kernel/relay.c

   1 /*
   2  * Public API and common code for kernel->userspace relay file support.
   3  *
   4  * See Documentation/filesystems/relay.txt 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  * close() vm_op implementation for relay file mapping.
  32  */
  33 static void relay_file_mmap_close(struct vm_area_struct *vma)
  34 {
  35         struct rchan_buf *buf = vma->vm_private_data;
  36         buf->chan->cb->buf_unmapped(buf, vma->vm_file);
  37 }
  38
  39 /*
  40  * fault() vm_op implementation for relay file mapping.
  41  */
  42 static int relay_buf_fault(struct vm_fault *vmf)
  43 {
  44         struct page *page;
  45         struct rchan_buf *buf = vmf->vma->vm_private_data;
  46         pgoff_t pgoff = vmf->pgoff;
  47
  48         if (!buf)
  49                 return VM_FAULT_OOM;
  50
  51         page = vmalloc_to_page(buf->start + (pgoff << PAGE_SHIFT));
  52         if (!page)
  53                 return VM_FAULT_SIGBUS;
  54         get_page(page);
  55         vmf->page = page;
  56
  57         return 0;
  58 }
  59
  60 /*
  61  * vm_ops for relay file mappings.
  62  */
  63 static const struct vm_operations_struct relay_file_mmap_ops = {
  64         .fault = relay_buf_fault,
  65         .close = relay_file_mmap_close,
  66 };
  67
  68 /*
  69  * allocate an array of pointers of struct page
  70  */
  71 static struct page **relay_alloc_page_array(unsigned int n_pages)
  72 {
  73         const size_t pa_size = n_pages * sizeof(struct page *);
  74         if (pa_size > PAGE_SIZE)
  75                 return vzalloc(pa_size);
  76         return kzalloc(pa_size, GFP_KERNEL);
  77 }
  78
  79 /*
  80  * free an array of pointers of struct page
  81  */
  82 static void relay_free_page_array(struct page **array)
  83 {
  84         kvfree(array);
  85 }
  86
  87 /**
  88  *      relay_mmap_buf: - mmap channel buffer to process address space
  89  *      @buf: relay channel buffer
  90  *      @vma: vm_area_struct describing memory to be mapped
  91  *
  92  *      Returns 0 if ok, negative on error
  93  *
  94  *      Caller should already have grabbed mmap_sem.
  95  */
  96 static int relay_mmap_buf(struct rchan_buf *buf, struct vm_area_struct *vma)
  97 {
  98         unsigned long length = vma->vm_end - vma->vm_start;
  99         struct file *filp = vma->vm_file;
 100
 101         if (!buf)
 102                 return -EBADF;
 103
 104         if (length != (unsigned long)buf->chan->alloc_size)
 105                 return -EINVAL;
 106
 107         vma->vm_ops = &relay_file_mmap_ops;
 108         vma->vm_flags |= VM_DONTEXPAND;
 109         vma->vm_private_data = buf;
 110         buf->chan->cb->buf_mapped(buf, filp);
 111
 112         return 0;
 113 }
 114
 115 /**
 116  *      relay_alloc_buf - allocate a channel buffer
 117  *      @buf: the buffer struct
 118  *      @size: total size of the buffer
 119  *
 120  *      Returns a pointer to the resulting buffer, %NULL if unsuccessful. The
 121  *      passed in size will get page aligned, if it isn't already.
 122  */
 123 static void *relay_alloc_buf(struct rchan_buf *buf, size_t *size)
 124 {
 125         void *mem;
 126         unsigned int i, j, n_pages;
 127
 128         *size = PAGE_ALIGN(*size);
 129         n_pages = *size >> PAGE_SHIFT;
 130
 131         buf->page_array = relay_alloc_page_array(n_pages);
 132         if (!buf->page_array)
 133                 return NULL;
 134
 135         for (i = 0; i < n_pages; i++) {
 136                 buf->page_array[i] = alloc_page(GFP_KERNEL);
 137                 if (unlikely(!buf->page_array[i]))
 138                         goto depopulate;
 139                 set_page_private(buf->page_array[i], (unsigned long)buf);
 140         }
 141         mem = vmap(buf->page_array, n_pages, VM_MAP, PAGE_KERNEL);
 142         if (!mem)
 143                 goto depopulate;
 144
 145         memset(mem, 0, *size);
 146         buf->page_count = n_pages;
 147         return mem;
 148
 149 depopulate:
 150         for (j = 0; j < i; j++)
 151                 __free_page(buf->page_array[j]);
 152         relay_free_page_array(buf->page_array);
 153         return NULL;
 154 }
 155
 156 /**
 157  *      relay_create_buf - allocate and initialize a channel buffer
 158  *      @chan: the relay channel
 159  *
 160  *      Returns channel buffer if successful, %NULL otherwise.
 161  */
 162 static struct rchan_buf *relay_create_buf(struct rchan *chan)
 163 {
 164         struct rchan_buf *buf;
 165
 166         if (chan->n_subbufs > KMALLOC_MAX_SIZE / sizeof(size_t *))
 167                 return NULL;
 168
 169         buf = kzalloc(sizeof(struct rchan_buf), GFP_KERNEL);
 170         if (!buf)
 171                 return NULL;
 172         buf->padding = kmalloc(chan->n_subbufs * sizeof(size_t *), GFP_KERNEL);
 173         if (!buf->padding)
 174                 goto free_buf;
 175
 176         buf->start = relay_alloc_buf(buf, &chan->alloc_size);
 177         if (!buf->start)
 178                 goto free_buf;
 179
 180         buf->chan = chan;
 181         kref_get(&buf->chan->kref);
 182         return buf;
 183
 184 free_buf:
 185         kfree(buf->padding);
 186         kfree(buf);
 187         return NULL;
 188 }
 189
 190 /**
 191  *      relay_destroy_channel - free the channel struct
 192  *      @kref: target kernel reference that contains the relay channel
 193  *
 194  *      Should only be called from kref_put().
