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/config.h>
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/smp_lock.h>
  46 #include <linux/file.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/kref.h>
  49
  50 #include <linux/nfs_fs.h>
  51 #include <linux/nfs_page.h>
  52 #include <linux/sunrpc/clnt.h>
  53
  54 #include <asm/system.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/atomic.h>
  57
  58 #include "iostat.h"
  59
  60 #define NFSDBG_FACILITY         NFSDBG_VFS
  61
  62 static kmem_cache_t *nfs_direct_cachep;
  63
  64 /*
  65  * This represents a set of asynchronous requests that we're waiting on
  66  */
  67 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct list_head        list,           /* nfs_read/write_data structs */
  72                                 rewrite_list;   /* saved nfs_write_data structs */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct kiocb *          iocb;           /* controlling i/o request */
  75         struct inode *          inode;          /* target file of i/o */
  76
  77         /* completion state */
  78         atomic_t                io_count;       /* i/os we're waiting for */
  79         spinlock_t              lock;           /* protect completion state */
  80         ssize_t                 count,          /* bytes actually processed */
  81                                 error;          /* any reported error */
  82         struct completion       completion;     /* wait for i/o completion */
  83
  84         /* commit state */
  85         struct nfs_write_data * commit_data;    /* special write_data for commits */
  86         int                     flags;
  87 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  88 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  89         struct nfs_writeverf    verf;           /* unstable write verifier */
  90 };
  91
  92 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  93 static const struct rpc_call_ops nfs_write_direct_ops;
  94
  95 static inline void get_dreq(struct nfs_direct_req *dreq)
  96 {
  97         atomic_inc(&dreq->io_count);
  98 }
  99
 100 static inline int put_dreq(struct nfs_direct_req *dreq)
 101 {
 102         return atomic_dec_and_test(&dreq->io_count);
 103 }
 104
 105 /*
 106  * "size" is never larger than rsize or wsize.
 107  */
 108 static inline int nfs_direct_count_pages(unsigned long user_addr, size_t size)
 109 {
 110         int page_count;
 111
 112         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 113         page_count -= user_addr >> PAGE_SHIFT;
 114         BUG_ON(page_count < 0);
 115
 116         return page_count;
 117 }
 118
 119 /**
 120  * nfs_direct_IO - NFS address space operation for direct I/O
 121  * @rw: direction (read or write)
 122  * @iocb: target I/O control block
 123  * @iov: array of vectors that define I/O buffer
 124  * @pos: offset in file to begin the operation
 125  * @nr_segs: size of iovec array
 126  *
 127  * The presence of this routine in the address space ops vector means
 128  * the NFS client supports direct I/O.  However, we shunt off direct
 129  * read and write requests before the VFS gets them, so this method
 130  * should never be called.
 131  */
 132 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 133 {
 134         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 135                         iocb->ki_filp->f_dentry->d_name.name,
 136                         (long long) pos, nr_segs);
 137
 138         return -EINVAL;
 139 }
 140
 141 static void nfs_direct_dirty_pages(struct page **pages, int npages)
 142 {
 143         int i;
 144         for (i = 0; i < npages; i++) {
 145                 struct page *page = pages[i];
 146                 if (!PageCompound(page))
 147                         set_page_dirty_lock(page);
 148         }
 149 }
 150
 151 static void nfs_direct_release_pages(struct page **pages, int npages)
 152 {
 153         int i;
 154         for (i = 0; i < npages; i++)
 155                 page_cache_release(pages[i]);
 156 }
 157
 158 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 159 {
 160         struct nfs_direct_req *dreq;
 161
 162         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 163         if (!dreq)
 164                 return NULL;
 165
 166         kref_init(&dreq->kref);
 167         init_completion(&dreq->completion);
 168         INIT_LIST_HEAD(&dreq->list);
 169         INIT_LIST_HEAD(&dreq->rewrite_list);
 170         dreq->iocb = NULL;
 171         dreq->ctx = NULL;
 172         spin_lock_init(&dreq->lock);
 173         atomic_set(&dreq->io_count, 0);
 174         dreq->count = 0;
 175         dreq->error = 0;
 176         dreq->flags = 0;
 177
 178         return dreq;
 179 }
 180
 181 static void nfs_direct_req_release(struct kref *kref)
 182 {
 183         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 184
 185         if (dreq->ctx != NULL)
 186                 put_nfs_open_context(dreq->ctx);
 187         kmem_cache_free(nfs_direct_cachep, dreq);
 188 }
 189
 190 /*
 191  * Collects and returns the final error value/byte-count.
