fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  *
  38  */
  39
  40 #include <linux/config.h>
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/smp_lock.h>
  45 #include <linux/file.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/kref.h>
  48
  49 #include <linux/nfs_fs.h>
  50 #include <linux/nfs_page.h>
  51 #include <linux/sunrpc/clnt.h>
  52
  53 #include <asm/system.h>
  54 #include <asm/uaccess.h>
  55 #include <asm/atomic.h>
  56
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60 #define MAX_DIRECTIO_SIZE       (4096UL << PAGE_SHIFT)
  61
  62 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty);
  63 static kmem_cache_t *nfs_direct_cachep;
  64
  65 /*
  66  * This represents a set of asynchronous requests that we're waiting on
  67  */
  68 struct nfs_direct_req {
  69         struct kref             kref;           /* release manager */
  70         struct list_head        list;           /* nfs_read_data structs */
  71         struct file *           filp;           /* file descriptor */
  72         struct kiocb *          iocb;           /* controlling i/o request */
  73         wait_queue_head_t       wait;           /* wait for i/o completion */
  74         struct inode *          inode;          /* target file of I/O */
  75         struct page **          pages;          /* pages in our buffer */
  76         unsigned int            npages;         /* count of pages */
  77         atomic_t                complete,       /* i/os we're waiting for */
  78                                 count,          /* bytes actually processed */
  79                                 error;          /* any reported error */
  80 };
  81
  82
  83 /**
  84  * nfs_direct_IO - NFS address space operation for direct I/O
  85  * @rw: direction (read or write)
  86  * @iocb: target I/O control block
  87  * @iov: array of vectors that define I/O buffer
  88  * @pos: offset in file to begin the operation
  89  * @nr_segs: size of iovec array
  90  *
  91  * The presence of this routine in the address space ops vector means
  92  * the NFS client supports direct I/O.  However, we shunt off direct
  93  * read and write requests before the VFS gets them, so this method
  94  * should never be called.
  95  */
  96 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
  97 {
  98         struct dentry *dentry = iocb->ki_filp->f_dentry;
  99
 100         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 101                         dentry->d_name.name, (long long) pos, nr_segs);
 102
 103         return -EINVAL;
 104 }
 105
 106 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 107 {
 108         int result = -ENOMEM;
 109         unsigned long page_count;
 110         size_t array_size;
 111
 112         /* set an arbitrary limit to prevent type overflow */
 113         /* XXX: this can probably be as large as INT_MAX */
 114         if (size > MAX_DIRECTIO_SIZE) {
 115                 *pages = NULL;
 116                 return -EFBIG;
 117         }
 118
 119         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 120         page_count -= user_addr >> PAGE_SHIFT;
 121
 122         array_size = (page_count * sizeof(struct page *));
 123         *pages = kmalloc(array_size, GFP_KERNEL);
 124         if (*pages) {
 125                 down_read(&current->mm->mmap_sem);
 126                 result = get_user_pages(current, current->mm, user_addr,
 127                                         page_count, (rw == READ), 0,
 128                                         *pages, NULL);
 129                 up_read(&current->mm->mmap_sem);
 130                 /*
 131                  * If we got fewer pages than expected from get_user_pages(),
 132                  * the user buffer runs off the end of a mapping; return EFAULT.
 133                  */
 134                 if (result >= 0 && result < page_count) {
 135                         nfs_free_user_pages(*pages, result, 0);
 136                         *pages = NULL;
 137                         result = -EFAULT;
 138                 }
 139         }
 140         return result;
 141 }
 142
 143 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 144 {
 145         int i;
 146         for (i = 0; i < npages; i++) {
 147                 struct page *page = pages[i];
 148                 if (do_dirty && !PageCompound(page))
 149                         set_page_dirty_lock(page);
 150                 page_cache_release(page);
 151         }
 152         kfree(pages);
 153 }
 154
 155 static void nfs_direct_req_release(struct kref *kref)
 156 {
 157         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 158         kmem_cache_free(nfs_direct_cachep, dreq);
 159 }
 160
 161 /*
 162  * Note we also set the number of requests we have in the dreq when we are
 163  * done.  This prevents races with I/O completion so we will always wait
 164  * until all requests have been dispatched and completed.
