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         if (put_dreq(dreq))
 411                 nfs_direct_complete(dreq);
 412
 413         if (requested_bytes != 0)
 414                 return 0;
 415
 416         if (result < 0)
 417                 return result;
 418         return -EIO;
 419 }
 420
 421 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 422                                unsigned long nr_segs, loff_t pos)
 423 {
 424         ssize_t result = -ENOMEM;
 425         struct inode *inode = iocb->ki_filp->f_mapping->host;
 426         struct nfs_direct_req *dreq;
 427
 428         dreq = nfs_direct_req_alloc();
 429         if (dreq == NULL)
 430                 goto out;
 431
 432         dreq->inode = inode;
 433         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 434         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 435         if (dreq->l_ctx == NULL)
 436                 goto out_release;
 437         if (!is_sync_kiocb(iocb))
 438                 dreq->iocb = iocb;
 439
 440         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 441         if (!result)
 442                 result = nfs_direct_wait(dreq);
 443 out_release:
 444         nfs_direct_req_release(dreq);
 445 out:
 446         return result;
 447 }
 448
 449 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 450 {
 451         while (!list_empty(&dreq->rewrite_list)) {
 452                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 453                 list_del(&data->pages);
 454                 nfs_direct_release_pages(data->pagevec, data->npages);
 455                 nfs_writedata_free(data);
 456         }
 457 }
 458
 459 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 460 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 461 {
 462         struct inode *inode = dreq->inode;
 463         struct list_head *p;
 464         struct nfs_write_data *data;
 465         struct rpc_task *task;
 466         struct rpc_message msg = {
 467                 .rpc_cred = dreq->ctx->cred,
 468         };
 469         struct rpc_task_setup task_setup_data = {
 470                 .rpc_client = NFS_CLIENT(inode),
 471                 .rpc_message = &msg,
 472                 .callback_ops = &nfs_write_direct_ops,
 473                 .workqueue = nfsiod_workqueue,
 474                 .flags = RPC_TASK_ASYNC,
 475         };
 476
 477         dreq->count = 0;
 478         get_dreq(dreq);
 479
 480         list_for_each(p, &dreq->rewrite_list) {
 481                 data = list_entry(p, struct nfs_write_data, pages);
 482
 483                 get_dreq(dreq);
 484
 485                 /* Use stable writes */
 486                 data->args.stable = NFS_FILE_SYNC;
 487
 488                 /*
 489                  * Reset data->res.
 490                  */
 491                 nfs_fattr_init(&data->fattr);
 492                 data->res.count = data->args.count;
 493                 memset(&data->verf, 0, sizeof(data->verf));
 494
 495                 /*
 496                  * Reuse data->task; data->args should not have changed
 497                  * since the original request was sent.
 498                  */
 499                 task_setup_data.task = &data->task;
 500                 task_setup_data.callback_data = data;
 501                 msg.rpc_argp = &data->args;
 502                 msg.rpc_resp = &data->res;
 503                 NFS_PROTO(inode)->write_setup(data, &msg);
 504
 505                 /*
 506                  * We're called via an RPC callback, so BKL is already held.