 192  */
 193 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 194 {
 195         ssize_t result = -EIOCBQUEUED;
 196
 197         /* Async requests don't wait here */
 198         if (dreq->iocb)
 199                 goto out;
 200
 201         result = wait_for_completion_interruptible(&dreq->completion);
 202
 203         if (!result)
 204                 result = dreq->error;
 205         if (!result)
 206                 result = dreq->count;
 207
 208 out:
 209         kref_put(&dreq->kref, nfs_direct_req_release);
 210         return (ssize_t) result;
 211 }
 212
 213 /*
 214  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 215  * the iocb is still valid here if this is a synchronous request.
 216  */
 217 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 218 {
 219         if (dreq->iocb) {
 220                 long res = (long) dreq->error;
 221                 if (!res)
 222                         res = (long) dreq->count;
 223                 aio_complete(dreq->iocb, res, 0);
 224         }
 225         complete_all(&dreq->completion);
 226
 227         kref_put(&dreq->kref, nfs_direct_req_release);
 228 }
 229
 230 /*
 231  * Note we also set the number of requests we have in the dreq when we are
 232  * done.  This prevents races with I/O completion so we will always wait
 233  * until all requests have been dispatched and completed.
 234  */
 235 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 236 {
 237         struct list_head *list;
 238         struct nfs_direct_req *dreq;
 239         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 240
 241         dreq = nfs_direct_req_alloc();
 242         if (!dreq)
 243                 return NULL;
 244
 245         list = &dreq->list;
 246         for(;;) {
 247                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 248
 249                 if (unlikely(!data)) {
 250                         while (!list_empty(list)) {
 251                                 data = list_entry(list->next,
 252                                                   struct nfs_read_data, pages);
 253                                 list_del(&data->pages);
 254                                 nfs_readdata_free(data);
 255                         }
 256                         kref_put(&dreq->kref, nfs_direct_req_release);
 257                         return NULL;
 258                 }
 259
 260                 INIT_LIST_HEAD(&data->pages);
 261                 list_add(&data->pages, list);
 262
 263                 data->req = (struct nfs_page *) dreq;
 264                 get_dreq(dreq);
 265                 if (nbytes <= rsize)
 266                         break;
 267                 nbytes -= rsize;
 268         }
 269         kref_get(&dreq->kref);
 270         return dreq;
 271 }
 272
 273 /*
 274  * We must hold a reference to all the pages in this direct read request
 275  * until the RPCs complete.  This could be long *after* we are woken up in
 276  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 277  */
 278 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 279 {
 280         struct nfs_read_data *data = calldata;
 281         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 282
 283         if (nfs_readpage_result(task, data) != 0)
 284                 return;
 285
 286         nfs_direct_dirty_pages(data->pagevec, data->npages);
 287         nfs_direct_release_pages(data->pagevec, data->npages);
 288
 289         spin_lock(&dreq->lock);
 290
 291         if (likely(task->tk_status >= 0))
 292                 dreq->count += data->res.count;
 293         else
 294                 dreq->error = task->tk_status;
 295
 296         spin_unlock(&dreq->lock);
 297
 298         if (put_dreq(dreq))
 299                 nfs_direct_complete(dreq);
 300 }
 301
 302 static const struct rpc_call_ops nfs_read_direct_ops = {
 303         .rpc_call_done = nfs_direct_read_result,
 304         .rpc_release = nfs_readdata_release,
 305 };
 306
 307 /*
 308  * For each nfs_read_data struct that was allocated on the list, dispatch
 309  * an NFS READ operation.  If get_user_pages() fails, we stop sending reads.
 310  * Read length accounting is handled by nfs_direct_read_result().
 311  * Otherwise, if no requests have been sent, just return an error.