 165  */
 166 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 167 {
 168         struct list_head *list;
 169         struct nfs_direct_req *dreq;
 170         unsigned int reads = 0;
 171         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 172
 173         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 174         if (!dreq)
 175                 return NULL;
 176
 177         kref_init(&dreq->kref);
 178         init_waitqueue_head(&dreq->wait);
 179         INIT_LIST_HEAD(&dreq->list);
 180         atomic_set(&dreq->count, 0);
 181         atomic_set(&dreq->error, 0);
 182
 183         list = &dreq->list;
 184         for(;;) {
 185                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 186
 187                 if (unlikely(!data)) {
 188                         while (!list_empty(list)) {
 189                                 data = list_entry(list->next,
 190                                                   struct nfs_read_data, pages);
 191                                 list_del(&data->pages);
 192                                 nfs_readdata_free(data);
 193                         }
 194                         kref_put(&dreq->kref, nfs_direct_req_release);
 195                         return NULL;
 196                 }
 197
 198                 INIT_LIST_HEAD(&data->pages);
 199                 list_add(&data->pages, list);
 200
 201                 data->req = (struct nfs_page *) dreq;
 202                 reads++;
 203                 if (nbytes <= rsize)
 204                         break;
 205                 nbytes -= rsize;
 206         }
 207         kref_get(&dreq->kref);
 208         atomic_set(&dreq->complete, reads);
 209         return dreq;
 210 }
 211
 212 /*
 213  * We must hold a reference to all the pages in this direct read request
 214  * until the RPCs complete.  This could be long *after* we are woken up in
 215  * nfs_direct_read_wait (for instance, if someone hits ^C on a slow server).
 216  */
 217 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 218 {
 219         struct nfs_read_data *data = calldata;
 220         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 221
 222         if (nfs_readpage_result(task, data) != 0)
 223                 return;
 224         if (likely(task->tk_status >= 0))
 225                 atomic_add(data->res.count, &dreq->count);
 226         else
 227                 atomic_set(&dreq->error, task->tk_status);
 228
 229         if (unlikely(atomic_dec_and_test(&dreq->complete))) {
 230                 nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 231                 wake_up(&dreq->wait);
 232                 kref_put(&dreq->kref, nfs_direct_req_release);
 233         }
 234 }
 235
 236 static const struct rpc_call_ops nfs_read_direct_ops = {
 237         .rpc_call_done = nfs_direct_read_result,
 238         .rpc_release = nfs_readdata_release,
 239 };
 240
 241 /*
 242  * For each nfs_read_data struct that was allocated on the list, dispatch
 243  * an NFS READ operation
 244  */
 245 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t file_offset)
 246 {
 247         struct file *file = dreq->filp;
 248         struct inode *inode = file->f_mapping->host;
 249         struct nfs_open_context *ctx = (struct nfs_open_context *)
 250                                                         file->private_data;
 251         struct list_head *list = &dreq->list;
 252         struct page **pages = dreq->pages;
 253         size_t rsize = NFS_SERVER(inode)->rsize;
 254         unsigned int curpage, pgbase;
 255
 256         curpage = 0;
 257         pgbase = user_addr & ~PAGE_MASK;
 258         do {
 259                 struct nfs_read_data *data;
 260                 size_t bytes;
 261
 262                 bytes = rsize;
 263                 if (count < rsize)
 264                         bytes = count;
 265
 266                 data = list_entry(list->next, struct nfs_read_data, pages);
 267                 list_del_init(&data->pages);
 268
 269                 data->inode = inode;
 270                 data->cred = ctx->cred;
 271                 data->args.fh = NFS_FH(inode);
 272                 data->args.context = ctx;
 273                 data->args.offset = file_offset;
 274                 data->args.pgbase = pgbase;
 275                 data->args.pages = &pages[curpage];
 276                 data->args.count = bytes;
 277                 data->res.fattr = &data->fattr;
 278                 data->res.eof = 0;
 279                 data->res.count = bytes;
 280
 281                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 282                                 &nfs_read_direct_ops, data);
 283                 NFS_PROTO(inode)->read_setup(data);
 284
 285                 data->task.tk_cookie = (unsigned long) inode;
 286
 287                 lock_kernel();
 288                 rpc_execute(&data->task);
 289                 unlock_kernel();
 290
 291                 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 292                                 data->task.tk_pid,
 293                                 inode->i_sb->s_id,
 294                                 (long long)NFS_FILEID(inode),
 295                                 bytes,
 296                                 (unsigned long long)data->args.offset);
 297
 298                 file_offset += bytes;
 299                 pgbase += bytes;
 300                 curpage += pgbase >> PAGE_SHIFT;
 301                 pgbase &= ~PAGE_MASK;
 302
 303                 count -= bytes;
 304         } while (count != 0);
 305 }
 306
 307 /*
 308  * Collects and returns the final error value/byte-count.
