ust: add kernel files: relay-alloc.c (ltt-relay-alloc.c), relay.c (ltt-relay.c)
authorPierre-Marc Fournier <pierre-marc.fournier@polymtl.ca>
Tue, 10 Feb 2009 17:11:58 +0000 (12:11 -0500)
committerPierre-Marc Fournier <pierre-marc.fournier@polymtl.ca>
Tue, 10 Feb 2009 17:11:58 +0000 (12:11 -0500)
libtracing/relay-alloc.c [new file with mode: 0644]
libtracing/relay.c [new file with mode: 0644]

diff --git a/libtracing/relay-alloc.c b/libtracing/relay-alloc.c
new file mode 100644 (file)
index 0000000..0f92518
--- /dev/null
@@ -0,0 +1,657 @@
+/*
+ * Public API and common code for kernel->userspace relay file support.
+ *
+ * Copyright (C) 2002-2005 - Tom Zanussi (zanussi@us.ibm.com), IBM Corp
+ * Copyright (C) 1999-2005 - Karim Yaghmour (karim@opersys.com)
+ * Copyright (C) 2008 - Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca)
+ *
+ * Moved to kernel/relay.c by Paul Mundt, 2006.
+ * November 2006 - CPU hotplug support by Mathieu Desnoyers
+ *     (mathieu.desnoyers@polymtl.ca)
+ *
+ * This file is released under the GPL.
+ */
+#include <linux/errno.h>
+#include <linux/stddef.h>
+#include <linux/slab.h>
+#include <linux/module.h>
+#include <linux/string.h>
+#include <linux/ltt-relay.h>
+#include <linux/vmalloc.h>
+#include <linux/mm.h>
+#include <linux/cpu.h>
+#include <linux/splice.h>
+#include <linux/bitops.h>
+
+/* list of open channels, for cpu hotplug */
+static DEFINE_MUTEX(relay_channels_mutex);
+static LIST_HEAD(relay_channels);
+
+/**
+ *     relay_alloc_buf - allocate a channel buffer
+ *     @buf: the buffer struct
+ *     @size: total size of the buffer
+ */
+static int relay_alloc_buf(struct rchan_buf *buf, size_t *size)
+{
+       unsigned int i, n_pages;
+       struct buf_page *buf_page, *n;
+
+       *size = PAGE_ALIGN(*size);
+       n_pages = *size >> PAGE_SHIFT;
+
+       INIT_LIST_HEAD(&buf->pages);
+
+       for (i = 0; i < n_pages; i++) {
+               buf_page = kmalloc_node(sizeof(*buf_page), GFP_KERNEL,
+                       cpu_to_node(buf->cpu));
+               if (unlikely(!buf_page))
+                       goto depopulate;
+               buf_page->page = alloc_pages_node(cpu_to_node(buf->cpu),
+                       GFP_KERNEL | __GFP_ZERO, 0);
+               if (unlikely(!buf_page->page)) {
+                       kfree(buf_page);
+                       goto depopulate;
+               }
+               list_add_tail(&buf_page->list, &buf->pages);
+               buf_page->offset = (size_t)i << PAGE_SHIFT;
+               buf_page->buf = buf;
+               set_page_private(buf_page->page, (unsigned long)buf_page);
+               if (i == 0) {
+                       buf->wpage = buf_page;
+                       buf->hpage[0] = buf_page;
+                       buf->hpage[1] = buf_page;
+                       buf->rpage = buf_page;
+               }
+       }
+       buf->page_count = n_pages;
+       return 0;
+
+depopulate:
+       list_for_each_entry_safe(buf_page, n, &buf->pages, list) {
+               list_del_init(&buf_page->list);
+               __free_page(buf_page->page);
+               kfree(buf_page);
+       }
+       return -ENOMEM;
+}
+
+/**
+ *     relay_create_buf - allocate and initialize a channel buffer
+ *     @chan: the relay channel
+ *     @cpu: cpu the buffer belongs to
+ *
+ *     Returns channel buffer if successful, %NULL otherwise.
+ */
+static struct rchan_buf *relay_create_buf(struct rchan *chan, int cpu)
+{
+       int ret;
+       struct rchan_buf *buf = kzalloc(sizeof(struct rchan_buf), GFP_KERNEL);
+       if (!buf)
+               return NULL;
+
+       buf->cpu = cpu;
+       ret = relay_alloc_buf(buf, &chan->alloc_size);
+       if (ret)
+               goto free_buf;
+
+       buf->chan = chan;
+       kref_get(&buf->chan->kref);
+       return buf;
+
+free_buf:
+       kfree(buf);
+       return NULL;
+}
+
+/**
+ *     relay_destroy_channel - free the channel struct
+ *     @kref: target kernel reference that contains the relay channel
+ *
+ *     Should only be called from kref_put().
+ */
+static void relay_destroy_channel(struct kref *kref)
+{
+       struct rchan *chan = container_of(kref, struct rchan, kref);
+       kfree(chan);
+}
+
+/**
+ *     relay_destroy_buf - destroy an rchan_buf struct and associated buffer
+ *     @buf: the buffer struct
+ */
+static void relay_destroy_buf(struct rchan_buf *buf)
+{
+       struct rchan *chan = buf->chan;
+       struct buf_page *buf_page, *n;
+
+       list_for_each_entry_safe(buf_page, n, &buf->pages, list) {
+               list_del_init(&buf_page->list);
+               __free_page(buf_page->page);
+               kfree(buf_page);
+       }
+       chan->buf[buf->cpu] = NULL;
+       kfree(buf);
+       kref_put(&chan->kref, relay_destroy_channel);
+}
+
+/**
+ *     relay_remove_buf - remove a channel buffer
+ *     @kref: target kernel reference that contains the relay buffer
+ *
+ *     Removes the file from the fileystem, which also frees the
+ *     rchan_buf_struct and the channel buffer.  Should only be called from
+ *     kref_put().
