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 /**
 148  * nfs_folio_find_head_request - find head request associated with a folio
 149  * @folio: pointer to folio
 150  *
 151  * must be called while holding the inode lock.
 152  *
 153  * returns matching head request with reference held, or NULL if not found.
 154  */
 155 static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 156 {
 157         struct address_space *mapping = folio->mapping;
 158         struct nfs_page *req;
 159
 160         if (!folio_test_private(folio))
 161                 return NULL;
 162         spin_lock(&mapping->i_private_lock);
 163         req = folio->private;
 164         if (req) {
 165                 WARN_ON_ONCE(req->wb_head != req);
 166                 kref_get(&req->wb_kref);
 167         }
 168         spin_unlock(&mapping->i_private_lock);
 169         return req;
 170 }
 171
 172 /* Adjust the file length if we're writing beyond the end */
 173 static void nfs_grow_file(struct folio *folio, unsigned int offset,
 174                           unsigned int count)
 175 {
 176         struct inode *inode = folio->mapping->host;
 177         loff_t end, i_size;
 178         pgoff_t end_index;
 179
 180         spin_lock(&inode->i_lock);
 181         i_size = i_size_read(inode);
 182         end_index = ((i_size - 1) >> folio_shift(folio)) << folio_order(folio);
 183         if (i_size > 0 && folio->index < end_index)
 184                 goto out;
 185         end = folio_pos(folio) + (loff_t)offset + (loff_t)count;
 186         if (i_size >= end)
 187                 goto out;
 188         trace_nfs_size_grow(inode, end);
 189         i_size_write(inode, end);
 190         NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_SIZE;
 191         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 192 out:
 193         /* Atomically update timestamps if they are delegated to us. */
 194         nfs_update_delegated_mtime_locked(inode);
 195         spin_unlock(&inode->i_lock);
 196         nfs_fscache_invalidate(inode, 0);
 197 }
 198
 199 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 200 static void nfs_set_pageerror(struct address_space *mapping)
 201 {
 202         struct inode *inode = mapping->host;
 203
 204         nfs_zap_mapping(mapping->host, mapping);
 205         /* Force file size revalidation */
 206         spin_lock(&inode->i_lock);
 207         nfs_set_cache_invalid(inode, NFS_INO_REVAL_FORCED |
 208                                              NFS_INO_INVALID_CHANGE |
 209                                              NFS_INO_INVALID_SIZE);
 210         spin_unlock(&inode->i_lock);
 211 }
 212
 213 static void nfs_mapping_set_error(struct folio *folio, int error)
 214 {
 215         struct address_space *mapping = folio->mapping;
 216
 217         filemap_set_wb_err(mapping, error);
 218         if (mapping->host)
 219                 errseq_set(&mapping->host->i_sb->s_wb_err,
 220                            error == -ENOSPC ? -ENOSPC : -EIO);
 221         nfs_set_pageerror(mapping);
 222 }
 223
 224 /*
 225  * nfs_page_group_search_locked
 226  * @head - head request of page group
 227  * @page_offset - offset into page
 228  *
 229  * Search page group with head @head to find a request that contains the
 230  * page offset @page_offset.
 231  *
 232  * Returns a pointer to the first matching nfs request, or NULL if no
 233  * match is found.
 234  *
 235  * Must be called with the page group lock held
 236  */
 237 static struct nfs_page *
 238 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 239 {
 240         struct nfs_page *req;
 241
 242         req = head;
 243         do {
 244                 if (page_offset >= req->wb_pgbase &&
 245                     page_offset < (req->wb_pgbase + req->wb_bytes))
 246                         return req;
 247
 248                 req = req->wb_this_page;
 249         } while (req != head);
 250
 251         return NULL;
 252 }
 253
 254 /*
 255  * nfs_page_group_covers_page
 256  * @head - head request of page group
 257  *
 258  * Return true if the page group with head @head covers the whole page,
 259  * returns false otherwise
 260  */
 261 static bool nfs_page_group_covers_page(struct nfs_page *req)
 262 {
 263         unsigned int len = nfs_folio_length(nfs_page_to_folio(req));
 264         struct nfs_page *tmp;
 265         unsigned int pos = 0;
 266
 267         nfs_page_group_lock(req);
 268
 269         for (;;) {
 270                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 271                 if (!tmp)
 272                         break;
 273                 pos = tmp->wb_pgbase + tmp->wb_bytes;
 274         }
 275
 276         nfs_page_group_unlock(req);
 277         return pos >= len;
 278 }
 279
 280 /* We can set the PG_uptodate flag if we see that a write request
 281  * covers the full page.
 282  */
 283 static void nfs_mark_uptodate(struct nfs_page *req)
 284 {
 285         struct folio *folio = nfs_page_to_folio(req);
 286
 287         if (folio_test_uptodate(folio))
 288                 return;
 289         if (!nfs_page_group_covers_page(req))
 290                 return;
 291         folio_mark_uptodate(folio);
 292 }
 293
 294 static int wb_priority(struct writeback_control *wbc)
 295 {
 296         int ret = 0;
 297
 298         if (wbc->sync_mode == WB_SYNC_ALL)
 299                 ret = FLUSH_COND_STABLE;
 300         return ret;
 301 }
 302
 303 /*
 304  * NFS congestion control
 305  */
 306
 307 int nfs_congestion_kb;
 308
 309 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 310 #define NFS_CONGESTION_OFF_THRESH       \
 311         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 312
 313 static void nfs_folio_set_writeback(struct folio *folio)
 314 {
 315         struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 316
 317         folio_start_writeback(folio);
 318         if (atomic_long_inc_return(&nfss->writeback) > NFS_CONGESTION_ON_THRESH)
 319                 nfss->write_congested = 1;
 320 }
 321
 322 static void nfs_folio_end_writeback(struct folio *folio)
 323 {
 324         struct nfs_server *nfss = NFS_SERVER(folio->mapping->host);
 325
 326         folio_end_writeback(folio);
 327         if (atomic_long_dec_return(&nfss->writeback) <
 328             NFS_CONGESTION_OFF_THRESH) {
 329                 nfss->write_congested = 0;
 330                 wake_up_all(&nfss->write_congestion_wait);
 331         }
 332 }
 333
 334 static void nfs_page_end_writeback(struct nfs_page *req)
 335 {
 336         if (nfs_page_group_sync_on_bit(req, PG_WB_END)) {
 337                 nfs_unlock_request(req);
 338                 nfs_folio_end_writeback(nfs_page_to_folio(req));
 339         } else
 340                 nfs_unlock_request(req);
 341 }
 342
 343 /*
 344  * nfs_destroy_unlinked_subrequests - destroy recently unlinked subrequests
 345  *
 346  * @destroy_list - request list (using wb_this_page) terminated by @old_head
 347  * @old_head - the old head of the list
 348  *
 349  * All subrequests must be locked and removed from all lists, so at this point
 350  * they are only "active" in this function, and possibly in nfs_wait_on_request
 351  * with a reference held by some other context.
 352  */
 353 static void
 354 nfs_destroy_unlinked_subrequests(struct nfs_page *destroy_list,
 355                                  struct nfs_page *old_head,
 356                                  struct inode *inode)
 357 {
 358         while (destroy_list) {
 359                 struct nfs_page *subreq = destroy_list;
 360
 361                 destroy_list = (subreq->wb_this_page == old_head) ?
 362                                    NULL : subreq->wb_this_page;
 363
 364                 /* Note: lock subreq in order to change subreq->wb_head */
 365                 nfs_page_set_headlock(subreq);
 366                 WARN_ON_ONCE(old_head != subreq->wb_head);
 367
 368                 /* make sure old group is not used */
 369                 subreq->wb_this_page = subreq;
 370                 subreq->wb_head = subreq;
 371
 372                 clear_bit(PG_REMOVE, &subreq->wb_flags);
 373
 374                 /* Note: races with nfs_page_group_destroy() */
 375                 if (!kref_read(&subreq->wb_kref)) {
 376                         /* Check if we raced with nfs_page_group_destroy() */
 377                         if (test_and_clear_bit(PG_TEARDOWN, &subreq->wb_flags)) {
 378                                 nfs_page_clear_headlock(subreq);
 379                                 nfs_free_request(subreq);
 380                         } else
 381                                 nfs_page_clear_headlock(subreq);
 382                         continue;
 383                 }
 384                 nfs_page_clear_headlock(subreq);
 385
 386                 nfs_release_request(old_head);
 387
 388                 if (test_and_clear_bit(PG_INODE_REF, &subreq->wb_flags)) {
 389                         nfs_release_request(subreq);
 390                         atomic_long_dec(&NFS_I(inode)->nrequests);
 391                 }
 392
 393                 /* subreq is now totally disconnected from page group or any
 394                  * write / commit lists. last chance to wake any waiters */
 395                 nfs_unlock_and_release_request(subreq);
 396         }
 397 }
 398
 399 /*
 400  * nfs_join_page_group - destroy subrequests of the head req
 401  * @head: the page used to lookup the "page group" of nfs_page structures
 402  * @inode: Inode to which the request belongs.
