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         free_percpu(chan->buf);
 201         kfree(chan);
 202 }
 203
 204 /**
 205  *      relay_destroy_buf - destroy an rchan_buf struct and associated buffer
 206  *      @buf: the buffer struct
 207  */
 208 static void relay_destroy_buf(struct rchan_buf *buf)
 209 {
 210         struct rchan *chan = buf->chan;
 211         unsigned int i;
 212
 213         if (likely(buf->start)) {
 214                 vunmap(buf->start);
 215                 for (i = 0; i < buf->page_count; i++)
 216                         __free_page(buf->page_array[i]);
 217                 relay_free_page_array(buf->page_array);
 218         }
 219         *per_cpu_ptr(chan->buf, buf->cpu) = NULL;
 220         kfree(buf->padding);
 221         kfree(buf);
 222         kref_put(&chan->kref, relay_destroy_channel);
 223 }
 224
 225 /**
 226  *      relay_remove_buf - remove a channel buffer
 227  *      @kref: target kernel reference that contains the relay buffer
 228  *
 229  *      Removes the file from the filesystem, which also frees the
 230  *      rchan_buf_struct and the channel buffer.  Should only be called from
 231  *      kref_put().
 232  */
 233 static void relay_remove_buf(struct kref *kref)
 234 {
 235         struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
 236         relay_destroy_buf(buf);
 237 }
 238
 239 /**
 240  *      relay_buf_empty - boolean, is the channel buffer empty?
 241  *      @buf: channel buffer
 242  *
 243  *      Returns 1 if the buffer is empty, 0 otherwise.
 244  */
 245 static int relay_buf_empty(struct rchan_buf *buf)
 246 {
 247         return (buf->subbufs_produced - buf->subbufs_consumed) ? 0 : 1;
 248 }
 249
 250 /**
 251  *      relay_buf_full - boolean, is the channel buffer full?
 252  *      @buf: channel buffer
 253  *
 254  *      Returns 1 if the buffer is full, 0 otherwise.
 255  */
 256 int relay_buf_full(struct rchan_buf *buf)
 257 {
 258         size_t ready = buf->subbufs_produced - buf->subbufs_consumed;
 259         return (ready >= buf->chan->n_subbufs) ? 1 : 0;
 260 }
 261 EXPORT_SYMBOL_GPL(relay_buf_full);
 262
 263 /*
 264  * High-level relay kernel API and associated functions.
 265  */
 266
 267 /*
 268  * rchan_callback implementations defining default channel behavior.  Used
 269  * in place of corresponding NULL values in client callback struct.
 270  */
 271
 272 /*
 273  * subbuf_start() default callback.  Does nothing.
 274  */
 275 static int subbuf_start_default_callback (struct rchan_buf *buf,
 276                                           void *subbuf,
 277                                           void *prev_subbuf,
 278                                           size_t prev_padding)
 279 {
 280         if (relay_buf_full(buf))
 281                 return 0;
 282
 283         return 1;
 284 }
 285
 286 /*
 287  * buf_mapped() default callback.  Does nothing.
 288  */
 289 static void buf_mapped_default_callback(struct rchan_buf *buf,
 290                                         struct file *filp)
 291 {
 292 }
 293
 294 /*
 295  * buf_unmapped() default callback.  Does nothing.
 296  */
 297 static void buf_unmapped_default_callback(struct rchan_buf *buf,
 298                                           struct file *filp)
 299 {
 300 }
 301
 302 /*
 303  * create_buf_file_create() default callback.  Does nothing.
 304  */
 305 static struct dentry *create_buf_file_default_callback(const char *filename,
 306                                                        struct dentry *parent,
 307                                                        umode_t mode,
 308                                                        struct rchan_buf *buf,
 309                                                        int *is_global)
 310 {
 311         return NULL;
 312 }
 313
 314 /*
 315  * remove_buf_file() default callback.  Does nothing.
 316  */
 317 static int remove_buf_file_default_callback(struct dentry *dentry)
 318 {
 319         return -EINVAL;
 320 }
 321
 322 /* relay channel default callbacks */
 323 static struct rchan_callbacks default_channel_callbacks = {
 324         .subbuf_start = subbuf_start_default_callback,
 325         .buf_mapped = buf_mapped_default_callback,
 326         .buf_unmapped = buf_unmapped_default_callback,
 327         .create_buf_file = create_buf_file_default_callback,
 328         .remove_buf_file = remove_buf_file_default_callback,
 329 };
 330
 331 /**
 332  *      wakeup_readers - wake up readers waiting on a channel
 333  *      @work: contains the channel buffer
 334  *
 335  *      This is the function used to defer reader waking
 336  */
 337 static void wakeup_readers(struct irq_work *work)
 338 {
 339         struct rchan_buf *buf;
 340
 341         buf = container_of(work, struct rchan_buf, wakeup_work);
 342         wake_up_interruptible(&buf->read_wait);
 343 }
 344
 345 /**
 346  *      __relay_reset - reset a channel buffer
 347  *      @buf: the channel buffer
 348  *      @init: 1 if this is a first-time initialization
 349  *
 350  *      See relay_reset() for description of effect.