+ */
+static void relay_remove_buf(struct kref *kref)
+{
+       struct rchan_buf *buf = container_of(kref, struct rchan_buf, kref);
+       buf->chan->cb->remove_buf_file(buf->dentry);
+       relay_destroy_buf(buf);
+}
+
+/*
+ * High-level relay kernel API and associated functions.
+ */
+
+/*
+ * rchan_callback implementations defining default channel behavior.  Used
+ * in place of corresponding NULL values in client callback struct.
+ */
+
+/*
+ * create_buf_file_create() default callback.  Does nothing.
+ */
+static struct dentry *create_buf_file_default_callback(const char *filename,
+                                                      struct dentry *parent,
+                                                      int mode,
+                                                      struct rchan_buf *buf)
+{
+       return NULL;
+}
+
+/*
+ * remove_buf_file() default callback.  Does nothing.
+ */
+static int remove_buf_file_default_callback(struct dentry *dentry)
+{
+       return -EINVAL;
+}
+
+/* relay channel default callbacks */
+static struct rchan_callbacks default_channel_callbacks = {
+       .create_buf_file = create_buf_file_default_callback,
+       .remove_buf_file = remove_buf_file_default_callback,
+};
+
+/**
+ *     wakeup_readers - wake up readers waiting on a channel
+ *     @data: contains the channel buffer
+ *
+ *     This is the timer function used to defer reader waking.
+ */
+static void wakeup_readers(unsigned long data)
+{
+       struct rchan_buf *buf = (struct rchan_buf *)data;
+       wake_up_interruptible(&buf->read_wait);
+}
+
+/**
+ *     __relay_reset - reset a channel buffer
+ *     @buf: the channel buffer
+ *     @init: 1 if this is a first-time initialization
+ *
+ *     See relay_reset() for description of effect.
+ */
+static void __relay_reset(struct rchan_buf *buf, unsigned int init)
+{
+       if (init) {
+               init_waitqueue_head(&buf->read_wait);
+               kref_init(&buf->kref);
+               setup_timer(&buf->timer, wakeup_readers, (unsigned long)buf);
+       } else
+               del_timer_sync(&buf->timer);
+
+       buf->finalized = 0;
+}
+
+/*
+ *     relay_open_buf - create a new relay channel buffer
+ *
+ *     used by relay_open() and CPU hotplug.
+ */
+static struct rchan_buf *relay_open_buf(struct rchan *chan, unsigned int cpu)
+{
+       struct rchan_buf *buf = NULL;
+       struct dentry *dentry;
+       char *tmpname;
+
+       tmpname = kzalloc(NAME_MAX + 1, GFP_KERNEL);
+       if (!tmpname)
+               goto end;
+       snprintf(tmpname, NAME_MAX, "%s%d", chan->base_filename, cpu);
+
+       buf = relay_create_buf(chan, cpu);
+       if (!buf)
+               goto free_name;
+
+       __relay_reset(buf, 1);
+
+       /* Create file in fs */
+       dentry = chan->cb->create_buf_file(tmpname, chan->parent, S_IRUSR,
+                                          buf);
+       if (!dentry)
+               goto free_buf;
+
+       buf->dentry = dentry;
+
+       goto free_name;
+
+free_buf:
+       relay_destroy_buf(buf);
+       buf = NULL;
+free_name:
+       kfree(tmpname);
+end:
+       return buf;
+}
+
+/**
+ *     relay_close_buf - close a channel buffer
+ *     @buf: channel buffer
+ *
+ *     Marks the buffer finalized and restores the default callbacks.
+ *     The channel buffer and channel buffer data structure are then freed
+ *     automatically when the last reference is given up.
+ */
+static void relay_close_buf(struct rchan_buf *buf)
+{
+       del_timer_sync(&buf->timer);
+       kref_put(&buf->kref, relay_remove_buf);
+}
+
+static void setup_callbacks(struct rchan *chan,
+                                  struct rchan_callbacks *cb)
+{
+       if (!cb) {
+               chan->cb = &default_channel_callbacks;
+               return;
+       }
+
+       if (!cb->create_buf_file)
+               cb->create_buf_file = create_buf_file_default_callback;
+       if (!cb->remove_buf_file)
+               cb->remove_buf_file = remove_buf_file_default_callback;
+       chan->cb = cb;
+}
+
+/**
+ *     relay_hotcpu_callback - CPU hotplug callback
+ *     @nb: notifier block
+ *     @action: hotplug action to take
+ *     @hcpu: CPU number
+ *
+ *     Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD)
+ */
+static int __cpuinit relay_hotcpu_callback(struct notifier_block *nb,
+                               unsigned long action,
+                               void *hcpu)
+{
+       unsigned int hotcpu = (unsigned long)hcpu;
+       struct rchan *chan;
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+               mutex_lock(&relay_channels_mutex);
+               list_for_each_entry(chan, &relay_channels, list) {
+                       if (chan->buf[hotcpu])
+                               continue;
+                       chan->buf[hotcpu] = relay_open_buf(chan, hotcpu);
+                       if (!chan->buf[hotcpu]) {
+                               printk(KERN_ERR
+                                       "relay_hotcpu_callback: cpu %d buffer "
+                                       "creation failed\n", hotcpu);
+                               mutex_unlock(&relay_channels_mutex);
+                               return NOTIFY_BAD;
+                       }
+               }
+               mutex_unlock(&relay_channels_mutex);
+               break;
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               /* No need to flush the cpu : will be flushed upon
+                * final relay_flush() call. */
+               break;
+       }
+       return NOTIFY_OK;
+}
+
+/**
+ *     ltt_relay_open - create a new relay channel
+ *     @base_filename: base name of files to create
+ *     @parent: dentry of parent directory, %NULL for root directory
+ *     @subbuf_size: size of sub-buffers
+ *     @n_subbufs: number of sub-buffers
+ *     @cb: client callback functions
+ *     @private_data: user-defined data
+ *
+ *     Returns channel pointer if successful, %NULL otherwise.
