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  * 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 vm_fault_t 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_lock.
  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_array(chan->n_subbufs, sizeof(size_t *),
 173                                      GFP_KERNEL);
 174         if (!buf->padding)
 175                 goto free_buf;
 176
 177         buf->start = relay_alloc_buf(buf, &chan->alloc_size);
 178         if (!buf->start)
 179                 goto free_buf;
 180
 181         buf->chan = chan;
 182         kref_get(&buf->chan->kref);
 183         return buf;
 184
 185 free_buf:
 186         kfree(buf->padding);
 187         kfree(buf);
 188         return NULL;
 189 }
 190
 191 /**
 192  *      relay_destroy_channel - free the channel struct
 193  *      @kref: target kernel reference that contains the relay channel
 194  *
 195  *      Should only be called from kref_put().
 196  */
 197 static void relay_destroy_channel(struct kref *kref)
 198 {
 199         struct rchan *chan = container_of(kref, struct rchan, kref);
 200         kfree(chan);
 201 }
 202
 203 /**
 204  *      relay_destroy_buf - destroy an rchan_buf struct and associated buffer
 205  *      @buf: the buffer struct
 206  */
 207 static void relay_destroy_buf(struct rchan_buf *buf)
 208 {
 209         struct rchan *chan = buf->chan;
 210         unsigned int i;
 211
 212         if (likely(buf->start)) {
 213                 vunmap(buf->start);
 214                 for (i = 0; i < buf->page_count; i++)
 215                         __free_page(buf->page_array[i]);
 216                 relay_free_page_array(buf->page_array);
 217         }
 218         *per_cpu_ptr(chan->buf, buf->cpu) = NULL;
 219         kfree(buf->padding);
 220         kfree(buf);
 221         kref_put(&chan->kref, relay_destroy_channel);
 222 }
 223
 224 /**
 225  *      relay_remove_buf - remove a channel buffer
 226  *      @kref: target kernel reference that contains the relay buffer
 227  *
 228  *      Removes the file from the filesystem, which also frees the
 229  *      rchan_buf_struct and the channel buffer.  Should only be called from
 230  *      kref_put().
 231  */
 232 static void relay_remove_buf(struct kref *kref)
 233 {
 234         struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
 235         relay_destroy_buf(buf);
 236 }
 237
 238 /**
 239  *      relay_buf_empty - boolean, is the channel buffer empty?
 240  *      @buf: channel buffer
 241  *
 242  *      Returns 1 if the buffer is empty, 0 otherwise.
 243  */
 244 static int relay_buf_empty(struct rchan_buf *buf)
 245 {
 246         return (buf->subbufs_produced - buf->subbufs_consumed) ? 0 : 1;
 247 }
 248
 249 /**
 250  *      relay_buf_full - boolean, is the channel buffer full?
 251  *      @buf: channel buffer
 252  *
 253  *      Returns 1 if the buffer is full, 0 otherwise.
 254  */
 255 int relay_buf_full(struct rchan_buf *buf)
 256 {
 257         size_t ready = buf->subbufs_produced - buf->subbufs_consumed;
 258         return (ready >= buf->chan->n_subbufs) ? 1 : 0;
 259 }
 260 EXPORT_SYMBOL_GPL(relay_buf_full);
 261
 262 /*
 263  * High-level relay kernel API and associated functions.
 264  */
 265
 266 /*
 267  * rchan_callback implementations defining default channel behavior.  Used
 268  * in place of corresponding NULL values in client callback struct.
 269  */
 270
 271 /*
 272  * subbuf_start() default callback.  Does nothing.
 273  */
 274 static int subbuf_start_default_callback (struct rchan_buf *buf,
 275                                           void *subbuf,
 276                                           void *prev_subbuf,
 277                                           size_t prev_padding)
 278 {
 279         if (relay_buf_full(buf))
 280                 return 0;
 281
 282         return 1;
 283 }
 284
 285 /*
 286  * buf_mapped() default callback.  Does nothing.
 287  */
 288 static void buf_mapped_default_callback(struct rchan_buf *buf,
 289                                         struct file *filp)
 290 {
 291 }
 292
 293 /*
 294  * buf_unmapped() default callback.  Does nothing.
 295  */
 296 static void buf_unmapped_default_callback(struct rchan_buf *buf,
 297                                           struct file *filp)
 298 {
 299 }
 300
 301 /*
 302  * create_buf_file_create() default callback.  Does nothing.
 303  */
 304 static struct dentry *create_buf_file_default_callback(const char *filename,
 305                                                        struct dentry *parent,
 306                                                        umode_t mode,
 307                                                        struct rchan_buf *buf,
 308                                                        int *is_global)
 309 {
 310         return NULL;
 311 }
 312
 313 /*
 314  * remove_buf_file() default callback.  Does nothing.