 403  *
 404  * This function joins all sub requests to the head request by first
 405  * locking all requests in the group, cancelling any pending operations
 406  * and finally updating the head request to cover the whole range covered by
 407  * the (former) group.  All subrequests are removed from any write or commit
 408  * lists, unlinked from the group and destroyed.
 409  */
 410 void nfs_join_page_group(struct nfs_page *head, struct nfs_commit_info *cinfo,
 411                          struct inode *inode)
 412 {
 413         struct nfs_page *subreq;
 414         struct nfs_page *destroy_list = NULL;
 415         unsigned int pgbase, off, bytes;
 416
 417         pgbase = head->wb_pgbase;
 418         bytes = head->wb_bytes;
 419         off = head->wb_offset;
 420         for (subreq = head->wb_this_page; subreq != head;
 421                         subreq = subreq->wb_this_page) {
 422                 /* Subrequests should always form a contiguous range */
 423                 if (pgbase > subreq->wb_pgbase) {
 424                         off -= pgbase - subreq->wb_pgbase;
 425                         bytes += pgbase - subreq->wb_pgbase;
 426                         pgbase = subreq->wb_pgbase;
 427                 }
 428                 bytes = max(subreq->wb_pgbase + subreq->wb_bytes
 429                                 - pgbase, bytes);
 430         }
 431
 432         /* Set the head request's range to cover the former page group */
 433         head->wb_pgbase = pgbase;
 434         head->wb_bytes = bytes;
 435         head->wb_offset = off;
 436
 437         /* Now that all requests are locked, make sure they aren't on any list.
 438          * Commit list removal accounting is done after locks are dropped */
 439         subreq = head;
 440         do {
 441                 nfs_clear_request_commit(cinfo, subreq);
 442                 subreq = subreq->wb_this_page;
 443         } while (subreq != head);
 444
 445         /* unlink subrequests from head, destroy them later */
 446         if (head->wb_this_page != head) {
 447                 /* destroy list will be terminated by head */
 448                 destroy_list = head->wb_this_page;
 449                 head->wb_this_page = head;
 450         }
 451
 452         nfs_destroy_unlinked_subrequests(destroy_list, head, inode);
 453 }
 454
 455 /**
 456  * nfs_wait_on_request - Wait for a request to complete.
 457  * @req: request to wait upon.
 458  *
 459  * Interruptible by fatal signals only.
 460  * The user is responsible for holding a count on the request.
 461  */
 462 static int nfs_wait_on_request(struct nfs_page *req)
 463 {
 464         if (!test_bit(PG_BUSY, &req->wb_flags))
 465                 return 0;
 466         set_bit(PG_CONTENDED2, &req->wb_flags);
 467         smp_mb__after_atomic();
 468         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 469                               TASK_UNINTERRUPTIBLE);
 470 }
 471
 472 /*
 473  * nfs_unroll_locks -  unlock all newly locked reqs and wait on @req
 474  * @head: head request of page group, must be holding head lock
 475  * @req: request that couldn't lock and needs to wait on the req bit lock
 476  *
 477  * This is a helper function for nfs_lock_and_join_requests
 478  * returns 0 on success, < 0 on error.
 479  */
 480 static void
 481 nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
 482 {
 483         struct nfs_page *tmp;
 484
 485         /* relinquish all the locks successfully grabbed this run */
 486         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 487                 if (!kref_read(&tmp->wb_kref))
 488                         continue;
 489                 nfs_unlock_and_release_request(tmp);
 490         }
 491 }
 492
 493 /*
 494  * nfs_page_group_lock_subreq -  try to lock a subrequest
 495  * @head: head request of page group
 496  * @subreq: request to lock
 497  *
 498  * This is a helper function for nfs_lock_and_join_requests which
 499  * must be called with the head request and page group both locked.
 500  * On error, it returns with the page group unlocked.
 501  */
 502 static int
 503 nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
 504 {
 505         int ret;
 506
 507         if (!kref_get_unless_zero(&subreq->wb_kref))
 508                 return 0;
 509         while (!nfs_lock_request(subreq)) {
 510                 nfs_page_group_unlock(head);
 511                 ret = nfs_wait_on_request(subreq);
 512                 if (!ret)
 513                         ret = nfs_page_group_lock(head);
 514                 if (ret < 0) {
 515                         nfs_unroll_locks(head, subreq);
 516                         nfs_release_request(subreq);
 517                         return ret;
 518                 }
 519         }
 520         return 0;
 521 }
 522
 523 /*
 524  * nfs_lock_and_join_requests - join all subreqs to the head req
 525  * @folio: the folio used to lookup the "page group" of nfs_page structures
 526  *
 527  * This function joins all sub requests to the head request by first
 528  * locking all requests in the group, cancelling any pending operations
 529  * and finally updating the head request to cover the whole range covered by
 530  * the (former) group.  All subrequests are removed from any write or commit
 531  * lists, unlinked from the group and destroyed.
 532  *
 533  * Returns a locked, referenced pointer to the head request - which after
 534  * this call is guaranteed to be the only request associated with the page.
 535  * Returns NULL if no requests are found for @folio, or a ERR_PTR if an
 536  * error was encountered.
 537  */
 538 static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 539 {
 540         struct inode *inode = folio->mapping->host;
 541         struct nfs_page *head, *subreq;
 542         struct nfs_commit_info cinfo;
 543         bool removed;
 544         int ret;
 545
 546         /*
 547          * A reference is taken only on the head request which acts as a
 548          * reference to the whole page group - the group will not be destroyed
 549          * until the head reference is released.
 550          */
 551 retry:
 552         head = nfs_folio_find_head_request(folio);
 553         if (!head)
 554                 return NULL;
 555
 556         while (!nfs_lock_request(head)) {
 557                 ret = nfs_wait_on_request(head);
 558                 if (ret < 0)
 559                         return ERR_PTR(ret);
 560         }
 561
 562         /* Ensure that nobody removed the request before we locked it */
 563         if (head != folio->private) {
 564                 nfs_unlock_and_release_request(head);
 565                 goto retry;
 566         }
 567
 568         ret = nfs_page_group_lock(head);
 569         if (ret < 0)
 570                 goto out_unlock;
 571
 572         removed = test_bit(PG_REMOVE, &head->wb_flags);
 573
 574         /* lock each request in the page group */
 575         for (subreq = head->wb_this_page;
 576              subreq != head;
 577              subreq = subreq->wb_this_page) {
 578                 if (test_bit(PG_REMOVE, &subreq->wb_flags))
 579                         removed = true;
 580                 ret = nfs_page_group_lock_subreq(head, subreq);
 581                 if (ret < 0)
 582                         goto out_unlock;
 583         }
 584
 585         nfs_page_group_unlock(head);
 586
 587         /*
 588          * If PG_REMOVE is set on any request, I/O on that request has
 589          * completed, but some requests were still under I/O at the time
 590          * we locked the head request.
 591          *
 592          * In that case the above wait for all requests means that all I/O
 593          * has now finished, and we can restart from a clean slate.  Let the
 594          * old requests go away and start from scratch instead.
 595          */
 596         if (removed) {
 597                 nfs_unroll_locks(head, head);
 598                 nfs_unlock_and_release_request(head);
 599                 goto retry;
 600         }
 601
 602         nfs_init_cinfo_from_inode(&cinfo, inode);
 603         nfs_join_page_group(head, &cinfo, inode);
 604         return head;
 605
 606 out_unlock:
 607         nfs_unlock_and_release_request(head);
 608         return ERR_PTR(ret);
 609 }
 610
 611 static void nfs_write_error(struct nfs_page *req, int error)
 612 {
 613         trace_nfs_write_error(nfs_page_to_inode(req), req, error);
 614         nfs_mapping_set_error(nfs_page_to_folio(req), error);
 615         nfs_inode_remove_request(req);
 616         nfs_page_end_writeback(req);
 617         nfs_release_request(req);
 618 }
 619
 620 /*
 621  * Find an associated nfs write request, and prepare to flush it out
 622  * May return an error if the user signalled nfs_wait_on_request().
 623  */
 624 static int nfs_page_async_flush(struct folio *folio,
 625                                 struct writeback_control *wbc,
 626                                 struct nfs_pageio_descriptor *pgio)
 627 {
 628         struct nfs_page *req;
 629         int ret = 0;
 630
 631         req = nfs_lock_and_join_requests(folio);
 632         if (!req)
 633                 goto out;
 634         ret = PTR_ERR(req);
 635         if (IS_ERR(req))
 636                 goto out;
 637
 638         nfs_folio_set_writeback(folio);
 639         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 640
 641         /* If there is a fatal error that covers this write, just exit */
 642         ret = pgio->pg_error;
 643         if (nfs_error_is_fatal_on_server(ret))
 644                 goto out_launder;
 645
 646         ret = 0;
 647         if (!nfs_pageio_add_request(pgio, req)) {
 648                 ret = pgio->pg_error;
 649                 /*
 650                  * Remove the problematic req upon fatal errors on the server
 651                  */
 652                 if (nfs_error_is_fatal_on_server(ret))
 653                         goto out_launder;
 654                 if (wbc->sync_mode == WB_SYNC_NONE)
 655                         ret = AOP_WRITEPAGE_ACTIVATE;
 656                 folio_redirty_for_writepage(wbc, folio);
 657                 nfs_redirty_request(req);
 658                 pgio->pg_error = 0;
 659         } else
 660                 nfs_add_stats(folio->mapping->host,
 661                               NFSIOS_WRITEPAGES, 1);
 662 out:
 663         return ret;
 664 out_launder:
 665         nfs_write_error(req, ret);
 666         return 0;
 667 }
 668
 669 static int nfs_do_writepage(struct folio *folio, struct writeback_control *wbc,
 670                             struct nfs_pageio_descriptor *pgio)
 671 {
 672         nfs_pageio_cond_complete(pgio, folio->index);
 673         return nfs_page_async_flush(folio, wbc, pgio);
 674 }
 675
 676 /*
 677  * Write an mmapped page to the server.
