fs/nfs/write.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/write.c
   4  *
   5  * Write file data over NFS.
   6  *
   7  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/mm.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/file.h>
  15 #include <linux/writeback.h>
  16 #include <linux/swap.h>
  17 #include <linux/migrate.h>
  18
  19 #include <linux/sunrpc/clnt.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/export.h>
  25 #include <linux/freezer.h>
  26 #include <linux/wait.h>
  27 #include <linux/iversion.h>
  28 #include <linux/filelock.h>
  29
  30 #include <linux/uaccess.h>
  31 #include <linux/sched/mm.h>
  32
  33 #include "delegation.h"
  34 #include "internal.h"
  35 #include "iostat.h"
  36 #include "nfs4_fs.h"
  37 #include "fscache.h"
  38 #include "pnfs.h"
  39
  40 #include "nfstrace.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  43
  44 #define MIN_POOL_WRITE          (32)
  45 #define MIN_POOL_COMMIT         (4)
  46
  47 struct nfs_io_completion {
  48         void (*complete)(void *data);
  49         void *data;
  50         struct kref refcount;
  51 };
  52
  53 /*
  54  * Local function declarations
  55  */
  56 static void nfs_redirty_request(struct nfs_page *req);
  57 static const struct rpc_call_ops nfs_commit_ops;
  58 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  59 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  60 static const struct nfs_rw_ops nfs_rw_write_ops;
  61 static void nfs_inode_remove_request(struct nfs_page *req);
  62 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
  63                                      struct nfs_page *req);
  64 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  65                                       struct inode *inode);
  66 static struct nfs_page *
  67 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  68                                                 struct folio *folio);
  69
  70 static struct kmem_cache *nfs_wdata_cachep;
  71 static mempool_t *nfs_wdata_mempool;
  72 static struct kmem_cache *nfs_cdata_cachep;
  73 static mempool_t *nfs_commit_mempool;
  74
  75 struct nfs_commit_data *nfs_commitdata_alloc(void)
  76 {
  77         struct nfs_commit_data *p;
  78
  79         p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  80         if (!p) {
  81                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  82                 if (!p)
  83                         return NULL;
  84                 memset(p, 0, sizeof(*p));
  85         }
  86         INIT_LIST_HEAD(&p->pages);
  87         return p;
  88 }
  89 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  90
  91 void nfs_commit_free(struct nfs_commit_data *p)
  92 {
  93         mempool_free(p, nfs_commit_mempool);
  94 }
  95 EXPORT_SYMBOL_GPL(nfs_commit_free);
  96
  97 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  98 {
  99         struct nfs_pgio_header *p;
 100
 101         p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
 102         if (!p) {
 103                 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 104                 if (!p)
 105                         return NULL;
 106                 memset(p, 0, sizeof(*p));
 107         }
 108         p->rw_mode = FMODE_WRITE;
 109         return p;
 110 }
 111
 112 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 113 {
 114         mempool_free(hdr, nfs_wdata_mempool);
 115 }
 116
 117 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 118 {
 119         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 120 }
 121
 122 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 123                 void (*complete)(void *), void *data)
 124 {
 125         ioc->complete = complete;
 126         ioc->data = data;
 127         kref_init(&ioc->refcount);
 128 }
 129
 130 static void nfs_io_completion_release(struct kref *kref)
 131 {
 132         struct nfs_io_completion *ioc = container_of(kref,
 133                         struct nfs_io_completion, refcount);
 134         ioc->complete(ioc->data);
 135         kfree(ioc);
 136 }
 137
 138 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 139 {
 140         if (ioc != NULL)
 141                 kref_get(&ioc->refcount);
 142 }
 143
 144 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 145 {
 146         if (ioc != NULL)
 147                 kref_put(&ioc->refcount, nfs_io_completion_release);
 148 }
 149
 150 static void
 151 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 152 {
 153         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 154                 kref_get(&req->wb_kref);
 155                 atomic_long_inc(&NFS_I(inode)->nrequests);
 156         }
 157 }
 158
 159 static int
 160 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 161 {
 162         int ret;
 163
 164         if (!test_bit(PG_REMOVE, &req->wb_flags))
 165                 return 0;
 166         ret = nfs_page_group_lock(req);
 167         if (ret)
 168                 return ret;
 169         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 170                 nfs_page_set_inode_ref(req, inode);
 171         nfs_page_group_unlock(req);
 172         return 0;
 173 }
 174
 175 static struct nfs_page *nfs_folio_private_request(struct folio *folio)
 176 {
 177         return folio_get_private(folio);
 178 }
 179
 180 /**
 181  * nfs_folio_find_private_request - find head request associated with a folio
 182  * @folio: pointer to folio
 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 *nfs_folio_find_private_request(struct folio *folio)
 189 {
 190         struct address_space *mapping = folio_file_mapping(folio);
 191         struct nfs_page *req;
 192
 193         if (!folio_test_private(folio))
 194                 return NULL;
 195         spin_lock(&mapping->private_lock);
 196         req = nfs_folio_private_request(folio);
 197         if (req) {
 198                 WARN_ON_ONCE(req->wb_head != req);
 199                 kref_get(&req->wb_kref);
 200         }
 201         spin_unlock(&mapping->private_lock);
 202         return req;
 203 }
 204
 205 static struct nfs_page *nfs_folio_find_swap_request(struct folio *folio)
 206 {
 207         struct inode *inode = folio_file_mapping(folio)->host;
 208         struct nfs_inode *nfsi = NFS_I(inode);
 209         struct nfs_page *req = NULL;
 210         if (!folio_test_swapcache(folio))
 211                 return NULL;
 212         mutex_lock(&nfsi->commit_mutex);
 213         if (folio_test_swapcache(folio)) {
 214                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 215                                                                       folio);
 216                 if (req) {
 217                         WARN_ON_ONCE(req->wb_head != req);
 218                         kref_get(&req->wb_kref);
 219                 }
 220         }
 221         mutex_unlock(&nfsi->commit_mutex);
 222         return req;
 223 }
 224
 225 /**
 226  * nfs_folio_find_head_request - find head request associated with a folio
 227  * @folio: pointer to folio
 228  *
 229  * returns matching head request with reference held, or NULL if not found.
