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  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/config.h>
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/smp_lock.h>
  46 #include <linux/file.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/kref.h>
  49
  50 #include <linux/nfs_fs.h>
  51 #include <linux/nfs_page.h>
  52 #include <linux/sunrpc/clnt.h>
  53
  54 #include <asm/system.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/atomic.h>
  57
  58 #include "iostat.h"
  59
  60 #define NFSDBG_FACILITY         NFSDBG_VFS
  61
  62 static kmem_cache_t *nfs_direct_cachep;
  63
  64 /*
  65  * This represents a set of asynchronous requests that we're waiting on
  66  */
  67 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct nfs_open_context *ctx;           /* file open context info */
  72         struct kiocb *          iocb;           /* controlling i/o request */
  73         struct inode *          inode;          /* target file of i/o */
  74
  75         /* completion state */
  76         atomic_t                io_count;       /* i/os we're waiting for */
  77         spinlock_t              lock;           /* protect completion state */
  78         ssize_t                 count,          /* bytes actually processed */
  79                                 error;          /* any reported error */
  80         struct completion       completion;     /* wait for i/o completion */
  81
  82         /* commit state */
  83         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  84         struct nfs_write_data * commit_data;    /* special write_data for commits */
  85         int                     flags;
  86 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  87 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  88         struct nfs_writeverf    verf;           /* unstable write verifier */
  89 };
  90
  91 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  92 static const struct rpc_call_ops nfs_write_direct_ops;
  93
  94 static inline void get_dreq(struct nfs_direct_req *dreq)
  95 {
  96         atomic_inc(&dreq->io_count);
  97 }
  98
  99 static inline int put_dreq(struct nfs_direct_req *dreq)
 100 {
 101         return atomic_dec_and_test(&dreq->io_count);
 102 }
 103
 104 /*
 105  * "size" is never larger than rsize or wsize.
 106  */
 107 static inline int nfs_direct_count_pages(unsigned long user_addr, size_t size)
 108 {
 109         int page_count;
 110
 111         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 112         page_count -= user_addr >> PAGE_SHIFT;
 113         BUG_ON(page_count < 0);
 114
 115         return page_count;
 116 }
 117
 118 static inline unsigned int nfs_max_pages(unsigned int size)
 119 {
 120         return (size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 121 }
 122
 123 /**
 124  * nfs_direct_IO - NFS address space operation for direct I/O
 125  * @rw: direction (read or write)
 126  * @iocb: target I/O control block
 127  * @iov: array of vectors that define I/O buffer
 128  * @pos: offset in file to begin the operation
 129  * @nr_segs: size of iovec array
 130  *
 131  * The presence of this routine in the address space ops vector means
 132  * the NFS client supports direct I/O.  However, we shunt off direct
 133  * read and write requests before the VFS gets them, so this method
 134  * should never be called.
 135  */
 136 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 137 {
 138         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 139                         iocb->ki_filp->f_dentry->d_name.name,
 140                         (long long) pos, nr_segs);
 141
 142         return -EINVAL;
 143 }
 144
 145 static void nfs_direct_dirty_pages(struct page **pages, int npages)
 146 {
 147         int i;
 148         for (i = 0; i < npages; i++) {
 149                 struct page *page = pages[i];
 150                 if (!PageCompound(page))
 151                         set_page_dirty_lock(page);
 152         }
 153 }
 154
 155 static void nfs_direct_release_pages(struct page **pages, int npages)
 156 {
 157         int i;
 158         for (i = 0; i < npages; i++)
 159                 page_cache_release(pages[i]);
 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_alloc(nfs_direct_cachep, SLAB_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->rewrite_list);
 174         dreq->iocb = NULL;
 175         dreq->ctx = NULL;
 176         spin_lock_init(&dreq->lock);
 177         atomic_set(&dreq->io_count, 0);
 178         dreq->count = 0;
 179         dreq->error = 0;
 180         dreq->flags = 0;
 181
 182         return dreq;
 183 }
 184
 185 static void nfs_direct_req_release(struct kref *kref)
 186 {
 187         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 188
 189         if (dreq->ctx != NULL)
 190                 put_nfs_open_context(dreq->ctx);
 191         kmem_cache_free(nfs_direct_cachep, dreq);
 192 }
 193
 194 /*
 195  * Collects and returns the final error value/byte-count.
