fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #include "nfstrace.h"
  35
  36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  37
  38 #define MIN_POOL_WRITE          (32)
  39 #define MIN_POOL_COMMIT         (4)
  40
  41 /*
  42  * Local function declarations
  43  */
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_commit_ops;
  46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  48 static const struct nfs_rw_ops nfs_rw_write_ops;
  49
  50 static struct kmem_cache *nfs_wdata_cachep;
  51 static mempool_t *nfs_wdata_mempool;
  52 static struct kmem_cache *nfs_cdata_cachep;
  53 static mempool_t *nfs_commit_mempool;
  54
  55 struct nfs_commit_data *nfs_commitdata_alloc(void)
  56 {
  57         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  58
  59         if (p) {
  60                 memset(p, 0, sizeof(*p));
  61                 INIT_LIST_HEAD(&p->pages);
  62         }
  63         return p;
  64 }
  65 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  66
  67 void nfs_commit_free(struct nfs_commit_data *p)
  68 {
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71 EXPORT_SYMBOL_GPL(nfs_commit_free);
  72
  73 static struct nfs_rw_header *nfs_writehdr_alloc(void)
  74 {
  75         struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  76
  77         if (p)
  78                 memset(p, 0, sizeof(*p));
  79         return p;
  80 }
  81
  82 static void nfs_writehdr_free(struct nfs_rw_header *whdr)
  83 {
  84         mempool_free(whdr, nfs_wdata_mempool);
  85 }
  86
  87 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  88 {
  89         ctx->error = error;
  90         smp_wmb();
  91         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
  92 }
  93
  94 static struct nfs_page *
  95 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
  96 {
  97         struct nfs_page *req = NULL;
  98
  99         if (PagePrivate(page))
 100                 req = (struct nfs_page *)page_private(page);
 101         else if (unlikely(PageSwapCache(page))) {
 102                 struct nfs_page *freq, *t;
 103
 104                 /* Linearly search the commit list for the correct req */
 105                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 106                         if (freq->wb_page == page) {
 107                                 req = freq;
 108                                 break;
 109                         }
 110                 }
 111         }
 112
 113         if (req)
 114                 kref_get(&req->wb_kref);
 115
 116         return req;
 117 }
 118
 119 static struct nfs_page *nfs_page_find_request(struct page *page)
 120 {
 121         struct inode *inode = page_file_mapping(page)->host;
 122         struct nfs_page *req = NULL;
 123
 124         spin_lock(&inode->i_lock);
 125         req = nfs_page_find_request_locked(NFS_I(inode), page);
 126         spin_unlock(&inode->i_lock);
 127         return req;
 128 }
 129
 130 /* Adjust the file length if we're writing beyond the end */
 131 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 132 {
 133         struct inode *inode = page_file_mapping(page)->host;
 134         loff_t end, i_size;
 135         pgoff_t end_index;
 136
 137         spin_lock(&inode->i_lock);
 138         i_size = i_size_read(inode);
 139         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 140         if (i_size > 0 && page_file_index(page) < end_index)
 141                 goto out;
 142         end = page_file_offset(page) + ((loff_t)offset+count);
 143         if (i_size >= end)
 144                 goto out;
 145         i_size_write(inode, end);
 146         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 147 out:
 148         spin_unlock(&inode->i_lock);
 149 }
 150
 151 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 152 static void nfs_set_pageerror(struct page *page)
 153 {
 154         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 155 }
 156
 157 /*
 158  * nfs_page_group_search_locked
 159  * @head - head request of page group
 160  * @page_offset - offset into page
 161  *
 162  * Search page group with head @head to find a request that contains the
 163  * page offset @page_offset.
 164  *
 165  * Returns a pointer to the first matching nfs request, or NULL if no
 166  * match is found.
 167  *
 168  * Must be called with the page group lock held
 169  */
 170 static struct nfs_page *
 171 nfs_page_group_search_locked(struct nfs_page *head, unsigned int page_offset)
 172 {
 173         struct nfs_page *req;
 174
 175         WARN_ON_ONCE(head != head->wb_head);
 176         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_head->wb_flags));
 177
 178         req = head;
 179         do {
 180                 if (page_offset >= req->wb_pgbase &&
 181                     page_offset < (req->wb_pgbase + req->wb_bytes))
 182                         return req;
 183
 184                 req = req->wb_this_page;
 185         } while (req != head);
 186
 187         return NULL;
 188 }
 189
 190 /*
 191  * nfs_page_group_covers_page
 192  * @head - head request of page group
 193  *
 194  * Return true if the page group with head @head covers the whole page,
 195  * returns false otherwise
 196  */
 197 static bool nfs_page_group_covers_page(struct nfs_page *req)
 198 {
 199         struct nfs_page *tmp;
 200         unsigned int pos = 0;
 201         unsigned int len = nfs_page_length(req->wb_page);
 202
 203         nfs_page_group_lock(req);
 204
 205         do {
 206                 tmp = nfs_page_group_search_locked(req->wb_head, pos);
 207                 if (tmp) {
 208                         /* no way this should happen */
 209                         WARN_ON_ONCE(tmp->wb_pgbase != pos);
 210                         pos += tmp->wb_bytes - (pos - tmp->wb_pgbase);
 211                 }
 212         } while (tmp && pos < len);
 213
 214         nfs_page_group_unlock(req);
 215         WARN_ON_ONCE(pos > len);
 216         return pos == len;
 217 }
 218
 219 /* We can set the PG_uptodate flag if we see that a write request
 220  * covers the full page.
