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_page *req);
  63 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
  64                                       struct inode *inode);
  65 static struct nfs_page *
  66 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
  67                                                 struct folio *folio);
  68
  69 static struct kmem_cache *nfs_wdata_cachep;
  70 static mempool_t *nfs_wdata_mempool;
  71 static struct kmem_cache *nfs_cdata_cachep;
  72 static mempool_t *nfs_commit_mempool;
  73
  74 struct nfs_commit_data *nfs_commitdata_alloc(void)
  75 {
  76         struct nfs_commit_data *p;
  77
  78         p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  79         if (!p) {
  80                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  81                 if (!p)
  82                         return NULL;
  83                 memset(p, 0, sizeof(*p));
  84         }
  85         INIT_LIST_HEAD(&p->pages);
  86         return p;
  87 }
  88 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  89
  90 void nfs_commit_free(struct nfs_commit_data *p)
  91 {
  92         mempool_free(p, nfs_commit_mempool);
  93 }
  94 EXPORT_SYMBOL_GPL(nfs_commit_free);
  95
  96 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  97 {
  98         struct nfs_pgio_header *p;
  99
 100         p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
 101         if (!p) {
 102                 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 103                 if (!p)
 104                         return NULL;
 105                 memset(p, 0, sizeof(*p));
 106         }
 107         p->rw_mode = FMODE_WRITE;
 108         return p;
 109 }
 110
 111 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 112 {
 113         mempool_free(hdr, nfs_wdata_mempool);
 114 }
 115
 116 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 117 {
 118         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 119 }
 120
 121 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 122                 void (*complete)(void *), void *data)
 123 {
 124         ioc->complete = complete;
 125         ioc->data = data;
 126         kref_init(&ioc->refcount);
 127 }
 128
 129 static void nfs_io_completion_release(struct kref *kref)
 130 {
 131         struct nfs_io_completion *ioc = container_of(kref,
 132                         struct nfs_io_completion, refcount);
 133         ioc->complete(ioc->data);
 134         kfree(ioc);
 135 }
 136
 137 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 138 {
 139         if (ioc != NULL)
 140                 kref_get(&ioc->refcount);
 141 }
 142
 143 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 144 {
 145         if (ioc != NULL)
 146                 kref_put(&ioc->refcount, nfs_io_completion_release);
 147 }
 148
 149 static void
 150 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 151 {
 152         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 153                 kref_get(&req->wb_kref);
 154                 atomic_long_inc(&NFS_I(inode)->nrequests);
 155         }
 156 }
 157
 158 static int
 159 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 160 {
 161         int ret;
 162
 163         if (!test_bit(PG_REMOVE, &req->wb_flags))
 164                 return 0;
 165         ret = nfs_page_group_lock(req);
 166         if (ret)
 167                 return ret;
 168         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 169                 nfs_page_set_inode_ref(req, inode);
 170         nfs_page_group_unlock(req);
 171         return 0;
 172 }
 173
 174 static struct nfs_page *nfs_folio_private_request(struct folio *folio)
 175 {
 176         return folio_get_private(folio);
 177 }
 178
 179 /**
 180  * nfs_folio_find_private_request - find head request associated with a folio
 181  * @folio: pointer to folio
 182  *
 183  * must be called while holding the inode lock.
 184  *
 185  * returns matching head request with reference held, or NULL if not found.
 186  */
 187 static struct nfs_page *nfs_folio_find_private_request(struct folio *folio)
 188 {
 189         struct address_space *mapping = folio_file_mapping(folio);
 190         struct nfs_page *req;
 191
 192         if (!folio_test_private(folio))
 193                 return NULL;
 194         spin_lock(&mapping->private_lock);
 195         req = nfs_folio_private_request(folio);
 196         if (req) {
 197                 WARN_ON_ONCE(req->wb_head != req);
 198                 kref_get(&req->wb_kref);
 199         }
 200         spin_unlock(&mapping->private_lock);
 201         return req;
 202 }
 203
 204 static struct nfs_page *nfs_folio_find_swap_request(struct folio *folio)
 205 {
 206         struct inode *inode = folio_file_mapping(folio)->host;
 207         struct nfs_inode *nfsi = NFS_I(inode);
 208         struct nfs_page *req = NULL;
 209         if (!folio_test_swapcache(folio))
 210                 return NULL;
 211         mutex_lock(&nfsi->commit_mutex);
 212         if (folio_test_swapcache(folio)) {
 213                 req = nfs_page_search_commits_for_head_request_locked(nfsi,
 214                                                                       folio);
 215                 if (req) {
 216                         WARN_ON_ONCE(req->wb_head != req);
 217                         kref_get(&req->wb_kref);
 218                 }
 219         }
 220         mutex_unlock(&nfsi->commit_mutex);
 221         return req;
 222 }
 223
 224 /**
 225  * nfs_folio_find_head_request - find head request associated with a folio
 226  * @folio: pointer to folio
 227  *
 228  * returns matching head request with reference held, or NULL if not found.
 229  */
 230 static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 231 {
 232         struct nfs_page *req;
 233
 234         req = nfs_folio_find_private_request(folio);
 235         if (!req)
 236                 req = nfs_folio_find_swap_request(folio);
 237         return req;
 238 }
 239
 240 static struct nfs_page *nfs_folio_find_and_lock_request(struct folio *folio)
 241 {
 242         struct inode *inode = folio_file_mapping(folio)->host;
 243         struct nfs_page *req, *head;
 244         int ret;
 245
 246         for (;;) {
 247                 req = nfs_folio_find_head_request(folio);
 248                 if (!req)
 249                         return req;
 250                 head = nfs_page_group_lock_head(req);
 251                 if (head != req)
 252                         nfs_release_request(req);
 253                 if (IS_ERR(head))
 254                         return head;
 255                 ret = nfs_cancel_remove_inode(head, inode);
 256                 if (ret < 0) {
 257                         nfs_unlock_and_release_request(head);
 258                         return ERR_PTR(ret);
 259                 }
 260                 /* Ensure that nobody removed the request before we locked it */
 261                 if (head == nfs_folio_private_request(folio))
 262                         break;
 263                 if (folio_test_swapcache(folio))
 264                         break;
 265                 nfs_unlock_and_release_request(head);
 266         }
 267         return head;
 268 }
 269
 270 /* Adjust the file length if we're writing beyond the end */
 271 static void nfs_grow_file(struct folio *folio, unsigned int offset,
 272                           unsigned int count)
 273 {
 274         struct inode *inode = folio_file_mapping(folio)->host;
 275         loff_t end, i_size;
 276         pgoff_t end_index;
 277
 278         spin_lock(&inode->i_lock);
 279         i_size = i_size_read(inode);
 280         end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
 281         if (i_size > 0 && folio_index(folio) < end_index)
 282                 goto out;
 283         end = folio_file_pos(folio) + (loff_t)offset + (loff_t)count;
 284         if (i_size >= end)
 285                 goto out;
 286         trace_nfs_size_grow(inode, end);
 287         i_size_write(inode, end);
 288         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 289         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 290 out:
 291         spin_unlock(&inode->i_lock);
 292         nfs_fscache_invalidate(inode, 0);
 293 }
 294
 295 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 296 static void nfs_set_pageerror(struct address_space *mapping)
 297 {
 298         struct inode *inode = mapping->host;
 299
 300         nfs_zap_mapping(mapping->host, mapping);
 301         /* Force file size revalidation */
 302         spin_lock(&inode->i_lock);
 303         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 304                                              NFS_INO_INVALID_CHANGE |
 305                                              NFS_INO_INVALID_SIZE);
 306         spin_unlock(&inode->i_lock);
 307 }
 308
 309 static void nfs_mapping_set_error(struct folio *folio, int error)
 310 {
 311         struct address_space *mapping = folio_file_mapping(folio);
 312
 313         folio_set_error(folio);
 314         filemap_set_wb_err(mapping, error);
 315         if (mapping->host)
 316                 errseq_set(&mapping->host->i_sb->s_wb_err,
 317                            error == -ENOSPC ? -ENOSPC : -EIO);
 318         nfs_set_pageerror(mapping);
 319 }
 320
 321 /*
 322  * nfs_page_group_search_locked
 323  * @head - head request of page group
 324  * @page_offset - offset into page
 325  *
 326  * Search page group with head @head to find a request that contains the
 327  * page offset @page_offset.
 328  *
 329  * Returns a pointer to the first matching nfs request, or NULL if no
 330  * match is found.
