SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 "internal.h"
  58 #include "iostat.h"
  59
  60 #define NFSDBG_FACILITY         NFSDBG_VFS
  61
  62 static struct kmem_cache *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  * nfs_direct_IO - NFS address space operation for direct I/O
 106  * @rw: direction (read or write)
 107  * @iocb: target I/O control block
 108  * @iov: array of vectors that define I/O buffer
 109  * @pos: offset in file to begin the operation
 110  * @nr_segs: size of iovec array
 111  *
 112  * The presence of this routine in the address space ops vector means
 113  * the NFS client supports direct I/O.  However, we shunt off direct
 114  * read and write requests before the VFS gets them, so this method
 115  * should never be called.
 116  */
 117 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 118 {
 119         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 120                         iocb->ki_filp->f_path.dentry->d_name.name,
 121                         (long long) pos, nr_segs);
 122
 123         return -EINVAL;
 124 }
 125
 126 static void nfs_direct_dirty_pages(struct page **pages, int npages)
 127 {
 128         int i;
 129         for (i = 0; i < npages; i++) {
 130                 struct page *page = pages[i];
 131                 if (!PageCompound(page))
 132                         set_page_dirty_lock(page);
 133         }
 134 }
 135
 136 static void nfs_direct_release_pages(struct page **pages, int npages)
 137 {
 138         int i;
 139         for (i = 0; i < npages; i++)
 140                 page_cache_release(pages[i]);
 141 }
 142
 143 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 144 {
 145         struct nfs_direct_req *dreq;
 146
 147         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 148         if (!dreq)
 149                 return NULL;
 150
 151         kref_init(&dreq->kref);
 152         kref_get(&dreq->kref);
 153         init_completion(&dreq->completion);
 154         INIT_LIST_HEAD(&dreq->rewrite_list);
 155         dreq->iocb = NULL;
 156         dreq->ctx = NULL;
 157         spin_lock_init(&dreq->lock);
 158         atomic_set(&dreq->io_count, 0);
 159         dreq->count = 0;
 160         dreq->error = 0;
 161         dreq->flags = 0;
 162
 163         return dreq;
 164 }
 165
 166 static void nfs_direct_req_release(struct kref *kref)
 167 {
 168         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 169
 170         if (dreq->ctx != NULL)
 171                 put_nfs_open_context(dreq->ctx);
 172         kmem_cache_free(nfs_direct_cachep, dreq);
 173 }
 174
 175 /*
 176  * Collects and returns the final error value/byte-count.
 177  */
 178 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 179 {
 180         ssize_t result = -EIOCBQUEUED;
 181
 182         /* Async requests don't wait here */
 183         if (dreq->iocb)
 184                 goto out;
 185
 186         result = wait_for_completion_interruptible(&dreq->completion);
 187
 188         if (!result)
 189                 result = dreq->error;
 190         if (!result)
 191                 result = dreq->count;
 192
 193 out:
 194         kref_put(&dreq->kref, nfs_direct_req_release);
 195         return (ssize_t) result;
 196 }
 197
 198 /*
 199  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 200  * the iocb is still valid here if this is a synchronous request.
 201  */
 202 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 203 {
 204         if (dreq->iocb) {
 205                 long res = (long) dreq->error;
 206                 if (!res)
 207                         res = (long) dreq->count;
 208                 aio_complete(dreq->iocb, res, 0);
 209         }
 210         complete_all(&dreq->completion);
 211
 212         kref_put(&dreq->kref, nfs_direct_req_release);
 213 }
 214
 215 /*
 216  * We must hold a reference to all the pages in this direct read request
 217  * until the RPCs complete.  This could be long *after* we are woken up in
 218  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 219  */
 220 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 221 {
 222         struct nfs_read_data *data = calldata;
 223         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 224
 225         if (nfs_readpage_result(task, data) != 0)
 226                 return;
 227
 228         nfs_direct_dirty_pages(data->pagevec, data->npages);
 229         nfs_direct_release_pages(data->pagevec, data->npages);
 230
 231         spin_lock(&dreq->lock);
 232
 233         if (likely(task->tk_status >= 0))
 234                 dreq->count += data->res.count;
 235         else
 236                 dreq->error = task->tk_status;
 237
 238         spin_unlock(&dreq->lock);
 239
 240         if (put_dreq(dreq))
 241                 nfs_direct_complete(dreq);
 242 }
 243
 244 static const struct rpc_call_ops nfs_read_direct_ops = {
 245         .rpc_call_done = nfs_direct_read_result,
 246         .rpc_release = nfs_readdata_release,
 247 };
 248
 249 /*
 250  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 251  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 252  * bail and stop sending more reads.  Read length accounting is
 253  * handled automatically by nfs_direct_read_result().  Otherwise, if
 254  * no requests have been sent, just return an error.
