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         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 116                         iocb->ki_filp->f_dentry->d_name.name,
 117                         (long long) pos, nr_segs);
 118
 119         return -EINVAL;
 120 }
 121
 122 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 123 {
 124         int i;
 125         for (i = 0; i < npages; i++) {
 126                 struct page *page = pages[i];
 127                 if (do_dirty && !PageCompound(page))
 128                         set_page_dirty_lock(page);
 129                 page_cache_release(page);
 130         }
 131         kfree(pages);
 132 }
 133
 134 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 135 {
 136         int result = -ENOMEM;
 137         unsigned long page_count;
 138         size_t array_size;
 139
 140         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 141         page_count -= user_addr >> PAGE_SHIFT;
 142
 143         array_size = (page_count * sizeof(struct page *));
 144         *pages = kmalloc(array_size, GFP_KERNEL);
 145         if (*pages) {
 146                 down_read(&current->mm->mmap_sem);
 147                 result = get_user_pages(current, current->mm, user_addr,
 148                                         page_count, (rw == READ), 0,
 149                                         *pages, NULL);
 150                 up_read(&current->mm->mmap_sem);
 151                 if (result != page_count) {
 152                         /*
 153                          * If we got fewer pages than expected from
 154                          * get_user_pages(), the user buffer runs off the
 155                          * end of a mapping; return EFAULT.
 156                          */
 157                         if (result >= 0) {
 158                                 nfs_free_user_pages(*pages, result, 0);
 159                                 result = -EFAULT;
 160                         } else
 161                                 kfree(*pages);
 162                         *pages = NULL;
 163                 }
 164         }
 165         return result;
 166 }
 167
 168 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 169 {
 170         struct nfs_direct_req *dreq;
 171
 172         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 173         if (!dreq)
 174                 return NULL;
 175
 176         kref_init(&dreq->kref);
 177         init_completion(&dreq->completion);
 178         INIT_LIST_HEAD(&dreq->list);
 179         INIT_LIST_HEAD(&dreq->rewrite_list);
 180         dreq->iocb = NULL;
 181         dreq->ctx = NULL;
 182         spin_lock_init(&dreq->lock);
 183         dreq->outstanding = 0;
 184         dreq->count = 0;
 185         dreq->error = 0;
 186         dreq->flags = 0;
 187
 188         return dreq;
 189 }
 190
 191 static void nfs_direct_req_release(struct kref *kref)
 192 {
 193         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 194
 195         if (dreq->ctx != NULL)
 196                 put_nfs_open_context(dreq->ctx);
 197         kmem_cache_free(nfs_direct_cachep, dreq);
 198 }
 199
 200 /*
 201  * Collects and returns the final error value/byte-count.
 202  */
 203 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 204 {
 205         ssize_t result = -EIOCBQUEUED;
 206
 207         /* Async requests don't wait here */
 208         if (dreq->iocb)
 209                 goto out;
 210
 211         result = wait_for_completion_interruptible(&dreq->completion);
 212
 213         if (!result)
 214                 result = dreq->error;
 215         if (!result)
 216                 result = dreq->count;
 217
 218 out:
 219         kref_put(&dreq->kref, nfs_direct_req_release);
 220         return (ssize_t) result;
 221 }
 222
 223 /*
 224  * We must hold a reference to all the pages in this direct read request
 225  * until the RPCs complete.  This could be long *after* we are woken up in
 226  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 227  *
 228  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 229  * can't trust the iocb is still valid here if this is a synchronous
 230  * request.  If the waiter is woken prematurely, the iocb is long gone.
 231  */
 232 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 233 {
 234         nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 235
 236         if (dreq->iocb) {
 237                 long res = (long) dreq->error;
 238                 if (!res)
 239                         res = (long) dreq->count;
 240                 aio_complete(dreq->iocb, res, 0);
 241         }
 242         complete_all(&dreq->completion);
 243
 244         kref_put(&dreq->kref, nfs_direct_req_release);
 245 }
 246
 247 /*
 248  * Note we also set the number of requests we have in the dreq when we are
 249  * done.  This prevents races with I/O completion so we will always wait
 250  * until all requests have been dispatched and completed.
