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                 folio_redirty_for_writepage(wbc, folio);
 673                 return AOP_WRITEPAGE_ACTIVATE;
 674         }
 675
 676         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 677         nfs_pageio_init_write(&pgio, inode, 0, false,
 678                               &nfs_async_write_completion_ops);
 679         err = nfs_do_writepage(folio, wbc, &pgio);
 680         pgio.pg_error = 0;
 681         nfs_pageio_complete(&pgio);
 682         return err;
 683 }
 684
 685 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 686 {
 687         struct folio *folio = page_folio(page);
 688         int ret;
 689
 690         ret = nfs_writepage_locked(folio, wbc);
 691         if (ret != AOP_WRITEPAGE_ACTIVATE)
 692                 unlock_page(page);
 693         return ret;
 694 }
 695
 696 static int nfs_writepages_callback(struct folio *folio,
 697                                    struct writeback_control *wbc, void *data)
 698 {
 699         int ret;
 700
 701         ret = nfs_do_writepage(folio, wbc, data);
 702         if (ret != AOP_WRITEPAGE_ACTIVATE)
 703                 folio_unlock(folio);
 704         return ret;
 705 }
 706
 707 static void nfs_io_completion_commit(void *inode)
 708 {
 709         nfs_commit_inode(inode, 0);
 710 }
 711
 712 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 713 {
 714         struct inode *inode = mapping->host;
 715         struct nfs_pageio_descriptor pgio;
 716         struct nfs_io_completion *ioc = NULL;
 717         unsigned int mntflags = NFS_SERVER(inode)->flags;
 718         int priority = 0;
 719         int err;
 720
 721         if (wbc->sync_mode == WB_SYNC_NONE &&
 722             NFS_SERVER(inode)->write_congested)
 723                 return 0;
 724
 725         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 726
 727         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 728             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 729                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 730                 if (ioc)
 731                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 732                                                inode);
 733                 priority = wb_priority(wbc);
 734         }
 735
 736         do {
 737                 nfs_pageio_init_write(&pgio, inode, priority, false,
 738                                       &nfs_async_write_completion_ops);
 739                 pgio.pg_io_completion = ioc;
 740                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 741                                         &pgio);
 742                 pgio.pg_error = 0;
 743                 nfs_pageio_complete(&pgio);
 744                 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
 745                         break;
 746         } while (err < 0 && !nfs_error_is_fatal(err));
 747         nfs_io_completion_put(ioc);
 748
 749         if (err < 0)
 750                 goto out_err;
 751         return 0;
 752 out_err:
 753         return err;
 754 }
 755
 756 /*
 757  * Insert a write request into an inode
 758  */
 759 static void nfs_inode_add_request(struct nfs_page *req)
 760 {
 761         struct folio *folio = nfs_page_to_folio(req);
 762         struct address_space *mapping = folio_file_mapping(folio);
 763         struct nfs_inode *nfsi = NFS_I(mapping->host);
 764
 765         WARN_ON_ONCE(req->wb_this_page != req);
 766
 767         /* Lock the request! */
 768         nfs_lock_request(req);
 769
 770         /*
 771          * Swap-space should not get truncated. Hence no need to plug the race
 772          * with invalidate/truncate.
 773          */
 774         spin_lock(&mapping->private_lock);
 775         if (likely(!folio_test_swapcache(folio))) {
 776                 set_bit(PG_MAPPED, &req->wb_flags);
 777                 folio_set_private(folio);
 778                 folio->private = req;
 779         }
 780         spin_unlock(&mapping->private_lock);
 781         atomic_long_inc(&nfsi->nrequests);
 782         /* this a head request for a page group - mark it as having an
 783          * extra reference so sub groups can follow suit.
 784          * This flag also informs pgio layer when to bump nrequests when
 785          * adding subrequests. */
 786         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 787         kref_get(&req->wb_kref);
 788 }
 789
 790 /*
 791  * Remove a write request from an inode
 792  */
 793 static void nfs_inode_remove_request(struct nfs_page *req)
 794 {
 795         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 796
 797         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 798                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 799                 struct address_space *mapping = folio_file_mapping(folio);
 800
 801                 spin_lock(&mapping->private_lock);
 802                 if (likely(folio && !folio_test_swapcache(folio))) {
 803                         folio->private = NULL;
 804                         folio_clear_private(folio);
 805                         clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 806                 }
 807                 spin_unlock(&mapping->private_lock);
 808         }
 809
 810         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 811                 atomic_long_dec(&nfsi->nrequests);
 812                 nfs_release_request(req);
 813         }
 814 }
 815
 816 static void nfs_mark_request_dirty(struct nfs_page *req)
 817 {
 818         struct folio *folio = nfs_page_to_folio(req);
 819         if (folio)
 820                 filemap_dirty_folio(folio_mapping(folio), folio);
 821 }
 822
 823 /*
 824  * nfs_page_search_commits_for_head_request_locked
 825  *
 826  * Search through commit lists on @inode for the head request for @folio.
 827  * Must be called while holding the inode (which is cinfo) lock.
 828  *
 829  * Returns the head request if found, or NULL if not found.
 830  */
 831 static struct nfs_page *
 832 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 833                                                 struct folio *folio)
 834 {
 835         struct nfs_page *freq, *t;
 836         struct nfs_commit_info cinfo;
 837         struct inode *inode = &nfsi->vfs_inode;
 838
 839         nfs_init_cinfo_from_inode(&cinfo, inode);
 840
 841         /* search through pnfs commit lists */
 842         freq = pnfs_search_commit_reqs(inode, &cinfo, folio);
 843         if (freq)
 844                 return freq->wb_head;
 845
 846         /* Linearly search the commit list for the correct request */
 847         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 848                 if (nfs_page_to_folio(freq) == folio)
 849                         return freq->wb_head;
 850         }
 851
 852         return NULL;
 853 }
 854
 855 /**
 856  * nfs_request_add_commit_list_locked - add request to a commit list
 857  * @req: pointer to a struct nfs_page
 858  * @dst: commit list head
 859  * @cinfo: holds list lock and accounting info
 860  *
 861  * This sets the PG_CLEAN bit, updates the cinfo count of
 862  * number of outstanding requests requiring a commit as well as
 863  * the MM page stats.
