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         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         atomic_t                io_count;       /* i/os we're waiting for */
  84         spinlock_t              lock;           /* protect completion state */
  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 static inline void get_dreq(struct nfs_direct_req *dreq)
 101 {
 102         atomic_inc(&dreq->io_count);
 103 }
 104
 105 static inline int put_dreq(struct nfs_direct_req *dreq)
 106 {
 107         return atomic_dec_and_test(&dreq->io_count);
 108 }
 109
 110 /**
 111  * nfs_direct_IO - NFS address space operation for direct I/O
 112  * @rw: direction (read or write)
 113  * @iocb: target I/O control block
 114  * @iov: array of vectors that define I/O buffer
 115  * @pos: offset in file to begin the operation
 116  * @nr_segs: size of iovec array
 117  *
 118  * The presence of this routine in the address space ops vector means
 119  * the NFS client supports direct I/O.  However, we shunt off direct
 120  * read and write requests before the VFS gets them, so this method
 121  * should never be called.
 122  */
 123 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 124 {
 125         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 126                         iocb->ki_filp->f_dentry->d_name.name,
 127                         (long long) pos, nr_segs);
 128
 129         return -EINVAL;
 130 }
 131
 132 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
 133 {
 134         int i;
 135         for (i = 0; i < npages; i++) {
 136                 struct page *page = pages[i];
 137                 if (do_dirty && !PageCompound(page))
 138                         set_page_dirty_lock(page);
 139                 page_cache_release(page);
 140         }
 141         kfree(pages);
 142 }
 143
 144 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
 145 {
 146         int result = -ENOMEM;
 147         unsigned long page_count;
 148         size_t array_size;
 149
 150         page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 151         page_count -= user_addr >> PAGE_SHIFT;
 152
 153         array_size = (page_count * sizeof(struct page *));
 154         *pages = kmalloc(array_size, GFP_KERNEL);
 155         if (*pages) {
 156                 down_read(&current->mm->mmap_sem);
 157                 result = get_user_pages(current, current->mm, user_addr,
 158                                         page_count, (rw == READ), 0,
 159                                         *pages, NULL);
 160                 up_read(&current->mm->mmap_sem);
 161                 if (result != page_count) {
 162                         /*
 163                          * If we got fewer pages than expected from
 164                          * get_user_pages(), the user buffer runs off the
 165                          * end of a mapping; return EFAULT.
 166                          */
 167                         if (result >= 0) {
 168                                 nfs_free_user_pages(*pages, result, 0);
 169                                 result = -EFAULT;
 170                         } else
 171                                 kfree(*pages);
 172                         *pages = NULL;
 173                 }
 174         }
 175         return result;
 176 }
 177
 178 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 179 {
 180         struct nfs_direct_req *dreq;
 181
 182         dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 183         if (!dreq)
 184                 return NULL;
 185
 186         kref_init(&dreq->kref);
 187         init_completion(&dreq->completion);
 188         INIT_LIST_HEAD(&dreq->list);
 189         INIT_LIST_HEAD(&dreq->rewrite_list);
 190         dreq->iocb = NULL;
 191         dreq->ctx = NULL;
 192         spin_lock_init(&dreq->lock);
 193         atomic_set(&dreq->io_count, 0);
 194         dreq->count = 0;
 195         dreq->error = 0;
 196         dreq->flags = 0;
 197
 198         return dreq;
 199 }
 200
 201 static void nfs_direct_req_release(struct kref *kref)
 202 {
 203         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 204
 205         if (dreq->ctx != NULL)
 206                 put_nfs_open_context(dreq->ctx);
 207         kmem_cache_free(nfs_direct_cachep, dreq);
 208 }
 209
 210 /*
 211  * Collects and returns the final error value/byte-count.
