kernel/relay.c

   1 /*
   2  * Public API and common code for kernel->userspace relay file support.
   3  *
   4  * See Documentation/filesystems/relay.rst for an overview.
   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/export.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 #include <linux/splice.h>
  25
  26 /* list of open channels, for cpu hotplug */
  27 static DEFINE_MUTEX(relay_channels_mutex);
  28 static LIST_HEAD(relay_channels);
  29
  30 /*
  31  * fault() vm_op implementation for relay file mapping.
  32  */
  33 static vm_fault_t relay_buf_fault(struct vm_fault *vmf)
  34 {
  35         struct page *page;
  36         struct rchan_buf *buf = vmf->vma->vm_private_data;
  37         pgoff_t pgoff = vmf->pgoff;
  38
  39         if (!buf)
  40                 return VM_FAULT_OOM;
  41
  42         page = vmalloc_to_page(buf->start + (pgoff << PAGE_SHIFT));
  43         if (!page)
  44                 return VM_FAULT_SIGBUS;
  45         get_page(page);
  46         vmf->page = page;
  47
  48         return 0;
  49 }
  50
  51 /*
  52  * vm_ops for relay file mappings.
  53  */
  54 static const struct vm_operations_struct relay_file_mmap_ops = {
  55         .fault = relay_buf_fault,
  56 };
  57
  58 /*
  59  * allocate an array of pointers of struct page
  60  */
  61 static struct page **relay_alloc_page_array(unsigned int n_pages)
  62 {
  63         return kvcalloc(n_pages, sizeof(struct page *), GFP_KERNEL);
  64 }
  65
  66 /*
  67  * free an array of pointers of struct page
  68  */
  69 static void relay_free_page_array(struct page **array)
  70 {
  71         kvfree(array);
  72 }
  73
  74 /**
  75  *      relay_mmap_buf: - mmap channel buffer to process address space
  76  *      @buf: relay channel buffer
  77  *      @vma: vm_area_struct describing memory to be mapped
  78  *
  79  *      Returns 0 if ok, negative on error
  80  *
  81  *      Caller should already have grabbed mmap_lock.
  82  */
  83 static int relay_mmap_buf(struct rchan_buf *buf, struct vm_area_struct *vma)
  84 {
  85         unsigned long length = vma->vm_end - vma->vm_start;
  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         vm_flags_set(vma, VM_DONTEXPAND);
  95         vma->vm_private_data = buf;
  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 = relay_alloc_page_array(n_pages);
 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                 set_page_private(buf->page_array[i], (unsigned long)buf);
 125         }
 126         mem = vmap(buf->page_array, n_pages, VM_MAP, PAGE_KERNEL);
 127         if (!mem)
 128                 goto depopulate;
 129
 130         memset(mem, 0, *size);
 131         buf->page_count = n_pages;
 132         return mem;
 133
 134 depopulate:
 135         for (j = 0; j < i; j++)
 136                 __free_page(buf->page_array[j]);
 137         relay_free_page_array(buf->page_array);
 138         return NULL;
 139 }
 140
 141 /**
 142  *      relay_create_buf - allocate and initialize a channel buffer
 143  *      @chan: the relay channel
 144  *
 145  *      Returns channel buffer if successful, %NULL otherwise.
 146  */
 147 static struct rchan_buf *relay_create_buf(struct rchan *chan)
 148 {
 149         struct rchan_buf *buf;
 150
 151         if (chan->n_subbufs > KMALLOC_MAX_SIZE / sizeof(size_t))
 152                 return NULL;
 153
 154         buf = kzalloc(sizeof(struct rchan_buf), GFP_KERNEL);
 155         if (!buf)
 156                 return NULL;
 157         buf->padding = kmalloc_array(chan->n_subbufs, sizeof(size_t),
 158                                      GFP_KERNEL);
 159         if (!buf->padding)
 160                 goto free_buf;
 161
 162         buf->start = relay_alloc_buf(buf, &chan->alloc_size);
 163         if (!buf->start)
 164                 goto free_buf;
 165
 166         buf->chan = chan;
 167         kref_get(&buf->chan->kref);
 168         return buf;
 169
 170 free_buf:
 171         kfree(buf->padding);
 172         kfree(buf);
 173         return NULL;
 174 }
 175
 176 /**
 177  *      relay_destroy_channel - free the channel struct
 178  *      @kref: target kernel reference that contains the relay channel
 179  *
 180  *      Should only be called from kref_put().
