fs/nfs/write.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/write.c
   4  *
   5  * Write file data over NFS.
   6  *
   7  * Copyright (C) 1996, 1997, Olaf Kirch <[email protected]>
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/mm.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/file.h>
  15 #include <linux/writeback.h>
  16 #include <linux/swap.h>
  17 #include <linux/migrate.h>
  18
  19 #include <linux/sunrpc/clnt.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/nfs_page.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/export.h>
  25 #include <linux/freezer.h>
  26 #include <linux/wait.h>
  27 #include <linux/iversion.h>
  28 #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->i_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->i_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 static int nfs_writepages_callback(struct folio *folio,
 684                                    struct writeback_control *wbc, void *data)
 685 {
 686         int ret;
 687
 688         ret = nfs_do_writepage(folio, wbc, data);
 689         if (ret != AOP_WRITEPAGE_ACTIVATE)
 690                 folio_unlock(folio);
 691         return ret;
 692 }
 693
 694 static void nfs_io_completion_commit(void *inode)
 695 {
 696         nfs_commit_inode(inode, 0);
 697 }
 698
 699 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 700 {
 701         struct inode *inode = mapping->host;
 702         struct nfs_pageio_descriptor pgio;
 703         struct nfs_io_completion *ioc = NULL;
 704         unsigned int mntflags = NFS_SERVER(inode)->flags;
 705         int priority = 0;
 706         int err;
 707
 708         if (wbc->sync_mode == WB_SYNC_NONE &&
 709             NFS_SERVER(inode)->write_congested)
 710                 return 0;
 711
 712         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 713
 714         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 715             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 716                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 717                 if (ioc)
 718                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 719                                                inode);
 720                 priority = wb_priority(wbc);
 721         }
 722
 723         do {
 724                 nfs_pageio_init_write(&pgio, inode, priority, false,
 725                                       &nfs_async_write_completion_ops);
 726                 pgio.pg_io_completion = ioc;
 727                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 728                                         &pgio);
 729                 pgio.pg_error = 0;
 730                 nfs_pageio_complete(&pgio);
 731                 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
 732                         break;
 733         } while (err < 0 && !nfs_error_is_fatal(err));
 734         nfs_io_completion_put(ioc);
 735
 736         if (err < 0)
 737                 goto out_err;
 738         return 0;
 739 out_err:
 740         return err;
 741 }
 742
 743 /*
 744  * Insert a write request into an inode
 745  */
 746 static void nfs_inode_add_request(struct nfs_page *req)
 747 {
 748         struct folio *folio = nfs_page_to_folio(req);
 749         struct address_space *mapping = folio_file_mapping(folio);
 750         struct nfs_inode *nfsi = NFS_I(mapping->host);
 751
 752         WARN_ON_ONCE(req->wb_this_page != req);
 753
 754         /* Lock the request! */
 755         nfs_lock_request(req);
 756
 757         /*
 758          * Swap-space should not get truncated. Hence no need to plug the race
 759          * with invalidate/truncate.
 760          */
 761         spin_lock(&mapping->i_private_lock);
 762         if (likely(!folio_test_swapcache(folio))) {
 763                 set_bit(PG_MAPPED, &req->wb_flags);
 764                 folio_set_private(folio);
 765                 folio->private = req;
 766         }
 767         spin_unlock(&mapping->i_private_lock);
 768         atomic_long_inc(&nfsi->nrequests);
 769         /* this a head request for a page group - mark it as having an
 770          * extra reference so sub groups can follow suit.
 771          * This flag also informs pgio layer when to bump nrequests when
 772          * adding subrequests. */
 773         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 774         kref_get(&req->wb_kref);
 775 }
 776
 777 /*
 778  * Remove a write request from an inode
 779  */
 780 static void nfs_inode_remove_request(struct nfs_page *req)
 781 {
 782         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 783
 784         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 785                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 786                 struct address_space *mapping = folio_file_mapping(folio);
 787
 788                 spin_lock(&mapping->i_private_lock);
 789                 if (likely(folio && !folio_test_swapcache(folio))) {
 790                         folio->private = NULL;
 791                         folio_clear_private(folio);
 792                         clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 793                 }
 794                 spin_unlock(&mapping->i_private_lock);
 795         }
 796
 797         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 798                 atomic_long_dec(&nfsi->nrequests);
 799                 nfs_release_request(req);
 800         }
 801 }
 802
 803 static void nfs_mark_request_dirty(struct nfs_page *req)
 804 {
 805         struct folio *folio = nfs_page_to_folio(req);
 806         if (folio)
 807                 filemap_dirty_folio(folio_mapping(folio), folio);
 808 }
 809
 810 /*
 811  * nfs_page_search_commits_for_head_request_locked
 812  *
 813  * Search through commit lists on @inode for the head request for @folio.
 814  * Must be called while holding the inode (which is cinfo) lock.
 815  *
 816  * Returns the head request if found, or NULL if not found.
 817  */
 818 static struct nfs_page *
 819 nfs_page_search_commits_for_head_request_locked(struct nfs_inode *nfsi,
 820                                                 struct folio *folio)
 821 {
 822         struct nfs_page *freq, *t;
 823         struct nfs_commit_info cinfo;
 824         struct inode *inode = &nfsi->vfs_inode;
 825
 826         nfs_init_cinfo_from_inode(&cinfo, inode);
 827
 828         /* search through pnfs commit lists */
 829         freq = pnfs_search_commit_reqs(inode, &cinfo, folio);
 830         if (freq)
 831                 return freq->wb_head;
 832
 833         /* Linearly search the commit list for the correct request */
 834         list_for_each_entry_safe(freq, t, &cinfo.mds->list, wb_list) {
 835                 if (nfs_page_to_folio(freq) == folio)
 836                         return freq->wb_head;
 837         }
 838
 839         return NULL;
 840 }
 841
 842 /**
 843  * nfs_request_add_commit_list_locked - add request to a commit list
 844  * @req: pointer to a struct nfs_page
 845  * @dst: commit list head
 846  * @cinfo: holds list lock and accounting info
 847  *
 848  * This sets the PG_CLEAN bit, updates the cinfo count of
 849  * number of outstanding requests requiring a commit as well as
 850  * the MM page stats.