 351  */
 352 static void __relay_reset(struct rchan_buf *buf, unsigned int init)
 353 {
 354         size_t i;
 355
 356         if (init) {
 357                 init_waitqueue_head(&buf->read_wait);
 358                 kref_init(&buf->kref);
 359                 init_irq_work(&buf->wakeup_work, wakeup_readers);
 360         } else {
 361                 irq_work_sync(&buf->wakeup_work);
 362         }
 363
 364         buf->subbufs_produced = 0;
 365         buf->subbufs_consumed = 0;
 366         buf->bytes_consumed = 0;
 367         buf->finalized = 0;
 368         buf->data = buf->start;
 369         buf->offset = 0;
 370
 371         for (i = 0; i < buf->chan->n_subbufs; i++)
 372                 buf->padding[i] = 0;
 373
 374         buf->chan->cb->subbuf_start(buf, buf->data, NULL, 0);
 375 }
 376
 377 /**
 378  *      relay_reset - reset the channel
 379  *      @chan: the channel
 380  *
 381  *      This has the effect of erasing all data from all channel buffers
 382  *      and restarting the channel in its initial state.  The buffers
 383  *      are not freed, so any mappings are still in effect.
 384  *
 385  *      NOTE. Care should be taken that the channel isn't actually
 386  *      being used by anything when this call is made.
 387  */
 388 void relay_reset(struct rchan *chan)
 389 {
 390         struct rchan_buf *buf;
 391         unsigned int i;
 392
 393         if (!chan)
 394                 return;
 395
 396         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 397                 __relay_reset(buf, 0);
 398                 return;
 399         }
 400
 401         mutex_lock(&relay_channels_mutex);
 402         for_each_possible_cpu(i)
 403                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 404                         __relay_reset(buf, 0);
 405         mutex_unlock(&relay_channels_mutex);
 406 }
 407 EXPORT_SYMBOL_GPL(relay_reset);
 408
 409 static inline void relay_set_buf_dentry(struct rchan_buf *buf,
 410                                         struct dentry *dentry)
 411 {
 412         buf->dentry = dentry;
 413         d_inode(buf->dentry)->i_size = buf->early_bytes;
 414 }
 415
 416 static struct dentry *relay_create_buf_file(struct rchan *chan,
 417                                             struct rchan_buf *buf,
 418                                             unsigned int cpu)
 419 {
 420         struct dentry *dentry;
 421         char *tmpname;
 422
 423         tmpname = kzalloc(NAME_MAX + 1, GFP_KERNEL);
 424         if (!tmpname)
 425                 return NULL;
 426         snprintf(tmpname, NAME_MAX, "%s%d", chan->base_filename, cpu);
 427
 428         /* Create file in fs */
 429         dentry = chan->cb->create_buf_file(tmpname, chan->parent,
 430                                            S_IRUSR, buf,
 431                                            &chan->is_global);
 432         if (IS_ERR(dentry))
 433                 dentry = NULL;
 434
 435         kfree(tmpname);
 436
 437         return dentry;
 438 }
 439
 440 /*
 441  *      relay_open_buf - create a new relay channel buffer
 442  *
 443  *      used by relay_open() and CPU hotplug.
 444  */
 445 static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
 446 {
 447         struct rchan_buf *buf = NULL;
 448         struct dentry *dentry;
 449
 450         if (chan->is_global)
 451                 return *per_cpu_ptr(chan->buf, 0);
 452
 453         buf = relay_create_buf(chan);
 454         if (!buf)
 455                 return NULL;
 456
 457         if (chan->has_base_filename) {
 458                 dentry = relay_create_buf_file(chan, buf, cpu);
 459                 if (!dentry)
 460                         goto free_buf;
 461                 relay_set_buf_dentry(buf, dentry);
 462         } else {
 463                 /* Only retrieve global info, nothing more, nothing less */
 464                 dentry = chan->cb->create_buf_file(NULL, NULL,
 465                                                    S_IRUSR, buf,
 466                                                    &chan->is_global);
 467                 if (IS_ERR_OR_NULL(dentry))
 468                         goto free_buf;
 469         }
 470
 471         buf->cpu = cpu;
 472         __relay_reset(buf, 1);
 473
 474         if(chan->is_global) {
 475                 *per_cpu_ptr(chan->buf, 0) = buf;
 476                 buf->cpu = 0;
 477         }
 478
 479         return buf;
 480
 481 free_buf:
 482         relay_destroy_buf(buf);
 483         return NULL;
 484 }
 485
 486 /**
 487  *      relay_close_buf - close a channel buffer
 488  *      @buf: channel buffer
 489  *
 490  *      Marks the buffer finalized and restores the default callbacks.
 491  *      The channel buffer and channel buffer data structure are then freed
 492  *      automatically when the last reference is given up.
