                 }
1648
1649                 /* For non rpc-based layout drivers, retry-through-MDS */
1650                 if (!task->tk_ops) {
1651                         hdr->pnfs_error = -EAGAIN;
1652                         return;
1653                 }
1654
1655                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1656                 if (resp->verf->committed != NFS_UNSTABLE) {
1657                         /* Resend from where the server left off */
1658                         hdr->mds_offset += resp->count;
1659                         argp->offset += resp->count;
1660                         argp->pgbase += resp->count;
1661                         argp->count -= resp->count;
1662                 } else {
1663                         /* Resend as a stable write in order to avoid
1664                          * headaches in the case of a server crash.
1665                          */
1666                         argp->stable = NFS_FILE_SYNC;
1667                 }
1668                 resp->count = 0;
1669                 resp->verf->committed = 0;
1670                 rpc_restart_call_prepare(task);
1671         }
1672 }
1673
1674 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1675 {
1676         return wait_var_event_killable(&cinfo->rpcs_out,
1677                                        !atomic_read(&cinfo->rpcs_out));
1678 }
1679
1680 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1681 {
1682         atomic_inc(&cinfo->rpcs_out);
1683 }
1684
1685 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1686 {
1687         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1688                 wake_up_var(&cinfo->rpcs_out);
1689                 return true;
1690         }
1691         return false;
1692 }
1693
1694 void nfs_commitdata_release(struct nfs_commit_data *data)
1695 {
1696         put_nfs_open_context(data->context);
1697         nfs_commit_free(data);
1698 }
1699 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1700
1701 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1702                         const struct nfs_rpc_ops *nfs_ops,
1703                         const struct rpc_call_ops *call_ops,
1704                         int how, int flags)
1705 {
1706         struct rpc_task *task;
1707         int priority = flush_task_priority(how);
1708         struct rpc_message msg = {
1709                 .rpc_argp = &data->args,
1710                 .rpc_resp = &data->res,
1711                 .rpc_cred = data->cred,
1712         };
1713         struct rpc_task_setup task_setup_data = {
1714                 .task = &data->task,
1715                 .rpc_client = clnt,
1716                 .rpc_message = &msg,
1717                 .callback_ops = call_ops,
1718                 .callback_data = data,
1719                 .workqueue = nfsiod_workqueue,
1720                 .flags = RPC_TASK_ASYNC | flags,
1721                 .priority = priority,
1722         };
1723
1724         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1725                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1726
1727         /* Set up the initial task struct.  */
1728         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1729         trace_nfs_initiate_commit(data);
1730
1731         dprintk("NFS: initiated commit call\n");
1732
1733         task = rpc_run_task(&task_setup_data);
1734         if (IS_ERR(task))
1735                 return PTR_ERR(task);
1736         if (how & FLUSH_SYNC)
1737                 rpc_wait_for_completion_task(task);
1738         rpc_put_task(task);
1739         return 0;
1740 }
1741 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1742
1743 static loff_t nfs_get_lwb(struct list_head *head)
1744 {
1745         loff_t lwb = 0;
1746         struct nfs_page *req;
1747
1748         list_for_each_entry(req, head, wb_list)
1749                 if (lwb < (req_offset(req) + req->wb_bytes))
1750                         lwb = req_offset(req) + req->wb_bytes;
1751
1752         return lwb;
1753 }
1754
1755 /*
1756  * Set up the argument/result storage required for the RPC call.
