fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #include "nfstrace.h"
  35
  36 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  37
  38 #define MIN_POOL_WRITE          (32)
  39 #define MIN_POOL_COMMIT         (4)
  40
  41 /*
  42  * Local function declarations
  43  */
  44 static void nfs_redirty_request(struct nfs_page *req);
  45 static const struct rpc_call_ops nfs_commit_ops;
  46 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  47 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  48 static const struct nfs_rw_ops nfs_rw_write_ops;
  49
  50 static struct kmem_cache *nfs_wdata_cachep;
  51 static mempool_t *nfs_wdata_mempool;
  52 static struct kmem_cache *nfs_cdata_cachep;
  53 static mempool_t *nfs_commit_mempool;
  54
  55 struct nfs_commit_data *nfs_commitdata_alloc(void)
  56 {
  57         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  58
  59         if (p) {
  60                 memset(p, 0, sizeof(*p));
  61                 INIT_LIST_HEAD(&p->pages);
  62         }
  63         return p;
  64 }
  65 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  66
  67 void nfs_commit_free(struct nfs_commit_data *p)
  68 {
  69         mempool_free(p, nfs_commit_mempool);
  70 }
  71 EXPORT_SYMBOL_GPL(nfs_commit_free);
  72
  73 static struct nfs_rw_header *nfs_writehdr_alloc(void)
  74 {
  75         struct nfs_rw_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  76
  77         if (p)
  78                 memset(p, 0, sizeof(*p));
  79         return p;
  80 }
  81
  82 static void nfs_writehdr_free(struct nfs_rw_header *whdr)
  83 {
  84         mempool_free(whdr, nfs_wdata_mempool);
  85 }
  86
  87 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
  88 {
  89         ctx->error = error;
  90         smp_wmb();
  91         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
  92 }
  93
  94 static struct nfs_page *
  95 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
  96 {
  97         struct nfs_page *req = NULL;
  98
  99         if (PagePrivate(page))
 100                 req = (struct nfs_page *)page_private(page);
 101         else if (unlikely(PageSwapCache(page))) {
 102                 struct nfs_page *freq, *t;
 103
 104                 /* Linearly search the commit list for the correct req */
 105                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 106                         if (freq->wb_page == page) {
 107                                 req = freq;
 108                                 break;
 109                         }
 110                 }
 111         }
 112
 113         if (req)
 114                 kref_get(&req->wb_kref);
 115
 116         return req;
 117 }
 118
 119 static struct nfs_page *nfs_page_find_request(struct page *page)
 120 {
 121         struct inode *inode = page_file_mapping(page)->host;
 122         struct nfs_page *req = NULL;
 123
 124         spin_lock(&inode->i_lock);
 125         req = nfs_page_find_request_locked(NFS_I(inode), page);
 126         spin_unlock(&inode->i_lock);
 127         return req;
 128 }
 129
 130 /* Adjust the file length if we're writing beyond the end */
 131 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 132 {
 133         struct inode *inode = page_file_mapping(page)->host;
 134         loff_t end, i_size;
 135         pgoff_t end_index;
 136
 137         spin_lock(&inode->i_lock);
 138         i_size = i_size_read(inode);
 139         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 140         if (i_size > 0 && page_file_index(page) < end_index)
 141                 goto out;
 142         end = page_file_offset(page) + ((loff_t)offset+count);
 143         if (i_size >= end)
 144                 goto out;
 145         i_size_write(inode, end);
 146         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 147 out:
 148         spin_unlock(&inode->i_lock);
 149 }
 150
 151 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 152 static void nfs_set_pageerror(struct page *page)
 153 {
 154         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 155 }
 156
 157 /* We can set the PG_uptodate flag if we see that a write request
 158  * covers the full page.
 159  */
 160 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 161 {
 162         if (PageUptodate(page))
 163                 return;
 164         if (base != 0)
 165                 return;
 166         if (count != nfs_page_length(page))
 167                 return;
 168         SetPageUptodate(page);
 169 }
 170
 171 static int wb_priority(struct writeback_control *wbc)
 172 {
 173         if (wbc->for_reclaim)
 174                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 175         if (wbc->for_kupdate || wbc->for_background)
 176                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 177         return FLUSH_COND_STABLE;
 178 }
 179
 180 /*
 181  * NFS congestion control
 182  */
 183
 184 int nfs_congestion_kb;
 185
 186 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 187 #define NFS_CONGESTION_OFF_THRESH       \
 188         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 189
 190 static void nfs_set_page_writeback(struct page *page)
 191 {
 192         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 193         int ret = test_set_page_writeback(page);
 194
 195         WARN_ON_ONCE(ret != 0);
 196
 197         if (atomic_long_inc_return(&nfss->writeback) >
 198                         NFS_CONGESTION_ON_THRESH) {
 199                 set_bdi_congested(&nfss->backing_dev_info,
 200                                         BLK_RW_ASYNC);
 201         }
 202 }
 203
 204 static void nfs_end_page_writeback(struct page *page)
 205 {
 206         struct inode *inode = page_file_mapping(page)->host;
 207         struct nfs_server *nfss = NFS_SERVER(inode);
 208
 209         end_page_writeback(page);
 210         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 211                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 212 }
 213
 214 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 215 {
 216         struct inode *inode = page_file_mapping(page)->host;
 217         struct nfs_page *req;
 218         int ret;
 219
 220         spin_lock(&inode->i_lock);
 221         for (;;) {
 222                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 223                 if (req == NULL)
 224                         break;
 225                 if (nfs_lock_request(req))
 226                         break;
 227                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 228                  *       then the call to nfs_lock_request() will always
 229                  *       succeed provided that someone hasn't already marked the
 230                  *       request as dirty (in which case we don't care).
 231                  */
 232                 spin_unlock(&inode->i_lock);
 233                 if (!nonblock)
 234                         ret = nfs_wait_on_request(req);
 235                 else
 236                         ret = -EAGAIN;
 237                 nfs_release_request(req);
 238                 if (ret != 0)
 239                         return ERR_PTR(ret);
 240                 spin_lock(&inode->i_lock);
 241         }
 242         spin_unlock(&inode->i_lock);
 243         return req;
 244 }
 245
 246 /*
 247  * Find an associated nfs write request, and prepare to flush it out
 248  * May return an error if the user signalled nfs_wait_on_request().
 249  */
 250 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 251                                 struct page *page, bool nonblock)
 252 {
 253         struct nfs_page *req;
 254         int ret = 0;
 255
 256         req = nfs_find_and_lock_request(page, nonblock);
 257         if (!req)
 258                 goto out;
 259         ret = PTR_ERR(req);
 260         if (IS_ERR(req))
 261                 goto out;
 262
 263         nfs_set_page_writeback(page);
 264         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 265
 266         ret = 0;
 267         if (!nfs_pageio_add_request(pgio, req)) {
 268                 nfs_redirty_request(req);
 269                 ret = pgio->pg_error;
 270         }
 271 out:
 272         return ret;
 273 }
 274
 275 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 276 {
 277         struct inode *inode = page_file_mapping(page)->host;
 278         int ret;
 279
 280         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 281         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 282
 283         nfs_pageio_cond_complete(pgio, page_file_index(page));
 284         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 285         if (ret == -EAGAIN) {
 286                 redirty_page_for_writepage(wbc, page);
 287                 ret = 0;
 288         }
 289         return ret;
 290 }
 291
 292 /*
 293  * Write an mmapped page to the server.
