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