+ *
+ *     Creates a channel buffer for each cpu using the sizes and
+ *     attributes specified.  The created channel buffer files
+ *     will be named base_filename0...base_filenameN-1.  File
+ *     permissions will be %S_IRUSR.
+ */
+struct rchan *ltt_relay_open(const char *base_filename,
+                        struct dentry *parent,
+                        size_t subbuf_size,
+                        size_t n_subbufs,
+                        struct rchan_callbacks *cb,
+                        void *private_data)
+{
+       unsigned int i;
+       struct rchan *chan;
+       if (!base_filename)
+               return NULL;
+
+       if (!(subbuf_size && n_subbufs))
+               return NULL;
+
+       chan = kzalloc(sizeof(struct rchan), GFP_KERNEL);
+       if (!chan)
+               return NULL;
+
+       chan->version = LTT_RELAY_CHANNEL_VERSION;
+       chan->n_subbufs = n_subbufs;
+       chan->subbuf_size = subbuf_size;
+       chan->subbuf_size_order = get_count_order(subbuf_size);
+       chan->alloc_size = FIX_SIZE(subbuf_size * n_subbufs);
+       chan->parent = parent;
+       chan->private_data = private_data;
+       strlcpy(chan->base_filename, base_filename, NAME_MAX);
+       setup_callbacks(chan, cb);
+       kref_init(&chan->kref);
+
+       mutex_lock(&relay_channels_mutex);
+       for_each_online_cpu(i) {
+               chan->buf[i] = relay_open_buf(chan, i);
+               if (!chan->buf[i])
+                       goto free_bufs;
+       }
+       list_add(&chan->list, &relay_channels);
+       mutex_unlock(&relay_channels_mutex);
+
+       return chan;
+
+free_bufs:
+       for_each_possible_cpu(i) {
+               if (!chan->buf[i])
+                       break;
+               relay_close_buf(chan->buf[i]);
+       }
+
+       kref_put(&chan->kref, relay_destroy_channel);
+       mutex_unlock(&relay_channels_mutex);
+       return NULL;
+}
+EXPORT_SYMBOL_GPL(ltt_relay_open);
+
+/**
+ *     ltt_relay_close - close the channel
+ *     @chan: the channel
+ *
+ *     Closes all channel buffers and frees the channel.
+ */
+void ltt_relay_close(struct rchan *chan)
+{
+       unsigned int i;
+
+       if (!chan)
+               return;
+
+       mutex_lock(&relay_channels_mutex);
+       for_each_possible_cpu(i)
+               if (chan->buf[i])
+                       relay_close_buf(chan->buf[i]);
+
+       list_del(&chan->list);
+       kref_put(&chan->kref, relay_destroy_channel);
+       mutex_unlock(&relay_channels_mutex);
+}
+EXPORT_SYMBOL_GPL(ltt_relay_close);
+
+/*
+ * Start iteration at the previous element. Skip the real list head.
+ */
+struct buf_page *ltt_relay_find_prev_page(struct rchan_buf *buf,
+       struct buf_page *page, size_t offset, ssize_t diff_offset)
+{
+       struct buf_page *iter;
+       size_t orig_iter_off;
+       unsigned int i = 0;
+
+       orig_iter_off = page->offset;
+       list_for_each_entry_reverse(iter, &page->list, list) {
+               /*
+                * Skip the real list head.
+                */
+               if (&iter->list == &buf->pages)
+                       continue;
+               i++;
+               if (offset >= iter->offset
+                       && offset < iter->offset + PAGE_SIZE) {
+#ifdef CONFIG_LTT_RELAY_CHECK_RANDOM_ACCESS
+                       if (i > 1) {
+                               printk(KERN_WARNING
+                                       "Backward random access detected in "
+                                       "ltt_relay. Iterations %u, "
+                                       "offset %zu, orig iter->off %zu, "
+                                       "iter->off %zu diff_offset %zd.\n", i,
+                                       offset, orig_iter_off, iter->offset,
+                                       diff_offset);
+                               WARN_ON(1);
+                       }
+#endif
+                       return iter;
+               }
+       }
+       WARN_ON(1);
+       return NULL;
+}
+EXPORT_SYMBOL_GPL(ltt_relay_find_prev_page);
+
+/*
+ * Start iteration at the next element. Skip the real list head.
+ */
+struct buf_page *ltt_relay_find_next_page(struct rchan_buf *buf,
+       struct buf_page *page, size_t offset, ssize_t diff_offset)
+{
+       struct buf_page *iter;
+       unsigned int i = 0;
+       size_t orig_iter_off;
+
+       orig_iter_off = page->offset;
+       list_for_each_entry(iter, &page->list, list) {
+               /*
+                * Skip the real list head.
+                */
+               if (&iter->list == &buf->pages)
+                       continue;
+               i++;
+               if (offset >= iter->offset
+                       && offset < iter->offset + PAGE_SIZE) {
+#ifdef CONFIG_LTT_RELAY_CHECK_RANDOM_ACCESS
+                       if (i > 1) {
+                               printk(KERN_WARNING
+                                       "Forward random access detected in "
+                                       "ltt_relay. Iterations %u, "
+                                       "offset %zu, orig iter->off %zu, "
+                                       "iter->off %zu diff_offset %zd.\n", i,
+                                       offset, orig_iter_off, iter->offset,
+                                       diff_offset);
+                               WARN_ON(1);
+                       }
+#endif
+                       return iter;
+               }
+       }
+       WARN_ON(1);
+       return NULL;
+}
+EXPORT_SYMBOL_GPL(ltt_relay_find_next_page);
+
+/**
+ * ltt_relay_write - write data to a ltt_relay buffer.