 230  */
 231 static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 232 {
 233         struct nfs_page *req;
 234
 235         req = nfs_folio_find_private_request(folio);
 236         if (!req)
 237                 req = nfs_folio_find_swap_request(folio);
 238         return req;
 239 }
 240
 241 static struct nfs_page *nfs_folio_find_and_lock_request(struct folio *folio)
 242 {
 243         struct inode *inode = folio_file_mapping(folio)->host;
 244         struct nfs_page *req, *head;
 245         int ret;
 246
 247         for (;;) {
 248                 req = nfs_folio_find_head_request(folio);
 249                 if (!req)
 250                         return req;
 251                 head = nfs_page_group_lock_head(req);
 252                 if (head != req)
 253                         nfs_release_request(req);
 254                 if (IS_ERR(head))
 255                         return head;
 256                 ret = nfs_cancel_remove_inode(head, inode);
 257                 if (ret < 0) {
 258                         nfs_unlock_and_release_request(head);
 259                         return ERR_PTR(ret);
 260                 }
 261                 /* Ensure that nobody removed the request before we locked it */
 262                 if (head == nfs_folio_private_request(folio))
 263                         break;
 264                 if (folio_test_swapcache(folio))
 265                         break;
 266                 nfs_unlock_and_release_request(head);
 267         }
 268         return head;
 269 }
 270
 271 /* Adjust the file length if we're writing beyond the end */
 272 static void nfs_grow_file(struct folio *folio, unsigned int offset,
 273                           unsigned int count)
 274 {
 275         struct inode *inode = folio_file_mapping(folio)->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) >> folio_shift(folio)) << folio_order(folio);
 282         if (i_size > 0 && folio_index(folio) < end_index)
 283                 goto out;
 284         end = folio_file_pos(folio) + (loff_t)offset + (loff_t)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 folio *folio, int error)
 311 {
 312         struct address_space *mapping = folio_file_mapping(folio);
 313
 314         folio_set_error(folio);
 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         unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
 362         struct nfs_page *tmp;
 363         unsigned int pos = 0;
 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         struct folio *folio = nfs_page_to_folio(req);
 384
 385         if (folio_test_uptodate(folio))
 386                 return;
 387         if (!nfs_page_group_covers_page(req))
 388                 return;
 389         folio_mark_uptodate(folio);
 390 }
 391
 392 static int wb_priority(struct writeback_control *wbc)
 393 {
 394         int ret = 0;
 395
 396         if (wbc->sync_mode == WB_SYNC_ALL)
 397                 ret = FLUSH_COND_STABLE;
 398         return ret;
 399 }
 400
 401 /*
 402  * NFS congestion control
 403  */
 404
 405 int nfs_congestion_kb;
 406
 407 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 408 #define NFS_CONGESTION_OFF_THRESH       \
 409         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 410
 411 static void nfs_folio_set_writeback(struct folio *folio)
 412 {
 413         struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
 414
 415         folio_start_writeback(folio);
 416         if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
 417                 nfss->write_congested = 1;
 418 }
 419
 420 static void nfs_folio_end_writeback(struct folio *folio)
 421 {
 422         struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
 423
 424         folio_end_writeback(folio);
 425         if (atomic_long_dec_return(&nfss->writeback) <
 426             NFS_CONGESTION_OFF_THRESH)
 427                 nfss->write_congested = 0;
 428 }
 429
 430 static void nfs_page_end_writeback(struct nfs_page *req)
 431 {
 432         if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
 433                 nfs_unlock_request(req);
 434                 nfs_folio_end_writeback(nfs_page_to_folio(req));
 435         } else
 436                 nfs_unlock_request(req);
 437 }
 438
 439 /*
 440  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 441  *
 442  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 443  * @old_head - the old head of the list
 444  *
 445  * All subrequests must be locked and removed from all lists, so at this point
 446  * they are only "active" in this function, and possibly in nfs_wait_on_request
 447  * with a reference held by some other context.
 448  */
 449 static void
 450 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 451                                  struct nfs_page *old_head,
 452                                  struct inode *inode)
 453 {
 454         while (destroy_list) {
 455                 struct nfs_page *subreq = destroy_list;
 456
 457                 destroy_list = (subreq->wb_this_page == old_head) ?
 458                                    NULL : subreq->wb_this_page;
 459
 460                 /* Note: lock subreq in order to change subreq->wb_head */
 461                 nfs_page_set_headlock(subreq);
 462                 WARN_ON_ONCE(old_head != subreq->wb_head);
 463
 464                 /* make sure old group is not used */
 465                 subreq->wb_this_page = subreq;
 466                 subreq->wb_head = subreq;
 467
 468                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 469
 470                 /* Note: races with nfs_page_group_destroy() */
 471                 if (!kref_read(&subreq->wb_kref)) {
 472                         /* Check if we raced with nfs_page_group_destroy() */
 473                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 474                                 nfs_page_clear_headlock(subreq);
 475                                 nfs_free_request(subreq);
 476                         } else
 477                                 nfs_page_clear_headlock(subreq);
 478                         continue;
 479                 }
 480                 nfs_page_clear_headlock(subreq);
 481
 482                 nfs_release_request(old_head);
 483
 484                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 485                         nfs_release_request(subreq);
 486                         atomic_long_dec(&NFS_I(inode)->nrequests);
 487                 }
 488
 489                 /* subreq is now totally disconnected from page group or any
 490                  * write / commit lists. last chance to wake any waiters */
 491                 nfs_unlock_and_release_request(subreq);
 492         }
 493 }
 494
 495 /*
 496  * nfs_join_page_group - destroy subrequests of the head req
 497  * @head: the page used to lookup the "page group" of nfs_page structures
 498  * @inode: Inode to which the request belongs.
 499  *
 500  * This function joins all sub requests to the head request by first
 501  * locking all requests in the group, cancelling any pending operations
 502  * and finally updating the head request to cover the whole range covered by
 503  * the (former) group.  All subrequests are removed from any write or commit
 504  * lists, unlinked from the group and destroyed.
 505  */
 506 void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
 507                          struct inode *inode)
 508 {
 509         struct nfs_page *subreq;
 510         struct nfs_page *destroy_list = NULL;
 511         unsigned int pgbase, off, bytes;
 512
 513         pgbase = head->wb_pgbase;
 514         bytes = head->wb_bytes;
 515         off = head->wb_offset;
 516         for (subreq = head->wb_this_page; subreq != head;
 517                         subreq = subreq->wb_this_page) {
 518                 /* Subrequests should always form a contiguous range */
 519                 if (pgbase > subreq->wb_pgbase) {
 520                         off -= pgbase - subreq->wb_pgbase;
 521                         bytes += pgbase - subreq->wb_pgbase;
 522                         pgbase = subreq->wb_pgbase;
 523                 }
 524                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 525                                 - pgbase, bytes);
 526         }
 527
 528         /* Set the head request's range to cover the former page group */
 529         head->wb_pgbase = pgbase;
 530         head->wb_bytes = bytes;
 531         head->wb_offset = off;
 532
 533         /* Now that all requests are locked, make sure they aren't on any list.
 534          * Commit list removal accounting is done after locks are dropped */
 535         subreq = head;
 536         do {
 537                 nfs_clear_request_commit(cinfo, subreq);
 538                 subreq = subreq->wb_this_page;
 539         } while (subreq != head);
 540
 541         /* unlink subrequests from head, destroy them later */
 542         if (head->wb_this_page != head) {
 543                 /* destroy list will be terminated by head */
 544                 destroy_list = head->wb_this_page;
 545                 head->wb_this_page = head;
 546         }
 547
 548         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 549 }
 550
 551 /*
 552  * nfs_lock_and_join_requests - join all subreqs to the head req
 553  * @folio: the folio used to lookup the "page group" of nfs_page structures
 554  *
 555  * This function joins all sub requests to the head request by first
 556  * locking all requests in the group, cancelling any pending operations
 557  * and finally updating the head request to cover the whole range covered by
 558  * the (former) group.  All subrequests are removed from any write or commit
 559  * lists, unlinked from the group and destroyed.
 560  *
 561  * Returns a locked, referenced pointer to the head request - which after
 562  * this call is guaranteed to be the only request associated with the page.
 563  * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
 564  * error was encountered.
 565  */
 566 static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 567 {
 568         struct inode *inode = folio_file_mapping(folio)->host;
 569         struct nfs_page *head;
 570         struct nfs_commit_info cinfo;
 571         int ret;
 572
 573         nfs_init_cinfo_from_inode(&cinfo, inode);
 574         /*
 575          * A reference is taken only on the head request which acts as a
 576          * reference to the whole page group - the group will not be destroyed
 577          * until the head reference is released.