1757  */
1758 void nfs_init_commit(struct nfs_commit_data *data,
1759                      struct list_head *head,
1760                      struct pnfs_layout_segment *lseg,
1761                      struct nfs_commit_info *cinfo)
1762 {
1763         struct nfs_page *first;
1764         struct nfs_open_context *ctx;
1765         struct inode *inode;
1766
1767         /* Set up the RPC argument and reply structs
1768          * NB: take care not to mess about with data->commit et al. */
1769
1770         if (head)
1771                 list_splice_init(head, &data->pages);
1772
1773         first = nfs_list_entry(data->pages.next);
1774         ctx = nfs_req_openctx(first);
1775         inode = d_inode(ctx->dentry);
1776
1777         data->inode       = inode;
1778         data->cred        = ctx->cred;
1779         data->lseg        = lseg; /* reference transferred */
1780         /* only set lwb for pnfs commit */
1781         if (lseg)
1782                 data->lwb = nfs_get_lwb(&data->pages);
1783         data->mds_ops     = &nfs_commit_ops;
1784         data->completion_ops = cinfo->completion_ops;
1785         data->dreq        = cinfo->dreq;
1786
1787         data->args.fh     = NFS_FH(data->inode);
1788         /* Note: we always request a commit of the entire inode */
1789         data->args.offset = 0;
1790         data->args.count  = 0;
1791         data->context     = get_nfs_open_context(ctx);
1792         data->res.fattr   = &data->fattr;
1793         data->res.verf    = &data->verf;
1794         nfs_fattr_init(&data->fattr);
1795         nfs_commit_begin(cinfo->mds);
1796 }
1797 EXPORT_SYMBOL_GPL(nfs_init_commit);
1798
1799 void nfs_retry_commit(struct list_head *page_list,
1800                       struct pnfs_layout_segment *lseg,
1801                       struct nfs_commit_info *cinfo,
1802                       u32 ds_commit_idx)
1803 {
1804         struct nfs_page *req;
1805
1806         while (!list_empty(page_list)) {
1807                 req = nfs_list_entry(page_list->next);
1808                 nfs_list_remove_request(req);
1809                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1810                 if (!cinfo->dreq)
1811                         nfs_clear_page_commit(req->wb_page);
1812                 nfs_unlock_and_release_request(req);
1813         }
1814 }
1815 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1816
1817 static void
1818 nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1819                 struct nfs_page *req)
1820 {
1821         __set_page_dirty_nobuffers(req->wb_page);
1822 }
1823
1824 /*
1825  * Commit dirty pages
1826  */
1827 static int
1828 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1829                 struct nfs_commit_info *cinfo)
1830 {
1831         struct nfs_commit_data  *data;
1832         unsigned short task_flags = 0;
1833
1834         /* another commit raced with us */
1835         if (list_empty(head))
1836                 return 0;
1837
1838         data = nfs_commitdata_alloc();
1839         if (!data) {
1840                 nfs_retry_commit(head, NULL, cinfo, -1);
1841                 return -ENOMEM;
1842         }
1843
1844         /* Set up the argument struct */
1845         nfs_init_commit(data, head, NULL, cinfo);
1846         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1847                 task_flags = RPC_TASK_MOVEABLE;
1848         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1849                                    data->mds_ops, how,
1850                                    RPC_TASK_CRED_NOREF | task_flags);
1851 }
1852
1853 /*
1854  * COMMIT call returned
1855  */
1856 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1857 {
1858         struct nfs_commit_data  *data = calldata;
1859
1860         /* Call the NFS version-specific code */
1861         NFS_PROTO(data->inode)->commit_done(task, data);
1862         trace_nfs_commit_done(task, data);
1863 }
1864
1865 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1866 {
1867         const struct nfs_writeverf *verf = data->res.verf;
1868         struct nfs_page *req;
1869         int status = data->task.tk_status;
1870         struct nfs_commit_info cinfo;
1871         struct nfs_server *nfss;
1872
1873         while (!list_empty(&data->pages)) {
1874                 req = nfs_list_entry(data->pages.next);
1875                 nfs_list_remove_request(req);
1876                 if (req->wb_page)
1877                         nfs_clear_page_commit(req->wb_page);
1878
1879                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1880                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1881                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1882                         req->wb_bytes,
1883                         (long long)req_offset(req));
1884                 if (status < 0) {
1885                         if (req->wb_page) {
1886                                 trace_nfs_commit_error(data->inode, req,
1887                                                        status);
1888                                 nfs_mapping_set_error(req->wb_page, status);
1889                                 nfs_inode_remove_request(req);
1890                         }
1891                         dprintk_cont(", error = %d\n", status);
1892                         goto next;
1893                 }
1894
1895                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1896                  * returned by the server against all stored verfs. */
1897                 if (nfs_write_match_verf(verf, req)) {
1898                         /* We have a match */
1899                         if (req->wb_page)
1900                                 nfs_inode_remove_request(req);
1901                         dprintk_cont(" OK\n");
1902                         goto next;
1903                 }
1904                 /* We have a mismatch. Write the page again */
1905                 dprintk_cont(" mismatch\n");
1906                 nfs_mark_request_dirty(req);
1907                 set_bit(NFS_CONTEXT_RESEND_WRITES, &nfs_req_openctx(req)->flags);
1908         next:
1909                 nfs_unlock_and_release_request(req);
1910                 /* Latency breaker */
1911                 cond_resched();
1912         }
1913         nfss = NFS_SERVER(data->inode);