 181  */
 182 static void relay_destroy_channel(struct kref *kref)
 183 {
 184         struct rchan *chan = container_of(kref, struct rchan, kref);
 185         free_percpu(chan->buf);
 186         kfree(chan);
 187 }
 188
 189 /**
 190  *      relay_destroy_buf - destroy an rchan_buf struct and associated buffer
 191  *      @buf: the buffer struct
 192  */
 193 static void relay_destroy_buf(struct rchan_buf *buf)
 194 {
 195         struct rchan *chan = buf->chan;
 196         unsigned int i;
 197
 198         if (likely(buf->start)) {
 199                 vunmap(buf->start);
 200                 for (i = 0; i < buf->page_count; i++)
 201                         __free_page(buf->page_array[i]);
 202                 relay_free_page_array(buf->page_array);
 203         }
 204         *per_cpu_ptr(chan->buf, buf->cpu) = NULL;
 205         kfree(buf->padding);
 206         kfree(buf);
 207         kref_put(&chan->kref, relay_destroy_channel);
 208 }
 209
 210 /**
 211  *      relay_remove_buf - remove a channel buffer
 212  *      @kref: target kernel reference that contains the relay buffer
 213  *
 214  *      Removes the file from the filesystem, which also frees the
 215  *      rchan_buf_struct and the channel buffer.  Should only be called from
 216  *      kref_put().
 217  */
 218 static void relay_remove_buf(struct kref *kref)
 219 {
 220         struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
 221         relay_destroy_buf(buf);
 222 }
 223
 224 /**
 225  *      relay_buf_empty - boolean, is the channel buffer empty?
 226  *      @buf: channel buffer
 227  *
 228  *      Returns 1 if the buffer is empty, 0 otherwise.
 229  */
 230 static int relay_buf_empty(struct rchan_buf *buf)
 231 {
 232         return (buf->subbufs_produced - buf->subbufs_consumed) ? 0 : 1;
 233 }
 234
 235 /**
 236  *      relay_buf_full - boolean, is the channel buffer full?
 237  *      @buf: channel buffer
 238  *
 239  *      Returns 1 if the buffer is full, 0 otherwise.
 240  */
 241 int relay_buf_full(struct rchan_buf *buf)
 242 {
 243         size_t ready = buf->subbufs_produced - buf->subbufs_consumed;
 244         return (ready >= buf->chan->n_subbufs) ? 1 : 0;
 245 }
 246 EXPORT_SYMBOL_GPL(relay_buf_full);
 247
 248 /*
 249  * High-level relay kernel API and associated functions.
 250  */
 251
 252 static int relay_subbuf_start(struct rchan_buf *buf, void *subbuf,
 253                               void *prev_subbuf, size_t prev_padding)
 254 {
 255         if (!buf->chan->cb->subbuf_start)
 256                 return !relay_buf_full(buf);
 257
 258         return buf->chan->cb->subbuf_start(buf, subbuf,
 259                                            prev_subbuf, prev_padding);
 260 }
 261
 262 /**
 263  *      wakeup_readers - wake up readers waiting on a channel
 264  *      @work: contains the channel buffer
 265  *
 266  *      This is the function used to defer reader waking
 267  */
 268 static void wakeup_readers(struct irq_work *work)
 269 {
 270         struct rchan_buf *buf;
 271
 272         buf = container_of(work, struct rchan_buf, wakeup_work);
 273         wake_up_interruptible(&buf->read_wait);
 274 }
 275
 276 /**
 277  *      __relay_reset - reset a channel buffer
 278  *      @buf: the channel buffer
 279  *      @init: 1 if this is a first-time initialization
 280  *
 281  *      See relay_reset() for description of effect.