 196  */
 197 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 198 {
 199         ssize_t result = -EIOCBQUEUED;
 200
 201         /* Async requests don't wait here */
 202         if (dreq->iocb)
 203                 goto out;
 204
 205         result = wait_for_completion_interruptible(&dreq->completion);
 206
 207         if (!result)
 208                 result = dreq->error;
 209         if (!result)
 210                 result = dreq->count;
 211
 212 out:
 213         kref_put(&dreq->kref, nfs_direct_req_release);
 214         return (ssize_t) result;
 215 }
 216
 217 /*
 218  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 219  * the iocb is still valid here if this is a synchronous request.
 220  */
 221 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 222 {
 223         if (dreq->iocb) {
 224                 long res = (long) dreq->error;
 225                 if (!res)
 226                         res = (long) dreq->count;
 227                 aio_complete(dreq->iocb, res, 0);
 228         }
 229         complete_all(&dreq->completion);
 230
 231         kref_put(&dreq->kref, nfs_direct_req_release);
 232 }
 233
 234 /*
 235  * We must hold a reference to all the pages in this direct read request
 236  * until the RPCs complete.  This could be long *after* we are woken up in
 237  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 238  */
 239 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 240 {
 241         struct nfs_read_data *data = calldata;
 242         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 243
 244         if (nfs_readpage_result(task, data) != 0)
 245                 return;
 246
 247         nfs_direct_dirty_pages(data->pagevec, data->npages);
 248         nfs_direct_release_pages(data->pagevec, data->npages);
 249
 250         spin_lock(&dreq->lock);
 251
 252         if (likely(task->tk_status >= 0))
 253                 dreq->count += data->res.count;
 254         else
 255                 dreq->error = task->tk_status;
 256
 257         spin_unlock(&dreq->lock);
 258
 259         if (put_dreq(dreq))
 260                 nfs_direct_complete(dreq);
 261 }
 262
 263 static const struct rpc_call_ops nfs_read_direct_ops = {
 264         .rpc_call_done = nfs_direct_read_result,
 265         .rpc_release = nfs_readdata_release,
 266 };
 267
 268 /*
 269  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 270  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 271  * bail and stop sending more reads.  Read length accounting is
 272  * handled automatically by nfs_direct_read_result().  Otherwise, if
 273  * no requests have been sent, just return an error.
 274  */
 275 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 276 {
 277         struct nfs_open_context *ctx = dreq->ctx;
 278         struct inode *inode = ctx->dentry->d_inode;
 279         size_t rsize = NFS_SERVER(inode)->rsize;
 280         unsigned int rpages = nfs_max_pages(rsize);
 281         unsigned int pgbase;
 282         int result;
 283         ssize_t started = 0;
 284
 285         get_dreq(dreq);
 286
 287         pgbase = user_addr & ~PAGE_MASK;
 288         do {
 289                 struct nfs_read_data *data;
 290                 size_t bytes;
 291
 292                 result = -ENOMEM;
 293                 data = nfs_readdata_alloc(rpages);
 294                 if (unlikely(!data))
 295                         break;
 296
 297                 bytes = rsize;
 298                 if (count < rsize)
 299                         bytes = count;
 300
 301                 data->npages = nfs_direct_count_pages(user_addr, bytes);
 302                 down_read(&current->mm->mmap_sem);
 303                 result = get_user_pages(current, current->mm, user_addr,
 304                                         data->npages, 1, 0, data->pagevec, NULL);
 305                 up_read(&current->mm->mmap_sem);
 306                 if (unlikely(result < data->npages)) {
 307                         if (result > 0)
 308                                 nfs_direct_release_pages(data->pagevec, result);
 309                         nfs_readdata_release(data);
 310                         break;
 311                 }
 312
 313                 get_dreq(dreq);
 314
 315                 data->req = (struct nfs_page *) dreq;
 316                 data->inode = inode;
 317                 data->cred = ctx->cred;
 318                 data->args.fh = NFS_FH(inode);
