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_write_common_ops;
  46 static const struct rpc_call_ops nfs_commit_ops;
  47 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  48 static const struct nfs_commit_completion_ops nfs_commit_completion_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 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                 struct nfs_pgio_header *hdr = &p->header;
  79
  80                 memset(p, 0, sizeof(*p));
  81                 INIT_LIST_HEAD(&hdr->pages);
  82                 INIT_LIST_HEAD(&hdr->rpc_list);
  83                 spin_lock_init(&hdr->lock);
  84                 atomic_set(&hdr->refcnt, 0);
  85         }
  86         return p;
  87 }
  88 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
  89
  90 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
  91 {
  92         struct nfs_rw_header *whdr = container_of(hdr, struct nfs_rw_header, header);
  93         mempool_free(whdr, nfs_wdata_mempool);
  94 }
  95 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
  96
  97 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  98 {
  99         ctx->error = error;
 100         smp_wmb();
 101         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 102 }
 103
 104 static struct nfs_page *
 105 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
 106 {
 107         struct nfs_page *req = NULL;
 108
 109         if (PagePrivate(page))
 110                 req = (struct nfs_page *)page_private(page);
 111         else if (unlikely(PageSwapCache(page))) {
 112                 struct nfs_page *freq, *t;
 113
 114                 /* Linearly search the commit list for the correct req */
 115                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 116                         if (freq->wb_page == page) {
 117                                 req = freq;
 118                                 break;
 119                         }
 120                 }
 121         }
 122
 123         if (req)
 124                 kref_get(&req->wb_kref);
 125
 126         return req;
 127 }
 128
 129 static struct nfs_page *nfs_page_find_request(struct page *page)
 130 {
 131         struct inode *inode = page_file_mapping(page)->host;
 132         struct nfs_page *req = NULL;
 133
 134         spin_lock(&inode->i_lock);
 135         req = nfs_page_find_request_locked(NFS_I(inode), page);
 136         spin_unlock(&inode->i_lock);
 137         return req;
 138 }
 139
 140 /* Adjust the file length if we're writing beyond the end */
 141 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 142 {
 143         struct inode *inode = page_file_mapping(page)->host;
 144         loff_t end, i_size;
 145         pgoff_t end_index;
 146
 147         spin_lock(&inode->i_lock);
 148         i_size = i_size_read(inode);
 149         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 150         if (i_size > 0 && page_file_index(page) < end_index)
 151                 goto out;
 152         end = page_file_offset(page) + ((loff_t)offset+count);
 153         if (i_size >= end)
 154                 goto out;
 155         i_size_write(inode, end);
 156         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 157 out:
 158         spin_unlock(&inode->i_lock);
 159 }
 160
 161 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 162 static void nfs_set_pageerror(struct page *page)
 163 {
 164         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 165 }
 166
 167 /* We can set the PG_uptodate flag if we see that a write request
 168  * covers the full page.
 169  */
 170 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 171 {
 172         if (PageUptodate(page))
 173                 return;
 174         if (base != 0)
 175                 return;
 176         if (count != nfs_page_length(page))
 177                 return;
 178         SetPageUptodate(page);
 179 }
 180
 181 static int wb_priority(struct writeback_control *wbc)
 182 {
 183         if (wbc->for_reclaim)
 184                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 185         if (wbc->for_kupdate || wbc->for_background)
 186                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 187         return FLUSH_COND_STABLE;
 188 }
 189
 190 /*
 191  * NFS congestion control
 192  */
 193
 194 int nfs_congestion_kb;
 195
 196 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 197 #define NFS_CONGESTION_OFF_THRESH       \
 198         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 199
 200 static void nfs_set_page_writeback(struct page *page)
 201 {
 202         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 203         int ret = test_set_page_writeback(page);
 204
 205         WARN_ON_ONCE(ret != 0);
 206
 207         if (atomic_long_inc_return(&nfss->writeback) >
 208                         NFS_CONGESTION_ON_THRESH) {
 209                 set_bdi_congested(&nfss->backing_dev_info,
 210                                         BLK_RW_ASYNC);
 211         }
 212 }
 213
 214 static void nfs_end_page_writeback(struct page *page)
 215 {
 216         struct inode *inode = page_file_mapping(page)->host;
 217         struct nfs_server *nfss = NFS_SERVER(inode);
 218
 219         end_page_writeback(page);
 220         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 221                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 222 }
 223
 224 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 225 {
 226         struct inode *inode = page_file_mapping(page)->host;
 227         struct nfs_page *req;
 228         int ret;
 229
 230         spin_lock(&inode->i_lock);
 231         for (;;) {
 232                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 233                 if (req == NULL)
 234                         break;
 235                 if (nfs_lock_request(req))
 236                         break;
 237                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 238                  *       then the call to nfs_lock_request() will always
 239                  *       succeed provided that someone hasn't already marked the
 240                  *       request as dirty (in which case we don't care).
 241                  */
 242                 spin_unlock(&inode->i_lock);
 243                 if (!nonblock)
 244                         ret = nfs_wait_on_request(req);
 245                 else
 246                         ret = -EAGAIN;
 247                 nfs_release_request(req);
 248                 if (ret != 0)
 249                         return ERR_PTR(ret);
 250                 spin_lock(&inode->i_lock);
 251         }
 252         spin_unlock(&inode->i_lock);
 253         return req;
 254 }
 255
 256 /*
 257  * Find an associated nfs write request, and prepare to flush it out
 258  * May return an error if the user signalled nfs_wait_on_request().
 259  */
 260 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 261                                 struct page *page, bool nonblock)
 262 {
 263         struct nfs_page *req;
 264         int ret = 0;
 265
 266         req = nfs_find_and_lock_request(page, nonblock);
 267         if (!req)
 268                 goto out;
 269         ret = PTR_ERR(req);
 270         if (IS_ERR(req))
 271                 goto out;
 272
 273         nfs_set_page_writeback(page);
 274         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 275
 276         ret = 0;
 277         if (!nfs_pageio_add_request(pgio, req)) {
 278                 nfs_redirty_request(req);
 279                 ret = pgio->pg_error;
 280         }
 281 out:
 282         return ret;
 283 }
 284
 285 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 286 {
 287         struct inode *inode = page_file_mapping(page)->host;
 288         int ret;
 289
 290         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 291         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 292
 293         nfs_pageio_cond_complete(pgio, page_file_index(page));
 294         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 295         if (ret == -EAGAIN) {
 296                 redirty_page_for_writepage(wbc, page);
 297                 ret = 0;
 298         }
 299         return ret;
 300 }
 301
 302 /*
 303  * Write an mmapped page to the server.
