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                 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                 kfree(wdata);
 131         else
 132                 wdata->header = NULL;
 133         if (atomic_dec_and_test(&hdr->refcnt))
 134                 hdr->completion_ops->completion(hdr);
 135 }
 136 EXPORT_SYMBOL_GPL(nfs_writedata_release);
 137
 138 static void nfs_context_set_write_error(struct nfs_open_context *ctx, int error)
 139 {
 140         ctx->error = error;
 141         smp_wmb();
 142         set_bit(NFS_CONTEXT_ERROR_WRITE, &ctx->flags);
 143 }
 144
 145 static struct nfs_page *nfs_page_find_request_locked(struct page *page)
 146 {
 147         struct nfs_page *req = NULL;
 148
 149         if (PagePrivate(page)) {
 150                 req = (struct nfs_page *)page_private(page);
 151                 if (req != NULL)
 152                         kref_get(&req->wb_kref);
 153         }
 154         return req;
 155 }
 156
 157 static struct nfs_page *nfs_page_find_request(struct page *page)
 158 {
 159         struct inode *inode = page->mapping->host;
 160         struct nfs_page *req = NULL;
 161
 162         spin_lock(&inode->i_lock);
 163         req = nfs_page_find_request_locked(page);
 164         spin_unlock(&inode->i_lock);
 165         return req;
 166 }
 167
 168 /* Adjust the file length if we're writing beyond the end */
 169 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 170 {
 171         struct inode *inode = page->mapping->host;
 172         loff_t end, i_size;
 173         pgoff_t end_index;
 174
 175         spin_lock(&inode->i_lock);
 176         i_size = i_size_read(inode);
 177         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 178         if (i_size > 0 && page->index < end_index)
 179                 goto out;
 180         end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count);
 181         if (i_size >= end)
 182                 goto out;
 183         i_size_write(inode, end);
 184         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 185 out:
 186         spin_unlock(&inode->i_lock);
 187 }
 188
 189 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 190 static void nfs_set_pageerror(struct page *page)
 191 {
 192         SetPageError(page);
 193         nfs_zap_mapping(page->mapping->host, page->mapping);
 194 }
 195
 196 /* We can set the PG_uptodate flag if we see that a write request
 197  * covers the full page.
 198  */
 199 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 200 {
 201         if (PageUptodate(page))
 202                 return;
 203         if (base != 0)
 204                 return;
 205         if (count != nfs_page_length(page))
 206                 return;
 207         SetPageUptodate(page);
 208 }
 209
 210 static int wb_priority(struct writeback_control *wbc)
 211 {
 212         if (wbc->for_reclaim)
 213                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 214         if (wbc->for_kupdate || wbc->for_background)
 215                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 216         return FLUSH_COND_STABLE;
 217 }
 218
 219 /*
 220  * NFS congestion control
 221  */
 222
 223 int nfs_congestion_kb;
 224
 225 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 226 #define NFS_CONGESTION_OFF_THRESH       \
 227         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 228
 229 static int nfs_set_page_writeback(struct page *page)
 230 {
 231         int ret = test_set_page_writeback(page);
 232
 233         if (!ret) {
 234                 struct inode *inode = page->mapping->host;
 235                 struct nfs_server *nfss = NFS_SERVER(inode);
 236
 237                 if (atomic_long_inc_return(&nfss->writeback) >
 238                                 NFS_CONGESTION_ON_THRESH) {
 239                         set_bdi_congested(&nfss->backing_dev_info,
 240                                                 BLK_RW_ASYNC);
 241                 }
 242         }
 243         return ret;
 244 }
 245
 246 static void nfs_end_page_writeback(struct page *page)
 247 {
 248         struct inode *inode = page->mapping->host;
 249         struct nfs_server *nfss = NFS_SERVER(inode);
 250
 251         end_page_writeback(page);
 252         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 253                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 254 }
 255
 256 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 257 {
 258         struct inode *inode = page->mapping->host;
 259         struct nfs_page *req;
 260         int ret;
 261
 262         spin_lock(&inode->i_lock);
 263         for (;;) {
 264                 req = nfs_page_find_request_locked(page);
 265                 if (req == NULL)
 266                         break;
 267                 if (nfs_lock_request(req))
 268                         break;
 269                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 270                  *       then the call to nfs_lock_request() will always
 271                  *       succeed provided that someone hasn't already marked the
 272                  *       request as dirty (in which case we don't care).
 273                  */
 274                 spin_unlock(&inode->i_lock);
 275                 if (!nonblock)
 276                         ret = nfs_wait_on_request(req);
 277                 else
 278                         ret = -EAGAIN;
 279                 nfs_release_request(req);
 280                 if (ret != 0)
 281                         return ERR_PTR(ret);
 282                 spin_lock(&inode->i_lock);
 283         }
 284         spin_unlock(&inode->i_lock);
 285         return req;
 286 }
 287
 288 /*
 289  * Find an associated nfs write request, and prepare to flush it out
 290  * May return an error if the user signalled nfs_wait_on_request().
