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
  67 static struct kmem_cache *nfs_wdata_cachep;
  68 static mempool_t *nfs_wdata_mempool;
  69 static struct kmem_cache *nfs_cdata_cachep;
  70 static mempool_t *nfs_commit_mempool;
  71
  72 struct nfs_commit_data *nfs_commitdata_alloc(void)
  73 {
  74         struct nfs_commit_data *p;
  75
  76         p = kmem_cache_zalloc(nfs_cdata_cachep, nfs_io_gfp_mask());
  77         if (!p) {
  78                 p = mempool_alloc(nfs_commit_mempool, GFP_NOWAIT);
  79                 if (!p)
  80                         return NULL;
  81                 memset(p, 0, sizeof(*p));
  82         }
  83         INIT_LIST_HEAD(&p->pages);
  84         return p;
  85 }
  86 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  87
  88 void nfs_commit_free(struct nfs_commit_data *p)
  89 {
  90         mempool_free(p, nfs_commit_mempool);
  91 }
  92 EXPORT_SYMBOL_GPL(nfs_commit_free);
  93
  94 static struct nfs_pgio_header *nfs_writehdr_alloc(void)
  95 {
  96         struct nfs_pgio_header *p;
  97
  98         p = kmem_cache_zalloc(nfs_wdata_cachep, nfs_io_gfp_mask());
  99         if (!p) {
 100                 p = mempool_alloc(nfs_wdata_mempool, GFP_NOWAIT);
 101                 if (!p)
 102                         return NULL;
 103                 memset(p, 0, sizeof(*p));
 104         }
 105         p->rw_mode = FMODE_WRITE;
 106         return p;
 107 }
 108
 109 static void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 110 {
 111         mempool_free(hdr, nfs_wdata_mempool);
 112 }
 113
 114 static struct nfs_io_completion *nfs_io_completion_alloc(gfp_t gfp_flags)
 115 {
 116         return kmalloc(sizeof(struct nfs_io_completion), gfp_flags);
 117 }
 118
 119 static void nfs_io_completion_init(struct nfs_io_completion *ioc,
 120                 void (*complete)(void *), void *data)
 121 {
 122         ioc->complete = complete;
 123         ioc->data = data;
 124         kref_init(&ioc->refcount);
 125 }
 126
 127 static void nfs_io_completion_release(struct kref *kref)
 128 {
 129         struct nfs_io_completion *ioc = container_of(kref,
 130                         struct nfs_io_completion, refcount);
 131         ioc->complete(ioc->data);
 132         kfree(ioc);
 133 }
 134
 135 static void nfs_io_completion_get(struct nfs_io_completion *ioc)
 136 {
 137         if (ioc != NULL)
 138                 kref_get(&ioc->refcount);
 139 }
 140
 141 static void nfs_io_completion_put(struct nfs_io_completion *ioc)
 142 {
 143         if (ioc != NULL)
 144                 kref_put(&ioc->refcount, nfs_io_completion_release);
 145 }
 146
 147 static void
 148 nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 149 {
 150         if (!test_and_set_bit(PG_INODE_REF, &req->wb_flags)) {
 151                 kref_get(&req->wb_kref);
 152                 atomic_long_inc(&NFS_I(inode)->nrequests);
 153         }
 154 }
 155
 156 static int
 157 nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 158 {
 159         int ret;
 160
 161         if (!test_bit(PG_REMOVE, &req->wb_flags))
 162                 return 0;
 163         ret = nfs_page_group_lock(req);
 164         if (ret)
 165                 return ret;
 166         if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 167                 nfs_page_set_inode_ref(req, inode);
 168         nfs_page_group_unlock(req);
 169         return 0;
 170 }
 171
 172 /**
 173  * nfs_folio_find_head_request - find head request associated with a folio
 174  * @folio: pointer to folio
 175  *
 176  * must be called while holding the inode lock.
 177  *
 178  * returns matching head request with reference held, or NULL if not found.
 179  */
 180 static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 181 {
 182         struct address_space *mapping = folio->mapping;
 183         struct nfs_page *req;
 184
 185         if (!folio_test_private(folio))
 186                 return NULL;
 187         spin_lock(&mapping->i_private_lock);
 188         req = folio->private;
 189         if (req) {
 190                 WARN_ON_ONCE(req->wb_head != req);
 191                 kref_get(&req->wb_kref);
 192         }
 193         spin_unlock(&mapping->i_private_lock);
 194         return req;
 195 }
 196
 197 /* Adjust the file length if we're writing beyond the end */
 198 static void nfs_grow_file(struct folio *folio, unsigned int offset,
 199                           unsigned int count)
 200 {
 201         struct inode *inode = folio->mapping->host;
 202         loff_t end, i_size;
 203         pgoff_t end_index;
 204
 205         spin_lock(&inode->i_lock);
 206         i_size = i_size_read(inode);
 207         end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
 208         if (i_size > 0 && folio->index < end_index)
 209                 goto out;
 210         end = folio_pos(folio) + (loff_t)offset + (loff_t)count;
 211         if (i_size >= end)
 212                 goto out;
 213         trace_nfs_size_grow(inode, end);
 214         i_size_write(inode, end);
 215         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 216         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 217 out:
 218         /* Atomically update timestamps if they are delegated to us. */
 219         nfs_update_delegated_mtime_locked(inode);
 220         spin_unlock(&inode->i_lock);
 221         nfs_fscache_invalidate(inode, 0);
 222 }
 223
 224 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 225 static void nfs_set_pageerror(struct address_space *mapping)
 226 {
 227         struct inode *inode = mapping->host;
 228
 229         nfs_zap_mapping(mapping->host, mapping);
 230         /* Force file size revalidation */
 231         spin_lock(&inode->i_lock);
 232         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 233                                              NFS_INO_INVALID_CHANGE |
 234                                              NFS_INO_INVALID_SIZE);
 235         spin_unlock(&inode->i_lock);
 236 }
 237
 238 static void nfs_mapping_set_error(struct folio *folio, int error)
 239 {
 240         struct address_space *mapping = folio->mapping;
 241
 242         filemap_set_wb_err(mapping, error);
 243         if (mapping->host)
 244                 errseq_set(&mapping->host->i_sb->s_wb_err,
 245                            error == -ENOSPC ? -ENOSPC : -EIO);
 246         nfs_set_pageerror(mapping);
 247 }
 248
 249 /*
 250  * nfs_page_group_search_locked
 251  * @head - head request of page group
 252  * @page_offset - offset into page
 253  *
 254  * Search page group with head @head to find a request that contains the
 255  * page offset @page_offset.
 256  *
 257  * Returns a pointer to the first matching nfs request, or NULL if no
 258  * match is found.
 259  *
 260  * Must be called with the page group lock held
 261  */
 262 static struct nfs_page *
 263 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 264 {
 265         struct nfs_page *req;
 266
 267         req = head;
 268         do {
 269                 if (page_offset >= req->wb_pgbase &&
 270                     page_offset < (req->wb_pgbase + req->wb_bytes))
 271                         return req;
 272
 273                 req = req->wb_this_page;
 274         } while (req != head);
 275
 276         return NULL;
 277 }
 278
 279 /*
 280  * nfs_page_group_covers_page
 281  * @head - head request of page group
 282  *
 283  * Return true if the page group with head @head covers the whole page,
 284  * returns false otherwise
 285  */
 286 static bool nfs_page_group_covers_page(struct nfs_page *req)
 287 {
 288         unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
 289         struct nfs_page *tmp;
 290         unsigned int pos = 0;
 291
 292         nfs_page_group_lock(req);
 293
 294         for (;;) {
 295                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 296                 if (!tmp)
 297                         break;
 298                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 299         }
 300
 301         nfs_page_group_unlock(req);
 302         return pos >= len;
 303 }
 304
 305 /* We can set the PG_uptodate flag if we see that a write request
 306  * covers the full page.