 195  */
 196 static void relay_destroy_channel(struct kref *kref)
 197 {
 198         struct rchan *chan = container_of(kref, struct rchan, kref);
 199         kfree(chan);
 200 }
 201
 202 /**
 203  *      relay_destroy_buf - destroy an rchan_buf struct and associated buffer
 204  *      @buf: the buffer struct
 205  */
 206 static void relay_destroy_buf(struct rchan_buf *buf)
 207 {
 208         struct rchan *chan = buf->chan;
 209         unsigned int i;
 210
 211         if (likely(buf->start)) {
 212                 vunmap(buf->start);
 213                 for (i = 0; i < buf->page_count; i++)
 214                         __free_page(buf->page_array[i]);
 215                 relay_free_page_array(buf->page_array);
 216         }
 217         *per_cpu_ptr(chan->buf, buf->cpu) = NULL;
 218         kfree(buf->padding);
 219         kfree(buf);
 220         kref_put(&chan->kref, relay_destroy_channel);
 221 }
 222
 223 /**
 224  *      relay_remove_buf - remove a channel buffer
 225  *      @kref: target kernel reference that contains the relay buffer
 226  *
 227  *      Removes the file from the filesystem, which also frees the
 228  *      rchan_buf_struct and the channel buffer.  Should only be called from
 229  *      kref_put().
 230  */
 231 static void relay_remove_buf(struct kref *kref)
 232 {
 233         struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
 234         relay_destroy_buf(buf);
 235 }
 236
 237 /**
 238  *      relay_buf_empty - boolean, is the channel buffer empty?
 239  *      @buf: channel buffer
 240  *
 241  *      Returns 1 if the buffer is empty, 0 otherwise.
 242  */
 243 static int relay_buf_empty(struct rchan_buf *buf)
 244 {
 245         return (buf->subbufs_produced - buf->subbufs_consumed) ? 0 : 1;
 246 }
 247
 248 /**
 249  *      relay_buf_full - boolean, is the channel buffer full?
 250  *      @buf: channel buffer
 251  *
 252  *      Returns 1 if the buffer is full, 0 otherwise.
 253  */
 254 int relay_buf_full(struct rchan_buf *buf)
 255 {
 256         size_t ready = buf->subbufs_produced - buf->subbufs_consumed;
 257         return (ready >= buf->chan->n_subbufs) ? 1 : 0;
 258 }
 259 EXPORT_SYMBOL_GPL(relay_buf_full);
 260
 261 /*
 262  * High-level relay kernel API and associated functions.
 263  */
 264
 265 /*
 266  * rchan_callback implementations defining default channel behavior.  Used
 267  * in place of corresponding NULL values in client callback struct.
 268  */
 269
 270 /*
 271  * subbuf_start() default callback.  Does nothing.
 272  */
 273 static int subbuf_start_default_callback (struct rchan_buf *buf,
 274                                           void *subbuf,
 275                                           void *prev_subbuf,
 276                                           size_t prev_padding)
 277 {
 278         if (relay_buf_full(buf))
 279                 return 0;
 280
 281         return 1;
 282 }
 283
 284 /*
 285  * buf_mapped() default callback.  Does nothing.
 286  */
 287 static void buf_mapped_default_callback(struct rchan_buf *buf,
 288                                         struct file *filp)
 289 {
 290 }
 291
 292 /*
 293  * buf_unmapped() default callback.  Does nothing.
 294  */
 295 static void buf_unmapped_default_callback(struct rchan_buf *buf,
 296                                           struct file *filp)
 297 {
 298 }
 299
 300 /*
 301  * create_buf_file_create() default callback.  Does nothing.
 302  */
 303 static struct dentry *create_buf_file_default_callback(const char *filename,
 304                                                        struct dentry *parent,
 305                                                        umode_t mode,
 306                                                        struct rchan_buf *buf,
 307                                                        int *is_global)
 308 {
 309         return NULL;
 310 }
 311
 312 /*
 313  * remove_buf_file() default callback.  Does nothing.
 314  */
 315 static int remove_buf_file_default_callback(struct dentry *dentry)
 316 {
 317         return -EINVAL;
 318 }
 319
 320 /* relay channel default callbacks */
 321 static struct rchan_callbacks default_channel_callbacks = {
 322         .subbuf_start = subbuf_start_default_callback,
 323         .buf_mapped = buf_mapped_default_callback,
 324         .buf_unmapped = buf_unmapped_default_callback,
 325         .create_buf_file = create_buf_file_default_callback,
 326         .remove_buf_file = remove_buf_file_default_callback,
 327 };
 328
 329 /**
 330  *      wakeup_readers - wake up readers waiting on a channel
 331  *      @work: contains the channel buffer
 332  *
 333  *      This is the function used to defer reader waking
 334  */
 335 static void wakeup_readers(struct irq_work *work)
 336 {
 337         struct rchan_buf *buf;
 338
 339         buf = container_of(work, struct rchan_buf, wakeup_work);
 340         wake_up_interruptible(&buf->read_wait);
 341 }
 342
 343 /**
 344  *      __relay_reset - reset a channel buffer
 345  *      @buf: the channel buffer
 346  *      @init: 1 if this is a first-time initialization
 347  *
 348  *      See relay_reset() for description of effect.
 349  */
 350 static void __relay_reset(struct rchan_buf *buf, unsigned int init)
 351 {
 352         size_t i;
 353
 354         if (init) {
 355                 init_waitqueue_head(&buf->read_wait);
 356                 kref_init(&buf->kref);
 357                 init_irq_work(&buf->wakeup_work, wakeup_readers);
 358         } else {
 359                 irq_work_sync(&buf->wakeup_work);
 360         }
 361
 362         buf->subbufs_produced = 0;
 363         buf->subbufs_consumed = 0;
 364         buf->bytes_consumed = 0;
 365         buf->finalized = 0;
 366         buf->data = buf->start;
 367         buf->offset = 0;
 368
 369         for (i = 0; i < buf->chan->n_subbufs; i++)
 370                 buf->padding[i] = 0;
 371
 372         buf->chan->cb->subbuf_start(buf, buf->data, NULL, 0);
 373 }
 374
 375 /**
 376  *      relay_reset - reset the channel
 377  *      @chan: the channel
 378  *
 379  *      This has the effect of erasing all data from all channel buffers
 380  *      and restarting the channel in its initial state.  The buffers
 381  *      are not freed, so any mappings are still in effect.