 864  *
 865  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 866  * nfs_page lock.
 867  */
 868 void
 869 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 870                             struct nfs_commit_info *cinfo)
 871 {
 872         set_bit(PG_CLEAN, &req->wb_flags);
 873         nfs_list_add_request(req, dst);
 874         atomic_long_inc(&cinfo->mds->ncommit);
 875 }
 876 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 877
 878 /**
 879  * nfs_request_add_commit_list - add request to a commit list
 880  * @req: pointer to a struct nfs_page
 881  * @cinfo: holds list lock and accounting info
 882  *
 883  * This sets the PG_CLEAN bit, updates the cinfo count of
 884  * number of outstanding requests requiring a commit as well as
 885  * the MM page stats.
 886  *
 887  * The caller must _not_ hold the cinfo->lock, but must be
 888  * holding the nfs_page lock.
 889  */
 890 void
 891 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 892 {
 893         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 894         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 895         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 896         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 897 }
 898 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 899
 900 /**
 901  * nfs_request_remove_commit_list - Remove request from a commit list
 902  * @req: pointer to a nfs_page
 903  * @cinfo: holds list lock and accounting info
 904  *
 905  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 906  * number of outstanding requests requiring a commit
 907  * It does not update the MM page stats.
 908  *
 909  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 910  */
 911 void
 912 nfs_request_remove_commit_list(struct nfs_page *req,
 913                                struct nfs_commit_info *cinfo)
 914 {
 915         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 916                 return;
 917         nfs_list_remove_request(req);
 918         atomic_long_dec(&cinfo->mds->ncommit);
 919 }
 920 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 921
 922 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 923                                       struct inode *inode)
 924 {
 925         cinfo->inode = inode;
 926         cinfo->mds = &NFS_I(inode)->commit_info;
 927         cinfo->ds = pnfs_get_ds_info(inode);
 928         cinfo->dreq = NULL;
 929         cinfo->completion_ops = &nfs_commit_completion_ops;
 930 }
 931
 932 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 933                     struct inode *inode,
 934                     struct nfs_direct_req *dreq)
 935 {
 936         if (dreq)
 937                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 938         else
 939                 nfs_init_cinfo_from_inode(cinfo, inode);
 940 }
 941 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 942
 943 /*
 944  * Add a request to the inode's commit list.
 945  */
 946 void
 947 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 948                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 949 {
 950         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 951                 return;
 952         nfs_request_add_commit_list(req, cinfo);
 953 }
 954
 955 static void nfs_folio_clear_commit(struct folio *folio)
 956 {
 957         if (folio) {
 958                 long nr = folio_nr_pages(folio);
 959
 960                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 961                 wb_stat_mod(&inode_to_bdi(folio_file_mapping(folio)->host)->wb,
 962                             WB_WRITEBACK, -nr);
 963         }
 964 }
 965
 966 /* Called holding the request lock on @req */
 967 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 968                                      struct nfs_page *req)
 969 {
 970         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 971                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 972                 struct inode *inode = d_inode(ctx->dentry);
 973
 974                 mutex_lock(&NFS_I(inode)->commit_mutex);
 975                 if (!pnfs_clear_request_commit(req, cinfo)) {
 976                         nfs_request_remove_commit_list(req, cinfo);
 977                 }
 978                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 979                 nfs_folio_clear_commit(nfs_page_to_folio(req));
 980         }
 981 }
 982
 983 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 984 {
 985         if (hdr->verf.committed == NFS_DATA_SYNC)
 986                 return hdr->lseg == NULL;
 987         return hdr->verf.committed != NFS_FILE_SYNC;
 988 }
 989
 990 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 991 {
 992         nfs_io_completion_get(hdr->io_completion);
 993 }
 994
 995 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 996 {
 997         struct nfs_commit_info cinfo;
 998         unsigned long bytes = 0;
 999
1000         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
1001                 goto out;
1002         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
1003         while (!list_empty(&hdr->pages)) {
1004                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1005
1006                 bytes += req->wb_bytes;
1007                 nfs_list_remove_request(req);
1008                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1009                     (hdr->good_bytes < bytes)) {
1010                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
1011                         nfs_mapping_set_error(nfs_page_to_folio(req),
1012                                               hdr->error);
1013                         goto remove_req;
1014                 }
1015                 if (nfs_write_need_commit(hdr)) {
1016                         /* Reset wb_nio, since the write was successful. */
1017                         req->wb_nio = 0;
1018                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1019                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1020                                 hdr->pgio_mirror_idx);
1021                         goto next;
1022                 }
1023 remove_req:
1024                 nfs_inode_remove_request(req);
1025 next:
1026                 nfs_page_end_writeback(req);
1027                 nfs_release_request(req);
1028         }
1029 out:
1030         nfs_io_completion_put(hdr->io_completion);
1031         hdr->release(hdr);
1032 }
1033
1034 unsigned long
1035 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1036 {
1037         return atomic_long_read(&cinfo->mds->ncommit);
1038 }
1039
1040 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1041 int
1042 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1043                      struct nfs_commit_info *cinfo, int max)
1044 {
1045         struct nfs_page *req, *tmp;
1046         int ret = 0;
1047
1048         list_for_each_entry_safe(req, tmp, src, wb_list) {
1049                 kref_get(&req->wb_kref);
1050                 if (!nfs_lock_request(req)) {
1051                         nfs_release_request(req);
1052                         continue;
1053                 }
1054                 nfs_request_remove_commit_list(req, cinfo);
1055                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1056                 nfs_list_add_request(req, dst);
1057                 ret++;
1058                 if ((ret == max) && !cinfo->dreq)
1059                         break;
1060                 cond_resched();
1061         }
1062         return ret;
1063 }
1064 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1065
1066 /*
1067  * nfs_scan_commit - Scan an inode for commit requests
1068  * @inode: NFS inode to scan
1069  * @dst: mds destination list
1070  * @cinfo: mds and ds lists of reqs ready to commit
1071  *
1072  * Moves requests from the inode's 'commit' request list.