 307  */
 308 static void nfs_mark_uptodate(struct nfs_page *req)
 309 {
 310         struct folio *folio = nfs_page_to_folio(req);
 311
 312         if (folio_test_uptodate(folio))
 313                 return;
 314         if (!nfs_page_group_covers_page(req))
 315                 return;
 316         folio_mark_uptodate(folio);
 317 }
 318
 319 static int wb_priority(struct writeback_control *wbc)
 320 {
 321         int ret = 0;
 322
 323         if (wbc->sync_mode == WB_SYNC_ALL)
 324                 ret = FLUSH_COND_STABLE;
 325         return ret;
 326 }
 327
 328 /*
 329  * NFS congestion control
 330  */
 331
 332 int nfs_congestion_kb;
 333
 334 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 335 #define NFS_CONGESTION_OFF_THRESH       \
 336         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 337
 338 static void nfs_folio_set_writeback(struct folio *folio)
 339 {
 340         struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 341
 342         folio_start_writeback(folio);
 343         if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
 344                 nfss->write_congested = 1;
 345 }
 346
 347 static void nfs_folio_end_writeback(struct folio *folio)
 348 {
 349         struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 350
 351         folio_end_writeback(folio);
 352         if (atomic_long_dec_return(&nfss->writeback) <
 353             NFS_CONGESTION_OFF_THRESH) {
 354                 nfss->write_congested = 0;
 355                 wake_up_all(&nfss->write_congestion_wait);
 356         }
 357 }
 358
 359 static void nfs_page_end_writeback(struct nfs_page *req)
 360 {
 361         if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
 362                 nfs_unlock_request(req);
 363                 nfs_folio_end_writeback(nfs_page_to_folio(req));
 364         } else
 365                 nfs_unlock_request(req);
 366 }
 367
 368 /*
 369  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 370  *
 371  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 372  * @old_head - the old head of the list
 373  *
 374  * All subrequests must be locked and removed from all lists, so at this point
 375  * they are only "active" in this function, and possibly in nfs_wait_on_request
 376  * with a reference held by some other context.
 377  */
 378 static void
 379 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 380                                  struct nfs_page *old_head,
 381                                  struct inode *inode)
 382 {
 383         while (destroy_list) {
 384                 struct nfs_page *subreq = destroy_list;
 385
 386                 destroy_list = (subreq->wb_this_page == old_head) ?
 387                                    NULL : subreq->wb_this_page;
 388
 389                 /* Note: lock subreq in order to change subreq->wb_head */
 390                 nfs_page_set_headlock(subreq);
 391                 WARN_ON_ONCE(old_head != subreq->wb_head);
 392
 393                 /* make sure old group is not used */
 394                 subreq->wb_this_page = subreq;
 395                 subreq->wb_head = subreq;
 396
 397                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 398
 399                 /* Note: races with nfs_page_group_destroy() */
 400                 if (!kref_read(&subreq->wb_kref)) {
 401                         /* Check if we raced with nfs_page_group_destroy() */
 402                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 403                                 nfs_page_clear_headlock(subreq);
 404                                 nfs_free_request(subreq);
 405                         } else
 406                                 nfs_page_clear_headlock(subreq);
 407                         continue;
 408                 }
 409                 nfs_page_clear_headlock(subreq);
 410
 411                 nfs_release_request(old_head);
 412
 413                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 414                         nfs_release_request(subreq);
 415                         atomic_long_dec(&NFS_I(inode)->nrequests);
 416                 }
 417
 418                 /* subreq is now totally disconnected from page group or any
 419                  * write / commit lists. last chance to wake any waiters */
 420                 nfs_unlock_and_release_request(subreq);
 421         }
 422 }
 423
 424 /*
 425  * nfs_join_page_group - destroy subrequests of the head req
 426  * @head: the page used to lookup the "page group" of nfs_page structures
 427  * @inode: Inode to which the request belongs.
 428  *
 429  * This function joins all sub requests to the head request by first
 430  * locking all requests in the group, cancelling any pending operations
 431  * and finally updating the head request to cover the whole range covered by
 432  * the (former) group.  All subrequests are removed from any write or commit
 433  * lists, unlinked from the group and destroyed.
 434  */
 435 void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
 436                          struct inode *inode)
 437 {
 438         struct nfs_page *subreq;
 439         struct nfs_page *destroy_list = NULL;
 440         unsigned int pgbase, off, bytes;
 441
 442         pgbase = head->wb_pgbase;
 443         bytes = head->wb_bytes;
 444         off = head->wb_offset;
 445         for (subreq = head->wb_this_page; subreq != head;
 446                         subreq = subreq->wb_this_page) {
 447                 /* Subrequests should always form a contiguous range */
 448                 if (pgbase > subreq->wb_pgbase) {
 449                         off -= pgbase - subreq->wb_pgbase;
 450                         bytes += pgbase - subreq->wb_pgbase;
 451                         pgbase = subreq->wb_pgbase;
 452                 }
 453                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 454                                 - pgbase, bytes);
 455         }
 456
 457         /* Set the head request's range to cover the former page group */
 458         head->wb_pgbase = pgbase;
 459         head->wb_bytes = bytes;
 460         head->wb_offset = off;
 461
 462         /* Now that all requests are locked, make sure they aren't on any list.
 463          * Commit list removal accounting is done after locks are dropped */
 464         subreq = head;
 465         do {
 466                 nfs_clear_request_commit(cinfo, subreq);
 467                 subreq = subreq->wb_this_page;
 468         } while (subreq != head);
 469
 470         /* unlink subrequests from head, destroy them later */
 471         if (head->wb_this_page != head) {
 472                 /* destroy list will be terminated by head */
 473                 destroy_list = head->wb_this_page;
 474                 head->wb_this_page = head;
 475         }
 476
 477         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 478 }
 479
 480 /**
 481  * nfs_wait_on_request - Wait for a request to complete.
 482  * @req: request to wait upon.
 483  *
 484  * Interruptible by fatal signals only.
 485  * The user is responsible for holding a count on the request.
 486  */
 487 static int nfs_wait_on_request(struct nfs_page *req)
 488 {
 489         if (!test_bit(PG_BUSY, &req->wb_flags))
 490                 return 0;
 491         set_bit(PG_CONTENDED2, &req->wb_flags);
 492         smp_mb__after_atomic();
 493         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 494                               TASK_UNINTERRUPTIBLE);
 495 }
 496
 497 /*
 498  * nfs_unroll_locks -  unlock all newly locked reqs and wait on @req
 499  * @head: head request of page group, must be holding head lock
 500  * @req: request that couldn't lock and needs to wait on the req bit lock
 501  *
 502  * This is a helper function for nfs_lock_and_join_requests
 503  * returns 0 on success, < 0 on error.
 504  */
 505 static void
 506 nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
 507 {
 508         struct nfs_page *tmp;
 509
 510         /* relinquish all the locks successfully grabbed this run */
 511         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 512                 if (!kref_read(&tmp->wb_kref))
 513                         continue;
 514                 nfs_unlock_and_release_request(tmp);
 515         }
 516 }
 517
 518 /*
 519  * nfs_page_group_lock_subreq -  try to lock a subrequest
 520  * @head: head request of page group
 521  * @subreq: request to lock
 522  *
 523  * This is a helper function for nfs_lock_and_join_requests which
 524  * must be called with the head request and page group both locked.
 525  * On error, it returns with the page group unlocked.
 526  */
 527 static int
 528 nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
 529 {
 530         int ret;
 531
 532         if (!kref_get_unless_zero(&subreq->wb_kref))
 533                 return 0;
 534         while (!nfs_lock_request(subreq)) {
 535                 nfs_page_group_unlock(head);
 536                 ret = nfs_wait_on_request(subreq);
 537                 if (!ret)
 538                         ret = nfs_page_group_lock(head);
 539                 if (ret < 0) {
 540                         nfs_unroll_locks(head, subreq);
 541                         nfs_release_request(subreq);
 542                         return ret;
 543                 }
 544         }
 545         return 0;
 546 }
 547
 548 /*
 549  * nfs_lock_and_join_requests - join all subreqs to the head req
 550  * @folio: the folio used to lookup the "page group" of nfs_page structures
 551  *
 552  * This function joins all sub requests to the head request by first
 553  * locking all requests in the group, cancelling any pending operations
 554  * and finally updating the head request to cover the whole range covered by
 555  * the (former) group.  All subrequests are removed from any write or commit
 556  * lists, unlinked from the group and destroyed.
 557  *
 558  * Returns a locked, referenced pointer to the head request - which after
 559  * this call is guaranteed to be the only request associated with the page.
 560  * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
 561  * error was encountered.
 562  */
 563 static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 564 {
 565         struct inode *inode = folio->mapping->host;
 566         struct nfs_page *head, *subreq;
 567         struct nfs_commit_info cinfo;
 568         int ret;
 569
 570         /*
 571          * A reference is taken only on the head request which acts as a
 572          * reference to the whole page group - the group will not be destroyed
 573          * until the head reference is released.