 382  *
 383  *      NOTE. Care should be taken that the channel isn't actually
 384  *      being used by anything when this call is made.
 385  */
 386 void relay_reset(struct rchan *chan)
 387 {
 388         struct rchan_buf *buf;
 389         unsigned int i;
 390
 391         if (!chan)
 392                 return;
 393
 394         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 395                 __relay_reset(buf, 0);
 396                 return;
 397         }
 398
 399         mutex_lock(&relay_channels_mutex);
 400         for_each_possible_cpu(i)
 401                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 402                         __relay_reset(buf, 0);
 403         mutex_unlock(&relay_channels_mutex);
 404 }
 405 EXPORT_SYMBOL_GPL(relay_reset);
 406
 407 static inline void relay_set_buf_dentry(struct rchan_buf *buf,
 408                                         struct dentry *dentry)
 409 {
 410         buf->dentry = dentry;
 411         d_inode(buf->dentry)->i_size = buf->early_bytes;
 412 }
 413
 414 static struct dentry *relay_create_buf_file(struct rchan *chan,
 415                                             struct rchan_buf *buf,
 416                                             unsigned int cpu)
 417 {
 418         struct dentry *dentry;
 419         char *tmpname;
 420
 421         tmpname = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 422         if (!tmpname)
 423                 return NULL;
 424         snprintf(tmpname, NAME_MAX, "%s%d", chan->base_filename, cpu);
 425
 426         /* Create file in fs */
 427         dentry = chan->cb->create_buf_file(tmpname, chan->parent,
 428                                            S_IRUSR, buf,
 429                                            &chan->is_global);
 430
 431         kfree(tmpname);
 432
 433         return dentry;
 434 }
 435
 436 /*
 437  *      relay_open_buf - create a new relay channel buffer
 438  *
 439  *      used by relay_open() and CPU hotplug.
 440  */
 441 static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
 442 {
 443         struct rchan_buf *buf = NULL;
 444         struct dentry *dentry;
 445
 446         if (chan->is_global)
 447                 return *per_cpu_ptr(chan->buf, 0);
 448
 449         buf = relay_create_buf(chan);
 450         if (!buf)
 451                 return NULL;
 452
 453         if (chan->has_base_filename) {
 454                 dentry = relay_create_buf_file(chan, buf, cpu);
 455                 if (!dentry)
 456                         goto free_buf;
 457                 relay_set_buf_dentry(buf, dentry);
 458         } else {
 459                 /* Only retrieve global info, nothing more, nothing less */
 460                 dentry = chan->cb->create_buf_file(NULL, NULL,
 461                                                    S_IRUSR, buf,
 462                                                    &chan->is_global);
 463                 if (WARN_ON(dentry))
 464                         goto free_buf;
 465         }
 466
 467         buf->cpu = cpu;
 468         __relay_reset(buf, 1);
 469
 470         if(chan->is_global) {
 471                 *per_cpu_ptr(chan->buf, 0) = buf;
 472                 buf->cpu = 0;
 473         }
 474
 475         return buf;
 476
 477 free_buf:
 478         relay_destroy_buf(buf);
 479         return NULL;
 480 }
 481
 482 /**
 483  *      relay_close_buf - close a channel buffer
 484  *      @buf: channel buffer
 485  *
 486  *      Marks the buffer finalized and restores the default callbacks.
 487  *      The channel buffer and channel buffer data structure are then freed
 488  *      automatically when the last reference is given up.
 489  */
 490 static void relay_close_buf(struct rchan_buf *buf)
 491 {
 492         buf->finalized = 1;
 493         irq_work_sync(&buf->wakeup_work);
 494         buf->chan->cb->remove_buf_file(buf->dentry);
 495         kref_put(&buf->kref, relay_remove_buf);
 496 }
 497
 498 static void setup_callbacks(struct rchan *chan,
 499                                    struct rchan_callbacks *cb)
 500 {
 501         if (!cb) {
 502                 chan->cb = &default_channel_callbacks;
 503                 return;
 504         }
 505
 506         if (!cb->subbuf_start)
 507                 cb->subbuf_start = subbuf_start_default_callback;
 508         if (!cb->buf_mapped)
 509                 cb->buf_mapped = buf_mapped_default_callback;
 510         if (!cb->buf_unmapped)
 511                 cb->buf_unmapped = buf_unmapped_default_callback;
 512         if (!cb->create_buf_file)
 513                 cb->create_buf_file = create_buf_file_default_callback;
 514         if (!cb->remove_buf_file)
 515                 cb->remove_buf_file = remove_buf_file_default_callback;
 516         chan->cb = cb;
 517 }
 518
 519 int relay_prepare_cpu(unsigned int cpu)
 520 {
 521         struct rchan *chan;
 522         struct rchan_buf *buf;
 523
 524         mutex_lock(&relay_channels_mutex);
 525         list_for_each_entry(chan, &relay_channels, list) {
 526                 if ((buf = *per_cpu_ptr(chan->buf, cpu)))
 527                         continue;
 528                 buf = relay_open_buf(chan, cpu);
 529                 if (!buf) {
 530                         pr_err("relay: cpu %d buffer creation failed\n", cpu);
 531                         mutex_unlock(&relay_channels_mutex);
 532                         return -ENOMEM;
 533                 }
 534                 *per_cpu_ptr(chan->buf, cpu) = buf;
 535         }
 536         mutex_unlock(&relay_channels_mutex);
 537         return 0;
 538 }
 539
 540 /**
 541  *      relay_open - create a new relay channel
 542  *      @base_filename: base name of files to create, %NULL for buffering only
 543  *      @parent: dentry of parent directory, %NULL for root directory or buffer
 544  *      @subbuf_size: size of sub-buffers
 545  *      @n_subbufs: number of sub-buffers
 546  *      @cb: client callback functions
 547  *      @private_data: user-defined data
 548  *
 549  *      Returns channel pointer if successful, %NULL otherwise.
