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_killable(&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
 233         nfs_readpage_result(task, data);
 234 }
 235
 236 static void nfs_direct_read_release(void *calldata)
 237 {
 238
 239         struct nfs_read_data *data = calldata;
 240         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 241         int status = data->task.tk_status;
 242
 243         spin_lock(&dreq->lock);
 244         if (unlikely(status < 0)) {
 245                 dreq->error = status;
 246                 spin_unlock(&dreq->lock);
 247         } else {
 248                 dreq->count += data->res.count;
 249                 spin_unlock(&dreq->lock);
 250                 nfs_direct_dirty_pages(data->pagevec,
 251                                 data->args.pgbase,
 252                                 data->res.count);
 253         }
 254         nfs_direct_release_pages(data->pagevec, data->npages);
 255
 256         if (put_dreq(dreq))
 257                 nfs_direct_complete(dreq);
 258         nfs_readdata_release(calldata);
 259 }
 260
 261 static const struct rpc_call_ops nfs_read_direct_ops = {
 262 #if defined(CONFIG_NFS_V4_1)
 263         .rpc_call_prepare = nfs_read_prepare,
 264 #endif /* CONFIG_NFS_V4_1 */
 265         .rpc_call_done = nfs_direct_read_result,
 266         .rpc_release = nfs_direct_read_release,
 267 };
 268
 269 /*
 270  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 271  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 272  * bail and stop sending more reads.  Read length accounting is
 273  * handled automatically by nfs_direct_read_result().  Otherwise, if
 274  * no requests have been sent, just return an error.
 275  */
 276 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
 277                                                 const struct iovec *iov,
 278                                                 loff_t pos)
 279 {
 280         struct nfs_open_context *ctx = dreq->ctx;
 281         struct inode *inode = ctx->path.dentry->d_inode;
 282         unsigned long user_addr = (unsigned long)iov->iov_base;
 283         size_t count = iov->iov_len;
 284         size_t rsize = NFS_SERVER(inode)->rsize;
 285         struct rpc_task *task;
 286         struct rpc_message msg = {
 287                 .rpc_cred = ctx->cred,
 288         };
 289         struct rpc_task_setup task_setup_data = {
 290                 .rpc_client = NFS_CLIENT(inode),
 291                 .rpc_message = &msg,
 292                 .callback_ops = &nfs_read_direct_ops,
 293                 .workqueue = nfsiod_workqueue,
 294                 .flags = RPC_TASK_ASYNC,
 295         };
 296         unsigned int pgbase;
 297         int result;
 298         ssize_t started = 0;
 299
 300         do {
 301                 struct nfs_read_data *data;
 302                 size_t bytes;
 303
 304                 pgbase = user_addr & ~PAGE_MASK;
 305                 bytes = min(rsize,count);
 306
 307                 result = -ENOMEM;
 308                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 309                 if (unlikely(!data))
 310                         break;
 311
 312                 down_read(&current->mm->mmap_sem);
 313                 result = get_user_pages(current, current->mm, user_addr,
 314                                         data->npages, 1, 0, data->pagevec, NULL);
 315                 up_read(&current->mm->mmap_sem);
 316                 if (result < 0) {
 317                         nfs_readdata_release(data);
 318                         break;
 319                 }
 320                 if ((unsigned)result < data->npages) {
 321                         bytes = result * PAGE_SIZE;
 322                         if (bytes <= pgbase) {
 323                                 nfs_direct_release_pages(data->pagevec, result);
 324                                 nfs_readdata_release(data);
 325                                 break;
 326                         }
 327                         bytes -= pgbase;
 328                         data->npages = result;
 329                 }
 330
 331                 get_dreq(dreq);
 332
 333                 data->req = (struct nfs_page *) dreq;
 334                 data->inode = inode;
 335                 data->cred = msg.rpc_cred;
 336                 data->args.fh = NFS_FH(inode);
 337                 data->args.context = get_nfs_open_context(ctx);
 338                 data->args.offset = pos;
 339                 data->args.pgbase = pgbase;
 340                 data->args.pages = data->pagevec;
 341                 data->args.count = bytes;
 342                 data->res.fattr = &data->fattr;
 343                 data->res.eof = 0;
 344                 data->res.count = bytes;
 345                 msg.rpc_argp = &data->args;
 346                 msg.rpc_resp = &data->res;
 347
 348                 task_setup_data.task = &data->task;
 349                 task_setup_data.callback_data = data;