 251  */
 252 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 253 {
 254         struct list_head *list;
 255         struct nfs_direct_req *dreq;
 256         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 257
 258         dreq = nfs_direct_req_alloc();
 259         if (!dreq)
 260                 return NULL;
 261
 262         list = &dreq->list;
 263         for(;;) {
 264                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 265
 266                 if (unlikely(!data)) {
 267                         while (!list_empty(list)) {
 268                                 data = list_entry(list->next,
 269                                                   struct nfs_read_data, pages);
 270                                 list_del(&data->pages);
 271                                 nfs_readdata_free(data);
 272                         }
 273                         kref_put(&dreq->kref, nfs_direct_req_release);
 274                         return NULL;
 275                 }
 276
 277                 INIT_LIST_HEAD(&data->pages);
 278                 list_add(&data->pages, list);
 279
 280                 data->req = (struct nfs_page *) dreq;
 281                 dreq->outstanding++;
 282                 if (nbytes <= rsize)
 283                         break;
 284                 nbytes -= rsize;
 285         }
 286         kref_get(&dreq->kref);
 287         return dreq;
 288 }
 289
 290 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 291 {
 292         struct nfs_read_data *data = calldata;
 293         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 294
 295         if (nfs_readpage_result(task, data) != 0)
 296                 return;
 297
 298         spin_lock(&dreq->lock);
 299
 300         if (likely(task->tk_status >= 0))
 301                 dreq->count += data->res.count;
 302         else
 303                 dreq->error = task->tk_status;
 304
 305         if (--dreq->outstanding) {
 306                 spin_unlock(&dreq->lock);
 307                 return;
 308         }
 309
 310         spin_unlock(&dreq->lock);
 311         nfs_direct_complete(dreq);
 312 }
 313
 314 static const struct rpc_call_ops nfs_read_direct_ops = {
 315         .rpc_call_done = nfs_direct_read_result,
 316         .rpc_release = nfs_readdata_release,
 317 };
 318
 319 /*
 320  * For each nfs_read_data struct that was allocated on the list, dispatch
 321  * an NFS READ operation
 322  */
 323 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq)
 324 {
 325         struct nfs_open_context *ctx = dreq->ctx;
 326         struct inode *inode = ctx->dentry->d_inode;
 327         struct list_head *list = &dreq->list;
 328         struct page **pages = dreq->pages;
 329         size_t count = dreq->user_count;
 330         loff_t pos = dreq->pos;
 331         size_t rsize = NFS_SERVER(inode)->rsize;
 332         unsigned int curpage, pgbase;
 333
 334         curpage = 0;
 335         pgbase = dreq->user_addr & ~PAGE_MASK;
 336         do {
 337                 struct nfs_read_data *data;
 338                 size_t bytes;
 339
 340                 bytes = rsize;
 341                 if (count < rsize)
 342                         bytes = count;
 343
 344                 BUG_ON(list_empty(list));
 345                 data = list_entry(list->next, struct nfs_read_data, pages);
 346                 list_del_init(&data->pages);
 347
 348                 data->inode = inode;
 349                 data->cred = ctx->cred;
 350                 data->args.fh = NFS_FH(inode);
 351                 data->args.context = ctx;
 352                 data->args.offset = pos;
 353                 data->args.pgbase = pgbase;
 354                 data->args.pages = &pages[curpage];
 355                 data->args.count = bytes;
 356                 data->res.fattr = &data->fattr;
 357                 data->res.eof = 0;
 358                 data->res.count = bytes;
 359
 360                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 361                                 &nfs_read_direct_ops, data);
 362                 NFS_PROTO(inode)->read_setup(data);
 363
 364                 data->task.tk_cookie = (unsigned long) inode;