1073  * The requests are *not* checked to ensure that they form a contiguous set.
1074  */
1075 int
1076 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1077                 struct nfs_commit_info *cinfo)
1078 {
1079         int ret = 0;
1080
1081         if (!atomic_long_read(&cinfo->mds->ncommit))
1082                 return 0;
1083         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1084         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1085                 const int max = INT_MAX;
1086
1087                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1088                                            cinfo, max);
1089                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1090         }
1091         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1092         return ret;
1093 }
1094
1095 /*
1096  * Search for an existing write request, and attempt to update
1097  * it to reflect a new dirty region on a given page.
1098  *
1099  * If the attempt fails, then the existing request is flushed out
1100  * to disk.
1101  */
1102 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1103                                                   unsigned int offset,
1104                                                   unsigned int bytes)
1105 {
1106         struct nfs_page *req;
1107         unsigned int rqend;
1108         unsigned int end;
1109         int error;
1110
1111         end = offset + bytes;
1112
1113         req = nfs_lock_and_join_requests(folio);
1114         if (IS_ERR_OR_NULL(req))
1115                 return req;
1116
1117         rqend = req->wb_offset + req->wb_bytes;
1118         /*
1119          * Tell the caller to flush out the request if
1120          * the offsets are non-contiguous.
1121          * Note: nfs_flush_incompatible() will already
1122          * have flushed out requests having wrong owners.
1123          */
1124         if (offset > rqend || end < req->wb_offset)
1125                 goto out_flushme;
1126
1127         /* Okay, the request matches. Update the region */
1128         if (offset < req->wb_offset) {
1129                 req->wb_offset = offset;
1130                 req->wb_pgbase = offset;
1131         }
1132         if (end > rqend)
1133                 req->wb_bytes = end - req->wb_offset;
1134         else
1135                 req->wb_bytes = rqend - req->wb_offset;
1136         req->wb_nio = 0;
1137         return req;
1138 out_flushme:
1139         /*
1140          * Note: we mark the request dirty here because
1141          * nfs_lock_and_join_requests() cannot preserve
1142          * commit flags, so we have to replay the write.
1143          */
1144         nfs_mark_request_dirty(req);
1145         nfs_unlock_and_release_request(req);
1146         error = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1147         return (error < 0) ? ERR_PTR(error) : NULL;
1148 }
1149
1150 /*
1151  * Try to update an existing write request, or create one if there is none.
1152  *
1153  * Note: Should always be called with the Page Lock held to prevent races
1154  * if we have to add a new request. Also assumes that the caller has
1155  * already called nfs_flush_incompatible() if necessary.
1156  */
1157 static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1158                                                 struct folio *folio,
1159                                                 unsigned int offset,
1160                                                 unsigned int bytes)
1161 {
1162         struct nfs_page *req;
1163
1164         req = nfs_try_to_update_request(folio, offset, bytes);
1165         if (req != NULL)
1166                 goto out;
1167         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1168         if (IS_ERR(req))
1169                 goto out;
1170         nfs_inode_add_request(req);
1171 out:
1172         return req;
1173 }
1174
1175 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1176                                struct folio *folio, unsigned int offset,
1177                                unsigned int count)
1178 {
1179         struct nfs_page *req;
1180
1181         req = nfs_setup_write_request(ctx, folio, offset, count);
1182         if (IS_ERR(req))
1183                 return PTR_ERR(req);
1184         /* Update file length */
1185         nfs_grow_file(folio, offset, count);
1186         nfs_mark_uptodate(req);
1187         nfs_mark_request_dirty(req);
1188         nfs_unlock_and_release_request(req);
1189         return 0;
1190 }
1191
1192 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1193 {
1194         struct nfs_open_context *ctx = nfs_file_open_context(file);
1195         struct nfs_lock_context *l_ctx;
1196         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1197         struct nfs_page *req;
1198         int do_flush, status;
1199         /*
1200          * Look for a request corresponding to this page. If there
1201          * is one, and it belongs to another file, we flush it out
1202          * before we try to copy anything into the page. Do this
1203          * due to the lack of an ACCESS-type call in NFSv2.
1204          * Also do the same if we find a request from an existing
1205          * dropped page.
1206          */
1207         do {
1208                 req = nfs_folio_find_head_request(folio);
1209                 if (req == NULL)
1210                         return 0;
1211                 l_ctx = req->wb_lock_context;
1212                 do_flush = nfs_page_to_folio(req) != folio ||
1213                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1214                 if (l_ctx && flctx &&
1215                     !(list_empty_careful(&flctx->flc_posix) &&
1216                       list_empty_careful(&flctx->flc_flock))) {
1217                         do_flush |= l_ctx->lockowner != current->files;
1218                 }
1219                 nfs_release_request(req);
1220                 if (!do_flush)
1221                         return 0;
1222                 status = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1223         } while (status == 0);
1224         return status;
1225 }
1226
1227 /*
1228  * Avoid buffered writes when a open context credential's key would
1229  * expire soon.
1230  *
1231  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1232  *
1233  * Return 0 and set a credential flag which triggers the inode to flush
1234  * and performs  NFS_FILE_SYNC writes if the key will expired within
1235  * RPC_KEY_EXPIRE_TIMEO.
1236  */
1237 int
1238 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1239 {
1240         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1241
1242         if (nfs_ctx_key_to_expire(ctx, inode) &&
1243             !rcu_access_pointer(ctx->ll_cred))
1244                 /* Already expired! */
1245                 return -EACCES;
1246         return 0;
1247 }
1248
1249 /*
1250  * Test if the open context credential key is marked to expire soon.
