SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/config.h>
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/smp_lock.h>
  46 #include <linux/file.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/kref.h>
  49
  50 #include <linux/nfs_fs.h>
  51 #include <linux/nfs_page.h>
  52 #include <linux/sunrpc/clnt.h>
  53
  54 #include <asm/system.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/atomic.h>
  57
  58 #include "iostat.h"
  59
  60 #define NFSDBG_FACILITY         NFSDBG_VFS
  61
  62 static kmem_cache_t *nfs_direct_cachep;
  63
  64 /*
  65  * This represents a set of asynchronous requests that we're waiting on
  66  */
  67 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct list_head        list,           /* nfs_read/write_data structs */
  72                                 rewrite_list;   /* saved nfs_write_data structs */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct kiocb *          iocb;           /* controlling i/o request */
  75         struct inode *          inode;          /* target file of i/o */
  76         struct page **          pages;          /* pages in our buffer */
  77         unsigned int            npages;         /* count of pages */
  78
  79         /* completion state */
  80         atomic_t                io_count;       /* i/os we're waiting for */
  81         spinlock_t              lock;           /* protect completion state */
  82         ssize_t                 count,          /* bytes actually processed */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_write_data * commit_data;    /* special write_data for commits */
  88         int                     flags;
  89 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  90 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  91         struct nfs_writeverf    verf;           /* unstable write verifier */
  92 };
  93
  94 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  95 static const struct rpc_call_ops nfs_write_direct_ops;
  96
  97 static inline void get_dreq(struct nfs_direct_req *dreq)
  98 {
  99         atomic_inc(&dreq->io_count);
 100 }
 101
 102 static inline int put_dreq(struct nfs_direct_req *dreq)
 103 {
 104         return atomic_dec_and_test(&dreq->io_count);
 105 }
 106
 107 /**
 108  * nfs_direct_IO - NFS address space operation for direct I/O
 109  * @rw: direction (read or write)
 110  * @iocb: target I/O control block
 111  * @iov: array of vectors that define I/O buffer
 112  * @pos: offset in file to begin the operation
 113  * @nr_segs: size of iovec array
 114  *
 115  * The presence of this routine in the address space ops vector means
 116  * the NFS client supports direct I/O.  However, we shunt off direct
 117  * read and write requests before the VFS gets them, so this method
 118  * should never be called.
 119  */
 120 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 121 {
 122         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 123                         iocb->ki_filp->f_dentry->d_name.name,
 124                         (long long) pos, nr_segs);
 125
 126         return -EINVAL;
 127 }
 128
 129 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 130 {
 131         int i;
 132         for (i = 0; i < npages; i++) {
 133                 struct page *page = pages[i];
 134                 if (do_dirty && !PageCompound(page))
 135                         set_page_dirty_lock(page);
 136                 page_cache_release(page);
 137         }
 138         kfree(pages);
 139 }
 140
 141 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 142 {
 143         int result = -ENOMEM;
 144         unsigned long page_count;
 145         size_t array_size;
 146
 147         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 148         page_count -= user_addr >> PAGE_SHIFT;
 149
 150         array_size = (page_count * sizeof(struct page *));
 151         *pages = kmalloc(array_size, GFP_KERNEL);
 152         if (*pages) {
 153                 down_read(&current->mm->mmap_sem);
 154                 result = get_user_pages(current, current->mm, user_addr,
 155                                         page_count, (rw == READ), 0,
 156                                         *pages, NULL);
 157                 up_read(&current->mm->mmap_sem);
 158                 if (result != page_count) {
 159                         /*
 160                          * If we got fewer pages than expected from
 161                          * get_user_pages(), the user buffer runs off the
 162                          * end of a mapping; return EFAULT.