 212  */
 213 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 214 {
 215         ssize_t result = -EIOCBQUEUED;
 216
 217         /* Async requests don't wait here */
 218         if (dreq->iocb)
 219                 goto out;
 220
 221         result = wait_for_completion_interruptible(&dreq->completion);
 222
 223         if (!result)
 224                 result = dreq->error;
 225         if (!result)
 226                 result = dreq->count;
 227
 228 out:
 229         kref_put(&dreq->kref, nfs_direct_req_release);
 230         return (ssize_t) result;
 231 }
 232
 233 /*
 234  * We must hold a reference to all the pages in this direct read request
 235  * until the RPCs complete.  This could be long *after* we are woken up in
 236  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 237  *
 238  * In addition, synchronous I/O uses a stack-allocated iocb.  Thus we
 239  * can't trust the iocb is still valid here if this is a synchronous
 240  * request.  If the waiter is woken prematurely, the iocb is long gone.
 241  */
 242 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 243 {
 244         nfs_free_user_pages(dreq->pages, dreq->npages, 1);
 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_complete(dreq);
 319 }
 320
 321 static const struct rpc_call_ops nfs_read_direct_ops = {
 322         .rpc_call_done = nfs_direct_read_result,
 323         .rpc_release = nfs_readdata_release,
 324 };
 325
 326 /*
 327  * For each nfs_read_data struct that was allocated on the list, dispatch
 328  * an NFS READ operation
 329  */
 330 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq)
 331 {
 332         struct nfs_open_context *ctx = dreq->ctx;
 333         struct inode *inode = ctx->dentry->d_inode;
 334         struct list_head *list = &dreq->list;
 335         struct page **pages = dreq->pages;
 336         size_t count = dreq->user_count;
 337         loff_t pos = dreq->pos;
 338         size_t rsize = NFS_SERVER(inode)->rsize;
 339         unsigned int curpage, pgbase;
 340
 341         curpage = 0;
 342         pgbase = dreq->user_addr & ~PAGE_MASK;
 343         do {
 344                 struct nfs_read_data *data;
 345                 size_t bytes;
 346
 347                 bytes = rsize;
 348                 if (count < rsize)
 349                         bytes = count;
 350
 351                 BUG_ON(list_empty(list));
 352                 data = list_entry(list->next, struct nfs_read_data, pages);
 353                 list_del_init(&data->pages);
 354
 355                 data->inode = inode;
 356                 data->cred = ctx->cred;
 357                 data->args.fh = NFS_FH(inode);
 358                 data->args.context = ctx;
 359                 data->args.offset = pos;
 360                 data->args.pgbase = pgbase;
 361                 data->args.pages = &pages[curpage];
 362                 data->args.count = bytes;
 363                 data->res.fattr = &data->fattr;
 364                 data->res.eof = 0;
 365                 data->res.count = bytes;
 366
 367                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 368                                 &nfs_read_direct_ops, data);
 369                 NFS_PROTO(inode)->read_setup(data);
 370
 371                 data->task.tk_cookie = (unsigned long) inode;
 372
 373                 lock_kernel();
 374                 rpc_execute(&data->task);
 375                 unlock_kernel();
 376
 377                 dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 378                                 data->task.tk_pid,
 379                                 inode->i_sb->s_id,
 380                                 (long long)NFS_FILEID(inode),
 381                                 bytes,
 382                                 (unsigned long long)data->args.offset);
 383
 384                 pos += bytes;
 385                 pgbase += bytes;
 386                 curpage += pgbase >> PAGE_SHIFT;
 387                 pgbase &= ~PAGE_MASK;
 388
 389                 count -= bytes;
 390         } while (count != 0);
 391         BUG_ON(!list_empty(list));
 392 }
 393
 394 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)
 395 {
 396         ssize_t result;
 397         sigset_t oldset;
 398         struct inode *inode = iocb->ki_filp->f_mapping->host;
 399         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 400         struct nfs_direct_req *dreq;
 401
 402         dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
 403         if (!dreq)
 404                 return -ENOMEM;
 405
