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                 hdr->verf = &p->verf;
  86         }
  87         return p;
  88 }
  89 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
  90
  91 static struct nfs_pgio_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
  92                                                   unsigned int pagecount)
  93 {
  94         struct nfs_pgio_data *data, *prealloc;
  95
  96         prealloc = &container_of(hdr, struct nfs_rw_header, header)->rpc_data;
  97         if (prealloc->header == NULL)
  98                 data = prealloc;
  99         else
 100                 data = kzalloc(sizeof(*data), GFP_KERNEL);
 101         if (!data)
 102                 goto out;
 103
 104         if (nfs_pgarray_set(&data->pages, pagecount)) {
 105                 data->header = hdr;
 106                 atomic_inc(&hdr->refcnt);
 107         } else {
 108                 if (data != prealloc)
 109                         kfree(data);
 110                 data = NULL;
 111         }
 112 out:
 113         return data;
 114 }
 115
 116 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 117 {
 118         struct nfs_rw_header *whdr = container_of(hdr, struct nfs_rw_header, header);
 119         mempool_free(whdr, nfs_wdata_mempool);
 120 }
 121 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
 122
 123 void nfs_writedata_release(struct nfs_pgio_data *wdata)
 124 {
 125         struct nfs_pgio_header *hdr = wdata->header;
 126         struct nfs_rw_header *write_header = container_of(hdr, struct nfs_rw_header, header);
 127
 128         put_nfs_open_context(wdata->args.context);
 129         if (wdata->pages.pagevec != wdata->pages.page_array)
 130                 kfree(wdata->pages.pagevec);
 131         if (wdata == &write_header->rpc_data) {
 132                 wdata->header = NULL;
 133                 wdata = NULL;
 134         }
 135         if (atomic_dec_and_test(&hdr->refcnt))
 136                 hdr->completion_ops->completion(hdr);
 137         /* Note: we only free the rpc_task after callbacks are done.
 138          * See the comment in rpc_free_task() for why
 139          */
 140         kfree(wdata);
 141 }
 142 EXPORT_SYMBOL_GPL(nfs_writedata_release);
 143
 144 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 145 {
 146         ctx->error = error;
 147         smp_wmb();
 148         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 149 }
 150
 151 static struct nfs_page *
 152 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
 153 {
 154         struct nfs_page *req = NULL;
 155
 156         if (PagePrivate(page))
 157                 req = (struct nfs_page *)page_private(page);
 158         else if (unlikely(PageSwapCache(page))) {
 159                 struct nfs_page *freq, *t;
 160
 161                 /* Linearly search the commit list for the correct req */
 162                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 163                         if (freq->wb_page == page) {
 164                                 req = freq;
 165                                 break;
 166                         }
 167                 }
 168         }
 169
 170         if (req)
 171                 kref_get(&req->wb_kref);
 172
 173         return req;
 174 }
 175
 176 static struct nfs_page *nfs_page_find_request(struct page *page)
 177 {
 178         struct inode *inode = page_file_mapping(page)->host;
 179         struct nfs_page *req = NULL;
 180
 181         spin_lock(&inode->i_lock);
 182         req = nfs_page_find_request_locked(NFS_I(inode), page);
 183         spin_unlock(&inode->i_lock);
 184         return req;
 185 }
 186
 187 /* Adjust the file length if we're writing beyond the end */
 188 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 189 {
 190         struct inode *inode = page_file_mapping(page)->host;
 191         loff_t end, i_size;
 192         pgoff_t end_index;
 193
 194         spin_lock(&inode->i_lock);
 195         i_size = i_size_read(inode);
 196         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 197         if (i_size > 0 && page_file_index(page) < end_index)
 198                 goto out;
 199         end = page_file_offset(page) + ((loff_t)offset+count);
 200         if (i_size >= end)
 201                 goto out;
 202         i_size_write(inode, end);
 203         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 204 out:
 205         spin_unlock(&inode->i_lock);
 206 }
 207
 208 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 209 static void nfs_set_pageerror(struct page *page)
 210 {
 211         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 212 }
 213
 214 /* We can set the PG_uptodate flag if we see that a write request
 215  * covers the full page.
 216  */
 217 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 218 {
 219         if (PageUptodate(page))
 220                 return;
 221         if (base != 0)
 222                 return;
 223         if (count != nfs_page_length(page))
 224                 return;
 225         SetPageUptodate(page);
 226 }
 227
 228 static int wb_priority(struct writeback_control *wbc)
 229 {
 230         if (wbc->for_reclaim)
 231                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 232         if (wbc->for_kupdate || wbc->for_background)
 233                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 234         return FLUSH_COND_STABLE;
 235 }
 236
 237 /*
 238  * NFS congestion control
 239  */
 240
 241 int nfs_congestion_kb;
 242
 243 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 244 #define NFS_CONGESTION_OFF_THRESH       \
 245         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 246
 247 static void nfs_set_page_writeback(struct page *page)
 248 {
 249         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 250         int ret = test_set_page_writeback(page);
 251
 252         WARN_ON_ONCE(ret != 0);
 253
 254         if (atomic_long_inc_return(&nfss->writeback) >
 255                         NFS_CONGESTION_ON_THRESH) {
 256                 set_bdi_congested(&nfss->backing_dev_info,
 257                                         BLK_RW_ASYNC);
 258         }
 259 }
 260
 261 static void nfs_end_page_writeback(struct page *page)
 262 {
 263         struct inode *inode = page_file_mapping(page)->host;
 264         struct nfs_server *nfss = NFS_SERVER(inode);
 265
 266         end_page_writeback(page);
 267         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 268                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 269 }
 270
 271 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 272 {
 273         struct inode *inode = page_file_mapping(page)->host;
 274         struct nfs_page *req;
 275         int ret;
 276
 277         spin_lock(&inode->i_lock);
 278         for (;;) {
 279                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 280                 if (req == NULL)
 281                         break;
 282                 if (nfs_lock_request(req))
 283                         break;
 284                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 285                  *       then the call to nfs_lock_request() will always
 286                  *       succeed provided that someone hasn't already marked the
 287                  *       request as dirty (in which case we don't care).
 288                  */
 289                 spin_unlock(&inode->i_lock);
 290                 if (!nonblock)
 291                         ret = nfs_wait_on_request(req);
 292                 else
 293                         ret = -EAGAIN;
 294                 nfs_release_request(req);
 295                 if (ret != 0)
 296                         return ERR_PTR(ret);
 297                 spin_lock(&inode->i_lock);
 298         }
 299         spin_unlock(&inode->i_lock);
 300         return req;
 301 }
 302
 303 /*
 304  * Find an associated nfs write request, and prepare to flush it out
 305  * May return an error if the user signalled nfs_wait_on_request().
