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 #include <linux/slab.h>
  48 #include <linux/task_io_accounting_ops.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 <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60
  61 #define NFSDBG_FACILITY         NFSDBG_VFS
  62
  63 static struct kmem_cache *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 nfs_open_context *ctx;           /* file open context info */
  73         struct nfs_lock_context *l_ctx;         /* Lock context info */
  74         struct kiocb *          iocb;           /* controlling i/o request */
  75         struct inode *          inode;          /* target file of i/o */
  76
  77         /* completion state */
  78         atomic_t                io_count;       /* i/os we're waiting for */
  79         spinlock_t              lock;           /* protect completion state */
  80         ssize_t                 count,          /* bytes actually processed */
  81                                 error;          /* any reported error */
  82         struct completion       completion;     /* wait for i/o completion */
  83
  84         /* commit state */
  85         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  86         struct nfs_write_data * commit_data;    /* special write_data for commits */
  87         int                     flags;
  88 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  89 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  90         struct nfs_writeverf    verf;           /* unstable write verifier */
  91 };
  92
  93 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  94 static const struct rpc_call_ops nfs_write_direct_ops;
  95
  96 static inline void get_dreq(struct nfs_direct_req *dreq)
  97 {
  98         atomic_inc(&dreq->io_count);
  99 }
 100
 101 static inline int put_dreq(struct nfs_direct_req *dreq)
 102 {
 103         return atomic_dec_and_test(&dreq->io_count);
 104 }
 105
 106 /**
 107  * nfs_direct_IO - NFS address space operation for direct I/O
 108  * @rw: direction (read or write)
 109  * @iocb: target I/O control block
 110  * @iov: array of vectors that define I/O buffer
 111  * @pos: offset in file to begin the operation
 112  * @nr_segs: size of iovec array
 113  *
 114  * The presence of this routine in the address space ops vector means
 115  * the NFS client supports direct I/O.  However, we shunt off direct
 116  * read and write requests before the VFS gets them, so this method
 117  * should never be called.
 118  */
 119 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 120 {
 121         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 122                         iocb->ki_filp->f_path.dentry->d_name.name,
 123                         (long long) pos, nr_segs);
 124
 125         return -EINVAL;
 126 }
 127
 128 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
 129 {
 130         unsigned int npages;
 131         unsigned int i;
 132
 133         if (count == 0)
 134                 return;
 135         pages += (pgbase >> PAGE_SHIFT);
 136         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 137         for (i = 0; i < npages; i++) {
 138                 struct page *page = pages[i];
 139                 if (!PageCompound(page))
 140                         set_page_dirty(page);
 141         }
 142 }
 143
 144 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 145 {
 146         unsigned int i;
 147         for (i = 0; i < npages; i++)
 148                 page_cache_release(pages[i]);
 149 }
 150
 151 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 152 {
 153         struct nfs_direct_req *dreq;
 154
 155         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 156         if (!dreq)
 157                 return NULL;
 158
 159         kref_init(&dreq->kref);
 160         kref_get(&dreq->kref);
 161         init_completion(&dreq->completion);
 162         INIT_LIST_HEAD(&dreq->rewrite_list);
 163         dreq->iocb = NULL;
 164         dreq->ctx = NULL;
 165         dreq->l_ctx = NULL;
 166         spin_lock_init(&dreq->lock);
 167         atomic_set(&dreq->io_count, 0);
 168         dreq->count = 0;
 169         dreq->error = 0;
 170         dreq->flags = 0;
 171
 172         return dreq;
 173 }
 174
 175 static void nfs_direct_req_free(struct kref *kref)
 176 {
 177         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 178
 179         if (dreq->l_ctx != NULL)
 180                 nfs_put_lock_context(dreq->l_ctx);
 181         if (dreq->ctx != NULL)
 182                 put_nfs_open_context(dreq->ctx);
 183         kmem_cache_free(nfs_direct_cachep, dreq);
 184 }
 185
 186 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 187 {
 188         kref_put(&dreq->kref, nfs_direct_req_free);
 189 }
 190
 191 /*
 192  * Collects and returns the final error value/byte-count.