 255  */
 256 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 257 {
 258         struct nfs_open_context *ctx = dreq->ctx;
 259         struct inode *inode = ctx->dentry->d_inode;
 260         size_t rsize = NFS_SERVER(inode)->rsize;
 261         unsigned int pgbase;
 262         int result;
 263         ssize_t started = 0;
 264
 265         get_dreq(dreq);
 266
 267         do {
 268                 struct nfs_read_data *data;
 269                 size_t bytes;
 270
 271                 pgbase = user_addr & ~PAGE_MASK;
 272                 bytes = min(rsize,count);
 273
 274                 result = -ENOMEM;
 275                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 276                 if (unlikely(!data))
 277                         break;
 278
 279                 down_read(&current->mm->mmap_sem);
 280                 result = get_user_pages(current, current->mm, user_addr,
 281                                         data->npages, 1, 0, data->pagevec, NULL);
 282                 up_read(&current->mm->mmap_sem);
 283                 if (unlikely(result < data->npages)) {
 284                         if (result > 0)
 285                                 nfs_direct_release_pages(data->pagevec, result);
 286                         nfs_readdata_release(data);
 287                         break;
 288                 }
 289
 290                 get_dreq(dreq);
 291
 292                 data->req = (struct nfs_page *) dreq;
 293                 data->inode = inode;
 294                 data->cred = ctx->cred;
 295                 data->args.fh = NFS_FH(inode);
 296                 data->args.context = ctx;
 297                 data->args.offset = pos;
 298                 data->args.pgbase = pgbase;
 299                 data->args.pages = data->pagevec;
 300                 data->args.count = bytes;
 301                 data->res.fattr = &data->fattr;
 302                 data->res.eof = 0;
 303                 data->res.count = bytes;
 304
 305                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 306                                 &nfs_read_direct_ops, data);
 307                 NFS_PROTO(inode)->read_setup(data);
 308
 309                 data->task.tk_cookie = (unsigned long) inode;
 310
 311                 rpc_execute(&data->task);
 312
 313                 dprintk("NFS: %5u initiated direct read call "
 314                         "(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                 dprintk("NFS: %5u commit failed with error %d.\n",
 437                                 task->tk_pid, task->tk_status);
 438                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 439         } else 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         rpc_execute(&data->task);
 479 }
 480
 481 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 482 {
 483         int flags = dreq->flags;
 484
 485         dreq->flags = 0;
 486         switch (flags) {
 487                 case NFS_ODIRECT_DO_COMMIT:
 488                         nfs_direct_commit_schedule(dreq);
 489                         break;
 490                 case NFS_ODIRECT_RESCHED_WRITES:
 491                         nfs_direct_write_reschedule(dreq);
 492                         break;
 493                 default:
 494                         nfs_end_data_update(inode);
 495                         if (dreq->commit_data != NULL)
 496                                 nfs_commit_free(dreq->commit_data);
 497                         nfs_direct_free_writedata(dreq);
 498                         nfs_zap_mapping(inode, inode->i_mapping);
 499                         nfs_direct_complete(dreq);
 500         }
 501 }
 502
 503 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 504 {
 505         dreq->commit_data = nfs_commit_alloc();
 506         if (dreq->commit_data != NULL)
 507                 dreq->commit_data->req = (struct nfs_page *) dreq;
 508 }
 509 #else
 510 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 511 {
 512         dreq->commit_data = NULL;
 513 }
 514
 515 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 516 {
 517         nfs_end_data_update(inode);
 518         nfs_direct_free_writedata(dreq);
 519         nfs_zap_mapping(inode, inode->i_mapping);
 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 (unlikely(dreq->error != 0))
 536                 goto out_unlock;
 537         if (unlikely(status < 0)) {
 538                 /* An error has occured, so we should not commit */
 539                 dreq->flags = 0;
 540                 dreq->error = status;
 541         }
 542
 543         dreq->count += data->res.count;
 544
 545         if (data->res.verf->committed != NFS_FILE_SYNC) {
 546                 switch (dreq->flags) {
 547                         case 0:
 548                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 549                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 550                                 break;
 551                         case NFS_ODIRECT_DO_COMMIT:
 552                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 553                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 554                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 555                                 }
 556                 }
 557         }
 558 out_unlock:
 559         spin_unlock(&dreq->lock);
 560 }
 561
 562 /*
 563  * NB: Return the value of the first error return code.  Subsequent
 564  *     errors after the first one are ignored.
 565  */
 566 static void nfs_direct_write_release(void *calldata)
 567 {
 568         struct nfs_write_data *data = calldata;
 569         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 570
 571         if (put_dreq(dreq))
 572                 nfs_direct_write_complete(dreq, data->inode);
 573 }
 574
 575 static const struct rpc_call_ops nfs_write_direct_ops = {
 576         .rpc_call_done = nfs_direct_write_result,
 577         .rpc_release = nfs_direct_write_release,
 578 };
 579
 580 /*
 581  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 582  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 583  * bail and stop sending more writes.  Write length accounting is
 584  * handled automatically by nfs_direct_write_result().  Otherwise, if
 585  * no requests have been sent, just return an error.