 406         dreq->user_addr = user_addr;
 407         dreq->user_count = count;
 408         dreq->pos = pos;
 409         dreq->pages = pages;
 410         dreq->npages = nr_pages;
 411         dreq->inode = inode;
 412         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 413         if (!is_sync_kiocb(iocb))
 414                 dreq->iocb = iocb;
 415
 416         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 417         rpc_clnt_sigmask(clnt, &oldset);
 418         nfs_direct_read_schedule(dreq);
 419         result = nfs_direct_wait(dreq);
 420         rpc_clnt_sigunmask(clnt, &oldset);
 421
 422         return result;
 423 }
 424
 425 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 426 {
 427         list_splice_init(&dreq->rewrite_list, &dreq->list);
 428         while (!list_empty(&dreq->list)) {
 429                 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
 430                 list_del(&data->pages);
 431                 nfs_writedata_release(data);
 432         }
 433 }
 434
 435 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 436 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 437 {
 438         struct list_head *pos;
 439
 440         list_splice_init(&dreq->rewrite_list, &dreq->list);
 441         list_for_each(pos, &dreq->list)
 442                 get_dreq(dreq);
 443         dreq->count = 0;
 444
 445         nfs_direct_write_schedule(dreq, FLUSH_STABLE);
 446 }
 447
 448 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 449 {
 450         struct nfs_write_data *data = calldata;
 451         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 452
 453         /* Call the NFS version-specific code */
 454         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 455                 return;
 456         if (unlikely(task->tk_status < 0)) {
 457                 dreq->error = task->tk_status;
 458                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 459         }
 460         if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 461                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 462                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 463         }
 464
 465         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 466         nfs_direct_write_complete(dreq, data->inode);
 467 }
 468
 469 static const struct rpc_call_ops nfs_commit_direct_ops = {
 470         .rpc_call_done = nfs_direct_commit_result,
 471         .rpc_release = nfs_commit_release,
 472 };
 473
 474 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 475 {
 476         struct nfs_write_data *data = dreq->commit_data;
 477
 478         data->inode = dreq->inode;
 479         data->cred = dreq->ctx->cred;
 480
 481         data->args.fh = NFS_FH(data->inode);
 482         data->args.offset = dreq->pos;
 483         data->args.count = dreq->user_count;
 484         data->res.count = 0;
 485         data->res.fattr = &data->fattr;
 486         data->res.verf = &data->verf;
 487
 488         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 489                                 &nfs_commit_direct_ops, data);
 490         NFS_PROTO(data->inode)->commit_setup(data, 0);
 491
 492         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 493         data->task.tk_cookie = (unsigned long)data->inode;
 494         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 495         dreq->commit_data = NULL;
 496
 497         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 498
 499         lock_kernel();
 500         rpc_execute(&data->task);
 501         unlock_kernel();
 502 }
 503
 504 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 505 {
 506         int flags = dreq->flags;
 507
 508         dreq->flags = 0;
 509         switch (flags) {
 510                 case NFS_ODIRECT_DO_COMMIT:
 511                         nfs_direct_commit_schedule(dreq);
 512                         break;
 513                 case NFS_ODIRECT_RESCHED_WRITES:
 514                         nfs_direct_write_reschedule(dreq);
 515                         break;
 516                 default:
 517                         nfs_end_data_update(inode);
 518                         if (dreq->commit_data != NULL)
 519                                 nfs_commit_free(dreq->commit_data);
 520                         nfs_direct_free_writedata(dreq);
 521                         nfs_direct_complete(dreq);
 522         }
 523 }
 524
 525 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 526 {
 527         dreq->commit_data = nfs_commit_alloc(0);
