SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47
  48 #include <linux/nfs_fs.h>
  49 #include <linux/nfs_page.h>
  50 #include <linux/sunrpc/clnt.h>
  51
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54 #include <asm/atomic.h>
  55
  56 #include "internal.h"
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60
  61 static struct kmem_cache *nfs_direct_cachep;
  62
  63 /*
  64  * This represents a set of asynchronous requests that we're waiting on
  65  */
  66 struct nfs_direct_req {
  67         struct kref             kref;           /* release manager */
  68
  69         /* I/O parameters */
  70         struct nfs_open_context *ctx;           /* file open context info */
  71         struct kiocb *          iocb;           /* controlling i/o request */
  72         struct inode *          inode;          /* target file of i/o */
  73
  74         /* completion state */
  75         atomic_t                io_count;       /* i/os we're waiting for */
  76         spinlock_t              lock;           /* protect completion state */
  77         ssize_t                 count,          /* bytes actually processed */
  78                                 error;          /* any reported error */
  79         struct completion       completion;     /* wait for i/o completion */
  80
  81         /* commit state */
  82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  83         struct nfs_write_data * commit_data;    /* special write_data for commits */
  84         int                     flags;
  85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  87         struct nfs_writeverf    verf;           /* unstable write verifier */
  88 };
  89
  90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  91 static const struct rpc_call_ops nfs_write_direct_ops;
  92
  93 static inline void get_dreq(struct nfs_direct_req *dreq)
  94 {
  95         atomic_inc(&dreq->io_count);
  96 }
  97
  98 static inline int put_dreq(struct nfs_direct_req *dreq)
  99 {
 100         return atomic_dec_and_test(&dreq->io_count);
 101 }
 102
 103 /**
 104  * nfs_direct_IO - NFS address space operation for direct I/O
 105  * @rw: direction (read or write)
 106  * @iocb: target I/O control block
 107  * @iov: array of vectors that define I/O buffer
 108  * @pos: offset in file to begin the operation
 109  * @nr_segs: size of iovec array
 110  *
 111  * The presence of this routine in the address space ops vector means
 112  * the NFS client supports direct I/O.  However, we shunt off direct
 113  * read and write requests before the VFS gets them, so this method
 114  * should never be called.
 115  */
 116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 117 {
 118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 119                         iocb->ki_filp->f_path.dentry->d_name.name,
 120                         (long long) pos, nr_segs);
 121
 122         return -EINVAL;
 123 }
 124
 125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
 126 {
 127         unsigned int npages;
 128         unsigned int i;
 129
 130         if (count == 0)
 131                 return;
 132         pages += (pgbase >> PAGE_SHIFT);
 133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 134         for (i = 0; i < npages; i++) {
 135                 struct page *page = pages[i];
 136                 if (!PageCompound(page))
 137                         set_page_dirty(page);
 138         }
 139 }
 140
 141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 142 {
 143         unsigned int i;
 144         for (i = 0; i < npages; i++)
 145                 page_cache_release(pages[i]);
 146 }
 147
 148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 149 {
 150         struct nfs_direct_req *dreq;
 151
 152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 153         if (!dreq)
 154                 return NULL;
 155
 156         kref_init(&dreq->kref);
 157         kref_get(&dreq->kref);
 158         init_completion(&dreq->completion);
 159         INIT_LIST_HEAD(&dreq->rewrite_list);
 160         dreq->iocb = NULL;
 161         dreq->ctx = NULL;
 162         spin_lock_init(&dreq->lock);
 163         atomic_set(&dreq->io_count, 0);
 164         dreq->count = 0;
 165         dreq->error = 0;
 166         dreq->flags = 0;
 167
 168         return dreq;
 169 }
 170
 171 static void nfs_direct_req_free(struct kref *kref)
 172 {
 173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 174
 175         if (dreq->ctx != NULL)
 176                 put_nfs_open_context(dreq->ctx);
 177         kmem_cache_free(nfs_direct_cachep, dreq);
 178 }
 179
 180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 181 {
 182         kref_put(&dreq->kref, nfs_direct_req_free);
 183 }
 184
 185 /*
 186  * Collects and returns the final error value/byte-count.
