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                 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_page_create_from_page(dreq->ctx, pagevec[i],
 347                                                         pgbase, pos, req_len);
 348                         if (IS_ERR(req)) {
 349                                 result = PTR_ERR(req);
 350                                 break;
 351                         }
 352                         if (!nfs_pageio_add_request(&desc, req)) {
 353                                 result = desc.pg_error;
 354                                 nfs_release_request(req);
 355                                 break;
 356                         }
 357                         pgbase = 0;
 358                         bytes -= req_len;
 359                         requested_bytes += req_len;
 360                         pos += req_len;
 361                         dreq->bytes_left -= req_len;
 362                 }
 363                 nfs_direct_release_pages(pagevec, npages);
 364                 kvfree(pagevec);
 365                 if (result < 0)
 366                         break;
 367         }
 368
 369         nfs_pageio_complete(&desc);
 370
 371         /*
 372          * If no bytes were started, return the error, and let the
 373          * generic layer handle the completion.
 374          */
 375         if (requested_bytes == 0) {
 376                 inode_dio_end(inode);
 377                 nfs_direct_req_release(dreq);
 378                 return result < 0 ? result : -EIO;
 379         }
 380
 381         if (put_dreq(dreq))
 382                 nfs_direct_complete(dreq);
 383         return requested_bytes;
 384 }
 385
 386 /**
 387  * nfs_file_direct_read - file direct read operation for NFS files
 388  * @iocb: target I/O control block
 389  * @iter: vector of user buffers into which to read data
 390  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 391  *
 392  * We use this function for direct reads instead of calling
 393  * generic_file_aio_read() in order to avoid gfar's check to see if
 394  * the request starts before the end of the file.  For that check
 395  * to work, we must generate a GETATTR before each direct read, and
 396  * even then there is a window between the GETATTR and the subsequent
 397  * READ where the file size could change.  Our preference is simply
 398  * to do all reads the application wants, and the server will take
 399  * care of managing the end of file boundary.
 400  *
 401  * This function also eliminates unnecessarily updating the file's
 402  * atime locally, as the NFS server sets the file's atime, and this
 403  * client must read the updated atime from the server back into its
 404  * cache.
 405  */
 406 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 407                              bool swap)
 408 {
 409         struct file *file = iocb->ki_filp;
 410         struct address_space *mapping = file->f_mapping;
 411         struct inode *inode = mapping->host;
 412         struct nfs_direct_req *dreq;
 413         struct nfs_lock_context *l_ctx;
 414         ssize_t result, requested;
 415         size_t count = iov_iter_count(iter);
 416         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 417
 418         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 419                 file, count, (long long) iocb->ki_pos);
 420
 421         result = 0;
 422         if (!count)
 423                 goto out;
 424
 425         task_io_account_read(count);
 426
 427         result = -ENOMEM;
 428         dreq = nfs_direct_req_alloc();
 429         if (dreq == NULL)
 430                 goto out;
 431
 432         dreq->inode = inode;
 433         dreq->bytes_left = dreq->max_count = count;
 434         dreq->io_start = iocb->ki_pos;
 435         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 436         l_ctx = nfs_get_lock_context(dreq->ctx);
 437         if (IS_ERR(l_ctx)) {
 438                 result = PTR_ERR(l_ctx);
 439                 nfs_direct_req_release(dreq);
 440                 goto out_release;
 441         }
 442         dreq->l_ctx = l_ctx;
 443         if (!is_sync_kiocb(iocb))
 444                 dreq->iocb = iocb;
 445
 446         if (user_backed_iter(iter))
 447                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 448
 449         if (!swap)
 450                 nfs_start_io_direct(inode);
 451
 452         NFS_I(inode)->read_io += count;
 453         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 454
 455         if (!swap)
 456                 nfs_end_io_direct(inode);
 457
 458         if (requested > 0) {
 459                 result = nfs_direct_wait(dreq);
 460                 if (result > 0) {
 461                         requested -= result;
