fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47 #include <linux/slab.h>
  48
  49 #include <linux/nfs_fs.h>
  50 #include <linux/nfs_page.h>
  51 #include <linux/sunrpc/clnt.h>
  52
  53 #include <asm/system.h>
  54 #include <asm/uaccess.h>
  55 #include <asm/atomic.h>
  56
  57 #include "internal.h"
  58 #include "iostat.h"
  59
  60 #define NFSDBG_FACILITY         NFSDBG_VFS
  61
  62 static struct kmem_cache *nfs_direct_cachep;
  63
  64 /*
  65  * This represents a set of asynchronous requests that we're waiting on
  66  */
  67 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct nfs_open_context *ctx;           /* file open context info */
  72         struct nfs_lock_context *l_ctx;         /* Lock context info */
  73         struct kiocb *          iocb;           /* controlling i/o request */
  74         struct inode *          inode;          /* target file of i/o */
  75
  76         /* completion state */
  77         atomic_t                io_count;       /* i/os we're waiting for */
  78         spinlock_t              lock;           /* protect completion state */
  79         ssize_t                 count,          /* bytes actually processed */
  80                                 error;          /* any reported error */
  81         struct completion       completion;     /* wait for i/o completion */
  82
  83         /* commit state */
  84         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
  85         struct nfs_write_data * commit_data;    /* special write_data for commits */
  86         int                     flags;
  87 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  88 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  89         struct nfs_writeverf    verf;           /* unstable write verifier */
  90 };
  91
  92 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  93 static const struct rpc_call_ops nfs_write_direct_ops;
  94
  95 static inline void get_dreq(struct nfs_direct_req *dreq)
  96 {
  97         atomic_inc(&dreq->io_count);
  98 }
  99
 100 static inline int put_dreq(struct nfs_direct_req *dreq)
 101 {
 102         return atomic_dec_and_test(&dreq->io_count);
 103 }
 104
 105 /**
 106  * nfs_direct_IO - NFS address space operation for direct I/O
 107  * @rw: direction (read or write)
 108  * @iocb: target I/O control block
 109  * @iov: array of vectors that define I/O buffer
 110  * @pos: offset in file to begin the operation
 111  * @nr_segs: size of iovec array
 112  *
 113  * The presence of this routine in the address space ops vector means
 114  * the NFS client supports direct I/O.  However, we shunt off direct
 115  * read and write requests before the VFS gets them, so this method
 116  * should never be called.
 117  */
 118 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 119 {
 120         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 121                         iocb->ki_filp->f_path.dentry->d_name.name,
 122                         (long long) pos, nr_segs);
 123
 124         return -EINVAL;
 125 }
 126
 127 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
 128 {
 129         unsigned int npages;
 130         unsigned int i;
 131
 132         if (count == 0)
 133                 return;
 134         pages += (pgbase >> PAGE_SHIFT);
 135         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 136         for (i = 0; i < npages; i++) {
 137                 struct page *page = pages[i];
 138                 if (!PageCompound(page))
 139                         set_page_dirty(page);
 140         }
 141 }
 142
 143 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 144 {
 145         unsigned int i;
 146         for (i = 0; i < npages; i++)
 147                 page_cache_release(pages[i]);
 148 }
 149
 150 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 151 {
 152         struct nfs_direct_req *dreq;
 153
 154         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
 155         if (!dreq)
 156                 return NULL;
 157
 158         kref_init(&dreq->kref);
 159         kref_get(&dreq->kref);
 160         init_completion(&dreq->completion);
 161         INIT_LIST_HEAD(&dreq->rewrite_list);
 162         dreq->iocb = NULL;
 163         dreq->ctx = NULL;
 164         dreq->l_ctx = NULL;
 165         spin_lock_init(&dreq->lock);
 166         atomic_set(&dreq->io_count, 0);
 167         dreq->count = 0;
 168         dreq->error = 0;
 169         dreq->flags = 0;
 170
 171         return dreq;
 172 }
 173
 174 static void nfs_direct_req_free(struct kref *kref)
 175 {
 176         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 177
 178         if (dreq->l_ctx != NULL)
 179                 nfs_put_lock_context(dreq->l_ctx);
 180         if (dreq->ctx != NULL)
 181                 put_nfs_open_context(dreq->ctx);
 182         kmem_cache_free(nfs_direct_cachep, dreq);
 183 }
 184
 185 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 186 {
 187         kref_put(&dreq->kref, nfs_direct_req_free);
 188 }
 189
 190 /*
 191  * Collects and returns the final error value/byte-count.