 315  */
 316 static int remove_buf_file_default_callback(struct dentry *dentry)
 317 {
 318         return -EINVAL;
 319 }
 320
 321 /* relay channel default callbacks */
 322 static struct rchan_callbacks default_channel_callbacks = {
 323         .subbuf_start = subbuf_start_default_callback,
 324         .buf_mapped = buf_mapped_default_callback,
 325         .buf_unmapped = buf_unmapped_default_callback,
 326         .create_buf_file = create_buf_file_default_callback,
 327         .remove_buf_file = remove_buf_file_default_callback,
 328 };
 329
 330 /**
 331  *      wakeup_readers - wake up readers waiting on a channel
 332  *      @work: contains the channel buffer
 333  *
 334  *      This is the function used to defer reader waking
 335  */
 336 static void wakeup_readers(struct irq_work *work)
 337 {
 338         struct rchan_buf *buf;
 339
 340         buf = container_of(work, struct rchan_buf, wakeup_work);
 341         wake_up_interruptible(&buf->read_wait);
 342 }
 343
 344 /**
 345  *      __relay_reset - reset a channel buffer
 346  *      @buf: the channel buffer
 347  *      @init: 1 if this is a first-time initialization
 348  *
 349  *      See relay_reset() for description of effect.
 350  */
 351 static void __relay_reset(struct rchan_buf *buf, unsigned int init)
 352 {
 353         size_t i;
 354
 355         if (init) {
 356                 init_waitqueue_head(&buf->read_wait);
 357                 kref_init(&buf->kref);
 358                 init_irq_work(&buf->wakeup_work, wakeup_readers);
 359         } else {
 360                 irq_work_sync(&buf->wakeup_work);
 361         }
 362
 363         buf->subbufs_produced = 0;
 364         buf->subbufs_consumed = 0;
 365         buf->bytes_consumed = 0;
 366         buf->finalized = 0;
 367         buf->data = buf->start;
 368         buf->offset = 0;
 369
 370         for (i = 0; i < buf->chan->n_subbufs; i++)
 371                 buf->padding[i] = 0;
 372
 373         buf->chan->cb->subbuf_start(buf, buf->data, NULL, 0);
 374 }
 375
 376 /**
 377  *      relay_reset - reset the channel
 378  *      @chan: the channel
 379  *
 380  *      This has the effect of erasing all data from all channel buffers
 381  *      and restarting the channel in its initial state.  The buffers
 382  *      are not freed, so any mappings are still in effect.
 383  *
 384  *      NOTE. Care should be taken that the channel isn't actually
 385  *      being used by anything when this call is made.
 386  */
 387 void relay_reset(struct rchan *chan)
 388 {
 389         struct rchan_buf *buf;
 390         unsigned int i;
 391
 392         if (!chan)
 393                 return;
 394
 395         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 396                 __relay_reset(buf, 0);
 397                 return;
 398         }
 399
 400         mutex_lock(&relay_channels_mutex);
 401         for_each_possible_cpu(i)
 402                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 403                         __relay_reset(buf, 0);
 404         mutex_unlock(&relay_channels_mutex);
 405 }
 406 EXPORT_SYMBOL_GPL(relay_reset);
 407
 408 static inline void relay_set_buf_dentry(struct rchan_buf *buf,
 409                                         struct dentry *dentry)
 410 {
 411         buf->dentry = dentry;
 412         d_inode(buf->dentry)->i_size = buf->early_bytes;
 413 }
 414
 415 static struct dentry *relay_create_buf_file(struct rchan *chan,
 416                                             struct rchan_buf *buf,
 417                                             unsigned int cpu)
 418 {
 419         struct dentry *dentry;
 420         char *tmpname;
 421
 422         tmpname = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 423         if (!tmpname)
 424                 return NULL;
 425         snprintf(tmpname, NAME_MAX, "%s%d", chan->base_filename, cpu);
 426
 427         /* Create file in fs */
 428         dentry = chan->cb->create_buf_file(tmpname, chan->parent,
 429                                            S_IRUSR, buf,
 430                                            &chan->is_global);
 431         if (IS_ERR(dentry))
 432                 dentry = NULL;
 433
 434         kfree(tmpname);
 435
 436         return dentry;
 437 }
 438
 439 /*
 440  *      relay_open_buf - create a new relay channel buffer
 441  *
 442  *      used by relay_open() and CPU hotplug.
 443  */
 444 static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
 445 {
 446         struct rchan_buf *buf = NULL;
 447         struct dentry *dentry;
 448
 449         if (chan->is_global)
 450                 return *per_cpu_ptr(chan->buf, 0);
 451
 452         buf = relay_create_buf(chan);
 453         if (!buf)
 454                 return NULL;
 455
 456         if (chan->has_base_filename) {
 457                 dentry = relay_create_buf_file(chan, buf, cpu);
 458                 if (!dentry)
 459                         goto free_buf;
 460                 relay_set_buf_dentry(buf, dentry);
 461         } else {
 462                 /* Only retrieve global info, nothing more, nothing less */
 463                 dentry = chan->cb->create_buf_file(NULL, NULL,
 464                                                    S_IRUSR, buf,
 465                                                    &chan->is_global);
 466                 if (IS_ERR_OR_NULL(dentry))
 467                         goto free_buf;
 468         }
 469
 470         buf->cpu = cpu;
 471         __relay_reset(buf, 1);
 472
 473         if(chan->is_global) {
 474                 *per_cpu_ptr(chan->buf, 0) = buf;
 475                 buf->cpu = 0;
 476         }
 477
 478         return buf;
 479
 480 free_buf:
 481         relay_destroy_buf(buf);
 482         return NULL;
 483 }
 484
 485 /**
 486  *      relay_close_buf - close a channel buffer
 487  *      @buf: channel buffer
 488  *
 489  *      Marks the buffer finalized and restores the default callbacks.