 462                         iocb->ki_pos += result;
 463                 }
 464                 iov_iter_revert(iter, requested);
 465         } else {
 466                 result = requested;
 467         }
 468
 469 out_release:
 470         nfs_direct_req_release(dreq);
 471 out:
 472         return result;
 473 }
 474
 475 static void
 476 nfs_direct_join_group(struct list_head *list, struct inode *inode)
 477 {
 478         struct nfs_page *req, *next;
 479
 480         list_for_each_entry(req, list, wb_list) {
 481                 if (req->wb_head != req || req->wb_this_page == req)
 482                         continue;
 483                 for (next = req->wb_this_page;
 484                                 next != req->wb_head;
 485                                 next = next->wb_this_page) {
 486                         nfs_list_remove_request(next);
 487                         nfs_release_request(next);
 488                 }
 489                 nfs_join_page_group(req, inode);
 490         }
 491 }
 492
 493 static void
 494 nfs_direct_write_scan_commit_list(struct inode *inode,
 495                                   struct list_head *list,
 496                                   struct nfs_commit_info *cinfo)
 497 {
 498         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 499         pnfs_recover_commit_reqs(list, cinfo);
 500         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 501         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 502 }
 503
 504 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 505 {
 506         struct nfs_pageio_descriptor desc;
 507         struct nfs_page *req, *tmp;
 508         LIST_HEAD(reqs);
 509         struct nfs_commit_info cinfo;
 510         LIST_HEAD(failed);
 511
 512         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 513         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 514
 515         nfs_direct_join_group(&reqs, dreq->inode);
 516
 517         dreq->count = 0;
 518         dreq->max_count = 0;
 519         list_for_each_entry(req, &reqs, wb_list)
 520                 dreq->max_count += req->wb_bytes;
 521         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 522         get_dreq(dreq);
 523
 524         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 525                               &nfs_direct_write_completion_ops);
 526         desc.pg_dreq = dreq;
 527
 528         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 529                 /* Bump the transmission count */
 530                 req->wb_nio++;
 531                 if (!nfs_pageio_add_request(&desc, req)) {
 532                         nfs_list_move_request(req, &failed);
 533                         spin_lock(&cinfo.inode->i_lock);
 534                         dreq->flags = 0;
 535                         if (desc.pg_error < 0)
 536                                 dreq->error = desc.pg_error;
 537                         else
 538                                 dreq->error = -EIO;
 539                         spin_unlock(&cinfo.inode->i_lock);
 540                 }
 541                 nfs_release_request(req);
 542         }
 543         nfs_pageio_complete(&desc);
 544
 545         while (!list_empty(&failed)) {
 546                 req = nfs_list_entry(failed.next);
 547                 nfs_list_remove_request(req);
 548                 nfs_unlock_and_release_request(req);
 549         }
 550
 551         if (put_dreq(dreq))
 552                 nfs_direct_write_complete(dreq);
 553 }
 554
 555 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 556 {
 557         const struct nfs_writeverf *verf = data->res.verf;
 558         struct nfs_direct_req *dreq = data->dreq;
 559         struct nfs_commit_info cinfo;
 560         struct nfs_page *req;
 561         int status = data->task.tk_status;
 562
 563         trace_nfs_direct_commit_complete(dreq);
 564
 565         if (status < 0) {
 566                 /* Errors in commit are fatal */
 567                 dreq->error = status;
 568                 dreq->max_count = 0;
 569                 dreq->count = 0;
 570                 dreq->flags = NFS_ODIRECT_DONE;
 571         } else {
 572                 status = dreq->error;
 573         }
 574
 575         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 576
 577         while (!list_empty(&data->pages)) {
 578                 req = nfs_list_entry(data->pages.next);
 579                 nfs_list_remove_request(req);
 580                 if (status >= 0 && !nfs_write_match_verf(verf, req)) {
 581                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 582                         /*
 583                          * Despite the reboot, the write was successful,
 584                          * so reset wb_nio.