 528         if (dreq->commit_data != NULL)
 529                 dreq->commit_data->req = (struct nfs_page *) dreq;
 530 }
 531 #else
 532 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 533 {
 534         dreq->commit_data = NULL;
 535 }
 536
 537 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 538 {
 539         nfs_end_data_update(inode);
 540         nfs_direct_free_writedata(dreq);
 541         nfs_direct_complete(dreq);
 542 }
 543 #endif
 544
 545 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
 546 {
 547         struct list_head *list;
 548         struct nfs_direct_req *dreq;
 549         unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 550
 551         dreq = nfs_direct_req_alloc();
 552         if (!dreq)
 553                 return NULL;
 554
 555         list = &dreq->list;
 556         for(;;) {
 557                 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
 558
 559                 if (unlikely(!data)) {
 560                         while (!list_empty(list)) {
 561                                 data = list_entry(list->next,
 562                                                   struct nfs_write_data, pages);
 563                                 list_del(&data->pages);
 564                                 nfs_writedata_free(data);
 565                         }
 566                         kref_put(&dreq->kref, nfs_direct_req_release);
 567                         return NULL;
 568                 }
 569
 570                 INIT_LIST_HEAD(&data->pages);
 571                 list_add(&data->pages, list);
 572
 573                 data->req = (struct nfs_page *) dreq;
 574                 get_dreq(dreq);
 575                 if (nbytes <= wsize)
 576                         break;
 577                 nbytes -= wsize;
 578         }
 579
 580         nfs_alloc_commit_data(dreq);
 581
 582         kref_get(&dreq->kref);
 583         return dreq;
 584 }
 585
 586 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 587 {
 588         struct nfs_write_data *data = calldata;
 589         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 590         int status = task->tk_status;
 591
 592         if (nfs_writeback_done(task, data) != 0)
 593                 return;
 594
 595         spin_lock(&dreq->lock);
 596
 597         if (likely(status >= 0))
 598                 dreq->count += data->res.count;
 599         else
 600                 dreq->error = task->tk_status;
 601
 602         if (data->res.verf->committed != NFS_FILE_SYNC) {
 603                 switch (dreq->flags) {
 604                         case 0:
 605                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 606                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 607                                 break;
 608                         case NFS_ODIRECT_DO_COMMIT:
 609                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 610                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 611                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 612                                 }
 613                 }
 614         }
 615         /* In case we have to resend */
 616         data->args.stable = NFS_FILE_SYNC;
 617
 618         spin_unlock(&dreq->lock);
 619 }
 620
 621 /*
 622  * NB: Return the value of the first error return code.  Subsequent
 623  *     errors after the first one are ignored.
 624  */
 625 static void nfs_direct_write_release(void *calldata)
 626 {
 627         struct nfs_write_data *data = calldata;
 628         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 629
 630         if (put_dreq(dreq))
 631                 nfs_direct_write_complete(dreq, data->inode);
 632 }
 633
 634 static const struct rpc_call_ops nfs_write_direct_ops = {
 635         .rpc_call_done = nfs_direct_write_result,
 636         .rpc_release = nfs_direct_write_release,
 637 };
 638
 639 /*
 640  * For each nfs_write_data struct that was allocated on the list, dispatch
 641  * an NFS WRITE operation
 642  */
 643 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync)
 644 {
 645         struct nfs_open_context *ctx = dreq->ctx;
 646         struct inode *inode = ctx->dentry->d_inode;
 647         struct list_head *list = &dreq->list;
 648         struct page **pages = dreq->pages;
 649         size_t count = dreq->user_count;
 650         loff_t pos = dreq->pos;
 651         size_t wsize = NFS_SERVER(inode)->wsize;
 652         unsigned int curpage, pgbase;
 653
 654         curpage = 0;
