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/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 pgbase, size_t count)
 126 {
 127         unsigned int npages;
 128         unsigned int i;
 129
 130         if (count == 0)
 131                 return;
 132         pages += (pgbase >> PAGE_SHIFT);
 133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 134         for (i = 0; i < npages; i++) {
 135                 struct page *page = pages[i];
 136                 if (!PageCompound(page))
 137                         set_page_dirty(page);
 138         }
 139 }
 140
 141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 142 {
 143         unsigned int i;
 144         for (i = 0; i < npages; i++)
 145                 page_cache_release(pages[i]);
 146 }
 147
 148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 149 {
 150         struct nfs_direct_req *dreq;
 151
 152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 153         if (!dreq)
 154                 return NULL;
 155
 156         kref_init(&dreq->kref);
 157         kref_get(&dreq->kref);
 158         init_completion(&dreq->completion);
 159         INIT_LIST_HEAD(&dreq->rewrite_list);
 160         dreq->iocb = NULL;
 161         dreq->ctx = NULL;
 162         spin_lock_init(&dreq->lock);
 163         atomic_set(&dreq->io_count, 0);
 164         dreq->count = 0;
 165         dreq->error = 0;
 166         dreq->flags = 0;
 167
 168         return dreq;
 169 }
 170
 171 static void nfs_direct_req_free(struct kref *kref)
 172 {
 173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 174
 175         if (dreq->ctx != NULL)
 176                 put_nfs_open_context(dreq->ctx);
 177         kmem_cache_free(nfs_direct_cachep, dreq);
 178 }
 179
 180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 181 {
 182         kref_put(&dreq->kref, nfs_direct_req_free);
 183 }
 184
 185 /*
 186  * Collects and returns the final error value/byte-count.
 187  */
 188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 189 {
 190         ssize_t result = -EIOCBQUEUED;
 191
 192         /* Async requests don't wait here */
 193         if (dreq->iocb)
 194                 goto out;
 195
 196         result = wait_for_completion_interruptible(&dreq->completion);
 197
 198         if (!result)
 199                 result = dreq->error;
 200         if (!result)
 201                 result = dreq->count;
 202
 203 out:
 204         return (ssize_t) result;
 205 }
 206
 207 /*
 208  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 209  * the iocb is still valid here if this is a synchronous request.
 210  */
 211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 212 {
 213         if (dreq->iocb) {
 214                 long res = (long) dreq->error;
 215                 if (!res)
 216                         res = (long) dreq->count;
 217                 aio_complete(dreq->iocb, res, 0);
 218         }
 219         complete_all(&dreq->completion);
 220
 221         nfs_direct_req_release(dreq);
 222 }
 223
 224 /*
 225  * We must hold a reference to all the pages in this direct read request
 226  * until the RPCs complete.  This could be long *after* we are woken up in
 227  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 228  */
 229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 230 {
 231         struct nfs_read_data *data = calldata;
 232         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 233
 234         if (nfs_readpage_result(task, data) != 0)
 235                 return;
 236
 237         spin_lock(&dreq->lock);
 238         if (unlikely(task->tk_status < 0)) {
 239                 dreq->error = task->tk_status;
 240                 spin_unlock(&dreq->lock);
 241         } else {
 242                 dreq->count += data->res.count;
 243                 spin_unlock(&dreq->lock);
 244                 nfs_direct_dirty_pages(data->pagevec,
 245                                 data->args.pgbase,
 246                                 data->res.count);
 247         }
 248         nfs_direct_release_pages(data->pagevec, data->npages);
 249
 250         if (put_dreq(dreq))
 251                 nfs_direct_complete(dreq);
 252 }
 253
 254 static const struct rpc_call_ops nfs_read_direct_ops = {
 255         .rpc_call_done = nfs_direct_read_result,
 256         .rpc_release = nfs_readdata_release,
 257 };
 258
 259 /*
 260  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 261  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 262  * bail and stop sending more reads.  Read length accounting is
 263  * handled automatically by nfs_direct_read_result().  Otherwise, if
 264  * no requests have been sent, just return an error.