 192  */
 193 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 194 {
 195         ssize_t result = -EIOCBQUEUED;
 196
 197         /* Async requests don't wait here */
 198         if (dreq->iocb)
 199                 goto out;
 200
 201         result = wait_for_completion_killable(&dreq->completion);
 202
 203         if (!result)
 204                 result = dreq->error;
 205         if (!result)
 206                 result = dreq->count;
 207
 208 out:
 209         return (ssize_t) result;
 210 }
 211
 212 /*
 213  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 214  * the iocb is still valid here if this is a synchronous request.
 215  */
 216 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 217 {
 218         if (dreq->iocb) {
 219                 long res = (long) dreq->error;
 220                 if (!res)
 221                         res = (long) dreq->count;
 222                 aio_complete(dreq->iocb, res, 0);
 223         }
 224         complete_all(&dreq->completion);
 225
 226         nfs_direct_req_release(dreq);
 227 }
 228
 229 /*
 230  * We must hold a reference to all the pages in this direct read request
 231  * until the RPCs complete.  This could be long *after* we are woken up in
 232  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
 233  */
 234 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
 235 {
 236         struct nfs_read_data *data = calldata;
 237
 238         nfs_readpage_result(task, data);
 239 }
 240
 241 static void nfs_direct_read_release(void *calldata)
 242 {
 243
 244         struct nfs_read_data *data = calldata;
 245         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 246         int status = data->task.tk_status;
 247
 248         spin_lock(&dreq->lock);
 249         if (unlikely(status < 0)) {
 250                 dreq->error = status;
 251                 spin_unlock(&dreq->lock);
 252         } else {
 253                 dreq->count += data->res.count;
 254                 spin_unlock(&dreq->lock);
 255                 nfs_direct_dirty_pages(data->pagevec,
 256                                 data->args.pgbase,
 257                                 data->res.count);
 258         }
 259         nfs_direct_release_pages(data->pagevec, data->npages);
 260
 261         if (put_dreq(dreq))
 262                 nfs_direct_complete(dreq);
 263         nfs_readdata_free(data);
 264 }
 265
 266 static const struct rpc_call_ops nfs_read_direct_ops = {
 267 #if defined(CONFIG_NFS_V4_1)
 268         .rpc_call_prepare = nfs_read_prepare,
 269 #endif /* CONFIG_NFS_V4_1 */
 270         .rpc_call_done = nfs_direct_read_result,
 271         .rpc_release = nfs_direct_read_release,
 272 };
 273
 274 /*
 275  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 276  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 277  * bail and stop sending more reads.  Read length accounting is
 278  * handled automatically by nfs_direct_read_result().  Otherwise, if
 279  * no requests have been sent, just return an error.