 365
 366                 lock_kernel();
 367                 rpc_execute(&data->task);
 368                 unlock_kernel();
 369
 370                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 371                                 data->task.tk_pid,
 372                                 inode->i_sb->s_id,
 373                                 (long long)NFS_FILEID(inode),
 374                                 bytes,
 375                                 (unsigned long long)data->args.offset);
 376
 377                 pos += bytes;
 378                 pgbase += bytes;
 379                 curpage += pgbase >> PAGE_SHIFT;
 380                 pgbase &= ~PAGE_MASK;
 381
 382                 count -= bytes;
 383         } while (count != 0);
 384         BUG_ON(!list_empty(list));
 385 }
 386
 387 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)
 388 {
 389         ssize_t result;
 390         sigset_t oldset;
 391         struct inode *inode = iocb->ki_filp->f_mapping->host;
 392         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 393         struct nfs_direct_req *dreq;
 394
 395         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 396         if (!dreq)
 397                 return -ENOMEM;
 398
 399         dreq->user_addr = user_addr;
 400         dreq->user_count = count;
 401         dreq->pos = pos;
 402         dreq->pages = pages;
 403         dreq->npages = nr_pages;
 404         dreq->inode = inode;
 405         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 406         if (!is_sync_kiocb(iocb))
 407                 dreq->iocb = iocb;
 408
 409         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 410         rpc_clnt_sigmask(clnt, &oldset);
 411         nfs_direct_read_schedule(dreq);
 412         result = nfs_direct_wait(dreq);
 413         rpc_clnt_sigunmask(clnt, &oldset);
 414
 415         return result;
 416 }
 417
 418 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 419 {
 420         list_splice_init(&dreq->rewrite_list, &dreq->list);
 421         while (!list_empty(&dreq->list)) {
 422                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 423                 list_del(&data->pages);
 424                 nfs_writedata_release(data);
 425         }
 426 }
 427
 428 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 429 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 430 {
 431         struct list_head *pos;
 432
 433         list_splice_init(&dreq->rewrite_list, &dreq->list);
 434         list_for_each(pos, &dreq->list)
 435                 dreq->outstanding++;
 436         dreq->count = 0;
 437
 438         nfs_direct_write_schedule(dreq, FLUSH_STABLE);
 439 }
 440
 441 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 442 {
 443         struct nfs_write_data *data = calldata;
 444         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 445
 446         /* Call the NFS version-specific code */
 447         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 448                 return;
 449         if (unlikely(task->tk_status < 0)) {
 450                 dreq->error = task->tk_status;
 451                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 452         }
 453         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 454                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 455                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 456         }
 457
 458         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 459         nfs_direct_write_complete(dreq, data->inode);
 460 }
 461
 462 static const struct rpc_call_ops nfs_commit_direct_ops = {
 463         .rpc_call_done = nfs_direct_commit_result,
 464         .rpc_release = nfs_commit_release,
 465 };
 466
 467 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 468 {
 469         struct nfs_write_data *data = dreq->commit_data;
 470
 471         data->inode = dreq->inode;
 472         data->cred = dreq->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", data->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 nfs_open_context *ctx = dreq->ctx;
