fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/smp_lock.h>
  45 #include <linux/file.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/kref.h>
  48
  49 #include <linux/nfs_fs.h>
  50 #include <linux/nfs_page.h>
  51 #include <linux/sunrpc/clnt.h>
  52
  53 #include <asm/system.h>
  54 #include <asm/uaccess.h>
  55 #include <asm/atomic.h>
  56
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60
  61 static kmem_cache_t *nfs_direct_cachep;
  62
  63 /*
  64  * This represents a set of asynchronous requests that we're waiting on
  65  */
  66 struct nfs_direct_req {
  67         struct kref             kref;           /* release manager */
  68
  69         /* I/O parameters */
  70         struct list_head        list,           /* nfs_read/write_data structs */
  71                                 rewrite_list;   /* saved nfs_write_data structs */
  72         struct nfs_open_context *ctx;           /* file open context info */
  73         struct kiocb *          iocb;           /* controlling i/o request */
  74         struct inode *          inode;          /* target file of i/o */
  75         unsigned long           user_addr;      /* location of user's buffer */
  76         size_t                  user_count;     /* total bytes to move */
  77         loff_t                  pos;            /* starting offset in file */
  78         struct page **          pages;          /* pages in our buffer */
  79         unsigned int            npages;         /* count of pages */
  80
  81         /* completion state */
  82         spinlock_t              lock;           /* protect completion state */
  83         int                     outstanding;    /* i/os we're waiting for */
  84         ssize_t                 count,          /* bytes actually processed */
  85                                 error;          /* any reported error */
  86         struct completion       completion;     /* wait for i/o completion */
  87
  88         /* commit state */
  89         struct nfs_write_data * commit_data;    /* special write_data for commits */
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93         struct nfs_writeverf    verf;           /* unstable write verifier */
  94 };
  95
  96 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync);
  97 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  98
  99 /**
 100  * nfs_direct_IO - NFS address space operation for direct I/O
 101  * @rw: direction (read or write)
 102  * @iocb: target I/O control block
 103  * @iov: array of vectors that define I/O buffer
 104  * @pos: offset in file to begin the operation
 105  * @nr_segs: size of iovec array
 106  *
 107  * The presence of this routine in the address space ops vector means
 108  * the NFS client supports direct I/O.  However, we shunt off direct
 109  * read and write requests before the VFS gets them, so this method
 110  * should never be called.
 111  */
 112 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 113 {
 114         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 115                         iocb->ki_filp->f_dentry->d_name.name,
 116                         (long long) pos, nr_segs);
 117
 118         return -EINVAL;
 119 }
 120
 121 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 122 {
 123         int i;
 124         for (i = 0; i < npages; i++) {
 125                 struct page *page = pages[i];
 126                 if (do_dirty && !PageCompound(page))
 127                         set_page_dirty_lock(page);
 128                 page_cache_release(page);
 129         }
 130         kfree(pages);
 131 }
 132
 133 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 134 {
 135         int result = -ENOMEM;
 136         unsigned long page_count;
 137         size_t array_size;
 138
 139         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 140         page_count -= user_addr >> PAGE_SHIFT;
 141
 142         array_size = (page_count * sizeof(struct page *));
 143         *pages = kmalloc(array_size, GFP_KERNEL);
 144         if (*pages) {
 145                 down_read(&current->mm->mmap_sem);
 146                 result = get_user_pages(current, current->mm, user_addr,
 147                                         page_count, (rw == READ), 0,
 148                                         *pages, NULL);
 149                 up_read(&current->mm->mmap_sem);
 150                 if (result != page_count) {
 151                         /*
 152                          * If we got fewer pages than expected from
 153                          * get_user_pages(), the user buffer runs off the
 154                          * end of a mapping; return EFAULT.