 350                 NFS_PROTO(inode)->read_setup(data, &msg);
 351
 352                 task = rpc_run_task(&task_setup_data);
 353                 if (IS_ERR(task))
 354                         break;
 355                 rpc_put_task(task);
 356
 357                 dprintk("NFS: %5u initiated direct read call "
 358                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 359                                 data->task.tk_pid,
 360                                 inode->i_sb->s_id,
 361                                 (long long)NFS_FILEID(inode),
 362                                 bytes,
 363                                 (unsigned long long)data->args.offset);
 364
 365                 started += bytes;
 366                 user_addr += bytes;
 367                 pos += bytes;
 368                 /* FIXME: Remove this unnecessary math from final patch */
 369                 pgbase += bytes;
 370                 pgbase &= ~PAGE_MASK;
 371                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 372
 373                 count -= bytes;
 374         } while (count != 0);
 375
 376         if (started)
 377                 return started;
 378         return result < 0 ? (ssize_t) result : -EFAULT;
 379 }
 380
 381 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 382                                               const struct iovec *iov,
 383                                               unsigned long nr_segs,
 384                                               loff_t pos)
 385 {
 386         ssize_t result = -EINVAL;
 387         size_t requested_bytes = 0;
 388         unsigned long seg;
 389
 390         get_dreq(dreq);
 391
 392         for (seg = 0; seg < nr_segs; seg++) {
 393                 const struct iovec *vec = &iov[seg];
 394                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
 395                 if (result < 0)
 396                         break;
 397                 requested_bytes += result;
 398                 if ((size_t)result < vec->iov_len)
 399                         break;
 400                 pos += vec->iov_len;
 401         }
 402
 403         if (put_dreq(dreq))
 404                 nfs_direct_complete(dreq);
 405
 406         if (requested_bytes != 0)
 407                 return 0;
 408
 409         if (result < 0)
 410                 return result;
 411         return -EIO;
 412 }
 413
 414 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 415                                unsigned long nr_segs, loff_t pos)
 416 {
 417         ssize_t result = 0;
 418         struct inode *inode = iocb->ki_filp->f_mapping->host;
 419         struct nfs_direct_req *dreq;
 420
 421         dreq = nfs_direct_req_alloc();
 422         if (!dreq)
 423                 return -ENOMEM;
 424
 425         dreq->inode = inode;
 426         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 427         if (!is_sync_kiocb(iocb))
 428                 dreq->iocb = iocb;
 429
 430         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 431         if (!result)
 432                 result = nfs_direct_wait(dreq);
 433         nfs_direct_req_release(dreq);
 434
 435         return result;
 436 }
 437
 438 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 439 {
 440         while (!list_empty(&dreq->rewrite_list)) {
 441                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 442                 list_del(&data->pages);
 443                 nfs_direct_release_pages(data->pagevec, data->npages);
 444                 nfs_writedata_release(data);
 445         }
 446 }
 447
 448 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 449 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 450 {
 451         struct inode *inode = dreq->inode;
 452         struct list_head *p;
 453         struct nfs_write_data *data;
 454         struct rpc_task *task;
 455         struct rpc_message msg = {
 456                 .rpc_cred = dreq->ctx->cred,
 457         };
 458         struct rpc_task_setup task_setup_data = {
 459                 .rpc_client = NFS_CLIENT(inode),
 460                 .callback_ops = &nfs_write_direct_ops,
 461                 .workqueue = nfsiod_workqueue,
 462                 .flags = RPC_TASK_ASYNC,
 463         };
 464
 465         dreq->count = 0;
 466         get_dreq(dreq);
 467
 468         list_for_each(p, &dreq->rewrite_list) {
 469                 data = list_entry(p, struct nfs_write_data, pages);
 470
 471                 get_dreq(dreq);
 472
 473                 /* Use stable writes */
 474                 data->args.stable = NFS_FILE_SYNC;
 475
 476                 /*
 477                  * Reset data->res.
 478                  */
 479                 nfs_fattr_init(&data->fattr);
 480                 data->res.count = data->args.count;
 481                 memset(&data->verf, 0, sizeof(data->verf));
 482
 483                 /*
 484                  * Reuse data->task; data->args should not have changed
 485                  * since the original request was sent.