 221  */
 222 static void nfs_mark_uptodate(struct nfs_page *req)
 223 {
 224         if (PageUptodate(req->wb_page))
 225                 return;
 226         if (!nfs_page_group_covers_page(req))
 227                 return;
 228         SetPageUptodate(req->wb_page);
 229 }
 230
 231 static int wb_priority(struct writeback_control *wbc)
 232 {
 233         if (wbc->for_reclaim)
 234                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 235         if (wbc->for_kupdate || wbc->for_background)
 236                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 237         return FLUSH_COND_STABLE;
 238 }
 239
 240 /*
 241  * NFS congestion control
 242  */
 243
 244 int nfs_congestion_kb;
 245
 246 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 247 #define NFS_CONGESTION_OFF_THRESH       \
 248         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 249
 250 static void nfs_set_page_writeback(struct page *page)
 251 {
 252         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 253         int ret = test_set_page_writeback(page);
 254
 255         WARN_ON_ONCE(ret != 0);
 256
 257         if (atomic_long_inc_return(&nfss->writeback) >
 258                         NFS_CONGESTION_ON_THRESH) {
 259                 set_bdi_congested(&nfss->backing_dev_info,
 260                                         BLK_RW_ASYNC);
 261         }
 262 }
 263
 264 static void nfs_end_page_writeback(struct nfs_page *req)
 265 {
 266         struct inode *inode = page_file_mapping(req->wb_page)->host;
 267         struct nfs_server *nfss = NFS_SERVER(inode);
 268
 269         if (!nfs_page_group_sync_on_bit(req, PG_WB_END))
 270                 return;
 271
 272         end_page_writeback(req->wb_page);
 273         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 274                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 275 }
 276
 277 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 278 {
 279         struct inode *inode = page_file_mapping(page)->host;
 280         struct nfs_page *req;
 281         int ret;
 282
 283         spin_lock(&inode->i_lock);
 284         for (;;) {
 285                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 286                 if (req == NULL)
 287                         break;
 288                 if (nfs_lock_request(req))
 289                         break;
 290                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 291                  *       then the call to nfs_lock_request() will always
 292                  *       succeed provided that someone hasn't already marked the
 293                  *       request as dirty (in which case we don't care).
 294                  */
 295                 spin_unlock(&inode->i_lock);
 296                 if (!nonblock)
 297                         ret = nfs_wait_on_request(req);
 298                 else
 299                         ret = -EAGAIN;
 300                 nfs_release_request(req);
 301                 if (ret != 0)
 302                         return ERR_PTR(ret);
 303                 spin_lock(&inode->i_lock);
 304         }
 305         spin_unlock(&inode->i_lock);
 306         return req;
 307 }
 308
 309 /*
 310  * Find an associated nfs write request, and prepare to flush it out
 311  * May return an error if the user signalled nfs_wait_on_request().
 312  */
 313 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 314                                 struct page *page, bool nonblock)
 315 {
 316         struct nfs_page *req;
 317         int ret = 0;
 318
 319         req = nfs_find_and_lock_request(page, nonblock);
 320         if (!req)
 321                 goto out;
 322         ret = PTR_ERR(req);
 323         if (IS_ERR(req))
 324                 goto out;
 325
 326         nfs_set_page_writeback(page);
 327         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 328
 329         ret = 0;
 330         if (!nfs_pageio_add_request(pgio, req)) {
 331                 nfs_redirty_request(req);
 332                 ret = pgio->pg_error;
 333         }
 334 out:
 335         return ret;
 336 }
 337
 338 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 339 {
 340         struct inode *inode = page_file_mapping(page)->host;
 341         int ret;
 342
 343         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 344         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 345
 346         nfs_pageio_cond_complete(pgio, page_file_index(page));
 347         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 348         if (ret == -EAGAIN) {
 349                 redirty_page_for_writepage(wbc, page);
 350                 ret = 0;
 351         }
 352         return ret;
 353 }
 354
 355 /*
 356  * Write an mmapped page to the server.
 357  */
 358 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 359 {
 360         struct nfs_pageio_descriptor pgio;
 361         int err;
 362
 363         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 364                                 false, &nfs_async_write_completion_ops);
 365         err = nfs_do_writepage(page, wbc, &pgio);
 366         nfs_pageio_complete(&pgio);
 367         if (err < 0)
 368                 return err;
 369         if (pgio.pg_error < 0)
 370                 return pgio.pg_error;
 371         return 0;
 372 }
 373
 374 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 375 {
 376         int ret;
 377
 378         ret = nfs_writepage_locked(page, wbc);
 379         unlock_page(page);
 380         return ret;
 381 }
 382
 383 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 384 {
 385         int ret;
 386
 387         ret = nfs_do_writepage(page, wbc, data);
 388         unlock_page(page);
 389         return ret;
 390 }
 391
 392 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 393 {
 394         struct inode *inode = mapping->host;
 395         unsigned long *bitlock = &NFS_I(inode)->flags;
 396         struct nfs_pageio_descriptor pgio;
 397         int err;
 398
 399         /* Stop dirtying of new pages while we sync */
 400         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 401                         nfs_wait_bit_killable, TASK_KILLABLE);
 402         if (err)
 403                 goto out_err;
 404
 405         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 406
 407         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 408                                 &nfs_async_write_completion_ops);
 409         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 410         nfs_pageio_complete(&pgio);
 411
 412         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 413         smp_mb__after_atomic();
 414         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 415
 416         if (err < 0)
 417                 goto out_err;
 418         err = pgio.pg_error;
 419         if (err < 0)
 420                 goto out_err;
 421         return 0;
 422 out_err:
 423         return err;
 424 }
 425
 426 /*
 427  * Insert a write request into an inode
 428  */
 429 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 430 {
 431         struct nfs_inode *nfsi = NFS_I(inode);
 432
 433         WARN_ON_ONCE(req->wb_this_page != req);
 434
 435         /* Lock the request! */
 436         nfs_lock_request(req);
 437
 438         spin_lock(&inode->i_lock);
 439         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 440                 inode->i_version++;
 441         /*
 442          * Swap-space should not get truncated. Hence no need to plug the race
 443          * with invalidate/truncate.
 444          */
 445         if (likely(!PageSwapCache(req->wb_page))) {
 446                 set_bit(PG_MAPPED, &req->wb_flags);
 447                 SetPagePrivate(req->wb_page);
 448                 set_page_private(req->wb_page, (unsigned long)req);
 449         }
 450         nfsi->npages++;
 451         set_bit(PG_INODE_REF, &req->wb_flags);
 452         kref_get(&req->wb_kref);
 453         spin_unlock(&inode->i_lock);
 454 }
 455
 456 /*
 457  * Remove a write request from an inode
 458  */
 459 static void nfs_inode_remove_request(struct nfs_page *req)
 460 {
 461         struct inode *inode = req->wb_context->dentry->d_inode;
 462         struct nfs_inode *nfsi = NFS_I(inode);
 463         struct nfs_page *head;
 464
 465         if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
 466                 head = req->wb_head;
 467
 468                 spin_lock(&inode->i_lock);
 469                 if (likely(!PageSwapCache(head->wb_page))) {
 470                         set_page_private(head->wb_page, 0);
 471                         ClearPagePrivate(head->wb_page);
 472                         clear_bit(PG_MAPPED, &head->wb_flags);
 473                 }
 474                 nfsi->npages--;
 475                 spin_unlock(&inode->i_lock);
 476         }
 477         nfs_release_request(req);
 478 }
 479
 480 static void
 481 nfs_mark_request_dirty(struct nfs_page *req)
 482 {
 483         __set_page_dirty_nobuffers(req->wb_page);
 484 }
 485
 486 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 487 /**
 488  * nfs_request_add_commit_list - add request to a commit list
 489  * @req: pointer to a struct nfs_page
 490  * @dst: commit list head
 491  * @cinfo: holds list lock and accounting info
 492  *
 493  * This sets the PG_CLEAN bit, updates the cinfo count of
 494  * number of outstanding requests requiring a commit as well as
 495  * the MM page stats.
