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/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
  61 static kmem_cache_t *nfs_direct_cachep;
  62
  63 /*
  64  * This represents a set of asynchronous requests that we're waiting on
  65  */
  66 struct nfs_direct_req {
  67         struct kref             kref;           /* release manager */
  68
  69         /* I/O parameters */
  70         struct nfs_open_context *ctx;           /* file open context info */
  71         struct kiocb *          iocb;           /* controlling i/o request */
  72         struct inode *          inode;          /* target file of i/o */
  73
  74         /* completion state */
  75         atomic_t                io_count;       /* i/os we're waiting for */
  76         spinlock_t              lock;           /* protect completion state */
  77         ssize_t                 count,          /* bytes actually processed */
  78                                 error;          /* any reported error */
  79         struct completion       completion;     /* wait for i/o completion */
  80
  81         /* commit state */
  82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  83         struct nfs_write_data * commit_data;    /* special write_data for commits */
  84         int                     flags;
  85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  87         struct nfs_writeverf    verf;           /* unstable write verifier */
  88 };
  89
  90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  91 static const struct rpc_call_ops nfs_write_direct_ops;
  92
  93 static inline void get_dreq(struct nfs_direct_req *dreq)
  94 {
  95         atomic_inc(&dreq->io_count);
  96 }
  97
  98 static inline int put_dreq(struct nfs_direct_req *dreq)
  99 {
 100         return atomic_dec_and_test(&dreq->io_count);
 101 }
 102
 103 /**
 104  * nfs_direct_IO - NFS address space operation for direct I/O
 105  * @rw: direction (read or write)
 106  * @iocb: target I/O control block
 107  * @iov: array of vectors that define I/O buffer
 108  * @pos: offset in file to begin the operation
 109  * @nr_segs: size of iovec array
 110  *
 111  * The presence of this routine in the address space ops vector means
 112  * the NFS client supports direct I/O.  However, we shunt off direct
 113  * read and write requests before the VFS gets them, so this method
 114  * should never be called.
 115  */
 116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 117 {
 118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 119                         iocb->ki_filp->f_dentry->d_name.name,
 120                         (long long) pos, nr_segs);
 121
 122         return -EINVAL;
 123 }
 124
 125 static void nfs_direct_dirty_pages(struct page **pages, int npages)
 126 {
 127         int i;
 128         for (i = 0; i < npages; i++) {
 129                 struct page *page = pages[i];
 130                 if (!PageCompound(page))
 131                         set_page_dirty_lock(page);
 132         }
 133 }
 134
 135 static void nfs_direct_release_pages(struct page **pages, int npages)
 136 {
 137         int i;
 138         for (i = 0; i < npages; i++)
 139                 page_cache_release(pages[i]);
 140 }
 141
 142 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 143 {
 144         struct nfs_direct_req *dreq;
 145
 146         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 147         if (!dreq)
 148                 return NULL;
 149
 150         kref_init(&dreq->kref);
 151         kref_get(&dreq->kref);
 152         init_completion(&dreq->completion);
 153         INIT_LIST_HEAD(&dreq->rewrite_list);
 154         dreq->iocb = NULL;
 155         dreq->ctx = NULL;
 156         spin_lock_init(&dreq->lock);
 157         atomic_set(&dreq->io_count, 0);
 158         dreq->count = 0;
 159         dreq->error = 0;
 160         dreq->flags = 0;
 161
 162         return dreq;
 163 }
 164
 165 static void nfs_direct_req_release(struct kref *kref)
 166 {
 167         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 168
 169         if (dreq->ctx != NULL)
 170                 put_nfs_open_context(dreq->ctx);
 171         kmem_cache_free(nfs_direct_cachep, dreq);
 172 }
 173
 174 /*
 175  * Collects and returns the final error value/byte-count.
 176  */
 177 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 178 {
 179         ssize_t result = -EIOCBQUEUED;
 180
 181         /* Async requests don't wait here */
 182         if (dreq->iocb)
 183                 goto out;
 184
 185         result = wait_for_completion_interruptible(&dreq->completion);
 186
 187         if (!result)
 188                 result = dreq->error;
 189         if (!result)
 190                 result = dreq->count;
 191
 192 out:
 193         kref_put(&dreq->kref, nfs_direct_req_release);
 194         return (ssize_t) result;
 195 }
 196
 197 /*
 198  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 199  * the iocb is still valid here if this is a synchronous request.
