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

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47
  48 #include <linux/nfs_fs.h>
  49 #include <linux/nfs_page.h>
  50 #include <linux/sunrpc/clnt.h>
  51
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54 #include <asm/atomic.h>
  55
  56 #include "internal.h"
  57 #include "iostat.h"
  58
  59 #define NFSDBG_FACILITY         NFSDBG_VFS
  60
  61 static struct kmem_cache *nfs_direct_cachep;
  62
  63 /*
  64  * This represents a set of asynchronous requests that we're waiting on
  65  */
  66 struct nfs_direct_req {
  67         struct kref             kref;           /* release manager */
  68
  69         /* I/O parameters */
  70         struct nfs_open_context *ctx;           /* file open context info */
  71         struct kiocb *          iocb;           /* controlling i/o request */
  72         struct inode *          inode;          /* target file of i/o */
  73
  74         /* completion state */
  75         atomic_t                io_count;       /* i/os we're waiting for */
  76         spinlock_t              lock;           /* protect completion state */
  77         ssize_t                 count,          /* bytes actually processed */
  78                                 error;          /* any reported error */
  79         struct completion       completion;     /* wait for i/o completion */
  80
  81         /* commit state */
  82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  83         struct nfs_write_data * commit_data;    /* special write_data for commits */
  84         int                     flags;
  85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  87         struct nfs_writeverf    verf;           /* unstable write verifier */
  88 };
  89
  90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  91 static const struct rpc_call_ops nfs_write_direct_ops;
  92
  93 static inline void get_dreq(struct nfs_direct_req *dreq)
  94 {
  95         atomic_inc(&dreq->io_count);
  96 }
  97
  98 static inline int put_dreq(struct nfs_direct_req *dreq)
  99 {
 100         return atomic_dec_and_test(&dreq->io_count);
 101 }
 102
 103 /**
 104  * nfs_direct_IO - NFS address space operation for direct I/O
 105  * @rw: direction (read or write)
 106  * @iocb: target I/O control block
 107  * @iov: array of vectors that define I/O buffer
 108  * @pos: offset in file to begin the operation
 109  * @nr_segs: size of iovec array
 110  *
 111  * The presence of this routine in the address space ops vector means
 112  * the NFS client supports direct I/O.  However, we shunt off direct
 113  * read and write requests before the VFS gets them, so this method
 114  * should never be called.
 115  */
 116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 117 {
 118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 119                         iocb->ki_filp->f_path.dentry->d_name.name,
 120                         (long long) pos, nr_segs);
 121
 122         return -EINVAL;
 123 }
 124
 125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
 126 {
 127         unsigned int npages;
 128         unsigned int i;
 129
 130         if (count == 0)
 131                 return;
 132         pages += (pgbase >> PAGE_SHIFT);
 133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 134         for (i = 0; i < npages; i++) {
 135                 struct page *page = pages[i];
 136                 if (!PageCompound(page))
 137                         set_page_dirty(page);
 138         }
 139 }
 140
 141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 142 {
 143         unsigned int i;
 144         for (i = 0; i < npages; i++)
 145                 page_cache_release(pages[i]);
 146 }
 147
 148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 149 {
 150         struct nfs_direct_req *dreq;
 151
 152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 153         if (!dreq)
 154                 return NULL;
 155
 156         kref_init(&dreq->kref);
 157         kref_get(&dreq->kref);
 158         init_completion(&dreq->completion);
 159         INIT_LIST_HEAD(&dreq->rewrite_list);
 160         dreq->iocb = NULL;
 161         dreq->ctx = NULL;
 162         spin_lock_init(&dreq->lock);
 163         atomic_set(&dreq->io_count, 0);
 164         dreq->count = 0;
 165         dreq->error = 0;
 166         dreq->flags = 0;
 167
 168         return dreq;
 169 }
 170
 171 static void nfs_direct_req_free(struct kref *kref)
 172 {
 173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 174
 175         if (dreq->ctx != NULL)
 176                 put_nfs_open_context(dreq->ctx);
 177         kmem_cache_free(nfs_direct_cachep, dreq);
 178 }
 179
 180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 181 {
 182         kref_put(&dreq->kref, nfs_direct_req_free);
 183 }
 184
 185 /*
 186  * Collects and returns the final error value/byte-count.
