kernel/relay.c

   1 /*
   2  * Public API and common code for kernel->userspace relay file support.
   3  *
   4  * See Documentation/filesystems/relayfs.txt for an overview of relayfs.
   5  *
   6  * Copyright (C) 2002-2005 - Tom Zanussi (zanussi@us.ibm.com), IBM Corp
   7  * Copyright (C) 1999-2005 - Karim Yaghmour (karim@opersys.com)
   8  *
   9  * Moved to kernel/relay.c by Paul Mundt, 2006.
  10  * November 2006 - CPU hotplug support by Mathieu Desnoyers
  11  *      (mathieu.desnoyers@polymtl.ca)
  12  *
  13  * This file is released under the GPL.
  14  */
  15 #include <linux/errno.h>
  16 #include <linux/stddef.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <linux/string.h>
  20 #include <linux/relay.h>
  21 #include <linux/vmalloc.h>
  22 #include <linux/mm.h>
  23 #include <linux/cpu.h>
  24
  25 /* list of open channels, for cpu hotplug */
  26 static DEFINE_MUTEX(relay_channels_mutex);
  27 static LIST_HEAD(relay_channels);
  28
  29 /*
  30  * close() vm_op implementation for relay file mapping.
  31  */
  32 static void relay_file_mmap_close(struct vm_area_struct *vma)
  33 {
  34         struct rchan_buf *buf = vma->vm_private_data;
  35         buf->chan->cb->buf_unmapped(buf, vma->vm_file);
  36 }
  37
  38 /*
  39  * nopage() vm_op implementation for relay file mapping.
  40  */
  41 static struct page *relay_buf_nopage(struct vm_area_struct *vma,
  42                                      unsigned long address,
  43                                      int *type)
  44 {
  45         struct page *page;
  46         struct rchan_buf *buf = vma->vm_private_data;
  47         unsigned long offset = address - vma->vm_start;
  48
  49         if (address > vma->vm_end)
  50                 return NOPAGE_SIGBUS; /* Disallow mremap */
  51         if (!buf)
  52                 return NOPAGE_OOM;
  53
  54         page = vmalloc_to_page(buf->start + offset);
  55         if (!page)
  56                 return NOPAGE_OOM;
  57         get_page(page);
  58
  59         if (type)
  60                 *type = VM_FAULT_MINOR;
  61
  62         return page;
  63 }
  64
  65 /*
  66  * vm_ops for relay file mappings.
  67  */
  68 static struct vm_operations_struct relay_file_mmap_ops = {
  69         .nopage = relay_buf_nopage,
  70         .close = relay_file_mmap_close,
  71 };
  72
  73 /**
  74  *      relay_mmap_buf: - mmap channel buffer to process address space
  75  *      @buf: relay channel buffer
  76  *      @vma: vm_area_struct describing memory to be mapped
  77  *
  78  *      Returns 0 if ok, negative on error
  79  *
  80  *      Caller should already have grabbed mmap_sem.
  81  */
  82 int relay_mmap_buf(struct rchan_buf *buf, struct vm_area_struct *vma)
  83 {
  84         unsigned long length = vma->vm_end - vma->vm_start;
  85         struct file *filp = vma->vm_file;
  86
  87         if (!buf)
  88                 return -EBADF;
  89
  90         if (length != (unsigned long)buf->chan->alloc_size)
  91                 return -EINVAL;
  92
  93         vma->vm_ops = &relay_file_mmap_ops;
  94         vma->vm_private_data = buf;
  95         buf->chan->cb->buf_mapped(buf, filp);
  96
  97         return 0;
  98 }
  99
 100 /**
 101  *      relay_alloc_buf - allocate a channel buffer
 102  *      @buf: the buffer struct
 103  *      @size: total size of the buffer
 104  *
 105  *      Returns a pointer to the resulting buffer, %NULL if unsuccessful. The
 106  *      passed in size will get page aligned, if it isn't already.
 107  */
 108 static void *relay_alloc_buf(struct rchan_buf *buf, size_t *size)
 109 {
 110         void *mem;
 111         unsigned int i, j, n_pages;
 112
 113         *size = PAGE_ALIGN(*size);
 114         n_pages = *size >> PAGE_SHIFT;
 115
 116         buf->page_array = kcalloc(n_pages, sizeof(struct page *), GFP_KERNEL);
 117         if (!buf->page_array)
 118                 return NULL;
 119
 120         for (i = 0; i < n_pages; i++) {
 121                 buf->page_array[i] = alloc_page(GFP_KERNEL);
 122                 if (unlikely(!buf->page_array[i]))
 123                         goto depopulate;
 124         }
 125         mem = vmap(buf->page_array, n_pages, VM_MAP, PAGE_KERNEL);
 126         if (!mem)
 127                 goto depopulate;
 128
 129         memset(mem, 0, *size);
 130         buf->page_count = n_pages;
 131         return mem;
 132
 133 depopulate:
 134         for (j = 0; j < i; j++)
 135                 __free_page(buf->page_array[j]);
 136         kfree(buf->page_array);
 137         return NULL;
 138 }
 139
 140 /**
 141  *      relay_create_buf - allocate and initialize a channel buffer
 142  *      @chan: the relay channel
 143  *
 144  *      Returns channel buffer if successful, %NULL otherwise.