 490  *      The channel buffer and channel buffer data structure are then freed
 491  *      automatically when the last reference is given up.
 492  */
 493 static void relay_close_buf(struct rchan_buf *buf)
 494 {
 495         buf->finalized = 1;
 496         irq_work_sync(&buf->wakeup_work);
 497         buf->chan->cb->remove_buf_file(buf->dentry);
 498         kref_put(&buf->kref, relay_remove_buf);
 499 }
 500
 501 static void setup_callbacks(struct rchan *chan,
 502                                    struct rchan_callbacks *cb)
 503 {
 504         if (!cb) {
 505                 chan->cb = &default_channel_callbacks;
 506                 return;
 507         }
 508
 509         if (!cb->subbuf_start)
 510                 cb->subbuf_start = subbuf_start_default_callback;
 511         if (!cb->buf_mapped)
 512                 cb->buf_mapped = buf_mapped_default_callback;
 513         if (!cb->buf_unmapped)
 514                 cb->buf_unmapped = buf_unmapped_default_callback;
 515         if (!cb->create_buf_file)
 516                 cb->create_buf_file = create_buf_file_default_callback;
 517         if (!cb->remove_buf_file)
 518                 cb->remove_buf_file = remove_buf_file_default_callback;
 519         chan->cb = cb;
 520 }
 521
 522 int relay_prepare_cpu(unsigned int cpu)
 523 {
 524         struct rchan *chan;
 525         struct rchan_buf *buf;
 526
 527         mutex_lock(&relay_channels_mutex);
 528         list_for_each_entry(chan, &relay_channels, list) {
 529                 if ((buf = *per_cpu_ptr(chan->buf, cpu)))
 530                         continue;
 531                 buf = relay_open_buf(chan, cpu);
 532                 if (!buf) {
 533                         pr_err("relay: cpu %d buffer creation failed\n", cpu);
 534                         mutex_unlock(&relay_channels_mutex);
 535                         return -ENOMEM;
 536                 }
 537                 *per_cpu_ptr(chan->buf, cpu) = buf;
 538         }
 539         mutex_unlock(&relay_channels_mutex);
 540         return 0;
 541 }
 542
 543 /**
 544  *      relay_open - create a new relay channel
 545  *      @base_filename: base name of files to create, %NULL for buffering only
 546  *      @parent: dentry of parent directory, %NULL for root directory or buffer
 547  *      @subbuf_size: size of sub-buffers
 548  *      @n_subbufs: number of sub-buffers
 549  *      @cb: client callback functions
 550  *      @private_data: user-defined data
 551  *
 552  *      Returns channel pointer if successful, %NULL otherwise.
 553  *
 554  *      Creates a channel buffer for each cpu using the sizes and
 555  *      attributes specified.  The created channel buffer files
 556  *      will be named base_filename0...base_filenameN-1.  File
 557  *      permissions will be %S_IRUSR.
 558  *
 559  *      If opening a buffer (@parent = NULL) that you later wish to register
 560  *      in a filesystem, call relay_late_setup_files() once the @parent dentry
 561  *      is available.
 562  */
 563 struct rchan *relay_open(const char *base_filename,
 564                          struct dentry *parent,
 565                          size_t subbuf_size,
 566                          size_t n_subbufs,
 567                          struct rchan_callbacks *cb,
 568                          void *private_data)
 569 {
 570         unsigned int i;
 571         struct rchan *chan;
 572         struct rchan_buf *buf;
 573
 574         if (!(subbuf_size && n_subbufs))
 575                 return NULL;
 576         if (subbuf_size > UINT_MAX / n_subbufs)
 577                 return NULL;
 578
 579         chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
 580         if (!chan)
 581                 return NULL;
 582
 583         chan->buf = alloc_percpu(struct rchan_buf *);
 584         if (!chan->buf) {
 585                 kfree(chan);
 586                 return NULL;
 587         }
 588
 589         chan->version = RELAYFS_CHANNEL_VERSION;
 590         chan->n_subbufs = n_subbufs;
 591         chan->subbuf_size = subbuf_size;
 592         chan->alloc_size = PAGE_ALIGN(subbuf_size * n_subbufs);
 593         chan->parent = parent;
 594         chan->private_data = private_data;
 595         if (base_filename) {
 596                 chan->has_base_filename = 1;
 597                 strlcpy(chan->base_filename, base_filename, NAME_MAX);
 598         }
 599         setup_callbacks(chan, cb);
 600         kref_init(&chan->kref);
 601
 602         mutex_lock(&relay_channels_mutex);
 603         for_each_online_cpu(i) {
 604                 buf = relay_open_buf(chan, i);
 605                 if (!buf)
 606                         goto free_bufs;
 607                 *per_cpu_ptr(chan->buf, i) = buf;
 608         }
 609         list_add(&chan->list, &relay_channels);
 610         mutex_unlock(&relay_channels_mutex);
 611
 612         return chan;
 613
 614 free_bufs:
 615         for_each_possible_cpu(i) {
 616                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 617                         relay_close_buf(buf);
 618         }
 619
 620         kref_put(&chan->kref, relay_destroy_channel);
 621         mutex_unlock(&relay_channels_mutex);
 622         return NULL;
 623 }
 624 EXPORT_SYMBOL_GPL(relay_open);
 625
 626 struct rchan_percpu_buf_dispatcher {
 627         struct rchan_buf *buf;
 628         struct dentry *dentry;
 629 };
 630
 631 /* Called in atomic context. */
 632 static void __relay_set_buf_dentry(void *info)
 633 {
 634         struct rchan_percpu_buf_dispatcher *p = info;
 635
 636         relay_set_buf_dentry(p->buf, p->dentry);
 637 }
 638
 639 /**
 640  *      relay_late_setup_files - triggers file creation
 641  *      @chan: channel to operate on
 642  *      @base_filename: base name of files to create
 643  *      @parent: dentry of parent directory, %NULL for root directory
 644  *
 645  *      Returns 0 if successful, non-zero otherwise.