 312  */
 313 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 314 {
 315         struct nfs_open_context *ctx = dreq->ctx;
 316         struct inode *inode = ctx->dentry->d_inode;
 317         struct list_head *list = &dreq->list;
 318         size_t rsize = NFS_SERVER(inode)->rsize;
 319         unsigned int pgbase;
 320         int result;
 321         ssize_t started = 0;
 322         struct nfs_read_data *data;
 323
 324         pgbase = user_addr & ~PAGE_MASK;
 325         do {
 326                 size_t bytes;
 327
 328                 bytes = rsize;
 329                 if (count < rsize)
 330                         bytes = count;
 331
 332                 BUG_ON(list_empty(list));
 333                 data = list_entry(list->next, struct nfs_read_data, pages);
 334                 list_del_init(&data->pages);
 335
 336                 data->npages = nfs_direct_count_pages(user_addr, bytes);
 337                 down_read(&current->mm->mmap_sem);
 338                 result = get_user_pages(current, current->mm, user_addr,
 339                                         data->npages, 1, 0, data->pagevec, NULL);
 340                 up_read(&current->mm->mmap_sem);
 341                 if (unlikely(result < data->npages))
 342                         goto out_err;
 343
 344                 data->inode = inode;
 345                 data->cred = ctx->cred;
 346                 data->args.fh = NFS_FH(inode);
 347                 data->args.context = ctx;
 348                 data->args.offset = pos;
 349                 data->args.pgbase = pgbase;
 350                 data->args.pages = data->pagevec;
 351                 data->args.count = bytes;
 352                 data->res.fattr = &data->fattr;
 353                 data->res.eof = 0;
 354                 data->res.count = bytes;
 355
 356                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 357                                 &nfs_read_direct_ops, data);
 358                 NFS_PROTO(inode)->read_setup(data);
 359
 360                 data->task.tk_cookie = (unsigned long) inode;
 361
 362                 lock_kernel();
 363                 rpc_execute(&data->task);
 364                 unlock_kernel();
 365
 366                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 367                                 data->task.tk_pid,
 368                                 inode->i_sb->s_id,
 369                                 (long long)NFS_FILEID(inode),
 370                                 bytes,
 371                                 (unsigned long long)data->args.offset);
 372
 373                 started += bytes;
 374                 user_addr += bytes;
 375                 pos += bytes;
 376                 pgbase += bytes;
 377                 pgbase &= ~PAGE_MASK;
 378
 379                 count -= bytes;
 380         } while (count != 0);
 381         BUG_ON(!list_empty(list));
 382         return 0;
 383
 384 out_err:
 385         if (result > 0)
 386                 nfs_direct_release_pages(data->pagevec, result);
 387
 388         list_add(&data->pages, list);
 389         while (!list_empty(list)) {
 390                 data = list_entry(list->next, struct nfs_read_data, pages);
 391                 list_del(&data->pages);
 392                 nfs_readdata_free(data);
 393                 if (put_dreq(dreq))
 394                         nfs_direct_complete(dreq);
 395         }
 396
 397         if (started)
 398                 return 0;
 399         return result < 0 ? (ssize_t) result : -EFAULT;
 400 }
 401
 402 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 403 {
 404         ssize_t result;
 405         sigset_t oldset;
 406         struct inode *inode = iocb->ki_filp->f_mapping->host;
 407         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 408         struct nfs_direct_req *dreq;
 409
 410         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 411         if (!dreq)
 412                 return -ENOMEM;
 413
 414         dreq->inode = inode;
 415         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 416         if (!is_sync_kiocb(iocb))
 417                 dreq->iocb = iocb;
 418
 419         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 420         rpc_clnt_sigmask(clnt, &oldset);
 421         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 422         if (!result)
 423                 result = nfs_direct_wait(dreq);
 424         rpc_clnt_sigunmask(clnt, &oldset);
 425
 426         return result;
 427 }
 428
 429 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 430 {
 431         list_splice_init(&dreq->rewrite_list, &dreq->list);
 432         while (!list_empty(&dreq->list)) {
 433                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 434                 list_del(&data->pages);
 435                 nfs_direct_release_pages(data->pagevec, data->npages);
 436                 nfs_writedata_release(data);
 437         }
 438 }
 439
 440 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 441 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 442 {
 443         struct inode *inode = dreq->inode;
 444         struct list_head *p;
 445         struct nfs_write_data *data;
 446
 447         dreq->count = 0;
 448         get_dreq(dreq);
 449
 450         list_for_each(p, &dreq->rewrite_list) {
 451                 data = list_entry(p, struct nfs_write_data, pages);
 452
 453                 get_dreq(dreq);
 454
 455                 /*
 456                  * Reset data->res.