 645         struct inode *inode = ctx->dentry->d_inode;
 646         struct list_head *list = &dreq->list;
 647         struct page **pages = dreq->pages;
 648         size_t count = dreq->user_count;
 649         loff_t pos = dreq->pos;
 650         size_t wsize = NFS_SERVER(inode)->wsize;
 651         unsigned int curpage, pgbase;
 652
 653         curpage = 0;
 654         pgbase = dreq->user_addr & ~PAGE_MASK;
 655         do {
 656                 struct nfs_write_data *data;
 657                 size_t bytes;
 658
 659                 bytes = wsize;
 660                 if (count < wsize)
 661                         bytes = count;
 662
 663                 BUG_ON(list_empty(list));
 664                 data = list_entry(list->next, struct nfs_write_data, pages);
 665                 list_move_tail(&data->pages, &dreq->rewrite_list);
 666
 667                 data->inode = inode;
 668                 data->cred = ctx->cred;
 669                 data->args.fh = NFS_FH(inode);
 670                 data->args.context = ctx;
 671                 data->args.offset = pos;
 672                 data->args.pgbase = pgbase;
 673                 data->args.pages = &pages[curpage];
 674                 data->args.count = bytes;
 675                 data->res.fattr = &data->fattr;
 676                 data->res.count = bytes;
 677                 data->res.verf = &data->verf;
 678
 679                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 680                                 &nfs_write_direct_ops, data);
 681                 NFS_PROTO(inode)->write_setup(data, sync);
 682
 683                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 684                 data->task.tk_cookie = (unsigned long) inode;
 685
 686                 lock_kernel();
 687                 rpc_execute(&data->task);
 688                 unlock_kernel();
 689
 690                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 691                                 data->task.tk_pid,
 692                                 inode->i_sb->s_id,
 693                                 (long long)NFS_FILEID(inode),
 694                                 bytes,
 695                                 (unsigned long long)data->args.offset);
 696
 697                 pos += bytes;
 698                 pgbase += bytes;
 699                 curpage += pgbase >> PAGE_SHIFT;
 700                 pgbase &= ~PAGE_MASK;
 701
 702                 count -= bytes;
 703         } while (count != 0);
 704         BUG_ON(!list_empty(list));
 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         dreq->inode = inode;
 729         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 730         if (!is_sync_kiocb(iocb))
 731                 dreq->iocb = iocb;
 732
 733         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 734
 735         nfs_begin_data_update(inode);
 736
 737         rpc_clnt_sigmask(clnt, &oldset);
 738         nfs_direct_write_schedule(dreq, sync);
 739         result = nfs_direct_wait(dreq);
 740         rpc_clnt_sigunmask(clnt, &oldset);
 741
 742         return result;
 743 }
 744
 745 /**
 746  * nfs_file_direct_read - file direct read operation for NFS files
 747  * @iocb: target I/O control block
 748  * @buf: user's buffer into which to read data
 749  * @count: number of bytes to read
 750  * @pos: byte offset in file where reading starts
 751  *
 752  * We use this function for direct reads instead of calling
 753  * generic_file_aio_read() in order to avoid gfar's check to see if
 754  * the request starts before the end of the file.  For that check
 755  * to work, we must generate a GETATTR before each direct read, and
 756  * even then there is a window between the GETATTR and the subsequent
 757  * READ where the file size could change.  Our preference is simply
 758  * to do all reads the application wants, and the server will take
 759  * care of managing the end of file boundary.
 760  *
 761  * This function also eliminates unnecessarily updating the file's
 762  * atime locally, as the NFS server sets the file's atime, and this
 763  * client must read the updated atime from the server back into its
 764  * cache.