 486                  */
 487                 task_setup_data.task = &data->task;
 488                 task_setup_data.callback_data = data;
 489                 msg.rpc_argp = &data->args;
 490                 msg.rpc_resp = &data->res;
 491                 NFS_PROTO(inode)->write_setup(data, &msg);
 492
 493                 /*
 494                  * We're called via an RPC callback, so BKL is already held.
 495                  */
 496                 task = rpc_run_task(&task_setup_data);
 497                 if (!IS_ERR(task))
 498                         rpc_put_task(task);
 499
 500                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 501                                 data->task.tk_pid,
 502                                 inode->i_sb->s_id,
 503                                 (long long)NFS_FILEID(inode),
 504                                 data->args.count,
 505                                 (unsigned long long)data->args.offset);
 506         }
 507
 508         if (put_dreq(dreq))
 509                 nfs_direct_write_complete(dreq, inode);
 510 }
 511
 512 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 513 {
 514         struct nfs_write_data *data = calldata;
 515
 516         /* Call the NFS version-specific code */
 517         NFS_PROTO(data->inode)->commit_done(task, data);
 518 }
 519
 520 static void nfs_direct_commit_release(void *calldata)
 521 {
 522         struct nfs_write_data *data = calldata;
 523         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 524         int status = data->task.tk_status;
 525
 526         if (status < 0) {
 527                 dprintk("NFS: %5u commit failed with error %d.\n",
 528                                 data->task.tk_pid, status);
 529                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 530         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 531                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 532                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 533         }
 534
 535         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 536         nfs_direct_write_complete(dreq, data->inode);
 537         nfs_commitdata_release(calldata);
 538 }
 539
 540 static const struct rpc_call_ops nfs_commit_direct_ops = {
 541         .rpc_call_done = nfs_direct_commit_result,
 542         .rpc_release = nfs_direct_commit_release,
 543 };
 544
 545 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 546 {
 547         struct nfs_write_data *data = dreq->commit_data;
 548         struct rpc_task *task;
 549         struct rpc_message msg = {
 550                 .rpc_argp = &data->args,
 551                 .rpc_resp = &data->res,
 552                 .rpc_cred = dreq->ctx->cred,
 553         };
 554         struct rpc_task_setup task_setup_data = {
 555                 .task = &data->task,
 556                 .rpc_client = NFS_CLIENT(dreq->inode),
 557                 .rpc_message = &msg,
 558                 .callback_ops = &nfs_commit_direct_ops,
 559                 .callback_data = data,
 560                 .workqueue = nfsiod_workqueue,
 561                 .flags = RPC_TASK_ASYNC,
 562         };
 563
 564         data->inode = dreq->inode;
 565         data->cred = msg.rpc_cred;
 566
 567         data->args.fh = NFS_FH(data->inode);
 568         data->args.offset = 0;
 569         data->args.count = 0;
 570         data->args.context = get_nfs_open_context(dreq->ctx);
 571         data->res.count = 0;
 572         data->res.fattr = &data->fattr;
 573         data->res.verf = &data->verf;
 574
 575         NFS_PROTO(data->inode)->commit_setup(data, &msg);
 576
 577         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 578         dreq->commit_data = NULL;
 579
 580         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 581
 582         task = rpc_run_task(&task_setup_data);
 583         if (!IS_ERR(task))
 584                 rpc_put_task(task);
 585 }
 586
 587 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 588 {
 589         int flags = dreq->flags;
 590
 591         dreq->flags = 0;
 592         switch (flags) {
 593                 case NFS_ODIRECT_DO_COMMIT:
 594                         nfs_direct_commit_schedule(dreq);
 595                         break;
 596                 case NFS_ODIRECT_RESCHED_WRITES:
 597                         nfs_direct_write_reschedule(dreq);
 598                         break;
 599                 default:
 600                         if (dreq->commit_data != NULL)
 601                                 nfs_commit_free(dreq->commit_data);
 602                         nfs_direct_free_writedata(dreq);
 603                         nfs_zap_mapping(inode, inode->i_mapping);
 604                         nfs_direct_complete(dreq);
 605         }
 606 }
 607
 608 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 609 {
 610         dreq->commit_data = nfs_commitdata_alloc();
 611         if (dreq->commit_data != NULL)
 612                 dreq->commit_data->req = (struct nfs_page *) dreq;
 613 }
 614 #else
 615 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 616 {
 617         dreq->commit_data = NULL;
 618 }
 619
 620 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 621 {
 622         nfs_direct_free_writedata(dreq);
 623         nfs_zap_mapping(inode, inode->i_mapping);
 624         nfs_direct_complete(dreq);
 625 }
 626 #endif
 627
 628 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 629 {
 630         struct nfs_write_data *data = calldata;
 631
 632         if (nfs_writeback_done(task, data) != 0)
 633                 return;
 634 }
 635
 636 /*
 637  * NB: Return the value of the first error return code.  Subsequent
 638  *     errors after the first one are ignored.
