err.no Git - linux-2.6/blob - fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/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 void nfs_free_user_pages(struct page **pages, int npages, int do_dirty);
  63 static kmem_cache_t *nfs_direct_cachep;
  64
  65 /*
  66  * This represents a set of asynchronous requests that we're waiting on
  67  */
  68 struct nfs_direct_req {
  69         struct kref             kref;           /* release manager */
  70
  71         /* I/O parameters */
  72         struct list_head        list,           /* nfs_read/write_data structs */
  73                                 rewrite_list;   /* saved nfs_write_data structs */
  74         struct file *           filp;           /* file descriptor */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         wait_queue_head_t       wait;           /* wait for i/o completion */
  77         struct inode *          inode;          /* target file of i/o */
  78         unsigned long           user_addr;      /* location of user's buffer */
  79         size_t                  user_count;     /* total bytes to move */
  80         loff_t                  pos;            /* starting offset in file */
  81         struct page **          pages;          /* pages in our buffer */
  82         unsigned int            npages;         /* count of pages */
  83
  84         /* completion state */
  85         spinlock_t              lock;           /* protect completion state */
  86         int                     outstanding;    /* i/os we're waiting for */
  87         ssize_t                 count,          /* bytes actually processed */
  88                                 error;          /* any reported error */
  89
  90         /* commit state */
  91         struct nfs_write_data * commit_data;    /* special write_data for commits */
  92         int                     flags;
  93 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  94 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  95         struct nfs_writeverf    verf;           /* unstable write verifier */
  96 };
  97
  98 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync);
  99 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 100
 101 /**
 102  * nfs_direct_IO - NFS address space operation for direct I/O
 103  * @rw: direction (read or write)
 104  * @iocb: target I/O control block
 105  * @iov: array of vectors that define I/O buffer
 106  * @pos: offset in file to begin the operation
 107  * @nr_segs: size of iovec array
 108  *
 109  * The presence of this routine in the address space ops vector means
 110  * the NFS client supports direct I/O.  However, we shunt off direct
 111  * read and write requests before the VFS gets them, so this method
 112  * should never be called.
 113  */
 114 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 115 {
 116         struct dentry *dentry = iocb->ki_filp->f_dentry;
 117
 118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 119                         dentry->d_name.name, (long long) pos, nr_segs);
 120
 121         return -EINVAL;
 122 }
 123
 124 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 125 {
 126         int result = -ENOMEM;
 127         unsigned long page_count;
 128         size_t array_size;
 129
 130         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 131         page_count -= user_addr >> PAGE_SHIFT;
 132
 133         array_size = (page_count * sizeof(struct page *));
 134         *pages = kmalloc(array_size, GFP_KERNEL);
 135         if (*pages) {
 136                 down_read(&current->mm->mmap_sem);
 137                 result = get_user_pages(current, current->mm, user_addr,
 138                                         page_count, (rw == READ), 0,
 139                                         *pages, NULL);
 140                 up_read(&current->mm->mmap_sem);
 141                 /*
 142                  * If we got fewer pages than expected from get_user_pages(),
 143                  * the user buffer runs off the end of a mapping; return EFAULT.
 144                  */
 145                 if (result >= 0 && result < page_count) {
 146                         nfs_free_user_pages(*pages, result, 0);
 147                         *pages = NULL;
 148                         result = -EFAULT;
 149                 }
 150         }
 151         return result;
 152 }
 153
 154 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 155 {
 156         int i;
 157         for (i = 0; i < npages; i++) {
 158                 struct page *page = pages[i];
 159                 if (do_dirty && !PageCompound(page))
 160                         set_page_dirty_lock(page);
 161                 page_cache_release(page);
 162         }
 163         kfree(pages);
 164 }
 165
 166 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 167 {
 168         struct nfs_direct_req *dreq;
 169
 170         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 171         if (!dreq)
 172                 return NULL;
 173
 174         kref_init(&dreq->kref);
 175         init_waitqueue_head(&dreq->wait);
 176         INIT_LIST_HEAD(&dreq->list);
 177         INIT_LIST_HEAD(&dreq->rewrite_list);
 178         dreq->iocb = NULL;
 179         spin_lock_init(&dreq->lock);
 180         dreq->outstanding = 0;
 181         dreq->count = 0;
 182         dreq->error = 0;
 183         dreq->flags = 0;
 184
 185         return dreq;
 186 }
 187
 188 static void nfs_direct_req_release(struct kref *kref)
 189 {
 190         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 191         kmem_cache_free(nfs_direct_cachep, dreq);
 192 }
 193
 194 /*
 195  * Collects and returns the final error value/byte-count.