 574          */
 575 retry:
 576         head = nfs_folio_find_head_request(folio);
 577         if (!head)
 578                 return NULL;
 579
 580         while (!nfs_lock_request(head)) {
 581                 ret = nfs_wait_on_request(head);
 582                 if (ret < 0)
 583                         return ERR_PTR(ret);
 584         }
 585
 586         /* Ensure that nobody removed the request before we locked it */
 587         if (head != folio->private) {
 588                 nfs_unlock_and_release_request(head);
 589                 goto retry;
 590         }
 591
 592         ret = nfs_cancel_remove_inode(head, inode);
 593         if (ret < 0)
 594                 goto out_unlock;
 595
 596         ret = nfs_page_group_lock(head);
 597         if (ret < 0)
 598                 goto out_unlock;
 599
 600         /* lock each request in the page group */
 601         for (subreq = head->wb_this_page;
 602              subreq != head;
 603              subreq = subreq->wb_this_page) {
 604                 ret = nfs_page_group_lock_subreq(head, subreq);
 605                 if (ret < 0)
 606                         goto out_unlock;
 607         }
 608
 609         nfs_page_group_unlock(head);
 610
 611         nfs_init_cinfo_from_inode(&cinfo, inode);
 612         nfs_join_page_group(head, &cinfo, inode);
 613         return head;
 614
 615 out_unlock:
 616         nfs_unlock_and_release_request(head);
 617         return ERR_PTR(ret);
 618 }
 619
 620 static void nfs_write_error(struct nfs_page *req, int error)
 621 {
 622         trace_nfs_write_error(nfs_page_to_inode(req), req, error);
 623         nfs_mapping_set_error(nfs_page_to_folio(req), error);
 624         nfs_inode_remove_request(req);
 625         nfs_page_end_writeback(req);
 626         nfs_release_request(req);
 627 }
 628
 629 /*
 630  * Find an associated nfs write request, and prepare to flush it out
 631  * May return an error if the user signalled nfs_wait_on_request().
 632  */
 633 static int nfs_page_async_flush(struct folio *folio,
 634                                 struct writeback_control *wbc,
 635                                 struct nfs_pageio_descriptor *pgio)
 636 {
 637         struct nfs_page *req;
 638         int ret = 0;
 639
 640         req = nfs_lock_and_join_requests(folio);
 641         if (!req)
 642                 goto out;
 643         ret = PTR_ERR(req);
 644         if (IS_ERR(req))
 645                 goto out;
 646
 647         nfs_folio_set_writeback(folio);
 648         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 649
 650         /* If there is a fatal error that covers this write, just exit */
 651         ret = pgio->pg_error;
 652         if (nfs_error_is_fatal_on_server(ret))
 653                 goto out_launder;
 654
 655         ret = 0;
 656         if (!nfs_pageio_add_request(pgio, req)) {
 657                 ret = pgio->pg_error;
 658                 /*
 659                  * Remove the problematic req upon fatal errors on the server
 660                  */
 661                 if (nfs_error_is_fatal_on_server(ret))
 662                         goto out_launder;
 663                 if (wbc->sync_mode == WB_SYNC_NONE)
 664                         ret = AOP_WRITEPAGE_ACTIVATE;
 665                 folio_redirty_for_writepage(wbc, folio);
 666                 nfs_redirty_request(req);
 667                 pgio->pg_error = 0;
 668         } else
 669                 nfs_add_stats(folio->mapping->host,
 670                               NFSIOS_WRITEPAGES, 1);
 671 out:
 672         return ret;
 673 out_launder:
 674         nfs_write_error(req, ret);
 675         return 0;
 676 }
 677
 678 static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
 679                             struct nfs_pageio_descriptor *pgio)
 680 {
 681         nfs_pageio_cond_complete(pgio, folio->index);
 682         return nfs_page_async_flush(folio, wbc, pgio);
 683 }
 684
 685 /*
 686  * Write an mmapped page to the server.
 687  */
 688 static int nfs_writepage_locked(struct folio *folio,
 689                                 struct writeback_control *wbc)
 690 {
 691         struct nfs_pageio_descriptor pgio;
 692         struct inode *inode = folio->mapping->host;
 693         int err;
 694
 695         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 696         nfs_pageio_init_write(&pgio, inode, 0, false,
 697                               &nfs_async_write_completion_ops);
 698         err = nfs_do_writepage(folio, wbc, &pgio);
 699         pgio.pg_error = 0;
 700         nfs_pageio_complete(&pgio);
 701         return err;
 702 }
 703
 704 static int nfs_writepages_callback(struct folio *folio,
 705                                    struct writeback_control *wbc, void *data)
 706 {
 707         int ret;
 708
 709         ret = nfs_do_writepage(folio, wbc, data);
 710         if (ret != AOP_WRITEPAGE_ACTIVATE)
 711                 folio_unlock(folio);
 712         return ret;
 713 }
 714
 715 static void nfs_io_completion_commit(void *inode)
 716 {
 717         nfs_commit_inode(inode, 0);
 718 }
 719
 720 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 721 {
 722         struct inode *inode = mapping->host;
 723         struct nfs_pageio_descriptor pgio;
 724         struct nfs_io_completion *ioc = NULL;
 725         unsigned int mntflags = NFS_SERVER(inode)->flags;
 726         struct nfs_server *nfss = NFS_SERVER(inode);
 727         int priority = 0;
 728         int err;
 729
 730         /* Wait with writeback until write congestion eases */
 731         if (wbc->sync_mode == WB_SYNC_NONE && nfss->write_congested) {
 732                 err = wait_event_killable(nfss->write_congestion_wait,
 733                                           nfss->write_congested == 0);
 734                 if (err)
 735                         return err;
 736         }
 737
 738         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 739
 740         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 741             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 742                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 743                 if (ioc)
 744                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 745                                                inode);
 746                 priority = wb_priority(wbc);
 747         }
 748
 749         do {
 750                 nfs_pageio_init_write(&pgio, inode, priority, false,
 751                                       &nfs_async_write_completion_ops);
 752                 pgio.pg_io_completion = ioc;
 753                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 754                                         &pgio);
 755                 pgio.pg_error = 0;
 756                 nfs_pageio_complete(&pgio);
 757                 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
 758                         break;
 759         } while (err < 0 && !nfs_error_is_fatal(err));
 760         nfs_io_completion_put(ioc);
 761
 762         if (err < 0)
 763                 goto out_err;
 764         return 0;
 765 out_err:
 766         return err;
 767 }
 768
 769 /*
 770  * Insert a write request into an inode
 771  */
 772 static void nfs_inode_add_request(struct nfs_page *req)
 773 {
 774         struct folio *folio = nfs_page_to_folio(req);
 775         struct address_space *mapping = folio->mapping;
 776         struct nfs_inode *nfsi = NFS_I(mapping->host);
 777
 778         WARN_ON_ONCE(req->wb_this_page != req);
 779
 780         /* Lock the request! */
 781         nfs_lock_request(req);
 782         spin_lock(&mapping->i_private_lock);
 783         set_bit(PG_MAPPED, &req->wb_flags);
 784         folio_set_private(folio);
 785         folio->private = req;
 786         spin_unlock(&mapping->i_private_lock);
 787         atomic_long_inc(&nfsi->nrequests);
 788         /* this a head request for a page group - mark it as having an
 789          * extra reference so sub groups can follow suit.
 790          * This flag also informs pgio layer when to bump nrequests when
 791          * adding subrequests. */
 792         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 793         kref_get(&req->wb_kref);
 794 }
 795
 796 /*
 797  * Remove a write request from an inode
 798  */
 799 static void nfs_inode_remove_request(struct nfs_page *req)
 800 {
 801         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 802
 803         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 804                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 805                 struct address_space *mapping = folio->mapping;
 806
 807                 spin_lock(&mapping->i_private_lock);
 808                 if (likely(folio)) {
 809                         folio->private = NULL;
 810                         folio_clear_private(folio);
 811                         clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 812                 }
 813                 spin_unlock(&mapping->i_private_lock);
 814         }
 815
 816         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 817                 atomic_long_dec(&nfsi->nrequests);
 818                 nfs_release_request(req);
 819         }
 820 }
 821
 822 static void nfs_mark_request_dirty(struct nfs_page *req)
 823 {
 824         struct folio *folio = nfs_page_to_folio(req);
 825         if (folio)
 826                 filemap_dirty_folio(folio_mapping(folio), folio);
 827 }
 828
 829 /**
 830  * nfs_request_add_commit_list_locked - add request to a commit list
 831  * @req: pointer to a struct nfs_page
 832  * @dst: commit list head
 833  * @cinfo: holds list lock and accounting info
 834  *
 835  * This sets the PG_CLEAN bit, updates the cinfo count of
 836  * number of outstanding requests requiring a commit as well as
 837  * the MM page stats.