 145  */
 146 struct rchan_buf *relay_create_buf(struct rchan *chan)
 147 {
 148         struct rchan_buf *buf = kzalloc(sizeof(struct rchan_buf), GFP_KERNEL);
 149         if (!buf)
 150                 return NULL;
 151
 152         buf->padding = kmalloc(chan->n_subbufs * sizeof(size_t *), GFP_KERNEL);
 153         if (!buf->padding)
 154                 goto free_buf;
 155
 156         buf->start = relay_alloc_buf(buf, &chan->alloc_size);
 157         if (!buf->start)
 158                 goto free_buf;
 159
 160         buf->chan = chan;
 161         kref_get(&buf->chan->kref);
 162         return buf;
 163
 164 free_buf:
 165         kfree(buf->padding);
 166         kfree(buf);
 167         return NULL;
 168 }
 169
 170 /**
 171  *      relay_destroy_channel - free the channel struct
 172  *      @kref: target kernel reference that contains the relay channel
 173  *
 174  *      Should only be called from kref_put().
 175  */
 176 void relay_destroy_channel(struct kref *kref)
 177 {
 178         struct rchan *chan = container_of(kref, struct rchan, kref);
 179         kfree(chan);
 180 }
 181
 182 /**
 183  *      relay_destroy_buf - destroy an rchan_buf struct and associated buffer
 184  *      @buf: the buffer struct
 185  */
 186 void relay_destroy_buf(struct rchan_buf *buf)
 187 {
 188         struct rchan *chan = buf->chan;
 189         unsigned int i;
 190
 191         if (likely(buf->start)) {
 192                 vunmap(buf->start);
 193                 for (i = 0; i < buf->page_count; i++)
 194                         __free_page(buf->page_array[i]);
 195                 kfree(buf->page_array);
 196         }
 197         chan->buf[buf->cpu] = NULL;
 198         kfree(buf->padding);
 199         kfree(buf);
 200         kref_put(&chan->kref, relay_destroy_channel);
 201 }
 202
 203 /**
 204  *      relay_remove_buf - remove a channel buffer
 205  *      @kref: target kernel reference that contains the relay buffer
 206  *
 207  *      Removes the file from the fileystem, which also frees the
 208  *      rchan_buf_struct and the channel buffer.  Should only be called from
 209  *      kref_put().
 210  */
 211 void relay_remove_buf(struct kref *kref)
 212 {
 213         struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
 214         buf->chan->cb->remove_buf_file(buf->dentry);
 215         relay_destroy_buf(buf);
 216 }
 217
 218 /**
 219  *      relay_buf_empty - boolean, is the channel buffer empty?
 220  *      @buf: channel buffer
 221  *
 222  *      Returns 1 if the buffer is empty, 0 otherwise.
 223  */
 224 int relay_buf_empty(struct rchan_buf *buf)
 225 {
 226         return (buf->subbufs_produced - buf->subbufs_consumed) ? 0 : 1;
 227 }
 228 EXPORT_SYMBOL_GPL(relay_buf_empty);
 229
 230 /**
 231  *      relay_buf_full - boolean, is the channel buffer full?
 232  *      @buf: channel buffer
 233  *
 234  *      Returns 1 if the buffer is full, 0 otherwise.
 235  */
 236 int relay_buf_full(struct rchan_buf *buf)
 237 {
 238         size_t ready = buf->subbufs_produced - buf->subbufs_consumed;
 239         return (ready >= buf->chan->n_subbufs) ? 1 : 0;
 240 }
 241 EXPORT_SYMBOL_GPL(relay_buf_full);
 242
 243 /*
 244  * High-level relay kernel API and associated functions.
 245  */
 246
 247 /*
 248  * rchan_callback implementations defining default channel behavior.  Used
 249  * in place of corresponding NULL values in client callback struct.
 250  */
 251
 252 /*
 253  * subbuf_start() default callback.  Does nothing.
 254  */
 255 static int subbuf_start_default_callback (struct rchan_buf *buf,
 256                                           void *subbuf,
 257                                           void *prev_subbuf,
 258                                           size_t prev_padding)
 259 {
 260         if (relay_buf_full(buf))
 261                 return 0;
 262
 263         return 1;
 264 }
 265
 266 /*
 267  * buf_mapped() default callback.  Does nothing.
 268  */
 269 static void buf_mapped_default_callback(struct rchan_buf *buf,
 270                                         struct file *filp)
 271 {
 272 }
 273
 274 /*
 275  * buf_unmapped() default callback.  Does nothing.