 331  *
 332  * Must be called with the page group lock held
 333  */
 334 static struct nfs_page *
 335 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 336 {
 337         struct nfs_page *req;
 338
 339         req = head;
 340         do {
 341                 if (page_offset >= req->wb_pgbase &&
 342                     page_offset < (req->wb_pgbase + req->wb_bytes))
 343                         return req;
 344
 345                 req = req->wb_this_page;
 346         } while (req != head);
 347
 348         return NULL;
 349 }
 350
 351 /*
 352  * nfs_page_group_covers_page
 353  * @head - head request of page group
 354  *
 355  * Return true if the page group with head @head covers the whole page,
 356  * returns false otherwise
 357  */
 358 static bool nfs_page_group_covers_page(struct nfs_page *req)
 359 {
 360         unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
 361         struct nfs_page *tmp;
 362         unsigned int pos = 0;
 363
 364         nfs_page_group_lock(req);
 365
 366         for (;;) {
 367                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 368                 if (!tmp)
 369                         break;
 370                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 371         }
 372
 373         nfs_page_group_unlock(req);
 374         return pos >= len;
 375 }
 376
 377 /* We can set the PG_uptodate flag if we see that a write request
 378  * covers the full page.
 379  */
 380 static void nfs_mark_uptodate(struct nfs_page *req)
 381 {
 382         struct folio *folio = nfs_page_to_folio(req);
 383
 384         if (folio_test_uptodate(folio))
 385                 return;
 386         if (!nfs_page_group_covers_page(req))
 387                 return;
 388         folio_mark_uptodate(folio);
 389 }
 390
 391 static int wb_priority(struct writeback_control *wbc)
 392 {
 393         int ret = 0;
 394
 395         if (wbc->sync_mode == WB_SYNC_ALL)
 396                 ret = FLUSH_COND_STABLE;
 397         return ret;
 398 }
 399
 400 /*
 401  * NFS congestion control
 402  */
 403
 404 int nfs_congestion_kb;
 405
 406 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 407 #define NFS_CONGESTION_OFF_THRESH       \
 408         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 409
 410 static void nfs_folio_set_writeback(struct folio *folio)
 411 {
 412         struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
 413
 414         folio_start_writeback(folio);
 415         if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
 416                 nfss->write_congested = 1;
 417 }
 418
 419 static void nfs_folio_end_writeback(struct folio *folio)
 420 {
 421         struct nfs_server *nfss = NFS_SERVER(folio_file_mapping(folio)->host);
 422
 423         folio_end_writeback(folio);
 424         if (atomic_long_dec_return(&nfss->writeback) <
 425             NFS_CONGESTION_OFF_THRESH)
 426                 nfss->write_congested = 0;
 427 }
 428
 429 static void nfs_page_end_writeback(struct nfs_page *req)
 430 {
 431         if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
 432                 nfs_unlock_request(req);
 433                 nfs_folio_end_writeback(nfs_page_to_folio(req));
 434         } else
 435                 nfs_unlock_request(req);
 436 }
 437
 438 /*
 439  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 440  *
 441  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 442  * @old_head - the old head of the list
 443  *
 444  * All subrequests must be locked and removed from all lists, so at this point
 445  * they are only "active" in this function, and possibly in nfs_wait_on_request
 446  * with a reference held by some other context.
 447  */
 448 static void
 449 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 450                                  struct nfs_page *old_head,
 451                                  struct inode *inode)
 452 {
 453         while (destroy_list) {
 454                 struct nfs_page *subreq = destroy_list;
 455
 456                 destroy_list = (subreq->wb_this_page == old_head) ?
 457                                    NULL : subreq->wb_this_page;
 458
 459                 /* Note: lock subreq in order to change subreq->wb_head */
 460                 nfs_page_set_headlock(subreq);
 461                 WARN_ON_ONCE(old_head != subreq->wb_head);
 462
 463                 /* make sure old group is not used */
 464                 subreq->wb_this_page = subreq;
 465                 subreq->wb_head = subreq;
 466
 467                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 468
 469                 /* Note: races with nfs_page_group_destroy() */
 470                 if (!kref_read(&subreq->wb_kref)) {
 471                         /* Check if we raced with nfs_page_group_destroy() */
 472                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 473                                 nfs_page_clear_headlock(subreq);
 474                                 nfs_free_request(subreq);
 475                         } else
 476                                 nfs_page_clear_headlock(subreq);
 477                         continue;
 478                 }
 479                 nfs_page_clear_headlock(subreq);
 480
 481                 nfs_release_request(old_head);
 482
 483                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 484                         nfs_release_request(subreq);
 485                         atomic_long_dec(&NFS_I(inode)->nrequests);
 486                 }
 487
 488                 /* subreq is now totally disconnected from page group or any
 489                  * write / commit lists. last chance to wake any waiters */
 490                 nfs_unlock_and_release_request(subreq);
 491         }
 492 }
 493
 494 /*
 495  * nfs_join_page_group - destroy subrequests of the head req
 496  * @head: the page used to lookup the "page group" of nfs_page structures
 497  * @inode: Inode to which the request belongs.
 498  *
 499  * This function joins all sub requests to the head request by first
 500  * locking all requests in the group, cancelling any pending operations
 501  * and finally updating the head request to cover the whole range covered by
 502  * the (former) group.  All subrequests are removed from any write or commit
 503  * lists, unlinked from the group and destroyed.
 504  */
 505 void
 506 nfs_join_page_group(struct nfs_page *head, struct inode *inode)
 507 {
 508         struct nfs_page *subreq;
 509         struct nfs_page *destroy_list = NULL;
 510         unsigned int pgbase, off, bytes;
 511
 512         pgbase = head->wb_pgbase;
 513         bytes = head->wb_bytes;
 514         off = head->wb_offset;
 515         for (subreq = head->wb_this_page; subreq != head;
 516                         subreq = subreq->wb_this_page) {
 517                 /* Subrequests should always form a contiguous range */
 518                 if (pgbase > subreq->wb_pgbase) {
 519                         off -= pgbase - subreq->wb_pgbase;
 520                         bytes += pgbase - subreq->wb_pgbase;
 521                         pgbase = subreq->wb_pgbase;
 522                 }
 523                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 524                                 - pgbase, bytes);
 525         }
 526
 527         /* Set the head request's range to cover the former page group */
 528         head->wb_pgbase = pgbase;
 529         head->wb_bytes = bytes;
 530         head->wb_offset = off;
 531
 532         /* Now that all requests are locked, make sure they aren't on any list.
 533          * Commit list removal accounting is done after locks are dropped */
 534         subreq = head;
 535         do {
 536                 nfs_clear_request_commit(subreq);
 537                 subreq = subreq->wb_this_page;
 538         } while (subreq != head);
 539
 540         /* unlink subrequests from head, destroy them later */
 541         if (head->wb_this_page != head) {
 542                 /* destroy list will be terminated by head */
 543                 destroy_list = head->wb_this_page;
 544                 head->wb_this_page = head;
 545         }
 546
 547         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 548 }
 549
 550 /*
 551  * nfs_lock_and_join_requests - join all subreqs to the head req
 552  * @folio: the folio used to lookup the "page group" of nfs_page structures
 553  *
 554  * This function joins all sub requests to the head request by first
 555  * locking all requests in the group, cancelling any pending operations
 556  * and finally updating the head request to cover the whole range covered by
 557  * the (former) group.  All subrequests are removed from any write or commit
 558  * lists, unlinked from the group and destroyed.
 559  *
 560  * Returns a locked, referenced pointer to the head request - which after
 561  * this call is guaranteed to be the only request associated with the page.
 562  * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
 563  * error was encountered.
 564  */
 565 static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 566 {
 567         struct inode *inode = folio_file_mapping(folio)->host;
 568         struct nfs_page *head;
 569         int ret;
 570
 571         /*
 572          * A reference is taken only on the head request which acts as a
 573          * reference to the whole page group - the group will not be destroyed
 574          * until the head reference is released.
 575          */
 576         head = nfs_folio_find_and_lock_request(folio);
 577         if (IS_ERR_OR_NULL(head))
 578                 return head;
 579
 580         /* lock each request in the page group */
 581         ret = nfs_page_group_lock_subrequests(head);
 582         if (ret < 0) {
 583                 nfs_unlock_and_release_request(head);
 584                 return ERR_PTR(ret);
 585         }
 586
 587         nfs_join_page_group(head, inode);
 588
 589         return head;
 590 }
 591
 592 static void nfs_write_error(struct nfs_page *req, int error)
 593 {
 594         trace_nfs_write_error(nfs_page_to_inode(req), req, error);
 595         nfs_mapping_set_error(nfs_page_to_folio(req), error);
 596         nfs_inode_remove_request(req);
 597         nfs_page_end_writeback(req);
 598         nfs_release_request(req);
 599 }
 600
 601 /*
 602  * Find an associated nfs write request, and prepare to flush it out
 603  * May return an error if the user signalled nfs_wait_on_request().