 294  */
 295 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 296 {
 297         struct nfs_pageio_descriptor pgio;
 298         int err;
 299
 300         nfs_pageio_init_write(&pgio, page->mapping->host, wb_priority(wbc),
 301                                 false, &nfs_async_write_completion_ops);
 302         err = nfs_do_writepage(page, wbc, &pgio);
 303         nfs_pageio_complete(&pgio);
 304         if (err < 0)
 305                 return err;
 306         if (pgio.pg_error < 0)
 307                 return pgio.pg_error;
 308         return 0;
 309 }
 310
 311 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 312 {
 313         int ret;
 314
 315         ret = nfs_writepage_locked(page, wbc);
 316         unlock_page(page);
 317         return ret;
 318 }
 319
 320 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 321 {
 322         int ret;
 323
 324         ret = nfs_do_writepage(page, wbc, data);
 325         unlock_page(page);
 326         return ret;
 327 }
 328
 329 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 330 {
 331         struct inode *inode = mapping->host;
 332         unsigned long *bitlock = &NFS_I(inode)->flags;
 333         struct nfs_pageio_descriptor pgio;
 334         int err;
 335
 336         /* Stop dirtying of new pages while we sync */
 337         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 338                         nfs_wait_bit_killable, TASK_KILLABLE);
 339         if (err)
 340                 goto out_err;
 341
 342         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 343
 344         nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
 345                                 &nfs_async_write_completion_ops);
 346         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 347         nfs_pageio_complete(&pgio);
 348
 349         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 350         smp_mb__after_clear_bit();
 351         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 352
 353         if (err < 0)
 354                 goto out_err;
 355         err = pgio.pg_error;
 356         if (err < 0)
 357                 goto out_err;
 358         return 0;
 359 out_err:
 360         return err;
 361 }
 362
 363 /*
 364  * Insert a write request into an inode
 365  */
 366 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 367 {
 368         struct nfs_inode *nfsi = NFS_I(inode);
 369
 370         /* Lock the request! */
 371         nfs_lock_request(req);
 372
 373         spin_lock(&inode->i_lock);
 374         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 375                 inode->i_version++;
 376         /*
 377          * Swap-space should not get truncated. Hence no need to plug the race
 378          * with invalidate/truncate.
 379          */
 380         if (likely(!PageSwapCache(req->wb_page))) {
 381                 set_bit(PG_MAPPED, &req->wb_flags);
 382                 SetPagePrivate(req->wb_page);
 383                 set_page_private(req->wb_page, (unsigned long)req);
 384         }
 385         nfsi->npages++;
 386         kref_get(&req->wb_kref);
 387         spin_unlock(&inode->i_lock);
 388 }
 389
 390 /*
 391  * Remove a write request from an inode
 392  */
 393 static void nfs_inode_remove_request(struct nfs_page *req)
 394 {
 395         struct inode *inode = req->wb_context->dentry->d_inode;
 396         struct nfs_inode *nfsi = NFS_I(inode);
 397
 398         spin_lock(&inode->i_lock);
 399         if (likely(!PageSwapCache(req->wb_page))) {
 400                 set_page_private(req->wb_page, 0);
 401                 ClearPagePrivate(req->wb_page);
 402                 clear_bit(PG_MAPPED, &req->wb_flags);
 403         }
 404         nfsi->npages--;
 405         spin_unlock(&inode->i_lock);
 406         nfs_release_request(req);
 407 }
 408
 409 static void
 410 nfs_mark_request_dirty(struct nfs_page *req)
 411 {
 412         __set_page_dirty_nobuffers(req->wb_page);
 413 }
 414
 415 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 416 /**
 417  * nfs_request_add_commit_list - add request to a commit list
 418  * @req: pointer to a struct nfs_page
 419  * @dst: commit list head
 420  * @cinfo: holds list lock and accounting info
 421  *
 422  * This sets the PG_CLEAN bit, updates the cinfo count of
 423  * number of outstanding requests requiring a commit as well as
 424  * the MM page stats.
 425  *
 426  * The caller must _not_ hold the cinfo->lock, but must be
 427  * holding the nfs_page lock.
 428  */
 429 void
 430 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 431                             struct nfs_commit_info *cinfo)
 432 {
 433         set_bit(PG_CLEAN, &(req)->wb_flags);
 434         spin_lock(cinfo->lock);
 435         nfs_list_add_request(req, dst);
 436         cinfo->mds->ncommit++;
 437         spin_unlock(cinfo->lock);
 438         if (!cinfo->dreq) {
 439                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 440                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 441                              BDI_RECLAIMABLE);
 442                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 443                                    I_DIRTY_DATASYNC);
 444         }
 445 }
 446 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 447
 448 /**
 449  * nfs_request_remove_commit_list - Remove request from a commit list
 450  * @req: pointer to a nfs_page
 451  * @cinfo: holds list lock and accounting info
 452  *
 453  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 454  * number of outstanding requests requiring a commit
 455  * It does not update the MM page stats.
 456  *
 457  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 458  */
 459 void
 460 nfs_request_remove_commit_list(struct nfs_page *req,
 461                                struct nfs_commit_info *cinfo)
 462 {
 463         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 464                 return;
 465         nfs_list_remove_request(req);
 466         cinfo->mds->ncommit--;
 467 }
 468 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 469
 470 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 471                                       struct inode *inode)
 472 {
 473         cinfo->lock = &inode->i_lock;
 474         cinfo->mds = &NFS_I(inode)->commit_info;
 475         cinfo->ds = pnfs_get_ds_info(inode);
 476         cinfo->dreq = NULL;
 477         cinfo->completion_ops = &nfs_commit_completion_ops;
 478 }
 479
 480 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 481                     struct inode *inode,
 482                     struct nfs_direct_req *dreq)
 483 {
 484         if (dreq)
 485                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 486         else
 487                 nfs_init_cinfo_from_inode(cinfo, inode);
 488 }
 489 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 490
 491 /*
 492  * Add a request to the inode's commit list.