 276  */
 277 static void buf_unmapped_default_callback(struct rchan_buf *buf,
 278                                           struct file *filp)
 279 {
 280 }
 281
 282 /*
 283  * create_buf_file_create() default callback.  Does nothing.
 284  */
 285 static struct dentry *create_buf_file_default_callback(const char *filename,
 286                                                        struct dentry *parent,
 287                                                        int mode,
 288                                                        struct rchan_buf *buf,
 289                                                        int *is_global)
 290 {
 291         return NULL;
 292 }
 293
 294 /*
 295  * remove_buf_file() default callback.  Does nothing.
 296  */
 297 static int remove_buf_file_default_callback(struct dentry *dentry)
 298 {
 299         return -EINVAL;
 300 }
 301
 302 /* relay channel default callbacks */
 303 static struct rchan_callbacks default_channel_callbacks = {
 304         .subbuf_start = subbuf_start_default_callback,
 305         .buf_mapped = buf_mapped_default_callback,
 306         .buf_unmapped = buf_unmapped_default_callback,
 307         .create_buf_file = create_buf_file_default_callback,
 308         .remove_buf_file = remove_buf_file_default_callback,
 309 };
 310
 311 /**
 312  *      wakeup_readers - wake up readers waiting on a channel
 313  *      @work: work struct that contains the the channel buffer
 314  *
 315  *      This is the work function used to defer reader waking.  The
 316  *      reason waking is deferred is that calling directly from write
 317  *      causes problems if you're writing from say the scheduler.
 318  */
 319 static void wakeup_readers(struct work_struct *work)
 320 {
 321         struct rchan_buf *buf =
 322                 container_of(work, struct rchan_buf, wake_readers.work);
 323         wake_up_interruptible(&buf->read_wait);
 324 }
 325
 326 /**
 327  *      __relay_reset - reset a channel buffer
 328  *      @buf: the channel buffer
 329  *      @init: 1 if this is a first-time initialization
 330  *
 331  *      See relay_reset() for description of effect.
 332  */
 333 static void __relay_reset(struct rchan_buf *buf, unsigned int init)
 334 {
 335         size_t i;
 336
 337         if (init) {
 338                 init_waitqueue_head(&buf->read_wait);
 339                 kref_init(&buf->kref);
 340                 INIT_DELAYED_WORK(&buf->wake_readers, NULL);
 341         } else {
 342                 cancel_delayed_work(&buf->wake_readers);
 343                 flush_scheduled_work();
 344         }
 345
 346         buf->subbufs_produced = 0;
 347         buf->subbufs_consumed = 0;
 348         buf->bytes_consumed = 0;
 349         buf->finalized = 0;
 350         buf->data = buf->start;
 351         buf->offset = 0;
 352
 353         for (i = 0; i < buf->chan->n_subbufs; i++)
 354                 buf->padding[i] = 0;
 355
 356         buf->chan->cb->subbuf_start(buf, buf->data, NULL, 0);
 357 }
 358
 359 /**
 360  *      relay_reset - reset the channel
 361  *      @chan: the channel
 362  *
 363  *      This has the effect of erasing all data from all channel buffers
 364  *      and restarting the channel in its initial state.  The buffers
 365  *      are not freed, so any mappings are still in effect.
 366  *
 367  *      NOTE. Care should be taken that the channel isn't actually
 368  *      being used by anything when this call is made.
 369  */
 370 void relay_reset(struct rchan *chan)
 371 {
 372         unsigned int i;
 373
 374         if (!chan)
 375                 return;
 376
 377         if (chan->is_global && chan->buf[0]) {
 378                 __relay_reset(chan->buf[0], 0);
 379                 return;
 380         }
 381
 382         mutex_lock(&relay_channels_mutex);
 383         for_each_online_cpu(i)
 384                 if (chan->buf[i])
 385                         __relay_reset(chan->buf[i], 0);
 386         mutex_unlock(&relay_channels_mutex);
 387 }
 388 EXPORT_SYMBOL_GPL(relay_reset);
 389
 390 /*
 391  *      relay_open_buf - create a new relay channel buffer
 392  *
 393  *      used by relay_open() and CPU hotplug.
 394  */
 395 static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
 396 {
 397         struct rchan_buf *buf = NULL;
 398         struct dentry *dentry;
 399         char *tmpname;
 400
 401         if (chan->is_global)
 402                 return chan->buf[0];
 403
 404         tmpname = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 405         if (!tmpname)
 406                 goto end;
 407         snprintf(tmpname, NAME_MAX, "%s%d", chan->base_filename, cpu);
 408
 409         buf = relay_create_buf(chan);
 410         if (!buf)
 411                 goto free_name;
 412
 413         buf->cpu = cpu;
 414         __relay_reset(buf, 1);
 415
 416         /* Create file in fs */
 417         dentry = chan->cb->create_buf_file(tmpname, chan->parent, S_IRUSR,
 418                                            buf, &chan->is_global);
 419         if (!dentry)
 420                 goto free_buf;
 421
 422         buf->dentry = dentry;
 423
 424         if(chan->is_global) {
 425                 chan->buf[0] = buf;
 426                 buf->cpu = 0;
 427         }
 428
 429         goto free_name;
 430
 431 free_buf:
 432         relay_destroy_buf(buf);
 433 free_name:
 434         kfree(tmpname);
 435 end:
 436         return buf;
 437 }
 438
 439 /**
 440  *      relay_close_buf - close a channel buffer
 441  *      @buf: channel buffer
 442  *
 443  *      Marks the buffer finalized and restores the default callbacks.