 187  */
 188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 189 {
 190         ssize_t result = -EIOCBQUEUED;
 191
 192         /* Async requests don't wait here */
 193         if (dreq->iocb)
 194                 goto out;
 195
 196         result = wait_for_completion_interruptible(&dreq->completion);
 197
 198         if (!result)
 199                 result = dreq->error;
 200         if (!result)
 201                 result = dreq->count;
 202
 203 out:
 204         return (ssize_t) result;
 205 }
 206
 207 /*
 208  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 209  * the iocb is still valid here if this is a synchronous request.
 210  */
 211 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 212 {
 213         if (dreq->iocb) {
 214                 long res = (long) dreq->error;
 215                 if (!res)
 216                         res = (long) dreq->count;
 217                 aio_complete(dreq->iocb, res, 0);
 218         }
 219         complete_all(&dreq->completion);
 220
 221         nfs_direct_req_release(dreq);
 222 }
 223
 224 /*
 225  * We must hold a reference to all the pages in this direct read request
 226  * until the RPCs complete.  This could be long *after* we are woken up in
 227  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 228  */
 229 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 230 {
 231         struct nfs_read_data *data = calldata;
 232         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 233
 234         if (nfs_readpage_result(task, data) != 0)
 235                 return;
 236
 237         spin_lock(&dreq->lock);
 238         if (unlikely(task->tk_status < 0)) {
 239                 dreq->error = task->tk_status;
 240                 spin_unlock(&dreq->lock);
 241         } else {
 242                 dreq->count += data->res.count;
 243                 spin_unlock(&dreq->lock);
 244                 nfs_direct_dirty_pages(data->pagevec,
 245                                 data->args.pgbase,
 246                                 data->res.count);
 247         }
 248         nfs_direct_release_pages(data->pagevec, data->npages);
 249
 250         if (put_dreq(dreq))
 251                 nfs_direct_complete(dreq);
 252 }
 253
 254 static const struct rpc_call_ops nfs_read_direct_ops = {
 255         .rpc_call_done = nfs_direct_read_result,
 256         .rpc_release = nfs_readdata_release,
 257 };
 258
 259 /*
 260  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 261  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 262  * bail and stop sending more reads.  Read length accounting is
 263  * handled automatically by nfs_direct_read_result().  Otherwise, if
 264  * no requests have been sent, just return an error.
 265  */
 266 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 267 {
 268         struct nfs_open_context *ctx = dreq->ctx;
 269         struct inode *inode = ctx->path.dentry->d_inode;
 270         size_t rsize = NFS_SERVER(inode)->rsize;
 271         unsigned int pgbase;
 272         int result;
 273         ssize_t started = 0;
 274
 275         get_dreq(dreq);
 276
 277         do {
 278                 struct nfs_read_data *data;
 279                 size_t bytes;
 280
 281                 pgbase = user_addr & ~PAGE_MASK;
 282                 bytes = min(rsize,count);
 283
 284                 result = -ENOMEM;
 285                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 286                 if (unlikely(!data))
 287                         break;
 288
 289                 down_read(&current->mm->mmap_sem);
 290                 result = get_user_pages(current, current->mm, user_addr,
 291                                         data->npages, 1, 0, data->pagevec, NULL);
 292                 up_read(&current->mm->mmap_sem);
 293                 if (result < 0) {
 294                         nfs_readdata_release(data);
 295                         break;
 296                 }
 297                 if ((unsigned)result < data->npages) {
 298                         bytes = result * PAGE_SIZE;
 299                         if (bytes <= pgbase) {
 300                                 nfs_direct_release_pages(data->pagevec, result);
 301                                 nfs_readdata_release(data);
 302                                 break;
 303                         }
 304                         bytes -= pgbase;
 305                         data->npages = result;
 306                 }
 307
 308                 get_dreq(dreq);
 309
 310                 data->req = (struct nfs_page *) dreq;
 311                 data->inode = inode;
 312                 data->cred = ctx->cred;
 313                 data->args.fh = NFS_FH(inode);
 314                 data->args.context = ctx;
 315                 data->args.offset = pos;
 316                 data->args.pgbase = pgbase;