 291  */
 292 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 293                                 struct page *page, bool nonblock)
 294 {
 295         struct nfs_page *req;
 296         int ret = 0;
 297
 298         req = nfs_find_and_lock_request(page, nonblock);
 299         if (!req)
 300                 goto out;
 301         ret = PTR_ERR(req);
 302         if (IS_ERR(req))
 303                 goto out;
 304
 305         ret = nfs_set_page_writeback(page);
 306         BUG_ON(ret != 0);
 307         BUG_ON(test_bit(PG_CLEAN, &req->wb_flags));
 308
 309         if (!nfs_pageio_add_request(pgio, req)) {
 310                 nfs_redirty_request(req);
 311                 ret = pgio->pg_error;
 312         }
 313 out:
 314         return ret;
 315 }
 316
 317 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 318 {
 319         struct inode *inode = page->mapping->host;
 320         int ret;
 321
 322         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 323         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 324
 325         nfs_pageio_cond_complete(pgio, page->index);
 326         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 327         if (ret == -EAGAIN) {
 328                 redirty_page_for_writepage(wbc, page);
 329                 ret = 0;
 330         }
 331         return ret;
 332 }
 333
 334 /*
 335  * Write an mmapped page to the server.
 336  */
 337 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 338 {
 339         struct nfs_pageio_descriptor pgio;
 340         int err;
 341
 342         NFS_PROTO(page->mapping->host)->write_pageio_init(&pgio,
 343                                                           page->mapping->host,
 344                                                           wb_priority(wbc),
 345                                                           &nfs_async_write_completion_ops);
 346         err = nfs_do_writepage(page, wbc, &pgio);
 347         nfs_pageio_complete(&pgio);
 348         if (err < 0)
 349                 return err;
 350         if (pgio.pg_error < 0)
 351                 return pgio.pg_error;
 352         return 0;
 353 }
 354
 355 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 356 {
 357         int ret;
 358
 359         ret = nfs_writepage_locked(page, wbc);
 360         unlock_page(page);
 361         return ret;
 362 }
 363
 364 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 365 {
 366         int ret;
 367
 368         ret = nfs_do_writepage(page, wbc, data);
 369         unlock_page(page);
 370         return ret;
 371 }
 372
 373 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 374 {
 375         struct inode *inode = mapping->host;
 376         unsigned long *bitlock = &NFS_I(inode)->flags;
 377         struct nfs_pageio_descriptor pgio;
 378         int err;
 379
 380         /* Stop dirtying of new pages while we sync */
 381         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 382                         nfs_wait_bit_killable, TASK_KILLABLE);
 383         if (err)
 384                 goto out_err;
 385
 386         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 387
 388         NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
 389         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 390         nfs_pageio_complete(&pgio);
 391
 392         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 393         smp_mb__after_clear_bit();
 394         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 395
 396         if (err < 0)
 397                 goto out_err;
 398         err = pgio.pg_error;
 399         if (err < 0)
 400                 goto out_err;
 401         return 0;
 402 out_err:
 403         return err;
 404 }
 405
 406 /*
 407  * Insert a write request into an inode
 408  */
 409 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 410 {
 411         struct nfs_inode *nfsi = NFS_I(inode);
 412
 413         /* Lock the request! */
 414         nfs_lock_request(req);
 415
 416         spin_lock(&inode->i_lock);
 417         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 418                 inode->i_version++;
 419         set_bit(PG_MAPPED, &req->wb_flags);
 420         SetPagePrivate(req->wb_page);
 421         set_page_private(req->wb_page, (unsigned long)req);
 422         nfsi->npages++;
 423         kref_get(&req->wb_kref);
 424         spin_unlock(&inode->i_lock);
 425 }
 426
 427 /*
 428  * Remove a write request from an inode
 429  */
 430 static void nfs_inode_remove_request(struct nfs_page *req)
 431 {
 432         struct inode *inode = req->wb_context->dentry->d_inode;
 433         struct nfs_inode *nfsi = NFS_I(inode);
 434
 435         BUG_ON (!NFS_WBACK_BUSY(req));
 436
 437         spin_lock(&inode->i_lock);
 438         set_page_private(req->wb_page, 0);
 439         ClearPagePrivate(req->wb_page);
 440         clear_bit(PG_MAPPED, &req->wb_flags);
 441         nfsi->npages--;
 442         spin_unlock(&inode->i_lock);
 443         nfs_release_request(req);
 444 }
 445
 446 static void
 447 nfs_mark_request_dirty(struct nfs_page *req)
 448 {
 449         __set_page_dirty_nobuffers(req->wb_page);
 450 }
 451
 452 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 453 /**
 454  * nfs_request_add_commit_list - add request to a commit list
 455  * @req: pointer to a struct nfs_page
 456  * @dst: commit list head
 457  * @cinfo: holds list lock and accounting info
 458  *
 459  * This sets the PG_CLEAN bit, updates the cinfo count of
 460  * number of outstanding requests requiring a commit as well as
 461  * the MM page stats.