 304  */
 305 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 306 {
 307         struct nfs_pageio_descriptor pgio;
 308         int err;
 309
 310         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 311                                 false, &nfs_async_write_completion_ops);
 312         err = nfs_do_writepage(page, wbc, &pgio);
 313         nfs_pageio_complete(&pgio);
 314         if (err < 0)
 315                 return err;
 316         if (pgio.pg_error < 0)
 317                 return pgio.pg_error;
 318         return 0;
 319 }
 320
 321 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 322 {
 323         int ret;
 324
 325         ret = nfs_writepage_locked(page, wbc);
 326         unlock_page(page);
 327         return ret;
 328 }
 329
 330 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 331 {
 332         int ret;
 333
 334         ret = nfs_do_writepage(page, wbc, data);
 335         unlock_page(page);
 336         return ret;
 337 }
 338
 339 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 340 {
 341         struct inode *inode = mapping->host;
 342         unsigned long *bitlock = &NFS_I(inode)->flags;
 343         struct nfs_pageio_descriptor pgio;
 344         int err;
 345
 346         /* Stop dirtying of new pages while we sync */
 347         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 348                         nfs_wait_bit_killable, TASK_KILLABLE);
 349         if (err)
 350                 goto out_err;
 351
 352         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 353
 354         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 355                                 &nfs_async_write_completion_ops);
 356         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 357         nfs_pageio_complete(&pgio);
 358
 359         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 360         smp_mb__after_clear_bit();
 361         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 362
 363         if (err < 0)
 364                 goto out_err;
 365         err = pgio.pg_error;
 366         if (err < 0)
 367                 goto out_err;
 368         return 0;
 369 out_err:
 370         return err;
 371 }
 372
 373 /*
 374  * Insert a write request into an inode
 375  */
 376 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 377 {
 378         struct nfs_inode *nfsi = NFS_I(inode);
 379
 380         /* Lock the request! */
 381         nfs_lock_request(req);
 382
 383         spin_lock(&inode->i_lock);
 384         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 385                 inode->i_version++;
 386         /*
 387          * Swap-space should not get truncated. Hence no need to plug the race
 388          * with invalidate/truncate.
 389          */
 390         if (likely(!PageSwapCache(req->wb_page))) {
 391                 set_bit(PG_MAPPED, &req->wb_flags);
 392                 SetPagePrivate(req->wb_page);
 393                 set_page_private(req->wb_page, (unsigned long)req);
 394         }
 395         nfsi->npages++;
 396         kref_get(&req->wb_kref);
 397         spin_unlock(&inode->i_lock);
 398 }
 399
 400 /*
 401  * Remove a write request from an inode
 402  */
 403 static void nfs_inode_remove_request(struct nfs_page *req)
 404 {
 405         struct inode *inode = req->wb_context->dentry->d_inode;
 406         struct nfs_inode *nfsi = NFS_I(inode);
 407
 408         spin_lock(&inode->i_lock);
 409         if (likely(!PageSwapCache(req->wb_page))) {
 410                 set_page_private(req->wb_page, 0);
 411                 ClearPagePrivate(req->wb_page);
 412                 clear_bit(PG_MAPPED, &req->wb_flags);
 413         }
 414         nfsi->npages--;
 415         spin_unlock(&inode->i_lock);
 416         nfs_release_request(req);
 417 }
 418
 419 static void
 420 nfs_mark_request_dirty(struct nfs_page *req)
 421 {
 422         __set_page_dirty_nobuffers(req->wb_page);
 423 }
 424
 425 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 426 /**
 427  * nfs_request_add_commit_list - add request to a commit list
 428  * @req: pointer to a struct nfs_page
 429  * @dst: commit list head
 430  * @cinfo: holds list lock and accounting info
 431  *
 432  * This sets the PG_CLEAN bit, updates the cinfo count of
 433  * number of outstanding requests requiring a commit as well as
 434  * the MM page stats.
 435  *
 436  * The caller must _not_ hold the cinfo->lock, but must be
 437  * holding the nfs_page lock.
 438  */
 439 void
 440 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 441                             struct nfs_commit_info *cinfo)
 442 {
 443         set_bit(PG_CLEAN, &(req)->wb_flags);
 444         spin_lock(cinfo->lock);
 445         nfs_list_add_request(req, dst);
 446         cinfo->mds->ncommit++;
 447         spin_unlock(cinfo->lock);
 448         if (!cinfo->dreq) {
 449                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 450                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 451                              BDI_RECLAIMABLE);
 452                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 453                                    I_DIRTY_DATASYNC);
 454         }
 455 }
 456 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 457
 458 /**
 459  * nfs_request_remove_commit_list - Remove request from a commit list
 460  * @req: pointer to a nfs_page
 461  * @cinfo: holds list lock and accounting info
 462  *
 463  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 464  * number of outstanding requests requiring a commit
 465  * It does not update the MM page stats.
 466  *
 467  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 468  */
 469 void
 470 nfs_request_remove_commit_list(struct nfs_page *req,
 471                                struct nfs_commit_info *cinfo)
 472 {
 473         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 474                 return;
 475         nfs_list_remove_request(req);
 476         cinfo->mds->ncommit--;
 477 }
 478 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 479
 480 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 481                                       struct inode *inode)
 482 {
 483         cinfo->lock = &inode->i_lock;
 484         cinfo->mds = &NFS_I(inode)->commit_info;
 485         cinfo->ds = pnfs_get_ds_info(inode);
 486         cinfo->dreq = NULL;
 487         cinfo->completion_ops = &nfs_commit_completion_ops;
 488 }
 489
 490 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 491                     struct inode *inode,
 492                     struct nfs_direct_req *dreq)
 493 {
 494         if (dreq)
 495                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 496         else
 497                 nfs_init_cinfo_from_inode(cinfo, inode);
 498 }
 499 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 500
 501 /*
 502  * Add a request to the inode's commit list.