 851  *
 852  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 853  * nfs_page lock.
 854  */
 855 void
 856 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 857                             struct nfs_commit_info *cinfo)
 858 {
 859         set_bit(PG_CLEAN, &req->wb_flags);
 860         nfs_list_add_request(req, dst);
 861         atomic_long_inc(&cinfo->mds->ncommit);
 862 }
 863 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 864
 865 /**
 866  * nfs_request_add_commit_list - add request to a commit list
 867  * @req: pointer to a struct nfs_page
 868  * @cinfo: holds list lock and accounting info
 869  *
 870  * This sets the PG_CLEAN bit, updates the cinfo count of
 871  * number of outstanding requests requiring a commit as well as
 872  * the MM page stats.
 873  *
 874  * The caller must _not_ hold the cinfo->lock, but must be
 875  * holding the nfs_page lock.
 876  */
 877 void
 878 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 879 {
 880         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 881         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 882         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 883         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 884 }
 885 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 886
 887 /**
 888  * nfs_request_remove_commit_list - Remove request from a commit list
 889  * @req: pointer to a nfs_page
 890  * @cinfo: holds list lock and accounting info
 891  *
 892  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 893  * number of outstanding requests requiring a commit
 894  * It does not update the MM page stats.
 895  *
 896  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 897  */
 898 void
 899 nfs_request_remove_commit_list(struct nfs_page *req,
 900                                struct nfs_commit_info *cinfo)
 901 {
 902         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 903                 return;
 904         nfs_list_remove_request(req);
 905         atomic_long_dec(&cinfo->mds->ncommit);
 906 }
 907 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 908
 909 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 910                                       struct inode *inode)
 911 {
 912         cinfo->inode = inode;
 913         cinfo->mds = &NFS_I(inode)->commit_info;
 914         cinfo->ds = pnfs_get_ds_info(inode);
 915         cinfo->dreq = NULL;
 916         cinfo->completion_ops = &nfs_commit_completion_ops;
 917 }
 918
 919 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 920                     struct inode *inode,
 921                     struct nfs_direct_req *dreq)
 922 {
 923         if (dreq)
 924                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 925         else
 926                 nfs_init_cinfo_from_inode(cinfo, inode);
 927 }
 928 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 929
 930 /*
 931  * Add a request to the inode's commit list.
 932  */
 933 void
 934 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 935                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 936 {
 937         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 938                 return;
 939         nfs_request_add_commit_list(req, cinfo);
 940 }
 941
 942 static void nfs_folio_clear_commit(struct folio *folio)
 943 {
 944         if (folio) {
 945                 long nr = folio_nr_pages(folio);
 946
 947                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 948                 wb_stat_mod(&inode_to_bdi(folio_file_mapping(folio)->host)->wb,
 949                             WB_WRITEBACK, -nr);
 950         }
 951 }
 952
 953 /* Called holding the request lock on @req */
 954 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 955                                      struct nfs_page *req)
 956 {
 957         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 958                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 959                 struct inode *inode = d_inode(ctx->dentry);
 960
 961                 mutex_lock(&NFS_I(inode)->commit_mutex);
 962                 if (!pnfs_clear_request_commit(req, cinfo)) {
 963                         nfs_request_remove_commit_list(req, cinfo);
 964                 }
 965                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 966                 nfs_folio_clear_commit(nfs_page_to_folio(req));
 967         }
 968 }
 969
 970 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 971 {
 972         if (hdr->verf.committed == NFS_DATA_SYNC)
 973                 return hdr->lseg == NULL;
 974         return hdr->verf.committed != NFS_FILE_SYNC;
 975 }
 976
 977 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 978 {
 979         nfs_io_completion_get(hdr->io_completion);
 980 }
 981
 982 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 983 {
 984         struct nfs_commit_info cinfo;
 985         unsigned long bytes = 0;
 986
 987         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 988                 goto out;
 989         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 990         while (!list_empty(&hdr->pages)) {
 991                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 992
 993                 bytes += req->wb_bytes;
 994                 nfs_list_remove_request(req);
 995                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 996                     (hdr->good_bytes < bytes)) {
 997                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
 998                         nfs_mapping_set_error(nfs_page_to_folio(req),
 999                                               hdr->error);
1000                         goto remove_req;
1001                 }
1002                 if (nfs_write_need_commit(hdr)) {
1003                         /* Reset wb_nio, since the write was successful. */
1004                         req->wb_nio = 0;
1005                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
1006                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
1007                                 hdr->pgio_mirror_idx);
1008                         goto next;
1009                 }
1010 remove_req:
1011                 nfs_inode_remove_request(req);
1012 next:
1013                 nfs_page_end_writeback(req);
1014                 nfs_release_request(req);
1015         }
1016 out:
1017         nfs_io_completion_put(hdr->io_completion);
1018         hdr->release(hdr);
1019 }
1020
1021 unsigned long
1022 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1023 {
1024         return atomic_long_read(&cinfo->mds->ncommit);
1025 }
1026
1027 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1028 int
1029 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1030                      struct nfs_commit_info *cinfo, int max)
1031 {
1032         struct nfs_page *req, *tmp;
1033         int ret = 0;
1034
1035         list_for_each_entry_safe(req, tmp, src, wb_list) {
1036                 kref_get(&req->wb_kref);
1037                 if (!nfs_lock_request(req)) {
1038                         nfs_release_request(req);
1039                         continue;
1040                 }
1041                 nfs_request_remove_commit_list(req, cinfo);
1042                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1043                 nfs_list_add_request(req, dst);
1044                 ret++;
1045                 if ((ret == max) && !cinfo->dreq)
1046                         break;
1047                 cond_resched();
1048         }
1049         return ret;
1050 }
1051 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1052
1053 /*
1054  * nfs_scan_commit - Scan an inode for commit requests
1055  * @inode: NFS inode to scan
1056  * @dst: mds destination list
1057  * @cinfo: mds and ds lists of reqs ready to commit
1058  *
1059  * Moves requests from the inode's 'commit' request list.
