fs/nfs/write.c

   1 /*
   2  * linux/fs/nfs/write.c
   3  *
   4  * Write file data over NFS.
   5  *
   6  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/slab.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/file.h>
  14 #include <linux/writeback.h>
  15 #include <linux/swap.h>
  16 #include <linux/migrate.h>
  17
  18 #include <linux/sunrpc/clnt.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/export.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #include "delegation.h"
  28 #include "internal.h"
  29 #include "iostat.h"
  30 #include "nfs4_fs.h"
  31 #include "fscache.h"
  32 #include "pnfs.h"
  33
  34 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  35
  36 #define MIN_POOL_WRITE          (32)
  37 #define MIN_POOL_COMMIT         (4)
  38
  39 /*
  40  * Local function declarations
  41  */
  42 static void nfs_redirty_request(struct nfs_page *req);
  43 static const struct rpc_call_ops nfs_write_common_ops;
  44 static const struct rpc_call_ops nfs_commit_ops;
  45 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops;
  46 static const struct nfs_commit_completion_ops nfs_commit_completion_ops;
  47
  48 static struct kmem_cache *nfs_wdata_cachep;
  49 static mempool_t *nfs_wdata_mempool;
  50 static struct kmem_cache *nfs_cdata_cachep;
  51 static mempool_t *nfs_commit_mempool;
  52
  53 struct nfs_commit_data *nfs_commitdata_alloc(void)
  54 {
  55         struct nfs_commit_data *p = mempool_alloc(nfs_commit_mempool, GFP_NOIO);
  56
  57         if (p) {
  58                 memset(p, 0, sizeof(*p));
  59                 INIT_LIST_HEAD(&p->pages);
  60         }
  61         return p;
  62 }
  63 EXPORT_SYMBOL_GPL(nfs_commitdata_alloc);
  64
  65 void nfs_commit_free(struct nfs_commit_data *p)
  66 {
  67         mempool_free(p, nfs_commit_mempool);
  68 }
  69 EXPORT_SYMBOL_GPL(nfs_commit_free);
  70
  71 struct nfs_write_header *nfs_writehdr_alloc(void)
  72 {
  73         struct nfs_write_header *p = mempool_alloc(nfs_wdata_mempool, GFP_NOIO);
  74
  75         if (p) {
  76                 struct nfs_pgio_header *hdr = &p->header;
  77
  78                 memset(p, 0, sizeof(*p));
  79                 INIT_LIST_HEAD(&hdr->pages);
  80                 INIT_LIST_HEAD(&hdr->rpc_list);
  81                 spin_lock_init(&hdr->lock);
  82                 atomic_set(&hdr->refcnt, 0);
  83                 hdr->verf = &p->verf;
  84         }
  85         return p;
  86 }
  87 EXPORT_SYMBOL_GPL(nfs_writehdr_alloc);
  88
  89 static struct nfs_write_data *nfs_writedata_alloc(struct nfs_pgio_header *hdr,
  90                                                   unsigned int pagecount)
  91 {
  92         struct nfs_write_data *data, *prealloc;
  93
  94         prealloc = &container_of(hdr, struct nfs_write_header, header)->rpc_data;
  95         if (prealloc->header == NULL)
  96                 data = prealloc;
  97         else
  98                 data = kzalloc(sizeof(*data), GFP_KERNEL);
  99         if (!data)
 100                 goto out;
 101
 102         if (nfs_pgarray_set(&data->pages, pagecount)) {
 103                 data->header = hdr;
 104                 atomic_inc(&hdr->refcnt);
 105         } else {
 106                 if (data != prealloc)
 107                         kfree(data);
 108                 data = NULL;
 109         }
 110 out:
 111         return data;
 112 }
 113
 114 void nfs_writehdr_free(struct nfs_pgio_header *hdr)
 115 {
 116         struct nfs_write_header *whdr = container_of(hdr, struct nfs_write_header, header);
 117         mempool_free(whdr, nfs_wdata_mempool);
 118 }
 119 EXPORT_SYMBOL_GPL(nfs_writehdr_free);
 120
 121 void nfs_writedata_release(struct nfs_write_data *wdata)
 122 {
 123         struct nfs_pgio_header *hdr = wdata->header;
 124         struct nfs_write_header *write_header = container_of(hdr, struct nfs_write_header, header);
 125
 126         put_nfs_open_context(wdata->args.context);
 127         if (wdata->pages.pagevec != wdata->pages.page_array)
 128                 kfree(wdata->pages.pagevec);
 129         if (wdata != &write_header->rpc_data)
 130                 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 *
 146 nfs_page_find_request_locked(struct nfs_inode *nfsi, struct page *page)
 147 {
 148         struct nfs_page *req = NULL;
 149
 150         if (PagePrivate(page))
 151                 req = (struct nfs_page *)page_private(page);
 152         else if (unlikely(PageSwapCache(page))) {
 153                 struct nfs_page *freq, *t;
 154
 155                 /* Linearly search the commit list for the correct req */
 156                 list_for_each_entry_safe(freq, t, &nfsi->commit_info.list, wb_list) {
 157                         if (freq->wb_page == page) {
 158                                 req = freq;
 159                                 break;
 160                         }
 161                 }
 162         }
 163
 164         if (req)
 165                 kref_get(&req->wb_kref);
 166
 167         return req;
 168 }
 169
 170 static struct nfs_page *nfs_page_find_request(struct page *page)
 171 {
 172         struct inode *inode = page_file_mapping(page)->host;
 173         struct nfs_page *req = NULL;
 174
 175         spin_lock(&inode->i_lock);
 176         req = nfs_page_find_request_locked(NFS_I(inode), page);
 177         spin_unlock(&inode->i_lock);
 178         return req;
 179 }
 180
 181 /* Adjust the file length if we're writing beyond the end */
 182 static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count)
 183 {
 184         struct inode *inode = page_file_mapping(page)->host;
 185         loff_t end, i_size;
 186         pgoff_t end_index;
 187
 188         spin_lock(&inode->i_lock);
 189         i_size = i_size_read(inode);
 190         end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
 191         if (i_size > 0 && page_file_index(page) < end_index)
 192                 goto out;
 193         end = page_file_offset(page) + ((loff_t)offset+count);
 194         if (i_size >= end)
 195                 goto out;
 196         i_size_write(inode, end);
 197         nfs_inc_stats(inode, NFSIOS_EXTENDWRITE);
 198 out:
 199         spin_unlock(&inode->i_lock);
 200 }
 201
 202 /* A writeback failed: mark the page as bad, and invalidate the page cache */
 203 static void nfs_set_pageerror(struct page *page)
 204 {
 205         nfs_zap_mapping(page_file_mapping(page)->host, page_file_mapping(page));
 206 }
 207
 208 /* We can set the PG_uptodate flag if we see that a write request
 209  * covers the full page.
 210  */
 211 static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count)
 212 {
 213         if (PageUptodate(page))
 214                 return;
 215         if (base != 0)
 216                 return;
 217         if (count != nfs_page_length(page))
 218                 return;
 219         SetPageUptodate(page);
 220 }
 221
 222 static int wb_priority(struct writeback_control *wbc)
 223 {
 224         if (wbc->for_reclaim)
 225                 return FLUSH_HIGHPRI | FLUSH_STABLE;
 226         if (wbc->for_kupdate || wbc->for_background)
 227                 return FLUSH_LOWPRI | FLUSH_COND_STABLE;
 228         return FLUSH_COND_STABLE;
 229 }
 230
 231 /*
 232  * NFS congestion control
 233  */
 234
 235 int nfs_congestion_kb;
 236
 237 #define NFS_CONGESTION_ON_THRESH        (nfs_congestion_kb >> (PAGE_SHIFT-10))
 238 #define NFS_CONGESTION_OFF_THRESH       \
 239         (NFS_CONGESTION_ON_THRESH - (NFS_CONGESTION_ON_THRESH >> 2))
 240
 241 static void nfs_set_page_writeback(struct page *page)
 242 {
 243         struct nfs_server *nfss = NFS_SERVER(page_file_mapping(page)->host);
 244         int ret = test_set_page_writeback(page);
 245
 246         WARN_ON_ONCE(ret != 0);
 247
 248         if (atomic_long_inc_return(&nfss->writeback) >
 249                         NFS_CONGESTION_ON_THRESH) {
 250                 set_bdi_congested(&nfss->backing_dev_info,
 251                                         BLK_RW_ASYNC);
 252         }
 253 }
 254
 255 static void nfs_end_page_writeback(struct page *page)
 256 {
 257         struct inode *inode = page_file_mapping(page)->host;
 258         struct nfs_server *nfss = NFS_SERVER(inode);
 259
 260         end_page_writeback(page);
 261         if (atomic_long_dec_return(&nfss->writeback) < NFS_CONGESTION_OFF_THRESH)
 262                 clear_bdi_congested(&nfss->backing_dev_info, BLK_RW_ASYNC);
 263 }
 264
 265 static struct nfs_page *nfs_find_and_lock_request(struct page *page, bool nonblock)
 266 {
 267         struct inode *inode = page_file_mapping(page)->host;
 268         struct nfs_page *req;
 269         int ret;
 270
 271         spin_lock(&inode->i_lock);
 272         for (;;) {
 273                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 274                 if (req == NULL)
 275                         break;
 276                 if (nfs_lock_request(req))
 277                         break;
 278                 /* Note: If we hold the page lock, as is the case in nfs_writepage,
 279                  *       then the call to nfs_lock_request() will always
 280                  *       succeed provided that someone hasn't already marked the
 281                  *       request as dirty (in which case we don't care).