 319                 data->args.context = ctx;
 320                 data->args.offset = pos;
 321                 data->args.pgbase = pgbase;
 322                 data->args.pages = data->pagevec;
 323                 data->args.count = bytes;
 324                 data->res.fattr = &data->fattr;
 325                 data->res.eof = 0;
 326                 data->res.count = bytes;
 327
 328                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 329                                 &nfs_read_direct_ops, data);
 330                 NFS_PROTO(inode)->read_setup(data);
 331
 332                 data->task.tk_cookie = (unsigned long) inode;
 333
 334                 lock_kernel();
 335                 rpc_execute(&data->task);
 336                 unlock_kernel();
 337
 338                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 339                                 data->task.tk_pid,
 340                                 inode->i_sb->s_id,
 341                                 (long long)NFS_FILEID(inode),
 342                                 bytes,
 343                                 (unsigned long long)data->args.offset);
 344
 345                 started += bytes;
 346                 user_addr += bytes;
 347                 pos += bytes;
 348                 pgbase += bytes;
 349                 pgbase &= ~PAGE_MASK;
 350
 351                 count -= bytes;
 352         } while (count != 0);
 353
 354         if (put_dreq(dreq))
 355                 nfs_direct_complete(dreq);
 356
 357         if (started)
 358                 return 0;
 359         return result < 0 ? (ssize_t) result : -EFAULT;
 360 }
 361
 362 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 363 {
 364         ssize_t result = 0;
 365         sigset_t oldset;
 366         struct inode *inode = iocb->ki_filp->f_mapping->host;
 367         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 368         struct nfs_direct_req *dreq;
 369
 370         dreq = nfs_direct_req_alloc();
 371         if (!dreq)
 372                 return -ENOMEM;
 373
 374         dreq->inode = inode;
 375         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 376         if (!is_sync_kiocb(iocb))
 377                 dreq->iocb = iocb;
 378
 379         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 380         rpc_clnt_sigmask(clnt, &oldset);
 381         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 382         if (!result)
 383                 result = nfs_direct_wait(dreq);
 384         rpc_clnt_sigunmask(clnt, &oldset);
 385
 386         return result;
 387 }
 388
 389 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 390 {
 391         while (!list_empty(&dreq->rewrite_list)) {
 392                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 393                 list_del(&data->pages);
 394                 nfs_direct_release_pages(data->pagevec, data->npages);
 395                 nfs_writedata_release(data);
 396         }
 397 }
 398
 399 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 400 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 401 {
 402         struct inode *inode = dreq->inode;
 403         struct list_head *p;
 404         struct nfs_write_data *data;
 405
 406         dreq->count = 0;
 407         get_dreq(dreq);
 408
 409         list_for_each(p, &dreq->rewrite_list) {
 410                 data = list_entry(p, struct nfs_write_data, pages);
 411
 412                 get_dreq(dreq);
 413
 414                 /*
 415                  * Reset data->res.
 416                  */
 417                 nfs_fattr_init(&data->fattr);
 418                 data->res.count = data->args.count;
 419                 memset(&data->verf, 0, sizeof(data->verf));
 420
 421                 /*
 422                  * Reuse data->task; data->args should not have changed
 423                  * since the original request was sent.
 424                  */
 425                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 426                                 &nfs_write_direct_ops, data);
 427                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 428
 429                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 430                 data->task.tk_cookie = (unsigned long) inode;
 431
 432                 /*
 433                  * We're called via an RPC callback, so BKL is already held.
