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 static const struct nfs_rw_ops nfs_rw_write_ops;
  50
  51 static struct kmem_cache *nfs_wdata_cachep;
  52 static mempool_t *nfs_wdata_mempool;
  53 static struct kmem_cache *nfs_cdata_cachep;
  54 static mempool_t *nfs_commit_mempool;
  55
  56 struct nfs_commit_data *nfs_commitdata_alloc(void)
  57 {
  58         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  59
  60         if (p) {
  61                 memset(p, 0, sizeof(*p));
  62                 INIT_LIST_HEAD(&p->pages);
  63         }
  64         return p;
  65 }
  66 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  67
  68 void nfs_commit_free(struct nfs_commit_data *p)
  69 {
  70         mempool_free(p, nfs_commit_mempool);
  71 }
  72 EXPORT_SYMBOL_GPL(nfs_commit_free);
  73
  74 static struct nfs_rw_header *nfs_writehdr_alloc(void)
  75 {
  76         struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  77
  78         if (p)
  79                 memset(p, 0, sizeof(*p));
  80         return p;
  81 }
  82
  83 static void nfs_writehdr_free(struct nfs_rw_header *whdr)
  84 {
  85         mempool_free(whdr, nfs_wdata_mempool);
  86 }
  87
  88 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  89 {
  90         ctx->error = error;
  91         smp_wmb();
  92         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
  93 }
  94
  95 static struct nfs_page *
  96 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
  97 {
  98         struct nfs_page *req = NULL;
  99
 100         if (PagePrivate(page))
 101                 req = (struct nfs_page *)page_private(page);
 102         else if (unlikely(PageSwapCache(page))) {
 103                 struct nfs_page *freq, *t;
 104
 105                 /* Linearly search the commit list for the correct req */
 106                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 107                         if (freq->wb_page == page) {
 108                                 req = freq;
 109                                 break;
 110                         }
 111                 }
 112         }
 113
 114         if (req)
 115                 kref_get(&req->wb_kref);
 116
 117         return req;
 118 }
 119
 120 static struct nfs_page *nfs_page_find_request(struct page *page)
 121 {
 122         struct inode *inode = page_file_mapping(page)->host;
 123         struct nfs_page *req = NULL;
 124
 125         spin_lock(&inode->i_lock);
 126         req = nfs_page_find_request_locked(NFS_I(inode), page);
 127         spin_unlock(&inode->i_lock);
 128         return req;
 129 }
 130
 131 /* Adjust the file length if we're writing beyond the end */
 132 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 133 {
 134         struct inode *inode = page_file_mapping(page)->host;
 135         loff_t end, i_size;
 136         pgoff_t end_index;
 137
 138         spin_lock(&inode->i_lock);
 139         i_size = i_size_read(inode);
 140         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 141         if (i_size > 0 && page_file_index(page) < end_index)
 142                 goto out;
 143         end = page_file_offset(page) + ((loff_t)offset+count);
 144         if (i_size >= end)
 145                 goto out;
 146         i_size_write(inode, end);
 147         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 148 out:
 149         spin_unlock(&inode->i_lock);
 150 }
 151
 152 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 153 static void nfs_set_pageerror(struct page *page)
 154 {
 155         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 156 }
 157
 158 /* We can set the PG_uptodate flag if we see that a write request
 159  * covers the full page.
 160  */
 161 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 162 {
 163         if (PageUptodate(page))
 164                 return;
 165         if (base != 0)
 166                 return;
 167         if (count != nfs_page_length(page))
 168                 return;
 169         SetPageUptodate(page);
 170 }
 171
 172 static int wb_priority(struct writeback_control *wbc)
 173 {
 174         if (wbc->for_reclaim)
 175                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 176         if (wbc->for_kupdate || wbc->for_background)
 177                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 178         return FLUSH_COND_STABLE;
 179 }
 180
 181 /*
 182  * NFS congestion control
 183  */
 184
 185 int nfs_congestion_kb;
 186
 187 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 188 #define NFS_CONGESTION_OFF_THRESH       \
 189         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 190
 191 static void nfs_set_page_writeback(struct page *page)
 192 {
 193         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 194         int ret = test_set_page_writeback(page);
 195
 196         WARN_ON_ONCE(ret != 0);
 197
 198         if (atomic_long_inc_return(&nfss->writeback) >
 199                         NFS_CONGESTION_ON_THRESH) {
 200                 set_bdi_congested(&nfss->backing_dev_info,
 201                                         BLK_RW_ASYNC);
 202         }
 203 }
 204
 205 static void nfs_end_page_writeback(struct page *page)
 206 {
 207         struct inode *inode = page_file_mapping(page)->host;
 208         struct nfs_server *nfss = NFS_SERVER(inode);
 209
 210         end_page_writeback(page);
 211         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 212                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 213 }
 214
 215 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 216 {
 217         struct inode *inode = page_file_mapping(page)->host;
 218         struct nfs_page *req;
 219         int ret;
 220
 221         spin_lock(&inode->i_lock);
 222         for (;;) {
 223                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 224                 if (req == NULL)
 225                         break;
 226                 if (nfs_lock_request(req))
 227                         break;
 228                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 229                  *       then the call to nfs_lock_request() will always
 230                  *       succeed provided that someone hasn't already marked the
 231                  *       request as dirty (in which case we don't care).
 232                  */
 233                 spin_unlock(&inode->i_lock);
 234                 if (!nonblock)
 235                         ret = nfs_wait_on_request(req);
 236                 else
 237                         ret = -EAGAIN;
 238                 nfs_release_request(req);
 239                 if (ret != 0)
 240                         return ERR_PTR(ret);
 241                 spin_lock(&inode->i_lock);
 242         }
 243         spin_unlock(&inode->i_lock);
 244         return req;
 245 }
 246
 247 /*
 248  * Find an associated nfs write request, and prepare to flush it out
 249  * May return an error if the user signalled nfs_wait_on_request().
 250  */
 251 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 252                                 struct page *page, bool nonblock)
 253 {
 254         struct nfs_page *req;
 255         int ret = 0;
 256
 257         req = nfs_find_and_lock_request(page, nonblock);
 258         if (!req)
 259                 goto out;
 260         ret = PTR_ERR(req);
 261         if (IS_ERR(req))
 262                 goto out;
 263
 264         nfs_set_page_writeback(page);
 265         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 266
 267         ret = 0;
 268         if (!nfs_pageio_add_request(pgio, req)) {
 269                 nfs_redirty_request(req);
 270                 ret = pgio->pg_error;
 271         }
 272 out:
 273         return ret;
 274 }
 275
 276 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 277 {
 278         struct inode *inode = page_file_mapping(page)->host;
 279         int ret;
 280
 281         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 282         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 283
 284         nfs_pageio_cond_complete(pgio, page_file_index(page));
 285         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 286         if (ret == -EAGAIN) {
 287                 redirty_page_for_writepage(wbc, page);
 288                 ret = 0;
 289         }
 290         return ret;
 291 }
 292
 293 /*
 294  * Write an mmapped page to the server.