 462  *
 463  * The caller must _not_ hold the cinfo->lock, but must be
 464  * holding the nfs_page lock.
 465  */
 466 void
 467 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 468                             struct nfs_commit_info *cinfo)
 469 {
 470         set_bit(PG_CLEAN, &(req)->wb_flags);
 471         spin_lock(cinfo->lock);
 472         nfs_list_add_request(req, dst);
 473         cinfo->mds->ncommit++;
 474         spin_unlock(cinfo->lock);
 475         if (!cinfo->dreq) {
 476                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 477                 inc_bdi_stat(req->wb_page->mapping->backing_dev_info,
 478                              BDI_RECLAIMABLE);
 479                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 480                                    I_DIRTY_DATASYNC);
 481         }
 482 }
 483 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 484
 485 /**
 486  * nfs_request_remove_commit_list - Remove request from a commit list
 487  * @req: pointer to a nfs_page
 488  * @cinfo: holds list lock and accounting info
 489  *
 490  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 491  * number of outstanding requests requiring a commit
 492  * It does not update the MM page stats.
 493  *
 494  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 495  */
 496 void
 497 nfs_request_remove_commit_list(struct nfs_page *req,
 498                                struct nfs_commit_info *cinfo)
 499 {
 500         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 501                 return;
 502         nfs_list_remove_request(req);
 503         cinfo->mds->ncommit--;
 504 }
 505 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 506
 507 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 508                                       struct inode *inode)
 509 {
 510         cinfo->lock = &inode->i_lock;
 511         cinfo->mds = &NFS_I(inode)->commit_info;
 512         cinfo->ds = pnfs_get_ds_info(inode);
 513         cinfo->dreq = NULL;
 514         cinfo->completion_ops = &nfs_commit_completion_ops;
 515 }
 516
 517 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 518                     struct inode *inode,
 519                     struct nfs_direct_req *dreq)
 520 {
 521         if (dreq)
 522                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 523         else
 524                 nfs_init_cinfo_from_inode(cinfo, inode);
 525 }
 526 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 527
 528 /*
 529  * Add a request to the inode's commit list.
 530  */
 531 void
 532 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 533                         struct nfs_commit_info *cinfo)
 534 {
 535         if (pnfs_mark_request_commit(req, lseg, cinfo))
 536                 return;
 537         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 538 }
 539
 540 static void
 541 nfs_clear_page_commit(struct page *page)
 542 {
 543         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 544         dec_bdi_stat(page->mapping->backing_dev_info, BDI_RECLAIMABLE);
 545 }
 546
 547 static void
 548 nfs_clear_request_commit(struct nfs_page *req)
 549 {
 550         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 551                 struct inode *inode = req->wb_context->dentry->d_inode;
 552                 struct nfs_commit_info cinfo;
 553
 554                 nfs_init_cinfo_from_inode(&cinfo, inode);
 555                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 556                         spin_lock(cinfo.lock);
 557                         nfs_request_remove_commit_list(req, &cinfo);
 558                         spin_unlock(cinfo.lock);
 559                 }
 560                 nfs_clear_page_commit(req->wb_page);
 561         }
 562 }
 563
 564 static inline
 565 int nfs_write_need_commit(struct nfs_write_data *data)
 566 {
 567         if (data->verf.committed == NFS_DATA_SYNC)
 568                 return data->header->lseg == NULL;
 569         return data->verf.committed != NFS_FILE_SYNC;
 570 }
 571
 572 #else
 573 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 574                                       struct inode *inode)
 575 {
 576 }
 577
 578 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 579                     struct inode *inode,
 580                     struct nfs_direct_req *dreq)
 581 {
 582 }
 583
 584 void
 585 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 586                         struct nfs_commit_info *cinfo)
 587 {
 588 }
 589
 590 static void
 591 nfs_clear_request_commit(struct nfs_page *req)
 592 {
 593 }
 594
 595 static inline
 596 int nfs_write_need_commit(struct nfs_write_data *data)
 597 {
 598         return 0;
 599 }
 600
 601 #endif
 602
 603 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 604 {
 605         struct nfs_commit_info cinfo;
 606         unsigned long bytes = 0;
 607
 608         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 609                 goto out;
 610         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 611         while (!list_empty(&hdr->pages)) {
 612                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 613
 614                 bytes += req->wb_bytes;
 615                 nfs_list_remove_request(req);
 616                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 617                     (hdr->good_bytes < bytes)) {
 618                         nfs_set_pageerror(req->wb_page);
 619                         nfs_context_set_write_error(req->wb_context, hdr->error);
 620                         goto remove_req;
 621                 }
 622                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 623                         nfs_mark_request_dirty(req);
 624                         goto next;
 625                 }
 626                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 627                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
 628                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 629                         goto next;
 630                 }
 631 remove_req:
 632                 nfs_inode_remove_request(req);
 633 next:
 634                 nfs_unlock_request(req);
 635                 nfs_end_page_writeback(req->wb_page);
 636                 nfs_release_request(req);
 637         }
 638 out:
 639         hdr->release(hdr);
 640 }
 641
 642 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 643 static unsigned long
 644 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 645 {
 646         return cinfo->mds->ncommit;
 647 }
 648
 649 /* cinfo->lock held by caller */
 650 int
 651 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 652                      struct nfs_commit_info *cinfo, int max)
 653 {
 654         struct nfs_page *req, *tmp;
 655         int ret = 0;
 656
 657         list_for_each_entry_safe(req, tmp, src, wb_list) {
 658                 if (!nfs_lock_request(req))
 659                         continue;
 660                 kref_get(&req->wb_kref);
 661                 if (cond_resched_lock(cinfo->lock))
 662                         list_safe_reset_next(req, tmp, wb_list);
 663                 nfs_request_remove_commit_list(req, cinfo);
 664                 nfs_list_add_request(req, dst);
 665                 ret++;
 666                 if ((ret == max) && !cinfo->dreq)
 667                         break;
 668         }
 669         return ret;
 670 }
 671
 672 /*
 673  * nfs_scan_commit - Scan an inode for commit requests
 674  * @inode: NFS inode to scan
 675  * @dst: mds destination list
 676  * @cinfo: mds and ds lists of reqs ready to commit
 677  *
 678  * Moves requests from the inode's 'commit' request list.
