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_direct_dirty_pages(struct page **pages, int npages)
 130 {
 131         int i;
 132         for (i = 0; i < npages; i++) {
 133                 struct page *page = pages[i];
 134                 if (!PageCompound(page))
 135                         set_page_dirty_lock(page);
 136         }
 137 }
 138
 139 static void nfs_direct_release_pages(struct page **pages, int npages)
 140 {
 141         int i;
 142         for (i = 0; i < npages; i++)
 143                 page_cache_release(pages[i]);
 144 }
 145
 146 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 147 {
 148         int result = -ENOMEM;
 149         unsigned long page_count;
 150         size_t array_size;
 151
 152         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 153         page_count -= user_addr >> PAGE_SHIFT;
 154
 155         array_size = (page_count * sizeof(struct page *));
 156         *pages = kmalloc(array_size, GFP_KERNEL);
 157         if (*pages) {
 158                 down_read(&current->mm->mmap_sem);
 159                 result = get_user_pages(current, current->mm, user_addr,
 160                                         page_count, (rw == READ), 0,
 161                                         *pages, NULL);
 162                 up_read(&current->mm->mmap_sem);
 163                 if (result != page_count) {
 164                         /*
 165                          * If we got fewer pages than expected from
 166                          * get_user_pages(), the user buffer runs off the
 167                          * end of a mapping; return EFAULT.
 168                          */
 169                         if (result >= 0) {
 170                                 nfs_direct_release_pages(*pages, result);
 171                                 result = -EFAULT;
 172                         } else
 173                                 kfree(*pages);
 174                         *pages = NULL;
 175                 }
 176         }
 177         return result;
 178 }
 179
 180 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 181 {
 182         struct nfs_direct_req *dreq;
 183
 184         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 185         if (!dreq)
 186                 return NULL;
 187
 188         kref_init(&dreq->kref);
 189         init_completion(&dreq->completion);
 190         INIT_LIST_HEAD(&dreq->list);
 191         INIT_LIST_HEAD(&dreq->rewrite_list);
 192         dreq->iocb = NULL;
 193         dreq->ctx = NULL;
 194         spin_lock_init(&dreq->lock);
 195         atomic_set(&dreq->io_count, 0);
 196         dreq->count = 0;
 197         dreq->error = 0;
 198         dreq->flags = 0;
 199
 200         return dreq;
 201 }
 202
 203 static void nfs_direct_req_release(struct kref *kref)
 204 {
 205         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 206
 207         if (dreq->ctx != NULL)
 208                 put_nfs_open_context(dreq->ctx);
 209         kmem_cache_free(nfs_direct_cachep, dreq);
 210 }
 211
 212 /*
 213  * Collects and returns the final error value/byte-count.
 214  */
 215 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 216 {
 217         ssize_t result = -EIOCBQUEUED;
 218
 219         /* Async requests don't wait here */
 220         if (dreq->iocb)
 221                 goto out;
 222
 223         result = wait_for_completion_interruptible(&dreq->completion);
 224
 225         if (!result)
 226                 result = dreq->error;
 227         if (!result)
 228                 result = dreq->count;
 229
 230 out:
 231         kref_put(&dreq->kref, nfs_direct_req_release);
 232         return (ssize_t) result;
 233 }
 234
 235 /*
 236  * We must hold a reference to all the pages in this direct read request
 237  * until the RPCs complete.  This could be long *after* we are woken up in
 238  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 239  *
 240  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 241  * can't trust the iocb is still valid here if this is a synchronous
 242  * request.  If the waiter is woken prematurely, the iocb is long gone.
 243  */
 244 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 245 {
 246         if (dreq->iocb) {
 247                 long res = (long) dreq->error;
 248                 if (!res)
 249                         res = (long) dreq->count;
 250                 aio_complete(dreq->iocb, res, 0);
 251         }
 252         complete_all(&dreq->completion);
 253
 254         kref_put(&dreq->kref, nfs_direct_req_release);
 255 }
 256
 257 /*
 258  * Note we also set the number of requests we have in the dreq when we are
 259  * done.  This prevents races with I/O completion so we will always wait
 260  * until all requests have been dispatched and completed.