 434                  */
 435                 rpc_execute(&data->task);
 436
 437                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 438                                 data->task.tk_pid,
 439                                 inode->i_sb->s_id,
 440                                 (long long)NFS_FILEID(inode),
 441                                 data->args.count,
 442                                 (unsigned long long)data->args.offset);
 443         }
 444
 445         if (put_dreq(dreq))
 446                 nfs_direct_write_complete(dreq, inode);
 447 }
 448
 449 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 450 {
 451         struct nfs_write_data *data = calldata;
 452         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 453
 454         /* Call the NFS version-specific code */
 455         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 456                 return;
 457         if (unlikely(task->tk_status < 0)) {
 458                 dreq->error = task->tk_status;
 459                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 460         }
 461         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 462                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 463                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 464         }
 465
 466         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 467         nfs_direct_write_complete(dreq, data->inode);
 468 }
 469
 470 static const struct rpc_call_ops nfs_commit_direct_ops = {
 471         .rpc_call_done = nfs_direct_commit_result,
 472         .rpc_release = nfs_commit_release,
 473 };
 474
 475 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 476 {
 477         struct nfs_write_data *data = dreq->commit_data;
 478
 479         data->inode = dreq->inode;
 480         data->cred = dreq->ctx->cred;
 481
 482         data->args.fh = NFS_FH(data->inode);
 483         data->args.offset = 0;
 484         data->args.count = 0;
 485         data->res.count = 0;
 486         data->res.fattr = &data->fattr;
 487         data->res.verf = &data->verf;
 488
 489         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 490                                 &nfs_commit_direct_ops, data);
 491         NFS_PROTO(data->inode)->commit_setup(data, 0);
 492
 493         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 494         data->task.tk_cookie = (unsigned long)data->inode;
 495         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 496         dreq->commit_data = NULL;
 497
 498         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 499
 500         lock_kernel();
 501         rpc_execute(&data->task);
 502         unlock_kernel();
 503 }
 504
 505 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 506 {
 507         int flags = dreq->flags;
 508
 509         dreq->flags = 0;
 510         switch (flags) {
 511                 case NFS_ODIRECT_DO_COMMIT:
 512                         nfs_direct_commit_schedule(dreq);
 513                         break;
 514                 case NFS_ODIRECT_RESCHED_WRITES:
 515                         nfs_direct_write_reschedule(dreq);
 516                         break;
 517                 default:
 518                         nfs_end_data_update(inode);
 519                         if (dreq->commit_data != NULL)
 520                                 nfs_commit_free(dreq->commit_data);
 521                         nfs_direct_free_writedata(dreq);
 522                         nfs_direct_complete(dreq);
 523         }
 524 }
 525
 526 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 527 {
 528         dreq->commit_data = nfs_commit_alloc(0);
 529         if (dreq->commit_data != NULL)
 530                 dreq->commit_data->req = (struct nfs_page *) dreq;
 531 }
 532 #else
 533 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 534 {
 535         dreq->commit_data = NULL;
 536 }
 537
 538 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 539 {
 540         nfs_end_data_update(inode);
 541         nfs_direct_free_writedata(dreq);
 542         nfs_direct_complete(dreq);
 543 }
 544 #endif
 545
 546 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 547 {
 548         struct nfs_write_data *data = calldata;
 549         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 550         int status = task->tk_status;
 551
 552         if (nfs_writeback_done(task, data) != 0)
 553                 return;
 554
 555         spin_lock(&dreq->lock);
 556
 557         if (likely(status >= 0))
 558                 dreq->count += data->res.count;
 559         else
 560                 dreq->error = task->tk_status;
 561
 562         if (data->res.verf->committed != NFS_FILE_SYNC) {
 563                 switch (dreq->flags) {
 564                         case 0:
 565                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 566                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 567                                 break;
 568                         case NFS_ODIRECT_DO_COMMIT:
 569                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 570                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 571                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 572                                 }
 573                 }
 574         }
 575
 576         spin_unlock(&dreq->lock);
 577 }
 578
 579 /*
 580  * NB: Return the value of the first error return code.  Subsequent
 581  *     errors after the first one are ignored.
 582  */
 583 static void nfs_direct_write_release(void *calldata)
 584 {
 585         struct nfs_write_data *data = calldata;
 586         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 587
 588         if (put_dreq(dreq))
 589                 nfs_direct_write_complete(dreq, data->inode);
 590 }
 591
 592 static const struct rpc_call_ops nfs_write_direct_ops = {
 593         .rpc_call_done = nfs_direct_write_result,
 594         .rpc_release = nfs_direct_write_release,
 595 };
 596
 597 /*
 598  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 599  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 600  * bail and stop sending more writes.  Write length accounting is
 601  * handled automatically by nfs_direct_write_result().  Otherwise, if
 602  * no requests have been sent, just return an error.