 196  */
 197 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 198 {
 199         ssize_t result = -EIOCBQUEUED;
 200
 201         /* Async requests don't wait here */
 202         if (dreq->iocb)
 203                 goto out;
 204
 205         result = wait_event_interruptible(dreq->wait, (dreq->outstanding == 0));
 206
 207         if (!result)
 208                 result = dreq->error;
 209         if (!result)
 210                 result = dreq->count;
 211
 212 out:
 213         kref_put(&dreq->kref, nfs_direct_req_release);
 214         return (ssize_t) result;
 215 }
 216
 217 /*
 218  * We must hold a reference to all the pages in this direct read request
 219  * until the RPCs complete.  This could be long *after* we are woken up in
 220  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 221  *
 222  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 223  * can't trust the iocb is still valid here if this is a synchronous
 224  * request.  If the waiter is woken prematurely, the iocb is long gone.
 225  */
 226 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 227 {
 228         nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 229
 230         if (dreq->iocb) {
 231                 long res = (long) dreq->error;
 232                 if (!res)
 233                         res = (long) dreq->count;
 234                 aio_complete(dreq->iocb, res, 0);
 235         } else
 236                 wake_up(&dreq->wait);
 237
 238         iput(dreq->inode);
 239         kref_put(&dreq->kref, nfs_direct_req_release);
 240 }
 241
 242 /*
 243  * Note we also set the number of requests we have in the dreq when we are
 244  * done.  This prevents races with I/O completion so we will always wait
 245  * until all requests have been dispatched and completed.
 246  */
 247 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 248 {
 249         struct list_head *list;
 250         struct nfs_direct_req *dreq;
 251         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 252
 253         dreq = nfs_direct_req_alloc();
 254         if (!dreq)
 255                 return NULL;
 256
 257         list = &dreq->list;
 258         for(;;) {
 259                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 260
 261                 if (unlikely(!data)) {
 262                         while (!list_empty(list)) {
 263                                 data = list_entry(list->next,
 264                                                   struct nfs_read_data, pages);
 265                                 list_del(&data->pages);
 266                                 nfs_readdata_free(data);
 267                         }
 268                         kref_put(&dreq->kref, nfs_direct_req_release);
 269                         return NULL;
 270                 }
 271
 272                 INIT_LIST_HEAD(&data->pages);
 273                 list_add(&data->pages, list);
 274
 275                 data->req = (struct nfs_page *) dreq;
 276                 dreq->outstanding++;
 277                 if (nbytes <= rsize)
 278                         break;
 279                 nbytes -= rsize;
 280         }
 281         kref_get(&dreq->kref);
 282         return dreq;
 283 }
 284
 285 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 286 {
 287         struct nfs_read_data *data = calldata;
 288         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 289
 290         if (nfs_readpage_result(task, data) != 0)
 291                 return;
 292
 293         spin_lock(&dreq->lock);
 294
 295         if (likely(task->tk_status >= 0))
 296                 dreq->count += data->res.count;
 297         else
 298                 dreq->error = task->tk_status;
 299
 300         if (--dreq->outstanding) {
 301                 spin_unlock(&dreq->lock);
 302                 return;
 303         }
 304
 305         spin_unlock(&dreq->lock);
 306         nfs_direct_complete(dreq);
 307 }
 308
 309 static const struct rpc_call_ops nfs_read_direct_ops = {
 310         .rpc_call_done = nfs_direct_read_result,
 311         .rpc_release = nfs_readdata_release,
 312 };
 313
 314 /*
 315  * For each nfs_read_data struct that was allocated on the list, dispatch
 316  * an NFS READ operation
 317  */
 318 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq)
 319 {
 320         struct file *file = dreq->filp;
 321         struct inode *inode = file->f_mapping->host;
 322         struct nfs_open_context *ctx = (struct nfs_open_context *)
 323                                                         file->private_data;
 324         struct list_head *list = &dreq->list;
 325         struct page **pages = dreq->pages;
 326         size_t count = dreq->user_count;
 327         loff_t pos = dreq->pos;
 328         size_t rsize = NFS_SERVER(inode)->rsize;
 329         unsigned int curpage, pgbase;
 330
 331         curpage = 0;
 332         pgbase = dreq->user_addr & ~PAGE_MASK;
 333         do {
 334                 struct nfs_read_data *data;
 335                 size_t bytes;
 336
 337                 bytes = rsize;
 338                 if (count < rsize)