 503  */
 504 void
 505 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 506                         struct nfs_commit_info *cinfo)
 507 {
 508         if (pnfs_mark_request_commit(req, lseg, cinfo))
 509                 return;
 510         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 511 }
 512
 513 static void
 514 nfs_clear_page_commit(struct page *page)
 515 {
 516         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 517         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 518 }
 519
 520 static void
 521 nfs_clear_request_commit(struct nfs_page *req)
 522 {
 523         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 524                 struct inode *inode = req->wb_context->dentry->d_inode;
 525                 struct nfs_commit_info cinfo;
 526
 527                 nfs_init_cinfo_from_inode(&cinfo, inode);
 528                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 529                         spin_lock(cinfo.lock);
 530                         nfs_request_remove_commit_list(req, &cinfo);
 531                         spin_unlock(cinfo.lock);
 532                 }
 533                 nfs_clear_page_commit(req->wb_page);
 534         }
 535 }
 536
 537 static inline
 538 int nfs_write_need_commit(struct nfs_pgio_data *data)
 539 {
 540         if (data->verf.committed == NFS_DATA_SYNC)
 541                 return data->header->lseg == NULL;
 542         return data->verf.committed != NFS_FILE_SYNC;
 543 }
 544
 545 #else
 546 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 547                                       struct inode *inode)
 548 {
 549 }
 550
 551 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 552                     struct inode *inode,
 553                     struct nfs_direct_req *dreq)
 554 {
 555 }
 556
 557 void
 558 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 559                         struct nfs_commit_info *cinfo)
 560 {
 561 }
 562
 563 static void
 564 nfs_clear_request_commit(struct nfs_page *req)
 565 {
 566 }
 567
 568 static inline
 569 int nfs_write_need_commit(struct nfs_pgio_data *data)
 570 {
 571         return 0;
 572 }
 573
 574 #endif
 575
 576 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 577 {
 578         struct nfs_commit_info cinfo;
 579         unsigned long bytes = 0;
 580
 581         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 582                 goto out;
 583         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 584         while (!list_empty(&hdr->pages)) {
 585                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 586
 587                 bytes += req->wb_bytes;
 588                 nfs_list_remove_request(req);
 589                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 590                     (hdr->good_bytes < bytes)) {
 591                         nfs_set_pageerror(req->wb_page);
 592                         nfs_context_set_write_error(req->wb_context, hdr->error);
 593                         goto remove_req;
 594                 }
 595                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 596                         nfs_mark_request_dirty(req);
 597                         goto next;
 598                 }
 599                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 600                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 601                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 602                         goto next;
 603                 }
 604 remove_req:
 605                 nfs_inode_remove_request(req);
 606 next:
 607                 nfs_unlock_request(req);
 608                 nfs_end_page_writeback(req->wb_page);
 609                 nfs_release_request(req);
 610         }
 611 out:
 612         hdr->release(hdr);
 613 }
 614
 615 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 616 static unsigned long
 617 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 618 {
 619         return cinfo->mds->ncommit;
 620 }
 621
 622 /* cinfo->lock held by caller */
 623 int
 624 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 625                      struct nfs_commit_info *cinfo, int max)
 626 {
 627         struct nfs_page *req, *tmp;
 628         int ret = 0;
 629
 630         list_for_each_entry_safe(req, tmp, src, wb_list) {
 631                 if (!nfs_lock_request(req))
 632                         continue;
 633                 kref_get(&req->wb_kref);
 634                 if (cond_resched_lock(cinfo->lock))
 635                         list_safe_reset_next(req, tmp, wb_list);
 636                 nfs_request_remove_commit_list(req, cinfo);
 637                 nfs_list_add_request(req, dst);
 638                 ret++;
 639                 if ((ret == max) && !cinfo->dreq)
 640                         break;
 641         }
 642         return ret;
 643 }
 644
 645 /*
 646  * nfs_scan_commit - Scan an inode for commit requests
 647  * @inode: NFS inode to scan
 648  * @dst: mds destination list
 649  * @cinfo: mds and ds lists of reqs ready to commit
 650  *
 651  * Moves requests from the inode's 'commit' request list.
 652  * The requests are *not* checked to ensure that they form a contiguous set.
 653  */
 654 int
 655 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 656                 struct nfs_commit_info *cinfo)
 657 {
 658         int ret = 0;
 659
 660         spin_lock(cinfo->lock);
 661         if (cinfo->mds->ncommit > 0) {
 662                 const int max = INT_MAX;
 663
 664                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 665                                            cinfo, max);
 666                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 667         }
 668         spin_unlock(cinfo->lock);
 669         return ret;
 670 }
 671
 672 #else
 673 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 674 {
 675         return 0;
 676 }
 677
 678 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 679                     struct nfs_commit_info *cinfo)
 680 {
 681         return 0;
 682 }
 683 #endif
 684
 685 /*
 686  * Search for an existing write request, and attempt to update
 687  * it to reflect a new dirty region on a given page.
 688  *
 689  * If the attempt fails, then the existing request is flushed out
 690  * to disk.
 691  */
 692 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 693                 struct page *page,
 694                 unsigned int offset,
 695                 unsigned int bytes)
 696 {
 697         struct nfs_page *req;
 698         unsigned int rqend;
 699         unsigned int end;
 700         int error;
 701
 702         if (!PagePrivate(page))
 703                 return NULL;
 704
 705         end = offset + bytes;
 706         spin_lock(&inode->i_lock);
 707
 708         for (;;) {
 709                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 710                 if (req == NULL)
 711                         goto out_unlock;
 712
 713                 rqend = req->wb_offset + req->wb_bytes;
 714                 /*
 715                  * Tell the caller to flush out the request if
 716                  * the offsets are non-contiguous.
 717                  * Note: nfs_flush_incompatible() will already
 718                  * have flushed out requests having wrong owners.
 719                  */
 720                 if (offset > rqend
 721                     || end < req->wb_offset)
 722                         goto out_flushme;
 723
 724                 if (nfs_lock_request(req))
 725                         break;
 726
 727                 /* The request is locked, so wait and then retry */
 728                 spin_unlock(&inode->i_lock);
 729                 error = nfs_wait_on_request(req);
 730                 nfs_release_request(req);
 731                 if (error != 0)
 732                         goto out_err;
 733                 spin_lock(&inode->i_lock);
 734         }
 735
 736         /* Okay, the request matches. Update the region */
 737         if (offset < req->wb_offset) {
 738                 req->wb_offset = offset;
 739                 req->wb_pgbase = offset;
 740         }
 741         if (end > rqend)
 742                 req->wb_bytes = end - req->wb_offset;
 743         else
 744                 req->wb_bytes = rqend - req->wb_offset;
 745 out_unlock:
 746         spin_unlock(&inode->i_lock);
 747         if (req)
 748                 nfs_clear_request_commit(req);
 749         return req;
 750 out_flushme:
 751         spin_unlock(&inode->i_lock);
 752         nfs_release_request(req);
 753         error = nfs_wb_page(inode, page);
 754 out_err:
 755         return ERR_PTR(error);
 756 }
 757
 758 /*
 759  * Try to update an existing write request, or create one if there is none.
 760  *
 761  * Note: Should always be called with the Page Lock held to prevent races
 762  * if we have to add a new request. Also assumes that the caller has
 763  * already called nfs_flush_incompatible() if necessary.
 764  */
 765 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 766                 struct page *page, unsigned int offset, unsigned int bytes)
 767 {
 768         struct inode *inode = page_file_mapping(page)->host;
 769         struct nfs_page *req;
 770
 771         req = nfs_try_to_update_request(inode, page, offset, bytes);
 772         if (req != NULL)
 773                 goto out;
 774         req = nfs_create_request(ctx, inode, page, offset, bytes);
 775         if (IS_ERR(req))
 776                 goto out;
 777         nfs_inode_add_request(inode, req);
 778 out:
 779         return req;
 780 }
 781
 782 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 783                 unsigned int offset, unsigned int count)
 784 {
 785         struct nfs_page *req;
 786
 787         req = nfs_setup_write_request(ctx, page, offset, count);
 788         if (IS_ERR(req))
 789                 return PTR_ERR(req);
 790         /* Update file length */
 791         nfs_grow_file(page, offset, count);
 792         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 793         nfs_mark_request_dirty(req);
 794         nfs_unlock_and_release_request(req);
 795         return 0;
 796 }
 797
 798 int nfs_flush_incompatible(struct file *file, struct page *page)
 799 {
 800         struct nfs_open_context *ctx = nfs_file_open_context(file);
 801         struct nfs_lock_context *l_ctx;
 802         struct nfs_page *req;
 803         int do_flush, status;
 804         /*
 805          * Look for a request corresponding to this page. If there
 806          * is one, and it belongs to another file, we flush it out
 807          * before we try to copy anything into the page. Do this
 808          * due to the lack of an ACCESS-type call in NFSv2.
