fs/nfs/direct.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/direct.c
   4  *
   5  * Copyright (C) 2003 by Chuck Lever <[email protected]>
   6  *
   7  * High-performance uncached I/O for the Linux NFS client
   8  *
   9  * There are important applications whose performance or correctness
  10  * depends on uncached access to file data.  Database clusters
  11  * (multiple copies of the same instance running on separate hosts)
  12  * implement their own cache coherency protocol that subsumes file
  13  * system cache protocols.  Applications that process datasets
  14  * considerably larger than the client's memory do not always benefit
  15  * from a local cache.  A streaming video server, for instance, has no
  16  * need to cache the contents of a file.
  17  *
  18  * When an application requests uncached I/O, all read and write requests
  19  * are made directly to the server; data stored or fetched via these
  20  * requests is not cached in the Linux page cache.  The client does not
  21  * correct unaligned requests from applications.  All requested bytes are
  22  * held on permanent storage before a direct write system call returns to
  23  * an application.
  24  *
  25  * Solaris implements an uncached I/O facility called directio() that
  26  * is used for backups and sequential I/O to very large files.  Solaris
  27  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  28  * an undocumented mount option.
  29  *
  30  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  31  * help from Andrew Morton.
  32  *
  33  * 18 Dec 2001  Initial implementation for 2.4  --cel
  34  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  35  * 08 Jun 2003  Port to 2.5 APIs  --cel
  36  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  37  * 15 Sep 2004  Parallel async reads  --cel
  38  * 04 May 2005  support O_DIRECT with aio  --cel
  39  *
  40  */
  41
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/file.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/kref.h>
  48 #include <linux/slab.h>
  49 #include <linux/task_io_accounting_ops.h>
  50 #include <linux/module.h>
  51
  52 #include <linux/nfs_fs.h>
  53 #include <linux/nfs_page.h>
  54 #include <linux/sunrpc/clnt.h>
  55
  56 #include <linux/uaccess.h>
  57 #include <linux/atomic.h>
  58
  59 #include "internal.h"
  60 #include "iostat.h"
  61 #include "pnfs.h"
  62 #include "fscache.h"
  63 #include "nfstrace.h"
  64
  65 #define NFSDBG_FACILITY         NFSDBG_VFS
  66
  67 static struct kmem_cache *nfs_direct_cachep;
  68
  69 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  70 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  71 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
  72 static void nfs_direct_write_schedule_work(struct work_struct *work);
  73
  74 static inline void get_dreq(struct nfs_direct_req *dreq)
  75 {
  76         atomic_inc(&dreq->io_count);
  77 }
  78
  79 static inline int put_dreq(struct nfs_direct_req *dreq)
  80 {
  81         return atomic_dec_and_test(&dreq->io_count);
  82 }
  83
  84 static void
  85 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
  86                             const struct nfs_pgio_header *hdr,
  87                             ssize_t dreq_len)
  88 {
  89         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
  90               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
  91                 return;
  92         if (dreq->max_count >= dreq_len) {
  93                 dreq->max_count = dreq_len;
  94                 if (dreq->count > dreq_len)
  95                         dreq->count = dreq_len;
  96         }
  97
  98         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
  99                 dreq->error = hdr->error;
 100 }
 101
 102 static void
 103 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 104                        const struct nfs_pgio_header *hdr)
 105 {
 106         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 107         ssize_t dreq_len = 0;
 108
 109         if (hdr_end > dreq->io_start)
 110                 dreq_len = hdr_end - dreq->io_start;
 111
 112         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 113
 114         if (dreq_len > dreq->max_count)
 115                 dreq_len = dreq->max_count;
 116
 117         if (dreq->count < dreq_len)
 118                 dreq->count = dreq_len;
 119 }
 120
 121 static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
 122                                         struct nfs_page *req)
 123 {
 124         loff_t offs = req_offset(req);
 125         size_t req_start = (size_t)(offs - dreq->io_start);
 126
 127         if (req_start < dreq->max_count)
 128                 dreq->max_count = req_start;
 129         if (req_start < dreq->count)
 130                 dreq->count = req_start;
 131 }
 132
 133 /**
 134  * nfs_swap_rw - NFS address space operation for swap I/O
 135  * @iocb: target I/O control block
 136  * @iter: I/O buffer
 137  *
 138  * Perform IO to the swap-file.  This is much like direct IO.