 155                          */
 156                         if (result >= 0) {
 157                                 nfs_free_user_pages(*pages, result, 0);
 158                                 result = -EFAULT;
 159                         } else
 160                                 kfree(*pages);
 161                         *pages = NULL;
 162                 }
 163         }
 164         return result;
 165 }
 166
 167 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 168 {
 169         struct nfs_direct_req *dreq;
 170
 171         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 172         if (!dreq)
 173                 return NULL;
 174
 175         kref_init(&dreq->kref);
 176         init_completion(&dreq->completion);
 177         INIT_LIST_HEAD(&dreq->list);
 178         INIT_LIST_HEAD(&dreq->rewrite_list);
 179         dreq->iocb = NULL;
 180         dreq->ctx = NULL;
 181         spin_lock_init(&dreq->lock);
 182         dreq->outstanding = 0;
 183         dreq->count = 0;
 184         dreq->error = 0;
 185         dreq->flags = 0;
 186
 187         return dreq;
 188 }
 189
 190 static void nfs_direct_req_release(struct kref *kref)
 191 {
 192         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 193
 194         if (dreq->ctx != NULL)
 195                 put_nfs_open_context(dreq->ctx);
 196         kmem_cache_free(nfs_direct_cachep, dreq);
 197 }
 198
 199 /*
 200  * Collects and returns the final error value/byte-count.
 201  */
 202 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 203 {
 204         ssize_t result = -EIOCBQUEUED;
 205
 206         /* Async requests don't wait here */
 207         if (dreq->iocb)
 208                 goto out;
 209
 210         result = wait_for_completion_interruptible(&dreq->completion);
 211
 212         if (!result)
 213                 result = dreq->error;
 214         if (!result)
 215                 result = dreq->count;
 216
 217 out:
 218         kref_put(&dreq->kref, nfs_direct_req_release);
 219         return (ssize_t) result;
 220 }
 221
 222 /*
 223  * We must hold a reference to all the pages in this direct read request
 224  * until the RPCs complete.  This could be long *after* we are woken up in
 225  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 226  *
 227  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 228  * can't trust the iocb is still valid here if this is a synchronous
 229  * request.  If the waiter is woken prematurely, the iocb is long gone.
 230  */
 231 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 232 {
 233         nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 234
 235         if (dreq->iocb) {
 236                 long res = (long) dreq->error;
 237                 if (!res)
 238                         res = (long) dreq->count;
 239                 aio_complete(dreq->iocb, res, 0);
 240         }
 241         complete_all(&dreq->completion);
 242
 243         kref_put(&dreq->kref, nfs_direct_req_release);
 244 }
 245
 246 /*
 247  * Note we also set the number of requests we have in the dreq when we are
 248  * done.  This prevents races with I/O completion so we will always wait
 249  * until all requests have been dispatched and completed.
 250  */
 251 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 252 {
 253         struct list_head *list;
 254         struct nfs_direct_req *dreq;
 255         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 256
 257         dreq = nfs_direct_req_alloc();
 258         if (!dreq)
 259                 return NULL;
 260
 261         list = &dreq->list;
 262         for(;;) {
 263                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 264
 265                 if (unlikely(!data)) {
 266                         while (!list_empty(list)) {
 267                                 data = list_entry(list->next,
 268                                                   struct nfs_read_data, pages);
 269                                 list_del(&data->pages);
 270                                 nfs_readdata_free(data);
 271                         }
 272                         kref_put(&dreq->kref, nfs_direct_req_release);
 273                         return NULL;
 274                 }
 275
 276                 INIT_LIST_HEAD(&data->pages);
 277                 list_add(&data->pages, list);
 278
 279                 data->req = (struct nfs_page *) dreq;
 280                 dreq->outstanding++;
 281                 if (nbytes <= rsize)
 282                         break;