 306  */
 307 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 308                                 struct page *page, bool nonblock)
 309 {
 310         struct nfs_page *req;
 311         int ret = 0;
 312
 313         req = nfs_find_and_lock_request(page, nonblock);
 314         if (!req)
 315                 goto out;
 316         ret = PTR_ERR(req);
 317         if (IS_ERR(req))
 318                 goto out;
 319
 320         nfs_set_page_writeback(page);
 321         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 322
 323         ret = 0;
 324         if (!nfs_pageio_add_request(pgio, req)) {
 325                 nfs_redirty_request(req);
 326                 ret = pgio->pg_error;
 327         }
 328 out:
 329         return ret;
 330 }
 331
 332 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 333 {
 334         struct inode *inode = page_file_mapping(page)->host;
 335         int ret;
 336
 337         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 338         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 339
 340         nfs_pageio_cond_complete(pgio, page_file_index(page));
 341         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 342         if (ret == -EAGAIN) {
 343                 redirty_page_for_writepage(wbc, page);
 344                 ret = 0;
 345         }
 346         return ret;
 347 }
 348
 349 /*
 350  * Write an mmapped page to the server.
 351  */
 352 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 353 {
 354         struct nfs_pageio_descriptor pgio;
 355         int err;
 356
 357         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 358                                 false, &nfs_async_write_completion_ops);
 359         err = nfs_do_writepage(page, wbc, &pgio);
 360         nfs_pageio_complete(&pgio);
 361         if (err < 0)
 362                 return err;
 363         if (pgio.pg_error < 0)
 364                 return pgio.pg_error;
 365         return 0;
 366 }
 367
 368 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 369 {
 370         int ret;
 371
 372         ret = nfs_writepage_locked(page, wbc);
 373         unlock_page(page);
 374         return ret;
 375 }
 376
 377 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 378 {
 379         int ret;
 380
 381         ret = nfs_do_writepage(page, wbc, data);
 382         unlock_page(page);
 383         return ret;
 384 }
 385
 386 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 387 {
 388         struct inode *inode = mapping->host;
 389         unsigned long *bitlock = &NFS_I(inode)->flags;
 390         struct nfs_pageio_descriptor pgio;
 391         int err;
 392
 393         /* Stop dirtying of new pages while we sync */
 394         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 395                         nfs_wait_bit_killable, TASK_KILLABLE);
 396         if (err)
 397                 goto out_err;
 398
 399         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 400
 401         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 402                                 &nfs_async_write_completion_ops);
 403         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 404         nfs_pageio_complete(&pgio);
 405
 406         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 407         smp_mb__after_clear_bit();
 408         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 409
 410         if (err < 0)
 411                 goto out_err;
 412         err = pgio.pg_error;
 413         if (err < 0)
 414                 goto out_err;
 415         return 0;
 416 out_err:
 417         return err;
 418 }
 419
 420 /*
 421  * Insert a write request into an inode
 422  */
 423 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 424 {
 425         struct nfs_inode *nfsi = NFS_I(inode);
 426
 427         /* Lock the request! */
 428         nfs_lock_request(req);
 429
 430         spin_lock(&inode->i_lock);
 431         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 432                 inode->i_version++;
 433         /*
 434          * Swap-space should not get truncated. Hence no need to plug the race
 435          * with invalidate/truncate.
 436          */
 437         if (likely(!PageSwapCache(req->wb_page))) {
 438                 set_bit(PG_MAPPED, &req->wb_flags);
 439                 SetPagePrivate(req->wb_page);
 440                 set_page_private(req->wb_page, (unsigned long)req);
 441         }
 442         nfsi->npages++;
 443         kref_get(&req->wb_kref);
 444         spin_unlock(&inode->i_lock);
 445 }
 446
 447 /*
 448  * Remove a write request from an inode
 449  */
 450 static void nfs_inode_remove_request(struct nfs_page *req)
 451 {
 452         struct inode *inode = req->wb_context->dentry->d_inode;
 453         struct nfs_inode *nfsi = NFS_I(inode);
 454
 455         spin_lock(&inode->i_lock);
 456         if (likely(!PageSwapCache(req->wb_page))) {
 457                 set_page_private(req->wb_page, 0);
 458                 ClearPagePrivate(req->wb_page);
 459                 clear_bit(PG_MAPPED, &req->wb_flags);
 460         }
 461         nfsi->npages--;
 462         spin_unlock(&inode->i_lock);
 463         nfs_release_request(req);
 464 }
 465
 466 static void
 467 nfs_mark_request_dirty(struct nfs_page *req)
 468 {
 469         __set_page_dirty_nobuffers(req->wb_page);
 470 }
 471
 472 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 473 /**
 474  * nfs_request_add_commit_list - add request to a commit list
 475  * @req: pointer to a struct nfs_page
 476  * @dst: commit list head
 477  * @cinfo: holds list lock and accounting info
 478  *
 479  * This sets the PG_CLEAN bit, updates the cinfo count of
 480  * number of outstanding requests requiring a commit as well as
 481  * the MM page stats.
 482  *
 483  * The caller must _not_ hold the cinfo->lock, but must be
 484  * holding the nfs_page lock.
 485  */
 486 void
 487 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 488                             struct nfs_commit_info *cinfo)
 489 {
 490         set_bit(PG_CLEAN, &(req)->wb_flags);
 491         spin_lock(cinfo->lock);
 492         nfs_list_add_request(req, dst);
 493         cinfo->mds->ncommit++;
 494         spin_unlock(cinfo->lock);
 495         if (!cinfo->dreq) {
 496                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 497                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 498                              BDI_RECLAIMABLE);
 499                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 500                                    I_DIRTY_DATASYNC);
 501         }
 502 }
 503 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 504
 505 /**
 506  * nfs_request_remove_commit_list - Remove request from a commit list
 507  * @req: pointer to a nfs_page
 508  * @cinfo: holds list lock and accounting info
 509  *
 510  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 511  * number of outstanding requests requiring a commit
 512  * It does not update the MM page stats.
 513  *
 514  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 515  */
 516 void
 517 nfs_request_remove_commit_list(struct nfs_page *req,
 518                                struct nfs_commit_info *cinfo)
 519 {
 520         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 521                 return;
 522         nfs_list_remove_request(req);
 523         cinfo->mds->ncommit--;
 524 }
 525 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 526
 527 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 528                                       struct inode *inode)
 529 {
 530         cinfo->lock = &inode->i_lock;
 531         cinfo->mds = &NFS_I(inode)->commit_info;
 532         cinfo->ds = pnfs_get_ds_info(inode);
 533         cinfo->dreq = NULL;
 534         cinfo->completion_ops = &nfs_commit_completion_ops;
 535 }
 536
 537 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 538                     struct inode *inode,
 539                     struct nfs_direct_req *dreq)
 540 {
 541         if (dreq)
 542                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 543         else
 544                 nfs_init_cinfo_from_inode(cinfo, inode);
 545 }
 546 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 547
 548 /*
 549  * Add a request to the inode's commit list.