 496  *
 497  * The caller must _not_ hold the cinfo->lock, but must be
 498  * holding the nfs_page lock.
 499  */
 500 void
 501 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 502                             struct nfs_commit_info *cinfo)
 503 {
 504         set_bit(PG_CLEAN, &(req)->wb_flags);
 505         spin_lock(cinfo->lock);
 506         nfs_list_add_request(req, dst);
 507         cinfo->mds->ncommit++;
 508         spin_unlock(cinfo->lock);
 509         if (!cinfo->dreq) {
 510                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 511                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 512                              BDI_RECLAIMABLE);
 513                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 514                                    I_DIRTY_DATASYNC);
 515         }
 516 }
 517 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 518
 519 /**
 520  * nfs_request_remove_commit_list - Remove request from a commit list
 521  * @req: pointer to a nfs_page
 522  * @cinfo: holds list lock and accounting info
 523  *
 524  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 525  * number of outstanding requests requiring a commit
 526  * It does not update the MM page stats.
 527  *
 528  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 529  */
 530 void
 531 nfs_request_remove_commit_list(struct nfs_page *req,
 532                                struct nfs_commit_info *cinfo)
 533 {
 534         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 535                 return;
 536         nfs_list_remove_request(req);
 537         cinfo->mds->ncommit--;
 538 }
 539 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 540
 541 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 542                                       struct inode *inode)
 543 {
 544         cinfo->lock = &inode->i_lock;
 545         cinfo->mds = &NFS_I(inode)->commit_info;
 546         cinfo->ds = pnfs_get_ds_info(inode);
 547         cinfo->dreq = NULL;
 548         cinfo->completion_ops = &nfs_commit_completion_ops;
 549 }
 550
 551 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 552                     struct inode *inode,
 553                     struct nfs_direct_req *dreq)
 554 {
 555         if (dreq)
 556                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 557         else
 558                 nfs_init_cinfo_from_inode(cinfo, inode);
 559 }
 560 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 561
 562 /*
 563  * Add a request to the inode's commit list.
 564  */
 565 void
 566 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 567                         struct nfs_commit_info *cinfo)
 568 {
 569         if (pnfs_mark_request_commit(req, lseg, cinfo))
 570                 return;
 571         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 572 }
 573
 574 static void
 575 nfs_clear_page_commit(struct page *page)
 576 {
 577         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 578         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 579 }
 580
 581 static void
 582 nfs_clear_request_commit(struct nfs_page *req)
 583 {
 584         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 585                 struct inode *inode = req->wb_context->dentry->d_inode;
 586                 struct nfs_commit_info cinfo;
 587
 588                 nfs_init_cinfo_from_inode(&cinfo, inode);
 589                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 590                         spin_lock(cinfo.lock);
 591                         nfs_request_remove_commit_list(req, &cinfo);
 592                         spin_unlock(cinfo.lock);
 593                 }
 594                 nfs_clear_page_commit(req->wb_page);
 595         }
 596 }
 597
 598 static inline
 599 int nfs_write_need_commit(struct nfs_pgio_data *data)
 600 {
 601         if (data->verf.committed == NFS_DATA_SYNC)
 602                 return data->header->lseg == NULL;
 603         return data->verf.committed != NFS_FILE_SYNC;
 604 }
 605
 606 #else
 607 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 608                                       struct inode *inode)
 609 {
 610 }
 611
 612 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 613                     struct inode *inode,
 614                     struct nfs_direct_req *dreq)
 615 {
 616 }
 617
 618 void
 619 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 620                         struct nfs_commit_info *cinfo)
 621 {
 622 }
 623
 624 static void
 625 nfs_clear_request_commit(struct nfs_page *req)
 626 {
 627 }
 628
 629 static inline
 630 int nfs_write_need_commit(struct nfs_pgio_data *data)
 631 {
 632         return 0;
 633 }
 634
 635 #endif
 636
 637 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 638 {
 639         struct nfs_commit_info cinfo;
 640         unsigned long bytes = 0;
 641         bool do_destroy;
 642
 643         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 644                 goto out;
 645         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 646         while (!list_empty(&hdr->pages)) {
 647                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 648
 649                 bytes += req->wb_bytes;
 650                 nfs_list_remove_request(req);
 651                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 652                     (hdr->good_bytes < bytes)) {
 653                         nfs_set_pageerror(req->wb_page);
 654                         nfs_context_set_write_error(req->wb_context, hdr->error);
 655                         goto remove_req;
 656                 }
 657                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 658                         nfs_mark_request_dirty(req);
 659                         goto next;
 660                 }
 661                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 662                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 663                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 664                         goto next;
 665                 }
 666 remove_req:
 667                 nfs_inode_remove_request(req);
 668 next:
 669                 nfs_unlock_request(req);
 670                 nfs_end_page_writeback(req);
 671                 do_destroy = !test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags);
 672                 nfs_release_request(req);
 673         }
 674 out:
 675         hdr->release(hdr);
 676 }
 677
 678 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 679 unsigned long
 680 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 681 {
 682         return cinfo->mds->ncommit;
 683 }
 684
 685 /* cinfo->lock held by caller */
 686 int
 687 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 688                      struct nfs_commit_info *cinfo, int max)
 689 {
 690         struct nfs_page *req, *tmp;
 691         int ret = 0;
 692
 693         list_for_each_entry_safe(req, tmp, src, wb_list) {
 694                 if (!nfs_lock_request(req))
 695                         continue;
 696                 kref_get(&req->wb_kref);
 697                 if (cond_resched_lock(cinfo->lock))
 698                         list_safe_reset_next(req, tmp, wb_list);
 699                 nfs_request_remove_commit_list(req, cinfo);
 700                 nfs_list_add_request(req, dst);
 701                 ret++;
 702                 if ((ret == max) && !cinfo->dreq)
 703                         break;
 704         }
 705         return ret;
 706 }
 707
 708 /*
 709  * nfs_scan_commit - Scan an inode for commit requests
 710  * @inode: NFS inode to scan
 711  * @dst: mds destination list
 712  * @cinfo: mds and ds lists of reqs ready to commit
 713  *
 714  * Moves requests from the inode's 'commit' request list.