 578          */
 579         head = nfs_folio_find_and_lock_request(folio);
 580         if (IS_ERR_OR_NULL(head))
 581                 return head;
 582
 583         /* lock each request in the page group */
 584         ret = nfs_page_group_lock_subrequests(head);
 585         if (ret < 0) {
 586                 nfs_unlock_and_release_request(head);
 587                 return ERR_PTR(ret);
 588         }
 589
 590         nfs_join_page_group(head, &cinfo, inode);
 591
 592         return head;
 593 }
 594
 595 static void nfs_write_error(struct nfs_page *req, int error)
 596 {
 597         trace_nfs_write_error(nfs_page_to_inode(req), req, error);
 598         nfs_mapping_set_error(nfs_page_to_folio(req), error);
 599         nfs_inode_remove_request(req);
 600         nfs_page_end_writeback(req);
 601         nfs_release_request(req);
 602 }
 603
 604 /*
 605  * Find an associated nfs write request, and prepare to flush it out
 606  * May return an error if the user signalled nfs_wait_on_request().
 607  */
 608 static int nfs_page_async_flush(struct folio *folio,
 609                                 struct writeback_control *wbc,
 610                                 struct nfs_pageio_descriptor *pgio)
 611 {
 612         struct nfs_page *req;
 613         int ret = 0;
 614
 615         req = nfs_lock_and_join_requests(folio);
 616         if (!req)
 617                 goto out;
 618         ret = PTR_ERR(req);
 619         if (IS_ERR(req))
 620                 goto out;
 621
 622         nfs_folio_set_writeback(folio);
 623         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 624
 625         /* If there is a fatal error that covers this write, just exit */
 626         ret = pgio->pg_error;
 627         if (nfs_error_is_fatal_on_server(ret))
 628                 goto out_launder;
 629
 630         ret = 0;
 631         if (!nfs_pageio_add_request(pgio, req)) {
 632                 ret = pgio->pg_error;
 633                 /*
 634                  * Remove the problematic req upon fatal errors on the server
 635                  */
 636                 if (nfs_error_is_fatal_on_server(ret))
 637                         goto out_launder;
 638                 if (wbc->sync_mode == WB_SYNC_NONE)
 639                         ret = AOP_WRITEPAGE_ACTIVATE;
 640                 folio_redirty_for_writepage(wbc, folio);
 641                 nfs_redirty_request(req);
 642                 pgio->pg_error = 0;
 643         } else
 644                 nfs_add_stats(folio_file_mapping(folio)->host,
 645                               NFSIOS_WRITEPAGES, 1);
 646 out:
 647         return ret;
 648 out_launder:
 649         nfs_write_error(req, ret);
 650         return 0;
 651 }
 652
 653 static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
 654                             struct nfs_pageio_descriptor *pgio)
 655 {
 656         nfs_pageio_cond_complete(pgio, folio_index(folio));
 657         return nfs_page_async_flush(folio, wbc, pgio);
 658 }
 659
 660 /*
 661  * Write an mmapped page to the server.
 662  */
 663 static int nfs_writepage_locked(struct folio *folio,
 664                                 struct writeback_control *wbc)
 665 {
 666         struct nfs_pageio_descriptor pgio;
 667         struct inode *inode = folio_file_mapping(folio)->host;
 668         int err;
 669
 670         if (wbc->sync_mode == WB_SYNC_NONE &&
 671             NFS_SERVER(inode)->write_congested)
 672                 return AOP_WRITEPAGE_ACTIVATE;
 673
 674         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 675         nfs_pageio_init_write(&pgio, inode, 0, false,
 676                               &nfs_async_write_completion_ops);
 677         err = nfs_do_writepage(folio, wbc, &pgio);
 678         pgio.pg_error = 0;
 679         nfs_pageio_complete(&pgio);
 680         return err;
 681 }
 682
 683 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 684 {
 685         struct folio *folio = page_folio(page);
 686         int ret;
 687
 688         ret = nfs_writepage_locked(folio, wbc);
 689         if (ret != AOP_WRITEPAGE_ACTIVATE)
 690                 unlock_page(page);
 691         return ret;
 692 }
 693
 694 static int nfs_writepages_callback(struct folio *folio,
 695                                    struct writeback_control *wbc, void *data)
 696 {
 697         int ret;
 698
 699         ret = nfs_do_writepage(folio, wbc, data);
 700         if (ret != AOP_WRITEPAGE_ACTIVATE)
 701                 folio_unlock(folio);
 702         return ret;
 703 }
 704
 705 static void nfs_io_completion_commit(void *inode)
 706 {
 707         nfs_commit_inode(inode, 0);
 708 }
 709
 710 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 711 {
 712         struct inode *inode = mapping->host;
 713         struct nfs_pageio_descriptor pgio;
 714         struct nfs_io_completion *ioc = NULL;
 715         unsigned int mntflags = NFS_SERVER(inode)->flags;
 716         int priority = 0;
 717         int err;
 718
 719         if (wbc->sync_mode == WB_SYNC_NONE &&
 720             NFS_SERVER(inode)->write_congested)
 721                 return 0;
 722
 723         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 724
 725         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 726             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 727                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 728                 if (ioc)
 729                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 730                                                inode);
 731                 priority = wb_priority(wbc);
 732         }
 733
 734         do {
 735                 nfs_pageio_init_write(&pgio, inode, priority, false,
 736                                       &nfs_async_write_completion_ops);
 737                 pgio.pg_io_completion = ioc;
 738                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 739                                         &pgio);
 740                 pgio.pg_error = 0;
 741                 nfs_pageio_complete(&pgio);
 742         } while (err < 0 && !nfs_error_is_fatal(err));
 743         nfs_io_completion_put(ioc);
 744
 745         if (err < 0)
 746                 goto out_err;
 747         return 0;
 748 out_err:
 749         return err;
 750 }
 751
 752 /*
 753  * Insert a write request into an inode
 754  */
 755 static void nfs_inode_add_request(struct nfs_page *req)
 756 {
 757         struct folio *folio = nfs_page_to_folio(req);
 758         struct address_space *mapping = folio_file_mapping(folio);
 759         struct nfs_inode *nfsi = NFS_I(mapping->host);
 760
 761         WARN_ON_ONCE(req->wb_this_page != req);
 762
 763         /* Lock the request! */
 764         nfs_lock_request(req);
 765
 766         /*
 767          * Swap-space should not get truncated. Hence no need to plug the race
 768          * with invalidate/truncate.
 769          */
 770         spin_lock(&mapping->private_lock);
 771         if (likely(!folio_test_swapcache(folio))) {
 772                 set_bit(PG_MAPPED, &req->wb_flags);
 773                 folio_set_private(folio);
 774                 folio->private = req;
 775         }
 776         spin_unlock(&mapping->private_lock);
 777         atomic_long_inc(&nfsi->nrequests);
 778         /* this a head request for a page group - mark it as having an
 779          * extra reference so sub groups can follow suit.
 780          * This flag also informs pgio layer when to bump nrequests when
 781          * adding subrequests. */
 782         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 783         kref_get(&req->wb_kref);
 784 }
 785
 786 /*
 787  * Remove a write request from an inode
 788  */
 789 static void nfs_inode_remove_request(struct nfs_page *req)
 790 {
 791         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 792                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 793                 struct address_space *mapping = folio_file_mapping(folio);
 794
 795                 spin_lock(&mapping->private_lock);
 796                 if (likely(folio && !folio_test_swapcache(folio))) {
 797                         folio->private = NULL;
 798                         folio_clear_private(folio);
 799                         clear_bit(PG_MAPPED, &req->wb_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                 atomic_long_dec(&NFS_I(nfs_page_to_inode(req))->nrequests);
 806                 nfs_release_request(req);
 807         }
 808 }
 809
 810 static void nfs_mark_request_dirty(struct nfs_page *req)
 811 {
 812         struct folio *folio = nfs_page_to_folio(req);
 813         if (folio)
 814                 filemap_dirty_folio(folio_mapping(folio), folio);
 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 @folio.