 444  *      The channel buffer and channel buffer data structure are then freed
 445  *      automatically when the last reference is given up.
 446  */
 447 static void relay_close_buf(struct rchan_buf *buf)
 448 {
 449         buf->finalized = 1;
 450         cancel_delayed_work(&buf->wake_readers);
 451         flush_scheduled_work();
 452         kref_put(&buf->kref, relay_remove_buf);
 453 }
 454
 455 static void setup_callbacks(struct rchan *chan,
 456                                    struct rchan_callbacks *cb)
 457 {
 458         if (!cb) {
 459                 chan->cb = &default_channel_callbacks;
 460                 return;
 461         }
 462
 463         if (!cb->subbuf_start)
 464                 cb->subbuf_start = subbuf_start_default_callback;
 465         if (!cb->buf_mapped)
 466                 cb->buf_mapped = buf_mapped_default_callback;
 467         if (!cb->buf_unmapped)
 468                 cb->buf_unmapped = buf_unmapped_default_callback;
 469         if (!cb->create_buf_file)
 470                 cb->create_buf_file = create_buf_file_default_callback;
 471         if (!cb->remove_buf_file)
 472                 cb->remove_buf_file = remove_buf_file_default_callback;
 473         chan->cb = cb;
 474 }
 475
 476 /**
 477  *
 478  *      relay_hotcpu_callback - CPU hotplug callback
 479  *      @nb: notifier block
 480  *      @action: hotplug action to take
 481  *      @hcpu: CPU number
 482  *
 483  *      Returns the success/failure of the operation. (NOTIFY_OK, NOTIFY_BAD)
 484  */
 485 static int __cpuinit relay_hotcpu_callback(struct notifier_block *nb,
 486                                 unsigned long action,
 487                                 void *hcpu)
 488 {
 489         unsigned int hotcpu = (unsigned long)hcpu;
 490         struct rchan *chan;
 491
 492         switch(action) {
 493         case CPU_UP_PREPARE:
 494                 mutex_lock(&relay_channels_mutex);
 495                 list_for_each_entry(chan, &relay_channels, list) {
 496                         if (chan->buf[hotcpu])
 497                                 continue;
 498                         chan->buf[hotcpu] = relay_open_buf(chan, hotcpu);
 499                         if(!chan->buf[hotcpu]) {
 500                                 printk(KERN_ERR
 501                                         "relay_hotcpu_callback: cpu %d buffer "
 502                                         "creation failed\n", hotcpu);
 503                                 mutex_unlock(&relay_channels_mutex);
 504                                 return NOTIFY_BAD;
 505                         }
 506                 }
 507                 mutex_unlock(&relay_channels_mutex);
 508                 break;
 509         case CPU_DEAD:
 510                 /* No need to flush the cpu : will be flushed upon
 511                  * final relay_flush() call. */
 512                 break;
 513         }
 514         return NOTIFY_OK;
 515 }
 516
 517 /**
 518  *      relay_open - create a new relay channel
 519  *      @base_filename: base name of files to create
 520  *      @parent: dentry of parent directory, %NULL for root directory
 521  *      @subbuf_size: size of sub-buffers
 522  *      @n_subbufs: number of sub-buffers
 523  *      @cb: client callback functions
 524  *      @private_data: user-defined data
 525  *
 526  *      Returns channel pointer if successful, %NULL otherwise.
 527  *
 528  *      Creates a channel buffer for each cpu using the sizes and
 529  *      attributes specified.  The created channel buffer files
 530  *      will be named base_filename0...base_filenameN-1.  File
 531  *      permissions will be %S_IRUSR.