 715  * The requests are *not* checked to ensure that they form a contiguous set.
 716  */
 717 int
 718 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 719                 struct nfs_commit_info *cinfo)
 720 {
 721         int ret = 0;
 722
 723         spin_lock(cinfo->lock);
 724         if (cinfo->mds->ncommit > 0) {
 725                 const int max = INT_MAX;
 726
 727                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 728                                            cinfo, max);
 729                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 730         }
 731         spin_unlock(cinfo->lock);
 732         return ret;
 733 }
 734
 735 #else
 736 unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 737 {
 738         return 0;
 739 }
 740
 741 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 742                     struct nfs_commit_info *cinfo)
 743 {
 744         return 0;
 745 }
 746 #endif
 747
 748 /*
 749  * Search for an existing write request, and attempt to update
 750  * it to reflect a new dirty region on a given page.
 751  *
 752  * If the attempt fails, then the existing request is flushed out
 753  * to disk.
 754  */
 755 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 756                 struct page *page,
 757                 unsigned int offset,
 758                 unsigned int bytes)
 759 {
 760         struct nfs_page *req;
 761         unsigned int rqend;
 762         unsigned int end;
 763         int error;
 764
 765         if (!PagePrivate(page))
 766                 return NULL;
 767
 768         end = offset + bytes;
 769         spin_lock(&inode->i_lock);
 770
 771         for (;;) {
 772                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 773                 if (req == NULL)
 774                         goto out_unlock;
 775
 776                 /* should be handled by nfs_flush_incompatible */
 777                 WARN_ON_ONCE(req->wb_head != req);
 778                 WARN_ON_ONCE(req->wb_this_page != req);
 779
 780                 rqend = req->wb_offset + req->wb_bytes;
 781                 /*
 782                  * Tell the caller to flush out the request if
 783                  * the offsets are non-contiguous.
 784                  * Note: nfs_flush_incompatible() will already
 785                  * have flushed out requests having wrong owners.
 786                  */
 787                 if (offset > rqend
 788                     || end < req->wb_offset)
 789                         goto out_flushme;
 790
 791                 if (nfs_lock_request(req))
 792                         break;
 793
 794                 /* The request is locked, so wait and then retry */
 795                 spin_unlock(&inode->i_lock);
 796                 error = nfs_wait_on_request(req);
 797                 nfs_release_request(req);
 798                 if (error != 0)
 799                         goto out_err;
 800                 spin_lock(&inode->i_lock);
 801         }
 802
 803         /* Okay, the request matches. Update the region */
 804         if (offset < req->wb_offset) {
 805                 req->wb_offset = offset;
 806                 req->wb_pgbase = offset;
 807         }
 808         if (end > rqend)
 809                 req->wb_bytes = end - req->wb_offset;
 810         else
 811                 req->wb_bytes = rqend - req->wb_offset;
 812 out_unlock:
 813         spin_unlock(&inode->i_lock);
 814         if (req)
 815                 nfs_clear_request_commit(req);
 816         return req;
 817 out_flushme:
 818         spin_unlock(&inode->i_lock);
 819         nfs_release_request(req);
 820         error = nfs_wb_page(inode, page);
 821 out_err:
 822         return ERR_PTR(error);
 823 }
 824
 825 /*
 826  * Try to update an existing write request, or create one if there is none.
 827  *
 828  * Note: Should always be called with the Page Lock held to prevent races
 829  * if we have to add a new request. Also assumes that the caller has
 830  * already called nfs_flush_incompatible() if necessary.
 831  */
 832 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 833                 struct page *page, unsigned int offset, unsigned int bytes)
 834 {
 835         struct inode *inode = page_file_mapping(page)->host;
 836         struct nfs_page *req;
 837
 838         req = nfs_try_to_update_request(inode, page, offset, bytes);
 839         if (req != NULL)
 840                 goto out;
 841         req = nfs_create_request(ctx, page, NULL, offset, bytes);
 842         if (IS_ERR(req))
 843                 goto out;
 844         nfs_inode_add_request(inode, req);
 845 out:
 846         return req;
 847 }
 848
 849 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 850                 unsigned int offset, unsigned int count)
 851 {
 852         struct nfs_page *req;
 853
 854         req = nfs_setup_write_request(ctx, page, offset, count);
 855         if (IS_ERR(req))
 856                 return PTR_ERR(req);
 857         /* Update file length */
 858         nfs_grow_file(page, offset, count);
 859         nfs_mark_uptodate(req);
 860         nfs_mark_request_dirty(req);
 861         nfs_unlock_and_release_request(req);
 862         return 0;
 863 }
 864
 865 int nfs_flush_incompatible(struct file *file, struct page *page)
 866 {
 867         struct nfs_open_context *ctx = nfs_file_open_context(file);
 868         struct nfs_lock_context *l_ctx;
 869         struct nfs_page *req;
 870         int do_flush, status;
 871         /*
 872          * Look for a request corresponding to this page. If there
 873          * is one, and it belongs to another file, we flush it out
 874          * before we try to copy anything into the page. Do this
 875          * due to the lack of an ACCESS-type call in NFSv2.
 876          * Also do the same if we find a request from an existing
 877          * dropped page.
 878          */
 879         do {
 880                 req = nfs_page_find_request(page);
 881                 if (req == NULL)
 882                         return 0;
 883                 l_ctx = req->wb_lock_context;
 884                 do_flush = req->wb_page != page || req->wb_context != ctx;
 885                 /* for now, flush if more than 1 request in page_group */
 886                 do_flush |= req->wb_this_page != req;
 887                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 888                         do_flush |= l_ctx->lockowner.l_owner != current->files
 889                                 || l_ctx->lockowner.l_pid != current->tgid;
 890                 }
 891                 nfs_release_request(req);
 892                 if (!do_flush)
 893                         return 0;
 894                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 895         } while (status == 0);
 896         return status;
 897 }
 898
 899 /*
 900  * Avoid buffered writes when a open context credential's key would
 901  * expire soon.
 902  *
 903  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 904  *
 905  * Return 0 and set a credential flag which triggers the inode to flush
 906  * and performs  NFS_FILE_SYNC writes if the key will expired within
 907  * RPC_KEY_EXPIRE_TIMEO.