 838  *
 839  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 840  * nfs_page lock.
 841  */
 842 void
 843 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 844                             struct nfs_commit_info *cinfo)
 845 {
 846         set_bit(PG_CLEAN, &req->wb_flags);
 847         nfs_list_add_request(req, dst);
 848         atomic_long_inc(&cinfo->mds->ncommit);
 849 }
 850 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 851
 852 /**
 853  * nfs_request_add_commit_list - add request to a commit list
 854  * @req: pointer to a struct nfs_page
 855  * @cinfo: holds list lock and accounting info
 856  *
 857  * This sets the PG_CLEAN bit, updates the cinfo count of
 858  * number of outstanding requests requiring a commit as well as
 859  * the MM page stats.
 860  *
 861  * The caller must _not_ hold the cinfo->lock, but must be
 862  * holding the nfs_page lock.
 863  */
 864 void
 865 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 866 {
 867         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 868         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 869         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 870         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 871 }
 872 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 873
 874 /**
 875  * nfs_request_remove_commit_list - Remove request from a commit list
 876  * @req: pointer to a nfs_page
 877  * @cinfo: holds list lock and accounting info
 878  *
 879  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 880  * number of outstanding requests requiring a commit
 881  * It does not update the MM page stats.
 882  *
 883  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 884  */
 885 void
 886 nfs_request_remove_commit_list(struct nfs_page *req,
 887                                struct nfs_commit_info *cinfo)
 888 {
 889         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 890                 return;
 891         nfs_list_remove_request(req);
 892         atomic_long_dec(&cinfo->mds->ncommit);
 893 }
 894 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 895
 896 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 897                                       struct inode *inode)
 898 {
 899         cinfo->inode = inode;
 900         cinfo->mds = &NFS_I(inode)->commit_info;
 901         cinfo->ds = pnfs_get_ds_info(inode);
 902         cinfo->dreq = NULL;
 903         cinfo->completion_ops = &nfs_commit_completion_ops;
 904 }
 905
 906 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 907                     struct inode *inode,
 908                     struct nfs_direct_req *dreq)
 909 {
 910         if (dreq)
 911                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 912         else
 913                 nfs_init_cinfo_from_inode(cinfo, inode);
 914 }
 915 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 916
 917 /*
 918  * Add a request to the inode's commit list.
 919  */
 920 void
 921 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 922                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 923 {
 924         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 925                 return;
 926         nfs_request_add_commit_list(req, cinfo);
 927 }
 928
 929 static void nfs_folio_clear_commit(struct folio *folio)
 930 {
 931         if (folio) {
 932                 long nr = folio_nr_pages(folio);
 933
 934                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 935                 wb_stat_mod(&inode_to_bdi(folio->mapping->host)->wb,
 936                             WB_WRITEBACK, -nr);
 937         }
 938 }
 939
 940 /* Called holding the request lock on @req */
 941 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 942                                      struct nfs_page *req)
 943 {
 944         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 945                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 946                 struct inode *inode = d_inode(ctx->dentry);
 947
 948                 mutex_lock(&NFS_I(inode)->commit_mutex);
 949                 if (!pnfs_clear_request_commit(req, cinfo)) {
 950                         nfs_request_remove_commit_list(req, cinfo);
 951                 }
 952                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 953                 nfs_folio_clear_commit(nfs_page_to_folio(req));
 954         }
 955 }
 956
 957 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 958 {
 959         if (hdr->verf.committed == NFS_DATA_SYNC)
 960                 return hdr->lseg == NULL;
 961         return hdr->verf.committed != NFS_FILE_SYNC;
 962 }
 963
 964 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 965 {
 966         nfs_io_completion_get(hdr->io_completion);
 967 }
 968
 969 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 970 {
 971         struct nfs_commit_info cinfo;
 972         unsigned long bytes = 0;
 973
 974         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 975                 goto out;
 976         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 977         while (!list_empty(&hdr->pages)) {
 978                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 979
 980                 bytes += req->wb_bytes;
 981                 nfs_list_remove_request(req);
 982                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 983                     (hdr->good_bytes < bytes)) {
 984                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
 985                         nfs_mapping_set_error(nfs_page_to_folio(req),
 986                                               hdr->error);
 987                         goto remove_req;
 988                 }
 989                 if (nfs_write_need_commit(hdr)) {
 990                         /* Reset wb_nio, since the write was successful. */
 991                         req->wb_nio = 0;
 992                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 993                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 994                                 hdr->pgio_mirror_idx);
 995                         goto next;
 996                 }
 997 remove_req:
 998                 nfs_inode_remove_request(req);
 999 next:
1000                 nfs_page_end_writeback(req);
1001                 nfs_release_request(req);
1002         }
1003 out:
1004         nfs_io_completion_put(hdr->io_completion);
1005         hdr->release(hdr);
1006 }
1007
1008 unsigned long
1009 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1010 {
1011         return atomic_long_read(&cinfo->mds->ncommit);
1012 }
1013
1014 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1015 int
1016 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1017                      struct nfs_commit_info *cinfo, int max)
1018 {
1019         struct nfs_page *req, *tmp;
1020         int ret = 0;
1021
1022         list_for_each_entry_safe(req, tmp, src, wb_list) {
1023                 kref_get(&req->wb_kref);
1024                 if (!nfs_lock_request(req)) {
1025                         nfs_release_request(req);
1026                         continue;
1027                 }
1028                 nfs_request_remove_commit_list(req, cinfo);
1029                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1030                 nfs_list_add_request(req, dst);
1031                 ret++;
1032                 if ((ret == max) && !cinfo->dreq)
1033                         break;
1034                 cond_resched();
1035         }
1036         return ret;
1037 }
1038 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1039
1040 /*
1041  * nfs_scan_commit - Scan an inode for commit requests
1042  * @inode: NFS inode to scan
1043  * @dst: mds destination list
1044  * @cinfo: mds and ds lists of reqs ready to commit
1045  *
1046  * Moves requests from the inode's 'commit' request list.
1047  * The requests are *not* checked to ensure that they form a contiguous set.
1048  */
1049 int
1050 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1051                 struct nfs_commit_info *cinfo)
1052 {
1053         int ret = 0;
1054
1055         if (!atomic_long_read(&cinfo->mds->ncommit))
1056                 return 0;
1057         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1058         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1059                 const int max = INT_MAX;
1060
1061                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1062                                            cinfo, max);
1063                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1064         }
1065         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1066         return ret;
1067 }
1068
1069 /*
1070  * Search for an existing write request, and attempt to update
1071  * it to reflect a new dirty region on a given page.
1072  *
1073  * If the attempt fails, then the existing request is flushed out
1074  * to disk.
1075  */
1076 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1077                                                   unsigned int offset,
1078                                                   unsigned int bytes)
1079 {
1080         struct nfs_page *req;
1081         unsigned int rqend;
1082         unsigned int end;
1083         int error;
1084
1085         end = offset + bytes;
1086
1087         req = nfs_lock_and_join_requests(folio);
1088         if (IS_ERR_OR_NULL(req))
1089                 return req;
1090
1091         rqend = req->wb_offset + req->wb_bytes;
1092         /*
1093          * Tell the caller to flush out the request if
1094          * the offsets are non-contiguous.
1095          * Note: nfs_flush_incompatible() will already
1096          * have flushed out requests having wrong owners.
1097          */
1098         if (offset > rqend || end < req->wb_offset)
1099                 goto out_flushme;
1100
1101         /* Okay, the request matches. Update the region */
1102         if (offset < req->wb_offset) {
1103                 req->wb_offset = offset;
1104                 req->wb_pgbase = offset;
1105         }
1106         if (end > rqend)
1107                 req->wb_bytes = end - req->wb_offset;
1108         else
1109                 req->wb_bytes = rqend - req->wb_offset;
1110         req->wb_nio = 0;
1111         return req;
1112 out_flushme:
1113         /*
1114          * Note: we mark the request dirty here because
1115          * nfs_lock_and_join_requests() cannot preserve
1116          * commit flags, so we have to replay the write.
1117          */
1118         nfs_mark_request_dirty(req);
1119         nfs_unlock_and_release_request(req);
1120         error = nfs_wb_folio(folio->mapping->host, folio);
1121         return (error < 0) ? ERR_PTR(error) : NULL;
1122 }
1123
1124 /*
1125  * Try to update an existing write request, or create one if there is none.