 639  */
 640 static void nfs_direct_write_release(void *calldata)
 641 {
 642         struct nfs_write_data *data = calldata;
 643         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 644         int status = data->task.tk_status;
 645
 646         spin_lock(&dreq->lock);
 647
 648         if (unlikely(status < 0)) {
 649                 /* An error has occurred, so we should not commit */
 650                 dreq->flags = 0;
 651                 dreq->error = status;
 652         }
 653         if (unlikely(dreq->error != 0))
 654                 goto out_unlock;
 655
 656         dreq->count += data->res.count;
 657
 658         if (data->res.verf->committed != NFS_FILE_SYNC) {
 659                 switch (dreq->flags) {
 660                         case 0:
 661                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 662                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 663                                 break;
 664                         case NFS_ODIRECT_DO_COMMIT:
 665                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 666                                         dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
 667                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 668                                 }
 669                 }
 670         }
 671 out_unlock:
 672         spin_unlock(&dreq->lock);
 673
 674         if (put_dreq(dreq))
 675                 nfs_direct_write_complete(dreq, data->inode);
 676 }
 677
 678 static const struct rpc_call_ops nfs_write_direct_ops = {
 679 #if defined(CONFIG_NFS_V4_1)
 680         .rpc_call_prepare = nfs_write_prepare,
 681 #endif /* CONFIG_NFS_V4_1 */
 682         .rpc_call_done = nfs_direct_write_result,
 683         .rpc_release = nfs_direct_write_release,
 684 };
 685
 686 /*
 687  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 688  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 689  * bail and stop sending more writes.  Write length accounting is
 690  * handled automatically by nfs_direct_write_result().  Otherwise, if
 691  * no requests have been sent, just return an error.
 692  */
 693 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
 694                                                  const struct iovec *iov,
 695                                                  loff_t pos, int sync)
 696 {
 697         struct nfs_open_context *ctx = dreq->ctx;
 698         struct inode *inode = ctx->path.dentry->d_inode;
 699         unsigned long user_addr = (unsigned long)iov->iov_base;
 700         size_t count = iov->iov_len;
 701         struct rpc_task *task;
 702         struct rpc_message msg = {
 703                 .rpc_cred = ctx->cred,
 704         };
 705         struct rpc_task_setup task_setup_data = {
 706                 .rpc_client = NFS_CLIENT(inode),
 707                 .rpc_message = &msg,
 708                 .callback_ops = &nfs_write_direct_ops,
 709                 .workqueue = nfsiod_workqueue,
 710                 .flags = RPC_TASK_ASYNC,
 711         };
 712         size_t wsize = NFS_SERVER(inode)->wsize;
 713         unsigned int pgbase;
 714         int result;
 715         ssize_t started = 0;
 716
 717         do {
 718                 struct nfs_write_data *data;
 719                 size_t bytes;
 720
 721                 pgbase = user_addr & ~PAGE_MASK;
 722                 bytes = min(wsize,count);
 723
 724                 result = -ENOMEM;
 725                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 726                 if (unlikely(!data))
 727                         break;
 728
 729                 down_read(&current->mm->mmap_sem);
 730                 result = get_user_pages(current, current->mm, user_addr,
 731                                         data->npages, 0, 0, data->pagevec, NULL);
 732                 up_read(&current->mm->mmap_sem);
 733                 if (result < 0) {
 734                         nfs_writedata_release(data);
 735                         break;
 736                 }
 737                 if ((unsigned)result < data->npages) {
 738                         bytes = result * PAGE_SIZE;
 739                         if (bytes <= pgbase) {
 740                                 nfs_direct_release_pages(data->pagevec, result);
 741                                 nfs_writedata_release(data);
 742                                 break;
 743                         }
 744                         bytes -= pgbase;
 745                         data->npages = result;
 746                 }
 747
 748                 get_dreq(dreq);
 749
 750                 list_move_tail(&data->pages, &dreq->rewrite_list);
 751
 752                 data->req = (struct nfs_page *) dreq;
 753                 data->inode = inode;
 754                 data->cred = msg.rpc_cred;
 755                 data->args.fh = NFS_FH(inode);
 756                 data->args.context = get_nfs_open_context(ctx);
 757                 data->args.offset = pos;