 646  *
 647  *      Use to setup files for a previously buffer-only channel created
 648  *      by relay_open() with a NULL parent dentry.
 649  *
 650  *      For example, this is useful for perfomring early tracing in kernel,
 651  *      before VFS is up and then exposing the early results once the dentry
 652  *      is available.
 653  */
 654 int relay_late_setup_files(struct rchan *chan,
 655                            const char *base_filename,
 656                            struct dentry *parent)
 657 {
 658         int err = 0;
 659         unsigned int i, curr_cpu;
 660         unsigned long flags;
 661         struct dentry *dentry;
 662         struct rchan_buf *buf;
 663         struct rchan_percpu_buf_dispatcher disp;
 664
 665         if (!chan || !base_filename)
 666                 return -EINVAL;
 667
 668         strlcpy(chan->base_filename, base_filename, NAME_MAX);
 669
 670         mutex_lock(&relay_channels_mutex);
 671         /* Is chan already set up? */
 672         if (unlikely(chan->has_base_filename)) {
 673                 mutex_unlock(&relay_channels_mutex);
 674                 return -EEXIST;
 675         }
 676         chan->has_base_filename = 1;
 677         chan->parent = parent;
 678
 679         if (chan->is_global) {
 680                 err = -EINVAL;
 681                 buf = *per_cpu_ptr(chan->buf, 0);
 682                 if (!WARN_ON_ONCE(!buf)) {
 683                         dentry = relay_create_buf_file(chan, buf, 0);
 684                         if (dentry && !WARN_ON_ONCE(!chan->is_global)) {
 685                                 relay_set_buf_dentry(buf, dentry);
 686                                 err = 0;
 687                         }
 688                 }
 689                 mutex_unlock(&relay_channels_mutex);
 690                 return err;
 691         }
 692
 693         curr_cpu = get_cpu();
 694         /*
 695          * The CPU hotplug notifier ran before us and created buffers with
 696          * no files associated. So it's safe to call relay_setup_buf_file()
 697          * on all currently online CPUs.
 698          */
 699         for_each_online_cpu(i) {
 700                 buf = *per_cpu_ptr(chan->buf, i);
 701                 if (unlikely(!buf)) {
 702                         WARN_ONCE(1, KERN_ERR "CPU has no buffer!\n");
 703                         err = -EINVAL;
 704                         break;
 705                 }
 706
 707                 dentry = relay_create_buf_file(chan, buf, i);
 708                 if (unlikely(!dentry)) {
 709                         err = -EINVAL;
 710                         break;
 711                 }
 712
 713                 if (curr_cpu == i) {
 714                         local_irq_save(flags);
 715                         relay_set_buf_dentry(buf, dentry);
 716                         local_irq_restore(flags);
 717                 } else {
 718                         disp.buf = buf;
 719                         disp.dentry = dentry;
 720                         smp_mb();
 721                         /* relay_channels_mutex must be held, so wait. */
 722                         err = smp_call_function_single(i,
 723                                                        __relay_set_buf_dentry,
 724                                                        &disp, 1);
 725                 }
 726                 if (unlikely(err))
 727                         break;
 728         }
 729         put_cpu();
 730         mutex_unlock(&relay_channels_mutex);
 731
 732         return err;
 733 }
 734 EXPORT_SYMBOL_GPL(relay_late_setup_files);
 735
 736 /**
 737  *      relay_switch_subbuf - switch to a new sub-buffer
 738  *      @buf: channel buffer
 739  *      @length: size of current event
 740  *
 741  *      Returns either the length passed in or 0 if full.
 742  *
 743  *      Performs sub-buffer-switch tasks such as invoking callbacks,
 744  *      updating padding counts, waking up readers, etc.
 745  */
 746 size_t relay_switch_subbuf(struct rchan_buf *buf, size_t length)
 747 {
 748         void *old, *new;
 749         size_t old_subbuf, new_subbuf;
 750
 751         if (unlikely(length > buf->chan->subbuf_size))
 752                 goto toobig;
 753
 754         if (buf->offset != buf->chan->subbuf_size + 1) {
 755                 buf->prev_padding = buf->chan->subbuf_size - buf->offset;
 756                 old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 757                 buf->padding[old_subbuf] = buf->prev_padding;
 758                 buf->subbufs_produced++;
 759                 if (buf->dentry)
 760                         d_inode(buf->dentry)->i_size +=
 761                                 buf->chan->subbuf_size -
 762                                 buf->padding[old_subbuf];
 763                 else
 764                         buf->early_bytes += buf->chan->subbuf_size -
 765                                             buf->padding[old_subbuf];
 766                 smp_mb();
 767                 if (waitqueue_active(&buf->read_wait)) {
 768                         /*
 769                          * Calling wake_up_interruptible() from here
 770                          * will deadlock if we happen to be logging
 771                          * from the scheduler (trying to re-grab
 772                          * rq->lock), so defer it.