1126  *
1127  * Note: Should always be called with the Page Lock held to prevent races
1128  * if we have to add a new request. Also assumes that the caller has
1129  * already called nfs_flush_incompatible() if necessary.
1130  */
1131 static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1132                                                 struct folio *folio,
1133                                                 unsigned int offset,
1134                                                 unsigned int bytes)
1135 {
1136         struct nfs_page *req;
1137
1138         req = nfs_try_to_update_request(folio, offset, bytes);
1139         if (req != NULL)
1140                 goto out;
1141         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1142         if (IS_ERR(req))
1143                 goto out;
1144         nfs_inode_add_request(req);
1145 out:
1146         return req;
1147 }
1148
1149 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1150                                struct folio *folio, unsigned int offset,
1151                                unsigned int count)
1152 {
1153         struct nfs_page *req;
1154
1155         req = nfs_setup_write_request(ctx, folio, offset, count);
1156         if (IS_ERR(req))
1157                 return PTR_ERR(req);
1158         /* Update file length */
1159         nfs_grow_file(folio, offset, count);
1160         nfs_mark_uptodate(req);
1161         nfs_mark_request_dirty(req);
1162         nfs_unlock_and_release_request(req);
1163         return 0;
1164 }
1165
1166 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1167 {
1168         struct nfs_open_context *ctx = nfs_file_open_context(file);
1169         struct nfs_lock_context *l_ctx;
1170         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1171         struct nfs_page *req;
1172         int do_flush, status;
1173         /*
1174          * Look for a request corresponding to this page. If there
1175          * is one, and it belongs to another file, we flush it out
1176          * before we try to copy anything into the page. Do this
1177          * due to the lack of an ACCESS-type call in NFSv2.
1178          * Also do the same if we find a request from an existing
1179          * dropped page.
1180          */
1181         do {
1182                 req = nfs_folio_find_head_request(folio);
1183                 if (req == NULL)
1184                         return 0;
1185                 l_ctx = req->wb_lock_context;
1186                 do_flush = nfs_page_to_folio(req) != folio ||
1187                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1188                 if (l_ctx && flctx &&
1189                     !(list_empty_careful(&flctx->flc_posix) &&
1190                       list_empty_careful(&flctx->flc_flock))) {
1191                         do_flush |= l_ctx->lockowner != current->files;
1192                 }
1193                 nfs_release_request(req);
1194                 if (!do_flush)
1195                         return 0;
1196                 status = nfs_wb_folio(folio->mapping->host, folio);
1197         } while (status == 0);
1198         return status;
1199 }
1200
1201 /*
1202  * Avoid buffered writes when a open context credential's key would
1203  * expire soon.
1204  *
1205  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1206  *
1207  * Return 0 and set a credential flag which triggers the inode to flush
1208  * and performs  NFS_FILE_SYNC writes if the key will expired within
1209  * RPC_KEY_EXPIRE_TIMEO.
1210  */
1211 int
1212 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1213 {
1214         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1215
1216         if (nfs_ctx_key_to_expire(ctx, inode) &&
1217             !rcu_access_pointer(ctx->ll_cred))
1218                 /* Already expired! */
1219                 return -EACCES;
1220         return 0;
1221 }
1222
1223 /*
1224  * Test if the open context credential key is marked to expire soon.
1225  */
1226 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1227 {
1228         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1229         struct rpc_cred *cred, *new, *old = NULL;
1230         struct auth_cred acred = {
1231                 .cred = ctx->cred,
1232         };
1233         bool ret = false;
1234
1235         rcu_read_lock();
1236         cred = rcu_dereference(ctx->ll_cred);
1237         if (cred && !(cred->cr_ops->crkey_timeout &&
1238                       cred->cr_ops->crkey_timeout(cred)))
1239                 goto out;
1240         rcu_read_unlock();
1241
1242         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1243         if (new == cred) {
1244                 put_rpccred(new);
1245                 return true;
1246         }
1247         if (IS_ERR_OR_NULL(new)) {
1248                 new = NULL;
1249                 ret = true;
1250         } else if (new->cr_ops->crkey_timeout &&
1251                    new->cr_ops->crkey_timeout(new))
1252                 ret = true;
1253
1254         rcu_read_lock();
1255         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1256                                              RCU_INITIALIZER(new)), 1);
1257 out:
1258         rcu_read_unlock();
1259         put_rpccred(old);
1260         return ret;
1261 }
1262
1263 /*
1264  * If the page cache is marked as unsafe or invalid, then we can't rely on
1265  * the PageUptodate() flag. In this case, we will need to turn off
1266  * write optimisations that depend on the page contents being correct.
1267  */
1268 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1269 {
1270         struct inode *inode = folio->mapping->host;
1271         struct nfs_inode *nfsi = NFS_I(inode);
1272
1273         if (nfs_have_delegated_attributes(inode))
1274                 goto out;
1275         if (nfsi->cache_validity &
1276             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1277                 return false;
1278         smp_rmb();
1279         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1280                 return false;
1281 out:
1282         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1283                 return false;
1284         return folio_test_uptodate(folio) != 0;
1285 }
1286
1287 static bool
1288 is_whole_file_wrlock(struct file_lock *fl)
1289 {
1290         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1291                         lock_is_write(fl);
1292 }
1293
1294 /* If we know the page is up to date, and we're not using byte range locks (or
1295  * if we have the whole file locked for writing), it may be more efficient to
1296  * extend the write to cover the entire page in order to avoid fragmentation
1297  * inefficiencies.
1298  *
1299  * If the file is opened for synchronous writes then we can just skip the rest
1300  * of the checks.
1301  */
1302 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1303                                 unsigned int pagelen)
1304 {
1305         struct inode *inode = file_inode(file);
1306         struct file_lock_context *flctx = locks_inode_context(inode);
1307         struct file_lock *fl;
1308         int ret;
1309         unsigned int mntflags = NFS_SERVER(inode)->flags;
1310
1311         if (mntflags & NFS_MOUNT_NO_ALIGNWRITE)
1312                 return 0;
1313         if (file->f_flags & O_DSYNC)
1314                 return 0;
1315         if (!nfs_folio_write_uptodate(folio, pagelen))
1316                 return 0;
1317         if (nfs_have_write_delegation(inode))
1318                 return 1;
1319         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1320                        list_empty_careful(&flctx->flc_posix)))
1321                 return 1;
1322
1323         /* Check to see if there are whole file write locks */
1324         ret = 0;
1325         spin_lock(&flctx->flc_lock);
1326         if (!list_empty(&flctx->flc_posix)) {
1327                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1328                                         c.flc_list);
1329                 if (is_whole_file_wrlock(fl))
1330                         ret = 1;
1331         } else if (!list_empty(&flctx->flc_flock)) {
1332                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1333                                         c.flc_list);
1334                 if (lock_is_write(fl))
1335                         ret = 1;
1336         }
1337         spin_unlock(&flctx->flc_lock);
1338         return ret;
1339 }
1340
1341 /*
1342  * Update and possibly write a cached page of an NFS file.
1343  *
1344  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1345  * things with a page scheduled for an RPC call (e.g. invalidate it).
1346  */
1347 int nfs_update_folio(struct file *file, struct folio *folio,
1348                      unsigned int offset, unsigned int count)
1349 {
1350         struct nfs_open_context *ctx = nfs_file_open_context(file);
1351         struct address_space *mapping = folio->mapping;
1352         struct inode *inode = mapping->host;
1353         unsigned int pagelen = nfs_folio_length(folio);
1354         int             status = 0;
1355
1356         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1357
1358         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1359                 (long long)(folio_pos(folio) + offset));
1360
1361         if (!count)
1362                 goto out;
1363
1364         if (nfs_can_extend_write(file, folio, pagelen)) {
1365                 unsigned int end = count + offset;
1366
1367                 offset = round_down(offset, PAGE_SIZE);
1368                 if (end < pagelen)
1369                         end = min(round_up(end, PAGE_SIZE), pagelen);
1370                 count = end - offset;
1371         }
1372
1373         status = nfs_writepage_setup(ctx, folio, offset, count);
1374         if (status < 0)
1375                 nfs_set_pageerror(mapping);
1376 out:
1377         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1378                         status, (long long)i_size_read(inode));
1379         return status;
1380 }
1381
1382 static int flush_task_priority(int how)
1383 {
1384         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1385                 case FLUSH_HIGHPRI:
1386                         return RPC_PRIORITY_HIGH;
1387                 case FLUSH_LOWPRI:
1388                         return RPC_PRIORITY_LOW;
1389         }
1390         return RPC_PRIORITY_NORMAL;
1391 }
1392
1393 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1394                                struct rpc_message *msg,
1395                                const struct nfs_rpc_ops *rpc_ops,
1396                                struct rpc_task_setup *task_setup_data, int how)
1397 {
1398         int priority = flush_task_priority(how);
1399
1400         if (IS_SWAPFILE(hdr->inode))
1401                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1402         task_setup_data->priority = priority;
1403         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1404         trace_nfs_initiate_write(hdr);
1405 }
1406
1407 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1408  * call this on each, which will prepare them to be retried on next
1409  * writeback using standard nfs.