 261  */
 262 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
 263 {
 264         struct list_head *list;
 265         struct nfs_direct_req *dreq;
 266         unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 267
 268         dreq = nfs_direct_req_alloc();
 269         if (!dreq)
 270                 return NULL;
 271
 272         list = &dreq->list;
 273         for(;;) {
 274                 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 275
 276                 if (unlikely(!data)) {
 277                         while (!list_empty(list)) {
 278                                 data = list_entry(list->next,
 279                                                   struct nfs_read_data, pages);
 280                                 list_del(&data->pages);
 281                                 nfs_readdata_free(data);
 282                         }
 283                         kref_put(&dreq->kref, nfs_direct_req_release);
 284                         return NULL;
 285                 }
 286
 287                 INIT_LIST_HEAD(&data->pages);
 288                 list_add(&data->pages, list);
 289
 290                 data->req = (struct nfs_page *) dreq;
 291                 get_dreq(dreq);
 292                 if (nbytes <= rsize)
 293                         break;
 294                 nbytes -= rsize;
 295         }
 296         kref_get(&dreq->kref);
 297         return dreq;
 298 }
 299
 300 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 301 {
 302         struct nfs_read_data *data = calldata;
 303         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 304
 305         if (nfs_readpage_result(task, data) != 0)
 306                 return;
 307
 308         spin_lock(&dreq->lock);
 309
 310         if (likely(task->tk_status >= 0))
 311                 dreq->count += data->res.count;
 312         else
 313                 dreq->error = task->tk_status;
 314
 315         spin_unlock(&dreq->lock);
 316
 317         if (put_dreq(dreq)) {
 318                 nfs_direct_dirty_pages(dreq->pages, dreq->npages);
 319                 nfs_direct_release_pages(dreq->pages, dreq->npages);
 320                 nfs_direct_complete(dreq);
 321         }
 322 }
 323
 324 static const struct rpc_call_ops nfs_read_direct_ops = {
 325         .rpc_call_done = nfs_direct_read_result,
 326         .rpc_release = nfs_readdata_release,
 327 };
 328
 329 /*
 330  * For each nfs_read_data struct that was allocated on the list, dispatch
 331  * an NFS READ operation
 332  */
 333 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 334 {
 335         struct nfs_open_context *ctx = dreq->ctx;
 336         struct inode *inode = ctx->dentry->d_inode;
 337         struct list_head *list = &dreq->list;
 338         struct page **pages = dreq->pages;
 339         size_t rsize = NFS_SERVER(inode)->rsize;
 340         unsigned int curpage, pgbase;
 341
 342         curpage = 0;
 343         pgbase = user_addr & ~PAGE_MASK;
 344         do {
 345                 struct nfs_read_data *data;
 346                 size_t bytes;
 347
 348                 bytes = rsize;
 349                 if (count < rsize)
 350                         bytes = count;
 351
 352                 BUG_ON(list_empty(list));
 353                 data = list_entry(list->next, struct nfs_read_data, pages);
 354                 list_del_init(&data->pages);
 355
 356                 data->inode = inode;
 357                 data->cred = ctx->cred;
 358                 data->args.fh = NFS_FH(inode);
 359                 data->args.context = ctx;
 360                 data->args.offset = pos;
 361                 data->args.pgbase = pgbase;
 362                 data->args.pages = &pages[curpage];
 363                 data->args.count = bytes;
 364                 data->res.fattr = &data->fattr;
 365                 data->res.eof = 0;
 366                 data->res.count = bytes;
 367
 368                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 369                                 &nfs_read_direct_ops, data);
 370                 NFS_PROTO(inode)->read_setup(data);
 371
 372                 data->task.tk_cookie = (unsigned long) inode;
 373
 374                 lock_kernel();
 375                 rpc_execute(&data->task);
 376                 unlock_kernel();
 377
 378                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 379                                 data->task.tk_pid,
 380                                 inode->i_sb->s_id,
 381                                 (long long)NFS_FILEID(inode),
 382                                 bytes,
 383                                 (unsigned long long)data->args.offset);
 384
 385                 pos += bytes;
 386                 pgbase += bytes;
 387                 curpage += pgbase >> PAGE_SHIFT;
 388                 pgbase &= ~PAGE_MASK;
 389
 390                 count -= bytes;
 391         } while (count != 0);
 392         BUG_ON(!list_empty(list));