 265  */
 266 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 267 {
 268         struct nfs_open_context *ctx = dreq->ctx;
 269         struct inode *inode = ctx->path.dentry->d_inode;
 270         size_t rsize = NFS_SERVER(inode)->rsize;
 271         unsigned int pgbase;
 272         int result;
 273         ssize_t started = 0;
 274
 275         get_dreq(dreq);
 276
 277         do {
 278                 struct nfs_read_data *data;
 279                 size_t bytes;
 280
 281                 pgbase = user_addr & ~PAGE_MASK;
 282                 bytes = min(rsize,count);
 283
 284                 result = -ENOMEM;
 285                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 286                 if (unlikely(!data))
 287                         break;
 288
 289                 down_read(&current->mm->mmap_sem);
 290                 result = get_user_pages(current, current->mm, user_addr,
 291                                         data->npages, 1, 0, data->pagevec, NULL);
 292                 up_read(&current->mm->mmap_sem);
 293                 if (result < 0) {
 294                         nfs_readdata_release(data);
 295                         break;
 296                 }
 297                 if ((unsigned)result < data->npages) {
 298                         bytes = result * PAGE_SIZE;
 299                         if (bytes <= pgbase) {
 300                                 nfs_direct_release_pages(data->pagevec, result);
 301                                 nfs_readdata_release(data);
 302                                 break;
 303                         }
 304                         bytes -= pgbase;
 305                         data->npages = result;
 306                 }
 307
 308                 get_dreq(dreq);
 309
 310                 data->req = (struct nfs_page *) dreq;
 311                 data->inode = inode;
 312                 data->cred = ctx->cred;
 313                 data->args.fh = NFS_FH(inode);
 314                 data->args.context = ctx;
 315                 data->args.offset = pos;
 316                 data->args.pgbase = pgbase;
 317                 data->args.pages = data->pagevec;
 318                 data->args.count = bytes;
 319                 data->res.fattr = &data->fattr;
 320                 data->res.eof = 0;
 321                 data->res.count = bytes;
 322
 323                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 324                                 &nfs_read_direct_ops, data);
 325                 NFS_PROTO(inode)->read_setup(data);
 326
 327                 data->task.tk_cookie = (unsigned long) inode;
 328
 329                 rpc_execute(&data->task);
 330
 331                 dprintk("NFS: %5u initiated direct read call "
 332                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 333                                 data->task.tk_pid,
 334                                 inode->i_sb->s_id,
 335                                 (long long)NFS_FILEID(inode),
 336                                 bytes,
 337                                 (unsigned long long)data->args.offset);
 338
 339                 started += bytes;
 340                 user_addr += bytes;
 341                 pos += bytes;
 342                 /* FIXME: Remove this unnecessary math from final patch */
 343                 pgbase += bytes;
 344                 pgbase &= ~PAGE_MASK;
 345                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 346
 347                 count -= bytes;
 348         } while (count != 0);
 349
 350         if (put_dreq(dreq))
 351                 nfs_direct_complete(dreq);
 352
 353         if (started)
 354                 return 0;
 355         return result < 0 ? (ssize_t) result : -EFAULT;
 356 }
 357
 358 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 359 {
 360         ssize_t result = 0;
 361         sigset_t oldset;
 362         struct inode *inode = iocb->ki_filp->f_mapping->host;
 363         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 364         struct nfs_direct_req *dreq;
 365
 366         dreq = nfs_direct_req_alloc();
 367         if (!dreq)
 368                 return -ENOMEM;
 369
 370         dreq->inode = inode;
 371         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 372         if (!is_sync_kiocb(iocb))
 373                 dreq->iocb = iocb;
 374
 375         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 376         rpc_clnt_sigmask(clnt, &oldset);
 377         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 378         if (!result)
 379                 result = nfs_direct_wait(dreq);
 380         rpc_clnt_sigunmask(clnt, &oldset);
 381         nfs_direct_req_release(dreq);
 382
 383         return result;
 384 }
 385
 386 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 387 {
 388         while (!list_empty(&dreq->rewrite_list)) {
 389                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 390                 list_del(&data->pages);
 391                 nfs_direct_release_pages(data->pagevec, data->npages);
 392                 nfs_writedata_release(data);
 393         }
 394 }
 395
 396 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 397 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 398 {
 399         struct inode *inode = dreq->inode;
 400         struct list_head *p;
 401         struct nfs_write_data *data;
 402
 403         dreq->count = 0;
 404         get_dreq(dreq);
 405
 406         list_for_each(p, &dreq->rewrite_list) {
 407                 data = list_entry(p, struct nfs_write_data, pages);
 408
 409                 get_dreq(dreq);
 410
 411                 /*
 412                  * Reset data->res.