 187  */
 188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 189 {
 190         ssize_t result = -EIOCBQUEUED;
 191
 192         /* Async requests don't wait here */
 193         if (dreq->iocb)
 194                 goto out;
 195
 196         result = wait_for_completion_interruptible(&dreq->completion);
 197
 198         if (!result)
 199                 result = dreq->error;
 200         if (!result)
 201                 result = dreq->count;
 202
 203 out:
 204         return (ssize_t) result;
 205 }
 206
 207 /*
 208  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 209  * the iocb is still valid here if this is a synchronous request.
 210  */
 211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 212 {
 213         if (dreq->iocb) {
 214                 long res = (long) dreq->error;
 215                 if (!res)
 216                         res = (long) dreq->count;
 217                 aio_complete(dreq->iocb, res, 0);
 218         }
 219         complete_all(&dreq->completion);
 220
 221         nfs_direct_req_release(dreq);
 222 }
 223
 224 /*
 225  * We must hold a reference to all the pages in this direct read request
 226  * until the RPCs complete.  This could be long *after* we are woken up in
 227  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 228  */
 229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 230 {
 231         struct nfs_read_data *data = calldata;
 232         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 233
 234         if (nfs_readpage_result(task, data) != 0)
 235                 return;
 236
 237         spin_lock(&dreq->lock);
 238         if (unlikely(task->tk_status < 0)) {
 239                 dreq->error = task->tk_status;
 240                 spin_unlock(&dreq->lock);
 241         } else {
 242                 dreq->count += data->res.count;
 243                 spin_unlock(&dreq->lock);
 244                 nfs_direct_dirty_pages(data->pagevec,
 245                                 data->args.pgbase,
 246                                 data->res.count);
 247         }
 248         nfs_direct_release_pages(data->pagevec, data->npages);
 249
 250         if (put_dreq(dreq))
 251                 nfs_direct_complete(dreq);
 252 }
 253
 254 static const struct rpc_call_ops nfs_read_direct_ops = {
 255         .rpc_call_done = nfs_direct_read_result,
 256         .rpc_release = nfs_readdata_release,
 257 };
 258
 259 /*
 260  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 261  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 262  * bail and stop sending more reads.  Read length accounting is
 263  * handled automatically by nfs_direct_read_result().  Otherwise, if
 264  * no requests have been sent, just return an error.
 265  */
 266 static ssize_t nfs_direct_read_schedule_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_message msg = {
 276                 .rpc_cred = ctx->cred,
 277         };
 278         struct rpc_task_setup task_setup_data = {
 279                 .rpc_client = NFS_CLIENT(inode),
 280                 .rpc_message = &msg,
 281                 .callback_ops = &nfs_read_direct_ops,
 282                 .flags = RPC_TASK_ASYNC,
 283         };
 284         unsigned int pgbase;
 285         int result;
 286         ssize_t started = 0;
 287
 288         do {
 289                 struct nfs_read_data *data;
 290                 size_t bytes;
 291
 292                 pgbase = user_addr & ~PAGE_MASK;
 293                 bytes = min(rsize,count);
 294
 295                 result = -ENOMEM;
 296                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 297                 if (unlikely(!data))
 298                         break;
 299
 300                 down_read(&current->mm->mmap_sem);
 301                 result = get_user_pages(current, current->mm, user_addr,
 302                                         data->npages, 1, 0, data->pagevec, NULL);
 303                 up_read(&current->mm->mmap_sem);
 304                 if (result < 0) {
 305                         nfs_readdata_release(data);
 306                         break;
 307                 }
 308                 if ((unsigned)result < data->npages) {
 309                         bytes = result * PAGE_SIZE;
 310                         if (bytes <= pgbase) {
 311                                 nfs_direct_release_pages(data->pagevec, result);
 312                                 nfs_readdata_release(data);
 313                                 break;
 314                         }
 315                         bytes -= pgbase;
 316                         data->npages = result;
 317                 }
 318
 319                 get_dreq(dreq);
 320
 321                 data->req = (struct nfs_page *) dreq;
 322                 data->inode = inode;
 323                 data->cred = msg.rpc_cred;
 324                 data->args.fh = NFS_FH(inode);