 193  */
 194 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 195 {
 196         ssize_t result = -EIOCBQUEUED;
 197
 198         /* Async requests don't wait here */
 199         if (dreq->iocb)
 200                 goto out;
 201
 202         result = wait_for_completion_killable(&dreq->completion);
 203
 204         if (!result)
 205                 result = dreq->error;
 206         if (!result)
 207                 result = dreq->count;
 208
 209 out:
 210         return (ssize_t) result;
 211 }
 212
 213 /*
 214  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 215  * the iocb is still valid here if this is a synchronous request.
 216  */
 217 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 218 {
 219         if (dreq->iocb) {
 220                 long res = (long) dreq->error;
 221                 if (!res)
 222                         res = (long) dreq->count;
 223                 aio_complete(dreq->iocb, res, 0);
 224         }
 225         complete_all(&dreq->completion);
 226
 227         nfs_direct_req_release(dreq);
 228 }
 229
 230 /*
 231  * We must hold a reference to all the pages in this direct read request
 232  * until the RPCs complete.  This could be long *after* we are woken up in
 233  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 234  */
 235 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 236 {
 237         struct nfs_read_data *data = calldata;
 238
 239         nfs_readpage_result(task, data);
 240 }
 241
 242 static void nfs_direct_read_release(void *calldata)
 243 {
 244
 245         struct nfs_read_data *data = calldata;
 246         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 247         int status = data->task.tk_status;
 248
 249         spin_lock(&dreq->lock);
 250         if (unlikely(status < 0)) {
 251                 dreq->error = status;
 252                 spin_unlock(&dreq->lock);
 253         } else {
 254                 dreq->count += data->res.count;
 255                 spin_unlock(&dreq->lock);
 256                 nfs_direct_dirty_pages(data->pagevec,
 257                                 data->args.pgbase,
 258                                 data->res.count);
 259         }
 260         nfs_direct_release_pages(data->pagevec, data->npages);
 261
 262         if (put_dreq(dreq))
 263                 nfs_direct_complete(dreq);
 264         nfs_readdata_free(data);
 265 }
 266
 267 static const struct rpc_call_ops nfs_read_direct_ops = {
 268         .rpc_call_prepare = nfs_read_prepare,
 269         .rpc_call_done = nfs_direct_read_result,
 270         .rpc_release = nfs_direct_read_release,
 271 };
 272
 273 /*
 274  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 275  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 276  * bail and stop sending more reads.  Read length accounting is
 277  * handled automatically by nfs_direct_read_result().  Otherwise, if
 278  * no requests have been sent, just return an error.
 279  */
 280 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
 281                                                 const struct iovec *iov,
 282                                                 loff_t pos)
 283 {
 284         struct nfs_open_context *ctx = dreq->ctx;
 285         struct inode *inode = ctx->dentry->d_inode;
 286         unsigned long user_addr = (unsigned long)iov->iov_base;
 287         size_t count = iov->iov_len;
 288         size_t rsize = NFS_SERVER(inode)->rsize;
 289         struct rpc_task *task;
 290         struct rpc_message msg = {
 291                 .rpc_cred = ctx->cred,
 292         };
 293         struct rpc_task_setup task_setup_data = {
 294                 .rpc_client = NFS_CLIENT(inode),
 295                 .rpc_message = &msg,
 296                 .callback_ops = &nfs_read_direct_ops,
 297                 .workqueue = nfsiod_workqueue,
 298                 .flags = RPC_TASK_ASYNC,
 299         };
 300         unsigned int pgbase;
 301         int result;
 302         ssize_t started = 0;
 303
 304         do {
 305                 struct nfs_read_data *data;
 306                 size_t bytes;
 307
 308                 pgbase = user_addr & ~PAGE_MASK;
 309                 bytes = min(rsize,count);
 310
 311                 result = -ENOMEM;