 413                  */
 414                 nfs_fattr_init(&data->fattr);
 415                 data->res.count = data->args.count;
 416                 memset(&data->verf, 0, sizeof(data->verf));
 417
 418                 /*
 419                  * Reuse data->task; data->args should not have changed
 420                  * since the original request was sent.
 421                  */
 422                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 423                                 &nfs_write_direct_ops, data);
 424                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 425
 426                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 427                 data->task.tk_cookie = (unsigned long) inode;
 428
 429                 /*
 430                  * We're called via an RPC callback, so BKL is already held.
 431                  */
 432                 rpc_execute(&data->task);
 433
 434                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 435                                 data->task.tk_pid,
 436                                 inode->i_sb->s_id,
 437                                 (long long)NFS_FILEID(inode),
 438                                 data->args.count,
 439                                 (unsigned long long)data->args.offset);
 440         }
 441
 442         if (put_dreq(dreq))
 443                 nfs_direct_write_complete(dreq, inode);
 444 }
 445
 446 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 447 {
 448         struct nfs_write_data *data = calldata;
 449         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 450
 451         /* Call the NFS version-specific code */
 452         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 453                 return;
 454         if (unlikely(task->tk_status < 0)) {
 455                 dprintk("NFS: %5u commit failed with error %d.\n",
 456                                 task->tk_pid, task->tk_status);
 457                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 458         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 459                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 460                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 461         }
 462
 463         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 464         nfs_direct_write_complete(dreq, data->inode);
 465 }
 466
 467 static const struct rpc_call_ops nfs_commit_direct_ops = {
 468         .rpc_call_done = nfs_direct_commit_result,
 469         .rpc_release = nfs_commit_release,
 470 };
 471
 472 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 473 {
 474         struct nfs_write_data *data = dreq->commit_data;
 475
 476         data->inode = dreq->inode;
 477         data->cred = dreq->ctx->cred;
 478
 479         data->args.fh = NFS_FH(data->inode);
 480         data->args.offset = 0;
 481         data->args.count = 0;
 482         data->res.count = 0;
 483         data->res.fattr = &data->fattr;
 484         data->res.verf = &data->verf;
 485
 486         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 487                                 &nfs_commit_direct_ops, data);
 488         NFS_PROTO(data->inode)->commit_setup(data, 0);
 489
 490         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 491         data->task.tk_cookie = (unsigned long)data->inode;
 492         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 493         dreq->commit_data = NULL;
 494
 495         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 496
 497         rpc_execute(&data->task);
 498 }
 499
 500 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 501 {
 502         int flags = dreq->flags;
 503
 504         dreq->flags = 0;
 505         switch (flags) {
 506                 case NFS_ODIRECT_DO_COMMIT:
 507                         nfs_direct_commit_schedule(dreq);
 508                         break;
 509                 case NFS_ODIRECT_RESCHED_WRITES:
 510                         nfs_direct_write_reschedule(dreq);
 511                         break;
 512                 default:
 513                         nfs_end_data_update(inode);
 514                         if (dreq->commit_data != NULL)
 515                                 nfs_commit_free(dreq->commit_data);
 516                         nfs_direct_free_writedata(dreq);
 517                         nfs_zap_mapping(inode, inode->i_mapping);
 518                         nfs_direct_complete(dreq);
 519         }
 520 }
 521
 522 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 523 {
 524         dreq->commit_data = nfs_commit_alloc();