 163                          */
 164                         if (result >= 0) {
 165                                 nfs_free_user_pages(*pages, result, 0);
 166                                 result = -EFAULT;
 167                         } else
 168                                 kfree(*pages);
 169                         *pages = NULL;
 170                 }
 171         }
 172         return result;
 173 }
 174
 175 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 176 {
 177         struct nfs_direct_req *dreq;
 178
 179         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 180         if (!dreq)
 181                 return NULL;
 182
 183         kref_init(&dreq->kref);
 184         init_completion(&dreq->completion);
 185         INIT_LIST_HEAD(&dreq->list);
 186         INIT_LIST_HEAD(&dreq->rewrite_list);
 187         dreq->iocb = NULL;
 188         dreq->ctx = NULL;
 189         spin_lock_init(&dreq->lock);
 190         atomic_set(&dreq->io_count, 0);
 191         dreq->count = 0;
 192         dreq->error = 0;
 193         dreq->flags = 0;
 194
 195         return dreq;
 196 }
 197
 198 static void nfs_direct_req_release(struct kref *kref)
 199 {
 200         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 201
 202         if (dreq->ctx != NULL)
 203                 put_nfs_open_context(dreq->ctx);
 204         kmem_cache_free(nfs_direct_cachep, dreq);
 205 }
 206
 207 /*
 208  * Collects and returns the final error value/byte-count.
 209  */
 210 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 211 {
 212         ssize_t result = -EIOCBQUEUED;
 213
 214         /* Async requests don't wait here */
 215         if (dreq->iocb)
 216                 goto out;
 217
 218         result = wait_for_completion_interruptible(&dreq->completion);
 219
 220         if (!result)
 221                 result = dreq->error;
 222         if (!result)
 223                 result = dreq->count;
 224
 225 out:
 226         kref_put(&dreq->kref, nfs_direct_req_release);
 227         return (ssize_t) result;
 228 }
 229
 230 /*
 231  * We must hold a reference to all the pages in this direct read request
 232  * until the RPCs complete.  This could be long *after* we are woken up in
 233  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 234  *
 235  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 236  * can't trust the iocb is still valid here if this is a synchronous
 237  * request.  If the waiter is woken prematurely, the iocb is long gone.
 238  */
 239 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 240 {
 241         nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 242
 243         if (dreq->iocb) {
 244                 long res = (long) dreq->error;
 245                 if (!res)
 246                         res = (long) dreq->count;
 247                 aio_complete(dreq->iocb, res, 0);
 248         }
 249         complete_all(&dreq->completion);
 250
 251         kref_put(&dreq->kref, nfs_direct_req_release);
 252 }
 253
 254 /*
 255  * Note we also set the number of requests we have in the dreq when we are
 256  * done.  This prevents races with I/O completion so we will always wait
 257  * until all requests have been dispatched and completed.
 258  */
 259 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 260 {
 261         struct list_head *list;
 262         struct nfs_direct_req *dreq;
 263         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 264
 265         dreq = nfs_direct_req_alloc();
 266         if (!dreq)
 267                 return NULL;
 268
 269         list = &dreq->list;
 270         for(;;) {
 271                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 272
 273                 if (unlikely(!data)) {
 274                         while (!list_empty(list)) {
 275                                 data = list_entry(list->next,
 276                                                   struct nfs_read_data, pages);
 277                                 list_del(&data->pages);
 278                                 nfs_readdata_free(data);
 279                         }
 280                         kref_put(&dreq->kref, nfs_direct_req_release);
 281                         return NULL;
 282                 }
 283
 284                 INIT_LIST_HEAD(&data->pages);
 285                 list_add(&data->pages, list);
 286
 287                 data->req = (struct nfs_page *) dreq;
 288                 get_dreq(dreq);
 289                 if (nbytes <= rsize)
 290                         break;
 291                 nbytes -= rsize;
 292         }
 293         kref_get(&dreq->kref);
 294         return dreq;
 295 }
 296
 297 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 298 {
 299         struct nfs_read_data *data = calldata;
 300         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 301
 302         if (nfs_readpage_result(task, data) != 0)
 303                 return;
 304
 305         spin_lock(&dreq->lock);
 306