 603  */
 604 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 605 {
 606         struct nfs_open_context *ctx = dreq->ctx;
 607         struct inode *inode = ctx->dentry->d_inode;
 608         size_t wsize = NFS_SERVER(inode)->wsize;
 609         unsigned int wpages = nfs_max_pages(wsize);
 610         unsigned int pgbase;
 611         int result;
 612         ssize_t started = 0;
 613
 614         get_dreq(dreq);
 615
 616         pgbase = user_addr & ~PAGE_MASK;
 617         do {
 618                 struct nfs_write_data *data;
 619                 size_t bytes;
 620
 621                 result = -ENOMEM;
 622                 data = nfs_writedata_alloc(wpages);
 623                 if (unlikely(!data))
 624                         break;
 625
 626                 bytes = wsize;
 627                 if (count < wsize)
 628                         bytes = count;
 629
 630                 data->npages = nfs_direct_count_pages(user_addr, bytes);
 631                 down_read(&current->mm->mmap_sem);
 632                 result = get_user_pages(current, current->mm, user_addr,
 633                                         data->npages, 0, 0, data->pagevec, NULL);
 634                 up_read(&current->mm->mmap_sem);
 635                 if (unlikely(result < data->npages)) {
 636                         if (result > 0)
 637                                 nfs_direct_release_pages(data->pagevec, result);
 638                         nfs_writedata_release(data);
 639                         break;
 640                 }
 641
 642                 get_dreq(dreq);
 643
 644                 list_move_tail(&data->pages, &dreq->rewrite_list);
 645
 646                 data->req = (struct nfs_page *) dreq;
 647                 data->inode = inode;
 648                 data->cred = ctx->cred;
 649                 data->args.fh = NFS_FH(inode);
 650                 data->args.context = ctx;
 651                 data->args.offset = pos;
 652                 data->args.pgbase = pgbase;
 653                 data->args.pages = data->pagevec;
 654                 data->args.count = bytes;
 655                 data->res.fattr = &data->fattr;
 656                 data->res.count = bytes;
 657                 data->res.verf = &data->verf;
 658
 659                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 660                                 &nfs_write_direct_ops, data);
 661                 NFS_PROTO(inode)->write_setup(data, sync);
 662
 663                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 664                 data->task.tk_cookie = (unsigned long) inode;
 665
 666                 lock_kernel();
 667                 rpc_execute(&data->task);
 668                 unlock_kernel();
 669
 670                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 671                                 data->task.tk_pid,
 672                                 inode->i_sb->s_id,
 673                                 (long long)NFS_FILEID(inode),
 674                                 bytes,
 675                                 (unsigned long long)data->args.offset);
 676
 677                 started += bytes;
 678                 user_addr += bytes;
 679                 pos += bytes;
 680                 pgbase += bytes;
 681                 pgbase &= ~PAGE_MASK;
 682
 683                 count -= bytes;
 684         } while (count != 0);
 685
 686         if (put_dreq(dreq))
 687                 nfs_direct_write_complete(dreq, inode);
 688
 689         if (started)
 690                 return 0;
 691         return result < 0 ? (ssize_t) result : -EFAULT;
 692 }
 693
 694 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 695 {
 696         ssize_t result = 0;
 697         sigset_t oldset;
 698         struct inode *inode = iocb->ki_filp->f_mapping->host;
 699         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 700         struct nfs_direct_req *dreq;
 701         size_t wsize = NFS_SERVER(inode)->wsize;
 702         int sync = 0;
 703
 704         dreq = nfs_direct_req_alloc();
 705         if (!dreq)
 706                 return -ENOMEM;
 707         nfs_alloc_commit_data(dreq);
 708
 709         if (dreq->commit_data == NULL || count < wsize)
 710                 sync = FLUSH_STABLE;
 711
 712         dreq->inode = inode;
 713         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 714         if (!is_sync_kiocb(iocb))
 715                 dreq->iocb = iocb;
 716
 717         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 718
 719         nfs_begin_data_update(inode);
 720
 721         rpc_clnt_sigmask(clnt, &oldset);
 722         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 723         if (!result)
 724                 result = nfs_direct_wait(dreq);
 725         rpc_clnt_sigunmask(clnt, &oldset);
 726
 727         return result;
 728 }
 729
 730 /**
 731  * nfs_file_direct_read - file direct read operation for NFS files
 732  * @iocb: target I/O control block
 733  * @buf: user's buffer into which to read data
 734  * @count: number of bytes to read
 735  * @pos: byte offset in file where reading starts
 736  *
 737  * We use this function for direct reads instead of calling
 738  * generic_file_aio_read() in order to avoid gfar's check to see if
 739  * the request starts before the end of the file.  For that check
 740  * to work, we must generate a GETATTR before each direct read, and
 741  * even then there is a window between the GETATTR and the subsequent
 742  * READ where the file size could change.  Our preference is simply
 743  * to do all reads the application wants, and the server will take
 744  * care of managing the end of file boundary.