 309  */
 310 static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
 311 {
 312         int result = 0;
 313
 314         if (intr) {
 315                 result = wait_event_interruptible(dreq->wait,
 316                                         (atomic_read(&dreq->complete) == 0));
 317         } else {
 318                 wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
 319         }
 320
 321         if (!result)
 322                 result = atomic_read(&dreq->error);
 323         if (!result)
 324                 result = atomic_read(&dreq->count);
 325
 326         kref_put(&dreq->kref, nfs_direct_req_release);
 327         return (ssize_t) result;
 328 }
 329
 330 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t file_offset, struct page **pages, unsigned int nr_pages)
 331 {
 332         ssize_t result;
 333         sigset_t oldset;
 334         struct inode *inode = iocb->ki_filp->f_mapping->host;
 335         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 336         struct nfs_direct_req *dreq;
 337
 338         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 339         if (!dreq)
 340                 return -ENOMEM;
 341
 342         dreq->pages = pages;
 343         dreq->npages = nr_pages;
 344         dreq->inode = inode;
 345         dreq->filp = iocb->ki_filp;
 346
 347         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 348         rpc_clnt_sigmask(clnt, &oldset);
 349         nfs_direct_read_schedule(dreq, user_addr, count, file_offset);
 350         result = nfs_direct_read_wait(dreq, clnt->cl_intr);
 351         rpc_clnt_sigunmask(clnt, &oldset);
 352
 353         return result;
 354 }
 355
 356 static ssize_t nfs_direct_write_seg(struct inode *inode, struct nfs_open_context *ctx, unsigned long user_addr, size_t count, loff_t file_offset, struct page **pages, int nr_pages)
 357 {
 358         const unsigned int wsize = NFS_SERVER(inode)->wsize;
 359         size_t request;
 360         int curpage, need_commit;
 361         ssize_t result, tot_bytes;
 362         struct nfs_writeverf first_verf;
 363         struct nfs_write_data *wdata;
 364
 365         wdata = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
 366         if (!wdata)
 367                 return -ENOMEM;
 368
 369         wdata->inode = inode;
 370         wdata->cred = ctx->cred;
 371         wdata->args.fh = NFS_FH(inode);
 372         wdata->args.context = ctx;
 373         wdata->args.stable = NFS_UNSTABLE;
 374         if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
 375                 wdata->args.stable = NFS_FILE_SYNC;
 376         wdata->res.fattr = &wdata->fattr;
 377         wdata->res.verf = &wdata->verf;
 378
 379         nfs_begin_data_update(inode);
 380 retry:
 381         need_commit = 0;
 382         tot_bytes = 0;
 383         curpage = 0;
 384         request = count;
 385         wdata->args.pgbase = user_addr & ~PAGE_MASK;
 386         wdata->args.offset = file_offset;
 387         do {
 388                 wdata->args.count = request;
 389                 if (wdata->args.count > wsize)
 390                         wdata->args.count = wsize;
 391                 wdata->args.pages = &pages[curpage];
 392
 393                 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
 394                         wdata->args.count, (long long) wdata->args.offset,
 395                         user_addr + tot_bytes, wdata->args.pgbase, curpage);
 396
 397                 lock_kernel();
 398                 result = NFS_PROTO(inode)->write(wdata);
 399                 unlock_kernel();
 400
 401                 if (result <= 0) {
 402                         if (tot_bytes > 0)
 403                                 break;
 404                         goto out;
 405                 }
 406
 407                 if (tot_bytes == 0)
 408                         memcpy(&first_verf.verifier, &wdata->verf.verifier,
 409                                                 sizeof(first_verf.verifier));
 410                 if (wdata->verf.committed != NFS_FILE_SYNC) {
 411                         need_commit = 1;
 412                         if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
 413                                         sizeof(first_verf.verifier)))
 414                                 goto sync_retry;
 415                 }
 416
 417                 tot_bytes += result;
 418