1060  * The requests are *not* checked to ensure that they form a contiguous set.
1061  */
1062 int
1063 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1064                 struct nfs_commit_info *cinfo)
1065 {
1066         int ret = 0;
1067
1068         if (!atomic_long_read(&cinfo->mds->ncommit))
1069                 return 0;
1070         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1071         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1072                 const int max = INT_MAX;
1073
1074                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1075                                            cinfo, max);
1076                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1077         }
1078         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1079         return ret;
1080 }
1081
1082 /*
1083  * Search for an existing write request, and attempt to update
1084  * it to reflect a new dirty region on a given page.
1085  *
1086  * If the attempt fails, then the existing request is flushed out
1087  * to disk.
1088  */
1089 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1090                                                   unsigned int offset,
1091                                                   unsigned int bytes)
1092 {
1093         struct nfs_page *req;
1094         unsigned int rqend;
1095         unsigned int end;
1096         int error;
1097
1098         end = offset + bytes;
1099
1100         req = nfs_lock_and_join_requests(folio);
1101         if (IS_ERR_OR_NULL(req))
1102                 return req;
1103
1104         rqend = req->wb_offset + req->wb_bytes;
1105         /*
1106          * Tell the caller to flush out the request if
1107          * the offsets are non-contiguous.
1108          * Note: nfs_flush_incompatible() will already
1109          * have flushed out requests having wrong owners.
1110          */
1111         if (offset > rqend || end < req->wb_offset)
1112                 goto out_flushme;
1113
1114         /* Okay, the request matches. Update the region */
1115         if (offset < req->wb_offset) {
1116                 req->wb_offset = offset;
1117                 req->wb_pgbase = offset;
1118         }
1119         if (end > rqend)
1120                 req->wb_bytes = end - req->wb_offset;
1121         else
1122                 req->wb_bytes = rqend - req->wb_offset;
1123         req->wb_nio = 0;
1124         return req;
1125 out_flushme:
1126         /*
1127          * Note: we mark the request dirty here because
1128          * nfs_lock_and_join_requests() cannot preserve
1129          * commit flags, so we have to replay the write.
1130          */
1131         nfs_mark_request_dirty(req);
1132         nfs_unlock_and_release_request(req);
1133         error = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1134         return (error < 0) ? ERR_PTR(error) : NULL;
1135 }
1136
1137 /*
1138  * Try to update an existing write request, or create one if there is none.
1139  *
1140  * Note: Should always be called with the Page Lock held to prevent races
1141  * if we have to add a new request. Also assumes that the caller has
1142  * already called nfs_flush_incompatible() if necessary.
1143  */
1144 static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1145                                                 struct folio *folio,
1146                                                 unsigned int offset,
1147                                                 unsigned int bytes)
1148 {
1149         struct nfs_page *req;
1150
1151         req = nfs_try_to_update_request(folio, offset, bytes);
1152         if (req != NULL)
1153                 goto out;
1154         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1155         if (IS_ERR(req))
1156                 goto out;
1157         nfs_inode_add_request(req);
1158 out:
1159         return req;
1160 }
1161
1162 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1163                                struct folio *folio, unsigned int offset,
1164                                unsigned int count)
1165 {
1166         struct nfs_page *req;
1167
1168         req = nfs_setup_write_request(ctx, folio, offset, count);
1169         if (IS_ERR(req))
1170                 return PTR_ERR(req);
1171         /* Update file length */
1172         nfs_grow_file(folio, offset, count);
1173         nfs_mark_uptodate(req);
1174         nfs_mark_request_dirty(req);
1175         nfs_unlock_and_release_request(req);
1176         return 0;
1177 }
1178
1179 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1180 {
1181         struct nfs_open_context *ctx = nfs_file_open_context(file);
1182         struct nfs_lock_context *l_ctx;
1183         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1184         struct nfs_page *req;
1185         int do_flush, status;
1186         /*
1187          * Look for a request corresponding to this page. If there
1188          * is one, and it belongs to another file, we flush it out
1189          * before we try to copy anything into the page. Do this
1190          * due to the lack of an ACCESS-type call in NFSv2.
1191          * Also do the same if we find a request from an existing
1192          * dropped page.
1193          */
1194         do {
1195                 req = nfs_folio_find_head_request(folio);
1196                 if (req == NULL)
1197                         return 0;
1198                 l_ctx = req->wb_lock_context;
1199                 do_flush = nfs_page_to_folio(req) != folio ||
1200                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1201                 if (l_ctx && flctx &&
1202                     !(list_empty_careful(&flctx->flc_posix) &&
1203                       list_empty_careful(&flctx->flc_flock))) {
1204                         do_flush |= l_ctx->lockowner != current->files;
1205                 }
1206                 nfs_release_request(req);
1207                 if (!do_flush)
1208                         return 0;
1209                 status = nfs_wb_folio(folio_file_mapping(folio)->host, folio);
1210         } while (status == 0);
1211         return status;
1212 }
1213
1214 /*
1215  * Avoid buffered writes when a open context credential's key would
1216  * expire soon.
1217  *
1218  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1219  *
1220  * Return 0 and set a credential flag which triggers the inode to flush
1221  * and performs  NFS_FILE_SYNC writes if the key will expired within
1222  * RPC_KEY_EXPIRE_TIMEO.
1223  */
1224 int
1225 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1226 {
1227         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1228
1229         if (nfs_ctx_key_to_expire(ctx, inode) &&
1230             !rcu_access_pointer(ctx->ll_cred))
1231                 /* Already expired! */
1232                 return -EACCES;
1233         return 0;
1234 }
1235
1236 /*
1237  * Test if the open context credential key is marked to expire soon.