 585                          */
 586                         req->wb_nio = 0;
 587                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 588                 } else /* Error or match */
 589                         nfs_release_request(req);
 590                 nfs_unlock_and_release_request(req);
 591         }
 592
 593         if (nfs_commit_end(cinfo.mds))
 594                 nfs_direct_write_complete(dreq);
 595 }
 596
 597 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 598                 struct nfs_page *req)
 599 {
 600         struct nfs_direct_req *dreq = cinfo->dreq;
 601
 602         trace_nfs_direct_resched_write(dreq);
 603
 604         spin_lock(&dreq->lock);
 605         if (dreq->flags != NFS_ODIRECT_DONE)
 606                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 607         spin_unlock(&dreq->lock);
 608         nfs_mark_request_commit(req, NULL, cinfo, 0);
 609 }
 610
 611 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 612         .completion = nfs_direct_commit_complete,
 613         .resched_write = nfs_direct_resched_write,
 614 };
 615
 616 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 617 {
 618         int res;
 619         struct nfs_commit_info cinfo;
 620         LIST_HEAD(mds_list);
 621
 622         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 623         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 624         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 625         if (res < 0) /* res == -ENOMEM */
 626                 nfs_direct_write_reschedule(dreq);
 627 }
 628
 629 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 630 {
 631         struct nfs_commit_info cinfo;
 632         struct nfs_page *req;
 633         LIST_HEAD(reqs);
 634
 635         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 636         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 637
 638         while (!list_empty(&reqs)) {
 639                 req = nfs_list_entry(reqs.next);
 640                 nfs_list_remove_request(req);
 641                 nfs_release_request(req);
 642                 nfs_unlock_and_release_request(req);
 643         }
 644 }
 645
 646 static void nfs_direct_write_schedule_work(struct work_struct *work)
 647 {
 648         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 649         int flags = dreq->flags;
 650
 651         dreq->flags = 0;
 652         switch (flags) {
 653                 case NFS_ODIRECT_DO_COMMIT:
 654                         nfs_direct_commit_schedule(dreq);
 655                         break;
 656                 case NFS_ODIRECT_RESCHED_WRITES:
 657                         nfs_direct_write_reschedule(dreq);
 658                         break;
 659                 default:
 660                         nfs_direct_write_clear_reqs(dreq);
 661                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 662                         nfs_direct_complete(dreq);
 663         }
 664 }
 665
 666 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 667 {
 668         trace_nfs_direct_write_complete(dreq);
 669         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 670 }
 671
 672 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 673 {
 674         struct nfs_direct_req *dreq = hdr->dreq;
 675         struct nfs_commit_info cinfo;
 676         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 677         int flags = NFS_ODIRECT_DONE;
 678
 679         trace_nfs_direct_write_completion(dreq);
 680
 681         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 682
 683         spin_lock(&dreq->lock);
 684         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 685                 spin_unlock(&dreq->lock);
 686                 goto out_put;
 687         }
 688
 689         nfs_direct_count_bytes(dreq, hdr);
 690         if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags)) {
 691                 if (!dreq->flags)
 692                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
 693                 flags = dreq->flags;
 694         }
 695         spin_unlock(&dreq->lock);
 696
 697         while (!list_empty(&hdr->pages)) {
 698
 699                 req = nfs_list_entry(hdr->pages.next);
 700                 nfs_list_remove_request(req);
 701                 if (flags == NFS_ODIRECT_DO_COMMIT) {
 702                         kref_get(&req->wb_kref);
 703                         memcpy(&req->wb_verf, &hdr->verf.verifier,
 704                                sizeof(req->wb_verf));
 705                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 706                                 hdr->ds_commit_idx);
 707                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 708                         kref_get(&req->wb_kref);
 709                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 710                 }
 711                 nfs_unlock_and_release_request(req);
 712         }
 713
 714 out_put:
 715         if (put_dreq(dreq))
 716                 nfs_direct_write_complete(dreq);
 717         hdr->release(hdr);
 718 }
 719
 720 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 721 {
 722         struct nfs_page *req;
 723
 724         while (!list_empty(head)) {
 725                 req = nfs_list_entry(head->next);
 726                 nfs_list_remove_request(req);
 727                 nfs_unlock_and_release_request(req);
 728         }
 729 }
 730
 731 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 732 {
 733         struct nfs_direct_req *dreq = hdr->dreq;
 734
 735         trace_nfs_direct_write_reschedule_io(dreq);
 736
 737         spin_lock(&dreq->lock);
 738         if (dreq->error == 0) {
 739                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 740                 /* fake unstable write to let common nfs resend pages */
 741                 hdr->verf.committed = NFS_UNSTABLE;
 742                 hdr->good_bytes = hdr->args.offset + hdr->args.count -
 743                         hdr->io_start;
 744         }
 745         spin_unlock(&dreq->lock);
 746 }
 747
 748 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 749         .error_cleanup = nfs_write_sync_pgio_error,
 750         .init_hdr = nfs_direct_pgio_init,
 751         .completion = nfs_direct_write_completion,
 752         .reschedule_io = nfs_direct_write_reschedule_io,
 753 };
 754
 755
 756 /*
 757  * NB: Return the value of the first error return code.  Subsequent
 758  *     errors after the first one are ignored.