 307         if (likely(task->tk_status >= 0))
 308                 dreq->count += data->res.count;
 309         else
 310                 dreq->error = task->tk_status;
 311
 312         spin_unlock(&dreq->lock);
 313
 314         if (put_dreq(dreq))
 315                 nfs_direct_complete(dreq);
 316 }
 317
 318 static const struct rpc_call_ops nfs_read_direct_ops = {
 319         .rpc_call_done = nfs_direct_read_result,
 320         .rpc_release = nfs_readdata_release,
 321 };
 322
 323 /*
 324  * For each nfs_read_data struct that was allocated on the list, dispatch
 325  * an NFS READ operation
 326  */
 327 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 328 {
 329         struct nfs_open_context *ctx = dreq->ctx;
 330         struct inode *inode = ctx->dentry->d_inode;
 331         struct list_head *list = &dreq->list;
 332         struct page **pages = dreq->pages;
 333         size_t rsize = NFS_SERVER(inode)->rsize;
 334         unsigned int curpage, pgbase;
 335
 336         curpage = 0;
 337         pgbase = user_addr & ~PAGE_MASK;
 338         do {
 339                 struct nfs_read_data *data;
 340                 size_t bytes;
 341
 342                 bytes = rsize;
 343                 if (count < rsize)
 344                         bytes = count;
 345
 346                 BUG_ON(list_empty(list));
 347                 data = list_entry(list->next, struct nfs_read_data, pages);
 348                 list_del_init(&data->pages);
 349
 350                 data->inode = inode;
 351                 data->cred = ctx->cred;
 352                 data->args.fh = NFS_FH(inode);
 353                 data->args.context = ctx;
 354                 data->args.offset = pos;
 355                 data->args.pgbase = pgbase;
 356                 data->args.pages = &pages[curpage];
 357                 data->args.count = bytes;
 358                 data->res.fattr = &data->fattr;
 359                 data->res.eof = 0;
 360                 data->res.count = bytes;
 361
 362                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 363                                 &nfs_read_direct_ops, data);
 364                 NFS_PROTO(inode)->read_setup(data);
 365
 366                 data->task.tk_cookie = (unsigned long) inode;
 367
 368                 lock_kernel();
 369                 rpc_execute(&data->task);
 370                 unlock_kernel();
 371
 372                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 373                                 data->task.tk_pid,
 374                                 inode->i_sb->s_id,
 375                                 (long long)NFS_FILEID(inode),
 376                                 bytes,
 377                                 (unsigned long long)data->args.offset);
 378
 379                 pos += bytes;
 380                 pgbase += bytes;
 381                 curpage += pgbase >> PAGE_SHIFT;
 382                 pgbase &= ~PAGE_MASK;
 383
 384                 count -= bytes;
 385         } while (count != 0);
 386         BUG_ON(!list_empty(list));
 387 }
 388
 389 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)
 390 {
 391         ssize_t result;
 392         sigset_t oldset;
 393         struct inode *inode = iocb->ki_filp->f_mapping->host;
 394         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 395         struct nfs_direct_req *dreq;
 396
 397         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 398         if (!dreq)
 399                 return -ENOMEM;
 400
 401         dreq->pages = pages;
 402         dreq->npages = nr_pages;
 403         dreq->inode = inode;
 404         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 405         if (!is_sync_kiocb(iocb))
 406                 dreq->iocb = iocb;
 407
 408         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 409         rpc_clnt_sigmask(clnt, &oldset);
 410         nfs_direct_read_schedule(dreq, user_addr, count, pos);
 411         result = nfs_direct_wait(dreq);
 412         rpc_clnt_sigunmask(clnt, &oldset);
 413
 414         return result;
 415 }
 416
 417 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 418 {
 419         list_splice_init(&dreq->rewrite_list, &dreq->list);
 420         while (!list_empty(&dreq->list)) {
 421                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 422                 list_del(&data->pages);
 423                 nfs_writedata_release(data);
 424         }
 425 }
 426
 427 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 428 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 429 {
 430         struct inode *inode = dreq->inode;
 431         struct list_head *p;
 432         struct nfs_write_data *data;
 433
 434         dreq->count = 0;
 435         get_dreq(dreq);
 436
 437         list_for_each(p, &dreq->rewrite_list) {
 438                 data = list_entry(p, struct nfs_write_data, pages);
 439
 440                 get_dreq(dreq);
 441
 442                 /*
 443                  * Reset data->res.