 550  */
 551 void
 552 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 553                         struct nfs_commit_info *cinfo)
 554 {
 555         if (pnfs_mark_request_commit(req, lseg, cinfo))
 556                 return;
 557         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 558 }
 559
 560 static void
 561 nfs_clear_page_commit(struct page *page)
 562 {
 563         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 564         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 565 }
 566
 567 static void
 568 nfs_clear_request_commit(struct nfs_page *req)
 569 {
 570         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 571                 struct inode *inode = req->wb_context->dentry->d_inode;
 572                 struct nfs_commit_info cinfo;
 573
 574                 nfs_init_cinfo_from_inode(&cinfo, inode);
 575                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 576                         spin_lock(cinfo.lock);
 577                         nfs_request_remove_commit_list(req, &cinfo);
 578                         spin_unlock(cinfo.lock);
 579                 }
 580                 nfs_clear_page_commit(req->wb_page);
 581         }
 582 }
 583
 584 static inline
 585 int nfs_write_need_commit(struct nfs_pgio_data *data)
 586 {
 587         if (data->verf.committed == NFS_DATA_SYNC)
 588                 return data->header->lseg == NULL;
 589         return data->verf.committed != NFS_FILE_SYNC;
 590 }
 591
 592 #else
 593 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 594                                       struct inode *inode)
 595 {
 596 }
 597
 598 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 599                     struct inode *inode,
 600                     struct nfs_direct_req *dreq)
 601 {
 602 }
 603
 604 void
 605 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 606                         struct nfs_commit_info *cinfo)
 607 {
 608 }
 609
 610 static void
 611 nfs_clear_request_commit(struct nfs_page *req)
 612 {
 613 }
 614
 615 static inline
 616 int nfs_write_need_commit(struct nfs_pgio_data *data)
 617 {
 618         return 0;
 619 }
 620
 621 #endif
 622
 623 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 624 {
 625         struct nfs_commit_info cinfo;
 626         unsigned long bytes = 0;
 627
 628         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 629                 goto out;
 630         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 631         while (!list_empty(&hdr->pages)) {
 632                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 633
 634                 bytes += req->wb_bytes;
 635                 nfs_list_remove_request(req);
 636                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 637                     (hdr->good_bytes < bytes)) {
 638                         nfs_set_pageerror(req->wb_page);
 639                         nfs_context_set_write_error(req->wb_context, hdr->error);
 640                         goto remove_req;
 641                 }
 642                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 643                         nfs_mark_request_dirty(req);
 644                         goto next;
 645                 }
 646                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 647                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
 648                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 649                         goto next;
 650                 }
 651 remove_req:
 652                 nfs_inode_remove_request(req);
 653 next:
 654                 nfs_unlock_request(req);
 655                 nfs_end_page_writeback(req->wb_page);
 656                 nfs_release_request(req);
 657         }
 658 out:
 659         hdr->release(hdr);
 660 }
 661
 662 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 663 static unsigned long
 664 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 665 {
 666         return cinfo->mds->ncommit;
 667 }
 668
 669 /* cinfo->lock held by caller */
 670 int
 671 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 672                      struct nfs_commit_info *cinfo, int max)
 673 {
 674         struct nfs_page *req, *tmp;
 675         int ret = 0;
 676
 677         list_for_each_entry_safe(req, tmp, src, wb_list) {
 678                 if (!nfs_lock_request(req))
 679                         continue;
 680                 kref_get(&req->wb_kref);
 681                 if (cond_resched_lock(cinfo->lock))
 682                         list_safe_reset_next(req, tmp, wb_list);
 683                 nfs_request_remove_commit_list(req, cinfo);
 684                 nfs_list_add_request(req, dst);
 685                 ret++;
 686                 if ((ret == max) && !cinfo->dreq)
 687                         break;
 688         }
 689         return ret;
 690 }
 691
 692 /*
 693  * nfs_scan_commit - Scan an inode for commit requests
 694  * @inode: NFS inode to scan
 695  * @dst: mds destination list
 696  * @cinfo: mds and ds lists of reqs ready to commit
 697  *
 698  * Moves requests from the inode's 'commit' request list.
 699  * The requests are *not* checked to ensure that they form a contiguous set.
 700  */
 701 int
 702 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 703                 struct nfs_commit_info *cinfo)
 704 {
 705         int ret = 0;
 706
 707         spin_lock(cinfo->lock);
 708         if (cinfo->mds->ncommit > 0) {
 709                 const int max = INT_MAX;
 710
 711                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 712                                            cinfo, max);
 713                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 714         }
 715         spin_unlock(cinfo->lock);
 716         return ret;
 717 }
 718
 719 #else
 720 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 721 {
 722         return 0;
 723 }
 724
 725 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 726                     struct nfs_commit_info *cinfo)
 727 {
 728         return 0;
 729 }
 730 #endif
 731
 732 /*
 733  * Search for an existing write request, and attempt to update
 734  * it to reflect a new dirty region on a given page.
 735  *
 736  * If the attempt fails, then the existing request is flushed out
 737  * to disk.
 738  */
 739 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 740                 struct page *page,
 741                 unsigned int offset,
 742                 unsigned int bytes)
 743 {
 744         struct nfs_page *req;
 745         unsigned int rqend;
 746         unsigned int end;
 747         int error;
 748
 749         if (!PagePrivate(page))
 750                 return NULL;
 751
 752         end = offset + bytes;
 753         spin_lock(&inode->i_lock);
 754
 755         for (;;) {
 756                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 757                 if (req == NULL)
 758                         goto out_unlock;
 759
 760                 rqend = req->wb_offset + req->wb_bytes;
 761                 /*
 762                  * Tell the caller to flush out the request if
 763                  * the offsets are non-contiguous.
 764                  * Note: nfs_flush_incompatible() will already
 765                  * have flushed out requests having wrong owners.
 766                  */
 767                 if (offset > rqend
 768                     || end < req->wb_offset)
 769                         goto out_flushme;
 770
 771                 if (nfs_lock_request(req))
 772                         break;
 773
 774                 /* The request is locked, so wait and then retry */
 775                 spin_unlock(&inode->i_lock);
 776                 error = nfs_wait_on_request(req);
 777                 nfs_release_request(req);
 778                 if (error != 0)
 779                         goto out_err;
 780                 spin_lock(&inode->i_lock);
 781         }
 782
 783         /* Okay, the request matches. Update the region */
 784         if (offset < req->wb_offset) {
 785                 req->wb_offset = offset;
 786                 req->wb_pgbase = offset;
 787         }
 788         if (end > rqend)
 789                 req->wb_bytes = end - req->wb_offset;
 790         else
 791                 req->wb_bytes = rqend - req->wb_offset;
 792 out_unlock:
 793         spin_unlock(&inode->i_lock);
 794         if (req)
 795                 nfs_clear_request_commit(req);
 796         return req;
 797 out_flushme:
 798         spin_unlock(&inode->i_lock);
 799         nfs_release_request(req);
 800         error = nfs_wb_page(inode, page);
 801 out_err:
 802         return ERR_PTR(error);
 803 }
 804
 805 /*
 806  * Try to update an existing write request, or create one if there is none.
 807  *
 808  * Note: Should always be called with the Page Lock held to prevent races
 809  * if we have to add a new request. Also assumes that the caller has
 810  * already called nfs_flush_incompatible() if necessary.