 759  */
 760 /*
 761  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 762  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 763  * bail and stop sending more writes.  Write length accounting is
 764  * handled automatically by nfs_direct_write_result().  Otherwise, if
 765  * no requests have been sent, just return an error.
 766  */
 767 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 768                                                struct iov_iter *iter,
 769                                                loff_t pos, int ioflags)
 770 {
 771         struct nfs_pageio_descriptor desc;
 772         struct inode *inode = dreq->inode;
 773         ssize_t result = 0;
 774         size_t requested_bytes = 0;
 775         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 776
 777         trace_nfs_direct_write_schedule_iovec(dreq);
 778
 779         nfs_pageio_init_write(&desc, inode, ioflags, false,
 780                               &nfs_direct_write_completion_ops);
 781         desc.pg_dreq = dreq;
 782         get_dreq(dreq);
 783         inode_dio_begin(inode);
 784
 785         NFS_I(inode)->write_io += iov_iter_count(iter);
 786         while (iov_iter_count(iter)) {
 787                 struct page **pagevec;
 788                 size_t bytes;
 789                 size_t pgbase;
 790                 unsigned npages, i;
 791
 792                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
 793                                                   wsize, &pgbase);
 794                 if (result < 0)
 795                         break;
 796
 797                 bytes = result;
 798                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 799                 for (i = 0; i < npages; i++) {
 800                         struct nfs_page *req;
 801                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 802
 803                         req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 804                                                         pgbase, pos, req_len);
 805                         if (IS_ERR(req)) {
 806                                 result = PTR_ERR(req);
 807                                 break;
 808                         }
 809
 810                         if (desc.pg_error < 0) {
 811                                 nfs_free_request(req);
 812                                 result = desc.pg_error;
 813                                 break;
 814                         }
 815
 816                         nfs_lock_request(req);
 817                         if (!nfs_pageio_add_request(&desc, req)) {
 818                                 result = desc.pg_error;
 819                                 nfs_unlock_and_release_request(req);
 820                                 break;
 821                         }
 822                         pgbase = 0;
 823                         bytes -= req_len;
 824                         requested_bytes += req_len;
 825                         pos += req_len;
 826                         dreq->bytes_left -= req_len;
 827                 }
 828                 nfs_direct_release_pages(pagevec, npages);
 829                 kvfree(pagevec);
 830                 if (result < 0)
 831                         break;
 832         }
 833         nfs_pageio_complete(&desc);
 834
 835         /*
 836          * If no bytes were started, return the error, and let the
 837          * generic layer handle the completion.