 773                          */
 774                         irq_work_queue(&buf->wakeup_work);
 775                 }
 776         }
 777
 778         old = buf->data;
 779         new_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 780         new = buf->start + new_subbuf * buf->chan->subbuf_size;
 781         buf->offset = 0;
 782         if (!buf->chan->cb->subbuf_start(buf, new, old, buf->prev_padding)) {
 783                 buf->offset = buf->chan->subbuf_size + 1;
 784                 return 0;
 785         }
 786         buf->data = new;
 787         buf->padding[new_subbuf] = 0;
 788
 789         if (unlikely(length + buf->offset > buf->chan->subbuf_size))
 790                 goto toobig;
 791
 792         return length;
 793
 794 toobig:
 795         buf->chan->last_toobig = length;
 796         return 0;
 797 }
 798 EXPORT_SYMBOL_GPL(relay_switch_subbuf);
 799
 800 /**
 801  *      relay_subbufs_consumed - update the buffer's sub-buffers-consumed count
 802  *      @chan: the channel
 803  *      @cpu: the cpu associated with the channel buffer to update
 804  *      @subbufs_consumed: number of sub-buffers to add to current buf's count
 805  *
 806  *      Adds to the channel buffer's consumed sub-buffer count.
 807  *      subbufs_consumed should be the number of sub-buffers newly consumed,
 808  *      not the total consumed.
 809  *
 810  *      NOTE. Kernel clients don't need to call this function if the channel
 811  *      mode is 'overwrite'.
 812  */
 813 void relay_subbufs_consumed(struct rchan *chan,
 814                             unsigned int cpu,
 815                             size_t subbufs_consumed)
 816 {
 817         struct rchan_buf *buf;
 818
 819         if (!chan || cpu >= NR_CPUS)
 820                 return;
 821
 822         buf = *per_cpu_ptr(chan->buf, cpu);
 823         if (!buf || subbufs_consumed > chan->n_subbufs)
 824                 return;
 825
 826         if (subbufs_consumed > buf->subbufs_produced - buf->subbufs_consumed)
 827                 buf->subbufs_consumed = buf->subbufs_produced;
 828         else
 829                 buf->subbufs_consumed += subbufs_consumed;
 830 }
 831 EXPORT_SYMBOL_GPL(relay_subbufs_consumed);
 832
 833 /**
 834  *      relay_close - close the channel
 835  *      @chan: the channel
 836  *
 837  *      Closes all channel buffers and frees the channel.
 838  */
 839 void relay_close(struct rchan *chan)
 840 {
 841         struct rchan_buf *buf;
 842         unsigned int i;
 843
 844         if (!chan)
 845                 return;
 846
 847         mutex_lock(&relay_channels_mutex);
 848         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0)))
 849                 relay_close_buf(buf);
 850         else
 851                 for_each_possible_cpu(i)
 852                         if ((buf = *per_cpu_ptr(chan->buf, i)))
 853                                 relay_close_buf(buf);
 854
 855         if (chan->last_toobig)
 856                 printk(KERN_WARNING "relay: one or more items not logged "
 857                        "[item size (%zd) > sub-buffer size (%zd)]\n",
 858                        chan->last_toobig, chan->subbuf_size);
 859
 860         list_del(&chan->list);
 861         kref_put(&chan->kref, relay_destroy_channel);
 862         mutex_unlock(&relay_channels_mutex);
 863 }
 864 EXPORT_SYMBOL_GPL(relay_close);
 865
 866 /**
 867  *      relay_flush - close the channel
 868  *      @chan: the channel
 869  *
 870  *      Flushes all channel buffers, i.e. forces buffer switch.
 871  */
 872 void relay_flush(struct rchan *chan)
 873 {
 874         struct rchan_buf *buf;
 875         unsigned int i;
 876
 877         if (!chan)
 878                 return;
 879
 880         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 881                 relay_switch_subbuf(buf, 0);
 882                 return;
 883         }
 884
 885         mutex_lock(&relay_channels_mutex);
 886         for_each_possible_cpu(i)
 887                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 888                         relay_switch_subbuf(buf, 0);
 889         mutex_unlock(&relay_channels_mutex);
 890 }
 891 EXPORT_SYMBOL_GPL(relay_flush);
 892
 893 /**
 894  *      relay_file_open - open file op for relay files
 895  *      @inode: the inode
 896  *      @filp: the file
 897  *
 898  *      Increments the channel buffer refcount.
 899  */
 900 static int relay_file_open(struct inode *inode, struct file *filp)
 901 {
 902         struct rchan_buf *buf = inode->i_private;
 903         kref_get(&buf->kref);
 904         filp->private_data = buf;
 905
 906         return nonseekable_open(inode, filp);
 907 }
 908
 909 /**
 910  *      relay_file_mmap - mmap file op for relay files
 911  *      @filp: the file
 912  *      @vma: the vma describing what to map
 913  *
 914  *      Calls upon relay_mmap_buf() to map the file into user space.