 282                  */
 283                 spin_unlock(&inode->i_lock);
 284                 if (!nonblock)
 285                         ret = nfs_wait_on_request(req);
 286                 else
 287                         ret = -EAGAIN;
 288                 nfs_release_request(req);
 289                 if (ret != 0)
 290                         return ERR_PTR(ret);
 291                 spin_lock(&inode->i_lock);
 292         }
 293         spin_unlock(&inode->i_lock);
 294         return req;
 295 }
 296
 297 /*
 298  * Find an associated nfs write request, and prepare to flush it out
 299  * May return an error if the user signalled nfs_wait_on_request().
 300  */
 301 static int nfs_page_async_flush(struct nfs_pageio_descriptor *pgio,
 302                                 struct page *page, bool nonblock)
 303 {
 304         struct nfs_page *req;
 305         int ret = 0;
 306
 307         req = nfs_find_and_lock_request(page, nonblock);
 308         if (!req)
 309                 goto out;
 310         ret = PTR_ERR(req);
 311         if (IS_ERR(req))
 312                 goto out;
 313
 314         nfs_set_page_writeback(page);
 315         WARN_ON_ONCE(test_bit(PG_CLEAN, &req->wb_flags));
 316
 317         ret = 0;
 318         if (!nfs_pageio_add_request(pgio, req)) {
 319                 nfs_redirty_request(req);
 320                 ret = pgio->pg_error;
 321         }
 322 out:
 323         return ret;
 324 }
 325
 326 static int nfs_do_writepage(struct page *page, struct writeback_control *wbc, struct nfs_pageio_descriptor *pgio)
 327 {
 328         struct inode *inode = page_file_mapping(page)->host;
 329         int ret;
 330
 331         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE);
 332         nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1);
 333
 334         nfs_pageio_cond_complete(pgio, page_file_index(page));
 335         ret = nfs_page_async_flush(pgio, page, wbc->sync_mode == WB_SYNC_NONE);
 336         if (ret == -EAGAIN) {
 337                 redirty_page_for_writepage(wbc, page);
 338                 ret = 0;
 339         }
 340         return ret;
 341 }
 342
 343 /*
 344  * Write an mmapped page to the server.
 345  */
 346 static int nfs_writepage_locked(struct page *page, struct writeback_control *wbc)
 347 {
 348         struct nfs_pageio_descriptor pgio;
 349         int err;
 350
 351         NFS_PROTO(page_file_mapping(page)->host)->write_pageio_init(&pgio,
 352                                                           page->mapping->host,
 353                                                           wb_priority(wbc),
 354                                                           &nfs_async_write_completion_ops);
 355         err = nfs_do_writepage(page, wbc, &pgio);
 356         nfs_pageio_complete(&pgio);
 357         if (err < 0)
 358                 return err;
 359         if (pgio.pg_error < 0)
 360                 return pgio.pg_error;
 361         return 0;
 362 }
 363
 364 int nfs_writepage(struct page *page, struct writeback_control *wbc)
 365 {
 366         int ret;
 367
 368         ret = nfs_writepage_locked(page, wbc);
 369         unlock_page(page);
 370         return ret;
 371 }
 372
 373 static int nfs_writepages_callback(struct page *page, struct writeback_control *wbc, void *data)
 374 {
 375         int ret;
 376
 377         ret = nfs_do_writepage(page, wbc, data);
 378         unlock_page(page);
 379         return ret;
 380 }
 381
 382 int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
 383 {
 384         struct inode *inode = mapping->host;
 385         unsigned long *bitlock = &NFS_I(inode)->flags;
 386         struct nfs_pageio_descriptor pgio;
 387         int err;
 388
 389         /* Stop dirtying of new pages while we sync */
 390         err = wait_on_bit_lock(bitlock, NFS_INO_FLUSHING,
 391                         nfs_wait_bit_killable, TASK_KILLABLE);
 392         if (err)
 393                 goto out_err;
 394
 395         nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
 396
 397         NFS_PROTO(inode)->write_pageio_init(&pgio, inode, wb_priority(wbc), &nfs_async_write_completion_ops);
 398         err = write_cache_pages(mapping, wbc, nfs_writepages_callback, &pgio);
 399         nfs_pageio_complete(&pgio);
 400
 401         clear_bit_unlock(NFS_INO_FLUSHING, bitlock);
 402         smp_mb__after_clear_bit();
 403         wake_up_bit(bitlock, NFS_INO_FLUSHING);
 404
 405         if (err < 0)
 406                 goto out_err;
 407         err = pgio.pg_error;
 408         if (err < 0)
 409                 goto out_err;
 410         return 0;
 411 out_err:
 412         return err;
 413 }
 414
 415 /*
 416  * Insert a write request into an inode
 417  */
 418 static void nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
 419 {
 420         struct nfs_inode *nfsi = NFS_I(inode);
 421
 422         /* Lock the request! */
 423         nfs_lock_request(req);
 424
 425         spin_lock(&inode->i_lock);
 426         if (!nfsi->npages && NFS_PROTO(inode)->have_delegation(inode, FMODE_WRITE))
 427                 inode->i_version++;
 428         /*
 429          * Swap-space should not get truncated. Hence no need to plug the race
 430          * with invalidate/truncate.
 431          */
 432         if (likely(!PageSwapCache(req->wb_page))) {
 433                 set_bit(PG_MAPPED, &req->wb_flags);
 434                 SetPagePrivate(req->wb_page);
 435                 set_page_private(req->wb_page, (unsigned long)req);
 436         }
 437         nfsi->npages++;
 438         kref_get(&req->wb_kref);
 439         spin_unlock(&inode->i_lock);
 440 }
 441
 442 /*
 443  * Remove a write request from an inode
 444  */
 445 static void nfs_inode_remove_request(struct nfs_page *req)
 446 {
 447         struct inode *inode = req->wb_context->dentry->d_inode;
 448         struct nfs_inode *nfsi = NFS_I(inode);
 449
 450         spin_lock(&inode->i_lock);
 451         if (likely(!PageSwapCache(req->wb_page))) {
 452                 set_page_private(req->wb_page, 0);
 453                 ClearPagePrivate(req->wb_page);
 454                 clear_bit(PG_MAPPED, &req->wb_flags);
 455         }
 456         nfsi->npages--;
 457         spin_unlock(&inode->i_lock);
 458         nfs_release_request(req);
 459 }
 460
 461 static void
 462 nfs_mark_request_dirty(struct nfs_page *req)
 463 {
 464         __set_page_dirty_nobuffers(req->wb_page);
 465 }
 466
 467 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 468 /**
 469  * nfs_request_add_commit_list - add request to a commit list
 470  * @req: pointer to a struct nfs_page
 471  * @dst: commit list head
 472  * @cinfo: holds list lock and accounting info
 473  *
 474  * This sets the PG_CLEAN bit, updates the cinfo count of
 475  * number of outstanding requests requiring a commit as well as
 476  * the MM page stats.
 477  *
 478  * The caller must _not_ hold the cinfo->lock, but must be
 479  * holding the nfs_page lock.