 550  *
 551  *      Creates a channel buffer for each cpu using the sizes and
 552  *      attributes specified.  The created channel buffer files
 553  *      will be named base_filename0...base_filenameN-1.  File
 554  *      permissions will be %S_IRUSR.
 555  *
 556  *      If opening a buffer (@parent = NULL) that you later wish to register
 557  *      in a filesystem, call relay_late_setup_files() once the @parent dentry
 558  *      is available.
 559  */
 560 struct rchan *relay_open(const char *base_filename,
 561                          struct dentry *parent,
 562                          size_t subbuf_size,
 563                          size_t n_subbufs,
 564                          struct rchan_callbacks *cb,
 565                          void *private_data)
 566 {
 567         unsigned int i;
 568         struct rchan *chan;
 569         struct rchan_buf *buf;
 570
 571         if (!(subbuf_size && n_subbufs))
 572                 return NULL;
 573         if (subbuf_size > UINT_MAX / n_subbufs)
 574                 return NULL;
 575
 576         chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
 577         if (!chan)
 578                 return NULL;
 579
 580         chan->buf = alloc_percpu(struct rchan_buf *);
 581         chan->version = RELAYFS_CHANNEL_VERSION;
 582         chan->n_subbufs = n_subbufs;
 583         chan->subbuf_size = subbuf_size;
 584         chan->alloc_size = PAGE_ALIGN(subbuf_size * n_subbufs);
 585         chan->parent = parent;
 586         chan->private_data = private_data;
 587         if (base_filename) {
 588                 chan->has_base_filename = 1;
 589                 strlcpy(chan->base_filename, base_filename, NAME_MAX);
 590         }
 591         setup_callbacks(chan, cb);
 592         kref_init(&chan->kref);
 593
 594         mutex_lock(&relay_channels_mutex);
 595         for_each_online_cpu(i) {
 596                 buf = relay_open_buf(chan, i);
 597                 if (!buf)
 598                         goto free_bufs;
 599                 *per_cpu_ptr(chan->buf, i) = buf;
 600         }
 601         list_add(&chan->list, &relay_channels);
 602         mutex_unlock(&relay_channels_mutex);
 603
 604         return chan;
 605
 606 free_bufs:
 607         for_each_possible_cpu(i) {
 608                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 609                         relay_close_buf(buf);
 610         }
 611
 612         kref_put(&chan->kref, relay_destroy_channel);
 613         mutex_unlock(&relay_channels_mutex);
 614         return NULL;
 615 }
 616 EXPORT_SYMBOL_GPL(relay_open);
 617
 618 struct rchan_percpu_buf_dispatcher {
 619         struct rchan_buf *buf;
 620         struct dentry *dentry;
 621 };
 622
 623 /* Called in atomic context. */
 624 static void __relay_set_buf_dentry(void *info)
 625 {
 626         struct rchan_percpu_buf_dispatcher *p = info;
 627
 628         relay_set_buf_dentry(p->buf, p->dentry);
 629 }
 630
 631 /**
 632  *      relay_late_setup_files - triggers file creation
 633  *      @chan: channel to operate on
 634  *      @base_filename: base name of files to create
 635  *      @parent: dentry of parent directory, %NULL for root directory
 636  *
 637  *      Returns 0 if successful, non-zero otherwise.
 638  *
 639  *      Use to setup files for a previously buffer-only channel created
 640  *      by relay_open() with a NULL parent dentry.
 641  *
 642  *      For example, this is useful for perfomring early tracing in kernel,
 643  *      before VFS is up and then exposing the early results once the dentry
 644  *      is available.
 645  */
 646 int relay_late_setup_files(struct rchan *chan,
 647                            const char *base_filename,
 648                            struct dentry *parent)
 649 {
 650         int err = 0;
 651         unsigned int i, curr_cpu;
 652         unsigned long flags;
 653         struct dentry *dentry;
 654         struct rchan_buf *buf;
 655         struct rchan_percpu_buf_dispatcher disp;
 656
 657         if (!chan || !base_filename)
 658                 return -EINVAL;
 659
 660         strlcpy(chan->base_filename, base_filename, NAME_MAX);
 661
 662         mutex_lock(&relay_channels_mutex);
 663         /* Is chan already set up? */
 664         if (unlikely(chan->has_base_filename)) {
 665                 mutex_unlock(&relay_channels_mutex);
 666                 return -EEXIST;
 667         }
 668         chan->has_base_filename = 1;
 669         chan->parent = parent;
 670
 671         if (chan->is_global) {
 672                 err = -EINVAL;
 673                 buf = *per_cpu_ptr(chan->buf, 0);
 674                 if (!WARN_ON_ONCE(!buf)) {
 675                         dentry = relay_create_buf_file(chan, buf, 0);
 676                         if (dentry && !WARN_ON_ONCE(!chan->is_global)) {
 677                                 relay_set_buf_dentry(buf, dentry);
 678                                 err = 0;
 679                         }
 680                 }
 681                 mutex_unlock(&relay_channels_mutex);
 682                 return err;
 683         }
 684
 685         curr_cpu = get_cpu();
 686         /*
 687          * The CPU hotplug notifier ran before us and created buffers with
 688          * no files associated. So it's safe to call relay_setup_buf_file()
 689          * on all currently online CPUs.
