git.oblomov.eu Git - 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/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47
  48 #include <linux/nfs_fs.h>
  49 #include <linux/nfs_page.h>
  50 #include <linux/sunrpc/clnt.h>
  51
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54 #include <asm/atomic.h>
  55
  56 #include "internal.h"
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60
  61 static struct kmem_cache *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_path.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, unsigned int npages)
 126 {
 127         unsigned 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, unsigned int npages)
 136 {
 137         unsigned 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, GFP_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(nfs_page_array_len(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 (result < 0) {
 283                         nfs_readdata_release(data);
 284                         break;
 285                 }
 286                 if ((unsigned)result < data->npages) {
 287                         nfs_direct_release_pages(data->pagevec, result);
 288                         nfs_readdata_release(data);
 289                         break;
 290                 }
 291
 292                 get_dreq(dreq);
 293
 294                 data->req = (struct nfs_page *) dreq;
 295                 data->inode = inode;
 296                 data->cred = ctx->cred;
 297                 data->args.fh = NFS_FH(inode);
 298                 data->args.context = ctx;
 299                 data->args.offset = pos;
 300                 data->args.pgbase = pgbase;
 301                 data->args.pages = data->pagevec;
 302                 data->args.count = bytes;
 303                 data->res.fattr = &data->fattr;
 304                 data->res.eof = 0;
 305                 data->res.count = bytes;
 306
 307                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 308                                 &nfs_read_direct_ops, data);
 309                 NFS_PROTO(inode)->read_setup(data);
 310
 311                 data->task.tk_cookie = (unsigned long) inode;
 312
 313                 rpc_execute(&data->task);
 314
 315                 dprintk("NFS: %5u initiated direct read call "
 316                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 317                                 data->task.tk_pid,
 318                                 inode->i_sb->s_id,
 319                                 (long long)NFS_FILEID(inode),
 320                                 bytes,
 321                                 (unsigned long long)data->args.offset);
 322
 323                 started += bytes;
 324                 user_addr += bytes;
 325                 pos += bytes;
 326                 /* FIXME: Remove this unnecessary math from final patch */
 327                 pgbase += bytes;
 328                 pgbase &= ~PAGE_MASK;
 329                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 330
 331                 count -= bytes;
 332         } while (count != 0);
 333
 334         if (put_dreq(dreq))
 335                 nfs_direct_complete(dreq);
 336
 337         if (started)
 338                 return 0;
 339         return result < 0 ? (ssize_t) result : -EFAULT;
 340 }
 341
 342 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 343 {
 344         ssize_t result = 0;
 345         sigset_t oldset;
 346         struct inode *inode = iocb->ki_filp->f_mapping->host;
 347         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 348         struct nfs_direct_req *dreq;
 349
 350         dreq = nfs_direct_req_alloc();
 351         if (!dreq)
 352                 return -ENOMEM;
 353
 354         dreq->inode = inode;
 355         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 356         if (!is_sync_kiocb(iocb))
 357                 dreq->iocb = iocb;
 358
 359         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 360         rpc_clnt_sigmask(clnt, &oldset);
 361         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 362         if (!result)
 363                 result = nfs_direct_wait(dreq);
 364         rpc_clnt_sigunmask(clnt, &oldset);
 365
 366         return result;
 367 }
 368
 369 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 370 {
 371         while (!list_empty(&dreq->rewrite_list)) {
 372                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 373                 list_del(&data->pages);
 374                 nfs_direct_release_pages(data->pagevec, data->npages);
 375                 nfs_writedata_release(data);
 376         }
 377 }
 378
 379 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 380 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 381 {
 382         struct inode *inode = dreq->inode;
 383         struct list_head *p;
 384         struct nfs_write_data *data;
 385
 386         dreq->count = 0;
 387         get_dreq(dreq);
 388
 389         list_for_each(p, &dreq->rewrite_list) {
 390                 data = list_entry(p, struct nfs_write_data, pages);
 391
 392                 get_dreq(dreq);
 393
 394                 /*
 395                  * Reset data->res.
 396                  */
 397                 nfs_fattr_init(&data->fattr);
 398                 data->res.count = data->args.count;
 399                 memset(&data->verf, 0, sizeof(data->verf));
 400
 401                 /*
 402                  * Reuse data->task; data->args should not have changed
 403                  * since the original request was sent.
 404                  */
 405                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 406                                 &nfs_write_direct_ops, data);
 407                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 408
 409                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 410                 data->task.tk_cookie = (unsigned long) inode;
 411
 412                 /*
 413                  * We're called via an RPC callback, so BKL is already held.