 339                         bytes = count;
 340
 341                 data = list_entry(list->next, struct nfs_read_data, pages);
 342                 list_del_init(&data->pages);
 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 = &pages[curpage];
 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: %4d initiated direct read call (req %s/%Ld, %u 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                 pos += bytes;
 374                 pgbase += bytes;
 375                 curpage += pgbase >> PAGE_SHIFT;
 376                 pgbase &= ~PAGE_MASK;
 377
 378                 count -= bytes;
 379         } while (count != 0);
 380 }
 381
 382 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)
 383 {
 384         ssize_t result;
 385         sigset_t oldset;
 386         struct inode *inode = iocb->ki_filp->f_mapping->host;
 387         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 388         struct nfs_direct_req *dreq;
 389
 390         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 391         if (!dreq)
 392                 return -ENOMEM;
 393
 394         dreq->user_addr = user_addr;
 395         dreq->user_count = count;
 396         dreq->pos = pos;
 397         dreq->pages = pages;
 398         dreq->npages = nr_pages;
 399         igrab(inode);
 400         dreq->inode = inode;
 401         dreq->filp = iocb->ki_filp;
 402         if (!is_sync_kiocb(iocb))
 403                 dreq->iocb = iocb;
 404
 405         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 406         rpc_clnt_sigmask(clnt, &oldset);
 407         nfs_direct_read_schedule(dreq);
 408         result = nfs_direct_wait(dreq);
 409         rpc_clnt_sigunmask(clnt, &oldset);
 410
 411         return result;
 412 }
 413
 414 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 415 {
 416         list_splice_init(&dreq->rewrite_list, &dreq->list);
 417         while (!list_empty(&dreq->list)) {
 418                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 419                 list_del(&data->pages);
 420                 nfs_writedata_release(data);
 421         }
 422 }
 423
 424 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 425 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 426 {
 427         struct list_head *pos;
 428
 429         list_splice_init(&dreq->rewrite_list, &dreq->list);
 430         list_for_each(pos, &dreq->list)
 431                 dreq->outstanding++;
 432         dreq->count = 0;
 433
 434         nfs_direct_write_schedule(dreq, FLUSH_STABLE);
 435 }
 436
 437 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 438 {
 439         struct nfs_write_data *data = calldata;
 440         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 441
 442         /* Call the NFS version-specific code */
 443         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 444                 return;
 445         if (unlikely(task->tk_status < 0)) {
 446                 dreq->error = task->tk_status;
 447                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 448         }
 449         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 450                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 451                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 452         }
 453
 454         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 455         nfs_direct_write_complete(dreq, data->inode);
 456 }
 457
 458 static const struct rpc_call_ops nfs_commit_direct_ops = {
 459         .rpc_call_done = nfs_direct_commit_result,
 460         .rpc_release = nfs_commit_release,
 461 };
 462
 463 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 464 {
 465         struct file *file = dreq->filp;
 466         struct nfs_open_context *ctx = (struct nfs_open_context *)
 467                                                         file->private_data;
 468         struct nfs_write_data *data = dreq->commit_data;
 469         struct rpc_task *task = &data->task;
 470
 471         data->inode = dreq->inode;
 472         data->cred = ctx->cred;
 473
 474         data->args.fh = NFS_FH(data->inode);
 475         data->args.offset = dreq->pos;
 476         data->args.count = dreq->user_count;
 477         data->res.count = 0;
 478         data->res.fattr = &data->fattr;
 479         data->res.verf = &data->verf;
 480
 481         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 482                                 &nfs_commit_direct_ops, data);
 483         NFS_PROTO(data->inode)->commit_setup(data, 0);
 484
 485         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 486         data->task.tk_cookie = (unsigned long)data->inode;
 487         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 488         dreq->commit_data = NULL;
 489
 490         dprintk("NFS: %5u initiated commit call\n", task->tk_pid);
 491
 492         lock_kernel();
 493         rpc_execute(&data->task);
 494         unlock_kernel();
 495 }
 496
 497 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 498 {
 499         int flags = dreq->flags;
 500