 758                 data->args.pgbase = pgbase;
 759                 data->args.pages = data->pagevec;
 760                 data->args.count = bytes;
 761                 data->args.stable = sync;
 762                 data->res.fattr = &data->fattr;
 763                 data->res.count = bytes;
 764                 data->res.verf = &data->verf;
 765
 766                 task_setup_data.task = &data->task;
 767                 task_setup_data.callback_data = data;
 768                 msg.rpc_argp = &data->args;
 769                 msg.rpc_resp = &data->res;
 770                 NFS_PROTO(inode)->write_setup(data, &msg);
 771
 772                 task = rpc_run_task(&task_setup_data);
 773                 if (IS_ERR(task))
 774                         break;
 775                 rpc_put_task(task);
 776
 777                 dprintk("NFS: %5u initiated direct write call "
 778                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 779                                 data->task.tk_pid,
 780                                 inode->i_sb->s_id,
 781                                 (long long)NFS_FILEID(inode),
 782                                 bytes,
 783                                 (unsigned long long)data->args.offset);
 784
 785                 started += bytes;
 786                 user_addr += bytes;
 787                 pos += bytes;
 788
 789                 /* FIXME: Remove this useless math from the final patch */
 790                 pgbase += bytes;
 791                 pgbase &= ~PAGE_MASK;
 792                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 793
 794                 count -= bytes;
 795         } while (count != 0);
 796
 797         if (started)
 798                 return started;
 799         return result < 0 ? (ssize_t) result : -EFAULT;
 800 }
 801
 802 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 803                                                const struct iovec *iov,
 804                                                unsigned long nr_segs,
 805                                                loff_t pos, int sync)
 806 {
 807         ssize_t result = 0;
 808         size_t requested_bytes = 0;
 809         unsigned long seg;
 810
 811         get_dreq(dreq);
 812
 813         for (seg = 0; seg < nr_segs; seg++) {
 814                 const struct iovec *vec = &iov[seg];
 815                 result = nfs_direct_write_schedule_segment(dreq, vec,
 816                                                            pos, sync);
 817                 if (result < 0)
 818                         break;
 819                 requested_bytes += result;
 820                 if ((size_t)result < vec->iov_len)
 821                         break;
 822                 pos += vec->iov_len;
 823         }
 824
 825         if (put_dreq(dreq))
 826                 nfs_direct_write_complete(dreq, dreq->inode);
 827
 828         if (requested_bytes != 0)
 829                 return 0;
 830
 831         if (result < 0)
 832                 return result;
 833         return -EIO;
 834 }
 835
 836 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 837                                 unsigned long nr_segs, loff_t pos,
 838                                 size_t count)
 839 {
 840         ssize_t result = 0;
 841         struct inode *inode = iocb->ki_filp->f_mapping->host;
 842         struct nfs_direct_req *dreq;
 843         size_t wsize = NFS_SERVER(inode)->wsize;
 844         int sync = NFS_UNSTABLE;
 845
 846         dreq = nfs_direct_req_alloc();
 847         if (!dreq)
 848                 return -ENOMEM;
 849         nfs_alloc_commit_data(dreq);
 850
 851         if (dreq->commit_data == NULL || count < wsize)
 852                 sync = NFS_FILE_SYNC;
 853
 854         dreq->inode = inode;
 855         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 856         if (!is_sync_kiocb(iocb))
 857                 dreq->iocb = iocb;
 858
 859         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 860         if (!result)
 861                 result = nfs_direct_wait(dreq);
 862         nfs_direct_req_release(dreq);
 863
 864         return result;
 865 }
 866
 867 /**
 868  * nfs_file_direct_read - file direct read operation for NFS files
 869  * @iocb: target I/O control block
 870  * @iov: vector of user buffers into which to read data
 871  * @nr_segs: size of iov vector
 872  * @pos: byte offset in file where reading starts
 873  *
 874  * We use this function for direct reads instead of calling
 875  * generic_file_aio_read() in order to avoid gfar's check to see if
 876  * the request starts before the end of the file.  For that check
 877  * to work, we must generate a GETATTR before each direct read, and
 878  * even then there is a window between the GETATTR and the subsequent
 879  * READ where the file size could change.  Our preference is simply
 880  * to do all reads the application wants, and the server will take
 881  * care of managing the end of file boundary.
