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/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47
  48 #include <linux/nfs_fs.h>
  49 #include <linux/nfs_page.h>
  50 #include <linux/sunrpc/clnt.h>
  51
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54 #include <asm/atomic.h>
  55
  56 #include "internal.h"
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60
  61 static struct kmem_cache *nfs_direct_cachep;
  62
  63 /*
  64  * This represents a set of asynchronous requests that we're waiting on
  65  */
  66 struct nfs_direct_req {
  67         struct kref             kref;           /* release manager */
  68
  69         /* I/O parameters */
  70         struct nfs_open_context *ctx;           /* file open context info */
  71         struct kiocb *          iocb;           /* controlling i/o request */
  72         struct inode *          inode;          /* target file of i/o */
  73
  74         /* completion state */
  75         atomic_t                io_count;       /* i/os we're waiting for */
  76         spinlock_t              lock;           /* protect completion state */
  77         ssize_t                 count,          /* bytes actually processed */
  78                                 error;          /* any reported error */
  79         struct completion       completion;     /* wait for i/o completion */
  80
  81         /* commit state */
  82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  83         struct nfs_write_data * commit_data;    /* special write_data for commits */
  84         int                     flags;
  85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  87         struct nfs_writeverf    verf;           /* unstable write verifier */
  88 };
  89
  90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  91 static const struct rpc_call_ops nfs_write_direct_ops;
  92
  93 static inline void get_dreq(struct nfs_direct_req *dreq)
  94 {
  95         atomic_inc(&dreq->io_count);
  96 }
  97
  98 static inline int put_dreq(struct nfs_direct_req *dreq)
  99 {
 100         return atomic_dec_and_test(&dreq->io_count);
 101 }
 102
 103 /**
 104  * nfs_direct_IO - NFS address space operation for direct I/O
 105  * @rw: direction (read or write)
 106  * @iocb: target I/O control block
 107  * @iov: array of vectors that define I/O buffer
 108  * @pos: offset in file to begin the operation
 109  * @nr_segs: size of iovec array
 110  *
 111  * The presence of this routine in the address space ops vector means
 112  * the NFS client supports direct I/O.  However, we shunt off direct
 113  * read and write requests before the VFS gets them, so this method
 114  * should never be called.
 115  */
 116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 117 {
 118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 119                         iocb->ki_filp->f_path.dentry->d_name.name,
 120                         (long long) pos, nr_segs);
 121
 122         return -EINVAL;
 123 }
 124
 125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
 126 {
 127         unsigned int npages;
 128         unsigned int i;
 129
 130         if (count == 0)
 131                 return;
 132         pages += (pgbase >> PAGE_SHIFT);
 133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 134         for (i = 0; i < npages; i++) {
 135                 struct page *page = pages[i];
 136                 if (!PageCompound(page))
 137                         set_page_dirty(page);
 138         }
 139 }
 140
 141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 142 {
 143         unsigned int i;
 144         for (i = 0; i < npages; i++)
 145                 page_cache_release(pages[i]);
 146 }
 147
 148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 149 {
 150         struct nfs_direct_req *dreq;
 151
 152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 153         if (!dreq)
 154                 return NULL;
 155
 156         kref_init(&dreq->kref);
 157         kref_get(&dreq->kref);
 158         init_completion(&dreq->completion);
 159         INIT_LIST_HEAD(&dreq->rewrite_list);
 160         dreq->iocb = NULL;
 161         dreq->ctx = NULL;
 162         spin_lock_init(&dreq->lock);
 163         atomic_set(&dreq->io_count, 0);
 164         dreq->count = 0;
 165         dreq->error = 0;
 166         dreq->flags = 0;
 167
 168         return dreq;
 169 }
 170
 171 static void nfs_direct_req_free(struct kref *kref)
 172 {
 173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 174
 175         if (dreq->ctx != NULL)
 176                 put_nfs_open_context(dreq->ctx);
 177         kmem_cache_free(nfs_direct_cachep, dreq);
 178 }
 179
 180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 181 {
 182         kref_put(&dreq->kref, nfs_direct_req_free);
 183 }
 184
 185 /*
 186  * Collects and returns the final error value/byte-count.