 493  */
 494 void
 495 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 496                         struct nfs_commit_info *cinfo)
 497 {
 498         if (pnfs_mark_request_commit(req, lseg, cinfo))
 499                 return;
 500         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 501 }
 502
 503 static void
 504 nfs_clear_page_commit(struct page *page)
 505 {
 506         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 507         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 508 }
 509
 510 static void
 511 nfs_clear_request_commit(struct nfs_page *req)
 512 {
 513         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 514                 struct inode *inode = req->wb_context->dentry->d_inode;
 515                 struct nfs_commit_info cinfo;
 516
 517                 nfs_init_cinfo_from_inode(&cinfo, inode);
 518                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 519                         spin_lock(cinfo.lock);
 520                         nfs_request_remove_commit_list(req, &cinfo);
 521                         spin_unlock(cinfo.lock);
 522                 }
 523                 nfs_clear_page_commit(req->wb_page);
 524         }
 525 }
 526
 527 static inline
 528 int nfs_write_need_commit(struct nfs_pgio_data *data)
 529 {
 530         if (data->verf.committed == NFS_DATA_SYNC)
 531                 return data->header->lseg == NULL;
 532         return data->verf.committed != NFS_FILE_SYNC;
 533 }
 534
 535 #else
 536 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 537                                       struct inode *inode)
 538 {
 539 }
 540
 541 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 542                     struct inode *inode,
 543                     struct nfs_direct_req *dreq)
 544 {
 545 }
 546
 547 void
 548 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 549                         struct nfs_commit_info *cinfo)
 550 {
 551 }
 552
 553 static void
 554 nfs_clear_request_commit(struct nfs_page *req)
 555 {
 556 }
 557
 558 static inline
 559 int nfs_write_need_commit(struct nfs_pgio_data *data)
 560 {
 561         return 0;
 562 }
 563
 564 #endif
 565
 566 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 567 {
 568         struct nfs_commit_info cinfo;
 569         unsigned long bytes = 0;
 570
 571         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 572                 goto out;
 573         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 574         while (!list_empty(&hdr->pages)) {
 575                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 576
 577                 bytes += req->wb_bytes;
 578                 nfs_list_remove_request(req);
 579                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 580                     (hdr->good_bytes < bytes)) {
 581                         nfs_set_pageerror(req->wb_page);
 582                         nfs_context_set_write_error(req->wb_context, hdr->error);
 583                         goto remove_req;
 584                 }
 585                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 586                         nfs_mark_request_dirty(req);
 587                         goto next;
 588                 }
 589                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 590                         memcpy(&req->wb_verf, &hdr->verf.verifier, sizeof(req->wb_verf));
 591                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 592                         goto next;
 593                 }
 594 remove_req:
 595                 nfs_inode_remove_request(req);
 596 next:
 597                 nfs_unlock_request(req);
 598                 nfs_end_page_writeback(req->wb_page);
 599                 nfs_release_request(req);
 600         }
 601 out:
 602         hdr->release(hdr);
 603 }
 604
 605 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 606 unsigned long
 607 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 608 {
 609         return cinfo->mds->ncommit;
 610 }
 611
 612 /* cinfo->lock held by caller */
 613 int
 614 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 615                      struct nfs_commit_info *cinfo, int max)
 616 {
 617         struct nfs_page *req, *tmp;
 618         int ret = 0;
 619
 620         list_for_each_entry_safe(req, tmp, src, wb_list) {
 621                 if (!nfs_lock_request(req))
 622                         continue;
 623                 kref_get(&req->wb_kref);
 624                 if (cond_resched_lock(cinfo->lock))
 625                         list_safe_reset_next(req, tmp, wb_list);
 626                 nfs_request_remove_commit_list(req, cinfo);
 627                 nfs_list_add_request(req, dst);
 628                 ret++;
 629                 if ((ret == max) && !cinfo->dreq)
 630                         break;
 631         }
 632         return ret;
 633 }
 634
 635 /*
 636  * nfs_scan_commit - Scan an inode for commit requests
 637  * @inode: NFS inode to scan
 638  * @dst: mds destination list
 639  * @cinfo: mds and ds lists of reqs ready to commit
 640  *
 641  * Moves requests from the inode's 'commit' request list.
 642  * The requests are *not* checked to ensure that they form a contiguous set.
 643  */
 644 int
 645 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 646                 struct nfs_commit_info *cinfo)
 647 {
 648         int ret = 0;
 649
 650         spin_lock(cinfo->lock);
 651         if (cinfo->mds->ncommit > 0) {
 652                 const int max = INT_MAX;
 653
 654                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 655                                            cinfo, max);
 656                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 657         }
 658         spin_unlock(cinfo->lock);
 659         return ret;
 660 }
 661
 662 #else
 663 unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 664 {
 665         return 0;
 666 }
 667
 668 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 669                     struct nfs_commit_info *cinfo)
 670 {
 671         return 0;
 672 }
 673 #endif
 674
 675 /*
 676  * Search for an existing write request, and attempt to update
 677  * it to reflect a new dirty region on a given page.
 678  *
 679  * If the attempt fails, then the existing request is flushed out
 680  * to disk.
 681  */
 682 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 683                 struct page *page,
 684                 unsigned int offset,
 685                 unsigned int bytes)
 686 {
 687         struct nfs_page *req;
 688         unsigned int rqend;
 689         unsigned int end;
 690         int error;
 691
 692         if (!PagePrivate(page))
 693                 return NULL;
 694
 695         end = offset + bytes;
 696         spin_lock(&inode->i_lock);
 697
 698         for (;;) {
 699                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 700                 if (req == NULL)
 701                         goto out_unlock;
 702
 703                 rqend = req->wb_offset + req->wb_bytes;
 704                 /*
 705                  * Tell the caller to flush out the request if
 706                  * the offsets are non-contiguous.
 707                  * Note: nfs_flush_incompatible() will already
 708                  * have flushed out requests having wrong owners.
 709                  */
 710                 if (offset > rqend
 711                     || end < req->wb_offset)
 712                         goto out_flushme;
 713
 714                 if (nfs_lock_request(req))
 715                         break;
 716
 717                 /* The request is locked, so wait and then retry */
 718                 spin_unlock(&inode->i_lock);
 719                 error = nfs_wait_on_request(req);
 720                 nfs_release_request(req);
 721                 if (error != 0)
 722                         goto out_err;
 723                 spin_lock(&inode->i_lock);
 724         }
 725
 726         /* Okay, the request matches. Update the region */
 727         if (offset < req->wb_offset) {
 728                 req->wb_offset = offset;
 729                 req->wb_pgbase = offset;
 730         }
 731         if (end > rqend)
 732                 req->wb_bytes = end - req->wb_offset;
 733         else
 734                 req->wb_bytes = rqend - req->wb_offset;
 735 out_unlock:
 736         spin_unlock(&inode->i_lock);
 737         if (req)
 738                 nfs_clear_request_commit(req);
 739         return req;
 740 out_flushme:
 741         spin_unlock(&inode->i_lock);
 742         nfs_release_request(req);
 743         error = nfs_wb_page(inode, page);
 744 out_err:
 745         return ERR_PTR(error);
 746 }
 747
 748 /*
 749  * Try to update an existing write request, or create one if there is none.
 750  *
 751  * Note: Should always be called with the Page Lock held to prevent races
 752  * if we have to add a new request. Also assumes that the caller has
 753  * already called nfs_flush_incompatible() if necessary.