 679  * The requests are *not* checked to ensure that they form a contiguous set.
 680  */
 681 int
 682 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 683                 struct nfs_commit_info *cinfo)
 684 {
 685         int ret = 0;
 686
 687         spin_lock(cinfo->lock);
 688         if (cinfo->mds->ncommit > 0) {
 689                 const int max = INT_MAX;
 690
 691                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 692                                            cinfo, max);
 693                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 694         }
 695         spin_unlock(cinfo->lock);
 696         return ret;
 697 }
 698
 699 #else
 700 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 701 {
 702         return 0;
 703 }
 704
 705 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 706                     struct nfs_commit_info *cinfo)
 707 {
 708         return 0;
 709 }
 710 #endif
 711
 712 /*
 713  * Search for an existing write request, and attempt to update
 714  * it to reflect a new dirty region on a given page.
 715  *
 716  * If the attempt fails, then the existing request is flushed out
 717  * to disk.
 718  */
 719 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 720                 struct page *page,
 721                 unsigned int offset,
 722                 unsigned int bytes)
 723 {
 724         struct nfs_page *req;
 725         unsigned int rqend;
 726         unsigned int end;
 727         int error;
 728
 729         if (!PagePrivate(page))
 730                 return NULL;
 731
 732         end = offset + bytes;
 733         spin_lock(&inode->i_lock);
 734
 735         for (;;) {
 736                 req = nfs_page_find_request_locked(page);
 737                 if (req == NULL)
 738                         goto out_unlock;
 739
 740                 rqend = req->wb_offset + req->wb_bytes;
 741                 /*
 742                  * Tell the caller to flush out the request if
 743                  * the offsets are non-contiguous.
 744                  * Note: nfs_flush_incompatible() will already
 745                  * have flushed out requests having wrong owners.
 746                  */
 747                 if (offset > rqend
 748                     || end < req->wb_offset)
 749                         goto out_flushme;
 750
 751                 if (nfs_lock_request(req))
 752                         break;
 753
 754                 /* The request is locked, so wait and then retry */
 755                 spin_unlock(&inode->i_lock);
 756                 error = nfs_wait_on_request(req);
 757                 nfs_release_request(req);
 758                 if (error != 0)
 759                         goto out_err;
 760                 spin_lock(&inode->i_lock);
 761         }
 762
 763         /* Okay, the request matches. Update the region */
 764         if (offset < req->wb_offset) {
 765                 req->wb_offset = offset;
 766                 req->wb_pgbase = offset;
 767         }
 768         if (end > rqend)
 769                 req->wb_bytes = end - req->wb_offset;
 770         else
 771                 req->wb_bytes = rqend - req->wb_offset;
 772 out_unlock:
 773         spin_unlock(&inode->i_lock);
 774         if (req)
 775                 nfs_clear_request_commit(req);
 776         return req;
 777 out_flushme:
 778         spin_unlock(&inode->i_lock);
 779         nfs_release_request(req);
 780         error = nfs_wb_page(inode, page);
 781 out_err:
 782         return ERR_PTR(error);
 783 }
 784
 785 /*
 786  * Try to update an existing write request, or create one if there is none.
 787  *
 788  * Note: Should always be called with the Page Lock held to prevent races
 789  * if we have to add a new request. Also assumes that the caller has
 790  * already called nfs_flush_incompatible() if necessary.