 586  */
 587 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 588 {
 589         struct nfs_open_context *ctx = dreq->ctx;
 590         struct inode *inode = ctx->dentry->d_inode;
 591         size_t wsize = NFS_SERVER(inode)->wsize;
 592         unsigned int pgbase;
 593         int result;
 594         ssize_t started = 0;
 595
 596         get_dreq(dreq);
 597
 598         do {
 599                 struct nfs_write_data *data;
 600                 size_t bytes;
 601
 602                 pgbase = user_addr & ~PAGE_MASK;
 603                 bytes = min(wsize,count);
 604
 605                 result = -ENOMEM;
 606                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 607                 if (unlikely(!data))
 608                         break;
 609
 610                 down_read(&current->mm->mmap_sem);
 611                 result = get_user_pages(current, current->mm, user_addr,
 612                                         data->npages, 0, 0, data->pagevec, NULL);
 613                 up_read(&current->mm->mmap_sem);
 614                 if (unlikely(result < data->npages)) {
 615                         if (result > 0)
 616                                 nfs_direct_release_pages(data->pagevec, result);
 617                         nfs_writedata_release(data);
 618                         break;
 619                 }
 620
 621                 get_dreq(dreq);
 622
 623                 list_move_tail(&data->pages, &dreq->rewrite_list);
 624
 625                 data->req = (struct nfs_page *) dreq;
 626                 data->inode = inode;
 627                 data->cred = ctx->cred;
 628                 data->args.fh = NFS_FH(inode);
 629                 data->args.context = ctx;
 630                 data->args.offset = pos;
 631                 data->args.pgbase = pgbase;
 632                 data->args.pages = data->pagevec;
 633                 data->args.count = bytes;
 634                 data->res.fattr = &data->fattr;
 635                 data->res.count = bytes;
 636                 data->res.verf = &data->verf;
 637
 638                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 639                                 &nfs_write_direct_ops, data);
 640                 NFS_PROTO(inode)->write_setup(data, sync);
 641
 642                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 643                 data->task.tk_cookie = (unsigned long) inode;
 644
 645                 rpc_execute(&data->task);
 646
 647                 dprintk("NFS: %5u initiated direct write call "
 648                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 649                                 data->task.tk_pid,
 650                                 inode->i_sb->s_id,
 651                                 (long long)NFS_FILEID(inode),
 652                                 bytes,
 653                                 (unsigned long long)data->args.offset);
 654
 655                 started += bytes;
 656                 user_addr += bytes;
 657                 pos += bytes;
 658
 659                 /* FIXME: Remove this useless math from the final patch */
 660                 pgbase += bytes;
 661                 pgbase &= ~PAGE_MASK;
 662                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 663
 664                 count -= bytes;
 665         } while (count != 0);
 666
 667         if (put_dreq(dreq))
 668                 nfs_direct_write_complete(dreq, inode);
 669
 670         if (started)
 671                 return 0;
 672         return result < 0 ? (ssize_t) result : -EFAULT;
 673 }
 674
 675 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 676 {
 677         ssize_t result = 0;
 678         sigset_t oldset;
 679         struct inode *inode = iocb->ki_filp->f_mapping->host;
 680         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 681         struct nfs_direct_req *dreq;
 682         size_t wsize = NFS_SERVER(inode)->wsize;
 683         int sync = 0;
 684
 685         dreq = nfs_direct_req_alloc();
 686         if (!dreq)
 687                 return -ENOMEM;
 688         nfs_alloc_commit_data(dreq);
 689
 690         if (dreq->commit_data == NULL || count < wsize)
 691                 sync = FLUSH_STABLE;
 692
 693         dreq->inode = inode;
 694         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 695         if (!is_sync_kiocb(iocb))
 696                 dreq->iocb = iocb;
 697
 698         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 699
 700         nfs_begin_data_update(inode);
 701
 702         rpc_clnt_sigmask(clnt, &oldset);
 703         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 704         if (!result)
 705                 result = nfs_direct_wait(dreq);
 706         rpc_clnt_sigunmask(clnt, &oldset);
 707
 708         return result;
 709 }
 710
 711 /**
 712  * nfs_file_direct_read - file direct read operation for NFS files
 713  * @iocb: target I/O control block
 714  * @iov: vector of user buffers into which to read data
 715  * @nr_segs: size of iov vector
 716  * @pos: byte offset in file where reading starts
 717  *
 718  * We use this function for direct reads instead of calling
 719  * generic_file_aio_read() in order to avoid gfar's check to see if
 720  * the request starts before the end of the file.  For that check
 721  * to work, we must generate a GETATTR before each direct read, and
 722  * even then there is a window between the GETATTR and the subsequent
 723  * READ where the file size could change.  Our preference is simply
 724  * to do all reads the application wants, and the server will take
 725  * care of managing the end of file boundary.