 393 }
 394
 395 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)
 396 {
 397         ssize_t result;
 398         sigset_t oldset;
 399         struct inode *inode = iocb->ki_filp->f_mapping->host;
 400         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 401         struct nfs_direct_req *dreq;
 402
 403         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 404         if (!dreq)
 405                 return -ENOMEM;
 406
 407         dreq->pages = pages;
 408         dreq->npages = nr_pages;
 409         dreq->inode = inode;
 410         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 411         if (!is_sync_kiocb(iocb))
 412                 dreq->iocb = iocb;
 413
 414         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 415         rpc_clnt_sigmask(clnt, &oldset);
 416         nfs_direct_read_schedule(dreq, user_addr, count, pos);
 417         result = nfs_direct_wait(dreq);
 418         rpc_clnt_sigunmask(clnt, &oldset);
 419
 420         return result;
 421 }
 422
 423 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 424 {
 425         list_splice_init(&dreq->rewrite_list, &dreq->list);
 426         while (!list_empty(&dreq->list)) {
 427                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 428                 list_del(&data->pages);
 429                 nfs_writedata_release(data);
 430         }
 431         nfs_direct_release_pages(dreq->pages, dreq->npages);
 432 }
 433
 434 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 435 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 436 {
 437         struct inode *inode = dreq->inode;
 438         struct list_head *p;
 439         struct nfs_write_data *data;
 440
 441         dreq->count = 0;
 442         get_dreq(dreq);
 443
 444         list_for_each(p, &dreq->rewrite_list) {
 445                 data = list_entry(p, struct nfs_write_data, pages);
 446
 447                 get_dreq(dreq);
 448
 449                 /*
 450                  * Reset data->res.
 451                  */
 452                 nfs_fattr_init(&data->fattr);
 453                 data->res.count = data->args.count;
 454                 memset(&data->verf, 0, sizeof(data->verf));
 455
 456                 /*
 457                  * Reuse data->task; data->args should not have changed
 458                  * since the original request was sent.
 459                  */
 460                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 461                                 &nfs_write_direct_ops, data);
 462                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 463
 464                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 465                 data->task.tk_cookie = (unsigned long) inode;
 466
 467                 /*
 468                  * We're called via an RPC callback, so BKL is already held.
 469                  */
 470                 rpc_execute(&data->task);
 471
 472                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 473                                 data->task.tk_pid,
 474                                 inode->i_sb->s_id,
 475                                 (long long)NFS_FILEID(inode),
 476                                 data->args.count,
 477                                 (unsigned long long)data->args.offset);
 478         }
 479
 480         if (put_dreq(dreq))
 481                 nfs_direct_write_complete(dreq, inode);
 482 }
 483
 484 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 485 {
 486         struct nfs_write_data *data = calldata;
 487         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 488
 489         /* Call the NFS version-specific code */
 490         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 491                 return;
 492         if (unlikely(task->tk_status < 0)) {
 493                 dreq->error = task->tk_status;
 494                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 495         }
 496         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 497                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 498                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 499         }
 500
 501         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 502         nfs_direct_write_complete(dreq, data->inode);
 503 }
 504
 505 static const struct rpc_call_ops nfs_commit_direct_ops = {
 506         .rpc_call_done = nfs_direct_commit_result,
 507         .rpc_release = nfs_commit_release,
 508 };
 509
 510 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 511 {
 512         struct nfs_write_data *data = dreq->commit_data;
 513
 514         data->inode = dreq->inode;
 515         data->cred = dreq->ctx->cred;
 516
 517         data->args.fh = NFS_FH(data->inode);
 518         data->args.offset = 0;
 519         data->args.count = 0;
 520         data->res.count = 0;
 521         data->res.fattr = &data->fattr;
