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