 525         if (dreq->commit_data != NULL)
 526                 dreq->commit_data->req = (struct nfs_page *) dreq;
 527 }
 528 #else
 529 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 530 {
 531         dreq->commit_data = NULL;
 532 }
 533
 534 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 535 {
 536         nfs_end_data_update(inode);
 537         nfs_direct_free_writedata(dreq);
 538         nfs_zap_mapping(inode, inode->i_mapping);
 539         nfs_direct_complete(dreq);
 540 }
 541 #endif
 542
 543 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 544 {
 545         struct nfs_write_data *data = calldata;
 546         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 547         int status = task->tk_status;
 548
 549         if (nfs_writeback_done(task, data) != 0)
 550                 return;
 551
 552         spin_lock(&dreq->lock);
 553
 554         if (unlikely(dreq->error != 0))
 555                 goto out_unlock;
 556         if (unlikely(status < 0)) {
 557                 /* An error has occured, so we should not commit */
 558                 dreq->flags = 0;
 559                 dreq->error = status;
 560         }
 561
 562         dreq->count += data->res.count;
 563
 564         if (data->res.verf->committed != NFS_FILE_SYNC) {
 565                 switch (dreq->flags) {
 566                         case 0:
 567                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 568                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 569                                 break;
 570                         case NFS_ODIRECT_DO_COMMIT:
 571                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 572                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 573                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 574                                 }
 575                 }
 576         }
 577 out_unlock:
 578         spin_unlock(&dreq->lock);
 579 }
 580
 581 /*
 582  * NB: Return the value of the first error return code.  Subsequent
 583  *     errors after the first one are ignored.
 584  */
 585 static void nfs_direct_write_release(void *calldata)
 586 {
 587         struct nfs_write_data *data = calldata;
 588         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 589
 590         if (put_dreq(dreq))
 591                 nfs_direct_write_complete(dreq, data->inode);
 592 }
 593
 594 static const struct rpc_call_ops nfs_write_direct_ops = {
 595         .rpc_call_done = nfs_direct_write_result,
 596         .rpc_release = nfs_direct_write_release,
 597 };
 598
 599 /*
 600  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 601  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 602  * bail and stop sending more writes.  Write length accounting is
 603  * handled automatically by nfs_direct_write_result().  Otherwise, if
 604  * no requests have been sent, just return an error.
 605  */
 606 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 607 {
 608         struct nfs_open_context *ctx = dreq->ctx;
 609         struct inode *inode = ctx->path.dentry->d_inode;
 610         size_t wsize = NFS_SERVER(inode)->wsize;
 611         unsigned int pgbase;
 612         int result;
 613         ssize_t started = 0;
 614
 615         get_dreq(dreq);
 616
 617         do {
 618                 struct nfs_write_data *data;
 619                 size_t bytes;
 620
 621                 pgbase = user_addr & ~PAGE_MASK;
 622                 bytes = min(wsize,count);
 623
 624                 result = -ENOMEM;
 625                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 626                 if (unlikely(!data))
 627                         break;
 628
 629                 down_read(&current->mm->mmap_sem);
 630                 result = get_user_pages(current, current->mm, user_addr,
 631                                         data->npages, 0, 0, data->pagevec, NULL);
 632                 up_read(&current->mm->mmap_sem);
 633                 if (result < 0) {
 634                         nfs_writedata_release(data);
 635                         break;
 636                 }
 637                 if ((unsigned)result < data->npages) {
 638                         bytes = result * PAGE_SIZE;
 639                         if (bytes <= pgbase) {
 640                                 nfs_direct_release_pages(data->pagevec, result);
 641                                 nfs_writedata_release(data);
 642                                 break;
 643                         }
 644                         bytes -= pgbase;
 645                         data->npages = result;