 444                  */
 445                 nfs_fattr_init(&data->fattr);
 446                 data->res.count = data->args.count;
 447                 memset(&data->verf, 0, sizeof(data->verf));
 448
 449                 /*
 450                  * Reuse data->task; data->args should not have changed
 451                  * since the original request was sent.
 452                  */
 453                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 454                                 &nfs_write_direct_ops, data);
 455                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 456
 457                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 458                 data->task.tk_cookie = (unsigned long) inode;
 459
 460                 /*
 461                  * We're called via an RPC callback, so BKL is already held.
 462                  */
 463                 rpc_execute(&data->task);
 464
 465                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 466                                 data->task.tk_pid,
 467                                 inode->i_sb->s_id,
 468                                 (long long)NFS_FILEID(inode),
 469                                 data->args.count,
 470                                 (unsigned long long)data->args.offset);
 471         }
 472
 473         if (put_dreq(dreq))
 474                 nfs_direct_write_complete(dreq, inode);
 475 }
 476
 477 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 478 {
 479         struct nfs_write_data *data = calldata;
 480         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 481
 482         /* Call the NFS version-specific code */
 483         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 484                 return;
 485         if (unlikely(task->tk_status < 0)) {
 486                 dreq->error = task->tk_status;
 487                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 488         }
 489         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 490                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 491                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 492         }
 493
 494         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 495         nfs_direct_write_complete(dreq, data->inode);
 496 }
 497
 498 static const struct rpc_call_ops nfs_commit_direct_ops = {
 499         .rpc_call_done = nfs_direct_commit_result,
 500         .rpc_release = nfs_commit_release,
 501 };
 502
 503 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 504 {
 505         struct nfs_write_data *data = dreq->commit_data;
 506
 507         data->inode = dreq->inode;
 508         data->cred = dreq->ctx->cred;
 509
 510         data->args.fh = NFS_FH(data->inode);
 511         data->args.offset = 0;
 512         data->args.count = 0;
 513         data->res.count = 0;
 514         data->res.fattr = &data->fattr;
 515         data->res.verf = &data->verf;
 516
 517         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 518                                 &nfs_commit_direct_ops, data);
 519         NFS_PROTO(data->inode)->commit_setup(data, 0);
 520
 521         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 522         data->task.tk_cookie = (unsigned long)data->inode;
 523         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 524         dreq->commit_data = NULL;
 525
 526         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 527
 528         lock_kernel();
 529         rpc_execute(&data->task);
 530         unlock_kernel();
 531 }
 532
 533 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 534 {
 535         int flags = dreq->flags;
 536
 537         dreq->flags = 0;
 538         switch (flags) {
 539                 case NFS_ODIRECT_DO_COMMIT:
 540                         nfs_direct_commit_schedule(dreq);
 541                         break;
 542                 case NFS_ODIRECT_RESCHED_WRITES:
 543                         nfs_direct_write_reschedule(dreq);
 544                         break;
 545                 default:
 546                         nfs_end_data_update(inode);
 547                         if (dreq->commit_data != NULL)
 548                                 nfs_commit_free(dreq->commit_data);
 549                         nfs_direct_free_writedata(dreq);
 550                         nfs_direct_complete(dreq);
 551         }
 552 }
 553