 754  */
 755 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 756                 struct page *page, unsigned int offset, unsigned int bytes)
 757 {
 758         struct inode *inode = page_file_mapping(page)->host;
 759         struct nfs_page *req;
 760
 761         req = nfs_try_to_update_request(inode, page, offset, bytes);
 762         if (req != NULL)
 763                 goto out;
 764         req = nfs_create_request(ctx, inode, page, offset, bytes);
 765         if (IS_ERR(req))
 766                 goto out;
 767         nfs_inode_add_request(inode, req);
 768 out:
 769         return req;
 770 }
 771
 772 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 773                 unsigned int offset, unsigned int count)
 774 {
 775         struct nfs_page *req;
 776
 777         req = nfs_setup_write_request(ctx, page, offset, count);
 778         if (IS_ERR(req))
 779                 return PTR_ERR(req);
 780         /* Update file length */
 781         nfs_grow_file(page, offset, count);
 782         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 783         nfs_mark_request_dirty(req);
 784         nfs_unlock_and_release_request(req);
 785         return 0;
 786 }
 787
 788 int nfs_flush_incompatible(struct file *file, struct page *page)
 789 {
 790         struct nfs_open_context *ctx = nfs_file_open_context(file);
 791         struct nfs_lock_context *l_ctx;
 792         struct nfs_page *req;
 793         int do_flush, status;
 794         /*
 795          * Look for a request corresponding to this page. If there
 796          * is one, and it belongs to another file, we flush it out
 797          * before we try to copy anything into the page. Do this
 798          * due to the lack of an ACCESS-type call in NFSv2.
 799          * Also do the same if we find a request from an existing
 800          * dropped page.
 801          */
 802         do {
 803                 req = nfs_page_find_request(page);
 804                 if (req == NULL)
 805                         return 0;
 806                 l_ctx = req->wb_lock_context;
 807                 do_flush = req->wb_page != page || req->wb_context != ctx;
 808                 if (l_ctx && ctx->dentry->d_inode->i_flock != NULL) {
 809                         do_flush |= l_ctx->lockowner.l_owner != current->files
 810                                 || l_ctx->lockowner.l_pid != current->tgid;
 811                 }
 812                 nfs_release_request(req);
 813                 if (!do_flush)
 814                         return 0;
 815                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 816         } while (status == 0);
 817         return status;
 818 }
 819
 820 /*
 821  * Avoid buffered writes when a open context credential's key would
 822  * expire soon.
 823  *
 824  * Returns -EACCES if the key will expire within RPC_KEY_EXPIRE_FAIL.
 825  *
 826  * Return 0 and set a credential flag which triggers the inode to flush
 827  * and performs  NFS_FILE_SYNC writes if the key will expired within
 828  * RPC_KEY_EXPIRE_TIMEO.
 829  */
 830 int
 831 nfs_key_timeout_notify(struct file *filp, struct inode *inode)
 832 {
 833         struct nfs_open_context *ctx = nfs_file_open_context(filp);
 834         struct rpc_auth *auth = NFS_SERVER(inode)->client->cl_auth;
 835
 836         return rpcauth_key_timeout_notify(auth, ctx->cred);
 837 }
 838
 839 /*
 840  * Test if the open context credential key is marked to expire soon.
 841  */
 842 bool nfs_ctx_key_to_expire(struct nfs_open_context *ctx)
 843 {
 844         return rpcauth_cred_key_to_expire(ctx->cred);
 845 }
 846
 847 /*
 848  * If the page cache is marked as unsafe or invalid, then we can't rely on
 849  * the PageUptodate() flag. In this case, we will need to turn off
 850  * write optimisations that depend on the page contents being correct.
 851  */
 852 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 853 {
 854         struct nfs_inode *nfsi = NFS_I(inode);
 855
 856         if (nfs_have_delegated_attributes(inode))
 857                 goto out;
 858         if (nfsi->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 859                 return false;
 860         smp_rmb();
 861         if (test_bit(NFS_INO_INVALIDATING, &nfsi->flags))
 862                 return false;
 863 out:
 864         return PageUptodate(page) != 0;
 865 }
 866
 867 /* If we know the page is up to date, and we're not using byte range locks (or
 868  * if we have the whole file locked for writing), it may be more efficient to
 869  * extend the write to cover the entire page in order to avoid fragmentation
 870  * inefficiencies.
 871  *
 872  * If the file is opened for synchronous writes then we can just skip the rest
 873  * of the checks.
 874  */
 875 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 876 {
 877         if (file->f_flags & O_DSYNC)
 878                 return 0;
 879         if (!nfs_write_pageuptodate(page, inode))
 880                 return 0;
 881         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 882                 return 1;
 883         if (inode->i_flock == NULL || (inode->i_flock->fl_start == 0 &&
 884                         inode->i_flock->fl_end == OFFSET_MAX &&
 885                         inode->i_flock->fl_type != F_RDLCK))
 886                 return 1;
 887         return 0;
 888 }
 889
 890 /*
 891  * Update and possibly write a cached page of an NFS file.
 892  *
 893  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 894  * things with a page scheduled for an RPC call (e.g. invalidate it).
 895  */
 896 int nfs_updatepage(struct file *file, struct page *page,
 897                 unsigned int offset, unsigned int count)
 898 {
 899         struct nfs_open_context *ctx = nfs_file_open_context(file);
 900         struct inode    *inode = page_file_mapping(page)->host;
 901         int             status = 0;
 902
 903         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 904
 905         dprintk("NFS:       nfs_updatepage(%pD2 %d@%lld)\n",
 906                 file, count, (long long)(page_file_offset(page) + offset));
 907
 908         if (nfs_can_extend_write(file, page, inode)) {
 909                 count = max(count + offset, nfs_page_length(page));
 910                 offset = 0;
 911         }
 912
 913         status = nfs_writepage_setup(ctx, page, offset, count);
 914         if (status < 0)
 915                 nfs_set_pageerror(page);
 916         else
 917                 __set_page_dirty_nobuffers(page);
 918
 919         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 920                         status, (long long)i_size_read(inode));
 921         return status;
 922 }
 923
 924 static int flush_task_priority(int how)
 925 {
 926         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 927                 case FLUSH_HIGHPRI:
 928                         return RPC_PRIORITY_HIGH;
 929                 case FLUSH_LOWPRI:
 930                         return RPC_PRIORITY_LOW;
 931         }
 932         return RPC_PRIORITY_NORMAL;
 933 }
 934
 935 int nfs_initiate_write(struct rpc_clnt *clnt,
 936                        struct nfs_pgio_data *data,
 937                        const struct rpc_call_ops *call_ops,
 938                        int how, int flags)
 939 {
 940         struct inode *inode = data->header->inode;
 941         int priority = flush_task_priority(how);
 942         struct rpc_task *task;
 943         struct rpc_message msg = {
 944                 .rpc_argp = &data->args,
 945                 .rpc_resp = &data->res,
 946                 .rpc_cred = data->header->cred,
 947         };
 948         struct rpc_task_setup task_setup_data = {
 949                 .rpc_client = clnt,
 950                 .task = &data->task,
 951                 .rpc_message = &msg,
 952                 .callback_ops = call_ops,
 953                 .callback_data = data,
 954                 .workqueue = nfsiod_workqueue,
 955                 .flags = RPC_TASK_ASYNC | flags,
 956                 .priority = priority,
 957         };
 958         int ret = 0;
 959
 960         /* Set up the initial task struct.  */
 961         NFS_PROTO(inode)->write_setup(data, &msg);
 962
 963         dprintk("NFS: %5u initiated write call "
 964                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 965                 data->task.tk_pid,
 966                 inode->i_sb->s_id,
 967                 (unsigned long long)NFS_FILEID(inode),
 968                 data->args.count,
 969                 (unsigned long long)data->args.offset);
 970
 971         nfs4_state_protect_write(NFS_SERVER(inode)->nfs_client,
 972                                  &task_setup_data.rpc_client, &msg, data);
 973
 974         task = rpc_run_task(&task_setup_data);
 975         if (IS_ERR(task)) {
 976                 ret = PTR_ERR(task);
 977                 goto out;
 978         }
 979         if (how & FLUSH_SYNC) {
 980                 ret = rpc_wait_for_completion_task(task);
 981                 if (ret == 0)
 982                         ret = task->tk_status;
 983         }
 984         rpc_put_task(task);
 985 out:
 986         return ret;
 987 }
 988 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 989
 990 static int nfs_do_write(struct nfs_pgio_data *data,
 991                 const struct rpc_call_ops *call_ops,
 992                 int how)
 993 {
 994         struct inode *inode = data->header->inode;
 995
 996         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
 997 }
 998
 999 static int nfs_do_multiple_writes(struct list_head *head,
1000                 const struct rpc_call_ops *call_ops,
1001                 int how)
1002 {
1003         struct nfs_pgio_data *data;
1004         int ret = 0;
1005
1006         while (!list_empty(head)) {
1007                 int ret2;
1008
1009                 data = list_first_entry(head, struct nfs_pgio_data, list);
1010                 list_del_init(&data->list);
1011
1012                 ret2 = nfs_do_write(data, call_ops, how);
1013                  if (ret == 0)
1014                          ret = ret2;
1015         }
1016         return ret;
1017 }
1018
1019 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1020  * call this on each, which will prepare them to be retried on next
1021  * writeback using standard nfs.