 809          * Also do the same if we find a request from an existing
 810          * dropped page.
 811          */
 812         do {
 813                 req = nfs_page_find_request(page);
 814                 if (req == NULL)
 815                         return 0;
 816                 l_ctx = req->wb_lock_context;
 817                 do_flush = req->wb_page != page || req->wb_context != ctx;
 818                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 819                         do_flush |= l_ctx->lockowner.l_owner != current->files
 820                                 || l_ctx->lockowner.l_pid != current->tgid;
 821                 }
 822                 nfs_release_request(req);
 823                 if (!do_flush)
 824                         return 0;
 825                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 826         } while (status == 0);
 827         return status;
 828 }
 829
 830 /*
 831  * Avoid buffered writes when a open context credential's key would
 832  * expire soon.
 833  *
 834  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 835  *
 836  * Return 0 and set a credential flag which triggers the inode to flush
 837  * and performs  NFS_FILE_SYNC writes if the key will expired within
 838  * RPC_KEY_EXPIRE_TIMEO.
 839  */
 840 int
 841 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 842 {
 843         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 844         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 845
 846         return rpcauth_key_timeout_notify(auth, ctx->cred);
 847 }
 848
 849 /*
 850  * Test if the open context credential key is marked to expire soon.
 851  */
 852 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 853 {
 854         return rpcauth_cred_key_to_expire(ctx->cred);
 855 }
 856
 857 /*
 858  * If the page cache is marked as unsafe or invalid, then we can't rely on
 859  * the PageUptodate() flag. In this case, we will need to turn off
 860  * write optimisations that depend on the page contents being correct.
 861  */
 862 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 863 {
 864         struct nfs_inode *nfsi = NFS_I(inode);
 865
 866         if (nfs_have_delegated_attributes(inode))
 867                 goto out;
 868         if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 869                 return false;
 870         smp_rmb();
 871         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 872                 return false;
 873 out:
 874         return PageUptodate(page) != 0;
 875 }
 876
 877 /* If we know the page is up to date, and we're not using byte range locks (or
 878  * if we have the whole file locked for writing), it may be more efficient to
 879  * extend the write to cover the entire page in order to avoid fragmentation
 880  * inefficiencies.
 881  *
 882  * If the file is opened for synchronous writes then we can just skip the rest
 883  * of the checks.
 884  */
 885 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 886 {
 887         if (file->f_flags & O_DSYNC)
 888                 return 0;
 889         if (!nfs_write_pageuptodate(page, inode))
 890                 return 0;
 891         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 892                 return 1;
 893         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 894                         inode->i_flock->fl_end == OFFSET_MAX &&
 895                         inode->i_flock->fl_type != F_RDLCK))
 896                 return 1;
 897         return 0;
 898 }
 899
 900 /*
 901  * Update and possibly write a cached page of an NFS file.
 902  *
 903  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 904  * things with a page scheduled for an RPC call (e.g. invalidate it).
 905  */
 906 int nfs_updatepage(struct file *file, struct page *page,
 907                 unsigned int offset, unsigned int count)
 908 {
 909         struct nfs_open_context *ctx = nfs_file_open_context(file);
 910         struct inode    *inode = page_file_mapping(page)->host;
 911         int             status = 0;
 912
 913         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 914
 915         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 916                 file, count, (long long)(page_file_offset(page) + offset));
 917
 918         if (nfs_can_extend_write(file, page, inode)) {
 919                 count = max(count + offset, nfs_page_length(page));
 920                 offset = 0;
 921         }
 922
 923         status = nfs_writepage_setup(ctx, page, offset, count);
 924         if (status < 0)
 925                 nfs_set_pageerror(page);
 926         else
 927                 __set_page_dirty_nobuffers(page);
 928
 929         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 930                         status, (long long)i_size_read(inode));
 931         return status;
 932 }
 933
 934 static int flush_task_priority(int how)
 935 {
 936         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 937                 case FLUSH_HIGHPRI:
 938                         return RPC_PRIORITY_HIGH;
 939                 case FLUSH_LOWPRI:
 940                         return RPC_PRIORITY_LOW;
 941         }
 942         return RPC_PRIORITY_NORMAL;
 943 }
 944
 945 int nfs_initiate_write(struct rpc_clnt *clnt,
 946                        struct nfs_pgio_data *data,
 947                        const struct rpc_call_ops *call_ops,
 948                        int how, int flags)
 949 {
 950         struct inode *inode = data->header->inode;
 951         int priority = flush_task_priority(how);
 952         struct rpc_task *task;
 953         struct rpc_message msg = {
 954                 .rpc_argp = &data->args,
 955                 .rpc_resp = &data->res,
 956                 .rpc_cred = data->header->cred,
 957         };
 958         struct rpc_task_setup task_setup_data = {
 959                 .rpc_client = clnt,
 960                 .task = &data->task,
 961                 .rpc_message = &msg,
 962                 .callback_ops = call_ops,
 963                 .callback_data = data,
 964                 .workqueue = nfsiod_workqueue,
 965                 .flags = RPC_TASK_ASYNC | flags,
 966                 .priority = priority,
 967         };
 968         int ret = 0;
 969
 970         /* Set up the initial task struct.  */
 971         NFS_PROTO(inode)->write_setup(data, &msg);
 972
 973         dprintk("NFS: %5u initiated write call "
 974                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 975                 data->task.tk_pid,
 976                 inode->i_sb->s_id,
 977                 (unsigned long long)NFS_FILEID(inode),
 978                 data->args.count,
 979                 (unsigned long long)data->args.offset);
 980
 981         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
 982                                  &task_setup_data.rpc_client, &msg, data);
 983
 984         task = rpc_run_task(&task_setup_data);
 985         if (IS_ERR(task)) {
 986                 ret = PTR_ERR(task);
 987                 goto out;
 988         }
 989         if (how & FLUSH_SYNC) {
 990                 ret = rpc_wait_for_completion_task(task);
 991                 if (ret == 0)
 992                         ret = task->tk_status;
 993         }
 994         rpc_put_task(task);
 995 out:
 996         return ret;
 997 }
 998 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 999
1000 /*
1001  * Set up the argument/result storage required for the RPC call.