 317                 data->args.pages = data->pagevec;
 318                 data->args.count = bytes;
 319                 data->res.fattr = &data->fattr;
 320                 data->res.eof = 0;
 321                 data->res.count = bytes;
 322
 323                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 324                                 &nfs_read_direct_ops, data);
 325                 NFS_PROTO(inode)->read_setup(data);
 326
 327                 data->task.tk_cookie = (unsigned long) inode;
 328
 329                 rpc_execute(&data->task);
 330
 331                 dprintk("NFS: %5u initiated direct read call "
 332                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 333                                 data->task.tk_pid,
 334                                 inode->i_sb->s_id,
 335                                 (long long)NFS_FILEID(inode),
 336                                 bytes,
 337                                 (unsigned long long)data->args.offset);
 338
 339                 started += bytes;
 340                 user_addr += bytes;
 341                 pos += bytes;
 342                 /* FIXME: Remove this unnecessary math from final patch */
 343                 pgbase += bytes;
 344                 pgbase &= ~PAGE_MASK;
 345                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 346
 347                 count -= bytes;
 348         } while (count != 0);
 349
 350         if (put_dreq(dreq))
 351                 nfs_direct_complete(dreq);
 352
 353         if (started)
 354                 return 0;
 355         return result < 0 ? (ssize_t) result : -EFAULT;
 356 }
 357
 358 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 359                                               const struct iovec *iov,
 360                                               unsigned long nr_segs,
 361                                               loff_t pos)
 362 {
 363         ssize_t result = -EINVAL;
 364         size_t requested_bytes = 0;
 365         unsigned long seg;
 366
 367         get_dreq(dreq);
 368
 369         for (seg = 0; seg < nr_segs; seg++) {
 370                 const struct iovec *vec = &iov[seg];
 371                 result = nfs_direct_read_schedule(dreq,
 372                                         (unsigned long)vec->iov_base,
 373                                                   vec->iov_len, pos);
 374                 if (result < 0)
 375                         break;
 376                 requested_bytes += result;
 377                 if ((size_t)result < vec->iov_len)
 378                         break;
 379                 pos += vec->iov_len;
 380         }
 381
 382         if (put_dreq(dreq))
 383                 nfs_direct_complete(dreq);
 384
 385         if (requested_bytes != 0)
 386                 return 0;
 387
 388         if (result < 0)
 389                 return result;
 390         return -EIO;
 391 }
 392
 393 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 394 {
 395         ssize_t result = 0;
 396         sigset_t oldset;
 397         struct inode *inode = iocb->ki_filp->f_mapping->host;
 398         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 399         struct nfs_direct_req *dreq;
 400
 401         dreq = nfs_direct_req_alloc();
 402         if (!dreq)
 403                 return -ENOMEM;
 404
 405         dreq->inode = inode;
 406         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 407         if (!is_sync_kiocb(iocb))
 408                 dreq->iocb = iocb;
 409
 410         nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
 411         rpc_clnt_sigmask(clnt, &oldset);
 412         result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
 413         if (!result)
 414                 result = nfs_direct_wait(dreq);
 415         rpc_clnt_sigunmask(clnt, &oldset);
 416         nfs_direct_req_release(dreq);
 417
 418         return result;
 419 }
 420
 421 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 422 {
 423         while (!list_empty(&dreq->rewrite_list)) {
 424                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 425                 list_del(&data->pages);
 426                 nfs_direct_release_pages(data->pagevec, data->npages);
 427                 nfs_writedata_release(data);
 428         }
 429 }
 430
 431 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 432 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 433 {
 434         struct inode *inode = dreq->inode;
 435         struct list_head *p;
 436         struct nfs_write_data *data;
 437
 438         dreq->count = 0;
 439         get_dreq(dreq);
 440
 441         list_for_each(p, &dreq->rewrite_list) {
 442                 data = list_entry(p, struct nfs_write_data, pages);
 443
 444                 get_dreq(dreq);
 445
 446                 /*
 447                  * Reset data->res.