 312                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 313                 if (unlikely(!data))
 314                         break;
 315
 316                 down_read(&current->mm->mmap_sem);
 317                 result = get_user_pages(current, current->mm, user_addr,
 318                                         data->npages, 1, 0, data->pagevec, NULL);
 319                 up_read(&current->mm->mmap_sem);
 320                 if (result < 0) {
 321                         nfs_readdata_free(data);
 322                         break;
 323                 }
 324                 if ((unsigned)result < data->npages) {
 325                         bytes = result * PAGE_SIZE;
 326                         if (bytes <= pgbase) {
 327                                 nfs_direct_release_pages(data->pagevec, result);
 328                                 nfs_readdata_free(data);
 329                                 break;
 330                         }
 331                         bytes -= pgbase;
 332                         data->npages = result;
 333                 }
 334
 335                 get_dreq(dreq);
 336
 337                 data->req = (struct nfs_page *) dreq;
 338                 data->inode = inode;
 339                 data->cred = msg.rpc_cred;
 340                 data->args.fh = NFS_FH(inode);
 341                 data->args.context = ctx;
 342                 data->args.lock_context = dreq->l_ctx;
 343                 data->args.offset = pos;
 344                 data->args.pgbase = pgbase;
 345                 data->args.pages = data->pagevec;
 346                 data->args.count = bytes;
 347                 data->res.fattr = &data->fattr;
 348                 data->res.eof = 0;
 349                 data->res.count = bytes;
 350                 nfs_fattr_init(&data->fattr);
 351                 msg.rpc_argp = &data->args;
 352                 msg.rpc_resp = &data->res;
 353
 354                 task_setup_data.task = &data->task;
 355                 task_setup_data.callback_data = data;
 356                 NFS_PROTO(inode)->read_setup(data, &msg);
 357
 358                 task = rpc_run_task(&task_setup_data);
 359                 if (IS_ERR(task))
 360                         break;
 361                 rpc_put_task(task);
 362
 363                 dprintk("NFS: %5u initiated direct read call "
 364                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 365                                 data->task.tk_pid,
 366                                 inode->i_sb->s_id,
 367                                 (long long)NFS_FILEID(inode),
 368                                 bytes,
 369                                 (unsigned long long)data->args.offset);
 370
 371                 started += bytes;
 372                 user_addr += bytes;
 373                 pos += bytes;
 374                 /* FIXME: Remove this unnecessary math from final patch */
 375                 pgbase += bytes;
 376                 pgbase &= ~PAGE_MASK;
 377                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 378
 379                 count -= bytes;
 380         } while (count != 0);
 381
 382         if (started)
 383                 return started;
 384         return result < 0 ? (ssize_t) result : -EFAULT;
 385 }
 386
 387 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 388                                               const struct iovec *iov,
 389                                               unsigned long nr_segs,
 390                                               loff_t pos)
 391 {
 392         ssize_t result = -EINVAL;
 393         size_t requested_bytes = 0;
 394         unsigned long seg;
 395
 396         get_dreq(dreq);
 397
 398         for (seg = 0; seg < nr_segs; seg++) {
 399                 const struct iovec *vec = &iov[seg];
 400                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
 401                 if (result < 0)
 402                         break;
 403                 requested_bytes += result;
 404                 if ((size_t)result < vec->iov_len)
 405                         break;
 406                 pos += vec->iov_len;
 407         }
 408
 409         /*
 410          * If no bytes were started, return the error, and let the
 411          * generic layer handle the completion.