 507                  */
 508                 task = rpc_run_task(&task_setup_data);
 509                 if (!IS_ERR(task))
 510                         rpc_put_task(task);
 511
 512                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
 513                                 data->task.tk_pid,
 514                                 inode->i_sb->s_id,
 515                                 (long long)NFS_FILEID(inode),
 516                                 data->args.count,
 517                                 (unsigned long long)data->args.offset);
 518         }
 519
 520         if (put_dreq(dreq))
 521                 nfs_direct_write_complete(dreq, inode);
 522 }
 523
 524 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
 525 {
 526         struct nfs_write_data *data = calldata;
 527
 528         /* Call the NFS version-specific code */
 529         NFS_PROTO(data->inode)->commit_done(task, data);
 530 }
 531
 532 static void nfs_direct_commit_release(void *calldata)
 533 {
 534         struct nfs_write_data *data = calldata;
 535         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 536         int status = data->task.tk_status;
 537
 538         if (status < 0) {
 539                 dprintk("NFS: %5u commit failed with error %d.\n",
 540                                 data->task.tk_pid, status);
 541                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 542         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 543                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 544                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 545         }
 546
 547         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 548         nfs_direct_write_complete(dreq, data->inode);
 549         nfs_commit_free(data);
 550 }
 551
 552 static const struct rpc_call_ops nfs_commit_direct_ops = {
 553 #if defined(CONFIG_NFS_V4_1)
 554         .rpc_call_prepare = nfs_write_prepare,
 555 #endif /* CONFIG_NFS_V4_1 */
 556         .rpc_call_done = nfs_direct_commit_result,
 557         .rpc_release = nfs_direct_commit_release,
 558 };
 559
 560 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 561 {
 562         struct nfs_write_data *data = dreq->commit_data;
 563         struct rpc_task *task;
 564         struct rpc_message msg = {
 565                 .rpc_argp = &data->args,
 566                 .rpc_resp = &data->res,
 567                 .rpc_cred = dreq->ctx->cred,
 568         };
 569         struct rpc_task_setup task_setup_data = {
 570                 .task = &data->task,
 571                 .rpc_client = NFS_CLIENT(dreq->inode),
 572                 .rpc_message = &msg,
 573                 .callback_ops = &nfs_commit_direct_ops,
 574                 .callback_data = data,
 575                 .workqueue = nfsiod_workqueue,
 576                 .flags = RPC_TASK_ASYNC,
 577         };
 578
 579         data->inode = dreq->inode;
 580         data->cred = msg.rpc_cred;
 581
 582         data->args.fh = NFS_FH(data->inode);
 583         data->args.offset = 0;
 584         data->args.count = 0;
 585         data->args.context = dreq->ctx;
 586         data->args.lock_context = dreq->l_ctx;
 587         data->res.count = 0;
 588         data->res.fattr = &data->fattr;
 589         data->res.verf = &data->verf;
 590         nfs_fattr_init(&data->fattr);
 591
 592         NFS_PROTO(data->inode)->commit_setup(data, &msg);
 593
 594         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
 595         dreq->commit_data = NULL;
 596
 597         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
 598
 599         task = rpc_run_task(&task_setup_data);
 600         if (!IS_ERR(task))
 601                 rpc_put_task(task);
 602 }
 603
 604 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 605 {
 606         int flags = dreq->flags;
 607
 608         dreq->flags = 0;
 609         switch (flags) {
 610                 case NFS_ODIRECT_DO_COMMIT:
 611                         nfs_direct_commit_schedule(dreq);
 612                         break;
 613                 case NFS_ODIRECT_RESCHED_WRITES:
 614                         nfs_direct_write_reschedule(dreq);
 615                         break;
 616                 default:
 617                         if (dreq->commit_data != NULL)
 618                                 nfs_commit_free(dreq->commit_data);
 619                         nfs_direct_free_writedata(dreq);
 620                         nfs_zap_mapping(inode, inode->i_mapping);
 621                         nfs_direct_complete(dreq);
 622         }
 623 }
 624
 625 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 626 {
 627         dreq->commit_data = nfs_commitdata_alloc();
 628         if (dreq->commit_data != NULL)
 629                 dreq->commit_data->req = (struct nfs_page *) dreq;
 630 }
 631 #else
 632 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
 633 {
 634         dreq->commit_data = NULL;
 635 }
 636
 637 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 638 {
 639         nfs_direct_free_writedata(dreq);
 640         nfs_zap_mapping(inode, inode->i_mapping);
 641         nfs_direct_complete(dreq);
 642 }
 643 #endif
 644
 645 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 646 {
 647         struct nfs_write_data *data = calldata;
 648
 649         if (nfs_writeback_done(task, data) != 0)
 650                 return;
 651 }
 652
 653 /*
 654  * NB: Return the value of the first error return code.  Subsequent
 655  *     errors after the first one are ignored.