1251  */
1252 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1253 {
1254         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1255         struct rpc_cred *cred, *new, *old = NULL;
1256         struct auth_cred acred = {
1257                 .cred = ctx->cred,
1258         };
1259         bool ret = false;
1260
1261         rcu_read_lock();
1262         cred = rcu_dereference(ctx->ll_cred);
1263         if (cred && !(cred->cr_ops->crkey_timeout &&
1264                       cred->cr_ops->crkey_timeout(cred)))
1265                 goto out;
1266         rcu_read_unlock();
1267
1268         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1269         if (new == cred) {
1270                 put_rpccred(new);
1271                 return true;
1272         }
1273         if (IS_ERR_OR_NULL(new)) {
1274                 new = NULL;
1275                 ret = true;
1276         } else if (new->cr_ops->crkey_timeout &&
1277                    new->cr_ops->crkey_timeout(new))
1278                 ret = true;
1279
1280         rcu_read_lock();
1281         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1282                                              RCU_INITIALIZER(new)), 1);
1283 out:
1284         rcu_read_unlock();
1285         put_rpccred(old);
1286         return ret;
1287 }
1288
1289 /*
1290  * If the page cache is marked as unsafe or invalid, then we can't rely on
1291  * the PageUptodate() flag. In this case, we will need to turn off
1292  * write optimisations that depend on the page contents being correct.
1293  */
1294 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1295 {
1296         struct inode *inode = folio_file_mapping(folio)->host;
1297         struct nfs_inode *nfsi = NFS_I(inode);
1298
1299         if (nfs_have_delegated_attributes(inode))
1300                 goto out;
1301         if (nfsi->cache_validity &
1302             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1303                 return false;
1304         smp_rmb();
1305         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1306                 return false;
1307 out:
1308         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1309                 return false;
1310         return folio_test_uptodate(folio) != 0;
1311 }
1312
1313 static bool
1314 is_whole_file_wrlock(struct file_lock *fl)
1315 {
1316         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1317                         fl->fl_type == F_WRLCK;
1318 }
1319
1320 /* If we know the page is up to date, and we're not using byte range locks (or
1321  * if we have the whole file locked for writing), it may be more efficient to
1322  * extend the write to cover the entire page in order to avoid fragmentation
1323  * inefficiencies.
1324  *
1325  * If the file is opened for synchronous writes then we can just skip the rest
1326  * of the checks.
1327  */
1328 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1329                                 unsigned int pagelen)
1330 {
1331         struct inode *inode = file_inode(file);
1332         struct file_lock_context *flctx = locks_inode_context(inode);
1333         struct file_lock *fl;
1334         int ret;
1335
1336         if (file->f_flags & O_DSYNC)
1337                 return 0;
1338         if (!nfs_folio_write_uptodate(folio, pagelen))
1339                 return 0;
1340         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1341                 return 1;
1342         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1343                        list_empty_careful(&flctx->flc_posix)))
1344                 return 1;
1345
1346         /* Check to see if there are whole file write locks */
1347         ret = 0;
1348         spin_lock(&flctx->flc_lock);
1349         if (!list_empty(&flctx->flc_posix)) {
1350                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1351                                         fl_list);
1352                 if (is_whole_file_wrlock(fl))
1353                         ret = 1;
1354         } else if (!list_empty(&flctx->flc_flock)) {
1355                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1356                                         fl_list);
1357                 if (fl->fl_type == F_WRLCK)
1358                         ret = 1;
1359         }
1360         spin_unlock(&flctx->flc_lock);
1361         return ret;
1362 }
1363
1364 /*
1365  * Update and possibly write a cached page of an NFS file.
1366  *
1367  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1368  * things with a page scheduled for an RPC call (e.g. invalidate it).
1369  */
1370 int nfs_update_folio(struct file *file, struct folio *folio,
1371                      unsigned int offset, unsigned int count)
1372 {
1373         struct nfs_open_context *ctx = nfs_file_open_context(file);
1374         struct address_space *mapping = folio_file_mapping(folio);
1375         struct inode *inode = mapping->host;
1376         unsigned int pagelen = nfs_folio_length(folio);
1377         int             status = 0;
1378
1379         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1380
1381         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1382                 (long long)(folio_file_pos(folio) + offset));
1383
1384         if (!count)
1385                 goto out;
1386
1387         if (nfs_can_extend_write(file, folio, pagelen)) {
1388                 count = max(count + offset, pagelen);
1389                 offset = 0;
1390         }
1391
1392         status = nfs_writepage_setup(ctx, folio, offset, count);
1393         if (status < 0)
1394                 nfs_set_pageerror(mapping);
1395 out:
1396         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1397                         status, (long long)i_size_read(inode));
1398         return status;
1399 }
1400
1401 static int flush_task_priority(int how)
1402 {
1403         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1404                 case FLUSH_HIGHPRI:
1405                         return RPC_PRIORITY_HIGH;
1406                 case FLUSH_LOWPRI:
1407                         return RPC_PRIORITY_LOW;
1408         }
1409         return RPC_PRIORITY_NORMAL;
1410 }
1411
1412 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1413                                struct rpc_message *msg,
1414                                const struct nfs_rpc_ops *rpc_ops,
1415                                struct rpc_task_setup *task_setup_data, int how)
1416 {
1417         int priority = flush_task_priority(how);
1418
1419         if (IS_SWAPFILE(hdr->inode))
1420                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1421         task_setup_data->priority = priority;
1422         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1423         trace_nfs_initiate_write(hdr);
1424 }
1425
1426 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1427  * call this on each, which will prepare them to be retried on next
1428  * writeback using standard nfs.