 457                  */
 458                 nfs_fattr_init(&data->fattr);
 459                 data->res.count = data->args.count;
 460                 memset(&data->verf, 0, sizeof(data->verf));
 461
 462                 /*
 463                  * Reuse data->task; data->args should not have changed
 464                  * since the original request was sent.
 465                  */
 466                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 467                                 &nfs_write_direct_ops, data);
 468                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 469
 470                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 471                 data->task.tk_cookie = (unsigned long) inode;
 472
 473                 /*
 474                  * We're called via an RPC callback, so BKL is already held.
 475                  */
 476                 rpc_execute(&data->task);
 477
 478                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 479                                 data->task.tk_pid,
 480                                 inode->i_sb->s_id,
 481                                 (long long)NFS_FILEID(inode),
 482                                 data->args.count,
 483                                 (unsigned long long)data->args.offset);
 484         }
 485
 486         if (put_dreq(dreq))
 487                 nfs_direct_write_complete(dreq, inode);
 488 }
 489
 490 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 491 {
 492         struct nfs_write_data *data = calldata;
 493         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 494
 495         /* Call the NFS version-specific code */
 496         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 497                 return;
 498         if (unlikely(task->tk_status < 0)) {
 499                 dreq->error = task->tk_status;
 500                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 501         }
 502         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 503                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 504                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 505         }
 506
 507         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 508         nfs_direct_write_complete(dreq, data->inode);
 509 }
 510
 511 static const struct rpc_call_ops nfs_commit_direct_ops = {
 512         .rpc_call_done = nfs_direct_commit_result,
 513         .rpc_release = nfs_commit_release,
 514 };
 515
 516 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 517 {
 518         struct nfs_write_data *data = dreq->commit_data;
 519
 520         data->inode = dreq->inode;
 521         data->cred = dreq->ctx->cred;
 522
 523         data->args.fh = NFS_FH(data->inode);
 524         data->args.offset = 0;
 525         data->args.count = 0;
 526         data->res.count = 0;
 527         data->res.fattr = &data->fattr;
 528         data->res.verf = &data->verf;
 529
 530         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 531                                 &nfs_commit_direct_ops, data);
 532         NFS_PROTO(data->inode)->commit_setup(data, 0);
 533
 534         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 535         data->task.tk_cookie = (unsigned long)data->inode;
 536         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 537         dreq->commit_data = NULL;
 538
 539         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 540
 541         lock_kernel();
 542         rpc_execute(&data->task);
 543         unlock_kernel();
 544 }
 545
 546 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 547 {
 548         int flags = dreq->flags;
 549
 550         dreq->flags = 0;
 551         switch (flags) {
 552                 case NFS_ODIRECT_DO_COMMIT:
 553                         nfs_direct_commit_schedule(dreq);
 554                         break;
 555                 case NFS_ODIRECT_RESCHED_WRITES:
 556                         nfs_direct_write_reschedule(dreq);
 557                         break;
 558                 default:
 559                         nfs_end_data_update(inode);
 560                         if (dreq->commit_data != NULL)
 561                                 nfs_commit_free(dreq->commit_data);
 562                         nfs_direct_free_writedata(dreq);