 678  */
 679 static int nfs_writepage_locked(struct folio *folio,
 680                                 struct writeback_control *wbc)
 681 {
 682         struct nfs_pageio_descriptor pgio;
 683         struct inode *inode = folio->mapping->host;
 684         int err;
 685
 686         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 687         nfs_pageio_init_write(&pgio, inode, 0, false,
 688                               &nfs_async_write_completion_ops);
 689         err = nfs_do_writepage(folio, wbc, &pgio);
 690         pgio.pg_error = 0;
 691         nfs_pageio_complete(&pgio);
 692         return err;
 693 }
 694
 695 static int nfs_writepages_callback(struct folio *folio,
 696                                    struct writeback_control *wbc, void *data)
 697 {
 698         int ret;
 699
 700         ret = nfs_do_writepage(folio, wbc, data);
 701         if (ret != AOP_WRITEPAGE_ACTIVATE)
 702                 folio_unlock(folio);
 703         return ret;
 704 }
 705
 706 static void nfs_io_completion_commit(void *inode)
 707 {
 708         nfs_commit_inode(inode, 0);
 709 }
 710
 711 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 712 {
 713         struct inode *inode = mapping->host;
 714         struct nfs_pageio_descriptor pgio;
 715         struct nfs_io_completion *ioc = NULL;
 716         unsigned int mntflags = NFS_SERVER(inode)->flags;
 717         struct nfs_server *nfss = NFS_SERVER(inode);
 718         int priority = 0;
 719         int err;
 720
 721         /* Wait with writeback until write congestion eases */
 722         if (wbc->sync_mode == WB_SYNC_NONE && nfss->write_congested) {
 723                 err = wait_event_killable(nfss->write_congestion_wait,
 724                                           nfss->write_congested == 0);
 725                 if (err)
 726                         return err;
 727         }
 728
 729         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 730
 731         if (!(mntflags & NFS_MOUNT_WRITE_EAGER) || wbc->for_kupdate ||
 732             wbc->for_background || wbc->for_sync || wbc->for_reclaim) {
 733                 ioc = nfs_io_completion_alloc(GFP_KERNEL);
 734                 if (ioc)
 735                         nfs_io_completion_init(ioc, nfs_io_completion_commit,
 736                                                inode);
 737                 priority = wb_priority(wbc);
 738         }
 739
 740         do {
 741                 nfs_pageio_init_write(&pgio, inode, priority, false,
 742                                       &nfs_async_write_completion_ops);
 743                 pgio.pg_io_completion = ioc;
 744                 err = write_cache_pages(mapping, wbc, nfs_writepages_callback,
 745                                         &pgio);
 746                 pgio.pg_error = 0;
 747                 nfs_pageio_complete(&pgio);
 748                 if (err == -EAGAIN && mntflags & NFS_MOUNT_SOFTERR)
 749                         break;
 750         } while (err < 0 && !nfs_error_is_fatal(err));
 751         nfs_io_completion_put(ioc);
 752
 753         if (err < 0)
 754                 goto out_err;
 755         return 0;
 756 out_err:
 757         return err;
 758 }
 759
 760 /*
 761  * Insert a write request into an inode
 762  */
 763 static void nfs_inode_add_request(struct nfs_page *req)
 764 {
 765         struct folio *folio = nfs_page_to_folio(req);
 766         struct address_space *mapping = folio->mapping;
 767         struct nfs_inode *nfsi = NFS_I(mapping->host);
 768
 769         WARN_ON_ONCE(req->wb_this_page != req);
 770
 771         /* Lock the request! */
 772         nfs_lock_request(req);
 773         spin_lock(&mapping->i_private_lock);
 774         set_bit(PG_MAPPED, &req->wb_flags);
 775         folio_set_private(folio);
 776         folio->private = req;
 777         spin_unlock(&mapping->i_private_lock);
 778         atomic_long_inc(&nfsi->nrequests);
 779         /* this a head request for a page group - mark it as having an
 780          * extra reference so sub groups can follow suit.
 781          * This flag also informs pgio layer when to bump nrequests when
 782          * adding subrequests. */
 783         WARN_ON(test_and_set_bit(PG_INODE_REF, &req->wb_flags));
 784         kref_get(&req->wb_kref);
 785 }
 786
 787 /*
 788  * Remove a write request from an inode
 789  */
 790 static void nfs_inode_remove_request(struct nfs_page *req)
 791 {
 792         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 793
 794         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 795                 struct folio *folio = nfs_page_to_folio(req->wb_head);
 796                 struct address_space *mapping = folio->mapping;
 797
 798                 spin_lock(&mapping->i_private_lock);
 799                 if (likely(folio)) {
 800                         folio->private = NULL;
 801                         folio_clear_private(folio);
 802                         clear_bit(PG_MAPPED, &req->wb_head->wb_flags);
 803                 }
 804                 spin_unlock(&mapping->i_private_lock);
 805         }
 806
 807         if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 808                 atomic_long_dec(&nfsi->nrequests);
 809                 nfs_release_request(req);
 810         }
 811 }
 812
 813 static void nfs_mark_request_dirty(struct nfs_page *req)
 814 {
 815         struct folio *folio = nfs_page_to_folio(req);
 816         if (folio)
 817                 filemap_dirty_folio(folio_mapping(folio), folio);
 818 }
 819
 820 /**
 821  * nfs_request_add_commit_list_locked - add request to a commit list
 822  * @req: pointer to a struct nfs_page
 823  * @dst: commit list head
 824  * @cinfo: holds list lock and accounting info
 825  *
 826  * This sets the PG_CLEAN bit, updates the cinfo count of
 827  * number of outstanding requests requiring a commit as well as
 828  * the MM page stats.
 829  *
 830  * The caller must hold NFS_I(cinfo->inode)->commit_mutex, and the
 831  * nfs_page lock.
 832  */
 833 void
 834 nfs_request_add_commit_list_locked(struct nfs_page *req, struct list_head *dst,
 835                             struct nfs_commit_info *cinfo)
 836 {
 837         set_bit(PG_CLEAN, &req->wb_flags);
 838         nfs_list_add_request(req, dst);
 839         atomic_long_inc(&cinfo->mds->ncommit);
 840 }
 841 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list_locked);
 842
 843 /**
 844  * nfs_request_add_commit_list - add request to a commit list
 845  * @req: pointer to a struct nfs_page
 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 _not_ hold the cinfo->lock, but must be
 853  * holding the nfs_page lock.
 854  */
 855 void
 856 nfs_request_add_commit_list(struct nfs_page *req, struct nfs_commit_info *cinfo)
 857 {
 858         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 859         nfs_request_add_commit_list_locked(req, &cinfo->mds->list, cinfo);
 860         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 861         nfs_folio_mark_unstable(nfs_page_to_folio(req), cinfo);
 862 }
 863 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 864
 865 /**
 866  * nfs_request_remove_commit_list - Remove request from a commit list
 867  * @req: pointer to a nfs_page
 868  * @cinfo: holds list lock and accounting info
 869  *
 870  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 871  * number of outstanding requests requiring a commit
 872  * It does not update the MM page stats.
 873  *
 874  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 875  */
 876 void
 877 nfs_request_remove_commit_list(struct nfs_page *req,
 878                                struct nfs_commit_info *cinfo)
 879 {
 880         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 881                 return;
 882         nfs_list_remove_request(req);
 883         atomic_long_dec(&cinfo->mds->ncommit);
 884 }
 885 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 886
 887 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 888                                       struct inode *inode)
 889 {
 890         cinfo->inode = inode;
 891         cinfo->mds = &NFS_I(inode)->commit_info;
 892         cinfo->ds = pnfs_get_ds_info(inode);
 893         cinfo->dreq = NULL;
 894         cinfo->completion_ops = &nfs_commit_completion_ops;
 895 }
 896
 897 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 898                     struct inode *inode,
 899                     struct nfs_direct_req *dreq)
 900 {
 901         if (dreq)
 902                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 903         else
 904                 nfs_init_cinfo_from_inode(cinfo, inode);
 905 }
 906 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 907
 908 /*
 909  * Add a request to the inode's commit list.