 646                 }
 647
 648                 get_dreq(dreq);
 649
 650                 list_move_tail(&data->pages, &dreq->rewrite_list);
 651
 652                 data->req = (struct nfs_page *) dreq;
 653                 data->inode = inode;
 654                 data->cred = ctx->cred;
 655                 data->args.fh = NFS_FH(inode);
 656                 data->args.context = ctx;
 657                 data->args.offset = pos;
 658                 data->args.pgbase = pgbase;
 659                 data->args.pages = data->pagevec;
 660                 data->args.count = bytes;
 661                 data->res.fattr = &data->fattr;
 662                 data->res.count = bytes;
 663                 data->res.verf = &data->verf;
 664
 665                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 666                                 &nfs_write_direct_ops, data);
 667                 NFS_PROTO(inode)->write_setup(data, sync);
 668
 669                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 670                 data->task.tk_cookie = (unsigned long) inode;
 671
 672                 rpc_execute(&data->task);
 673
 674                 dprintk("NFS: %5u initiated direct write call "
 675                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 676                                 data->task.tk_pid,
 677                                 inode->i_sb->s_id,
 678                                 (long long)NFS_FILEID(inode),
 679                                 bytes,
 680                                 (unsigned long long)data->args.offset);
 681
 682                 started += bytes;
 683                 user_addr += bytes;
 684                 pos += bytes;
 685
 686                 /* FIXME: Remove this useless math from the final patch */
 687                 pgbase += bytes;
 688                 pgbase &= ~PAGE_MASK;
 689                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 690
 691                 count -= bytes;
 692         } while (count != 0);
 693
 694         if (put_dreq(dreq))
 695                 nfs_direct_write_complete(dreq, inode);
 696
 697         if (started)
 698                 return 0;
 699         return result < 0 ? (ssize_t) result : -EFAULT;
 700 }
 701
 702 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 703 {
 704         ssize_t result = 0;
 705         sigset_t oldset;
 706         struct inode *inode = iocb->ki_filp->f_mapping->host;
 707         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 708         struct nfs_direct_req *dreq;
 709         size_t wsize = NFS_SERVER(inode)->wsize;
 710         int sync = 0;
 711
 712         dreq = nfs_direct_req_alloc();
 713         if (!dreq)
 714                 return -ENOMEM;
 715         nfs_alloc_commit_data(dreq);
 716
 717         if (dreq->commit_data == NULL || count < wsize)
 718                 sync = FLUSH_STABLE;
 719
 720         dreq->inode = inode;
 721         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 722         if (!is_sync_kiocb(iocb))
 723                 dreq->iocb = iocb;
 724
 725         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 726
 727         nfs_begin_data_update(inode);
 728
 729         rpc_clnt_sigmask(clnt, &oldset);
 730         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 731         if (!result)
 732                 result = nfs_direct_wait(dreq);
 733         rpc_clnt_sigunmask(clnt, &oldset);
 734         nfs_direct_req_release(dreq);
 735
 736         return result;
 737 }
 738
 739 /**
 740  * nfs_file_direct_read - file direct read operation for NFS files
 741  * @iocb: target I/O control block
 742  * @iov: vector of user buffers into which to read data
 743  * @nr_segs: size of iov vector
 744  * @pos: byte offset in file where reading starts
 745  *
 746  * We use this function for direct reads instead of calling
 747  * generic_file_aio_read() in order to avoid gfar's check to see if
 748  * the request starts before the end of the file.  For that check
 749  * to work, we must generate a GETATTR before each direct read, and
 750  * even then there is a window between the GETATTR and the subsequent
 751  * READ where the file size could change.  Our preference is simply
 752  * to do all reads the application wants, and the server will take
 753  * care of managing the end of file boundary.
 754  *
 755  * This function also eliminates unnecessarily updating the file's
 756  * atime locally, as the NFS server sets the file's atime, and this
 757  * client must read the updated atime from the server back into its
 758  * cache.