 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 folio *folio)
 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, folio);
 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 (nfs_page_to_folio(freq) == folio)
 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         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 891 }
 892 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 893
 894 /**
 895  * nfs_request_remove_commit_list - Remove request from a commit list
 896  * @req: pointer to a nfs_page
 897  * @cinfo: holds list lock and accounting info
 898  *
 899  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 900  * number of outstanding requests requiring a commit
 901  * It does not update the MM page stats.
 902  *
 903  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 904  */
 905 void
 906 nfs_request_remove_commit_list(struct nfs_page *req,
 907                                struct nfs_commit_info *cinfo)
 908 {
 909         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 910                 return;
 911         nfs_list_remove_request(req);
 912         atomic_long_dec(&cinfo->mds->ncommit);
 913 }
 914 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 915
 916 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 917                                       struct inode *inode)
 918 {
 919         cinfo->inode = inode;
 920         cinfo->mds = &NFS_I(inode)->commit_info;
 921         cinfo->ds = pnfs_get_ds_info(inode);
 922         cinfo->dreq = NULL;
 923         cinfo->completion_ops = &nfs_commit_completion_ops;
 924 }
 925
 926 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 927                     struct inode *inode,
 928                     struct nfs_direct_req *dreq)
 929 {
 930         if (dreq)
 931                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 932         else
 933                 nfs_init_cinfo_from_inode(cinfo, inode);
 934 }
 935 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 936
 937 /*
 938  * Add a request to the inode's commit list.
 939  */
 940 void
 941 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 942                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 943 {
 944         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 945                 return;
 946         nfs_request_add_commit_list(req, cinfo);
 947 }
 948
 949 static void nfs_folio_clear_commit(struct folio *folio)
 950 {
 951         if (folio) {
 952                 long nr = folio_nr_pages(folio);
 953
 954                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 955                 wb_stat_mod(&inode_to_bdi(folio_file_mapping(folio)->host)->wb,
 956                             WB_WRITEBACK, -nr);
 957         }
 958 }
 959
 960 /* Called holding the request lock on @req */
 961 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 962                                      struct nfs_page *req)
 963 {
 964         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 965                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 966                 struct inode *inode = d_inode(ctx->dentry);
 967
 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_folio_clear_commit(nfs_page_to_folio(req));
 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(nfs_page_to_folio(req),
1006                                               hdr->error);
1007                         goto remove_req;
1008                 }
1009                 if (nfs_write_need_commit(hdr)) {
1010                         /* Reset wb_nio, since the write was successful. */
1011                         req->wb_nio = 0;
1012                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1013                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1014                                 hdr->pgio_mirror_idx);
1015                         goto next;
1016                 }
1017 remove_req:
1018                 nfs_inode_remove_request(req);
1019 next:
1020                 nfs_page_end_writeback(req);
1021                 nfs_release_request(req);
1022         }
1023 out:
1024         nfs_io_completion_put(hdr->io_completion);
1025         hdr->release(hdr);
1026 }
1027
1028 unsigned long
1029 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1030 {
1031         return atomic_long_read(&cinfo->mds->ncommit);
1032 }
1033
1034 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1035 int
1036 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1037                      struct nfs_commit_info *cinfo, int max)
1038 {
1039         struct nfs_page *req, *tmp;
1040         int ret = 0;
1041
1042         list_for_each_entry_safe(req, tmp, src, wb_list) {
1043                 kref_get(&req->wb_kref);
1044                 if (!nfs_lock_request(req)) {
1045                         nfs_release_request(req);
1046                         continue;
1047                 }
1048                 nfs_request_remove_commit_list(req, cinfo);
1049                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1050                 nfs_list_add_request(req, dst);
1051                 ret++;
1052                 if ((ret == max) && !cinfo->dreq)
1053                         break;
1054                 cond_resched();
1055         }
1056         return ret;
1057 }
1058 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1059
1060 /*
1061  * nfs_scan_commit - Scan an inode for commit requests
1062  * @inode: NFS inode to scan
1063  * @dst: mds destination list
1064  * @cinfo: mds and ds lists of reqs ready to commit
1065  *
1066  * Moves requests from the inode's 'commit' request list.
1067  * The requests are *not* checked to ensure that they form a contiguous set.
1068  */
1069 int
1070 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1071                 struct nfs_commit_info *cinfo)
1072 {
1073         int ret = 0;
1074
1075         if (!atomic_long_read(&cinfo->mds->ncommit))
1076                 return 0;
1077         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1078         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1079                 const int max = INT_MAX;
1080
1081                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1082                                            cinfo, max);
1083                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1084         }
1085         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1086         return ret;
1087 }
1088
1089 /*
1090  * Search for an existing write request, and attempt to update
1091  * it to reflect a new dirty region on a given page.
1092  *
1093  * If the attempt fails, then the existing request is flushed out
1094  * to disk.
1095  */
1096 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
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(folio);
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_folio(folio_file_mapping(folio)->host, folio);
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 folio *folio,
1153                                                 unsigned int offset,
1154                                                 unsigned int bytes)
1155 {
1156         struct nfs_page *req;
1157
1158         req = nfs_try_to_update_request(folio, offset, bytes);
1159         if (req != NULL)
1160                 goto out;
1161         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1162         if (IS_ERR(req))
1163                 goto out;
1164         nfs_inode_add_request(req);
1165 out:
1166         return req;
1167 }
1168
1169 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1170                                struct folio *folio, unsigned int offset,
1171                                unsigned int count)
1172 {
1173         struct nfs_page *req;
1174
1175         req = nfs_setup_write_request(ctx, folio, offset, count);
1176         if (IS_ERR(req))
1177                 return PTR_ERR(req);
1178         /* Update file length */
1179         nfs_grow_file(folio, offset, count);
1180         nfs_mark_uptodate(req);
1181         nfs_mark_request_dirty(req);
1182         nfs_unlock_and_release_request(req);
1183         return 0;
1184 }
1185
1186 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1187 {
1188         struct nfs_open_context *ctx = nfs_file_open_context(file);
1189         struct nfs_lock_context *l_ctx;
1190         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1191         struct nfs_page *req;
1192         int do_flush, status;
1193         /*
1194          * Look for a request corresponding to this page. If there
1195          * is one, and it belongs to another file, we flush it out
1196          * before we try to copy anything into the page. Do this
1197          * due to the lack of an ACCESS-type call in NFSv2.
1198          * Also do the same if we find a request from an existing
1199          * dropped page.
1200          */
1201         do {
1202                 req = nfs_folio_find_head_request(folio);
1203                 if (req == NULL)
1204                         return 0;
1205                 l_ctx = req->wb_lock_context;
1206                 do_flush = nfs_page_to_folio(req) != folio ||
1207                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1208                 if (l_ctx && flctx &&
1209                     !(list_empty_careful(&flctx->flc_posix) &&
1210                       list_empty_careful(&flctx->flc_flock))) {
1211                         do_flush |= l_ctx->lockowner != current->files;
1212                 }
1213                 nfs_release_request(req);
1214                 if (!do_flush)
1215                         return 0;
1216                 status = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1217         } while (status == 0);
1218         return status;
1219 }
1220
1221 /*
1222  * Avoid buffered writes when a open context credential's key would
1223  * expire soon.
1224  *
1225  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1226  *
1227  * Return 0 and set a credential flag which triggers the inode to flush
1228  * and performs  NFS_FILE_SYNC writes if the key will expired within
1229  * RPC_KEY_EXPIRE_TIMEO.
1230  */
1231 int
1232 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1233 {
1234         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1235
1236         if (nfs_ctx_key_to_expire(ctx, inode) &&
1237             !rcu_access_pointer(ctx->ll_cred))
1238                 /* Already expired! */
1239                 return -EACCES;
1240         return 0;
1241 }
1242
1243 /*
1244  * Test if the open context credential key is marked to expire soon.