 200  */
 201 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 202 {
 203         if (dreq->iocb) {
 204                 long res = (long) dreq->error;
 205                 if (!res)
 206                         res = (long) dreq->count;
 207                 aio_complete(dreq->iocb, res, 0);
 208         }
 209         complete_all(&dreq->completion);
 210
 211         kref_put(&dreq->kref, nfs_direct_req_release);
 212 }
 213
 214 /*
 215  * We must hold a reference to all the pages in this direct read request
 216  * until the RPCs complete.  This could be long *after* we are woken up in
 217  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 218  */
 219 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 220 {
 221         struct nfs_read_data *data = calldata;
 222         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 223
 224         if (nfs_readpage_result(task, data) != 0)
 225                 return;
 226
 227         nfs_direct_dirty_pages(data->pagevec, data->npages);
 228         nfs_direct_release_pages(data->pagevec, data->npages);
 229
 230         spin_lock(&dreq->lock);
 231
 232         if (likely(task->tk_status >= 0))
 233                 dreq->count += data->res.count;
 234         else
 235                 dreq->error = task->tk_status;
 236
 237         spin_unlock(&dreq->lock);
 238
 239         if (put_dreq(dreq))
 240                 nfs_direct_complete(dreq);
 241 }
 242
 243 static const struct rpc_call_ops nfs_read_direct_ops = {
 244         .rpc_call_done = nfs_direct_read_result,
 245         .rpc_release = nfs_readdata_release,
 246 };
 247
 248 /*
 249  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 250  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 251  * bail and stop sending more reads.  Read length accounting is
 252  * handled automatically by nfs_direct_read_result().  Otherwise, if
 253  * no requests have been sent, just return an error.
 254  */
 255 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 256 {
 257         struct nfs_open_context *ctx = dreq->ctx;
 258         struct inode *inode = ctx->dentry->d_inode;
 259         size_t rsize = NFS_SERVER(inode)->rsize;
 260         unsigned int pgbase;
 261         int result;
 262         ssize_t started = 0;
 263
 264         get_dreq(dreq);
 265
 266         do {
 267                 struct nfs_read_data *data;
 268                 size_t bytes;
 269
 270                 pgbase = user_addr & ~PAGE_MASK;
 271                 bytes = min(rsize,count);
 272
 273                 result = -ENOMEM;
 274                 data = nfs_readdata_alloc(pgbase + bytes);
 275                 if (unlikely(!data))
 276                         break;
 277
 278                 down_read(&current->mm->mmap_sem);
 279                 result = get_user_pages(current, current->mm, user_addr,
 280                                         data->npages, 1, 0, data->pagevec, NULL);
 281                 up_read(&current->mm->mmap_sem);
 282                 if (unlikely(result < data->npages)) {
 283                         if (result > 0)
 284                                 nfs_direct_release_pages(data->pagevec, result);
 285                         nfs_readdata_release(data);
 286                         break;
 287                 }
 288
 289                 get_dreq(dreq);
 290
 291                 data->req = (struct nfs_page *) dreq;
 292                 data->inode = inode;
 293                 data->cred = ctx->cred;
 294                 data->args.fh = NFS_FH(inode);
 295                 data->args.context = ctx;
 296                 data->args.offset = pos;
 297                 data->args.pgbase = pgbase;
 298                 data->args.pages = data->pagevec;
 299                 data->args.count = bytes;
 300                 data->res.fattr = &data->fattr;
 301                 data->res.eof = 0;
 302                 data->res.count = bytes;
 303
 304                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 305                                 &nfs_read_direct_ops, data);
 306                 NFS_PROTO(inode)->read_setup(data);
 307
 308                 data->task.tk_cookie = (unsigned long) inode;
 309
 310                 lock_kernel();
 311                 rpc_execute(&data->task);
 312                 unlock_kernel();
 313
 314                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 315                                 data->task.tk_pid,
 316                                 inode->i_sb->s_id,
 317                                 (long long)NFS_FILEID(inode),
 318                                 bytes,
 319                                 (unsigned long long)data->args.offset);
 320
 321                 started += bytes;
 322                 user_addr += bytes;
 323                 pos += bytes;