 532  */
 533 struct rchan *relay_open(const char *base_filename,
 534                          struct dentry *parent,
 535                          size_t subbuf_size,
 536                          size_t n_subbufs,
 537                          struct rchan_callbacks *cb,
 538                          void *private_data)
 539 {
 540         unsigned int i;
 541         struct rchan *chan;
 542         if (!base_filename)
 543                 return NULL;
 544
 545         if (!(subbuf_size && n_subbufs))
 546                 return NULL;
 547
 548         chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
 549         if (!chan)
 550                 return NULL;
 551
 552         chan->version = RELAYFS_CHANNEL_VERSION;
 553         chan->n_subbufs = n_subbufs;
 554         chan->subbuf_size = subbuf_size;
 555         chan->alloc_size = FIX_SIZE(subbuf_size * n_subbufs);
 556         chan->parent = parent;
 557         chan->private_data = private_data;
 558         strlcpy(chan->base_filename, base_filename, NAME_MAX);
 559         setup_callbacks(chan, cb);
 560         kref_init(&chan->kref);
 561
 562         mutex_lock(&relay_channels_mutex);
 563         for_each_online_cpu(i) {
 564                 chan->buf[i] = relay_open_buf(chan, i);
 565                 if (!chan->buf[i])
 566                         goto free_bufs;
 567         }
 568         list_add(&chan->list, &relay_channels);
 569         mutex_unlock(&relay_channels_mutex);
 570
 571         return chan;
 572
 573 free_bufs:
 574         for_each_online_cpu(i) {
 575                 if (!chan->buf[i])
 576                         break;
 577                 relay_close_buf(chan->buf[i]);
 578         }
 579
 580         kref_put(&chan->kref, relay_destroy_channel);
 581         mutex_unlock(&relay_channels_mutex);
 582         return NULL;
 583 }
 584 EXPORT_SYMBOL_GPL(relay_open);
 585
 586 /**
 587  *      relay_switch_subbuf - switch to a new sub-buffer
 588  *      @buf: channel buffer
 589  *      @length: size of current event
 590  *
 591  *      Returns either the length passed in or 0 if full.
 592  *
 593  *      Performs sub-buffer-switch tasks such as invoking callbacks,
 594  *      updating padding counts, waking up readers, etc.
 595  */
 596 size_t relay_switch_subbuf(struct rchan_buf *buf, size_t length)
 597 {
 598         void *old, *new;
 599         size_t old_subbuf, new_subbuf;
 600
 601         if (unlikely(length > buf->chan->subbuf_size))
 602                 goto toobig;
 603
 604         if (buf->offset != buf->chan->subbuf_size + 1) {
 605                 buf->prev_padding = buf->chan->subbuf_size - buf->offset;
 606                 old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 607                 buf->padding[old_subbuf] = buf->prev_padding;
 608                 buf->subbufs_produced++;
 609                 buf->dentry->d_inode->i_size += buf->chan->subbuf_size -
 610                         buf->padding[old_subbuf];
 611                 smp_mb();
 612                 if (waitqueue_active(&buf->read_wait)) {
 613                         PREPARE_DELAYED_WORK(&buf->wake_readers,
 614                                              wakeup_readers);
 615                         schedule_delayed_work(&buf->wake_readers, 1);
 616                 }
 617         }
 618
 619         old = buf->data;
 620         new_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 621         new = buf->start + new_subbuf * buf->chan->subbuf_size;
 622         buf->offset = 0;
 623         if (!buf->chan->cb->subbuf_start(buf, new, old, buf->prev_padding)) {
 624                 buf->offset = buf->chan->subbuf_size + 1;
 625                 return 0;
 626         }
 627         buf->data = new;
 628         buf->padding[new_subbuf] = 0;
 629
 630         if (unlikely(length + buf->offset > buf->chan->subbuf_size))
 631                 goto toobig;
 632
 633         return length;
 634
 635 toobig:
 636         buf->chan->last_toobig = length;
 637         return 0;
 638 }
 639 EXPORT_SYMBOL_GPL(relay_switch_subbuf);
 640
 641 /**
 642  *      relay_subbufs_consumed - update the buffer's sub-buffers-consumed count
 643  *      @chan: the channel
 644  *      @cpu: the cpu associated with the channel buffer to update
 645  *      @subbufs_consumed: number of sub-buffers to add to current buf's count
 646  *
 647  *      Adds to the channel buffer's consumed sub-buffer count.
 648  *      subbufs_consumed should be the number of sub-buffers newly consumed,
 649  *      not the total consumed.
 650  *
 651  *      NOTE. Kernel clients don't need to call this function if the channel
 652  *      mode is 'overwrite'.
 653  */
 654 void relay_subbufs_consumed(struct rchan *chan,
 655                             unsigned int cpu,
 656                             size_t subbufs_consumed)
 657 {
 658         struct rchan_buf *buf;
 659
 660         if (!chan)
 661                 return;
 662
 663         if (cpu >= NR_CPUS || !chan->buf[cpu])
 664                 return;
 665
 666         buf = chan->buf[cpu];
 667         buf->subbufs_consumed += subbufs_consumed;
 668         if (buf->subbufs_consumed > buf->subbufs_produced)
 669                 buf->subbufs_consumed = buf->subbufs_produced;
 670 }
 671 EXPORT_SYMBOL_GPL(relay_subbufs_consumed);
 672
 673 /**
 674  *      relay_close - close the channel
 675  *      @chan: the channel
 676  *
 677  *      Closes all channel buffers and frees the channel.