 811  */
 812 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 813                 struct page *page, unsigned int offset, unsigned int bytes)
 814 {
 815         struct inode *inode = page_file_mapping(page)->host;
 816         struct nfs_page *req;
 817
 818         req = nfs_try_to_update_request(inode, page, offset, bytes);
 819         if (req != NULL)
 820                 goto out;
 821         req = nfs_create_request(ctx, inode, page, offset, bytes);
 822         if (IS_ERR(req))
 823                 goto out;
 824         nfs_inode_add_request(inode, req);
 825 out:
 826         return req;
 827 }
 828
 829 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 830                 unsigned int offset, unsigned int count)
 831 {
 832         struct nfs_page *req;
 833
 834         req = nfs_setup_write_request(ctx, page, offset, count);
 835         if (IS_ERR(req))
 836                 return PTR_ERR(req);
 837         /* Update file length */
 838         nfs_grow_file(page, offset, count);
 839         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 840         nfs_mark_request_dirty(req);
 841         nfs_unlock_and_release_request(req);
 842         return 0;
 843 }
 844
 845 int nfs_flush_incompatible(struct file *file, struct page *page)
 846 {
 847         struct nfs_open_context *ctx = nfs_file_open_context(file);
 848         struct nfs_lock_context *l_ctx;
 849         struct nfs_page *req;
 850         int do_flush, status;
 851         /*
 852          * Look for a request corresponding to this page. If there
 853          * is one, and it belongs to another file, we flush it out
 854          * before we try to copy anything into the page. Do this
 855          * due to the lack of an ACCESS-type call in NFSv2.
 856          * Also do the same if we find a request from an existing
 857          * dropped page.
 858          */
 859         do {
 860                 req = nfs_page_find_request(page);
 861                 if (req == NULL)
 862                         return 0;
 863                 l_ctx = req->wb_lock_context;
 864                 do_flush = req->wb_page != page || req->wb_context != ctx;
 865                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 866                         do_flush |= l_ctx->lockowner.l_owner != current->files
 867                                 || l_ctx->lockowner.l_pid != current->tgid;
 868                 }
 869                 nfs_release_request(req);
 870                 if (!do_flush)
 871                         return 0;
 872                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 873         } while (status == 0);
 874         return status;
 875 }
 876
 877 /*
 878  * Avoid buffered writes when a open context credential's key would
 879  * expire soon.
 880  *
 881  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 882  *
 883  * Return 0 and set a credential flag which triggers the inode to flush
 884  * and performs  NFS_FILE_SYNC writes if the key will expired within
 885  * RPC_KEY_EXPIRE_TIMEO.
 886  */
 887 int
 888 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 889 {
 890         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 891         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 892
 893         return rpcauth_key_timeout_notify(auth, ctx->cred);
 894 }
 895
 896 /*
 897  * Test if the open context credential key is marked to expire soon.
 898  */
 899 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 900 {
 901         return rpcauth_cred_key_to_expire(ctx->cred);
 902 }
 903
 904 /*
 905  * If the page cache is marked as unsafe or invalid, then we can't rely on
 906  * the PageUptodate() flag. In this case, we will need to turn off
 907  * write optimisations that depend on the page contents being correct.
 908  */
 909 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 910 {
 911         struct nfs_inode *nfsi = NFS_I(inode);
 912
 913         if (nfs_have_delegated_attributes(inode))
 914                 goto out;
 915         if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 916                 return false;
 917         smp_rmb();
 918         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 919                 return false;
 920 out:
 921         return PageUptodate(page) != 0;
 922 }
 923
 924 /* If we know the page is up to date, and we're not using byte range locks (or
 925  * if we have the whole file locked for writing), it may be more efficient to
 926  * extend the write to cover the entire page in order to avoid fragmentation
 927  * inefficiencies.
 928  *
 929  * If the file is opened for synchronous writes then we can just skip the rest
 930  * of the checks.
 931  */
 932 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 933 {
 934         if (file->f_flags & O_DSYNC)
 935                 return 0;
 936         if (!nfs_write_pageuptodate(page, inode))
 937                 return 0;
 938         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 939                 return 1;
 940         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 941                         inode->i_flock->fl_end == OFFSET_MAX &&
 942                         inode->i_flock->fl_type != F_RDLCK))
 943                 return 1;
 944         return 0;
 945 }
 946
 947 /*
 948  * Update and possibly write a cached page of an NFS file.
 949  *
 950  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 951  * things with a page scheduled for an RPC call (e.g. invalidate it).
 952  */
 953 int nfs_updatepage(struct file *file, struct page *page,
 954                 unsigned int offset, unsigned int count)
 955 {
 956         struct nfs_open_context *ctx = nfs_file_open_context(file);
 957         struct inode    *inode = page_file_mapping(page)->host;
 958         int             status = 0;
 959
 960         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 961
 962         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 963                 file, count, (long long)(page_file_offset(page) + offset));
 964
 965         if (nfs_can_extend_write(file, page, inode)) {
 966                 count = max(count + offset, nfs_page_length(page));
 967                 offset = 0;
 968         }
 969
 970         status = nfs_writepage_setup(ctx, page, offset, count);
 971         if (status < 0)
 972                 nfs_set_pageerror(page);
 973         else
 974                 __set_page_dirty_nobuffers(page);
 975
 976         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 977                         status, (long long)i_size_read(inode));
 978         return status;
 979 }
 980
 981 static int flush_task_priority(int how)
 982 {
 983         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 984                 case FLUSH_HIGHPRI:
 985                         return RPC_PRIORITY_HIGH;
 986                 case FLUSH_LOWPRI:
 987                         return RPC_PRIORITY_LOW;
 988         }
 989         return RPC_PRIORITY_NORMAL;
 990 }
 991
 992 int nfs_initiate_write(struct rpc_clnt *clnt,
 993                        struct nfs_pgio_data *data,
 994                        const struct rpc_call_ops *call_ops,
 995                        int how, int flags)
 996 {
 997         struct inode *inode = data->header->inode;
 998         int priority = flush_task_priority(how);
 999         struct rpc_task *task;
1000         struct rpc_message msg = {
1001                 .rpc_argp = &data->args,
1002                 .rpc_resp = &data->res,
1003                 .rpc_cred = data->header->cred,
1004         };
1005         struct rpc_task_setup task_setup_data = {
1006                 .rpc_client = clnt,
1007                 .task = &data->task,
1008                 .rpc_message = &msg,
1009                 .callback_ops = call_ops,
1010                 .callback_data = data,
1011                 .workqueue = nfsiod_workqueue,
1012                 .flags = RPC_TASK_ASYNC | flags,
1013                 .priority = priority,
1014         };
1015         int ret = 0;
1016
1017         /* Set up the initial task struct.  */
1018         NFS_PROTO(inode)->write_setup(data, &msg);
1019
1020         dprintk("NFS: %5u initiated write call "
1021                 "(req %s/%llu, %u bytes @ offset %llu)\n",
1022                 data->task.tk_pid,
1023                 inode->i_sb->s_id,
1024                 (unsigned long long)NFS_FILEID(inode),
1025                 data->args.count,
1026                 (unsigned long long)data->args.offset);
1027
1028         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
1029                                  &task_setup_data.rpc_client, &msg, data);
1030
1031         task = rpc_run_task(&task_setup_data);
1032         if (IS_ERR(task)) {
1033                 ret = PTR_ERR(task);
1034                 goto out;
1035         }
1036         if (how & FLUSH_SYNC) {
1037                 ret = rpc_wait_for_completion_task(task);
1038                 if (ret == 0)
1039                         ret = task->tk_status;
1040         }
1041         rpc_put_task(task);
1042 out:
1043         return ret;
1044 }
1045 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1046
1047 /*
1048  * Set up the argument/result storage required for the RPC call.