 325                 data->args.context = ctx;
 326                 data->args.offset = pos;
 327                 data->args.pgbase = pgbase;
 328                 data->args.pages = data->pagevec;
 329                 data->args.count = bytes;
 330                 data->res.fattr = &data->fattr;
 331                 data->res.eof = 0;
 332                 data->res.count = bytes;
 333                 msg.rpc_argp = &data->args;
 334                 msg.rpc_resp = &data->res;
 335
 336                 task_setup_data.callback_data = data;
 337                 NFS_PROTO(inode)->read_setup(data, &msg);
 338                 rpc_init_task(&data->task, &task_setup_data);
 339
 340                 rpc_execute(&data->task);
 341
 342                 dprintk("NFS: %5u initiated direct read call "
 343                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 344                                 data->task.tk_pid,
 345                                 inode->i_sb->s_id,
 346                                 (long long)NFS_FILEID(inode),
 347                                 bytes,
 348                                 (unsigned long long)data->args.offset);
 349
 350                 started += bytes;
 351                 user_addr += bytes;
 352                 pos += bytes;
 353                 /* FIXME: Remove this unnecessary math from final patch */
 354                 pgbase += bytes;
 355                 pgbase &= ~PAGE_MASK;
 356                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 357
 358                 count -= bytes;
 359         } while (count != 0);
 360
 361         if (started)
 362                 return started;
 363         return result < 0 ? (ssize_t) result : -EFAULT;
 364 }
 365
 366 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 367                                               const struct iovec *iov,
 368                                               unsigned long nr_segs,
 369                                               loff_t pos)
 370 {
 371         ssize_t result = -EINVAL;
 372         size_t requested_bytes = 0;
 373         unsigned long seg;
 374
 375         get_dreq(dreq);
 376
 377         for (seg = 0; seg < nr_segs; seg++) {
 378                 const struct iovec *vec = &iov[seg];
 379                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
 380                 if (result < 0)
 381                         break;
 382                 requested_bytes += result;
 383                 if ((size_t)result < vec->iov_len)
 384                         break;
 385                 pos += vec->iov_len;
 386         }
 387
 388         if (put_dreq(dreq))
 389                 nfs_direct_complete(dreq);
 390
 391         if (requested_bytes != 0)
 392                 return 0;
 393
 394         if (result < 0)
 395                 return result;
 396         return -EIO;
 397 }
 398
 399 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 400                                unsigned long nr_segs, loff_t pos)
 401 {
 402         ssize_t result = 0;
 403         sigset_t oldset;
 404         struct inode *inode = iocb->ki_filp->f_mapping->host;
 405         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 406         struct nfs_direct_req *dreq;
 407
 408         dreq = nfs_direct_req_alloc();
 409         if (!dreq)
 410                 return -ENOMEM;
 411
 412         dreq->inode = inode;
 413         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 414         if (!is_sync_kiocb(iocb))
 415                 dreq->iocb = iocb;
 416
 417         rpc_clnt_sigmask(clnt, &oldset);
 418         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 419         if (!result)
 420                 result = nfs_direct_wait(dreq);
 421         rpc_clnt_sigunmask(clnt, &oldset);
 422         nfs_direct_req_release(dreq);
 423
 424         return result;
 425 }
 426
 427 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 428 {
 429         while (!list_empty(&dreq->rewrite_list)) {
 430                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 431                 list_del(&data->pages);
 432                 nfs_direct_release_pages(data->pagevec, data->npages);
 433                 nfs_writedata_release(data);
 434         }
 435 }
 436
 437 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 438 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 439 {
 440         struct inode *inode = dreq->inode;
 441         struct list_head *p;
 442         struct nfs_write_data *data;
 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.callback_data = data;
 475                 msg.rpc_argp = &data->args;
 476                 msg.rpc_resp = &data->res;
 477                 NFS_PROTO(inode)->write_setup(data, &msg);
 478                 rpc_init_task(&data->task, &task_setup_data);
 479
 480                 /*
 481                  * We're called via an RPC callback, so BKL is already held.