 501         dreq->flags = 0;
 502         switch (flags) {
 503                 case NFS_ODIRECT_DO_COMMIT:
 504                         nfs_direct_commit_schedule(dreq);
 505                         break;
 506                 case NFS_ODIRECT_RESCHED_WRITES:
 507                         nfs_direct_write_reschedule(dreq);
 508                         break;
 509                 default:
 510                         nfs_end_data_update(inode);
 511                         if (dreq->commit_data != NULL)
 512                                 nfs_commit_free(dreq->commit_data);
 513                         nfs_direct_free_writedata(dreq);
 514                         nfs_direct_complete(dreq);
 515         }
 516 }
 517
 518 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 519 {
 520         dreq->commit_data = nfs_commit_alloc(0);
 521         if (dreq->commit_data != NULL)
 522                 dreq->commit_data->req = (struct nfs_page *) dreq;
 523 }
 524 #else
 525 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 526 {
 527         dreq->commit_data = NULL;
 528 }
 529
 530 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 531 {
 532         nfs_end_data_update(inode);
 533         nfs_direct_free_writedata(dreq);
 534         nfs_direct_complete(dreq);
 535 }
 536 #endif
 537
 538 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 539 {
 540         struct list_head *list;
 541         struct nfs_direct_req *dreq;
 542         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 543
 544         dreq = nfs_direct_req_alloc();
 545         if (!dreq)
 546                 return NULL;
 547
 548         list = &dreq->list;
 549         for(;;) {
 550                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 551
 552                 if (unlikely(!data)) {
 553                         while (!list_empty(list)) {
 554                                 data = list_entry(list->next,
 555                                                   struct nfs_write_data, pages);
 556                                 list_del(&data->pages);
 557                                 nfs_writedata_free(data);
 558                         }
 559                         kref_put(&dreq->kref, nfs_direct_req_release);
 560                         return NULL;
 561                 }
 562
 563                 INIT_LIST_HEAD(&data->pages);
 564                 list_add(&data->pages, list);
 565
 566                 data->req = (struct nfs_page *) dreq;
 567                 dreq->outstanding++;
 568                 if (nbytes <= wsize)
 569                         break;
 570                 nbytes -= wsize;
 571         }
 572
 573         nfs_alloc_commit_data(dreq);
 574
 575         kref_get(&dreq->kref);
 576         return dreq;
 577 }
 578
 579 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 580 {
 581         struct nfs_write_data *data = calldata;
 582         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 583         int status = task->tk_status;
 584
 585         if (nfs_writeback_done(task, data) != 0)
 586                 return;
 587
 588         spin_lock(&dreq->lock);
 589
 590         if (likely(status >= 0))
 591                 dreq->count += data->res.count;
 592         else
 593                 dreq->error = task->tk_status;
 594
 595         if (data->res.verf->committed != NFS_FILE_SYNC) {
 596                 switch (dreq->flags) {
 597                         case 0:
 598                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 599                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 600                                 break;
 601                         case NFS_ODIRECT_DO_COMMIT:
 602                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 603                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 604                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 605                                 }
 606                 }
 607         }
 608         /* In case we have to resend */
 609         data->args.stable = NFS_FILE_SYNC;
 610
 611         spin_unlock(&dreq->lock);
 612 }
 613
 614 /*
 615  * NB: Return the value of the first error return code.  Subsequent
 616  *     errors after the first one are ignored.
 617  */
 618 static void nfs_direct_write_release(void *calldata)
 619 {
 620         struct nfs_write_data *data = calldata;
 621         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 622
 623         spin_lock(&dreq->lock);
 624         if (--dreq->outstanding) {
 625                 spin_unlock(&dreq->lock);
 626                 return;
 627         }
 628         spin_unlock(&dreq->lock);
 629
 630         nfs_direct_write_complete(dreq, data->inode);
 631 }
 632
 633 static const struct rpc_call_ops nfs_write_direct_ops = {
 634         .rpc_call_done = nfs_direct_write_result,
 635         .rpc_release = nfs_direct_write_release,
 636 };
 637
 638 /*
 639  * For each nfs_write_data struct that was allocated on the list, dispatch
 640  * an NFS WRITE operation
 641  */
 642 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync)