 522         data->res.verf = &data->verf;
 523
 524         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 525                                 &nfs_commit_direct_ops, data);
 526         NFS_PROTO(data->inode)->commit_setup(data, 0);
 527
 528         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 529         data->task.tk_cookie = (unsigned long)data->inode;
 530         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 531         dreq->commit_data = NULL;
 532
 533         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 534
 535         lock_kernel();
 536         rpc_execute(&data->task);
 537         unlock_kernel();
 538 }
 539
 540 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 541 {
 542         int flags = dreq->flags;
 543
 544         dreq->flags = 0;
 545         switch (flags) {
 546                 case NFS_ODIRECT_DO_COMMIT:
 547                         nfs_direct_commit_schedule(dreq);
 548                         break;
 549                 case NFS_ODIRECT_RESCHED_WRITES:
 550                         nfs_direct_write_reschedule(dreq);
 551                         break;
 552                 default:
 553                         nfs_end_data_update(inode);
 554                         if (dreq->commit_data != NULL)
 555                                 nfs_commit_free(dreq->commit_data);
 556                         nfs_direct_free_writedata(dreq);
 557                         nfs_direct_complete(dreq);
 558         }
 559 }
 560
 561 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 562 {
 563         dreq->commit_data = nfs_commit_alloc(0);
 564         if (dreq->commit_data != NULL)
 565                 dreq->commit_data->req = (struct nfs_page *) dreq;
 566 }
 567 #else
 568 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 569 {
 570         dreq->commit_data = NULL;
 571 }
 572
 573 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 574 {
 575         nfs_end_data_update(inode);
 576         nfs_direct_free_writedata(dreq);
 577         nfs_direct_complete(dreq);
 578 }
 579 #endif
 580
 581 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 582 {
 583         struct list_head *list;
 584         struct nfs_direct_req *dreq;
 585         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 586
 587         dreq = nfs_direct_req_alloc();
 588         if (!dreq)
 589                 return NULL;
 590
 591         list = &dreq->list;
 592         for(;;) {
 593                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 594
 595                 if (unlikely(!data)) {
 596                         while (!list_empty(list)) {
 597                                 data = list_entry(list->next,
 598                                                   struct nfs_write_data, pages);
 599                                 list_del(&data->pages);
 600                                 nfs_writedata_free(data);
 601                         }
 602                         kref_put(&dreq->kref, nfs_direct_req_release);
 603                         return NULL;
 604                 }
 605
 606                 INIT_LIST_HEAD(&data->pages);
 607                 list_add(&data->pages, list);
 608
 609                 data->req = (struct nfs_page *) dreq;
 610                 get_dreq(dreq);
 611                 if (nbytes <= wsize)
 612                         break;
 613                 nbytes -= wsize;
 614         }
 615
 616         nfs_alloc_commit_data(dreq);
 617
 618         kref_get(&dreq->kref);
 619         return dreq;
 620 }
 621
 622 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 623 {
 624         struct nfs_write_data *data = calldata;
 625         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 626         int status = task->tk_status;
 627
 628         if (nfs_writeback_done(task, data) != 0)
 629                 return;
 630
 631         spin_lock(&dreq->lock);
 632
 633         if (likely(status >= 0))
 634                 dreq->count += data->res.count;
 635         else
 636                 dreq->error = task->tk_status;
 637
 638         if (data->res.verf->committed != NFS_FILE_SYNC) {
 639                 switch (dreq->flags) {
 640                         case 0:
 641                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 642                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 643                                 break;
 644                         case NFS_ODIRECT_DO_COMMIT:
 645                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 646                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 647                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 648                                 }
 649                 }
 650         }
 651
 652         spin_unlock(&dreq->lock);
 653 }
 654
 655 /*
 656  * NB: Return the value of the first error return code.  Subsequent
 657  *     errors after the first one are ignored.