1238  */
1239 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1240 {
1241         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1242         struct rpc_cred *cred, *new, *old = NULL;
1243         struct auth_cred acred = {
1244                 .cred = ctx->cred,
1245         };
1246         bool ret = false;
1247
1248         rcu_read_lock();
1249         cred = rcu_dereference(ctx->ll_cred);
1250         if (cred && !(cred->cr_ops->crkey_timeout &&
1251                       cred->cr_ops->crkey_timeout(cred)))
1252                 goto out;
1253         rcu_read_unlock();
1254
1255         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1256         if (new == cred) {
1257                 put_rpccred(new);
1258                 return true;
1259         }
1260         if (IS_ERR_OR_NULL(new)) {
1261                 new = NULL;
1262                 ret = true;
1263         } else if (new->cr_ops->crkey_timeout &&
1264                    new->cr_ops->crkey_timeout(new))
1265                 ret = true;
1266
1267         rcu_read_lock();
1268         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1269                                              RCU_INITIALIZER(new)), 1);
1270 out:
1271         rcu_read_unlock();
1272         put_rpccred(old);
1273         return ret;
1274 }
1275
1276 /*
1277  * If the page cache is marked as unsafe or invalid, then we can't rely on
1278  * the PageUptodate() flag. In this case, we will need to turn off
1279  * write optimisations that depend on the page contents being correct.
1280  */
1281 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1282 {
1283         struct inode *inode = folio_file_mapping(folio)->host;
1284         struct nfs_inode *nfsi = NFS_I(inode);
1285
1286         if (nfs_have_delegated_attributes(inode))
1287                 goto out;
1288         if (nfsi->cache_validity &
1289             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1290                 return false;
1291         smp_rmb();
1292         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1293                 return false;
1294 out:
1295         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1296                 return false;
1297         return folio_test_uptodate(folio) != 0;
1298 }
1299
1300 static bool
1301 is_whole_file_wrlock(struct file_lock *fl)
1302 {
1303         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1304                         fl->fl_type == F_WRLCK;
1305 }
1306
1307 /* If we know the page is up to date, and we're not using byte range locks (or
1308  * if we have the whole file locked for writing), it may be more efficient to
1309  * extend the write to cover the entire page in order to avoid fragmentation
1310  * inefficiencies.
1311  *
1312  * If the file is opened for synchronous writes then we can just skip the rest
1313  * of the checks.
1314  */
1315 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1316                                 unsigned int pagelen)
1317 {
1318         struct inode *inode = file_inode(file);
1319         struct file_lock_context *flctx = locks_inode_context(inode);
1320         struct file_lock *fl;
1321         int ret;
1322
1323         if (file->f_flags & O_DSYNC)
1324                 return 0;
1325         if (!nfs_folio_write_uptodate(folio, pagelen))
1326                 return 0;
1327         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
1328                 return 1;
1329         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1330                        list_empty_careful(&flctx->flc_posix)))
1331                 return 1;
1332
1333         /* Check to see if there are whole file write locks */
1334         ret = 0;
1335         spin_lock(&flctx->flc_lock);
1336         if (!list_empty(&flctx->flc_posix)) {
1337                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1338                                         fl_list);
1339                 if (is_whole_file_wrlock(fl))
1340                         ret = 1;
1341         } else if (!list_empty(&flctx->flc_flock)) {
1342                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1343                                         fl_list);
1344                 if (fl->fl_type == F_WRLCK)
1345                         ret = 1;
1346         }
1347         spin_unlock(&flctx->flc_lock);
1348         return ret;
1349 }
1350
1351 /*
1352  * Update and possibly write a cached page of an NFS file.
1353  *
1354  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1355  * things with a page scheduled for an RPC call (e.g. invalidate it).
1356  */
1357 int nfs_update_folio(struct file *file, struct folio *folio,
1358                      unsigned int offset, unsigned int count)
1359 {
1360         struct nfs_open_context *ctx = nfs_file_open_context(file);
1361         struct address_space *mapping = folio_file_mapping(folio);
1362         struct inode *inode = mapping->host;
1363         unsigned int pagelen = nfs_folio_length(folio);
1364         int             status = 0;
1365
1366         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1367
1368         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1369                 (long long)(folio_file_pos(folio) + offset));
1370
1371         if (!count)
1372                 goto out;
1373
1374         if (nfs_can_extend_write(file, folio, pagelen)) {
1375                 count = max(count + offset, pagelen);
1376                 offset = 0;
1377         }
1378
1379         status = nfs_writepage_setup(ctx, folio, offset, count);
1380         if (status < 0)
1381                 nfs_set_pageerror(mapping);
1382 out:
1383         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1384                         status, (long long)i_size_read(inode));
1385         return status;
1386 }
1387
1388 static int flush_task_priority(int how)
1389 {
1390         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1391                 case FLUSH_HIGHPRI:
1392                         return RPC_PRIORITY_HIGH;
1393                 case FLUSH_LOWPRI:
1394                         return RPC_PRIORITY_LOW;
1395         }
1396         return RPC_PRIORITY_NORMAL;
1397 }
1398
1399 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1400                                struct rpc_message *msg,
1401                                const struct nfs_rpc_ops *rpc_ops,
1402                                struct rpc_task_setup *task_setup_data, int how)
1403 {
1404         int priority = flush_task_priority(how);
1405
1406         if (IS_SWAPFILE(hdr->inode))
1407                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1408         task_setup_data->priority = priority;
1409         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1410         trace_nfs_initiate_write(hdr);
1411 }
1412
1413 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1414  * call this on each, which will prepare them to be retried on next
1415  * writeback using standard nfs.