 726  *
 727  * This function also eliminates unnecessarily updating the file's
 728  * atime locally, as the NFS server sets the file's atime, and this
 729  * client must read the updated atime from the server back into its
 730  * cache.
 731  */
 732 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 733                                 unsigned long nr_segs, loff_t pos)
 734 {
 735         ssize_t retval = -EINVAL;
 736         struct file *file = iocb->ki_filp;
 737         struct address_space *mapping = file->f_mapping;
 738         /* XXX: temporary */
 739         const char __user *buf = iov[0].iov_base;
 740         size_t count = iov[0].iov_len;
 741
 742         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 743                 file->f_path.dentry->d_parent->d_name.name,
 744                 file->f_path.dentry->d_name.name,
 745                 (unsigned long) count, (long long) pos);
 746
 747         if (nr_segs != 1)
 748                 return -EINVAL;
 749
 750         if (count < 0)
 751                 goto out;
 752         retval = -EFAULT;
 753         if (!access_ok(VERIFY_WRITE, buf, count))
 754                 goto out;
 755         retval = 0;
 756         if (!count)
 757                 goto out;
 758
 759         retval = nfs_sync_mapping(mapping);
 760         if (retval)
 761                 goto out;
 762
 763         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 764         if (retval > 0)
 765                 iocb->ki_pos = pos + retval;
 766
 767 out:
 768         return retval;
 769 }
 770
 771 /**
 772  * nfs_file_direct_write - file direct write operation for NFS files
 773  * @iocb: target I/O control block
 774  * @iov: vector of user buffers from which to write data
 775  * @nr_segs: size of iov vector
 776  * @pos: byte offset in file where writing starts
 777  *
 778  * We use this function for direct writes instead of calling
 779  * generic_file_aio_write() in order to avoid taking the inode
 780  * semaphore and updating the i_size.  The NFS server will set
 781  * the new i_size and this client must read the updated size
 782  * back into its cache.  We let the server do generic write
 783  * parameter checking and report problems.
 784  *
 785  * We also avoid an unnecessary invocation of generic_osync_inode(),
 786  * as it is fairly meaningless to sync the metadata of an NFS file.
 787  *
 788  * We eliminate local atime updates, see direct read above.
 789  *
 790  * We avoid unnecessary page cache invalidations for normal cached
 791  * readers of this file.
 792  *
 793  * Note that O_APPEND is not supported for NFS direct writes, as there
 794  * is no atomic O_APPEND write facility in the NFS protocol.
 795  */
 796 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 797                                 unsigned long nr_segs, loff_t pos)
 798 {
 799         ssize_t retval;
 800         struct file *file = iocb->ki_filp;
 801         struct address_space *mapping = file->f_mapping;
 802         /* XXX: temporary */
 803         const char __user *buf = iov[0].iov_base;
 804         size_t count = iov[0].iov_len;
 805
 806         dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
 807                 file->f_path.dentry->d_parent->d_name.name,
 808                 file->f_path.dentry->d_name.name,
 809                 (unsigned long) count, (long long) pos);
 810
 811         if (nr_segs != 1)
 812                 return -EINVAL;
 813
 814         retval = generic_write_checks(file, &pos, &count, 0);
 815         if (retval)
 816                 goto out;
 817
 818         retval = -EINVAL;
 819         if ((ssize_t) count < 0)
 820                 goto out;
 821         retval = 0;
 822         if (!count)
 823                 goto out;
 824
 825         retval = -EFAULT;
 826         if (!access_ok(VERIFY_READ, buf, count))
 827                 goto out;
 828
 829         retval = nfs_sync_mapping(mapping);
 830         if (retval)
 831                 goto out;
 832
 833         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 834
 835         if (retval > 0)
 836                 iocb->ki_pos = pos + retval;
 837
 838 out:
 839         return retval;
 840 }
 841
 842 /**
 843  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 844  *
 845  */
 846 int __init nfs_init_directcache(void)
 847 {
 848         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 849                                                 sizeof(struct nfs_direct_req),
 850                                                 0, (SLAB_RECLAIM_ACCOUNT|
 851                                                         SLAB_MEM_SPREAD),
 852                                                 NULL, NULL);
 853         if (nfs_direct_cachep == NULL)
 854                 return -ENOMEM;
 855
 856         return 0;
 857 }
 858
 859 /**
 860  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 861  *
 862  */
 863 void nfs_destroy_directcache(void)
 864 {
 865         kmem_cache_destroy(nfs_direct_cachep);
 866 }