1245  */
1246 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1247 {
1248         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1249         struct rpc_cred *cred, *new, *old = NULL;
1250         struct auth_cred acred = {
1251                 .cred = ctx->cred,
1252         };
1253         bool ret = false;
1254
1255         rcu_read_lock();
1256         cred = rcu_dereference(ctx->ll_cred);
1257         if (cred && !(cred->cr_ops->crkey_timeout &&
1258                       cred->cr_ops->crkey_timeout(cred)))
1259                 goto out;
1260         rcu_read_unlock();
1261
1262         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1263         if (new == cred) {
1264                 put_rpccred(new);
1265                 return true;
1266         }
1267         if (IS_ERR_OR_NULL(new)) {
1268                 new = NULL;
1269                 ret = true;
1270         } else if (new->cr_ops->crkey_timeout &&
1271                    new->cr_ops->crkey_timeout(new))
1272                 ret = true;
1273
1274         rcu_read_lock();
1275         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1276                                              RCU_INITIALIZER(new)), 1);
1277 out:
1278         rcu_read_unlock();
1279         put_rpccred(old);
1280         return ret;
1281 }
1282
1283 /*
1284  * If the page cache is marked as unsafe or invalid, then we can't rely on
1285  * the PageUptodate() flag. In this case, we will need to turn off
1286  * write optimisations that depend on the page contents being correct.
1287  */
1288 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1289 {
1290         struct inode *inode = folio_file_mapping(folio)->host;
1291         struct nfs_inode *nfsi = NFS_I(inode);
1292
1293         if (nfs_have_delegated_attributes(inode))
1294                 goto out;
1295         if (nfsi->cache_validity &
1296             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1297                 return false;
1298         smp_rmb();
1299         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1300                 return false;
1301 out:
1302         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1303                 return false;
1304         return folio_test_uptodate(folio) != 0;
1305 }
1306
1307 static bool
1308 is_whole_file_wrlock(struct file_lock *fl)
1309 {
1310         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1311                         fl->fl_type == F_WRLCK;
1312 }
1313
1314 /* If we know the page is up to date, and we're not using byte range locks (or
1315  * if we have the whole file locked for writing), it may be more efficient to
1316  * extend the write to cover the entire page in order to avoid fragmentation
1317  * inefficiencies.
1318  *
1319  * If the file is opened for synchronous writes then we can just skip the rest
1320  * of the checks.
1321  */
1322 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1323                                 unsigned int pagelen)
1324 {
1325         struct inode *inode = file_inode(file);
1326         struct file_lock_context *flctx = locks_inode_context(inode);
1327         struct file_lock *fl;
1328         int ret;
1329
1330         if (file->f_flags & O_DSYNC)
1331                 return 0;
1332         if (!nfs_folio_write_uptodate(folio, pagelen))
1333                 return 0;
1334         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1335                 return 1;
1336         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1337                        list_empty_careful(&flctx->flc_posix)))
1338                 return 1;
1339
1340         /* Check to see if there are whole file write locks */
1341         ret = 0;
1342         spin_lock(&flctx->flc_lock);
1343         if (!list_empty(&flctx->flc_posix)) {
1344                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1345                                         fl_list);
1346                 if (is_whole_file_wrlock(fl))
1347                         ret = 1;
1348         } else if (!list_empty(&flctx->flc_flock)) {
1349                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1350                                         fl_list);
1351                 if (fl->fl_type == F_WRLCK)
1352                         ret = 1;
1353         }
1354         spin_unlock(&flctx->flc_lock);
1355         return ret;
1356 }
1357
1358 /*
1359  * Update and possibly write a cached page of an NFS file.
1360  *
1361  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1362  * things with a page scheduled for an RPC call (e.g. invalidate it).
1363  */
1364 int nfs_update_folio(struct file *file, struct folio *folio,
1365                      unsigned int offset, unsigned int count)
1366 {
1367         struct nfs_open_context *ctx = nfs_file_open_context(file);
1368         struct address_space *mapping = folio_file_mapping(folio);
1369         struct inode *inode = mapping->host;
1370         unsigned int pagelen = nfs_folio_length(folio);
1371         int             status = 0;
1372
1373         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1374
1375         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1376                 (long long)(folio_file_pos(folio) + offset));
1377
1378         if (!count)
1379                 goto out;
1380
1381         if (nfs_can_extend_write(file, folio, pagelen)) {
1382                 count = max(count + offset, pagelen);
1383                 offset = 0;
1384         }
1385
1386         status = nfs_writepage_setup(ctx, folio, offset, count);
1387         if (status < 0)
1388                 nfs_set_pageerror(mapping);
1389 out:
1390         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1391                         status, (long long)i_size_read(inode));
1392         return status;
1393 }
1394
1395 static int flush_task_priority(int how)
1396 {
1397         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1398                 case FLUSH_HIGHPRI:
1399                         return RPC_PRIORITY_HIGH;
1400                 case FLUSH_LOWPRI:
1401                         return RPC_PRIORITY_LOW;
1402         }
1403         return RPC_PRIORITY_NORMAL;
1404 }
1405
1406 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1407                                struct rpc_message *msg,
1408                                const struct nfs_rpc_ops *rpc_ops,
1409                                struct rpc_task_setup *task_setup_data, int how)
1410 {
1411         int priority = flush_task_priority(how);
1412
1413         if (IS_SWAPFILE(hdr->inode))
1414                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1415         task_setup_data->priority = priority;
1416         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1417         trace_nfs_initiate_write(hdr);
1418 }
1419
1420 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1421  * call this on each, which will prepare them to be retried on next
1422  * writeback using standard nfs.
1423  */
1424 static void nfs_redirty_request(struct nfs_page *req)
1425 {
1426         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1427
1428         /* Bump the transmission count */
1429         req->wb_nio++;
1430         nfs_mark_request_dirty(req);
1431         atomic_long_inc(&nfsi->redirtied_pages);
1432         nfs_page_end_writeback(req);
1433         nfs_release_request(req);
1434 }
1435
1436 static void nfs_async_write_error(struct list_head *head, int error)
1437 {
1438         struct nfs_page *req;
1439
1440         while (!list_empty(head)) {
1441                 req = nfs_list_entry(head->next);
1442                 nfs_list_remove_request(req);
1443                 if (nfs_error_is_fatal_on_server(error))
1444                         nfs_write_error(req, error);
1445                 else
1446                         nfs_redirty_request(req);
1447         }
1448 }
1449
1450 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1451 {
1452         nfs_async_write_error(&hdr->pages, 0);
1453 }
1454
1455 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1456         .init_hdr = nfs_async_write_init,
1457         .error_cleanup = nfs_async_write_error,
1458         .completion = nfs_write_completion,
1459         .reschedule_io = nfs_async_write_reschedule_io,
1460 };
1461
1462 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1463                                struct inode *inode, int ioflags, bool force_mds,
1464                                const struct nfs_pgio_completion_ops *compl_ops)
1465 {
1466         struct nfs_server *server = NFS_SERVER(inode);
1467         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1468
1469 #ifdef CONFIG_NFS_V4_1
1470         if (server->pnfs_curr_ld && !force_mds)
1471                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1472 #endif
1473         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1474                         server->wsize, ioflags);
1475 }
1476 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1477
1478 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1479 {
1480         struct nfs_pgio_mirror *mirror;
1481
1482         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1483                 pgio->pg_ops->pg_cleanup(pgio);
1484
1485         pgio->pg_ops = &nfs_pgio_rw_ops;
1486
1487         nfs_pageio_stop_mirroring(pgio);
1488
1489         mirror = &pgio->pg_mirrors[0];
1490         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1491 }
1492 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1493
1494
1495 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1496 {
1497         struct nfs_commit_data *data = calldata;
1498
1499         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1500 }
1501
1502 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1503                 struct nfs_fattr *fattr)
1504 {
1505         struct nfs_pgio_args *argp = &hdr->args;
1506         struct nfs_pgio_res *resp = &hdr->res;
1507         u64 size = argp->offset + resp->count;
1508
1509         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1510                 fattr->size = size;
1511         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1512                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1513                 return;
1514         }
1515         if (size != fattr->size)
1516                 return;
1517         /* Set attribute barrier */
1518         nfs_fattr_set_barrier(fattr);
1519         /* ...and update size */
1520         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1521 }
1522
1523 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1524 {
1525         struct nfs_fattr *fattr = &hdr->fattr;
1526         struct inode *inode = hdr->inode;
1527
1528         spin_lock(&inode->i_lock);
1529         nfs_writeback_check_extend(hdr, fattr);
1530         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1531         spin_unlock(&inode->i_lock);
1532 }
1533 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1534
1535 /*
1536  * This function is called when the WRITE call is complete.