1022  */
1023 static void nfs_redirty_request(struct nfs_page *req)
1024 {
1025         nfs_mark_request_dirty(req);
1026         nfs_unlock_request(req);
1027         nfs_end_page_writeback(req->wb_page);
1028         nfs_release_request(req);
1029 }
1030
1031 static void nfs_async_write_error(struct list_head *head)
1032 {
1033         struct nfs_page *req;
1034
1035         while (!list_empty(head)) {
1036                 req = nfs_list_entry(head->next);
1037                 nfs_list_remove_request(req);
1038                 nfs_redirty_request(req);
1039         }
1040 }
1041
1042 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1043         .error_cleanup = nfs_async_write_error,
1044         .completion = nfs_write_completion,
1045 };
1046
1047 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1048                 struct nfs_pgio_header *hdr)
1049 {
1050         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1051         while (!list_empty(&hdr->rpc_list)) {
1052                 struct nfs_pgio_data *data = list_first_entry(&hdr->rpc_list,
1053                                 struct nfs_pgio_data, list);
1054                 list_del(&data->list);
1055                 nfs_pgio_data_release(data);
1056         }
1057         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1058 }
1059
1060 /*
1061  * Generate multiple small requests to write out a single
1062  * contiguous dirty area on one page.
1063  */
1064 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1065                            struct nfs_pgio_header *hdr)
1066 {
1067         struct nfs_page *req = hdr->req;
1068         struct page *page = req->wb_page;
1069         struct nfs_pgio_data *data;
1070         size_t wsize = desc->pg_bsize, nbytes;
1071         unsigned int offset;
1072         int requests = 0;
1073         struct nfs_commit_info cinfo;
1074
1075         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1076
1077         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1078             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1079              desc->pg_count > wsize))
1080                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1081
1082
1083         offset = 0;
1084         nbytes = desc->pg_count;
1085         do {
1086                 size_t len = min(nbytes, wsize);
1087
1088                 data = nfs_pgio_data_alloc(hdr, 1);
1089                 if (!data) {
1090                         nfs_flush_error(desc, hdr);
1091                         return -ENOMEM;
1092                 }
1093                 data->pages.pagevec[0] = page;
1094                 nfs_pgio_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1095                 list_add(&data->list, &hdr->rpc_list);
1096                 requests++;
1097                 nbytes -= len;
1098                 offset += len;
1099         } while (nbytes != 0);
1100         nfs_list_remove_request(req);
1101         nfs_list_add_request(req, &hdr->pages);
1102         desc->pg_rpc_callops = &nfs_pgio_common_ops;
1103         return 0;
1104 }
1105
1106 /*
1107  * Create an RPC task for the given write request and kick it.
1108  * The page must have been locked by the caller.
1109  *
1110  * It may happen that the page we're passed is not marked dirty.
1111  * This is the case if nfs_updatepage detects a conflicting request
1112  * that has been written but not committed.
1113  */
1114 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1115                          struct nfs_pgio_header *hdr)
1116 {
1117         struct nfs_page         *req;
1118         struct page             **pages;
1119         struct nfs_pgio_data    *data;
1120         struct list_head *head = &desc->pg_list;
1121         struct nfs_commit_info cinfo;
1122
1123         data = nfs_pgio_data_alloc(hdr, nfs_page_array_len(desc->pg_base,
1124                                                            desc->pg_count));
1125         if (!data) {
1126                 nfs_flush_error(desc, hdr);
1127                 return -ENOMEM;
1128         }
1129
1130         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1131         pages = data->pages.pagevec;
1132         while (!list_empty(head)) {
1133                 req = nfs_list_entry(head->next);
1134                 nfs_list_remove_request(req);
1135                 nfs_list_add_request(req, &hdr->pages);
1136                 *pages++ = req->wb_page;
1137         }
1138
1139         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1140             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1141                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1142
1143         /* Set up the argument struct */
1144         nfs_pgio_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1145         list_add(&data->list, &hdr->rpc_list);
1146         desc->pg_rpc_callops = &nfs_pgio_common_ops;
1147         return 0;
1148 }
1149
1150 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1151                       struct nfs_pgio_header *hdr)
1152 {
1153         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1154                 return nfs_flush_multi(desc, hdr);
1155         return nfs_flush_one(desc, hdr);
1156 }
1157 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1158
1159 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1160 {
1161         struct nfs_rw_header *whdr;
1162         struct nfs_pgio_header *hdr;
1163         int ret;
1164
1165         whdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1166         if (!whdr) {
1167                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1168                 return -ENOMEM;
1169         }
1170         hdr = &whdr->header;
1171         nfs_pgheader_init(desc, hdr, nfs_rw_header_free);
1172         atomic_inc(&hdr->refcnt);
1173         ret = nfs_generic_flush(desc, hdr);
1174         if (ret == 0)
1175                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1176                                              desc->pg_rpc_callops,
1177                                              desc->pg_ioflags);
1178         if (atomic_dec_and_test(&hdr->refcnt))
1179                 hdr->completion_ops->completion(hdr);
1180         return ret;
1181 }
1182
1183 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1184         .pg_test = nfs_generic_pg_test,
1185         .pg_doio = nfs_generic_pg_writepages,
1186 };
1187
1188 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1189                                struct inode *inode, int ioflags, bool force_mds,
1190                                const struct nfs_pgio_completion_ops *compl_ops)
1191 {
1192         struct nfs_server *server = NFS_SERVER(inode);
1193         const struct nfs_pageio_ops *pg_ops = &nfs_pageio_write_ops;
1194
1195 #ifdef CONFIG_NFS_V4_1
1196         if (server->pnfs_curr_ld && !force_mds)
1197                 pg_ops = server->pnfs_curr_ld->pg_write_ops;
1198 #endif
1199         nfs_pageio_init(pgio, inode, pg_ops, compl_ops, &nfs_rw_write_ops,
1200                         server->wsize, ioflags);
1201 }
1202 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1203
1204 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1205 {
1206         pgio->pg_ops = &nfs_pageio_write_ops;
1207         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1208 }
1209 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1210
1211
1212 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1213 {
1214         struct nfs_commit_data *data = calldata;
1215
1216         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1217 }
1218
1219 static void nfs_writeback_release_common(struct nfs_pgio_data *data)
1220 {
1221         struct nfs_pgio_header *hdr = data->header;
1222         int status = data->task.tk_status;
1223
1224         if ((status >= 0) && nfs_write_need_commit(data)) {
1225                 spin_lock(&hdr->lock);
1226                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1227                         ; /* Do nothing */
1228                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1229                         memcpy(&hdr->verf, &data->verf, sizeof(hdr->verf));
1230                 else if (memcmp(&hdr->verf, &data->verf, sizeof(hdr->verf)))
1231                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1232                 spin_unlock(&hdr->lock);
1233         }
1234 }
1235
1236 /*
1237  * Special version of should_remove_suid() that ignores capabilities.