 280  */
 281 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
 282                                                 const struct iovec *iov,
 283                                                 loff_t pos)
 284 {
 285         struct nfs_open_context *ctx = dreq->ctx;
 286         struct inode *inode = ctx->path.dentry->d_inode;
 287         unsigned long user_addr = (unsigned long)iov->iov_base;
 288         size_t count = iov->iov_len;
 289         size_t rsize = NFS_SERVER(inode)->rsize;
 290         struct rpc_task *task;
 291         struct rpc_message msg = {
 292                 .rpc_cred = ctx->cred,
 293         };
 294         struct rpc_task_setup task_setup_data = {
 295                 .rpc_client = NFS_CLIENT(inode),
 296                 .rpc_message = &msg,
 297                 .callback_ops = &nfs_read_direct_ops,
 298                 .workqueue = nfsiod_workqueue,
 299                 .flags = RPC_TASK_ASYNC,
 300         };
 301         unsigned int pgbase;
 302         int result;
 303         ssize_t started = 0;
 304
 305         do {
 306                 struct nfs_read_data *data;
 307                 size_t bytes;
 308
 309                 pgbase = user_addr & ~PAGE_MASK;
 310                 bytes = min(rsize,count);
 311
 312                 result = -ENOMEM;
 313                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
 314                 if (unlikely(!data))
 315                         break;
 316
 317                 down_read(&current->mm->mmap_sem);
 318                 result = get_user_pages(current, current->mm, user_addr,
 319                                         data->npages, 1, 0, data->pagevec, NULL);
 320                 up_read(&current->mm->mmap_sem);
 321                 if (result < 0) {
 322                         nfs_readdata_free(data);
 323                         break;
 324                 }
 325                 if ((unsigned)result < data->npages) {
 326                         bytes = result * PAGE_SIZE;
 327                         if (bytes <= pgbase) {
 328                                 nfs_direct_release_pages(data->pagevec, result);
 329                                 nfs_readdata_free(data);
 330                                 break;
 331                         }
 332                         bytes -= pgbase;
 333                         data->npages = result;
 334                 }
 335
 336                 get_dreq(dreq);
 337
 338                 data->req = (struct nfs_page *) dreq;
 339                 data->inode = inode;
 340                 data->cred = msg.rpc_cred;
 341                 data->args.fh = NFS_FH(inode);
 342                 data->args.context = ctx;
 343                 data->args.lock_context = dreq->l_ctx;
 344                 data->args.offset = pos;
 345                 data->args.pgbase = pgbase;
 346                 data->args.pages = data->pagevec;
 347                 data->args.count = bytes;
 348                 data->res.fattr = &data->fattr;
 349                 data->res.eof = 0;
 350                 data->res.count = bytes;
 351                 nfs_fattr_init(&data->fattr);
 352                 msg.rpc_argp = &data->args;
 353                 msg.rpc_resp = &data->res;
 354
 355                 task_setup_data.task = &data->task;
 356                 task_setup_data.callback_data = data;
 357                 NFS_PROTO(inode)->read_setup(data, &msg);
 358
 359                 task = rpc_run_task(&task_setup_data);
 360                 if (IS_ERR(task))
 361                         break;
 362                 rpc_put_task(task);
 363
 364                 dprintk("NFS: %5u initiated direct read call "
 365                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 366                                 data->task.tk_pid,
 367                                 inode->i_sb->s_id,
 368                                 (long long)NFS_FILEID(inode),
 369                                 bytes,
 370                                 (unsigned long long)data->args.offset);
 371
 372                 started += bytes;
 373                 user_addr += bytes;
 374                 pos += bytes;
 375                 /* FIXME: Remove this unnecessary math from final patch */
 376                 pgbase += bytes;
 377                 pgbase &= ~PAGE_MASK;
 378                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 379
 380                 count -= bytes;
 381         } while (count != 0);
 382
 383         if (started)
 384                 return started;
 385         return result < 0 ? (ssize_t) result : -EFAULT;
 386 }
 387
 388 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 389                                               const struct iovec *iov,
 390                                               unsigned long nr_segs,
 391                                               loff_t pos)
 392 {
 393         ssize_t result = -EINVAL;
 394         size_t requested_bytes = 0;
 395         unsigned long seg;
 396
 397         get_dreq(dreq);
 398
 399         for (seg = 0; seg < nr_segs; seg++) {
 400                 const struct iovec *vec = &iov[seg];
 401                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
 402                 if (result < 0)
 403                         break;
 404                 requested_bytes += result;
 405                 if ((size_t)result < vec->iov_len)
 406                         break;
 407                 pos += vec->iov_len;
 408         }
 409
 410         /*
 411          * If no bytes were started, return the error, and let the
 412          * generic layer handle the completion.