 554 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 555 {
 556         dreq->commit_data = nfs_commit_alloc(0);
 557         if (dreq->commit_data != NULL)
 558                 dreq->commit_data->req = (struct nfs_page *) dreq;
 559 }
 560 #else
 561 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 562 {
 563         dreq->commit_data = NULL;
 564 }
 565
 566 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 567 {
 568         nfs_end_data_update(inode);
 569         nfs_direct_free_writedata(dreq);
 570         nfs_direct_complete(dreq);
 571 }
 572 #endif
 573
 574 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 575 {
 576         struct list_head *list;
 577         struct nfs_direct_req *dreq;
 578         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 579
 580         dreq = nfs_direct_req_alloc();
 581         if (!dreq)
 582                 return NULL;
 583
 584         list = &dreq->list;
 585         for(;;) {
 586                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 587
 588                 if (unlikely(!data)) {
 589                         while (!list_empty(list)) {
 590                                 data = list_entry(list->next,
 591                                                   struct nfs_write_data, pages);
 592                                 list_del(&data->pages);
 593                                 nfs_writedata_free(data);
 594                         }
 595                         kref_put(&dreq->kref, nfs_direct_req_release);
 596                         return NULL;
 597                 }
 598
 599                 INIT_LIST_HEAD(&data->pages);
 600                 list_add(&data->pages, list);
 601
 602                 data->req = (struct nfs_page *) dreq;
 603                 get_dreq(dreq);
 604                 if (nbytes <= wsize)
 605                         break;
 606                 nbytes -= wsize;
 607         }
 608
 609         nfs_alloc_commit_data(dreq);
 610
 611         kref_get(&dreq->kref);
 612         return dreq;
 613 }
 614
 615 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 616 {
 617         struct nfs_write_data *data = calldata;
 618         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 619         int status = task->tk_status;
 620
 621         if (nfs_writeback_done(task, data) != 0)
 622                 return;
 623
 624         spin_lock(&dreq->lock);
 625
 626         if (likely(status >= 0))
 627                 dreq->count += data->res.count;
 628         else
 629                 dreq->error = task->tk_status;
 630
 631         if (data->res.verf->committed != NFS_FILE_SYNC) {
 632                 switch (dreq->flags) {
 633                         case 0:
 634                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 635                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 636                                 break;
 637                         case NFS_ODIRECT_DO_COMMIT:
 638                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 639                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 640                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 641                                 }
 642                 }
 643         }
 644
 645         spin_unlock(&dreq->lock);
 646 }
 647
 648 /*
 649  * NB: Return the value of the first error return code.  Subsequent
 650  *     errors after the first one are ignored.
 651  */
 652 static void nfs_direct_write_release(void *calldata)
 653 {
 654         struct nfs_write_data *data = calldata;
 655         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 656
 657         if (put_dreq(dreq))
 658                 nfs_direct_write_complete(dreq, data->inode);
 659 }
 660
 661 static const struct rpc_call_ops nfs_write_direct_ops = {
 662         .rpc_call_done = nfs_direct_write_result,
 663         .rpc_release = nfs_direct_write_release,
 664 };
 665
 666 /*
 667  * For each nfs_write_data struct that was allocated on the list, dispatch
 668  * an NFS WRITE operation
 669  */
 670 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 671 {
 672         struct nfs_open_context *ctx = dreq->ctx;
 673         struct inode *inode = ctx->dentry->d_inode;
 674         struct list_head *list = &dreq->list;
 675         struct page **pages = dreq->pages;
 676         size_t wsize = NFS_SERVER(inode)->wsize;
 677         unsigned int curpage, pgbase;
 678
 679         curpage = 0;