 419                 /* in case of a short write: stop now, let the app recover */
 420                 if (result < wdata->args.count)
 421                         break;
 422
 423                 wdata->args.offset += result;
 424                 wdata->args.pgbase += result;
 425                 curpage += wdata->args.pgbase >> PAGE_SHIFT;
 426                 wdata->args.pgbase &= ~PAGE_MASK;
 427                 request -= result;
 428         } while (request != 0);
 429
 430         /*
 431          * Commit data written so far, even in the event of an error
 432          */
 433         if (need_commit) {
 434                 wdata->args.count = tot_bytes;
 435                 wdata->args.offset = file_offset;
 436
 437                 lock_kernel();
 438                 result = NFS_PROTO(inode)->commit(wdata);
 439                 unlock_kernel();
 440
 441                 if (result < 0 || memcmp(&first_verf.verifier,
 442                                          &wdata->verf.verifier,
 443                                          sizeof(first_verf.verifier)) != 0)
 444                         goto sync_retry;
 445         }
 446         result = tot_bytes;
 447
 448 out:
 449         nfs_end_data_update(inode);
 450         nfs_writedata_free(wdata);
 451         return result;
 452
 453 sync_retry:
 454         wdata->args.stable = NFS_FILE_SYNC;
 455         goto retry;
 456 }
 457
 458 /*
 459  * Upon return, generic_file_direct_IO invalidates any cached pages
 460  * that non-direct readers might access, so they will pick up these
 461  * writes immediately.
 462  */
 463 static ssize_t nfs_direct_write(struct inode *inode, struct nfs_open_context *ctx, const struct iovec *iov, loff_t file_offset, unsigned long nr_segs)
 464 {
 465         ssize_t tot_bytes = 0;
 466         unsigned long seg = 0;
 467
 468         while ((seg < nr_segs) && (tot_bytes >= 0)) {
 469                 ssize_t result;
 470                 int page_count;
 471                 struct page **pages;
 472                 const struct iovec *vec = &iov[seg++];
 473                 unsigned long user_addr = (unsigned long) vec->iov_base;
 474                 size_t size = vec->iov_len;
 475
 476                 page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
 477                 if (page_count < 0) {
 478                         nfs_free_user_pages(pages, 0, 0);
 479                         if (tot_bytes > 0)
 480                                 break;
 481                         return page_count;
 482                 }
 483
 484                 nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, size);
 485                 result = nfs_direct_write_seg(inode, ctx, user_addr, size,
 486                                 file_offset, pages, page_count);
 487                 nfs_free_user_pages(pages, page_count, 0);
 488
 489                 if (result <= 0) {
 490                         if (tot_bytes > 0)
 491                                 break;
 492                         return result;
 493                 }
 494                 nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result);
 495                 tot_bytes += result;
 496                 file_offset += result;
 497                 if (result < size)
 498                         break;
 499         }
 500         return tot_bytes;
 501 }
 502
 503 /**
 504  * nfs_file_direct_read - file direct read operation for NFS files
 505  * @iocb: target I/O control block
 506  * @buf: user's buffer into which to read data
 507  * count: number of bytes to read
 508  * pos: byte offset in file where reading starts
 509  *
 510  * We use this function for direct reads instead of calling
 511  * generic_file_aio_read() in order to avoid gfar's check to see if
 512  * the request starts before the end of the file.  For that check
 513  * to work, we must generate a GETATTR before each direct read, and
 514  * even then there is a window between the GETATTR and the subsequent
 515  * READ where the file size could change.  So our preference is simply
 516  * to do all reads the application wants, and the server will take
 517  * care of managing the end of file boundary.
 518  *
 519  * This function also eliminates unnecessarily updating the file's
 520  * atime locally, as the NFS server sets the file's atime, and this
 521  * client must read the updated atime from the server back into its
 522  * cache.