 910  */
 911 void
 912 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 913                         struct nfs_commit_info *cinfo, u32 ds_commit_idx)
 914 {
 915         if (pnfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx))
 916                 return;
 917         nfs_request_add_commit_list(req, cinfo);
 918 }
 919
 920 static void nfs_folio_clear_commit(struct folio *folio)
 921 {
 922         if (folio) {
 923                 long nr = folio_nr_pages(folio);
 924
 925                 node_stat_mod_folio(folio, NR_WRITEBACK, -nr);
 926                 wb_stat_mod(&inode_to_bdi(folio->mapping->host)->wb,
 927                             WB_WRITEBACK, -nr);
 928         }
 929 }
 930
 931 /* Called holding the request lock on @req */
 932 static void nfs_clear_request_commit(struct nfs_commit_info *cinfo,
 933                                      struct nfs_page *req)
 934 {
 935         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 936                 struct nfs_open_context *ctx = nfs_req_openctx(req);
 937                 struct inode *inode = d_inode(ctx->dentry);
 938
 939                 mutex_lock(&NFS_I(inode)->commit_mutex);
 940                 if (!pnfs_clear_request_commit(req, cinfo)) {
 941                         nfs_request_remove_commit_list(req, cinfo);
 942                 }
 943                 mutex_unlock(&NFS_I(inode)->commit_mutex);
 944                 nfs_folio_clear_commit(nfs_page_to_folio(req));
 945         }
 946 }
 947
 948 int nfs_write_need_commit(struct nfs_pgio_header *hdr)
 949 {
 950         if (hdr->verf.committed == NFS_DATA_SYNC)
 951                 return hdr->lseg == NULL;
 952         return hdr->verf.committed != NFS_FILE_SYNC;
 953 }
 954
 955 static void nfs_async_write_init(struct nfs_pgio_header *hdr)
 956 {
 957         nfs_io_completion_get(hdr->io_completion);
 958 }
 959
 960 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 961 {
 962         struct nfs_commit_info cinfo;
 963         unsigned long bytes = 0;
 964
 965         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 966                 goto out;
 967         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 968         while (!list_empty(&hdr->pages)) {
 969                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 970
 971                 bytes += req->wb_bytes;
 972                 nfs_list_remove_request(req);
 973                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 974                     (hdr->good_bytes < bytes)) {
 975                         trace_nfs_comp_error(hdr->inode, req, hdr->error);
 976                         nfs_mapping_set_error(nfs_page_to_folio(req),
 977                                               hdr->error);
 978                         goto remove_req;
 979                 }
 980                 if (nfs_write_need_commit(hdr)) {
 981                         /* Reset wb_nio, since the write was successful. */
 982                         req->wb_nio = 0;
 983                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 984                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 985                                 hdr->pgio_mirror_idx);
 986                         goto next;
 987                 }
 988 remove_req:
 989                 nfs_inode_remove_request(req);
 990 next:
 991                 nfs_page_end_writeback(req);
 992                 nfs_release_request(req);
 993         }
 994 out:
 995         nfs_io_completion_put(hdr->io_completion);
 996         hdr->release(hdr);
 997 }
 998
 999 unsigned long
1000 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
1001 {
1002         return atomic_long_read(&cinfo->mds->ncommit);
1003 }
1004
1005 /* NFS_I(cinfo->inode)->commit_mutex held by caller */
1006 int
1007 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
1008                      struct nfs_commit_info *cinfo, int max)
1009 {
1010         struct nfs_page *req, *tmp;
1011         int ret = 0;
1012
1013         list_for_each_entry_safe(req, tmp, src, wb_list) {
1014                 kref_get(&req->wb_kref);
1015                 if (!nfs_lock_request(req)) {
1016                         nfs_release_request(req);
1017                         continue;
1018                 }
1019                 nfs_request_remove_commit_list(req, cinfo);
1020                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1021                 nfs_list_add_request(req, dst);
1022                 ret++;
1023                 if ((ret == max) && !cinfo->dreq)
1024                         break;
1025                 cond_resched();
1026         }
1027         return ret;
1028 }
1029 EXPORT_SYMBOL_GPL(nfs_scan_commit_list);
1030
1031 /*
1032  * nfs_scan_commit - Scan an inode for commit requests
1033  * @inode: NFS inode to scan
1034  * @dst: mds destination list
1035  * @cinfo: mds and ds lists of reqs ready to commit
1036  *
1037  * Moves requests from the inode's 'commit' request list.
1038  * The requests are *not* checked to ensure that they form a contiguous set.
1039  */
1040 int
1041 nfs_scan_commit(struct inode *inode, struct list_head *dst,
1042                 struct nfs_commit_info *cinfo)
1043 {
1044         int ret = 0;
1045
1046         if (!atomic_long_read(&cinfo->mds->ncommit))
1047                 return 0;
1048         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
1049         if (atomic_long_read(&cinfo->mds->ncommit) > 0) {
1050                 const int max = INT_MAX;
1051
1052                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
1053                                            cinfo, max);
1054                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
1055         }
1056         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
1057         return ret;
1058 }
1059
1060 /*
1061  * Search for an existing write request, and attempt to update
1062  * it to reflect a new dirty region on a given page.
1063  *
1064  * If the attempt fails, then the existing request is flushed out
1065  * to disk.
1066  */
1067 static struct nfs_page *nfs_try_to_update_request(struct folio *folio,
1068                                                   unsigned int offset,
1069                                                   unsigned int bytes)
1070 {
1071         struct nfs_page *req;
1072         unsigned int rqend;
1073         unsigned int end;
1074         int error;
1075
1076         end = offset + bytes;
1077
1078         req = nfs_lock_and_join_requests(folio);
1079         if (IS_ERR_OR_NULL(req))
1080                 return req;
1081
1082         rqend = req->wb_offset + req->wb_bytes;
1083         /*
1084          * Tell the caller to flush out the request if
1085          * the offsets are non-contiguous.
1086          * Note: nfs_flush_incompatible() will already
1087          * have flushed out requests having wrong owners.
1088          */
1089         if (offset > rqend || end < req->wb_offset)
1090                 goto out_flushme;
1091
1092         /* Okay, the request matches. Update the region */
1093         if (offset < req->wb_offset) {
1094                 req->wb_offset = offset;
1095                 req->wb_pgbase = offset;
1096         }
1097         if (end > rqend)
1098                 req->wb_bytes = end - req->wb_offset;
1099         else
1100                 req->wb_bytes = rqend - req->wb_offset;
1101         req->wb_nio = 0;
1102         return req;
1103 out_flushme:
1104         /*
1105          * Note: we mark the request dirty here because
1106          * nfs_lock_and_join_requests() cannot preserve
1107          * commit flags, so we have to replay the write.
1108          */
1109         nfs_mark_request_dirty(req);
1110         nfs_unlock_and_release_request(req);
1111         error = nfs_wb_folio(folio->mapping->host, folio);
1112         return (error < 0) ? ERR_PTR(error) : NULL;
1113 }
1114
1115 /*
1116  * Try to update an existing write request, or create one if there is none.
1117  *
1118  * Note: Should always be called with the Page Lock held to prevent races
1119  * if we have to add a new request. Also assumes that the caller has
1120  * already called nfs_flush_incompatible() if necessary.
1121  */
1122 static struct nfs_page *nfs_setup_write_request(struct nfs_open_context *ctx,
1123                                                 struct folio *folio,
1124                                                 unsigned int offset,
1125                                                 unsigned int bytes)
1126 {
1127         struct nfs_page *req;
1128
1129         req = nfs_try_to_update_request(folio, offset, bytes);
1130         if (req != NULL)
1131                 goto out;
1132         req = nfs_page_create_from_folio(ctx, folio, offset, bytes);
1133         if (IS_ERR(req))
1134                 goto out;
1135         nfs_inode_add_request(req);
1136 out:
1137         return req;
1138 }
1139
1140 static int nfs_writepage_setup(struct nfs_open_context *ctx,
1141                                struct folio *folio, unsigned int offset,
1142                                unsigned int count)
1143 {
1144         struct nfs_page *req;
1145
1146         req = nfs_setup_write_request(ctx, folio, offset, count);
1147         if (IS_ERR(req))
1148                 return PTR_ERR(req);
1149         /* Update file length */
1150         nfs_grow_file(folio, offset, count);
1151         nfs_mark_uptodate(req);
1152         nfs_mark_request_dirty(req);
1153         nfs_unlock_and_release_request(req);
1154         return 0;
1155 }
1156
1157 int nfs_flush_incompatible(struct file *file, struct folio *folio)
1158 {
1159         struct nfs_open_context *ctx = nfs_file_open_context(file);
1160         struct nfs_lock_context *l_ctx;
1161         struct file_lock_context *flctx = locks_inode_context(file_inode(file));
1162         struct nfs_page *req;
1163         int do_flush, status;
1164         /*
1165          * Look for a request corresponding to this page. If there
1166          * is one, and it belongs to another file, we flush it out
1167          * before we try to copy anything into the page. Do this
1168          * due to the lack of an ACCESS-type call in NFSv2.
1169          * Also do the same if we find a request from an existing
1170          * dropped page.