 295  */
 296 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 297 {
 298         struct nfs_pageio_descriptor pgio;
 299         int err;
 300
 301         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 302                                 false, &nfs_async_write_completion_ops);
 303         err = nfs_do_writepage(page, wbc, &pgio);
 304         nfs_pageio_complete(&pgio);
 305         if (err < 0)
 306                 return err;
 307         if (pgio.pg_error < 0)
 308                 return pgio.pg_error;
 309         return 0;
 310 }
 311
 312 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 313 {
 314         int ret;
 315
 316         ret = nfs_writepage_locked(page, wbc);
 317         unlock_page(page);
 318         return ret;
 319 }
 320
 321 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 322 {
 323         int ret;
 324
 325         ret = nfs_do_writepage(page, wbc, data);
 326         unlock_page(page);
 327         return ret;
 328 }
 329
 330 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 331 {
 332         struct inode *inode = mapping->host;
 333         unsigned long *bitlock = &NFS_I(inode)->flags;
 334         struct nfs_pageio_descriptor pgio;
 335         int err;
 336
 337         /* Stop dirtying of new pages while we sync */
 338         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 339                         nfs_wait_bit_killable, TASK_KILLABLE);
 340         if (err)
 341                 goto out_err;
 342
 343         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 344
 345         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 346                                 &nfs_async_write_completion_ops);
 347         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 348         nfs_pageio_complete(&pgio);
 349
 350         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 351         smp_mb__after_clear_bit();
 352         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 353
 354         if (err < 0)
 355                 goto out_err;
 356         err = pgio.pg_error;
 357         if (err < 0)
 358                 goto out_err;
 359         return 0;
 360 out_err:
 361         return err;
 362 }
 363
 364 /*
 365  * Insert a write request into an inode
 366  */
 367 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 368 {
 369         struct nfs_inode *nfsi = NFS_I(inode);
 370
 371         /* Lock the request! */
 372         nfs_lock_request(req);
 373
 374         spin_lock(&inode->i_lock);
 375         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 376                 inode->i_version++;
 377         /*
 378          * Swap-space should not get truncated. Hence no need to plug the race
 379          * with invalidate/truncate.
 380          */
 381         if (likely(!PageSwapCache(req->wb_page))) {
 382                 set_bit(PG_MAPPED, &req->wb_flags);
 383                 SetPagePrivate(req->wb_page);
 384                 set_page_private(req->wb_page, (unsigned long)req);
 385         }
 386         nfsi->npages++;
 387         kref_get(&req->wb_kref);
 388         spin_unlock(&inode->i_lock);
 389 }
 390
 391 /*
 392  * Remove a write request from an inode
 393  */
 394 static void nfs_inode_remove_request(struct nfs_page *req)
 395 {
 396         struct inode *inode = req->wb_context->dentry->d_inode;
 397         struct nfs_inode *nfsi = NFS_I(inode);
 398
 399         spin_lock(&inode->i_lock);
 400         if (likely(!PageSwapCache(req->wb_page))) {
 401                 set_page_private(req->wb_page, 0);
 402                 ClearPagePrivate(req->wb_page);
 403                 clear_bit(PG_MAPPED, &req->wb_flags);
 404         }
 405         nfsi->npages--;
 406         spin_unlock(&inode->i_lock);
 407         nfs_release_request(req);
 408 }
 409
 410 static void
 411 nfs_mark_request_dirty(struct nfs_page *req)
 412 {
 413         __set_page_dirty_nobuffers(req->wb_page);
 414 }
 415
 416 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 417 /**
 418  * nfs_request_add_commit_list - add request to a commit list
 419  * @req: pointer to a struct nfs_page
 420  * @dst: commit list head
 421  * @cinfo: holds list lock and accounting info
 422  *
 423  * This sets the PG_CLEAN bit, updates the cinfo count of
 424  * number of outstanding requests requiring a commit as well as
 425  * the MM page stats.
 426  *
 427  * The caller must _not_ hold the cinfo->lock, but must be
 428  * holding the nfs_page lock.
 429  */
 430 void
 431 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 432                             struct nfs_commit_info *cinfo)
 433 {
 434         set_bit(PG_CLEAN, &(req)->wb_flags);
 435         spin_lock(cinfo->lock);
 436         nfs_list_add_request(req, dst);
 437         cinfo->mds->ncommit++;
 438         spin_unlock(cinfo->lock);
 439         if (!cinfo->dreq) {
 440                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 441                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 442                              BDI_RECLAIMABLE);
 443                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 444                                    I_DIRTY_DATASYNC);
 445         }
 446 }
 447 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 448
 449 /**
 450  * nfs_request_remove_commit_list - Remove request from a commit list
 451  * @req: pointer to a nfs_page
 452  * @cinfo: holds list lock and accounting info
 453  *
 454  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 455  * number of outstanding requests requiring a commit
 456  * It does not update the MM page stats.
 457  *
 458  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 459  */
 460 void
 461 nfs_request_remove_commit_list(struct nfs_page *req,
 462                                struct nfs_commit_info *cinfo)
 463 {
 464         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 465                 return;
 466         nfs_list_remove_request(req);
 467         cinfo->mds->ncommit--;
 468 }
 469 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 470
 471 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 472                                       struct inode *inode)
 473 {
 474         cinfo->lock = &inode->i_lock;
 475         cinfo->mds = &NFS_I(inode)->commit_info;
 476         cinfo->ds = pnfs_get_ds_info(inode);
 477         cinfo->dreq = NULL;
 478         cinfo->completion_ops = &nfs_commit_completion_ops;
 479 }
 480
 481 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 482                     struct inode *inode,
 483                     struct nfs_direct_req *dreq)
 484 {
 485         if (dreq)
 486                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 487         else
 488                 nfs_init_cinfo_from_inode(cinfo, inode);
 489 }
 490 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 491
 492 /*
 493  * Add a request to the inode's commit list.