 480  */
 481 void
 482 nfs_request_add_commit_list(struct nfs_page *req, struct list_head *dst,
 483                             struct nfs_commit_info *cinfo)
 484 {
 485         set_bit(PG_CLEAN, &(req)->wb_flags);
 486         spin_lock(cinfo->lock);
 487         nfs_list_add_request(req, dst);
 488         cinfo->mds->ncommit++;
 489         spin_unlock(cinfo->lock);
 490         if (!cinfo->dreq) {
 491                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
 492                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
 493                              BDI_RECLAIMABLE);
 494                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
 495                                    I_DIRTY_DATASYNC);
 496         }
 497 }
 498 EXPORT_SYMBOL_GPL(nfs_request_add_commit_list);
 499
 500 /**
 501  * nfs_request_remove_commit_list - Remove request from a commit list
 502  * @req: pointer to a nfs_page
 503  * @cinfo: holds list lock and accounting info
 504  *
 505  * This clears the PG_CLEAN bit, and updates the cinfo's count of
 506  * number of outstanding requests requiring a commit
 507  * It does not update the MM page stats.
 508  *
 509  * The caller _must_ hold the cinfo->lock and the nfs_page lock.
 510  */
 511 void
 512 nfs_request_remove_commit_list(struct nfs_page *req,
 513                                struct nfs_commit_info *cinfo)
 514 {
 515         if (!test_and_clear_bit(PG_CLEAN, &(req)->wb_flags))
 516                 return;
 517         nfs_list_remove_request(req);
 518         cinfo->mds->ncommit--;
 519 }
 520 EXPORT_SYMBOL_GPL(nfs_request_remove_commit_list);
 521
 522 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 523                                       struct inode *inode)
 524 {
 525         cinfo->lock = &inode->i_lock;
 526         cinfo->mds = &NFS_I(inode)->commit_info;
 527         cinfo->ds = pnfs_get_ds_info(inode);
 528         cinfo->dreq = NULL;
 529         cinfo->completion_ops = &nfs_commit_completion_ops;
 530 }
 531
 532 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 533                     struct inode *inode,
 534                     struct nfs_direct_req *dreq)
 535 {
 536         if (dreq)
 537                 nfs_init_cinfo_from_dreq(cinfo, dreq);
 538         else
 539                 nfs_init_cinfo_from_inode(cinfo, inode);
 540 }
 541 EXPORT_SYMBOL_GPL(nfs_init_cinfo);
 542
 543 /*
 544  * Add a request to the inode's commit list.
 545  */
 546 void
 547 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 548                         struct nfs_commit_info *cinfo)
 549 {
 550         if (pnfs_mark_request_commit(req, lseg, cinfo))
 551                 return;
 552         nfs_request_add_commit_list(req, &cinfo->mds->list, cinfo);
 553 }
 554
 555 static void
 556 nfs_clear_page_commit(struct page *page)
 557 {
 558         dec_zone_page_state(page, NR_UNSTABLE_NFS);
 559         dec_bdi_stat(page_file_mapping(page)->backing_dev_info, BDI_RECLAIMABLE);
 560 }
 561
 562 static void
 563 nfs_clear_request_commit(struct nfs_page *req)
 564 {
 565         if (test_bit(PG_CLEAN, &req->wb_flags)) {
 566                 struct inode *inode = req->wb_context->dentry->d_inode;
 567                 struct nfs_commit_info cinfo;
 568
 569                 nfs_init_cinfo_from_inode(&cinfo, inode);
 570                 if (!pnfs_clear_request_commit(req, &cinfo)) {
 571                         spin_lock(cinfo.lock);
 572                         nfs_request_remove_commit_list(req, &cinfo);
 573                         spin_unlock(cinfo.lock);
 574                 }
 575                 nfs_clear_page_commit(req->wb_page);
 576         }
 577 }
 578
 579 static inline
 580 int nfs_write_need_commit(struct nfs_write_data *data)
 581 {
 582         if (data->verf.committed == NFS_DATA_SYNC)
 583                 return data->header->lseg == NULL;
 584         return data->verf.committed != NFS_FILE_SYNC;
 585 }
 586
 587 #else
 588 static void nfs_init_cinfo_from_inode(struct nfs_commit_info *cinfo,
 589                                       struct inode *inode)
 590 {
 591 }
 592
 593 void nfs_init_cinfo(struct nfs_commit_info *cinfo,
 594                     struct inode *inode,
 595                     struct nfs_direct_req *dreq)
 596 {
 597 }
 598
 599 void
 600 nfs_mark_request_commit(struct nfs_page *req, struct pnfs_layout_segment *lseg,
 601                         struct nfs_commit_info *cinfo)
 602 {
 603 }
 604
 605 static void
 606 nfs_clear_request_commit(struct nfs_page *req)
 607 {
 608 }
 609
 610 static inline
 611 int nfs_write_need_commit(struct nfs_write_data *data)
 612 {
 613         return 0;
 614 }
 615
 616 #endif
 617
 618 static void nfs_write_completion(struct nfs_pgio_header *hdr)
 619 {
 620         struct nfs_commit_info cinfo;
 621         unsigned long bytes = 0;
 622
 623         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 624                 goto out;
 625         nfs_init_cinfo_from_inode(&cinfo, hdr->inode);
 626         while (!list_empty(&hdr->pages)) {
 627                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 628
 629                 bytes += req->wb_bytes;
 630                 nfs_list_remove_request(req);
 631                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) &&
 632                     (hdr->good_bytes < bytes)) {
 633                         nfs_set_pageerror(req->wb_page);
 634                         nfs_context_set_write_error(req->wb_context, hdr->error);
 635                         goto remove_req;
 636                 }
 637                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 638                         nfs_mark_request_dirty(req);
 639                         goto next;
 640                 }
 641                 if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 642                         memcpy(&req->wb_verf, &hdr->verf->verifier, sizeof(req->wb_verf));
 643                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 644                         goto next;
 645                 }
 646 remove_req:
 647                 nfs_inode_remove_request(req);
 648 next:
 649                 nfs_unlock_request(req);
 650                 nfs_end_page_writeback(req->wb_page);
 651                 nfs_release_request(req);
 652         }
 653 out:
 654         hdr->release(hdr);
 655 }
 656
 657 #if  IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 658 static unsigned long
 659 nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 660 {
 661         return cinfo->mds->ncommit;
 662 }
 663
 664 /* cinfo->lock held by caller */
 665 int
 666 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
 667                      struct nfs_commit_info *cinfo, int max)
 668 {
 669         struct nfs_page *req, *tmp;
 670         int ret = 0;
 671
 672         list_for_each_entry_safe(req, tmp, src, wb_list) {
 673                 if (!nfs_lock_request(req))
 674                         continue;
 675                 kref_get(&req->wb_kref);
 676                 if (cond_resched_lock(cinfo->lock))
 677                         list_safe_reset_next(req, tmp, wb_list);
 678                 nfs_request_remove_commit_list(req, cinfo);
 679                 nfs_list_add_request(req, dst);
 680                 ret++;
 681                 if ((ret == max) && !cinfo->dreq)
 682                         break;
 683         }
 684         return ret;
 685 }
 686
 687 /*
 688  * nfs_scan_commit - Scan an inode for commit requests
 689  * @inode: NFS inode to scan
 690  * @dst: mds destination list
 691  * @cinfo: mds and ds lists of reqs ready to commit
 692  *
 693  * Moves requests from the inode's 'commit' request list.
 694  * The requests are *not* checked to ensure that they form a contiguous set.
 695  */
 696 int
 697 nfs_scan_commit(struct inode *inode, struct list_head *dst,
 698                 struct nfs_commit_info *cinfo)
 699 {
 700         int ret = 0;
 701
 702         spin_lock(cinfo->lock);
 703         if (cinfo->mds->ncommit > 0) {
 704                 const int max = INT_MAX;
 705
 706                 ret = nfs_scan_commit_list(&cinfo->mds->list, dst,
 707                                            cinfo, max);
 708                 ret += pnfs_scan_commit_lists(inode, cinfo, max - ret);
 709         }
 710         spin_unlock(cinfo->lock);
 711         return ret;
 712 }
 713
 714 #else
 715 static unsigned long nfs_reqs_to_commit(struct nfs_commit_info *cinfo)
 716 {
 717         return 0;
 718 }
 719
 720 int nfs_scan_commit(struct inode *inode, struct list_head *dst,
 721                     struct nfs_commit_info *cinfo)
 722 {
 723         return 0;
 724 }
 725 #endif
 726
 727 /*
 728  * Search for an existing write request, and attempt to update
 729  * it to reflect a new dirty region on a given page.
 730  *
 731  * If the attempt fails, then the existing request is flushed out
 732  * to disk.