 882  *
 883  * This function also eliminates unnecessarily updating the file's
 884  * atime locally, as the NFS server sets the file's atime, and this
 885  * client must read the updated atime from the server back into its
 886  * cache.
 887  */
 888 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 889                                 unsigned long nr_segs, loff_t pos)
 890 {
 891         ssize_t retval = -EINVAL;
 892         struct file *file = iocb->ki_filp;
 893         struct address_space *mapping = file->f_mapping;
 894         size_t count;
 895
 896         count = iov_length(iov, nr_segs);
 897         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 898
 899         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 900                 file->f_path.dentry->d_parent->d_name.name,
 901                 file->f_path.dentry->d_name.name,
 902                 count, (long long) pos);
 903
 904         retval = 0;
 905         if (!count)
 906                 goto out;
 907
 908         retval = nfs_sync_mapping(mapping);
 909         if (retval)
 910                 goto out;
 911
 912         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 913         if (retval > 0)
 914                 iocb->ki_pos = pos + retval;
 915
 916 out:
 917         return retval;
 918 }
 919
 920 /**
 921  * nfs_file_direct_write - file direct write operation for NFS files
 922  * @iocb: target I/O control block
 923  * @iov: vector of user buffers from which to write data
 924  * @nr_segs: size of iov vector
 925  * @pos: byte offset in file where writing starts
 926  *
 927  * We use this function for direct writes instead of calling
 928  * generic_file_aio_write() in order to avoid taking the inode
 929  * semaphore and updating the i_size.  The NFS server will set
 930  * the new i_size and this client must read the updated size
 931  * back into its cache.  We let the server do generic write
 932  * parameter checking and report problems.
 933  *
 934  * We also avoid an unnecessary invocation of generic_osync_inode(),
 935  * as it is fairly meaningless to sync the metadata of an NFS file.
 936  *
 937  * We eliminate local atime updates, see direct read above.
 938  *
 939  * We avoid unnecessary page cache invalidations for normal cached
 940  * readers of this file.
 941  *
 942  * Note that O_APPEND is not supported for NFS direct writes, as there
 943  * is no atomic O_APPEND write facility in the NFS protocol.
 944  */
 945 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 946                                 unsigned long nr_segs, loff_t pos)
 947 {
 948         ssize_t retval = -EINVAL;
 949         struct file *file = iocb->ki_filp;
 950         struct address_space *mapping = file->f_mapping;
 951         size_t count;
 952
 953         count = iov_length(iov, nr_segs);
 954         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 955
 956         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 957                 file->f_path.dentry->d_parent->d_name.name,
 958                 file->f_path.dentry->d_name.name,
 959                 count, (long long) pos);
 960
 961         retval = generic_write_checks(file, &pos, &count, 0);
 962         if (retval)
 963                 goto out;
 964
 965         retval = -EINVAL;
 966         if ((ssize_t) count < 0)
 967                 goto out;
 968         retval = 0;
 969         if (!count)
 970                 goto out;
 971
 972         retval = nfs_sync_mapping(mapping);
 973         if (retval)
 974                 goto out;
 975
 976         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
 977
 978         if (retval > 0)
 979                 iocb->ki_pos = pos + retval;
 980
 981 out:
 982         return retval;
 983 }
 984
 985 /**
 986  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 987  *
 988  */
 989 int __init nfs_init_directcache(void)
 990 {
 991         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 992                                                 sizeof(struct nfs_direct_req),
 993                                                 0, (SLAB_RECLAIM_ACCOUNT|
 994                                                         SLAB_MEM_SPREAD),
 995                                                 NULL);
 996         if (nfs_direct_cachep == NULL)
 997                 return -ENOMEM;
 998
 999         return 0;
1000 }
1001
1002 /**
1003  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1004  *
1005  */
1006 void nfs_destroy_directcache(void)
1007 {
1008         kmem_cache_destroy(nfs_direct_cachep);
1009 }