 908  */
 909 int
 910 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 911 {
 912         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 913         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 914
 915         return rpcauth_key_timeout_notify(auth, ctx->cred);
 916 }
 917
 918 /*
 919  * Test if the open context credential key is marked to expire soon.
 920  */
 921 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 922 {
 923         return rpcauth_cred_key_to_expire(ctx->cred);
 924 }
 925
 926 /*
 927  * If the page cache is marked as unsafe or invalid, then we can't rely on
 928  * the PageUptodate() flag. In this case, we will need to turn off
 929  * write optimisations that depend on the page contents being correct.
 930  */
 931 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 932 {
 933         struct nfs_inode *nfsi = NFS_I(inode);
 934
 935         if (nfs_have_delegated_attributes(inode))
 936                 goto out;
 937         if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 938                 return false;
 939         smp_rmb();
 940         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 941                 return false;
 942 out:
 943         return PageUptodate(page) != 0;
 944 }
 945
 946 /* If we know the page is up to date, and we're not using byte range locks (or
 947  * if we have the whole file locked for writing), it may be more efficient to
 948  * extend the write to cover the entire page in order to avoid fragmentation
 949  * inefficiencies.
 950  *
 951  * If the file is opened for synchronous writes then we can just skip the rest
 952  * of the checks.
 953  */
 954 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 955 {
 956         if (file->f_flags & O_DSYNC)
 957                 return 0;
 958         if (!nfs_write_pageuptodate(page, inode))
 959                 return 0;
 960         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 961                 return 1;
 962         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 963                         inode->i_flock->fl_end == OFFSET_MAX &&
 964                         inode->i_flock->fl_type != F_RDLCK))
 965                 return 1;
 966         return 0;
 967 }
 968
 969 /*
 970  * Update and possibly write a cached page of an NFS file.
 971  *
 972  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 973  * things with a page scheduled for an RPC call (e.g. invalidate it).
 974  */
 975 int nfs_updatepage(struct file *file, struct page *page,
 976                 unsigned int offset, unsigned int count)
 977 {
 978         struct nfs_open_context *ctx = nfs_file_open_context(file);
 979         struct inode    *inode = page_file_mapping(page)->host;
 980         int             status = 0;
 981
 982         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 983
 984         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 985                 file, count, (long long)(page_file_offset(page) + offset));
 986
 987         if (nfs_can_extend_write(file, page, inode)) {
 988                 count = max(count + offset, nfs_page_length(page));
 989                 offset = 0;
 990         }
 991
 992         status = nfs_writepage_setup(ctx, page, offset, count);
 993         if (status < 0)
 994                 nfs_set_pageerror(page);
 995         else
 996                 __set_page_dirty_nobuffers(page);
 997
 998         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 999                         status, (long long)i_size_read(inode));
1000         return status;
1001 }
1002
1003 static int flush_task_priority(int how)
1004 {
1005         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
1006                 case FLUSH_HIGHPRI:
1007                         return RPC_PRIORITY_HIGH;
1008                 case FLUSH_LOWPRI:
1009                         return RPC_PRIORITY_LOW;
1010         }
1011         return RPC_PRIORITY_NORMAL;
1012 }
1013
1014 static void nfs_initiate_write(struct nfs_pgio_data *data, struct rpc_message *msg,
1015                                struct rpc_task_setup *task_setup_data, int how)
1016 {
1017         struct inode *inode = data->header->inode;
1018         int priority = flush_task_priority(how);
1019
1020         task_setup_data->priority = priority;
1021         NFS_PROTO(inode)->write_setup(data, msg);
1022
1023         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1024                                  &task_setup_data->rpc_client, msg, data);
1025 }
1026
1027 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1028  * call this on each, which will prepare them to be retried on next
1029  * writeback using standard nfs.
1030  */
1031 static void nfs_redirty_request(struct nfs_page *req)
1032 {
1033         nfs_mark_request_dirty(req);
1034         nfs_unlock_request(req);
1035         nfs_end_page_writeback(req);
1036         nfs_release_request(req);
1037 }
1038
1039 static void nfs_async_write_error(struct list_head *head)
1040 {
1041         struct nfs_page *req;
1042
1043         while (!list_empty(head)) {
1044                 req = nfs_list_entry(head->next);
1045                 nfs_list_remove_request(req);
1046                 nfs_redirty_request(req);
1047         }
1048 }
1049
1050 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1051         .error_cleanup = nfs_async_write_error,
1052         .completion = nfs_write_completion,
1053 };
1054
1055 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1056                                struct inode *inode, int ioflags, bool force_mds,
1057                                const struct nfs_pgio_completion_ops *compl_ops)
1058 {
1059         struct nfs_server *server = NFS_SERVER(inode);
1060         const struct nfs_pageio_ops *pg_ops = &nfs_pgio_rw_ops;
1061
1062 #ifdef CONFIG_NFS_V4_1
1063         if (server->pnfs_curr_ld && !force_mds)
1064                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1065 #endif
1066         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1067                         server->wsize, ioflags);
1068 }
1069 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1070
1071 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1072 {
1073         pgio->pg_ops = &nfs_pgio_rw_ops;
1074         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1075 }
1076 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1077
1078
1079 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1080 {
1081         struct nfs_commit_data *data = calldata;
1082
1083         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1084 }
1085
1086 static void nfs_writeback_release_common(struct nfs_pgio_data *data)
1087 {
1088         struct nfs_pgio_header *hdr = data->header;
1089         int status = data->task.tk_status;
1090
1091         if ((status >= 0) && nfs_write_need_commit(data)) {
1092                 spin_lock(&hdr->lock);
1093                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1094                         ; /* Do nothing */
1095                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1096                         memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1097                 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1098                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1099                 spin_unlock(&hdr->lock);
1100         }
1101 }
1102
1103 /*
1104  * Special version of should_remove_suid() that ignores capabilities.
1105  */
1106 static int nfs_should_remove_suid(const struct inode *inode)
1107 {
1108         umode_t mode = inode->i_mode;
1109         int kill = 0;
1110
1111         /* suid always must be killed */
1112         if (unlikely(mode & S_ISUID))
1113                 kill = ATTR_KILL_SUID;
1114
1115         /*
1116          * sgid without any exec bits is just a mandatory locking mark; leave
1117          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1118          */
1119         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1120                 kill |= ATTR_KILL_SGID;
1121
1122         if (unlikely(kill && S_ISREG(mode)))
1123                 return kill;
1124
1125         return 0;
1126 }
1127
1128 /*
1129  * This function is called when the WRITE call is complete.