 493  */
 494 static void relay_close_buf(struct rchan_buf *buf)
 495 {
 496         buf->finalized = 1;
 497         irq_work_sync(&buf->wakeup_work);
 498         buf->chan->cb->remove_buf_file(buf->dentry);
 499         kref_put(&buf->kref, relay_remove_buf);
 500 }
 501
 502 static void setup_callbacks(struct rchan *chan,
 503                                    struct rchan_callbacks *cb)
 504 {
 505         if (!cb) {
 506                 chan->cb = &default_channel_callbacks;
 507                 return;
 508         }
 509
 510         if (!cb->subbuf_start)
 511                 cb->subbuf_start = subbuf_start_default_callback;
 512         if (!cb->buf_mapped)
 513                 cb->buf_mapped = buf_mapped_default_callback;
 514         if (!cb->buf_unmapped)
 515                 cb->buf_unmapped = buf_unmapped_default_callback;
 516         if (!cb->create_buf_file)
 517                 cb->create_buf_file = create_buf_file_default_callback;
 518         if (!cb->remove_buf_file)
 519                 cb->remove_buf_file = remove_buf_file_default_callback;
 520         chan->cb = cb;
 521 }
 522
 523 int relay_prepare_cpu(unsigned int cpu)
 524 {
 525         struct rchan *chan;
 526         struct rchan_buf *buf;
 527
 528         mutex_lock(&relay_channels_mutex);
 529         list_for_each_entry(chan, &relay_channels, list) {
 530                 if ((buf = *per_cpu_ptr(chan->buf, cpu)))
 531                         continue;
 532                 buf = relay_open_buf(chan, cpu);
 533                 if (!buf) {
 534                         pr_err("relay: cpu %d buffer creation failed\n", cpu);
 535                         mutex_unlock(&relay_channels_mutex);
 536                         return -ENOMEM;
 537                 }
 538                 *per_cpu_ptr(chan->buf, cpu) = buf;
 539         }
 540         mutex_unlock(&relay_channels_mutex);
 541         return 0;
 542 }
 543
 544 /**
 545  *      relay_open - create a new relay channel
 546  *      @base_filename: base name of files to create, %NULL for buffering only
 547  *      @parent: dentry of parent directory, %NULL for root directory or buffer
 548  *      @subbuf_size: size of sub-buffers
 549  *      @n_subbufs: number of sub-buffers
 550  *      @cb: client callback functions
 551  *      @private_data: user-defined data
 552  *
 553  *      Returns channel pointer if successful, %NULL otherwise.
 554  *
 555  *      Creates a channel buffer for each cpu using the sizes and
 556  *      attributes specified.  The created channel buffer files
 557  *      will be named base_filename0...base_filenameN-1.  File
 558  *      permissions will be %S_IRUSR.
 559  *
 560  *      If opening a buffer (@parent = NULL) that you later wish to register
 561  *      in a filesystem, call relay_late_setup_files() once the @parent dentry
 562  *      is available.
 563  */
 564 struct rchan *relay_open(const char *base_filename,
 565                          struct dentry *parent,
 566                          size_t subbuf_size,
 567                          size_t n_subbufs,
 568                          struct rchan_callbacks *cb,
 569                          void *private_data)
 570 {
 571         unsigned int i;
 572         struct rchan *chan;
 573         struct rchan_buf *buf;
 574
 575         if (!(subbuf_size && n_subbufs))
 576                 return NULL;
 577         if (subbuf_size > UINT_MAX / n_subbufs)
 578                 return NULL;
 579
 580         chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
 581         if (!chan)
 582                 return NULL;
 583
 584         chan->buf = alloc_percpu(struct rchan_buf *);
 585         if (!chan->buf) {
 586                 kfree(chan);
 587                 return NULL;
 588         }
 589
 590         chan->version = RELAYFS_CHANNEL_VERSION;
 591         chan->n_subbufs = n_subbufs;
 592         chan->subbuf_size = subbuf_size;
 593         chan->alloc_size = PAGE_ALIGN(subbuf_size * n_subbufs);
 594         chan->parent = parent;
 595         chan->private_data = private_data;
 596         if (base_filename) {
 597                 chan->has_base_filename = 1;
 598                 strlcpy(chan->base_filename, base_filename, NAME_MAX);
 599         }
 600         setup_callbacks(chan, cb);
 601         kref_init(&chan->kref);
 602
 603         mutex_lock(&relay_channels_mutex);
 604         for_each_online_cpu(i) {
 605                 buf = relay_open_buf(chan, i);
 606                 if (!buf)
 607                         goto free_bufs;
 608                 *per_cpu_ptr(chan->buf, i) = buf;
 609         }
 610         list_add(&chan->list, &relay_channels);
 611         mutex_unlock(&relay_channels_mutex);
 612
 613         return chan;
 614
 615 free_bufs:
 616         for_each_possible_cpu(i) {
 617                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 618                         relay_close_buf(buf);
 619         }
 620
 621         kref_put(&chan->kref, relay_destroy_channel);
 622         mutex_unlock(&relay_channels_mutex);
 623         return NULL;
 624 }
 625 EXPORT_SYMBOL_GPL(relay_open);
 626
 627 struct rchan_percpu_buf_dispatcher {
 628         struct rchan_buf *buf;
 629         struct dentry *dentry;
 630 };
 631
 632 /* Called in atomic context. */
 633 static void __relay_set_buf_dentry(void *info)
 634 {
 635         struct rchan_percpu_buf_dispatcher *p = info;
 636
 637         relay_set_buf_dentry(p->buf, p->dentry);
 638 }
 639
 640 /**
 641  *      relay_late_setup_files - triggers file creation
 642  *      @chan: channel to operate on
 643  *      @base_filename: base name of files to create
 644  *      @parent: dentry of parent directory, %NULL for root directory
 645  *
 646  *      Returns 0 if successful, non-zero otherwise.
