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 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  35
  36 #define MIN_POOL_WRITE          (32)
  37 #define MIN_POOL_COMMIT         (4)
  38
  39 /*
  40  * Local function declarations
  41  */
  42 static void nfs_redirty_request(struct nfs_page *req);
  43 static const struct rpc_call_ops nfs_write_common_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  46 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  47
  48 static struct kmem_cache *nfs_wdata_cachep;
  49 static mempool_t *nfs_wdata_mempool;
  50 static struct kmem_cache *nfs_cdata_cachep;
  51 static mempool_t *nfs_commit_mempool;
  52
  53 struct nfs_commit_data *nfs_commitdata_alloc(void)
  54 {
  55         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  56
  57         if (p) {
  58                 memset(p, 0, sizeof(*p));
  59                 INIT_LIST_HEAD(&p->pages);
  60         }
  61         return p;
  62 }
  63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  64
  65 void nfs_commit_free(struct nfs_commit_data *p)
  66 {
  67         mempool_free(p, nfs_commit_mempool);
  68 }
  69 EXPORT_SYMBOL_GPL(nfs_commit_free);
  70
  71 struct nfs_write_header *nfs_writehdr_alloc(void)
  72 {
  73         struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  74
  75         if (p) {
  76                 struct nfs_pgio_header *hdr = &p->header;
  77
  78                 memset(p, 0, sizeof(*p));
  79                 INIT_LIST_HEAD(&hdr->pages);
  80                 INIT_LIST_HEAD(&hdr->rpc_list);
  81                 spin_lock_init(&hdr->lock);
  82                 atomic_set(&hdr->refcnt, 0);
  83                 hdr->verf = &p->verf;
  84         }
  85         return p;
  86 }
  87 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
  88
  89 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
  90                                                   unsigned int pagecount)
  91 {
  92         struct nfs_write_data *data, *prealloc;
  93
  94         prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
  95         if (prealloc->header == NULL)
  96                 data = prealloc;
  97         else
  98                 data = kzalloc(sizeof(*data), GFP_KERNEL);
  99         if (!data)
 100                 goto out;
 101
 102         if (nfs_pgarray_set(&data->pages, pagecount)) {
 103                 data->header = hdr;
 104                 atomic_inc(&hdr->refcnt);
 105         } else {
 106                 if (data != prealloc)
 107                         kfree(data);
 108                 data = NULL;
 109         }
 110 out:
 111         return data;
 112 }
 113
 114 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 115 {
 116         struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
 117         mempool_free(whdr, nfs_wdata_mempool);
 118 }
 119 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
 120
 121 void nfs_writedata_release(struct nfs_write_data *wdata)
 122 {
 123         struct nfs_pgio_header *hdr = wdata->header;
 124         struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
 125
 126         put_nfs_open_context(wdata->args.context);
 127         if (wdata->pages.pagevec != wdata->pages.page_array)
 128                 kfree(wdata->pages.pagevec);
 129         if (wdata == &write_header->rpc_data) {
 130                 wdata->header = NULL;
 131                 wdata = NULL;
 132         }
 133         if (atomic_dec_and_test(&hdr->refcnt))
 134                 hdr->completion_ops->completion(hdr);
 135         /* Note: we only free the rpc_task after callbacks are done.
 136          * See the comment in rpc_free_task() for why
 137          */
 138         kfree(wdata);
 139 }
 140 EXPORT_SYMBOL_GPL(nfs_writedata_release);
 141
 142 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 143 {
 144         ctx->error = error;
 145         smp_wmb();
 146         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 147 }
 148
 149 static struct nfs_page *
 150 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
 151 {
 152         struct nfs_page *req = NULL;
 153
 154         if (PagePrivate(page))
 155                 req = (struct nfs_page *)page_private(page);
 156         else if (unlikely(PageSwapCache(page))) {
 157                 struct nfs_page *freq, *t;
 158
 159                 /* Linearly search the commit list for the correct req */
 160                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 161                         if (freq->wb_page == page) {
 162                                 req = freq;
 163                                 break;
 164                         }
 165                 }
 166         }
 167
 168         if (req)
 169                 kref_get(&req->wb_kref);
 170
 171         return req;
 172 }
 173
 174 static struct nfs_page *nfs_page_find_request(struct page *page)
 175 {
 176         struct inode *inode = page_file_mapping(page)->host;
 177         struct nfs_page *req = NULL;
 178
 179         spin_lock(&inode->i_lock);
 180         req = nfs_page_find_request_locked(NFS_I(inode), page);
 181         spin_unlock(&inode->i_lock);
 182         return req;
 183 }
 184
 185 /* Adjust the file length if we're writing beyond the end */
 186 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 187 {
 188         struct inode *inode = page_file_mapping(page)->host;
 189         loff_t end, i_size;
 190         pgoff_t end_index;
 191
 192         spin_lock(&inode->i_lock);
 193         i_size = i_size_read(inode);
 194         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 195         if (i_size > 0 && page_file_index(page) < end_index)
 196                 goto out;
 197         end = page_file_offset(page) + ((loff_t)offset+count);
 198         if (i_size >= end)
 199                 goto out;
 200         i_size_write(inode, end);
 201         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 202 out:
 203         spin_unlock(&inode->i_lock);
 204 }
 205
 206 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 207 static void nfs_set_pageerror(struct page *page)
 208 {
 209         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 210 }
 211
 212 /* We can set the PG_uptodate flag if we see that a write request
 213  * covers the full page.
 214  */
 215 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 216 {
 217         if (PageUptodate(page))
 218                 return;
 219         if (base != 0)
 220                 return;
 221         if (count != nfs_page_length(page))
 222                 return;
 223         SetPageUptodate(page);
 224 }
 225
 226 static int wb_priority(struct writeback_control *wbc)
 227 {
 228         if (wbc->for_reclaim)
 229                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 230         if (wbc->for_kupdate || wbc->for_background)
 231                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 232         return FLUSH_COND_STABLE;
 233 }
 234
 235 /*
 236  * NFS congestion control
 237  */
 238
 239 int nfs_congestion_kb;
 240
 241 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 242 #define NFS_CONGESTION_OFF_THRESH       \
 243         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 244
 245 static void nfs_set_page_writeback(struct page *page)
 246 {
 247         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 248         int ret = test_set_page_writeback(page);
 249
 250         WARN_ON_ONCE(ret != 0);
 251
 252         if (atomic_long_inc_return(&nfss->writeback) >
 253                         NFS_CONGESTION_ON_THRESH) {
 254                 set_bdi_congested(&nfss->backing_dev_info,
 255                                         BLK_RW_ASYNC);
 256         }
 257 }
 258
 259 static void nfs_end_page_writeback(struct page *page)
 260 {
 261         struct inode *inode = page_file_mapping(page)->host;
 262         struct nfs_server *nfss = NFS_SERVER(inode);
 263
 264         end_page_writeback(page);
 265         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 266                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 267 }
 268
 269 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 270 {
 271         struct inode *inode = page_file_mapping(page)->host;
 272         struct nfs_page *req;
 273         int ret;
 274
 275         spin_lock(&inode->i_lock);
 276         for (;;) {
 277                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 278                 if (req == NULL)
 279                         break;
 280                 if (nfs_lock_request(req))
 281                         break;
 282                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 283                  *       then the call to nfs_lock_request() will always
 284                  *       succeed provided that someone hasn't already marked the
 285                  *       request as dirty (in which case we don't care).