1914         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1915                 nfss->write_congested = 0;
1916
1917         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1918         nfs_commit_end(cinfo.mds);
1919 }
1920
1921 static void nfs_commit_release(void *calldata)
1922 {
1923         struct nfs_commit_data *data = calldata;
1924
1925         data->completion_ops->completion(data);
1926         nfs_commitdata_release(calldata);
1927 }
1928
1929 static const struct rpc_call_ops nfs_commit_ops = {
1930         .rpc_call_prepare = nfs_commit_prepare,
1931         .rpc_call_done = nfs_commit_done,
1932         .rpc_release = nfs_commit_release,
1933 };
1934
1935 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1936         .completion = nfs_commit_release_pages,
1937         .resched_write = nfs_commit_resched_write,
1938 };
1939
1940 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1941                             int how, struct nfs_commit_info *cinfo)
1942 {
1943         int status;
1944
1945         status = pnfs_commit_list(inode, head, how, cinfo);
1946         if (status == PNFS_NOT_ATTEMPTED)
1947                 status = nfs_commit_list(inode, head, how, cinfo);
1948         return status;
1949 }
1950
1951 static int __nfs_commit_inode(struct inode *inode, int how,
1952                 struct writeback_control *wbc)
1953 {
1954         LIST_HEAD(head);
1955         struct nfs_commit_info cinfo;
1956         int may_wait = how & FLUSH_SYNC;
1957         int ret, nscan;
1958
1959         how &= ~FLUSH_SYNC;
1960         nfs_init_cinfo_from_inode(&cinfo, inode);
1961         nfs_commit_begin(cinfo.mds);
1962         for (;;) {
1963                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1964                 if (ret <= 0)
1965                         break;
1966                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1967                 if (ret < 0)
1968                         break;
1969                 ret = 0;
1970                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1971                         if (nscan < wbc->nr_to_write)
1972                                 wbc->nr_to_write -= nscan;
1973                         else
1974                                 wbc->nr_to_write = 0;
1975                 }
1976                 if (nscan < INT_MAX)
1977                         break;
1978                 cond_resched();
1979         }
1980         nfs_commit_end(cinfo.mds);
1981         if (ret || !may_wait)
1982                 return ret;
1983         return wait_on_commit(cinfo.mds);
1984 }
1985
1986 int nfs_commit_inode(struct inode *inode, int how)
1987 {
1988         return __nfs_commit_inode(inode, how, NULL);
1989 }
1990 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1991
1992 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1993 {
1994         struct nfs_inode *nfsi = NFS_I(inode);
1995         int flags = FLUSH_SYNC;
1996         int ret = 0;
1997
1998         if (wbc->sync_mode == WB_SYNC_NONE) {
1999                 /* no commits means nothing needs to be done */
2000                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
2001                         goto check_requests_outstanding;
2002
2003                 /* Don't commit yet if this is a non-blocking flush and there
2004                  * are a lot of outstanding writes for this mapping.
2005                  */
2006                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
2007                         goto out_mark_dirty;
2008
2009                 /* don't wait for the COMMIT response */
2010                 flags = 0;
2011         }
2012
2013         ret = __nfs_commit_inode(inode, flags, wbc);
2014         if (!ret) {
2015                 if (flags & FLUSH_SYNC)
2016                         return 0;
2017         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
2018                 goto out_mark_dirty;
2019
2020 check_requests_outstanding:
2021         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2022                 return ret;
2023 out_mark_dirty:
2024         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2025         return ret;
2026 }
2027 EXPORT_SYMBOL_GPL(nfs_write_inode);
2028
2029 /*
2030  * Wrapper for filemap_write_and_wait_range()
2031  *
2032  * Needed for pNFS in order to ensure data becomes visible to the
2033  * client.
2034  */
2035 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2036                 loff_t lstart, loff_t lend)
2037 {
2038         int ret;
2039
2040         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2041         if (ret == 0)
2042                 ret = pnfs_sync_inode(mapping->host, true);
2043         return ret;
2044 }
2045 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2046
2047 /*
2048  * flush the inode to disk.
2049  */
2050 int nfs_wb_all(struct inode *inode)
2051 {
2052         int ret;
2053
2054         trace_nfs_writeback_inode_enter(inode);
2055
2056         ret = filemap_write_and_wait(inode->i_mapping);
2057         if (ret)
2058                 goto out;
2059         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2060         if (ret < 0)
2061                 goto out;
2062         pnfs_sync_inode(inode, true);
2063         ret = 0;
2064
2065 out:
2066         trace_nfs_writeback_inode_exit(inode, ret);
2067         return ret;
2068 }
2069 EXPORT_SYMBOL_GPL(nfs_wb_all);
2070
2071 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2072 {
2073         struct nfs_page *req;
2074         int ret = 0;
2075
2076         folio_wait_writeback(folio);
2077
2078         /* blocking call to cancel all requests and join to a single (head)
2079          * request */
2080         req = nfs_lock_and_join_requests(&folio->page);
2081
2082         if (IS_ERR(req)) {
2083                 ret = PTR_ERR(req);
2084         } else if (req) {
2085                 /* all requests from this folio have been cancelled by
2086                  * nfs_lock_and_join_requests, so just remove the head
2087                  * request from the inode / page_private pointer and
2088                  * release it */
2089                 nfs_inode_remove_request(req);
2090                 nfs_unlock_and_release_request(req);
2091         }
2092
2093         return ret;
2094 }
2095
2096 /*
2097  * Write back all requests on one page - we do this before reading it.