 915  */
 916 static int relay_file_mmap(struct file *filp, struct vm_area_struct *vma)
 917 {
 918         struct rchan_buf *buf = filp->private_data;
 919         return relay_mmap_buf(buf, vma);
 920 }
 921
 922 /**
 923  *      relay_file_poll - poll file op for relay files
 924  *      @filp: the file
 925  *      @wait: poll table
 926  *
 927  *      Poll implemention.
 928  */
 929 static __poll_t relay_file_poll(struct file *filp, poll_table *wait)
 930 {
 931         __poll_t mask = 0;
 932         struct rchan_buf *buf = filp->private_data;
 933
 934         if (buf->finalized)
 935                 return EPOLLERR;
 936
 937         if (filp->f_mode & FMODE_READ) {
 938                 poll_wait(filp, &buf->read_wait, wait);
 939                 if (!relay_buf_empty(buf))
 940                         mask |= EPOLLIN | EPOLLRDNORM;
 941         }
 942
 943         return mask;
 944 }
 945
 946 /**
 947  *      relay_file_release - release file op for relay files
 948  *      @inode: the inode
 949  *      @filp: the file
 950  *
 951  *      Decrements the channel refcount, as the filesystem is
 952  *      no longer using it.
 953  */
 954 static int relay_file_release(struct inode *inode, struct file *filp)
 955 {
 956         struct rchan_buf *buf = filp->private_data;
 957         kref_put(&buf->kref, relay_remove_buf);
 958
 959         return 0;
 960 }
 961
 962 /*
 963  *      relay_file_read_consume - update the consumed count for the buffer
 964  */
 965 static void relay_file_read_consume(struct rchan_buf *buf,
 966                                     size_t read_pos,
 967                                     size_t bytes_consumed)
 968 {
 969         size_t subbuf_size = buf->chan->subbuf_size;
 970         size_t n_subbufs = buf->chan->n_subbufs;
 971         size_t read_subbuf;
 972
 973         if (buf->subbufs_produced == buf->subbufs_consumed &&
 974             buf->offset == buf->bytes_consumed)
 975                 return;
 976
 977         if (buf->bytes_consumed + bytes_consumed > subbuf_size) {
 978                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 979                 buf->bytes_consumed = 0;
 980         }
 981
 982         buf->bytes_consumed += bytes_consumed;
 983         if (!read_pos)
 984                 read_subbuf = buf->subbufs_consumed % n_subbufs;
 985         else
 986                 read_subbuf = read_pos / buf->chan->subbuf_size;
 987         if (buf->bytes_consumed + buf->padding[read_subbuf] == subbuf_size) {
 988                 if ((read_subbuf == buf->subbufs_produced % n_subbufs) &&
 989                     (buf->offset == subbuf_size))
 990                         return;
 991                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 992                 buf->bytes_consumed = 0;
 993         }
 994 }
 995
 996 /*
 997  *      relay_file_read_avail - boolean, are there unconsumed bytes available?
 998  */
 999 static int relay_file_read_avail(struct rchan_buf *buf)
1000 {
1001         size_t subbuf_size = buf->chan->subbuf_size;
1002         size_t n_subbufs = buf->chan->n_subbufs;
1003         size_t produced = buf->subbufs_produced;
1004         size_t consumed = buf->subbufs_consumed;
1005
1006         relay_file_read_consume(buf, 0, 0);
1007
1008         consumed = buf->subbufs_consumed;
1009
1010         if (unlikely(buf->offset > subbuf_size)) {
1011                 if (produced == consumed)
1012                         return 0;
1013                 return 1;
1014         }
1015
1016         if (unlikely(produced - consumed >= n_subbufs)) {
1017                 consumed = produced - n_subbufs + 1;
1018                 buf->subbufs_consumed = consumed;
1019                 buf->bytes_consumed = 0;
1020         }
1021
1022         produced = (produced % n_subbufs) * subbuf_size + buf->offset;
1023         consumed = (consumed % n_subbufs) * subbuf_size + buf->bytes_consumed;
1024
1025         if (consumed > produced)
1026                 produced += n_subbufs * subbuf_size;
1027
1028         if (consumed == produced) {
1029                 if (buf->offset == subbuf_size &&
1030                     buf->subbufs_produced > buf->subbufs_consumed)
1031                         return 1;
1032                 return 0;
1033         }
1034
1035         return 1;
1036 }
1037
1038 /**
1039  *      relay_file_read_subbuf_avail - return bytes available in sub-buffer
1040  *      @read_pos: file read position
1041  *      @buf: relay channel buffer
1042  */
1043 static size_t relay_file_read_subbuf_avail(size_t read_pos,
1044                                            struct rchan_buf *buf)
1045 {
1046         size_t padding, avail = 0;
1047         size_t read_subbuf, read_offset, write_subbuf, write_offset;
1048         size_t subbuf_size = buf->chan->subbuf_size;
1049
1050         write_subbuf = (buf->data - buf->start) / subbuf_size;
1051         write_offset = buf->offset > subbuf_size ? subbuf_size : buf->offset;
1052         read_subbuf = read_pos / subbuf_size;
1053         read_offset = read_pos % subbuf_size;
1054         padding = buf->padding[read_subbuf];
1055
1056         if (read_subbuf == write_subbuf) {
1057                 if (read_offset + padding < write_offset)
1058                         avail = write_offset - (read_offset + padding);
1059         } else
1060                 avail = (subbuf_size - padding) - read_offset;
1061
1062         return avail;
1063 }
1064
1065 /**
1066  *      relay_file_read_start_pos - find the first available byte to read
1067  *      @buf: relay channel buffer
1068  *
1069  *      If the read_pos is in the middle of padding, return the
1070  *      position of the first actually available byte, otherwise
1071  *      return the original value.