 733  */
 734 static struct nfs_page *nfs_try_to_update_request(struct inode *inode,
 735                 struct page *page,
 736                 unsigned int offset,
 737                 unsigned int bytes)
 738 {
 739         struct nfs_page *req;
 740         unsigned int rqend;
 741         unsigned int end;
 742         int error;
 743
 744         if (!PagePrivate(page))
 745                 return NULL;
 746
 747         end = offset + bytes;
 748         spin_lock(&inode->i_lock);
 749
 750         for (;;) {
 751                 req = nfs_page_find_request_locked(NFS_I(inode), page);
 752                 if (req == NULL)
 753                         goto out_unlock;
 754
 755                 rqend = req->wb_offset + req->wb_bytes;
 756                 /*
 757                  * Tell the caller to flush out the request if
 758                  * the offsets are non-contiguous.
 759                  * Note: nfs_flush_incompatible() will already
 760                  * have flushed out requests having wrong owners.
 761                  */
 762                 if (offset > rqend
 763                     || end < req->wb_offset)
 764                         goto out_flushme;
 765
 766                 if (nfs_lock_request(req))
 767                         break;
 768
 769                 /* The request is locked, so wait and then retry */
 770                 spin_unlock(&inode->i_lock);
 771                 error = nfs_wait_on_request(req);
 772                 nfs_release_request(req);
 773                 if (error != 0)
 774                         goto out_err;
 775                 spin_lock(&inode->i_lock);
 776         }
 777
 778         /* Okay, the request matches. Update the region */
 779         if (offset < req->wb_offset) {
 780                 req->wb_offset = offset;
 781                 req->wb_pgbase = offset;
 782         }
 783         if (end > rqend)
 784                 req->wb_bytes = end - req->wb_offset;
 785         else
 786                 req->wb_bytes = rqend - req->wb_offset;
 787 out_unlock:
 788         spin_unlock(&inode->i_lock);
 789         if (req)
 790                 nfs_clear_request_commit(req);
 791         return req;
 792 out_flushme:
 793         spin_unlock(&inode->i_lock);
 794         nfs_release_request(req);
 795         error = nfs_wb_page(inode, page);
 796 out_err:
 797         return ERR_PTR(error);
 798 }
 799
 800 /*
 801  * Try to update an existing write request, or create one if there is none.
 802  *
 803  * Note: Should always be called with the Page Lock held to prevent races
 804  * if we have to add a new request. Also assumes that the caller has
 805  * already called nfs_flush_incompatible() if necessary.
 806  */
 807 static struct nfs_page * nfs_setup_write_request(struct nfs_open_context* ctx,
 808                 struct page *page, unsigned int offset, unsigned int bytes)
 809 {
 810         struct inode *inode = page_file_mapping(page)->host;
 811         struct nfs_page *req;
 812
 813         req = nfs_try_to_update_request(inode, page, offset, bytes);
 814         if (req != NULL)
 815                 goto out;
 816         req = nfs_create_request(ctx, inode, page, offset, bytes);
 817         if (IS_ERR(req))
 818                 goto out;
 819         nfs_inode_add_request(inode, req);
 820 out:
 821         return req;
 822 }
 823
 824 static int nfs_writepage_setup(struct nfs_open_context *ctx, struct page *page,
 825                 unsigned int offset, unsigned int count)
 826 {
 827         struct nfs_page *req;
 828
 829         req = nfs_setup_write_request(ctx, page, offset, count);
 830         if (IS_ERR(req))
 831                 return PTR_ERR(req);
 832         /* Update file length */
 833         nfs_grow_file(page, offset, count);
 834         nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes);
 835         nfs_mark_request_dirty(req);
 836         nfs_unlock_and_release_request(req);
 837         return 0;
 838 }
 839
 840 int nfs_flush_incompatible(struct file *file, struct page *page)
 841 {
 842         struct nfs_open_context *ctx = nfs_file_open_context(file);
 843         struct nfs_lock_context *l_ctx;
 844         struct nfs_page *req;
 845         int do_flush, status;
 846         /*
 847          * Look for a request corresponding to this page. If there
 848          * is one, and it belongs to another file, we flush it out
 849          * before we try to copy anything into the page. Do this
 850          * due to the lack of an ACCESS-type call in NFSv2.
 851          * Also do the same if we find a request from an existing
 852          * dropped page.
 853          */
 854         do {
 855                 req = nfs_page_find_request(page);
 856                 if (req == NULL)
 857                         return 0;
 858                 l_ctx = req->wb_lock_context;
 859                 do_flush = req->wb_page != page || req->wb_context != ctx;
 860                 if (l_ctx) {
 861                         do_flush |= l_ctx->lockowner.l_owner != current->files
 862                                 || l_ctx->lockowner.l_pid != current->tgid;
 863                 }
 864                 nfs_release_request(req);
 865                 if (!do_flush)
 866                         return 0;
 867                 status = nfs_wb_page(page_file_mapping(page)->host, page);
 868         } while (status == 0);
 869         return status;
 870 }
 871
 872 /*
 873  * If the page cache is marked as unsafe or invalid, then we can't rely on
 874  * the PageUptodate() flag. In this case, we will need to turn off
 875  * write optimisations that depend on the page contents being correct.
 876  */
 877 static bool nfs_write_pageuptodate(struct page *page, struct inode *inode)
 878 {
 879         if (nfs_have_delegated_attributes(inode))
 880                 goto out;
 881         if (NFS_I(inode)->cache_validity & (NFS_INO_INVALID_DATA|NFS_INO_REVAL_PAGECACHE))
 882                 return false;
 883 out:
 884         return PageUptodate(page) != 0;
 885 }
 886
 887 /*
 888  * Update and possibly write a cached page of an NFS file.
 889  *
 890  * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad
 891  * things with a page scheduled for an RPC call (e.g. invalidate it).
 892  */
 893 int nfs_updatepage(struct file *file, struct page *page,
 894                 unsigned int offset, unsigned int count)
 895 {
 896         struct nfs_open_context *ctx = nfs_file_open_context(file);
 897         struct inode    *inode = page_file_mapping(page)->host;
 898         int             status = 0;
 899
 900         nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE);
 901
 902         dprintk("NFS:       nfs_updatepage(%s/%s %d@%lld)\n",
 903                 file->f_path.dentry->d_parent->d_name.name,
 904                 file->f_path.dentry->d_name.name, count,
 905                 (long long)(page_file_offset(page) + offset));
 906
 907         /* If we're not using byte range locks, and we know the page
 908          * is up to date, it may be more efficient to extend the write
 909          * to cover the entire page in order to avoid fragmentation
 910          * inefficiencies.
 911          */
 912         if (nfs_write_pageuptodate(page, inode) &&
 913                         inode->i_flock == NULL &&
 914                         !(file->f_flags & O_DSYNC)) {
 915                 count = max(count + offset, nfs_page_length(page));
 916                 offset = 0;
 917         }
 918
 919         status = nfs_writepage_setup(ctx, page, offset, count);
 920         if (status < 0)
 921                 nfs_set_pageerror(page);
 922         else
 923                 __set_page_dirty_nobuffers(page);
 924
 925         dprintk("NFS:       nfs_updatepage returns %d (isize %lld)\n",
 926                         status, (long long)i_size_read(inode));
 927         return status;
 928 }
 929
 930 static int flush_task_priority(int how)
 931 {
 932         switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) {
 933                 case FLUSH_HIGHPRI:
 934                         return RPC_PRIORITY_HIGH;
 935                 case FLUSH_LOWPRI:
 936                         return RPC_PRIORITY_LOW;
 937         }
 938         return RPC_PRIORITY_NORMAL;
 939 }
 940
 941 int nfs_initiate_write(struct rpc_clnt *clnt,
 942                        struct nfs_write_data *data,
 943                        const struct rpc_call_ops *call_ops,
 944                        int how, int flags)
 945 {
 946         struct inode *inode = data->header->inode;
 947         int priority = flush_task_priority(how);
 948         struct rpc_task *task;
 949         struct rpc_message msg = {
 950                 .rpc_argp = &data->args,
 951                 .rpc_resp = &data->res,
 952                 .rpc_cred = data->header->cred,
 953         };
 954         struct rpc_task_setup task_setup_data = {
 955                 .rpc_client = clnt,
 956                 .task = &data->task,
 957                 .rpc_message = &msg,
 958                 .callback_ops = call_ops,
 959                 .callback_data = data,
 960                 .workqueue = nfsiod_workqueue,
 961                 .flags = RPC_TASK_ASYNC | flags,
 962                 .priority = priority,
 963         };
 964         int ret = 0;
 965
 966         /* Set up the initial task struct.  */
 967         NFS_PROTO(inode)->write_setup(data, &msg);
 968
 969         dprintk("NFS: %5u initiated write call "
 970                 "(req %s/%lld, %u bytes @ offset %llu)\n",
 971                 data->task.tk_pid,
 972                 inode->i_sb->s_id,
 973                 (long long)NFS_FILEID(inode),
 974                 data->args.count,
 975                 (unsigned long long)data->args.offset);
 976
 977         task = rpc_run_task(&task_setup_data);
 978         if (IS_ERR(task)) {
 979                 ret = PTR_ERR(task);
 980                 goto out;
 981         }
 982         if (how & FLUSH_SYNC) {
 983                 ret = rpc_wait_for_completion_task(task);
 984                 if (ret == 0)
 985                         ret = task->tk_status;
 986         }
 987         rpc_put_task(task);
 988 out:
 989         return ret;
 990 }
 991 EXPORT_SYMBOL_GPL(nfs_initiate_write);
 992
 993 /*
 994  * Set up the argument/result storage required for the RPC call.