 187  */
 188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 189 {
 190         ssize_t result = -EIOCBQUEUED;
 191
 192         /* Async requests don't wait here */
 193         if (dreq->iocb)
 194                 goto out;
 195
 196         result = wait_for_completion_interruptible(&dreq->completion);
 197
 198         if (!result)
 199                 result = dreq->error;
 200         if (!result)
 201                 result = dreq->count;
 202
 203 out:
 204         return (ssize_t) result;
 205 }
 206
 207 /*
 208  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 209  * the iocb is still valid here if this is a synchronous request.
 210  */
 211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 212 {
 213         if (dreq->iocb) {
 214                 long res = (long) dreq->error;
 215                 if (!res)
 216                         res = (long) dreq->count;
 217                 aio_complete(dreq->iocb, res, 0);
 218         }
 219         complete_all(&dreq->completion);
 220
 221         nfs_direct_req_release(dreq);
 222 }
 223
 224 /*
 225  * We must hold a reference to all the pages in this direct read request
 226  * until the RPCs complete.  This could be long *after* we are woken up in
 227  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 228  */
 229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 230 {
 231         struct nfs_read_data *data = calldata;
 232         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 233
 234         if (nfs_readpage_result(task, data) != 0)
 235                 return;
 236
 237         spin_lock(&dreq->lock);
 238         if (unlikely(task->tk_status < 0)) {
 239                 dreq->error = task->tk_status;
 240                 spin_unlock(&dreq->lock);
 241         } else {
 242                 dreq->count += data->res.count;
 243                 spin_unlock(&dreq->lock);
 244                 nfs_direct_dirty_pages(data->pagevec,
 245                                 data->args.pgbase,
 246                                 data->res.count);
 247         }
 248         nfs_direct_release_pages(data->pagevec, data->npages);
 249
 250         if (put_dreq(dreq))
 251                 nfs_direct_complete(dreq);
 252 }
 253
 254 static const struct rpc_call_ops nfs_read_direct_ops = {
 255         .rpc_call_done = nfs_direct_read_result,
 256         .rpc_release = nfs_readdata_release,
 257 };
 258
 259 /*
 260  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 261  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 262  * bail and stop sending more reads.  Read length accounting is
 263  * handled automatically by nfs_direct_read_result().  Otherwise, if
 264  * no requests have been sent, just return an error.
 265  */
 266 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 267 {
 268         struct nfs_open_context *ctx = dreq->ctx;
 269         struct inode *inode = ctx->path.dentry->d_inode;
 270         size_t rsize = NFS_SERVER(inode)->rsize;
 271         unsigned int pgbase;
 272         int result;
 273         ssize_t started = 0;
 274
 275         do {
 276                 struct nfs_read_data *data;
 277                 size_t bytes;
 278
 279                 pgbase = user_addr & ~PAGE_MASK;
 280                 bytes = min(rsize,count);
 281
 282                 result = -ENOMEM;
 283                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 284                 if (unlikely(!data))
 285                         break;
 286
 287                 down_read(&current->mm->mmap_sem);
 288                 result = get_user_pages(current, current->mm, user_addr,
 289                                         data->npages, 1, 0, data->pagevec, NULL);
 290                 up_read(&current->mm->mmap_sem);
 291                 if (result < 0) {
 292                         nfs_readdata_release(data);
 293                         break;
 294                 }
 295                 if ((unsigned)result < data->npages) {
 296                         bytes = result * PAGE_SIZE;
 297                         if (bytes <= pgbase) {
 298                                 nfs_direct_release_pages(data->pagevec, result);
 299                                 nfs_readdata_release(data);
 300                                 break;
 301                         }
 302                         bytes -= pgbase;
 303                         data->npages = result;
 304                 }
 305
 306                 get_dreq(dreq);
 307
 308                 data->req = (struct nfs_page *) dreq;
 309                 data->inode = inode;
 310                 data->cred = ctx->cred;
 311                 data->args.fh = NFS_FH(inode);
 312                 data->args.context = ctx;
 313                 data->args.offset = pos;
 314                 data->args.pgbase = pgbase;