 690          */
 691         for_each_online_cpu(i) {
 692                 buf = *per_cpu_ptr(chan->buf, i);
 693                 if (unlikely(!buf)) {
 694                         WARN_ONCE(1, KERN_ERR "CPU has no buffer!\n");
 695                         err = -EINVAL;
 696                         break;
 697                 }
 698
 699                 dentry = relay_create_buf_file(chan, buf, i);
 700                 if (unlikely(!dentry)) {
 701                         err = -EINVAL;
 702                         break;
 703                 }
 704
 705                 if (curr_cpu == i) {
 706                         local_irq_save(flags);
 707                         relay_set_buf_dentry(buf, dentry);
 708                         local_irq_restore(flags);
 709                 } else {
 710                         disp.buf = buf;
 711                         disp.dentry = dentry;
 712                         smp_mb();
 713                         /* relay_channels_mutex must be held, so wait. */
 714                         err = smp_call_function_single(i,
 715                                                        __relay_set_buf_dentry,
 716                                                        &disp, 1);
 717                 }
 718                 if (unlikely(err))
 719                         break;
 720         }
 721         put_cpu();
 722         mutex_unlock(&relay_channels_mutex);
 723
 724         return err;
 725 }
 726 EXPORT_SYMBOL_GPL(relay_late_setup_files);
 727
 728 /**
 729  *      relay_switch_subbuf - switch to a new sub-buffer
 730  *      @buf: channel buffer
 731  *      @length: size of current event
 732  *
 733  *      Returns either the length passed in or 0 if full.
 734  *
 735  *      Performs sub-buffer-switch tasks such as invoking callbacks,
 736  *      updating padding counts, waking up readers, etc.
 737  */
 738 size_t relay_switch_subbuf(struct rchan_buf *buf, size_t length)
 739 {
 740         void *old, *new;
 741         size_t old_subbuf, new_subbuf;
 742
 743         if (unlikely(length > buf->chan->subbuf_size))
 744                 goto toobig;
 745
 746         if (buf->offset != buf->chan->subbuf_size + 1) {
 747                 buf->prev_padding = buf->chan->subbuf_size - buf->offset;
 748                 old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 749                 buf->padding[old_subbuf] = buf->prev_padding;
 750                 buf->subbufs_produced++;
 751                 if (buf->dentry)
 752                         d_inode(buf->dentry)->i_size +=
 753                                 buf->chan->subbuf_size -
 754                                 buf->padding[old_subbuf];
 755                 else
 756                         buf->early_bytes += buf->chan->subbuf_size -
 757                                             buf->padding[old_subbuf];
 758                 smp_mb();
 759                 if (waitqueue_active(&buf->read_wait)) {
 760                         /*
 761                          * Calling wake_up_interruptible() from here
 762                          * will deadlock if we happen to be logging
 763                          * from the scheduler (trying to re-grab
 764                          * rq->lock), so defer it.
 765                          */
 766                         irq_work_queue(&buf->wakeup_work);
 767                 }
 768         }
 769
 770         old = buf->data;
 771         new_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 772         new = buf->start + new_subbuf * buf->chan->subbuf_size;
 773         buf->offset = 0;
 774         if (!buf->chan->cb->subbuf_start(buf, new, old, buf->prev_padding)) {
 775                 buf->offset = buf->chan->subbuf_size + 1;
 776                 return 0;
 777         }
 778         buf->data = new;
 779         buf->padding[new_subbuf] = 0;
 780
 781         if (unlikely(length + buf->offset > buf->chan->subbuf_size))
 782                 goto toobig;
 783
 784         return length;
 785
 786 toobig:
 787         buf->chan->last_toobig = length;
 788         return 0;
 789 }
 790 EXPORT_SYMBOL_GPL(relay_switch_subbuf);
 791
 792 /**
 793  *      relay_subbufs_consumed - update the buffer's sub-buffers-consumed count
 794  *      @chan: the channel
 795  *      @cpu: the cpu associated with the channel buffer to update
 796  *      @subbufs_consumed: number of sub-buffers to add to current buf's count
 797  *
 798  *      Adds to the channel buffer's consumed sub-buffer count.
 799  *      subbufs_consumed should be the number of sub-buffers newly consumed,
 800  *      not the total consumed.
 801  *
 802  *      NOTE. Kernel clients don't need to call this function if the channel
 803  *      mode is 'overwrite'.
 804  */
 805 void relay_subbufs_consumed(struct rchan *chan,
 806                             unsigned int cpu,
 807                             size_t subbufs_consumed)
 808 {
 809         struct rchan_buf *buf;
 810
 811         if (!chan || cpu >= NR_CPUS)
 812                 return;
 813
 814         buf = *per_cpu_ptr(chan->buf, cpu);
 815         if (!buf || subbufs_consumed > chan->n_subbufs)
 816                 return;
 817
 818         if (subbufs_consumed > buf->subbufs_produced - buf->subbufs_consumed)
 819                 buf->subbufs_consumed = buf->subbufs_produced;
 820         else
 821                 buf->subbufs_consumed += subbufs_consumed;
 822 }
 823 EXPORT_SYMBOL_GPL(relay_subbufs_consumed);
 824
 825 /**
 826  *      relay_close - close the channel
 827  *      @chan: the channel
 828  *
 829  *      Closes all channel buffers and frees the channel.
 830  */
 831 void relay_close(struct rchan *chan)
 832 {
 833         struct rchan_buf *buf;
 834         unsigned int i;
 835
 836         if (!chan)
 837                 return;
 838
 839         mutex_lock(&relay_channels_mutex);
 840         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0)))
 841                 relay_close_buf(buf);
 842         else
 843                 for_each_possible_cpu(i)
 844                         if ((buf = *per_cpu_ptr(chan->buf, i)))
 845                                 relay_close_buf(buf);
 846
 847         if (chan->last_toobig)
 848                 printk(KERN_WARNING "relay: one or more items not logged "
 849                        "[item size (%zd) > sub-buffer size (%zd)]\n",
 850                        chan->last_toobig, chan->subbuf_size);
 851
 852         list_del(&chan->list);
 853         kref_put(&chan->kref, relay_destroy_channel);
 854         mutex_unlock(&relay_channels_mutex);
 855 }
 856 EXPORT_SYMBOL_GPL(relay_close);
 857
 858 /**
 859  *      relay_flush - close the channel
 860  *      @chan: the channel
 861  *
 862  *      Flushes all channel buffers, i.e. forces buffer switch.
