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