+ * @buf : buffer
+ * @offset : offset within the buffer
+ * @src : source address
+ * @len : length to write
+ * @page : cached buffer page
+ * @pagecpy : page size copied so far
+ */
+void _ltt_relay_write(struct rchan_buf *buf, size_t offset,
+       const void *src, size_t len, struct buf_page *page, ssize_t pagecpy)
+{
+       do {
+               len -= pagecpy;
+               src += pagecpy;
+               offset += pagecpy;
+               /*
+                * Underlying layer should never ask for writes across
+                * subbuffers.
+                */
+               WARN_ON(offset >= buf->chan->alloc_size);
+
+               page = ltt_relay_cache_page(buf, &buf->wpage, page, offset);
+               pagecpy = min_t(size_t, len, PAGE_SIZE - (offset & ~PAGE_MASK));
+               ltt_relay_do_copy(page_address(page->page)
+                       + (offset & ~PAGE_MASK), src, pagecpy);
+       } while (unlikely(len != pagecpy));
+}
+EXPORT_SYMBOL_GPL(_ltt_relay_write);
+
+/**
+ * ltt_relay_read - read data from ltt_relay_buffer.
+ * @buf : buffer
+ * @offset : offset within the buffer
+ * @dest : destination address
+ * @len : length to write
+ */
+int ltt_relay_read(struct rchan_buf *buf, size_t offset,
+       void *dest, size_t len)
+{
+       struct buf_page *page;
+       ssize_t pagecpy, orig_len;
+
+       orig_len = len;
+       offset &= buf->chan->alloc_size - 1;
+       page = buf->rpage;
+       if (unlikely(!len))
+               return 0;
+       for (;;) {
+               page = ltt_relay_cache_page(buf, &buf->rpage, page, offset);
+               pagecpy = min_t(size_t, len, PAGE_SIZE - (offset & ~PAGE_MASK));
+               memcpy(dest, page_address(page->page) + (offset & ~PAGE_MASK),
+                       pagecpy);
+               len -= pagecpy;
+               if (likely(!len))
+                       break;
+               dest += pagecpy;
+               offset += pagecpy;
+               /*
+                * Underlying layer should never ask for reads across
+                * subbuffers.
+                */
+               WARN_ON(offset >= buf->chan->alloc_size);
+       }
+       return orig_len;
+}
+EXPORT_SYMBOL_GPL(ltt_relay_read);
+
+/**
+ * ltt_relay_read_get_page - Get a whole page to read from
+ * @buf : buffer
+ * @offset : offset within the buffer
+ */
+struct buf_page *ltt_relay_read_get_page(struct rchan_buf *buf, size_t offset)
+{
+       struct buf_page *page;
+
+       offset &= buf->chan->alloc_size - 1;
+       page = buf->rpage;
+       page = ltt_relay_cache_page(buf, &buf->rpage, page, offset);
+       return page;
+}
+EXPORT_SYMBOL_GPL(ltt_relay_read_get_page);
+
+/**
+ * ltt_relay_offset_address - get address of a location within the buffer
+ * @buf : buffer
+ * @offset : offset within the buffer.
+ *
+ * Return the address where a given offset is located.
+ * Should be used to get the current subbuffer header pointer. Given we know
+ * it's never on a page boundary, it's safe to write directly to this address,
+ * as long as the write is never bigger than a page size.
+ */
+void *ltt_relay_offset_address(struct rchan_buf *buf, size_t offset)
+{
+       struct buf_page *page;
+       unsigned int odd;
+
+       offset &= buf->chan->alloc_size - 1;
+       odd = !!(offset & buf->chan->subbuf_size);
+       page = buf->hpage[odd];
+       if (offset < page->offset || offset >= page->offset + PAGE_SIZE)
+               buf->hpage[odd] = page = buf->wpage;
+       page = ltt_relay_cache_page(buf, &buf->hpage[odd], page, offset);
+       return page_address(page->page) + (offset & ~PAGE_MASK);
+}
+EXPORT_SYMBOL_GPL(ltt_relay_offset_address);
+
+/**
+ *     relay_file_open - open file op for relay files
+ *     @inode: the inode
+ *     @filp: the file
+ *
+ *     Increments the channel buffer refcount.
+ */
+static int relay_file_open(struct inode *inode, struct file *filp)
+{
+       struct rchan_buf *buf = inode->i_private;
+       kref_get(&buf->kref);
+       filp->private_data = buf;
+
+       return nonseekable_open(inode, filp);
+}
+
+/**
+ *     relay_file_release - release file op for relay files
+ *     @inode: the inode
+ *     @filp: the file
+ *
+ *     Decrements the channel refcount, as the filesystem is
+ *     no longer using it.
+ */
+static int relay_file_release(struct inode *inode, struct file *filp)
+{
+       struct rchan_buf *buf = filp->private_data;
+       kref_put(&buf->kref, relay_remove_buf);
+
+       return 0;
+}
+
+const struct file_operations ltt_relay_file_operations = {
+       .open           = relay_file_open,
+       .release        = relay_file_release,
+};
+EXPORT_SYMBOL_GPL(ltt_relay_file_operations);
+
+static __init int relay_init(void)
+{
+       hotcpu_notifier(relay_hotcpu_callback, 5);
+       return 0;
+}
+
+module_init(relay_init);
diff --git a/libtracing/relay.c b/libtracing/relay.c
new file mode 100644 (file)
index 0000000..f8bff41
--- /dev/null
@@ -0,0 +1,1671 @@
+/*
+ * ltt/ltt-relay.c
+ *
+ * (C) Copyright 2005-2008 - Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca)
+ *
+ * LTTng lockless buffer space management (reader/writer).
+ *
+ * Author:
+ *     Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca)
+ *
+ * Inspired from LTT :
+ *  Karim Yaghmour (karim@opersys.com)
+ *  Tom Zanussi (zanussi@us.ibm.com)
+ *  Bob Wisniewski (bob@watson.ibm.com)
+ * And from K42 :
+ *  Bob Wisniewski (bob@watson.ibm.com)
+ *
+ * Changelog:
+ *  08/10/08, Cleanup.