1171          */
1172         do {
1173                 req = nfs_folio_find_head_request(folio);
1174                 if (req == NULL)
1175                         return 0;
1176                 l_ctx = req->wb_lock_context;
1177                 do_flush = nfs_page_to_folio(req) != folio ||
1178                            !nfs_match_open_context(nfs_req_openctx(req), ctx);
1179                 if (l_ctx && flctx &&
1180                     !(list_empty_careful(&flctx->flc_posix) &&
1181                       list_empty_careful(&flctx->flc_flock))) {
1182                         do_flush |= l_ctx->lockowner != current->files;
1183                 }
1184                 nfs_release_request(req);
1185                 if (!do_flush)
1186                         return 0;
1187                 status = nfs_wb_folio(folio->mapping->host, folio);
1188         } while (status == 0);
1189         return status;
1190 }
1191
1192 /*
1193  * Avoid buffered writes when a open context credential's key would
1194  * expire soon.
1195  *
1196  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
1197  *
1198  * Return 0 and set a credential flag which triggers the inode to flush
1199  * and performs  NFS_FILE_SYNC writes if the key will expired within
1200  * RPC_KEY_EXPIRE_TIMEO.
1201  */
1202 int
1203 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
1204 {
1205         struct nfs_open_context *ctx = nfs_file_open_context(filp);
1206
1207         if (nfs_ctx_key_to_expire(ctx, inode) &&
1208             !rcu_access_pointer(ctx->ll_cred))
1209                 /* Already expired! */
1210                 return -EACCES;
1211         return 0;
1212 }
1213
1214 /*
1215  * Test if the open context credential key is marked to expire soon.
1216  */
1217 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx, struct inode *inode)
1218 {
1219         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
1220         struct rpc_cred *cred, *new, *old = NULL;
1221         struct auth_cred acred = {
1222                 .cred = ctx->cred,
1223         };
1224         bool ret = false;
1225
1226         rcu_read_lock();
1227         cred = rcu_dereference(ctx->ll_cred);
1228         if (cred && !(cred->cr_ops->crkey_timeout &&
1229                       cred->cr_ops->crkey_timeout(cred)))
1230                 goto out;
1231         rcu_read_unlock();
1232
1233         new = auth->au_ops->lookup_cred(auth, &acred, 0);
1234         if (new == cred) {
1235                 put_rpccred(new);
1236                 return true;
1237         }
1238         if (IS_ERR_OR_NULL(new)) {
1239                 new = NULL;
1240                 ret = true;
1241         } else if (new->cr_ops->crkey_timeout &&
1242                    new->cr_ops->crkey_timeout(new))
1243                 ret = true;
1244
1245         rcu_read_lock();
1246         old = rcu_dereference_protected(xchg(&ctx->ll_cred,
1247                                              RCU_INITIALIZER(new)), 1);
1248 out:
1249         rcu_read_unlock();
1250         put_rpccred(old);
1251         return ret;
1252 }
1253
1254 /*
1255  * If the page cache is marked as unsafe or invalid, then we can't rely on
1256  * the PageUptodate() flag. In this case, we will need to turn off
1257  * write optimisations that depend on the page contents being correct.
1258  */
1259 static bool nfs_folio_write_uptodate(struct folio *folio, unsigned int pagelen)
1260 {
1261         struct inode *inode = folio->mapping->host;
1262         struct nfs_inode *nfsi = NFS_I(inode);
1263
1264         if (nfs_have_delegated_attributes(inode))
1265                 goto out;
1266         if (nfsi->cache_validity &
1267             (NFS_INO_INVALID_CHANGE | NFS_INO_INVALID_SIZE))
1268                 return false;
1269         smp_rmb();
1270         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags) && pagelen != 0)
1271                 return false;
1272 out:
1273         if (nfsi->cache_validity & NFS_INO_INVALID_DATA && pagelen != 0)
1274                 return false;
1275         return folio_test_uptodate(folio) != 0;
1276 }
1277
1278 static bool
1279 is_whole_file_wrlock(struct file_lock *fl)
1280 {
1281         return fl->fl_start == 0 && fl->fl_end == OFFSET_MAX &&
1282                         lock_is_write(fl);
1283 }
1284
1285 /* If we know the page is up to date, and we're not using byte range locks (or
1286  * if we have the whole file locked for writing), it may be more efficient to
1287  * extend the write to cover the entire page in order to avoid fragmentation
1288  * inefficiencies.
1289  *
1290  * If the file is opened for synchronous writes then we can just skip the rest
1291  * of the checks.
1292  */
1293 static int nfs_can_extend_write(struct file *file, struct folio *folio,
1294                                 unsigned int pagelen)
1295 {
1296         struct inode *inode = file_inode(file);
1297         struct file_lock_context *flctx = locks_inode_context(inode);
1298         struct file_lock *fl;
1299         int ret;
1300
1301         if (file->f_flags & O_DSYNC)
1302                 return 0;
1303         if (!nfs_folio_write_uptodate(folio, pagelen))
1304                 return 0;
1305         if (nfs_have_write_delegation(inode))
1306                 return 1;
1307         if (!flctx || (list_empty_careful(&flctx->flc_flock) &&
1308                        list_empty_careful(&flctx->flc_posix)))
1309                 return 1;
1310
1311         /* Check to see if there are whole file write locks */
1312         ret = 0;
1313         spin_lock(&flctx->flc_lock);
1314         if (!list_empty(&flctx->flc_posix)) {
1315                 fl = list_first_entry(&flctx->flc_posix, struct file_lock,
1316                                         c.flc_list);
1317                 if (is_whole_file_wrlock(fl))
1318                         ret = 1;
1319         } else if (!list_empty(&flctx->flc_flock)) {
1320                 fl = list_first_entry(&flctx->flc_flock, struct file_lock,
1321                                         c.flc_list);
1322                 if (lock_is_write(fl))
1323                         ret = 1;
1324         }
1325         spin_unlock(&flctx->flc_lock);
1326         return ret;
1327 }
1328
1329 /*
1330  * Update and possibly write a cached page of an NFS file.
1331  *
1332  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
1333  * things with a page scheduled for an RPC call (e.g. invalidate it).
1334  */
1335 int nfs_update_folio(struct file *file, struct folio *folio,
1336                      unsigned int offset, unsigned int count)
1337 {
1338         struct nfs_open_context *ctx = nfs_file_open_context(file);
1339         struct address_space *mapping = folio->mapping;
1340         struct inode *inode = mapping->host;
1341         unsigned int pagelen = nfs_folio_length(folio);
1342         int             status = 0;
1343
1344         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
1345
1346         dprintk("NFS:       nfs_update_folio(%pD2 %d@%lld)\n", file, count,
1347                 (long long)(folio_pos(folio) + offset));
1348
1349         if (!count)
1350                 goto out;
1351
1352         if (nfs_can_extend_write(file, folio, pagelen)) {
1353                 unsigned int end = count + offset;
1354
1355                 offset = round_down(offset, PAGE_SIZE);
1356                 if (end < pagelen)
1357                         end = min(round_up(end, PAGE_SIZE), pagelen);
1358                 count = end - offset;
1359         }
1360
1361         status = nfs_writepage_setup(ctx, folio, offset, count);
1362         if (status < 0)
1363                 nfs_set_pageerror(mapping);
1364 out:
1365         dprintk("NFS:       nfs_update_folio returns %d (isize %lld)\n",
1366                         status, (long long)i_size_read(inode));
1367         return status;
1368 }
1369
1370 static int flush_task_priority(int how)
1371 {
1372         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1373                 case FLUSH_HIGHPRI:
1374                         return RPC_PRIORITY_HIGH;
1375                 case FLUSH_LOWPRI:
1376                         return RPC_PRIORITY_LOW;
1377         }
1378         return RPC_PRIORITY_NORMAL;
1379 }
1380
1381 static void nfs_initiate_write(struct nfs_pgio_header *hdr,
1382                                struct rpc_message *msg,
1383                                const struct nfs_rpc_ops *rpc_ops,
1384                                struct rpc_task_setup *task_setup_data, int how)
1385 {
1386         int priority = flush_task_priority(how);
1387
1388         if (IS_SWAPFILE(hdr->inode))
1389                 task_setup_data->flags |= RPC_TASK_SWAPPER;
1390         task_setup_data->priority = priority;
1391         rpc_ops->write_setup(hdr, msg, &task_setup_data->rpc_client);
1392         trace_nfs_initiate_write(hdr);
1393 }
1394
1395 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1396  * call this on each, which will prepare them to be retried on next
1397  * writeback using standard nfs.