 863  */
 864 void relay_flush(struct rchan *chan)
 865 {
 866         struct rchan_buf *buf;
 867         unsigned int i;
 868
 869         if (!chan)
 870                 return;
 871
 872         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 873                 relay_switch_subbuf(buf, 0);
 874                 return;
 875         }
 876
 877         mutex_lock(&relay_channels_mutex);
 878         for_each_possible_cpu(i)
 879                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 880                         relay_switch_subbuf(buf, 0);
 881         mutex_unlock(&relay_channels_mutex);
 882 }
 883 EXPORT_SYMBOL_GPL(relay_flush);
 884
 885 /**
 886  *      relay_file_open - open file op for relay files
 887  *      @inode: the inode
 888  *      @filp: the file
 889  *
 890  *      Increments the channel buffer refcount.
 891  */
 892 static int relay_file_open(struct inode *inode, struct file *filp)
 893 {
 894         struct rchan_buf *buf = inode->i_private;
 895         kref_get(&buf->kref);
 896         filp->private_data = buf;
 897
 898         return nonseekable_open(inode, filp);
 899 }
 900
 901 /**
 902  *      relay_file_mmap - mmap file op for relay files
 903  *      @filp: the file
 904  *      @vma: the vma describing what to map
 905  *
 906  *      Calls upon relay_mmap_buf() to map the file into user space.
 907  */
 908 static int relay_file_mmap(struct file *filp, struct vm_area_struct *vma)
 909 {
 910         struct rchan_buf *buf = filp->private_data;
 911         return relay_mmap_buf(buf, vma);
 912 }
 913
 914 /**
 915  *      relay_file_poll - poll file op for relay files
 916  *      @filp: the file
 917  *      @wait: poll table
 918  *
 919  *      Poll implemention.
 920  */
 921 static __poll_t relay_file_poll(struct file *filp, poll_table *wait)
 922 {
 923         __poll_t mask = 0;
 924         struct rchan_buf *buf = filp->private_data;
 925
 926         if (buf->finalized)
 927                 return EPOLLERR;
 928
 929         if (filp->f_mode & FMODE_READ) {
 930                 poll_wait(filp, &buf->read_wait, wait);
 931                 if (!relay_buf_empty(buf))
 932                         mask |= EPOLLIN | EPOLLRDNORM;
 933         }
 934
 935         return mask;
 936 }
 937
 938 /**
 939  *      relay_file_release - release file op for relay files
 940  *      @inode: the inode
 941  *      @filp: the file
 942  *
 943  *      Decrements the channel refcount, as the filesystem is
 944  *      no longer using it.
 945  */
 946 static int relay_file_release(struct inode *inode, struct file *filp)
 947 {
 948         struct rchan_buf *buf = filp->private_data;
 949         kref_put(&buf->kref, relay_remove_buf);
 950
 951         return 0;
 952 }
 953
 954 /*
 955  *      relay_file_read_consume - update the consumed count for the buffer
 956  */
 957 static void relay_file_read_consume(struct rchan_buf *buf,
 958                                     size_t read_pos,
 959                                     size_t bytes_consumed)
 960 {
 961         size_t subbuf_size = buf->chan->subbuf_size;
 962         size_t n_subbufs = buf->chan->n_subbufs;
 963         size_t read_subbuf;
 964
 965         if (buf->subbufs_produced == buf->subbufs_consumed &&
 966             buf->offset == buf->bytes_consumed)
 967                 return;
 968
 969         if (buf->bytes_consumed + bytes_consumed > subbuf_size) {
 970                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 971                 buf->bytes_consumed = 0;
 972         }
 973
 974         buf->bytes_consumed += bytes_consumed;
 975         if (!read_pos)
 976                 read_subbuf = buf->subbufs_consumed % n_subbufs;
 977         else
 978                 read_subbuf = read_pos / buf->chan->subbuf_size;
 979         if (buf->bytes_consumed + buf->padding[read_subbuf] == subbuf_size) {
 980                 if ((read_subbuf == buf->subbufs_produced % n_subbufs) &&
 981                     (buf->offset == subbuf_size))
 982                         return;
 983                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 984                 buf->bytes_consumed = 0;
 985         }
 986 }
 987
 988 /*
 989  *      relay_file_read_avail - boolean, are there unconsumed bytes available?