 658  */
 659 static void nfs_direct_write_release(void *calldata)
 660 {
 661         struct nfs_write_data *data = calldata;
 662         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 663
 664         if (put_dreq(dreq))
 665                 nfs_direct_write_complete(dreq, data->inode);
 666 }
 667
 668 static const struct rpc_call_ops nfs_write_direct_ops = {
 669         .rpc_call_done = nfs_direct_write_result,
 670         .rpc_release = nfs_direct_write_release,
 671 };
 672
 673 /*
 674  * For each nfs_write_data struct that was allocated on the list, dispatch
 675  * an NFS WRITE operation
 676  */
 677 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 678 {
 679         struct nfs_open_context *ctx = dreq->ctx;
 680         struct inode *inode = ctx->dentry->d_inode;
 681         struct list_head *list = &dreq->list;
 682         struct page **pages = dreq->pages;
 683         size_t wsize = NFS_SERVER(inode)->wsize;
 684         unsigned int curpage, pgbase;
 685
 686         curpage = 0;
 687         pgbase = user_addr & ~PAGE_MASK;
 688         do {
 689                 struct nfs_write_data *data;
 690                 size_t bytes;
 691
 692                 bytes = wsize;
 693                 if (count < wsize)
 694                         bytes = count;
 695
 696                 BUG_ON(list_empty(list));
 697                 data = list_entry(list->next, struct nfs_write_data, pages);
 698                 list_move_tail(&data->pages, &dreq->rewrite_list);
 699
 700                 data->inode = inode;
 701                 data->cred = ctx->cred;
 702                 data->args.fh = NFS_FH(inode);
 703                 data->args.context = ctx;
 704                 data->args.offset = pos;
 705                 data->args.pgbase = pgbase;
 706                 data->args.pages = &pages[curpage];
 707                 data->args.count = bytes;
 708                 data->res.fattr = &data->fattr;
 709                 data->res.count = bytes;
 710                 data->res.verf = &data->verf;
 711
 712                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 713                                 &nfs_write_direct_ops, data);
 714                 NFS_PROTO(inode)->write_setup(data, sync);
 715
 716                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 717                 data->task.tk_cookie = (unsigned long) inode;
 718
 719                 lock_kernel();
 720                 rpc_execute(&data->task);
 721                 unlock_kernel();
 722
 723                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 724                                 data->task.tk_pid,
 725                                 inode->i_sb->s_id,
 726                                 (long long)NFS_FILEID(inode),
 727                                 bytes,
 728                                 (unsigned long long)data->args.offset);
 729
 730                 pos += bytes;
 731                 pgbase += bytes;
 732                 curpage += pgbase >> PAGE_SHIFT;
 733                 pgbase &= ~PAGE_MASK;
 734
 735                 count -= bytes;
 736         } while (count != 0);
 737         BUG_ON(!list_empty(list));
 738 }
 739
 740 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)
 741 {
 742         ssize_t result;
 743         sigset_t oldset;
 744         struct inode *inode = iocb->ki_filp->f_mapping->host;
 745         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 746         struct nfs_direct_req *dreq;
 747         size_t wsize = NFS_SERVER(inode)->wsize;
 748         int sync = 0;
 749
 750         dreq = nfs_direct_write_alloc(count, wsize);
 751         if (!dreq)
 752                 return -ENOMEM;
 753         if (dreq->commit_data == NULL || count < wsize)
 754                 sync = FLUSH_STABLE;
 755
 756         dreq->pages = pages;
 757         dreq->npages = nr_pages;
 758         dreq->inode = inode;
 759         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 760         if (!is_sync_kiocb(iocb))
 761                 dreq->iocb = iocb;
 762
 763         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 764
 765         nfs_begin_data_update(inode);
 766
 767         rpc_clnt_sigmask(clnt, &oldset);
 768         nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 769         result = nfs_direct_wait(dreq);
 770         rpc_clnt_sigunmask(clnt, &oldset);
 771
 772         return result;
 773 }
 774
 775 /**
 776  * nfs_file_direct_read - file direct read operation for NFS files
 777  * @iocb: target I/O control block
 778  * @buf: user's buffer into which to read data
 779  * @count: number of bytes to read
 780  * @pos: byte offset in file where reading starts
 781  *
 782  * We use this function for direct reads instead of calling
 783  * generic_file_aio_read() in order to avoid gfar's check to see if
 784  * the request starts before the end of the file.  For that check
 785  * to work, we must generate a GETATTR before each direct read, and
 786  * even then there is a window between the GETATTR and the subsequent
 787  * READ where the file size could change.  Our preference is simply
 788  * to do all reads the application wants, and the server will take
 789  * care of managing the end of file boundary.
 790  *
 791  * This function also eliminates unnecessarily updating the file's
 792  * atime locally, as the NFS server sets the file's atime, and this
 793  * client must read the updated atime from the server back into its
 794  * cache.