 759  */
 760 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 761                                 unsigned long nr_segs, loff_t pos)
 762 {
 763         ssize_t retval = -EINVAL;
 764         struct file *file = iocb->ki_filp;
 765         struct address_space *mapping = file->f_mapping;
 766         /* XXX: temporary */
 767         const char __user *buf = iov[0].iov_base;
 768         size_t count = iov[0].iov_len;
 769
 770         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 771                 file->f_path.dentry->d_parent->d_name.name,
 772                 file->f_path.dentry->d_name.name,
 773                 (unsigned long) count, (long long) pos);
 774
 775         if (nr_segs != 1)
 776                 goto out;
 777
 778         retval = -EFAULT;
 779         if (!access_ok(VERIFY_WRITE, buf, count))
 780                 goto out;
 781         retval = 0;
 782         if (!count)
 783                 goto out;
 784
 785         retval = nfs_sync_mapping(mapping);
 786         if (retval)
 787                 goto out;
 788
 789         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 790         if (retval > 0)
 791                 iocb->ki_pos = pos + retval;
 792
 793 out:
 794         return retval;
 795 }
 796
 797 /**
 798  * nfs_file_direct_write - file direct write operation for NFS files
 799  * @iocb: target I/O control block
 800  * @iov: vector of user buffers from which to write data
 801  * @nr_segs: size of iov vector
 802  * @pos: byte offset in file where writing starts
 803  *
 804  * We use this function for direct writes instead of calling
 805  * generic_file_aio_write() in order to avoid taking the inode
 806  * semaphore and updating the i_size.  The NFS server will set
 807  * the new i_size and this client must read the updated size
 808  * back into its cache.  We let the server do generic write
 809  * parameter checking and report problems.
 810  *
 811  * We also avoid an unnecessary invocation of generic_osync_inode(),
 812  * as it is fairly meaningless to sync the metadata of an NFS file.
 813  *
 814  * We eliminate local atime updates, see direct read above.
 815  *
 816  * We avoid unnecessary page cache invalidations for normal cached
 817  * readers of this file.
 818  *
 819  * Note that O_APPEND is not supported for NFS direct writes, as there
 820  * is no atomic O_APPEND write facility in the NFS protocol.
 821  */
 822 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 823                                 unsigned long nr_segs, loff_t pos)
 824 {
 825         ssize_t retval = -EINVAL;
 826         struct file *file = iocb->ki_filp;
 827         struct address_space *mapping = file->f_mapping;
 828         /* XXX: temporary */
 829         const char __user *buf = iov[0].iov_base;
 830         size_t count = iov[0].iov_len;
 831
 832         dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
 833                 file->f_path.dentry->d_parent->d_name.name,
 834                 file->f_path.dentry->d_name.name,
 835                 (unsigned long) count, (long long) pos);
 836
 837         if (nr_segs != 1)
 838                 goto out;
 839
 840         retval = generic_write_checks(file, &pos, &count, 0);
 841         if (retval)
 842                 goto out;
 843
 844         retval = -EINVAL;
 845         if ((ssize_t) count < 0)
 846                 goto out;
 847         retval = 0;
 848         if (!count)
 849                 goto out;
 850
 851         retval = -EFAULT;
 852         if (!access_ok(VERIFY_READ, buf, count))
 853                 goto out;
 854
 855         retval = nfs_sync_mapping(mapping);
 856         if (retval)
 857                 goto out;
 858
 859         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 860
 861         if (retval > 0)
 862                 iocb->ki_pos = pos + retval;
 863
 864 out:
 865         return retval;
 866 }
 867
 868 /**
 869  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 870  *
 871  */
 872 int __init nfs_init_directcache(void)
 873 {
 874         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 875                                                 sizeof(struct nfs_direct_req),
 876                                                 0, (SLAB_RECLAIM_ACCOUNT|
 877                                                         SLAB_MEM_SPREAD),
 878                                                 NULL);
 879         if (nfs_direct_cachep == NULL)
 880                 return -ENOMEM;
 881
 882         return 0;
 883 }
 884
 885 /**
 886  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 887  *
 888  */
 889 void nfs_destroy_directcache(void)
 890 {
 891         kmem_cache_destroy(nfs_direct_cachep);
 892 }