1429  */
1430 static void nfs_redirty_request(struct nfs_page *req)
1431 {
1432         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1433
1434         /* Bump the transmission count */
1435         req->wb_nio++;
1436         nfs_mark_request_dirty(req);
1437         atomic_long_inc(&nfsi->redirtied_pages);
1438         nfs_page_end_writeback(req);
1439         nfs_release_request(req);
1440 }
1441
1442 static void nfs_async_write_error(struct list_head *head, int error)
1443 {
1444         struct nfs_page *req;
1445
1446         while (!list_empty(head)) {
1447                 req = nfs_list_entry(head->next);
1448                 nfs_list_remove_request(req);
1449                 if (nfs_error_is_fatal_on_server(error))
1450                         nfs_write_error(req, error);
1451                 else
1452                         nfs_redirty_request(req);
1453         }
1454 }
1455
1456 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1457 {
1458         nfs_async_write_error(&hdr->pages, 0);
1459 }
1460
1461 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1462         .init_hdr = nfs_async_write_init,
1463         .error_cleanup = nfs_async_write_error,
1464         .completion = nfs_write_completion,
1465         .reschedule_io = nfs_async_write_reschedule_io,
1466 };
1467
1468 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1469                                struct inode *inode, int ioflags, bool force_mds,
1470                                const struct nfs_pgio_completion_ops *compl_ops)
1471 {
1472         struct nfs_server *server = NFS_SERVER(inode);
1473         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1474
1475 #ifdef CONFIG_NFS_V4_1
1476         if (server->pnfs_curr_ld && !force_mds)
1477                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1478 #endif
1479         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1480                         server->wsize, ioflags);
1481 }
1482 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1483
1484 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1485 {
1486         struct nfs_pgio_mirror *mirror;
1487
1488         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1489                 pgio->pg_ops->pg_cleanup(pgio);
1490
1491         pgio->pg_ops = &nfs_pgio_rw_ops;
1492
1493         nfs_pageio_stop_mirroring(pgio);
1494
1495         mirror = &pgio->pg_mirrors[0];
1496         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1497 }
1498 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1499
1500
1501 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1502 {
1503         struct nfs_commit_data *data = calldata;
1504
1505         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1506 }
1507
1508 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1509                 struct nfs_fattr *fattr)
1510 {
1511         struct nfs_pgio_args *argp = &hdr->args;
1512         struct nfs_pgio_res *resp = &hdr->res;
1513         u64 size = argp->offset + resp->count;
1514
1515         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1516                 fattr->size = size;
1517         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1518                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1519                 return;
1520         }
1521         if (size != fattr->size)
1522                 return;
1523         /* Set attribute barrier */
1524         nfs_fattr_set_barrier(fattr);
1525         /* ...and update size */
1526         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1527 }
1528
1529 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1530 {
1531         struct nfs_fattr *fattr = &hdr->fattr;
1532         struct inode *inode = hdr->inode;
1533
1534         spin_lock(&inode->i_lock);
1535         nfs_writeback_check_extend(hdr, fattr);
1536         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1537         spin_unlock(&inode->i_lock);
1538 }
1539 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1540
1541 /*
1542  * This function is called when the WRITE call is complete.
1543  */
1544 static int nfs_writeback_done(struct rpc_task *task,
1545                               struct nfs_pgio_header *hdr,
1546                               struct inode *inode)
1547 {
1548         int status;
1549
1550         /*
1551          * ->write_done will attempt to use post-op attributes to detect
1552          * conflicting writes by other clients.  A strict interpretation
1553          * of close-to-open would allow us to continue caching even if
1554          * another writer had changed the file, but some applications
1555          * depend on tighter cache coherency when writing.
1556          */
1557         status = NFS_PROTO(inode)->write_done(task, hdr);
1558         if (status != 0)
1559                 return status;
1560
1561         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1562         trace_nfs_writeback_done(task, hdr);
1563
1564         if (task->tk_status >= 0) {
1565                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1566
1567                 if (committed == NFS_UNSTABLE) {
1568                         /*
1569                          * We have some uncommitted data on the server at
1570                          * this point, so ensure that we keep track of that
1571                          * fact irrespective of what later writes do.
1572                          */
1573                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1574                 }
1575
1576                 if (committed < hdr->args.stable) {
1577                         /* We tried a write call, but the server did not
1578                          * commit data to stable storage even though we
1579                          * requested it.
1580                          * Note: There is a known bug in Tru64 < 5.0 in which
1581                          *       the server reports NFS_DATA_SYNC, but performs
1582                          *       NFS_FILE_SYNC. We therefore implement this checking
1583                          *       as a dprintk() in order to avoid filling syslog.
1584                          */
1585                         static unsigned long    complain;
1586
1587                         /* Note this will print the MDS for a DS write */
1588                         if (time_before(complain, jiffies)) {
1589                                 dprintk("NFS:       faulty NFS server %s:"
1590                                         " (committed = %d) != (stable = %d)\n",
1591                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1592                                         committed, hdr->args.stable);
1593                                 complain = jiffies + 300 * HZ;
1594                         }
1595                 }
1596         }
1597
1598         /* Deal with the suid/sgid bit corner case */
1599         if (nfs_should_remove_suid(inode)) {
1600                 spin_lock(&inode->i_lock);
1601                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1602                 spin_unlock(&inode->i_lock);
1603         }
1604         return 0;
1605 }
1606
1607 /*
1608  * This function is called when the WRITE call is complete.
1609  */
1610 static void nfs_writeback_result(struct rpc_task *task,
1611                                  struct nfs_pgio_header *hdr)
1612 {
1613         struct nfs_pgio_args    *argp = &hdr->args;
1614         struct nfs_pgio_res     *resp = &hdr->res;
1615
1616         if (resp->count < argp->count) {
1617                 static unsigned long    complain;
1618
1619                 /* This a short write! */
1620                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1621
1622                 /* Has the server at least made some progress? */
1623                 if (resp->count == 0) {
1624                         if (time_before(complain, jiffies)) {
1625                                 printk(KERN_WARNING
1626                                        "NFS: Server wrote zero bytes, expected %u.\n",
1627                                        argp->count);
1628                                 complain = jiffies + 300 * HZ;
1629                         }
1630                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1631                         task->tk_status = -EIO;
1632                         return;
1633                 }
1634
1635                 /* For non rpc-based layout drivers, retry-through-MDS */
1636                 if (!task->tk_ops) {
1637                         hdr->pnfs_error = -EAGAIN;
1638                         return;
1639                 }
1640
1641                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1642                 if (resp->verf->committed != NFS_UNSTABLE) {
1643                         /* Resend from where the server left off */
1644                         hdr->mds_offset += resp->count;
1645                         argp->offset += resp->count;
1646                         argp->pgbase += resp->count;
1647                         argp->count -= resp->count;
1648                 } else {
1649                         /* Resend as a stable write in order to avoid
1650                          * headaches in the case of a server crash.