+ *  19/10/05, Complete lockless mechanism.
+ *  27/05/05, Modular redesign and rewrite.
+ *
+ * Userspace reader semantic :
+ * while (poll fd != POLLHUP) {
+ *   - ioctl RELAY_GET_SUBBUF_SIZE
+ *   while (1) {
+ *     - ioctl GET_SUBBUF
+ *     - splice 1 subbuffer worth of data to a pipe
+ *     - splice the data from pipe to disk/network
+ *     - ioctl PUT_SUBBUF, check error value
+ *       if err val < 0, previous subbuffer was corrupted.
+ *   }
+ * }
+ */
+
+#include <linux/time.h>
+#include <linux/ltt-tracer.h>
+#include <linux/ltt-relay.h>
+#include <linux/module.h>
+#include <linux/string.h>
+#include <linux/slab.h>
+#include <linux/init.h>
+#include <linux/rcupdate.h>
+#include <linux/sched.h>
+#include <linux/bitops.h>
+#include <linux/fs.h>
+#include <linux/smp_lock.h>
+#include <linux/debugfs.h>
+#include <linux/stat.h>
+#include <linux/cpu.h>
+#include <linux/pipe_fs_i.h>
+#include <linux/splice.h>
+#include <asm/atomic.h>
+#include <asm/local.h>
+
+#if 0
+#define printk_dbg(fmt, args...) printk(fmt, args)
+#else
+#define printk_dbg(fmt, args...)
+#endif
+
+/* LTTng lockless logging buffer info */
+struct ltt_channel_buf_struct {
+       /* First 32 bytes cache-hot cacheline */
+       local_t offset;                 /* Current offset in the buffer */
+       local_t *commit_count;          /* Commit count per sub-buffer */
+       atomic_long_t consumed;         /*
+                                        * Current offset in the buffer
+                                        * standard atomic access (shared)
+                                        */
+       unsigned long last_tsc;         /*
+                                        * Last timestamp written in the buffer.
+                                        */
+       /* End of first 32 bytes cacheline */
+       atomic_long_t active_readers;   /*
+                                        * Active readers count
+                                        * standard atomic access (shared)
+                                        */
+       local_t events_lost;
+       local_t corrupted_subbuffers;
+       spinlock_t full_lock;           /*
+                                        * buffer full condition spinlock, only
+                                        * for userspace tracing blocking mode
+                                        * synchronization with reader.
+                                        */
+       wait_queue_head_t write_wait;   /*
+                                        * Wait queue for blocking user space
+                                        * writers
+                                        */
+       atomic_t wakeup_readers;        /* Boolean : wakeup readers waiting ? */
+} ____cacheline_aligned;
+
+/*
+ * Last TSC comparison functions. Check if the current TSC overflows
+ * LTT_TSC_BITS bits from the last TSC read. Reads and writes last_tsc
+ * atomically.
+ */
+
+#if (BITS_PER_LONG == 32)
+static inline void save_last_tsc(struct ltt_channel_buf_struct *ltt_buf,
+                                       u64 tsc)
+{
+       ltt_buf->last_tsc = (unsigned long)(tsc >> LTT_TSC_BITS);
+}
+
+static inline int last_tsc_overflow(struct ltt_channel_buf_struct *ltt_buf,
+                                       u64 tsc)
+{
+       unsigned long tsc_shifted = (unsigned long)(tsc >> LTT_TSC_BITS);
+
+       if (unlikely((tsc_shifted - ltt_buf->last_tsc)))
+               return 1;
+       else
+               return 0;
+}
+#else
+static inline void save_last_tsc(struct ltt_channel_buf_struct *ltt_buf,
+                                       u64 tsc)
+{
+       ltt_buf->last_tsc = (unsigned long)tsc;
+}
+
+static inline int last_tsc_overflow(struct ltt_channel_buf_struct *ltt_buf,
+                                       u64 tsc)
+{
+       if (unlikely((tsc - ltt_buf->last_tsc) >> LTT_TSC_BITS))
+               return 1;
+       else
+               return 0;
+}
+#endif
+
+static struct file_operations ltt_file_operations;
+
+/*
+ * A switch is done during tracing or as a final flush after tracing (so it
+ * won't write in the new sub-buffer).
+ */
+enum force_switch_mode { FORCE_ACTIVE, FORCE_FLUSH };
+
+static int ltt_relay_create_buffer(struct ltt_trace_struct *trace,
+               struct ltt_channel_struct *ltt_chan,
+               struct rchan_buf *buf,
+               unsigned int cpu,
+               unsigned int n_subbufs);
+
+static void ltt_relay_destroy_buffer(struct ltt_channel_struct *ltt_chan,
+               unsigned int cpu);
+
+static void ltt_force_switch(struct rchan_buf *buf,
+               enum force_switch_mode mode);
+
+/*
+ * Trace callbacks
+ */
+static void ltt_buffer_begin_callback(struct rchan_buf *buf,
+                       u64 tsc, unsigned int subbuf_idx)
+{
+       struct ltt_channel_struct *channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_subbuffer_header *header =
+               (struct ltt_subbuffer_header *)
+                       ltt_relay_offset_address(buf,
+                               subbuf_idx * buf->chan->subbuf_size);
+
+       header->cycle_count_begin = tsc;
+       header->lost_size = 0xFFFFFFFF; /* for debugging */
+       header->buf_size = buf->chan->subbuf_size;
+       ltt_write_trace_header(channel->trace, header);
+}
+
+/*
+ * offset is assumed to never be 0 here : never deliver a completely empty
+ * subbuffer. The lost size is between 0 and subbuf_size-1.