 995  */
 996 static void nfs_write_rpcsetup(struct nfs_write_data *data,
 997                 unsigned int count, unsigned int offset,
 998                 int how, struct nfs_commit_info *cinfo)
 999 {
1000         struct nfs_page *req = data->header->req;
1001
1002         /* Set up the RPC argument and reply structs
1003          * NB: take care not to mess about with data->commit et al. */
1004
1005         data->args.fh     = NFS_FH(data->header->inode);
1006         data->args.offset = req_offset(req) + offset;
1007         /* pnfs_set_layoutcommit needs this */
1008         data->mds_offset = data->args.offset;
1009         data->args.pgbase = req->wb_pgbase + offset;
1010         data->args.pages  = data->pages.pagevec;
1011         data->args.count  = count;
1012         data->args.context = get_nfs_open_context(req->wb_context);
1013         data->args.lock_context = req->wb_lock_context;
1014         data->args.stable  = NFS_UNSTABLE;
1015         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
1016         case 0:
1017                 break;
1018         case FLUSH_COND_STABLE:
1019                 if (nfs_reqs_to_commit(cinfo))
1020                         break;
1021         default:
1022                 data->args.stable = NFS_FILE_SYNC;
1023         }
1024
1025         data->res.fattr   = &data->fattr;
1026         data->res.count   = count;
1027         data->res.verf    = &data->verf;
1028         nfs_fattr_init(&data->fattr);
1029 }
1030
1031 static int nfs_do_write(struct nfs_write_data *data,
1032                 const struct rpc_call_ops *call_ops,
1033                 int how)
1034 {
1035         struct inode *inode = data->header->inode;
1036
1037         return nfs_initiate_write(NFS_CLIENT(inode), data, call_ops, how, 0);
1038 }
1039
1040 static int nfs_do_multiple_writes(struct list_head *head,
1041                 const struct rpc_call_ops *call_ops,
1042                 int how)
1043 {
1044         struct nfs_write_data *data;
1045         int ret = 0;
1046
1047         while (!list_empty(head)) {
1048                 int ret2;
1049
1050                 data = list_first_entry(head, struct nfs_write_data, list);
1051                 list_del_init(&data->list);
1052
1053                 ret2 = nfs_do_write(data, call_ops, how);
1054                  if (ret == 0)
1055                          ret = ret2;
1056         }
1057         return ret;
1058 }
1059
1060 /* If a nfs_flush_* function fails, it should remove reqs from @head and
1061  * call this on each, which will prepare them to be retried on next
1062  * writeback using standard nfs.
1063  */
1064 static void nfs_redirty_request(struct nfs_page *req)
1065 {
1066         nfs_mark_request_dirty(req);
1067         nfs_unlock_request(req);
1068         nfs_end_page_writeback(req->wb_page);
1069         nfs_release_request(req);
1070 }
1071
1072 static void nfs_async_write_error(struct list_head *head)
1073 {
1074         struct nfs_page *req;
1075
1076         while (!list_empty(head)) {
1077                 req = nfs_list_entry(head->next);
1078                 nfs_list_remove_request(req);
1079                 nfs_redirty_request(req);
1080         }
1081 }
1082
1083 static const struct nfs_pgio_completion_ops nfs_async_write_completion_ops = {
1084         .error_cleanup = nfs_async_write_error,
1085         .completion = nfs_write_completion,
1086 };
1087
1088 static void nfs_flush_error(struct nfs_pageio_descriptor *desc,
1089                 struct nfs_pgio_header *hdr)
1090 {
1091         set_bit(NFS_IOHDR_REDO, &hdr->flags);
1092         while (!list_empty(&hdr->rpc_list)) {
1093                 struct nfs_write_data *data = list_first_entry(&hdr->rpc_list,
1094                                 struct nfs_write_data, list);
1095                 list_del(&data->list);
1096                 nfs_writedata_release(data);
1097         }
1098         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1099 }
1100
1101 /*
1102  * Generate multiple small requests to write out a single
1103  * contiguous dirty area on one page.
1104  */
1105 static int nfs_flush_multi(struct nfs_pageio_descriptor *desc,
1106                            struct nfs_pgio_header *hdr)
1107 {
1108         struct nfs_page *req = hdr->req;
1109         struct page *page = req->wb_page;
1110         struct nfs_write_data *data;
1111         size_t wsize = desc->pg_bsize, nbytes;
1112         unsigned int offset;
1113         int requests = 0;
1114         struct nfs_commit_info cinfo;
1115
1116         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1117
1118         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1119             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo) ||
1120              desc->pg_count > wsize))
1121                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1122
1123
1124         offset = 0;
1125         nbytes = desc->pg_count;
1126         do {
1127                 size_t len = min(nbytes, wsize);
1128
1129                 data = nfs_writedata_alloc(hdr, 1);
1130                 if (!data) {
1131                         nfs_flush_error(desc, hdr);
1132                         return -ENOMEM;
1133                 }
1134                 data->pages.pagevec[0] = page;
1135                 nfs_write_rpcsetup(data, len, offset, desc->pg_ioflags, &cinfo);
1136                 list_add(&data->list, &hdr->rpc_list);
1137                 requests++;
1138                 nbytes -= len;
1139                 offset += len;
1140         } while (nbytes != 0);
1141         nfs_list_remove_request(req);
1142         nfs_list_add_request(req, &hdr->pages);
1143         desc->pg_rpc_callops = &nfs_write_common_ops;
1144         return 0;
1145 }
1146
1147 /*
1148  * Create an RPC task for the given write request and kick it.
1149  * The page must have been locked by the caller.
1150  *
1151  * It may happen that the page we're passed is not marked dirty.
1152  * This is the case if nfs_updatepage detects a conflicting request
1153  * that has been written but not committed.