1002  */
1003 static void nfs_write_rpcsetup(struct nfs_pgio_data *data,
1004                 unsigned int count, unsigned int offset,
1005                 int how, struct nfs_commit_info *cinfo)
1006 {
1007         struct nfs_page *req = data->header->req;
1008
1009         /* Set up the RPC argument and reply structs
1010          * NB: take care not to mess about with data->commit et al. */
1011
1012         data->args.fh     = NFS_FH(data->header->inode);
1013         data->args.offset = req_offset(req) + offset;
1014         /* pnfs_set_layoutcommit needs this */
1015         data->mds_offset = data->args.offset;
1016         data->args.pgbase = req->wb_pgbase + offset;
1017         data->args.pages  = data->pages.pagevec;
1018         data->args.count  = count;
1019         data->args.context = get_nfs_open_context(req->wb_context);
1020         data->args.lock_context = req->wb_lock_context;
1021         data->args.stable  = NFS_UNSTABLE;
1022         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1023         case 0:
1024                 break;
1025         case FLUSH_COND_STABLE:
1026                 if (nfs_reqs_to_commit(cinfo))
1027                         break;
1028         default:
1029                 data->args.stable = NFS_FILE_SYNC;
1030         }
1031
1032         data->res.fattr   = &data->fattr;
1033         data->res.count   = count;
1034         data->res.verf    = &data->verf;
1035         nfs_fattr_init(&data->fattr);
1036 }
1037
1038 static int nfs_do_write(struct nfs_pgio_data *data,
1039                 const struct rpc_call_ops *call_ops,
1040                 int how)
1041 {
1042         struct inode *inode = data->header->inode;
1043
1044         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1045 }
1046
1047 static int nfs_do_multiple_writes(struct list_head *head,
1048                 const struct rpc_call_ops *call_ops,
1049                 int how)
1050 {
1051         struct nfs_pgio_data *data;
1052         int ret = 0;
1053
1054         while (!list_empty(head)) {
1055                 int ret2;
1056
1057                 data = list_first_entry(head, struct nfs_pgio_data, list);
1058                 list_del_init(&data->list);
1059
1060                 ret2 = nfs_do_write(data, call_ops, how);
1061                  if (ret == 0)
1062                          ret = ret2;
1063         }
1064         return ret;
1065 }
1066
1067 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1068  * call this on each, which will prepare them to be retried on next
1069  * writeback using standard nfs.
1070  */
1071 static void nfs_redirty_request(struct nfs_page *req)
1072 {
1073         nfs_mark_request_dirty(req);
1074         nfs_unlock_request(req);
1075         nfs_end_page_writeback(req->wb_page);
1076         nfs_release_request(req);
1077 }
1078
1079 static void nfs_async_write_error(struct list_head *head)
1080 {
1081         struct nfs_page *req;
1082
1083         while (!list_empty(head)) {
1084                 req = nfs_list_entry(head->next);
1085                 nfs_list_remove_request(req);
1086                 nfs_redirty_request(req);
1087         }
1088 }
1089
1090 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1091         .error_cleanup = nfs_async_write_error,
1092         .completion = nfs_write_completion,
1093 };
1094
1095 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1096                 struct nfs_pgio_header *hdr)
1097 {
1098         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1099         while (!list_empty(&hdr->rpc_list)) {
1100                 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
1101                                 struct nfs_pgio_data, list);
1102                 list_del(&data->list);
1103                 nfs_pgio_data_release(data);
1104         }
1105         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1106 }
1107
1108 /*
1109  * Generate multiple small requests to write out a single
1110  * contiguous dirty area on one page.
1111  */
1112 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1113                            struct nfs_pgio_header *hdr)
1114 {
1115         struct nfs_page *req = hdr->req;
1116         struct page *page = req->wb_page;
1117         struct nfs_pgio_data *data;
1118         size_t wsize = desc->pg_bsize, nbytes;
1119         unsigned int offset;
1120         int requests = 0;
1121         struct nfs_commit_info cinfo;
1122
1123         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1124
1125         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1126             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1127              desc->pg_count > wsize))
1128                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1129
1130
1131         offset = 0;
1132         nbytes = desc->pg_count;
1133         do {
1134                 size_t len = min(nbytes, wsize);
1135
1136                 data = nfs_pgio_data_alloc(hdr, 1);
1137                 if (!data) {
1138                         nfs_flush_error(desc, hdr);
1139                         return -ENOMEM;
1140                 }
1141                 data->pages.pagevec[0] = page;
1142                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1143                 list_add(&data->list, &hdr->rpc_list);
1144                 requests++;
1145                 nbytes -= len;
1146                 offset += len;
1147         } while (nbytes != 0);
1148         nfs_list_remove_request(req);
1149         nfs_list_add_request(req, &hdr->pages);
1150         desc->pg_rpc_callops = &nfs_write_common_ops;
1151         return 0;
1152 }
1153
1154 /*
1155  * Create an RPC task for the given write request and kick it.
1156  * The page must have been locked by the caller.
1157  *
1158  * It may happen that the page we're passed is not marked dirty.
1159  * This is the case if nfs_updatepage detects a conflicting request
1160  * that has been written but not committed.
1161  */
1162 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1163                          struct nfs_pgio_header *hdr)
1164 {
1165         struct nfs_page         *req;
1166         struct page             **pages;
1167         struct nfs_pgio_data    *data;
1168         struct list_head *head = &desc->pg_list;
1169         struct nfs_commit_info cinfo;
1170
1171         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
1172                                                            desc->pg_count));
1173         if (!data) {
1174                 nfs_flush_error(desc, hdr);
1175                 return -ENOMEM;
1176         }
1177
1178         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1179         pages = data->pages.pagevec;
1180         while (!list_empty(head)) {
1181                 req = nfs_list_entry(head->next);
1182                 nfs_list_remove_request(req);
1183                 nfs_list_add_request(req, &hdr->pages);
1184                 *pages++ = req->wb_page;
1185         }
1186
1187         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1188             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1189                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1190
1191         /* Set up the argument struct */
1192         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1193         list_add(&data->list, &hdr->rpc_list);
1194         desc->pg_rpc_callops = &nfs_write_common_ops;
1195         return 0;
1196 }
1197
1198 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1199                       struct nfs_pgio_header *hdr)
1200 {
1201         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1202                 return nfs_flush_multi(desc, hdr);
1203         return nfs_flush_one(desc, hdr);
1204 }
1205 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1206
1207 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1208 {
1209         struct nfs_rw_header *whdr;
1210         struct nfs_pgio_header *hdr;
1211         int ret;
1212
1213         whdr = nfs_writehdr_alloc();
1214         if (!whdr) {
1215                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1216                 return -ENOMEM;
1217         }
1218         hdr = &whdr->header;
1219         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1220         atomic_inc(&hdr->refcnt);
1221         ret = nfs_generic_flush(desc, hdr);
1222         if (ret == 0)
1223                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1224                                              desc->pg_rpc_callops,
1225                                              desc->pg_ioflags);
1226         if (atomic_dec_and_test(&hdr->refcnt))
1227                 hdr->completion_ops->completion(hdr);
1228         return ret;
1229 }
1230
1231 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1232         .pg_test = nfs_generic_pg_test,
1233         .pg_doio = nfs_generic_pg_writepages,
1234 };
1235
1236 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1237                                struct inode *inode, int ioflags, bool force_mds,
1238                                const struct nfs_pgio_completion_ops *compl_ops)
1239 {
1240         struct nfs_server *server = NFS_SERVER(inode);
1241         const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1242
1243 #ifdef CONFIG_NFS_V4_1
1244         if (server->pnfs_curr_ld && !force_mds)
1245                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1246 #endif
1247         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, server->wsize, ioflags);
1248 }
1249 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1250
1251 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1252 {
1253         pgio->pg_ops = &nfs_pageio_write_ops;
1254         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1255 }
1256 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1257
1258
1259 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1260 {
1261         struct nfs_pgio_data *data = calldata;
1262         int err;
1263         err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1264         if (err)
1265                 rpc_exit(task, err);
1266 }
1267
1268 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1269 {
1270         struct nfs_commit_data *data = calldata;
1271
1272         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1273 }
1274
1275 /*
1276  * Handle a write reply that flushes a whole page.