 315                 data->args.pages = data->pagevec;
 316                 data->args.count = bytes;
 317                 data->res.fattr = &data->fattr;
 318                 data->res.eof = 0;
 319                 data->res.count = bytes;
 320
 321                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 322                                 &nfs_read_direct_ops, data);
 323                 NFS_PROTO(inode)->read_setup(data);
 324
 325                 data->task.tk_cookie = (unsigned long) inode;
 326
 327                 rpc_execute(&data->task);
 328
 329                 dprintk("NFS: %5u initiated direct read call "
 330                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 331                                 data->task.tk_pid,
 332                                 inode->i_sb->s_id,
 333                                 (long long)NFS_FILEID(inode),
 334                                 bytes,
 335                                 (unsigned long long)data->args.offset);
 336
 337                 started += bytes;
 338                 user_addr += bytes;
 339                 pos += bytes;
 340                 /* FIXME: Remove this unnecessary math from final patch */
 341                 pgbase += bytes;
 342                 pgbase &= ~PAGE_MASK;
 343                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 344
 345                 count -= bytes;
 346         } while (count != 0);
 347
 348         if (started)
 349                 return started;
 350         return result < 0 ? (ssize_t) result : -EFAULT;
 351 }
 352
 353 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 354                                               const struct iovec *iov,
 355                                               unsigned long nr_segs,
 356                                               loff_t pos)
 357 {
 358         ssize_t result = -EINVAL;
 359         size_t requested_bytes = 0;
 360         unsigned long seg;
 361
 362         get_dreq(dreq);
 363
 364         for (seg = 0; seg < nr_segs; seg++) {
 365                 const struct iovec *vec = &iov[seg];
 366                 result = nfs_direct_read_schedule(dreq,
 367                                         (unsigned long)vec->iov_base,
 368                                                   vec->iov_len, pos);
 369                 if (result < 0)
 370                         break;
 371                 requested_bytes += result;
 372                 if ((size_t)result < vec->iov_len)
 373                         break;
 374                 pos += vec->iov_len;
 375         }
 376
 377         if (put_dreq(dreq))
 378                 nfs_direct_complete(dreq);
 379
 380         if (requested_bytes != 0)
 381                 return 0;
 382
 383         if (result < 0)
 384                 return result;
 385         return -EIO;
 386 }
 387
 388 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 389                                unsigned long nr_segs, loff_t pos)
 390 {
 391         ssize_t result = 0;
 392         sigset_t oldset;
 393         struct inode *inode = iocb->ki_filp->f_mapping->host;
 394         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 395         struct nfs_direct_req *dreq;
 396
 397         dreq = nfs_direct_req_alloc();
 398         if (!dreq)
 399                 return -ENOMEM;
 400
 401         dreq->inode = inode;
 402         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 403         if (!is_sync_kiocb(iocb))
 404                 dreq->iocb = iocb;
 405
 406         rpc_clnt_sigmask(clnt, &oldset);
 407         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 408         if (!result)
 409                 result = nfs_direct_wait(dreq);
 410         rpc_clnt_sigunmask(clnt, &oldset);
 411         nfs_direct_req_release(dreq);
 412
 413         return result;
 414 }
 415
 416 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 417 {
 418         while (!list_empty(&dreq->rewrite_list)) {
 419                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 420                 list_del(&data->pages);
 421                 nfs_direct_release_pages(data->pagevec, data->npages);
 422                 nfs_writedata_release(data);
 423         }
 424 }
 425
 426 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 427 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 428 {
 429         struct inode *inode = dreq->inode;
 430         struct list_head *p;
 431         struct nfs_write_data *data;
 432
 433         dreq->count = 0;
 434         get_dreq(dreq);
 435
 436         list_for_each(p, &dreq->rewrite_list) {
 437                 data = list_entry(p, struct nfs_write_data, pages);
 438
 439                 get_dreq(dreq);
 440
 441                 /*
 442                  * Reset data->res.