 482                  */
 483                 rpc_execute(&data->task);
 484
 485                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 486                                 data->task.tk_pid,
 487                                 inode->i_sb->s_id,
 488                                 (long long)NFS_FILEID(inode),
 489                                 data->args.count,
 490                                 (unsigned long long)data->args.offset);
 491         }
 492
 493         if (put_dreq(dreq))
 494                 nfs_direct_write_complete(dreq, inode);
 495 }
 496
 497 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 498 {
 499         struct nfs_write_data *data = calldata;
 500         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 501
 502         /* Call the NFS version-specific code */
 503         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 504                 return;
 505         if (unlikely(task->tk_status < 0)) {
 506                 dprintk("NFS: %5u commit failed with error %d.\n",
 507                                 task->tk_pid, task->tk_status);
 508                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 509         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 510                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 511                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 512         }
 513
 514         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 515         nfs_direct_write_complete(dreq, data->inode);
 516 }
 517
 518 static const struct rpc_call_ops nfs_commit_direct_ops = {
 519         .rpc_call_done = nfs_direct_commit_result,
 520         .rpc_release = nfs_commit_release,
 521 };
 522
 523 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 524 {
 525         struct nfs_write_data *data = dreq->commit_data;
 526         struct rpc_message msg = {
 527                 .rpc_argp = &data->args,
 528                 .rpc_resp = &data->res,
 529                 .rpc_cred = dreq->ctx->cred,
 530         };
 531         struct rpc_task_setup task_setup_data = {
 532                 .rpc_client = NFS_CLIENT(dreq->inode),
 533                 .rpc_message = &msg,
 534                 .callback_ops = &nfs_commit_direct_ops,
 535                 .callback_data = data,
 536                 .flags = RPC_TASK_ASYNC,
 537         };
 538
 539         data->inode = dreq->inode;
 540         data->cred = msg.rpc_cred;
 541
 542         data->args.fh = NFS_FH(data->inode);
 543         data->args.offset = 0;
 544         data->args.count = 0;
 545         data->res.count = 0;
 546         data->res.fattr = &data->fattr;
 547         data->res.verf = &data->verf;
 548
 549         NFS_PROTO(data->inode)->commit_setup(data, &msg);
 550         rpc_init_task(&data->task, &task_setup_data);
 551
 552         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 553         dreq->commit_data = NULL;
 554
 555         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 556
 557         rpc_execute(&data->task);
 558 }
 559
 560 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 561 {
 562         int flags = dreq->flags;
 563
 564         dreq->flags = 0;
 565         switch (flags) {
 566                 case NFS_ODIRECT_DO_COMMIT:
 567                         nfs_direct_commit_schedule(dreq);
 568                         break;
 569                 case NFS_ODIRECT_RESCHED_WRITES:
 570                         nfs_direct_write_reschedule(dreq);
 571                         break;
 572                 default:
 573                         if (dreq->commit_data != NULL)
 574                                 nfs_commit_free(dreq->commit_data);
 575                         nfs_direct_free_writedata(dreq);
 576                         nfs_zap_mapping(inode, inode->i_mapping);
 577                         nfs_direct_complete(dreq);
 578         }
 579 }
 580
 581 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 582 {
 583         dreq->commit_data = nfs_commit_alloc();
 584         if (dreq->commit_data != NULL)
 585                 dreq->commit_data->req = (struct nfs_page *) dreq;
 586 }
 587 #else
 588 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 589 {
 590         dreq->commit_data = NULL;
 591 }
 592
 593 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 594 {
 595         nfs_direct_free_writedata(dreq);
 596         nfs_zap_mapping(inode, inode->i_mapping);
 597         nfs_direct_complete(dreq);
 598 }
 599 #endif
 600
 601 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 602 {
 603         struct nfs_write_data *data = calldata;
 604         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 605         int status = task->tk_status;
 606
 607         if (nfs_writeback_done(task, data) != 0)
 608                 return;
 609
 610         spin_lock(&dreq->lock);
 611
 612         if (unlikely(status < 0)) {
 613                 /* An error has occurred, so we should not commit */
 614                 dreq->flags = 0;
 615                 dreq->error = status;
 616         }
 617         if (unlikely(dreq->error != 0))
 618                 goto out_unlock;
 619
 620         dreq->count += data->res.count;
 621
 622         if (data->res.verf->committed != NFS_FILE_SYNC) {
 623                 switch (dreq->flags) {
 624                         case 0:
 625                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 626                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 627                                 break;
 628                         case NFS_ODIRECT_DO_COMMIT:
 629                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 630                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 631                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 632                                 }
 633                 }
 634         }
 635 out_unlock:
 636         spin_unlock(&dreq->lock);
 637 }
 638
 639 /*
 640  * NB: Return the value of the first error return code.  Subsequent
 641  *     errors after the first one are ignored.