 647  *
 648  *      Use to setup files for a previously buffer-only channel created
 649  *      by relay_open() with a NULL parent dentry.
 650  *
 651  *      For example, this is useful for perfomring early tracing in kernel,
 652  *      before VFS is up and then exposing the early results once the dentry
 653  *      is available.
 654  */
 655 int relay_late_setup_files(struct rchan *chan,
 656                            const char *base_filename,
 657                            struct dentry *parent)
 658 {
 659         int err = 0;
 660         unsigned int i, curr_cpu;
 661         unsigned long flags;
 662         struct dentry *dentry;
 663         struct rchan_buf *buf;
 664         struct rchan_percpu_buf_dispatcher disp;
 665
 666         if (!chan || !base_filename)
 667                 return -EINVAL;
 668
 669         strlcpy(chan->base_filename, base_filename, NAME_MAX);
 670
 671         mutex_lock(&relay_channels_mutex);
 672         /* Is chan already set up? */
 673         if (unlikely(chan->has_base_filename)) {
 674                 mutex_unlock(&relay_channels_mutex);
 675                 return -EEXIST;
 676         }
 677         chan->has_base_filename = 1;
 678         chan->parent = parent;
 679
 680         if (chan->is_global) {
 681                 err = -EINVAL;
 682                 buf = *per_cpu_ptr(chan->buf, 0);
 683                 if (!WARN_ON_ONCE(!buf)) {
 684                         dentry = relay_create_buf_file(chan, buf, 0);
 685                         if (dentry && !WARN_ON_ONCE(!chan->is_global)) {
 686                                 relay_set_buf_dentry(buf, dentry);
 687                                 err = 0;
 688                         }
 689                 }
 690                 mutex_unlock(&relay_channels_mutex);
 691                 return err;
 692         }
 693
 694         curr_cpu = get_cpu();
 695         /*
 696          * The CPU hotplug notifier ran before us and created buffers with
 697          * no files associated. So it's safe to call relay_setup_buf_file()
 698          * on all currently online CPUs.
 699          */
 700         for_each_online_cpu(i) {
 701                 buf = *per_cpu_ptr(chan->buf, i);
 702                 if (unlikely(!buf)) {
 703                         WARN_ONCE(1, KERN_ERR "CPU has no buffer!\n");
 704                         err = -EINVAL;
 705                         break;
 706                 }
 707
 708                 dentry = relay_create_buf_file(chan, buf, i);
 709                 if (unlikely(!dentry)) {
 710                         err = -EINVAL;
 711                         break;
 712                 }
 713
 714                 if (curr_cpu == i) {
 715                         local_irq_save(flags);
 716                         relay_set_buf_dentry(buf, dentry);
 717                         local_irq_restore(flags);
 718                 } else {
 719                         disp.buf = buf;
 720                         disp.dentry = dentry;
 721                         smp_mb();
 722                         /* relay_channels_mutex must be held, so wait. */
 723                         err = smp_call_function_single(i,
 724                                                        __relay_set_buf_dentry,
 725                                                        &disp, 1);
 726                 }
 727                 if (unlikely(err))
 728                         break;
 729         }
 730         put_cpu();
 731         mutex_unlock(&relay_channels_mutex);
 732
 733         return err;
 734 }
 735 EXPORT_SYMBOL_GPL(relay_late_setup_files);
 736
 737 /**
 738  *      relay_switch_subbuf - switch to a new sub-buffer
 739  *      @buf: channel buffer
 740  *      @length: size of current event
 741  *
 742  *      Returns either the length passed in or 0 if full.
 743  *
 744  *      Performs sub-buffer-switch tasks such as invoking callbacks,
 745  *      updating padding counts, waking up readers, etc.
 746  */
 747 size_t relay_switch_subbuf(struct rchan_buf *buf, size_t length)
 748 {
 749         void *old, *new;
 750         size_t old_subbuf, new_subbuf;
 751
 752         if (unlikely(length > buf->chan->subbuf_size))
 753                 goto toobig;
 754
 755         if (buf->offset != buf->chan->subbuf_size + 1) {
 756                 buf->prev_padding = buf->chan->subbuf_size - buf->offset;
 757                 old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 758                 buf->padding[old_subbuf] = buf->prev_padding;
 759                 buf->subbufs_produced++;
 760                 if (buf->dentry)
 761                         d_inode(buf->dentry)->i_size +=
 762                                 buf->chan->subbuf_size -
 763                                 buf->padding[old_subbuf];
 764                 else
 765                         buf->early_bytes += buf->chan->subbuf_size -
 766                                             buf->padding[old_subbuf];
 767                 smp_mb();
 768                 if (waitqueue_active(&buf->read_wait)) {
 769                         /*
 770                          * Calling wake_up_interruptible() from here
 771                          * will deadlock if we happen to be logging
 772                          * from the scheduler (trying to re-grab
 773                          * rq->lock), so defer it.