1049  */
1050 static void nfs_write_rpcsetup(struct nfs_pgio_data *data,
1051                 unsigned int count, unsigned int offset,
1052                 int how, struct nfs_commit_info *cinfo)
1053 {
1054         struct nfs_page *req = data->header->req;
1055
1056         /* Set up the RPC argument and reply structs
1057          * NB: take care not to mess about with data->commit et al. */
1058
1059         data->args.fh     = NFS_FH(data->header->inode);
1060         data->args.offset = req_offset(req) + offset;
1061         /* pnfs_set_layoutcommit needs this */
1062         data->mds_offset = data->args.offset;
1063         data->args.pgbase = req->wb_pgbase + offset;
1064         data->args.pages  = data->pages.pagevec;
1065         data->args.count  = count;
1066         data->args.context = get_nfs_open_context(req->wb_context);
1067         data->args.lock_context = req->wb_lock_context;
1068         data->args.stable  = NFS_UNSTABLE;
1069         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1070         case 0:
1071                 break;
1072         case FLUSH_COND_STABLE:
1073                 if (nfs_reqs_to_commit(cinfo))
1074                         break;
1075         default:
1076                 data->args.stable = NFS_FILE_SYNC;
1077         }
1078
1079         data->res.fattr   = &data->fattr;
1080         data->res.count   = count;
1081         data->res.verf    = &data->verf;
1082         nfs_fattr_init(&data->fattr);
1083 }
1084
1085 static int nfs_do_write(struct nfs_pgio_data *data,
1086                 const struct rpc_call_ops *call_ops,
1087                 int how)
1088 {
1089         struct inode *inode = data->header->inode;
1090
1091         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1092 }
1093
1094 static int nfs_do_multiple_writes(struct list_head *head,
1095                 const struct rpc_call_ops *call_ops,
1096                 int how)
1097 {
1098         struct nfs_pgio_data *data;
1099         int ret = 0;
1100
1101         while (!list_empty(head)) {
1102                 int ret2;
1103
1104                 data = list_first_entry(head, struct nfs_pgio_data, list);
1105                 list_del_init(&data->list);
1106
1107                 ret2 = nfs_do_write(data, call_ops, how);
1108                  if (ret == 0)
1109                          ret = ret2;
1110         }
1111         return ret;
1112 }
1113
1114 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1115  * call this on each, which will prepare them to be retried on next
1116  * writeback using standard nfs.
1117  */
1118 static void nfs_redirty_request(struct nfs_page *req)
1119 {
1120         nfs_mark_request_dirty(req);
1121         nfs_unlock_request(req);
1122         nfs_end_page_writeback(req->wb_page);
1123         nfs_release_request(req);
1124 }
1125
1126 static void nfs_async_write_error(struct list_head *head)
1127 {
1128         struct nfs_page *req;
1129
1130         while (!list_empty(head)) {
1131                 req = nfs_list_entry(head->next);
1132                 nfs_list_remove_request(req);
1133                 nfs_redirty_request(req);
1134         }
1135 }
1136
1137 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1138         .error_cleanup = nfs_async_write_error,
1139         .completion = nfs_write_completion,
1140 };
1141
1142 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1143                 struct nfs_pgio_header *hdr)
1144 {
1145         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1146         while (!list_empty(&hdr->rpc_list)) {
1147                 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
1148                                 struct nfs_pgio_data, list);
1149                 list_del(&data->list);
1150                 nfs_writedata_release(data);
1151         }
1152         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1153 }
1154
1155 /*
1156  * Generate multiple small requests to write out a single
1157  * contiguous dirty area on one page.
1158  */
1159 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1160                            struct nfs_pgio_header *hdr)
1161 {
1162         struct nfs_page *req = hdr->req;
1163         struct page *page = req->wb_page;
1164         struct nfs_pgio_data *data;
1165         size_t wsize = desc->pg_bsize, nbytes;
1166         unsigned int offset;
1167         int requests = 0;
1168         struct nfs_commit_info cinfo;
1169
1170         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1171
1172         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1173             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1174              desc->pg_count > wsize))
1175                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1176
1177
1178         offset = 0;
1179         nbytes = desc->pg_count;
1180         do {
1181                 size_t len = min(nbytes, wsize);
1182
1183                 data = nfs_writedata_alloc(hdr, 1);
1184                 if (!data) {
1185                         nfs_flush_error(desc, hdr);
1186                         return -ENOMEM;
1187                 }
1188                 data->pages.pagevec[0] = page;
1189                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1190                 list_add(&data->list, &hdr->rpc_list);
1191                 requests++;
1192                 nbytes -= len;
1193                 offset += len;
1194         } while (nbytes != 0);
1195         nfs_list_remove_request(req);
1196         nfs_list_add_request(req, &hdr->pages);
1197         desc->pg_rpc_callops = &nfs_write_common_ops;
1198         return 0;
1199 }
1200
1201 /*
1202  * Create an RPC task for the given write request and kick it.
1203  * The page must have been locked by the caller.
1204  *
1205  * It may happen that the page we're passed is not marked dirty.
1206  * This is the case if nfs_updatepage detects a conflicting request
1207  * that has been written but not committed.
1208  */
1209 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1210                          struct nfs_pgio_header *hdr)
1211 {
1212         struct nfs_page         *req;
1213         struct page             **pages;
1214         struct nfs_pgio_data    *data;
1215         struct list_head *head = &desc->pg_list;
1216         struct nfs_commit_info cinfo;
1217
1218         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1219                                                            desc->pg_count));
1220         if (!data) {
1221                 nfs_flush_error(desc, hdr);
1222                 return -ENOMEM;
1223         }
1224
1225         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1226         pages = data->pages.pagevec;
1227         while (!list_empty(head)) {
1228                 req = nfs_list_entry(head->next);
1229                 nfs_list_remove_request(req);
1230                 nfs_list_add_request(req, &hdr->pages);
1231                 *pages++ = req->wb_page;
1232         }
1233
1234         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1235             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1236                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1237
1238         /* Set up the argument struct */
1239         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1240         list_add(&data->list, &hdr->rpc_list);
1241         desc->pg_rpc_callops = &nfs_write_common_ops;
1242         return 0;
1243 }
1244
1245 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1246                       struct nfs_pgio_header *hdr)
1247 {
1248         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1249                 return nfs_flush_multi(desc, hdr);
1250         return nfs_flush_one(desc, hdr);
1251 }
1252 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1253
1254 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1255 {
1256         struct nfs_rw_header *whdr;
1257         struct nfs_pgio_header *hdr;
1258         int ret;
1259
1260         whdr = nfs_writehdr_alloc();
1261         if (!whdr) {
1262                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1263                 return -ENOMEM;
1264         }
1265         hdr = &whdr->header;
1266         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1267         atomic_inc(&hdr->refcnt);
1268         ret = nfs_generic_flush(desc, hdr);
1269         if (ret == 0)
1270                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1271                                              desc->pg_rpc_callops,
1272                                              desc->pg_ioflags);
1273         if (atomic_dec_and_test(&hdr->refcnt))
1274                 hdr->completion_ops->completion(hdr);
1275         return ret;
1276 }
1277
1278 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1279         .pg_test = nfs_generic_pg_test,
1280         .pg_doio = nfs_generic_pg_writepages,
1281 };
1282
1283 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1284                                struct inode *inode, int ioflags, bool force_mds,
1285                                const struct nfs_pgio_completion_ops *compl_ops)
1286 {
1287         struct nfs_server *server = NFS_SERVER(inode);
1288         const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1289
1290 #ifdef CONFIG_NFS_V4_1
1291         if (server->pnfs_curr_ld && !force_mds)
1292                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1293 #endif
1294         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, server->wsize, ioflags);
1295 }
1296 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1297
1298 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1299 {
1300         pgio->pg_ops = &nfs_pageio_write_ops;
1301         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1302 }
1303 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1304
1305
1306 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1307 {
1308         struct nfs_pgio_data *data = calldata;
1309         int err;
1310         err = NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1311         if (err)
1312                 rpc_exit(task, err);
1313 }
1314
1315 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1316 {
1317         struct nfs_commit_data *data = calldata;
1318
1319         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1320 }
1321
1322 /*
1323  * Handle a write reply that flushes a whole page.