 643 {
 644         struct file *file = dreq->filp;
 645         struct inode *inode = file->f_mapping->host;
 646         struct nfs_open_context *ctx = (struct nfs_open_context *)
 647                                                         file->private_data;
 648         struct list_head *list = &dreq->list;
 649         struct page **pages = dreq->pages;
 650         size_t count = dreq->user_count;
 651         loff_t pos = dreq->pos;
 652         size_t wsize = NFS_SERVER(inode)->wsize;
 653         unsigned int curpage, pgbase;
 654
 655         curpage = 0;
 656         pgbase = dreq->user_addr & ~PAGE_MASK;
 657         do {
 658                 struct nfs_write_data *data;
 659                 size_t bytes;
 660
 661                 bytes = wsize;
 662                 if (count < wsize)
 663                         bytes = count;
 664
 665                 data = list_entry(list->next, struct nfs_write_data, pages);
 666                 list_move_tail(&data->pages, &dreq->rewrite_list);
 667
 668                 data->inode = inode;
 669                 data->cred = ctx->cred;
 670                 data->args.fh = NFS_FH(inode);
 671                 data->args.context = ctx;
 672                 data->args.offset = pos;
 673                 data->args.pgbase = pgbase;
 674                 data->args.pages = &pages[curpage];
 675                 data->args.count = bytes;
 676                 data->res.fattr = &data->fattr;
 677                 data->res.count = bytes;
 678                 data->res.verf = &data->verf;
 679
 680                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 681                                 &nfs_write_direct_ops, data);
 682                 NFS_PROTO(inode)->write_setup(data, sync);
 683
 684                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 685                 data->task.tk_cookie = (unsigned long) inode;
 686
 687                 lock_kernel();
 688                 rpc_execute(&data->task);
 689                 unlock_kernel();
 690
 691                 dfprintk(VFS, "NFS: %4d initiated direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 692                                 data->task.tk_pid,
 693                                 inode->i_sb->s_id,
 694                                 (long long)NFS_FILEID(inode),
 695                                 bytes,
 696                                 (unsigned long long)data->args.offset);
 697
 698                 pos += bytes;
 699                 pgbase += bytes;
 700                 curpage += pgbase >> PAGE_SHIFT;
 701                 pgbase &= ~PAGE_MASK;
 702
 703                 count -= bytes;
 704         } while (count != 0);
 705 }
 706
 707 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)
 708 {
 709         ssize_t result;
 710         sigset_t oldset;
 711         struct inode *inode = iocb->ki_filp->f_mapping->host;
 712         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 713         struct nfs_direct_req *dreq;
 714         size_t wsize = NFS_SERVER(inode)->wsize;
 715         int sync = 0;
 716
 717         dreq = nfs_direct_write_alloc(count, wsize);
 718         if (!dreq)
 719                 return -ENOMEM;
 720         if (dreq->commit_data == NULL || count < wsize)
 721                 sync = FLUSH_STABLE;
 722
 723         dreq->user_addr = user_addr;
 724         dreq->user_count = count;
 725         dreq->pos = pos;
 726         dreq->pages = pages;
 727         dreq->npages = nr_pages;
 728         igrab(inode);
 729         dreq->inode = inode;
 730         dreq->filp = iocb->ki_filp;
 731         if (!is_sync_kiocb(iocb))
 732                 dreq->iocb = iocb;
 733
 734         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 735
 736         nfs_begin_data_update(inode);
 737
 738         rpc_clnt_sigmask(clnt, &oldset);
 739         nfs_direct_write_schedule(dreq, sync);
 740         result = nfs_direct_wait(dreq);
 741         rpc_clnt_sigunmask(clnt, &oldset);
 742
 743         return result;
 744 }
 745
 746 /**
 747  * nfs_file_direct_read - file direct read operation for NFS files
 748  * @iocb: target I/O control block
 749  * @buf: user's buffer into which to read data
 750  * @count: number of bytes to read
 751  * @pos: byte offset in file where reading starts
 752  *
 753  * We use this function for direct reads instead of calling
 754  * generic_file_aio_read() in order to avoid gfar's check to see if
 755  * the request starts before the end of the file.  For that check
 756  * to work, we must generate a GETATTR before each direct read, and
 757  * even then there is a window between the GETATTR and the subsequent
 758  * READ where the file size could change.  Our preference is simply
 759  * to do all reads the application wants, and the server will take
 760  * care of managing the end of file boundary.
 761  *
 762  * This function also eliminates unnecessarily updating the file's
 763  * atime locally, as the NFS server sets the file's atime, and this
 764  * client must read the updated atime from the server back into its
 765  * cache.