 286                  */
 287                 spin_unlock(&inode->i_lock);
 288                 if (!nonblock)
 289                         ret = nfs_wait_on_request(req);
 290                 else
 291                         ret = -EAGAIN;
 292                 nfs_release_request(req);
 293                 if (ret != 0)
 294                         return ERR_PTR(ret);
 295                 spin_lock(&inode->i_lock);
 296         }
 297         spin_unlock(&inode->i_lock);
 298         return req;
 299 }
 300
 301 /*
 302  * Find an associated nfs write request, and prepare to flush it out
 303  * May return an error if the user signalled nfs_wait_on_request().
 304  */
 305 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 306                                 struct page *page, bool nonblock)
 307 {
 308         struct nfs_page *req;
 309         int ret = 0;
 310
 311         req = nfs_find_and_lock_request(page, nonblock);
 312         if (!req)
 313                 goto out;
 314         ret = PTR_ERR(req);
 315         if (IS_ERR(req))
 316                 goto out;
 317
 318         nfs_set_page_writeback(page);
 319         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 320
 321         ret = 0;
 322         if (!nfs_pageio_add_request(pgio, req)) {
 323                 nfs_redirty_request(req);
 324                 ret = pgio->pg_error;
 325         }
 326 out:
 327         return ret;
 328 }
 329
 330 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 331 {
 332         struct inode *inode = page_file_mapping(page)->host;
 333         int ret;
 334
 335         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 336         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 337
 338         nfs_pageio_cond_complete(pgio, page_file_index(page));
 339         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 340         if (ret == -EAGAIN) {
 341                 redirty_page_for_writepage(wbc, page);
 342                 ret = 0;
 343         }
 344         return ret;
 345 }
 346
 347 /*
 348  * Write an mmapped page to the server.
 349  */
 350 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 351 {
 352         struct nfs_pageio_descriptor pgio;
 353         int err;
 354
 355         NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
 356                                                           page->mapping->host,
 357                                                           wb_priority(wbc),
 358                                                           &nfs_async_write_completion_ops);
 359         err = nfs_do_writepage(page, wbc, &pgio);
 360         nfs_pageio_complete(&pgio);
 361         if (err < 0)
 362                 return err;
 363         if (pgio.pg_error < 0)
 364                 return pgio.pg_error;
 365         return 0;
 366 }
 367
 368 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 369 {
 370         int ret;
 371
 372         ret = nfs_writepage_locked(page, wbc);
 373         unlock_page(page);
 374         return ret;
 375 }
 376
 377 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 378 {
 379         int ret;
 380
 381         ret = nfs_do_writepage(page, wbc, data);
 382         unlock_page(page);
 383         return ret;
 384 }
 385
 386 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 387 {
 388         struct inode *inode = mapping->host;
 389         unsigned long *bitlock = &NFS_I(inode)->flags;
 390         struct nfs_pageio_descriptor pgio;
 391         int err;
 392
 393         /* Stop dirtying of new pages while we sync */
 394         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 395                         nfs_wait_bit_killable, TASK_KILLABLE);
 396         if (err)
 397                 goto out_err;
 398
 399         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 400
 401         NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
 402         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 403         nfs_pageio_complete(&pgio);
 404
 405         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 406         smp_mb__after_clear_bit();
 407         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 408
 409         if (err < 0)
 410                 goto out_err;
 411         err = pgio.pg_error;
 412         if (err < 0)
 413                 goto out_err;
 414         return 0;
 415 out_err:
 416         return err;
 417 }
 418
 419 /*
 420  * Insert a write request into an inode
 421  */
 422 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 423 {
 424         struct nfs_inode *nfsi = NFS_I(inode);
 425
 426         /* Lock the request! */
 427         nfs_lock_request(req);
 428
 429         spin_lock(&inode->i_lock);
 430         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 431                 inode->i_version++;
 432         /*
 433          * Swap-space should not get truncated. Hence no need to plug the race
 434          * with invalidate/truncate.
 435          */
 436         if (likely(!PageSwapCache(req->wb_page))) {
 437                 set_bit(PG_MAPPED, &req->wb_flags);
 438                 SetPagePrivate(req->wb_page);
 439                 set_page_private(req->wb_page, (unsigned long)req);
 440         }
 441         nfsi->npages++;
 442         kref_get(&req->wb_kref);
 443         spin_unlock(&inode->i_lock);
 444 }
 445
 446 /*
 447  * Remove a write request from an inode
 448  */
 449 static void nfs_inode_remove_request(struct nfs_page *req)
 450 {
 451         struct inode *inode = req->wb_context->dentry->d_inode;
 452         struct nfs_inode *nfsi = NFS_I(inode);
 453
 454         spin_lock(&inode->i_lock);
 455         if (likely(!PageSwapCache(req->wb_page))) {
 456                 set_page_private(req->wb_page, 0);
 457                 ClearPagePrivate(req->wb_page);
 458                 clear_bit(PG_MAPPED, &req->wb_flags);
 459         }
 460         nfsi->npages--;
 461         spin_unlock(&inode->i_lock);
 462         nfs_release_request(req);
 463 }
 464
 465 static void
 466 nfs_mark_request_dirty(struct nfs_page *req)
 467 {
 468         __set_page_dirty_nobuffers(req->wb_page);
 469 }
 470
 471 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 472 /**
 473  * nfs_request_add_commit_list - add request to a commit list
 474  * @req: pointer to a struct nfs_page
 475  * @dst: commit list head
 476  * @cinfo: holds list lock and accounting info
 477  *
 478  * This sets the PG_CLEAN bit, updates the cinfo count of
 479  * number of outstanding requests requiring a commit as well as
 480  * the MM page stats.
 481  *
 482  * The caller must _not_ hold the cinfo->lock, but must be
 483  * holding the nfs_page lock.
 484  */
 485 void
 486 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 487                             struct nfs_commit_info *cinfo)
 488 {
 489         set_bit(PG_CLEAN, &(req)->wb_flags);
 490         spin_lock(cinfo->lock);
 491         nfs_list_add_request(req, dst);
 492         cinfo->mds->ncommit++;
 493         spin_unlock(cinfo->lock);
 494         if (!cinfo->dreq) {
 495                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 496                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 497                              BDI_RECLAIMABLE);
 498                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 499                                    I_DIRTY_DATASYNC);
 500         }
 501 }
 502 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 503
 504 /**
 505  * nfs_request_remove_commit_list - Remove request from a commit list
 506  * @req: pointer to a nfs_page
 507  * @cinfo: holds list lock and accounting info
 508  *
 509  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 510  * number of outstanding requests requiring a commit
 511  * It does not update the MM page stats.