1238  */
1239 static int nfs_should_remove_suid(const struct inode *inode)
1240 {
1241         umode_t mode = inode->i_mode;
1242         int kill = 0;
1243
1244         /* suid always must be killed */
1245         if (unlikely(mode & S_ISUID))
1246                 kill = ATTR_KILL_SUID;
1247
1248         /*
1249          * sgid without any exec bits is just a mandatory locking mark; leave
1250          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1251          */
1252         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1253                 kill |= ATTR_KILL_SGID;
1254
1255         if (unlikely(kill && S_ISREG(mode)))
1256                 return kill;
1257
1258         return 0;
1259 }
1260
1261 /*
1262  * This function is called when the WRITE call is complete.
1263  */
1264 static int nfs_writeback_done(struct rpc_task *task, struct nfs_pgio_data *data,
1265                               struct inode *inode)
1266 {
1267         int status;
1268
1269         /*
1270          * ->write_done will attempt to use post-op attributes to detect
1271          * conflicting writes by other clients.  A strict interpretation
1272          * of close-to-open would allow us to continue caching even if
1273          * another writer had changed the file, but some applications
1274          * depend on tighter cache coherency when writing.
1275          */
1276         status = NFS_PROTO(inode)->write_done(task, data);
1277         if (status != 0)
1278                 return status;
1279         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, data->res.count);
1280
1281 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1282         if (data->res.verf->committed < data->args.stable && task->tk_status >= 0) {
1283                 /* We tried a write call, but the server did not
1284                  * commit data to stable storage even though we
1285                  * requested it.
1286                  * Note: There is a known bug in Tru64 < 5.0 in which
1287                  *       the server reports NFS_DATA_SYNC, but performs
1288                  *       NFS_FILE_SYNC. We therefore implement this checking
1289                  *       as a dprintk() in order to avoid filling syslog.
1290                  */
1291                 static unsigned long    complain;
1292
1293                 /* Note this will print the MDS for a DS write */
1294                 if (time_before(complain, jiffies)) {
1295                         dprintk("NFS:       faulty NFS server %s:"
1296                                 " (committed = %d) != (stable = %d)\n",
1297                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1298                                 data->res.verf->committed, data->args.stable);
1299                         complain = jiffies + 300 * HZ;
1300                 }
1301         }
1302 #endif
1303
1304         /* Deal with the suid/sgid bit corner case */
1305         if (nfs_should_remove_suid(inode))
1306                 nfs_mark_for_revalidate(inode);
1307         return 0;
1308 }
1309
1310 /*
1311  * This function is called when the WRITE call is complete.
1312  */
1313 static void nfs_writeback_result(struct rpc_task *task, struct nfs_pgio_data *data)
1314 {
1315         struct nfs_pgio_args    *argp = &data->args;
1316         struct nfs_pgio_res     *resp = &data->res;
1317
1318         if (resp->count < argp->count) {
1319                 static unsigned long    complain;
1320
1321                 /* This a short write! */
1322                 nfs_inc_stats(data->header->inode, NFSIOS_SHORTWRITE);
1323
1324                 /* Has the server at least made some progress? */
1325                 if (resp->count == 0) {
1326                         if (time_before(complain, jiffies)) {
1327                                 printk(KERN_WARNING
1328                                        "NFS: Server wrote zero bytes, expected %u.\n",
1329                                        argp->count);
1330                                 complain = jiffies + 300 * HZ;
1331                         }
1332                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1333                         task->tk_status = -EIO;
1334                         return;
1335                 }
1336                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1337                 if (resp->verf->committed != NFS_UNSTABLE) {
1338                         /* Resend from where the server left off */
1339                         data->mds_offset += resp->count;
1340                         argp->offset += resp->count;
1341                         argp->pgbase += resp->count;
1342                         argp->count -= resp->count;
1343                 } else {
1344                         /* Resend as a stable write in order to avoid
1345                          * headaches in the case of a server crash.
1346                          */
1347                         argp->stable = NFS_FILE_SYNC;
1348                 }
1349                 rpc_restart_call_prepare(task);
1350         }
1351 }
1352
1353
1354 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1355 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1356 {
1357         int ret;
1358
1359         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1360                 return 1;
1361         if (!may_wait)
1362                 return 0;
1363         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1364                                 NFS_INO_COMMIT,
1365                                 nfs_wait_bit_killable,
1366                                 TASK_KILLABLE);
1367         return (ret < 0) ? ret : 1;
1368 }
1369
1370 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1371 {
1372         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1373         smp_mb__after_clear_bit();
1374         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1375 }
1376
1377 void nfs_commitdata_release(struct nfs_commit_data *data)
1378 {
1379         put_nfs_open_context(data->context);
1380         nfs_commit_free(data);
1381 }
1382 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1383
1384 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1385                         const struct rpc_call_ops *call_ops,
1386                         int how, int flags)
1387 {
1388         struct rpc_task *task;
1389         int priority = flush_task_priority(how);
1390         struct rpc_message msg = {
1391                 .rpc_argp = &data->args,
1392                 .rpc_resp = &data->res,
1393                 .rpc_cred = data->cred,
1394         };
1395         struct rpc_task_setup task_setup_data = {
1396                 .task = &data->task,
1397                 .rpc_client = clnt,
1398                 .rpc_message = &msg,
1399                 .callback_ops = call_ops,
1400                 .callback_data = data,
1401                 .workqueue = nfsiod_workqueue,
1402                 .flags = RPC_TASK_ASYNC | flags,
1403                 .priority = priority,
1404         };
1405         /* Set up the initial task struct.  */
1406         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1407
1408         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1409
1410         nfs4_state_protect(NFS_SERVER(data->inode)->nfs_client,
1411                 NFS_SP4_MACH_CRED_COMMIT, &task_setup_data.rpc_client, &msg);
1412
1413         task = rpc_run_task(&task_setup_data);
1414         if (IS_ERR(task))
1415                 return PTR_ERR(task);
1416         if (how & FLUSH_SYNC)
1417                 rpc_wait_for_completion_task(task);
1418         rpc_put_task(task);
1419         return 0;
1420 }
1421 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1422
1423 /*
1424  * Set up the argument/result storage required for the RPC call.