+ */
+static notrace void ltt_buffer_end_callback(struct rchan_buf *buf,
+               u64 tsc, unsigned int offset, unsigned int subbuf_idx)
+{
+       struct ltt_channel_struct *channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(channel->buf, buf->cpu);
+       struct ltt_subbuffer_header *header =
+               (struct ltt_subbuffer_header *)
+                       ltt_relay_offset_address(buf,
+                               subbuf_idx * buf->chan->subbuf_size);
+
+       header->lost_size = SUBBUF_OFFSET((buf->chan->subbuf_size - offset),
+                               buf->chan);
+       header->cycle_count_end = tsc;
+       header->events_lost = local_read(&ltt_buf->events_lost);
+       header->subbuf_corrupt = local_read(&ltt_buf->corrupted_subbuffers);
+}
+
+static notrace void ltt_deliver(struct rchan_buf *buf, unsigned int subbuf_idx,
+               void *subbuf)
+{
+       struct ltt_channel_struct *channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(channel->buf, buf->cpu);
+
+       atomic_set(&ltt_buf->wakeup_readers, 1);
+}
+
+static struct dentry *ltt_create_buf_file_callback(const char *filename,
+               struct dentry *parent, int mode,
+               struct rchan_buf *buf)
+{
+       struct ltt_channel_struct *ltt_chan;
+       int err;
+       struct dentry *dentry;
+
+       ltt_chan = buf->chan->private_data;
+       err = ltt_relay_create_buffer(ltt_chan->trace, ltt_chan,
+                                       buf, buf->cpu,
+                                       buf->chan->n_subbufs);
+       if (err)
+               return ERR_PTR(err);
+
+       dentry = debugfs_create_file(filename, mode, parent, buf,
+                       &ltt_file_operations);
+       if (!dentry)
+               goto error;
+       return dentry;
+error:
+       ltt_relay_destroy_buffer(ltt_chan, buf->cpu);
+       return NULL;
+}
+
+static int ltt_remove_buf_file_callback(struct dentry *dentry)
+{
+       struct rchan_buf *buf = dentry->d_inode->i_private;
+       struct ltt_channel_struct *ltt_chan = buf->chan->private_data;
+
+       debugfs_remove(dentry);
+       ltt_relay_destroy_buffer(ltt_chan, buf->cpu);
+
+       return 0;
+}
+
+/*
+ * Wake writers :
+ *
+ * This must be done after the trace is removed from the RCU list so that there
+ * are no stalled writers.
+ */
+static void ltt_relay_wake_writers(struct ltt_channel_buf_struct *ltt_buf)
+{
+
+       if (waitqueue_active(&ltt_buf->write_wait))
+               wake_up_interruptible(&ltt_buf->write_wait);
+}
+
+/*
+ * This function should not be called from NMI interrupt context
+ */
+static notrace void ltt_buf_unfull(struct rchan_buf *buf,
+               unsigned int subbuf_idx,
+               long offset)
+{
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+
+       ltt_relay_wake_writers(ltt_buf);
+}
+
+/**
+ *     ltt_open - open file op for ltt files
+ *     @inode: opened inode
+ *     @file: opened file
+ *
+ *     Open implementation. Makes sure only one open instance of a buffer is
+ *     done at a given moment.
+ */
+static int ltt_open(struct inode *inode, struct file *file)
+{
+       struct rchan_buf *buf = inode->i_private;
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+
+       if (!atomic_long_add_unless(&ltt_buf->active_readers, 1, 1))
+               return -EBUSY;
+       return ltt_relay_file_operations.open(inode, file);
+}
+
+/**
+ *     ltt_release - release file op for ltt files
+ *     @inode: opened inode
+ *     @file: opened file
+ *
+ *     Release implementation.
+ */
+static int ltt_release(struct inode *inode, struct file *file)
+{
+       struct rchan_buf *buf = inode->i_private;
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+       int ret;
+
+       WARN_ON(atomic_long_read(&ltt_buf->active_readers) != 1);
+       atomic_long_dec(&ltt_buf->active_readers);
+       ret = ltt_relay_file_operations.release(inode, file);
+       WARN_ON(ret);
+       return ret;
+}
+
+/**
+ *     ltt_poll - file op for ltt files
+ *     @filp: the file
+ *     @wait: poll table
+ *
+ *     Poll implementation.
+ */
+static unsigned int ltt_poll(struct file *filp, poll_table *wait)
+{
+       unsigned int mask = 0;
+       struct inode *inode = filp->f_dentry->d_inode;
+       struct rchan_buf *buf = inode->i_private;
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+
+       if (filp->f_mode & FMODE_READ) {
+               poll_wait_set_exclusive(wait);
+               poll_wait(filp, &buf->read_wait, wait);
+
+               WARN_ON(atomic_long_read(&ltt_buf->active_readers) != 1);
+               if (SUBBUF_TRUNC(local_read(&ltt_buf->offset),
+                                                       buf->chan)
+                 - SUBBUF_TRUNC(atomic_long_read(&ltt_buf->consumed),
+                                                       buf->chan)
+                 == 0) {
+                       if (buf->finalized)
+                               return POLLHUP;
+                       else
+                               return 0;
+               } else {
+                       struct rchan *rchan =
+                               ltt_channel->trans_channel_data;
+                       if (SUBBUF_TRUNC(local_read(&ltt_buf->offset),
+                                       buf->chan)
+                         - SUBBUF_TRUNC(atomic_long_read(
+                                               &ltt_buf->consumed),
+                                       buf->chan)
+                         >= rchan->alloc_size)
+                               return POLLPRI | POLLRDBAND;
+                       else
+                               return POLLIN | POLLRDNORM;
+               }
+       }
+       return mask;
+}
+
+static int ltt_do_get_subbuf(struct rchan_buf *buf, struct ltt_channel_buf_struct *ltt_buf, long *pconsumed_old)
+{
+       long consumed_old, consumed_idx, commit_count, write_offset;
+       consumed_old = atomic_long_read(&ltt_buf->consumed);
+       consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan);
+       commit_count = local_read(&ltt_buf->commit_count[consumed_idx]);
+       /*
+        * Make sure we read the commit count before reading the buffer
+        * data and the write offset. Correct consumed offset ordering
+        * wrt commit count is insured by the use of cmpxchg to update
+        * the consumed offset.