 443                  */
 444                 nfs_fattr_init(&data->fattr);
 445                 data->res.count = data->args.count;
 446                 memset(&data->verf, 0, sizeof(data->verf));
 447
 448                 /*
 449                  * Reuse data->task; data->args should not have changed
 450                  * since the original request was sent.
 451                  */
 452                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 453                                 &nfs_write_direct_ops, data);
 454                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 455
 456                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 457                 data->task.tk_cookie = (unsigned long) inode;
 458
 459                 /*
 460                  * We're called via an RPC callback, so BKL is already held.
 461                  */
 462                 rpc_execute(&data->task);
 463
 464                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 465                                 data->task.tk_pid,
 466                                 inode->i_sb->s_id,
 467                                 (long long)NFS_FILEID(inode),
 468                                 data->args.count,
 469                                 (unsigned long long)data->args.offset);
 470         }
 471
 472         if (put_dreq(dreq))
 473                 nfs_direct_write_complete(dreq, inode);
 474 }
 475
 476 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 477 {
 478         struct nfs_write_data *data = calldata;
 479         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 480
 481         /* Call the NFS version-specific code */
 482         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 483                 return;
 484         if (unlikely(task->tk_status < 0)) {
 485                 dprintk("NFS: %5u commit failed with error %d.\n",
 486                                 task->tk_pid, task->tk_status);
 487                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 488         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 489                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 490                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 491         }
 492
 493         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 494         nfs_direct_write_complete(dreq, data->inode);
 495 }
 496
 497 static const struct rpc_call_ops nfs_commit_direct_ops = {
 498         .rpc_call_done = nfs_direct_commit_result,
 499         .rpc_release = nfs_commit_release,
 500 };
 501
 502 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 503 {
 504         struct nfs_write_data *data = dreq->commit_data;
 505
 506         data->inode = dreq->inode;
 507         data->cred = dreq->ctx->cred;
 508
 509         data->args.fh = NFS_FH(data->inode);
 510         data->args.offset = 0;
 511         data->args.count = 0;
 512         data->res.count = 0;
 513         data->res.fattr = &data->fattr;
 514         data->res.verf = &data->verf;
 515
 516         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 517                                 &nfs_commit_direct_ops, data);
 518         NFS_PROTO(data->inode)->commit_setup(data, 0);
 519
 520         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 521         data->task.tk_cookie = (unsigned long)data->inode;
 522         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 523         dreq->commit_data = NULL;
 524
 525         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 526
 527         rpc_execute(&data->task);
 528 }
 529
 530 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 531 {
 532         int flags = dreq->flags;
 533
 534         dreq->flags = 0;
 535         switch (flags) {
 536                 case NFS_ODIRECT_DO_COMMIT:
 537                         nfs_direct_commit_schedule(dreq);
 538                         break;
 539                 case NFS_ODIRECT_RESCHED_WRITES:
 540                         nfs_direct_write_reschedule(dreq);
 541                         break;
 542                 default:
 543                         if (dreq->commit_data != NULL)
 544                                 nfs_commit_free(dreq->commit_data);
 545                         nfs_direct_free_writedata(dreq);
 546                         nfs_zap_mapping(inode, inode->i_mapping);
 547                         nfs_direct_complete(dreq);
 548         }
 549 }
 550
 551 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 552 {
 553         dreq->commit_data = nfs_commit_alloc();
 554         if (dreq->commit_data != NULL)
 555                 dreq->commit_data->req = (struct nfs_page *) dreq;
 556 }
 557 #else
 558 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 559 {
 560         dreq->commit_data = NULL;
 561 }
 562
 563 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 564 {
 565         nfs_direct_free_writedata(dreq);
 566         nfs_zap_mapping(inode, inode->i_mapping);
 567         nfs_direct_complete(dreq);
 568 }
 569 #endif
 570
 571 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 572 {
 573         struct nfs_write_data *data = calldata;
 574         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 575         int status = task->tk_status;
 576
 577         if (nfs_writeback_done(task, data) != 0)
 578                 return;
 579
 580         spin_lock(&dreq->lock);
 581
 582         if (unlikely(status < 0)) {
 583                 /* An error has occurred, so we should not commit */
 584                 dreq->flags = 0;
 585                 dreq->error = status;
 586         }
 587         if (unlikely(dreq->error != 0))
 588                 goto out_unlock;
 589
 590         dreq->count += data->res.count;
 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 out_unlock:
 606         spin_unlock(&dreq->lock);
 607 }
 608
 609 /*
 610  * NB: Return the value of the first error return code.  Subsequent
 611  *     errors after the first one are ignored.