1410  */
1411 static void nfs_redirty_request(struct nfs_page *req)
1412 {
1413         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1414
1415         /* Bump the transmission count */
1416         req->wb_nio++;
1417         nfs_mark_request_dirty(req);
1418         atomic_long_inc(&nfsi->redirtied_pages);
1419         nfs_page_end_writeback(req);
1420         nfs_release_request(req);
1421 }
1422
1423 static void nfs_async_write_error(struct list_head *head, int error)
1424 {
1425         struct nfs_page *req;
1426
1427         while (!list_empty(head)) {
1428                 req = nfs_list_entry(head->next);
1429                 nfs_list_remove_request(req);
1430                 if (nfs_error_is_fatal_on_server(error))
1431                         nfs_write_error(req, error);
1432                 else
1433                         nfs_redirty_request(req);
1434         }
1435 }
1436
1437 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1438 {
1439         nfs_async_write_error(&hdr->pages, 0);
1440 }
1441
1442 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1443         .init_hdr = nfs_async_write_init,
1444         .error_cleanup = nfs_async_write_error,
1445         .completion = nfs_write_completion,
1446         .reschedule_io = nfs_async_write_reschedule_io,
1447 };
1448
1449 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1450                                struct inode *inode, int ioflags, bool force_mds,
1451                                const struct nfs_pgio_completion_ops *compl_ops)
1452 {
1453         struct nfs_server *server = NFS_SERVER(inode);
1454         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1455
1456 #ifdef CONFIG_NFS_V4_1
1457         if (server->pnfs_curr_ld && !force_mds)
1458                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1459 #endif
1460         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1461                         server->wsize, ioflags);
1462 }
1463 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1464
1465 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1466 {
1467         struct nfs_pgio_mirror *mirror;
1468
1469         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1470                 pgio->pg_ops->pg_cleanup(pgio);
1471
1472         pgio->pg_ops = &nfs_pgio_rw_ops;
1473
1474         nfs_pageio_stop_mirroring(pgio);
1475
1476         mirror = &pgio->pg_mirrors[0];
1477         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1478 }
1479 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1480
1481
1482 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1483 {
1484         struct nfs_commit_data *data = calldata;
1485
1486         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1487 }
1488
1489 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1490                 struct nfs_fattr *fattr)
1491 {
1492         struct nfs_pgio_args *argp = &hdr->args;
1493         struct nfs_pgio_res *resp = &hdr->res;
1494         u64 size = argp->offset + resp->count;
1495
1496         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1497                 fattr->size = size;
1498         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1499                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1500                 return;
1501         }
1502         if (size != fattr->size)
1503                 return;
1504         /* Set attribute barrier */
1505         nfs_fattr_set_barrier(fattr);
1506         /* ...and update size */
1507         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1508 }
1509
1510 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1511 {
1512         struct nfs_fattr *fattr = &hdr->fattr;
1513         struct inode *inode = hdr->inode;
1514
1515         if (nfs_have_delegated_mtime(inode)) {
1516                 spin_lock(&inode->i_lock);
1517                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_BLOCKS);
1518                 spin_unlock(&inode->i_lock);
1519                 return;
1520         }
1521
1522         spin_lock(&inode->i_lock);
1523         nfs_writeback_check_extend(hdr, fattr);
1524         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1525         spin_unlock(&inode->i_lock);
1526 }
1527 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1528
1529 /*
1530  * This function is called when the WRITE call is complete.
1531  */
1532 static int nfs_writeback_done(struct rpc_task *task,
1533                               struct nfs_pgio_header *hdr,
1534                               struct inode *inode)
1535 {
1536         int status;
1537
1538         /*
1539          * ->write_done will attempt to use post-op attributes to detect
1540          * conflicting writes by other clients.  A strict interpretation
1541          * of close-to-open would allow us to continue caching even if
1542          * another writer had changed the file, but some applications
1543          * depend on tighter cache coherency when writing.
1544          */
1545         status = NFS_PROTO(inode)->write_done(task, hdr);
1546         if (status != 0)
1547                 return status;
1548
1549         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1550         trace_nfs_writeback_done(task, hdr);
1551
1552         if (task->tk_status >= 0) {
1553                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1554
1555                 if (committed == NFS_UNSTABLE) {
1556                         /*
1557                          * We have some uncommitted data on the server at
1558                          * this point, so ensure that we keep track of that
1559                          * fact irrespective of what later writes do.
1560                          */
1561                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1562                 }
1563
1564                 if (committed < hdr->args.stable) {
1565                         /* We tried a write call, but the server did not
1566                          * commit data to stable storage even though we
1567                          * requested it.
1568                          * Note: There is a known bug in Tru64 < 5.0 in which
1569                          *       the server reports NFS_DATA_SYNC, but performs
1570                          *       NFS_FILE_SYNC. We therefore implement this checking
1571                          *       as a dprintk() in order to avoid filling syslog.
1572                          */
1573                         static unsigned long    complain;
1574
1575                         /* Note this will print the MDS for a DS write */
1576                         if (time_before(complain, jiffies)) {
1577                                 dprintk("NFS:       faulty NFS server %s:"
1578                                         " (committed = %d) != (stable = %d)\n",
1579                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1580                                         committed, hdr->args.stable);
1581                                 complain = jiffies + 300 * HZ;
1582                         }
1583                 }
1584         }
1585
1586         /* Deal with the suid/sgid bit corner case */
1587         if (nfs_should_remove_suid(inode)) {
1588                 spin_lock(&inode->i_lock);
1589                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1590                 spin_unlock(&inode->i_lock);
1591         }
1592         return 0;
1593 }
1594
1595 /*
1596  * This function is called when the WRITE call is complete.
1597  */
1598 static void nfs_writeback_result(struct rpc_task *task,
1599                                  struct nfs_pgio_header *hdr)
1600 {
1601         struct nfs_pgio_args    *argp = &hdr->args;
1602         struct nfs_pgio_res     *resp = &hdr->res;
1603
1604         if (resp->count < argp->count) {
1605                 static unsigned long    complain;
1606
1607                 /* This a short write! */
1608                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1609
1610                 /* Has the server at least made some progress? */
1611                 if (resp->count == 0) {
1612                         if (time_before(complain, jiffies)) {
1613                                 printk(KERN_WARNING
1614                                        "NFS: Server wrote zero bytes, expected %u.\n",
1615                                        argp->count);
1616                                 complain = jiffies + 300 * HZ;
1617                         }
1618                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1619                         task->tk_status = -EIO;
1620                         return;
1621                 }
1622
1623                 /* For non rpc-based layout drivers, retry-through-MDS */
1624                 if (!task->tk_ops) {
1625                         hdr->pnfs_error = -EAGAIN;
1626                         return;
1627                 }
1628
1629                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1630                 if (resp->verf->committed != NFS_UNSTABLE) {
1631                         /* Resend from where the server left off */
1632                         hdr->mds_offset += resp->count;
1633                         argp->offset += resp->count;
1634                         argp->pgbase += resp->count;
1635                         argp->count -= resp->count;
1636                 } else {
1637                         /* Resend as a stable write in order to avoid
1638                          * headaches in the case of a server crash.