 604  */
 605 static int nfs_page_async_flush(struct folio *folio,
 606                                 struct writeback_control *wbc,
 607                                 struct nfs_pageio_descriptor *pgio)
 608 {
 609         struct nfs_page *req;
 610         int ret = 0;
 611
 612         req = nfs_lock_and_join_requests(folio);
 613         if (!req)
 614                 goto out;
 615         ret = PTR_ERR(req);
 616         if (IS_ERR(req))
 617                 goto out;
 618
 619         nfs_folio_set_writeback(folio);
 620         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 621
 622         /* If there is a fatal error that covers this write, just exit */
 623         ret = pgio->pg_error;
 624         if (nfs_error_is_fatal_on_server(ret))
 625                 goto out_launder;
 626
 627         ret = 0;
 628         if (!nfs_pageio_add_request(pgio, req)) {
 629                 ret = pgio->pg_error;
 630                 /*
 631                  * Remove the problematic req upon fatal errors on the server
 632                  */
 633                 if (nfs_error_is_fatal_on_server(ret))
 634                         goto out_launder;
 635                 if (wbc->sync_mode == WB_SYNC_NONE)
 636                         ret = AOP_WRITEPAGE_ACTIVATE;
 637                 folio_redirty_for_writepage(wbc, folio);
 638                 nfs_redirty_request(req);
 639                 pgio->pg_error = 0;
 640         } else
 641                 nfs_add_stats(folio_file_mapping(folio)->host,
 642                               NFSIOS_WRITEPAGES, 1);
 643 out:
 644         return ret;
 645 out_launder:
 646         nfs_write_error(req, ret);
 647         return 0;
 648 }
 649
 650 static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
 651                             struct nfs_pageio_descriptor *pgio)
 652 {
 653         nfs_pageio_cond_complete(pgio, folio_index(folio));
 654         return nfs_page_async_flush(folio, wbc, pgio);
 655 }
 656
 657 /*
 658  * Write an mmapped page to the server.
 659  */
 660 static int nfs_writepage_locked(struct folio *folio,
 661                                 struct writeback_control *wbc)
 662 {
 663         struct nfs_pageio_descriptor pgio;
 664         struct inode *inode = folio_file_mapping(folio)->host;
 665         int err;
 666
 667         if (wbc->sync_mode == WB_SYNC_NONE &&
 668             NFS_SERVER(inode)->write_congested)
 669                 return AOP_WRITEPAGE_ACTIVATE;
 670
 671         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 672         nfs_pageio_init_write(&pgio, inode, 0, false,
 673                               &nfs_async_write_completion_ops);
 674         err = nfs_do_writepage(folio, wbc, &pgio);
 675         pgio.pg_error = 0;
 676         nfs_pageio_complete(&pgio);
 677         return err;
 678 }
 679
 680 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 681 {
 682         struct folio *folio = page_folio(page);
 683         int ret;
 684
 685         ret = nfs_writepage_locked(folio, wbc);
 686         if (ret != AOP_WRITEPAGE_ACTIVATE)
 687                 unlock_page(page);
 688         return ret;
 689 }
 690
 691 static int nfs_writepages_callback(struct folio *folio,
 692                                    struct writeback_control *wbc, void *data)
 693 {
 694         int ret;
 695
 696         ret = nfs_do_writepage(folio, wbc, data);
 697         if (ret != AOP_WRITEPAGE_ACTIVATE)
 698                 folio_unlock(folio);
 699         return ret;
 700 }
 701
 702 static void nfs_io_completion_commit(void *inode)
 703 {
 704         nfs_commit_inode(inode, 0);
 705 }
 706
 707 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 708 {
 709         struct inode *inode = mapping->host;
 710         struct nfs_pageio_descriptor pgio;
 711         struct nfs_io_completion *ioc = NULL;
 712         unsigned int mntflags = NFS_SERVER(inode)->flags;
 713         int priority = 0;
 714         int err;
 715
 716         if (wbc->sync_mode == WB_SYNC_NONE &&
 717             NFS_SERVER(inode)->write_congested)
 718                 return 0;
 719
 720         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 721
 722         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 723             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 724                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 725                 if (ioc)
 726                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 727                                                inode);
 728                 priority = wb_priority(wbc);
 729         }
 730
 731         do {
 732                 nfs_pageio_init_write(&pgio, inode, priority, false,
 733                                       &nfs_async_write_completion_ops);
 734                 pgio.pg_io_completion = ioc;
 735                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 736                                         &pgio);
 737                 pgio.pg_error = 0;
 738                 nfs_pageio_complete(&pgio);
 739         } while (err < 0 && !nfs_error_is_fatal(err));
 740         nfs_io_completion_put(ioc);
 741
 742         if (err < 0)
 743                 goto out_err;
 744         return 0;
 745 out_err:
 746         return err;
 747 }
 748
 749 /*
 750  * Insert a write request into an inode
 751  */
 752 static void nfs_inode_add_request(struct nfs_page *req)
 753 {
 754         struct folio *folio = nfs_page_to_folio(req);
 755         struct address_space *mapping = folio_file_mapping(folio);
 756         struct nfs_inode *nfsi = NFS_I(mapping->host);
 757
 758         WARN_ON_ONCE(req->wb_this_page != req);
 759
 760         /* Lock the request! */
 761         nfs_lock_request(req);
 762
 763         /*
 764          * Swap-space should not get truncated. Hence no need to plug the race
 765          * with invalidate/truncate.
 766          */
 767         spin_lock(&mapping->private_lock);
 768         if (likely(!folio_test_swapcache(folio))) {
 769                 set_bit(PG_MAPPED, &req->wb_flags);
 770                 folio_set_private(folio);
 771                 folio->private = req;
 772         }
 773         spin_unlock(&mapping->private_lock);
 774         atomic_long_inc(&nfsi->nrequests);
 775         /* this a head request for a page group - mark it as having an
 776          * extra reference so sub groups can follow suit.
 777          * This flag also informs pgio layer when to bump nrequests when
 778          * adding subrequests. */
 779         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 780         kref_get(&req->wb_kref);
 781 }
 782
 783 /*
 784  * Remove a write request from an inode
 785  */
 786 static void nfs_inode_remove_request(struct nfs_page *req)
 787 {
 788         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 789                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 790                 struct address_space *mapping = folio_file_mapping(folio);
 791
 792                 spin_lock(&mapping->private_lock);
 793                 if (likely(folio && !folio_test_swapcache(folio))) {
 794                         folio->private = NULL;
 795                         folio_clear_private(folio);
 796                         clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 797                 }
 798                 spin_unlock(&mapping->private_lock);
 799         }
 800
 801         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 802                 nfs_release_request(req);
 803                 atomic_long_dec(&NFS_I(nfs_page_to_inode(req))->nrequests);
 804         }
 805 }
 806
 807 static void nfs_mark_request_dirty(struct nfs_page *req)
 808 {
 809         struct folio *folio = nfs_page_to_folio(req);
 810         if (folio)
 811                 filemap_dirty_folio(folio_mapping(folio), folio);
 812 }
 813
 814 /*
 815  * nfs_page_search_commits_for_head_request_locked
 816  *
 817  * Search through commit lists on @inode for the head request for @folio.
 818  * Must be called while holding the inode (which is cinfo) lock.
 819  *
 820  * Returns the head request if found, or NULL if not found.
 821  */
 822 static struct nfs_page *
 823 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 824                                                 struct folio *folio)
 825 {
 826         struct nfs_page *freq, *t;
 827         struct nfs_commit_info cinfo;
 828         struct inode *inode = &nfsi->vfs_inode;
 829
 830         nfs_init_cinfo_from_inode(&cinfo, inode);
 831
 832         /* search through pnfs commit lists */
 833         freq = pnfs_search_commit_reqs(inode, &cinfo, folio);
 834         if (freq)
 835                 return freq->wb_head;
 836
 837         /* Linearly search the commit list for the correct request */
 838         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 839                 if (nfs_page_to_folio(freq) == folio)
 840                         return freq->wb_head;
 841         }
 842
 843         return NULL;
 844 }
 845
 846 /**
 847  * nfs_request_add_commit_list_locked - add request to a commit list
 848  * @req: pointer to a struct nfs_page
 849  * @dst: commit list head
 850  * @cinfo: holds list lock and accounting info
 851  *
 852  * This sets the PG_CLEAN bit, updates the cinfo count of
 853  * number of outstanding requests requiring a commit as well as
 854  * the MM page stats.