1398  */
1399 static void nfs_redirty_request(struct nfs_page *req)
1400 {
1401         struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
1402
1403         /* Bump the transmission count */
1404         req->wb_nio++;
1405         nfs_mark_request_dirty(req);
1406         atomic_long_inc(&nfsi->redirtied_pages);
1407         nfs_page_end_writeback(req);
1408         nfs_release_request(req);
1409 }
1410
1411 static void nfs_async_write_error(struct list_head *head, int error)
1412 {
1413         struct nfs_page *req;
1414
1415         while (!list_empty(head)) {
1416                 req = nfs_list_entry(head->next);
1417                 nfs_list_remove_request(req);
1418                 if (nfs_error_is_fatal_on_server(error))
1419                         nfs_write_error(req, error);
1420                 else
1421                         nfs_redirty_request(req);
1422         }
1423 }
1424
1425 static void nfs_async_write_reschedule_io(struct nfs_pgio_header *hdr)
1426 {
1427         nfs_async_write_error(&hdr->pages, 0);
1428 }
1429
1430 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1431         .init_hdr = nfs_async_write_init,
1432         .error_cleanup = nfs_async_write_error,
1433         .completion = nfs_write_completion,
1434         .reschedule_io = nfs_async_write_reschedule_io,
1435 };
1436
1437 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1438                                struct inode *inode, int ioflags, bool force_mds,
1439                                const struct nfs_pgio_completion_ops *compl_ops)
1440 {
1441         struct nfs_server *server = NFS_SERVER(inode);
1442         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1443
1444 #ifdef CONFIG_NFS_V4_1
1445         if (server->pnfs_curr_ld && !force_mds)
1446                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1447 #endif
1448         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1449                         server->wsize, ioflags);
1450 }
1451 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1452
1453 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1454 {
1455         struct nfs_pgio_mirror *mirror;
1456
1457         if (pgio->pg_ops && pgio->pg_ops->pg_cleanup)
1458                 pgio->pg_ops->pg_cleanup(pgio);
1459
1460         pgio->pg_ops = &nfs_pgio_rw_ops;
1461
1462         nfs_pageio_stop_mirroring(pgio);
1463
1464         mirror = &pgio->pg_mirrors[0];
1465         mirror->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1466 }
1467 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1468
1469
1470 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1471 {
1472         struct nfs_commit_data *data = calldata;
1473
1474         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1475 }
1476
1477 static void nfs_writeback_check_extend(struct nfs_pgio_header *hdr,
1478                 struct nfs_fattr *fattr)
1479 {
1480         struct nfs_pgio_args *argp = &hdr->args;
1481         struct nfs_pgio_res *resp = &hdr->res;
1482         u64 size = argp->offset + resp->count;
1483
1484         if (!(fattr->valid & NFS_ATTR_FATTR_SIZE))
1485                 fattr->size = size;
1486         if (nfs_size_to_loff_t(fattr->size) < i_size_read(hdr->inode)) {
1487                 fattr->valid &= ~NFS_ATTR_FATTR_SIZE;
1488                 return;
1489         }
1490         if (size != fattr->size)
1491                 return;
1492         /* Set attribute barrier */
1493         nfs_fattr_set_barrier(fattr);
1494         /* ...and update size */
1495         fattr->valid |= NFS_ATTR_FATTR_SIZE;
1496 }
1497
1498 void nfs_writeback_update_inode(struct nfs_pgio_header *hdr)
1499 {
1500         struct nfs_fattr *fattr = &hdr->fattr;
1501         struct inode *inode = hdr->inode;
1502
1503         if (nfs_have_delegated_mtime(inode)) {
1504                 spin_lock(&inode->i_lock);
1505                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_BLOCKS);
1506                 spin_unlock(&inode->i_lock);
1507                 return;
1508         }
1509
1510         spin_lock(&inode->i_lock);
1511         nfs_writeback_check_extend(hdr, fattr);
1512         nfs_post_op_update_inode_force_wcc_locked(inode, fattr);
1513         spin_unlock(&inode->i_lock);
1514 }
1515 EXPORT_SYMBOL_GPL(nfs_writeback_update_inode);
1516
1517 /*
1518  * This function is called when the WRITE call is complete.
1519  */
1520 static int nfs_writeback_done(struct rpc_task *task,
1521                               struct nfs_pgio_header *hdr,
1522                               struct inode *inode)
1523 {
1524         int status;
1525
1526         /*
1527          * ->write_done will attempt to use post-op attributes to detect
1528          * conflicting writes by other clients.  A strict interpretation
1529          * of close-to-open would allow us to continue caching even if
1530          * another writer had changed the file, but some applications
1531          * depend on tighter cache coherency when writing.
1532          */
1533         status = NFS_PROTO(inode)->write_done(task, hdr);
1534         if (status != 0)
1535                 return status;
1536
1537         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, hdr->res.count);
1538         trace_nfs_writeback_done(task, hdr);
1539
1540         if (task->tk_status >= 0) {
1541                 enum nfs3_stable_how committed = hdr->res.verf->committed;
1542
1543                 if (committed == NFS_UNSTABLE) {
1544                         /*
1545                          * We have some uncommitted data on the server at
1546                          * this point, so ensure that we keep track of that
1547                          * fact irrespective of what later writes do.
1548                          */
1549                         set_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags);
1550                 }
1551
1552                 if (committed < hdr->args.stable) {
1553                         /* We tried a write call, but the server did not
1554                          * commit data to stable storage even though we
1555                          * requested it.
1556                          * Note: There is a known bug in Tru64 < 5.0 in which
1557                          *       the server reports NFS_DATA_SYNC, but performs
1558                          *       NFS_FILE_SYNC. We therefore implement this checking
1559                          *       as a dprintk() in order to avoid filling syslog.
1560                          */
1561                         static unsigned long    complain;
1562
1563                         /* Note this will print the MDS for a DS write */
1564                         if (time_before(complain, jiffies)) {
1565                                 dprintk("NFS:       faulty NFS server %s:"
1566                                         " (committed = %d) != (stable = %d)\n",
1567                                         NFS_SERVER(inode)->nfs_client->cl_hostname,
1568                                         committed, hdr->args.stable);
1569                                 complain = jiffies + 300 * HZ;
1570                         }
1571                 }
1572         }
1573
1574         /* Deal with the suid/sgid bit corner case */
1575         if (nfs_should_remove_suid(inode)) {
1576                 spin_lock(&inode->i_lock);
1577                 nfs_set_cache_invalid(inode, NFS_INO_INVALID_MODE);
1578                 spin_unlock(&inode->i_lock);
1579         }
1580         return 0;
1581 }
1582
1583 /*
1584  * This function is called when the WRITE call is complete.
1585  */
1586 static void nfs_writeback_result(struct rpc_task *task,
1587                                  struct nfs_pgio_header *hdr)
1588 {
1589         struct nfs_pgio_args    *argp = &hdr->args;
1590         struct nfs_pgio_res     *resp = &hdr->res;
1591
1592         if (resp->count < argp->count) {
1593                 static unsigned long    complain;
1594
1595                 /* This a short write! */
1596                 nfs_inc_stats(hdr->inode, NFSIOS_SHORTWRITE);
1597
1598                 /* Has the server at least made some progress? */
1599                 if (resp->count == 0) {
1600                         if (time_before(complain, jiffies)) {
1601                                 printk(KERN_WARNING
1602                                        "NFS: Server wrote zero bytes, expected %u.\n",
1603                                        argp->count);
1604                                 complain = jiffies + 300 * HZ;
1605                         }
1606                         nfs_set_pgio_error(hdr, -EIO, argp->offset);
1607                         task->tk_status = -EIO;
1608                         return;
1609                 }
1610
1611                 /* For non rpc-based layout drivers, retry-through-MDS */
1612                 if (!task->tk_ops) {
1613                         hdr->pnfs_error = -EAGAIN;
1614                         return;
1615                 }
1616
1617                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1618                 if (resp->verf->committed != NFS_UNSTABLE) {
1619                         /* Resend from where the server left off */
1620                         hdr->mds_offset += resp->count;
1621                         argp->offset += resp->count;
1622                         argp->pgbase += resp->count;
1623                         argp->count -= resp->count;
1624                 } else {
1625                         /* Resend as a stable write in order to avoid
1626                          * headaches in the case of a server crash.
1627                          */
1628                         argp->stable = NFS_FILE_SYNC;
1629                 }
1630                 resp->count = 0;
1631                 resp->verf->committed = 0;
1632                 rpc_restart_call_prepare(task);
1633         }
1634 }
1635
1636 static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
1637 {
1638         return wait_var_event_killable(&cinfo->rpcs_out,
1639                                        !atomic_read(&cinfo->rpcs_out));
1640 }
1641
1642 void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
1643 {
1644         atomic_inc(&cinfo->rpcs_out);
1645 }
1646
1647 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo)
1648 {
1649         if (atomic_dec_and_test(&cinfo->rpcs_out)) {
1650                 wake_up_var(&cinfo->rpcs_out);
1651                 return true;
1652         }
1653         return false;
1654 }
1655
1656 void nfs_commitdata_release(struct nfs_commit_data *data)
1657 {
1658         put_nfs_open_context(data->context);
1659         nfs_commit_free(data);
1660 }
1661 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1662
1663 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1664                         const struct nfs_rpc_ops *nfs_ops,
1665                         const struct rpc_call_ops *call_ops,
1666                         int how, int flags)
1667 {
1668         struct rpc_task *task;
1669         int priority = flush_task_priority(how);
1670         struct rpc_message msg = {
1671                 .rpc_argp = &data->args,
1672                 .rpc_resp = &data->res,
1673                 .rpc_cred = data->cred,
1674         };
1675         struct rpc_task_setup task_setup_data = {
1676                 .task = &data->task,
1677                 .rpc_client = clnt,
1678                 .rpc_message = &msg,
1679                 .callback_ops = call_ops,
1680                 .callback_data = data,
1681                 .workqueue = nfsiod_workqueue,
1682                 .flags = RPC_TASK_ASYNC | flags,
1683                 .priority = priority,
1684         };
1685
1686         if (nfs_server_capable(data->inode, NFS_CAP_MOVEABLE))
1687                 task_setup_data.flags |= RPC_TASK_MOVEABLE;
1688
1689         /* Set up the initial task struct.  */
1690         nfs_ops->commit_setup(data, &msg, &task_setup_data.rpc_client);
1691         trace_nfs_initiate_commit(data);
1692
1693         dprintk("NFS: initiated commit call\n");
1694
1695         task = rpc_run_task(&task_setup_data);
1696         if (IS_ERR(task))
1697                 return PTR_ERR(task);
1698         if (how & FLUSH_SYNC)
1699                 rpc_wait_for_completion_task(task);
1700         rpc_put_task(task);
1701         return 0;
1702 }
1703 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1704
1705 static loff_t nfs_get_lwb(struct list_head *head)
1706 {
1707         loff_t lwb = 0;
1708         struct nfs_page *req;
1709
1710         list_for_each_entry(req, head, wb_list)
1711                 if (lwb < (req_offset(req) + req->wb_bytes))
1712                         lwb = req_offset(req) + req->wb_bytes;
1713
1714         return lwb;
1715 }
1716
1717 /*
1718  * Set up the argument/result storage required for the RPC call.
