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