1277  *
1278  * FIXME: There is an inherent race with invalidate_inode_pages and
1279  *        writebacks since the page->count is kept > 1 for as long
1280  *        as the page has a write request pending.
1281  */
1282 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1283 {
1284         struct nfs_pgio_data    *data = calldata;
1285
1286         nfs_writeback_done(task, data);
1287 }
1288
1289 static void nfs_writeback_release_common(void *calldata)
1290 {
1291         struct nfs_pgio_data    *data = calldata;
1292         struct nfs_pgio_header *hdr = data->header;
1293         int status = data->task.tk_status;
1294
1295         if ((status >= 0) && nfs_write_need_commit(data)) {
1296                 spin_lock(&hdr->lock);
1297                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1298                         ; /* Do nothing */
1299                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1300                         memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1301                 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1302                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1303                 spin_unlock(&hdr->lock);
1304         }
1305         nfs_pgio_data_release(data);
1306 }
1307
1308 static const struct rpc_call_ops nfs_write_common_ops = {
1309         .rpc_call_prepare = nfs_write_prepare,
1310         .rpc_call_done = nfs_writeback_done_common,
1311         .rpc_release = nfs_writeback_release_common,
1312 };
1313
1314 /*
1315  * Special version of should_remove_suid() that ignores capabilities.
1316  */
1317 static int nfs_should_remove_suid(const struct inode *inode)
1318 {
1319         umode_t mode = inode->i_mode;
1320         int kill = 0;
1321
1322         /* suid always must be killed */
1323         if (unlikely(mode & S_ISUID))
1324                 kill = ATTR_KILL_SUID;
1325
1326         /*
1327          * sgid without any exec bits is just a mandatory locking mark; leave
1328          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1329          */
1330         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1331                 kill |= ATTR_KILL_SGID;
1332
1333         if (unlikely(kill && S_ISREG(mode)))
1334                 return kill;
1335
1336         return 0;
1337 }
1338
1339 /*
1340  * This function is called when the WRITE call is complete.
1341  */
1342 void nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data)
1343 {
1344         struct nfs_pgio_args    *argp = &data->args;
1345         struct nfs_pgio_res     *resp = &data->res;
1346         struct inode            *inode = data->header->inode;
1347         int status;
1348
1349         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1350                 task->tk_pid, task->tk_status);
1351
1352         /*
1353          * ->write_done will attempt to use post-op attributes to detect
1354          * conflicting writes by other clients.  A strict interpretation
1355          * of close-to-open would allow us to continue caching even if
1356          * another writer had changed the file, but some applications
1357          * depend on tighter cache coherency when writing.
1358          */
1359         status = NFS_PROTO(inode)->write_done(task, data);
1360         if (status != 0)
1361                 return;
1362         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1363
1364 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1365         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1366                 /* We tried a write call, but the server did not
1367                  * commit data to stable storage even though we
1368                  * requested it.
1369                  * Note: There is a known bug in Tru64 < 5.0 in which
1370                  *       the server reports NFS_DATA_SYNC, but performs
1371                  *       NFS_FILE_SYNC. We therefore implement this checking
1372                  *       as a dprintk() in order to avoid filling syslog.
1373                  */
1374                 static unsigned long    complain;
1375
1376                 /* Note this will print the MDS for a DS write */
1377                 if (time_before(complain, jiffies)) {
1378                         dprintk("NFS:       faulty NFS server %s:"
1379                                 " (committed = %d) != (stable = %d)\n",
1380                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1381                                 resp->verf->committed, argp->stable);
1382                         complain = jiffies + 300 * HZ;
1383                 }
1384         }
1385 #endif
1386         if (task->tk_status < 0) {
1387                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1388                 return;
1389         }
1390
1391         /* Deal with the suid/sgid bit corner case */
1392         if (nfs_should_remove_suid(inode))
1393                 nfs_mark_for_revalidate(inode);
1394
1395         if (resp->count < argp->count) {
1396                 static unsigned long    complain;
1397
1398                 /* This a short write! */
1399                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1400
1401                 /* Has the server at least made some progress? */
1402                 if (resp->count == 0) {
1403                         if (time_before(complain, jiffies)) {
1404                                 printk(KERN_WARNING
1405                                        "NFS: Server wrote zero bytes, expected %u.\n",
1406                                        argp->count);
1407                                 complain = jiffies + 300 * HZ;
1408                         }
1409                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1410                         task->tk_status = -EIO;
1411                         return;
1412                 }
1413                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1414                 if (resp->verf->committed != NFS_UNSTABLE) {
1415                         /* Resend from where the server left off */
1416                         data->mds_offset += resp->count;
1417                         argp->offset += resp->count;
1418                         argp->pgbase += resp->count;
1419                         argp->count -= resp->count;
1420                 } else {
1421                         /* Resend as a stable write in order to avoid
1422                          * headaches in the case of a server crash.
1423                          */
1424                         argp->stable = NFS_FILE_SYNC;
1425                 }
1426                 rpc_restart_call_prepare(task);
1427         }
1428 }
1429
1430
1431 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1432 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1433 {
1434         int ret;
1435
1436         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1437                 return 1;
1438         if (!may_wait)
1439                 return 0;
1440         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1441                                 NFS_INO_COMMIT,
1442                                 nfs_wait_bit_killable,
1443                                 TASK_KILLABLE);
1444         return (ret < 0) ? ret : 1;
1445 }
1446
1447 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1448 {
1449         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1450         smp_mb__after_clear_bit();
1451         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1452 }
1453
1454 void nfs_commitdata_release(struct nfs_commit_data *data)
1455 {
1456         put_nfs_open_context(data->context);
1457         nfs_commit_free(data);
1458 }
1459 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1460
1461 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1462                         const struct rpc_call_ops *call_ops,
1463                         int how, int flags)
1464 {
1465         struct rpc_task *task;
1466         int priority = flush_task_priority(how);
1467         struct rpc_message msg = {
1468                 .rpc_argp = &data->args,
1469                 .rpc_resp = &data->res,
1470                 .rpc_cred = data->cred,
1471         };
1472         struct rpc_task_setup task_setup_data = {
1473                 .task = &data->task,
1474                 .rpc_client = clnt,
1475                 .rpc_message = &msg,
1476                 .callback_ops = call_ops,
1477                 .callback_data = data,
1478                 .workqueue = nfsiod_workqueue,
1479                 .flags = RPC_TASK_ASYNC | flags,
1480                 .priority = priority,
1481         };
1482         /* Set up the initial task struct.  */
1483         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1484
1485         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1486
1487         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1488                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1489
1490         task = rpc_run_task(&task_setup_data);
1491         if (IS_ERR(task))
1492                 return PTR_ERR(task);
1493         if (how & FLUSH_SYNC)
1494                 rpc_wait_for_completion_task(task);
1495         rpc_put_task(task);
1496         return 0;
1497 }
1498 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1499
1500 /*
1501  * Set up the argument/result storage required for the RPC call.