1537  */
1538 static int nfs_writeback_done(struct rpc_task *task,
1539                               struct nfs_pgio_header *hdr,
1540                               struct inode *inode)
1541 {
1542         int status;
1543
1544         /*
1545          * ->write_done will attempt to use post-op attributes to detect
1546          * conflicting writes by other clients.  A strict interpretation
1547          * of close-to-open would allow us to continue caching even if
1548          * another writer had changed the file, but some applications
1549          * depend on tighter cache coherency when writing.
1550          */
1551         status = NFS_PROTO(inode)->write_done(task, hdr);
1552         if (status != 0)
1553                 return status;
1554
1555         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1556         trace_nfs_writeback_done(task, hdr);
1557
1558         if (task->tk_status >= 0) {
1559                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1560
1561                 if (committed == NFS_UNSTABLE) {
1562                         /*
1563                          * We have some uncommitted data on the server at
1564                          * this point, so ensure that we keep track of that
1565                          * fact irrespective of what later writes do.
1566                          */
1567                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1568                 }
1569
1570                 if (committed < hdr->args.stable) {
1571                         /* We tried a write call, but the server did not
1572                          * commit data to stable storage even though we
1573                          * requested it.
1574                          * Note: There is a known bug in Tru64 < 5.0 in which
1575                          *       the server reports NFS_DATA_SYNC, but performs
1576                          *       NFS_FILE_SYNC. We therefore implement this checking
1577                          *       as a dprintk() in order to avoid filling syslog.
1578                          */
1579                         static unsigned long    complain;
1580
1581                         /* Note this will print the MDS for a DS write */
1582                         if (time_before(complain, jiffies)) {
1583                                 dprintk("NFS:       faulty NFS server %s:"
1584                                         " (committed = %d) != (stable = %d)\n",
1585                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1586                                         committed, hdr->args.stable);
1587                                 complain = jiffies + 300 * HZ;
1588                         }
1589                 }
1590         }
1591
1592         /* Deal with the suid/sgid bit corner case */
1593         if (nfs_should_remove_suid(inode)) {
1594                 spin_lock(&inode->i_lock);
1595                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1596                 spin_unlock(&inode->i_lock);
1597         }
1598         return 0;
1599 }
1600
1601 /*
1602  * This function is called when the WRITE call is complete.
1603  */
1604 static void nfs_writeback_result(struct rpc_task *task,
1605                                  struct nfs_pgio_header *hdr)
1606 {
1607         struct nfs_pgio_args    *argp = &hdr->args;
1608         struct nfs_pgio_res     *resp = &hdr->res;
1609
1610         if (resp->count < argp->count) {
1611                 static unsigned long    complain;
1612
1613                 /* This a short write! */
1614                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1615
1616                 /* Has the server at least made some progress? */
1617                 if (resp->count == 0) {
1618                         if (time_before(complain, jiffies)) {
1619                                 printk(KERN_WARNING
1620                                        "NFS: Server wrote zero bytes, expected %u.\n",
1621                                        argp->count);
1622                                 complain = jiffies + 300 * HZ;
1623                         }
1624                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1625                         task->tk_status = -EIO;
1626                         return;
1627                 }
1628
1629                 /* For non rpc-based layout drivers, retry-through-MDS */
1630                 if (!task->tk_ops) {
1631                         hdr->pnfs_error = -EAGAIN;
1632                         return;
1633                 }
1634
1635                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1636                 if (resp->verf->committed != NFS_UNSTABLE) {
1637                         /* Resend from where the server left off */
1638                         hdr->mds_offset += resp->count;
1639                         argp->offset += resp->count;
1640                         argp->pgbase += resp->count;
1641                         argp->count -= resp->count;
1642                 } else {
1643                         /* Resend as a stable write in order to avoid
1644                          * headaches in the case of a server crash.
1645                          */
1646                         argp->stable = NFS_FILE_SYNC;
1647                 }
1648                 resp->count = 0;
1649                 resp->verf->committed = 0;
1650                 rpc_restart_call_prepare(task);
1651         }
1652 }
1653
1654 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1655 {
1656         return wait_var_event_killable(&cinfo->rpcs_out,
1657                                        !atomic_read(&cinfo->rpcs_out));
1658 }
1659
1660 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1661 {
1662         atomic_inc(&cinfo->rpcs_out);
1663 }
1664
1665 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1666 {
1667         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1668                 wake_up_var(&cinfo->rpcs_out);
1669                 return true;
1670         }
1671         return false;
1672 }
1673
1674 void nfs_commitdata_release(struct nfs_commit_data *data)
1675 {
1676         put_nfs_open_context(data->context);
1677         nfs_commit_free(data);
1678 }
1679 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1680
1681 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1682                         const struct nfs_rpc_ops *nfs_ops,
1683                         const struct rpc_call_ops *call_ops,
1684                         int how, int flags)
1685 {
1686         struct rpc_task *task;
1687         int priority = flush_task_priority(how);
1688         struct rpc_message msg = {
1689                 .rpc_argp = &data->args,
1690                 .rpc_resp = &data->res,
1691                 .rpc_cred = data->cred,
1692         };
1693         struct rpc_task_setup task_setup_data = {
1694                 .task = &data->task,
1695                 .rpc_client = clnt,
1696                 .rpc_message = &msg,
1697                 .callback_ops = call_ops,
1698                 .callback_data = data,
1699                 .workqueue = nfsiod_workqueue,
1700                 .flags = RPC_TASK_ASYNC | flags,
1701                 .priority = priority,
1702         };
1703
1704         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1705                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1706
1707         /* Set up the initial task struct.  */
1708         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1709         trace_nfs_initiate_commit(data);
1710
1711         dprintk("NFS: initiated commit call\n");
1712
1713         task = rpc_run_task(&task_setup_data);
1714         if (IS_ERR(task))
1715                 return PTR_ERR(task);
1716         if (how & FLUSH_SYNC)
1717                 rpc_wait_for_completion_task(task);
1718         rpc_put_task(task);
1719         return 0;
1720 }
1721 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1722
1723 static loff_t nfs_get_lwb(struct list_head *head)
1724 {
1725         loff_t lwb = 0;
1726         struct nfs_page *req;
1727
1728         list_for_each_entry(req, head, wb_list)
1729                 if (lwb < (req_offset(req) + req->wb_bytes))
1730                         lwb = req_offset(req) + req->wb_bytes;
1731
1732         return lwb;
1733 }
1734
1735 /*
1736  * Set up the argument/result storage required for the RPC call.