1154  */
1155 static int nfs_flush_one(struct nfs_pageio_descriptor *desc,
1156                          struct nfs_pgio_header *hdr)
1157 {
1158         struct nfs_page         *req;
1159         struct page             **pages;
1160         struct nfs_write_data   *data;
1161         struct list_head *head = &desc->pg_list;
1162         struct nfs_commit_info cinfo;
1163
1164         data = nfs_writedata_alloc(hdr, nfs_page_array_len(desc->pg_base,
1165                                                            desc->pg_count));
1166         if (!data) {
1167                 nfs_flush_error(desc, hdr);
1168                 return -ENOMEM;
1169         }
1170
1171         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
1172         pages = data->pages.pagevec;
1173         while (!list_empty(head)) {
1174                 req = nfs_list_entry(head->next);
1175                 nfs_list_remove_request(req);
1176                 nfs_list_add_request(req, &hdr->pages);
1177                 *pages++ = req->wb_page;
1178         }
1179
1180         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
1181             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
1182                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
1183
1184         /* Set up the argument struct */
1185         nfs_write_rpcsetup(data, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
1186         list_add(&data->list, &hdr->rpc_list);
1187         desc->pg_rpc_callops = &nfs_write_common_ops;
1188         return 0;
1189 }
1190
1191 int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
1192                       struct nfs_pgio_header *hdr)
1193 {
1194         if (desc->pg_bsize < PAGE_CACHE_SIZE)
1195                 return nfs_flush_multi(desc, hdr);
1196         return nfs_flush_one(desc, hdr);
1197 }
1198 EXPORT_SYMBOL_GPL(nfs_generic_flush);
1199
1200 static int nfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1201 {
1202         struct nfs_write_header *whdr;
1203         struct nfs_pgio_header *hdr;
1204         int ret;
1205
1206         whdr = nfs_writehdr_alloc();
1207         if (!whdr) {
1208                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1209                 return -ENOMEM;
1210         }
1211         hdr = &whdr->header;
1212         nfs_pgheader_init(desc, hdr, nfs_writehdr_free);
1213         atomic_inc(&hdr->refcnt);
1214         ret = nfs_generic_flush(desc, hdr);
1215         if (ret == 0)
1216                 ret = nfs_do_multiple_writes(&hdr->rpc_list,
1217                                              desc->pg_rpc_callops,
1218                                              desc->pg_ioflags);
1219         if (atomic_dec_and_test(&hdr->refcnt))
1220                 hdr->completion_ops->completion(hdr);
1221         return ret;
1222 }
1223
1224 static const struct nfs_pageio_ops nfs_pageio_write_ops = {
1225         .pg_test = nfs_generic_pg_test,
1226         .pg_doio = nfs_generic_pg_writepages,
1227 };
1228
1229 void nfs_pageio_init_write(struct nfs_pageio_descriptor *pgio,
1230                                struct inode *inode, int ioflags,
1231                                const struct nfs_pgio_completion_ops *compl_ops)
1232 {
1233         nfs_pageio_init(pgio, inode, &nfs_pageio_write_ops, compl_ops,
1234                                 NFS_SERVER(inode)->wsize, ioflags);
1235 }
1236 EXPORT_SYMBOL_GPL(nfs_pageio_init_write);
1237
1238 void nfs_pageio_reset_write_mds(struct nfs_pageio_descriptor *pgio)
1239 {
1240         pgio->pg_ops = &nfs_pageio_write_ops;
1241         pgio->pg_bsize = NFS_SERVER(pgio->pg_inode)->wsize;
1242 }
1243 EXPORT_SYMBOL_GPL(nfs_pageio_reset_write_mds);
1244
1245
1246 void nfs_write_prepare(struct rpc_task *task, void *calldata)
1247 {
1248         struct nfs_write_data *data = calldata;
1249         NFS_PROTO(data->header->inode)->write_rpc_prepare(task, data);
1250 }
1251
1252 void nfs_commit_prepare(struct rpc_task *task, void *calldata)
1253 {
1254         struct nfs_commit_data *data = calldata;
1255
1256         NFS_PROTO(data->inode)->commit_rpc_prepare(task, data);
1257 }
1258
1259 /*
1260  * Handle a write reply that flushes a whole page.
1261  *
1262  * FIXME: There is an inherent race with invalidate_inode_pages and
1263  *        writebacks since the page->count is kept > 1 for as long
1264  *        as the page has a write request pending.
1265  */
1266 static void nfs_writeback_done_common(struct rpc_task *task, void *calldata)
1267 {
1268         struct nfs_write_data   *data = calldata;
1269
1270         nfs_writeback_done(task, data);
1271 }
1272
1273 static void nfs_writeback_release_common(void *calldata)
1274 {
1275         struct nfs_write_data   *data = calldata;
1276         struct nfs_pgio_header *hdr = data->header;
1277         int status = data->task.tk_status;
1278
1279         if ((status >= 0) && nfs_write_need_commit(data)) {
1280                 spin_lock(&hdr->lock);
1281                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags))
1282                         ; /* Do nothing */
1283                 else if (!test_and_set_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags))
1284                         memcpy(hdr->verf, &data->verf, sizeof(*hdr->verf));
1285                 else if (memcmp(hdr->verf, &data->verf, sizeof(*hdr->verf)))
1286                         set_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags);
1287                 spin_unlock(&hdr->lock);
1288         }
1289         nfs_writedata_release(data);
1290 }
1291
1292 static const struct rpc_call_ops nfs_write_common_ops = {
1293         .rpc_call_prepare = nfs_write_prepare,
1294         .rpc_call_done = nfs_writeback_done_common,
1295         .rpc_release = nfs_writeback_release_common,
1296 };
1297
1298
1299 /*
1300  * This function is called when the WRITE call is complete.
1301  */
1302 void nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
1303 {
1304         struct nfs_writeargs    *argp = &data->args;
1305         struct nfs_writeres     *resp = &data->res;
1306         struct inode            *inode = data->header->inode;
1307         int status;
1308
1309         dprintk("NFS: %5u nfs_writeback_done (status %d)\n",
1310                 task->tk_pid, task->tk_status);
1311
1312         /*
1313          * ->write_done will attempt to use post-op attributes to detect
1314          * conflicting writes by other clients.  A strict interpretation
1315          * of close-to-open would allow us to continue caching even if
1316          * another writer had changed the file, but some applications
1317          * depend on tighter cache coherency when writing.
1318          */
1319         status = NFS_PROTO(inode)->write_done(task, data);
1320         if (status != 0)
1321                 return;
1322         nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, resp->count);
1323
1324 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1325         if (resp->verf->committed < argp->stable && task->tk_status >= 0) {
1326                 /* We tried a write call, but the server did not
1327                  * commit data to stable storage even though we
1328                  * requested it.
1329                  * Note: There is a known bug in Tru64 < 5.0 in which
1330                  *       the server reports NFS_DATA_SYNC, but performs
1331                  *       NFS_FILE_SYNC. We therefore implement this checking
1332                  *       as a dprintk() in order to avoid filling syslog.
1333                  */
1334                 static unsigned long    complain;
1335
1336                 /* Note this will print the MDS for a DS write */
1337                 if (time_before(complain, jiffies)) {
1338                         dprintk("NFS:       faulty NFS server %s:"
1339                                 " (committed = %d) != (stable = %d)\n",
1340                                 NFS_SERVER(inode)->nfs_client->cl_hostname,
1341                                 resp->verf->committed, argp->stable);
1342                         complain = jiffies + 300 * HZ;
1343                 }
1344         }
1345 #endif
1346         if (task->tk_status < 0)
1347                 nfs_set_pgio_error(data->header, task->tk_status, argp->offset);
1348         else if (resp->count < argp->count) {
1349                 static unsigned long    complain;
1350
1351                 /* This a short write! */
1352                 nfs_inc_stats(inode, NFSIOS_SHORTWRITE);
1353
1354                 /* Has the server at least made some progress? */
1355                 if (resp->count == 0) {
1356                         if (time_before(complain, jiffies)) {
1357                                 printk(KERN_WARNING
1358                                        "NFS: Server wrote zero bytes, expected %u.\n",
1359                                        argp->count);
1360                                 complain = jiffies + 300 * HZ;
1361                         }
1362                         nfs_set_pgio_error(data->header, -EIO, argp->offset);
1363                         task->tk_status = -EIO;
1364                         return;
1365                 }
1366                 /* Was this an NFSv2 write or an NFSv3 stable write? */
1367                 if (resp->verf->committed != NFS_UNSTABLE) {
1368                         /* Resend from where the server left off */
1369                         data->mds_offset += resp->count;
1370                         argp->offset += resp->count;
1371                         argp->pgbase += resp->count;
1372                         argp->count -= resp->count;
1373                 } else {
1374                         /* Resend as a stable write in order to avoid
1375                          * headaches in the case of a server crash.