 282  */
 283 static void __relay_reset(struct rchan_buf *buf, unsigned int init)
 284 {
 285         size_t i;
 286
 287         if (init) {
 288                 init_waitqueue_head(&buf->read_wait);
 289                 kref_init(&buf->kref);
 290                 init_irq_work(&buf->wakeup_work, wakeup_readers);
 291         } else {
 292                 irq_work_sync(&buf->wakeup_work);
 293         }
 294
 295         buf->subbufs_produced = 0;
 296         buf->subbufs_consumed = 0;
 297         buf->bytes_consumed = 0;
 298         buf->finalized = 0;
 299         buf->data = buf->start;
 300         buf->offset = 0;
 301
 302         for (i = 0; i < buf->chan->n_subbufs; i++)
 303                 buf->padding[i] = 0;
 304
 305         relay_subbuf_start(buf, buf->data, NULL, 0);
 306 }
 307
 308 /**
 309  *      relay_reset - reset the channel
 310  *      @chan: the channel
 311  *
 312  *      This has the effect of erasing all data from all channel buffers
 313  *      and restarting the channel in its initial state.  The buffers
 314  *      are not freed, so any mappings are still in effect.
 315  *
 316  *      NOTE. Care should be taken that the channel isn't actually
 317  *      being used by anything when this call is made.
 318  */
 319 void relay_reset(struct rchan *chan)
 320 {
 321         struct rchan_buf *buf;
 322         unsigned int i;
 323
 324         if (!chan)
 325                 return;
 326
 327         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 328                 __relay_reset(buf, 0);
 329                 return;
 330         }
 331
 332         mutex_lock(&relay_channels_mutex);
 333         for_each_possible_cpu(i)
 334                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 335                         __relay_reset(buf, 0);
 336         mutex_unlock(&relay_channels_mutex);
 337 }
 338 EXPORT_SYMBOL_GPL(relay_reset);
 339
 340 static inline void relay_set_buf_dentry(struct rchan_buf *buf,
 341                                         struct dentry *dentry)
 342 {
 343         buf->dentry = dentry;
 344         d_inode(buf->dentry)->i_size = buf->early_bytes;
 345 }
 346
 347 static struct dentry *relay_create_buf_file(struct rchan *chan,
 348                                             struct rchan_buf *buf,
 349                                             unsigned int cpu)
 350 {
 351         struct dentry *dentry;
 352         char *tmpname;
 353
 354         tmpname = kasprintf(GFP_KERNEL, "%s%d", chan->base_filename, cpu);
 355         if (!tmpname)
 356                 return NULL;
 357
 358         /* Create file in fs */
 359         dentry = chan->cb->create_buf_file(tmpname, chan->parent,
 360                                            S_IRUSR, buf,
 361                                            &chan->is_global);
 362         if (IS_ERR(dentry))
 363                 dentry = NULL;
 364
 365         kfree(tmpname);
 366
 367         return dentry;
 368 }
 369
 370 /*
 371  *      relay_open_buf - create a new relay channel buffer
 372  *
 373  *      used by relay_open() and CPU hotplug.
 374  */
 375 static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
 376 {
 377         struct rchan_buf *buf;
 378         struct dentry *dentry;
 379
 380         if (chan->is_global)
 381                 return *per_cpu_ptr(chan->buf, 0);
 382
 383         buf = relay_create_buf(chan);
 384         if (!buf)
 385                 return NULL;
 386
 387         if (chan->has_base_filename) {
 388                 dentry = relay_create_buf_file(chan, buf, cpu);
 389                 if (!dentry)
 390                         goto free_buf;
 391                 relay_set_buf_dentry(buf, dentry);
 392         } else {
 393                 /* Only retrieve global info, nothing more, nothing less */
 394                 dentry = chan->cb->create_buf_file(NULL, NULL,
 395                                                    S_IRUSR, buf,
 396                                                    &chan->is_global);
 397                 if (IS_ERR_OR_NULL(dentry))
 398                         goto free_buf;
 399         }
 400
 401         buf->cpu = cpu;
 402         __relay_reset(buf, 1);
 403
 404         if(chan->is_global) {
 405                 *per_cpu_ptr(chan->buf, 0) = buf;
 406                 buf->cpu = 0;
 407         }
 408
 409         return buf;
 410
 411 free_buf:
 412         relay_destroy_buf(buf);
 413         return NULL;
 414 }
 415
 416 /**
 417  *      relay_close_buf - close a channel buffer
 418  *      @buf: channel buffer
 419  *
 420  *      Marks the buffer finalized and restores the default callbacks.