 855  *
 856  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 857  * nfs_page lock.
 858  */
 859 void
 860 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 861                             struct nfs_commit_info *cinfo)
 862 {
 863         set_bit(PG_CLEAN, &req->wb_flags);
 864         nfs_list_add_request(req, dst);
 865         atomic_long_inc(&cinfo->mds->ncommit);
 866 }
 867 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 868
 869 /**
 870  * nfs_request_add_commit_list - add request to a commit list
 871  * @req: pointer to a struct nfs_page
 872  * @cinfo: holds list lock and accounting info
 873  *
 874  * This sets the PG_CLEAN bit, updates the cinfo count of
 875  * number of outstanding requests requiring a commit as well as
 876  * the MM page stats.
 877  *
 878  * The caller must _not_ hold the cinfo->lock, but must be
 879  * holding the nfs_page lock.
 880  */
 881 void
 882 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 883 {
 884         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 885         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 886         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 887         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 888 }
 889 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 890
 891 /**
 892  * nfs_request_remove_commit_list - Remove request from a commit list
 893  * @req: pointer to a nfs_page
 894  * @cinfo: holds list lock and accounting info
 895  *
 896  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 897  * number of outstanding requests requiring a commit
 898  * It does not update the MM page stats.
 899  *
 900  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 901  */
 902 void
 903 nfs_request_remove_commit_list(struct nfs_page *req,
 904                                struct nfs_commit_info *cinfo)
 905 {
 906         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 907                 return;
 908         nfs_list_remove_request(req);
 909         atomic_long_dec(&cinfo->mds->ncommit);
 910 }
 911 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 912
 913 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 914                                       struct inode *inode)
 915 {
 916         cinfo->inode = inode;
 917         cinfo->mds = &NFS_I(inode)->commit_info;
 918         cinfo->ds = pnfs_get_ds_info(inode);
 919         cinfo->dreq = NULL;
 920         cinfo->completion_ops = &nfs_commit_completion_ops;
 921 }
 922
 923 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 924                     struct inode *inode,
 925                     struct nfs_direct_req *dreq)
 926 {
 927         if (dreq)
 928                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 929         else
 930                 nfs_init_cinfo_from_inode(cinfo, inode);
 931 }
 932 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 933
 934 /*
 935  * Add a request to the inode's commit list.
 936  */
 937 void
 938 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 939                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 940 {
 941         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 942                 return;
 943         nfs_request_add_commit_list(req, cinfo);
 944 }
 945
 946 static void nfs_folio_clear_commit(struct folio *folio)
 947 {
 948         if (folio) {
 949                 long nr = folio_nr_pages(folio);
 950
 951                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 952                 wb_stat_mod(&inode_to_bdi(folio_file_mapping(folio)->host)->wb,
 953                             WB_WRITEBACK, -nr);
 954         }
 955 }
 956
 957 /* Called holding the request lock on @req */
 958 static void
 959 nfs_clear_request_commit(struct nfs_page *req)
 960 {
 961         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 962                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 963                 struct inode *inode = d_inode(ctx->dentry);
 964                 struct nfs_commit_info cinfo;
 965
 966                 nfs_init_cinfo_from_inode(&cinfo, inode);
 967                 mutex_lock(&NFS_I(inode)->commit_mutex);
 968                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 969                         nfs_request_remove_commit_list(req, &cinfo);
 970                 }
 971                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 972                 nfs_folio_clear_commit(nfs_page_to_folio(req));
 973         }
 974 }
 975
 976 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 977 {
 978         if (hdr->verf.committed == NFS_DATA_SYNC)
 979                 return hdr->lseg == NULL;
 980         return hdr->verf.committed != NFS_FILE_SYNC;
 981 }
 982
 983 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 984 {
 985         nfs_io_completion_get(hdr->io_completion);
 986 }
 987
 988 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 989 {
 990         struct nfs_commit_info cinfo;
 991         unsigned long bytes = 0;
 992
 993         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 994                 goto out;
 995         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 996         while (!list_empty(&hdr->pages)) {
 997                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 998
 999                 bytes += req->wb_bytes;
1000                 nfs_list_remove_request(req);
1001                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
1002                     (hdr->good_bytes < bytes)) {
1003                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
1004                         nfs_mapping_set_error(nfs_page_to_folio(req),
1005                                               hdr->error);
1006                         goto remove_req;
1007                 }
1008                 if (nfs_write_need_commit(hdr)) {
1009                         /* Reset wb_nio, since the write was successful. */
1010                         req->wb_nio = 0;
1011                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1012                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1013                                 hdr->pgio_mirror_idx);
1014                         goto next;
1015                 }
1016 remove_req:
1017                 nfs_inode_remove_request(req);
1018 next:
1019                 nfs_page_end_writeback(req);
1020                 nfs_release_request(req);
1021         }
1022 out:
1023         nfs_io_completion_put(hdr->io_completion);
1024         hdr->release(hdr);
1025 }
1026
1027 unsigned long
1028 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1029 {
1030         return atomic_long_read(&cinfo->mds->ncommit);
1031 }
1032
1033 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1034 int
1035 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1036                      struct nfs_commit_info *cinfo, int max)
1037 {
1038         struct nfs_page *req, *tmp;
1039         int ret = 0;
1040
1041         list_for_each_entry_safe(req, tmp, src, wb_list) {
1042                 kref_get(&req->wb_kref);
1043                 if (!nfs_lock_request(req)) {
1044                         nfs_release_request(req);
1045                         continue;
1046                 }
1047                 nfs_request_remove_commit_list(req, cinfo);
1048                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1049                 nfs_list_add_request(req, dst);
1050                 ret++;
1051                 if ((ret == max) && !cinfo->dreq)
1052                         break;
1053                 cond_resched();
1054         }
1055         return ret;
1056 }
1057 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1058
1059 /*
1060  * nfs_scan_commit - Scan an inode for commit requests
1061  * @inode: NFS inode to scan
1062  * @dst: mds destination list
1063  * @cinfo: mds and ds lists of reqs ready to commit
1064  *
1065  * Moves requests from the inode's 'commit' request list.
1066  * The requests are *not* checked to ensure that they form a contiguous set.
1067  */
1068 int
1069 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1070                 struct nfs_commit_info *cinfo)
1071 {
1072         int ret = 0;
1073
1074         if (!atomic_long_read(&cinfo->mds->ncommit))
1075                 return 0;
1076         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1077         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1078                 const int max = INT_MAX;
1079
1080                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1081                                            cinfo, max);
1082                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1083         }
1084         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1085         return ret;
1086 }
1087
1088 /*
1089  * Search for an existing write request, and attempt to update
1090  * it to reflect a new dirty region on a given page.
1091  *
1092  * If the attempt fails, then the existing request is flushed out
1093  * to disk.
1094  */
1095 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1096                                                   unsigned int offset,
1097                                                   unsigned int bytes)
1098 {
1099         struct nfs_page *req;
1100         unsigned int rqend;
1101         unsigned int end;
1102         int error;
1103
1104         end = offset + bytes;
1105
1106         req = nfs_lock_and_join_requests(folio);
1107         if (IS_ERR_OR_NULL(req))
1108                 return req;
1109
1110         rqend = req->wb_offset + req->wb_bytes;
1111         /*
1112          * Tell the caller to flush out the request if
1113          * the offsets are non-contiguous.
1114          * Note: nfs_flush_incompatible() will already
1115          * have flushed out requests having wrong owners.
1116          */
1117         if (offset > rqend || end < req->wb_offset)
1118                 goto out_flushme;
1119
1120         /* Okay, the request matches. Update the region */
1121         if (offset < req->wb_offset) {
1122                 req->wb_offset = offset;
1123                 req->wb_pgbase = offset;
1124         }
1125         if (end > rqend)
1126                 req->wb_bytes = end - req->wb_offset;
1127         else
1128                 req->wb_bytes = rqend - req->wb_offset;
1129         req->wb_nio = 0;
1130         return req;
1131 out_flushme:
1132         /*
1133          * Note: we mark the request dirty here because
1134          * nfs_lock_and_join_requests() cannot preserve
1135          * commit flags, so we have to replay the write.