 139  */
 140 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
 141 {
 142         ssize_t ret;
 143
 144         if (iov_iter_rw(iter) == READ)
 145                 ret = nfs_file_direct_read(iocb, iter, true);
 146         else
 147                 ret = nfs_file_direct_write(iocb, iter, true);
 148         if (ret < 0)
 149                 return ret;
 150         return 0;
 151 }
 152
 153 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 154 {
 155         unsigned int i;
 156         for (i = 0; i < npages; i++)
 157                 put_page(pages[i]);
 158 }
 159
 160 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 161                               struct nfs_direct_req *dreq)
 162 {
 163         cinfo->inode = dreq->inode;
 164         cinfo->mds = &dreq->mds_cinfo;
 165         cinfo->ds = &dreq->ds_cinfo;
 166         cinfo->dreq = dreq;
 167         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 168 }
 169
 170 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 171 {
 172         struct nfs_direct_req *dreq;
 173
 174         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 175         if (!dreq)
 176                 return NULL;
 177
 178         kref_init(&dreq->kref);
 179         kref_get(&dreq->kref);
 180         init_completion(&dreq->completion);
 181         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 182         pnfs_init_ds_commit_info(&dreq->ds_cinfo);
 183         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 184         spin_lock_init(&dreq->lock);
 185
 186         return dreq;
 187 }
 188
 189 static void nfs_direct_req_free(struct kref *kref)
 190 {
 191         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 192
 193         pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
 194         if (dreq->l_ctx != NULL)
 195                 nfs_put_lock_context(dreq->l_ctx);
 196         if (dreq->ctx != NULL)
 197                 put_nfs_open_context(dreq->ctx);
 198         kmem_cache_free(nfs_direct_cachep, dreq);
 199 }
 200
 201 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 202 {
 203         kref_put(&dreq->kref, nfs_direct_req_free);
 204 }
 205
 206 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
 207 {
 208         loff_t start = offset - dreq->io_start;
 209         return dreq->max_count - start;
 210 }
 211 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 212
 213 /*
 214  * Collects and returns the final error value/byte-count.
 215  */
 216 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 217 {
 218         ssize_t result = -EIOCBQUEUED;
 219
 220         /* Async requests don't wait here */
 221         if (dreq->iocb)
 222                 goto out;
 223
 224         result = wait_for_completion_killable(&dreq->completion);
 225
 226         if (!result) {
 227                 result = dreq->count;
 228                 WARN_ON_ONCE(dreq->count < 0);
 229         }
 230         if (!result)
 231                 result = dreq->error;
 232
 233 out:
 234         return (ssize_t) result;
 235 }
 236
 237 /*
 238  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 239  * the iocb is still valid here if this is a synchronous request.
 240  */
 241 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 242 {
 243         struct inode *inode = dreq->inode;
 244
 245         inode_dio_end(inode);
 246
 247         if (dreq->iocb) {
 248                 long res = (long) dreq->error;
 249                 if (dreq->count != 0) {
 250                         res = (long) dreq->count;
 251                         WARN_ON_ONCE(dreq->count < 0);
 252                 }
 253                 dreq->iocb->ki_complete(dreq->iocb, res);
 254         }
 255
 256         complete(&dreq->completion);
 257
 258         nfs_direct_req_release(dreq);
 259 }
 260
 261 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 262 {
 263         unsigned long bytes = 0;
 264         struct nfs_direct_req *dreq = hdr->dreq;
 265
 266         spin_lock(&dreq->lock);
 267         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 268                 spin_unlock(&dreq->lock);
 269                 goto out_put;
 270         }
 271
 272         nfs_direct_count_bytes(dreq, hdr);
 273         spin_unlock(&dreq->lock);
 274
 275         while (!list_empty(&hdr->pages)) {
 276                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 277                 struct page *page = req->wb_page;
 278
 279                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 280                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 281                         set_page_dirty(page);
 282                 bytes += req->wb_bytes;
 283                 nfs_list_remove_request(req);
 284                 nfs_release_request(req);
 285         }
 286 out_put:
 287         if (put_dreq(dreq))
 288                 nfs_direct_complete(dreq);
 289         hdr->release(hdr);
 290 }
 291
 292 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 293 {
 294         struct nfs_page *req;
 295
 296         while (!list_empty(head)) {
 297                 req = nfs_list_entry(head->next);
 298                 nfs_list_remove_request(req);
 299                 nfs_release_request(req);
 300         }
 301 }
 302
 303 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 304 {
 305         get_dreq(hdr->dreq);
 306 }
 307
 308 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 309         .error_cleanup = nfs_read_sync_pgio_error,
 310         .init_hdr = nfs_direct_pgio_init,
 311         .completion = nfs_direct_read_completion,
 312 };
 313
 314 /*
 315  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 316  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 317  * bail and stop sending more reads.  Read length accounting is
 318  * handled automatically by nfs_direct_read_result().  Otherwise, if
 319  * no requests have been sent, just return an error.