 791  */
 792 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 793                 struct page *page, unsigned int offset, unsigned int bytes)
 794 {
 795         struct inode *inode = page->mapping->host;
 796         struct nfs_page *req;
 797
 798         req = nfs_try_to_update_request(inode, page, offset, bytes);
 799         if (req != NULL)
 800                 goto out;
 801         req = nfs_create_request(ctx, inode, page, offset, bytes);
 802         if (IS_ERR(req))
 803                 goto out;
 804         nfs_inode_add_request(inode, req);
 805 out:
 806         return req;
 807 }
 808
 809 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 810                 unsigned int offset, unsigned int count)
 811 {
 812         struct nfs_page *req;
 813
 814         req = nfs_setup_write_request(ctx, page, offset, count);
 815         if (IS_ERR(req))
 816                 return PTR_ERR(req);
 817         /* Update file length */
 818         nfs_grow_file(page, offset, count);
 819         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 820         nfs_mark_request_dirty(req);
 821         nfs_unlock_and_release_request(req);
 822         return 0;
 823 }
 824
 825 int nfs_flush_incompatible(struct file *file, struct page *page)
 826 {
 827         struct nfs_open_context *ctx = nfs_file_open_context(file);
 828         struct nfs_page *req;
 829         int do_flush, status;
 830         /*
 831          * Look for a request corresponding to this page. If there
 832          * is one, and it belongs to another file, we flush it out
 833          * before we try to copy anything into the page. Do this
 834          * due to the lack of an ACCESS-type call in NFSv2.
 835          * Also do the same if we find a request from an existing
 836          * dropped page.
 837          */
 838         do {
 839                 req = nfs_page_find_request(page);
 840                 if (req == NULL)
 841                         return 0;
 842                 do_flush = req->wb_page != page || req->wb_context != ctx ||
 843                         req->wb_lock_context->lockowner != current->files ||
 844                         req->wb_lock_context->pid != current->tgid;
 845                 nfs_release_request(req);
 846                 if (!do_flush)
 847                         return 0;
 848                 status = nfs_wb_page(page->mapping->host, page);
 849         } while (status == 0);
 850         return status;
 851 }
 852
 853 /*
 854  * If the page cache is marked as unsafe or invalid, then we can't rely on
 855  * the PageUptodate() flag. In this case, we will need to turn off
 856  * write optimisations that depend on the page contents being correct.
 857  */
 858 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 859 {
 860         if (nfs_have_delegated_attributes(inode))
 861                 goto out;
 862         if (NFS_I(inode)->cache_validity & NFS_INO_REVAL_PAGECACHE)
 863                 return false;
 864 out:
 865         return PageUptodate(page) != 0;
 866 }
 867
 868 /*
 869  * Update and possibly write a cached page of an NFS file.
 870  *
 871  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 872  * things with a page scheduled for an RPC call (e.g. invalidate it).
 873  */
 874 int nfs_updatepage(struct file *file, struct page *page,
 875                 unsigned int offset, unsigned int count)
 876 {
 877         struct nfs_open_context *ctx = nfs_file_open_context(file);
 878         struct inode    *inode = page->mapping->host;
 879         int             status = 0;
 880
 881         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 882
 883         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 884                 file->f_path.dentry->d_parent->d_name.name,
 885                 file->f_path.dentry->d_name.name, count,
 886                 (long long)(page_offset(page) + offset));
 887
 888         /* If we're not using byte range locks, and we know the page
 889          * is up to date, it may be more efficient to extend the write
 890          * to cover the entire page in order to avoid fragmentation
 891          * inefficiencies.
 892          */
 893         if (nfs_write_pageuptodate(page, inode) &&
 894                         inode->i_flock == NULL &&
 895                         !(file->f_flags & O_DSYNC)) {
 896                 count = max(count + offset, nfs_page_length(page));
 897                 offset = 0;
 898         }
 899
 900         status = nfs_writepage_setup(ctx, page, offset, count);
 901         if (status < 0)
 902                 nfs_set_pageerror(page);
 903         else
 904                 __set_page_dirty_nobuffers(page);
 905
 906         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 907                         status, (long long)i_size_read(inode));
 908         return status;
 909 }
 910
 911 static int flush_task_priority(int how)
 912 {
 913         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 914                 case FLUSH_HIGHPRI:
 915                         return RPC_PRIORITY_HIGH;
 916                 case FLUSH_LOWPRI:
 917                         return RPC_PRIORITY_LOW;
 918         }
 919         return RPC_PRIORITY_NORMAL;
 920 }
 921
 922 int nfs_initiate_write(struct rpc_clnt *clnt,
 923                        struct nfs_write_data *data,
 924                        const struct rpc_call_ops *call_ops,
 925                        int how, int flags)
 926 {
 927         struct inode *inode = data->header->inode;
 928         int priority = flush_task_priority(how);
 929         struct rpc_task *task;
 930         struct rpc_message msg = {
 931                 .rpc_argp = &data->args,
 932                 .rpc_resp = &data->res,
 933                 .rpc_cred = data->header->cred,
 934         };
 935         struct rpc_task_setup task_setup_data = {
 936                 .rpc_client = clnt,
 937                 .task = &data->task,
 938                 .rpc_message = &msg,
 939                 .callback_ops = call_ops,
 940                 .callback_data = data,
 941                 .workqueue = nfsiod_workqueue,
 942                 .flags = RPC_TASK_ASYNC | flags,
 943                 .priority = priority,
 944         };
 945         int ret = 0;
 946
 947         /* Set up the initial task struct.  */
 948         NFS_PROTO(inode)->write_setup(data, &msg);
 949
 950         dprintk("NFS: %5u initiated write call "
 951                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 952                 data->task.tk_pid,
 953                 inode->i_sb->s_id,
 954                 (long long)NFS_FILEID(inode),
 955                 data->args.count,
 956                 (unsigned long long)data->args.offset);
 957
 958         task = rpc_run_task(&task_setup_data);
 959         if (IS_ERR(task)) {
 960                 ret = PTR_ERR(task);
 961                 goto out;
 962         }
 963         if (how & FLUSH_SYNC) {
 964                 ret = rpc_wait_for_completion_task(task);
 965                 if (ret == 0)
 966                         ret = task->tk_status;
 967         }
 968         rpc_put_task(task);
 969 out:
 970         return ret;
 971 }
 972 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 973
 974 /*
 975  * Set up the argument/result storage required for the RPC call.