 413          */
 414         if (requested_bytes == 0) {
 415                 nfs_direct_req_release(dreq);
 416                 return result < 0 ? result : -EIO;
 417         }
 418
 419         if (put_dreq(dreq))
 420                 nfs_direct_complete(dreq);
 421         return 0;
 422 }
 423
 424 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 425                                unsigned long nr_segs, loff_t pos)
 426 {
 427         ssize_t result = -ENOMEM;
 428         struct inode *inode = iocb->ki_filp->f_mapping->host;
 429         struct nfs_direct_req *dreq;
 430
 431         dreq = nfs_direct_req_alloc();
 432         if (dreq == NULL)
 433                 goto out;
 434
 435         dreq->inode = inode;
 436         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 437         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 438         if (dreq->l_ctx == NULL)
 439                 goto out_release;
 440         if (!is_sync_kiocb(iocb))
 441                 dreq->iocb = iocb;
 442
 443         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 444         if (!result)
 445                 result = nfs_direct_wait(dreq);
 446 out_release:
 447         nfs_direct_req_release(dreq);
 448 out:
 449         return result;
 450 }
 451
 452 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 453 {
 454         while (!list_empty(&dreq->rewrite_list)) {
 455                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 456                 list_del(&data->pages);
 457                 nfs_direct_release_pages(data->pagevec, data->npages);
 458                 nfs_writedata_free(data);
 459         }
 460 }
 461
 462 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 463 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 464 {
 465         struct inode *inode = dreq->inode;
 466         struct list_head *p;
 467         struct nfs_write_data *data;
 468         struct rpc_task *task;
 469         struct rpc_message msg = {
 470                 .rpc_cred = dreq->ctx->cred,
 471         };
 472         struct rpc_task_setup task_setup_data = {
 473                 .rpc_client = NFS_CLIENT(inode),
 474                 .rpc_message = &msg,
 475                 .callback_ops = &nfs_write_direct_ops,
 476                 .workqueue = nfsiod_workqueue,
 477                 .flags = RPC_TASK_ASYNC,
 478         };
 479
 480         dreq->count = 0;
 481         get_dreq(dreq);
 482
 483         list_for_each(p, &dreq->rewrite_list) {
 484                 data = list_entry(p, struct nfs_write_data, pages);
 485
 486                 get_dreq(dreq);
 487
 488                 /* Use stable writes */
 489                 data->args.stable = NFS_FILE_SYNC;
 490
 491                 /*
 492                  * Reset data->res.
 493                  */
 494                 nfs_fattr_init(&data->fattr);
 495                 data->res.count = data->args.count;
 496                 memset(&data->verf, 0, sizeof(data->verf));
 497
 498                 /*
 499                  * Reuse data->task; data->args should not have changed
 500                  * since the original request was sent.
 501                  */
 502                 task_setup_data.task = &data->task;
 503                 task_setup_data.callback_data = data;
 504                 msg.rpc_argp = &data->args;
 505                 msg.rpc_resp = &data->res;
 506                 NFS_PROTO(inode)->write_setup(data, &msg);
 507
 508                 /*
 509                  * We're called via an RPC callback, so BKL is already held.