 283                 nbytes -= rsize;
 284         }
 285         kref_get(&dreq->kref);
 286         return dreq;
 287 }
 288
 289 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 290 {
 291         struct nfs_read_data *data = calldata;
 292         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 293
 294         if (nfs_readpage_result(task, data) != 0)
 295                 return;
 296
 297         spin_lock(&dreq->lock);
 298
 299         if (likely(task->tk_status >= 0))
 300                 dreq->count += data->res.count;
 301         else
 302                 dreq->error = task->tk_status;
 303
 304         if (--dreq->outstanding) {
 305                 spin_unlock(&dreq->lock);
 306                 return;
 307         }
 308
 309         spin_unlock(&dreq->lock);
 310         nfs_direct_complete(dreq);
 311 }
 312
 313 static const struct rpc_call_ops nfs_read_direct_ops = {
 314         .rpc_call_done = nfs_direct_read_result,
 315         .rpc_release = nfs_readdata_release,
 316 };
 317
 318 /*
 319  * For each nfs_read_data struct that was allocated on the list, dispatch
 320  * an NFS READ operation
 321  */
 322 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq)
 323 {
 324         struct nfs_open_context *ctx = dreq->ctx;
 325         struct inode *inode = ctx->dentry->d_inode;
 326         struct list_head *list = &dreq->list;
 327         struct page **pages = dreq->pages;
 328         size_t count = dreq->user_count;
 329         loff_t pos = dreq->pos;
 330         size_t rsize = NFS_SERVER(inode)->rsize;
 331         unsigned int curpage, pgbase;
 332
 333         curpage = 0;
 334         pgbase = dreq->user_addr & ~PAGE_MASK;
 335         do {
 336                 struct nfs_read_data *data;
 337                 size_t bytes;
 338
 339                 bytes = rsize;
 340                 if (count < rsize)
 341                         bytes = count;
 342
 343                 BUG_ON(list_empty(list));
 344                 data = list_entry(list->next, struct nfs_read_data, pages);
 345                 list_del_init(&data->pages);
 346
 347                 data->inode = inode;
 348                 data->cred = ctx->cred;
 349                 data->args.fh = NFS_FH(inode);
 350                 data->args.context = ctx;
 351                 data->args.offset = pos;
 352                 data->args.pgbase = pgbase;
 353                 data->args.pages = &pages[curpage];
 354                 data->args.count = bytes;
 355                 data->res.fattr = &data->fattr;
 356                 data->res.eof = 0;
 357                 data->res.count = bytes;
 358
 359                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 360                                 &nfs_read_direct_ops, data);
 361                 NFS_PROTO(inode)->read_setup(data);
 362
 363                 data->task.tk_cookie = (unsigned long) inode;
 364
 365                 lock_kernel();
 366                 rpc_execute(&data->task);
 367                 unlock_kernel();
 368
 369                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 370                                 data->task.tk_pid,
 371                                 inode->i_sb->s_id,
 372                                 (long long)NFS_FILEID(inode),
 373                                 bytes,
 374                                 (unsigned long long)data->args.offset);
 375
 376                 pos += bytes;
 377                 pgbase += bytes;
 378                 curpage += pgbase >> PAGE_SHIFT;
 379                 pgbase &= ~PAGE_MASK;
 380
 381                 count -= bytes;
 382         } while (count != 0);
 383         BUG_ON(!list_empty(list));
 384 }
 385
 386 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, unsigned int nr_pages)
 387 {
 388         ssize_t result;
 389         sigset_t oldset;
 390         struct inode *inode = iocb->ki_filp->f_mapping->host;
 391         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 392         struct nfs_direct_req *dreq;
 393
 394         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 395         if (!dreq)
 396                 return -ENOMEM;
 397
 398         dreq->user_addr = user_addr;
 399         dreq->user_count = count;
 400         dreq->pos = pos;
 401         dreq->pages = pages;
 402         dreq->npages = nr_pages;
 403         dreq->inode = inode;
 404         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 405         if (!is_sync_kiocb(iocb))