 494  */
 495 void
 496 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 497                         struct nfs_commit_info *cinfo)
 498 {
 499         if (pnfs_mark_request_commit(req, lseg, cinfo))
 500                 return;
 501         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 502 }
 503
 504 static void
 505 nfs_clear_page_commit(struct page *page)
 506 {
 507         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 508         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 509 }
 510
 511 static void
 512 nfs_clear_request_commit(struct nfs_page *req)
 513 {
 514         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 515                 struct inode *inode = req->wb_context->dentry->d_inode;
 516                 struct nfs_commit_info cinfo;
 517
 518                 nfs_init_cinfo_from_inode(&cinfo, inode);
 519                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 520                         spin_lock(cinfo.lock);
 521                         nfs_request_remove_commit_list(req, &cinfo);
 522                         spin_unlock(cinfo.lock);
 523                 }
 524                 nfs_clear_page_commit(req->wb_page);
 525         }
 526 }
 527
 528 static inline
 529 int nfs_write_need_commit(struct nfs_pgio_data *data)
 530 {
 531         if (data->verf.committed == NFS_DATA_SYNC)
 532                 return data->header->lseg == NULL;
 533         return data->verf.committed != NFS_FILE_SYNC;
 534 }
 535
 536 #else
 537 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 538                                       struct inode *inode)
 539 {
 540 }
 541
 542 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 543                     struct inode *inode,
 544                     struct nfs_direct_req *dreq)
 545 {
 546 }
 547
 548 void
 549 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 550                         struct nfs_commit_info *cinfo)
 551 {
 552 }
 553
 554 static void
 555 nfs_clear_request_commit(struct nfs_page *req)
 556 {
 557 }
 558
 559 static inline
 560 int nfs_write_need_commit(struct nfs_pgio_data *data)
 561 {
 562         return 0;
 563 }
 564
 565 #endif
 566
 567 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 568 {
 569         struct nfs_commit_info cinfo;
 570         unsigned long bytes = 0;
 571
 572         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 573                 goto out;
 574         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 575         while (!list_empty(&hdr->pages)) {
 576                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 577
 578                 bytes += req->wb_bytes;
 579                 nfs_list_remove_request(req);
 580                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 581                     (hdr->good_bytes < bytes)) {
 582                         nfs_set_pageerror(req->wb_page);
 583                         nfs_context_set_write_error(req->wb_context, hdr->error);
 584                         goto remove_req;
 585                 }
 586                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 587                         nfs_mark_request_dirty(req);
 588                         goto next;
 589                 }
 590                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 591                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 592                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 593                         goto next;
 594                 }
 595 remove_req:
 596                 nfs_inode_remove_request(req);
 597 next:
 598                 nfs_unlock_request(req);
 599                 nfs_end_page_writeback(req->wb_page);
 600                 nfs_release_request(req);
 601         }
 602 out:
 603         hdr->release(hdr);
 604 }
 605
 606 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 607 static unsigned long
 608 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 609 {
 610         return cinfo->mds->ncommit;
 611 }
 612
 613 /* cinfo->lock held by caller */
 614 int
 615 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 616                      struct nfs_commit_info *cinfo, int max)
 617 {
 618         struct nfs_page *req, *tmp;
 619         int ret = 0;
 620
 621         list_for_each_entry_safe(req, tmp, src, wb_list) {
 622                 if (!nfs_lock_request(req))
 623                         continue;
 624                 kref_get(&req->wb_kref);
 625                 if (cond_resched_lock(cinfo->lock))
 626                         list_safe_reset_next(req, tmp, wb_list);
 627                 nfs_request_remove_commit_list(req, cinfo);
 628                 nfs_list_add_request(req, dst);
 629                 ret++;
 630                 if ((ret == max) && !cinfo->dreq)
 631                         break;
 632         }
 633         return ret;
 634 }
 635
 636 /*
 637  * nfs_scan_commit - Scan an inode for commit requests
 638  * @inode: NFS inode to scan
 639  * @dst: mds destination list
 640  * @cinfo: mds and ds lists of reqs ready to commit
 641  *
 642  * Moves requests from the inode's 'commit' request list.
 643  * The requests are *not* checked to ensure that they form a contiguous set.
 644  */
 645 int
 646 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 647                 struct nfs_commit_info *cinfo)
 648 {
 649         int ret = 0;
 650
 651         spin_lock(cinfo->lock);
 652         if (cinfo->mds->ncommit > 0) {
 653                 const int max = INT_MAX;
 654
 655                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 656                                            cinfo, max);
 657                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 658         }
 659         spin_unlock(cinfo->lock);
 660         return ret;
 661 }
 662
 663 #else
 664 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 665 {
 666         return 0;
 667 }
 668
 669 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 670                     struct nfs_commit_info *cinfo)
 671 {
 672         return 0;
 673 }
 674 #endif
 675
 676 /*
 677  * Search for an existing write request, and attempt to update
 678  * it to reflect a new dirty region on a given page.
 679  *
 680  * If the attempt fails, then the existing request is flushed out
 681  * to disk.
 682  */
 683 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 684                 struct page *page,
 685                 unsigned int offset,
 686                 unsigned int bytes)
 687 {
 688         struct nfs_page *req;
 689         unsigned int rqend;
 690         unsigned int end;
 691         int error;
 692
 693         if (!PagePrivate(page))
 694                 return NULL;
 695
 696         end = offset + bytes;
 697         spin_lock(&inode->i_lock);
 698
 699         for (;;) {
 700                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 701                 if (req == NULL)
 702                         goto out_unlock;
 703
 704                 rqend = req->wb_offset + req->wb_bytes;
 705                 /*
 706                  * Tell the caller to flush out the request if
 707                  * the offsets are non-contiguous.
 708                  * Note: nfs_flush_incompatible() will already
 709                  * have flushed out requests having wrong owners.
 710                  */
 711                 if (offset > rqend
 712                     || end < req->wb_offset)
 713                         goto out_flushme;
 714
 715                 if (nfs_lock_request(req))
 716                         break;
 717
 718                 /* The request is locked, so wait and then retry */
 719                 spin_unlock(&inode->i_lock);
 720                 error = nfs_wait_on_request(req);
 721                 nfs_release_request(req);
 722                 if (error != 0)
 723                         goto out_err;
 724                 spin_lock(&inode->i_lock);
 725         }
 726
 727         /* Okay, the request matches. Update the region */
 728         if (offset < req->wb_offset) {
 729                 req->wb_offset = offset;
 730                 req->wb_pgbase = offset;
 731         }
 732         if (end > rqend)
 733                 req->wb_bytes = end - req->wb_offset;
 734         else
 735                 req->wb_bytes = rqend - req->wb_offset;
 736 out_unlock:
 737         spin_unlock(&inode->i_lock);
 738         if (req)
 739                 nfs_clear_request_commit(req);
 740         return req;
 741 out_flushme:
 742         spin_unlock(&inode->i_lock);
 743         nfs_release_request(req);
 744         error = nfs_wb_page(inode, page);
 745 out_err:
 746         return ERR_PTR(error);
 747 }
 748
 749 /*
 750  * Try to update an existing write request, or create one if there is none.
 751  *
 752  * Note: Should always be called with the Page Lock held to prevent races
 753  * if we have to add a new request. Also assumes that the caller has
 754  * already called nfs_flush_incompatible() if necessary.
 755  */
 756 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 757                 struct page *page, unsigned int offset, unsigned int bytes)
 758 {
 759         struct inode *inode = page_file_mapping(page)->host;
 760         struct nfs_page *req;
 761
 762         req = nfs_try_to_update_request(inode, page, offset, bytes);
 763         if (req != NULL)
 764                 goto out;
 765         req = nfs_create_request(ctx, inode, page, offset, bytes);
 766         if (IS_ERR(req))
 767                 goto out;
 768         nfs_inode_add_request(inode, req);
 769 out:
 770         return req;
 771 }
 772
 773 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 774                 unsigned int offset, unsigned int count)
 775 {
 776         struct nfs_page *req;
 777
 778         req = nfs_setup_write_request(ctx, page, offset, count);
 779         if (IS_ERR(req))
 780                 return PTR_ERR(req);
 781         /* Update file length */
 782         nfs_grow_file(page, offset, count);
 783         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 784         nfs_mark_request_dirty(req);
 785         nfs_unlock_and_release_request(req);
 786         return 0;
 787 }
 788
 789 int nfs_flush_incompatible(struct file *file, struct page *page)
 790 {
 791         struct nfs_open_context *ctx = nfs_file_open_context(file);
 792         struct nfs_lock_context *l_ctx;
 793         struct nfs_page *req;
 794         int do_flush, status;
 795         /*
 796          * Look for a request corresponding to this page. If there
 797          * is one, and it belongs to another file, we flush it out
 798          * before we try to copy anything into the page. Do this
 799          * due to the lack of an ACCESS-type call in NFSv2.