2098  */
2099 int nfs_wb_page(struct inode *inode, struct page *page)
2100 {
2101         loff_t range_start = page_file_offset(page);
2102         loff_t range_end = range_start + (loff_t)(PAGE_SIZE - 1);
2103         struct writeback_control wbc = {
2104                 .sync_mode = WB_SYNC_ALL,
2105                 .nr_to_write = 0,
2106                 .range_start = range_start,
2107                 .range_end = range_end,
2108         };
2109         int ret;
2110
2111         trace_nfs_writeback_page_enter(inode);
2112
2113         for (;;) {
2114                 wait_on_page_writeback(page);
2115                 if (clear_page_dirty_for_io(page)) {
2116                         ret = nfs_writepage_locked(page, &wbc);
2117                         if (ret < 0)
2118                                 goto out_error;
2119                         continue;
2120                 }
2121                 ret = 0;
2122                 if (!PagePrivate(page))
2123                         break;
2124                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2125                 if (ret < 0)
2126                         goto out_error;
2127         }
2128 out_error:
2129         trace_nfs_writeback_page_exit(inode, ret);
2130         return ret;
2131 }
2132
2133 #ifdef CONFIG_MIGRATION
2134 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2135                 struct folio *src, enum migrate_mode mode)
2136 {
2137         /*
2138          * If the private flag is set, the folio is currently associated with
2139          * an in-progress read or write request. Don't try to migrate it.
2140          *
2141          * FIXME: we could do this in principle, but we'll need a way to ensure
2142          *        that we can safely release the inode reference while holding
2143          *        the folio lock.
2144          */
2145         if (folio_test_private(src))
2146                 return -EBUSY;
2147
2148         if (folio_test_fscache(src)) {
2149                 if (mode == MIGRATE_ASYNC)
2150                         return -EBUSY;
2151                 folio_wait_fscache(src);
2152         }
2153
2154         return migrate_folio(mapping, dst, src, mode);
2155 }
2156 #endif
2157
2158 int __init nfs_init_writepagecache(void)
2159 {
2160         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2161                                              sizeof(struct nfs_pgio_header),
2162                                              0, SLAB_HWCACHE_ALIGN,
2163                                              NULL);
2164         if (nfs_wdata_cachep == NULL)
2165                 return -ENOMEM;
2166
2167         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2168                                                      nfs_wdata_cachep);
2169         if (nfs_wdata_mempool == NULL)
2170                 goto out_destroy_write_cache;
2171
2172         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2173                                              sizeof(struct nfs_commit_data),
2174                                              0, SLAB_HWCACHE_ALIGN,
2175                                              NULL);
2176         if (nfs_cdata_cachep == NULL)
2177                 goto out_destroy_write_mempool;
2178
2179         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2180                                                       nfs_cdata_cachep);
2181         if (nfs_commit_mempool == NULL)
2182                 goto out_destroy_commit_cache;
2183
2184         /*
2185          * NFS congestion size, scale with available memory.
2186          *
2187          *  64MB:    8192k
2188          * 128MB:   11585k
2189          * 256MB:   16384k
2190          * 512MB:   23170k
2191          *   1GB:   32768k
2192          *   2GB:   46340k
2193          *   4GB:   65536k
2194          *   8GB:   92681k
2195          *  16GB:  131072k
2196          *
2197          * This allows larger machines to have larger/more transfers.
2198          * Limit the default to 256M
2199          */
2200         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2201         if (nfs_congestion_kb > 256*1024)
2202                 nfs_congestion_kb = 256*1024;
2203
2204         return 0;
2205
2206 out_destroy_commit_cache:
2207         kmem_cache_destroy(nfs_cdata_cachep);
2208 out_destroy_write_mempool:
2209         mempool_destroy(nfs_wdata_mempool);
2210 out_destroy_write_cache:
2211         kmem_cache_destroy(nfs_wdata_cachep);
2212         return -ENOMEM;
2213 }
2214
2215 void nfs_destroy_writepagecache(void)
2216 {
2217         mempool_destroy(nfs_commit_mempool);
2218         kmem_cache_destroy(nfs_cdata_cachep);
2219         mempool_destroy(nfs_wdata_mempool);
2220         kmem_cache_destroy(nfs_wdata_cachep);
2221 }
2222
2223 static const struct nfs_rw_ops nfs_rw_write_ops = {
2224         .rw_alloc_header        = nfs_writehdr_alloc,
2225         .rw_free_header         = nfs_writehdr_free,
2226         .rw_done                = nfs_writeback_done,
2227         .rw_result              = nfs_writeback_result,
2228         .rw_initiate            = nfs_initiate_write,
2229 };