1072  */
1073 static size_t relay_file_read_start_pos(struct rchan_buf *buf)
1074 {
1075         size_t read_subbuf, padding, padding_start, padding_end;
1076         size_t subbuf_size = buf->chan->subbuf_size;
1077         size_t n_subbufs = buf->chan->n_subbufs;
1078         size_t consumed = buf->subbufs_consumed % n_subbufs;
1079         size_t read_pos = consumed * subbuf_size + buf->bytes_consumed;
1080
1081         read_subbuf = read_pos / subbuf_size;
1082         padding = buf->padding[read_subbuf];
1083         padding_start = (read_subbuf + 1) * subbuf_size - padding;
1084         padding_end = (read_subbuf + 1) * subbuf_size;
1085         if (read_pos >= padding_start && read_pos < padding_end) {
1086                 read_subbuf = (read_subbuf + 1) % n_subbufs;
1087                 read_pos = read_subbuf * subbuf_size;
1088         }
1089
1090         return read_pos;
1091 }
1092
1093 /**
1094  *      relay_file_read_end_pos - return the new read position
1095  *      @read_pos: file read position
1096  *      @buf: relay channel buffer
1097  *      @count: number of bytes to be read
1098  */
1099 static size_t relay_file_read_end_pos(struct rchan_buf *buf,
1100                                       size_t read_pos,
1101                                       size_t count)
1102 {
1103         size_t read_subbuf, padding, end_pos;
1104         size_t subbuf_size = buf->chan->subbuf_size;
1105         size_t n_subbufs = buf->chan->n_subbufs;
1106
1107         read_subbuf = read_pos / subbuf_size;
1108         padding = buf->padding[read_subbuf];
1109         if (read_pos % subbuf_size + count + padding == subbuf_size)
1110                 end_pos = (read_subbuf + 1) * subbuf_size;
1111         else
1112                 end_pos = read_pos + count;
1113         if (end_pos >= subbuf_size * n_subbufs)
1114                 end_pos = 0;
1115
1116         return end_pos;
1117 }
1118
1119 static ssize_t relay_file_read(struct file *filp,
1120                                char __user *buffer,
1121                                size_t count,
1122                                loff_t *ppos)
1123 {
1124         struct rchan_buf *buf = filp->private_data;
1125         size_t read_start, avail;
1126         size_t written = 0;
1127         int ret;
1128
1129         if (!count)
1130                 return 0;
1131
1132         inode_lock(file_inode(filp));
1133         do {
1134                 void *from;
1135
1136                 if (!relay_file_read_avail(buf))
1137                         break;
1138
1139                 read_start = relay_file_read_start_pos(buf);
1140                 avail = relay_file_read_subbuf_avail(read_start, buf);
1141                 if (!avail)
1142                         break;
1143
1144                 avail = min(count, avail);
1145                 from = buf->start + read_start;
1146                 ret = avail;
1147                 if (copy_to_user(buffer, from, avail))
1148                         break;
1149
1150                 buffer += ret;
1151                 written += ret;
1152                 count -= ret;
1153
1154                 relay_file_read_consume(buf, read_start, ret);
1155                 *ppos = relay_file_read_end_pos(buf, read_start, ret);
1156         } while (count);
1157         inode_unlock(file_inode(filp));
1158
1159         return written;
1160 }
1161
1162 static void relay_consume_bytes(struct rchan_buf *rbuf, int bytes_consumed)
1163 {
1164         rbuf->bytes_consumed += bytes_consumed;
1165
1166         if (rbuf->bytes_consumed >= rbuf->chan->subbuf_size) {
1167                 relay_subbufs_consumed(rbuf->chan, rbuf->cpu, 1);
1168                 rbuf->bytes_consumed %= rbuf->chan->subbuf_size;
1169         }
1170 }
1171
1172 static void relay_pipe_buf_release(struct pipe_inode_info *pipe,
1173                                    struct pipe_buffer *buf)
1174 {
1175         struct rchan_buf *rbuf;
1176
1177         rbuf = (struct rchan_buf *)page_private(buf->page);
1178         relay_consume_bytes(rbuf, buf->private);
1179 }
1180
1181 static const struct pipe_buf_operations relay_pipe_buf_ops = {
1182         .release        = relay_pipe_buf_release,
1183         .try_steal      = generic_pipe_buf_try_steal,
1184         .get            = generic_pipe_buf_get,
1185 };
1186
1187 static void relay_page_release(struct splice_pipe_desc *spd, unsigned int i)
1188 {
1189 }
1190
1191 /*
1192  *      subbuf_splice_actor - splice up to one subbuf's worth of data
1193  */
1194 static ssize_t subbuf_splice_actor(struct file *in,
1195                                loff_t *ppos,
1196                                struct pipe_inode_info *pipe,
1197                                size_t len,
1198                                unsigned int flags,
1199                                int *nonpad_ret)
1200 {
1201         unsigned int pidx, poff, total_len, subbuf_pages, nr_pages;
1202         struct rchan_buf *rbuf = in->private_data;