 320  */
 321
 322 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 323                                               struct iov_iter *iter,
 324                                               loff_t pos)
 325 {
 326         struct nfs_pageio_descriptor desc;
 327         struct inode *inode = dreq->inode;
 328         ssize_t result = -EINVAL;
 329         size_t requested_bytes = 0;
 330         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 331
 332         nfs_pageio_init_read(&desc, dreq->inode, false,
 333                              &nfs_direct_read_completion_ops);
 334         get_dreq(dreq);
 335         desc.pg_dreq = dreq;
 336         inode_dio_begin(inode);
 337
 338         while (iov_iter_count(iter)) {
 339                 struct page **pagevec;
 340                 size_t bytes;
 341                 size_t pgbase;
 342                 unsigned npages, i;
 343
 344                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
 345                                                   rsize, &pgbase);
 346                 if (result < 0)
 347                         break;
 348
 349                 bytes = result;
 350                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 351                 for (i = 0; i < npages; i++) {
 352                         struct nfs_page *req;
 353                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 354                         /* XXX do we need to do the eof zeroing found in async_filler? */
 355                         req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 356                                                         pgbase, pos, req_len);
 357                         if (IS_ERR(req)) {
 358                                 result = PTR_ERR(req);
 359                                 break;
 360                         }
 361                         if (!nfs_pageio_add_request(&desc, req)) {
 362                                 result = desc.pg_error;
 363                                 nfs_release_request(req);
 364                                 break;
 365                         }
 366                         pgbase = 0;
 367                         bytes -= req_len;
 368                         requested_bytes += req_len;
 369                         pos += req_len;
 370                 }
 371                 nfs_direct_release_pages(pagevec, npages);
 372                 kvfree(pagevec);
 373                 if (result < 0)
 374                         break;
 375         }
 376
 377         nfs_pageio_complete(&desc);
 378
 379         /*
 380          * If no bytes were started, return the error, and let the
 381          * generic layer handle the completion.
 382          */
 383         if (requested_bytes == 0) {
 384                 inode_dio_end(inode);
 385                 nfs_direct_req_release(dreq);
 386                 return result < 0 ? result : -EIO;
 387         }
 388
 389         if (put_dreq(dreq))
 390                 nfs_direct_complete(dreq);
 391         return requested_bytes;
 392 }
 393
 394 /**
 395  * nfs_file_direct_read - file direct read operation for NFS files
 396  * @iocb: target I/O control block
 397  * @iter: vector of user buffers into which to read data
 398  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 399  *
 400  * We use this function for direct reads instead of calling
 401  * generic_file_aio_read() in order to avoid gfar's check to see if
 402  * the request starts before the end of the file.  For that check
 403  * to work, we must generate a GETATTR before each direct read, and
 404  * even then there is a window between the GETATTR and the subsequent
 405  * READ where the file size could change.  Our preference is simply
 406  * to do all reads the application wants, and the server will take
 407  * care of managing the end of file boundary.
 408  *
 409  * This function also eliminates unnecessarily updating the file's
 410  * atime locally, as the NFS server sets the file's atime, and this
 411  * client must read the updated atime from the server back into its
 412  * cache.