 800          * Also do the same if we find a request from an existing
 801          * dropped page.
 802          */
 803         do {
 804                 req = nfs_page_find_request(page);
 805                 if (req == NULL)
 806                         return 0;
 807                 l_ctx = req->wb_lock_context;
 808                 do_flush = req->wb_page != page || req->wb_context != ctx;
 809                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 810                         do_flush |= l_ctx->lockowner.l_owner != current->files
 811                                 || l_ctx->lockowner.l_pid != current->tgid;
 812                 }
 813                 nfs_release_request(req);
 814                 if (!do_flush)
 815                         return 0;
 816                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 817         } while (status == 0);
 818         return status;
 819 }
 820
 821 /*
 822  * Avoid buffered writes when a open context credential's key would
 823  * expire soon.
 824  *
 825  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 826  *
 827  * Return 0 and set a credential flag which triggers the inode to flush
 828  * and performs  NFS_FILE_SYNC writes if the key will expired within
 829  * RPC_KEY_EXPIRE_TIMEO.
 830  */
 831 int
 832 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 833 {
 834         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 835         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 836
 837         return rpcauth_key_timeout_notify(auth, ctx->cred);
 838 }
 839
 840 /*
 841  * Test if the open context credential key is marked to expire soon.
 842  */
 843 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 844 {
 845         return rpcauth_cred_key_to_expire(ctx->cred);
 846 }
 847
 848 /*
 849  * If the page cache is marked as unsafe or invalid, then we can't rely on
 850  * the PageUptodate() flag. In this case, we will need to turn off
 851  * write optimisations that depend on the page contents being correct.
 852  */
 853 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 854 {
 855         struct nfs_inode *nfsi = NFS_I(inode);
 856
 857         if (nfs_have_delegated_attributes(inode))
 858                 goto out;
 859         if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 860                 return false;
 861         smp_rmb();
 862         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 863                 return false;
 864 out:
 865         return PageUptodate(page) != 0;
 866 }
 867
 868 /* If we know the page is up to date, and we're not using byte range locks (or
 869  * if we have the whole file locked for writing), it may be more efficient to
 870  * extend the write to cover the entire page in order to avoid fragmentation
 871  * inefficiencies.
 872  *
 873  * If the file is opened for synchronous writes then we can just skip the rest
 874  * of the checks.
 875  */
 876 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 877 {
 878         if (file->f_flags & O_DSYNC)
 879                 return 0;
 880         if (!nfs_write_pageuptodate(page, inode))
 881                 return 0;
 882         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 883                 return 1;
 884         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 885                         inode->i_flock->fl_end == OFFSET_MAX &&
 886                         inode->i_flock->fl_type != F_RDLCK))
 887                 return 1;
 888         return 0;
 889 }
 890
 891 /*
 892  * Update and possibly write a cached page of an NFS file.
 893  *
 894  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 895  * things with a page scheduled for an RPC call (e.g. invalidate it).
 896  */
 897 int nfs_updatepage(struct file *file, struct page *page,
 898                 unsigned int offset, unsigned int count)
 899 {
 900         struct nfs_open_context *ctx = nfs_file_open_context(file);
 901         struct inode    *inode = page_file_mapping(page)->host;
 902         int             status = 0;
 903
 904         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 905
 906         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 907                 file, count, (long long)(page_file_offset(page) + offset));
 908
 909         if (nfs_can_extend_write(file, page, inode)) {
 910                 count = max(count + offset, nfs_page_length(page));
 911                 offset = 0;
 912         }
 913
 914         status = nfs_writepage_setup(ctx, page, offset, count);
 915         if (status < 0)
 916                 nfs_set_pageerror(page);
 917         else
 918                 __set_page_dirty_nobuffers(page);
 919
 920         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 921                         status, (long long)i_size_read(inode));
 922         return status;
 923 }
 924
 925 static int flush_task_priority(int how)
 926 {
 927         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 928                 case FLUSH_HIGHPRI:
 929                         return RPC_PRIORITY_HIGH;
 930                 case FLUSH_LOWPRI:
 931                         return RPC_PRIORITY_LOW;
 932         }
 933         return RPC_PRIORITY_NORMAL;
 934 }
 935
 936 int nfs_initiate_write(struct rpc_clnt *clnt,
 937                        struct nfs_pgio_data *data,
 938                        const struct rpc_call_ops *call_ops,
 939                        int how, int flags)
 940 {
 941         struct inode *inode = data->header->inode;
 942         int priority = flush_task_priority(how);
 943         struct rpc_task *task;
 944         struct rpc_message msg = {
 945                 .rpc_argp = &data->args,
 946                 .rpc_resp = &data->res,
 947                 .rpc_cred = data->header->cred,
 948         };
 949         struct rpc_task_setup task_setup_data = {
 950                 .rpc_client = clnt,
 951                 .task = &data->task,
 952                 .rpc_message = &msg,
 953                 .callback_ops = call_ops,
 954                 .callback_data = data,
 955                 .workqueue = nfsiod_workqueue,
 956                 .flags = RPC_TASK_ASYNC | flags,
 957                 .priority = priority,
 958         };
 959         int ret = 0;
 960
 961         /* Set up the initial task struct.  */
 962         NFS_PROTO(inode)->write_setup(data, &msg);
 963
 964         dprintk("NFS: %5u initiated write call "
 965                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 966                 data->task.tk_pid,
 967                 inode->i_sb->s_id,
 968                 (unsigned long long)NFS_FILEID(inode),
 969                 data->args.count,
 970                 (unsigned long long)data->args.offset);
 971
 972         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
 973                                  &task_setup_data.rpc_client, &msg, data);
 974
 975         task = rpc_run_task(&task_setup_data);
 976         if (IS_ERR(task)) {
 977                 ret = PTR_ERR(task);
 978                 goto out;
 979         }
 980         if (how & FLUSH_SYNC) {
 981                 ret = rpc_wait_for_completion_task(task);
 982                 if (ret == 0)
 983                         ret = task->tk_status;
 984         }
 985         rpc_put_task(task);
 986 out:
 987         return ret;
 988 }
 989 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 990
 991 /*
 992  * Set up the argument/result storage required for the RPC call.