 406                 dreq->iocb = iocb;
 407
 408         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 409         rpc_clnt_sigmask(clnt, &oldset);
 410         nfs_direct_read_schedule(dreq);
 411         result = nfs_direct_wait(dreq);
 412         rpc_clnt_sigunmask(clnt, &oldset);
 413
 414         return result;
 415 }
 416
 417 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 418 {
 419         list_splice_init(&dreq->rewrite_list, &dreq->list);
 420         while (!list_empty(&dreq->list)) {
 421                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 422                 list_del(&data->pages);
 423                 nfs_writedata_release(data);
 424         }
 425 }
 426
 427 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 428 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 429 {
 430         struct list_head *pos;
 431
 432         list_splice_init(&dreq->rewrite_list, &dreq->list);
 433         list_for_each(pos, &dreq->list)
 434                 dreq->outstanding++;
 435         dreq->count = 0;
 436
 437         nfs_direct_write_schedule(dreq, FLUSH_STABLE);
 438 }
 439
 440 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 441 {
 442         struct nfs_write_data *data = calldata;
 443         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 444
 445         /* Call the NFS version-specific code */
 446         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 447                 return;
 448         if (unlikely(task->tk_status < 0)) {
 449                 dreq->error = task->tk_status;
 450                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 451         }
 452         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 453                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 454                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 455         }
 456
 457         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 458         nfs_direct_write_complete(dreq, data->inode);
 459 }
 460
 461 static const struct rpc_call_ops nfs_commit_direct_ops = {
 462         .rpc_call_done = nfs_direct_commit_result,
 463         .rpc_release = nfs_commit_release,
 464 };
 465
 466 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 467 {
 468         struct nfs_write_data *data = dreq->commit_data;
 469
 470         data->inode = dreq->inode;
 471         data->cred = dreq->ctx->cred;
 472
 473         data->args.fh = NFS_FH(data->inode);
 474         data->args.offset = dreq->pos;
 475         data->args.count = dreq->user_count;
 476         data->res.count = 0;
 477         data->res.fattr = &data->fattr;
 478         data->res.verf = &data->verf;
 479
 480         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 481                                 &nfs_commit_direct_ops, data);
 482         NFS_PROTO(data->inode)->commit_setup(data, 0);
 483
 484         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 485         data->task.tk_cookie = (unsigned long)data->inode;
 486         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 487         dreq->commit_data = NULL;
 488
 489         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 490
 491         lock_kernel();
 492         rpc_execute(&data->task);
 493         unlock_kernel();
 494 }
 495
 496 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 497 {
 498         int flags = dreq->flags;
 499
 500         dreq->flags = 0;
 501         switch (flags) {
 502                 case NFS_ODIRECT_DO_COMMIT:
 503                         nfs_direct_commit_schedule(dreq);
 504                         break;
 505                 case NFS_ODIRECT_RESCHED_WRITES:
 506                         nfs_direct_write_reschedule(dreq);
 507                         break;
 508                 default:
 509                         nfs_end_data_update(inode);
 510                         if (dreq->commit_data != NULL)
 511                                 nfs_commit_free(dreq->commit_data);
 512                         nfs_direct_free_writedata(dreq);
 513                         nfs_direct_complete(dreq);
 514         }
 515 }
 516
 517 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 518 {
 519         dreq->commit_data = nfs_commit_alloc(0);
 520         if (dreq->commit_data != NULL)
 521                 dreq->commit_data->req = (struct nfs_page *) dreq;
 522 }
 523 #else
 524 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 525 {
 526         dreq->commit_data = NULL;
 527 }
 528