1376                          */
1377                         argp->stable = NFS_FILE_SYNC;
1378                 }
1379                 rpc_restart_call_prepare(task);
1380         }
1381 }
1382
1383
1384 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
1385 static int nfs_commit_set_lock(struct nfs_inode *nfsi, int may_wait)
1386 {
1387         int ret;
1388
1389         if (!test_and_set_bit(NFS_INO_COMMIT, &nfsi->flags))
1390                 return 1;
1391         if (!may_wait)
1392                 return 0;
1393         ret = out_of_line_wait_on_bit_lock(&nfsi->flags,
1394                                 NFS_INO_COMMIT,
1395                                 nfs_wait_bit_killable,
1396                                 TASK_KILLABLE);
1397         return (ret < 0) ? ret : 1;
1398 }
1399
1400 static void nfs_commit_clear_lock(struct nfs_inode *nfsi)
1401 {
1402         clear_bit(NFS_INO_COMMIT, &nfsi->flags);
1403         smp_mb__after_clear_bit();
1404         wake_up_bit(&nfsi->flags, NFS_INO_COMMIT);
1405 }
1406
1407 void nfs_commitdata_release(struct nfs_commit_data *data)
1408 {
1409         put_nfs_open_context(data->context);
1410         nfs_commit_free(data);
1411 }
1412 EXPORT_SYMBOL_GPL(nfs_commitdata_release);
1413
1414 int nfs_initiate_commit(struct rpc_clnt *clnt, struct nfs_commit_data *data,
1415                         const struct rpc_call_ops *call_ops,
1416                         int how, int flags)
1417 {
1418         struct rpc_task *task;
1419         int priority = flush_task_priority(how);
1420         struct rpc_message msg = {
1421                 .rpc_argp = &data->args,
1422                 .rpc_resp = &data->res,
1423                 .rpc_cred = data->cred,
1424         };
1425         struct rpc_task_setup task_setup_data = {
1426                 .task = &data->task,
1427                 .rpc_client = clnt,
1428                 .rpc_message = &msg,
1429                 .callback_ops = call_ops,
1430                 .callback_data = data,
1431                 .workqueue = nfsiod_workqueue,
1432                 .flags = RPC_TASK_ASYNC | flags,
1433                 .priority = priority,
1434         };
1435         /* Set up the initial task struct.  */
1436         NFS_PROTO(data->inode)->commit_setup(data, &msg);
1437
1438         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
1439
1440         task = rpc_run_task(&task_setup_data);
1441         if (IS_ERR(task))
1442                 return PTR_ERR(task);
1443         if (how & FLUSH_SYNC)
1444                 rpc_wait_for_completion_task(task);
1445         rpc_put_task(task);
1446         return 0;
1447 }
1448 EXPORT_SYMBOL_GPL(nfs_initiate_commit);
1449
1450 /*
1451  * Set up the argument/result storage required for the RPC call.
1452  */
1453 void nfs_init_commit(struct nfs_commit_data *data,
1454                      struct list_head *head,
1455                      struct pnfs_layout_segment *lseg,
1456                      struct nfs_commit_info *cinfo)
1457 {
1458         struct nfs_page *first = nfs_list_entry(head->next);
1459         struct inode *inode = first->wb_context->dentry->d_inode;
1460
1461         /* Set up the RPC argument and reply structs
1462          * NB: take care not to mess about with data->commit et al. */
1463
1464         list_splice_init(head, &data->pages);
1465
1466         data->inode       = inode;
1467         data->cred        = first->wb_context->cred;
1468         data->lseg        = lseg; /* reference transferred */
1469         data->mds_ops     = &nfs_commit_ops;
1470         data->completion_ops = cinfo->completion_ops;
1471         data->dreq        = cinfo->dreq;
1472
1473         data->args.fh     = NFS_FH(data->inode);
1474         /* Note: we always request a commit of the entire inode */
1475         data->args.offset = 0;
1476         data->args.count  = 0;
1477         data->context     = get_nfs_open_context(first->wb_context);
1478         data->res.fattr   = &data->fattr;
1479         data->res.verf    = &data->verf;
1480         nfs_fattr_init(&data->fattr);
1481 }
1482 EXPORT_SYMBOL_GPL(nfs_init_commit);
1483
1484 void nfs_retry_commit(struct list_head *page_list,
1485                       struct pnfs_layout_segment *lseg,
1486                       struct nfs_commit_info *cinfo)
1487 {
1488         struct nfs_page *req;
1489
1490         while (!list_empty(page_list)) {
1491                 req = nfs_list_entry(page_list->next);
1492                 nfs_list_remove_request(req);
1493                 nfs_mark_request_commit(req, lseg, cinfo);
1494                 if (!cinfo->dreq) {
1495                         dec_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1496                         dec_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1497                                      BDI_RECLAIMABLE);
1498                 }
1499                 nfs_unlock_and_release_request(req);
1500         }
1501 }
1502 EXPORT_SYMBOL_GPL(nfs_retry_commit);
1503
1504 /*
1505  * Commit dirty pages
1506  */
1507 static int
1508 nfs_commit_list(struct inode *inode, struct list_head *head, int how,
1509                 struct nfs_commit_info *cinfo)
1510 {
1511         struct nfs_commit_data  *data;
1512
1513         data = nfs_commitdata_alloc();
1514
1515         if (!data)
1516                 goto out_bad;
1517
1518         /* Set up the argument struct */
1519         nfs_init_commit(data, head, NULL, cinfo);
1520         atomic_inc(&cinfo->mds->rpcs_out);
1521         return nfs_initiate_commit(NFS_CLIENT(inode), data, data->mds_ops,
1522                                    how, 0);
1523  out_bad:
1524         nfs_retry_commit(head, NULL, cinfo);
1525         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1526         return -ENOMEM;
1527 }
1528
1529 /*
1530  * COMMIT call returned
1531  */
1532 static void nfs_commit_done(struct rpc_task *task, void *calldata)
1533 {
1534         struct nfs_commit_data  *data = calldata;
1535
1536         dprintk("NFS: %5u nfs_commit_done (status %d)\n",
1537                                 task->tk_pid, task->tk_status);
1538
1539         /* Call the NFS version-specific code */
1540         NFS_PROTO(data->inode)->commit_done(task, data);
1541 }
1542
1543 static void nfs_commit_release_pages(struct nfs_commit_data *data)
1544 {
1545         struct nfs_page *req;
1546         int status = data->task.tk_status;
1547         struct nfs_commit_info cinfo;
1548
1549         while (!list_empty(&data->pages)) {
1550                 req = nfs_list_entry(data->pages.next);
1551                 nfs_list_remove_request(req);
1552                 nfs_clear_page_commit(req->wb_page);
1553
1554                 dprintk("NFS:       commit (%s/%lld %d@%lld)",
1555                         req->wb_context->dentry->d_sb->s_id,
1556                         (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
1557                         req->wb_bytes,
1558                         (long long)req_offset(req));
1559                 if (status < 0) {
1560                         nfs_context_set_write_error(req->wb_context, status);
1561                         nfs_inode_remove_request(req);
1562                         dprintk(", error = %d\n", status);
1563                         goto next;
1564                 }
1565
1566                 /* Okay, COMMIT succeeded, apparently. Check the verifier
1567                  * returned by the server against all stored verfs. */
1568                 if (!memcmp(&req->wb_verf, &data->verf.verifier, sizeof(req->wb_verf))) {
1569                         /* We have a match */
1570                         nfs_inode_remove_request(req);
1571                         dprintk(" OK\n");
1572                         goto next;
1573                 }
1574                 /* We have a mismatch. Write the page again */
1575                 dprintk(" mismatch\n");
1576                 nfs_mark_request_dirty(req);
1577                 set_bit(NFS_CONTEXT_RESEND_WRITES, &req->wb_context->flags);
1578         next:
1579                 nfs_unlock_and_release_request(req);
1580         }
1581         nfs_init_cinfo(&cinfo, data->inode, data->dreq);
1582         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
1583                 nfs_commit_clear_lock(NFS_I(data->inode));
1584 }
1585
1586 static void nfs_commit_release(void *calldata)
1587 {
1588         struct nfs_commit_data *data = calldata;
1589
1590         data->completion_ops->completion(data);
1591         nfs_commitdata_release(calldata);
1592 }
1593
1594 static const struct rpc_call_ops nfs_commit_ops = {
1595         .rpc_call_prepare = nfs_commit_prepare,
1596         .rpc_call_done = nfs_commit_done,
1597         .rpc_release = nfs_commit_release,
1598 };
1599
1600 static const struct nfs_commit_completion_ops nfs_commit_completion_ops = {
1601         .completion = nfs_commit_release_pages,
1602         .error_cleanup = nfs_commit_clear_lock,
1603 };
1604
1605 int nfs_generic_commit_list(struct inode *inode, struct list_head *head,
1606                             int how, struct nfs_commit_info *cinfo)
1607 {
1608         int status;
1609
1610         status = pnfs_commit_list(inode, head, how, cinfo);
1611         if (status == PNFS_NOT_ATTEMPTED)
1612                 status = nfs_commit_list(inode, head, how, cinfo);
1613         return status;
1614 }
1615
1616 int nfs_commit_inode(struct inode *inode, int how)
1617 {
1618         LIST_HEAD(head);
1619         struct nfs_commit_info cinfo;
1620         int may_wait = how & FLUSH_SYNC;
1621         int res;
1622
1623         res = nfs_commit_set_lock(NFS_I(inode), may_wait);
1624         if (res <= 0)
1625                 goto out_mark_dirty;
1626         nfs_init_cinfo_from_inode(&cinfo, inode);
1627         res = nfs_scan_commit(inode, &head, &cinfo);
1628         if (res) {
1629                 int error;
1630
1631                 error = nfs_generic_commit_list(inode, &head, how, &cinfo);
1632                 if (error < 0)
1633                         return error;
1634                 if (!may_wait)
1635                         goto out_mark_dirty;
1636                 error = wait_on_bit(&NFS_I(inode)->flags,
1637                                 NFS_INO_COMMIT,
1638                                 nfs_wait_bit_killable,
1639                                 TASK_KILLABLE);
1640                 if (error < 0)
1641                         return error;
1642         } else
1643                 nfs_commit_clear_lock(NFS_I(inode));
1644         return res;
1645         /* Note: If we exit without ensuring that the commit is complete,
1646          * we must mark the inode as dirty. Otherwise, future calls to
1647          * sync_inode() with the WB_SYNC_ALL flag set will fail to ensure
1648          * that the data is on the disk.