 993  */
 994 static void nfs_write_rpcsetup(struct nfs_pgio_data *data,
 995                 unsigned int count, unsigned int offset,
 996                 int how, struct nfs_commit_info *cinfo)
 997 {
 998         struct nfs_page *req = data->header->req;
 999
1000         /* Set up the RPC argument and reply structs
1001          * NB: take care not to mess about with data->commit et al. */
1002
1003         data->args.fh     = NFS_FH(data->header->inode);
1004         data->args.offset = req_offset(req) + offset;
1005         /* pnfs_set_layoutcommit needs this */
1006         data->mds_offset = data->args.offset;
1007         data->args.pgbase = req->wb_pgbase + offset;
1008         data->args.pages  = data->pages.pagevec;
1009         data->args.count  = count;
1010         data->args.context = get_nfs_open_context(req->wb_context);
1011         data->args.lock_context = req->wb_lock_context;
1012         data->args.stable  = NFS_UNSTABLE;
1013         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1014         case 0:
1015                 break;
1016         case FLUSH_COND_STABLE:
1017                 if (nfs_reqs_to_commit(cinfo))
1018                         break;
1019         default:
1020                 data->args.stable = NFS_FILE_SYNC;
1021         }
1022
1023         data->res.fattr   = &data->fattr;
1024         data->res.count   = count;
1025         data->res.verf    = &data->verf;
1026         nfs_fattr_init(&data->fattr);
1027 }
1028
1029 static int nfs_do_write(struct nfs_pgio_data *data,
1030                 const struct rpc_call_ops *call_ops,
1031                 int how)
1032 {
1033         struct inode *inode = data->header->inode;
1034
1035         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1036 }
1037
1038 static int nfs_do_multiple_writes(struct list_head *head,
1039                 const struct rpc_call_ops *call_ops,
1040                 int how)
1041 {
1042         struct nfs_pgio_data *data;
1043         int ret = 0;
1044
1045         while (!list_empty(head)) {
1046                 int ret2;
1047
1048                 data = list_first_entry(head, struct nfs_pgio_data, list);
1049                 list_del_init(&data->list);
1050
1051                 ret2 = nfs_do_write(data, call_ops, how);
1052                  if (ret == 0)
1053                          ret = ret2;
1054         }
1055         return ret;
1056 }
1057
1058 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1059  * call this on each, which will prepare them to be retried on next
1060  * writeback using standard nfs.
1061  */
1062 static void nfs_redirty_request(struct nfs_page *req)
1063 {
1064         nfs_mark_request_dirty(req);
1065         nfs_unlock_request(req);
1066         nfs_end_page_writeback(req->wb_page);
1067         nfs_release_request(req);
1068 }
1069
1070 static void nfs_async_write_error(struct list_head *head)
1071 {
1072         struct nfs_page *req;
1073
1074         while (!list_empty(head)) {
1075                 req = nfs_list_entry(head->next);
1076                 nfs_list_remove_request(req);
1077                 nfs_redirty_request(req);
1078         }
1079 }
1080
1081 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1082         .error_cleanup = nfs_async_write_error,
1083         .completion = nfs_write_completion,
1084 };
1085
1086 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1087                 struct nfs_pgio_header *hdr)
1088 {
1089         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1090         while (!list_empty(&hdr->rpc_list)) {
1091                 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
1092                                 struct nfs_pgio_data, list);
1093                 list_del(&data->list);
1094                 nfs_pgio_data_release(data);
1095         }
1096         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1097 }
1098
1099 /*
1100  * Generate multiple small requests to write out a single
1101  * contiguous dirty area on one page.
1102  */
1103 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1104                            struct nfs_pgio_header *hdr)
1105 {
1106         struct nfs_page *req = hdr->req;
1107         struct page *page = req->wb_page;
1108         struct nfs_pgio_data *data;
1109         size_t wsize = desc->pg_bsize, nbytes;
1110         unsigned int offset;
1111         int requests = 0;
1112         struct nfs_commit_info cinfo;
1113
1114         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1115
1116         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1117             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1118              desc->pg_count > wsize))
1119                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1120
1121
1122         offset = 0;
1123         nbytes = desc->pg_count;
1124         do {
1125                 size_t len = min(nbytes, wsize);
1126
1127                 data = nfs_pgio_data_alloc(hdr, 1);
1128                 if (!data) {
1129                         nfs_flush_error(desc, hdr);
1130                         return -ENOMEM;
1131                 }
1132                 data->pages.pagevec[0] = page;
1133                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1134                 list_add(&data->list, &hdr->rpc_list);
1135                 requests++;
1136                 nbytes -= len;
1137                 offset += len;
1138         } while (nbytes != 0);
1139         nfs_list_remove_request(req);
1140         nfs_list_add_request(req, &hdr->pages);
1141         desc->pg_rpc_callops = &nfs_write_common_ops;
1142         return 0;
1143 }
1144
1145 /*
1146  * Create an RPC task for the given write request and kick it.
1147  * The page must have been locked by the caller.
1148  *
1149  * It may happen that the page we're passed is not marked dirty.
1150  * This is the case if nfs_updatepage detects a conflicting request
1151  * that has been written but not committed.
1152  */
1153 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1154                          struct nfs_pgio_header *hdr)
1155 {
1156         struct nfs_page         *req;
1157         struct page             **pages;
1158         struct nfs_pgio_data    *data;
1159         struct list_head *head = &desc->pg_list;
1160         struct nfs_commit_info cinfo;
1161
1162         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
1163                                                            desc->pg_count));
1164         if (!data) {
1165                 nfs_flush_error(desc, hdr);
1166                 return -ENOMEM;
1167         }
1168
1169         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1170         pages = data->pages.pagevec;
1171         while (!list_empty(head)) {
1172                 req = nfs_list_entry(head->next);
1173                 nfs_list_remove_request(req);
1174                 nfs_list_add_request(req, &hdr->pages);
1175                 *pages++ = req->wb_page;
1176         }
1177
1178         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1179             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1180                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1181
1182         /* Set up the argument struct */
1183         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1184         list_add(&data->list, &hdr->rpc_list);
1185         desc->pg_rpc_callops = &nfs_write_common_ops;
1186         return 0;
1187 }
1188
1189 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1190                       struct nfs_pgio_header *hdr)
1191 {
1192         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1193                 return nfs_flush_multi(desc, hdr);
1194         return nfs_flush_one(desc, hdr);
1195 }
1196 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1197
1198 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1199 {
1200         struct nfs_rw_header *whdr;
1201         struct nfs_pgio_header *hdr;
1202         int ret;
1203
1204         whdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1205         if (!whdr) {
1206                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1207                 return -ENOMEM;
1208         }
1209         hdr = &whdr->header;
1210         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
1211         atomic_inc(&hdr->refcnt);
1212         ret = nfs_generic_flush(desc, hdr);
1213         if (ret == 0)
1214                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1215                                              desc->pg_rpc_callops,
1216                                              desc->pg_ioflags);
1217         if (atomic_dec_and_test(&hdr->refcnt))
1218                 hdr->completion_ops->completion(hdr);
1219         return ret;
1220 }
1221
1222 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1223         .pg_test = nfs_generic_pg_test,
1224         .pg_doio = nfs_generic_pg_writepages,
1225 };
1226
1227 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1228                                struct inode *inode, int ioflags, bool force_mds,
1229                                const struct nfs_pgio_completion_ops *compl_ops)
1230 {
1231         struct nfs_server *server = NFS_SERVER(inode);
1232         const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1233
1234 #ifdef CONFIG_NFS_V4_1
1235         if (server->pnfs_curr_ld && !force_mds)
1236                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1237 #endif
1238         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1239                         server->wsize, ioflags);
1240 }
1241 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1242
1243 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1244 {
1245         pgio->pg_ops = &nfs_pageio_write_ops;
1246         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1247 }
1248 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1249
1250
1251 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1252 {
1253         struct nfs_commit_data *data = calldata;
1254
1255         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1256 }
1257
1258 static void nfs_writeback_release_common(struct nfs_pgio_data *data)
1259 {
1260         struct nfs_pgio_header *hdr = data->header;
1261         int status = data->task.tk_status;
1262
1263         if ((status >= 0) && nfs_write_need_commit(data)) {
1264                 spin_lock(&hdr->lock);
1265                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1266                         ; /* Do nothing */
1267                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1268                         memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1269                 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1270                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1271                 spin_unlock(&hdr->lock);
1272         }
1273 }
1274
1275 static const struct rpc_call_ops nfs_write_common_ops = {
1276         .rpc_call_prepare = nfs_pgio_prepare,
1277         .rpc_call_done = nfs_pgio_result,
1278         .rpc_release = nfs_pgio_release,
1279 };
1280
1281 /*
1282  * Special version of should_remove_suid() that ignores capabilities.