 324                 /* FIXME: Remove this unnecessary math from final patch */
 325                 pgbase += bytes;
 326                 pgbase &= ~PAGE_MASK;
 327                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 328
 329                 count -= bytes;
 330         } while (count != 0);
 331
 332         if (put_dreq(dreq))
 333                 nfs_direct_complete(dreq);
 334
 335         if (started)
 336                 return 0;
 337         return result < 0 ? (ssize_t) result : -EFAULT;
 338 }
 339
 340 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 341 {
 342         ssize_t result = 0;
 343         sigset_t oldset;
 344         struct inode *inode = iocb->ki_filp->f_mapping->host;
 345         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 346         struct nfs_direct_req *dreq;
 347
 348         dreq = nfs_direct_req_alloc();
 349         if (!dreq)
 350                 return -ENOMEM;
 351
 352         dreq->inode = inode;
 353         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 354         if (!is_sync_kiocb(iocb))
 355                 dreq->iocb = iocb;
 356
 357         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 358         rpc_clnt_sigmask(clnt, &oldset);
 359         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 360         if (!result)
 361                 result = nfs_direct_wait(dreq);
 362         rpc_clnt_sigunmask(clnt, &oldset);
 363
 364         return result;
 365 }
 366
 367 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 368 {
 369         while (!list_empty(&dreq->rewrite_list)) {
 370                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 371                 list_del(&data->pages);
 372                 nfs_direct_release_pages(data->pagevec, data->npages);
 373                 nfs_writedata_release(data);
 374         }
 375 }
 376
 377 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 378 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 379 {
 380         struct inode *inode = dreq->inode;
 381         struct list_head *p;
 382         struct nfs_write_data *data;
 383
 384         dreq->count = 0;
 385         get_dreq(dreq);
 386
 387         list_for_each(p, &dreq->rewrite_list) {
 388                 data = list_entry(p, struct nfs_write_data, pages);
 389
 390                 get_dreq(dreq);
 391
 392                 /*
 393                  * Reset data->res.
 394                  */
 395                 nfs_fattr_init(&data->fattr);
 396                 data->res.count = data->args.count;
 397                 memset(&data->verf, 0, sizeof(data->verf));
 398
 399                 /*
 400                  * Reuse data->task; data->args should not have changed
 401                  * since the original request was sent.
 402                  */
 403                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 404                                 &nfs_write_direct_ops, data);
 405                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 406
 407                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 408                 data->task.tk_cookie = (unsigned long) inode;
 409
 410                 /*
 411                  * We're called via an RPC callback, so BKL is already held.
 412                  */
 413                 rpc_execute(&data->task);
 414
 415                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 416                                 data->task.tk_pid,
 417                                 inode->i_sb->s_id,
 418                                 (long long)NFS_FILEID(inode),
 419                                 data->args.count,
 420                                 (unsigned long long)data->args.offset);
 421         }
 422
 423         if (put_dreq(dreq))
 424                 nfs_direct_write_complete(dreq, inode);
 425 }
 426
 427 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 428 {
 429         struct nfs_write_data *data = calldata;
 430         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 431
 432         /* Call the NFS version-specific code */
 433         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 434                 return;
 435         if (unlikely(task->tk_status < 0)) {
 436                 dreq->error = task->tk_status;
 437                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 438         }
 439         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 440                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 441                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 442         }
 443
 444         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 445         nfs_direct_write_complete(dreq, data->inode);
 446 }
 447
 448 static const struct rpc_call_ops nfs_commit_direct_ops = {