1719  */
1720 void nfs_init_commit(struct nfs_commit_data *data,
1721                      struct list_head *head,
1722                      struct pnfs_layout_segment *lseg,
1723                      struct nfs_commit_info *cinfo)
1724 {
1725         struct nfs_page *first;
1726         struct nfs_open_context *ctx;
1727         struct inode *inode;
1728
1729         /* Set up the RPC argument and reply structs
1730          * NB: take care not to mess about with data->commit et al. */
1731
1732         if (head)
1733                 list_splice_init(head, &data->pages);
1734
1735         first = nfs_list_entry(data->pages.next);
1736         ctx = nfs_req_openctx(first);
1737         inode = d_inode(ctx->dentry);
1738
1739         data->inode       = inode;
1740         data->cred        = ctx->cred;
1741         data->lseg        = lseg; /* reference transferred */
1742         /* only set lwb for pnfs commit */
1743         if (lseg)
1744                 data->lwb = nfs_get_lwb(&data->pages);
1745         data->mds_ops     = &nfs_commit_ops;
1746         data->completion_ops = cinfo->completion_ops;
1747         data->dreq        = cinfo->dreq;
1748
1749         data->args.fh     = NFS_FH(data->inode);
1750         /* Note: we always request a commit of the entire inode */
1751         data->args.offset = 0;
1752         data->args.count  = 0;
1753         data->context     = get_nfs_open_context(ctx);
1754         data->res.fattr   = &data->fattr;
1755         data->res.verf    = &data->verf;
1756         nfs_fattr_init(&data->fattr);
1757         nfs_commit_begin(cinfo->mds);
1758 }
1759 EXPORT_SYMBOL_GPL(nfs_init_commit);
1760
1761 void nfs_retry_commit(struct list_head *page_list,
1762                       struct pnfs_layout_segment *lseg,
1763                       struct nfs_commit_info *cinfo,
1764                       u32 ds_commit_idx)
1765 {
1766         struct nfs_page *req;
1767
1768         while (!list_empty(page_list)) {
1769                 req = nfs_list_entry(page_list->next);
1770                 nfs_list_remove_request(req);
1771                 nfs_mark_request_commit(req, lseg, cinfo, ds_commit_idx);
1772                 nfs_folio_clear_commit(nfs_page_to_folio(req));
1773                 nfs_unlock_and_release_request(req);
1774         }
1775 }
1776 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1777
1778 static void nfs_commit_resched_write(struct nfs_commit_info *cinfo,
1779                                      struct nfs_page *req)
1780 {
1781         struct folio *folio = nfs_page_to_folio(req);
1782
1783         filemap_dirty_folio(folio_mapping(folio), folio);
1784 }
1785
1786 /*
1787  * Commit dirty pages
1788  */
1789 static int
1790 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1791                 struct nfs_commit_info *cinfo)
1792 {
1793         struct nfs_commit_data  *data;
1794         unsigned short task_flags = 0;
1795
1796         /* another commit raced with us */
1797         if (list_empty(head))
1798                 return 0;
1799
1800         data = nfs_commitdata_alloc();
1801         if (!data) {
1802                 nfs_retry_commit(head, NULL, cinfo, -1);
1803                 return -ENOMEM;
1804         }
1805
1806         /* Set up the argument struct */
1807         nfs_init_commit(data, head, NULL, cinfo);
1808         if (NFS_SERVER(inode)->nfs_client->cl_minorversion)
1809                 task_flags = RPC_TASK_MOVEABLE;
1810         return nfs_initiate_commit(NFS_CLIENT(inode), data, NFS_PROTO(inode),
1811                                    data->mds_ops, how,
1812                                    RPC_TASK_CRED_NOREF | task_flags);
1813 }
1814
1815 /*
1816  * COMMIT call returned
1817  */
1818 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1819 {
1820         struct nfs_commit_data  *data = calldata;
1821
1822         /* Call the NFS version-specific code */
1823         NFS_PROTO(data->inode)->commit_done(task, data);
1824         trace_nfs_commit_done(task, data);
1825 }
1826
1827 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1828 {
1829         const struct nfs_writeverf *verf = data->res.verf;
1830         struct nfs_page *req;
1831         int status = data->task.tk_status;
1832         struct nfs_commit_info cinfo;
1833         struct folio *folio;
1834
1835         while (!list_empty(&data->pages)) {
1836                 req = nfs_list_entry(data->pages.next);
1837                 nfs_list_remove_request(req);
1838                 folio = nfs_page_to_folio(req);
1839                 nfs_folio_clear_commit(folio);
1840
1841                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1842                         nfs_req_openctx(req)->dentry->d_sb->s_id,
1843                         (unsigned long long)NFS_FILEID(d_inode(nfs_req_openctx(req)->dentry)),
1844                         req->wb_bytes,
1845                         (long long)req_offset(req));
1846                 if (status < 0) {
1847                         if (folio) {
1848                                 trace_nfs_commit_error(data->inode, req,
1849                                                        status);
1850                                 nfs_mapping_set_error(folio, status);
1851                                 nfs_inode_remove_request(req);
1852                         }
1853                         dprintk_cont(", error = %d\n", status);
1854                         goto next;
1855                 }
1856
1857                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1858                  * returned by the server against all stored verfs. */
1859                 if (nfs_write_match_verf(verf, req)) {
1860                         /* We have a match */
1861                         if (folio)
1862                                 nfs_inode_remove_request(req);
1863                         dprintk_cont(" OK\n");
1864                         goto next;
1865                 }
1866                 /* We have a mismatch. Write the page again */
1867                 dprintk_cont(" mismatch\n");
1868                 nfs_mark_request_dirty(req);
1869                 atomic_long_inc(&NFS_I(data->inode)->redirtied_pages);
1870         next:
1871                 nfs_unlock_and_release_request(req);
1872                 /* Latency breaker */
1873                 cond_resched();
1874         }
1875
1876         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1877         nfs_commit_end(cinfo.mds);
1878 }
1879
1880 static void nfs_commit_release(void *calldata)
1881 {
1882         struct nfs_commit_data *data = calldata;
1883
1884         data->completion_ops->completion(data);
1885         nfs_commitdata_release(calldata);
1886 }
1887
1888 static const struct rpc_call_ops nfs_commit_ops = {
1889         .rpc_call_prepare = nfs_commit_prepare,
1890         .rpc_call_done = nfs_commit_done,
1891         .rpc_release = nfs_commit_release,
1892 };
1893
1894 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1895         .completion = nfs_commit_release_pages,
1896         .resched_write = nfs_commit_resched_write,
1897 };
1898
1899 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1900                             int how, struct nfs_commit_info *cinfo)
1901 {
1902         int status;
1903
1904         status = pnfs_commit_list(inode, head, how, cinfo);
1905         if (status == PNFS_NOT_ATTEMPTED)
1906                 status = nfs_commit_list(inode, head, how, cinfo);
1907         return status;
1908 }
1909
1910 static int __nfs_commit_inode(struct inode *inode, int how,
1911                 struct writeback_control *wbc)
1912 {
1913         LIST_HEAD(head);
1914         struct nfs_commit_info cinfo;
1915         int may_wait = how & FLUSH_SYNC;
1916         int ret, nscan;
1917
1918         how &= ~FLUSH_SYNC;
1919         nfs_init_cinfo_from_inode(&cinfo, inode);
1920         nfs_commit_begin(cinfo.mds);
1921         for (;;) {
1922                 ret = nscan = nfs_scan_commit(inode, &head, &cinfo);
1923                 if (ret <= 0)
1924                         break;
1925                 ret = nfs_generic_commit_list(inode, &head, how, &cinfo);
1926                 if (ret < 0)
1927                         break;
1928                 ret = 0;
1929                 if (wbc && wbc->sync_mode == WB_SYNC_NONE) {
1930                         if (nscan < wbc->nr_to_write)
1931                                 wbc->nr_to_write -= nscan;
1932                         else
1933                                 wbc->nr_to_write = 0;
1934                 }
1935                 if (nscan < INT_MAX)
1936                         break;
1937                 cond_resched();
1938         }
1939         nfs_commit_end(cinfo.mds);
1940         if (ret || !may_wait)
1941                 return ret;
1942         return wait_on_commit(cinfo.mds);
1943 }
1944
1945 int nfs_commit_inode(struct inode *inode, int how)
1946 {
1947         return __nfs_commit_inode(inode, how, NULL);
1948 }
1949 EXPORT_SYMBOL_GPL(nfs_commit_inode);
1950
1951 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1952 {
1953         struct nfs_inode *nfsi = NFS_I(inode);
1954         int flags = FLUSH_SYNC;
1955         int ret = 0;
1956
1957         if (wbc->sync_mode == WB_SYNC_NONE) {
1958                 /* no commits means nothing needs to be done */
1959                 if (!atomic_long_read(&nfsi->commit_info.ncommit))
1960                         goto check_requests_outstanding;
1961
1962                 /* Don't commit yet if this is a non-blocking flush and there
1963                  * are a lot of outstanding writes for this mapping.