 421  *      The channel buffer and channel buffer data structure are then freed
 422  *      automatically when the last reference is given up.
 423  */
 424 static void relay_close_buf(struct rchan_buf *buf)
 425 {
 426         buf->finalized = 1;
 427         irq_work_sync(&buf->wakeup_work);
 428         buf->chan->cb->remove_buf_file(buf->dentry);
 429         kref_put(&buf->kref, relay_remove_buf);
 430 }
 431
 432 int relay_prepare_cpu(unsigned int cpu)
 433 {
 434         struct rchan *chan;
 435         struct rchan_buf *buf;
 436
 437         mutex_lock(&relay_channels_mutex);
 438         list_for_each_entry(chan, &relay_channels, list) {
 439                 if (*per_cpu_ptr(chan->buf, cpu))
 440                         continue;
 441                 buf = relay_open_buf(chan, cpu);
 442                 if (!buf) {
 443                         pr_err("relay: cpu %d buffer creation failed\n", cpu);
 444                         mutex_unlock(&relay_channels_mutex);
 445                         return -ENOMEM;
 446                 }
 447                 *per_cpu_ptr(chan->buf, cpu) = buf;
 448         }
 449         mutex_unlock(&relay_channels_mutex);
 450         return 0;
 451 }
 452
 453 /**
 454  *      relay_open - create a new relay channel
 455  *      @base_filename: base name of files to create
 456  *      @parent: dentry of parent directory, %NULL for root directory or buffer
 457  *      @subbuf_size: size of sub-buffers
 458  *      @n_subbufs: number of sub-buffers
 459  *      @cb: client callback functions
 460  *      @private_data: user-defined data
 461  *
 462  *      Returns channel pointer if successful, %NULL otherwise.
 463  *
 464  *      Creates a channel buffer for each cpu using the sizes and
 465  *      attributes specified.  The created channel buffer files
 466  *      will be named base_filename0...base_filenameN-1.  File
 467  *      permissions will be %S_IRUSR.
 468  */
 469 struct rchan *relay_open(const char *base_filename,
 470                          struct dentry *parent,
 471                          size_t subbuf_size,
 472                          size_t n_subbufs,
 473                          const struct rchan_callbacks *cb,
 474                          void *private_data)
 475 {
 476         unsigned int i;
 477         struct rchan *chan;
 478         struct rchan_buf *buf;
 479
 480         if (!(subbuf_size && n_subbufs))
 481                 return NULL;
 482         if (subbuf_size > UINT_MAX / n_subbufs)
 483                 return NULL;
 484         if (!cb || !cb->create_buf_file || !cb->remove_buf_file)
 485                 return NULL;
 486
 487         chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
 488         if (!chan)
 489                 return NULL;
 490
 491         chan->buf = alloc_percpu(struct rchan_buf *);
 492         if (!chan->buf) {
 493                 kfree(chan);
 494                 return NULL;
 495         }
 496
 497         chan->version = RELAYFS_CHANNEL_VERSION;
 498         chan->n_subbufs = n_subbufs;
 499         chan->subbuf_size = subbuf_size;
 500         chan->alloc_size = PAGE_ALIGN(subbuf_size * n_subbufs);
 501         chan->parent = parent;
 502         chan->private_data = private_data;
 503         if (base_filename) {
 504                 chan->has_base_filename = 1;
 505                 strscpy(chan->base_filename, base_filename, NAME_MAX);
 506         }
 507         chan->cb = cb;
 508         kref_init(&chan->kref);
 509
 510         mutex_lock(&relay_channels_mutex);
 511         for_each_online_cpu(i) {
 512                 buf = relay_open_buf(chan, i);
 513                 if (!buf)
 514                         goto free_bufs;
 515                 *per_cpu_ptr(chan->buf, i) = buf;
 516         }
 517         list_add(&chan->list, &relay_channels);
 518         mutex_unlock(&relay_channels_mutex);
 519
 520         return chan;
 521
 522 free_bufs:
 523         for_each_possible_cpu(i) {
 524                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 525                         relay_close_buf(buf);
 526         }
 527
 528         kref_put(&chan->kref, relay_destroy_channel);
 529         mutex_unlock(&relay_channels_mutex);
 530         return NULL;
 531 }
 532 EXPORT_SYMBOL_GPL(relay_open);
 533
 534 struct rchan_percpu_buf_dispatcher {
 535         struct rchan_buf *buf;
 536         struct dentry *dentry;
 537 };
 538
 539 /**
 540  *      relay_switch_subbuf - switch to a new sub-buffer
 541  *      @buf: channel buffer
 542  *      @length: size of current event
 543  *
 544  *      Returns either the length passed in or 0 if full.
