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