1416  */
1417 static void nfs_redirty_request(struct nfs_page *req)
1418 {
1419         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1420
1421         /* Bump the transmission count */
1422         req->wb_nio++;
1423         nfs_mark_request_dirty(req);
1424         atomic_long_inc(&nfsi->redirtied_pages);
1425         nfs_page_end_writeback(req);
1426         nfs_release_request(req);
1427 }
1428
1429 static void nfs_async_write_error(struct list_head *head, int error)
1430 {
1431         struct nfs_page *req;
1432
1433         while (!list_empty(head)) {
1434                 req = nfs_list_entry(head->next);
1435                 nfs_list_remove_request(req);
1436                 if (nfs_error_is_fatal_on_server(error))
1437                         nfs_write_error(req, error);
1438                 else
1439                         nfs_redirty_request(req);
1440         }
1441 }
1442
1443 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1444 {
1445         nfs_async_write_error(&hdr->pages, 0);
1446 }
1447
1448 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1449         .init_hdr = nfs_async_write_init,
1450         .error_cleanup = nfs_async_write_error,
1451         .completion = nfs_write_completion,
1452         .reschedule_io = nfs_async_write_reschedule_io,
1453 };
1454
1455 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1456                                struct inode *inode, int ioflags, bool force_mds,
1457                                const struct nfs_pgio_completion_ops *compl_ops)
1458 {
1459         struct nfs_server *server = NFS_SERVER(inode);
1460         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1461
1462 #ifdef CONFIG_NFS_V4_1
1463         if (server->pnfs_curr_ld && !force_mds)
1464                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1465 #endif
1466         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1467                         server->wsize, ioflags);
1468 }
1469 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1470
1471 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1472 {
1473         struct nfs_pgio_mirror *mirror;
1474
1475         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1476                 pgio->pg_ops->pg_cleanup(pgio);
1477
1478         pgio->pg_ops = &nfs_pgio_rw_ops;
1479
1480         nfs_pageio_stop_mirroring(pgio);
1481
1482         mirror = &pgio->pg_mirrors[0];
1483         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1484 }
1485 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1486
1487
1488 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1489 {
1490         struct nfs_commit_data *data = calldata;
1491
1492         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1493 }
1494
1495 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1496                 struct nfs_fattr *fattr)
1497 {
1498         struct nfs_pgio_args *argp = &hdr->args;
1499         struct nfs_pgio_res *resp = &hdr->res;
1500         u64 size = argp->offset + resp->count;
1501
1502         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1503                 fattr->size = size;
1504         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1505                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1506                 return;
1507         }
1508         if (size != fattr->size)
1509                 return;
1510         /* Set attribute barrier */
1511         nfs_fattr_set_barrier(fattr);
1512         /* ...and update size */
1513         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1514 }
1515
1516 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1517 {
1518         struct nfs_fattr *fattr = &hdr->fattr;
1519         struct inode *inode = hdr->inode;
1520
1521         spin_lock(&inode->i_lock);
1522         nfs_writeback_check_extend(hdr, fattr);
1523         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1524         spin_unlock(&inode->i_lock);
1525 }
1526 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1527
1528 /*
1529  * This function is called when the WRITE call is complete.
1530  */
1531 static int nfs_writeback_done(struct rpc_task *task,
1532                               struct nfs_pgio_header *hdr,
1533                               struct inode *inode)
1534 {
1535         int status;
1536
1537         /*
1538          * ->write_done will attempt to use post-op attributes to detect
1539          * conflicting writes by other clients.  A strict interpretation
1540          * of close-to-open would allow us to continue caching even if
1541          * another writer had changed the file, but some applications
1542          * depend on tighter cache coherency when writing.
1543          */
1544         status = NFS_PROTO(inode)->write_done(task, hdr);
1545         if (status != 0)
1546                 return status;
1547
1548         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1549         trace_nfs_writeback_done(task, hdr);
1550
1551         if (task->tk_status >= 0) {
1552                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1553
1554                 if (committed == NFS_UNSTABLE) {
1555                         /*
1556                          * We have some uncommitted data on the server at
1557                          * this point, so ensure that we keep track of that
1558                          * fact irrespective of what later writes do.
1559                          */
1560                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1561                 }
1562
1563                 if (committed < hdr->args.stable) {
1564                         /* We tried a write call, but the server did not
1565                          * commit data to stable storage even though we
1566                          * requested it.
1567                          * Note: There is a known bug in Tru64 < 5.0 in which
1568                          *       the server reports NFS_DATA_SYNC, but performs
1569                          *       NFS_FILE_SYNC. We therefore implement this checking
1570                          *       as a dprintk() in order to avoid filling syslog.
1571                          */
1572                         static unsigned long    complain;
1573
1574                         /* Note this will print the MDS for a DS write */
1575                         if (time_before(complain, jiffies)) {
1576                                 dprintk("NFS:       faulty NFS server %s:"
1577                                         " (committed = %d) != (stable = %d)\n",
1578                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1579                                         committed, hdr->args.stable);
1580                                 complain = jiffies + 300 * HZ;
1581                         }
1582                 }
1583         }
1584
1585         /* Deal with the suid/sgid bit corner case */
1586         if (nfs_should_remove_suid(inode)) {
1587                 spin_lock(&inode->i_lock);
1588                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1589                 spin_unlock(&inode->i_lock);
1590         }
1591         return 0;
1592 }
1593
1594 /*
1595  * This function is called when the WRITE call is complete.
1596  */
1597 static void nfs_writeback_result(struct rpc_task *task,
1598                                  struct nfs_pgio_header *hdr)
1599 {
1600         struct nfs_pgio_args    *argp = &hdr->args;
1601         struct nfs_pgio_res     *resp = &hdr->res;
1602
1603         if (resp->count < argp->count) {
1604                 static unsigned long    complain;
1605
1606                 /* This a short write! */
1607                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1608
1609                 /* Has the server at least made some progress? */
1610                 if (resp->count == 0) {
1611                         if (time_before(complain, jiffies)) {
1612                                 printk(KERN_WARNING
1613                                        "NFS: Server wrote zero bytes, expected %u.\n",
1614                                        argp->count);
1615                                 complain = jiffies + 300 * HZ;
1616                         }
1617                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1618                         task->tk_status = -EIO;
1619                         return;
1620                 }
1621
1622                 /* For non rpc-based layout drivers, retry-through-MDS */
1623                 if (!task->tk_ops) {
1624                         hdr->pnfs_error = -EAGAIN;
1625                         return;
1626                 }
1627
1628                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1629                 if (resp->verf->committed != NFS_UNSTABLE) {
1630                         /* Resend from where the server left off */
1631                         hdr->mds_offset += resp->count;
1632                         argp->offset += resp->count;
1633                         argp->pgbase += resp->count;
1634                         argp->count -= resp->count;
1635                 } else {
1636                         /* Resend as a stable write in order to avoid
1637                          * headaches in the case of a server crash.