1324  *
1325  * FIXME: There is an inherent race with invalidate_inode_pages and
1326  *        writebacks since the page->count is kept > 1 for as long
1327  *        as the page has a write request pending.
1328  */
1329 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1330 {
1331         struct nfs_pgio_data    *data = calldata;
1332
1333         nfs_writeback_done(task, data);
1334 }
1335
1336 static void nfs_writeback_release_common(void *calldata)
1337 {
1338         struct nfs_pgio_data    *data = calldata;
1339         struct nfs_pgio_header *hdr = data->header;
1340         int status = data->task.tk_status;
1341
1342         if ((status >= 0) && nfs_write_need_commit(data)) {
1343                 spin_lock(&hdr->lock);
1344                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1345                         ; /* Do nothing */
1346                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1347                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1348                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1349                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1350                 spin_unlock(&hdr->lock);
1351         }
1352         nfs_writedata_release(data);
1353 }
1354
1355 static const struct rpc_call_ops nfs_write_common_ops = {
1356         .rpc_call_prepare = nfs_write_prepare,
1357         .rpc_call_done = nfs_writeback_done_common,
1358         .rpc_release = nfs_writeback_release_common,
1359 };
1360
1361 /*
1362  * Special version of should_remove_suid() that ignores capabilities.
1363  */
1364 static int nfs_should_remove_suid(const struct inode *inode)
1365 {
1366         umode_t mode = inode->i_mode;
1367         int kill = 0;
1368
1369         /* suid always must be killed */
1370         if (unlikely(mode & S_ISUID))
1371                 kill = ATTR_KILL_SUID;
1372
1373         /*
1374          * sgid without any exec bits is just a mandatory locking mark; leave
1375          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1376          */
1377         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1378                 kill |= ATTR_KILL_SGID;
1379
1380         if (unlikely(kill && S_ISREG(mode)))
1381                 return kill;
1382
1383         return 0;
1384 }
1385
1386 /*
1387  * This function is called when the WRITE call is complete.
1388  */
1389 void nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data)
1390 {
1391         struct nfs_pgio_args    *argp = &data->args;
1392         struct nfs_pgio_res     *resp = &data->res;
1393         struct inode            *inode = data->header->inode;
1394         int status;
1395
1396         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1397                 task->tk_pid, task->tk_status);
1398
1399         /*
1400          * ->write_done will attempt to use post-op attributes to detect
1401          * conflicting writes by other clients.  A strict interpretation
1402          * of close-to-open would allow us to continue caching even if
1403          * another writer had changed the file, but some applications
1404          * depend on tighter cache coherency when writing.
1405          */
1406         status = NFS_PROTO(inode)->write_done(task, data);
1407         if (status != 0)
1408                 return;
1409         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1410
1411 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1412         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1413                 /* We tried a write call, but the server did not
1414                  * commit data to stable storage even though we
1415                  * requested it.
1416                  * Note: There is a known bug in Tru64 < 5.0 in which
1417                  *       the server reports NFS_DATA_SYNC, but performs
1418                  *       NFS_FILE_SYNC. We therefore implement this checking
1419                  *       as a dprintk() in order to avoid filling syslog.
1420                  */
1421                 static unsigned long    complain;
1422
1423                 /* Note this will print the MDS for a DS write */
1424                 if (time_before(complain, jiffies)) {
1425                         dprintk("NFS:       faulty NFS server %s:"
1426                                 " (committed = %d) != (stable = %d)\n",
1427                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1428                                 resp->verf->committed, argp->stable);
1429                         complain = jiffies + 300 * HZ;
1430                 }
1431         }
1432 #endif
1433         if (task->tk_status < 0) {
1434                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1435                 return;
1436         }
1437
1438         /* Deal with the suid/sgid bit corner case */
1439         if (nfs_should_remove_suid(inode))
1440                 nfs_mark_for_revalidate(inode);
1441
1442         if (resp->count < argp->count) {
1443                 static unsigned long    complain;
1444
1445                 /* This a short write! */
1446                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1447
1448                 /* Has the server at least made some progress? */
1449                 if (resp->count == 0) {
1450                         if (time_before(complain, jiffies)) {
1451                                 printk(KERN_WARNING
1452                                        "NFS: Server wrote zero bytes, expected %u.\n",
1453                                        argp->count);
1454                                 complain = jiffies + 300 * HZ;
1455                         }
1456                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1457                         task->tk_status = -EIO;
1458                         return;
1459                 }
1460                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1461                 if (resp->verf->committed != NFS_UNSTABLE) {
1462                         /* Resend from where the server left off */
1463                         data->mds_offset += resp->count;
1464                         argp->offset += resp->count;
1465                         argp->pgbase += resp->count;
1466                         argp->count -= resp->count;
1467                 } else {
1468                         /* Resend as a stable write in order to avoid
1469                          * headaches in the case of a server crash.
1470                          */
1471                         argp->stable = NFS_FILE_SYNC;
1472                 }
1473                 rpc_restart_call_prepare(task);
1474         }
1475 }
1476
1477
1478 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1479 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1480 {
1481         int ret;
1482
1483         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1484                 return 1;
1485         if (!may_wait)
1486                 return 0;
1487         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1488                                 NFS_INO_COMMIT,
1489                                 nfs_wait_bit_killable,
1490                                 TASK_KILLABLE);
1491         return (ret < 0) ? ret : 1;
1492 }
1493
1494 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1495 {
1496         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1497         smp_mb__after_clear_bit();
1498         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1499 }
1500
1501 void nfs_commitdata_release(struct nfs_commit_data *data)
1502 {
1503         put_nfs_open_context(data->context);
1504         nfs_commit_free(data);
1505 }
1506 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1507
1508 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1509                         const struct rpc_call_ops *call_ops,
1510                         int how, int flags)
1511 {
1512         struct rpc_task *task;
1513         int priority = flush_task_priority(how);
1514         struct rpc_message msg = {
1515                 .rpc_argp = &data->args,
1516                 .rpc_resp = &data->res,
1517                 .rpc_cred = data->cred,
1518         };
1519         struct rpc_task_setup task_setup_data = {
1520                 .task = &data->task,
1521                 .rpc_client = clnt,
1522                 .rpc_message = &msg,
1523                 .callback_ops = call_ops,
1524                 .callback_data = data,
1525                 .workqueue = nfsiod_workqueue,
1526                 .flags = RPC_TASK_ASYNC | flags,
1527                 .priority = priority,
1528         };
1529         /* Set up the initial task struct.  */
1530         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1531
1532         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1533
1534         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1535                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1536
1537         task = rpc_run_task(&task_setup_data);
1538         if (IS_ERR(task))
1539                 return PTR_ERR(task);
1540         if (how & FLUSH_SYNC)
1541                 rpc_wait_for_completion_task(task);
1542         rpc_put_task(task);
1543         return 0;
1544 }
1545 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1546
1547 /*
1548  * Set up the argument/result storage required for the RPC call.