 642  */
 643 static void nfs_direct_write_release(void *calldata)
 644 {
 645         struct nfs_write_data *data = calldata;
 646         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 647
 648         if (put_dreq(dreq))
 649                 nfs_direct_write_complete(dreq, data->inode);
 650 }
 651
 652 static const struct rpc_call_ops nfs_write_direct_ops = {
 653         .rpc_call_done = nfs_direct_write_result,
 654         .rpc_release = nfs_direct_write_release,
 655 };
 656
 657 /*
 658  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 659  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 660  * bail and stop sending more writes.  Write length accounting is
 661  * handled automatically by nfs_direct_write_result().  Otherwise, if
 662  * no requests have been sent, just return an error.
 663  */
 664 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
 665                                                  const struct iovec *iov,
 666                                                  loff_t pos, int sync)
 667 {
 668         struct nfs_open_context *ctx = dreq->ctx;
 669         struct inode *inode = ctx->path.dentry->d_inode;
 670         unsigned long user_addr = (unsigned long)iov->iov_base;
 671         size_t count = iov->iov_len;
 672         struct rpc_message msg = {
 673                 .rpc_cred = ctx->cred,
 674         };
 675         struct rpc_task_setup task_setup_data = {
 676                 .rpc_client = NFS_CLIENT(inode),
 677                 .rpc_message = &msg,
 678                 .callback_ops = &nfs_write_direct_ops,
 679                 .flags = RPC_TASK_ASYNC,
 680         };
 681         size_t wsize = NFS_SERVER(inode)->wsize;
 682         unsigned int pgbase;
 683         int result;
 684         ssize_t started = 0;
 685
 686         do {
 687                 struct nfs_write_data *data;
 688                 size_t bytes;
 689
 690                 pgbase = user_addr & ~PAGE_MASK;
 691                 bytes = min(wsize,count);
 692
 693                 result = -ENOMEM;
 694                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 695                 if (unlikely(!data))
 696                         break;
 697
 698                 down_read(&current->mm->mmap_sem);
 699                 result = get_user_pages(current, current->mm, user_addr,
 700                                         data->npages, 0, 0, data->pagevec, NULL);
 701                 up_read(&current->mm->mmap_sem);
 702                 if (result < 0) {
 703                         nfs_writedata_release(data);
 704                         break;
 705                 }
 706                 if ((unsigned)result < data->npages) {
 707                         bytes = result * PAGE_SIZE;
 708                         if (bytes <= pgbase) {
 709                                 nfs_direct_release_pages(data->pagevec, result);
 710                                 nfs_writedata_release(data);
 711                                 break;
 712                         }
 713                         bytes -= pgbase;
 714                         data->npages = result;
 715                 }
 716
 717                 get_dreq(dreq);
 718
 719                 list_move_tail(&data->pages, &dreq->rewrite_list);
 720
 721                 data->req = (struct nfs_page *) dreq;
 722                 data->inode = inode;
 723                 data->cred = msg.rpc_cred;
 724                 data->args.fh = NFS_FH(inode);
 725                 data->args.context = ctx;
 726                 data->args.offset = pos;
 727                 data->args.pgbase = pgbase;
 728                 data->args.pages = data->pagevec;
 729                 data->args.count = bytes;
 730                 data->args.stable = sync;
 731                 data->res.fattr = &data->fattr;
 732                 data->res.count = bytes;
 733                 data->res.verf = &data->verf;
 734
 735                 task_setup_data.callback_data = data;
 736                 msg.rpc_argp = &data->args;
 737                 msg.rpc_resp = &data->res;
 738                 NFS_PROTO(inode)->write_setup(data, &msg);
 739                 rpc_init_task(&data->task, &task_setup_data);
 740
 741                 rpc_execute(&data->task);
 742
 743                 dprintk("NFS: %5u initiated direct write call "