 563                         nfs_direct_complete(dreq);
 564         }
 565 }
 566
 567 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 568 {
 569         dreq->commit_data = nfs_commit_alloc(0);
 570         if (dreq->commit_data != NULL)
 571                 dreq->commit_data->req = (struct nfs_page *) dreq;
 572 }
 573 #else
 574 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 575 {
 576         dreq->commit_data = NULL;
 577 }
 578
 579 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 580 {
 581         nfs_end_data_update(inode);
 582         nfs_direct_free_writedata(dreq);
 583         nfs_direct_complete(dreq);
 584 }
 585 #endif
 586
 587 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 588 {
 589         struct list_head *list;
 590         struct nfs_direct_req *dreq;
 591         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 592
 593         dreq = nfs_direct_req_alloc();
 594         if (!dreq)
 595                 return NULL;
 596
 597         list = &dreq->list;
 598         for(;;) {
 599                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 600
 601                 if (unlikely(!data)) {
 602                         while (!list_empty(list)) {
 603                                 data = list_entry(list->next,
 604                                                   struct nfs_write_data, pages);
 605                                 list_del(&data->pages);
 606                                 nfs_writedata_free(data);
 607                         }
 608                         kref_put(&dreq->kref, nfs_direct_req_release);
 609                         return NULL;
 610                 }
 611
 612                 INIT_LIST_HEAD(&data->pages);
 613                 list_add(&data->pages, list);
 614
 615                 data->req = (struct nfs_page *) dreq;
 616                 get_dreq(dreq);
 617                 if (nbytes <= wsize)
 618                         break;
 619                 nbytes -= wsize;
 620         }
 621
 622         nfs_alloc_commit_data(dreq);
 623
 624         kref_get(&dreq->kref);
 625         return dreq;
 626 }
 627
 628 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 629 {
 630         struct nfs_write_data *data = calldata;
 631         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 632         int status = task->tk_status;
 633
 634         if (nfs_writeback_done(task, data) != 0)
 635                 return;
 636
 637         spin_lock(&dreq->lock);
 638
 639         if (likely(status >= 0))
 640                 dreq->count += data->res.count;
 641         else
 642                 dreq->error = task->tk_status;
 643
 644         if (data->res.verf->committed != NFS_FILE_SYNC) {
 645                 switch (dreq->flags) {
 646                         case 0:
 647                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 648                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 649                                 break;
 650                         case NFS_ODIRECT_DO_COMMIT:
 651                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 652                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 653                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 654                                 }
 655                 }
 656         }
 657
 658         spin_unlock(&dreq->lock);
 659 }
 660
 661 /*
 662  * NB: Return the value of the first error return code.  Subsequent
 663  *     errors after the first one are ignored.
 664  */
 665 static void nfs_direct_write_release(void *calldata)
 666 {
 667         struct nfs_write_data *data = calldata;
 668         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 669
 670         if (put_dreq(dreq))
 671                 nfs_direct_write_complete(dreq, data->inode);
 672 }
 673
 674 static const struct rpc_call_ops nfs_write_direct_ops = {
 675         .rpc_call_done = nfs_direct_write_result,
 676         .rpc_release = nfs_direct_write_release,
 677 };
 678
 679 /*
 680  * For each nfs_write_data struct that was allocated on the list, dispatch
 681  * an NFS WRITE operation.  If get_user_pages() fails, we stop sending writes.
 682  * Write length accounting is handled by nfs_direct_write_result().
 683  * Otherwise, if no requests have been sent, just return an error.