 774                          */
 775                         irq_work_queue(&buf->wakeup_work);
 776                 }
 777         }
 778
 779         old = buf->data;
 780         new_subbuf = buf->subbufs_produced % buf->chan->n_subbufs;
 781         new = buf->start + new_subbuf * buf->chan->subbuf_size;
 782         buf->offset = 0;
 783         if (!buf->chan->cb->subbuf_start(buf, new, old, buf->prev_padding)) {
 784                 buf->offset = buf->chan->subbuf_size + 1;
 785                 return 0;
 786         }
 787         buf->data = new;
 788         buf->padding[new_subbuf] = 0;
 789
 790         if (unlikely(length + buf->offset > buf->chan->subbuf_size))
 791                 goto toobig;
 792
 793         return length;
 794
 795 toobig:
 796         buf->chan->last_toobig = length;
 797         return 0;
 798 }
 799 EXPORT_SYMBOL_GPL(relay_switch_subbuf);
 800
 801 /**
 802  *      relay_subbufs_consumed - update the buffer's sub-buffers-consumed count
 803  *      @chan: the channel
 804  *      @cpu: the cpu associated with the channel buffer to update
 805  *      @subbufs_consumed: number of sub-buffers to add to current buf's count
 806  *
 807  *      Adds to the channel buffer's consumed sub-buffer count.
 808  *      subbufs_consumed should be the number of sub-buffers newly consumed,
 809  *      not the total consumed.
 810  *
 811  *      NOTE. Kernel clients don't need to call this function if the channel
 812  *      mode is 'overwrite'.
 813  */
 814 void relay_subbufs_consumed(struct rchan *chan,
 815                             unsigned int cpu,
 816                             size_t subbufs_consumed)
 817 {
 818         struct rchan_buf *buf;
 819
 820         if (!chan || cpu >= NR_CPUS)
 821                 return;
 822
 823         buf = *per_cpu_ptr(chan->buf, cpu);
 824         if (!buf || subbufs_consumed > chan->n_subbufs)
 825                 return;
 826
 827         if (subbufs_consumed > buf->subbufs_produced - buf->subbufs_consumed)
 828                 buf->subbufs_consumed = buf->subbufs_produced;
 829         else
 830                 buf->subbufs_consumed += subbufs_consumed;
 831 }
 832 EXPORT_SYMBOL_GPL(relay_subbufs_consumed);
 833
 834 /**
 835  *      relay_close - close the channel
 836  *      @chan: the channel
 837  *
 838  *      Closes all channel buffers and frees the channel.
 839  */
 840 void relay_close(struct rchan *chan)
 841 {
 842         struct rchan_buf *buf;
 843         unsigned int i;
 844
 845         if (!chan)
 846                 return;
 847
 848         mutex_lock(&relay_channels_mutex);
 849         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0)))
 850                 relay_close_buf(buf);
 851         else
 852                 for_each_possible_cpu(i)
 853                         if ((buf = *per_cpu_ptr(chan->buf, i)))
 854                                 relay_close_buf(buf);
 855
 856         if (chan->last_toobig)
 857                 printk(KERN_WARNING "relay: one or more items not logged "
 858                        "[item size (%zd) > sub-buffer size (%zd)]\n",
 859                        chan->last_toobig, chan->subbuf_size);
 860
 861         list_del(&chan->list);
 862         kref_put(&chan->kref, relay_destroy_channel);
 863         mutex_unlock(&relay_channels_mutex);
 864 }
 865 EXPORT_SYMBOL_GPL(relay_close);
 866
 867 /**
 868  *      relay_flush - close the channel
 869  *      @chan: the channel
 870  *
 871  *      Flushes all channel buffers, i.e. forces buffer switch.
 872  */
 873 void relay_flush(struct rchan *chan)
 874 {
 875         struct rchan_buf *buf;
 876         unsigned int i;
 877
 878         if (!chan)
 879                 return;
 880
 881         if (chan->is_global && (buf = *per_cpu_ptr(chan->buf, 0))) {
 882                 relay_switch_subbuf(buf, 0);
 883                 return;
 884         }
 885
 886         mutex_lock(&relay_channels_mutex);
 887         for_each_possible_cpu(i)
 888                 if ((buf = *per_cpu_ptr(chan->buf, i)))
 889                         relay_switch_subbuf(buf, 0);
 890         mutex_unlock(&relay_channels_mutex);
 891 }
 892 EXPORT_SYMBOL_GPL(relay_flush);
 893
 894 /**
 895  *      relay_file_open - open file op for relay files
 896  *      @inode: the inode
 897  *      @filp: the file
 898  *
 899  *      Increments the channel buffer refcount.
 900  */
 901 static int relay_file_open(struct inode *inode, struct file *filp)
 902 {
 903         struct rchan_buf *buf = inode->i_private;
 904         kref_get(&buf->kref);
 905         filp->private_data = buf;
 906
 907         return nonseekable_open(inode, filp);
 908 }
 909
 910 /**
 911  *      relay_file_mmap - mmap file op for relay files
 912  *      @filp: the file
 913  *      @vma: the vma describing what to map
 914  *
 915  *      Calls upon relay_mmap_buf() to map the file into user space.