1651                          */
1652                         argp->stable = NFS_FILE_SYNC;
1653                 }
1654                 resp->count = 0;
1655                 resp->verf->committed = 0;
1656                 rpc_restart_call_prepare(task);
1657         }
1658 }
1659
1660 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1661 {
1662         return wait_var_event_killable(&cinfo->rpcs_out,
1663                                        !atomic_read(&cinfo->rpcs_out));
1664 }
1665
1666 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1667 {
1668         atomic_inc(&cinfo->rpcs_out);
1669 }
1670
1671 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1672 {
1673         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1674                 wake_up_var(&cinfo->rpcs_out);
1675                 return true;
1676         }
1677         return false;
1678 }
1679
1680 void nfs_commitdata_release(struct nfs_commit_data *data)
1681 {
1682         put_nfs_open_context(data->context);
1683         nfs_commit_free(data);
1684 }
1685 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1686
1687 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1688                         const struct nfs_rpc_ops *nfs_ops,
1689                         const struct rpc_call_ops *call_ops,
1690                         int how, int flags)
1691 {
1692         struct rpc_task *task;
1693         int priority = flush_task_priority(how);
1694         struct rpc_message msg = {
1695                 .rpc_argp = &data->args,
1696                 .rpc_resp = &data->res,
1697                 .rpc_cred = data->cred,
1698         };
1699         struct rpc_task_setup task_setup_data = {
1700                 .task = &data->task,
1701                 .rpc_client = clnt,
1702                 .rpc_message = &msg,
1703                 .callback_ops = call_ops,
1704                 .callback_data = data,
1705                 .workqueue = nfsiod_workqueue,
1706                 .flags = RPC_TASK_ASYNC | flags,
1707                 .priority = priority,
1708         };
1709
1710         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1711                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1712
1713         /* Set up the initial task struct.  */
1714         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1715         trace_nfs_initiate_commit(data);
1716
1717         dprintk("NFS: initiated commit call\n");
1718
1719         task = rpc_run_task(&task_setup_data);
1720         if (IS_ERR(task))
1721                 return PTR_ERR(task);
1722         if (how & FLUSH_SYNC)
1723                 rpc_wait_for_completion_task(task);
1724         rpc_put_task(task);
1725         return 0;
1726 }
1727 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1728
1729 static loff_t nfs_get_lwb(struct list_head *head)
1730 {
1731         loff_t lwb = 0;
1732         struct nfs_page *req;
1733
1734         list_for_each_entry(req, head, wb_list)
1735                 if (lwb < (req_offset(req) + req->wb_bytes))
1736                         lwb = req_offset(req) + req->wb_bytes;
1737
1738         return lwb;
1739 }
1740
1741 /*
1742  * Set up the argument/result storage required for the RPC call.
1743  */
1744 void nfs_init_commit(struct nfs_commit_data *data,
1745                      struct list_head *head,
1746                      struct pnfs_layout_segment *lseg,
1747                      struct nfs_commit_info *cinfo)
1748 {
1749         struct nfs_page *first;
1750         struct nfs_open_context *ctx;
1751         struct inode *inode;
1752
1753         /* Set up the RPC argument and reply structs
1754          * NB: take care not to mess about with data->commit et al. */
1755
1756         if (head)
1757                 list_splice_init(head, &data->pages);
1758
1759         first = nfs_list_entry(data->pages.next);
1760         ctx = nfs_req_openctx(first);
1761         inode = d_inode(ctx->dentry);
1762
1763         data->inode       = inode;
1764         data->cred        = ctx->cred;
1765         data->lseg        = lseg; /* reference transferred */
1766         /* only set lwb for pnfs commit */
1767         if (lseg)
1768                 data->lwb = nfs_get_lwb(&data->pages);
1769         data->mds_ops     = &nfs_commit_ops;
1770         data->completion_ops = cinfo->completion_ops;
1771         data->dreq        = cinfo->dreq;
1772
1773         data->args.fh     = NFS_FH(data->inode);
1774         /* Note: we always request a commit of the entire inode */
1775         data->args.offset = 0;
1776         data->args.count  = 0;
1777         data->context     = get_nfs_open_context(ctx);
1778         data->res.fattr   = &data->fattr;
1779         data->res.verf    = &data->verf;
1780         nfs_fattr_init(&data->fattr);
1781         nfs_commit_begin(cinfo->mds);
1782 }
1783 EXPORT_SYMBOL_GPL(nfs_init_commit);
1784
1785 void nfs_retry_commit(struct list_head *page_list,
1786                       struct pnfs_layout_segment *lseg,
1787                       struct nfs_commit_info *cinfo,
1788                       u32 ds_commit_idx)
1789 {
1790         struct nfs_page *req;
1791
1792         while (!list_empty(page_list)) {
1793                 req = nfs_list_entry(page_list->next);
1794                 nfs_list_remove_request(req);
1795                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1796                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1797                 nfs_unlock_and_release_request(req);
1798         }
1799 }
1800 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1801
1802 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1803                                      struct nfs_page *req)
1804 {
1805         struct folio *folio = nfs_page_to_folio(req);
1806
1807         filemap_dirty_folio(folio_mapping(folio), folio);
1808 }
1809
1810 /*
1811  * Commit dirty pages
1812  */
1813 static int
1814 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1815                 struct nfs_commit_info *cinfo)
1816 {
1817         struct nfs_commit_data  *data;
1818         unsigned short task_flags = 0;
1819
1820         /* another commit raced with us */
1821         if (list_empty(head))
1822                 return 0;
1823
1824         data = nfs_commitdata_alloc();
1825         if (!data) {
1826                 nfs_retry_commit(head, NULL, cinfo, -1);
1827                 return -ENOMEM;
1828         }
1829
1830         /* Set up the argument struct */
1831         nfs_init_commit(data, head, NULL, cinfo);
1832         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1833                 task_flags = RPC_TASK_MOVEABLE;
1834         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1835                                    data->mds_ops, how,
1836                                    RPC_TASK_CRED_NOREF | task_flags);
1837 }
1838
1839 /*
1840  * COMMIT call returned
1841  */
1842 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1843 {
1844         struct nfs_commit_data  *data = calldata;
1845
1846         /* Call the NFS version-specific code */
1847         NFS_PROTO(data->inode)->commit_done(task, data);
1848         trace_nfs_commit_done(task, data);
1849 }
1850