 412          */
 413         if (requested_bytes == 0) {
 414                 nfs_direct_req_release(dreq);
 415                 return result < 0 ? result : -EIO;
 416         }
 417
 418         if (put_dreq(dreq))
 419                 nfs_direct_complete(dreq);
 420         return 0;
 421 }
 422
 423 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 424                                unsigned long nr_segs, loff_t pos)
 425 {
 426         ssize_t result = -ENOMEM;
 427         struct inode *inode = iocb->ki_filp->f_mapping->host;
 428         struct nfs_direct_req *dreq;
 429
 430         dreq = nfs_direct_req_alloc();
 431         if (dreq == NULL)
 432                 goto out;
 433
 434         dreq->inode = inode;
 435         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 436         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 437         if (dreq->l_ctx == NULL)
 438                 goto out_release;
 439         if (!is_sync_kiocb(iocb))
 440                 dreq->iocb = iocb;
 441
 442         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 443         if (!result)
 444                 result = nfs_direct_wait(dreq);
 445 out_release:
 446         nfs_direct_req_release(dreq);
 447 out:
 448         return result;
 449 }
 450
 451 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 452 {
 453         while (!list_empty(&dreq->rewrite_list)) {
 454                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 455                 list_del(&data->pages);
 456                 nfs_direct_release_pages(data->pagevec, data->npages);
 457                 nfs_writedata_free(data);
 458         }
 459 }
 460
 461 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 462 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 463 {
 464         struct inode *inode = dreq->inode;
 465         struct list_head *p;
 466         struct nfs_write_data *data;
 467         struct rpc_task *task;
 468         struct rpc_message msg = {
 469                 .rpc_cred = dreq->ctx->cred,
 470         };
 471         struct rpc_task_setup task_setup_data = {
 472                 .rpc_client = NFS_CLIENT(inode),
 473                 .rpc_message = &msg,
 474                 .callback_ops = &nfs_write_direct_ops,
 475                 .workqueue = nfsiod_workqueue,
 476                 .flags = RPC_TASK_ASYNC,
 477         };
 478
 479         dreq->count = 0;
 480         get_dreq(dreq);
 481
 482         list_for_each(p, &dreq->rewrite_list) {
 483                 data = list_entry(p, struct nfs_write_data, pages);
 484
 485                 get_dreq(dreq);
 486
 487                 /* Use stable writes */
 488                 data->args.stable = NFS_FILE_SYNC;
 489
 490                 /*
 491                  * Reset data->res.
 492                  */
 493                 nfs_fattr_init(&data->fattr);
 494                 data->res.count = data->args.count;
 495                 memset(&data->verf, 0, sizeof(data->verf));
 496
 497                 /*
 498                  * Reuse data->task; data->args should not have changed
 499                  * since the original request was sent.
 500                  */
 501                 task_setup_data.task = &data->task;
 502                 task_setup_data.callback_data = data;
 503                 msg.rpc_argp = &data->args;
 504                 msg.rpc_resp = &data->res;
 505                 NFS_PROTO(inode)->write_setup(data, &msg);
 506
 507                 /*
 508                  * We're called via an RPC callback, so BKL is already held.
 509                  */
 510                 task = rpc_run_task(&task_setup_data);
 511                 if (!IS_ERR(task))
 512                         rpc_put_task(task);
 513
 514                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 515                                 data->task.tk_pid,
 516                                 inode->i_sb->s_id,
 517                                 (long long)NFS_FILEID(inode),
 518                                 data->args.count,
 519                                 (unsigned long long)data->args.offset);
 520         }
 521
 522         if (put_dreq(dreq))
 523                 nfs_direct_write_complete(dreq, inode);
 524 }
 525
 526 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 527 {
 528         struct nfs_write_data *data = calldata;
 529
 530         /* Call the NFS version-specific code */
 531         NFS_PROTO(data->inode)->commit_done(task, data);
 532 }
 533