 612  */
 613 static void nfs_direct_write_release(void *calldata)
 614 {
 615         struct nfs_write_data *data = calldata;
 616         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 617
 618         if (put_dreq(dreq))
 619                 nfs_direct_write_complete(dreq, data->inode);
 620 }
 621
 622 static const struct rpc_call_ops nfs_write_direct_ops = {
 623         .rpc_call_done = nfs_direct_write_result,
 624         .rpc_release = nfs_direct_write_release,
 625 };
 626
 627 /*
 628  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 629  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 630  * bail and stop sending more writes.  Write length accounting is
 631  * handled automatically by nfs_direct_write_result().  Otherwise, if
 632  * no requests have been sent, just return an error.
 633  */
 634 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 635 {
 636         struct nfs_open_context *ctx = dreq->ctx;
 637         struct inode *inode = ctx->path.dentry->d_inode;
 638         size_t wsize = NFS_SERVER(inode)->wsize;
 639         unsigned int pgbase;
 640         int result;
 641         ssize_t started = 0;
 642
 643         do {
 644                 struct nfs_write_data *data;
 645                 size_t bytes;
 646
 647                 pgbase = user_addr & ~PAGE_MASK;
 648                 bytes = min(wsize,count);
 649
 650                 result = -ENOMEM;
 651                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 652                 if (unlikely(!data))
 653                         break;
 654
 655                 down_read(&current->mm->mmap_sem);
 656                 result = get_user_pages(current, current->mm, user_addr,
 657                                         data->npages, 0, 0, data->pagevec, NULL);
 658                 up_read(&current->mm->mmap_sem);
 659                 if (result < 0) {
 660                         nfs_writedata_release(data);
 661                         break;
 662                 }
 663                 if ((unsigned)result < data->npages) {
 664                         bytes = result * PAGE_SIZE;
 665                         if (bytes <= pgbase) {
 666                                 nfs_direct_release_pages(data->pagevec, result);
 667                                 nfs_writedata_release(data);
 668                                 break;
 669                         }
 670                         bytes -= pgbase;
 671                         data->npages = result;
 672                 }
 673
 674                 get_dreq(dreq);
 675
 676                 list_move_tail(&data->pages, &dreq->rewrite_list);
 677
 678                 data->req = (struct nfs_page *) dreq;
 679                 data->inode = inode;
 680                 data->cred = ctx->cred;
 681                 data->args.fh = NFS_FH(inode);
 682                 data->args.context = ctx;
 683                 data->args.offset = pos;
 684                 data->args.pgbase = pgbase;
 685                 data->args.pages = data->pagevec;
 686                 data->args.count = bytes;
 687                 data->res.fattr = &data->fattr;
 688                 data->res.count = bytes;
 689                 data->res.verf = &data->verf;
 690
 691                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 692                                 &nfs_write_direct_ops, data);
 693                 NFS_PROTO(inode)->write_setup(data, sync);
 694
 695                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 696                 data->task.tk_cookie = (unsigned long) inode;
 697
 698                 rpc_execute(&data->task);
 699
 700                 dprintk("NFS: %5u initiated direct write call "
 701                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 702                                 data->task.tk_pid,
 703                                 inode->i_sb->s_id,
 704                                 (long long)NFS_FILEID(inode),
 705                                 bytes,
 706                                 (unsigned long long)data->args.offset);