1136          */
1137         nfs_mark_request_dirty(req);
1138         nfs_unlock_and_release_request(req);
1139         error = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1140         return (error < 0) ? ERR_PTR(error) : NULL;
1141 }
1142
1143 /*
1144  * Try to update an existing write request, or create one if there is none.
1145  *
1146  * Note: Should always be called with the Page Lock held to prevent races
1147  * if we have to add a new request. Also assumes that the caller has
1148  * already called nfs_flush_incompatible() if necessary.
1149  */
1150 static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1151                                                 struct folio *folio,
1152                                                 unsigned int offset,
1153                                                 unsigned int bytes)
1154 {
1155         struct nfs_page *req;
1156
1157         req = nfs_try_to_update_request(folio, offset, bytes);
1158         if (req != NULL)
1159                 goto out;
1160         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1161         if (IS_ERR(req))
1162                 goto out;
1163         nfs_inode_add_request(req);
1164 out:
1165         return req;
1166 }
1167
1168 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1169                                struct folio *folio, unsigned int offset,
1170                                unsigned int count)
1171 {
1172         struct nfs_page *req;
1173
1174         req = nfs_setup_write_request(ctx, folio, offset, count);
1175         if (IS_ERR(req))
1176                 return PTR_ERR(req);
1177         /* Update file length */
1178         nfs_grow_file(folio, offset, count);
1179         nfs_mark_uptodate(req);
1180         nfs_mark_request_dirty(req);
1181         nfs_unlock_and_release_request(req);
1182         return 0;
1183 }
1184
1185 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1186 {
1187         struct nfs_open_context *ctx = nfs_file_open_context(file);
1188         struct nfs_lock_context *l_ctx;
1189         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1190         struct nfs_page *req;
1191         int do_flush, status;
1192         /*
1193          * Look for a request corresponding to this page. If there
1194          * is one, and it belongs to another file, we flush it out
1195          * before we try to copy anything into the page. Do this
1196          * due to the lack of an ACCESS-type call in NFSv2.
1197          * Also do the same if we find a request from an existing
1198          * dropped page.
1199          */
1200         do {
1201                 req = nfs_folio_find_head_request(folio);
1202                 if (req == NULL)
1203                         return 0;
1204                 l_ctx = req->wb_lock_context;
1205                 do_flush = nfs_page_to_folio(req) != folio ||
1206                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1207                 if (l_ctx && flctx &&
1208                     !(list_empty_careful(&flctx->flc_posix) &&
1209                       list_empty_careful(&flctx->flc_flock))) {
1210                         do_flush |= l_ctx->lockowner != current->files;
1211                 }
1212                 nfs_release_request(req);
1213                 if (!do_flush)
1214                         return 0;
1215                 status = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1216         } while (status == 0);
1217         return status;
1218 }
1219
1220 /*
1221  * Avoid buffered writes when a open context credential's key would
1222  * expire soon.
1223  *
1224  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1225  *
1226  * Return 0 and set a credential flag which triggers the inode to flush
1227  * and performs  NFS_FILE_SYNC writes if the key will expired within
1228  * RPC_KEY_EXPIRE_TIMEO.
1229  */
1230 int
1231 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1232 {
1233         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1234
1235         if (nfs_ctx_key_to_expire(ctx, inode) &&
1236             !rcu_access_pointer(ctx->ll_cred))
1237                 /* Already expired! */
1238                 return -EACCES;
1239         return 0;
1240 }
1241
1242 /*
1243  * Test if the open context credential key is marked to expire soon.
1244  */
1245 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1246 {
1247         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1248         struct rpc_cred *cred, *new, *old = NULL;
1249         struct auth_cred acred = {
1250                 .cred = ctx->cred,
1251         };
1252         bool ret = false;
1253
1254         rcu_read_lock();
1255         cred = rcu_dereference(ctx->ll_cred);
1256         if (cred && !(cred->cr_ops->crkey_timeout &&
1257                       cred->cr_ops->crkey_timeout(cred)))
1258                 goto out;
1259         rcu_read_unlock();
1260
1261         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1262         if (new == cred) {
1263                 put_rpccred(new);
1264                 return true;
1265         }
1266         if (IS_ERR_OR_NULL(new)) {
1267                 new = NULL;
1268                 ret = true;
1269         } else if (new->cr_ops->crkey_timeout &&
1270                    new->cr_ops->crkey_timeout(new))
1271                 ret = true;
1272
1273         rcu_read_lock();
1274         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1275                                              RCU_INITIALIZER(new)), 1);
1276 out:
1277         rcu_read_unlock();
1278         put_rpccred(old);
1279         return ret;
1280 }
1281
1282 /*
1283  * If the page cache is marked as unsafe or invalid, then we can't rely on
1284  * the PageUptodate() flag. In this case, we will need to turn off
1285  * write optimisations that depend on the page contents being correct.
1286  */
1287 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1288 {
1289         struct inode *inode = folio_file_mapping(folio)->host;
1290         struct nfs_inode *nfsi = NFS_I(inode);
1291
1292         if (nfs_have_delegated_attributes(inode))
1293                 goto out;
1294         if (nfsi->cache_validity &
1295             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1296                 return false;
1297         smp_rmb();
1298         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1299                 return false;
1300 out:
1301         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1302                 return false;
1303         return folio_test_uptodate(folio) != 0;
1304 }
1305
1306 static bool
1307 is_whole_file_wrlock(struct file_lock *fl)
1308 {
1309         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1310                         fl->fl_type == F_WRLCK;
1311 }
1312
1313 /* If we know the page is up to date, and we're not using byte range locks (or
1314  * if we have the whole file locked for writing), it may be more efficient to
1315  * extend the write to cover the entire page in order to avoid fragmentation
1316  * inefficiencies.
1317  *
1318  * If the file is opened for synchronous writes then we can just skip the rest
1319  * of the checks.
1320  */
1321 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1322                                 unsigned int pagelen)
1323 {
1324         struct inode *inode = file_inode(file);
1325         struct file_lock_context *flctx = locks_inode_context(inode);
1326         struct file_lock *fl;
1327         int ret;
1328
1329         if (file->f_flags & O_DSYNC)
1330                 return 0;
1331         if (!nfs_folio_write_uptodate(folio, pagelen))
1332                 return 0;
1333         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1334                 return 1;
1335         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1336                        list_empty_careful(&flctx->flc_posix)))
1337                 return 1;
1338
1339         /* Check to see if there are whole file write locks */
1340         ret = 0;
1341         spin_lock(&flctx->flc_lock);
1342         if (!list_empty(&flctx->flc_posix)) {
1343                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1344                                         fl_list);
1345                 if (is_whole_file_wrlock(fl))
1346                         ret = 1;
1347         } else if (!list_empty(&flctx->flc_flock)) {
1348                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1349                                         fl_list);
1350                 if (fl->fl_type == F_WRLCK)
1351                         ret = 1;
1352         }
1353         spin_unlock(&flctx->flc_lock);
1354         return ret;
1355 }
1356
1357 /*
1358  * Update and possibly write a cached page of an NFS file.
1359  *
1360  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1361  * things with a page scheduled for an RPC call (e.g. invalidate it).
1362  */
1363 int nfs_update_folio(struct file *file, struct folio *folio,
1364                      unsigned int offset, unsigned int count)
1365 {
1366         struct nfs_open_context *ctx = nfs_file_open_context(file);
1367         struct address_space *mapping = folio_file_mapping(folio);
1368         struct inode *inode = mapping->host;
1369         unsigned int pagelen = nfs_folio_length(folio);
1370         int             status = 0;
1371
1372         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1373
1374         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1375                 (long long)(folio_file_pos(folio) + offset));
1376
1377         if (!count)
1378                 goto out;
1379
1380         if (nfs_can_extend_write(file, folio, pagelen)) {
1381                 count = max(count + offset, pagelen);
1382                 offset = 0;
1383         }
1384
1385         status = nfs_writepage_setup(ctx, folio, offset, count);
1386         if (status < 0)
1387                 nfs_set_pageerror(mapping);
1388 out:
1389         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1390                         status, (long long)i_size_read(inode));
1391         return status;
1392 }
1393
1394 static int flush_task_priority(int how)
1395 {
1396         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1397                 case FLUSH_HIGHPRI:
1398                         return RPC_PRIORITY_HIGH;
1399                 case FLUSH_LOWPRI:
1400                         return RPC_PRIORITY_LOW;
1401         }
1402         return RPC_PRIORITY_NORMAL;
1403 }
1404
1405 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1406                                struct rpc_message *msg,
1407                                const struct nfs_rpc_ops *rpc_ops,
1408                                struct rpc_task_setup *task_setup_data, int how)
1409 {
1410         int priority = flush_task_priority(how);
1411
1412         if (IS_SWAPFILE(hdr->inode))
1413                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1414         task_setup_data->priority = priority;
1415         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1416         trace_nfs_initiate_write(hdr);
1417 }
1418
1419 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1420  * call this on each, which will prepare them to be retried on next
1421  * writeback using standard nfs.