1130  */
1131 static int nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data,
1132                               struct inode *inode)
1133 {
1134         int status;
1135
1136         /*
1137          * ->write_done will attempt to use post-op attributes to detect
1138          * conflicting writes by other clients.  A strict interpretation
1139          * of close-to-open would allow us to continue caching even if
1140          * another writer had changed the file, but some applications
1141          * depend on tighter cache coherency when writing.
1142          */
1143         status = NFS_PROTO(inode)->write_done(task, data);
1144         if (status != 0)
1145                 return status;
1146         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, data->res.count);
1147
1148 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1149         if (data->res.verf->committed < data->args.stable && task->tk_status >= 0) {
1150                 /* We tried a write call, but the server did not
1151                  * commit data to stable storage even though we
1152                  * requested it.
1153                  * Note: There is a known bug in Tru64 < 5.0 in which
1154                  *       the server reports NFS_DATA_SYNC, but performs
1155                  *       NFS_FILE_SYNC. We therefore implement this checking
1156                  *       as a dprintk() in order to avoid filling syslog.
1157                  */
1158                 static unsigned long    complain;
1159
1160                 /* Note this will print the MDS for a DS write */
1161                 if (time_before(complain, jiffies)) {
1162                         dprintk("NFS:       faulty NFS server %s:"
1163                                 " (committed = %d) != (stable = %d)\n",
1164                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1165                                 data->res.verf->committed, data->args.stable);
1166                         complain = jiffies + 300 * HZ;
1167                 }
1168         }
1169 #endif
1170
1171         /* Deal with the suid/sgid bit corner case */
1172         if (nfs_should_remove_suid(inode))
1173                 nfs_mark_for_revalidate(inode);
1174         return 0;
1175 }
1176
1177 /*
1178  * This function is called when the WRITE call is complete.
1179  */
1180 static void nfs_writeback_result(struct rpc_task *task, struct nfs_pgio_data *data)
1181 {
1182         struct nfs_pgio_args    *argp = &data->args;
1183         struct nfs_pgio_res     *resp = &data->res;
1184
1185         if (resp->count < argp->count) {
1186                 static unsigned long    complain;
1187
1188                 /* This a short write! */
1189                 nfs_inc_stats(data->header->inode, NFSIOS_SHORTWRITE);
1190
1191                 /* Has the server at least made some progress? */
1192                 if (resp->count == 0) {
1193                         if (time_before(complain, jiffies)) {
1194                                 printk(KERN_WARNING
1195                                        "NFS: Server wrote zero bytes, expected %u.\n",
1196                                        argp->count);
1197                                 complain = jiffies + 300 * HZ;
1198                         }
1199                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1200                         task->tk_status = -EIO;
1201                         return;
1202                 }
1203                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1204                 if (resp->verf->committed != NFS_UNSTABLE) {
1205                         /* Resend from where the server left off */
1206                         data->mds_offset += resp->count;
1207                         argp->offset += resp->count;
1208                         argp->pgbase += resp->count;
1209                         argp->count -= resp->count;
1210                 } else {
1211                         /* Resend as a stable write in order to avoid
1212                          * headaches in the case of a server crash.
1213                          */
1214                         argp->stable = NFS_FILE_SYNC;
1215                 }
1216                 rpc_restart_call_prepare(task);
1217         }
1218 }
1219
1220
1221 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1222 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1223 {
1224         int ret;
1225
1226         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1227                 return 1;
1228         if (!may_wait)
1229                 return 0;
1230         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1231                                 NFS_INO_COMMIT,
1232                                 nfs_wait_bit_killable,
1233                                 TASK_KILLABLE);
1234         return (ret < 0) ? ret : 1;
1235 }
1236
1237 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1238 {
1239         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1240         smp_mb__after_atomic();
1241         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1242 }
1243
1244 void nfs_commitdata_release(struct nfs_commit_data *data)
1245 {
1246         put_nfs_open_context(data->context);
1247         nfs_commit_free(data);
1248 }
1249 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1250
1251 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1252                         const struct rpc_call_ops *call_ops,
1253                         int how, int flags)
1254 {
1255         struct rpc_task *task;
1256         int priority = flush_task_priority(how);
1257         struct rpc_message msg = {
1258                 .rpc_argp = &data->args,
1259                 .rpc_resp = &data->res,
1260                 .rpc_cred = data->cred,
1261         };
1262         struct rpc_task_setup task_setup_data = {
1263                 .task = &data->task,
1264                 .rpc_client = clnt,
1265                 .rpc_message = &msg,
1266                 .callback_ops = call_ops,
1267                 .callback_data = data,
1268                 .workqueue = nfsiod_workqueue,
1269                 .flags = RPC_TASK_ASYNC | flags,
1270                 .priority = priority,
1271         };
1272         /* Set up the initial task struct.  */
1273         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1274
1275         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1276
1277         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1278                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1279
1280         task = rpc_run_task(&task_setup_data);
1281         if (IS_ERR(task))
1282                 return PTR_ERR(task);
1283         if (how & FLUSH_SYNC)
1284                 rpc_wait_for_completion_task(task);
1285         rpc_put_task(task);
1286         return 0;
1287 }
1288 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1289
1290 /*
1291  * Set up the argument/result storage required for the RPC call.