 523  */
 524 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 525 {
 526         ssize_t retval = -EINVAL;
 527         int page_count;
 528         struct page **pages;
 529         struct file *file = iocb->ki_filp;
 530         struct address_space *mapping = file->f_mapping;
 531
 532         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 533                 file->f_dentry->d_parent->d_name.name,
 534                 file->f_dentry->d_name.name,
 535                 (unsigned long) count, (long long) pos);
 536
 537         if (!is_sync_kiocb(iocb))
 538                 goto out;
 539         if (count < 0)
 540                 goto out;
 541         retval = -EFAULT;
 542         if (!access_ok(VERIFY_WRITE, buf, count))
 543                 goto out;
 544         retval = 0;
 545         if (!count)
 546                 goto out;
 547
 548         retval = nfs_sync_mapping(mapping);
 549         if (retval)
 550                 goto out;
 551
 552         page_count = nfs_get_user_pages(READ, (unsigned long) buf,
 553                                                 count, &pages);
 554         if (page_count < 0) {
 555                 nfs_free_user_pages(pages, 0, 0);
 556                 retval = page_count;
 557                 goto out;
 558         }
 559
 560         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 561                                                 pages, page_count);
 562         if (retval > 0)
 563                 iocb->ki_pos = pos + retval;
 564
 565 out:
 566         return retval;
 567 }
 568
 569 /**
 570  * nfs_file_direct_write - file direct write operation for NFS files
 571  * @iocb: target I/O control block
 572  * @buf: user's buffer from which to write data
 573  * count: number of bytes to write
 574  * pos: byte offset in file where writing starts
 575  *
 576  * We use this function for direct writes instead of calling
 577  * generic_file_aio_write() in order to avoid taking the inode
 578  * semaphore and updating the i_size.  The NFS server will set
 579  * the new i_size and this client must read the updated size
 580  * back into its cache.  We let the server do generic write
 581  * parameter checking and report problems.
 582  *
 583  * We also avoid an unnecessary invocation of generic_osync_inode(),
 584  * as it is fairly meaningless to sync the metadata of an NFS file.
 585  *
 586  * We eliminate local atime updates, see direct read above.
 587  *
 588  * We avoid unnecessary page cache invalidations for normal cached
 589  * readers of this file.
 590  *
 591  * Note that O_APPEND is not supported for NFS direct writes, as there
 592  * is no atomic O_APPEND write facility in the NFS protocol.
 593  */
 594 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 595 {
 596         ssize_t retval;
 597         struct file *file = iocb->ki_filp;
 598         struct nfs_open_context *ctx =
 599                         (struct nfs_open_context *) file->private_data;
 600         struct address_space *mapping = file->f_mapping;
 601         struct inode *inode = mapping->host;
 602         struct iovec iov = {
 603                 .iov_base = (char __user *)buf,
 604         };
 605
 606         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 607                 file->f_dentry->d_parent->d_name.name,
 608                 file->f_dentry->d_name.name,
 609                 (unsigned long) count, (long long) pos);
 610
 611         retval = -EINVAL;
 612         if (!is_sync_kiocb(iocb))
 613                 goto out;
 614
 615         retval = generic_write_checks(file, &pos, &count, 0);
 616         if (retval)
 617                 goto out;
 618
 619         retval = -EINVAL;
 620         if ((ssize_t) count < 0)
 621                 goto out;
 622         retval = 0;
 623         if (!count)
 624                 goto out;
 625         iov.iov_len = count,
 626
 627         retval = -EFAULT;
 628         if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
 629                 goto out;
 630
 631         retval = nfs_sync_mapping(mapping);
 632         if (retval)
 633                 goto out;
 634
 635         retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
 636         if (mapping->nrpages)
 637                 invalidate_inode_pages2(mapping);
 638         if (retval > 0)
 639                 iocb->ki_pos = pos + retval;
 640
 641 out:
 642         return retval;
 643 }
 644
 645 int nfs_init_directcache(void)
 646 {
 647         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 648                                                 sizeof(struct nfs_direct_req),
 649                                                 0, SLAB_RECLAIM_ACCOUNT,
 650                                                 NULL, NULL);
 651         if (nfs_direct_cachep == NULL)
 652                 return -ENOMEM;
 653
 654         return 0;
 655 }
 656
 657 void nfs_destroy_directcache(void)
 658 {
 659         if (kmem_cache_destroy(nfs_direct_cachep))
 660                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 661 }