1964                  */
1965                 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))
1966                         goto out_mark_dirty;
1967
1968                 /* don't wait for the COMMIT response */
1969                 flags = 0;
1970         }
1971
1972         ret = __nfs_commit_inode(inode, flags, wbc);
1973         if (!ret) {
1974                 if (flags & FLUSH_SYNC)
1975                         return 0;
1976         } else if (atomic_long_read(&nfsi->commit_info.ncommit))
1977                 goto out_mark_dirty;
1978
1979 check_requests_outstanding:
1980         if (!atomic_read(&nfsi->commit_info.rpcs_out))
1981                 return ret;
1982 out_mark_dirty:
1983         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1984         return ret;
1985 }
1986 EXPORT_SYMBOL_GPL(nfs_write_inode);
1987
1988 /*
1989  * Wrapper for filemap_write_and_wait_range()
1990  *
1991  * Needed for pNFS in order to ensure data becomes visible to the
1992  * client.
1993  */
1994 int nfs_filemap_write_and_wait_range(struct address_space *mapping,
1995                 loff_t lstart, loff_t lend)
1996 {
1997         int ret;
1998
1999         ret = filemap_write_and_wait_range(mapping, lstart, lend);
2000         if (ret == 0)
2001                 ret = pnfs_sync_inode(mapping->host, true);
2002         return ret;
2003 }
2004 EXPORT_SYMBOL_GPL(nfs_filemap_write_and_wait_range);
2005
2006 /*
2007  * flush the inode to disk.
2008  */
2009 int nfs_wb_all(struct inode *inode)
2010 {
2011         int ret;
2012
2013         trace_nfs_writeback_inode_enter(inode);
2014
2015         ret = filemap_write_and_wait(inode->i_mapping);
2016         if (ret)
2017                 goto out;
2018         ret = nfs_commit_inode(inode, FLUSH_SYNC);
2019         if (ret < 0)
2020                 goto out;
2021         pnfs_sync_inode(inode, true);
2022         ret = 0;
2023
2024 out:
2025         trace_nfs_writeback_inode_exit(inode, ret);
2026         return ret;
2027 }
2028 EXPORT_SYMBOL_GPL(nfs_wb_all);
2029
2030 int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio)
2031 {
2032         struct nfs_page *req;
2033         int ret = 0;
2034
2035         folio_wait_writeback(folio);
2036
2037         /* blocking call to cancel all requests and join to a single (head)
2038          * request */
2039         req = nfs_lock_and_join_requests(folio);
2040
2041         if (IS_ERR(req)) {
2042                 ret = PTR_ERR(req);
2043         } else if (req) {
2044                 /* all requests from this folio have been cancelled by
2045                  * nfs_lock_and_join_requests, so just remove the head
2046                  * request from the inode / page_private pointer and
2047                  * release it */
2048                 nfs_inode_remove_request(req);
2049                 nfs_unlock_and_release_request(req);
2050         }
2051
2052         return ret;
2053 }
2054
2055 /**
2056  * nfs_wb_folio - Write back all requests on one page
2057  * @inode: pointer to page
2058  * @folio: pointer to folio
2059  *
2060  * Assumes that the folio has been locked by the caller, and will
2061  * not unlock it.
2062  */
2063 int nfs_wb_folio(struct inode *inode, struct folio *folio)
2064 {
2065         loff_t range_start = folio_pos(folio);
2066         size_t len = folio_size(folio);
2067         struct writeback_control wbc = {
2068                 .sync_mode = WB_SYNC_ALL,
2069                 .nr_to_write = 0,
2070                 .range_start = range_start,
2071                 .range_end = range_start + len - 1,
2072         };
2073         int ret;
2074
2075         trace_nfs_writeback_folio(inode, range_start, len);
2076
2077         for (;;) {
2078                 folio_wait_writeback(folio);
2079                 if (folio_clear_dirty_for_io(folio)) {
2080                         ret = nfs_writepage_locked(folio, &wbc);
2081                         if (ret < 0)
2082                                 goto out_error;
2083                         continue;
2084                 }
2085                 ret = 0;
2086                 if (!folio_test_private(folio))
2087                         break;
2088                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
2089                 if (ret < 0)
2090                         goto out_error;
2091         }
2092 out_error:
2093         trace_nfs_writeback_folio_done(inode, range_start, len, ret);
2094         return ret;
2095 }
2096
2097 #ifdef CONFIG_MIGRATION
2098 int nfs_migrate_folio(struct address_space *mapping, struct folio *dst,
2099                 struct folio *src, enum migrate_mode mode)
2100 {
2101         /*
2102          * If the private flag is set, the folio is currently associated with
2103          * an in-progress read or write request. Don't try to migrate it.
2104          *
2105          * FIXME: we could do this in principle, but we'll need a way to ensure
2106          *        that we can safely release the inode reference while holding
2107          *        the folio lock.
2108          */
2109         if (folio_test_private(src))
2110                 return -EBUSY;
2111
2112         if (folio_test_private_2(src)) { /* [DEPRECATED] */
2113                 if (mode == MIGRATE_ASYNC)
2114                         return -EBUSY;
2115                 folio_wait_private_2(src);
2116         }
2117
2118         return migrate_folio(mapping, dst, src, mode);
2119 }
2120 #endif
2121
2122 int __init nfs_init_writepagecache(void)
2123 {
2124         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
2125                                              sizeof(struct nfs_pgio_header),
2126                                              0, SLAB_HWCACHE_ALIGN,
2127                                              NULL);
2128         if (nfs_wdata_cachep == NULL)
2129                 return -ENOMEM;
2130
2131         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
2132                                                      nfs_wdata_cachep);
2133         if (nfs_wdata_mempool == NULL)
2134                 goto out_destroy_write_cache;
2135
2136         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
2137                                              sizeof(struct nfs_commit_data),
2138                                              0, SLAB_HWCACHE_ALIGN,
2139                                              NULL);
2140         if (nfs_cdata_cachep == NULL)
2141                 goto out_destroy_write_mempool;
2142
2143         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
2144                                                       nfs_cdata_cachep);
2145         if (nfs_commit_mempool == NULL)
2146                 goto out_destroy_commit_cache;
2147
2148         /*
2149          * NFS congestion size, scale with available memory.
2150          *
2151          *  64MB:    8192k
2152          * 128MB:   11585k
2153          * 256MB:   16384k
2154          * 512MB:   23170k
2155          *   1GB:   32768k
2156          *   2GB:   46340k
2157          *   4GB:   65536k
2158          *   8GB:   92681k
2159          *  16GB:  131072k
2160          *
2161          * This allows larger machines to have larger/more transfers.
2162          * Limit the default to 256M
2163          */
2164         nfs_congestion_kb = (16*int_sqrt(totalram_pages())) << (PAGE_SHIFT-10);
2165         if (nfs_congestion_kb > 256*1024)
2166                 nfs_congestion_kb = 256*1024;
2167
2168         return 0;
2169
2170 out_destroy_commit_cache:
2171         kmem_cache_destroy(nfs_cdata_cachep);
2172 out_destroy_write_mempool:
2173         mempool_destroy(nfs_wdata_mempool);
2174 out_destroy_write_cache:
2175         kmem_cache_destroy(nfs_wdata_cachep);
2176         return -ENOMEM;
2177 }
2178
2179 void nfs_destroy_writepagecache(void)
2180 {
2181         mempool_destroy(nfs_commit_mempool);
2182         kmem_cache_destroy(nfs_cdata_cachep);
2183         mempool_destroy(nfs_wdata_mempool);
2184         kmem_cache_destroy(nfs_wdata_cachep);
2185 }
2186
2187 static const struct nfs_rw_ops nfs_rw_write_ops = {
2188         .rw_alloc_header        = nfs_writehdr_alloc,
2189         .rw_free_header         = nfs_writehdr_free,
2190         .rw_done                = nfs_writeback_done,
2191         .rw_result              = nfs_writeback_result,
2192         .rw_initiate            = nfs_initiate_write,
2193 };