 656  */
 657 static void nfs_direct_write_release(void *calldata)
 658 {
 659         struct nfs_write_data *data = calldata;
 660         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
 661         int status = data->task.tk_status;
 662
 663         spin_lock(&dreq->lock);
 664
 665         if (unlikely(status < 0)) {
 666                 /* An error has occurred, so we should not commit */
 667                 dreq->flags = 0;
 668                 dreq->error = status;
 669         }
 670         if (unlikely(dreq->error != 0))
 671                 goto out_unlock;
 672
 673         dreq->count += data->res.count;
 674
 675         if (data->res.verf->committed != NFS_FILE_SYNC) {
 676                 switch (dreq->flags) {
 677                         case 0:
 678                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
 679                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 680                                 break;
 681                         case NFS_ODIRECT_DO_COMMIT:
 682                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
 683                                         dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
 684                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 685                                 }
 686                 }
 687         }
 688 out_unlock:
 689         spin_unlock(&dreq->lock);
 690
 691         if (put_dreq(dreq))
 692                 nfs_direct_write_complete(dreq, data->inode);
 693 }
 694
 695 static const struct rpc_call_ops nfs_write_direct_ops = {
 696 #if defined(CONFIG_NFS_V4_1)
 697         .rpc_call_prepare = nfs_write_prepare,
 698 #endif /* CONFIG_NFS_V4_1 */
 699         .rpc_call_done = nfs_direct_write_result,
 700         .rpc_release = nfs_direct_write_release,
 701 };
 702
 703 /*
 704  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 705  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 706  * bail and stop sending more writes.  Write length accounting is
 707  * handled automatically by nfs_direct_write_result().  Otherwise, if
 708  * no requests have been sent, just return an error.
 709  */
 710 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
 711                                                  const struct iovec *iov,
 712                                                  loff_t pos, int sync)
 713 {
 714         struct nfs_open_context *ctx = dreq->ctx;
 715         struct inode *inode = ctx->path.dentry->d_inode;
 716         unsigned long user_addr = (unsigned long)iov->iov_base;
 717         size_t count = iov->iov_len;
 718         struct rpc_task *task;
 719         struct rpc_message msg = {
 720                 .rpc_cred = ctx->cred,
 721         };
 722         struct rpc_task_setup task_setup_data = {
 723                 .rpc_client = NFS_CLIENT(inode),
 724                 .rpc_message = &msg,
 725                 .callback_ops = &nfs_write_direct_ops,
 726                 .workqueue = nfsiod_workqueue,
 727                 .flags = RPC_TASK_ASYNC,
 728         };
 729         size_t wsize = NFS_SERVER(inode)->wsize;
 730         unsigned int pgbase;
 731         int result;
 732         ssize_t started = 0;
 733
 734         do {
 735                 struct nfs_write_data *data;
 736                 size_t bytes;
 737
 738                 pgbase = user_addr & ~PAGE_MASK;
 739                 bytes = min(wsize,count);
 740
 741                 result = -ENOMEM;
 742                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
 743                 if (unlikely(!data))
 744                         break;
 745
 746                 down_read(&current->mm->mmap_sem);
 747                 result = get_user_pages(current, current->mm, user_addr,
 748                                         data->npages, 0, 0, data->pagevec, NULL);
 749                 up_read(&current->mm->mmap_sem);
 750                 if (result < 0) {
 751                         nfs_writedata_free(data);
 752                         break;
 753                 }
 754                 if ((unsigned)result < data->npages) {
 755                         bytes = result * PAGE_SIZE;
 756                         if (bytes <= pgbase) {
 757                                 nfs_direct_release_pages(data->pagevec, result);
 758                                 nfs_writedata_free(data);
 759                                 break;
 760                         }
 761                         bytes -= pgbase;
 762                         data->npages = result;
 763                 }
 764
 765                 get_dreq(dreq);
 766
 767                 list_move_tail(&data->pages, &dreq->rewrite_list);
 768
 769                 data->req = (struct nfs_page *) dreq;
 770                 data->inode = inode;
 771                 data->cred = msg.rpc_cred;
 772                 data->args.fh = NFS_FH(inode);
 773                 data->args.context = ctx;
 774                 data->args.lock_context = dreq->l_ctx;
 775                 data->args.offset = pos;
 776                 data->args.pgbase = pgbase;
 777                 data->args.pages = data->pagevec;