 545  *
 546  *      Performs sub-buffer-switch tasks such as invoking callbacks,
 547  *      updating padding counts, waking up readers, etc.
 548  */
 549 size_t relay_switch_subbuf(struct rchan_buf *buf, size_t length)
 550 {
 551         void *old, *new;
 552         size_t old_subbuf, new_subbuf;
 553
 554         if (unlikely(length > buf->chan->subbuf_size))
 555                 goto toobig;
 556
 557         if (buf->offset != buf->chan->subbuf_size + 1) {
 558                 buf->prev_padding = buf->chan->subbuf_size - buf->offset;
 559                 old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 560                 buf->padding[old_subbuf] = buf->prev_padding;
 561                 buf->subbufs_produced++;
 562                 if (buf->dentry)
 563                         d_inode(buf->dentry)->i_size +=
 564                                 buf->chan->subbuf_size -
 565                                 buf->padding[old_subbuf];
 566                 else
 567                         buf->early_bytes += buf->chan->subbuf_size -
 568                                             buf->padding[old_subbuf];
 569                 smp_mb();
 570                 if (waitqueue_active(&buf->read_wait)) {
 571                         /*
 572                          * Calling wake_up_interruptible() from here
 573                          * will deadlock if we happen to be logging
 574                          * from the scheduler (trying to re-grab
 575                          * rq->lock), so defer it.
 576                          */
 577                         irq_work_queue(&buf->wakeup_work);
 578                 }
 579         }
 580
 581         old = buf->data;
 582         new_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 583         new = buf->start + new_subbuf * buf->chan->subbuf_size;
 584         buf->offset = 0;
 585         if (!relay_subbuf_start(buf, new, old, buf->prev_padding)) {
 586                 buf->offset = buf->chan->subbuf_size + 1;
 587                 return 0;
 588         }
 589         buf->data = new;
 590         buf->padding[new_subbuf] = 0;
 591
 592         if (unlikely(length + buf->offset > buf->chan->subbuf_size))
 593                 goto toobig;
 594
 595         return length;
 596
 597 toobig:
 598         buf->chan->last_toobig = length;
 599         return 0;
 600 }
 601 EXPORT_SYMBOL_GPL(relay_switch_subbuf);
 602
 603 /**
 604  *      relay_subbufs_consumed - update the buffer's sub-buffers-consumed count
 605  *      @chan: the channel
 606  *      @cpu: the cpu associated with the channel buffer to update
 607  *      @subbufs_consumed: number of sub-buffers to add to current buf's count
 608  *
 609  *      Adds to the channel buffer's consumed sub-buffer count.
 610  *      subbufs_consumed should be the number of sub-buffers newly consumed,
 611  *      not the total consumed.
 612  *
 613  *      NOTE. Kernel clients don't need to call this function if the channel
 614  *      mode is 'overwrite'.
 615  */
 616 void relay_subbufs_consumed(struct rchan *chan,
 617                             unsigned int cpu,
 618                             size_t subbufs_consumed)
 619 {
 620         struct rchan_buf *buf;
 621
 622         if (!chan || cpu >= NR_CPUS)
 623                 return;
 624
 625         buf = *per_cpu_ptr(chan->buf, cpu);
 626         if (!buf || subbufs_consumed > chan->n_subbufs)
 627                 return;
 628
 629         if (subbufs_consumed > buf->subbufs_produced - buf->subbufs_consumed)
 630                 buf->subbufs_consumed = buf->subbufs_produced;
 631         else
 632                 buf->subbufs_consumed += subbufs_consumed;
 633 }
 634 EXPORT_SYMBOL_GPL(relay_subbufs_consumed);
 635
 636 /**
 637  *      relay_close - close the channel
 638  *      @chan: the channel
 639  *
 640  *      Closes all channel buffers and frees the channel.
