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