1422  */
1423 static void nfs_redirty_request(struct nfs_page *req)
1424 {
1425         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1426
1427         /* Bump the transmission count */
1428         req->wb_nio++;
1429         nfs_mark_request_dirty(req);
1430         atomic_long_inc(&nfsi->redirtied_pages);
1431         nfs_page_end_writeback(req);
1432         nfs_release_request(req);
1433 }
1434
1435 static void nfs_async_write_error(struct list_head *head, int error)
1436 {
1437         struct nfs_page *req;
1438
1439         while (!list_empty(head)) {
1440                 req = nfs_list_entry(head->next);
1441                 nfs_list_remove_request(req);
1442                 if (nfs_error_is_fatal_on_server(error))
1443                         nfs_write_error(req, error);
1444                 else
1445                         nfs_redirty_request(req);
1446         }
1447 }
1448
1449 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1450 {
1451         nfs_async_write_error(&hdr->pages, 0);
1452 }
1453
1454 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1455         .init_hdr = nfs_async_write_init,
1456         .error_cleanup = nfs_async_write_error,
1457         .completion = nfs_write_completion,
1458         .reschedule_io = nfs_async_write_reschedule_io,
1459 };
1460
1461 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1462                                struct inode *inode, int ioflags, bool force_mds,
1463                                const struct nfs_pgio_completion_ops *compl_ops)
1464 {
1465         struct nfs_server *server = NFS_SERVER(inode);
1466         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1467
1468 #ifdef CONFIG_NFS_V4_1
1469         if (server->pnfs_curr_ld && !force_mds)
1470                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1471 #endif
1472         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1473                         server->wsize, ioflags);
1474 }
1475 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1476
1477 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1478 {
1479         struct nfs_pgio_mirror *mirror;
1480
1481         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1482                 pgio->pg_ops->pg_cleanup(pgio);
1483
1484         pgio->pg_ops = &nfs_pgio_rw_ops;
1485
1486         nfs_pageio_stop_mirroring(pgio);
1487
1488         mirror = &pgio->pg_mirrors[0];
1489         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1490 }
1491 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1492
1493
1494 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1495 {
1496         struct nfs_commit_data *data = calldata;
1497
1498         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1499 }
1500
1501 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1502                 struct nfs_fattr *fattr)
1503 {
1504         struct nfs_pgio_args *argp = &hdr->args;
1505         struct nfs_pgio_res *resp = &hdr->res;
1506         u64 size = argp->offset + resp->count;
1507
1508         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1509                 fattr->size = size;
1510         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1511                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1512                 return;
1513         }
1514         if (size != fattr->size)
1515                 return;
1516         /* Set attribute barrier */
1517         nfs_fattr_set_barrier(fattr);
1518         /* ...and update size */
1519         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1520 }
1521
1522 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1523 {
1524         struct nfs_fattr *fattr = &hdr->fattr;
1525         struct inode *inode = hdr->inode;
1526
1527         spin_lock(&inode->i_lock);
1528         nfs_writeback_check_extend(hdr, fattr);
1529         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1530         spin_unlock(&inode->i_lock);
1531 }
1532 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1533
1534 /*
1535  * This function is called when the WRITE call is complete.
1536  */
1537 static int nfs_writeback_done(struct rpc_task *task,
1538                               struct nfs_pgio_header *hdr,
1539                               struct inode *inode)
1540 {
1541         int status;
1542
1543         /*
1544          * ->write_done will attempt to use post-op attributes to detect
1545          * conflicting writes by other clients.  A strict interpretation
1546          * of close-to-open would allow us to continue caching even if
1547          * another writer had changed the file, but some applications
1548          * depend on tighter cache coherency when writing.
1549          */
1550         status = NFS_PROTO(inode)->write_done(task, hdr);
1551         if (status != 0)
1552                 return status;
1553
1554         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1555         trace_nfs_writeback_done(task, hdr);
1556
1557         if (task->tk_status >= 0) {
1558                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1559
1560                 if (committed == NFS_UNSTABLE) {
1561                         /*
1562                          * We have some uncommitted data on the server at
1563                          * this point, so ensure that we keep track of that
1564                          * fact irrespective of what later writes do.
1565                          */
1566                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1567                 }
1568
1569                 if (committed < hdr->args.stable) {
1570                         /* We tried a write call, but the server did not
1571                          * commit data to stable storage even though we
1572                          * requested it.
1573                          * Note: There is a known bug in Tru64 < 5.0 in which
1574                          *       the server reports NFS_DATA_SYNC, but performs
1575                          *       NFS_FILE_SYNC. We therefore implement this checking
1576                          *       as a dprintk() in order to avoid filling syslog.
1577                          */
1578                         static unsigned long    complain;
1579
1580                         /* Note this will print the MDS for a DS write */
1581                         if (time_before(complain, jiffies)) {
1582                                 dprintk("NFS:       faulty NFS server %s:"
1583                                         " (committed = %d) != (stable = %d)\n",
1584                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1585                                         committed, hdr->args.stable);
1586                                 complain = jiffies + 300 * HZ;
1587                         }
1588                 }
1589         }
1590
1591         /* Deal with the suid/sgid bit corner case */
1592         if (nfs_should_remove_suid(inode)) {
1593                 spin_lock(&inode->i_lock);
1594                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1595                 spin_unlock(&inode->i_lock);
1596         }
1597         return 0;
1598 }
1599
1600 /*
1601  * This function is called when the WRITE call is complete.
1602  */
1603 static void nfs_writeback_result(struct rpc_task *task,
1604                                  struct nfs_pgio_header *hdr)
1605 {
1606         struct nfs_pgio_args    *argp = &hdr->args;
1607         struct nfs_pgio_res     *resp = &hdr->res;
1608
1609         if (resp->count < argp->count) {
1610                 static unsigned long    complain;
1611
1612                 /* This a short write! */
1613                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1614
1615                 /* Has the server at least made some progress? */
1616                 if (resp->count == 0) {
1617                         if (time_before(complain, jiffies)) {
1618                                 printk(KERN_WARNING
1619                                        "NFS: Server wrote zero bytes, expected %u.\n",
1620                                        argp->count);
1621                                 complain = jiffies + 300 * HZ;
1622                         }
1623                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1624                         task->tk_status = -EIO;
1625                         return;
1626                 }
1627
1628                 /* For non rpc-based layout drivers, retry-through-MDS */
1629                 if (!task->tk_ops) {
1630                         hdr->pnfs_error = -EAGAIN;
1631                         return;
1632                 }
1633
1634                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1635                 if (resp->verf->committed != NFS_UNSTABLE) {
1636                         /* Resend from where the server left off */
1637                         hdr->mds_offset += resp->count;
1638                         argp->offset += resp->count;
1639                         argp->pgbase += resp->count;
1640                         argp->count -= resp->count;
1641                 } else {
1642                         /* Resend as a stable write in order to avoid
1643                          * headaches in the case of a server crash.