 680         pgbase = user_addr & ~PAGE_MASK;
 681         do {
 682                 struct nfs_write_data *data;
 683                 size_t bytes;
 684
 685                 bytes = wsize;
 686                 if (count < wsize)
 687                         bytes = count;
 688
 689                 BUG_ON(list_empty(list));
 690                 data = list_entry(list->next, struct nfs_write_data, pages);
 691                 list_move_tail(&data->pages, &dreq->rewrite_list);
 692
 693                 data->inode = inode;
 694                 data->cred = ctx->cred;
 695                 data->args.fh = NFS_FH(inode);
 696                 data->args.context = ctx;
 697                 data->args.offset = pos;
 698                 data->args.pgbase = pgbase;
 699                 data->args.pages = &pages[curpage];
 700                 data->args.count = bytes;
 701                 data->res.fattr = &data->fattr;
 702                 data->res.count = bytes;
 703                 data->res.verf = &data->verf;
 704
 705                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 706                                 &nfs_write_direct_ops, data);
 707                 NFS_PROTO(inode)->write_setup(data, sync);
 708
 709                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 710                 data->task.tk_cookie = (unsigned long) inode;
 711
 712                 lock_kernel();
 713                 rpc_execute(&data->task);
 714                 unlock_kernel();
 715
 716                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 717                                 data->task.tk_pid,
 718                                 inode->i_sb->s_id,
 719                                 (long long)NFS_FILEID(inode),
 720                                 bytes,
 721                                 (unsigned long long)data->args.offset);
 722
 723                 pos += bytes;
 724                 pgbase += bytes;
 725                 curpage += pgbase >> PAGE_SHIFT;
 726                 pgbase &= ~PAGE_MASK;
 727
 728                 count -= bytes;
 729         } while (count != 0);
 730         BUG_ON(!list_empty(list));
 731 }
 732
 733 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)
 734 {
 735         ssize_t result;
 736         sigset_t oldset;
 737         struct inode *inode = iocb->ki_filp->f_mapping->host;
 738         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 739         struct nfs_direct_req *dreq;
 740         size_t wsize = NFS_SERVER(inode)->wsize;
 741         int sync = 0;
 742
 743         dreq = nfs_direct_write_alloc(count, wsize);
 744         if (!dreq)
 745                 return -ENOMEM;
 746         if (dreq->commit_data == NULL || count < wsize)
 747                 sync = FLUSH_STABLE;
 748
 749         dreq->pages = pages;
 750         dreq->npages = nr_pages;
 751         dreq->inode = inode;
 752         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 753         if (!is_sync_kiocb(iocb))
 754                 dreq->iocb = iocb;
 755
 756         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 757
 758         nfs_begin_data_update(inode);
 759
 760         rpc_clnt_sigmask(clnt, &oldset);
 761         nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 762         result = nfs_direct_wait(dreq);
 763         rpc_clnt_sigunmask(clnt, &oldset);
 764
 765         return result;
 766 }
 767
 768 /**
 769  * nfs_file_direct_read - file direct read operation for NFS files
 770  * @iocb: target I/O control block
 771  * @buf: user's buffer into which to read data
 772  * @count: number of bytes to read
 773  * @pos: byte offset in file where reading starts
 774  *
 775  * We use this function for direct reads instead of calling
 776  * generic_file_aio_read() in order to avoid gfar's check to see if
 777  * the request starts before the end of the file.  For that check
 778  * to work, we must generate a GETATTR before each direct read, and
 779  * even then there is a window between the GETATTR and the subsequent
 780  * READ where the file size could change.  Our preference is simply
 781  * to do all reads the application wants, and the server will take
 782  * care of managing the end of file boundary.
 783  *
 784  * This function also eliminates unnecessarily updating the file's
 785  * atime locally, as the NFS server sets the file's atime, and this
 786  * client must read the updated atime from the server back into its
 787  * cache.