 641  */
 642 void relay_close(struct rchan *chan)
 643 {
 644         struct rchan_buf *buf;
 645         unsigned int i;
 646
 647         if (!chan)
 648                 return;
 649
 650         mutex_lock(&relay_channels_mutex);
 651         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0)))
 652                 relay_close_buf(buf);
 653         else
 654                 for_each_possible_cpu(i)
 655                         if ((buf = *per_cpu_ptr(chan->buf, i)))
 656                                 relay_close_buf(buf);
 657
 658         if (chan->last_toobig)
 659                 printk(KERN_WARNING "relay: one or more items not logged "
 660                        "[item size (%zd) > sub-buffer size (%zd)]\n",
 661                        chan->last_toobig, chan->subbuf_size);
 662
 663         list_del(&chan->list);
 664         kref_put(&chan->kref, relay_destroy_channel);
 665         mutex_unlock(&relay_channels_mutex);
 666 }
 667 EXPORT_SYMBOL_GPL(relay_close);
 668
 669 /**
 670  *      relay_flush - close the channel
 671  *      @chan: the channel
 672  *
 673  *      Flushes all channel buffers, i.e. forces buffer switch.
 674  */
 675 void relay_flush(struct rchan *chan)
 676 {
 677         struct rchan_buf *buf;
 678         unsigned int i;
 679
 680         if (!chan)
 681                 return;
 682
 683         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 684                 relay_switch_subbuf(buf, 0);
 685                 return;
 686         }
 687
 688         mutex_lock(&relay_channels_mutex);
 689         for_each_possible_cpu(i)
 690                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 691                         relay_switch_subbuf(buf, 0);
 692         mutex_unlock(&relay_channels_mutex);
 693 }
 694 EXPORT_SYMBOL_GPL(relay_flush);
 695
 696 /**
 697  *      relay_file_open - open file op for relay files
 698  *      @inode: the inode
 699  *      @filp: the file
 700  *
 701  *      Increments the channel buffer refcount.
 702  */
 703 static int relay_file_open(struct inode *inode, struct file *filp)
 704 {
 705         struct rchan_buf *buf = inode->i_private;
 706         kref_get(&buf->kref);
 707         filp->private_data = buf;
 708
 709         return nonseekable_open(inode, filp);
 710 }
 711
 712 /**
 713  *      relay_file_mmap - mmap file op for relay files
 714  *      @filp: the file
 715  *      @vma: the vma describing what to map
 716  *
 717  *      Calls upon relay_mmap_buf() to map the file into user space.
 718  */
 719 static int relay_file_mmap(struct file *filp, struct vm_area_struct *vma)
 720 {
 721         struct rchan_buf *buf = filp->private_data;
 722         return relay_mmap_buf(buf, vma);
 723 }
 724
 725 /**
 726  *      relay_file_poll - poll file op for relay files
 727  *      @filp: the file
 728  *      @wait: poll table
 729  *
 730  *      Poll implemention.
 731  */
 732 static __poll_t relay_file_poll(struct file *filp, poll_table *wait)
 733 {
 734         __poll_t mask = 0;
 735         struct rchan_buf *buf = filp->private_data;
 736
 737         if (buf->finalized)
 738                 return EPOLLERR;
 739
 740         if (filp->f_mode & FMODE_READ) {
 741                 poll_wait(filp, &buf->read_wait, wait);
 742                 if (!relay_buf_empty(buf))
 743                         mask |= EPOLLIN | EPOLLRDNORM;
 744         }
 745
 746         return mask;
 747 }
 748
 749 /**
 750  *      relay_file_release - release file op for relay files
 751  *      @inode: the inode
 752  *      @filp: the file
 753  *
 754  *      Decrements the channel refcount, as the filesystem is
 755  *      no longer using it.