 512  *
 513  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 514  */
 515 void
 516 nfs_request_remove_commit_list(struct nfs_page *req,
 517                                struct nfs_commit_info *cinfo)
 518 {
 519         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 520                 return;
 521         nfs_list_remove_request(req);
 522         cinfo->mds->ncommit--;
 523 }
 524 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 525
 526 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 527                                       struct inode *inode)
 528 {
 529         cinfo->lock = &inode->i_lock;
 530         cinfo->mds = &NFS_I(inode)->commit_info;
 531         cinfo->ds = pnfs_get_ds_info(inode);
 532         cinfo->dreq = NULL;
 533         cinfo->completion_ops = &nfs_commit_completion_ops;
 534 }
 535
 536 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 537                     struct inode *inode,
 538                     struct nfs_direct_req *dreq)
 539 {
 540         if (dreq)
 541                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 542         else
 543                 nfs_init_cinfo_from_inode(cinfo, inode);
 544 }
 545 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 546
 547 /*
 548  * Add a request to the inode's commit list.
 549  */
 550 void
 551 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 552                         struct nfs_commit_info *cinfo)
 553 {
 554         if (pnfs_mark_request_commit(req, lseg, cinfo))
 555                 return;
 556         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 557 }
 558
 559 static void
 560 nfs_clear_page_commit(struct page *page)
 561 {
 562         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 563         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 564 }
 565
 566 static void
 567 nfs_clear_request_commit(struct nfs_page *req)
 568 {
 569         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 570                 struct inode *inode = req->wb_context->dentry->d_inode;
 571                 struct nfs_commit_info cinfo;
 572
 573                 nfs_init_cinfo_from_inode(&cinfo, inode);
 574                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 575                         spin_lock(cinfo.lock);
 576                         nfs_request_remove_commit_list(req, &cinfo);
 577                         spin_unlock(cinfo.lock);
 578                 }
 579                 nfs_clear_page_commit(req->wb_page);
 580         }
 581 }
 582
 583 static inline
 584 int nfs_write_need_commit(struct nfs_write_data *data)
 585 {
 586         if (data->verf.committed == NFS_DATA_SYNC)
 587                 return data->header->lseg == NULL;
 588         return data->verf.committed != NFS_FILE_SYNC;
 589 }
 590
 591 #else
 592 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 593                                       struct inode *inode)
 594 {
 595 }
 596
 597 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 598                     struct inode *inode,
 599                     struct nfs_direct_req *dreq)
 600 {
 601 }
 602
 603 void
 604 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 605                         struct nfs_commit_info *cinfo)
 606 {
 607 }
 608
 609 static void
 610 nfs_clear_request_commit(struct nfs_page *req)
 611 {
 612 }
 613
 614 static inline
 615 int nfs_write_need_commit(struct nfs_write_data *data)
 616 {
 617         return 0;
 618 }
 619
 620 #endif
 621
 622 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 623 {
 624         struct nfs_commit_info cinfo;
 625         unsigned long bytes = 0;
 626
 627         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 628                 goto out;
 629         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 630         while (!list_empty(&hdr->pages)) {
 631                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 632
 633                 bytes += req->wb_bytes;
 634                 nfs_list_remove_request(req);
 635                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 636                     (hdr->good_bytes < bytes)) {
 637                         nfs_set_pageerror(req->wb_page);
 638                         nfs_context_set_write_error(req->wb_context, hdr->error);
 639                         goto remove_req;
 640                 }
 641                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 642                         nfs_mark_request_dirty(req);
 643                         goto next;
 644                 }
 645                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 646                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
 647                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 648                         goto next;
 649                 }
 650 remove_req:
 651                 nfs_inode_remove_request(req);
 652 next:
 653                 nfs_unlock_request(req);
 654                 nfs_end_page_writeback(req->wb_page);
 655                 nfs_release_request(req);
 656         }
 657 out:
 658         hdr->release(hdr);
 659 }
 660
 661 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 662 static unsigned long
 663 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 664 {
 665         return cinfo->mds->ncommit;
 666 }
 667
 668 /* cinfo->lock held by caller */
 669 int
 670 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 671                      struct nfs_commit_info *cinfo, int max)
 672 {
 673         struct nfs_page *req, *tmp;
 674         int ret = 0;
 675
 676         list_for_each_entry_safe(req, tmp, src, wb_list) {
 677                 if (!nfs_lock_request(req))
 678                         continue;
 679                 kref_get(&req->wb_kref);
 680                 if (cond_resched_lock(cinfo->lock))
 681                         list_safe_reset_next(req, tmp, wb_list);
 682                 nfs_request_remove_commit_list(req, cinfo);
 683                 nfs_list_add_request(req, dst);
 684                 ret++;
 685                 if ((ret == max) && !cinfo->dreq)
 686                         break;
 687         }
 688         return ret;
 689 }
 690
 691 /*
 692  * nfs_scan_commit - Scan an inode for commit requests
 693  * @inode: NFS inode to scan
 694  * @dst: mds destination list
 695  * @cinfo: mds and ds lists of reqs ready to commit
 696  *
 697  * Moves requests from the inode's 'commit' request list.
 698  * The requests are *not* checked to ensure that they form a contiguous set.
 699  */
 700 int
 701 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 702                 struct nfs_commit_info *cinfo)
 703 {
 704         int ret = 0;
 705
 706         spin_lock(cinfo->lock);
 707         if (cinfo->mds->ncommit > 0) {
 708                 const int max = INT_MAX;
 709
 710                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 711                                            cinfo, max);
 712                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 713         }
 714         spin_unlock(cinfo->lock);
 715         return ret;
 716 }
 717
 718 #else
 719 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 720 {
 721         return 0;
 722 }
 723
 724 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 725                     struct nfs_commit_info *cinfo)
 726 {
 727         return 0;
 728 }
 729 #endif
 730
 731 /*
 732  * Search for an existing write request, and attempt to update
 733  * it to reflect a new dirty region on a given page.
 734  *
 735  * If the attempt fails, then the existing request is flushed out
 736  * to disk.
 737  */
 738 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 739                 struct page *page,
 740                 unsigned int offset,
 741                 unsigned int bytes)
 742 {
 743         struct nfs_page *req;
 744         unsigned int rqend;
 745         unsigned int end;
 746         int error;
 747
 748         if (!PagePrivate(page))
 749                 return NULL;
 750
 751         end = offset + bytes;
 752         spin_lock(&inode->i_lock);
 753
 754         for (;;) {
 755                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 756                 if (req == NULL)
 757                         goto out_unlock;
 758
 759                 rqend = req->wb_offset + req->wb_bytes;
 760                 /*
 761                  * Tell the caller to flush out the request if
 762                  * the offsets are non-contiguous.
 763                  * Note: nfs_flush_incompatible() will already
 764                  * have flushed out requests having wrong owners.