1203         unsigned int subbuf_size = rbuf->chan->subbuf_size;
1204         uint64_t pos = (uint64_t) *ppos;
1205         uint32_t alloc_size = (uint32_t) rbuf->chan->alloc_size;
1206         size_t read_start = (size_t) do_div(pos, alloc_size);
1207         size_t read_subbuf = read_start / subbuf_size;
1208         size_t padding = rbuf->padding[read_subbuf];
1209         size_t nonpad_end = read_subbuf * subbuf_size + subbuf_size - padding;
1210         struct page *pages[PIPE_DEF_BUFFERS];
1211         struct partial_page partial[PIPE_DEF_BUFFERS];
1212         struct splice_pipe_desc spd = {
1213                 .pages = pages,
1214                 .nr_pages = 0,
1215                 .nr_pages_max = PIPE_DEF_BUFFERS,
1216                 .partial = partial,
1217                 .ops = &relay_pipe_buf_ops,
1218                 .spd_release = relay_page_release,
1219         };
1220         ssize_t ret;
1221
1222         if (rbuf->subbufs_produced == rbuf->subbufs_consumed)
1223                 return 0;
1224         if (splice_grow_spd(pipe, &spd))
1225                 return -ENOMEM;
1226
1227         /*
1228          * Adjust read len, if longer than what is available
1229          */
1230         if (len > (subbuf_size - read_start % subbuf_size))
1231                 len = subbuf_size - read_start % subbuf_size;
1232
1233         subbuf_pages = rbuf->chan->alloc_size >> PAGE_SHIFT;
1234         pidx = (read_start / PAGE_SIZE) % subbuf_pages;
1235         poff = read_start & ~PAGE_MASK;
1236         nr_pages = min_t(unsigned int, subbuf_pages, spd.nr_pages_max);
1237
1238         for (total_len = 0; spd.nr_pages < nr_pages; spd.nr_pages++) {
1239                 unsigned int this_len, this_end, private;
1240                 unsigned int cur_pos = read_start + total_len;
1241
1242                 if (!len)
1243                         break;
1244
1245                 this_len = min_t(unsigned long, len, PAGE_SIZE - poff);
1246                 private = this_len;
1247
1248                 spd.pages[spd.nr_pages] = rbuf->page_array[pidx];
1249                 spd.partial[spd.nr_pages].offset = poff;
1250
1251                 this_end = cur_pos + this_len;
1252                 if (this_end >= nonpad_end) {
1253                         this_len = nonpad_end - cur_pos;
1254                         private = this_len + padding;
1255                 }
1256                 spd.partial[spd.nr_pages].len = this_len;
1257                 spd.partial[spd.nr_pages].private = private;
1258
1259                 len -= this_len;
1260                 total_len += this_len;
1261                 poff = 0;
1262                 pidx = (pidx + 1) % subbuf_pages;
1263
1264                 if (this_end >= nonpad_end) {
1265                         spd.nr_pages++;
1266                         break;
1267                 }
1268         }
1269
1270         ret = 0;
1271         if (!spd.nr_pages)
1272                 goto out;
1273
1274         ret = *nonpad_ret = splice_to_pipe(pipe, &spd);
1275         if (ret < 0 || ret < total_len)
1276                 goto out;
1277
1278         if (read_start + ret == nonpad_end)
1279                 ret += padding;
1280
1281 out:
1282         splice_shrink_spd(&spd);
1283         return ret;
1284 }
1285
1286 static ssize_t relay_file_splice_read(struct file *in,
1287                                       loff_t *ppos,
1288                                       struct pipe_inode_info *pipe,
1289                                       size_t len,
1290                                       unsigned int flags)
1291 {
1292         ssize_t spliced;
1293         int ret;
1294         int nonpad_ret = 0;
1295
1296         ret = 0;
1297         spliced = 0;
1298
1299         while (len && !spliced) {
1300                 ret = subbuf_splice_actor(in, ppos, pipe, len, flags, &nonpad_ret);
1301                 if (ret < 0)
1302                         break;
1303                 else if (!ret) {
1304                         if (flags & SPLICE_F_NONBLOCK)
1305                                 ret = -EAGAIN;
1306                         break;
1307                 }
1308
1309                 *ppos += ret;
1310                 if (ret > len)
1311                         len = 0;
1312                 else
1313                         len -= ret;
1314                 spliced += nonpad_ret;
1315                 nonpad_ret = 0;
1316         }
1317
1318         if (spliced)
1319                 return spliced;
1320
1321         return ret;
1322 }
1323
1324 const struct file_operations relay_file_operations = {
1325         .open           = relay_file_open,
1326         .poll           = relay_file_poll,
1327         .mmap           = relay_file_mmap,
1328         .read           = relay_file_read,
1329         .llseek         = no_llseek,
1330         .release        = relay_file_release,
1331         .splice_read    = relay_file_splice_read,
1332 };
1333 EXPORT_SYMBOL_GPL(relay_file_operations);