 756  */
 757 static int relay_file_release(struct inode *inode, struct file *filp)
 758 {
 759         struct rchan_buf *buf = filp->private_data;
 760         kref_put(&buf->kref, relay_remove_buf);
 761
 762         return 0;
 763 }
 764
 765 /*
 766  *      relay_file_read_consume - update the consumed count for the buffer
 767  */
 768 static void relay_file_read_consume(struct rchan_buf *buf,
 769                                     size_t read_pos,
 770                                     size_t bytes_consumed)
 771 {
 772         size_t subbuf_size = buf->chan->subbuf_size;
 773         size_t n_subbufs = buf->chan->n_subbufs;
 774         size_t read_subbuf;
 775
 776         if (buf->subbufs_produced == buf->subbufs_consumed &&
 777             buf->offset == buf->bytes_consumed)
 778                 return;
 779
 780         if (buf->bytes_consumed + bytes_consumed > subbuf_size) {
 781                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 782                 buf->bytes_consumed = 0;
 783         }
 784
 785         buf->bytes_consumed += bytes_consumed;
 786         if (!read_pos)
 787                 read_subbuf = buf->subbufs_consumed % n_subbufs;
 788         else
 789                 read_subbuf = read_pos / buf->chan->subbuf_size;
 790         if (buf->bytes_consumed + buf->padding[read_subbuf] == subbuf_size) {
 791                 if ((read_subbuf == buf->subbufs_produced % n_subbufs) &&
 792                     (buf->offset == subbuf_size))
 793                         return;
 794                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 795                 buf->bytes_consumed = 0;
 796         }
 797 }
 798
 799 /*
 800  *      relay_file_read_avail - boolean, are there unconsumed bytes available?
 801  */
 802 static int relay_file_read_avail(struct rchan_buf *buf)
 803 {
 804         size_t subbuf_size = buf->chan->subbuf_size;
 805         size_t n_subbufs = buf->chan->n_subbufs;
 806         size_t produced = buf->subbufs_produced;
 807         size_t consumed;
 808
 809         relay_file_read_consume(buf, 0, 0);
 810
 811         consumed = buf->subbufs_consumed;
 812
 813         if (unlikely(buf->offset > subbuf_size)) {
 814                 if (produced == consumed)
 815                         return 0;
 816                 return 1;
 817         }
 818
 819         if (unlikely(produced - consumed >= n_subbufs)) {
 820                 consumed = produced - n_subbufs + 1;
 821                 buf->subbufs_consumed = consumed;
 822                 buf->bytes_consumed = 0;
 823         }
 824
 825         produced = (produced % n_subbufs) * subbuf_size + buf->offset;
 826         consumed = (consumed % n_subbufs) * subbuf_size + buf->bytes_consumed;
 827
 828         if (consumed > produced)
 829                 produced += n_subbufs * subbuf_size;
 830
 831         if (consumed == produced) {
 832                 if (buf->offset == subbuf_size &&
 833                     buf->subbufs_produced > buf->subbufs_consumed)
 834                         return 1;
 835                 return 0;
 836         }
 837
 838         return 1;
 839 }
 840
 841 /**
 842  *      relay_file_read_subbuf_avail - return bytes available in sub-buffer
 843  *      @read_pos: file read position
 844  *      @buf: relay channel buffer
 845  */
 846 static size_t relay_file_read_subbuf_avail(size_t read_pos,
 847                                            struct rchan_buf *buf)
 848 {
 849         size_t padding, avail = 0;
 850         size_t read_subbuf, read_offset, write_subbuf, write_offset;
 851         size_t subbuf_size = buf->chan->subbuf_size;
 852
 853         write_subbuf = (buf->data - buf->start) / subbuf_size;
 854         write_offset = buf->offset > subbuf_size ? subbuf_size : buf->offset;
 855         read_subbuf = read_pos / subbuf_size;
 856         read_offset = read_pos % subbuf_size;
 857         padding = buf->padding[read_subbuf];
 858
 859         if (read_subbuf == write_subbuf) {
 860                 if (read_offset + padding < write_offset)
 861                         avail = write_offset - (read_offset + padding);
 862         } else
 863                 avail = (subbuf_size - padding) - read_offset;
 864
 865         return avail;
 866 }
 867
 868 /**
 869  *      relay_file_read_start_pos - find the first available byte to read
 870  *      @buf: relay channel buffer
 871  *
 872  *      If the read_pos is in the middle of padding, return the
 873  *      position of the first actually available byte, otherwise
 874  *      return the original value.