 534 static void nfs_direct_commit_release(void *calldata)
 535 {
 536         struct nfs_write_data *data = calldata;
 537         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 538         int status = data->task.tk_status;
 539
 540         if (status < 0) {
 541                 dprintk("NFS: %5u commit failed with error %d.\n",
 542                                 data->task.tk_pid, status);
 543                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 544         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 545                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 546                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 547         }
 548
 549         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 550         nfs_direct_write_complete(dreq, data->inode);
 551         nfs_commit_free(data);
 552 }
 553
 554 static const struct rpc_call_ops nfs_commit_direct_ops = {
 555         .rpc_call_prepare = nfs_write_prepare,
 556         .rpc_call_done = nfs_direct_commit_result,
 557         .rpc_release = nfs_direct_commit_release,
 558 };
 559
 560 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 561 {
 562         struct nfs_write_data *data = dreq->commit_data;
 563         struct rpc_task *task;
 564         struct rpc_message msg = {
 565                 .rpc_argp = &data->args,
 566                 .rpc_resp = &data->res,
 567                 .rpc_cred = dreq->ctx->cred,
 568         };
 569         struct rpc_task_setup task_setup_data = {
 570                 .task = &data->task,
 571                 .rpc_client = NFS_CLIENT(dreq->inode),
 572                 .rpc_message = &msg,
 573                 .callback_ops = &nfs_commit_direct_ops,
 574                 .callback_data = data,
 575                 .workqueue = nfsiod_workqueue,
 576                 .flags = RPC_TASK_ASYNC,
 577         };
 578
 579         data->inode = dreq->inode;
 580         data->cred = msg.rpc_cred;
 581
 582         data->args.fh = NFS_FH(data->inode);
 583         data->args.offset = 0;
 584         data->args.count = 0;
 585         data->args.context = dreq->ctx;
 586         data->args.lock_context = dreq->l_ctx;
 587         data->res.count = 0;
 588         data->res.fattr = &data->fattr;
 589         data->res.verf = &data->verf;
 590         nfs_fattr_init(&data->fattr);
 591
 592         NFS_PROTO(data->inode)->commit_setup(data, &msg);
 593
 594         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 595         dreq->commit_data = NULL;
 596
 597         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 598
 599         task = rpc_run_task(&task_setup_data);
 600         if (!IS_ERR(task))
 601                 rpc_put_task(task);
 602 }
 603
 604 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 605 {
 606         int flags = dreq->flags;
 607
 608         dreq->flags = 0;
 609         switch (flags) {
 610                 case NFS_ODIRECT_DO_COMMIT:
 611                         nfs_direct_commit_schedule(dreq);
 612                         break;
 613                 case NFS_ODIRECT_RESCHED_WRITES:
 614                         nfs_direct_write_reschedule(dreq);
 615                         break;
 616                 default:
 617                         if (dreq->commit_data != NULL)
 618                                 nfs_commit_free(dreq->commit_data);
 619                         nfs_direct_free_writedata(dreq);
 620                         nfs_zap_mapping(inode, inode->i_mapping);
 621                         nfs_direct_complete(dreq);
 622         }
 623 }
 624
 625 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 626 {
 627         dreq->commit_data = nfs_commitdata_alloc();
 628         if (dreq->commit_data != NULL)
 629                 dreq->commit_data->req = (struct nfs_page *) dreq;
 630 }
 631 #else
 632 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 633 {
 634         dreq->commit_data = NULL;
 635 }
 636
 637 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 638 {
 639         nfs_direct_free_writedata(dreq);
 640         nfs_zap_mapping(inode, inode->i_mapping);
 641         nfs_direct_complete(dreq);
 642 }
 643 #endif
 644
 645 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 646 {
 647         struct nfs_write_data *data = calldata;
 648
 649         nfs_writeback_done(task, data);
 650 }
 651
 652 /*
 653  * NB: Return the value of the first error return code.  Subsequent
 654  *     errors after the first one are ignored.