 684  */
 685 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 686 {
 687         struct nfs_open_context *ctx = dreq->ctx;
 688         struct inode *inode = ctx->dentry->d_inode;
 689         struct list_head *list = &dreq->list;
 690         size_t wsize = NFS_SERVER(inode)->wsize;
 691         unsigned int pgbase;
 692         int result;
 693         ssize_t started = 0;
 694         struct nfs_write_data *data;
 695
 696         pgbase = user_addr & ~PAGE_MASK;
 697         do {
 698                 size_t bytes;
 699
 700                 bytes = wsize;
 701                 if (count < wsize)
 702                         bytes = count;
 703
 704                 BUG_ON(list_empty(list));
 705                 data = list_entry(list->next, struct nfs_write_data, pages);
 706
 707                 data->npages = nfs_direct_count_pages(user_addr, bytes);
 708                 down_read(&current->mm->mmap_sem);
 709                 result = get_user_pages(current, current->mm, user_addr,
 710                                         data->npages, 0, 0, data->pagevec, NULL);
 711                 up_read(&current->mm->mmap_sem);
 712                 if (unlikely(result < data->npages))
 713                         goto out_err;
 714
 715                 list_move_tail(&data->pages, &dreq->rewrite_list);
 716
 717                 data->inode = inode;
 718                 data->cred = ctx->cred;
 719                 data->args.fh = NFS_FH(inode);
 720                 data->args.context = ctx;
 721                 data->args.offset = pos;
 722                 data->args.pgbase = pgbase;
 723                 data->args.pages = data->pagevec;
 724                 data->args.count = bytes;
 725                 data->res.fattr = &data->fattr;
 726                 data->res.count = bytes;
 727                 data->res.verf = &data->verf;
 728
 729                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 730                                 &nfs_write_direct_ops, data);
 731                 NFS_PROTO(inode)->write_setup(data, sync);
 732
 733                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 734                 data->task.tk_cookie = (unsigned long) inode;
 735
 736                 lock_kernel();
 737                 rpc_execute(&data->task);
 738                 unlock_kernel();
 739
 740                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 741                                 data->task.tk_pid,
 742                                 inode->i_sb->s_id,
 743                                 (long long)NFS_FILEID(inode),
 744                                 bytes,
 745                                 (unsigned long long)data->args.offset);
 746
 747                 started += bytes;
 748                 user_addr += bytes;
 749                 pos += bytes;
 750                 pgbase += bytes;
 751                 pgbase &= ~PAGE_MASK;
 752
 753                 count -= bytes;
 754         } while (count != 0);
 755         BUG_ON(!list_empty(list));
 756         return 0;
 757
 758 out_err:
 759         if (result > 0)
 760                 nfs_direct_release_pages(data->pagevec, result);
 761
 762         list_add(&data->pages, list);
 763         while (!list_empty(list)) {
 764                 data = list_entry(list->next, struct nfs_write_data, pages);
 765                 list_del(&data->pages);
 766                 nfs_writedata_free(data);
 767                 if (put_dreq(dreq))
 768                         nfs_direct_write_complete(dreq, inode);
 769         }
 770
 771         if (started)
 772                 return 0;
 773         return result < 0 ? (ssize_t) result : -EFAULT;
 774 }
 775
 776 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 777 {
 778         ssize_t result;
 779         sigset_t oldset;
 780         struct inode *inode = iocb->ki_filp->f_mapping->host;
 781         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 782         struct nfs_direct_req *dreq;
 783         size_t wsize = NFS_SERVER(inode)->wsize;
 784         int sync = 0;
 785
 786         dreq = nfs_direct_write_alloc(count, wsize);
 787         if (!dreq)
 788                 return -ENOMEM;
 789         if (dreq->commit_data == NULL || count < wsize)
 790                 sync = FLUSH_STABLE;
 791
 792         dreq->inode = inode;
 793         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 794         if (!is_sync_kiocb(iocb))
 795                 dreq->iocb = iocb;
 796
 797         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 798
 799         nfs_begin_data_update(inode);
 800
 801         rpc_clnt_sigmask(clnt, &oldset);
 802         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 803         if (!result)
 804                 result = nfs_direct_wait(dreq);
 805         rpc_clnt_sigunmask(clnt, &oldset);
 806
 807         return result;
 808 }
 809
 810 /**
 811  * nfs_file_direct_read - file direct read operation for NFS files
 812  * @iocb: target I/O control block
 813  * @buf: user's buffer into which to read data
 814  * @count: number of bytes to read
 815  * @pos: byte offset in file where reading starts
 816  *
 817  * We use this function for direct reads instead of calling
 818  * generic_file_aio_read() in order to avoid gfar's check to see if
 819  * the request starts before the end of the file.  For that check
 820  * to work, we must generate a GETATTR before each direct read, and
 821  * even then there is a window between the GETATTR and the subsequent
 822  * READ where the file size could change.  Our preference is simply
 823  * to do all reads the application wants, and the server will take
 824  * care of managing the end of file boundary.
 825  *
 826  * This function also eliminates unnecessarily updating the file's
 827  * atime locally, as the NFS server sets the file's atime, and this
 828  * client must read the updated atime from the server back into its
 829  * cache.