 916  */
 917 static int relay_file_mmap(struct file *filp, struct vm_area_struct *vma)
 918 {
 919         struct rchan_buf *buf = filp->private_data;
 920         return relay_mmap_buf(buf, vma);
 921 }
 922
 923 /**
 924  *      relay_file_poll - poll file op for relay files
 925  *      @filp: the file
 926  *      @wait: poll table
 927  *
 928  *      Poll implemention.
 929  */
 930 static __poll_t relay_file_poll(struct file *filp, poll_table *wait)
 931 {
 932         __poll_t mask = 0;
 933         struct rchan_buf *buf = filp->private_data;
 934
 935         if (buf->finalized)
 936                 return EPOLLERR;
 937
 938         if (filp->f_mode & FMODE_READ) {
 939                 poll_wait(filp, &buf->read_wait, wait);
 940                 if (!relay_buf_empty(buf))
 941                         mask |= EPOLLIN | EPOLLRDNORM;
 942         }
 943
 944         return mask;
 945 }
 946
 947 /**
 948  *      relay_file_release - release file op for relay files
 949  *      @inode: the inode
 950  *      @filp: the file
 951  *
 952  *      Decrements the channel refcount, as the filesystem is
 953  *      no longer using it.
 954  */
 955 static int relay_file_release(struct inode *inode, struct file *filp)
 956 {
 957         struct rchan_buf *buf = filp->private_data;
 958         kref_put(&buf->kref, relay_remove_buf);
 959
 960         return 0;
 961 }
 962
 963 /*
 964  *      relay_file_read_consume - update the consumed count for the buffer
 965  */
 966 static void relay_file_read_consume(struct rchan_buf *buf,
 967                                     size_t read_pos,
 968                                     size_t bytes_consumed)
 969 {
 970         size_t subbuf_size = buf->chan->subbuf_size;
 971         size_t n_subbufs = buf->chan->n_subbufs;
 972         size_t read_subbuf;
 973
 974         if (buf->subbufs_produced == buf->subbufs_consumed &&
 975             buf->offset == buf->bytes_consumed)
 976                 return;
 977
 978         if (buf->bytes_consumed + bytes_consumed > subbuf_size) {
 979                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 980                 buf->bytes_consumed = 0;
 981         }
 982
 983         buf->bytes_consumed += bytes_consumed;
 984         if (!read_pos)
 985                 read_subbuf = buf->subbufs_consumed % n_subbufs;
 986         else
 987                 read_subbuf = read_pos / buf->chan->subbuf_size;
 988         if (buf->bytes_consumed + buf->padding[read_subbuf] == subbuf_size) {
 989                 if ((read_subbuf == buf->subbufs_produced % n_subbufs) &&
 990                     (buf->offset == subbuf_size))
 991                         return;
 992                 relay_subbufs_consumed(buf->chan, buf->cpu, 1);
 993                 buf->bytes_consumed = 0;
 994         }
 995 }
 996
 997 /*
 998  *      relay_file_read_avail - boolean, are there unconsumed bytes available?
 999  */
1000 static int relay_file_read_avail(struct rchan_buf *buf)
1001 {
1002         size_t subbuf_size = buf->chan->subbuf_size;
1003         size_t n_subbufs = buf->chan->n_subbufs;
1004         size_t produced = buf->subbufs_produced;
1005         size_t consumed;
1006
1007         relay_file_read_consume(buf, 0, 0);
1008
1009         consumed = buf->subbufs_consumed;
1010
1011         if (unlikely(buf->offset > subbuf_size)) {
1012                 if (produced == consumed)
1013                         return 0;
1014                 return 1;
1015         }
1016
1017         if (unlikely(produced - consumed >= n_subbufs)) {
1018                 consumed = produced - n_subbufs + 1;
1019                 buf->subbufs_consumed = consumed;
1020                 buf->bytes_consumed = 0;
1021         }
1022
1023         produced = (produced % n_subbufs) * subbuf_size + buf->offset;
1024         consumed = (consumed % n_subbufs) * subbuf_size + buf->bytes_consumed;
1025
1026         if (consumed > produced)
1027                 produced += n_subbufs * subbuf_size;
1028
1029         if (consumed == produced) {
1030                 if (buf->offset == subbuf_size &&
1031                     buf->subbufs_produced > buf->subbufs_consumed)
1032                         return 1;
1033                 return 0;
1034         }
1035
1036         return 1;
1037 }
1038
1039 /**
1040  *      relay_file_read_subbuf_avail - return bytes available in sub-buffer
1041  *      @read_pos: file read position
1042  *      @buf: relay channel buffer
1043  */
1044 static size_t relay_file_read_subbuf_avail(size_t read_pos,
1045                                            struct rchan_buf *buf)
1046 {
1047         size_t padding, avail = 0;
1048         size_t read_subbuf, read_offset, write_subbuf, write_offset;
1049         size_t subbuf_size = buf->chan->subbuf_size;
1050
1051         write_subbuf = (buf->data - buf->start) / subbuf_size;
1052         write_offset = buf->offset > subbuf_size ? subbuf_size : buf->offset;
1053         read_subbuf = read_pos / subbuf_size;
1054         read_offset = read_pos % subbuf_size;
1055         padding = buf->padding[read_subbuf];
1056
1057         if (read_subbuf == write_subbuf) {
1058                 if (read_offset + padding < write_offset)
1059                         avail = write_offset - (read_offset + padding);
1060         } else
1061                 avail = (subbuf_size - padding) - read_offset;
1062
1063         return avail;
1064 }
1065
1066 /**
1067  *      relay_file_read_start_pos - find the first available byte to read
1068  *      @buf: relay channel buffer
1069  *
1070  *      If the read_pos is in the middle of padding, return the
1071  *      position of the first actually available byte, otherwise
1072  *      return the original value.