 678  */
 679 void relay_close(struct rchan *chan)
 680 {
 681         unsigned int i;
 682
 683         if (!chan)
 684                 return;
 685
 686         mutex_lock(&relay_channels_mutex);
 687         if (chan->is_global && chan->buf[0])
 688                 relay_close_buf(chan->buf[0]);
 689         else
 690                 for_each_possible_cpu(i)
 691                         if (chan->buf[i])
 692                                 relay_close_buf(chan->buf[i]);
 693
 694         if (chan->last_toobig)
 695                 printk(KERN_WARNING "relay: one or more items not logged "
 696                        "[item size (%Zd) > sub-buffer size (%Zd)]\n",
 697                        chan->last_toobig, chan->subbuf_size);
 698
 699         list_del(&chan->list);
 700         kref_put(&chan->kref, relay_destroy_channel);
 701         mutex_unlock(&relay_channels_mutex);
 702 }
 703 EXPORT_SYMBOL_GPL(relay_close);
 704
 705 /**
 706  *      relay_flush - close the channel
 707  *      @chan: the channel
 708  *
 709  *      Flushes all channel buffers, i.e. forces buffer switch.
 710  */
 711 void relay_flush(struct rchan *chan)
 712 {
 713         unsigned int i;
 714
 715         if (!chan)
 716                 return;
 717
 718         if (chan->is_global && chan->buf[0]) {
 719                 relay_switch_subbuf(chan->buf[0], 0);
 720                 return;
 721         }
 722
 723         mutex_lock(&relay_channels_mutex);
 724         for_each_possible_cpu(i)
 725                 if (chan->buf[i])
 726                         relay_switch_subbuf(chan->buf[i], 0);
 727         mutex_unlock(&relay_channels_mutex);
 728 }
 729 EXPORT_SYMBOL_GPL(relay_flush);
 730
 731 /**
 732  *      relay_file_open - open file op for relay files
 733  *      @inode: the inode
 734  *      @filp: the file
 735  *
 736  *      Increments the channel buffer refcount.
 737  */
 738 static int relay_file_open(struct inode *inode, struct file *filp)
 739 {
 740         struct rchan_buf *buf = inode->i_private;
 741         kref_get(&buf->kref);
 742         filp->private_data = buf;
 743
 744         return 0;
 745 }
 746
 747 /**
 748  *      relay_file_mmap - mmap file op for relay files
 749  *      @filp: the file
 750  *      @vma: the vma describing what to map
 751  *
 752  *      Calls upon relay_mmap_buf() to map the file into user space.
 753  */
 754 static int relay_file_mmap(struct file *filp, struct vm_area_struct *vma)
 755 {
 756         struct rchan_buf *buf = filp->private_data;
 757         return relay_mmap_buf(buf, vma);
 758 }
 759
 760 /**
 761  *      relay_file_poll - poll file op for relay files
 762  *      @filp: the file
 763  *      @wait: poll table
 764  *
 765  *      Poll implemention.
 766  */
 767 static unsigned int relay_file_poll(struct file *filp, poll_table *wait)
 768 {
 769         unsigned int mask = 0;
 770         struct rchan_buf *buf = filp->private_data;
 771
 772         if (buf->finalized)
 773                 return POLLERR;
 774
 775         if (filp->f_mode & FMODE_READ) {
 776                 poll_wait(filp, &buf->read_wait, wait);
 777                 if (!relay_buf_empty(buf))
 778                         mask |= POLLIN | POLLRDNORM;
 779         }
 780
 781         return mask;
 782 }
 783
 784 /**
 785  *      relay_file_release - release file op for relay files
 786  *      @inode: the inode
 787  *      @filp: the file
 788  *
 789  *      Decrements the channel refcount, as the filesystem is
 790  *      no longer using it.
 791  */
 792 static int relay_file_release(struct inode *inode, struct file *filp)
 793 {
 794         struct rchan_buf *buf = filp->private_data;
 795         kref_put(&buf->kref, relay_remove_buf);
 796
 797         return 0;
 798 }
 799
 800 /*
 801  *      relay_file_read_consume - update the consumed count for the buffer
 802  */
 803 static void relay_file_read_consume(struct rchan_buf *buf,
 804                                     size_t read_pos,
 805                                     size_t bytes_consumed)
 806 {
 807         size_t subbuf_size = buf->chan->subbuf_size;
 808         size_t n_subbufs = buf->chan->n_subbufs;
 809         size_t read_subbuf;
 810
 811         if (buf->bytes_consumed + bytes_consumed > subbuf_size) {
 812                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 813                 buf->bytes_consumed = 0;
 814         }
 815
 816         buf->bytes_consumed += bytes_consumed;
 817         read_subbuf = read_pos / buf->chan->subbuf_size;
 818         if (buf->bytes_consumed + buf->padding[read_subbuf] == subbuf_size) {
 819                 if ((read_subbuf == buf->subbufs_produced % n_subbufs) &&
 820                     (buf->offset == subbuf_size))
 821                         return;
 822                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 823                 buf->bytes_consumed = 0;
 824         }
 825 }
 826
 827 /*
 828  *      relay_file_read_avail - boolean, are there unconsumed bytes available?