1644                          */
1645                         argp->stable = NFS_FILE_SYNC;
1646                 }
1647                 resp->count = 0;
1648                 resp->verf->committed = 0;
1649                 rpc_restart_call_prepare(task);
1650         }
1651 }
1652
1653 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1654 {
1655         return wait_var_event_killable(&cinfo->rpcs_out,
1656                                        !atomic_read(&cinfo->rpcs_out));
1657 }
1658
1659 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1660 {
1661         atomic_inc(&cinfo->rpcs_out);
1662 }
1663
1664 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1665 {
1666         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1667                 wake_up_var(&cinfo->rpcs_out);
1668                 return true;
1669         }
1670         return false;
1671 }
1672
1673 void nfs_commitdata_release(struct nfs_commit_data *data)
1674 {
1675         put_nfs_open_context(data->context);
1676         nfs_commit_free(data);
1677 }
1678 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1679
1680 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1681                         const struct nfs_rpc_ops *nfs_ops,
1682                         const struct rpc_call_ops *call_ops,
1683                         int how, int flags)
1684 {
1685         struct rpc_task *task;
1686         int priority = flush_task_priority(how);
1687         struct rpc_message msg = {
1688                 .rpc_argp = &data->args,
1689                 .rpc_resp = &data->res,
1690                 .rpc_cred = data->cred,
1691         };
1692         struct rpc_task_setup task_setup_data = {
1693                 .task = &data->task,
1694                 .rpc_client = clnt,
1695                 .rpc_message = &msg,
1696                 .callback_ops = call_ops,
1697                 .callback_data = data,
1698                 .workqueue = nfsiod_workqueue,
1699                 .flags = RPC_TASK_ASYNC | flags,
1700                 .priority = priority,
1701         };
1702
1703         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1704                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1705
1706         /* Set up the initial task struct.  */
1707         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1708         trace_nfs_initiate_commit(data);
1709
1710         dprintk("NFS: initiated commit call\n");
1711
1712         task = rpc_run_task(&task_setup_data);
1713         if (IS_ERR(task))
1714                 return PTR_ERR(task);
1715         if (how & FLUSH_SYNC)
1716                 rpc_wait_for_completion_task(task);
1717         rpc_put_task(task);
1718         return 0;
1719 }
1720 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1721
1722 static loff_t nfs_get_lwb(struct list_head *head)
1723 {
1724         loff_t lwb = 0;
1725         struct nfs_page *req;
1726
1727         list_for_each_entry(req, head, wb_list)
1728                 if (lwb < (req_offset(req) + req->wb_bytes))
1729                         lwb = req_offset(req) + req->wb_bytes;
1730
1731         return lwb;
1732 }
1733
1734 /*
1735  * Set up the argument/result storage required for the RPC call.
1736  */
1737 void nfs_init_commit(struct nfs_commit_data *data,
1738                      struct list_head *head,
1739                      struct pnfs_layout_segment *lseg,
1740                      struct nfs_commit_info *cinfo)
1741 {
1742         struct nfs_page *first;
1743         struct nfs_open_context *ctx;
1744         struct inode *inode;
1745
1746         /* Set up the RPC argument and reply structs
1747          * NB: take care not to mess about with data->commit et al. */
1748
1749         if (head)
1750                 list_splice_init(head, &data->pages);
1751
1752         first = nfs_list_entry(data->pages.next);
1753         ctx = nfs_req_openctx(first);
1754         inode = d_inode(ctx->dentry);
1755
1756         data->inode       = inode;
1757         data->cred        = ctx->cred;
1758         data->lseg        = lseg; /* reference transferred */
1759         /* only set lwb for pnfs commit */
1760         if (lseg)
1761                 data->lwb = nfs_get_lwb(&data->pages);
1762         data->mds_ops     = &nfs_commit_ops;
1763         data->completion_ops = cinfo->completion_ops;
1764         data->dreq        = cinfo->dreq;
1765
1766         data->args.fh     = NFS_FH(data->inode);
1767         /* Note: we always request a commit of the entire inode */
1768         data->args.offset = 0;
1769         data->args.count  = 0;
1770         data->context     = get_nfs_open_context(ctx);
1771         data->res.fattr   = &data->fattr;
1772         data->res.verf    = &data->verf;
1773         nfs_fattr_init(&data->fattr);
1774         nfs_commit_begin(cinfo->mds);
1775 }
1776 EXPORT_SYMBOL_GPL(nfs_init_commit);
1777
1778 void nfs_retry_commit(struct list_head *page_list,
1779                       struct pnfs_layout_segment *lseg,
1780                       struct nfs_commit_info *cinfo,
1781                       u32 ds_commit_idx)
1782 {
1783         struct nfs_page *req;
1784
1785         while (!list_empty(page_list)) {
1786                 req = nfs_list_entry(page_list->next);
1787                 nfs_list_remove_request(req);
1788                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1789                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1790                 nfs_unlock_and_release_request(req);
1791         }
1792 }
1793 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1794
1795 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1796                                      struct nfs_page *req)
1797 {
1798         struct folio *folio = nfs_page_to_folio(req);
1799
1800         filemap_dirty_folio(folio_mapping(folio), folio);
1801 }
1802
1803 /*
1804  * Commit dirty pages
1805  */
1806 static int
1807 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1808                 struct nfs_commit_info *cinfo)
1809 {
1810         struct nfs_commit_data  *data;
1811         unsigned short task_flags = 0;
1812
1813         /* another commit raced with us */
1814         if (list_empty(head))
1815                 return 0;
1816
1817         data = nfs_commitdata_alloc();
1818         if (!data) {
1819                 nfs_retry_commit(head, NULL, cinfo, -1);
1820                 return -ENOMEM;
1821         }
1822
1823         /* Set up the argument struct */
1824         nfs_init_commit(data, head, NULL, cinfo);
1825         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1826                 task_flags = RPC_TASK_MOVEABLE;
1827         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1828                                    data->mds_ops, how,
1829                                    RPC_TASK_CRED_NOREF | task_flags);
1830 }
1831
1832 /*
1833  * COMMIT call returned
1834  */
1835 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1836 {
1837         struct nfs_commit_data  *data = calldata;
1838
1839         /* Call the NFS version-specific code */
1840         NFS_PROTO(data->inode)->commit_done(task, data);
1841         trace_nfs_commit_done(task, data);
1842 }
1843
1844 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1845 {
1846         const struct nfs_writeverf *verf = data->res.verf;
1847         struct nfs_page *req;
1848         int status = data->task.tk_status;
1849         struct nfs_commit_info cinfo;
1850         struct nfs_server *nfss;
1851         struct folio *folio;
1852
1853         while (!list_empty(&data->pages)) {
1854                 req = nfs_list_entry(data->pages.next);
1855                 nfs_list_remove_request(req);
1856                 folio = nfs_page_to_folio(req);
1857                 nfs_folio_clear_commit(folio);
1858
1859                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1860                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1861                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1862                         req->wb_bytes,
1863                         (long long)req_offset(req));
1864                 if (status < 0) {
1865                         if (folio) {
1866                                 trace_nfs_commit_error(data->inode, req,
1867                                                        status);
1868                                 nfs_mapping_set_error(folio, status);
1869                                 nfs_inode_remove_request(req);
1870                         }
1871                         dprintk_cont(", error = %d\n", status);
1872                         goto next;
1873                 }
1874
1875                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1876                  * returned by the server against all stored verfs. */
1877                 if (nfs_write_match_verf(verf, req)) {
1878                         /* We have a match */
1879                         if (folio)
1880                                 nfs_inode_remove_request(req);
1881                         dprintk_cont(" OK\n");
1882                         goto next;
1883                 }
1884                 /* We have a mismatch. Write the page again */
1885                 dprintk_cont(" mismatch\n");
1886                 nfs_mark_request_dirty(req);
1887                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1888         next:
1889                 nfs_unlock_and_release_request(req);
1890                 /* Latency breaker */
1891                 cond_resched();
1892         }
1893         nfss = NFS_SERVER(data->inode);
1894         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1895                 nfss->write_congested = 0;
1896
1897         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1898         nfs_commit_end(cinfo.mds);
1899 }
1900
1901 static void nfs_commit_release(void *calldata)
1902 {
1903         struct nfs_commit_data *data = calldata;
1904
1905         data->completion_ops->completion(data);
1906         nfs_commitdata_release(calldata);
1907 }
1908
1909 static const struct rpc_call_ops nfs_commit_ops = {
1910         .rpc_call_prepare = nfs_commit_prepare,
1911         .rpc_call_done = nfs_commit_done,
1912         .rpc_release = nfs_commit_release,
1913 };
1914
1915 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1916         .completion = nfs_commit_release_pages,
1917         .resched_write = nfs_commit_resched_write,
1918 };
1919
1920 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1921                             int how, struct nfs_commit_info *cinfo)
1922 {
1923         int status;
1924
1925         status = pnfs_commit_list(inode, head, how, cinfo);
1926         if (status == PNFS_NOT_ATTEMPTED)
1927                 status = nfs_commit_list(inode, head, how, cinfo);
1928         return status;
1929 }
1930
1931 static int __nfs_commit_inode(struct inode *inode, int how,
1932                 struct writeback_control *wbc)
1933 {
1934         LIST_HEAD(head);
1935         struct nfs_commit_info cinfo;
1936         int may_wait = how & FLUSH_SYNC;
1937         int ret, nscan;
1938
1939         how &= ~FLUSH_SYNC;
1940         nfs_init_cinfo_from_inode(&cinfo, inode);
1941         nfs_commit_begin(cinfo.mds);
1942         for (;;) {
1943                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1944                 if (ret <= 0)
1945                         break;
1946                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1947                 if (ret < 0)
1948                         break;
1949                 ret = 0;
1950                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1951                         if (nscan < wbc->nr_to_write)
1952                                 wbc->nr_to_write -= nscan;
1953                         else
1954                                 wbc->nr_to_write = 0;
1955                 }
1956                 if (nscan < INT_MAX)
1957                         break;
1958                 cond_resched();
1959         }
1960         nfs_commit_end(cinfo.mds);
1961         if (ret || !may_wait)
1962                 return ret;
1963         return wait_on_commit(cinfo.mds);
1964 }
1965
1966 int nfs_commit_inode(struct inode *inode, int how)
1967 {
1968         return __nfs_commit_inode(inode, how, NULL);
1969 }
1970 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1971
1972 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1973 {
1974         struct nfs_inode *nfsi = NFS_I(inode);
1975         int flags = FLUSH_SYNC;
1976         int ret = 0;
1977
1978         if (wbc->sync_mode == WB_SYNC_NONE) {
1979                 /* no commits means nothing needs to be done */
1980                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1981                         goto check_requests_outstanding;
1982
1983                 /* Don't commit yet if this is a non-blocking flush and there
1984                  * are a lot of outstanding writes for this mapping.