 990  */
 991 static int relay_file_read_avail(struct rchan_buf *buf, size_t read_pos)
 992 {
 993         size_t subbuf_size = buf->chan->subbuf_size;
 994         size_t n_subbufs = buf->chan->n_subbufs;
 995         size_t produced = buf->subbufs_produced;
 996         size_t consumed = buf->subbufs_consumed;
 997
 998         relay_file_read_consume(buf, read_pos, 0);
 999
1000         consumed = buf->subbufs_consumed;
1001
1002         if (unlikely(buf->offset > subbuf_size)) {
1003                 if (produced == consumed)
1004                         return 0;
1005                 return 1;
1006         }
1007
1008         if (unlikely(produced - consumed >= n_subbufs)) {
1009                 consumed = produced - n_subbufs + 1;
1010                 buf->subbufs_consumed = consumed;
1011                 buf->bytes_consumed = 0;
1012         }
1013
1014         produced = (produced % n_subbufs) * subbuf_size + buf->offset;
1015         consumed = (consumed % n_subbufs) * subbuf_size + buf->bytes_consumed;
1016
1017         if (consumed > produced)
1018                 produced += n_subbufs * subbuf_size;
1019
1020         if (consumed == produced) {
1021                 if (buf->offset == subbuf_size &&
1022                     buf->subbufs_produced > buf->subbufs_consumed)
1023                         return 1;
1024                 return 0;
1025         }
1026
1027         return 1;
1028 }
1029
1030 /**
1031  *      relay_file_read_subbuf_avail - return bytes available in sub-buffer
1032  *      @read_pos: file read position
1033  *      @buf: relay channel buffer
1034  */
1035 static size_t relay_file_read_subbuf_avail(size_t read_pos,
1036                                            struct rchan_buf *buf)
1037 {
1038         size_t padding, avail = 0;
1039         size_t read_subbuf, read_offset, write_subbuf, write_offset;
1040         size_t subbuf_size = buf->chan->subbuf_size;
1041
1042         write_subbuf = (buf->data - buf->start) / subbuf_size;
1043         write_offset = buf->offset > subbuf_size ? subbuf_size : buf->offset;
1044         read_subbuf = read_pos / subbuf_size;
1045         read_offset = read_pos % subbuf_size;
1046         padding = buf->padding[read_subbuf];
1047
1048         if (read_subbuf == write_subbuf) {
1049                 if (read_offset + padding < write_offset)
1050                         avail = write_offset - (read_offset + padding);
1051         } else
1052                 avail = (subbuf_size - padding) - read_offset;
1053
1054         return avail;
1055 }
1056
1057 /**
1058  *      relay_file_read_start_pos - find the first available byte to read
1059  *      @read_pos: file read position
1060  *      @buf: relay channel buffer
1061  *
1062  *      If the @read_pos is in the middle of padding, return the
1063  *      position of the first actually available byte, otherwise
1064  *      return the original value.
1065  */
1066 static size_t relay_file_read_start_pos(size_t read_pos,
1067                                         struct rchan_buf *buf)
1068 {
1069         size_t read_subbuf, padding, padding_start, padding_end;
1070         size_t subbuf_size = buf->chan->subbuf_size;
1071         size_t n_subbufs = buf->chan->n_subbufs;
1072         size_t consumed = buf->subbufs_consumed % n_subbufs;
1073
1074         if (!read_pos)
1075                 read_pos = consumed * subbuf_size + buf->bytes_consumed;
1076         read_subbuf = read_pos / subbuf_size;
1077         padding = buf->padding[read_subbuf];
1078         padding_start = (read_subbuf + 1) * subbuf_size - padding;
1079         padding_end = (read_subbuf + 1) * subbuf_size;
1080         if (read_pos >= padding_start && read_pos < padding_end) {
1081                 read_subbuf = (read_subbuf + 1) % n_subbufs;
1082                 read_pos = read_subbuf * subbuf_size;
1083         }
1084
1085         return read_pos;
1086 }
1087
1088 /**
1089  *      relay_file_read_end_pos - return the new read position
1090  *      @read_pos: file read position
1091  *      @buf: relay channel buffer
1092  *      @count: number of bytes to be read
1093  */
1094 static size_t relay_file_read_end_pos(struct rchan_buf *buf,
1095                                       size_t read_pos,
1096                                       size_t count)
1097 {
1098         size_t read_subbuf, padding, end_pos;
1099         size_t subbuf_size = buf->chan->subbuf_size;
1100         size_t n_subbufs = buf->chan->n_subbufs;
1101
1102         read_subbuf = read_pos / subbuf_size;
1103         padding = buf->padding[read_subbuf];
1104         if (read_pos % subbuf_size + count + padding == subbuf_size)
1105                 end_pos = (read_subbuf + 1) * subbuf_size;
1106         else
1107                 end_pos = read_pos + count;
1108         if (end_pos >= subbuf_size * n_subbufs)
1109                 end_pos = 0;
1110
1111         return end_pos;
1112 }
1113
1114 static ssize_t relay_file_read(struct file *filp,
1115                                char __user *buffer,
1116                                size_t count,
1117                                loff_t *ppos)
1118 {
1119         struct rchan_buf *buf = filp->private_data;
1120         size_t read_start, avail;
1121         size_t written = 0;
1122         int ret;
1123
1124         if (!count)
1125                 return 0;
1126
1127         inode_lock(file_inode(filp));
1128         do {
1129                 void *from;
1130
1131                 if (!relay_file_read_avail(buf, *ppos))
1132                         break;
1133
1134                 read_start = relay_file_read_start_pos(*ppos, buf);
1135                 avail = relay_file_read_subbuf_avail(read_start, buf);
1136                 if (!avail)
1137                         break;
1138
1139                 avail = min(count, avail);
1140                 from = buf->start + read_start;
1141                 ret = avail;
1142                 if (copy_to_user(buffer, from, avail))
1143                         break;
1144
1145                 buffer += ret;
1146                 written += ret;
1147                 count -= ret;
1148
1149                 relay_file_read_consume(buf, read_start, ret);
1150                 *ppos = relay_file_read_end_pos(buf, read_start, ret);
1151         } while (count);
1152         inode_unlock(file_inode(filp));
1153
1154         return written;
1155 }
1156
1157 static void relay_consume_bytes(struct rchan_buf *rbuf, int bytes_consumed)
1158 {
1159         rbuf->bytes_consumed += bytes_consumed;