 788  */
 789 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 790 {
 791         ssize_t retval = -EINVAL;
 792         int page_count;
 793         struct page **pages;
 794         struct file *file = iocb->ki_filp;
 795         struct address_space *mapping = file->f_mapping;
 796
 797         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 798                 file->f_dentry->d_parent->d_name.name,
 799                 file->f_dentry->d_name.name,
 800                 (unsigned long) count, (long long) pos);
 801
 802         if (count < 0)
 803                 goto out;
 804         retval = -EFAULT;
 805         if (!access_ok(VERIFY_WRITE, buf, count))
 806                 goto out;
 807         retval = 0;
 808         if (!count)
 809                 goto out;
 810
 811         retval = nfs_sync_mapping(mapping);
 812         if (retval)
 813                 goto out;
 814
 815         retval = nfs_get_user_pages(READ, (unsigned long) buf,
 816                                                 count, &pages);
 817         if (retval < 0)
 818                 goto out;
 819         page_count = retval;
 820
 821         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 822                                                 pages, page_count);
 823         if (retval > 0)
 824                 iocb->ki_pos = pos + retval;
 825
 826 out:
 827         return retval;
 828 }
 829
 830 /**
 831  * nfs_file_direct_write - file direct write operation for NFS files
 832  * @iocb: target I/O control block
 833  * @buf: user's buffer from which to write data
 834  * @count: number of bytes to write
 835  * @pos: byte offset in file where writing starts
 836  *
 837  * We use this function for direct writes instead of calling
 838  * generic_file_aio_write() in order to avoid taking the inode
 839  * semaphore and updating the i_size.  The NFS server will set
 840  * the new i_size and this client must read the updated size
 841  * back into its cache.  We let the server do generic write
 842  * parameter checking and report problems.
 843  *
 844  * We also avoid an unnecessary invocation of generic_osync_inode(),
 845  * as it is fairly meaningless to sync the metadata of an NFS file.
 846  *
 847  * We eliminate local atime updates, see direct read above.
 848  *
 849  * We avoid unnecessary page cache invalidations for normal cached
 850  * readers of this file.
 851  *
 852  * Note that O_APPEND is not supported for NFS direct writes, as there
 853  * is no atomic O_APPEND write facility in the NFS protocol.
 854  */
 855 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 856 {
 857         ssize_t retval;
 858         int page_count;
 859         struct page **pages;
 860         struct file *file = iocb->ki_filp;
 861         struct address_space *mapping = file->f_mapping;
 862
 863         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 864                 file->f_dentry->d_parent->d_name.name,
 865                 file->f_dentry->d_name.name,
 866                 (unsigned long) count, (long long) pos);
 867
 868         retval = generic_write_checks(file, &pos, &count, 0);
 869         if (retval)
 870                 goto out;
 871
 872         retval = -EINVAL;
 873         if ((ssize_t) count < 0)
 874                 goto out;
 875         retval = 0;
 876         if (!count)
 877                 goto out;
 878
 879         retval = -EFAULT;
 880         if (!access_ok(VERIFY_READ, buf, count))
 881                 goto out;
 882
 883         retval = nfs_sync_mapping(mapping);
 884         if (retval)
 885                 goto out;
 886
 887         retval = nfs_get_user_pages(WRITE, (unsigned long) buf,
 888                                                 count, &pages);
 889         if (retval < 0)
 890                 goto out;
 891         page_count = retval;
 892
 893         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 894                                         pos, pages, page_count);
 895
 896         /*
 897          * XXX: nfs_end_data_update() already ensures this file's
 898          *      cached data is subsequently invalidated.  Do we really
 899          *      need to call invalidate_inode_pages2() again here?
 900          *
 901          *      For aio writes, this invalidation will almost certainly
 902          *      occur before the writes complete.  Kind of racey.
 903          */
 904         if (mapping->nrpages)
 905                 invalidate_inode_pages2(mapping);
 906
 907         if (retval > 0)
 908                 iocb->ki_pos = pos + retval;
 909
 910 out:
 911         return retval;
 912 }
 913
 914 /**
 915  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 916  *
 917  */
 918 int __init nfs_init_directcache(void)
 919 {
 920         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 921                                                 sizeof(struct nfs_direct_req),
 922                                                 0, (SLAB_RECLAIM_ACCOUNT|
 923                                                         SLAB_MEM_SPREAD),
 924                                                 NULL, NULL);
 925         if (nfs_direct_cachep == NULL)
 926                 return -ENOMEM;
 927
 928         return 0;
 929 }
 930
 931 /**
 932  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 933  *
 934  */
 935 void __exit nfs_destroy_directcache(void)
 936 {
 937         if (kmem_cache_destroy(nfs_direct_cachep))
 938                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 939 }