 655  */
 656 static void nfs_direct_write_release(void *calldata)
 657 {
 658         struct nfs_write_data *data = calldata;
 659         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 660         int status = data->task.tk_status;
 661
 662         spin_lock(&dreq->lock);
 663
 664         if (unlikely(status < 0)) {
 665                 /* An error has occurred, so we should not commit */
 666                 dreq->flags = 0;
 667                 dreq->error = status;
 668         }
 669         if (unlikely(dreq->error != 0))
 670                 goto out_unlock;
 671
 672         dreq->count += data->res.count;
 673
 674         if (data->res.verf->committed != NFS_FILE_SYNC) {
 675                 switch (dreq->flags) {
 676                         case 0:
 677                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 678                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 679                                 break;
 680                         case NFS_ODIRECT_DO_COMMIT:
 681                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 682                                         dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
 683                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 684                                 }
 685                 }
 686         }
 687 out_unlock:
 688         spin_unlock(&dreq->lock);
 689
 690         if (put_dreq(dreq))
 691                 nfs_direct_write_complete(dreq, data->inode);
 692 }
 693
 694 static const struct rpc_call_ops nfs_write_direct_ops = {
 695         .rpc_call_prepare = nfs_write_prepare,
 696         .rpc_call_done = nfs_direct_write_result,
 697         .rpc_release = nfs_direct_write_release,
 698 };
 699
 700 /*
 701  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 702  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 703  * bail and stop sending more writes.  Write length accounting is
 704  * handled automatically by nfs_direct_write_result().  Otherwise, if
 705  * no requests have been sent, just return an error.
 706  */
 707 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
 708                                                  const struct iovec *iov,
 709                                                  loff_t pos, int sync)
 710 {
 711         struct nfs_open_context *ctx = dreq->ctx;
 712         struct inode *inode = ctx->dentry->d_inode;
 713         unsigned long user_addr = (unsigned long)iov->iov_base;
 714         size_t count = iov->iov_len;
 715         struct rpc_task *task;
 716         struct rpc_message msg = {
 717                 .rpc_cred = ctx->cred,
 718         };
 719         struct rpc_task_setup task_setup_data = {
 720                 .rpc_client = NFS_CLIENT(inode),
 721                 .rpc_message = &msg,
 722                 .callback_ops = &nfs_write_direct_ops,
 723                 .workqueue = nfsiod_workqueue,
 724                 .flags = RPC_TASK_ASYNC,
 725         };
 726         size_t wsize = NFS_SERVER(inode)->wsize;
 727         unsigned int pgbase;
 728         int result;
 729         ssize_t started = 0;
 730
 731         do {
 732                 struct nfs_write_data *data;
 733                 size_t bytes;
 734
 735                 pgbase = user_addr & ~PAGE_MASK;
 736                 bytes = min(wsize,count);
 737
 738                 result = -ENOMEM;
 739                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 740                 if (unlikely(!data))
 741                         break;
 742
 743                 down_read(&current->mm->mmap_sem);
 744                 result = get_user_pages(current, current->mm, user_addr,
 745                                         data->npages, 0, 0, data->pagevec, NULL);
 746                 up_read(&current->mm->mmap_sem);
 747                 if (result < 0) {
 748                         nfs_writedata_free(data);
 749                         break;
 750                 }
 751                 if ((unsigned)result < data->npages) {
 752                         bytes = result * PAGE_SIZE;
 753                         if (bytes <= pgbase) {
 754                                 nfs_direct_release_pages(data->pagevec, result);
 755                                 nfs_writedata_free(data);
 756                                 break;
 757                         }
 758                         bytes -= pgbase;
 759                         data->npages = result;
 760                 }
 761
 762                 get_dreq(dreq);
 763
 764                 list_move_tail(&data->pages, &dreq->rewrite_list);
 765
 766                 data->req = (struct nfs_page *) dreq;
 767                 data->inode = inode;
 768                 data->cred = msg.rpc_cred;
 769                 data->args.fh = NFS_FH(inode);
 770                 data->args.context = ctx;
 771                 data->args.lock_context = dreq->l_ctx;
 772                 data->args.offset = pos;
 773                 data->args.pgbase = pgbase;
 774                 data->args.pages = data->pagevec;
 775                 data->args.count = bytes;
 776                 data->args.stable = sync;