 765                  */
 766                 if (offset > rqend
 767                     || end < req->wb_offset)
 768                         goto out_flushme;
 769
 770                 if (nfs_lock_request(req))
 771                         break;
 772
 773                 /* The request is locked, so wait and then retry */
 774                 spin_unlock(&inode->i_lock);
 775                 error = nfs_wait_on_request(req);
 776                 nfs_release_request(req);
 777                 if (error != 0)
 778                         goto out_err;
 779                 spin_lock(&inode->i_lock);
 780         }
 781
 782         /* Okay, the request matches. Update the region */
 783         if (offset < req->wb_offset) {
 784                 req->wb_offset = offset;
 785                 req->wb_pgbase = offset;
 786         }
 787         if (end > rqend)
 788                 req->wb_bytes = end - req->wb_offset;
 789         else
 790                 req->wb_bytes = rqend - req->wb_offset;
 791 out_unlock:
 792         spin_unlock(&inode->i_lock);
 793         if (req)
 794                 nfs_clear_request_commit(req);
 795         return req;
 796 out_flushme:
 797         spin_unlock(&inode->i_lock);
 798         nfs_release_request(req);
 799         error = nfs_wb_page(inode, page);
 800 out_err:
 801         return ERR_PTR(error);
 802 }
 803
 804 /*
 805  * Try to update an existing write request, or create one if there is none.
 806  *
 807  * Note: Should always be called with the Page Lock held to prevent races
 808  * if we have to add a new request. Also assumes that the caller has
 809  * already called nfs_flush_incompatible() if necessary.
 810  */
 811 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 812                 struct page *page, unsigned int offset, unsigned int bytes)
 813 {
 814         struct inode *inode = page_file_mapping(page)->host;
 815         struct nfs_page *req;
 816
 817         req = nfs_try_to_update_request(inode, page, offset, bytes);
 818         if (req != NULL)
 819                 goto out;
 820         req = nfs_create_request(ctx, inode, page, offset, bytes);
 821         if (IS_ERR(req))
 822                 goto out;
 823         nfs_inode_add_request(inode, req);
 824 out:
 825         return req;
 826 }
 827
 828 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 829                 unsigned int offset, unsigned int count)
 830 {
 831         struct nfs_page *req;
 832
 833         req = nfs_setup_write_request(ctx, page, offset, count);
 834         if (IS_ERR(req))
 835                 return PTR_ERR(req);
 836         /* Update file length */
 837         nfs_grow_file(page, offset, count);
 838         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 839         nfs_mark_request_dirty(req);
 840         nfs_unlock_and_release_request(req);
 841         return 0;
 842 }
 843
 844 int nfs_flush_incompatible(struct file *file, struct page *page)
 845 {
 846         struct nfs_open_context *ctx = nfs_file_open_context(file);
 847         struct nfs_lock_context *l_ctx;
 848         struct nfs_page *req;
 849         int do_flush, status;
 850         /*
 851          * Look for a request corresponding to this page. If there
 852          * is one, and it belongs to another file, we flush it out
 853          * before we try to copy anything into the page. Do this
 854          * due to the lack of an ACCESS-type call in NFSv2.
 855          * Also do the same if we find a request from an existing
 856          * dropped page.
 857          */
 858         do {
 859                 req = nfs_page_find_request(page);
 860                 if (req == NULL)
 861                         return 0;
 862                 l_ctx = req->wb_lock_context;
 863                 do_flush = req->wb_page != page || req->wb_context != ctx;
 864                 if (l_ctx) {
 865                         do_flush |= l_ctx->lockowner.l_owner != current->files
 866                                 || l_ctx->lockowner.l_pid != current->tgid;
 867                 }
 868                 nfs_release_request(req);
 869                 if (!do_flush)
 870                         return 0;
 871                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 872         } while (status == 0);
 873         return status;
 874 }
 875
 876 /*
 877  * If the page cache is marked as unsafe or invalid, then we can't rely on
 878  * the PageUptodate() flag. In this case, we will need to turn off
 879  * write optimisations that depend on the page contents being correct.
 880  */
 881 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 882 {
 883         if (nfs_have_delegated_attributes(inode))
 884                 goto out;
 885         if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 886                 return false;
 887 out:
 888         return PageUptodate(page) != 0;
 889 }
 890
 891 /* If we know the page is up to date, and we're not using byte range locks (or
 892  * if we have the whole file locked for writing), it may be more efficient to
 893  * extend the write to cover the entire page in order to avoid fragmentation
 894  * inefficiencies.
 895  *
 896  * If the file is opened for synchronous writes or if we have a write delegation
 897  * from the server then we can just skip the rest of the checks.
 898  */
 899 static int nfs_can_extend_write(struct file *file, struct page *page, struct inode *inode)
 900 {
 901         if (file->f_flags & O_DSYNC)
 902                 return 0;
 903         if (NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 904                 return 1;
 905         if (nfs_write_pageuptodate(page, inode) && (inode->i_flock == NULL ||
 906                         (inode->i_flock->fl_start == 0 &&
 907                         inode->i_flock->fl_end == OFFSET_MAX &&
 908                         inode->i_flock->fl_type != F_RDLCK)))
 909                 return 1;
 910         return 0;
 911 }
 912
 913 /*
 914  * Update and possibly write a cached page of an NFS file.
 915  *
 916  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 917  * things with a page scheduled for an RPC call (e.g. invalidate it).
 918  */
 919 int nfs_updatepage(struct file *file, struct page *page,
 920                 unsigned int offset, unsigned int count)
 921 {
 922         struct nfs_open_context *ctx = nfs_file_open_context(file);
 923         struct inode    *inode = page_file_mapping(page)->host;
 924         int             status = 0;
 925
 926         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 927
 928         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 929                 file->f_path.dentry->d_parent->d_name.name,
 930                 file->f_path.dentry->d_name.name, count,
 931                 (long long)(page_file_offset(page) + offset));
 932
 933         if (nfs_can_extend_write(file, page, inode)) {
 934                 count = max(count + offset, nfs_page_length(page));
 935                 offset = 0;
 936         }
 937
 938         status = nfs_writepage_setup(ctx, page, offset, count);
 939         if (status < 0)
 940                 nfs_set_pageerror(page);
 941         else
 942                 __set_page_dirty_nobuffers(page);
 943
 944         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 945                         status, (long long)i_size_read(inode));
 946         return status;
 947 }
 948
 949 static int flush_task_priority(int how)
 950 {
 951         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 952                 case FLUSH_HIGHPRI:
 953                         return RPC_PRIORITY_HIGH;
 954                 case FLUSH_LOWPRI:
 955                         return RPC_PRIORITY_LOW;
 956         }
 957         return RPC_PRIORITY_NORMAL;
 958 }
 959
 960 int nfs_initiate_write(struct rpc_clnt *clnt,
 961                        struct nfs_write_data *data,
 962                        const struct rpc_call_ops *call_ops,
 963                        int how, int flags)
 964 {
 965         struct inode *inode = data->header->inode;
 966         int priority = flush_task_priority(how);
 967         struct rpc_task *task;
 968         struct rpc_message msg = {
 969                 .rpc_argp = &data->args,
 970                 .rpc_resp = &data->res,
 971                 .rpc_cred = data->header->cred,
 972         };
 973         struct rpc_task_setup task_setup_data = {
 974                 .rpc_client = clnt,
 975                 .task = &data->task,
 976                 .rpc_message = &msg,
 977                 .callback_ops = call_ops,
 978                 .callback_data = data,
 979                 .workqueue = nfsiod_workqueue,
 980                 .flags = RPC_TASK_ASYNC | flags,
 981                 .priority = priority,
 982         };
 983         int ret = 0;
 984
 985         /* Set up the initial task struct.  */
 986         NFS_PROTO(inode)->write_setup(data, &msg);
 987
 988         dprintk("NFS: %5u initiated write call "
 989                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 990                 data->task.tk_pid,
 991                 inode->i_sb->s_id,
 992                 (long long)NFS_FILEID(inode),
 993                 data->args.count,
 994                 (unsigned long long)data->args.offset);
 995
 996         task = rpc_run_task(&task_setup_data);
 997         if (IS_ERR(task)) {
 998                 ret = PTR_ERR(task);
 999                 goto out;
1000         }
1001         if (how & FLUSH_SYNC) {
1002                 ret = rpc_wait_for_completion_task(task);
1003                 if (ret == 0)
1004                         ret = task->tk_status;
1005         }
1006         rpc_put_task(task);
1007 out:
1008         return ret;
1009 }
1010 EXPORT_SYMBOL_GPL(nfs_initiate_write);
1011
1012 /*
1013  * Set up the argument/result storage required for the RPC call.