+        */
+       smp_rmb();
+       write_offset = local_read(&ltt_buf->offset);
+       /*
+        * Check that the subbuffer we are trying to consume has been
+        * already fully committed.
+        */
+       if (((commit_count - buf->chan->subbuf_size)
+            & ltt_channel->commit_count_mask)
+           - (BUFFER_TRUNC(consumed_old, buf->chan)
+              >> ltt_channel->n_subbufs_order)
+           != 0) {
+               return -EAGAIN;
+       }
+       /*
+        * Check that we are not about to read the same subbuffer in
+        * which the writer head is.
+        */
+       if ((SUBBUF_TRUNC(write_offset, buf->chan)
+          - SUBBUF_TRUNC(consumed_old, buf->chan))
+          == 0) {
+               return -EAGAIN;
+       }
+
+       *pconsumed_old = consumed_old;
+       return 0;
+}
+
+static int ltt_do_put_subbuf(struct rchan_buf *buf, struct ltt_channel_buf_struct *ltt_buf, u32 uconsumed_old)
+{
+       long consumed_new, consumed_old;
+
+       consumed_old = atomic_long_read(&ltt_buf->consumed);
+       consumed_old = consumed_old & (~0xFFFFFFFFL);
+       consumed_old = consumed_old | uconsumed_old;
+       consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan);
+
+       spin_lock(&ltt_buf->full_lock);
+       if (atomic_long_cmpxchg(&ltt_buf->consumed, consumed_old,
+                               consumed_new)
+           != consumed_old) {
+               /* We have been pushed by the writer : the last
+                * buffer read _is_ corrupted! It can also
+                * happen if this is a buffer we never got. */
+               spin_unlock(&ltt_buf->full_lock);
+               return -EIO;
+       } else {
+               /* tell the client that buffer is now unfull */
+               int index;
+               long data;
+               index = SUBBUF_INDEX(consumed_old, buf->chan);
+               data = BUFFER_OFFSET(consumed_old, buf->chan);
+               ltt_buf_unfull(buf, index, data);
+               spin_unlock(&ltt_buf->full_lock);
+       }
+       return 0;
+}
+
+/**
+ *     ltt_ioctl - control on the debugfs file
+ *
+ *     @inode: the inode
+ *     @filp: the file
+ *     @cmd: the command
+ *     @arg: command arg
+ *
+ *     This ioctl implements three commands necessary for a minimal
+ *     producer/consumer implementation :
+ *     RELAY_GET_SUBBUF
+ *             Get the next sub buffer that can be read. It never blocks.
+ *     RELAY_PUT_SUBBUF
+ *             Release the currently read sub-buffer. Parameter is the last
+ *             put subbuffer (returned by GET_SUBBUF).
+ *     RELAY_GET_N_BUBBUFS
+ *             returns the number of sub buffers in the per cpu channel.
+ *     RELAY_GET_SUBBUF_SIZE
+ *             returns the size of the sub buffers.
+ */
+static int ltt_ioctl(struct inode *inode, struct file *filp,
+               unsigned int cmd, unsigned long arg)
+{
+       struct rchan_buf *buf = inode->i_private;
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+       u32 __user *argp = (u32 __user *)arg;
+
+       WARN_ON(atomic_long_read(&ltt_buf->active_readers) != 1);
+       switch (cmd) {
+       case RELAY_GET_SUBBUF:
+       {
+               int ret;
+               ret = ltt_do_get_subbuf(buf, ltt_buf, &consumed_old);
+               if(ret < 0)
+                       return ret;
+               return put_user((u32)consumed_old, argp);
+       }
+       case RELAY_PUT_SUBBUF:
+       {
+               int ret;
+               u32 uconsumed_old;
+               ret = get_user(uconsumed_old, argp);
+               if (ret)
+                       return ret; /* will return -EFAULT */
+               return ltt_do_put_subbuf(buf, ltt_buf, uconsumed_old);
+       }
+       case RELAY_GET_N_SUBBUFS:
+               return put_user((u32)buf->chan->n_subbufs, argp);
+               break;
+       case RELAY_GET_SUBBUF_SIZE:
+               return put_user((u32)buf->chan->subbuf_size, argp);
+               break;
+       default:
+               return -ENOIOCTLCMD;
+       }
+       return 0;
+}
+
+#ifdef CONFIG_COMPAT
+static long ltt_compat_ioctl(struct file *file, unsigned int cmd,
+               unsigned long arg)
+{
+       long ret = -ENOIOCTLCMD;
+
+       lock_kernel();
+       ret = ltt_ioctl(file->f_dentry->d_inode, file, cmd, arg);
+       unlock_kernel();
+
+       return ret;
+}
+#endif
+
+static void ltt_relay_pipe_buf_release(struct pipe_inode_info *pipe,
+                                  struct pipe_buffer *pbuf)
+{
+}
+
+static struct pipe_buf_operations ltt_relay_pipe_buf_ops = {
+       .can_merge = 0,
+       .map = generic_pipe_buf_map,
+       .unmap = generic_pipe_buf_unmap,
+       .confirm = generic_pipe_buf_confirm,
+       .release = ltt_relay_pipe_buf_release,
+       .steal = generic_pipe_buf_steal,
+       .get = generic_pipe_buf_get,
+};
+
+static void ltt_relay_page_release(struct splice_pipe_desc *spd, unsigned int i)
+{
+}
+
+/*
+ *     subbuf_splice_actor - splice up to one subbuf's worth of data
+ */
+static int subbuf_splice_actor(struct file *in,
+                              loff_t *ppos,
+                              struct pipe_inode_info *pipe,
+                              size_t len,
+                              unsigned int flags)
+{
+       struct rchan_buf *buf = in->private_data;
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_channel->buf, buf->cpu);
+       unsigned int poff, subbuf_pages, nr_pages;
+       struct page *pages[PIPE_BUFFERS];
+       struct partial_page partial[PIPE_BUFFERS];
+       struct splice_pipe_desc spd = {
+               .pages = pages,
+               .nr_pages = 0,
+               .partial = partial,
+               .flags = flags,
+               .ops = &ltt_relay_pipe_buf_ops,
+               .spd_release = ltt_relay_page_release,
+       };
+       long consumed_old, consumed_idx, roffset;
+       unsigned long bytes_avail;
+
+       /*
+        * Check that a GET_SUBBUF ioctl has been done before.