 510                  */
 511                 task = rpc_run_task(&task_setup_data);
 512                 if (!IS_ERR(task))
 513                         rpc_put_task(task);
 514
 515                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 516                                 data->task.tk_pid,
 517                                 inode->i_sb->s_id,
 518                                 (long long)NFS_FILEID(inode),
 519                                 data->args.count,
 520                                 (unsigned long long)data->args.offset);
 521         }
 522
 523         if (put_dreq(dreq))
 524                 nfs_direct_write_complete(dreq, inode);
 525 }
 526
 527 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 528 {
 529         struct nfs_write_data *data = calldata;
 530
 531         /* Call the NFS version-specific code */
 532         NFS_PROTO(data->inode)->commit_done(task, data);
 533 }
 534
 535 static void nfs_direct_commit_release(void *calldata)
 536 {
 537         struct nfs_write_data *data = calldata;
 538         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 539         int status = data->task.tk_status;
 540
 541         if (status < 0) {
 542                 dprintk("NFS: %5u commit failed with error %d.\n",
 543                                 data->task.tk_pid, status);
 544                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 545         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 546                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 547                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 548         }
 549
 550         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 551         nfs_direct_write_complete(dreq, data->inode);
 552         nfs_commit_free(data);
 553 }
 554
 555 static const struct rpc_call_ops nfs_commit_direct_ops = {
 556 #if defined(CONFIG_NFS_V4_1)
 557         .rpc_call_prepare = nfs_write_prepare,
 558 #endif /* CONFIG_NFS_V4_1 */
 559         .rpc_call_done = nfs_direct_commit_result,
 560         .rpc_release = nfs_direct_commit_release,
 561 };
 562
 563 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 564 {
 565         struct nfs_write_data *data = dreq->commit_data;
 566         struct rpc_task *task;
 567         struct rpc_message msg = {
 568                 .rpc_argp = &data->args,
 569                 .rpc_resp = &data->res,
 570                 .rpc_cred = dreq->ctx->cred,
 571         };
 572         struct rpc_task_setup task_setup_data = {
 573                 .task = &data->task,
 574                 .rpc_client = NFS_CLIENT(dreq->inode),
 575                 .rpc_message = &msg,
 576                 .callback_ops = &nfs_commit_direct_ops,
 577                 .callback_data = data,
 578                 .workqueue = nfsiod_workqueue,
 579                 .flags = RPC_TASK_ASYNC,
 580         };
 581
 582         data->inode = dreq->inode;
 583         data->cred = msg.rpc_cred;
 584
 585         data->args.fh = NFS_FH(data->inode);
 586         data->args.offset = 0;
 587         data->args.count = 0;
 588         data->args.context = dreq->ctx;
 589         data->args.lock_context = dreq->l_ctx;
 590         data->res.count = 0;
 591         data->res.fattr = &data->fattr;
 592         data->res.verf = &data->verf;
 593         nfs_fattr_init(&data->fattr);
 594
 595         NFS_PROTO(data->inode)->commit_setup(data, &msg);
 596
 597         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 598         dreq->commit_data = NULL;
 599
 600         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 601
 602         task = rpc_run_task(&task_setup_data);
 603         if (!IS_ERR(task))
 604                 rpc_put_task(task);
 605 }
 606
 607 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 608 {
 609         int flags = dreq->flags;
 610
 611         dreq->flags = 0;
 612         switch (flags) {
 613                 case NFS_ODIRECT_DO_COMMIT:
 614                         nfs_direct_commit_schedule(dreq);
 615                         break;
 616                 case NFS_ODIRECT_RESCHED_WRITES:
 617                         nfs_direct_write_reschedule(dreq);
 618                         break;
 619                 default:
 620                         if (dreq->commit_data != NULL)
 621                                 nfs_commit_free(dreq->commit_data);
 622                         nfs_direct_free_writedata(dreq);
 623                         nfs_zap_mapping(inode, inode->i_mapping);
 624                         nfs_direct_complete(dreq);
 625         }
 626 }
 627
 628 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 629 {
 630         dreq->commit_data = nfs_commitdata_alloc();
 631         if (dreq->commit_data != NULL)
 632                 dreq->commit_data->req = (struct nfs_page *) dreq;
 633 }
 634 #else
 635 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 636 {
 637         dreq->commit_data = NULL;
 638 }
 639
 640 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 641 {
 642         nfs_direct_free_writedata(dreq);
 643         nfs_zap_mapping(inode, inode->i_mapping);
 644         nfs_direct_complete(dreq);
 645 }
 646 #endif
 647
 648 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 649 {
 650         struct nfs_write_data *data = calldata;
 651
 652         if (nfs_writeback_done(task, data) != 0)
 653                 return;
 654 }
 655
 656 /*
 657  * NB: Return the value of the first error return code.  Subsequent
 658  *     errors after the first one are ignored.