1502  */
1503 void nfs_init_commit(struct nfs_commit_data *data,
1504                      struct list_head *head,
1505                      struct pnfs_layout_segment *lseg,
1506                      struct nfs_commit_info *cinfo)
1507 {
1508         struct nfs_page *first = nfs_list_entry(head->next);
1509         struct inode *inode = first->wb_context->dentry->d_inode;
1510
1511         /* Set up the RPC argument and reply structs
1512          * NB: take care not to mess about with data->commit et al. */
1513
1514         list_splice_init(head, &data->pages);
1515
1516         data->inode       = inode;
1517         data->cred        = first->wb_context->cred;
1518         data->lseg        = lseg; /* reference transferred */
1519         data->mds_ops     = &nfs_commit_ops;
1520         data->completion_ops = cinfo->completion_ops;
1521         data->dreq        = cinfo->dreq;
1522
1523         data->args.fh     = NFS_FH(data->inode);
1524         /* Note: we always request a commit of the entire inode */
1525         data->args.offset = 0;
1526         data->args.count  = 0;
1527         data->context     = get_nfs_open_context(first->wb_context);
1528         data->res.fattr   = &data->fattr;
1529         data->res.verf    = &data->verf;
1530         nfs_fattr_init(&data->fattr);
1531 }
1532 EXPORT_SYMBOL_GPL(nfs_init_commit);
1533
1534 void nfs_retry_commit(struct list_head *page_list,
1535                       struct pnfs_layout_segment *lseg,
1536                       struct nfs_commit_info *cinfo)
1537 {
1538         struct nfs_page *req;
1539
1540         while (!list_empty(page_list)) {
1541                 req = nfs_list_entry(page_list->next);
1542                 nfs_list_remove_request(req);
1543                 nfs_mark_request_commit(req, lseg, cinfo);
1544                 if (!cinfo->dreq) {
1545                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1546                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1547                                      BDI_RECLAIMABLE);
1548                 }
1549                 nfs_unlock_and_release_request(req);
1550         }
1551 }
1552 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1553
1554 /*
1555  * Commit dirty pages
1556  */
1557 static int
1558 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1559                 struct nfs_commit_info *cinfo)
1560 {
1561         struct nfs_commit_data  *data;
1562
1563         data = nfs_commitdata_alloc();
1564
1565         if (!data)
1566                 goto out_bad;
1567
1568         /* Set up the argument struct */
1569         nfs_init_commit(data, head, NULL, cinfo);
1570         atomic_inc(&cinfo->mds->rpcs_out);
1571         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1572                                    how, 0);
1573  out_bad:
1574         nfs_retry_commit(head, NULL, cinfo);
1575         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1576         return -ENOMEM;
1577 }
1578
1579 /*
1580  * COMMIT call returned
1581  */
1582 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1583 {
1584         struct nfs_commit_data  *data = calldata;
1585
1586         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1587                                 task->tk_pid, task->tk_status);
1588
1589         /* Call the NFS version-specific code */
1590         NFS_PROTO(data->inode)->commit_done(task, data);
1591 }
1592
1593 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1594 {
1595         struct nfs_page *req;
1596         int status = data->task.tk_status;
1597         struct nfs_commit_info cinfo;
1598
1599         while (!list_empty(&data->pages)) {
1600                 req = nfs_list_entry(data->pages.next);
1601                 nfs_list_remove_request(req);
1602                 nfs_clear_page_commit(req->wb_page);
1603
1604                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1605                         req->wb_context->dentry->d_sb->s_id,
1606                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1607                         req->wb_bytes,
1608                         (long long)req_offset(req));
1609                 if (status < 0) {
1610                         nfs_context_set_write_error(req->wb_context, status);
1611                         nfs_inode_remove_request(req);
1612                         dprintk(", error = %d\n", status);
1613                         goto next;
1614                 }
1615
1616                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1617                  * returned by the server against all stored verfs. */
1618                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1619                         /* We have a match */
1620                         nfs_inode_remove_request(req);
1621                         dprintk(" OK\n");
1622                         goto next;
1623                 }
1624                 /* We have a mismatch. Write the page again */
1625                 dprintk(" mismatch\n");
1626                 nfs_mark_request_dirty(req);
1627                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1628         next:
1629                 nfs_unlock_and_release_request(req);
1630         }
1631         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1632         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1633                 nfs_commit_clear_lock(NFS_I(data->inode));
1634 }
1635
1636 static void nfs_commit_release(void *calldata)
1637 {
1638         struct nfs_commit_data *data = calldata;
1639
1640         data->completion_ops->completion(data);
1641         nfs_commitdata_release(calldata);
1642 }
1643
1644 static const struct rpc_call_ops nfs_commit_ops = {
1645         .rpc_call_prepare = nfs_commit_prepare,
1646         .rpc_call_done = nfs_commit_done,
1647         .rpc_release = nfs_commit_release,
1648 };
1649
1650 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1651         .completion = nfs_commit_release_pages,
1652         .error_cleanup = nfs_commit_clear_lock,
1653 };
1654
1655 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1656                             int how, struct nfs_commit_info *cinfo)
1657 {
1658         int status;
1659
1660         status = pnfs_commit_list(inode, head, how, cinfo);
1661         if (status == PNFS_NOT_ATTEMPTED)
1662                 status = nfs_commit_list(inode, head, how, cinfo);
1663         return status;
1664 }
1665
1666 int nfs_commit_inode(struct inode *inode, int how)
1667 {
1668         LIST_HEAD(head);
1669         struct nfs_commit_info cinfo;
1670         int may_wait = how & FLUSH_SYNC;
1671         int res;
1672
1673         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1674         if (res <= 0)
1675                 goto out_mark_dirty;
1676         nfs_init_cinfo_from_inode(&cinfo, inode);
1677         res = nfs_scan_commit(inode, &head, &cinfo);
1678         if (res) {
1679                 int error;
1680
1681                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1682                 if (error < 0)
1683                         return error;
1684                 if (!may_wait)
1685                         goto out_mark_dirty;
1686                 error = wait_on_bit(&NFS_I(inode)->flags,
1687                                 NFS_INO_COMMIT,
1688                                 nfs_wait_bit_killable,
1689                                 TASK_KILLABLE);
1690                 if (error < 0)
1691                         return error;
1692         } else
1693                 nfs_commit_clear_lock(NFS_I(inode));
1694         return res;
1695         /* Note: If we exit without ensuring that the commit is complete,
1696          * we must mark the inode as dirty. Otherwise, future calls to
1697          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1698          * that the data is on the disk.