 778                 data->args.count = bytes;
 779                 data->args.stable = sync;
 780                 data->res.fattr = &data->fattr;
 781                 data->res.count = bytes;
 782                 data->res.verf = &data->verf;
 783                 nfs_fattr_init(&data->fattr);
 784
 785                 task_setup_data.task = &data->task;
 786                 task_setup_data.callback_data = data;
 787                 msg.rpc_argp = &data->args;
 788                 msg.rpc_resp = &data->res;
 789                 NFS_PROTO(inode)->write_setup(data, &msg);
 790
 791                 task = rpc_run_task(&task_setup_data);
 792                 if (IS_ERR(task))
 793                         break;
 794                 rpc_put_task(task);
 795
 796                 dprintk("NFS: %5u initiated direct write call "
 797                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
 798                                 data->task.tk_pid,
 799                                 inode->i_sb->s_id,
 800                                 (long long)NFS_FILEID(inode),
 801                                 bytes,
 802                                 (unsigned long long)data->args.offset);
 803
 804                 started += bytes;
 805                 user_addr += bytes;
 806                 pos += bytes;
 807
 808                 /* FIXME: Remove this useless math from the final patch */
 809                 pgbase += bytes;
 810                 pgbase &= ~PAGE_MASK;
 811                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
 812
 813                 count -= bytes;
 814         } while (count != 0);
 815
 816         if (started)
 817                 return started;
 818         return result < 0 ? (ssize_t) result : -EFAULT;
 819 }
 820
 821 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 822                                                const struct iovec *iov,
 823                                                unsigned long nr_segs,
 824                                                loff_t pos, int sync)
 825 {
 826         ssize_t result = 0;
 827         size_t requested_bytes = 0;
 828         unsigned long seg;
 829
 830         get_dreq(dreq);
 831
 832         for (seg = 0; seg < nr_segs; seg++) {
 833                 const struct iovec *vec = &iov[seg];
 834                 result = nfs_direct_write_schedule_segment(dreq, vec,
 835                                                            pos, sync);
 836                 if (result < 0)
 837                         break;
 838                 requested_bytes += result;
 839                 if ((size_t)result < vec->iov_len)
 840                         break;
 841                 pos += vec->iov_len;
 842         }
 843
 844         if (put_dreq(dreq))
 845                 nfs_direct_write_complete(dreq, dreq->inode);
 846
 847         if (requested_bytes != 0)
 848                 return 0;
 849
 850         if (result < 0)
 851                 return result;
 852         return -EIO;
 853 }
 854
 855 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 856                                 unsigned long nr_segs, loff_t pos,
 857                                 size_t count)
 858 {
 859         ssize_t result = -ENOMEM;
 860         struct inode *inode = iocb->ki_filp->f_mapping->host;
 861         struct nfs_direct_req *dreq;
 862         size_t wsize = NFS_SERVER(inode)->wsize;
 863         int sync = NFS_UNSTABLE;
 864
 865         dreq = nfs_direct_req_alloc();
 866         if (!dreq)
 867                 goto out;
 868         nfs_alloc_commit_data(dreq);
 869
 870         if (dreq->commit_data == NULL || count < wsize)
 871                 sync = NFS_FILE_SYNC;
 872
 873         dreq->inode = inode;
 874         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 875         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 876         if (dreq->l_ctx == NULL)
 877                 goto out_release;
 878         if (!is_sync_kiocb(iocb))
 879                 dreq->iocb = iocb;
 880
 881         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
 882         if (!result)
 883                 result = nfs_direct_wait(dreq);
 884 out_release:
 885         nfs_direct_req_release(dreq);
 886 out:
 887         return result;
 888 }
 889
 890 /**
 891  * nfs_file_direct_read - file direct read operation for NFS files
 892  * @iocb: target I/O control block
 893  * @iov: vector of user buffers into which to read data
 894  * @nr_segs: size of iov vector
 895  * @pos: byte offset in file where reading starts
 896  *
 897  * We use this function for direct reads instead of calling
 898  * generic_file_aio_read() in order to avoid gfar's check to see if
 899  * the request starts before the end of the file.  For that check
 900  * to work, we must generate a GETATTR before each direct read, and
 901  * even then there is a window between the GETATTR and the subsequent
 902  * READ where the file size could change.  Our preference is simply
 903  * to do all reads the application wants, and the server will take
 904  * care of managing the end of file boundary.