 655         pgbase = dreq->user_addr & ~PAGE_MASK;
 656         do {
 657                 struct nfs_write_data *data;
 658                 size_t bytes;
 659
 660                 bytes = wsize;
 661                 if (count < wsize)
 662                         bytes = count;
 663
 664                 BUG_ON(list_empty(list));
 665                 data = list_entry(list->next, struct nfs_write_data, pages);
 666                 list_move_tail(&data->pages, &dreq->rewrite_list);
 667
 668                 data->inode = inode;
 669                 data->cred = ctx->cred;
 670                 data->args.fh = NFS_FH(inode);
 671                 data->args.context = ctx;
 672                 data->args.offset = pos;
 673                 data->args.pgbase = pgbase;
 674                 data->args.pages = &pages[curpage];
 675                 data->args.count = bytes;
 676                 data->res.fattr = &data->fattr;
 677                 data->res.count = bytes;
 678                 data->res.verf = &data->verf;
 679
 680                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 681                                 &nfs_write_direct_ops, data);
 682                 NFS_PROTO(inode)->write_setup(data, sync);
 683
 684                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 685                 data->task.tk_cookie = (unsigned long) inode;
 686
 687                 lock_kernel();
 688                 rpc_execute(&data->task);
 689                 unlock_kernel();
 690
 691                 dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
 692                                 data->task.tk_pid,
 693                                 inode->i_sb->s_id,
 694                                 (long long)NFS_FILEID(inode),
 695                                 bytes,
 696                                 (unsigned long long)data->args.offset);
 697
 698                 pos += bytes;
 699                 pgbase += bytes;
 700                 curpage += pgbase >> PAGE_SHIFT;
 701                 pgbase &= ~PAGE_MASK;
 702
 703                 count -= bytes;
 704         } while (count != 0);
 705         BUG_ON(!list_empty(list));
 706 }
 707
 708 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)
 709 {
 710         ssize_t result;
 711         sigset_t oldset;
 712         struct inode *inode = iocb->ki_filp->f_mapping->host;
 713         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 714         struct nfs_direct_req *dreq;
 715         size_t wsize = NFS_SERVER(inode)->wsize;
 716         int sync = 0;
 717
 718         dreq = nfs_direct_write_alloc(count, wsize);
 719         if (!dreq)
 720                 return -ENOMEM;
 721         if (dreq->commit_data == NULL || count < wsize)
 722                 sync = FLUSH_STABLE;
 723
 724         dreq->user_addr = user_addr;
 725         dreq->user_count = count;
 726         dreq->pos = pos;
 727         dreq->pages = pages;
 728         dreq->npages = nr_pages;
 729         dreq->inode = inode;
 730         dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
 731         if (!is_sync_kiocb(iocb))
 732                 dreq->iocb = iocb;
 733
 734         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 735
 736         nfs_begin_data_update(inode);
 737
 738         rpc_clnt_sigmask(clnt, &oldset);
 739         nfs_direct_write_schedule(dreq, sync);
 740         result = nfs_direct_wait(dreq);
 741         rpc_clnt_sigunmask(clnt, &oldset);
 742
 743         return result;
 744 }
 745
 746 /**
 747  * nfs_file_direct_read - file direct read operation for NFS files
 748  * @iocb: target I/O control block
 749  * @buf: user's buffer into which to read data
 750  * @count: number of bytes to read
 751  * @pos: byte offset in file where reading starts
 752  *
 753  * We use this function for direct reads instead of calling
 754  * generic_file_aio_read() in order to avoid gfar's check to see if
 755  * the request starts before the end of the file.  For that check
 756  * to work, we must generate a GETATTR before each direct read, and
 757  * even then there is a window between the GETATTR and the subsequent
 758  * READ where the file size could change.  Our preference is simply
 759  * to do all reads the application wants, and the server will take
 760  * care of managing the end of file boundary.
 761  *
 762  * This function also eliminates unnecessarily updating the file's
 763  * atime locally, as the NFS server sets the file's atime, and this
 764  * client must read the updated atime from the server back into its
 765  * cache.