 830  */
 831 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 832 {
 833         ssize_t retval = -EINVAL;
 834         struct file *file = iocb->ki_filp;
 835         struct address_space *mapping = file->f_mapping;
 836
 837         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 838                 file->f_dentry->d_parent->d_name.name,
 839                 file->f_dentry->d_name.name,
 840                 (unsigned long) count, (long long) pos);
 841
 842         if (count < 0)
 843                 goto out;
 844         retval = -EFAULT;
 845         if (!access_ok(VERIFY_WRITE, buf, count))
 846                 goto out;
 847         retval = 0;
 848         if (!count)
 849                 goto out;
 850
 851         retval = nfs_sync_mapping(mapping);
 852         if (retval)
 853                 goto out;
 854
 855         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 856         if (retval > 0)
 857                 iocb->ki_pos = pos + retval;
 858
 859 out:
 860         return retval;
 861 }
 862
 863 /**
 864  * nfs_file_direct_write - file direct write operation for NFS files
 865  * @iocb: target I/O control block
 866  * @buf: user's buffer from which to write data
 867  * @count: number of bytes to write
 868  * @pos: byte offset in file where writing starts
 869  *
 870  * We use this function for direct writes instead of calling
 871  * generic_file_aio_write() in order to avoid taking the inode
 872  * semaphore and updating the i_size.  The NFS server will set
 873  * the new i_size and this client must read the updated size
 874  * back into its cache.  We let the server do generic write
 875  * parameter checking and report problems.
 876  *
 877  * We also avoid an unnecessary invocation of generic_osync_inode(),
 878  * as it is fairly meaningless to sync the metadata of an NFS file.
 879  *
 880  * We eliminate local atime updates, see direct read above.
 881  *
 882  * We avoid unnecessary page cache invalidations for normal cached
 883  * readers of this file.
 884  *
 885  * Note that O_APPEND is not supported for NFS direct writes, as there
 886  * is no atomic O_APPEND write facility in the NFS protocol.
 887  */
 888 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 889 {
 890         ssize_t retval;
 891         struct file *file = iocb->ki_filp;
 892         struct address_space *mapping = file->f_mapping;
 893
 894         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 895                 file->f_dentry->d_parent->d_name.name,
 896                 file->f_dentry->d_name.name,
 897                 (unsigned long) count, (long long) pos);
 898
 899         retval = generic_write_checks(file, &pos, &count, 0);
 900         if (retval)
 901                 goto out;
 902
 903         retval = -EINVAL;
 904         if ((ssize_t) count < 0)
 905                 goto out;
 906         retval = 0;
 907         if (!count)
 908                 goto out;
 909
 910         retval = -EFAULT;
 911         if (!access_ok(VERIFY_READ, buf, count))
 912                 goto out;
 913
 914         retval = nfs_sync_mapping(mapping);
 915         if (retval)
 916                 goto out;
 917
 918         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 919
 920         /*
 921          * XXX: nfs_end_data_update() already ensures this file's
 922          *      cached data is subsequently invalidated.  Do we really
 923          *      need to call invalidate_inode_pages2() again here?
 924          *
 925          *      For aio writes, this invalidation will almost certainly
 926          *      occur before the writes complete.  Kind of racey.
 927          */
 928         if (mapping->nrpages)
 929                 invalidate_inode_pages2(mapping);
 930
 931         if (retval > 0)
 932                 iocb->ki_pos = pos + retval;
 933
 934 out:
 935         return retval;
 936 }
 937
 938 /**
 939  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 940  *
 941  */
 942 int __init nfs_init_directcache(void)
 943 {
 944         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 945                                                 sizeof(struct nfs_direct_req),
 946                                                 0, (SLAB_RECLAIM_ACCOUNT|
 947                                                         SLAB_MEM_SPREAD),
 948                                                 NULL, NULL);
 949         if (nfs_direct_cachep == NULL)
 950                 return -ENOMEM;
 951
 952         return 0;
 953 }
 954
 955 /**
 956  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 957  *
 958  */
 959 void __exit nfs_destroy_directcache(void)
 960 {
 961         if (kmem_cache_destroy(nfs_direct_cachep))
 962                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 963 }