1292  */
1293 void nfs_init_commit(struct nfs_commit_data *data,
1294                      struct list_head *head,
1295                      struct pnfs_layout_segment *lseg,
1296                      struct nfs_commit_info *cinfo)
1297 {
1298         struct nfs_page *first = nfs_list_entry(head->next);
1299         struct inode *inode = first->wb_context->dentry->d_inode;
1300
1301         /* Set up the RPC argument and reply structs
1302          * NB: take care not to mess about with data->commit et al. */
1303
1304         list_splice_init(head, &data->pages);
1305
1306         data->inode       = inode;
1307         data->cred        = first->wb_context->cred;
1308         data->lseg        = lseg; /* reference transferred */
1309         data->mds_ops     = &nfs_commit_ops;
1310         data->completion_ops = cinfo->completion_ops;
1311         data->dreq        = cinfo->dreq;
1312
1313         data->args.fh     = NFS_FH(data->inode);
1314         /* Note: we always request a commit of the entire inode */
1315         data->args.offset = 0;
1316         data->args.count  = 0;
1317         data->context     = get_nfs_open_context(first->wb_context);
1318         data->res.fattr   = &data->fattr;
1319         data->res.verf    = &data->verf;
1320         nfs_fattr_init(&data->fattr);
1321 }
1322 EXPORT_SYMBOL_GPL(nfs_init_commit);
1323
1324 void nfs_retry_commit(struct list_head *page_list,
1325                       struct pnfs_layout_segment *lseg,
1326                       struct nfs_commit_info *cinfo)
1327 {
1328         struct nfs_page *req;
1329
1330         while (!list_empty(page_list)) {
1331                 req = nfs_list_entry(page_list->next);
1332                 nfs_list_remove_request(req);
1333                 nfs_mark_request_commit(req, lseg, cinfo);
1334                 if (!cinfo->dreq) {
1335                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1336                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1337                                      BDI_RECLAIMABLE);
1338                 }
1339                 nfs_unlock_and_release_request(req);
1340         }
1341 }
1342 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1343
1344 /*
1345  * Commit dirty pages
1346  */
1347 static int
1348 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1349                 struct nfs_commit_info *cinfo)
1350 {
1351         struct nfs_commit_data  *data;
1352
1353         data = nfs_commitdata_alloc();
1354
1355         if (!data)
1356                 goto out_bad;
1357
1358         /* Set up the argument struct */
1359         nfs_init_commit(data, head, NULL, cinfo);
1360         atomic_inc(&cinfo->mds->rpcs_out);
1361         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1362                                    how, 0);
1363  out_bad:
1364         nfs_retry_commit(head, NULL, cinfo);
1365         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1366         return -ENOMEM;
1367 }
1368
1369 /*
1370  * COMMIT call returned
1371  */
1372 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1373 {
1374         struct nfs_commit_data  *data = calldata;
1375
1376         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1377                                 task->tk_pid, task->tk_status);
1378
1379         /* Call the NFS version-specific code */
1380         NFS_PROTO(data->inode)->commit_done(task, data);
1381 }
1382
1383 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1384 {
1385         struct nfs_page *req;
1386         int status = data->task.tk_status;
1387         struct nfs_commit_info cinfo;
1388
1389         while (!list_empty(&data->pages)) {
1390                 req = nfs_list_entry(data->pages.next);
1391                 nfs_list_remove_request(req);
1392                 nfs_clear_page_commit(req->wb_page);
1393
1394                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1395                         req->wb_context->dentry->d_sb->s_id,
1396                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1397                         req->wb_bytes,
1398                         (long long)req_offset(req));
1399                 if (status < 0) {
1400                         nfs_context_set_write_error(req->wb_context, status);
1401                         nfs_inode_remove_request(req);
1402                         dprintk(", error = %d\n", status);
1403                         goto next;
1404                 }
1405
1406                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1407                  * returned by the server against all stored verfs. */
1408                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1409                         /* We have a match */
1410                         nfs_inode_remove_request(req);
1411                         dprintk(" OK\n");
1412                         goto next;
1413                 }
1414                 /* We have a mismatch. Write the page again */
1415                 dprintk(" mismatch\n");
1416                 nfs_mark_request_dirty(req);
1417                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1418         next:
1419                 nfs_unlock_and_release_request(req);
1420         }
1421         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1422         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1423                 nfs_commit_clear_lock(NFS_I(data->inode));
1424 }
1425
1426 static void nfs_commit_release(void *calldata)
1427 {
1428         struct nfs_commit_data *data = calldata;
1429
1430         data->completion_ops->completion(data);
1431         nfs_commitdata_release(calldata);
1432 }
1433
1434 static const struct rpc_call_ops nfs_commit_ops = {
1435         .rpc_call_prepare = nfs_commit_prepare,
1436         .rpc_call_done = nfs_commit_done,
1437         .rpc_release = nfs_commit_release,
1438 };
1439
1440 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1441         .completion = nfs_commit_release_pages,
1442         .error_cleanup = nfs_commit_clear_lock,
1443 };
1444
1445 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1446                             int how, struct nfs_commit_info *cinfo)
1447 {
1448         int status;
1449
1450         status = pnfs_commit_list(inode, head, how, cinfo);
1451         if (status == PNFS_NOT_ATTEMPTED)
1452                 status = nfs_commit_list(inode, head, how, cinfo);
1453         return status;
1454 }
1455
1456 int nfs_commit_inode(struct inode *inode, int how)
1457 {
1458         LIST_HEAD(head);
1459         struct nfs_commit_info cinfo;
1460         int may_wait = how & FLUSH_SYNC;
1461         int res;
1462
1463         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1464         if (res <= 0)
1465                 goto out_mark_dirty;
1466         nfs_init_cinfo_from_inode(&cinfo, inode);
1467         res = nfs_scan_commit(inode, &head, &cinfo);
1468         if (res) {
1469                 int error;
1470
1471                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1472                 if (error < 0)
1473                         return error;
1474                 if (!may_wait)
1475                         goto out_mark_dirty;
1476                 error = wait_on_bit(&NFS_I(inode)->flags,
1477                                 NFS_INO_COMMIT,
1478                                 nfs_wait_bit_killable,
1479                                 TASK_KILLABLE);
1480                 if (error < 0)
1481                         return error;
1482         } else
1483                 nfs_commit_clear_lock(NFS_I(inode));
1484         return res;
1485         /* Note: If we exit without ensuring that the commit is complete,
1486          * we must mark the inode as dirty. Otherwise, future calls to
1487          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1488          * that the data is on the disk.
1489          */
1490 out_mark_dirty:
1491         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1492         return res;
1493 }
1494
1495 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1496 {
1497         struct nfs_inode *nfsi = NFS_I(inode);
1498         int flags = FLUSH_SYNC;
1499         int ret = 0;
1500
1501         /* no commits means nothing needs to be done */
1502         if (!nfsi->commit_info.ncommit)
1503                 return ret;
1504
1505         if (wbc->sync_mode == WB_SYNC_NONE) {
1506                 /* Don't commit yet if this is a non-blocking flush and there
1507                  * are a lot of outstanding writes for this mapping.
1508                  */
1509                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1510                         goto out_mark_dirty;
1511
1512                 /* don't wait for the COMMIT response */
1513                 flags = 0;
1514         }
1515
1516         ret = nfs_commit_inode(inode, flags);
1517         if (ret >= 0) {
1518                 if (wbc->sync_mode == WB_SYNC_NONE) {
1519                         if (ret < wbc->nr_to_write)
1520                                 wbc->nr_to_write -= ret;
1521                         else
1522                                 wbc->nr_to_write = 0;
1523                 }
1524                 return 0;
1525         }
1526 out_mark_dirty:
1527         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1528         return ret;
1529 }
1530 #else
1531 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1532 {
1533         return 0;
1534 }
1535 #endif
1536
1537 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1538 {
1539         return nfs_commit_unstable_pages(inode, wbc);
1540 }
1541 EXPORT_SYMBOL_GPL(nfs_write_inode);
1542
1543 /*
1544  * flush the inode to disk.