 976  */
 977 static void nfs_write_rpcsetup(struct nfs_write_data *data,
 978                 unsigned int count, unsigned int offset,
 979                 int how, struct nfs_commit_info *cinfo)
 980 {
 981         struct nfs_page *req = data->header->req;
 982
 983         /* Set up the RPC argument and reply structs
 984          * NB: take care not to mess about with data->commit et al. */
 985
 986         data->args.fh     = NFS_FH(data->header->inode);
 987         data->args.offset = req_offset(req) + offset;
 988         /* pnfs_set_layoutcommit needs this */
 989         data->mds_offset = data->args.offset;
 990         data->args.pgbase = req->wb_pgbase + offset;
 991         data->args.pages  = data->pages.pagevec;
 992         data->args.count  = count;
 993         data->args.context = get_nfs_open_context(req->wb_context);
 994         data->args.lock_context = req->wb_lock_context;
 995         data->args.stable  = NFS_UNSTABLE;
 996         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 997         case 0:
 998                 break;
 999         case FLUSH_COND_STABLE:
1000                 if (nfs_reqs_to_commit(cinfo))
1001                         break;
1002         default:
1003                 data->args.stable = NFS_FILE_SYNC;
1004         }
1005
1006         data->res.fattr   = &data->fattr;
1007         data->res.count   = count;
1008         data->res.verf    = &data->verf;
1009         nfs_fattr_init(&data->fattr);
1010 }
1011
1012 static int nfs_do_write(struct nfs_write_data *data,
1013                 const struct rpc_call_ops *call_ops,
1014                 int how)
1015 {
1016         struct inode *inode = data->header->inode;
1017
1018         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1019 }
1020
1021 static int nfs_do_multiple_writes(struct list_head *head,
1022                 const struct rpc_call_ops *call_ops,
1023                 int how)
1024 {
1025         struct nfs_write_data *data;
1026         int ret = 0;
1027
1028         while (!list_empty(head)) {
1029                 int ret2;
1030
1031                 data = list_first_entry(head, struct nfs_write_data, list);
1032                 list_del_init(&data->list);
1033
1034                 ret2 = nfs_do_write(data, call_ops, how);
1035                  if (ret == 0)
1036                          ret = ret2;
1037         }
1038         return ret;
1039 }
1040
1041 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1042  * call this on each, which will prepare them to be retried on next
1043  * writeback using standard nfs.
1044  */
1045 static void nfs_redirty_request(struct nfs_page *req)
1046 {
1047         nfs_mark_request_dirty(req);
1048         nfs_unlock_request(req);
1049         nfs_end_page_writeback(req->wb_page);
1050         nfs_release_request(req);
1051 }
1052
1053 static void nfs_async_write_error(struct list_head *head)
1054 {
1055         struct nfs_page *req;
1056
1057         while (!list_empty(head)) {
1058                 req = nfs_list_entry(head->next);
1059                 nfs_list_remove_request(req);
1060                 nfs_redirty_request(req);
1061         }
1062 }
1063
1064 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1065         .error_cleanup = nfs_async_write_error,
1066         .completion = nfs_write_completion,
1067 };
1068
1069 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1070                 struct nfs_pgio_header *hdr)
1071 {
1072         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1073         while (!list_empty(&hdr->rpc_list)) {
1074                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1075                                 struct nfs_write_data, list);
1076                 list_del(&data->list);
1077                 nfs_writedata_release(data);
1078         }
1079         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1080 }
1081
1082 /*
1083  * Generate multiple small requests to write out a single
1084  * contiguous dirty area on one page.
