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