 414                  */
 415                 rpc_execute(&data->task);
 416
 417                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 418                                 data->task.tk_pid,
 419                                 inode->i_sb->s_id,
 420                                 (long long)NFS_FILEID(inode),
 421                                 data->args.count,
 422                                 (unsigned long long)data->args.offset);
 423         }
 424
 425         if (put_dreq(dreq))
 426                 nfs_direct_write_complete(dreq, inode);
 427 }
 428
 429 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 430 {
 431         struct nfs_write_data *data = calldata;
 432         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 433
 434         /* Call the NFS version-specific code */
 435         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 436                 return;
 437         if (unlikely(task->tk_status < 0)) {
 438                 dprintk("NFS: %5u commit failed with error %d.\n",
 439                                 task->tk_pid, task->tk_status);
 440                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 441         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 442                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 443                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 444         }
 445
 446         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 447         nfs_direct_write_complete(dreq, data->inode);
 448 }
 449
 450 static const struct rpc_call_ops nfs_commit_direct_ops = {
 451         .rpc_call_done = nfs_direct_commit_result,
 452         .rpc_release = nfs_commit_release,
 453 };
 454
 455 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 456 {
 457         struct nfs_write_data *data = dreq->commit_data;
 458
 459         data->inode = dreq->inode;
 460         data->cred = dreq->ctx->cred;
 461
 462         data->args.fh = NFS_FH(data->inode);
 463         data->args.offset = 0;
 464         data->args.count = 0;
 465         data->res.count = 0;
 466         data->res.fattr = &data->fattr;
 467         data->res.verf = &data->verf;
 468
 469         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 470                                 &nfs_commit_direct_ops, data);
 471         NFS_PROTO(data->inode)->commit_setup(data, 0);
 472
 473         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 474         data->task.tk_cookie = (unsigned long)data->inode;
 475         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 476         dreq->commit_data = NULL;
 477
 478         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 479
 480         rpc_execute(&data->task);
 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_zap_mapping(inode, inode->i_mapping);
 501                         nfs_direct_complete(dreq);
 502         }
 503 }
 504
 505 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 506 {
 507         dreq->commit_data = nfs_commit_alloc();
 508         if (dreq->commit_data != NULL)
 509                 dreq->commit_data->req = (struct nfs_page *) dreq;
 510 }
 511 #else
 512 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 513 {
 514         dreq->commit_data = NULL;
 515 }
 516
 517 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 518 {
 519         nfs_end_data_update(inode);
 520         nfs_direct_free_writedata(dreq);
 521         nfs_zap_mapping(inode, inode->i_mapping);
 522         nfs_direct_complete(dreq);
 523 }
 524 #endif
 525
 526 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 527 {
 528         struct nfs_write_data *data = calldata;
 529         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 530         int status = task->tk_status;
 531
 532         if (nfs_writeback_done(task, data) != 0)
 533                 return;
 534
 535         spin_lock(&dreq->lock);
 536
 537         if (unlikely(dreq->error != 0))
 538                 goto out_unlock;
 539         if (unlikely(status < 0)) {
 540                 /* An error has occured, so we should not commit */
 541                 dreq->flags = 0;
 542                 dreq->error = status;
 543         }
 544
 545         dreq->count += data->res.count;
 546
 547         if (data->res.verf->committed != NFS_FILE_SYNC) {
 548                 switch (dreq->flags) {
 549                         case 0:
 550                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 551                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 552                                 break;
 553                         case NFS_ODIRECT_DO_COMMIT:
 554                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 555                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 556                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 557                                 }
 558                 }
 559         }
 560 out_unlock:
 561         spin_unlock(&dreq->lock);
 562 }
 563
 564 /*
 565  * NB: Return the value of the first error return code.  Subsequent
 566  *     errors after the first one are ignored.
 567  */
 568 static void nfs_direct_write_release(void *calldata)
 569 {
 570         struct nfs_write_data *data = calldata;
 571         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 572
 573         if (put_dreq(dreq))
 574                 nfs_direct_write_complete(dreq, data->inode);
 575 }
 576
 577 static const struct rpc_call_ops nfs_write_direct_ops = {
 578         .rpc_call_done = nfs_direct_write_result,
 579         .rpc_release = nfs_direct_write_release,
 580 };
 581
 582 /*
 583  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 584  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 585  * bail and stop sending more writes.  Write length accounting is
 586  * handled automatically by nfs_direct_write_result().  Otherwise, if
 587  * no requests have been sent, just return an error.