 529 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 530 {
 531         nfs_end_data_update(inode);
 532         nfs_direct_free_writedata(dreq);
 533         nfs_direct_complete(dreq);
 534 }
 535 #endif
 536
 537 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 538 {
 539         struct list_head *list;
 540         struct nfs_direct_req *dreq;
 541         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 542
 543         dreq = nfs_direct_req_alloc();
 544         if (!dreq)
 545                 return NULL;
 546
 547         list = &dreq->list;
 548         for(;;) {
 549                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 550
 551                 if (unlikely(!data)) {
 552                         while (!list_empty(list)) {
 553                                 data = list_entry(list->next,
 554                                                   struct nfs_write_data, pages);
 555                                 list_del(&data->pages);
 556                                 nfs_writedata_free(data);
 557                         }
 558                         kref_put(&dreq->kref, nfs_direct_req_release);
 559                         return NULL;
 560                 }
 561
 562                 INIT_LIST_HEAD(&data->pages);
 563                 list_add(&data->pages, list);
 564
 565                 data->req = (struct nfs_page *) dreq;
 566                 dreq->outstanding++;
 567                 if (nbytes <= wsize)
 568                         break;
 569                 nbytes -= wsize;
 570         }
 571
 572         nfs_alloc_commit_data(dreq);
 573
 574         kref_get(&dreq->kref);
 575         return dreq;
 576 }
 577
 578 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 579 {
 580         struct nfs_write_data *data = calldata;
 581         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 582         int status = task->tk_status;
 583
 584         if (nfs_writeback_done(task, data) != 0)
 585                 return;
 586
 587         spin_lock(&dreq->lock);
 588
 589         if (likely(status >= 0))
 590                 dreq->count += data->res.count;
 591         else
 592                 dreq->error = task->tk_status;
 593
 594         if (data->res.verf->committed != NFS_FILE_SYNC) {
 595                 switch (dreq->flags) {
 596                         case 0:
 597                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 598                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 599                                 break;
 600                         case NFS_ODIRECT_DO_COMMIT:
 601                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 602                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 603                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 604                                 }
 605                 }
 606         }
 607         /* In case we have to resend */
 608         data->args.stable = NFS_FILE_SYNC;
 609
 610         spin_unlock(&dreq->lock);
 611 }
 612
 613 /*
 614  * NB: Return the value of the first error return code.  Subsequent
 615  *     errors after the first one are ignored.
 616  */
 617 static void nfs_direct_write_release(void *calldata)
 618 {
 619         struct nfs_write_data *data = calldata;
 620         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 621
 622         spin_lock(&dreq->lock);
 623         if (--dreq->outstanding) {
 624                 spin_unlock(&dreq->lock);
 625                 return;
 626         }
 627         spin_unlock(&dreq->lock);
 628
 629         nfs_direct_write_complete(dreq, data->inode);
 630 }
 631
 632 static const struct rpc_call_ops nfs_write_direct_ops = {
 633         .rpc_call_done = nfs_direct_write_result,
 634         .rpc_release = nfs_direct_write_release,
 635 };
 636
 637 /*
 638  * For each nfs_write_data struct that was allocated on the list, dispatch
 639  * an NFS WRITE operation
 640  */
 641 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync)
 642 {
 643         struct nfs_open_context *ctx = dreq->ctx;
 644         struct inode *inode = ctx->dentry->d_inode;
 645         struct list_head *list = &dreq->list;
 646         struct page **pages = dreq->pages;
 647         size_t count = dreq->user_count;
 648         loff_t pos = dreq->pos;
 649         size_t wsize = NFS_SERVER(inode)->wsize;
 650         unsigned int curpage, pgbase;
 651
 652         curpage = 0;
 653         pgbase = dreq->user_addr & ~PAGE_MASK;
 654         do {