1549  */
1550 void nfs_init_commit(struct nfs_commit_data *data,
1551                      struct list_head *head,
1552                      struct pnfs_layout_segment *lseg,
1553                      struct nfs_commit_info *cinfo)
1554 {
1555         struct nfs_page *first = nfs_list_entry(head->next);
1556         struct inode *inode = first->wb_context->dentry->d_inode;
1557
1558         /* Set up the RPC argument and reply structs
1559          * NB: take care not to mess about with data->commit et al. */
1560
1561         list_splice_init(head, &data->pages);
1562
1563         data->inode       = inode;
1564         data->cred        = first->wb_context->cred;
1565         data->lseg        = lseg; /* reference transferred */
1566         data->mds_ops     = &nfs_commit_ops;
1567         data->completion_ops = cinfo->completion_ops;
1568         data->dreq        = cinfo->dreq;
1569
1570         data->args.fh     = NFS_FH(data->inode);
1571         /* Note: we always request a commit of the entire inode */
1572         data->args.offset = 0;
1573         data->args.count  = 0;
1574         data->context     = get_nfs_open_context(first->wb_context);
1575         data->res.fattr   = &data->fattr;
1576         data->res.verf    = &data->verf;
1577         nfs_fattr_init(&data->fattr);
1578 }
1579 EXPORT_SYMBOL_GPL(nfs_init_commit);
1580
1581 void nfs_retry_commit(struct list_head *page_list,
1582                       struct pnfs_layout_segment *lseg,
1583                       struct nfs_commit_info *cinfo)
1584 {
1585         struct nfs_page *req;
1586
1587         while (!list_empty(page_list)) {
1588                 req = nfs_list_entry(page_list->next);
1589                 nfs_list_remove_request(req);
1590                 nfs_mark_request_commit(req, lseg, cinfo);
1591                 if (!cinfo->dreq) {
1592                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1593                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1594                                      BDI_RECLAIMABLE);
1595                 }
1596                 nfs_unlock_and_release_request(req);
1597         }
1598 }
1599 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1600
1601 /*
1602  * Commit dirty pages
1603  */
1604 static int
1605 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1606                 struct nfs_commit_info *cinfo)
1607 {
1608         struct nfs_commit_data  *data;
1609
1610         data = nfs_commitdata_alloc();
1611
1612         if (!data)
1613                 goto out_bad;
1614
1615         /* Set up the argument struct */
1616         nfs_init_commit(data, head, NULL, cinfo);
1617         atomic_inc(&cinfo->mds->rpcs_out);
1618         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1619                                    how, 0);
1620  out_bad:
1621         nfs_retry_commit(head, NULL, cinfo);
1622         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1623         return -ENOMEM;
1624 }
1625
1626 /*
1627  * COMMIT call returned
1628  */
1629 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1630 {
1631         struct nfs_commit_data  *data = calldata;
1632
1633         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1634                                 task->tk_pid, task->tk_status);
1635
1636         /* Call the NFS version-specific code */
1637         NFS_PROTO(data->inode)->commit_done(task, data);
1638 }
1639
1640 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1641 {
1642         struct nfs_page *req;
1643         int status = data->task.tk_status;
1644         struct nfs_commit_info cinfo;
1645
1646         while (!list_empty(&data->pages)) {
1647                 req = nfs_list_entry(data->pages.next);
1648                 nfs_list_remove_request(req);
1649                 nfs_clear_page_commit(req->wb_page);
1650
1651                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1652                         req->wb_context->dentry->d_sb->s_id,
1653                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1654                         req->wb_bytes,
1655                         (long long)req_offset(req));
1656                 if (status < 0) {
1657                         nfs_context_set_write_error(req->wb_context, status);
1658                         nfs_inode_remove_request(req);
1659                         dprintk(", error = %d\n", status);
1660                         goto next;
1661                 }
1662
1663                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1664                  * returned by the server against all stored verfs. */
1665                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1666                         /* We have a match */
1667                         nfs_inode_remove_request(req);
1668                         dprintk(" OK\n");
1669                         goto next;
1670                 }
1671                 /* We have a mismatch. Write the page again */
1672                 dprintk(" mismatch\n");
1673                 nfs_mark_request_dirty(req);
1674                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1675         next:
1676                 nfs_unlock_and_release_request(req);
1677         }
1678         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1679         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1680                 nfs_commit_clear_lock(NFS_I(data->inode));
1681 }
1682
1683 static void nfs_commit_release(void *calldata)
1684 {
1685         struct nfs_commit_data *data = calldata;
1686
1687         data->completion_ops->completion(data);
1688         nfs_commitdata_release(calldata);
1689 }
1690
1691 static const struct rpc_call_ops nfs_commit_ops = {
1692         .rpc_call_prepare = nfs_commit_prepare,
1693         .rpc_call_done = nfs_commit_done,
1694         .rpc_release = nfs_commit_release,
1695 };
1696
1697 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1698         .completion = nfs_commit_release_pages,
1699         .error_cleanup = nfs_commit_clear_lock,
1700 };
1701
1702 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1703                             int how, struct nfs_commit_info *cinfo)
1704 {
1705         int status;
1706
1707         status = pnfs_commit_list(inode, head, how, cinfo);
1708         if (status == PNFS_NOT_ATTEMPTED)
1709                 status = nfs_commit_list(inode, head, how, cinfo);
1710         return status;
1711 }
1712
1713 int nfs_commit_inode(struct inode *inode, int how)
1714 {
1715         LIST_HEAD(head);
1716         struct nfs_commit_info cinfo;
1717         int may_wait = how & FLUSH_SYNC;
1718         int res;
1719
1720         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1721         if (res <= 0)
1722                 goto out_mark_dirty;
1723         nfs_init_cinfo_from_inode(&cinfo, inode);
1724         res = nfs_scan_commit(inode, &head, &cinfo);
1725         if (res) {
1726                 int error;
1727
1728                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1729                 if (error < 0)
1730                         return error;
1731                 if (!may_wait)
1732                         goto out_mark_dirty;
1733                 error = wait_on_bit(&NFS_I(inode)->flags,
1734                                 NFS_INO_COMMIT,
1735                                 nfs_wait_bit_killable,
1736                                 TASK_KILLABLE);
1737                 if (error < 0)
1738                         return error;
1739         } else
1740                 nfs_commit_clear_lock(NFS_I(inode));
1741         return res;
1742         /* Note: If we exit without ensuring that the commit is complete,
1743          * we must mark the inode as dirty. Otherwise, future calls to
1744          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1745          * that the data is on the disk.