1638                          */
1639                         argp->stable = NFS_FILE_SYNC;
1640                 }
1641                 resp->count = 0;
1642                 resp->verf->committed = 0;
1643                 rpc_restart_call_prepare(task);
1644         }
1645 }
1646
1647 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1648 {
1649         return wait_var_event_killable(&cinfo->rpcs_out,
1650                                        !atomic_read(&cinfo->rpcs_out));
1651 }
1652
1653 static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1654 {
1655         atomic_inc(&cinfo->rpcs_out);
1656 }
1657
1658 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1659 {
1660         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1661                 wake_up_var(&cinfo->rpcs_out);
1662                 return true;
1663         }
1664         return false;
1665 }
1666
1667 void nfs_commitdata_release(struct nfs_commit_data *data)
1668 {
1669         put_nfs_open_context(data->context);
1670         nfs_commit_free(data);
1671 }
1672 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1673
1674 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1675                         const struct nfs_rpc_ops *nfs_ops,
1676                         const struct rpc_call_ops *call_ops,
1677                         int how, int flags)
1678 {
1679         struct rpc_task *task;
1680         int priority = flush_task_priority(how);
1681         struct rpc_message msg = {
1682                 .rpc_argp = &data->args,
1683                 .rpc_resp = &data->res,
1684                 .rpc_cred = data->cred,
1685         };
1686         struct rpc_task_setup task_setup_data = {
1687                 .task = &data->task,
1688                 .rpc_client = clnt,
1689                 .rpc_message = &msg,
1690                 .callback_ops = call_ops,
1691                 .callback_data = data,
1692                 .workqueue = nfsiod_workqueue,
1693                 .flags = RPC_TASK_ASYNC | flags,
1694                 .priority = priority,
1695         };
1696
1697         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1698                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1699
1700         /* Set up the initial task struct.  */
1701         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1702         trace_nfs_initiate_commit(data);
1703
1704         dprintk("NFS: initiated commit call\n");
1705
1706         task = rpc_run_task(&task_setup_data);
1707         if (IS_ERR(task))
1708                 return PTR_ERR(task);
1709         if (how & FLUSH_SYNC)
1710                 rpc_wait_for_completion_task(task);
1711         rpc_put_task(task);
1712         return 0;
1713 }
1714 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1715
1716 static loff_t nfs_get_lwb(struct list_head *head)
1717 {
1718         loff_t lwb = 0;
1719         struct nfs_page *req;
1720
1721         list_for_each_entry(req, head, wb_list)
1722                 if (lwb < (req_offset(req) + req->wb_bytes))
1723                         lwb = req_offset(req) + req->wb_bytes;
1724
1725         return lwb;
1726 }
1727
1728 /*
1729  * Set up the argument/result storage required for the RPC call.
1730  */
1731 void nfs_init_commit(struct nfs_commit_data *data,
1732                      struct list_head *head,
1733                      struct pnfs_layout_segment *lseg,
1734                      struct nfs_commit_info *cinfo)
1735 {
1736         struct nfs_page *first;
1737         struct nfs_open_context *ctx;
1738         struct inode *inode;
1739
1740         /* Set up the RPC argument and reply structs
1741          * NB: take care not to mess about with data->commit et al. */
1742
1743         if (head)
1744                 list_splice_init(head, &data->pages);
1745
1746         first = nfs_list_entry(data->pages.next);
1747         ctx = nfs_req_openctx(first);
1748         inode = d_inode(ctx->dentry);
1749
1750         data->inode       = inode;
1751         data->cred        = ctx->cred;
1752         data->lseg        = lseg; /* reference transferred */
1753         /* only set lwb for pnfs commit */
1754         if (lseg)
1755                 data->lwb = nfs_get_lwb(&data->pages);
1756         data->mds_ops     = &nfs_commit_ops;
1757         data->completion_ops = cinfo->completion_ops;
1758         data->dreq        = cinfo->dreq;
1759
1760         data->args.fh     = NFS_FH(data->inode);
1761         /* Note: we always request a commit of the entire inode */
1762         data->args.offset = 0;
1763         data->args.count  = 0;
1764         data->context     = get_nfs_open_context(ctx);
1765         data->res.fattr   = &data->fattr;
1766         data->res.verf    = &data->verf;
1767         nfs_fattr_init(&data->fattr);
1768         nfs_commit_begin(cinfo->mds);
1769 }
1770 EXPORT_SYMBOL_GPL(nfs_init_commit);
1771
1772 void nfs_retry_commit(struct list_head *page_list,
1773                       struct pnfs_layout_segment *lseg,
1774                       struct nfs_commit_info *cinfo,
1775                       u32 ds_commit_idx)
1776 {
1777         struct nfs_page *req;
1778
1779         while (!list_empty(page_list)) {
1780                 req = nfs_list_entry(page_list->next);
1781                 nfs_list_remove_request(req);
1782                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1783                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1784                 nfs_unlock_and_release_request(req);
1785         }
1786 }
1787 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1788
1789 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1790                                      struct nfs_page *req)
1791 {
1792         struct folio *folio = nfs_page_to_folio(req);
1793
1794         filemap_dirty_folio(folio_mapping(folio), folio);
1795 }
1796
1797 /*
1798  * Commit dirty pages
1799  */
1800 static int
1801 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1802                 struct nfs_commit_info *cinfo)
1803 {
1804         struct nfs_commit_data  *data;
1805         unsigned short task_flags = 0;
1806
1807         /* another commit raced with us */
1808         if (list_empty(head))
1809                 return 0;
1810
1811         data = nfs_commitdata_alloc();
1812         if (!data) {
1813                 nfs_retry_commit(head, NULL, cinfo, -1);
1814                 return -ENOMEM;
1815         }
1816
1817         /* Set up the argument struct */
1818         nfs_init_commit(data, head, NULL, cinfo);
1819         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1820                 task_flags = RPC_TASK_MOVEABLE;
1821         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1822                                    data->mds_ops, how,
1823                                    RPC_TASK_CRED_NOREF | task_flags);
1824 }
1825
1826 /*
1827  * COMMIT call returned
1828  */
1829 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1830 {
1831         struct nfs_commit_data  *data = calldata;
1832
1833         /* Call the NFS version-specific code */
1834         NFS_PROTO(data->inode)->commit_done(task, data);
1835         trace_nfs_commit_done(task, data);
1836 }
1837