 795  */
 796 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 797 {
 798         ssize_t retval = -EINVAL;
 799         int page_count;
 800         struct page **pages;
 801         struct file *file = iocb->ki_filp;
 802         struct address_space *mapping = file->f_mapping;
 803
 804         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 805                 file->f_dentry->d_parent->d_name.name,
 806                 file->f_dentry->d_name.name,
 807                 (unsigned long) count, (long long) pos);
 808
 809         if (count < 0)
 810                 goto out;
 811         retval = -EFAULT;
 812         if (!access_ok(VERIFY_WRITE, buf, count))
 813                 goto out;
 814         retval = 0;
 815         if (!count)
 816                 goto out;
 817
 818         retval = nfs_sync_mapping(mapping);
 819         if (retval)
 820                 goto out;
 821
 822         retval = nfs_get_user_pages(READ, (unsigned long) buf,
 823                                                 count, &pages);
 824         if (retval < 0)
 825                 goto out;
 826         page_count = retval;
 827
 828         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 829                                                 pages, page_count);
 830         if (retval > 0)
 831                 iocb->ki_pos = pos + retval;
 832
 833 out:
 834         return retval;
 835 }
 836
 837 /**
 838  * nfs_file_direct_write - file direct write operation for NFS files
 839  * @iocb: target I/O control block
 840  * @buf: user's buffer from which to write data
 841  * @count: number of bytes to write
 842  * @pos: byte offset in file where writing starts
 843  *
 844  * We use this function for direct writes instead of calling
 845  * generic_file_aio_write() in order to avoid taking the inode
 846  * semaphore and updating the i_size.  The NFS server will set
 847  * the new i_size and this client must read the updated size
 848  * back into its cache.  We let the server do generic write
 849  * parameter checking and report problems.
 850  *
 851  * We also avoid an unnecessary invocation of generic_osync_inode(),
 852  * as it is fairly meaningless to sync the metadata of an NFS file.
 853  *
 854  * We eliminate local atime updates, see direct read above.
 855  *
 856  * We avoid unnecessary page cache invalidations for normal cached
 857  * readers of this file.
 858  *
 859  * Note that O_APPEND is not supported for NFS direct writes, as there
 860  * is no atomic O_APPEND write facility in the NFS protocol.
 861  */
 862 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 863 {
 864         ssize_t retval;
 865         int page_count;
 866         struct page **pages;
 867         struct file *file = iocb->ki_filp;
 868         struct address_space *mapping = file->f_mapping;
 869
 870         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 871                 file->f_dentry->d_parent->d_name.name,
 872                 file->f_dentry->d_name.name,
 873                 (unsigned long) count, (long long) pos);
 874
 875         retval = generic_write_checks(file, &pos, &count, 0);
 876         if (retval)
 877                 goto out;
 878
 879         retval = -EINVAL;
 880         if ((ssize_t) count < 0)
 881                 goto out;
 882         retval = 0;
 883         if (!count)
 884                 goto out;
 885
 886         retval = -EFAULT;
 887         if (!access_ok(VERIFY_READ, buf, count))
 888                 goto out;
 889
 890         retval = nfs_sync_mapping(mapping);
 891         if (retval)
 892                 goto out;
 893
 894         retval = nfs_get_user_pages(WRITE, (unsigned long) buf,
 895                                                 count, &pages);
 896         if (retval < 0)
 897                 goto out;
 898         page_count = retval;
 899
 900         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 901                                         pos, pages, page_count);
 902
 903         /*
 904          * XXX: nfs_end_data_update() already ensures this file's
 905          *      cached data is subsequently invalidated.  Do we really
 906          *      need to call invalidate_inode_pages2() again here?
 907          *
 908          *      For aio writes, this invalidation will almost certainly
 909          *      occur before the writes complete.  Kind of racey.
 910          */
 911         if (mapping->nrpages)
 912                 invalidate_inode_pages2(mapping);
 913
 914         if (retval > 0)
 915                 iocb->ki_pos = pos + retval;
 916
 917 out:
 918         return retval;
 919 }
 920
 921 /**
 922  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 923  *
 924  */
 925 int __init nfs_init_directcache(void)
 926 {
 927         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 928                                                 sizeof(struct nfs_direct_req),
 929                                                 0, (SLAB_RECLAIM_ACCOUNT|
 930                                                         SLAB_MEM_SPREAD),
 931                                                 NULL, NULL);
 932         if (nfs_direct_cachep == NULL)
 933                 return -ENOMEM;
 934
 935         return 0;
 936 }
 937
 938 /**
 939  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 940  *
 941  */
 942 void __exit nfs_destroy_directcache(void)
 943 {
 944         if (kmem_cache_destroy(nfs_direct_cachep))
 945                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 946 }