 829  */
 830 static int relay_file_read_avail(struct rchan_buf *buf, size_t read_pos)
 831 {
 832         size_t subbuf_size = buf->chan->subbuf_size;
 833         size_t n_subbufs = buf->chan->n_subbufs;
 834         size_t produced = buf->subbufs_produced;
 835         size_t consumed = buf->subbufs_consumed;
 836
 837         relay_file_read_consume(buf, read_pos, 0);
 838
 839         if (unlikely(buf->offset > subbuf_size)) {
 840                 if (produced == consumed)
 841                         return 0;
 842                 return 1;
 843         }
 844
 845         if (unlikely(produced - consumed >= n_subbufs)) {
 846                 consumed = (produced / n_subbufs) * n_subbufs;
 847                 buf->subbufs_consumed = consumed;
 848         }
 849
 850         produced = (produced % n_subbufs) * subbuf_size + buf->offset;
 851         consumed = (consumed % n_subbufs) * subbuf_size + buf->bytes_consumed;
 852
 853         if (consumed > produced)
 854                 produced += n_subbufs * subbuf_size;
 855
 856         if (consumed == produced)
 857                 return 0;
 858
 859         return 1;
 860 }
 861
 862 /**
 863  *      relay_file_read_subbuf_avail - return bytes available in sub-buffer
 864  *      @read_pos: file read position
 865  *      @buf: relay channel buffer
 866  */
 867 static size_t relay_file_read_subbuf_avail(size_t read_pos,
 868                                            struct rchan_buf *buf)
 869 {
 870         size_t padding, avail = 0;
 871         size_t read_subbuf, read_offset, write_subbuf, write_offset;
 872         size_t subbuf_size = buf->chan->subbuf_size;
 873
 874         write_subbuf = (buf->data - buf->start) / subbuf_size;
 875         write_offset = buf->offset > subbuf_size ? subbuf_size : buf->offset;
 876         read_subbuf = read_pos / subbuf_size;
 877         read_offset = read_pos % subbuf_size;
 878         padding = buf->padding[read_subbuf];
 879
 880         if (read_subbuf == write_subbuf) {
 881                 if (read_offset + padding < write_offset)
 882                         avail = write_offset - (read_offset + padding);
 883         } else
 884                 avail = (subbuf_size - padding) - read_offset;
 885
 886         return avail;
 887 }
 888
 889 /**
 890  *      relay_file_read_start_pos - find the first available byte to read
 891  *      @read_pos: file read position
 892  *      @buf: relay channel buffer
 893  *
 894  *      If the @read_pos is in the middle of padding, return the
 895  *      position of the first actually available byte, otherwise
 896  *      return the original value.
 897  */
 898 static size_t relay_file_read_start_pos(size_t read_pos,
 899                                         struct rchan_buf *buf)
 900 {
 901         size_t read_subbuf, padding, padding_start, padding_end;
 902         size_t subbuf_size = buf->chan->subbuf_size;
 903         size_t n_subbufs = buf->chan->n_subbufs;
 904
 905         read_subbuf = read_pos / subbuf_size;
 906         padding = buf->padding[read_subbuf];
 907         padding_start = (read_subbuf + 1) * subbuf_size - padding;
 908         padding_end = (read_subbuf + 1) * subbuf_size;
 909         if (read_pos >= padding_start && read_pos < padding_end) {
 910                 read_subbuf = (read_subbuf + 1) % n_subbufs;
 911                 read_pos = read_subbuf * subbuf_size;
 912         }
 913
 914         return read_pos;
 915 }
 916
 917 /**
 918  *      relay_file_read_end_pos - return the new read position
 919  *      @read_pos: file read position
 920  *      @buf: relay channel buffer
 921  *      @count: number of bytes to be read
 922  */
 923 static size_t relay_file_read_end_pos(struct rchan_buf *buf,
 924                                       size_t read_pos,
 925                                       size_t count)
 926 {
 927         size_t read_subbuf, padding, end_pos;
 928         size_t subbuf_size = buf->chan->subbuf_size;
 929         size_t n_subbufs = buf->chan->n_subbufs;
 930
 931         read_subbuf = read_pos / subbuf_size;
 932         padding = buf->padding[read_subbuf];
 933         if (read_pos % subbuf_size + count + padding == subbuf_size)
 934                 end_pos = (read_subbuf + 1) * subbuf_size;
 935         else
 936                 end_pos = read_pos + count;
 937         if (end_pos >= subbuf_size * n_subbufs)
 938                 end_pos = 0;
 939
 940         return end_pos;
 941 }
 942
 943 /*
 944  *      subbuf_read_actor - read up to one subbuf's worth of data
 945  */
 946 static int subbuf_read_actor(size_t read_start,
 947                              struct rchan_buf *buf,
 948                              size_t avail,
 949                              read_descriptor_t *desc,
 950                              read_actor_t actor)
 951 {
 952         void *from;
 953         int ret = 0;
 954
 955         from = buf->start + read_start;
 956         ret = avail;
 957         if (copy_to_user(desc->arg.buf, from, avail)) {
 958                 desc->error = -EFAULT;