 655                 struct nfs_write_data *data;
 656                 size_t bytes;
 657
 658                 bytes = wsize;
 659                 if (count < wsize)
 660                         bytes = count;
 661
 662                 BUG_ON(list_empty(list));
 663                 data = list_entry(list->next, struct nfs_write_data, pages);
 664                 list_move_tail(&data->pages, &dreq->rewrite_list);
 665
 666                 data->inode = inode;
 667                 data->cred = ctx->cred;
 668                 data->args.fh = NFS_FH(inode);
 669                 data->args.context = ctx;
 670                 data->args.offset = pos;
 671                 data->args.pgbase = pgbase;
 672                 data->args.pages = &pages[curpage];
 673                 data->args.count = bytes;
 674                 data->res.fattr = &data->fattr;
 675                 data->res.count = bytes;
 676                 data->res.verf = &data->verf;
 677
 678                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 679                                 &nfs_write_direct_ops, data);
 680                 NFS_PROTO(inode)->write_setup(data, sync);
 681
 682                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 683                 data->task.tk_cookie = (unsigned long) inode;
 684
 685                 lock_kernel();
 686                 rpc_execute(&data->task);
 687                 unlock_kernel();
 688
 689                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 690                                 data->task.tk_pid,
 691                                 inode->i_sb->s_id,
 692                                 (long long)NFS_FILEID(inode),
 693                                 bytes,
 694                                 (unsigned long long)data->args.offset);
 695
 696                 pos += bytes;
 697                 pgbase += bytes;
 698                 curpage += pgbase >> PAGE_SHIFT;
 699                 pgbase &= ~PAGE_MASK;
 700
 701                 count -= bytes;
 702         } while (count != 0);
 703         BUG_ON(!list_empty(list));
 704 }
 705
 706 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, int nr_pages)
 707 {
 708         ssize_t result;
 709         sigset_t oldset;
 710         struct inode *inode = iocb->ki_filp->f_mapping->host;
 711         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 712         struct nfs_direct_req *dreq;
 713         size_t wsize = NFS_SERVER(inode)->wsize;
 714         int sync = 0;
 715
 716         dreq = nfs_direct_write_alloc(count, wsize);
 717         if (!dreq)
 718                 return -ENOMEM;
 719         if (dreq->commit_data == NULL || count < wsize)
 720                 sync = FLUSH_STABLE;
 721
 722         dreq->user_addr = user_addr;
 723         dreq->user_count = count;
 724         dreq->pos = pos;
 725         dreq->pages = pages;
 726         dreq->npages = nr_pages;
 727         dreq->inode = inode;
 728         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 729         if (!is_sync_kiocb(iocb))
 730                 dreq->iocb = iocb;
 731
 732         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 733
 734         nfs_begin_data_update(inode);
 735
 736         rpc_clnt_sigmask(clnt, &oldset);
 737         nfs_direct_write_schedule(dreq, sync);
 738         result = nfs_direct_wait(dreq);
 739         rpc_clnt_sigunmask(clnt, &oldset);
 740
 741         return result;
 742 }
 743
 744 /**
 745  * nfs_file_direct_read - file direct read operation for NFS files
 746  * @iocb: target I/O control block
 747  * @buf: user's buffer into which to read data
 748  * @count: number of bytes to read
 749  * @pos: byte offset in file where reading starts
 750  *
 751  * We use this function for direct reads instead of calling
 752  * generic_file_aio_read() in order to avoid gfar's check to see if
 753  * the request starts before the end of the file.  For that check
 754  * to work, we must generate a GETATTR before each direct read, and
 755  * even then there is a window between the GETATTR and the subsequent
 756  * READ where the file size could change.  Our preference is simply
 757  * to do all reads the application wants, and the server will take
 758  * care of managing the end of file boundary.
 759  *
 760  * This function also eliminates unnecessarily updating the file's
 761  * atime locally, as the NFS server sets the file's atime, and this
 762  * client must read the updated atime from the server back into its
 763  * cache.