1283  */
1284 static int nfs_should_remove_suid(const struct inode *inode)
1285 {
1286         umode_t mode = inode->i_mode;
1287         int kill = 0;
1288
1289         /* suid always must be killed */
1290         if (unlikely(mode & S_ISUID))
1291                 kill = ATTR_KILL_SUID;
1292
1293         /*
1294          * sgid without any exec bits is just a mandatory locking mark; leave
1295          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1296          */
1297         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1298                 kill |= ATTR_KILL_SGID;
1299
1300         if (unlikely(kill && S_ISREG(mode)))
1301                 return kill;
1302
1303         return 0;
1304 }
1305
1306 /*
1307  * This function is called when the WRITE call is complete.
1308  */
1309 static int nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data,
1310                               struct inode *inode)
1311 {
1312         int status;
1313
1314         /*
1315          * ->write_done will attempt to use post-op attributes to detect
1316          * conflicting writes by other clients.  A strict interpretation
1317          * of close-to-open would allow us to continue caching even if
1318          * another writer had changed the file, but some applications
1319          * depend on tighter cache coherency when writing.
1320          */
1321         status = NFS_PROTO(inode)->write_done(task, data);
1322         if (status != 0)
1323                 return status;
1324         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, data->res.count);
1325
1326 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1327         if (data->res.verf->committed < data->args.stable && task->tk_status >= 0) {
1328                 /* We tried a write call, but the server did not
1329                  * commit data to stable storage even though we
1330                  * requested it.
1331                  * Note: There is a known bug in Tru64 < 5.0 in which
1332                  *       the server reports NFS_DATA_SYNC, but performs
1333                  *       NFS_FILE_SYNC. We therefore implement this checking
1334                  *       as a dprintk() in order to avoid filling syslog.
1335                  */
1336                 static unsigned long    complain;
1337
1338                 /* Note this will print the MDS for a DS write */
1339                 if (time_before(complain, jiffies)) {
1340                         dprintk("NFS:       faulty NFS server %s:"
1341                                 " (committed = %d) != (stable = %d)\n",
1342                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1343                                 data->res.verf->committed, data->args.stable);
1344                         complain = jiffies + 300 * HZ;
1345                 }
1346         }
1347 #endif
1348
1349         /* Deal with the suid/sgid bit corner case */
1350         if (nfs_should_remove_suid(inode))
1351                 nfs_mark_for_revalidate(inode);
1352         return 0;
1353 }
1354
1355 /*
1356  * This function is called when the WRITE call is complete.
1357  */
1358 static void nfs_writeback_result(struct rpc_task *task, struct nfs_pgio_data *data)
1359 {
1360         struct nfs_pgio_args    *argp = &data->args;
1361         struct nfs_pgio_res     *resp = &data->res;
1362
1363         if (resp->count < argp->count) {
1364                 static unsigned long    complain;
1365
1366                 /* This a short write! */
1367                 nfs_inc_stats(data->header->inode, NFSIOS_SHORTWRITE);
1368
1369                 /* Has the server at least made some progress? */
1370                 if (resp->count == 0) {
1371                         if (time_before(complain, jiffies)) {
1372                                 printk(KERN_WARNING
1373                                        "NFS: Server wrote zero bytes, expected %u.\n",
1374                                        argp->count);
1375                                 complain = jiffies + 300 * HZ;
1376                         }
1377                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1378                         task->tk_status = -EIO;
1379                         return;
1380                 }
1381                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1382                 if (resp->verf->committed != NFS_UNSTABLE) {
1383                         /* Resend from where the server left off */
1384                         data->mds_offset += resp->count;
1385                         argp->offset += resp->count;
1386                         argp->pgbase += resp->count;
1387                         argp->count -= resp->count;
1388                 } else {
1389                         /* Resend as a stable write in order to avoid
1390                          * headaches in the case of a server crash.
1391                          */
1392                         argp->stable = NFS_FILE_SYNC;
1393                 }
1394                 rpc_restart_call_prepare(task);
1395         }
1396 }
1397
1398
1399 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1400 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1401 {
1402         int ret;
1403
1404         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1405                 return 1;
1406         if (!may_wait)
1407                 return 0;
1408         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1409                                 NFS_INO_COMMIT,
1410                                 nfs_wait_bit_killable,
1411                                 TASK_KILLABLE);
1412         return (ret < 0) ? ret : 1;
1413 }
1414
1415 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1416 {
1417         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1418         smp_mb__after_clear_bit();
1419         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1420 }
1421
1422 void nfs_commitdata_release(struct nfs_commit_data *data)
1423 {
1424         put_nfs_open_context(data->context);
1425         nfs_commit_free(data);
1426 }
1427 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1428
1429 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1430                         const struct rpc_call_ops *call_ops,
1431                         int how, int flags)
1432 {
1433         struct rpc_task *task;
1434         int priority = flush_task_priority(how);
1435         struct rpc_message msg = {
1436                 .rpc_argp = &data->args,
1437                 .rpc_resp = &data->res,
1438                 .rpc_cred = data->cred,
1439         };
1440         struct rpc_task_setup task_setup_data = {
1441                 .task = &data->task,
1442                 .rpc_client = clnt,
1443                 .rpc_message = &msg,
1444                 .callback_ops = call_ops,
1445                 .callback_data = data,
1446                 .workqueue = nfsiod_workqueue,
1447                 .flags = RPC_TASK_ASYNC | flags,
1448                 .priority = priority,
1449         };
1450         /* Set up the initial task struct.  */
1451         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1452
1453         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1454
1455         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1456                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1457
1458         task = rpc_run_task(&task_setup_data);
1459         if (IS_ERR(task))
1460                 return PTR_ERR(task);
1461         if (how & FLUSH_SYNC)
1462                 rpc_wait_for_completion_task(task);
1463         rpc_put_task(task);
1464         return 0;
1465 }
1466 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1467
1468 /*
1469  * Set up the argument/result storage required for the RPC call.