 449         .rpc_call_done = nfs_direct_commit_result,
 450         .rpc_release = nfs_commit_release,
 451 };
 452
 453 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 454 {
 455         struct nfs_write_data *data = dreq->commit_data;
 456
 457         data->inode = dreq->inode;
 458         data->cred = dreq->ctx->cred;
 459
 460         data->args.fh = NFS_FH(data->inode);
 461         data->args.offset = 0;
 462         data->args.count = 0;
 463         data->res.count = 0;
 464         data->res.fattr = &data->fattr;
 465         data->res.verf = &data->verf;
 466
 467         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 468                                 &nfs_commit_direct_ops, data);
 469         NFS_PROTO(data->inode)->commit_setup(data, 0);
 470
 471         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 472         data->task.tk_cookie = (unsigned long)data->inode;
 473         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 474         dreq->commit_data = NULL;
 475
 476         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 477
 478         lock_kernel();
 479         rpc_execute(&data->task);
 480         unlock_kernel();
 481 }
 482
 483 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 484 {
 485         int flags = dreq->flags;
 486
 487         dreq->flags = 0;
 488         switch (flags) {
 489                 case NFS_ODIRECT_DO_COMMIT:
 490                         nfs_direct_commit_schedule(dreq);
 491                         break;
 492                 case NFS_ODIRECT_RESCHED_WRITES:
 493                         nfs_direct_write_reschedule(dreq);
 494                         break;
 495                 default:
 496                         nfs_end_data_update(inode);
 497                         if (dreq->commit_data != NULL)
 498                                 nfs_commit_free(dreq->commit_data);
 499                         nfs_direct_free_writedata(dreq);
 500                         nfs_direct_complete(dreq);
 501         }
 502 }
 503
 504 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 505 {
 506         dreq->commit_data = nfs_commit_alloc();
 507         if (dreq->commit_data != NULL)
 508                 dreq->commit_data->req = (struct nfs_page *) dreq;
 509 }
 510 #else
 511 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 512 {
 513         dreq->commit_data = NULL;
 514 }
 515
 516 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 517 {
 518         nfs_end_data_update(inode);
 519         nfs_direct_free_writedata(dreq);
 520         nfs_direct_complete(dreq);
 521 }
 522 #endif
 523
 524 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 525 {
 526         struct nfs_write_data *data = calldata;
 527         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 528         int status = task->tk_status;
 529
 530         if (nfs_writeback_done(task, data) != 0)
 531                 return;
 532
 533         spin_lock(&dreq->lock);
 534
 535         if (likely(status >= 0))
 536                 dreq->count += data->res.count;
 537         else
 538                 dreq->error = task->tk_status;
 539
 540         if (data->res.verf->committed != NFS_FILE_SYNC) {
 541                 switch (dreq->flags) {
 542                         case 0:
 543                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 544                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 545                                 break;
 546                         case NFS_ODIRECT_DO_COMMIT:
 547                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 548                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 549                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 550                                 }
 551                 }
 552         }
 553
 554         spin_unlock(&dreq->lock);
 555 }
 556
 557 /*
 558  * NB: Return the value of the first error return code.  Subsequent
 559  *     errors after the first one are ignored.
 560  */
 561 static void nfs_direct_write_release(void *calldata)
 562 {
 563         struct nfs_write_data *data = calldata;
 564         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 565
 566         if (put_dreq(dreq))
 567                 nfs_direct_write_complete(dreq, data->inode);
 568 }
 569
 570 static const struct rpc_call_ops nfs_write_direct_ops = {
 571         .rpc_call_done = nfs_direct_write_result,
 572         .rpc_release = nfs_direct_write_release,
 573 };
 574
 575 /*
 576  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 577  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 578  * bail and stop sending more writes.  Write length accounting is
 579  * handled automatically by nfs_direct_write_result().  Otherwise, if
 580  * no requests have been sent, just return an error.