 766  */
 767 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
 768 {
 769         ssize_t retval = -EINVAL;
 770         int page_count;
 771         struct page **pages;
 772         struct file *file = iocb->ki_filp;
 773         struct address_space *mapping = file->f_mapping;
 774
 775         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 776                 file->f_dentry->d_parent->d_name.name,
 777                 file->f_dentry->d_name.name,
 778                 (unsigned long) count, (long long) pos);
 779
 780         if (count < 0)
 781                 goto out;
 782         retval = -EFAULT;
 783         if (!access_ok(VERIFY_WRITE, buf, count))
 784                 goto out;
 785         retval = 0;
 786         if (!count)
 787                 goto out;
 788
 789         retval = nfs_sync_mapping(mapping);
 790         if (retval)
 791                 goto out;
 792
 793         retval = nfs_get_user_pages(READ, (unsigned long) buf,
 794                                                 count, &pages);
 795         if (retval < 0)
 796                 goto out;
 797         page_count = retval;
 798
 799         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
 800                                                 pages, page_count);
 801         if (retval > 0)
 802                 iocb->ki_pos = pos + retval;
 803
 804 out:
 805         return retval;
 806 }
 807
 808 /**
 809  * nfs_file_direct_write - file direct write operation for NFS files
 810  * @iocb: target I/O control block
 811  * @buf: user's buffer from which to write data
 812  * @count: number of bytes to write
 813  * @pos: byte offset in file where writing starts
 814  *
 815  * We use this function for direct writes instead of calling
 816  * generic_file_aio_write() in order to avoid taking the inode
 817  * semaphore and updating the i_size.  The NFS server will set
 818  * the new i_size and this client must read the updated size
 819  * back into its cache.  We let the server do generic write
 820  * parameter checking and report problems.
 821  *
 822  * We also avoid an unnecessary invocation of generic_osync_inode(),
 823  * as it is fairly meaningless to sync the metadata of an NFS file.
 824  *
 825  * We eliminate local atime updates, see direct read above.
 826  *
 827  * We avoid unnecessary page cache invalidations for normal cached
 828  * readers of this file.
 829  *
 830  * Note that O_APPEND is not supported for NFS direct writes, as there
 831  * is no atomic O_APPEND write facility in the NFS protocol.
 832  */
 833 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
 834 {
 835         ssize_t retval;
 836         int page_count;
 837         struct page **pages;
 838         struct file *file = iocb->ki_filp;
 839         struct address_space *mapping = file->f_mapping;
 840
 841         dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
 842                 file->f_dentry->d_parent->d_name.name,
 843                 file->f_dentry->d_name.name,
 844                 (unsigned long) count, (long long) pos);
 845
 846         retval = generic_write_checks(file, &pos, &count, 0);
 847         if (retval)
 848                 goto out;
 849
 850         retval = -EINVAL;
 851         if ((ssize_t) count < 0)
 852                 goto out;
 853         retval = 0;
 854         if (!count)
 855                 goto out;
 856
 857         retval = -EFAULT;
 858         if (!access_ok(VERIFY_READ, buf, count))
 859                 goto out;
 860
 861         retval = nfs_sync_mapping(mapping);
 862         if (retval)
 863                 goto out;
 864
 865         retval = nfs_get_user_pages(WRITE, (unsigned long) buf,
 866                                                 count, &pages);
 867         if (retval < 0)
 868                 goto out;
 869         page_count = retval;
 870
 871         retval = nfs_direct_write(iocb, (unsigned long) buf, count,
 872                                         pos, pages, page_count);
 873
 874         /*
 875          * XXX: nfs_end_data_update() already ensures this file's
 876          *      cached data is subsequently invalidated.  Do we really
 877          *      need to call invalidate_inode_pages2() again here?
 878          *
 879          *      For aio writes, this invalidation will almost certainly
 880          *      occur before the writes complete.  Kind of racey.
 881          */
 882         if (mapping->nrpages)
 883                 invalidate_inode_pages2(mapping);
 884
 885         if (retval > 0)
 886                 iocb->ki_pos = pos + retval;
 887
 888 out:
 889         return retval;
 890 }
 891
 892 /**
 893  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 894  *
 895  */
 896 int __init nfs_init_directcache(void)
 897 {
 898         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 899                                                 sizeof(struct nfs_direct_req),
 900                                                 0, (SLAB_RECLAIM_ACCOUNT|
 901                                                         SLAB_MEM_SPREAD),
 902                                                 NULL, NULL);
 903         if (nfs_direct_cachep == NULL)
 904                 return -ENOMEM;
 905
 906         return 0;
 907 }
 908
 909 /**
 910  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 911  *
 912  */
 913 void __exit nfs_destroy_directcache(void)
 914 {
 915         if (kmem_cache_destroy(nfs_direct_cachep))
 916                 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
 917 }