1014  */
1015 static void nfs_write_rpcsetup(struct nfs_write_data *data,
1016                 unsigned int count, unsigned int offset,
1017                 int how, struct nfs_commit_info *cinfo)
1018 {
1019         struct nfs_page *req = data->header->req;
1020
1021         /* Set up the RPC argument and reply structs
1022          * NB: take care not to mess about with data->commit et al. */
1023
1024         data->args.fh     = NFS_FH(data->header->inode);
1025         data->args.offset = req_offset(req) + offset;
1026         /* pnfs_set_layoutcommit needs this */
1027         data->mds_offset = data->args.offset;
1028         data->args.pgbase = req->wb_pgbase + offset;
1029         data->args.pages  = data->pages.pagevec;
1030         data->args.count  = count;
1031         data->args.context = get_nfs_open_context(req->wb_context);
1032         data->args.lock_context = req->wb_lock_context;
1033         data->args.stable  = NFS_UNSTABLE;
1034         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1035         case 0:
1036                 break;
1037         case FLUSH_COND_STABLE:
1038                 if (nfs_reqs_to_commit(cinfo))
1039                         break;
1040         default:
1041                 data->args.stable = NFS_FILE_SYNC;
1042         }
1043
1044         data->res.fattr   = &data->fattr;
1045         data->res.count   = count;
1046         data->res.verf    = &data->verf;
1047         nfs_fattr_init(&data->fattr);
1048 }
1049
1050 static int nfs_do_write(struct nfs_write_data *data,
1051                 const struct rpc_call_ops *call_ops,
1052                 int how)
1053 {
1054         struct inode *inode = data->header->inode;
1055
1056         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1057 }
1058
1059 static int nfs_do_multiple_writes(struct list_head *head,
1060                 const struct rpc_call_ops *call_ops,
1061                 int how)
1062 {
1063         struct nfs_write_data *data;
1064         int ret = 0;
1065
1066         while (!list_empty(head)) {
1067                 int ret2;
1068
1069                 data = list_first_entry(head, struct nfs_write_data, list);
1070                 list_del_init(&data->list);
1071
1072                 ret2 = nfs_do_write(data, call_ops, how);
1073                  if (ret == 0)
1074                          ret = ret2;
1075         }
1076         return ret;
1077 }
1078
1079 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1080  * call this on each, which will prepare them to be retried on next
1081  * writeback using standard nfs.
1082  */
1083 static void nfs_redirty_request(struct nfs_page *req)
1084 {
1085         nfs_mark_request_dirty(req);
1086         nfs_unlock_request(req);
1087         nfs_end_page_writeback(req->wb_page);
1088         nfs_release_request(req);
1089 }
1090
1091 static void nfs_async_write_error(struct list_head *head)
1092 {
1093         struct nfs_page *req;
1094
1095         while (!list_empty(head)) {
1096                 req = nfs_list_entry(head->next);
1097                 nfs_list_remove_request(req);
1098                 nfs_redirty_request(req);
1099         }
1100 }
1101
1102 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1103         .error_cleanup = nfs_async_write_error,
1104         .completion = nfs_write_completion,
1105 };
1106
1107 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1108                 struct nfs_pgio_header *hdr)
1109 {
1110         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1111         while (!list_empty(&hdr->rpc_list)) {
1112                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1113                                 struct nfs_write_data, list);
1114                 list_del(&data->list);
1115                 nfs_writedata_release(data);
1116         }
1117         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1118 }
1119
1120 /*
1121  * Generate multiple small requests to write out a single
1122  * contiguous dirty area on one page.
1123  */
1124 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1125                            struct nfs_pgio_header *hdr)
1126 {
1127         struct nfs_page *req = hdr->req;
1128         struct page *page = req->wb_page;
1129         struct nfs_write_data *data;
1130         size_t wsize = desc->pg_bsize, nbytes;
1131         unsigned int offset;
1132         int requests = 0;
1133         struct nfs_commit_info cinfo;
1134
1135         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1136
1137         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1138             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1139              desc->pg_count > wsize))
1140                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1141
1142
1143         offset = 0;
1144         nbytes = desc->pg_count;
1145         do {
1146                 size_t len = min(nbytes, wsize);
1147
1148                 data = nfs_writedata_alloc(hdr, 1);
1149                 if (!data) {
1150                         nfs_flush_error(desc, hdr);
1151                         return -ENOMEM;
1152                 }
1153                 data->pages.pagevec[0] = page;
1154                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1155                 list_add(&data->list, &hdr->rpc_list);
1156                 requests++;
1157                 nbytes -= len;
1158                 offset += len;
1159         } while (nbytes != 0);
1160         nfs_list_remove_request(req);
1161         nfs_list_add_request(req, &hdr->pages);
1162         desc->pg_rpc_callops = &nfs_write_common_ops;
1163         return 0;
1164 }
1165
1166 /*
1167  * Create an RPC task for the given write request and kick it.
1168  * The page must have been locked by the caller.
1169  *
1170  * It may happen that the page we're passed is not marked dirty.
1171  * This is the case if nfs_updatepage detects a conflicting request
1172  * that has been written but not committed.