1470  */
1471 void nfs_init_commit(struct nfs_commit_data *data,
1472                      struct list_head *head,
1473                      struct pnfs_layout_segment *lseg,
1474                      struct nfs_commit_info *cinfo)
1475 {
1476         struct nfs_page *first = nfs_list_entry(head->next);
1477         struct inode *inode = first->wb_context->dentry->d_inode;
1478
1479         /* Set up the RPC argument and reply structs
1480          * NB: take care not to mess about with data->commit et al. */
1481
1482         list_splice_init(head, &data->pages);
1483
1484         data->inode       = inode;
1485         data->cred        = first->wb_context->cred;
1486         data->lseg        = lseg; /* reference transferred */
1487         data->mds_ops     = &nfs_commit_ops;
1488         data->completion_ops = cinfo->completion_ops;
1489         data->dreq        = cinfo->dreq;
1490
1491         data->args.fh     = NFS_FH(data->inode);
1492         /* Note: we always request a commit of the entire inode */
1493         data->args.offset = 0;
1494         data->args.count  = 0;
1495         data->context     = get_nfs_open_context(first->wb_context);
1496         data->res.fattr   = &data->fattr;
1497         data->res.verf    = &data->verf;
1498         nfs_fattr_init(&data->fattr);
1499 }
1500 EXPORT_SYMBOL_GPL(nfs_init_commit);
1501
1502 void nfs_retry_commit(struct list_head *page_list,
1503                       struct pnfs_layout_segment *lseg,
1504                       struct nfs_commit_info *cinfo)
1505 {
1506         struct nfs_page *req;
1507
1508         while (!list_empty(page_list)) {
1509                 req = nfs_list_entry(page_list->next);
1510                 nfs_list_remove_request(req);
1511                 nfs_mark_request_commit(req, lseg, cinfo);
1512                 if (!cinfo->dreq) {
1513                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1514                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1515                                      BDI_RECLAIMABLE);
1516                 }
1517                 nfs_unlock_and_release_request(req);
1518         }
1519 }
1520 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1521
1522 /*
1523  * Commit dirty pages
1524  */
1525 static int
1526 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1527                 struct nfs_commit_info *cinfo)
1528 {
1529         struct nfs_commit_data  *data;
1530
1531         data = nfs_commitdata_alloc();
1532
1533         if (!data)
1534                 goto out_bad;
1535
1536         /* Set up the argument struct */
1537         nfs_init_commit(data, head, NULL, cinfo);
1538         atomic_inc(&cinfo->mds->rpcs_out);
1539         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1540                                    how, 0);
1541  out_bad:
1542         nfs_retry_commit(head, NULL, cinfo);
1543         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1544         return -ENOMEM;
1545 }
1546
1547 /*
1548  * COMMIT call returned
1549  */
1550 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1551 {
1552         struct nfs_commit_data  *data = calldata;
1553
1554         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1555                                 task->tk_pid, task->tk_status);
1556
1557         /* Call the NFS version-specific code */
1558         NFS_PROTO(data->inode)->commit_done(task, data);
1559 }
1560
1561 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1562 {
1563         struct nfs_page *req;
1564         int status = data->task.tk_status;
1565         struct nfs_commit_info cinfo;
1566
1567         while (!list_empty(&data->pages)) {
1568                 req = nfs_list_entry(data->pages.next);
1569                 nfs_list_remove_request(req);
1570                 nfs_clear_page_commit(req->wb_page);
1571
1572                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1573                         req->wb_context->dentry->d_sb->s_id,
1574                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1575                         req->wb_bytes,
1576                         (long long)req_offset(req));
1577                 if (status < 0) {
1578                         nfs_context_set_write_error(req->wb_context, status);
1579                         nfs_inode_remove_request(req);
1580                         dprintk(", error = %d\n", status);
1581                         goto next;
1582                 }
1583
1584                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1585                  * returned by the server against all stored verfs. */
1586                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1587                         /* We have a match */
1588                         nfs_inode_remove_request(req);
1589                         dprintk(" OK\n");
1590                         goto next;
1591                 }
1592                 /* We have a mismatch. Write the page again */
1593                 dprintk(" mismatch\n");
1594                 nfs_mark_request_dirty(req);
1595                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1596         next:
1597                 nfs_unlock_and_release_request(req);
1598         }
1599         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1600         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1601                 nfs_commit_clear_lock(NFS_I(data->inode));
1602 }
1603
1604 static void nfs_commit_release(void *calldata)
1605 {
1606         struct nfs_commit_data *data = calldata;
1607
1608         data->completion_ops->completion(data);
1609         nfs_commitdata_release(calldata);
1610 }
1611
1612 static const struct rpc_call_ops nfs_commit_ops = {
1613         .rpc_call_prepare = nfs_commit_prepare,
1614         .rpc_call_done = nfs_commit_done,
1615         .rpc_release = nfs_commit_release,
1616 };
1617
1618 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1619         .completion = nfs_commit_release_pages,
1620         .error_cleanup = nfs_commit_clear_lock,
1621 };
1622
1623 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1624                             int how, struct nfs_commit_info *cinfo)
1625 {
1626         int status;
1627
1628         status = pnfs_commit_list(inode, head, how, cinfo);
1629         if (status == PNFS_NOT_ATTEMPTED)
1630                 status = nfs_commit_list(inode, head, how, cinfo);
1631         return status;
1632 }
1633
1634 int nfs_commit_inode(struct inode *inode, int how)
1635 {
1636         LIST_HEAD(head);
1637         struct nfs_commit_info cinfo;
1638         int may_wait = how & FLUSH_SYNC;
1639         int res;
1640
1641         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1642         if (res <= 0)
1643                 goto out_mark_dirty;
1644         nfs_init_cinfo_from_inode(&cinfo, inode);
1645         res = nfs_scan_commit(inode, &head, &cinfo);
1646         if (res) {
1647                 int error;
1648
1649                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1650                 if (error < 0)
1651                         return error;
1652                 if (!may_wait)
1653                         goto out_mark_dirty;
1654                 error = wait_on_bit(&NFS_I(inode)->flags,
1655                                 NFS_INO_COMMIT,
1656                                 nfs_wait_bit_killable,
1657                                 TASK_KILLABLE);
1658                 if (error < 0)
1659                         return error;
1660         } else
1661                 nfs_commit_clear_lock(NFS_I(inode));
1662         return res;
1663         /* Note: If we exit without ensuring that the commit is complete,
1664          * we must mark the inode as dirty. Otherwise, future calls to
1665          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1666          * that the data is on the disk.
1667          */
1668 out_mark_dirty:
1669         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1670         return res;
1671 }
1672
1673 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1674 {
1675         struct nfs_inode *nfsi = NFS_I(inode);
1676         int flags = FLUSH_SYNC;
1677         int ret = 0;
1678
1679         /* no commits means nothing needs to be done */
1680         if (!nfsi->commit_info.ncommit)
1681                 return ret;
1682
1683         if (wbc->sync_mode == WB_SYNC_NONE) {
1684                 /* Don't commit yet if this is a non-blocking flush and there
1685                  * are a lot of outstanding writes for this mapping.