1699          */
1700 out_mark_dirty:
1701         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1702         return res;
1703 }
1704
1705 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1706 {
1707         struct nfs_inode *nfsi = NFS_I(inode);
1708         int flags = FLUSH_SYNC;
1709         int ret = 0;
1710
1711         /* no commits means nothing needs to be done */
1712         if (!nfsi->commit_info.ncommit)
1713                 return ret;
1714
1715         if (wbc->sync_mode == WB_SYNC_NONE) {
1716                 /* Don't commit yet if this is a non-blocking flush and there
1717                  * are a lot of outstanding writes for this mapping.
1718                  */
1719                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1720                         goto out_mark_dirty;
1721
1722                 /* don't wait for the COMMIT response */
1723                 flags = 0;
1724         }
1725
1726         ret = nfs_commit_inode(inode, flags);
1727         if (ret >= 0) {
1728                 if (wbc->sync_mode == WB_SYNC_NONE) {
1729                         if (ret < wbc->nr_to_write)
1730                                 wbc->nr_to_write -= ret;
1731                         else
1732                                 wbc->nr_to_write = 0;
1733                 }
1734                 return 0;
1735         }
1736 out_mark_dirty:
1737         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1738         return ret;
1739 }
1740 #else
1741 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1742 {
1743         return 0;
1744 }
1745 #endif
1746
1747 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1748 {
1749         return nfs_commit_unstable_pages(inode, wbc);
1750 }
1751 EXPORT_SYMBOL_GPL(nfs_write_inode);
1752
1753 /*
1754  * flush the inode to disk.
1755  */
1756 int nfs_wb_all(struct inode *inode)
1757 {
1758         struct writeback_control wbc = {
1759                 .sync_mode = WB_SYNC_ALL,
1760                 .nr_to_write = LONG_MAX,
1761                 .range_start = 0,
1762                 .range_end = LLONG_MAX,
1763         };
1764         int ret;
1765
1766         trace_nfs_writeback_inode_enter(inode);
1767
1768         ret = sync_inode(inode, &wbc);
1769
1770         trace_nfs_writeback_inode_exit(inode, ret);
1771         return ret;
1772 }
1773 EXPORT_SYMBOL_GPL(nfs_wb_all);
1774
1775 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1776 {
1777         struct nfs_page *req;
1778         int ret = 0;
1779
1780         for (;;) {
1781                 wait_on_page_writeback(page);
1782                 req = nfs_page_find_request(page);
1783                 if (req == NULL)
1784                         break;
1785                 if (nfs_lock_request(req)) {
1786                         nfs_clear_request_commit(req);
1787                         nfs_inode_remove_request(req);
1788                         /*
1789                          * In case nfs_inode_remove_request has marked the
1790                          * page as being dirty
1791                          */
1792                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1793                         nfs_unlock_and_release_request(req);
1794                         break;
1795                 }
1796                 ret = nfs_wait_on_request(req);
1797                 nfs_release_request(req);
1798                 if (ret < 0)
1799                         break;
1800         }
1801         return ret;
1802 }
1803
1804 /*
1805  * Write back all requests on one page - we do this before reading it.
1806  */
1807 int nfs_wb_page(struct inode *inode, struct page *page)
1808 {
1809         loff_t range_start = page_file_offset(page);
1810         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1811         struct writeback_control wbc = {
1812                 .sync_mode = WB_SYNC_ALL,
1813                 .nr_to_write = 0,
1814                 .range_start = range_start,
1815                 .range_end = range_end,
1816         };
1817         int ret;
1818
1819         trace_nfs_writeback_page_enter(inode);
1820
1821         for (;;) {
1822                 wait_on_page_writeback(page);
1823                 if (clear_page_dirty_for_io(page)) {
1824                         ret = nfs_writepage_locked(page, &wbc);
1825                         if (ret < 0)
1826                                 goto out_error;
1827                         continue;
1828                 }
1829                 ret = 0;
1830                 if (!PagePrivate(page))
1831                         break;
1832                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1833                 if (ret < 0)
1834                         goto out_error;
1835         }
1836 out_error:
1837         trace_nfs_writeback_page_exit(inode, ret);
1838         return ret;
1839 }
1840
1841 #ifdef CONFIG_MIGRATION
1842 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1843                 struct page *page, enum migrate_mode mode)
1844 {
1845         /*
1846          * If PagePrivate is set, then the page is currently associated with
1847          * an in-progress read or write request. Don't try to migrate it.
1848          *
1849          * FIXME: we could do this in principle, but we'll need a way to ensure
1850          *        that we can safely release the inode reference while holding
1851          *        the page lock.
1852          */
1853         if (PagePrivate(page))
1854                 return -EBUSY;
1855
1856         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1857                 return -EBUSY;
1858
1859         return migrate_page(mapping, newpage, page, mode);
1860 }
1861 #endif
1862
1863 int __init nfs_init_writepagecache(void)
1864 {
1865         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1866                                              sizeof(struct nfs_rw_header),
1867                                              0, SLAB_HWCACHE_ALIGN,
1868                                              NULL);
1869         if (nfs_wdata_cachep == NULL)
1870                 return -ENOMEM;
1871
1872         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1873                                                      nfs_wdata_cachep);
1874         if (nfs_wdata_mempool == NULL)
1875                 goto out_destroy_write_cache;
1876
1877         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1878                                              sizeof(struct nfs_commit_data),
1879                                              0, SLAB_HWCACHE_ALIGN,
1880                                              NULL);
1881         if (nfs_cdata_cachep == NULL)
1882                 goto out_destroy_write_mempool;
1883
1884         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1885                                                       nfs_cdata_cachep);
1886         if (nfs_commit_mempool == NULL)
1887                 goto out_destroy_commit_cache;
1888
1889         /*
1890          * NFS congestion size, scale with available memory.
1891          *
1892          *  64MB:    8192k
1893          * 128MB:   11585k
1894          * 256MB:   16384k
1895          * 512MB:   23170k
1896          *   1GB:   32768k
1897          *   2GB:   46340k
1898          *   4GB:   65536k
1899          *   8GB:   92681k
1900          *  16GB:  131072k
1901          *
1902          * This allows larger machines to have larger/more transfers.
1903          * Limit the default to 256M
1904          */
1905         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1906         if (nfs_congestion_kb > 256*1024)
1907                 nfs_congestion_kb = 256*1024;
1908
1909         return 0;
1910
1911 out_destroy_commit_cache:
1912         kmem_cache_destroy(nfs_cdata_cachep);
1913 out_destroy_write_mempool:
1914         mempool_destroy(nfs_wdata_mempool);
1915 out_destroy_write_cache:
1916         kmem_cache_destroy(nfs_wdata_cachep);
1917         return -ENOMEM;
1918 }
1919
1920 void nfs_destroy_writepagecache(void)
1921 {
1922         mempool_destroy(nfs_commit_mempool);
1923         kmem_cache_destroy(nfs_cdata_cachep);
1924         mempool_destroy(nfs_wdata_mempool);
1925         kmem_cache_destroy(nfs_wdata_cachep);
1926 }
1927