 764  */
 765 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 766 {
 767         ssize_t retval = -EINVAL;
 768         int page_count;
 769         struct page **pages;
 770         struct file *file = iocb->ki_filp;
 771         struct address_space *mapping = file->f_mapping;
 772
 773         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 774                 file->f_dentry->d_parent->d_name.name,
 775                 file->f_dentry->d_name.name,
 776                 (unsigned long) count, (long long) pos);
 777
 778         if (count < 0)
 779                 goto out;
 780         retval = -EFAULT;
 781         if (!access_ok(VERIFY_WRITE, buf, count))
 782                 goto out;
 783         retval = 0;
 784         if (!count)
 785                 goto out;
 786
 787         retval = nfs_sync_mapping(mapping);
 788         if (retval)
 789                 goto out;
 790
 791         retval = nfs_get_user_pages(READ, (unsigned long) buf,
 792                                                 count, &pages);
 793         if (retval < 0)
 794                 goto out;
 795         page_count = retval;
 796
 797         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 798                                                 pages, page_count);
 799         if (retval > 0)
 800                 iocb->ki_pos = pos + retval;
 801
 802 out:
 803         return retval;
 804 }
 805
 806 /**
 807  * nfs_file_direct_write - file direct write operation for NFS files
 808  * @iocb: target I/O control block
 809  * @buf: user's buffer from which to write data
 810  * @count: number of bytes to write
 811  * @pos: byte offset in file where writing starts
 812  *
 813  * We use this function for direct writes instead of calling
 814  * generic_file_aio_write() in order to avoid taking the inode
 815  * semaphore and updating the i_size.  The NFS server will set
 816  * the new i_size and this client must read the updated size
 817  * back into its cache.  We let the server do generic write
 818  * parameter checking and report problems.
 819  *
 820  * We also avoid an unnecessary invocation of generic_osync_inode(),
 821  * as it is fairly meaningless to sync the metadata of an NFS file.
 822  *
 823  * We eliminate local atime updates, see direct read above.
 824  *
 825  * We avoid unnecessary page cache invalidations for normal cached
 826  * readers of this file.
 827  *
 828  * Note that O_APPEND is not supported for NFS direct writes, as there
 829  * is no atomic O_APPEND write facility in the NFS protocol.
 830  */
 831 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 832 {
 833         ssize_t retval;
 834         int page_count;
 835         struct page **pages;
 836         struct file *file = iocb->ki_filp;
 837         struct address_space *mapping = file->f_mapping;
 838
 839         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 840                 file->f_dentry->d_parent->d_name.name,
 841                 file->f_dentry->d_name.name,
 842                 (unsigned long) count, (long long) pos);
 843
 844         retval = generic_write_checks(file, &pos, &count, 0);
 845         if (retval)
 846                 goto out;
 847
 848         retval = -EINVAL;
 849         if ((ssize_t) count < 0)
 850                 goto out;
 851         retval = 0;
 852         if (!count)
 853                 goto out;
 854
 855         retval = -EFAULT;
 856         if (!access_ok(VERIFY_READ, buf, count))
 857                 goto out;
 858
 859         retval = nfs_sync_mapping(mapping);
 860         if (retval)
 861                 goto out;
 862
 863         retval = nfs_get_user_pages(WRITE, (unsigned long) buf,
 864                                                 count, &pages);
 865         if (retval < 0)
 866                 goto out;
 867         page_count = retval;
 868
 869         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 870                                         pos, pages, page_count);
 871
 872         /*
 873          * XXX: nfs_end_data_update() already ensures this file's
 874          *      cached data is subsequently invalidated.  Do we really
 875          *      need to call invalidate_inode_pages2() again here?
 876          *
 877          *      For aio writes, this invalidation will almost certainly
 878          *      occur before the writes complete.  Kind of racey.
 879          */
 880         if (mapping->nrpages)
 881                 invalidate_inode_pages2(mapping);
 882
 883         if (retval > 0)
 884                 iocb->ki_pos = pos + retval;
 885
 886 out:
 887         return retval;
 888 }
 889
 890 /**
 891  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 892  *
 893  */
 894 int __init nfs_init_directcache(void)
 895 {
 896         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 897                                                 sizeof(struct nfs_direct_req),
 898                                                 0, (SLAB_RECLAIM_ACCOUNT|
 899                                                         SLAB_MEM_SPREAD),
 900                                                 NULL, NULL);
 901         if (nfs_direct_cachep == NULL)
 902                 return -ENOMEM;
 903
 904         return 0;
 905 }
 906
 907 /**
 908  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 909  *
 910  */
 911 void nfs_destroy_directcache(void)
 912 {
 913         if (kmem_cache_destroy(nfs_direct_cachep))
 914                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 915 }