 659  */
 660 static void nfs_direct_write_release(void *calldata)
 661 {
 662         struct nfs_write_data *data = calldata;
 663         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 664         int status = data->task.tk_status;
 665
 666         spin_lock(&dreq->lock);
 667
 668         if (unlikely(status < 0)) {
 669                 /* An error has occurred, so we should not commit */
 670                 dreq->flags = 0;
 671                 dreq->error = status;
 672         }
 673         if (unlikely(dreq->error != 0))
 674                 goto out_unlock;
 675
 676         dreq->count += data->res.count;
 677
 678         if (data->res.verf->committed != NFS_FILE_SYNC) {
 679                 switch (dreq->flags) {
 680                         case 0:
 681                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 682                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 683                                 break;
 684                         case NFS_ODIRECT_DO_COMMIT:
 685                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 686                                         dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
 687                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 688                                 }
 689                 }
 690         }
 691 out_unlock:
 692         spin_unlock(&dreq->lock);
 693
 694         if (put_dreq(dreq))
 695                 nfs_direct_write_complete(dreq, data->inode);
 696 }
 697
 698 static const struct rpc_call_ops nfs_write_direct_ops = {
 699 #if defined(CONFIG_NFS_V4_1)
 700         .rpc_call_prepare = nfs_write_prepare,
 701 #endif /* CONFIG_NFS_V4_1 */
 702         .rpc_call_done = nfs_direct_write_result,
 703         .rpc_release = nfs_direct_write_release,
 704 };
 705
 706 /*
 707  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 708  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 709  * bail and stop sending more writes.  Write length accounting is
 710  * handled automatically by nfs_direct_write_result().  Otherwise, if
 711  * no requests have been sent, just return an error.
 712  */
 713 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
 714                                                  const struct iovec *iov,
 715                                                  loff_t pos, int sync)
 716 {
 717         struct nfs_open_context *ctx = dreq->ctx;
 718         struct inode *inode = ctx->path.dentry->d_inode;
 719         unsigned long user_addr = (unsigned long)iov->iov_base;
 720         size_t count = iov->iov_len;
 721         struct rpc_task *task;
 722         struct rpc_message msg = {
 723                 .rpc_cred = ctx->cred,
 724         };
 725         struct rpc_task_setup task_setup_data = {
 726                 .rpc_client = NFS_CLIENT(inode),
 727                 .rpc_message = &msg,
 728                 .callback_ops = &nfs_write_direct_ops,
 729                 .workqueue = nfsiod_workqueue,
 730                 .flags = RPC_TASK_ASYNC,
 731         };
 732         size_t wsize = NFS_SERVER(inode)->wsize;
 733         unsigned int pgbase;
 734         int result;
 735         ssize_t started = 0;
 736
 737         do {
 738                 struct nfs_write_data *data;
 739                 size_t bytes;
 740
 741                 pgbase = user_addr & ~PAGE_MASK;
 742                 bytes = min(wsize,count);
 743
 744                 result = -ENOMEM;
 745                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 746                 if (unlikely(!data))
 747                         break;
 748
 749                 down_read(&current->mm->mmap_sem);
 750                 result = get_user_pages(current, current->mm, user_addr,
 751                                         data->npages, 0, 0, data->pagevec, NULL);
 752                 up_read(&current->mm->mmap_sem);
 753                 if (result < 0) {
 754                         nfs_writedata_free(data);
 755                         break;
 756                 }
 757                 if ((unsigned)result < data->npages) {
 758                         bytes = result * PAGE_SIZE;
 759                         if (bytes <= pgbase) {
 760                                 nfs_direct_release_pages(data->pagevec, result);
 761                                 nfs_writedata_free(data);
 762                                 break;
 763                         }
 764                         bytes -= pgbase;
 765                         data->npages = result;
 766                 }
 767
 768                 get_dreq(dreq);
 769
 770                 list_move_tail(&data->pages, &dreq->rewrite_list);
 771
 772                 data->req = (struct nfs_page *) dreq;
 773                 data->inode = inode;
 774                 data->cred = msg.rpc_cred;
 775                 data->args.fh = NFS_FH(inode);
 776                 data->args.context = ctx;
 777                 data->args.lock_context = dreq->l_ctx;
 778                 data->args.offset = pos;
 779                 data->args.pgbase = pgbase;
 780                 data->args.pages = data->pagevec;
 781                 data->args.count = bytes;
 782                 data->args.stable = sync;