1173  */
1174 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1175                          struct nfs_pgio_header *hdr)
1176 {
1177         struct nfs_page         *req;
1178         struct page             **pages;
1179         struct nfs_write_data   *data;
1180         struct list_head *head = &desc->pg_list;
1181         struct nfs_commit_info cinfo;
1182
1183         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1184                                                            desc->pg_count));
1185         if (!data) {
1186                 nfs_flush_error(desc, hdr);
1187                 return -ENOMEM;
1188         }
1189
1190         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1191         pages = data->pages.pagevec;
1192         while (!list_empty(head)) {
1193                 req = nfs_list_entry(head->next);
1194                 nfs_list_remove_request(req);
1195                 nfs_list_add_request(req, &hdr->pages);
1196                 *pages++ = req->wb_page;
1197         }
1198
1199         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1200             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1201                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1202
1203         /* Set up the argument struct */
1204         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1205         list_add(&data->list, &hdr->rpc_list);
1206         desc->pg_rpc_callops = &nfs_write_common_ops;
1207         return 0;
1208 }
1209
1210 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1211                       struct nfs_pgio_header *hdr)
1212 {
1213         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1214                 return nfs_flush_multi(desc, hdr);
1215         return nfs_flush_one(desc, hdr);
1216 }
1217 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1218
1219 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1220 {
1221         struct nfs_write_header *whdr;
1222         struct nfs_pgio_header *hdr;
1223         int ret;
1224
1225         whdr = nfs_writehdr_alloc();
1226         if (!whdr) {
1227                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1228                 return -ENOMEM;
1229         }
1230         hdr = &whdr->header;
1231         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1232         atomic_inc(&hdr->refcnt);
1233         ret = nfs_generic_flush(desc, hdr);
1234         if (ret == 0)
1235                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1236                                              desc->pg_rpc_callops,
1237                                              desc->pg_ioflags);
1238         if (atomic_dec_and_test(&hdr->refcnt))
1239                 hdr->completion_ops->completion(hdr);
1240         return ret;
1241 }
1242
1243 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1244         .pg_test = nfs_generic_pg_test,
1245         .pg_doio = nfs_generic_pg_writepages,
1246 };
1247
1248 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1249                                struct inode *inode, int ioflags,
1250                                const struct nfs_pgio_completion_ops *compl_ops)
1251 {
1252         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1253                                 NFS_SERVER(inode)->wsize, ioflags);
1254 }
1255 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1256
1257 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1258 {
1259         pgio->pg_ops = &nfs_pageio_write_ops;
1260         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1261 }
1262 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1263
1264
1265 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1266 {
1267         struct nfs_write_data *data = calldata;
1268         NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1269         if (unlikely(test_bit(NFS_CONTEXT_BAD, &data->args.context->flags)))
1270                 rpc_exit(task, -EIO);
1271 }
1272
1273 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1274 {
1275         struct nfs_commit_data *data = calldata;
1276
1277         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1278 }
1279
1280 /*
1281  * Handle a write reply that flushes a whole page.
1282  *
1283  * FIXME: There is an inherent race with invalidate_inode_pages and
1284  *        writebacks since the page->count is kept > 1 for as long
1285  *        as the page has a write request pending.
1286  */
1287 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1288 {
1289         struct nfs_write_data   *data = calldata;
1290
1291         nfs_writeback_done(task, data);
1292 }
1293
1294 static void nfs_writeback_release_common(void *calldata)
1295 {
1296         struct nfs_write_data   *data = calldata;
1297         struct nfs_pgio_header *hdr = data->header;
1298         int status = data->task.tk_status;
1299
1300         if ((status >= 0) && nfs_write_need_commit(data)) {
1301                 spin_lock(&hdr->lock);
1302                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1303                         ; /* Do nothing */
1304                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1305                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1306                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1307                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1308                 spin_unlock(&hdr->lock);
1309         }
1310         nfs_writedata_release(data);
1311 }
1312
1313 static const struct rpc_call_ops nfs_write_common_ops = {
1314         .rpc_call_prepare = nfs_write_prepare,
1315         .rpc_call_done = nfs_writeback_done_common,
1316         .rpc_release = nfs_writeback_release_common,
1317 };
1318
1319
1320 /*
1321  * This function is called when the WRITE call is complete.
1322  */
1323 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1324 {
1325         struct nfs_writeargs    *argp = &data->args;
1326         struct nfs_writeres     *resp = &data->res;
1327         struct inode            *inode = data->header->inode;
1328         int status;
1329
1330         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1331                 task->tk_pid, task->tk_status);
1332
1333         /*
1334          * ->write_done will attempt to use post-op attributes to detect
1335          * conflicting writes by other clients.  A strict interpretation
1336          * of close-to-open would allow us to continue caching even if
1337          * another writer had changed the file, but some applications
1338          * depend on tighter cache coherency when writing.
1339          */
1340         status = NFS_PROTO(inode)->write_done(task, data);
1341         if (status != 0)
1342                 return;
1343         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1344
1345 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1346         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1347                 /* We tried a write call, but the server did not
1348                  * commit data to stable storage even though we
1349                  * requested it.
1350                  * Note: There is a known bug in Tru64 < 5.0 in which
1351                  *       the server reports NFS_DATA_SYNC, but performs
1352                  *       NFS_FILE_SYNC. We therefore implement this checking
1353                  *       as a dprintk() in order to avoid filling syslog.
1354                  */
1355                 static unsigned long    complain;
1356
1357                 /* Note this will print the MDS for a DS write */
1358                 if (time_before(complain, jiffies)) {
1359                         dprintk("NFS:       faulty NFS server %s:"
1360                                 " (committed = %d) != (stable = %d)\n",
1361                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1362                                 resp->verf->committed, argp->stable);
1363                         complain = jiffies + 300 * HZ;
1364                 }
1365         }
1366 #endif
1367         if (task->tk_status < 0)
1368                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1369         else if (resp->count < argp->count) {
1370                 static unsigned long    complain;
1371
1372                 /* This a short write! */
1373                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1374
1375                 /* Has the server at least made some progress? */
1376                 if (resp->count == 0) {
1377                         if (time_before(complain, jiffies)) {
1378                                 printk(KERN_WARNING
1379                                        "NFS: Server wrote zero bytes, expected %u.\n",
1380                                        argp->count);
1381                                 complain = jiffies + 300 * HZ;
1382                         }
1383                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1384                         task->tk_status = -EIO;
1385                         return;
1386                 }
1387                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1388                 if (resp->verf->committed != NFS_UNSTABLE) {
1389                         /* Resend from where the server left off */
1390                         data->mds_offset += resp->count;
1391                         argp->offset += resp->count;
1392                         argp->pgbase += resp->count;
1393                         argp->count -= resp->count;
1394                 } else {
1395                         /* Resend as a stable write in order to avoid
1396                          * headaches in the case of a server crash.
1397                          */
1398                         argp->stable = NFS_FILE_SYNC;
1399                 }
1400                 rpc_restart_call_prepare(task);
1401         }
1402 }
1403
1404
1405 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1406 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1407 {
1408         int ret;
1409
1410         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1411                 return 1;
1412         if (!may_wait)
1413                 return 0;
1414         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1415                                 NFS_INO_COMMIT,
1416                                 nfs_wait_bit_killable,
1417                                 TASK_KILLABLE);
1418         return (ret < 0) ? ret : 1;
1419 }
1420
1421 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1422 {
1423         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1424         smp_mb__after_clear_bit();
1425         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1426 }
1427
1428 void nfs_commitdata_release(struct nfs_commit_data *data)
1429 {
1430         put_nfs_open_context(data->context);
1431         nfs_commit_free(data);
1432 }
1433 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1434
1435 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1436                         const struct rpc_call_ops *call_ops,
1437                         int how, int flags)
1438 {
1439         struct rpc_task *task;
1440         int priority = flush_task_priority(how);
1441         struct rpc_message msg = {
1442                 .rpc_argp = &data->args,
1443                 .rpc_resp = &data->res,
1444                 .rpc_cred = data->cred,
1445         };
1446         struct rpc_task_setup task_setup_data = {
1447                 .task = &data->task,
1448                 .rpc_client = clnt,
1449                 .rpc_message = &msg,
1450                 .callback_ops = call_ops,
1451                 .callback_data = data,
1452                 .workqueue = nfsiod_workqueue,
1453                 .flags = RPC_TASK_ASYNC | flags,
1454                 .priority = priority,
1455         };
1456         /* Set up the initial task struct.  */
1457         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1458
1459         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1460
1461         task = rpc_run_task(&task_setup_data);
1462         if (IS_ERR(task))
1463                 return PTR_ERR(task);
1464         if (how & FLUSH_SYNC)
1465                 rpc_wait_for_completion_task(task);
1466         rpc_put_task(task);
1467         return 0;
1468 }
1469 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1470
1471 /*
1472  * Set up the argument/result storage required for the RPC call.