1085  */
1086 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1087                            struct nfs_pgio_header *hdr)
1088 {
1089         struct nfs_page *req = hdr->req;
1090         struct page *page = req->wb_page;
1091         struct nfs_write_data *data;
1092         size_t wsize = desc->pg_bsize, nbytes;
1093         unsigned int offset;
1094         int requests = 0;
1095         struct nfs_commit_info cinfo;
1096
1097         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1098
1099         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1100             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1101              desc->pg_count > wsize))
1102                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1103
1104
1105         offset = 0;
1106         nbytes = desc->pg_count;
1107         do {
1108                 size_t len = min(nbytes, wsize);
1109
1110                 data = nfs_writedata_alloc(hdr, 1);
1111                 if (!data) {
1112                         nfs_flush_error(desc, hdr);
1113                         return -ENOMEM;
1114                 }
1115                 data->pages.pagevec[0] = page;
1116                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1117                 list_add(&data->list, &hdr->rpc_list);
1118                 requests++;
1119                 nbytes -= len;
1120                 offset += len;
1121         } while (nbytes != 0);
1122         nfs_list_remove_request(req);
1123         nfs_list_add_request(req, &hdr->pages);
1124         desc->pg_rpc_callops = &nfs_write_common_ops;
1125         return 0;
1126 }
1127
1128 /*
1129  * Create an RPC task for the given write request and kick it.
1130  * The page must have been locked by the caller.
1131  *
1132  * It may happen that the page we're passed is not marked dirty.
1133  * This is the case if nfs_updatepage detects a conflicting request
1134  * that has been written but not committed.
1135  */
1136 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1137                          struct nfs_pgio_header *hdr)
1138 {
1139         struct nfs_page         *req;
1140         struct page             **pages;
1141         struct nfs_write_data   *data;
1142         struct list_head *head = &desc->pg_list;
1143         struct nfs_commit_info cinfo;
1144
1145         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1146                                                            desc->pg_count));
1147         if (!data) {
1148                 nfs_flush_error(desc, hdr);
1149                 return -ENOMEM;
1150         }
1151
1152         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1153         pages = data->pages.pagevec;
1154         while (!list_empty(head)) {
1155                 req = nfs_list_entry(head->next);
1156                 nfs_list_remove_request(req);
1157                 nfs_list_add_request(req, &hdr->pages);
1158                 *pages++ = req->wb_page;
1159         }
1160
1161         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1162             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1163                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1164
1165         /* Set up the argument struct */
1166         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1167         list_add(&data->list, &hdr->rpc_list);
1168         desc->pg_rpc_callops = &nfs_write_common_ops;
1169         return 0;
1170 }
1171
1172 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1173                       struct nfs_pgio_header *hdr)
1174 {
1175         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1176                 return nfs_flush_multi(desc, hdr);
1177         return nfs_flush_one(desc, hdr);
1178 }
1179 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1180
1181 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1182 {
1183         struct nfs_write_header *whdr;
1184         struct nfs_pgio_header *hdr;
1185         int ret;
1186
1187         whdr = nfs_writehdr_alloc();
1188         if (!whdr) {
1189                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1190                 return -ENOMEM;
1191         }
1192         hdr = &whdr->header;
1193         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1194         atomic_inc(&hdr->refcnt);
1195         ret = nfs_generic_flush(desc, hdr);
1196         if (ret == 0)
1197                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1198                                              desc->pg_rpc_callops,
1199                                              desc->pg_ioflags);
1200         if (atomic_dec_and_test(&hdr->refcnt))
1201                 hdr->completion_ops->completion(hdr);
1202         return ret;
1203 }
1204
1205 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1206         .pg_test = nfs_generic_pg_test,
1207         .pg_doio = nfs_generic_pg_writepages,
1208 };
1209
1210 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1211                                struct inode *inode, int ioflags,
1212                                const struct nfs_pgio_completion_ops *compl_ops)
1213 {
1214         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1215                                 NFS_SERVER(inode)->wsize, ioflags);
1216 }
1217 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1218
1219 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1220 {
1221         pgio->pg_ops = &nfs_pageio_write_ops;
1222         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1223 }
1224 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1225
1226
1227 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1228 {
1229         struct nfs_write_data *data = calldata;
1230         NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1231 }
1232
1233 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1234 {
1235         struct nfs_commit_data *data = calldata;
1236
1237         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1238 }
1239
1240 /*
1241  * Handle a write reply that flushes a whole page.
1242  *
1243  * FIXME: There is an inherent race with invalidate_inode_pages and
1244  *        writebacks since the page->count is kept > 1 for as long
1245  *        as the page has a write request pending.
1246  */
1247 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1248 {
1249         struct nfs_write_data   *data = calldata;
1250
1251         nfs_writeback_done(task, data);
1252 }
1253
1254 static void nfs_writeback_release_common(void *calldata)
1255 {
1256         struct nfs_write_data   *data = calldata;
1257         struct nfs_pgio_header *hdr = data->header;
1258         int status = data->task.tk_status;
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(hdr->verf, &data->verf, sizeof(*hdr->verf));
1266                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->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 IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(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 IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(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, &data->verf.verifier, sizeof(req->wb_verf))) {
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         return nfs_commit_unstable_pages(inode, wbc);
1680 }
1681 EXPORT_SYMBOL_GPL(nfs_write_inode);
1682
1683 /*
1684  * flush the inode to disk.