 783                 data->res.fattr = &data->fattr;
 784                 data->res.count = bytes;
 785                 data->res.verf = &data->verf;
 786                 nfs_fattr_init(&data->fattr);
 787
 788                 task_setup_data.task = &data->task;
 789                 task_setup_data.callback_data = data;
 790                 msg.rpc_argp = &data->args;
 791                 msg.rpc_resp = &data->res;
 792                 NFS_PROTO(inode)->write_setup(data, &msg);
 793
 794                 task = rpc_run_task(&task_setup_data);
 795                 if (IS_ERR(task))
 796                         break;
 797                 rpc_put_task(task);
 798
 799                 dprintk("NFS: %5u initiated direct write call "
 800                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 801                                 data->task.tk_pid,
 802                                 inode->i_sb->s_id,
 803                                 (long long)NFS_FILEID(inode),
 804                                 bytes,
 805                                 (unsigned long long)data->args.offset);
 806
 807                 started += bytes;
 808                 user_addr += bytes;
 809                 pos += bytes;
 810
 811                 /* FIXME: Remove this useless math from the final patch */
 812                 pgbase += bytes;
 813                 pgbase &= ~PAGE_MASK;
 814                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 815
 816                 count -= bytes;
 817         } while (count != 0);
 818
 819         if (started)
 820                 return started;
 821         return result < 0 ? (ssize_t) result : -EFAULT;
 822 }
 823
 824 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 825                                                const struct iovec *iov,
 826                                                unsigned long nr_segs,
 827                                                loff_t pos, int sync)
 828 {
 829         ssize_t result = 0;
 830         size_t requested_bytes = 0;
 831         unsigned long seg;
 832
 833         get_dreq(dreq);
 834
 835         for (seg = 0; seg < nr_segs; seg++) {
 836                 const struct iovec *vec = &iov[seg];
 837                 result = nfs_direct_write_schedule_segment(dreq, vec,
 838                                                            pos, sync);
 839                 if (result < 0)
 840                         break;
 841                 requested_bytes += result;
 842                 if ((size_t)result < vec->iov_len)
 843                         break;
 844                 pos += vec->iov_len;
 845         }
 846
 847         /*
 848          * If no bytes were started, return the error, and let the
 849          * generic layer handle the completion.
 850          */
 851         if (requested_bytes == 0) {
 852                 nfs_direct_req_release(dreq);
 853                 return result < 0 ? result : -EIO;
 854         }
 855
 856         if (put_dreq(dreq))
 857                 nfs_direct_write_complete(dreq, dreq->inode);
 858         return 0;
 859 }
 860
 861 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 862                                 unsigned long nr_segs, loff_t pos,
 863                                 size_t count)
 864 {
 865         ssize_t result = -ENOMEM;
 866         struct inode *inode = iocb->ki_filp->f_mapping->host;
 867         struct nfs_direct_req *dreq;
 868         size_t wsize = NFS_SERVER(inode)->wsize;
 869         int sync = NFS_UNSTABLE;
 870
 871         dreq = nfs_direct_req_alloc();
 872         if (!dreq)
 873                 goto out;
 874         nfs_alloc_commit_data(dreq);
 875
 876         if (dreq->commit_data == NULL || count <= wsize)
 877                 sync = NFS_FILE_SYNC;
 878
 879         dreq->inode = inode;
 880         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 881         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 882         if (dreq->l_ctx == NULL)
 883                 goto out_release;
 884         if (!is_sync_kiocb(iocb))
 885                 dreq->iocb = iocb;
 886
 887         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 888         if (!result)
 889                 result = nfs_direct_wait(dreq);
 890 out_release:
 891         nfs_direct_req_release(dreq);
 892 out:
 893         return result;
 894 }
 895
 896 /**
 897  * nfs_file_direct_read - file direct read operation for NFS files
 898  * @iocb: target I/O control block
 899  * @iov: vector of user buffers into which to read data
 900  * @nr_segs: size of iov vector
 901  * @pos: byte offset in file where reading starts
 902  *
 903  * We use this function for direct reads instead of calling
 904  * generic_file_aio_read() in order to avoid gfar's check to see if
 905  * the request starts before the end of the file.  For that check
 906  * to work, we must generate a GETATTR before each direct read, and
 907  * even then there is a window between the GETATTR and the subsequent
 908  * READ where the file size could change.  Our preference is simply
 909  * to do all reads the application wants, and the server will take
 910  * care of managing the end of file boundary.