1851 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1852 {
1853         const struct nfs_writeverf *verf = data->res.verf;
1854         struct nfs_page *req;
1855         int status = data->task.tk_status;
1856         struct nfs_commit_info cinfo;
1857         struct nfs_server *nfss;
1858         struct folio *folio;
1859
1860         while (!list_empty(&data->pages)) {
1861                 req = nfs_list_entry(data->pages.next);
1862                 nfs_list_remove_request(req);
1863                 folio = nfs_page_to_folio(req);
1864                 nfs_folio_clear_commit(folio);
1865
1866                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1867                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1868                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1869                         req->wb_bytes,
1870                         (long long)req_offset(req));
1871                 if (status < 0) {
1872                         if (folio) {
1873                                 trace_nfs_commit_error(data->inode, req,
1874                                                        status);
1875                                 nfs_mapping_set_error(folio, status);
1876                                 nfs_inode_remove_request(req);
1877                         }
1878                         dprintk_cont(", error = %d\n", status);
1879                         goto next;
1880                 }
1881
1882                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1883                  * returned by the server against all stored verfs. */
1884                 if (nfs_write_match_verf(verf, req)) {
1885                         /* We have a match */
1886                         if (folio)
1887                                 nfs_inode_remove_request(req);
1888                         dprintk_cont(" OK\n");
1889                         goto next;
1890                 }
1891                 /* We have a mismatch. Write the page again */
1892                 dprintk_cont(" mismatch\n");
1893                 nfs_mark_request_dirty(req);
1894                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1895         next:
1896                 nfs_unlock_and_release_request(req);
1897                 /* Latency breaker */
1898                 cond_resched();
1899         }
1900         nfss = NFS_SERVER(data->inode);
1901         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1902                 nfss->write_congested = 0;
1903
1904         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1905         nfs_commit_end(cinfo.mds);
1906 }
1907
1908 static void nfs_commit_release(void *calldata)
1909 {
1910         struct nfs_commit_data *data = calldata;
1911
1912         data->completion_ops->completion(data);
1913         nfs_commitdata_release(calldata);
1914 }
1915
1916 static const struct rpc_call_ops nfs_commit_ops = {
1917         .rpc_call_prepare = nfs_commit_prepare,
1918         .rpc_call_done = nfs_commit_done,
1919         .rpc_release = nfs_commit_release,
1920 };
1921
1922 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1923         .completion = nfs_commit_release_pages,
1924         .resched_write = nfs_commit_resched_write,
1925 };
1926
1927 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1928                             int how, struct nfs_commit_info *cinfo)
1929 {
1930         int status;
1931
1932         status = pnfs_commit_list(inode, head, how, cinfo);
1933         if (status == PNFS_NOT_ATTEMPTED)
1934                 status = nfs_commit_list(inode, head, how, cinfo);
1935         return status;
1936 }
1937
1938 static int __nfs_commit_inode(struct inode *inode, int how,
1939                 struct writeback_control *wbc)
1940 {
1941         LIST_HEAD(head);
1942         struct nfs_commit_info cinfo;
1943         int may_wait = how & FLUSH_SYNC;
1944         int ret, nscan;
1945
1946         how &= ~FLUSH_SYNC;
1947         nfs_init_cinfo_from_inode(&cinfo, inode);
1948         nfs_commit_begin(cinfo.mds);
1949         for (;;) {
1950                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1951                 if (ret <= 0)
1952                         break;
1953                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1954                 if (ret < 0)
1955                         break;
1956                 ret = 0;
1957                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1958                         if (nscan < wbc->nr_to_write)
1959                                 wbc->nr_to_write -= nscan;
1960                         else
1961                                 wbc->nr_to_write = 0;
1962                 }
1963                 if (nscan < INT_MAX)
1964                         break;
1965                 cond_resched();
1966         }
1967         nfs_commit_end(cinfo.mds);
1968         if (ret || !may_wait)
1969                 return ret;
1970         return wait_on_commit(cinfo.mds);
1971 }
1972
1973 int nfs_commit_inode(struct inode *inode, int how)
1974 {
1975         return __nfs_commit_inode(inode, how, NULL);
1976 }
1977 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1978
1979 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1980 {
1981         struct nfs_inode *nfsi = NFS_I(inode);
1982         int flags = FLUSH_SYNC;
1983         int ret = 0;
1984
1985         if (wbc->sync_mode == WB_SYNC_NONE) {
1986                 /* no commits means nothing needs to be done */
1987                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1988                         goto check_requests_outstanding;
1989
1990                 /* Don't commit yet if this is a non-blocking flush and there
1991                  * are a lot of outstanding writes for this mapping.
1992                  */
1993                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1994                         goto out_mark_dirty;
1995
1996                 /* don't wait for the COMMIT response */
1997                 flags = 0;
1998         }
1999
2000         ret = __nfs_commit_inode(inode, flags, wbc);
2001         if (!ret) {
2002                 if (flags & FLUSH_SYNC)
2003                         return 0;
2004         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
2005                 goto out_mark_dirty;
2006
2007 check_requests_outstanding:
2008         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2009                 return ret;
2010 out_mark_dirty:
2011         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2012         return ret;
2013 }
2014 EXPORT_SYMBOL_GPL(nfs_write_inode);
2015
2016 /*
2017  * Wrapper for filemap_write_and_wait_range()
2018  *
2019  * Needed for pNFS in order to ensure data becomes visible to the
2020  * client.
2021  */
2022 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2023                 loff_t lstart, loff_t lend)
2024 {
2025         int ret;
2026
2027         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2028         if (ret == 0)
2029                 ret = pnfs_sync_inode(mapping->host, true);
2030         return ret;
2031 }
2032 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2033
2034 /*
2035  * flush the inode to disk.