 777                 data->res.fattr = &data->fattr;
 778                 data->res.count = bytes;
 779                 data->res.verf = &data->verf;
 780                 nfs_fattr_init(&data->fattr);
 781
 782                 task_setup_data.task = &data->task;
 783                 task_setup_data.callback_data = data;
 784                 msg.rpc_argp = &data->args;
 785                 msg.rpc_resp = &data->res;
 786                 NFS_PROTO(inode)->write_setup(data, &msg);
 787
 788                 task = rpc_run_task(&task_setup_data);
 789                 if (IS_ERR(task))
 790                         break;
 791                 rpc_put_task(task);
 792
 793                 dprintk("NFS: %5u initiated direct write call "
 794                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 795                                 data->task.tk_pid,
 796                                 inode->i_sb->s_id,
 797                                 (long long)NFS_FILEID(inode),
 798                                 bytes,
 799                                 (unsigned long long)data->args.offset);
 800
 801                 started += bytes;
 802                 user_addr += bytes;
 803                 pos += bytes;
 804
 805                 /* FIXME: Remove this useless math from the final patch */
 806                 pgbase += bytes;
 807                 pgbase &= ~PAGE_MASK;
 808                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 809
 810                 count -= bytes;
 811         } while (count != 0);
 812
 813         if (started)
 814                 return started;
 815         return result < 0 ? (ssize_t) result : -EFAULT;
 816 }
 817
 818 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 819                                                const struct iovec *iov,
 820                                                unsigned long nr_segs,
 821                                                loff_t pos, int sync)
 822 {
 823         ssize_t result = 0;
 824         size_t requested_bytes = 0;
 825         unsigned long seg;
 826
 827         get_dreq(dreq);
 828
 829         for (seg = 0; seg < nr_segs; seg++) {
 830                 const struct iovec *vec = &iov[seg];
 831                 result = nfs_direct_write_schedule_segment(dreq, vec,
 832                                                            pos, sync);
 833                 if (result < 0)
 834                         break;
 835                 requested_bytes += result;
 836                 if ((size_t)result < vec->iov_len)
 837                         break;
 838                 pos += vec->iov_len;
 839         }
 840
 841         /*
 842          * If no bytes were started, return the error, and let the
 843          * generic layer handle the completion.
 844          */
 845         if (requested_bytes == 0) {
 846                 nfs_direct_req_release(dreq);
 847                 return result < 0 ? result : -EIO;
 848         }
 849
 850         if (put_dreq(dreq))
 851                 nfs_direct_write_complete(dreq, dreq->inode);
 852         return 0;
 853 }
 854
 855 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 856                                 unsigned long nr_segs, loff_t pos,
 857                                 size_t count)
 858 {
 859         ssize_t result = -ENOMEM;
 860         struct inode *inode = iocb->ki_filp->f_mapping->host;
 861         struct nfs_direct_req *dreq;
 862         size_t wsize = NFS_SERVER(inode)->wsize;
 863         int sync = NFS_UNSTABLE;
 864
 865         dreq = nfs_direct_req_alloc();
 866         if (!dreq)
 867                 goto out;
 868         nfs_alloc_commit_data(dreq);
 869
 870         if (dreq->commit_data == NULL || count <= wsize)
 871                 sync = NFS_FILE_SYNC;
 872
 873         dreq->inode = inode;
 874         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 875         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 876         if (dreq->l_ctx == NULL)
 877                 goto out_release;
 878         if (!is_sync_kiocb(iocb))
 879                 dreq->iocb = iocb;
 880
 881         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 882         if (!result)
 883                 result = nfs_direct_wait(dreq);
 884 out_release:
 885         nfs_direct_req_release(dreq);
 886 out:
 887         return result;
 888 }
 889
 890 /**
 891  * nfs_file_direct_read - file direct read operation for NFS files
 892  * @iocb: target I/O control block
 893  * @iov: vector of user buffers into which to read data
 894  * @nr_segs: size of iov vector
 895  * @pos: byte offset in file where reading starts
 896  *
 897  * We use this function for direct reads instead of calling
 898  * generic_file_aio_read() in order to avoid gfar's check to see if
 899  * the request starts before the end of the file.  For that check
 900  * to work, we must generate a GETATTR before each direct read, and
 901  * even then there is a window between the GETATTR and the subsequent
 902  * READ where the file size could change.  Our preference is simply
 903  * to do all reads the application wants, and the server will take
 904  * care of managing the end of file boundary.