 766  */
 767 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 768 {
 769         ssize_t retval = -EINVAL;
 770         int page_count;
 771         struct page **pages;
 772         struct file *file = iocb->ki_filp;
 773         struct address_space *mapping = file->f_mapping;
 774
 775         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 776                 file->f_dentry->d_parent->d_name.name,
 777                 file->f_dentry->d_name.name,
 778                 (unsigned long) count, (long long) pos);
 779
 780         if (count < 0)
 781                 goto out;
 782         retval = -EFAULT;
 783         if (!access_ok(VERIFY_WRITE, buf, count))
 784                 goto out;
 785         retval = 0;
 786         if (!count)
 787                 goto out;
 788
 789         retval = nfs_sync_mapping(mapping);
 790         if (retval)
 791                 goto out;
 792
 793         page_count = nfs_get_user_pages(READ, (unsigned long) buf,
 794                                                 count, &pages);
 795         if (page_count < 0) {
 796                 nfs_free_user_pages(pages, 0, 0);
 797                 retval = page_count;
 798                 goto out;
 799         }
 800
 801         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 802                                                 pages, page_count);
 803         if (retval > 0)
 804                 iocb->ki_pos = pos + retval;
 805
 806 out:
 807         return retval;
 808 }
 809
 810 /**
 811  * nfs_file_direct_write - file direct write operation for NFS files
 812  * @iocb: target I/O control block
 813  * @buf: user's buffer from which to write data
 814  * @count: number of bytes to write
 815  * @pos: byte offset in file where writing starts
 816  *
 817  * We use this function for direct writes instead of calling
 818  * generic_file_aio_write() in order to avoid taking the inode
 819  * semaphore and updating the i_size.  The NFS server will set
 820  * the new i_size and this client must read the updated size
 821  * back into its cache.  We let the server do generic write
 822  * parameter checking and report problems.
 823  *
 824  * We also avoid an unnecessary invocation of generic_osync_inode(),
 825  * as it is fairly meaningless to sync the metadata of an NFS file.
 826  *
 827  * We eliminate local atime updates, see direct read above.
 828  *
 829  * We avoid unnecessary page cache invalidations for normal cached
 830  * readers of this file.
 831  *
 832  * Note that O_APPEND is not supported for NFS direct writes, as there
 833  * is no atomic O_APPEND write facility in the NFS protocol.
 834  */
 835 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 836 {
 837         ssize_t retval;
 838         int page_count;
 839         struct page **pages;
 840         struct file *file = iocb->ki_filp;
 841         struct address_space *mapping = file->f_mapping;
 842
 843         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 844                 file->f_dentry->d_parent->d_name.name,
 845                 file->f_dentry->d_name.name,
 846                 (unsigned long) count, (long long) pos);
 847
 848         retval = generic_write_checks(file, &pos, &count, 0);
 849         if (retval)
 850                 goto out;
 851
 852         retval = -EINVAL;
 853         if ((ssize_t) count < 0)
 854                 goto out;
 855         retval = 0;
 856         if (!count)
 857                 goto out;
 858
 859         retval = -EFAULT;
 860         if (!access_ok(VERIFY_READ, buf, count))
 861                 goto out;
 862
 863         retval = nfs_sync_mapping(mapping);
 864         if (retval)
 865                 goto out;
 866
 867         page_count = nfs_get_user_pages(WRITE, (unsigned long) buf,
 868                                                 count, &pages);
 869         if (page_count < 0) {
 870                 nfs_free_user_pages(pages, 0, 0);
 871                 retval = page_count;
 872                 goto out;
 873         }
 874
 875         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 876                                         pos, pages, page_count);
 877
 878         /*
 879          * XXX: nfs_end_data_update() already ensures this file's
 880          *      cached data is subsequently invalidated.  Do we really
 881          *      need to call invalidate_inode_pages2() again here?
 882          *
 883          *      For aio writes, this invalidation will almost certainly
 884          *      occur before the writes complete.  Kind of racey.
 885          */
 886         if (mapping->nrpages)
 887                 invalidate_inode_pages2(mapping);
 888
 889         if (retval > 0)
 890                 iocb->ki_pos = pos + retval;
 891
 892 out:
 893         return retval;
 894 }
 895
 896 /**
 897  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 898  *
 899  */
 900 int nfs_init_directcache(void)
 901 {
 902         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 903                                                 sizeof(struct nfs_direct_req),
 904                                                 0, SLAB_RECLAIM_ACCOUNT,
 905                                                 NULL, NULL);
 906         if (nfs_direct_cachep == NULL)
 907                 return -ENOMEM;
 908
 909         return 0;
 910 }
 911
 912 /**
 913  * nfs_init_directcache - destroy the slab cache for nfs_direct_req structures
 914  *
 915  */
 916 void nfs_destroy_directcache(void)
 917 {
 918         if (kmem_cache_destroy(nfs_direct_cachep))
 919                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 920 }