1473  */
1474 void nfs_init_commit(struct nfs_commit_data *data,
1475                      struct list_head *head,
1476                      struct pnfs_layout_segment *lseg,
1477                      struct nfs_commit_info *cinfo)
1478 {
1479         struct nfs_page *first = nfs_list_entry(head->next);
1480         struct inode *inode = first->wb_context->dentry->d_inode;
1481
1482         /* Set up the RPC argument and reply structs
1483          * NB: take care not to mess about with data->commit et al. */
1484
1485         list_splice_init(head, &data->pages);
1486
1487         data->inode       = inode;
1488         data->cred        = first->wb_context->cred;
1489         data->lseg        = lseg; /* reference transferred */
1490         data->mds_ops     = &nfs_commit_ops;
1491         data->completion_ops = cinfo->completion_ops;
1492         data->dreq        = cinfo->dreq;
1493
1494         data->args.fh     = NFS_FH(data->inode);
1495         /* Note: we always request a commit of the entire inode */
1496         data->args.offset = 0;
1497         data->args.count  = 0;
1498         data->context     = get_nfs_open_context(first->wb_context);
1499         data->res.fattr   = &data->fattr;
1500         data->res.verf    = &data->verf;
1501         nfs_fattr_init(&data->fattr);
1502 }
1503 EXPORT_SYMBOL_GPL(nfs_init_commit);
1504
1505 void nfs_retry_commit(struct list_head *page_list,
1506                       struct pnfs_layout_segment *lseg,
1507                       struct nfs_commit_info *cinfo)
1508 {
1509         struct nfs_page *req;
1510
1511         while (!list_empty(page_list)) {
1512                 req = nfs_list_entry(page_list->next);
1513                 nfs_list_remove_request(req);
1514                 nfs_mark_request_commit(req, lseg, cinfo);
1515                 if (!cinfo->dreq) {
1516                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1517                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1518                                      BDI_RECLAIMABLE);
1519                 }
1520                 nfs_unlock_and_release_request(req);
1521         }
1522 }
1523 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1524
1525 /*
1526  * Commit dirty pages
1527  */
1528 static int
1529 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1530                 struct nfs_commit_info *cinfo)
1531 {
1532         struct nfs_commit_data  *data;
1533
1534         data = nfs_commitdata_alloc();
1535
1536         if (!data)
1537                 goto out_bad;
1538
1539         /* Set up the argument struct */
1540         nfs_init_commit(data, head, NULL, cinfo);
1541         atomic_inc(&cinfo->mds->rpcs_out);
1542         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1543                                    how, 0);
1544  out_bad:
1545         nfs_retry_commit(head, NULL, cinfo);
1546         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1547         return -ENOMEM;
1548 }
1549
1550 /*
1551  * COMMIT call returned
1552  */
1553 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1554 {
1555         struct nfs_commit_data  *data = calldata;
1556
1557         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1558                                 task->tk_pid, task->tk_status);
1559
1560         /* Call the NFS version-specific code */
1561         NFS_PROTO(data->inode)->commit_done(task, data);
1562 }
1563
1564 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1565 {
1566         struct nfs_page *req;
1567         int status = data->task.tk_status;
1568         struct nfs_commit_info cinfo;
1569
1570         while (!list_empty(&data->pages)) {
1571                 req = nfs_list_entry(data->pages.next);
1572                 nfs_list_remove_request(req);
1573                 nfs_clear_page_commit(req->wb_page);
1574
1575                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1576                         req->wb_context->dentry->d_sb->s_id,
1577                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1578                         req->wb_bytes,
1579                         (long long)req_offset(req));
1580                 if (status < 0) {
1581                         nfs_context_set_write_error(req->wb_context, status);
1582                         nfs_inode_remove_request(req);
1583                         dprintk(", error = %d\n", status);
1584                         goto next;
1585                 }
1586
1587                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1588                  * returned by the server against all stored verfs. */
1589                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1590                         /* We have a match */
1591                         nfs_inode_remove_request(req);
1592                         dprintk(" OK\n");
1593                         goto next;
1594                 }
1595                 /* We have a mismatch. Write the page again */
1596                 dprintk(" mismatch\n");
1597                 nfs_mark_request_dirty(req);
1598                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1599         next:
1600                 nfs_unlock_and_release_request(req);
1601         }
1602         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1603         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1604                 nfs_commit_clear_lock(NFS_I(data->inode));
1605 }
1606
1607 static void nfs_commit_release(void *calldata)
1608 {
1609         struct nfs_commit_data *data = calldata;
1610
1611         data->completion_ops->completion(data);
1612         nfs_commitdata_release(calldata);
1613 }
1614
1615 static const struct rpc_call_ops nfs_commit_ops = {
1616         .rpc_call_prepare = nfs_commit_prepare,
1617         .rpc_call_done = nfs_commit_done,
1618         .rpc_release = nfs_commit_release,
1619 };
1620
1621 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1622         .completion = nfs_commit_release_pages,
1623         .error_cleanup = nfs_commit_clear_lock,
1624 };
1625
1626 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1627                             int how, struct nfs_commit_info *cinfo)
1628 {
1629         int status;
1630
1631         status = pnfs_commit_list(inode, head, how, cinfo);
1632         if (status == PNFS_NOT_ATTEMPTED)
1633                 status = nfs_commit_list(inode, head, how, cinfo);
1634         return status;
1635 }
1636
1637 int nfs_commit_inode(struct inode *inode, int how)
1638 {
1639         LIST_HEAD(head);
1640         struct nfs_commit_info cinfo;
1641         int may_wait = how & FLUSH_SYNC;
1642         int res;
1643
1644         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1645         if (res <= 0)
1646                 goto out_mark_dirty;
1647         nfs_init_cinfo_from_inode(&cinfo, inode);
1648         res = nfs_scan_commit(inode, &head, &cinfo);
1649         if (res) {
1650                 int error;
1651
1652                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1653                 if (error < 0)
1654                         return error;
1655                 if (!may_wait)
1656                         goto out_mark_dirty;
1657                 error = wait_on_bit(&NFS_I(inode)->flags,
1658                                 NFS_INO_COMMIT,
1659                                 nfs_wait_bit_killable,
1660                                 TASK_KILLABLE);
1661                 if (error < 0)
1662                         return error;
1663         } else
1664                 nfs_commit_clear_lock(NFS_I(inode));
1665         return res;
1666         /* Note: If we exit without ensuring that the commit is complete,
1667          * we must mark the inode as dirty. Otherwise, future calls to
1668          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1669          * that the data is on the disk.