 581  */
 582 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 583 {
 584         struct nfs_open_context *ctx = dreq->ctx;
 585         struct inode *inode = ctx->dentry->d_inode;
 586         size_t wsize = NFS_SERVER(inode)->wsize;
 587         unsigned int pgbase;
 588         int result;
 589         ssize_t started = 0;
 590
 591         get_dreq(dreq);
 592
 593         do {
 594                 struct nfs_write_data *data;
 595                 size_t bytes;
 596
 597                 pgbase = user_addr & ~PAGE_MASK;
 598                 bytes = min(wsize,count);
 599
 600                 result = -ENOMEM;
 601                 data = nfs_writedata_alloc(pgbase + bytes);
 602                 if (unlikely(!data))
 603                         break;
 604
 605                 down_read(&current->mm->mmap_sem);
 606                 result = get_user_pages(current, current->mm, user_addr,
 607                                         data->npages, 0, 0, data->pagevec, NULL);
 608                 up_read(&current->mm->mmap_sem);
 609                 if (unlikely(result < data->npages)) {
 610                         if (result > 0)
 611                                 nfs_direct_release_pages(data->pagevec, result);
 612                         nfs_writedata_release(data);
 613                         break;
 614                 }
 615
 616                 get_dreq(dreq);
 617
 618                 list_move_tail(&data->pages, &dreq->rewrite_list);
 619
 620                 data->req = (struct nfs_page *) dreq;
 621                 data->inode = inode;
 622                 data->cred = ctx->cred;
 623                 data->args.fh = NFS_FH(inode);
 624                 data->args.context = ctx;
 625                 data->args.offset = pos;
 626                 data->args.pgbase = pgbase;
 627                 data->args.pages = data->pagevec;
 628                 data->args.count = bytes;
 629                 data->res.fattr = &data->fattr;
 630                 data->res.count = bytes;
 631                 data->res.verf = &data->verf;
 632
 633                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 634                                 &nfs_write_direct_ops, data);
 635                 NFS_PROTO(inode)->write_setup(data, sync);
 636
 637                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 638                 data->task.tk_cookie = (unsigned long) inode;
 639
 640                 lock_kernel();
 641                 rpc_execute(&data->task);
 642                 unlock_kernel();
 643
 644                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 645                                 data->task.tk_pid,
 646                                 inode->i_sb->s_id,
 647                                 (long long)NFS_FILEID(inode),
 648                                 bytes,
 649                                 (unsigned long long)data->args.offset);
 650
 651                 started += bytes;
 652                 user_addr += bytes;
 653                 pos += bytes;
 654
 655                 /* FIXME: Remove this useless math from the final patch */
 656                 pgbase += bytes;
 657                 pgbase &= ~PAGE_MASK;
 658                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 659
 660                 count -= bytes;
 661         } while (count != 0);
 662
 663         if (put_dreq(dreq))
 664                 nfs_direct_write_complete(dreq, inode);
 665
 666         if (started)
 667                 return 0;
 668         return result < 0 ? (ssize_t) result : -EFAULT;
 669 }
 670
 671 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 672 {
 673         ssize_t result = 0;
 674         sigset_t oldset;
 675         struct inode *inode = iocb->ki_filp->f_mapping->host;
 676         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 677         struct nfs_direct_req *dreq;
 678         size_t wsize = NFS_SERVER(inode)->wsize;
 679         int sync = 0;
 680
 681         dreq = nfs_direct_req_alloc();
 682         if (!dreq)
 683                 return -ENOMEM;
 684         nfs_alloc_commit_data(dreq);
 685
 686         if (dreq->commit_data == NULL || count < wsize)
 687                 sync = FLUSH_STABLE;
 688
 689         dreq->inode = inode;
 690         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 691         if (!is_sync_kiocb(iocb))
 692                 dreq->iocb = iocb;
 693
 694         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 695
 696         nfs_begin_data_update(inode);
 697
 698         rpc_clnt_sigmask(clnt, &oldset);
 699         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 700         if (!result)
 701                 result = nfs_direct_wait(dreq);
 702         rpc_clnt_sigunmask(clnt, &oldset);
 703
 704         return result;
 705 }
 706
 707 /**
 708  * nfs_file_direct_read - file direct read operation for NFS files
 709  * @iocb: target I/O control block
 710  * @buf: user's buffer into which to read data
 711  * @count: number of bytes to read
 712  * @pos: byte offset in file where reading starts
 713  *
 714  * We use this function for direct reads instead of calling
 715  * generic_file_aio_read() in order to avoid gfar's check to see if
 716  * the request starts before the end of the file.  For that check
 717  * to work, we must generate a GETATTR before each direct read, and
 718  * even then there is a window between the GETATTR and the subsequent
 719  * READ where the file size could change.  Our preference is simply
 720  * to do all reads the application wants, and the server will take
 721  * care of managing the end of file boundary.
 722  *
 723  * This function also eliminates unnecessarily updating the file's
 724  * atime locally, as the NFS server sets the file's atime, and this
 725  * client must read the updated atime from the server back into its
 726  * cache.