1838 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1839 {
1840         const struct nfs_writeverf *verf = data->res.verf;
1841         struct nfs_page *req;
1842         int status = data->task.tk_status;
1843         struct nfs_commit_info cinfo;
1844         struct nfs_server *nfss;
1845         struct folio *folio;
1846
1847         while (!list_empty(&data->pages)) {
1848                 req = nfs_list_entry(data->pages.next);
1849                 nfs_list_remove_request(req);
1850                 folio = nfs_page_to_folio(req);
1851                 nfs_folio_clear_commit(folio);
1852
1853                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1854                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1855                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1856                         req->wb_bytes,
1857                         (long long)req_offset(req));
1858                 if (status < 0) {
1859                         if (folio) {
1860                                 trace_nfs_commit_error(data->inode, req,
1861                                                        status);
1862                                 nfs_mapping_set_error(folio, status);
1863                                 nfs_inode_remove_request(req);
1864                         }
1865                         dprintk_cont(", error = %d\n", status);
1866                         goto next;
1867                 }
1868
1869                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1870                  * returned by the server against all stored verfs. */
1871                 if (nfs_write_match_verf(verf, req)) {
1872                         /* We have a match */
1873                         if (folio)
1874                                 nfs_inode_remove_request(req);
1875                         dprintk_cont(" OK\n");
1876                         goto next;
1877                 }
1878                 /* We have a mismatch. Write the page again */
1879                 dprintk_cont(" mismatch\n");
1880                 nfs_mark_request_dirty(req);
1881                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1882         next:
1883                 nfs_unlock_and_release_request(req);
1884                 /* Latency breaker */
1885                 cond_resched();
1886         }
1887         nfss = NFS_SERVER(data->inode);
1888         if (atomic_long_read(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
1889                 nfss->write_congested = 0;
1890
1891         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1892         nfs_commit_end(cinfo.mds);
1893 }
1894
1895 static void nfs_commit_release(void *calldata)
1896 {
1897         struct nfs_commit_data *data = calldata;
1898
1899         data->completion_ops->completion(data);
1900         nfs_commitdata_release(calldata);
1901 }
1902
1903 static const struct rpc_call_ops nfs_commit_ops = {
1904         .rpc_call_prepare = nfs_commit_prepare,
1905         .rpc_call_done = nfs_commit_done,
1906         .rpc_release = nfs_commit_release,
1907 };
1908
1909 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1910         .completion = nfs_commit_release_pages,
1911         .resched_write = nfs_commit_resched_write,
1912 };
1913
1914 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1915                             int how, struct nfs_commit_info *cinfo)
1916 {
1917         int status;
1918
1919         status = pnfs_commit_list(inode, head, how, cinfo);
1920         if (status == PNFS_NOT_ATTEMPTED)
1921                 status = nfs_commit_list(inode, head, how, cinfo);
1922         return status;
1923 }
1924
1925 static int __nfs_commit_inode(struct inode *inode, int how,
1926                 struct writeback_control *wbc)
1927 {
1928         LIST_HEAD(head);
1929         struct nfs_commit_info cinfo;
1930         int may_wait = how & FLUSH_SYNC;
1931         int ret, nscan;
1932
1933         how &= ~FLUSH_SYNC;
1934         nfs_init_cinfo_from_inode(&cinfo, inode);
1935         nfs_commit_begin(cinfo.mds);
1936         for (;;) {
1937                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1938                 if (ret <= 0)
1939                         break;
1940                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1941                 if (ret < 0)
1942                         break;
1943                 ret = 0;
1944                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1945                         if (nscan < wbc->nr_to_write)
1946                                 wbc->nr_to_write -= nscan;
1947                         else
1948                                 wbc->nr_to_write = 0;
1949                 }
1950                 if (nscan < INT_MAX)
1951                         break;
1952                 cond_resched();
1953         }
1954         nfs_commit_end(cinfo.mds);
1955         if (ret || !may_wait)
1956                 return ret;
1957         return wait_on_commit(cinfo.mds);
1958 }
1959
1960 int nfs_commit_inode(struct inode *inode, int how)
1961 {
1962         return __nfs_commit_inode(inode, how, NULL);
1963 }
1964 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1965
1966 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1967 {
1968         struct nfs_inode *nfsi = NFS_I(inode);
1969         int flags = FLUSH_SYNC;
1970         int ret = 0;
1971
1972         if (wbc->sync_mode == WB_SYNC_NONE) {
1973                 /* no commits means nothing needs to be done */
1974                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1975                         goto check_requests_outstanding;
1976
1977                 /* Don't commit yet if this is a non-blocking flush and there
1978                  * are a lot of outstanding writes for this mapping.
1979                  */
1980                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1981                         goto out_mark_dirty;
1982
1983                 /* don't wait for the COMMIT response */
1984                 flags = 0;
1985         }
1986
1987         ret = __nfs_commit_inode(inode, flags, wbc);
1988         if (!ret) {
1989                 if (flags & FLUSH_SYNC)
1990                         return 0;
1991         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1992                 goto out_mark_dirty;
1993
1994 check_requests_outstanding:
1995         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1996                 return ret;
1997 out_mark_dirty:
1998         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1999         return ret;
2000 }
2001 EXPORT_SYMBOL_GPL(nfs_write_inode);
2002
2003 /*
2004  * Wrapper for filemap_write_and_wait_range()
2005  *
2006  * Needed for pNFS in order to ensure data becomes visible to the
2007  * client.
2008  */
2009 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2010                 loff_t lstart, loff_t lend)
2011 {
2012         int ret;
2013
2014         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2015         if (ret == 0)
2016                 ret = pnfs_sync_inode(mapping->host, true);
2017         return ret;
2018 }
2019 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2020
2021 /*
2022  * flush the inode to disk.