 588  */
 589 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 590 {
 591         struct nfs_open_context *ctx = dreq->ctx;
 592         struct inode *inode = ctx->dentry->d_inode;
 593         size_t wsize = NFS_SERVER(inode)->wsize;
 594         unsigned int pgbase;
 595         int result;
 596         ssize_t started = 0;
 597
 598         get_dreq(dreq);
 599
 600         do {
 601                 struct nfs_write_data *data;
 602                 size_t bytes;
 603
 604                 pgbase = user_addr & ~PAGE_MASK;
 605                 bytes = min(wsize,count);
 606
 607                 result = -ENOMEM;
 608                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 609                 if (unlikely(!data))
 610                         break;
 611
 612                 down_read(&current->mm->mmap_sem);
 613                 result = get_user_pages(current, current->mm, user_addr,
 614                                         data->npages, 0, 0, data->pagevec, NULL);
 615                 up_read(&current->mm->mmap_sem);
 616                 if (result < 0) {
 617                         nfs_writedata_release(data);
 618                         break;
 619                 }
 620                 if ((unsigned)result < data->npages) {
 621                         nfs_direct_release_pages(data->pagevec, result);
 622                         nfs_writedata_release(data);
 623                         break;
 624                 }
 625
 626                 get_dreq(dreq);
 627
 628                 list_move_tail(&data->pages, &dreq->rewrite_list);
 629
 630                 data->req = (struct nfs_page *) dreq;
 631                 data->inode = inode;
 632                 data->cred = ctx->cred;
 633                 data->args.fh = NFS_FH(inode);
 634                 data->args.context = ctx;
 635                 data->args.offset = pos;
 636                 data->args.pgbase = pgbase;
 637                 data->args.pages = data->pagevec;
 638                 data->args.count = bytes;
 639                 data->res.fattr = &data->fattr;
 640                 data->res.count = bytes;
 641                 data->res.verf = &data->verf;
 642
 643                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 644                                 &nfs_write_direct_ops, data);
 645                 NFS_PROTO(inode)->write_setup(data, sync);
 646
 647                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 648                 data->task.tk_cookie = (unsigned long) inode;
 649
 650                 rpc_execute(&data->task);
 651
 652                 dprintk("NFS: %5u initiated direct write call "
 653                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 654                                 data->task.tk_pid,
 655                                 inode->i_sb->s_id,
 656                                 (long long)NFS_FILEID(inode),
 657                                 bytes,
 658                                 (unsigned long long)data->args.offset);
 659
 660                 started += bytes;
 661                 user_addr += bytes;
 662                 pos += bytes;
 663
 664                 /* FIXME: Remove this useless math from the final patch */
 665                 pgbase += bytes;
 666                 pgbase &= ~PAGE_MASK;
 667                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 668
 669                 count -= bytes;
 670         } while (count != 0);
 671
 672         if (put_dreq(dreq))
 673                 nfs_direct_write_complete(dreq, inode);
 674
 675         if (started)
 676                 return 0;
 677         return result < 0 ? (ssize_t) result : -EFAULT;
 678 }
 679
 680 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 681 {
 682         ssize_t result = 0;
 683         sigset_t oldset;
 684         struct inode *inode = iocb->ki_filp->f_mapping->host;
 685         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 686         struct nfs_direct_req *dreq;
 687         size_t wsize = NFS_SERVER(inode)->wsize;
 688         int sync = 0;
 689
 690         dreq = nfs_direct_req_alloc();
 691         if (!dreq)
 692                 return -ENOMEM;
 693         nfs_alloc_commit_data(dreq);
 694
 695         if (dreq->commit_data == NULL || count < wsize)
 696                 sync = FLUSH_STABLE;
 697
 698         dreq->inode = inode;
 699         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 700         if (!is_sync_kiocb(iocb))
 701                 dreq->iocb = iocb;
 702
 703         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 704
 705         nfs_begin_data_update(inode);
 706
 707         rpc_clnt_sigmask(clnt, &oldset);
 708         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 709         if (!result)
 710                 result = nfs_direct_wait(dreq);
 711         rpc_clnt_sigunmask(clnt, &oldset);
 712
 713         return result;
 714 }
 715
 716 /**
 717  * nfs_file_direct_read - file direct read operation for NFS files
 718  * @iocb: target I/O control block
 719  * @iov: vector of user buffers into which to read data
 720  * @nr_segs: size of iov vector
 721  * @pos: byte offset in file where reading starts
 722  *
 723  * We use this function for direct reads instead of calling
 724  * generic_file_aio_read() in order to avoid gfar's check to see if
 725  * the request starts before the end of the file.  For that check
 726  * to work, we must generate a GETATTR before each direct read, and
 727  * even then there is a window between the GETATTR and the subsequent
 728  * READ where the file size could change.  Our preference is simply
 729  * to do all reads the application wants, and the server will take
 730  * care of managing the end of file boundary.