1425  */
1426 void nfs_init_commit(struct nfs_commit_data *data,
1427                      struct list_head *head,
1428                      struct pnfs_layout_segment *lseg,
1429                      struct nfs_commit_info *cinfo)
1430 {
1431         struct nfs_page *first = nfs_list_entry(head->next);
1432         struct inode *inode = first->wb_context->dentry->d_inode;
1433
1434         /* Set up the RPC argument and reply structs
1435          * NB: take care not to mess about with data->commit et al. */
1436
1437         list_splice_init(head, &data->pages);
1438
1439         data->inode       = inode;
1440         data->cred        = first->wb_context->cred;
1441         data->lseg        = lseg; /* reference transferred */
1442         data->mds_ops     = &nfs_commit_ops;
1443         data->completion_ops = cinfo->completion_ops;
1444         data->dreq        = cinfo->dreq;
1445
1446         data->args.fh     = NFS_FH(data->inode);
1447         /* Note: we always request a commit of the entire inode */
1448         data->args.offset = 0;
1449         data->args.count  = 0;
1450         data->context     = get_nfs_open_context(first->wb_context);
1451         data->res.fattr   = &data->fattr;
1452         data->res.verf    = &data->verf;
1453         nfs_fattr_init(&data->fattr);
1454 }
1455 EXPORT_SYMBOL_GPL(nfs_init_commit);
1456
1457 void nfs_retry_commit(struct list_head *page_list,
1458                       struct pnfs_layout_segment *lseg,
1459                       struct nfs_commit_info *cinfo)
1460 {
1461         struct nfs_page *req;
1462
1463         while (!list_empty(page_list)) {
1464                 req = nfs_list_entry(page_list->next);
1465                 nfs_list_remove_request(req);
1466                 nfs_mark_request_commit(req, lseg, cinfo);
1467                 if (!cinfo->dreq) {
1468                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1469                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1470                                      BDI_RECLAIMABLE);
1471                 }
1472                 nfs_unlock_and_release_request(req);
1473         }
1474 }
1475 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1476
1477 /*
1478  * Commit dirty pages
1479  */
1480 static int
1481 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1482                 struct nfs_commit_info *cinfo)
1483 {
1484         struct nfs_commit_data  *data;
1485
1486         data = nfs_commitdata_alloc();
1487
1488         if (!data)
1489                 goto out_bad;
1490
1491         /* Set up the argument struct */
1492         nfs_init_commit(data, head, NULL, cinfo);
1493         atomic_inc(&cinfo->mds->rpcs_out);
1494         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1495                                    how, 0);
1496  out_bad:
1497         nfs_retry_commit(head, NULL, cinfo);
1498         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1499         return -ENOMEM;
1500 }
1501
1502 /*
1503  * COMMIT call returned
1504  */
1505 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1506 {
1507         struct nfs_commit_data  *data = calldata;
1508
1509         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1510                                 task->tk_pid, task->tk_status);
1511
1512         /* Call the NFS version-specific code */
1513         NFS_PROTO(data->inode)->commit_done(task, data);
1514 }
1515
1516 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1517 {
1518         struct nfs_page *req;
1519         int status = data->task.tk_status;
1520         struct nfs_commit_info cinfo;
1521
1522         while (!list_empty(&data->pages)) {
1523                 req = nfs_list_entry(data->pages.next);
1524                 nfs_list_remove_request(req);
1525                 nfs_clear_page_commit(req->wb_page);
1526
1527                 dprintk("NFS:       commit (%s/%llu %d@%lld)",
1528                         req->wb_context->dentry->d_sb->s_id,
1529                         (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1530                         req->wb_bytes,
1531                         (long long)req_offset(req));
1532                 if (status < 0) {
1533                         nfs_context_set_write_error(req->wb_context, status);
1534                         nfs_inode_remove_request(req);
1535                         dprintk(", error = %d\n", status);
1536                         goto next;
1537                 }
1538
1539                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1540                  * returned by the server against all stored verfs. */
1541                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1542                         /* We have a match */
1543                         nfs_inode_remove_request(req);
1544                         dprintk(" OK\n");
1545                         goto next;
1546                 }
1547                 /* We have a mismatch. Write the page again */
1548                 dprintk(" mismatch\n");
1549                 nfs_mark_request_dirty(req);
1550                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1551         next:
1552                 nfs_unlock_and_release_request(req);
1553         }
1554         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1555         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1556                 nfs_commit_clear_lock(NFS_I(data->inode));
1557 }
1558
1559 static void nfs_commit_release(void *calldata)
1560 {
1561         struct nfs_commit_data *data = calldata;
1562
1563         data->completion_ops->completion(data);
1564         nfs_commitdata_release(calldata);
1565 }
1566
1567 static const struct rpc_call_ops nfs_commit_ops = {
1568         .rpc_call_prepare = nfs_commit_prepare,
1569         .rpc_call_done = nfs_commit_done,
1570         .rpc_release = nfs_commit_release,
1571 };
1572
1573 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1574         .completion = nfs_commit_release_pages,
1575         .error_cleanup = nfs_commit_clear_lock,
1576 };
1577
1578 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1579                             int how, struct nfs_commit_info *cinfo)
1580 {
1581         int status;
1582
1583         status = pnfs_commit_list(inode, head, how, cinfo);
1584         if (status == PNFS_NOT_ATTEMPTED)
1585                 status = nfs_commit_list(inode, head, how, cinfo);
1586         return status;
1587 }
1588
1589 int nfs_commit_inode(struct inode *inode, int how)
1590 {
1591         LIST_HEAD(head);
1592         struct nfs_commit_info cinfo;
1593         int may_wait = how & FLUSH_SYNC;
1594         int res;
1595
1596         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1597         if (res <= 0)
1598                 goto out_mark_dirty;
1599         nfs_init_cinfo_from_inode(&cinfo, inode);
1600         res = nfs_scan_commit(inode, &head, &cinfo);
1601         if (res) {
1602                 int error;
1603
1604                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1605                 if (error < 0)
1606                         return error;
1607                 if (!may_wait)
1608                         goto out_mark_dirty;
1609                 error = wait_on_bit(&NFS_I(inode)->flags,
1610                                 NFS_INO_COMMIT,
1611                                 nfs_wait_bit_killable,
1612                                 TASK_KILLABLE);
1613                 if (error < 0)
1614                         return error;
1615         } else
1616                 nfs_commit_clear_lock(NFS_I(inode));
1617         return res;
1618         /* Note: If we exit without ensuring that the commit is complete,
1619          * we must mark the inode as dirty. Otherwise, future calls to
1620          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1621          * that the data is on the disk.
1622          */
1623 out_mark_dirty:
1624         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1625         return res;
1626 }
1627
1628 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1629 {
1630         struct nfs_inode *nfsi = NFS_I(inode);
1631         int flags = FLUSH_SYNC;
1632         int ret = 0;
1633
1634         /* no commits means nothing needs to be done */
1635         if (!nfsi->commit_info.ncommit)
1636                 return ret;
1637
1638         if (wbc->sync_mode == WB_SYNC_NONE) {
1639                 /* Don't commit yet if this is a non-blocking flush and there
1640                  * are a lot of outstanding writes for this mapping.