1746          */
1747 out_mark_dirty:
1748         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1749         return res;
1750 }
1751
1752 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1753 {
1754         struct nfs_inode *nfsi = NFS_I(inode);
1755         int flags = FLUSH_SYNC;
1756         int ret = 0;
1757
1758         /* no commits means nothing needs to be done */
1759         if (!nfsi->commit_info.ncommit)
1760                 return ret;
1761
1762         if (wbc->sync_mode == WB_SYNC_NONE) {
1763                 /* Don't commit yet if this is a non-blocking flush and there
1764                  * are a lot of outstanding writes for this mapping.
1765                  */
1766                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1767                         goto out_mark_dirty;
1768
1769                 /* don't wait for the COMMIT response */
1770                 flags = 0;
1771         }
1772
1773         ret = nfs_commit_inode(inode, flags);
1774         if (ret >= 0) {
1775                 if (wbc->sync_mode == WB_SYNC_NONE) {
1776                         if (ret < wbc->nr_to_write)
1777                                 wbc->nr_to_write -= ret;
1778                         else
1779                                 wbc->nr_to_write = 0;
1780                 }
1781                 return 0;
1782         }
1783 out_mark_dirty:
1784         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1785         return ret;
1786 }
1787 #else
1788 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1789 {
1790         return 0;
1791 }
1792 #endif
1793
1794 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1795 {
1796         return nfs_commit_unstable_pages(inode, wbc);
1797 }
1798 EXPORT_SYMBOL_GPL(nfs_write_inode);
1799
1800 /*
1801  * flush the inode to disk.
1802  */
1803 int nfs_wb_all(struct inode *inode)
1804 {
1805         struct writeback_control wbc = {
1806                 .sync_mode = WB_SYNC_ALL,
1807                 .nr_to_write = LONG_MAX,
1808                 .range_start = 0,
1809                 .range_end = LLONG_MAX,
1810         };
1811         int ret;
1812
1813         trace_nfs_writeback_inode_enter(inode);
1814
1815         ret = sync_inode(inode, &wbc);
1816
1817         trace_nfs_writeback_inode_exit(inode, ret);
1818         return ret;
1819 }
1820 EXPORT_SYMBOL_GPL(nfs_wb_all);
1821
1822 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1823 {
1824         struct nfs_page *req;
1825         int ret = 0;
1826
1827         for (;;) {
1828                 wait_on_page_writeback(page);
1829                 req = nfs_page_find_request(page);
1830                 if (req == NULL)
1831                         break;
1832                 if (nfs_lock_request(req)) {
1833                         nfs_clear_request_commit(req);
1834                         nfs_inode_remove_request(req);
1835                         /*
1836                          * In case nfs_inode_remove_request has marked the
1837                          * page as being dirty
1838                          */
1839                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1840                         nfs_unlock_and_release_request(req);
1841                         break;
1842                 }
1843                 ret = nfs_wait_on_request(req);
1844                 nfs_release_request(req);
1845                 if (ret < 0)
1846                         break;
1847         }
1848         return ret;
1849 }
1850
1851 /*
1852  * Write back all requests on one page - we do this before reading it.
1853  */
1854 int nfs_wb_page(struct inode *inode, struct page *page)
1855 {
1856         loff_t range_start = page_file_offset(page);
1857         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1858         struct writeback_control wbc = {
1859                 .sync_mode = WB_SYNC_ALL,
1860                 .nr_to_write = 0,
1861                 .range_start = range_start,
1862                 .range_end = range_end,
1863         };
1864         int ret;
1865
1866         trace_nfs_writeback_page_enter(inode);
1867
1868         for (;;) {
1869                 wait_on_page_writeback(page);
1870                 if (clear_page_dirty_for_io(page)) {
1871                         ret = nfs_writepage_locked(page, &wbc);
1872                         if (ret < 0)
1873                                 goto out_error;
1874                         continue;
1875                 }
1876                 ret = 0;
1877                 if (!PagePrivate(page))
1878                         break;
1879                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1880                 if (ret < 0)
1881                         goto out_error;
1882         }
1883 out_error:
1884         trace_nfs_writeback_page_exit(inode, ret);
1885         return ret;
1886 }
1887
1888 #ifdef CONFIG_MIGRATION
1889 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1890                 struct page *page, enum migrate_mode mode)
1891 {
1892         /*
1893          * If PagePrivate is set, then the page is currently associated with
1894          * an in-progress read or write request. Don't try to migrate it.
1895          *
1896          * FIXME: we could do this in principle, but we'll need a way to ensure
1897          *        that we can safely release the inode reference while holding
1898          *        the page lock.
1899          */
1900         if (PagePrivate(page))
1901                 return -EBUSY;
1902
1903         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1904                 return -EBUSY;
1905
1906         return migrate_page(mapping, newpage, page, mode);
1907 }
1908 #endif
1909
1910 int __init nfs_init_writepagecache(void)
1911 {
1912         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1913                                              sizeof(struct nfs_rw_header),
1914                                              0, SLAB_HWCACHE_ALIGN,
1915                                              NULL);
1916         if (nfs_wdata_cachep == NULL)
1917                 return -ENOMEM;
1918
1919         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1920                                                      nfs_wdata_cachep);
1921         if (nfs_wdata_mempool == NULL)
1922                 goto out_destroy_write_cache;
1923
1924         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1925                                              sizeof(struct nfs_commit_data),
1926                                              0, SLAB_HWCACHE_ALIGN,
1927                                              NULL);
1928         if (nfs_cdata_cachep == NULL)
1929                 goto out_destroy_write_mempool;
1930
1931         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1932                                                       nfs_cdata_cachep);
1933         if (nfs_commit_mempool == NULL)
1934                 goto out_destroy_commit_cache;
1935
1936         /*
1937          * NFS congestion size, scale with available memory.
1938          *
1939          *  64MB:    8192k
1940          * 128MB:   11585k
1941          * 256MB:   16384k
1942          * 512MB:   23170k
1943          *   1GB:   32768k
1944          *   2GB:   46340k
1945          *   4GB:   65536k
1946          *   8GB:   92681k
1947          *  16GB:  131072k
1948          *
1949          * This allows larger machines to have larger/more transfers.
1950          * Limit the default to 256M
1951          */
1952         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1953         if (nfs_congestion_kb > 256*1024)
1954                 nfs_congestion_kb = 256*1024;
1955
1956         return 0;
1957
1958 out_destroy_commit_cache:
1959         kmem_cache_destroy(nfs_cdata_cachep);
1960 out_destroy_write_mempool:
1961         mempool_destroy(nfs_wdata_mempool);
1962 out_destroy_write_cache:
1963         kmem_cache_destroy(nfs_wdata_cachep);
1964         return -ENOMEM;
1965 }
1966
1967 void nfs_destroy_writepagecache(void)
1968 {
1969         mempool_destroy(nfs_commit_mempool);
1970         kmem_cache_destroy(nfs_cdata_cachep);
1971         mempool_destroy(nfs_wdata_mempool);
1972         kmem_cache_destroy(nfs_wdata_cachep);
1973 }
1974