 413  */
 414 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 415                              bool swap)
 416 {
 417         struct file *file = iocb->ki_filp;
 418         struct address_space *mapping = file->f_mapping;
 419         struct inode *inode = mapping->host;
 420         struct nfs_direct_req *dreq;
 421         struct nfs_lock_context *l_ctx;
 422         ssize_t result, requested;
 423         size_t count = iov_iter_count(iter);
 424         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 425
 426         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 427                 file, count, (long long) iocb->ki_pos);
 428
 429         result = 0;
 430         if (!count)
 431                 goto out;
 432
 433         task_io_account_read(count);
 434
 435         result = -ENOMEM;
 436         dreq = nfs_direct_req_alloc();
 437         if (dreq == NULL)
 438                 goto out;
 439
 440         dreq->inode = inode;
 441         dreq->max_count = count;
 442         dreq->io_start = iocb->ki_pos;
 443         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 444         l_ctx = nfs_get_lock_context(dreq->ctx);
 445         if (IS_ERR(l_ctx)) {
 446                 result = PTR_ERR(l_ctx);
 447                 nfs_direct_req_release(dreq);
 448                 goto out_release;
 449         }
 450         dreq->l_ctx = l_ctx;
 451         if (!is_sync_kiocb(iocb))
 452                 dreq->iocb = iocb;
 453
 454         if (user_backed_iter(iter))
 455                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 456
 457         if (!swap) {
 458                 result = nfs_start_io_direct(inode);
 459                 if (result) {
 460                         /* release the reference that would usually be
 461                          * consumed by nfs_direct_read_schedule_iovec()
 462                          */
 463                         nfs_direct_req_release(dreq);
 464                         goto out_release;
 465                 }
 466         }
 467
 468         NFS_I(inode)->read_io += count;
 469         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 470
 471         if (!swap)
 472                 nfs_end_io_direct(inode);
 473
 474         if (requested > 0) {
 475                 result = nfs_direct_wait(dreq);
 476                 if (result > 0) {
 477                         requested -= result;
 478                         iocb->ki_pos += result;
 479                 }
 480                 iov_iter_revert(iter, requested);
 481         } else {
 482                 result = requested;
 483         }
 484
 485 out_release:
 486         nfs_direct_req_release(dreq);
 487 out:
 488         return result;
 489 }
 490
 491 static void nfs_direct_add_page_head(struct list_head *list,
 492                                      struct nfs_page *req)
 493 {
 494         struct nfs_page *head = req->wb_head;
 495
 496         if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
 497                 return;
 498         if (!list_empty(&head->wb_list)) {
 499                 nfs_unlock_request(head);
 500                 return;
 501         }
 502         list_add(&head->wb_list, list);
 503         kref_get(&head->wb_kref);
 504         kref_get(&head->wb_kref);
 505 }
 506
 507 static void nfs_direct_join_group(struct list_head *list,
 508                                   struct nfs_commit_info *cinfo,
 509                                   struct inode *inode)
 510 {
 511         struct nfs_page *req, *subreq;
 512
 513         list_for_each_entry(req, list, wb_list) {
 514                 if (req->wb_head != req) {
 515                         nfs_direct_add_page_head(&req->wb_list, req);
 516                         continue;
 517                 }
 518                 subreq = req->wb_this_page;
 519                 if (subreq == req)
 520                         continue;
 521                 do {
 522                         /*
 523                          * Remove subrequests from this list before freeing
 524                          * them in the call to nfs_join_page_group().