 765  */
 766 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 767 {
 768         ssize_t retval = -EINVAL;
 769         int page_count;
 770         struct page **pages;
 771         struct file *file = iocb->ki_filp;
 772         struct address_space *mapping = file->f_mapping;
 773
 774         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 775                 file->f_dentry->d_parent->d_name.name,
 776                 file->f_dentry->d_name.name,
 777                 (unsigned long) count, (long long) pos);
 778
 779         if (count < 0)
 780                 goto out;
 781         retval = -EFAULT;
 782         if (!access_ok(VERIFY_WRITE, buf, count))
 783                 goto out;
 784         retval = 0;
 785         if (!count)
 786                 goto out;
 787
 788         retval = nfs_sync_mapping(mapping);
 789         if (retval)
 790                 goto out;
 791
 792         retval = nfs_get_user_pages(READ, (unsigned long) buf,
 793                                                 count, &pages);
 794         if (retval < 0)
 795                 goto out;
 796         page_count = retval;
 797
 798         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 799                                                 pages, page_count);
 800         if (retval > 0)
 801                 iocb->ki_pos = pos + retval;
 802
 803 out:
 804         return retval;
 805 }
 806
 807 /**
 808  * nfs_file_direct_write - file direct write operation for NFS files
 809  * @iocb: target I/O control block
 810  * @buf: user's buffer from which to write data
 811  * @count: number of bytes to write
 812  * @pos: byte offset in file where writing starts
 813  *
 814  * We use this function for direct writes instead of calling
 815  * generic_file_aio_write() in order to avoid taking the inode
 816  * semaphore and updating the i_size.  The NFS server will set
 817  * the new i_size and this client must read the updated size
 818  * back into its cache.  We let the server do generic write
 819  * parameter checking and report problems.
 820  *
 821  * We also avoid an unnecessary invocation of generic_osync_inode(),
 822  * as it is fairly meaningless to sync the metadata of an NFS file.
 823  *
 824  * We eliminate local atime updates, see direct read above.
 825  *
 826  * We avoid unnecessary page cache invalidations for normal cached
 827  * readers of this file.
 828  *
 829  * Note that O_APPEND is not supported for NFS direct writes, as there
 830  * is no atomic O_APPEND write facility in the NFS protocol.
 831  */
 832 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 833 {
 834         ssize_t retval;
 835         int page_count;
 836         struct page **pages;
 837         struct file *file = iocb->ki_filp;
 838         struct address_space *mapping = file->f_mapping;
 839
 840         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 841                 file->f_dentry->d_parent->d_name.name,
 842                 file->f_dentry->d_name.name,
 843                 (unsigned long) count, (long long) pos);
 844
 845         retval = generic_write_checks(file, &pos, &count, 0);
 846         if (retval)
 847                 goto out;
 848
 849         retval = -EINVAL;
 850         if ((ssize_t) count < 0)
 851                 goto out;
 852         retval = 0;
 853         if (!count)
 854                 goto out;
 855
 856         retval = -EFAULT;
 857         if (!access_ok(VERIFY_READ, buf, count))
 858                 goto out;
 859
 860         retval = nfs_sync_mapping(mapping);
 861         if (retval)
 862                 goto out;
 863
 864         retval = nfs_get_user_pages(WRITE, (unsigned long) buf,
 865                                                 count, &pages);
 866         if (retval < 0)
 867                 goto out;
 868         page_count = retval;
 869
 870         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 871                                         pos, pages, page_count);
 872
 873         /*
 874          * XXX: nfs_end_data_update() already ensures this file's
 875          *      cached data is subsequently invalidated.  Do we really
 876          *      need to call invalidate_inode_pages2() again here?
 877          *
 878          *      For aio writes, this invalidation will almost certainly
 879          *      occur before the writes complete.  Kind of racey.
 880          */
 881         if (mapping->nrpages)
 882                 invalidate_inode_pages2(mapping);
 883
 884         if (retval > 0)
 885                 iocb->ki_pos = pos + retval;
 886
 887 out:
 888         return retval;
 889 }
 890
 891 /**
 892  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 893  *
 894  */
 895 int __init nfs_init_directcache(void)
 896 {
 897         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 898                                                 sizeof(struct nfs_direct_req),
 899                                                 0, (SLAB_RECLAIM_ACCOUNT|
 900                                                         SLAB_MEM_SPREAD),
 901                                                 NULL, NULL);
 902         if (nfs_direct_cachep == NULL)
 903                 return -ENOMEM;
 904
 905         return 0;
 906 }
 907
 908 /**
 909  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 910  *
 911  */
 912 void __exit nfs_destroy_directcache(void)
 913 {
 914         if (kmem_cache_destroy(nfs_direct_cachep))
 915                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 916 }