1685  */
1686 int nfs_wb_all(struct inode *inode)
1687 {
1688         struct writeback_control wbc = {
1689                 .sync_mode = WB_SYNC_ALL,
1690                 .nr_to_write = LONG_MAX,
1691                 .range_start = 0,
1692                 .range_end = LLONG_MAX,
1693         };
1694
1695         return sync_inode(inode, &wbc);
1696 }
1697 EXPORT_SYMBOL_GPL(nfs_wb_all);
1698
1699 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1700 {
1701         struct nfs_page *req;
1702         int ret = 0;
1703
1704         BUG_ON(!PageLocked(page));
1705         for (;;) {
1706                 wait_on_page_writeback(page);
1707                 req = nfs_page_find_request(page);
1708                 if (req == NULL)
1709                         break;
1710                 if (nfs_lock_request(req)) {
1711                         nfs_clear_request_commit(req);
1712                         nfs_inode_remove_request(req);
1713                         /*
1714                          * In case nfs_inode_remove_request has marked the
1715                          * page as being dirty
1716                          */
1717                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1718                         nfs_unlock_and_release_request(req);
1719                         break;
1720                 }
1721                 ret = nfs_wait_on_request(req);
1722                 nfs_release_request(req);
1723                 if (ret < 0)
1724                         break;
1725         }
1726         return ret;
1727 }
1728
1729 /*
1730  * Write back all requests on one page - we do this before reading it.
1731  */
1732 int nfs_wb_page(struct inode *inode, struct page *page)
1733 {
1734         loff_t range_start = page_offset(page);
1735         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1736         struct writeback_control wbc = {
1737                 .sync_mode = WB_SYNC_ALL,
1738                 .nr_to_write = 0,
1739                 .range_start = range_start,
1740                 .range_end = range_end,
1741         };
1742         int ret;
1743
1744         for (;;) {
1745                 wait_on_page_writeback(page);
1746                 if (clear_page_dirty_for_io(page)) {
1747                         ret = nfs_writepage_locked(page, &wbc);
1748                         if (ret < 0)
1749                                 goto out_error;
1750                         continue;
1751                 }
1752                 if (!PagePrivate(page))
1753                         break;
1754                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1755                 if (ret < 0)
1756                         goto out_error;
1757         }
1758         return 0;
1759 out_error:
1760         return ret;
1761 }
1762
1763 #ifdef CONFIG_MIGRATION
1764 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1765                 struct page *page, enum migrate_mode mode)
1766 {
1767         /*
1768          * If PagePrivate is set, then the page is currently associated with
1769          * an in-progress read or write request. Don't try to migrate it.
1770          *
1771          * FIXME: we could do this in principle, but we'll need a way to ensure
1772          *        that we can safely release the inode reference while holding
1773          *        the page lock.
1774          */
1775         if (PagePrivate(page))
1776                 return -EBUSY;
1777
1778         nfs_fscache_release_page(page, GFP_KERNEL);
1779
1780         return migrate_page(mapping, newpage, page, mode);
1781 }
1782 #endif
1783
1784 int __init nfs_init_writepagecache(void)
1785 {
1786         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1787                                              sizeof(struct nfs_write_header),
1788                                              0, SLAB_HWCACHE_ALIGN,
1789                                              NULL);
1790         if (nfs_wdata_cachep == NULL)
1791                 return -ENOMEM;
1792
1793         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1794                                                      nfs_wdata_cachep);
1795         if (nfs_wdata_mempool == NULL)
1796                 return -ENOMEM;
1797
1798         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1799                                              sizeof(struct nfs_commit_data),
1800                                              0, SLAB_HWCACHE_ALIGN,
1801                                              NULL);
1802         if (nfs_cdata_cachep == NULL)
1803                 return -ENOMEM;
1804
1805         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1806                                                       nfs_wdata_cachep);
1807         if (nfs_commit_mempool == NULL)
1808                 return -ENOMEM;
1809
1810         /*
1811          * NFS congestion size, scale with available memory.
1812          *
1813          *  64MB:    8192k
1814          * 128MB:   11585k
1815          * 256MB:   16384k
1816          * 512MB:   23170k
1817          *   1GB:   32768k
1818          *   2GB:   46340k
1819          *   4GB:   65536k
1820          *   8GB:   92681k
1821          *  16GB:  131072k
1822          *
1823          * This allows larger machines to have larger/more transfers.
1824          * Limit the default to 256M
1825          */
1826         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1827         if (nfs_congestion_kb > 256*1024)
1828                 nfs_congestion_kb = 256*1024;
1829
1830         return 0;
1831 }
1832
1833 void nfs_destroy_writepagecache(void)
1834 {
1835         mempool_destroy(nfs_commit_mempool);
1836         mempool_destroy(nfs_wdata_mempool);
1837         kmem_cache_destroy(nfs_wdata_cachep);
1838 }
1839