1073  */
1074 static size_t relay_file_read_start_pos(struct rchan_buf *buf)
1075 {
1076         size_t read_subbuf, padding, padding_start, padding_end;
1077         size_t subbuf_size = buf->chan->subbuf_size;
1078         size_t n_subbufs = buf->chan->n_subbufs;
1079         size_t consumed = buf->subbufs_consumed % n_subbufs;
1080         size_t read_pos = consumed * subbuf_size + buf->bytes_consumed;
1081
1082         read_subbuf = read_pos / subbuf_size;
1083         padding = buf->padding[read_subbuf];
1084         padding_start = (read_subbuf + 1) * subbuf_size - padding;
1085         padding_end = (read_subbuf + 1) * subbuf_size;
1086         if (read_pos >= padding_start && read_pos < padding_end) {
1087                 read_subbuf = (read_subbuf + 1) % n_subbufs;
1088                 read_pos = read_subbuf * subbuf_size;
1089         }
1090
1091         return read_pos;
1092 }
1093
1094 /**
1095  *      relay_file_read_end_pos - return the new read position
1096  *      @read_pos: file read position
1097  *      @buf: relay channel buffer
1098  *      @count: number of bytes to be read
1099  */
1100 static size_t relay_file_read_end_pos(struct rchan_buf *buf,
1101                                       size_t read_pos,
1102                                       size_t count)
1103 {
1104         size_t read_subbuf, padding, end_pos;
1105         size_t subbuf_size = buf->chan->subbuf_size;
1106         size_t n_subbufs = buf->chan->n_subbufs;
1107
1108         read_subbuf = read_pos / subbuf_size;
1109         padding = buf->padding[read_subbuf];
1110         if (read_pos % subbuf_size + count + padding == subbuf_size)
1111                 end_pos = (read_subbuf + 1) * subbuf_size;
1112         else
1113                 end_pos = read_pos + count;
1114         if (end_pos >= subbuf_size * n_subbufs)
1115                 end_pos = 0;
1116
1117         return end_pos;
1118 }
1119
1120 static ssize_t relay_file_read(struct file *filp,
1121                                char __user *buffer,
1122                                size_t count,
1123                                loff_t *ppos)
1124 {
1125         struct rchan_buf *buf = filp->private_data;
1126         size_t read_start, avail;
1127         size_t written = 0;
1128         int ret;
1129
1130         if (!count)
1131                 return 0;
1132
1133         inode_lock(file_inode(filp));
1134         do {
1135                 void *from;
1136
1137                 if (!relay_file_read_avail(buf))
1138                         break;
1139
1140                 read_start = relay_file_read_start_pos(buf);
1141                 avail = relay_file_read_subbuf_avail(read_start, buf);
1142                 if (!avail)
1143                         break;
1144
1145                 avail = min(count, avail);
1146                 from = buf->start + read_start;
1147                 ret = avail;
1148                 if (copy_to_user(buffer, from, avail))
1149                         break;
1150
1151                 buffer += ret;
1152                 written += ret;
1153                 count -= ret;
1154
1155                 relay_file_read_consume(buf, read_start, ret);
1156                 *ppos = relay_file_read_end_pos(buf, read_start, ret);
1157         } while (count);
1158         inode_unlock(file_inode(filp));
1159
1160         return written;
1161 }
1162
1163 static void relay_consume_bytes(struct rchan_buf *rbuf, int bytes_consumed)
1164 {
1165         rbuf->bytes_consumed += bytes_consumed;
1166
1167         if (rbuf->bytes_consumed >= rbuf->chan->subbuf_size) {
1168                 relay_subbufs_consumed(rbuf->chan, rbuf->cpu, 1);
1169                 rbuf->bytes_consumed %= rbuf->chan->subbuf_size;
1170         }
1171 }
1172
1173 static void relay_pipe_buf_release(struct pipe_inode_info *pipe,
1174                                    struct pipe_buffer *buf)
1175 {
1176         struct rchan_buf *rbuf;
1177
1178         rbuf = (struct rchan_buf *)page_private(buf->page);
1179         relay_consume_bytes(rbuf, buf->private);
1180 }
1181
1182 static const struct pipe_buf_operations relay_pipe_buf_ops = {
1183         .release        = relay_pipe_buf_release,
1184         .try_steal      = generic_pipe_buf_try_steal,
1185         .get            = generic_pipe_buf_get,
1186 };
1187
1188 static void relay_page_release(struct splice_pipe_desc *spd, unsigned int i)
1189 {
1190 }
1191
1192 /*
1193  *      subbuf_splice_actor - splice up to one subbuf's worth of data
1194  */
1195 static ssize_t subbuf_splice_actor(struct file *in,
1196                                loff_t *ppos,
1197                                struct pipe_inode_info *pipe,
1198                                size_t len,
1199                                unsigned int flags,
1200                                int *nonpad_ret)
1201 {
1202         unsigned int pidx, poff, total_len, subbuf_pages, nr_pages;
1203         struct rchan_buf *rbuf = in->private_data;