1160
1161         if (rbuf->bytes_consumed >= rbuf->chan->subbuf_size) {
1162                 relay_subbufs_consumed(rbuf->chan, rbuf->cpu, 1);
1163                 rbuf->bytes_consumed %= rbuf->chan->subbuf_size;
1164         }
1165 }
1166
1167 static void relay_pipe_buf_release(struct pipe_inode_info *pipe,
1168                                    struct pipe_buffer *buf)
1169 {
1170         struct rchan_buf *rbuf;
1171
1172         rbuf = (struct rchan_buf *)page_private(buf->page);
1173         relay_consume_bytes(rbuf, buf->private);
1174 }
1175
1176 static const struct pipe_buf_operations relay_pipe_buf_ops = {
1177         .can_merge = 0,
1178         .confirm = generic_pipe_buf_confirm,
1179         .release = relay_pipe_buf_release,
1180         .steal = generic_pipe_buf_steal,
1181         .get = generic_pipe_buf_get,
1182 };
1183
1184 static void relay_page_release(struct splice_pipe_desc *spd, unsigned int i)
1185 {
1186 }
1187
1188 /*
1189  *      subbuf_splice_actor - splice up to one subbuf's worth of data
1190  */
1191 static ssize_t subbuf_splice_actor(struct file *in,
1192                                loff_t *ppos,
1193                                struct pipe_inode_info *pipe,
1194                                size_t len,
1195                                unsigned int flags,
1196                                int *nonpad_ret)
1197 {
1198         unsigned int pidx, poff, total_len, subbuf_pages, nr_pages;
1199         struct rchan_buf *rbuf = in->private_data;
1200         unsigned int subbuf_size = rbuf->chan->subbuf_size;
1201         uint64_t pos = (uint64_t) *ppos;
1202         uint32_t alloc_size = (uint32_t) rbuf->chan->alloc_size;
1203         size_t read_start = (size_t) do_div(pos, alloc_size);
1204         size_t read_subbuf = read_start / subbuf_size;
1205         size_t padding = rbuf->padding[read_subbuf];
1206         size_t nonpad_end = read_subbuf * subbuf_size + subbuf_size - padding;
1207         struct page *pages[PIPE_DEF_BUFFERS];
1208         struct partial_page partial[PIPE_DEF_BUFFERS];
1209         struct splice_pipe_desc spd = {
1210                 .pages = pages,
1211                 .nr_pages = 0,
1212                 .nr_pages_max = PIPE_DEF_BUFFERS,
1213                 .partial = partial,
1214                 .ops = &relay_pipe_buf_ops,
1215                 .spd_release = relay_page_release,
1216         };
1217         ssize_t ret;
1218
1219         if (rbuf->subbufs_produced == rbuf->subbufs_consumed)
1220                 return 0;
1221         if (splice_grow_spd(pipe, &spd))
1222                 return -ENOMEM;
1223
1224         /*
1225          * Adjust read len, if longer than what is available
1226          */
1227         if (len > (subbuf_size - read_start % subbuf_size))
1228                 len = subbuf_size - read_start % subbuf_size;
1229
1230         subbuf_pages = rbuf->chan->alloc_size >> PAGE_SHIFT;
1231         pidx = (read_start / PAGE_SIZE) % subbuf_pages;
1232         poff = read_start & ~PAGE_MASK;
1233         nr_pages = min_t(unsigned int, subbuf_pages, spd.nr_pages_max);
1234
1235         for (total_len = 0; spd.nr_pages < nr_pages; spd.nr_pages++) {
1236                 unsigned int this_len, this_end, private;
1237                 unsigned int cur_pos = read_start + total_len;
1238
1239                 if (!len)
1240                         break;
1241
1242                 this_len = min_t(unsigned long, len, PAGE_SIZE - poff);
1243                 private = this_len;
1244
1245                 spd.pages[spd.nr_pages] = rbuf->page_array[pidx];
1246                 spd.partial[spd.nr_pages].offset = poff;
1247
1248                 this_end = cur_pos + this_len;
1249                 if (this_end >= nonpad_end) {
1250                         this_len = nonpad_end - cur_pos;
1251                         private = this_len + padding;
1252                 }
1253                 spd.partial[spd.nr_pages].len = this_len;
1254                 spd.partial[spd.nr_pages].private = private;
1255
1256                 len -= this_len;
1257                 total_len += this_len;
1258                 poff = 0;
1259                 pidx = (pidx + 1) % subbuf_pages;
1260
1261                 if (this_end >= nonpad_end) {
1262                         spd.nr_pages++;
1263                         break;
1264                 }
1265         }
1266
1267         ret = 0;
1268         if (!spd.nr_pages)
1269                 goto out;
1270
1271         ret = *nonpad_ret = splice_to_pipe(pipe, &spd);
1272         if (ret < 0 || ret < total_len)
1273                 goto out;
1274
1275         if (read_start + ret == nonpad_end)
1276                 ret += padding;
1277
1278 out:
1279         splice_shrink_spd(&spd);
1280         return ret;
1281 }
1282
1283 static ssize_t relay_file_splice_read(struct file *in,
1284                                       loff_t *ppos,
1285                                       struct pipe_inode_info *pipe,
1286                                       size_t len,
1287                                       unsigned int flags)
1288 {
1289         ssize_t spliced;
1290         int ret;
1291         int nonpad_ret = 0;
1292
1293         ret = 0;
1294         spliced = 0;
1295
1296         while (len && !spliced) {
1297                 ret = subbuf_splice_actor(in, ppos, pipe, len, flags, &nonpad_ret);
1298                 if (ret < 0)
1299                         break;
1300                 else if (!ret) {
1301                         if (flags & SPLICE_F_NONBLOCK)
1302                                 ret = -EAGAIN;
1303                         break;
1304                 }
1305
1306                 *ppos += ret;
1307                 if (ret > len)
1308                         len = 0;
1309                 else
1310                         len -= ret;
1311                 spliced += nonpad_ret;
1312                 nonpad_ret = 0;
1313         }
1314
1315         if (spliced)
1316                 return spliced;
1317
1318         return ret;
1319 }
1320
1321 const struct file_operations relay_file_operations = {
1322         .open           = relay_file_open,
1323         .poll           = relay_file_poll,
1324         .mmap           = relay_file_mmap,
1325         .read           = relay_file_read,
1326         .llseek         = no_llseek,
1327         .release        = relay_file_release,
1328         .splice_read    = relay_file_splice_read,
1329 };
1330 EXPORT_SYMBOL_GPL(relay_file_operations);