1641                  */
1642                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1643                         goto out_mark_dirty;
1644
1645                 /* don't wait for the COMMIT response */
1646                 flags = 0;
1647         }
1648
1649         ret = nfs_commit_inode(inode, flags);
1650         if (ret >= 0) {
1651                 if (wbc->sync_mode == WB_SYNC_NONE) {
1652                         if (ret < wbc->nr_to_write)
1653                                 wbc->nr_to_write -= ret;
1654                         else
1655                                 wbc->nr_to_write = 0;
1656                 }
1657                 return 0;
1658         }
1659 out_mark_dirty:
1660         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1661         return ret;
1662 }
1663 #else
1664 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1665 {
1666         return 0;
1667 }
1668 #endif
1669
1670 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1671 {
1672         return nfs_commit_unstable_pages(inode, wbc);
1673 }
1674 EXPORT_SYMBOL_GPL(nfs_write_inode);
1675
1676 /*
1677  * flush the inode to disk.
1678  */
1679 int nfs_wb_all(struct inode *inode)
1680 {
1681         struct writeback_control wbc = {
1682                 .sync_mode = WB_SYNC_ALL,
1683                 .nr_to_write = LONG_MAX,
1684                 .range_start = 0,
1685                 .range_end = LLONG_MAX,
1686         };
1687         int ret;
1688
1689         trace_nfs_writeback_inode_enter(inode);
1690
1691         ret = sync_inode(inode, &wbc);
1692
1693         trace_nfs_writeback_inode_exit(inode, ret);
1694         return ret;
1695 }
1696 EXPORT_SYMBOL_GPL(nfs_wb_all);
1697
1698 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1699 {
1700         struct nfs_page *req;
1701         int ret = 0;
1702
1703         for (;;) {
1704                 wait_on_page_writeback(page);
1705                 req = nfs_page_find_request(page);
1706                 if (req == NULL)
1707                         break;
1708                 if (nfs_lock_request(req)) {
1709                         nfs_clear_request_commit(req);
1710                         nfs_inode_remove_request(req);
1711                         /*
1712                          * In case nfs_inode_remove_request has marked the
1713                          * page as being dirty
1714                          */
1715                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1716                         nfs_unlock_and_release_request(req);
1717                         break;
1718                 }
1719                 ret = nfs_wait_on_request(req);
1720                 nfs_release_request(req);
1721                 if (ret < 0)
1722                         break;
1723         }
1724         return ret;
1725 }
1726
1727 /*
1728  * Write back all requests on one page - we do this before reading it.
1729  */
1730 int nfs_wb_page(struct inode *inode, struct page *page)
1731 {
1732         loff_t range_start = page_file_offset(page);
1733         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1734         struct writeback_control wbc = {
1735                 .sync_mode = WB_SYNC_ALL,
1736                 .nr_to_write = 0,
1737                 .range_start = range_start,
1738                 .range_end = range_end,
1739         };
1740         int ret;
1741
1742         trace_nfs_writeback_page_enter(inode);
1743
1744         for (;;) {
1745                 wait_on_page_writeback(page);
1746                 if (clear_page_dirty_for_io(page)) {
1747                         ret = nfs_writepage_locked(page, &wbc);
1748                         if (ret < 0)
1749                                 goto out_error;
1750                         continue;
1751                 }
1752                 ret = 0;
1753                 if (!PagePrivate(page))
1754                         break;
1755                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1756                 if (ret < 0)
1757                         goto out_error;
1758         }
1759 out_error:
1760         trace_nfs_writeback_page_exit(inode, ret);
1761         return ret;
1762 }
1763
1764 #ifdef CONFIG_MIGRATION
1765 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1766                 struct page *page, enum migrate_mode mode)
1767 {
1768         /*
1769          * If PagePrivate is set, then the page is currently associated with
1770          * an in-progress read or write request. Don't try to migrate it.
1771          *
1772          * FIXME: we could do this in principle, but we'll need a way to ensure
1773          *        that we can safely release the inode reference while holding
1774          *        the page lock.
1775          */
1776         if (PagePrivate(page))
1777                 return -EBUSY;
1778
1779         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1780                 return -EBUSY;
1781
1782         return migrate_page(mapping, newpage, page, mode);
1783 }
1784 #endif
1785
1786 int __init nfs_init_writepagecache(void)
1787 {
1788         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1789                                              sizeof(struct nfs_rw_header),
1790                                              0, SLAB_HWCACHE_ALIGN,
1791                                              NULL);
1792         if (nfs_wdata_cachep == NULL)
1793                 return -ENOMEM;
1794
1795         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1796                                                      nfs_wdata_cachep);
1797         if (nfs_wdata_mempool == NULL)
1798                 goto out_destroy_write_cache;
1799
1800         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1801                                              sizeof(struct nfs_commit_data),
1802                                              0, SLAB_HWCACHE_ALIGN,
1803                                              NULL);
1804         if (nfs_cdata_cachep == NULL)
1805                 goto out_destroy_write_mempool;
1806
1807         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1808                                                       nfs_cdata_cachep);
1809         if (nfs_commit_mempool == NULL)
1810                 goto out_destroy_commit_cache;
1811
1812         /*
1813          * NFS congestion size, scale with available memory.
1814          *
1815          *  64MB:    8192k
1816          * 128MB:   11585k
1817          * 256MB:   16384k
1818          * 512MB:   23170k
1819          *   1GB:   32768k
1820          *   2GB:   46340k
1821          *   4GB:   65536k
1822          *   8GB:   92681k
1823          *  16GB:  131072k
1824          *
1825          * This allows larger machines to have larger/more transfers.
1826          * Limit the default to 256M
1827          */
1828         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1829         if (nfs_congestion_kb > 256*1024)
1830                 nfs_congestion_kb = 256*1024;
1831
1832         return 0;
1833
1834 out_destroy_commit_cache:
1835         kmem_cache_destroy(nfs_cdata_cachep);
1836 out_destroy_write_mempool:
1837         mempool_destroy(nfs_wdata_mempool);
1838 out_destroy_write_cache:
1839         kmem_cache_destroy(nfs_wdata_cachep);
1840         return -ENOMEM;
1841 }
1842
1843 void nfs_destroy_writepagecache(void)
1844 {
1845         mempool_destroy(nfs_commit_mempool);
1846         kmem_cache_destroy(nfs_cdata_cachep);
1847         mempool_destroy(nfs_wdata_mempool);
1848         kmem_cache_destroy(nfs_wdata_cachep);
1849 }
1850
1851 static const struct nfs_rw_ops nfs_rw_write_ops = {
1852         .rw_mode                = FMODE_WRITE,
1853         .rw_alloc_header        = nfs_writehdr_alloc,
1854         .rw_free_header         = nfs_writehdr_free,
1855         .rw_release             = nfs_writeback_release_common,
1856         .rw_done                = nfs_writeback_done,
1857         .rw_result              = nfs_writeback_result,
1858 };