 525                          */
 526                         if (!list_empty(&subreq->wb_list)) {
 527                                 nfs_list_remove_request(subreq);
 528                                 nfs_release_request(subreq);
 529                         }
 530                 } while ((subreq = subreq->wb_this_page) != req);
 531                 nfs_join_page_group(req, cinfo, inode);
 532         }
 533 }
 534
 535 static void
 536 nfs_direct_write_scan_commit_list(struct inode *inode,
 537                                   struct list_head *list,
 538                                   struct nfs_commit_info *cinfo)
 539 {
 540         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 541         pnfs_recover_commit_reqs(list, cinfo);
 542         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 543         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 544 }
 545
 546 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 547 {
 548         struct nfs_pageio_descriptor desc;
 549         struct nfs_page *req;
 550         LIST_HEAD(reqs);
 551         struct nfs_commit_info cinfo;
 552
 553         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 554         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 555
 556         nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
 557
 558         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 559         get_dreq(dreq);
 560
 561         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 562                               &nfs_direct_write_completion_ops);
 563         desc.pg_dreq = dreq;
 564
 565         while (!list_empty(&reqs)) {
 566                 req = nfs_list_entry(reqs.next);
 567                 /* Bump the transmission count */
 568                 req->wb_nio++;
 569                 if (!nfs_pageio_add_request(&desc, req)) {
 570                         spin_lock(&dreq->lock);
 571                         if (dreq->error < 0) {
 572                                 desc.pg_error = dreq->error;
 573                         } else if (desc.pg_error != -EAGAIN) {
 574                                 dreq->flags = 0;
 575                                 if (!desc.pg_error)
 576                                         desc.pg_error = -EIO;
 577                                 dreq->error = desc.pg_error;
 578                         } else
 579                                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 580                         spin_unlock(&dreq->lock);
 581                         break;
 582                 }
 583                 nfs_release_request(req);
 584         }
 585         nfs_pageio_complete(&desc);
 586
 587         while (!list_empty(&reqs)) {
 588                 req = nfs_list_entry(reqs.next);
 589                 nfs_list_remove_request(req);
 590                 nfs_unlock_and_release_request(req);
 591                 if (desc.pg_error == -EAGAIN) {
 592                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 593                 } else {
 594                         spin_lock(&dreq->lock);
 595                         nfs_direct_truncate_request(dreq, req);
 596                         spin_unlock(&dreq->lock);
 597                         nfs_release_request(req);
 598                 }
 599         }
 600
 601         if (put_dreq(dreq))
 602                 nfs_direct_write_complete(dreq);
 603 }
 604
 605 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 606 {
 607         const struct nfs_writeverf *verf = data->res.verf;
 608         struct nfs_direct_req *dreq = data->dreq;
 609         struct nfs_commit_info cinfo;
 610         struct nfs_page *req;
 611         int status = data->task.tk_status;
 612
 613         trace_nfs_direct_commit_complete(dreq);
 614
 615         spin_lock(&dreq->lock);
 616         if (status < 0) {
 617                 /* Errors in commit are fatal */
 618                 dreq->error = status;
 619                 dreq->flags = NFS_ODIRECT_DONE;
 620         } else {
 621                 status = dreq->error;
 622         }
 623         spin_unlock(&dreq->lock);
 624
 625         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 626
 627         while (!list_empty(&data->pages)) {
 628                 req = nfs_list_entry(data->pages.next);
 629                 nfs_list_remove_request(req);
 630                 if (status < 0) {
 631                         spin_lock(&dreq->lock);
 632                         nfs_direct_truncate_request(dreq, req);
 633                         spin_unlock(&dreq->lock);
 634                         nfs_release_request(req);
 635                 } else if (!nfs_write_match_verf(verf, req)) {
 636                         spin_lock(&dreq->lock);
 637                         if (dreq->flags == 0)
 638                                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 639                         spin_unlock(&dreq->lock);
 640                         /*
 641                          * Despite the reboot, the write was successful,
 642                          * so reset wb_nio.