 448                  */
 449                 nfs_fattr_init(&data->fattr);
 450                 data->res.count = data->args.count;
 451                 memset(&data->verf, 0, sizeof(data->verf));
 452
 453                 /*
 454                  * Reuse data->task; data->args should not have changed
 455                  * since the original request was sent.
 456                  */
 457                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 458                                 &nfs_write_direct_ops, data);
 459                 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
 460
 461                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 462                 data->task.tk_cookie = (unsigned long) inode;
 463
 464                 /*
 465                  * We're called via an RPC callback, so BKL is already held.
 466                  */
 467                 rpc_execute(&data->task);
 468
 469                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 470                                 data->task.tk_pid,
 471                                 inode->i_sb->s_id,
 472                                 (long long)NFS_FILEID(inode),
 473                                 data->args.count,
 474                                 (unsigned long long)data->args.offset);
 475         }
 476
 477         if (put_dreq(dreq))
 478                 nfs_direct_write_complete(dreq, inode);
 479 }
 480
 481 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 482 {
 483         struct nfs_write_data *data = calldata;
 484         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 485
 486         /* Call the NFS version-specific code */
 487         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
 488                 return;
 489         if (unlikely(task->tk_status < 0)) {
 490                 dprintk("NFS: %5u commit failed with error %d.\n",
 491                                 task->tk_pid, task->tk_status);
 492                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 493         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 494                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
 495                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 496         }
 497
 498         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
 499         nfs_direct_write_complete(dreq, data->inode);
 500 }
 501
 502 static const struct rpc_call_ops nfs_commit_direct_ops = {
 503         .rpc_call_done = nfs_direct_commit_result,
 504         .rpc_release = nfs_commit_release,
 505 };
 506
 507 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 508 {
 509         struct nfs_write_data *data = dreq->commit_data;
 510
 511         data->inode = dreq->inode;
 512         data->cred = dreq->ctx->cred;
 513
 514         data->args.fh = NFS_FH(data->inode);
 515         data->args.offset = 0;
 516         data->args.count = 0;
 517         data->res.count = 0;
 518         data->res.fattr = &data->fattr;
 519         data->res.verf = &data->verf;
 520
 521         rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
 522                                 &nfs_commit_direct_ops, data);
 523         NFS_PROTO(data->inode)->commit_setup(data, 0);
 524
 525         data->task.tk_priority = RPC_PRIORITY_NORMAL;
 526         data->task.tk_cookie = (unsigned long)data->inode;
 527         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 528         dreq->commit_data = NULL;
 529
 530         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 531
 532         rpc_execute(&data->task);
 533 }
 534
 535 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 536 {
 537         int flags = dreq->flags;
 538
 539         dreq->flags = 0;
 540         switch (flags) {
 541                 case NFS_ODIRECT_DO_COMMIT:
 542                         nfs_direct_commit_schedule(dreq);
 543                         break;
 544                 case NFS_ODIRECT_RESCHED_WRITES:
 545                         nfs_direct_write_reschedule(dreq);
 546                         break;
 547                 default:
 548                         if (dreq->commit_data != NULL)
 549                                 nfs_commit_free(dreq->commit_data);
 550                         nfs_direct_free_writedata(dreq);
 551                         nfs_zap_mapping(inode, inode->i_mapping);
 552                         nfs_direct_complete(dreq);
 553         }
 554 }
 555
 556 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 557 {
 558         dreq->commit_data = nfs_commit_alloc();
 559         if (dreq->commit_data != NULL)
 560                 dreq->commit_data->req = (struct nfs_page *) dreq;
 561 }
 562 #else
 563 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 564 {
 565         dreq->commit_data = NULL;
 566 }
 567
 568 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 569 {
 570         nfs_direct_free_writedata(dreq);
 571         nfs_zap_mapping(inode, inode->i_mapping);
 572         nfs_direct_complete(dreq);
 573 }
 574 #endif
 575
 576 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 577 {
 578         struct nfs_write_data *data = calldata;
 579         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 580         int status = task->tk_status;
 581
 582         if (nfs_writeback_done(task, data) != 0)
 583                 return;
 584
 585         spin_lock(&dreq->lock);
 586
 587         if (unlikely(status < 0)) {
 588                 /* An error has occurred, so we should not commit */
 589                 dreq->flags = 0;
 590                 dreq->error = status;
 591         }
 592         if (unlikely(dreq->error != 0))
 593                 goto out_unlock;
 594
 595         dreq->count += data->res.count;
 596
 597         if (data->res.verf->committed != NFS_FILE_SYNC) {
 598                 switch (dreq->flags) {
 599                         case 0:
 600                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 601                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 602                                 break;
 603                         case NFS_ODIRECT_DO_COMMIT:
 604                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 605                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
 606                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 607                                 }
 608                 }
 609         }
 610 out_unlock:
 611         spin_unlock(&dreq->lock);
 612 }
 613
 614 /*
 615  * NB: Return the value of the first error return code.  Subsequent
 616  *     errors after the first one are ignored.