1639                          */
1640                         argp->stable = NFS_FILE_SYNC;
1641                 }
1642                 resp->count = 0;
1643                 resp->verf->committed = 0;
1644                 rpc_restart_call_prepare(task);
1645         }
1646 }
1647
1648 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1649 {
1650         return wait_var_event_killable(&cinfo->rpcs_out,
1651                                        !atomic_read(&cinfo->rpcs_out));
1652 }
1653
1654 void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1655 {
1656         atomic_inc(&cinfo->rpcs_out);
1657 }
1658
1659 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1660 {
1661         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1662                 wake_up_var(&cinfo->rpcs_out);
1663                 return true;
1664         }
1665         return false;
1666 }
1667
1668 void nfs_commitdata_release(struct nfs_commit_data *data)
1669 {
1670         put_nfs_open_context(data->context);
1671         nfs_commit_free(data);
1672 }
1673 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1674
1675 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1676                         const struct nfs_rpc_ops *nfs_ops,
1677                         const struct rpc_call_ops *call_ops,
1678                         int how, int flags,
1679                         struct nfsd_file *localio)
1680 {
1681         struct rpc_task *task;
1682         int priority = flush_task_priority(how);
1683         struct rpc_message msg = {
1684                 .rpc_argp = &data->args,
1685                 .rpc_resp = &data->res,
1686                 .rpc_cred = data->cred,
1687         };
1688         struct rpc_task_setup task_setup_data = {
1689                 .task = &data->task,
1690                 .rpc_client = clnt,
1691                 .rpc_message = &msg,
1692                 .callback_ops = call_ops,
1693                 .callback_data = data,
1694                 .workqueue = nfsiod_workqueue,
1695                 .flags = RPC_TASK_ASYNC | flags,
1696                 .priority = priority,
1697         };
1698
1699         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1700                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1701
1702         /* Set up the initial task struct.  */
1703         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1704         trace_nfs_initiate_commit(data);
1705
1706         dprintk("NFS: initiated commit call\n");
1707
1708         if (localio)
1709                 return nfs_local_commit(localio, data, call_ops, how);
1710
1711         task = rpc_run_task(&task_setup_data);
1712         if (IS_ERR(task))
1713                 return PTR_ERR(task);
1714         if (how & FLUSH_SYNC)
1715                 rpc_wait_for_completion_task(task);
1716         rpc_put_task(task);
1717         return 0;
1718 }
1719 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1720
1721 static loff_t nfs_get_lwb(struct list_head *head)
1722 {
1723         loff_t lwb = 0;
1724         struct nfs_page *req;
1725
1726         list_for_each_entry(req, head, wb_list)
1727                 if (lwb < (req_offset(req) + req->wb_bytes))
1728                         lwb = req_offset(req) + req->wb_bytes;
1729
1730         return lwb;
1731 }
1732
1733 /*
1734  * Set up the argument/result storage required for the RPC call.
1735  */
1736 void nfs_init_commit(struct nfs_commit_data *data,
1737                      struct list_head *head,
1738                      struct pnfs_layout_segment *lseg,
1739                      struct nfs_commit_info *cinfo)
1740 {
1741         struct nfs_page *first;
1742         struct nfs_open_context *ctx;
1743         struct inode *inode;
1744
1745         /* Set up the RPC argument and reply structs
1746          * NB: take care not to mess about with data->commit et al. */
1747
1748         if (head)
1749                 list_splice_init(head, &data->pages);
1750
1751         first = nfs_list_entry(data->pages.next);
1752         ctx = nfs_req_openctx(first);
1753         inode = d_inode(ctx->dentry);
1754
1755         data->inode       = inode;
1756         data->cred        = ctx->cred;
1757         data->lseg        = lseg; /* reference transferred */
1758         /* only set lwb for pnfs commit */
1759         if (lseg)
1760                 data->lwb = nfs_get_lwb(&data->pages);
1761         data->mds_ops     = &nfs_commit_ops;
1762         data->completion_ops = cinfo->completion_ops;
1763         data->dreq        = cinfo->dreq;
1764
1765         data->args.fh     = NFS_FH(data->inode);
1766         /* Note: we always request a commit of the entire inode */
1767         data->args.offset = 0;
1768         data->args.count  = 0;
1769         data->context     = get_nfs_open_context(ctx);
1770         data->res.fattr   = &data->fattr;
1771         data->res.verf    = &data->verf;
1772         nfs_fattr_init(&data->fattr);
1773         nfs_commit_begin(cinfo->mds);
1774 }
1775 EXPORT_SYMBOL_GPL(nfs_init_commit);
1776
1777 void nfs_retry_commit(struct list_head *page_list,
1778                       struct pnfs_layout_segment *lseg,
1779                       struct nfs_commit_info *cinfo,
1780                       u32 ds_commit_idx)
1781 {
1782         struct nfs_page *req;
1783
1784         while (!list_empty(page_list)) {
1785                 req = nfs_list_entry(page_list->next);
1786                 nfs_list_remove_request(req);
1787                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1788                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1789                 nfs_unlock_and_release_request(req);
1790         }
1791 }
1792 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1793
1794 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1795                                      struct nfs_page *req)
1796 {
1797         struct folio *folio = nfs_page_to_folio(req);
1798
1799         filemap_dirty_folio(folio_mapping(folio), folio);
1800 }
1801
1802 /*
1803  * Commit dirty pages
1804  */
1805 static int
1806 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1807                 struct nfs_commit_info *cinfo)
1808 {
1809         struct nfs_commit_data  *data;
1810         struct nfsd_file *localio;
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
1828         localio = nfs_local_open_fh(NFS_SERVER(inode)->nfs_client, data->cred,
1829                                     data->args.fh, data->context->mode);
1830         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1831                                    data->mds_ops, how,
1832                                    RPC_TASK_CRED_NOREF | task_flags, localio);
1833 }
1834
1835 /*
1836  * COMMIT call returned
1837  */
1838 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1839 {
1840         struct nfs_commit_data  *data = calldata;
1841
1842         /* Call the NFS version-specific code */
1843         NFS_PROTO(data->inode)->commit_done(task, data);
1844         trace_nfs_commit_done(task, data);
1845 }
1846
1847 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1848 {
1849         const struct nfs_writeverf *verf = data->res.verf;
1850         struct nfs_page *req;
1851         int status = data->task.tk_status;
1852         struct nfs_commit_info cinfo;
1853         struct folio *folio;
1854
1855         while (!list_empty(&data->pages)) {
1856                 req = nfs_list_entry(data->pages.next);
1857                 nfs_list_remove_request(req);
1858                 folio = nfs_page_to_folio(req);
1859                 nfs_folio_clear_commit(folio);
1860
1861                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1862                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1863                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1864                         req->wb_bytes,
1865                         (long long)req_offset(req));
1866                 if (status < 0) {
1867                         if (folio) {
1868                                 trace_nfs_commit_error(data->inode, req,
1869                                                        status);
1870                                 nfs_mapping_set_error(folio, status);
1871                                 nfs_inode_remove_request(req);
1872                         }
1873                         dprintk_cont(", error = %d\n", status);
1874                         goto next;
1875                 }
1876
1877                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1878                  * returned by the server against all stored verfs. */
1879                 if (nfs_write_match_verf(verf, req)) {
1880                         /* We have a match */
1881                         if (folio)
1882                                 nfs_inode_remove_request(req);
1883                         dprintk_cont(" OK\n");
1884                         goto next;
1885                 }
1886                 /* We have a mismatch. Write the page again */
1887                 dprintk_cont(" mismatch\n");
1888                 nfs_mark_request_dirty(req);
1889                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1890         next:
1891                 nfs_unlock_and_release_request(req);
1892                 /* Latency breaker */
1893                 cond_resched();
1894         }
1895
1896         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1897         nfs_commit_end(cinfo.mds);
1898 }
1899
1900 static void nfs_commit_release(void *calldata)
1901 {
1902         struct nfs_commit_data *data = calldata;
1903
1904         data->completion_ops->completion(data);
1905         nfs_commitdata_release(calldata);
1906 }
1907
1908 static const struct rpc_call_ops nfs_commit_ops = {
1909         .rpc_call_prepare = nfs_commit_prepare,
1910         .rpc_call_done = nfs_commit_done,
1911         .rpc_release = nfs_commit_release,
1912 };
1913
1914 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1915         .completion = nfs_commit_release_pages,
1916         .resched_write = nfs_commit_resched_write,
1917 };
1918
1919 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1920                             int how, struct nfs_commit_info *cinfo)
1921 {
1922         int status;
1923
1924         status = pnfs_commit_list(inode, head, how, cinfo);
1925         if (status == PNFS_NOT_ATTEMPTED)
1926                 status = nfs_commit_list(inode, head, how, cinfo);
1927         return status;
1928 }
1929
1930 static int __nfs_commit_inode(struct inode *inode, int how,
1931                 struct writeback_control *wbc)
1932 {
1933         LIST_HEAD(head);
1934         struct nfs_commit_info cinfo;
1935         int may_wait = how & FLUSH_SYNC;
1936         int ret, nscan;
1937
1938         how &= ~FLUSH_SYNC;
1939         nfs_init_cinfo_from_inode(&cinfo, inode);
1940         nfs_commit_begin(cinfo.mds);
1941         for (;;) {
1942                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1943                 if (ret <= 0)
1944                         break;
1945                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1946                 if (ret < 0)
1947                         break;
1948                 ret = 0;
1949                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1950                         if (nscan < wbc->nr_to_write)
1951                                 wbc->nr_to_write -= nscan;
1952                         else
1953                                 wbc->nr_to_write = 0;
1954                 }
1955                 if (nscan < INT_MAX)
1956                         break;
1957                 cond_resched();
1958         }
1959         nfs_commit_end(cinfo.mds);
1960         if (ret || !may_wait)
1961                 return ret;
1962         return wait_on_commit(cinfo.mds);
1963 }
1964
1965 int nfs_commit_inode(struct inode *inode, int how)
1966 {
1967         return __nfs_commit_inode(inode, how, NULL);
1968 }
1969 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1970
1971 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1972 {
1973         struct nfs_inode *nfsi = NFS_I(inode);
1974         int flags = FLUSH_SYNC;
1975         int ret = 0;
1976
1977         if (wbc->sync_mode == WB_SYNC_NONE) {
1978                 /* no commits means nothing needs to be done */
1979                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1980                         goto check_requests_outstanding;
1981
1982                 /* Don't commit yet if this is a non-blocking flush and there
1983                  * are a lot of outstanding writes for this mapping.