 643                          */
 644                         req->wb_nio = 0;
 645                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 646                 } else
 647                         nfs_release_request(req);
 648                 nfs_unlock_and_release_request(req);
 649         }
 650
 651         if (nfs_commit_end(cinfo.mds))
 652                 nfs_direct_write_complete(dreq);
 653 }
 654
 655 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 656                 struct nfs_page *req)
 657 {
 658         struct nfs_direct_req *dreq = cinfo->dreq;
 659
 660         trace_nfs_direct_resched_write(dreq);
 661
 662         spin_lock(&dreq->lock);
 663         if (dreq->flags != NFS_ODIRECT_DONE)
 664                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 665         spin_unlock(&dreq->lock);
 666         nfs_mark_request_commit(req, NULL, cinfo, 0);
 667 }
 668
 669 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 670         .completion = nfs_direct_commit_complete,
 671         .resched_write = nfs_direct_resched_write,
 672 };
 673
 674 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 675 {
 676         int res;
 677         struct nfs_commit_info cinfo;
 678         LIST_HEAD(mds_list);
 679
 680         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 681         nfs_commit_begin(cinfo.mds);
 682         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 683         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 684         if (res < 0) { /* res == -ENOMEM */
 685                 spin_lock(&dreq->lock);
 686                 if (dreq->flags == 0)
 687                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 688                 spin_unlock(&dreq->lock);
 689         }
 690         if (nfs_commit_end(cinfo.mds))
 691                 nfs_direct_write_complete(dreq);
 692 }
 693
 694 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 695 {
 696         struct nfs_commit_info cinfo;
 697         struct nfs_page *req;
 698         LIST_HEAD(reqs);
 699
 700         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 701         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 702
 703         while (!list_empty(&reqs)) {
 704                 req = nfs_list_entry(reqs.next);
 705                 nfs_list_remove_request(req);
 706                 nfs_direct_truncate_request(dreq, req);
 707                 nfs_release_request(req);
 708                 nfs_unlock_and_release_request(req);
 709         }
 710 }
 711
 712 static void nfs_direct_write_schedule_work(struct work_struct *work)
 713 {
 714         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 715         int flags = dreq->flags;
 716
 717         dreq->flags = 0;
 718         switch (flags) {
 719                 case NFS_ODIRECT_DO_COMMIT:
 720                         nfs_direct_commit_schedule(dreq);
 721                         break;
 722                 case NFS_ODIRECT_RESCHED_WRITES:
 723                         nfs_direct_write_reschedule(dreq);
 724                         break;
 725                 default:
 726                         nfs_direct_write_clear_reqs(dreq);
 727                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 728                         nfs_direct_complete(dreq);
 729         }
 730 }
 731
 732 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 733 {
 734         trace_nfs_direct_write_complete(dreq);
 735         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 736 }
 737
 738 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 739 {
 740         struct nfs_direct_req *dreq = hdr->dreq;
 741         struct nfs_commit_info cinfo;
 742         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 743         int flags = NFS_ODIRECT_DONE;
 744
 745         trace_nfs_direct_write_completion(dreq);
 746
 747         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 748
 749         spin_lock(&dreq->lock);
 750         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 751                 spin_unlock(&dreq->lock);
 752                 goto out_put;
 753         }
 754
 755         nfs_direct_count_bytes(dreq, hdr);
 756         if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
 757             !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 758                 if (!dreq->flags)
 759                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
 760                 flags = dreq->flags;
 761         }
 762         spin_unlock(&dreq->lock);
 763
 764         while (!list_empty(&hdr->pages)) {
 765
 766                 req = nfs_list_entry(hdr->pages.next);
 767                 nfs_list_remove_request(req);
 768                 if (flags == NFS_ODIRECT_DO_COMMIT) {
 769                         kref_get(&req->wb_kref);
 770                         memcpy(&req->wb_verf, &hdr->verf.verifier,
 771                                sizeof(req->wb_verf));
 772                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 773                                 hdr->ds_commit_idx);
 774                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 775                         kref_get(&req->wb_kref);
 776                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 777                 }
 778                 nfs_unlock_and_release_request(req);
 779         }
 780
 781 out_put:
 782         if (put_dreq(dreq))
 783                 nfs_direct_write_complete(dreq);
 784         hdr->release(hdr);
 785 }
 786
 787 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 788 {
 789         struct nfs_page *req;
 790
 791         while (!list_empty(head)) {
 792                 req = nfs_list_entry(head->next);
 793                 nfs_list_remove_request(req);
 794                 nfs_unlock_and_release_request(req);
 795         }
 796 }
 797
 798 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 799 {
 800         struct nfs_direct_req *dreq = hdr->dreq;
 801         struct nfs_page *req;
 802         struct nfs_commit_info cinfo;
 803
 804         trace_nfs_direct_write_reschedule_io(dreq);
 805
 806         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 807         spin_lock(&dreq->lock);
 808         if (dreq->error == 0)
 809                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 810         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 811         spin_unlock(&dreq->lock);
 812         while (!list_empty(&hdr->pages)) {
 813                 req = nfs_list_entry(hdr->pages.next);
 814                 nfs_list_remove_request(req);
 815                 nfs_unlock_request(req);
 816                 nfs_mark_request_commit(req, NULL, &cinfo, 0);
 817         }
 818 }
 819
 820 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 821         .error_cleanup = nfs_write_sync_pgio_error,
 822         .init_hdr = nfs_direct_pgio_init,
 823         .completion = nfs_direct_write_completion,
 824         .reschedule_io = nfs_direct_write_reschedule_io,
 825 };
 826
 827
 828 /*
 829  * NB: Return the value of the first error return code.  Subsequent
 830  *     errors after the first one are ignored.