 905  *
 906  * This function also eliminates unnecessarily updating the file's
 907  * atime locally, as the NFS server sets the file's atime, and this
 908  * client must read the updated atime from the server back into its
 909  * cache.
 910  */
 911 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 912                                 unsigned long nr_segs, loff_t pos)
 913 {
 914         ssize_t retval = -EINVAL;
 915         struct file *file = iocb->ki_filp;
 916         struct address_space *mapping = file->f_mapping;
 917         size_t count;
 918
 919         count = iov_length(iov, nr_segs);
 920         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 921
 922         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 923                 file->f_path.dentry->d_parent->d_name.name,
 924                 file->f_path.dentry->d_name.name,
 925                 count, (long long) pos);
 926
 927         retval = 0;
 928         if (!count)
 929                 goto out;
 930
 931         retval = nfs_sync_mapping(mapping);
 932         if (retval)
 933                 goto out;
 934
 935         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 936         if (retval > 0)
 937                 iocb->ki_pos = pos + retval;
 938
 939 out:
 940         return retval;
 941 }
 942
 943 /**
 944  * nfs_file_direct_write - file direct write operation for NFS files
 945  * @iocb: target I/O control block
 946  * @iov: vector of user buffers from which to write data
 947  * @nr_segs: size of iov vector
 948  * @pos: byte offset in file where writing starts
 949  *
 950  * We use this function for direct writes instead of calling
 951  * generic_file_aio_write() in order to avoid taking the inode
 952  * semaphore and updating the i_size.  The NFS server will set
 953  * the new i_size and this client must read the updated size
 954  * back into its cache.  We let the server do generic write
 955  * parameter checking and report problems.
 956  *
 957  * We eliminate local atime updates, see direct read above.
 958  *
 959  * We avoid unnecessary page cache invalidations for normal cached
 960  * readers of this file.
 961  *
 962  * Note that O_APPEND is not supported for NFS direct writes, as there
 963  * is no atomic O_APPEND write facility in the NFS protocol.
 964  */
 965 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 966                                 unsigned long nr_segs, loff_t pos)
 967 {
 968         ssize_t retval = -EINVAL;
 969         struct file *file = iocb->ki_filp;
 970         struct address_space *mapping = file->f_mapping;
 971         size_t count;
 972
 973         count = iov_length(iov, nr_segs);
 974         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 975
 976         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 977                 file->f_path.dentry->d_parent->d_name.name,
 978                 file->f_path.dentry->d_name.name,
 979                 count, (long long) pos);
 980
 981         retval = generic_write_checks(file, &pos, &count, 0);
 982         if (retval)
 983                 goto out;
 984
 985         retval = -EINVAL;
 986         if ((ssize_t) count < 0)
 987                 goto out;
 988         retval = 0;
 989         if (!count)
 990                 goto out;
 991
 992         retval = nfs_sync_mapping(mapping);
 993         if (retval)
 994                 goto out;
 995
 996         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
 997
 998         if (retval > 0)
 999                 iocb->ki_pos = pos + retval;
1000
1001 out:
1002         return retval;
1003 }
1004
1005 /**
1006  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1007  *
1008  */
1009 int __init nfs_init_directcache(void)
1010 {
1011         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1012                                                 sizeof(struct nfs_direct_req),
1013                                                 0, (SLAB_RECLAIM_ACCOUNT|
1014                                                         SLAB_MEM_SPREAD),
1015                                                 NULL);
1016         if (nfs_direct_cachep == NULL)
1017                 return -ENOMEM;
1018
1019         return 0;
1020 }
1021
1022 /**
1023  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1024  *
1025  */
1026 void nfs_destroy_directcache(void)
1027 {
1028         kmem_cache_destroy(nfs_direct_cachep);
1029 }