 727  */
 728 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 729 {
 730         ssize_t retval = -EINVAL;
 731         struct file *file = iocb->ki_filp;
 732         struct address_space *mapping = file->f_mapping;
 733
 734         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 735                 file->f_dentry->d_parent->d_name.name,
 736                 file->f_dentry->d_name.name,
 737                 (unsigned long) count, (long long) pos);
 738
 739         if (count < 0)
 740                 goto out;
 741         retval = -EFAULT;
 742         if (!access_ok(VERIFY_WRITE, buf, count))
 743                 goto out;
 744         retval = 0;
 745         if (!count)
 746                 goto out;
 747
 748         retval = nfs_sync_mapping(mapping);
 749         if (retval)
 750                 goto out;
 751
 752         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 753         if (retval > 0)
 754                 iocb->ki_pos = pos + retval;
 755
 756 out:
 757         return retval;
 758 }
 759
 760 /**
 761  * nfs_file_direct_write - file direct write operation for NFS files
 762  * @iocb: target I/O control block
 763  * @buf: user's buffer from which to write data
 764  * @count: number of bytes to write
 765  * @pos: byte offset in file where writing starts
 766  *
 767  * We use this function for direct writes instead of calling
 768  * generic_file_aio_write() in order to avoid taking the inode
 769  * semaphore and updating the i_size.  The NFS server will set
 770  * the new i_size and this client must read the updated size
 771  * back into its cache.  We let the server do generic write
 772  * parameter checking and report problems.
 773  *
 774  * We also avoid an unnecessary invocation of generic_osync_inode(),
 775  * as it is fairly meaningless to sync the metadata of an NFS file.
 776  *
 777  * We eliminate local atime updates, see direct read above.
 778  *
 779  * We avoid unnecessary page cache invalidations for normal cached
 780  * readers of this file.
 781  *
 782  * Note that O_APPEND is not supported for NFS direct writes, as there
 783  * is no atomic O_APPEND write facility in the NFS protocol.
 784  */
 785 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 786 {
 787         ssize_t retval;
 788         struct file *file = iocb->ki_filp;
 789         struct address_space *mapping = file->f_mapping;
 790
 791         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 792                 file->f_dentry->d_parent->d_name.name,
 793                 file->f_dentry->d_name.name,
 794                 (unsigned long) count, (long long) pos);
 795
 796         retval = generic_write_checks(file, &pos, &count, 0);
 797         if (retval)
 798                 goto out;
 799
 800         retval = -EINVAL;
 801         if ((ssize_t) count < 0)
 802                 goto out;
 803         retval = 0;
 804         if (!count)
 805                 goto out;
 806
 807         retval = -EFAULT;
 808         if (!access_ok(VERIFY_READ, buf, count))
 809                 goto out;
 810
 811         retval = nfs_sync_mapping(mapping);
 812         if (retval)
 813                 goto out;
 814
 815         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 816
 817         /*
 818          * XXX: nfs_end_data_update() already ensures this file's
 819          *      cached data is subsequently invalidated.  Do we really
 820          *      need to call invalidate_inode_pages2() again here?
 821          *
 822          *      For aio writes, this invalidation will almost certainly
 823          *      occur before the writes complete.  Kind of racey.
 824          */
 825         if (mapping->nrpages)
 826                 invalidate_inode_pages2(mapping);
 827
 828         if (retval > 0)
 829                 iocb->ki_pos = pos + retval;
 830
 831 out:
 832         return retval;
 833 }
 834
 835 /**
 836  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 837  *
 838  */
 839 int __init nfs_init_directcache(void)
 840 {
 841         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 842                                                 sizeof(struct nfs_direct_req),
 843                                                 0, (SLAB_RECLAIM_ACCOUNT|
 844                                                         SLAB_MEM_SPREAD),
 845                                                 NULL, NULL);
 846         if (nfs_direct_cachep == NULL)
 847                 return -ENOMEM;
 848
 849         return 0;
 850 }
 851
 852 /**
 853  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 854  *
 855  */
 856 void nfs_destroy_directcache(void)
 857 {
 858         if (kmem_cache_destroy(nfs_direct_cachep))
 859                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 860 }