 744                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 745                                 data->task.tk_pid,
 746                                 inode->i_sb->s_id,
 747                                 (long long)NFS_FILEID(inode),
 748                                 bytes,
 749                                 (unsigned long long)data->args.offset);
 750
 751                 started += bytes;
 752                 user_addr += bytes;
 753                 pos += bytes;
 754
 755                 /* FIXME: Remove this useless math from the final patch */
 756                 pgbase += bytes;
 757                 pgbase &= ~PAGE_MASK;
 758                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 759
 760                 count -= bytes;
 761         } while (count != 0);
 762
 763         if (started)
 764                 return started;
 765         return result < 0 ? (ssize_t) result : -EFAULT;
 766 }
 767
 768 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 769                                                const struct iovec *iov,
 770                                                unsigned long nr_segs,
 771                                                loff_t pos, int sync)
 772 {
 773         ssize_t result = 0;
 774         size_t requested_bytes = 0;
 775         unsigned long seg;
 776
 777         get_dreq(dreq);
 778
 779         for (seg = 0; seg < nr_segs; seg++) {
 780                 const struct iovec *vec = &iov[seg];
 781                 result = nfs_direct_write_schedule_segment(dreq, vec,
 782                                                            pos, sync);
 783                 if (result < 0)
 784                         break;
 785                 requested_bytes += result;
 786                 if ((size_t)result < vec->iov_len)
 787                         break;
 788                 pos += vec->iov_len;
 789         }
 790
 791         if (put_dreq(dreq))
 792                 nfs_direct_write_complete(dreq, dreq->inode);
 793
 794         if (requested_bytes != 0)
 795                 return 0;
 796
 797         if (result < 0)
 798                 return result;
 799         return -EIO;
 800 }
 801
 802 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 803                                 unsigned long nr_segs, loff_t pos,
 804                                 size_t count)
 805 {
 806         ssize_t result = 0;
 807         sigset_t oldset;
 808         struct inode *inode = iocb->ki_filp->f_mapping->host;
 809         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 810         struct nfs_direct_req *dreq;
 811         size_t wsize = NFS_SERVER(inode)->wsize;
 812         int sync = NFS_UNSTABLE;
 813
 814         dreq = nfs_direct_req_alloc();
 815         if (!dreq)
 816                 return -ENOMEM;
 817         nfs_alloc_commit_data(dreq);
 818
 819         if (dreq->commit_data == NULL || count < wsize)
 820                 sync = NFS_FILE_SYNC;
 821
 822         dreq->inode = inode;
 823         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 824         if (!is_sync_kiocb(iocb))
 825                 dreq->iocb = iocb;
 826
 827         rpc_clnt_sigmask(clnt, &oldset);
 828         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 829         if (!result)
 830                 result = nfs_direct_wait(dreq);
 831         rpc_clnt_sigunmask(clnt, &oldset);
 832         nfs_direct_req_release(dreq);
 833
 834         return result;
 835 }
 836
 837 /**
 838  * nfs_file_direct_read - file direct read operation for NFS files
 839  * @iocb: target I/O control block
 840  * @iov: vector of user buffers into which to read data
 841  * @nr_segs: size of iov vector
 842  * @pos: byte offset in file where reading starts
 843  *
 844  * We use this function for direct reads instead of calling
 845  * generic_file_aio_read() in order to avoid gfar's check to see if
 846  * the request starts before the end of the file.  For that check
 847  * to work, we must generate a GETATTR before each direct read, and
 848  * even then there is a window between the GETATTR and the subsequent
 849  * READ where the file size could change.  Our preference is simply
 850  * to do all reads the application wants, and the server will take
 851  * care of managing the end of file boundary.