 905  *
 906  * This function also eliminates unnecessarily updating the file's
 907  * atime locally, as the NFS server sets the file's atime, and this
 908  * client must read the updated atime from the server back into its
 909  * cache.
 910  */
 911 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 912                                 unsigned long nr_segs, loff_t pos)
 913 {
 914         ssize_t retval = -EINVAL;
 915         struct file *file = iocb->ki_filp;
 916         struct address_space *mapping = file->f_mapping;
 917         size_t count;
 918
 919         count = iov_length(iov, nr_segs);
 920         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 921
 922         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 923                 file->f_path.dentry->d_parent->d_name.name,
 924                 file->f_path.dentry->d_name.name,
 925                 count, (long long) pos);
 926
 927         retval = 0;
 928         if (!count)
 929                 goto out;
 930
 931         retval = nfs_sync_mapping(mapping);
 932         if (retval)
 933                 goto out;
 934
 935         task_io_account_read(count);
 936
 937         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 938         if (retval > 0)
 939                 iocb->ki_pos = pos + retval;
 940
 941 out:
 942         return retval;
 943 }
 944
 945 /**
 946  * nfs_file_direct_write - file direct write operation for NFS files
 947  * @iocb: target I/O control block
 948  * @iov: vector of user buffers from which to write data
 949  * @nr_segs: size of iov vector
 950  * @pos: byte offset in file where writing starts
 951  *
 952  * We use this function for direct writes instead of calling
 953  * generic_file_aio_write() in order to avoid taking the inode
 954  * semaphore and updating the i_size.  The NFS server will set
 955  * the new i_size and this client must read the updated size
 956  * back into its cache.  We let the server do generic write
 957  * parameter checking and report problems.
 958  *
 959  * We eliminate local atime updates, see direct read above.
 960  *
 961  * We avoid unnecessary page cache invalidations for normal cached
 962  * readers of this file.
 963  *
 964  * Note that O_APPEND is not supported for NFS direct writes, as there
 965  * is no atomic O_APPEND write facility in the NFS protocol.
 966  */
 967 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 968                                 unsigned long nr_segs, loff_t pos)
 969 {
 970         ssize_t retval = -EINVAL;
 971         struct file *file = iocb->ki_filp;
 972         struct address_space *mapping = file->f_mapping;
 973         size_t count;
 974
 975         count = iov_length(iov, nr_segs);
 976         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 977
 978         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 979                 file->f_path.dentry->d_parent->d_name.name,
 980                 file->f_path.dentry->d_name.name,
 981                 count, (long long) pos);
 982
 983         retval = generic_write_checks(file, &pos, &count, 0);
 984         if (retval)
 985                 goto out;
 986
 987         retval = -EINVAL;
 988         if ((ssize_t) count < 0)
 989                 goto out;
 990         retval = 0;
 991         if (!count)
 992                 goto out;
 993
 994         retval = nfs_sync_mapping(mapping);
 995         if (retval)
 996                 goto out;
 997
 998         task_io_account_write(count);
 999
1000         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
1001
1002         if (retval > 0)
1003                 iocb->ki_pos = pos + retval;
1004
1005 out:
1006         return retval;
1007 }
1008
1009 /**
1010  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1011  *
1012  */
1013 int __init nfs_init_directcache(void)
1014 {
1015         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1016                                                 sizeof(struct nfs_direct_req),
1017                                                 0, (SLAB_RECLAIM_ACCOUNT|
1018                                                         SLAB_MEM_SPREAD),
1019                                                 NULL);
1020         if (nfs_direct_cachep == NULL)
1021                 return -ENOMEM;
1022
1023         return 0;
1024 }
1025
1026 /**
1027  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1028  *
1029  */
1030 void nfs_destroy_directcache(void)
1031 {
1032         kmem_cache_destroy(nfs_direct_cachep);
1033 }