 745  *
 746  * This function also eliminates unnecessarily updating the file's
 747  * atime locally, as the NFS server sets the file's atime, and this
 748  * client must read the updated atime from the server back into its
 749  * cache.
 750  */
 751 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 752 {
 753         ssize_t retval = -EINVAL;
 754         struct file *file = iocb->ki_filp;
 755         struct address_space *mapping = file->f_mapping;
 756
 757         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 758                 file->f_dentry->d_parent->d_name.name,
 759                 file->f_dentry->d_name.name,
 760                 (unsigned long) count, (long long) pos);
 761
 762         if (count < 0)
 763                 goto out;
 764         retval = -EFAULT;
 765         if (!access_ok(VERIFY_WRITE, buf, count))
 766                 goto out;
 767         retval = 0;
 768         if (!count)
 769                 goto out;
 770
 771         retval = nfs_sync_mapping(mapping);
 772         if (retval)
 773                 goto out;
 774
 775         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 776         if (retval > 0)
 777                 iocb->ki_pos = pos + retval;
 778
 779 out:
 780         return retval;
 781 }
 782
 783 /**
 784  * nfs_file_direct_write - file direct write operation for NFS files
 785  * @iocb: target I/O control block
 786  * @buf: user's buffer from which to write data
 787  * @count: number of bytes to write
 788  * @pos: byte offset in file where writing starts
 789  *
 790  * We use this function for direct writes instead of calling
 791  * generic_file_aio_write() in order to avoid taking the inode
 792  * semaphore and updating the i_size.  The NFS server will set
 793  * the new i_size and this client must read the updated size
 794  * back into its cache.  We let the server do generic write
 795  * parameter checking and report problems.
 796  *
 797  * We also avoid an unnecessary invocation of generic_osync_inode(),
 798  * as it is fairly meaningless to sync the metadata of an NFS file.
 799  *
 800  * We eliminate local atime updates, see direct read above.
 801  *
 802  * We avoid unnecessary page cache invalidations for normal cached
 803  * readers of this file.
 804  *
 805  * Note that O_APPEND is not supported for NFS direct writes, as there
 806  * is no atomic O_APPEND write facility in the NFS protocol.
 807  */
 808 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 809 {
 810         ssize_t retval;
 811         struct file *file = iocb->ki_filp;
 812         struct address_space *mapping = file->f_mapping;
 813
 814         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 815                 file->f_dentry->d_parent->d_name.name,
 816                 file->f_dentry->d_name.name,
 817                 (unsigned long) count, (long long) pos);
 818
 819         retval = generic_write_checks(file, &pos, &count, 0);
 820         if (retval)
 821                 goto out;
 822
 823         retval = -EINVAL;
 824         if ((ssize_t) count < 0)
 825                 goto out;
 826         retval = 0;
 827         if (!count)
 828                 goto out;
 829
 830         retval = -EFAULT;
 831         if (!access_ok(VERIFY_READ, buf, count))
 832                 goto out;
 833
 834         retval = nfs_sync_mapping(mapping);
 835         if (retval)
 836                 goto out;
 837
 838         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 839
 840         /*
 841          * XXX: nfs_end_data_update() already ensures this file's
 842          *      cached data is subsequently invalidated.  Do we really
 843          *      need to call invalidate_inode_pages2() again here?
 844          *
 845          *      For aio writes, this invalidation will almost certainly
 846          *      occur before the writes complete.  Kind of racey.
 847          */
 848         if (mapping->nrpages)
 849                 invalidate_inode_pages2(mapping);
 850
 851         if (retval > 0)
 852                 iocb->ki_pos = pos + retval;
 853
 854 out:
 855         return retval;
 856 }
 857
 858 /**
 859  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 860  *
 861  */
 862 int __init nfs_init_directcache(void)
 863 {
 864         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 865                                                 sizeof(struct nfs_direct_req),
 866                                                 0, (SLAB_RECLAIM_ACCOUNT|
 867                                                         SLAB_MEM_SPREAD),
 868                                                 NULL, NULL);
 869         if (nfs_direct_cachep == NULL)
 870                 return -ENOMEM;
 871
 872         return 0;
 873 }
 874
 875 /**
 876  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 877  *
 878  */
 879 void __exit nfs_destroy_directcache(void)
 880 {
 881         if (kmem_cache_destroy(nfs_direct_cachep))
 882                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 883 }