1737  */
1738 void nfs_init_commit(struct nfs_commit_data *data,
1739                      struct list_head *head,
1740                      struct pnfs_layout_segment *lseg,
1741                      struct nfs_commit_info *cinfo)
1742 {
1743         struct nfs_page *first;
1744         struct nfs_open_context *ctx;
1745         struct inode *inode;
1746
1747         /* Set up the RPC argument and reply structs
1748          * NB: take care not to mess about with data->commit et al. */
1749
1750         if (head)
1751                 list_splice_init(head, &data->pages);
1752
1753         first = nfs_list_entry(data->pages.next);
1754         ctx = nfs_req_openctx(first);
1755         inode = d_inode(ctx->dentry);
1756
1757         data->inode       = inode;
1758         data->cred        = ctx->cred;
1759         data->lseg        = lseg; /* reference transferred */
1760         /* only set lwb for pnfs commit */
1761         if (lseg)
1762                 data->lwb = nfs_get_lwb(&data->pages);
1763         data->mds_ops     = &nfs_commit_ops;
1764         data->completion_ops = cinfo->completion_ops;
1765         data->dreq        = cinfo->dreq;
1766
1767         data->args.fh     = NFS_FH(data->inode);
1768         /* Note: we always request a commit of the entire inode */
1769         data->args.offset = 0;
1770         data->args.count  = 0;
1771         data->context     = get_nfs_open_context(ctx);
1772         data->res.fattr   = &data->fattr;
1773         data->res.verf    = &data->verf;
1774         nfs_fattr_init(&data->fattr);
1775         nfs_commit_begin(cinfo->mds);
1776 }
1777 EXPORT_SYMBOL_GPL(nfs_init_commit);
1778
1779 void nfs_retry_commit(struct list_head *page_list,
1780                       struct pnfs_layout_segment *lseg,
1781                       struct nfs_commit_info *cinfo,
1782                       u32 ds_commit_idx)
1783 {
1784         struct nfs_page *req;
1785
1786         while (!list_empty(page_list)) {
1787                 req = nfs_list_entry(page_list->next);
1788                 nfs_list_remove_request(req);
1789                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1790                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1791                 nfs_unlock_and_release_request(req);
1792         }
1793 }
1794 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1795
1796 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1797                                      struct nfs_page *req)
1798 {
1799         struct folio *folio = nfs_page_to_folio(req);
1800
1801         filemap_dirty_folio(folio_mapping(folio), folio);
1802 }
1803
1804 /*
1805  * Commit dirty pages
1806  */
1807 static int
1808 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1809                 struct nfs_commit_info *cinfo)
1810 {
1811         struct nfs_commit_data  *data;
1812         unsigned short task_flags = 0;
1813
1814         /* another commit raced with us */
1815         if (list_empty(head))
1816                 return 0;
1817
1818         data = nfs_commitdata_alloc();
1819         if (!data) {
1820                 nfs_retry_commit(head, NULL, cinfo, -1);
1821                 return -ENOMEM;
1822         }
1823
1824         /* Set up the argument struct */
1825         nfs_init_commit(data, head, NULL, cinfo);
1826         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1827                 task_flags = RPC_TASK_MOVEABLE;
1828         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1829                                    data->mds_ops, how,
1830                                    RPC_TASK_CRED_NOREF | task_flags);
1831 }
1832
1833 /*
1834  * COMMIT call returned
1835  */
1836 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1837 {
1838         struct nfs_commit_data  *data = calldata;
1839
1840         /* Call the NFS version-specific code */
1841         NFS_PROTO(data->inode)->commit_done(task, data);
1842         trace_nfs_commit_done(task, data);
1843 }
1844
1845 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1846 {
1847         const struct nfs_writeverf *verf = data->res.verf;
1848         struct nfs_page *req;
1849         int status = data->task.tk_status;
1850         struct nfs_commit_info cinfo;
1851         struct nfs_server *nfss;
1852         struct folio *folio;
1853
1854         while (!list_empty(&data->pages)) {
1855                 req = nfs_list_entry(data->pages.next);
1856                 nfs_list_remove_request(req);
1857                 folio = nfs_page_to_folio(req);
1858                 nfs_folio_clear_commit(folio);
1859
1860                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1861                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1862                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1863                         req->wb_bytes,
1864                         (long long)req_offset(req));
1865                 if (status < 0) {
1866                         if (folio) {
1867                                 trace_nfs_commit_error(data->inode, req,
1868                                                        status);
1869                                 nfs_mapping_set_error(folio, status);
1870                                 nfs_inode_remove_request(req);
1871                         }
1872                         dprintk_cont(", error = %d\n", status);
1873                         goto next;
1874                 }
1875
1876                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1877                  * returned by the server against all stored verfs. */
1878                 if (nfs_write_match_verf(verf, req)) {
1879                         /* We have a match */
1880                         if (folio)
1881                                 nfs_inode_remove_request(req);
1882                         dprintk_cont(" OK\n");
1883                         goto next;
1884                 }
1885                 /* We have a mismatch. Write the page again */
1886                 dprintk_cont(" mismatch\n");
1887                 nfs_mark_request_dirty(req);
1888                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1889         next:
1890                 nfs_unlock_and_release_request(req);
1891                 /* Latency breaker */
1892                 cond_resched();
1893         }
1894         nfss = NFS_SERVER(data->inode);
1895         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1896                 nfss->write_congested = 0;
1897
1898         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1899         nfs_commit_end(cinfo.mds);
1900 }
1901
1902 static void nfs_commit_release(void *calldata)
1903 {
1904         struct nfs_commit_data *data = calldata;
1905
1906         data->completion_ops->completion(data);
1907         nfs_commitdata_release(calldata);
1908 }
1909
1910 static const struct rpc_call_ops nfs_commit_ops = {
1911         .rpc_call_prepare = nfs_commit_prepare,
1912         .rpc_call_done = nfs_commit_done,
1913         .rpc_release = nfs_commit_release,
1914 };
1915
1916 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1917         .completion = nfs_commit_release_pages,
1918         .resched_write = nfs_commit_resched_write,
1919 };
1920
1921 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1922                             int how, struct nfs_commit_info *cinfo)
1923 {
1924         int status;
1925
1926         status = pnfs_commit_list(inode, head, how, cinfo);
1927         if (status == PNFS_NOT_ATTEMPTED)
1928                 status = nfs_commit_list(inode, head, how, cinfo);
1929         return status;
1930 }
1931
1932 static int __nfs_commit_inode(struct inode *inode, int how,
1933                 struct writeback_control *wbc)
1934 {
1935         LIST_HEAD(head);
1936         struct nfs_commit_info cinfo;
1937         int may_wait = how & FLUSH_SYNC;
1938         int ret, nscan;
1939
1940         how &= ~FLUSH_SYNC;
1941         nfs_init_cinfo_from_inode(&cinfo, inode);
1942         nfs_commit_begin(cinfo.mds);
1943         for (;;) {
1944                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1945                 if (ret <= 0)
1946                         break;
1947                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1948                 if (ret < 0)
1949                         break;
1950                 ret = 0;
1951                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1952                         if (nscan < wbc->nr_to_write)
1953                                 wbc->nr_to_write -= nscan;
1954                         else
1955                                 wbc->nr_to_write = 0;
1956                 }
1957                 if (nscan < INT_MAX)
1958                         break;
1959                 cond_resched();
1960         }
1961         nfs_commit_end(cinfo.mds);
1962         if (ret || !may_wait)
1963                 return ret;
1964         return wait_on_commit(cinfo.mds);
1965 }
1966
1967 int nfs_commit_inode(struct inode *inode, int how)
1968 {
1969         return __nfs_commit_inode(inode, how, NULL);
1970 }
1971 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1972
1973 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1974 {
1975         struct nfs_inode *nfsi = NFS_I(inode);
1976         int flags = FLUSH_SYNC;
1977         int ret = 0;
1978
1979         if (wbc->sync_mode == WB_SYNC_NONE) {
1980                 /* no commits means nothing needs to be done */
1981                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1982                         goto check_requests_outstanding;
1983
1984                 /* Don't commit yet if this is a non-blocking flush and there
1985                  * are a lot of outstanding writes for this mapping.