2023  */
2024 int nfs_wb_all(struct inode *inode)
2025 {
2026         int ret;
2027
2028         trace_nfs_writeback_inode_enter(inode);
2029
2030         ret = filemap_write_and_wait(inode->i_mapping);
2031         if (ret)
2032                 goto out;
2033         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2034         if (ret < 0)
2035                 goto out;
2036         pnfs_sync_inode(inode, true);
2037         ret = 0;
2038
2039 out:
2040         trace_nfs_writeback_inode_exit(inode, ret);
2041         return ret;
2042 }
2043 EXPORT_SYMBOL_GPL(nfs_wb_all);
2044
2045 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2046 {
2047         struct nfs_page *req;
2048         int ret = 0;
2049
2050         folio_wait_writeback(folio);
2051
2052         /* blocking call to cancel all requests and join to a single (head)
2053          * request */
2054         req = nfs_lock_and_join_requests(folio);
2055
2056         if (IS_ERR(req)) {
2057                 ret = PTR_ERR(req);
2058         } else if (req) {
2059                 /* all requests from this folio have been cancelled by
2060                  * nfs_lock_and_join_requests, so just remove the head
2061                  * request from the inode / page_private pointer and
2062                  * release it */
2063                 nfs_inode_remove_request(req);
2064                 nfs_unlock_and_release_request(req);
2065         }
2066
2067         return ret;
2068 }
2069
2070 /**
2071  * nfs_wb_folio - Write back all requests on one page
2072  * @inode: pointer to page
2073  * @folio: pointer to folio
2074  *
2075  * Assumes that the folio has been locked by the caller, and will
2076  * not unlock it.
2077  */
2078 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2079 {
2080         loff_t range_start = folio_file_pos(folio);
2081         loff_t range_end = range_start + (loff_t)folio_size(folio) - 1;
2082         struct writeback_control wbc = {
2083                 .sync_mode = WB_SYNC_ALL,
2084                 .nr_to_write = 0,
2085                 .range_start = range_start,
2086                 .range_end = range_end,
2087         };
2088         int ret;
2089
2090         trace_nfs_writeback_folio(inode, folio);
2091
2092         for (;;) {
2093                 folio_wait_writeback(folio);
2094                 if (folio_clear_dirty_for_io(folio)) {
2095                         ret = nfs_writepage_locked(folio, &wbc);
2096                         if (ret < 0)
2097                                 goto out_error;
2098                         continue;
2099                 }
2100                 ret = 0;
2101                 if (!folio_test_private(folio))
2102                         break;
2103                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2104                 if (ret < 0)
2105                         goto out_error;
2106         }
2107 out_error:
2108         trace_nfs_writeback_folio_done(inode, folio, ret);
2109         return ret;
2110 }
2111
2112 #ifdef CONFIG_MIGRATION
2113 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2114                 struct folio *src, enum migrate_mode mode)
2115 {
2116         /*
2117          * If the private flag is set, the folio is currently associated with
2118          * an in-progress read or write request. Don't try to migrate it.
2119          *
2120          * FIXME: we could do this in principle, but we'll need a way to ensure
2121          *        that we can safely release the inode reference while holding
2122          *        the folio lock.
2123          */
2124         if (folio_test_private(src))
2125                 return -EBUSY;
2126
2127         if (folio_test_fscache(src)) {
2128                 if (mode == MIGRATE_ASYNC)
2129                         return -EBUSY;
2130                 folio_wait_fscache(src);
2131         }
2132
2133         return migrate_folio(mapping, dst, src, mode);
2134 }
2135 #endif
2136
2137 int __init nfs_init_writepagecache(void)
2138 {
2139         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2140                                              sizeof(struct nfs_pgio_header),
2141                                              0, SLAB_HWCACHE_ALIGN,
2142                                              NULL);
2143         if (nfs_wdata_cachep == NULL)
2144                 return -ENOMEM;
2145
2146         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2147                                                      nfs_wdata_cachep);
2148         if (nfs_wdata_mempool == NULL)
2149                 goto out_destroy_write_cache;
2150
2151         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2152                                              sizeof(struct nfs_commit_data),
2153                                              0, SLAB_HWCACHE_ALIGN,
2154                                              NULL);
2155         if (nfs_cdata_cachep == NULL)
2156                 goto out_destroy_write_mempool;
2157
2158         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2159                                                       nfs_cdata_cachep);
2160         if (nfs_commit_mempool == NULL)
2161                 goto out_destroy_commit_cache;
2162
2163         /*
2164          * NFS congestion size, scale with available memory.
2165          *
2166          *  64MB:    8192k
2167          * 128MB:   11585k
2168          * 256MB:   16384k
2169          * 512MB:   23170k
2170          *   1GB:   32768k
2171          *   2GB:   46340k
2172          *   4GB:   65536k
2173          *   8GB:   92681k
2174          *  16GB:  131072k
2175          *
2176          * This allows larger machines to have larger/more transfers.
2177          * Limit the default to 256M
2178          */
2179         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2180         if (nfs_congestion_kb > 256*1024)
2181                 nfs_congestion_kb = 256*1024;
2182
2183         return 0;
2184
2185 out_destroy_commit_cache:
2186         kmem_cache_destroy(nfs_cdata_cachep);
2187 out_destroy_write_mempool:
2188         mempool_destroy(nfs_wdata_mempool);
2189 out_destroy_write_cache:
2190         kmem_cache_destroy(nfs_wdata_cachep);
2191         return -ENOMEM;
2192 }
2193
2194 void nfs_destroy_writepagecache(void)
2195 {
2196         mempool_destroy(nfs_commit_mempool);
2197         kmem_cache_destroy(nfs_cdata_cachep);
2198         mempool_destroy(nfs_wdata_mempool);
2199         kmem_cache_destroy(nfs_wdata_cachep);
2200 }
2201
2202 static const struct nfs_rw_ops nfs_rw_write_ops = {
2203         .rw_alloc_header        = nfs_writehdr_alloc,
2204         .rw_free_header         = nfs_writehdr_free,
2205         .rw_done                = nfs_writeback_done,
2206         .rw_result              = nfs_writeback_result,
2207         .rw_initiate            = nfs_initiate_write,
2208 };