linux/fs/nfs/direct.c
<<
>>
Prefs
   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
  62static struct kmem_cache *nfs_direct_cachep;
  63
  64/*
  65 * This represents a set of asynchronous requests that we're waiting on
  66 */
  67struct 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
  92static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  93static const struct rpc_call_ops nfs_write_direct_ops;
  94
  95static inline void get_dreq(struct nfs_direct_req *dreq)
  96{
  97        atomic_inc(&dreq->io_count);
  98}
  99
 100static 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 */
 118ssize_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
 127static 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
 143static 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
 150static 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
 174static 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
 185static 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 */
 193static 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
 208out:
 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 */
 216static 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 */
 234static 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
 241static 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
 266static 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 */
 281static 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
 388static 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
 424static 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);
 446out_release:
 447        nfs_direct_req_release(dreq);
 448out:
 449        return result;
 450}
 451
 452static 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)
 463static 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
 527static 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
 535static 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
 555static 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
 563static 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
 607static 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
 628static 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
 635static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 636{
 637        dreq->commit_data = NULL;
 638}
 639
 640static 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
 648static 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 */
 660static 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        }
 691out_unlock:
 692        spin_unlock(&dreq->lock);
 693
 694        if (put_dreq(dreq))
 695                nfs_direct_write_complete(dreq, data->inode);
 696}
 697
 698static 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 */
 713static 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
 824static 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
 861static 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);
 890out_release:
 891        nfs_direct_req_release(dreq);
 892out:
 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 */
 917ssize_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
 945out:
 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 */
 971ssize_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
1007out:
1008        return retval;
1009}
1010
1011/**
1012 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1013 *
1014 */
1015int __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 */
1032void nfs_destroy_directcache(void)
1033{
1034        kmem_cache_destroy(nfs_direct_cachep);
1035}
1036