 852  *
 853  * This function also eliminates unnecessarily updating the file's
 854  * atime locally, as the NFS server sets the file's atime, and this
 855  * client must read the updated atime from the server back into its
 856  * cache.
 857  */
 858 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 859                                 unsigned long nr_segs, loff_t pos)
 860 {
 861         ssize_t retval = -EINVAL;
 862         struct file *file = iocb->ki_filp;
 863         struct address_space *mapping = file->f_mapping;
 864         size_t count;
 865
 866         count = iov_length(iov, nr_segs);
 867         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 868
 869         dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
 870                 file->f_path.dentry->d_parent->d_name.name,
 871                 file->f_path.dentry->d_name.name,
 872                 count, (long long) pos);
 873
 874         retval = 0;
 875         if (!count)
 876                 goto out;
 877
 878         retval = nfs_sync_mapping(mapping);
 879         if (retval)
 880                 goto out;
 881
 882         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 883         if (retval > 0)
 884                 iocb->ki_pos = pos + retval;
 885
 886 out:
 887         return retval;
 888 }
 889
 890 /**
 891  * nfs_file_direct_write - file direct write operation for NFS files
 892  * @iocb: target I/O control block
 893  * @iov: vector of user buffers from which to write data
 894  * @nr_segs: size of iov vector
 895  * @pos: byte offset in file where writing starts
 896  *
 897  * We use this function for direct writes instead of calling
 898  * generic_file_aio_write() in order to avoid taking the inode
 899  * semaphore and updating the i_size.  The NFS server will set
 900  * the new i_size and this client must read the updated size
 901  * back into its cache.  We let the server do generic write
 902  * parameter checking and report problems.
 903  *
 904  * We also avoid an unnecessary invocation of generic_osync_inode(),
 905  * as it is fairly meaningless to sync the metadata of an NFS file.
 906  *
 907  * We eliminate local atime updates, see direct read above.
 908  *
 909  * We avoid unnecessary page cache invalidations for normal cached
 910  * readers of this file.
 911  *
 912  * Note that O_APPEND is not supported for NFS direct writes, as there
 913  * is no atomic O_APPEND write facility in the NFS protocol.
 914  */
 915 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 916                                 unsigned long nr_segs, loff_t pos)
 917 {
 918         ssize_t retval = -EINVAL;
 919         struct file *file = iocb->ki_filp;
 920         struct address_space *mapping = file->f_mapping;
 921         size_t count;
 922
 923         count = iov_length(iov, nr_segs);
 924         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 925
 926         dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
 927                 file->f_path.dentry->d_parent->d_name.name,
 928                 file->f_path.dentry->d_name.name,
 929                 count, (long long) pos);
 930
 931         retval = generic_write_checks(file, &pos, &count, 0);
 932         if (retval)
 933                 goto out;
 934
 935         retval = -EINVAL;
 936         if ((ssize_t) count < 0)
 937                 goto out;
 938         retval = 0;
 939         if (!count)
 940                 goto out;
 941
 942         retval = nfs_sync_mapping(mapping);
 943         if (retval)
 944                 goto out;
 945
 946         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
 947
 948         if (retval > 0)
 949                 iocb->ki_pos = pos + retval;
 950
 951 out:
 952         return retval;
 953 }
 954
 955 /**
 956  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 957  *
 958  */
 959 int __init nfs_init_directcache(void)
 960 {
 961         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 962                                                 sizeof(struct nfs_direct_req),
 963                                                 0, (SLAB_RECLAIM_ACCOUNT|
 964                                                         SLAB_MEM_SPREAD),
 965                                                 NULL);
 966         if (nfs_direct_cachep == NULL)
 967                 return -ENOMEM;
 968
 969         return 0;
 970 }
 971
 972 /**
 973  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 974  *
 975  */
 976 void nfs_destroy_directcache(void)
 977 {
 978         kmem_cache_destroy(nfs_direct_cachep);
 979 }