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