1986                  */
1987                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1988                         goto out_mark_dirty;
1989
1990                 /* don't wait for the COMMIT response */
1991                 flags = 0;
1992         }
1993
1994         ret = __nfs_commit_inode(inode, flags, wbc);
1995         if (!ret) {
1996                 if (flags & FLUSH_SYNC)
1997                         return 0;
1998         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1999                 goto out_mark_dirty;
2000
2001 check_requests_outstanding:
2002         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2003                 return ret;
2004 out_mark_dirty:
2005         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2006         return ret;
2007 }
2008 EXPORT_SYMBOL_GPL(nfs_write_inode);
2009
2010 /*
2011  * Wrapper for filemap_write_and_wait_range()
2012  *
2013  * Needed for pNFS in order to ensure data becomes visible to the
2014  * client.
2015  */
2016 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2017                 loff_t lstart, loff_t lend)
2018 {
2019         int ret;
2020
2021         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2022         if (ret == 0)
2023                 ret = pnfs_sync_inode(mapping->host, true);
2024         return ret;
2025 }
2026 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2027
2028 /*
2029  * flush the inode to disk.
2030  */
2031 int nfs_wb_all(struct inode *inode)
2032 {
2033         int ret;
2034
2035         trace_nfs_writeback_inode_enter(inode);
2036
2037         ret = filemap_write_and_wait(inode->i_mapping);
2038         if (ret)
2039                 goto out;
2040         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2041         if (ret < 0)
2042                 goto out;
2043         pnfs_sync_inode(inode, true);
2044         ret = 0;
2045
2046 out:
2047         trace_nfs_writeback_inode_exit(inode, ret);
2048         return ret;
2049 }
2050 EXPORT_SYMBOL_GPL(nfs_wb_all);
2051
2052 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2053 {
2054         struct nfs_page *req;
2055         int ret = 0;
2056
2057         folio_wait_writeback(folio);
2058
2059         /* blocking call to cancel all requests and join to a single (head)
2060          * request */
2061         req = nfs_lock_and_join_requests(folio);
2062
2063         if (IS_ERR(req)) {
2064                 ret = PTR_ERR(req);
2065         } else if (req) {
2066                 /* all requests from this folio have been cancelled by
2067                  * nfs_lock_and_join_requests, so just remove the head
2068                  * request from the inode / page_private pointer and
2069                  * release it */
2070                 nfs_inode_remove_request(req);
2071                 nfs_unlock_and_release_request(req);
2072         }
2073
2074         return ret;
2075 }
2076
2077 /**
2078  * nfs_wb_folio - Write back all requests on one page
2079  * @inode: pointer to page
2080  * @folio: pointer to folio
2081  *
2082  * Assumes that the folio has been locked by the caller, and will
2083  * not unlock it.
2084  */
2085 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2086 {
2087         loff_t range_start = folio_file_pos(folio);
2088         loff_t range_end = range_start + (loff_t)folio_size(folio) - 1;
2089         struct writeback_control wbc = {
2090                 .sync_mode = WB_SYNC_ALL,
2091                 .nr_to_write = 0,
2092                 .range_start = range_start,
2093                 .range_end = range_end,
2094         };
2095         int ret;
2096
2097         trace_nfs_writeback_folio(inode, folio);
2098
2099         for (;;) {
2100                 folio_wait_writeback(folio);
2101                 if (folio_clear_dirty_for_io(folio)) {
2102                         ret = nfs_writepage_locked(folio, &wbc);
2103                         if (ret < 0)
2104                                 goto out_error;
2105                         continue;
2106                 }
2107                 ret = 0;
2108                 if (!folio_test_private(folio))
2109                         break;
2110                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2111                 if (ret < 0)
2112                         goto out_error;
2113         }
2114 out_error:
2115         trace_nfs_writeback_folio_done(inode, folio, ret);
2116         return ret;
2117 }
2118
2119 #ifdef CONFIG_MIGRATION
2120 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2121                 struct folio *src, enum migrate_mode mode)
2122 {
2123         /*
2124          * If the private flag is set, the folio is currently associated with
2125          * an in-progress read or write request. Don't try to migrate it.
2126          *
2127          * FIXME: we could do this in principle, but we'll need a way to ensure
2128          *        that we can safely release the inode reference while holding
2129          *        the folio lock.
2130          */
2131         if (folio_test_private(src))
2132                 return -EBUSY;
2133
2134         if (folio_test_fscache(src)) {
2135                 if (mode == MIGRATE_ASYNC)
2136                         return -EBUSY;
2137                 folio_wait_fscache(src);
2138         }
2139
2140         return migrate_folio(mapping, dst, src, mode);
2141 }
2142 #endif
2143
2144 int __init nfs_init_writepagecache(void)
2145 {
2146         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2147                                              sizeof(struct nfs_pgio_header),
2148                                              0, SLAB_HWCACHE_ALIGN,
2149                                              NULL);
2150         if (nfs_wdata_cachep == NULL)
2151                 return -ENOMEM;
2152
2153         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2154                                                      nfs_wdata_cachep);
2155         if (nfs_wdata_mempool == NULL)
2156                 goto out_destroy_write_cache;
2157
2158         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2159                                              sizeof(struct nfs_commit_data),
2160                                              0, SLAB_HWCACHE_ALIGN,
2161                                              NULL);
2162         if (nfs_cdata_cachep == NULL)
2163                 goto out_destroy_write_mempool;
2164
2165         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2166                                                       nfs_cdata_cachep);
2167         if (nfs_commit_mempool == NULL)
2168                 goto out_destroy_commit_cache;
2169
2170         /*
2171          * NFS congestion size, scale with available memory.
2172          *
2173          *  64MB:    8192k
2174          * 128MB:   11585k
2175          * 256MB:   16384k
2176          * 512MB:   23170k
2177          *   1GB:   32768k
2178          *   2GB:   46340k
2179          *   4GB:   65536k
2180          *   8GB:   92681k
2181          *  16GB:  131072k
2182          *
2183          * This allows larger machines to have larger/more transfers.
2184          * Limit the default to 256M
2185          */
2186         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2187         if (nfs_congestion_kb > 256*1024)
2188                 nfs_congestion_kb = 256*1024;
2189
2190         return 0;
2191
2192 out_destroy_commit_cache:
2193         kmem_cache_destroy(nfs_cdata_cachep);
2194 out_destroy_write_mempool:
2195         mempool_destroy(nfs_wdata_mempool);
2196 out_destroy_write_cache:
2197         kmem_cache_destroy(nfs_wdata_cachep);
2198         return -ENOMEM;
2199 }
2200
2201 void nfs_destroy_writepagecache(void)
2202 {
2203         mempool_destroy(nfs_commit_mempool);
2204         kmem_cache_destroy(nfs_cdata_cachep);
2205         mempool_destroy(nfs_wdata_mempool);
2206         kmem_cache_destroy(nfs_wdata_cachep);
2207 }
2208
2209 static const struct nfs_rw_ops nfs_rw_write_ops = {
2210         .rw_alloc_header        = nfs_writehdr_alloc,
2211         .rw_free_header         = nfs_writehdr_free,
2212         .rw_done                = nfs_writeback_done,
2213         .rw_result              = nfs_writeback_result,
2214         .rw_initiate            = nfs_initiate_write,
2215 };