 875  */
 876 static size_t relay_file_read_start_pos(struct rchan_buf *buf)
 877 {
 878         size_t read_subbuf, padding, padding_start, padding_end;
 879         size_t subbuf_size = buf->chan->subbuf_size;
 880         size_t n_subbufs = buf->chan->n_subbufs;
 881         size_t consumed = buf->subbufs_consumed % n_subbufs;
 882         size_t read_pos = (consumed * subbuf_size + buf->bytes_consumed)
 883                         % (n_subbufs * subbuf_size);
 884
 885         read_subbuf = read_pos / subbuf_size;
 886         padding = buf->padding[read_subbuf];
 887         padding_start = (read_subbuf + 1) * subbuf_size - padding;
 888         padding_end = (read_subbuf + 1) * subbuf_size;
 889         if (read_pos >= padding_start && read_pos < padding_end) {
 890                 read_subbuf = (read_subbuf + 1) % n_subbufs;
 891                 read_pos = read_subbuf * subbuf_size;
 892         }
 893
 894         return read_pos;
 895 }
 896
 897 /**
 898  *      relay_file_read_end_pos - return the new read position
 899  *      @read_pos: file read position
 900  *      @buf: relay channel buffer
 901  *      @count: number of bytes to be read
 902  */
 903 static size_t relay_file_read_end_pos(struct rchan_buf *buf,
 904                                       size_t read_pos,
 905                                       size_t count)
 906 {
 907         size_t read_subbuf, padding, end_pos;
 908         size_t subbuf_size = buf->chan->subbuf_size;
 909         size_t n_subbufs = buf->chan->n_subbufs;
 910
 911         read_subbuf = read_pos / subbuf_size;
 912         padding = buf->padding[read_subbuf];
 913         if (read_pos % subbuf_size + count + padding == subbuf_size)
 914                 end_pos = (read_subbuf + 1) * subbuf_size;
 915         else
 916                 end_pos = read_pos + count;
 917         if (end_pos >= subbuf_size * n_subbufs)
 918                 end_pos = 0;
 919
 920         return end_pos;
 921 }
 922
 923 static ssize_t relay_file_read(struct file *filp,
 924                                char __user *buffer,
 925                                size_t count,
 926                                loff_t *ppos)
 927 {
 928         struct rchan_buf *buf = filp->private_data;
 929         size_t read_start, avail;
 930         size_t written = 0;
 931         int ret;
 932
 933         if (!count)
 934                 return 0;
 935
 936         inode_lock(file_inode(filp));
 937         do {
 938                 void *from;
 939
 940                 if (!relay_file_read_avail(buf))
 941                         break;
 942
 943                 read_start = relay_file_read_start_pos(buf);
 944                 avail = relay_file_read_subbuf_avail(read_start, buf);
 945                 if (!avail)
 946                         break;
 947
 948                 avail = min(count, avail);
 949                 from = buf->start + read_start;
 950                 ret = avail;
 951                 if (copy_to_user(buffer, from, avail))
 952                         break;
 953
 954                 buffer += ret;
 955                 written += ret;
 956                 count -= ret;
 957
 958                 relay_file_read_consume(buf, read_start, ret);
 959                 *ppos = relay_file_read_end_pos(buf, read_start, ret);
 960         } while (count);
 961         inode_unlock(file_inode(filp));
 962
 963         return written;
 964 }
 965
 966
 967 const struct file_operations relay_file_operations = {
 968         .open           = relay_file_open,
 969         .poll           = relay_file_poll,
 970         .mmap           = relay_file_mmap,
 971         .read           = relay_file_read,
 972         .release        = relay_file_release,
 973 };
 974 EXPORT_SYMBOL_GPL(relay_file_operations);