 959                 ret = 0;
 960         }
 961         desc->arg.data += ret;
 962         desc->written += ret;
 963         desc->count -= ret;
 964
 965         return ret;
 966 }
 967
 968 /*
 969  *      subbuf_send_actor - send up to one subbuf's worth of data
 970  */
 971 static int subbuf_send_actor(size_t read_start,
 972                              struct rchan_buf *buf,
 973                              size_t avail,
 974                              read_descriptor_t *desc,
 975                              read_actor_t actor)
 976 {
 977         unsigned long pidx, poff;
 978         unsigned int subbuf_pages;
 979         int ret = 0;
 980
 981         subbuf_pages = buf->chan->alloc_size >> PAGE_SHIFT;
 982         pidx = (read_start / PAGE_SIZE) % subbuf_pages;
 983         poff = read_start & ~PAGE_MASK;
 984         while (avail) {
 985                 struct page *p = buf->page_array[pidx];
 986                 unsigned int len;
 987
 988                 len = PAGE_SIZE - poff;
 989                 if (len > avail)
 990                         len = avail;
 991
 992                 len = actor(desc, p, poff, len);
 993                 if (desc->error)
 994                         break;
 995
 996                 avail -= len;
 997                 ret += len;
 998                 poff = 0;
 999                 pidx = (pidx + 1) % subbuf_pages;
1000         }
1001
1002         return ret;
1003 }
1004
1005 typedef int (*subbuf_actor_t) (size_t read_start,
1006                                struct rchan_buf *buf,
1007                                size_t avail,
1008                                read_descriptor_t *desc,
1009                                read_actor_t actor);
1010
1011 /*
1012  *      relay_file_read_subbufs - read count bytes, bridging subbuf boundaries
1013  */
1014 static ssize_t relay_file_read_subbufs(struct file *filp, loff_t *ppos,
1015                                         subbuf_actor_t subbuf_actor,
1016                                         read_actor_t actor,
1017                                         read_descriptor_t *desc)
1018 {
1019         struct rchan_buf *buf = filp->private_data;
1020         size_t read_start, avail;
1021         int ret;
1022
1023         if (!desc->count)
1024                 return 0;
1025
1026         mutex_lock(&filp->f_path.dentry->d_inode->i_mutex);
1027         do {
1028                 if (!relay_file_read_avail(buf, *ppos))
1029                         break;
1030
1031                 read_start = relay_file_read_start_pos(*ppos, buf);
1032                 avail = relay_file_read_subbuf_avail(read_start, buf);
1033                 if (!avail)
1034                         break;
1035
1036                 avail = min(desc->count, avail);
1037                 ret = subbuf_actor(read_start, buf, avail, desc, actor);
1038                 if (desc->error < 0)
1039                         break;
1040
1041                 if (ret) {
1042                         relay_file_read_consume(buf, read_start, ret);
1043                         *ppos = relay_file_read_end_pos(buf, read_start, ret);
1044                 }
1045         } while (desc->count && ret);
1046         mutex_unlock(&filp->f_path.dentry->d_inode->i_mutex);
1047
1048         return desc->written;
1049 }
1050
1051 static ssize_t relay_file_read(struct file *filp,
1052                                char __user *buffer,
1053                                size_t count,
1054                                loff_t *ppos)
1055 {
1056         read_descriptor_t desc;
1057         desc.written = 0;
1058         desc.count = count;
1059         desc.arg.buf = buffer;
1060         desc.error = 0;
1061         return relay_file_read_subbufs(filp, ppos, subbuf_read_actor,
1062                                        NULL, &desc);
1063 }
1064
1065 static ssize_t relay_file_sendfile(struct file *filp,
1066                                    loff_t *ppos,
1067                                    size_t count,
1068                                    read_actor_t actor,
1069                                    void *target)
1070 {
1071         read_descriptor_t desc;
1072         desc.written = 0;
1073         desc.count = count;
1074         desc.arg.data = target;
1075         desc.error = 0;
1076         return relay_file_read_subbufs(filp, ppos, subbuf_send_actor,
1077                                        actor, &desc);
1078 }
1079
1080 const struct file_operations relay_file_operations = {
1081         .open           = relay_file_open,
1082         .poll           = relay_file_poll,
1083         .mmap           = relay_file_mmap,
1084         .read           = relay_file_read,
1085         .llseek         = no_llseek,
1086         .release        = relay_file_release,
1087         .sendfile       = relay_file_sendfile,
1088 };
1089 EXPORT_SYMBOL_GPL(relay_file_operations);
1090
1091 static __init int relay_init(void)
1092 {
1093
1094         hotcpu_notifier(relay_hotcpu_callback, 0);
1095         return 0;
1096 }
1097
1098 module_init(relay_init);