 617  */
 618 static void nfs_direct_write_release(void *calldata)
 619 {
 620         struct nfs_write_data *data = calldata;
 621         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 622
 623         if (put_dreq(dreq))
 624                 nfs_direct_write_complete(dreq, data->inode);
 625 }
 626
 627 static const struct rpc_call_ops nfs_write_direct_ops = {
 628         .rpc_call_done = nfs_direct_write_result,
 629         .rpc_release = nfs_direct_write_release,
 630 };
 631
 632 /*
 633  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 634  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 635  * bail and stop sending more writes.  Write length accounting is
 636  * handled automatically by nfs_direct_write_result().  Otherwise, if
 637  * no requests have been sent, just return an error.
 638  */
 639 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 640 {
 641         struct nfs_open_context *ctx = dreq->ctx;
 642         struct inode *inode = ctx->path.dentry->d_inode;
 643         size_t wsize = NFS_SERVER(inode)->wsize;
 644         unsigned int pgbase;
 645         int result;
 646         ssize_t started = 0;
 647
 648         get_dreq(dreq);
 649
 650         do {
 651                 struct nfs_write_data *data;
 652                 size_t bytes;
 653
 654                 pgbase = user_addr & ~PAGE_MASK;
 655                 bytes = min(wsize,count);
 656
 657                 result = -ENOMEM;
 658                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 659                 if (unlikely(!data))
 660                         break;
 661
 662                 down_read(&current->mm->mmap_sem);
 663                 result = get_user_pages(current, current->mm, user_addr,
 664                                         data->npages, 0, 0, data->pagevec, NULL);
 665                 up_read(&current->mm->mmap_sem);
 666                 if (result < 0) {
 667                         nfs_writedata_release(data);
 668                         break;
 669                 }
 670                 if ((unsigned)result < data->npages) {
 671                         bytes = result * PAGE_SIZE;
 672                         if (bytes <= pgbase) {
 673                                 nfs_direct_release_pages(data->pagevec, result);
 674                                 nfs_writedata_release(data);
 675                                 break;
 676                         }
 677                         bytes -= pgbase;
 678                         data->npages = result;
 679                 }
 680
 681                 get_dreq(dreq);
 682
 683                 list_move_tail(&data->pages, &dreq->rewrite_list);
 684
 685                 data->req = (struct nfs_page *) dreq;
 686                 data->inode = inode;
 687                 data->cred = ctx->cred;
 688                 data->args.fh = NFS_FH(inode);
 689                 data->args.context = ctx;
 690                 data->args.offset = pos;
 691                 data->args.pgbase = pgbase;
 692                 data->args.pages = data->pagevec;
 693                 data->args.count = bytes;
 694                 data->res.fattr = &data->fattr;
 695                 data->res.count = bytes;
 696                 data->res.verf = &data->verf;
 697
 698                 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
 699                                 &nfs_write_direct_ops, data);
 700                 NFS_PROTO(inode)->write_setup(data, sync);
 701
 702                 data->task.tk_priority = RPC_PRIORITY_NORMAL;
 703                 data->task.tk_cookie = (unsigned long) inode;
 704
 705                 rpc_execute(&data->task);
 706
 707                 dprintk("NFS: %5u initiated direct write call "
 708                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 709                                 data->task.tk_pid,
 710                                 inode->i_sb->s_id,
 711                                 (long long)NFS_FILEID(inode),
 712                                 bytes,
 713                                 (unsigned long long)data->args.offset);
 714
 715                 started += bytes;
 716                 user_addr += bytes;
 717                 pos += bytes;