1545  */
1546 int nfs_wb_all(struct inode *inode)
1547 {
1548         struct writeback_control wbc = {
1549                 .sync_mode = WB_SYNC_ALL,
1550                 .nr_to_write = LONG_MAX,
1551                 .range_start = 0,
1552                 .range_end = LLONG_MAX,
1553         };
1554         int ret;
1555
1556         trace_nfs_writeback_inode_enter(inode);
1557
1558         ret = sync_inode(inode, &wbc);
1559
1560         trace_nfs_writeback_inode_exit(inode, ret);
1561         return ret;
1562 }
1563 EXPORT_SYMBOL_GPL(nfs_wb_all);
1564
1565 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1566 {
1567         struct nfs_page *req;
1568         int ret = 0;
1569
1570         for (;;) {
1571                 wait_on_page_writeback(page);
1572                 req = nfs_page_find_request(page);
1573                 if (req == NULL)
1574                         break;
1575                 if (nfs_lock_request(req)) {
1576                         nfs_clear_request_commit(req);
1577                         nfs_inode_remove_request(req);
1578                         /*
1579                          * In case nfs_inode_remove_request has marked the
1580                          * page as being dirty
1581                          */
1582                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1583                         nfs_unlock_and_release_request(req);
1584                         break;
1585                 }
1586                 ret = nfs_wait_on_request(req);
1587                 nfs_release_request(req);
1588                 if (ret < 0)
1589                         break;
1590         }
1591         return ret;
1592 }
1593
1594 /*
1595  * Write back all requests on one page - we do this before reading it.
1596  */
1597 int nfs_wb_page(struct inode *inode, struct page *page)
1598 {
1599         loff_t range_start = page_file_offset(page);
1600         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1601         struct writeback_control wbc = {
1602                 .sync_mode = WB_SYNC_ALL,
1603                 .nr_to_write = 0,
1604                 .range_start = range_start,
1605                 .range_end = range_end,
1606         };
1607         int ret;
1608
1609         trace_nfs_writeback_page_enter(inode);
1610
1611         for (;;) {
1612                 wait_on_page_writeback(page);
1613                 if (clear_page_dirty_for_io(page)) {
1614                         ret = nfs_writepage_locked(page, &wbc);
1615                         if (ret < 0)
1616                                 goto out_error;
1617                         continue;
1618                 }
1619                 ret = 0;
1620                 if (!PagePrivate(page))
1621                         break;
1622                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1623                 if (ret < 0)
1624                         goto out_error;
1625         }
1626 out_error:
1627         trace_nfs_writeback_page_exit(inode, ret);
1628         return ret;
1629 }
1630
1631 #ifdef CONFIG_MIGRATION
1632 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1633                 struct page *page, enum migrate_mode mode)
1634 {
1635         /*
1636          * If PagePrivate is set, then the page is currently associated with
1637          * an in-progress read or write request. Don't try to migrate it.
1638          *
1639          * FIXME: we could do this in principle, but we'll need a way to ensure
1640          *        that we can safely release the inode reference while holding
1641          *        the page lock.
1642          */
1643         if (PagePrivate(page))
1644                 return -EBUSY;
1645
1646         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1647                 return -EBUSY;
1648
1649         return migrate_page(mapping, newpage, page, mode);
1650 }
1651 #endif
1652
1653 int __init nfs_init_writepagecache(void)
1654 {
1655         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1656                                              sizeof(struct nfs_rw_header),
1657                                              0, SLAB_HWCACHE_ALIGN,
1658                                              NULL);
1659         if (nfs_wdata_cachep == NULL)
1660                 return -ENOMEM;
1661
1662         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1663                                                      nfs_wdata_cachep);
1664         if (nfs_wdata_mempool == NULL)
1665                 goto out_destroy_write_cache;
1666
1667         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1668                                              sizeof(struct nfs_commit_data),
1669                                              0, SLAB_HWCACHE_ALIGN,
1670                                              NULL);
1671         if (nfs_cdata_cachep == NULL)
1672                 goto out_destroy_write_mempool;
1673
1674         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1675                                                       nfs_cdata_cachep);
1676         if (nfs_commit_mempool == NULL)
1677                 goto out_destroy_commit_cache;
1678
1679         /*
1680          * NFS congestion size, scale with available memory.
1681          *
1682          *  64MB:    8192k
1683          * 128MB:   11585k
1684          * 256MB:   16384k
1685          * 512MB:   23170k
1686          *   1GB:   32768k
1687          *   2GB:   46340k
1688          *   4GB:   65536k
1689          *   8GB:   92681k
1690          *  16GB:  131072k
1691          *
1692          * This allows larger machines to have larger/more transfers.
1693          * Limit the default to 256M
1694          */
1695         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1696         if (nfs_congestion_kb > 256*1024)
1697                 nfs_congestion_kb = 256*1024;
1698
1699         return 0;
1700
1701 out_destroy_commit_cache:
1702         kmem_cache_destroy(nfs_cdata_cachep);
1703 out_destroy_write_mempool:
1704         mempool_destroy(nfs_wdata_mempool);
1705 out_destroy_write_cache:
1706         kmem_cache_destroy(nfs_wdata_cachep);
1707         return -ENOMEM;
1708 }
1709
1710 void nfs_destroy_writepagecache(void)
1711 {
1712         mempool_destroy(nfs_commit_mempool);
1713         kmem_cache_destroy(nfs_cdata_cachep);
1714         mempool_destroy(nfs_wdata_mempool);
1715         kmem_cache_destroy(nfs_wdata_cachep);
1716 }
1717
1718 static const struct nfs_rw_ops nfs_rw_write_ops = {
1719         .rw_mode                = FMODE_WRITE,
1720         .rw_alloc_header        = nfs_writehdr_alloc,
1721         .rw_free_header         = nfs_writehdr_free,
1722         .rw_release             = nfs_writeback_release_common,
1723         .rw_done                = nfs_writeback_done,
1724         .rw_result              = nfs_writeback_result,
1725         .rw_initiate            = nfs_initiate_write,
1726 };