 707
 708                 started += bytes;
 709                 user_addr += bytes;
 710                 pos += bytes;
 711
 712                 /* FIXME: Remove this useless math from the final patch */
 713                 pgbase += bytes;
 714                 pgbase &= ~PAGE_MASK;
 715                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 716
 717                 count -= bytes;
 718         } while (count != 0);
 719
 720         if (started)
 721                 return started;
 722         return result < 0 ? (ssize_t) result : -EFAULT;
 723 }
 724
 725 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 726                                                const struct iovec *iov,
 727                                                unsigned long nr_segs,
 728                                                loff_t pos, int sync)
 729 {
 730         ssize_t result = 0;
 731         size_t requested_bytes = 0;
 732         unsigned long seg;
 733
 734         get_dreq(dreq);
 735
 736         for (seg = 0; seg < nr_segs; seg++) {
 737                 const struct iovec *vec = &iov[seg];
 738                 result = nfs_direct_write_schedule(dreq,
 739                                         (unsigned long)vec->iov_base,
 740                                                    vec->iov_len,
 741                                                    pos, sync);
 742                 if (result < 0)
 743                         break;
 744                 requested_bytes += result;
 745                 if ((size_t)result < vec->iov_len)
 746                         break;
 747                 pos += vec->iov_len;
 748         }
 749
 750         if (put_dreq(dreq))
 751                 nfs_direct_write_complete(dreq, dreq->inode);
 752
 753         if (requested_bytes != 0)
 754                 return 0;
 755
 756         if (result < 0)
 757                 return result;
 758         return -EIO;
 759 }
 760
 761 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 762                                 unsigned long nr_segs, loff_t pos,
 763                                 size_t count)
 764 {
 765         ssize_t result = 0;
 766         sigset_t oldset;
 767         struct inode *inode = iocb->ki_filp->f_mapping->host;
 768         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 769         struct nfs_direct_req *dreq;
 770         size_t wsize = NFS_SERVER(inode)->wsize;
 771         int sync = 0;
 772
 773         dreq = nfs_direct_req_alloc();
 774         if (!dreq)
 775                 return -ENOMEM;
 776         nfs_alloc_commit_data(dreq);
 777
 778         if (dreq->commit_data == NULL || count < wsize)
 779                 sync = FLUSH_STABLE;
 780
 781         dreq->inode = inode;
 782         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 783         if (!is_sync_kiocb(iocb))
 784                 dreq->iocb = iocb;
 785
 786         rpc_clnt_sigmask(clnt, &oldset);
 787         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 788         if (!result)
 789                 result = nfs_direct_wait(dreq);
 790         rpc_clnt_sigunmask(clnt, &oldset);
 791         nfs_direct_req_release(dreq);
 792
 793         return result;
 794 }
 795
 796 /**
 797  * nfs_file_direct_read - file direct read operation for NFS files
 798  * @iocb: target I/O control block
 799  * @iov: vector of user buffers into which to read data
 800  * @nr_segs: size of iov vector
 801  * @pos: byte offset in file where reading starts
 802  *
 803  * We use this function for direct reads instead of calling
 804  * generic_file_aio_read() in order to avoid gfar's check to see if
 805  * the request starts before the end of the file.  For that check
 806  * to work, we must generate a GETATTR before each direct read, and
 807  * even then there is a window between the GETATTR and the subsequent
 808  * READ where the file size could change.  Our preference is simply
 809  * to do all reads the application wants, and the server will take
 810  * care of managing the end of file boundary.