 718
 719                 /* FIXME: Remove this useless math from the final patch */
 720                 pgbase += bytes;
 721                 pgbase &= ~PAGE_MASK;
 722                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 723
 724                 count -= bytes;
 725         } while (count != 0);
 726
 727         if (put_dreq(dreq))
 728                 nfs_direct_write_complete(dreq, inode);
 729
 730         if (started)
 731                 return 0;
 732         return result < 0 ? (ssize_t) result : -EFAULT;
 733 }
 734
 735 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 736                                                const struct iovec *iov,
 737                                                unsigned long nr_segs,
 738                                                loff_t pos, int sync)
 739 {
 740         ssize_t result = 0;
 741         size_t requested_bytes = 0;
 742         unsigned long seg;
 743
 744         get_dreq(dreq);
 745
 746         for (seg = 0; seg < nr_segs; seg++) {
 747                 const struct iovec *vec = &iov[seg];
 748                 result = nfs_direct_write_schedule(dreq,
 749                                         (unsigned long)vec->iov_base,
 750                                                    vec->iov_len,
 751                                                    pos, sync);
 752                 if (result < 0)
 753                         break;
 754                 requested_bytes += result;
 755                 if ((size_t)result < vec->iov_len)
 756                         break;
 757                 pos += vec->iov_len;
 758         }
 759
 760         if (put_dreq(dreq))
 761                 nfs_direct_write_complete(dreq, dreq->inode);
 762
 763         if (requested_bytes != 0)
 764                 return 0;
 765
 766         if (result < 0)
 767                 return result;
 768         return -EIO;
 769 }
 770
 771 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 772 {
 773         ssize_t result = 0;
 774         sigset_t oldset;
 775         struct inode *inode = iocb->ki_filp->f_mapping->host;
 776         struct rpc_clnt *clnt = NFS_CLIENT(inode);
 777         struct nfs_direct_req *dreq;
 778         size_t wsize = NFS_SERVER(inode)->wsize;
 779         int sync = 0;
 780
 781         dreq = nfs_direct_req_alloc();
 782         if (!dreq)
 783                 return -ENOMEM;
 784         nfs_alloc_commit_data(dreq);
 785
 786         if (dreq->commit_data == NULL || count < wsize)
 787                 sync = FLUSH_STABLE;
 788
 789         dreq->inode = inode;
 790         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 791         if (!is_sync_kiocb(iocb))
 792                 dreq->iocb = iocb;
 793
 794         nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
 795
 796         rpc_clnt_sigmask(clnt, &oldset);
 797         result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
 798         if (!result)
 799                 result = nfs_direct_wait(dreq);
 800         rpc_clnt_sigunmask(clnt, &oldset);
 801         nfs_direct_req_release(dreq);
 802
 803         return result;
 804 }
 805
 806 /**
 807  * nfs_file_direct_read - file direct read operation for NFS files
 808  * @iocb: target I/O control block
 809  * @iov: vector of user buffers into which to read data
 810  * @nr_segs: size of iov vector
 811  * @pos: byte offset in file where reading starts
 812  *
 813  * We use this function for direct reads instead of calling
 814  * generic_file_aio_read() in order to avoid gfar's check to see if
 815  * the request starts before the end of the file.  For that check
 816  * to work, we must generate a GETATTR before each direct read, and
 817  * even then there is a window between the GETATTR and the subsequent
 818  * READ where the file size could change.  Our preference is simply
 819  * to do all reads the application wants, and the server will take
 820  * care of managing the end of file boundary.
 821  *
 822  * This function also eliminates unnecessarily updating the file's
 823  * atime locally, as the NFS server sets the file's atime, and this
 824  * client must read the updated atime from the server back into its
 825  * cache.