1985                  */
1986                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1987                         goto out_mark_dirty;
1988
1989                 /* don't wait for the COMMIT response */
1990                 flags = 0;
1991         }
1992
1993         ret = __nfs_commit_inode(inode, flags, wbc);
1994         if (!ret) {
1995                 if (flags & FLUSH_SYNC)
1996                         return 0;
1997         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1998                 goto out_mark_dirty;
1999
2000 check_requests_outstanding:
2001         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2002                 return ret;
2003 out_mark_dirty:
2004         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2005         return ret;
2006 }
2007 EXPORT_SYMBOL_GPL(nfs_write_inode);
2008
2009 /*
2010  * Wrapper for filemap_write_and_wait_range()
2011  *
2012  * Needed for pNFS in order to ensure data becomes visible to the
2013  * client.
2014  */
2015 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2016                 loff_t lstart, loff_t lend)
2017 {
2018         int ret;
2019
2020         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2021         if (ret == 0)
2022                 ret = pnfs_sync_inode(mapping->host, true);
2023         return ret;
2024 }
2025 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2026
2027 /*
2028  * flush the inode to disk.
2029  */
2030 int nfs_wb_all(struct inode *inode)
2031 {
2032         int ret;
2033
2034         trace_nfs_writeback_inode_enter(inode);
2035
2036         ret = filemap_write_and_wait(inode->i_mapping);
2037         if (ret)
2038                 goto out;
2039         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2040         if (ret < 0)
2041                 goto out;
2042         pnfs_sync_inode(inode, true);
2043         ret = 0;
2044
2045 out:
2046         trace_nfs_writeback_inode_exit(inode, ret);
2047         return ret;
2048 }
2049 EXPORT_SYMBOL_GPL(nfs_wb_all);
2050
2051 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2052 {
2053         struct nfs_page *req;
2054         int ret = 0;
2055
2056         folio_wait_writeback(folio);
2057
2058         /* blocking call to cancel all requests and join to a single (head)
2059          * request */
2060         req = nfs_lock_and_join_requests(folio);
2061
2062         if (IS_ERR(req)) {
2063                 ret = PTR_ERR(req);
2064         } else if (req) {
2065                 /* all requests from this folio have been cancelled by
2066                  * nfs_lock_and_join_requests, so just remove the head
2067                  * request from the inode / page_private pointer and
2068                  * release it */
2069                 nfs_inode_remove_request(req);
2070                 nfs_unlock_and_release_request(req);
2071         }
2072
2073         return ret;
2074 }
2075
2076 /**
2077  * nfs_wb_folio - Write back all requests on one page
2078  * @inode: pointer to page
2079  * @folio: pointer to folio
2080  *
2081  * Assumes that the folio has been locked by the caller, and will
2082  * not unlock it.
2083  */
2084 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2085 {
2086         loff_t range_start = folio_file_pos(folio);
2087         loff_t range_end = range_start + (loff_t)folio_size(folio) - 1;
2088         struct writeback_control wbc = {
2089                 .sync_mode = WB_SYNC_ALL,
2090                 .nr_to_write = 0,
2091                 .range_start = range_start,
2092                 .range_end = range_end,
2093         };
2094         int ret;
2095
2096         trace_nfs_writeback_folio(inode, folio);
2097
2098         for (;;) {
2099                 folio_wait_writeback(folio);
2100                 if (folio_clear_dirty_for_io(folio)) {
2101                         ret = nfs_writepage_locked(folio, &wbc);
2102                         if (ret < 0)
2103                                 goto out_error;
2104                         continue;
2105                 }
2106                 ret = 0;
2107                 if (!folio_test_private(folio))
2108                         break;
2109                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2110                 if (ret < 0)
2111                         goto out_error;
2112         }
2113 out_error:
2114         trace_nfs_writeback_folio_done(inode, folio, ret);
2115         return ret;
2116 }
2117
2118 #ifdef CONFIG_MIGRATION
2119 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2120                 struct folio *src, enum migrate_mode mode)
2121 {
2122         /*
2123          * If the private flag is set, the folio is currently associated with
2124          * an in-progress read or write request. Don't try to migrate it.
2125          *
2126          * FIXME: we could do this in principle, but we'll need a way to ensure
2127          *        that we can safely release the inode reference while holding
2128          *        the folio lock.
2129          */
2130         if (folio_test_private(src))
2131                 return -EBUSY;
2132
2133         if (folio_test_fscache(src)) {
2134                 if (mode == MIGRATE_ASYNC)
2135                         return -EBUSY;
2136                 folio_wait_fscache(src);
2137         }
2138
2139         return migrate_folio(mapping, dst, src, mode);
2140 }
2141 #endif
2142
2143 int __init nfs_init_writepagecache(void)
2144 {
2145         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2146                                              sizeof(struct nfs_pgio_header),
2147                                              0, SLAB_HWCACHE_ALIGN,
2148                                              NULL);
2149         if (nfs_wdata_cachep == NULL)
2150                 return -ENOMEM;
2151
2152         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2153                                                      nfs_wdata_cachep);
2154         if (nfs_wdata_mempool == NULL)
2155                 goto out_destroy_write_cache;
2156
2157         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2158                                              sizeof(struct nfs_commit_data),
2159                                              0, SLAB_HWCACHE_ALIGN,
2160                                              NULL);
2161         if (nfs_cdata_cachep == NULL)
2162                 goto out_destroy_write_mempool;
2163
2164         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2165                                                       nfs_cdata_cachep);
2166         if (nfs_commit_mempool == NULL)
2167                 goto out_destroy_commit_cache;
2168
2169         /*
2170          * NFS congestion size, scale with available memory.
2171          *
2172          *  64MB:    8192k
2173          * 128MB:   11585k
2174          * 256MB:   16384k
2175          * 512MB:   23170k
2176          *   1GB:   32768k
2177          *   2GB:   46340k
2178          *   4GB:   65536k
2179          *   8GB:   92681k
2180          *  16GB:  131072k
2181          *
2182          * This allows larger machines to have larger/more transfers.
2183          * Limit the default to 256M
2184          */
2185         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2186         if (nfs_congestion_kb > 256*1024)
2187                 nfs_congestion_kb = 256*1024;
2188
2189         return 0;
2190
2191 out_destroy_commit_cache:
2192         kmem_cache_destroy(nfs_cdata_cachep);
2193 out_destroy_write_mempool:
2194         mempool_destroy(nfs_wdata_mempool);
2195 out_destroy_write_cache:
2196         kmem_cache_destroy(nfs_wdata_cachep);
2197         return -ENOMEM;
2198 }
2199
2200 void nfs_destroy_writepagecache(void)
2201 {
2202         mempool_destroy(nfs_commit_mempool);
2203         kmem_cache_destroy(nfs_cdata_cachep);
2204         mempool_destroy(nfs_wdata_mempool);
2205         kmem_cache_destroy(nfs_wdata_cachep);
2206 }
2207
2208 static const struct nfs_rw_ops nfs_rw_write_ops = {
2209         .rw_alloc_header        = nfs_writehdr_alloc,
2210         .rw_free_header         = nfs_writehdr_free,
2211         .rw_done                = nfs_writeback_done,
2212         .rw_result              = nfs_writeback_result,
2213         .rw_initiate            = nfs_initiate_write,
2214 };