1984                  */
1985                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1986                         goto out_mark_dirty;
1987
1988                 /* don't wait for the COMMIT response */
1989                 flags = 0;
1990         }
1991
1992         ret = __nfs_commit_inode(inode, flags, wbc);
1993         if (!ret) {
1994                 if (flags & FLUSH_SYNC)
1995                         return 0;
1996         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1997                 goto out_mark_dirty;
1998
1999 check_requests_outstanding:
2000         if (!atomic_read(&nfsi->commit_info.rpcs_out))
2001                 return ret;
2002 out_mark_dirty:
2003         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
2004         return ret;
2005 }
2006 EXPORT_SYMBOL_GPL(nfs_write_inode);
2007
2008 /*
2009  * Wrapper for filemap_write_and_wait_range()
2010  *
2011  * Needed for pNFS in order to ensure data becomes visible to the
2012  * client.
2013  */
2014 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
2015                 loff_t lstart, loff_t lend)
2016 {
2017         int ret;
2018
2019         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2020         if (ret == 0)
2021                 ret = pnfs_sync_inode(mapping->host, true);
2022         return ret;
2023 }
2024 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2025
2026 /*
2027  * flush the inode to disk.
2028  */
2029 int nfs_wb_all(struct inode *inode)
2030 {
2031         int ret;
2032
2033         trace_nfs_writeback_inode_enter(inode);
2034
2035         ret = filemap_write_and_wait(inode->i_mapping);
2036         if (ret)
2037                 goto out;
2038         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2039         if (ret < 0)
2040                 goto out;
2041         pnfs_sync_inode(inode, true);
2042         ret = 0;
2043
2044 out:
2045         trace_nfs_writeback_inode_exit(inode, ret);
2046         return ret;
2047 }
2048 EXPORT_SYMBOL_GPL(nfs_wb_all);
2049
2050 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2051 {
2052         struct nfs_page *req;
2053         int ret = 0;
2054
2055         folio_wait_writeback(folio);
2056
2057         /* blocking call to cancel all requests and join to a single (head)
2058          * request */
2059         req = nfs_lock_and_join_requests(folio);
2060
2061         if (IS_ERR(req)) {
2062                 ret = PTR_ERR(req);
2063         } else if (req) {
2064                 /* all requests from this folio have been cancelled by
2065                  * nfs_lock_and_join_requests, so just remove the head
2066                  * request from the inode / page_private pointer and
2067                  * release it */
2068                 nfs_inode_remove_request(req);
2069                 nfs_unlock_and_release_request(req);
2070         }
2071
2072         return ret;
2073 }
2074
2075 /**
2076  * nfs_wb_folio - Write back all requests on one page
2077  * @inode: pointer to page
2078  * @folio: pointer to folio
2079  *
2080  * Assumes that the folio has been locked by the caller, and will
2081  * not unlock it.
2082  */
2083 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2084 {
2085         loff_t range_start = folio_pos(folio);
2086         size_t len = folio_size(folio);
2087         struct writeback_control wbc = {
2088                 .sync_mode = WB_SYNC_ALL,
2089                 .nr_to_write = 0,
2090                 .range_start = range_start,
2091                 .range_end = range_start + len - 1,
2092         };
2093         int ret;
2094
2095         trace_nfs_writeback_folio(inode, range_start, len);
2096
2097         for (;;) {
2098                 folio_wait_writeback(folio);
2099                 if (folio_clear_dirty_for_io(folio)) {
2100                         ret = nfs_writepage_locked(folio, &wbc);
2101                         if (ret < 0)
2102                                 goto out_error;
2103                         continue;
2104                 }
2105                 ret = 0;
2106                 if (!folio_test_private(folio))
2107                         break;
2108                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2109                 if (ret < 0)
2110                         goto out_error;
2111         }
2112 out_error:
2113         trace_nfs_writeback_folio_done(inode, range_start, len, ret);
2114         return ret;
2115 }
2116
2117 #ifdef CONFIG_MIGRATION
2118 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2119                 struct folio *src, enum migrate_mode mode)
2120 {
2121         /*
2122          * If the private flag is set, the folio is currently associated with
2123          * an in-progress read or write request. Don't try to migrate it.
2124          *
2125          * FIXME: we could do this in principle, but we'll need a way to ensure
2126          *        that we can safely release the inode reference while holding
2127          *        the folio lock.
2128          */
2129         if (folio_test_private(src))
2130                 return -EBUSY;
2131
2132         if (folio_test_private_2(src)) { /* [DEPRECATED] */
2133                 if (mode == MIGRATE_ASYNC)
2134                         return -EBUSY;
2135                 folio_wait_private_2(src);
2136         }
2137
2138         return migrate_folio(mapping, dst, src, mode);
2139 }
2140 #endif
2141
2142 int __init nfs_init_writepagecache(void)
2143 {
2144         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2145                                              sizeof(struct nfs_pgio_header),
2146                                              0, SLAB_HWCACHE_ALIGN,
2147                                              NULL);
2148         if (nfs_wdata_cachep == NULL)
2149                 return -ENOMEM;
2150
2151         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2152                                                      nfs_wdata_cachep);
2153         if (nfs_wdata_mempool == NULL)
2154                 goto out_destroy_write_cache;
2155
2156         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2157                                              sizeof(struct nfs_commit_data),
2158                                              0, SLAB_HWCACHE_ALIGN,
2159                                              NULL);
2160         if (nfs_cdata_cachep == NULL)
2161                 goto out_destroy_write_mempool;
2162
2163         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2164                                                       nfs_cdata_cachep);
2165         if (nfs_commit_mempool == NULL)
2166                 goto out_destroy_commit_cache;
2167
2168         /*
2169          * NFS congestion size, scale with available memory.
2170          *
2171          *  64MB:    8192k
2172          * 128MB:   11585k
2173          * 256MB:   16384k
2174          * 512MB:   23170k
2175          *   1GB:   32768k
2176          *   2GB:   46340k
2177          *   4GB:   65536k
2178          *   8GB:   92681k
2179          *  16GB:  131072k
2180          *
2181          * This allows larger machines to have larger/more transfers.
2182          * Limit the default to 256M
2183          */
2184         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2185         if (nfs_congestion_kb > 256*1024)
2186                 nfs_congestion_kb = 256*1024;
2187
2188         return 0;
2189
2190 out_destroy_commit_cache:
2191         kmem_cache_destroy(nfs_cdata_cachep);
2192 out_destroy_write_mempool:
2193         mempool_destroy(nfs_wdata_mempool);
2194 out_destroy_write_cache:
2195         kmem_cache_destroy(nfs_wdata_cachep);
2196         return -ENOMEM;
2197 }
2198
2199 void nfs_destroy_writepagecache(void)
2200 {
2201         mempool_destroy(nfs_commit_mempool);
2202         kmem_cache_destroy(nfs_cdata_cachep);
2203         mempool_destroy(nfs_wdata_mempool);
2204         kmem_cache_destroy(nfs_wdata_cachep);
2205 }
2206
2207 static const struct nfs_rw_ops nfs_rw_write_ops = {
2208         .rw_alloc_header        = nfs_writehdr_alloc,
2209         .rw_free_header         = nfs_writehdr_free,
2210         .rw_done                = nfs_writeback_done,
2211         .rw_result              = nfs_writeback_result,
2212         .rw_initiate            = nfs_initiate_write,
2213 };