 826  */
 827 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 828                                 unsigned long nr_segs, loff_t pos)
 829 {
 830         ssize_t retval = -EINVAL;
 831         struct file *file = iocb->ki_filp;
 832         struct address_space *mapping = file->f_mapping;
 833         /* XXX: temporary */
 834         const char __user *buf = iov[0].iov_base;
 835         size_t count = iov[0].iov_len;
 836
 837         dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
 838                 file->f_path.dentry->d_parent->d_name.name,
 839                 file->f_path.dentry->d_name.name,
 840                 (unsigned long) count, (long long) pos);
 841
 842         if (nr_segs != 1)
 843                 goto out;
 844
 845         retval = -EFAULT;
 846         if (!access_ok(VERIFY_WRITE, buf, count))
 847                 goto out;
 848         retval = 0;
 849         if (!count)
 850                 goto out;
 851
 852         retval = nfs_sync_mapping(mapping);
 853         if (retval)
 854                 goto out;
 855
 856         retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
 857         if (retval > 0)
 858                 iocb->ki_pos = pos + retval;
 859
 860 out:
 861         return retval;
 862 }
 863
 864 /**
 865  * nfs_file_direct_write - file direct write operation for NFS files
 866  * @iocb: target I/O control block
 867  * @iov: vector of user buffers from which to write data
 868  * @nr_segs: size of iov vector
 869  * @pos: byte offset in file where writing starts
 870  *
 871  * We use this function for direct writes instead of calling
 872  * generic_file_aio_write() in order to avoid taking the inode
 873  * semaphore and updating the i_size.  The NFS server will set
 874  * the new i_size and this client must read the updated size
 875  * back into its cache.  We let the server do generic write
 876  * parameter checking and report problems.
 877  *
 878  * We also avoid an unnecessary invocation of generic_osync_inode(),
 879  * as it is fairly meaningless to sync the metadata of an NFS file.
 880  *
 881  * We eliminate local atime updates, see direct read above.
 882  *
 883  * We avoid unnecessary page cache invalidations for normal cached
 884  * readers of this file.
 885  *
 886  * Note that O_APPEND is not supported for NFS direct writes, as there
 887  * is no atomic O_APPEND write facility in the NFS protocol.
 888  */
 889 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 890                                 unsigned long nr_segs, loff_t pos)
 891 {
 892         ssize_t retval = -EINVAL;
 893         struct file *file = iocb->ki_filp;
 894         struct address_space *mapping = file->f_mapping;
 895         /* XXX: temporary */
 896         const char __user *buf = iov[0].iov_base;
 897         size_t count = iov[0].iov_len;
 898
 899         dprintk("nfs: direct write(%s/%s, %lu@%Ld)\n",
 900                 file->f_path.dentry->d_parent->d_name.name,
 901                 file->f_path.dentry->d_name.name,
 902                 (unsigned long) count, (long long) pos);
 903
 904         if (nr_segs != 1)
 905                 goto out;
 906
 907         retval = generic_write_checks(file, &pos, &count, 0);
 908         if (retval)
 909                 goto out;
 910
 911         retval = -EINVAL;
 912         if ((ssize_t) count < 0)
 913                 goto out;
 914         retval = 0;
 915         if (!count)
 916                 goto out;
 917
 918         retval = -EFAULT;
 919         if (!access_ok(VERIFY_READ, buf, count))
 920                 goto out;
 921
 922         retval = nfs_sync_mapping(mapping);
 923         if (retval)
 924                 goto out;
 925
 926         retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
 927
 928         if (retval > 0)
 929                 iocb->ki_pos = pos + retval;
 930
 931 out:
 932         return retval;
 933 }
 934
 935 /**
 936  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 937  *
 938  */
 939 int __init nfs_init_directcache(void)
 940 {
 941         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 942                                                 sizeof(struct nfs_direct_req),
 943                                                 0, (SLAB_RECLAIM_ACCOUNT|
 944                                                         SLAB_MEM_SPREAD),
 945                                                 NULL);
 946         if (nfs_direct_cachep == NULL)
 947                 return -ENOMEM;
 948
 949         return 0;
 950 }
 951
 952 /**
 953  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 954  *
 955  */
 956 void nfs_destroy_directcache(void)
 957 {
 958         kmem_cache_destroy(nfs_direct_cachep);
 959 }