1686                  */
1687                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1688                         goto out_mark_dirty;
1689
1690                 /* don't wait for the COMMIT response */
1691                 flags = 0;
1692         }
1693
1694         ret = nfs_commit_inode(inode, flags);
1695         if (ret >= 0) {
1696                 if (wbc->sync_mode == WB_SYNC_NONE) {
1697                         if (ret < wbc->nr_to_write)
1698                                 wbc->nr_to_write -= ret;
1699                         else
1700                                 wbc->nr_to_write = 0;
1701                 }
1702                 return 0;
1703         }
1704 out_mark_dirty:
1705         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1706         return ret;
1707 }
1708 #else
1709 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1710 {
1711         return 0;
1712 }
1713 #endif
1714
1715 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1716 {
1717         return nfs_commit_unstable_pages(inode, wbc);
1718 }
1719 EXPORT_SYMBOL_GPL(nfs_write_inode);
1720
1721 /*
1722  * flush the inode to disk.
1723  */
1724 int nfs_wb_all(struct inode *inode)
1725 {
1726         struct writeback_control wbc = {
1727                 .sync_mode = WB_SYNC_ALL,
1728                 .nr_to_write = LONG_MAX,
1729                 .range_start = 0,
1730                 .range_end = LLONG_MAX,
1731         };
1732         int ret;
1733
1734         trace_nfs_writeback_inode_enter(inode);
1735
1736         ret = sync_inode(inode, &wbc);
1737
1738         trace_nfs_writeback_inode_exit(inode, ret);
1739         return ret;
1740 }
1741 EXPORT_SYMBOL_GPL(nfs_wb_all);
1742
1743 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1744 {
1745         struct nfs_page *req;
1746         int ret = 0;
1747
1748         for (;;) {
1749                 wait_on_page_writeback(page);
1750                 req = nfs_page_find_request(page);
1751                 if (req == NULL)
1752                         break;
1753                 if (nfs_lock_request(req)) {
1754                         nfs_clear_request_commit(req);
1755                         nfs_inode_remove_request(req);
1756                         /*
1757                          * In case nfs_inode_remove_request has marked the
1758                          * page as being dirty
1759                          */
1760                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1761                         nfs_unlock_and_release_request(req);
1762                         break;
1763                 }
1764                 ret = nfs_wait_on_request(req);
1765                 nfs_release_request(req);
1766                 if (ret < 0)
1767                         break;
1768         }
1769         return ret;
1770 }
1771
1772 /*
1773  * Write back all requests on one page - we do this before reading it.
1774  */
1775 int nfs_wb_page(struct inode *inode, struct page *page)
1776 {
1777         loff_t range_start = page_file_offset(page);
1778         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1779         struct writeback_control wbc = {
1780                 .sync_mode = WB_SYNC_ALL,
1781                 .nr_to_write = 0,
1782                 .range_start = range_start,
1783                 .range_end = range_end,
1784         };
1785         int ret;
1786
1787         trace_nfs_writeback_page_enter(inode);
1788
1789         for (;;) {
1790                 wait_on_page_writeback(page);
1791                 if (clear_page_dirty_for_io(page)) {
1792                         ret = nfs_writepage_locked(page, &wbc);
1793                         if (ret < 0)
1794                                 goto out_error;
1795                         continue;
1796                 }
1797                 ret = 0;
1798                 if (!PagePrivate(page))
1799                         break;
1800                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1801                 if (ret < 0)
1802                         goto out_error;
1803         }
1804 out_error:
1805         trace_nfs_writeback_page_exit(inode, ret);
1806         return ret;
1807 }
1808
1809 #ifdef CONFIG_MIGRATION
1810 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1811                 struct page *page, enum migrate_mode mode)
1812 {
1813         /*
1814          * If PagePrivate is set, then the page is currently associated with
1815          * an in-progress read or write request. Don't try to migrate it.
1816          *
1817          * FIXME: we could do this in principle, but we'll need a way to ensure
1818          *        that we can safely release the inode reference while holding
1819          *        the page lock.
1820          */
1821         if (PagePrivate(page))
1822                 return -EBUSY;
1823
1824         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1825                 return -EBUSY;
1826
1827         return migrate_page(mapping, newpage, page, mode);
1828 }
1829 #endif
1830
1831 int __init nfs_init_writepagecache(void)
1832 {
1833         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1834                                              sizeof(struct nfs_rw_header),
1835                                              0, SLAB_HWCACHE_ALIGN,
1836                                              NULL);
1837         if (nfs_wdata_cachep == NULL)
1838                 return -ENOMEM;
1839
1840         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1841                                                      nfs_wdata_cachep);
1842         if (nfs_wdata_mempool == NULL)
1843                 goto out_destroy_write_cache;
1844
1845         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1846                                              sizeof(struct nfs_commit_data),
1847                                              0, SLAB_HWCACHE_ALIGN,
1848                                              NULL);
1849         if (nfs_cdata_cachep == NULL)
1850                 goto out_destroy_write_mempool;
1851
1852         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1853                                                       nfs_cdata_cachep);
1854         if (nfs_commit_mempool == NULL)
1855                 goto out_destroy_commit_cache;
1856
1857         /*
1858          * NFS congestion size, scale with available memory.
1859          *
1860          *  64MB:    8192k
1861          * 128MB:   11585k
1862          * 256MB:   16384k
1863          * 512MB:   23170k
1864          *   1GB:   32768k
1865          *   2GB:   46340k
1866          *   4GB:   65536k
1867          *   8GB:   92681k
1868          *  16GB:  131072k
1869          *
1870          * This allows larger machines to have larger/more transfers.
1871          * Limit the default to 256M
1872          */
1873         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1874         if (nfs_congestion_kb > 256*1024)
1875                 nfs_congestion_kb = 256*1024;
1876
1877         return 0;
1878
1879 out_destroy_commit_cache:
1880         kmem_cache_destroy(nfs_cdata_cachep);
1881 out_destroy_write_mempool:
1882         mempool_destroy(nfs_wdata_mempool);
1883 out_destroy_write_cache:
1884         kmem_cache_destroy(nfs_wdata_cachep);
1885         return -ENOMEM;
1886 }
1887
1888 void nfs_destroy_writepagecache(void)
1889 {
1890         mempool_destroy(nfs_commit_mempool);
1891         kmem_cache_destroy(nfs_cdata_cachep);
1892         mempool_destroy(nfs_wdata_mempool);
1893         kmem_cache_destroy(nfs_wdata_cachep);
1894 }
1895
1896 static const struct nfs_rw_ops nfs_rw_write_ops = {
1897         .rw_mode                = FMODE_WRITE,
1898         .rw_alloc_header        = nfs_writehdr_alloc,
1899         .rw_free_header         = nfs_writehdr_free,
1900         .rw_release             = nfs_writeback_release_common,
1901         .rw_done                = nfs_writeback_done,
1902         .rw_result              = nfs_writeback_result,
1903 };