 838          */
 839         if (requested_bytes == 0) {
 840                 inode_dio_end(inode);
 841                 nfs_direct_req_release(dreq);
 842                 return result < 0 ? result : -EIO;
 843         }
 844
 845         if (put_dreq(dreq))
 846                 nfs_direct_write_complete(dreq);
 847         return requested_bytes;
 848 }
 849
 850 /**
 851  * nfs_file_direct_write - file direct write operation for NFS files
 852  * @iocb: target I/O control block
 853  * @iter: vector of user buffers from which to write data
 854  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 855  *
 856  * We use this function for direct writes instead of calling
 857  * generic_file_aio_write() in order to avoid taking the inode
 858  * semaphore and updating the i_size.  The NFS server will set
 859  * the new i_size and this client must read the updated size
 860  * back into its cache.  We let the server do generic write
 861  * parameter checking and report problems.
 862  *
 863  * We eliminate local atime updates, see direct read above.
 864  *
 865  * We avoid unnecessary page cache invalidations for normal cached
 866  * readers of this file.
 867  *
 868  * Note that O_APPEND is not supported for NFS direct writes, as there
 869  * is no atomic O_APPEND write facility in the NFS protocol.
 870  */
 871 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 872                               bool swap)
 873 {
 874         ssize_t result, requested;
 875         size_t count;
 876         struct file *file = iocb->ki_filp;
 877         struct address_space *mapping = file->f_mapping;
 878         struct inode *inode = mapping->host;
 879         struct nfs_direct_req *dreq;
 880         struct nfs_lock_context *l_ctx;
 881         loff_t pos, end;
 882
 883         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 884                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 885
 886         if (swap)
 887                 /* bypass generic checks */
 888                 result =  iov_iter_count(iter);
 889         else
 890                 result = generic_write_checks(iocb, iter);
 891         if (result <= 0)
 892                 return result;
 893         count = result;
 894         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 895
 896         pos = iocb->ki_pos;
 897         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 898
 899         task_io_account_write(count);
 900
 901         result = -ENOMEM;
 902         dreq = nfs_direct_req_alloc();
 903         if (!dreq)
 904                 goto out;
 905
 906         dreq->inode = inode;
 907         dreq->bytes_left = dreq->max_count = count;
 908         dreq->io_start = pos;
 909         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 910         l_ctx = nfs_get_lock_context(dreq->ctx);
 911         if (IS_ERR(l_ctx)) {
 912                 result = PTR_ERR(l_ctx);
 913                 nfs_direct_req_release(dreq);
 914                 goto out_release;
 915         }
 916         dreq->l_ctx = l_ctx;
 917         if (!is_sync_kiocb(iocb))
 918                 dreq->iocb = iocb;
 919         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
 920
 921         if (swap) {
 922                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 923                                                             FLUSH_STABLE);
 924         } else {
 925                 nfs_start_io_direct(inode);
 926
 927                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 928                                                             FLUSH_COND_STABLE);
 929
 930                 if (mapping->nrpages) {
 931                         invalidate_inode_pages2_range(mapping,
 932                                                       pos >> PAGE_SHIFT, end);
 933                 }
 934
 935                 nfs_end_io_direct(inode);
 936         }
 937
 938         if (requested > 0) {
 939                 result = nfs_direct_wait(dreq);
 940                 if (result > 0) {
 941                         requested -= result;
 942                         iocb->ki_pos = pos + result;
 943                         /* XXX: should check the generic_write_sync retval */
 944                         generic_write_sync(iocb, result);
 945                 }
 946                 iov_iter_revert(iter, requested);
 947         } else {
 948                 result = requested;
 949         }
 950         nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
 951 out_release:
 952         nfs_direct_req_release(dreq);
 953 out:
 954         return result;
 955 }
 956
 957 /**
 958  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 959  *
 960  */
 961 int __init nfs_init_directcache(void)
 962 {
 963         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 964                                                 sizeof(struct nfs_direct_req),
 965                                                 0, (SLAB_RECLAIM_ACCOUNT|
 966                                                         SLAB_MEM_SPREAD),
 967                                                 NULL);
 968         if (nfs_direct_cachep == NULL)
 969                 return -ENOMEM;
 970
 971         return 0;
 972 }
 973
 974 /**
 975  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 976  *
 977  */
 978 void nfs_destroy_directcache(void)
 979 {
 980         kmem_cache_destroy(nfs_direct_cachep);
 981 }