 831  */
 832 /*
 833  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 834  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 835  * bail and stop sending more writes.  Write length accounting is
 836  * handled automatically by nfs_direct_write_result().  Otherwise, if
 837  * no requests have been sent, just return an error.
 838  */
 839 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 840                                                struct iov_iter *iter,
 841                                                loff_t pos, int ioflags)
 842 {
 843         struct nfs_pageio_descriptor desc;
 844         struct inode *inode = dreq->inode;
 845         struct nfs_commit_info cinfo;
 846         ssize_t result = 0;
 847         size_t requested_bytes = 0;
 848         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 849         bool defer = false;
 850
 851         trace_nfs_direct_write_schedule_iovec(dreq);
 852
 853         nfs_pageio_init_write(&desc, inode, ioflags, false,
 854                               &nfs_direct_write_completion_ops);
 855         desc.pg_dreq = dreq;
 856         get_dreq(dreq);
 857         inode_dio_begin(inode);
 858
 859         NFS_I(inode)->write_io += iov_iter_count(iter);
 860         while (iov_iter_count(iter)) {
 861                 struct page **pagevec;
 862                 size_t bytes;
 863                 size_t pgbase;
 864                 unsigned npages, i;
 865
 866                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
 867                                                   wsize, &pgbase);
 868                 if (result < 0)
 869                         break;
 870
 871                 bytes = result;
 872                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 873                 for (i = 0; i < npages; i++) {
 874                         struct nfs_page *req;
 875                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 876
 877                         req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 878                                                         pgbase, pos, req_len);
 879                         if (IS_ERR(req)) {
 880                                 result = PTR_ERR(req);
 881                                 break;
 882                         }
 883
 884                         if (desc.pg_error < 0) {
 885                                 nfs_free_request(req);
 886                                 result = desc.pg_error;
 887                                 break;
 888                         }
 889
 890                         pgbase = 0;
 891                         bytes -= req_len;
 892                         requested_bytes += req_len;
 893                         pos += req_len;
 894
 895                         if (defer) {
 896                                 nfs_mark_request_commit(req, NULL, &cinfo, 0);
 897                                 continue;
 898                         }
 899
 900                         nfs_lock_request(req);
 901                         if (nfs_pageio_add_request(&desc, req))
 902                                 continue;
 903
 904                         /* Exit on hard errors */
 905                         if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
 906                                 result = desc.pg_error;
 907                                 nfs_unlock_and_release_request(req);
 908                                 break;
 909                         }
 910
 911                         /* If the error is soft, defer remaining requests */
 912                         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 913                         spin_lock(&dreq->lock);
 914                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 915                         spin_unlock(&dreq->lock);
 916                         nfs_unlock_request(req);
 917                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 918                         desc.pg_error = 0;
 919                         defer = true;
 920                 }
 921                 nfs_direct_release_pages(pagevec, npages);
 922                 kvfree(pagevec);
 923                 if (result < 0)
 924                         break;
 925         }
 926         nfs_pageio_complete(&desc);
 927
 928         /*
 929          * If no bytes were started, return the error, and let the
 930          * generic layer handle the completion.