 731  *
 732  * This function also eliminates unnecessarily updating the file's
 733  * atime locally, as the NFS server sets the file's atime, and this
 734  * client must read the updated atime from the server back into its
 735  * cache.
 736  */
 737 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 738                                 unsigned long nr_segs, loff_t pos)
 739 {
 740         ssize_t retval = -EINVAL;
 741         struct file *file = iocb->ki_filp;
 742         struct address_space *mapping = file->f_mapping;
 743         /* XXX: temporary */
 744         const char __user *buf = iov[0].iov_base;
 745         size_t count = iov[0].iov_len;
 746
 747         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 748                 file->f_path.dentry->d_parent->d_name.name,
 749                 file->f_path.dentry->d_name.name,
 750                 (unsigned long) count, (long long) pos);
 751
 752         if (nr_segs != 1)
 753                 return -EINVAL;
 754
 755         if (count < 0)
 756                 goto out;
 757         retval = -EFAULT;
 758         if (!access_ok(VERIFY_WRITE, buf, count))
 759                 goto out;
 760         retval = 0;
 761         if (!count)
 762                 goto out;
 763
 764         retval = nfs_sync_mapping(mapping);
 765         if (retval)
 766                 goto out;
 767
 768         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 769         if (retval > 0)
 770                 iocb->ki_pos = pos + retval;
 771
 772 out:
 773         return retval;
 774 }
 775
 776 /**
 777  * nfs_file_direct_write - file direct write operation for NFS files
 778  * @iocb: target I/O control block
 779  * @iov: vector of user buffers from which to write data
 780  * @nr_segs: size of iov vector
 781  * @pos: byte offset in file where writing starts
 782  *
 783  * We use this function for direct writes instead of calling
 784  * generic_file_aio_write() in order to avoid taking the inode
 785  * semaphore and updating the i_size.  The NFS server will set
 786  * the new i_size and this client must read the updated size
 787  * back into its cache.  We let the server do generic write
 788  * parameter checking and report problems.
 789  *
 790  * We also avoid an unnecessary invocation of generic_osync_inode(),
 791  * as it is fairly meaningless to sync the metadata of an NFS file.
 792  *
 793  * We eliminate local atime updates, see direct read above.
 794  *
 795  * We avoid unnecessary page cache invalidations for normal cached
 796  * readers of this file.
 797  *
 798  * Note that O_APPEND is not supported for NFS direct writes, as there
 799  * is no atomic O_APPEND write facility in the NFS protocol.
 800  */
 801 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 802                                 unsigned long nr_segs, loff_t pos)
 803 {
 804         ssize_t retval;
 805         struct file *file = iocb->ki_filp;
 806         struct address_space *mapping = file->f_mapping;
 807         /* XXX: temporary */
 808         const char __user *buf = iov[0].iov_base;
 809         size_t count = iov[0].iov_len;
 810
 811         dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
 812                 file->f_path.dentry->d_parent->d_name.name,
 813                 file->f_path.dentry->d_name.name,
 814                 (unsigned long) count, (long long) pos);
 815
 816         if (nr_segs != 1)
 817                 return -EINVAL;
 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         if (retval > 0)
 841                 iocb->ki_pos = pos + retval;
 842
 843 out:
 844         return retval;
 845 }
 846
 847 /**
 848  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 849  *
 850  */
 851 int __init nfs_init_directcache(void)
 852 {
 853         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 854                                                 sizeof(struct nfs_direct_req),
 855                                                 0, (SLAB_RECLAIM_ACCOUNT|
 856                                                         SLAB_MEM_SPREAD),
 857                                                 NULL, NULL);
 858         if (nfs_direct_cachep == NULL)
 859                 return -ENOMEM;
 860
 861         return 0;
 862 }
 863
 864 /**
 865  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 866  *
 867  */
 868 void nfs_destroy_directcache(void)
 869 {
 870         kmem_cache_destroy(nfs_direct_cachep);
 871 }