1649          */
1650 out_mark_dirty:
1651         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1652         return res;
1653 }
1654
1655 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1656 {
1657         struct nfs_inode *nfsi = NFS_I(inode);
1658         int flags = FLUSH_SYNC;
1659         int ret = 0;
1660
1661         /* no commits means nothing needs to be done */
1662         if (!nfsi->commit_info.ncommit)
1663                 return ret;
1664
1665         if (wbc->sync_mode == WB_SYNC_NONE) {
1666                 /* Don't commit yet if this is a non-blocking flush and there
1667                  * are a lot of outstanding writes for this mapping.
1668                  */
1669                 if (nfsi->commit_info.ncommit <= (nfsi->npages >> 1))
1670                         goto out_mark_dirty;
1671
1672                 /* don't wait for the COMMIT response */
1673                 flags = 0;
1674         }
1675
1676         ret = nfs_commit_inode(inode, flags);
1677         if (ret >= 0) {
1678                 if (wbc->sync_mode == WB_SYNC_NONE) {
1679                         if (ret < wbc->nr_to_write)
1680                                 wbc->nr_to_write -= ret;
1681                         else
1682                                 wbc->nr_to_write = 0;
1683                 }
1684                 return 0;
1685         }
1686 out_mark_dirty:
1687         __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1688         return ret;
1689 }
1690 #else
1691 static int nfs_commit_unstable_pages(struct inode *inode, struct writeback_control *wbc)
1692 {
1693         return 0;
1694 }
1695 #endif
1696
1697 int nfs_write_inode(struct inode *inode, struct writeback_control *wbc)
1698 {
1699         return nfs_commit_unstable_pages(inode, wbc);
1700 }
1701 EXPORT_SYMBOL_GPL(nfs_write_inode);
1702
1703 /*
1704  * flush the inode to disk.
1705  */
1706 int nfs_wb_all(struct inode *inode)
1707 {
1708         struct writeback_control wbc = {
1709                 .sync_mode = WB_SYNC_ALL,
1710                 .nr_to_write = LONG_MAX,
1711                 .range_start = 0,
1712                 .range_end = LLONG_MAX,
1713         };
1714
1715         return sync_inode(inode, &wbc);
1716 }
1717 EXPORT_SYMBOL_GPL(nfs_wb_all);
1718
1719 int nfs_wb_page_cancel(struct inode *inode, struct page *page)
1720 {
1721         struct nfs_page *req;
1722         int ret = 0;
1723
1724         for (;;) {
1725                 wait_on_page_writeback(page);
1726                 req = nfs_page_find_request(page);
1727                 if (req == NULL)
1728                         break;
1729                 if (nfs_lock_request(req)) {
1730                         nfs_clear_request_commit(req);
1731                         nfs_inode_remove_request(req);
1732                         /*
1733                          * In case nfs_inode_remove_request has marked the
1734                          * page as being dirty
1735                          */
1736                         cancel_dirty_page(page, PAGE_CACHE_SIZE);
1737                         nfs_unlock_and_release_request(req);
1738                         break;
1739                 }
1740                 ret = nfs_wait_on_request(req);
1741                 nfs_release_request(req);
1742                 if (ret < 0)
1743                         break;
1744         }
1745         return ret;
1746 }
1747
1748 /*
1749  * Write back all requests on one page - we do this before reading it.
1750  */
1751 int nfs_wb_page(struct inode *inode, struct page *page)
1752 {
1753         loff_t range_start = page_file_offset(page);
1754         loff_t range_end = range_start + (loff_t)(PAGE_CACHE_SIZE - 1);
1755         struct writeback_control wbc = {
1756                 .sync_mode = WB_SYNC_ALL,
1757                 .nr_to_write = 0,
1758                 .range_start = range_start,
1759                 .range_end = range_end,
1760         };
1761         int ret;
1762
1763         for (;;) {
1764                 wait_on_page_writeback(page);
1765                 if (clear_page_dirty_for_io(page)) {
1766                         ret = nfs_writepage_locked(page, &wbc);
1767                         if (ret < 0)
1768                                 goto out_error;
1769                         continue;
1770                 }
1771                 if (!PagePrivate(page))
1772                         break;
1773                 ret = nfs_commit_inode(inode, FLUSH_SYNC);
1774                 if (ret < 0)
1775                         goto out_error;
1776         }
1777         return 0;
1778 out_error:
1779         return ret;
1780 }
1781
1782 #ifdef CONFIG_MIGRATION
1783 int nfs_migrate_page(struct address_space *mapping, struct page *newpage,
1784                 struct page *page, enum migrate_mode mode)
1785 {
1786         /*
1787          * If PagePrivate is set, then the page is currently associated with
1788          * an in-progress read or write request. Don't try to migrate it.
1789          *
1790          * FIXME: we could do this in principle, but we'll need a way to ensure
1791          *        that we can safely release the inode reference while holding
1792          *        the page lock.
1793          */
1794         if (PagePrivate(page))
1795                 return -EBUSY;
1796
1797         if (!nfs_fscache_release_page(page, GFP_KERNEL))
1798                 return -EBUSY;
1799
1800         return migrate_page(mapping, newpage, page, mode);
1801 }
1802 #endif
1803
1804 int __init nfs_init_writepagecache(void)
1805 {
1806         nfs_wdata_cachep = kmem_cache_create("nfs_write_data",
1807                                              sizeof(struct nfs_write_header),
1808                                              0, SLAB_HWCACHE_ALIGN,
1809                                              NULL);
1810         if (nfs_wdata_cachep == NULL)
1811                 return -ENOMEM;
1812
1813         nfs_wdata_mempool = mempool_create_slab_pool(MIN_POOL_WRITE,
1814                                                      nfs_wdata_cachep);
1815         if (nfs_wdata_mempool == NULL)
1816                 goto out_destroy_write_cache;
1817
1818         nfs_cdata_cachep = kmem_cache_create("nfs_commit_data",
1819                                              sizeof(struct nfs_commit_data),
1820                                              0, SLAB_HWCACHE_ALIGN,
1821                                              NULL);
1822         if (nfs_cdata_cachep == NULL)
1823                 goto out_destroy_write_mempool;
1824
1825         nfs_commit_mempool = mempool_create_slab_pool(MIN_POOL_COMMIT,
1826                                                       nfs_cdata_cachep);
1827         if (nfs_commit_mempool == NULL)
1828                 goto out_destroy_commit_cache;
1829
1830         /*
1831          * NFS congestion size, scale with available memory.
1832          *
1833          *  64MB:    8192k
1834          * 128MB:   11585k
1835          * 256MB:   16384k
1836          * 512MB:   23170k
1837          *   1GB:   32768k
1838          *   2GB:   46340k
1839          *   4GB:   65536k
1840          *   8GB:   92681k
1841          *  16GB:  131072k
1842          *
1843          * This allows larger machines to have larger/more transfers.
1844          * Limit the default to 256M
1845          */
1846         nfs_congestion_kb = (16*int_sqrt(totalram_pages)) << (PAGE_SHIFT-10);
1847         if (nfs_congestion_kb > 256*1024)
1848                 nfs_congestion_kb = 256*1024;
1849
1850         return 0;
1851
1852 out_destroy_commit_cache:
1853         kmem_cache_destroy(nfs_cdata_cachep);
1854 out_destroy_write_mempool:
1855         mempool_destroy(nfs_wdata_mempool);
1856 out_destroy_write_cache:
1857         kmem_cache_destroy(nfs_wdata_cachep);
1858         return -ENOMEM;
1859 }
1860
1861 void nfs_destroy_writepagecache(void)
1862 {
1863         mempool_destroy(nfs_commit_mempool);
1864         kmem_cache_destroy(nfs_cdata_cachep);
1865         mempool_destroy(nfs_wdata_mempool);
1866         kmem_cache_destroy(nfs_wdata_cachep);
1867 }
1868