 931          */
 932         if (requested_bytes == 0) {
 933                 inode_dio_end(inode);
 934                 nfs_direct_req_release(dreq);
 935                 return result < 0 ? result : -EIO;
 936         }
 937
 938         if (put_dreq(dreq))
 939                 nfs_direct_write_complete(dreq);
 940         return requested_bytes;
 941 }
 942
 943 /**
 944  * nfs_file_direct_write - file direct write operation for NFS files
 945  * @iocb: target I/O control block
 946  * @iter: vector of user buffers from which to write data
 947  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 948  *
 949  * We use this function for direct writes instead of calling
 950  * generic_file_aio_write() in order to avoid taking the inode
 951  * semaphore and updating the i_size.  The NFS server will set
 952  * the new i_size and this client must read the updated size
 953  * back into its cache.  We let the server do generic write
 954  * parameter checking and report problems.
 955  *
 956  * We eliminate local atime updates, see direct read above.
 957  *
 958  * We avoid unnecessary page cache invalidations for normal cached
 959  * readers of this file.
 960  *
 961  * Note that O_APPEND is not supported for NFS direct writes, as there
 962  * is no atomic O_APPEND write facility in the NFS protocol.
 963  */
 964 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 965                               bool swap)
 966 {
 967         ssize_t result, requested;
 968         size_t count;
 969         struct file *file = iocb->ki_filp;
 970         struct address_space *mapping = file->f_mapping;
 971         struct inode *inode = mapping->host;
 972         struct nfs_direct_req *dreq;
 973         struct nfs_lock_context *l_ctx;
 974         loff_t pos, end;
 975
 976         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 977                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 978
 979         if (swap)
 980                 /* bypass generic checks */
 981                 result =  iov_iter_count(iter);
 982         else
 983                 result = generic_write_checks(iocb, iter);
 984         if (result <= 0)
 985                 return result;
 986         count = result;
 987         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 988
 989         pos = iocb->ki_pos;
 990         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 991
 992         task_io_account_write(count);
 993
 994         result = -ENOMEM;
 995         dreq = nfs_direct_req_alloc();
 996         if (!dreq)
 997                 goto out;
 998
 999         dreq->inode = inode;
1000         dreq->max_count = count;
1001         dreq->io_start = pos;
1002         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1003         l_ctx = nfs_get_lock_context(dreq->ctx);
1004         if (IS_ERR(l_ctx)) {
1005                 result = PTR_ERR(l_ctx);
1006                 nfs_direct_req_release(dreq);
1007                 goto out_release;
1008         }
1009         dreq->l_ctx = l_ctx;
1010         if (!is_sync_kiocb(iocb))
1011                 dreq->iocb = iocb;
1012         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
1013
1014         if (swap) {
1015                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1016                                                             FLUSH_STABLE);
1017         } else {
1018                 result = nfs_start_io_direct(inode);
1019                 if (result) {
1020                         /* release the reference that would usually be
1021                          * consumed by nfs_direct_write_schedule_iovec()
1022                          */
1023                         nfs_direct_req_release(dreq);
1024                         goto out_release;
1025                 }
1026
1027                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1028                                                             FLUSH_COND_STABLE);
1029
1030                 if (mapping->nrpages) {
1031                         invalidate_inode_pages2_range(mapping,
1032                                                       pos >> PAGE_SHIFT, end);
1033                 }
1034
1035                 nfs_end_io_direct(inode);
1036         }
1037
1038         if (requested > 0) {
1039                 result = nfs_direct_wait(dreq);
1040                 if (result > 0) {
1041                         requested -= result;
1042                         iocb->ki_pos = pos + result;
1043                         /* XXX: should check the generic_write_sync retval */
1044                         generic_write_sync(iocb, result);
1045                 }
1046                 iov_iter_revert(iter, requested);
1047         } else {
1048                 result = requested;
1049         }
1050         nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1051 out_release:
1052         nfs_direct_req_release(dreq);
1053 out:
1054         return result;
1055 }
1056
1057 /**
1058  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1059  *
1060  */
1061 int __init nfs_init_directcache(void)
1062 {
1063         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1064                                                 sizeof(struct nfs_direct_req),
1065                                                 0, SLAB_RECLAIM_ACCOUNT,
1066                                                 NULL);
1067         if (nfs_direct_cachep == NULL)
1068                 return -ENOMEM;
1069
1070         return 0;
1071 }
1072
1073 /**
1074  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1075  *
1076  */
1077 void nfs_destroy_directcache(void)
1078 {
1079         kmem_cache_destroy(nfs_direct_cachep);
1080 }