1670          */
1671 out_mark_dirty:
1672         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1673         return res;
1674 }
1675
1676 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1677 {
1678         struct nfs_inode *nfsi = NFS_I(inode);
1679         int flags = FLUSH_SYNC;
1680         int ret = 0;
1681
1682         /* no commits means nothing needs to be done */
1683         if (!nfsi->commit_info.ncommit)
1684                 return ret;
1685
1686         if (wbc->sync_mode == WB_SYNC_NONE) {
1687                 /* Don't commit yet if this is a non-blocking flush and there
1688                  * are a lot of outstanding writes for this mapping.
1689                  */
1690                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1691                         goto out_mark_dirty;
1692
1693                 /* don't wait for the COMMIT response */
1694                 flags = 0;
1695         }
1696
1697         ret = nfs_commit_inode(inode, flags);
1698         if (ret >= 0) {
1699                 if (wbc->sync_mode == WB_SYNC_NONE) {
1700                         if (ret < wbc->nr_to_write)
1701                                 wbc->nr_to_write -= ret;
1702                         else
1703                                 wbc->nr_to_write = 0;
1704                 }
1705                 return 0;
1706         }
1707 out_mark_dirty:
1708         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1709         return ret;
1710 }
1711 #else
1712 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1713 {
1714         return 0;
1715 }
1716 #endif
1717
1718 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1719 {
1720         return nfs_commit_unstable_pages(inode, wbc);
1721 }
1722 EXPORT_SYMBOL_GPL(nfs_write_inode);
1723
1724 /*
1725  * flush the inode to disk.
1726  */
1727 int nfs_wb_all(struct inode *inode)
1728 {
1729         struct writeback_control wbc = {
1730                 .sync_mode = WB_SYNC_ALL,
1731                 .nr_to_write = LONG_MAX,
1732                 .range_start = 0,
1733                 .range_end = LLONG_MAX,
1734         };
1735
1736         return sync_inode(inode, &wbc);
1737 }
1738 EXPORT_SYMBOL_GPL(nfs_wb_all);
1739
1740 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1741 {
1742         struct nfs_page *req;
1743         int ret = 0;
1744
1745         for (;;) {
1746                 wait_on_page_writeback(page);
1747                 req = nfs_page_find_request(page);
1748                 if (req == NULL)
1749                         break;
1750                 if (nfs_lock_request(req)) {
1751                         nfs_clear_request_commit(req);
1752                         nfs_inode_remove_request(req);
1753                         /*
1754                          * In case nfs_inode_remove_request has marked the
1755                          * page as being dirty
1756                          */
1757                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1758                         nfs_unlock_and_release_request(req);
1759                         break;
1760                 }
1761                 ret = nfs_wait_on_request(req);
1762                 nfs_release_request(req);
1763                 if (ret < 0)
1764                         break;
1765         }
1766         return ret;
1767 }
1768
1769 /*
1770  * Write back all requests on one page - we do this before reading it.
1771  */
1772 int nfs_wb_page(struct inode *inode, struct page *page)
1773 {
1774         loff_t range_start = page_file_offset(page);
1775         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1776         struct writeback_control wbc = {
1777                 .sync_mode = WB_SYNC_ALL,
1778                 .nr_to_write = 0,
1779                 .range_start = range_start,
1780                 .range_end = range_end,
1781         };
1782         int ret;
1783
1784         for (;;) {
1785                 wait_on_page_writeback(page);
1786                 if (clear_page_dirty_for_io(page)) {
1787                         ret = nfs_writepage_locked(page, &wbc);
1788                         if (ret < 0)
1789                                 goto out_error;
1790                         continue;
1791                 }
1792                 if (!PagePrivate(page))
1793                         break;
1794                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1795                 if (ret < 0)
1796                         goto out_error;
1797         }
1798         return 0;
1799 out_error:
1800         return ret;
1801 }
1802
1803 #ifdef CONFIG_MIGRATION
1804 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1805                 struct page *page, enum migrate_mode mode)
1806 {
1807         /*
1808          * If PagePrivate is set, then the page is currently associated with
1809          * an in-progress read or write request. Don't try to migrate it.
1810          *
1811          * FIXME: we could do this in principle, but we'll need a way to ensure
1812          *        that we can safely release the inode reference while holding
1813          *        the page lock.
1814          */
1815         if (PagePrivate(page))
1816                 return -EBUSY;
1817
1818         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1819                 return -EBUSY;
1820
1821         return migrate_page(mapping, newpage, page, mode);
1822 }
1823 #endif
1824
1825 int __init nfs_init_writepagecache(void)
1826 {
1827         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1828                                              sizeof(struct nfs_write_header),
1829                                              0, SLAB_HWCACHE_ALIGN,
1830                                              NULL);
1831         if (nfs_wdata_cachep == NULL)
1832                 return -ENOMEM;
1833
1834         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1835                                                      nfs_wdata_cachep);
1836         if (nfs_wdata_mempool == NULL)
1837                 goto out_destroy_write_cache;
1838
1839         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1840                                              sizeof(struct nfs_commit_data),
1841                                              0, SLAB_HWCACHE_ALIGN,
1842                                              NULL);
1843         if (nfs_cdata_cachep == NULL)
1844                 goto out_destroy_write_mempool;
1845
1846         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1847                                                       nfs_cdata_cachep);
1848         if (nfs_commit_mempool == NULL)
1849                 goto out_destroy_commit_cache;
1850
1851         /*
1852          * NFS congestion size, scale with available memory.
1853          *
1854          *  64MB:    8192k
1855          * 128MB:   11585k
1856          * 256MB:   16384k
1857          * 512MB:   23170k
1858          *   1GB:   32768k
1859          *   2GB:   46340k
1860          *   4GB:   65536k
1861          *   8GB:   92681k
1862          *  16GB:  131072k
1863          *
1864          * This allows larger machines to have larger/more transfers.
1865          * Limit the default to 256M
1866          */
1867         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1868         if (nfs_congestion_kb > 256*1024)
1869                 nfs_congestion_kb = 256*1024;
1870
1871         return 0;
1872
1873 out_destroy_commit_cache:
1874         kmem_cache_destroy(nfs_cdata_cachep);
1875 out_destroy_write_mempool:
1876         mempool_destroy(nfs_wdata_mempool);
1877 out_destroy_write_cache:
1878         kmem_cache_destroy(nfs_wdata_cachep);
1879         return -ENOMEM;
1880 }
1881
1882 void nfs_destroy_writepagecache(void)
1883 {
1884         mempool_destroy(nfs_commit_mempool);
1885         kmem_cache_destroy(nfs_cdata_cachep);
1886         mempool_destroy(nfs_wdata_mempool);
1887         kmem_cache_destroy(nfs_wdata_cachep);
1888 }
1889