1204         unsigned int subbuf_size = rbuf->chan->subbuf_size;
1205         uint64_t pos = (uint64_t) *ppos;
1206         uint32_t alloc_size = (uint32_t) rbuf->chan->alloc_size;
1207         size_t read_start = (size_t) do_div(pos, alloc_size);
1208         size_t read_subbuf = read_start / subbuf_size;
1209         size_t padding = rbuf->padding[read_subbuf];
1210         size_t nonpad_end = read_subbuf * subbuf_size + subbuf_size - padding;
1211         struct page *pages[PIPE_DEF_BUFFERS];
1212         struct partial_page partial[PIPE_DEF_BUFFERS];
1213         struct splice_pipe_desc spd = {
1214                 .pages = pages,
1215                 .nr_pages = 0,
1216                 .nr_pages_max = PIPE_DEF_BUFFERS,
1217                 .partial = partial,
1218                 .ops = &relay_pipe_buf_ops,
1219                 .spd_release = relay_page_release,
1220         };
1221         ssize_t ret;
1222
1223         if (rbuf->subbufs_produced == rbuf->subbufs_consumed)
1224                 return 0;
1225         if (splice_grow_spd(pipe, &spd))
1226                 return -ENOMEM;
1227
1228         /*
1229          * Adjust read len, if longer than what is available
1230          */
1231         if (len > (subbuf_size - read_start % subbuf_size))
1232                 len = subbuf_size - read_start % subbuf_size;
1233
1234         subbuf_pages = rbuf->chan->alloc_size >> PAGE_SHIFT;
1235         pidx = (read_start / PAGE_SIZE) % subbuf_pages;
1236         poff = read_start & ~PAGE_MASK;
1237         nr_pages = min_t(unsigned int, subbuf_pages, spd.nr_pages_max);
1238
1239         for (total_len = 0; spd.nr_pages < nr_pages; spd.nr_pages++) {
1240                 unsigned int this_len, this_end, private;
1241                 unsigned int cur_pos = read_start + total_len;
1242
1243                 if (!len)
1244                         break;
1245
1246                 this_len = min_t(unsigned long, len, PAGE_SIZE - poff);
1247                 private = this_len;
1248
1249                 spd.pages[spd.nr_pages] = rbuf->page_array[pidx];
1250                 spd.partial[spd.nr_pages].offset = poff;
1251
1252                 this_end = cur_pos + this_len;
1253                 if (this_end >= nonpad_end) {
1254                         this_len = nonpad_end - cur_pos;
1255                         private = this_len + padding;
1256                 }
1257                 spd.partial[spd.nr_pages].len = this_len;
1258                 spd.partial[spd.nr_pages].private = private;
1259
1260                 len -= this_len;
1261                 total_len += this_len;
1262                 poff = 0;
1263                 pidx = (pidx + 1) % subbuf_pages;
1264
1265                 if (this_end >= nonpad_end) {
1266                         spd.nr_pages++;
1267                         break;
1268                 }
1269         }
1270
1271         ret = 0;
1272         if (!spd.nr_pages)
1273                 goto out;
1274
1275         ret = *nonpad_ret = splice_to_pipe(pipe, &spd);
1276         if (ret < 0 || ret < total_len)
1277                 goto out;
1278
1279         if (read_start + ret == nonpad_end)
1280                 ret += padding;
1281
1282 out:
1283         splice_shrink_spd(&spd);
1284         return ret;
1285 }
1286
1287 static ssize_t relay_file_splice_read(struct file *in,
1288                                       loff_t *ppos,
1289                                       struct pipe_inode_info *pipe,
1290                                       size_t len,
1291                                       unsigned int flags)
1292 {
1293         ssize_t spliced;
1294         int ret;
1295         int nonpad_ret = 0;
1296
1297         ret = 0;
1298         spliced = 0;
1299
1300         while (len && !spliced) {
1301                 ret = subbuf_splice_actor(in, ppos, pipe, len, flags, &nonpad_ret);
1302                 if (ret < 0)
1303                         break;
1304                 else if (!ret) {
1305                         if (flags & SPLICE_F_NONBLOCK)
1306                                 ret = -EAGAIN;
1307                         break;
1308                 }
1309
1310                 *ppos += ret;
1311                 if (ret > len)
1312                         len = 0;
1313                 else
1314                         len -= ret;
1315                 spliced += nonpad_ret;
1316                 nonpad_ret = 0;
1317         }
1318
1319         if (spliced)
1320                 return spliced;
1321
1322         return ret;
1323 }
1324
1325 const struct file_operations relay_file_operations = {
1326         .open           = relay_file_open,
1327         .poll           = relay_file_poll,
1328         .mmap           = relay_file_mmap,
1329         .read           = relay_file_read,
1330         .llseek         = no_llseek,
1331         .release        = relay_file_release,
1332         .splice_read    = relay_file_splice_read,
1333 };
1334 EXPORT_SYMBOL_GPL(relay_file_operations);