2036  */
2037 int nfs_wb_all(struct inode *inode)
2038 {
2039         int ret;
2040
2041         trace_nfs_writeback_inode_enter(inode);
2042
2043         ret = filemap_write_and_wait(inode->i_mapping);
2044         if (ret)
2045                 goto out;
2046         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2047         if (ret < 0)
2048                 goto out;
2049         pnfs_sync_inode(inode, true);
2050         ret = 0;
2051
2052 out:
2053         trace_nfs_writeback_inode_exit(inode, ret);
2054         return ret;
2055 }
2056 EXPORT_SYMBOL_GPL(nfs_wb_all);
2057
2058 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2059 {
2060         struct nfs_page *req;
2061         int ret = 0;
2062
2063         folio_wait_writeback(folio);
2064
2065         /* blocking call to cancel all requests and join to a single (head)
2066          * request */
2067         req = nfs_lock_and_join_requests(folio);
2068
2069         if (IS_ERR(req)) {
2070                 ret = PTR_ERR(req);
2071         } else if (req) {
2072                 /* all requests from this folio have been cancelled by
2073                  * nfs_lock_and_join_requests, so just remove the head
2074                  * request from the inode / page_private pointer and
2075                  * release it */
2076                 nfs_inode_remove_request(req);
2077                 nfs_unlock_and_release_request(req);
2078         }
2079
2080         return ret;
2081 }
2082
2083 /**
2084  * nfs_wb_folio - Write back all requests on one page
2085  * @inode: pointer to page
2086  * @folio: pointer to folio
2087  *
2088  * Assumes that the folio has been locked by the caller, and will
2089  * not unlock it.
2090  */
2091 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2092 {
2093         loff_t range_start = folio_file_pos(folio);
2094         loff_t range_end = range_start + (loff_t)folio_size(folio) - 1;
2095         struct writeback_control wbc = {
2096                 .sync_mode = WB_SYNC_ALL,
2097                 .nr_to_write = 0,
2098                 .range_start = range_start,
2099                 .range_end = range_end,
2100         };
2101         int ret;
2102
2103         trace_nfs_writeback_folio(inode, folio);
2104
2105         for (;;) {
2106                 folio_wait_writeback(folio);
2107                 if (folio_clear_dirty_for_io(folio)) {
2108                         ret = nfs_writepage_locked(folio, &wbc);
2109                         if (ret < 0)
2110                                 goto out_error;
2111                         continue;
2112                 }
2113                 ret = 0;
2114                 if (!folio_test_private(folio))
2115                         break;
2116                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2117                 if (ret < 0)
2118                         goto out_error;
2119         }
2120 out_error:
2121         trace_nfs_writeback_folio_done(inode, folio, ret);
2122         return ret;
2123 }
2124
2125 #ifdef CONFIG_MIGRATION
2126 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2127                 struct folio *src, enum migrate_mode mode)
2128 {
2129         /*
2130          * If the private flag is set, the folio is currently associated with
2131          * an in-progress read or write request. Don't try to migrate it.
2132          *
2133          * FIXME: we could do this in principle, but we'll need a way to ensure
2134          *        that we can safely release the inode reference while holding
2135          *        the folio lock.
2136          */
2137         if (folio_test_private(src))
2138                 return -EBUSY;
2139
2140         if (folio_test_fscache(src)) {
2141                 if (mode == MIGRATE_ASYNC)
2142                         return -EBUSY;
2143                 folio_wait_fscache(src);
2144         }
2145
2146         return migrate_folio(mapping, dst, src, mode);
2147 }
2148 #endif
2149
2150 int __init nfs_init_writepagecache(void)
2151 {
2152         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2153                                              sizeof(struct nfs_pgio_header),
2154                                              0, SLAB_HWCACHE_ALIGN,
2155                                              NULL);
2156         if (nfs_wdata_cachep == NULL)
2157                 return -ENOMEM;
2158
2159         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2160                                                      nfs_wdata_cachep);
2161         if (nfs_wdata_mempool == NULL)
2162                 goto out_destroy_write_cache;
2163
2164         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2165                                              sizeof(struct nfs_commit_data),
2166                                              0, SLAB_HWCACHE_ALIGN,
2167                                              NULL);
2168         if (nfs_cdata_cachep == NULL)
2169                 goto out_destroy_write_mempool;
2170
2171         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2172                                                       nfs_cdata_cachep);
2173         if (nfs_commit_mempool == NULL)
2174                 goto out_destroy_commit_cache;
2175
2176         /*
2177          * NFS congestion size, scale with available memory.
2178          *
2179          *  64MB:    8192k
2180          * 128MB:   11585k
2181          * 256MB:   16384k
2182          * 512MB:   23170k
2183          *   1GB:   32768k
2184          *   2GB:   46340k
2185          *   4GB:   65536k
2186          *   8GB:   92681k
2187          *  16GB:  131072k
2188          *
2189          * This allows larger machines to have larger/more transfers.
2190          * Limit the default to 256M
2191          */
2192         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2193         if (nfs_congestion_kb > 256*1024)
2194                 nfs_congestion_kb = 256*1024;
2195
2196         return 0;
2197
2198 out_destroy_commit_cache:
2199         kmem_cache_destroy(nfs_cdata_cachep);
2200 out_destroy_write_mempool:
2201         mempool_destroy(nfs_wdata_mempool);
2202 out_destroy_write_cache:
2203         kmem_cache_destroy(nfs_wdata_cachep);
2204         return -ENOMEM;
2205 }
2206
2207 void nfs_destroy_writepagecache(void)
2208 {
2209         mempool_destroy(nfs_commit_mempool);
2210         kmem_cache_destroy(nfs_cdata_cachep);
2211         mempool_destroy(nfs_wdata_mempool);
2212         kmem_cache_destroy(nfs_wdata_cachep);
2213 }
2214
2215 static const struct nfs_rw_ops nfs_rw_write_ops = {
2216         .rw_alloc_header        = nfs_writehdr_alloc,
2217         .rw_free_header         = nfs_writehdr_free,
2218         .rw_done                = nfs_writeback_done,
2219         .rw_result              = nfs_writeback_result,
2220         .rw_initiate            = nfs_initiate_write,
2221 };