 911  *
 912  * This function also eliminates unnecessarily updating the file's
 913  * atime locally, as the NFS server sets the file's atime, and this
 914  * client must read the updated atime from the server back into its
 915  * cache.
 916  */
 917 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 918                                 unsigned long nr_segs, loff_t pos)
 919 {
 920         ssize_t retval = -EINVAL;
 921         struct file *file = iocb->ki_filp;
 922         struct address_space *mapping = file->f_mapping;
 923         size_t count;
 924
 925         count = iov_length(iov, nr_segs);
 926         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 927
 928         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 929                 file->f_path.dentry->d_parent->d_name.name,
 930                 file->f_path.dentry->d_name.name,
 931                 count, (long long) pos);
 932
 933         retval = 0;
 934         if (!count)
 935                 goto out;
 936
 937         retval = nfs_sync_mapping(mapping);
 938         if (retval)
 939                 goto out;
 940
 941         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 942         if (retval > 0)
 943                 iocb->ki_pos = pos + retval;
 944
 945 out:
 946         return retval;
 947 }
 948
 949 /**
 950  * nfs_file_direct_write - file direct write operation for NFS files
 951  * @iocb: target I/O control block
 952  * @iov: vector of user buffers from which to write data
 953  * @nr_segs: size of iov vector
 954  * @pos: byte offset in file where writing starts
 955  *
 956  * We use this function for direct writes instead of calling
 957  * generic_file_aio_write() in order to avoid taking the inode
 958  * semaphore and updating the i_size.  The NFS server will set
 959  * the new i_size and this client must read the updated size
 960  * back into its cache.  We let the server do generic write
 961  * parameter checking and report problems.
 962  *
 963  * We eliminate local atime updates, see direct read above.
 964  *
 965  * We avoid unnecessary page cache invalidations for normal cached
 966  * readers of this file.
 967  *
 968  * Note that O_APPEND is not supported for NFS direct writes, as there
 969  * is no atomic O_APPEND write facility in the NFS protocol.
 970  */
 971 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 972                                 unsigned long nr_segs, loff_t pos)
 973 {
 974         ssize_t retval = -EINVAL;
 975         struct file *file = iocb->ki_filp;
 976         struct address_space *mapping = file->f_mapping;
 977         size_t count;
 978
 979         count = iov_length(iov, nr_segs);
 980         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 981
 982         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 983                 file->f_path.dentry->d_parent->d_name.name,
 984                 file->f_path.dentry->d_name.name,
 985                 count, (long long) pos);
 986
 987         retval = generic_write_checks(file, &pos, &count, 0);
 988         if (retval)
 989                 goto out;
 990
 991         retval = -EINVAL;
 992         if ((ssize_t) count < 0)
 993                 goto out;
 994         retval = 0;
 995         if (!count)
 996                 goto out;
 997
 998         retval = nfs_sync_mapping(mapping);
 999         if (retval)
1000                 goto out;
1001
1002         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
1003
1004         if (retval > 0)
1005                 iocb->ki_pos = pos + retval;
1006
1007 out:
1008         return retval;
1009 }
1010
1011 /**
1012  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1013  *
1014  */
1015 int __init nfs_init_directcache(void)
1016 {
1017         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1018                                                 sizeof(struct nfs_direct_req),
1019                                                 0, (SLAB_RECLAIM_ACCOUNT|
1020                                                         SLAB_MEM_SPREAD),
1021                                                 NULL);
1022         if (nfs_direct_cachep == NULL)
1023                 return -ENOMEM;
1024
1025         return 0;
1026 }
1027
1028 /**
1029  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1030  *
1031  */
1032 void nfs_destroy_directcache(void)
1033 {
1034         kmem_cache_destroy(nfs_direct_cachep);
1035 }