 811  *
 812  * This function also eliminates unnecessarily updating the file's
 813  * atime locally, as the NFS server sets the file's atime, and this
 814  * client must read the updated atime from the server back into its
 815  * cache.
 816  */
 817 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 818                                 unsigned long nr_segs, loff_t pos)
 819 {
 820         ssize_t retval = -EINVAL;
 821         struct file *file = iocb->ki_filp;
 822         struct address_space *mapping = file->f_mapping;
 823         size_t count;
 824
 825         count = iov_length(iov, nr_segs);
 826         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 827
 828         dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
 829                 file->f_path.dentry->d_parent->d_name.name,
 830                 file->f_path.dentry->d_name.name,
 831                 count, (long long) pos);
 832
 833         retval = 0;
 834         if (!count)
 835                 goto out;
 836
 837         retval = nfs_sync_mapping(mapping);
 838         if (retval)
 839                 goto out;
 840
 841         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 842         if (retval > 0)
 843                 iocb->ki_pos = pos + retval;
 844
 845 out:
 846         return retval;
 847 }
 848
 849 /**
 850  * nfs_file_direct_write - file direct write operation for NFS files
 851  * @iocb: target I/O control block
 852  * @iov: vector of user buffers from which to write data
 853  * @nr_segs: size of iov vector
 854  * @pos: byte offset in file where writing starts
 855  *
 856  * We use this function for direct writes instead of calling
 857  * generic_file_aio_write() in order to avoid taking the inode
 858  * semaphore and updating the i_size.  The NFS server will set
 859  * the new i_size and this client must read the updated size
 860  * back into its cache.  We let the server do generic write
 861  * parameter checking and report problems.
 862  *
 863  * We also avoid an unnecessary invocation of generic_osync_inode(),
 864  * as it is fairly meaningless to sync the metadata of an NFS file.
 865  *
 866  * We eliminate local atime updates, see direct read above.
 867  *
 868  * We avoid unnecessary page cache invalidations for normal cached
 869  * readers of this file.
 870  *
 871  * Note that O_APPEND is not supported for NFS direct writes, as there
 872  * is no atomic O_APPEND write facility in the NFS protocol.
 873  */
 874 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 875                                 unsigned long nr_segs, loff_t pos)
 876 {
 877         ssize_t retval = -EINVAL;
 878         struct file *file = iocb->ki_filp;
 879         struct address_space *mapping = file->f_mapping;
 880         size_t count;
 881
 882         count = iov_length(iov, nr_segs);
 883         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 884
 885         dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
 886                 file->f_path.dentry->d_parent->d_name.name,
 887                 file->f_path.dentry->d_name.name,
 888                 count, (long long) pos);
 889
 890         retval = generic_write_checks(file, &pos, &count, 0);
 891         if (retval)
 892                 goto out;
 893         if (!count)
 894                 goto out;       /* return 0 */
 895
 896         retval = -EINVAL;
 897         if ((ssize_t) count < 0)
 898                 goto out;
 899         retval = 0;
 900         if (!count)
 901                 goto out;
 902
 903         retval = nfs_sync_mapping(mapping);
 904         if (retval)
 905                 goto out;
 906
 907         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
 908
 909         if (retval > 0)
 910                 iocb->ki_pos = pos + retval;
 911
 912 out:
 913         return retval;
 914 }
 915
 916 /**
 917  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 918  *
 919  */
 920 int __init nfs_init_directcache(void)
 921 {
 922         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 923                                                 sizeof(struct nfs_direct_req),
 924                                                 0, (SLAB_RECLAIM_ACCOUNT|
 925                                                         SLAB_MEM_SPREAD),
 926                                                 NULL);
 927         if (nfs_direct_cachep == NULL)
 928                 return -ENOMEM;
 929
 930         return 0;
 931 }
 932
 933 /**
 934  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 935  *
 936  */
 937 void nfs_destroy_directcache(void)
 938 {
 939         kmem_cache_destroy(nfs_direct_cachep);
 940 }