+        */
+       WARN_ON(atomic_long_read(&ltt_buf->active_readers) != 1);
+       consumed_old = atomic_long_read(&ltt_buf->consumed);
+       consumed_old += *ppos;
+       consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan);
+
+       /*
+        * Adjust read len, if longer than what is available
+        */
+       bytes_avail = SUBBUF_TRUNC(local_read(&ltt_buf->offset), buf->chan)
+                   - consumed_old;
+       WARN_ON(bytes_avail > buf->chan->alloc_size);
+       len = min_t(size_t, len, bytes_avail);
+       subbuf_pages = bytes_avail >> PAGE_SHIFT;
+       nr_pages = min_t(unsigned int, subbuf_pages, PIPE_BUFFERS);
+       roffset = consumed_old & PAGE_MASK;
+       poff = consumed_old & ~PAGE_MASK;
+       printk_dbg(KERN_DEBUG "SPLICE actor len %zu pos %zd write_pos %ld\n",
+               len, (ssize_t)*ppos, local_read(&ltt_buf->offset));
+
+       for (; spd.nr_pages < nr_pages; spd.nr_pages++) {
+               unsigned int this_len;
+               struct buf_page *page;
+
+               if (!len)
+                       break;
+               printk_dbg(KERN_DEBUG "SPLICE actor loop len %zu roffset %ld\n",
+                       len, roffset);
+
+               this_len = PAGE_SIZE - poff;
+               page = ltt_relay_read_get_page(buf, roffset);
+               spd.pages[spd.nr_pages] = page->page;
+               spd.partial[spd.nr_pages].offset = poff;
+               spd.partial[spd.nr_pages].len = this_len;
+
+               poff = 0;
+               roffset += PAGE_SIZE;
+               len -= this_len;
+       }
+
+       if (!spd.nr_pages)
+               return 0;
+
+       return splice_to_pipe(pipe, &spd);
+}
+
+static ssize_t ltt_relay_file_splice_read(struct file *in,
+                                     loff_t *ppos,
+                                     struct pipe_inode_info *pipe,
+                                     size_t len,
+                                     unsigned int flags)
+{
+       ssize_t spliced;
+       int ret;
+
+       ret = 0;
+       spliced = 0;
+
+       printk_dbg(KERN_DEBUG "SPLICE read len %zu pos %zd\n",
+               len, (ssize_t)*ppos);
+       while (len && !spliced) {
+               ret = subbuf_splice_actor(in, ppos, pipe, len, flags);
+               printk_dbg(KERN_DEBUG "SPLICE read loop ret %d\n", ret);
+               if (ret < 0)
+                       break;
+               else if (!ret) {
+                       if (flags & SPLICE_F_NONBLOCK)
+                               ret = -EAGAIN;
+                       break;
+               }
+
+               *ppos += ret;
+               if (ret > len)
+                       len = 0;
+               else
+                       len -= ret;
+               spliced += ret;
+       }
+
+       if (spliced)
+               return spliced;
+
+       return ret;
+}
+
+static void ltt_relay_print_subbuffer_errors(
+               struct ltt_channel_struct *ltt_chan,
+               long cons_off, unsigned int cpu)
+{
+       struct rchan *rchan = ltt_chan->trans_channel_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_chan->buf, cpu);
+       long cons_idx, commit_count, write_offset;
+
+       cons_idx = SUBBUF_INDEX(cons_off, rchan);
+       commit_count = local_read(&ltt_buf->commit_count[cons_idx]);
+       /*
+        * No need to order commit_count and write_offset reads because we
+        * execute after trace is stopped when there are no readers left.
+        */
+       write_offset = local_read(&ltt_buf->offset);
+       printk(KERN_WARNING
+               "LTT : unread channel %s offset is %ld "
+               "and cons_off : %ld (cpu %u)\n",
+               ltt_chan->channel_name, write_offset, cons_off, cpu);
+       /* Check each sub-buffer for non filled commit count */
+       if (((commit_count - rchan->subbuf_size) & ltt_chan->commit_count_mask)
+           - (BUFFER_TRUNC(cons_off, rchan) >> ltt_chan->n_subbufs_order)
+           != 0)
+               printk(KERN_ALERT
+                       "LTT : %s : subbuffer %lu has non filled "
+                       "commit count %lu.\n",
+                       ltt_chan->channel_name, cons_idx, commit_count);
+       printk(KERN_ALERT "LTT : %s : commit count : %lu, subbuf size %zd\n",
+                       ltt_chan->channel_name, commit_count,
+                       rchan->subbuf_size);
+}
+
+static void ltt_relay_print_errors(struct ltt_trace_struct *trace,
+               struct ltt_channel_struct *ltt_chan, int cpu)
+{
+       struct rchan *rchan = ltt_chan->trans_channel_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_chan->buf, cpu);
+       long cons_off;
+
+       for (cons_off = atomic_long_read(&ltt_buf->consumed);
+                       (SUBBUF_TRUNC(local_read(&ltt_buf->offset),
+                                     rchan)
+                        - cons_off) > 0;
+                       cons_off = SUBBUF_ALIGN(cons_off, rchan))
+               ltt_relay_print_subbuffer_errors(ltt_chan, cons_off, cpu);
+}
+
+static void ltt_relay_print_buffer_errors(struct ltt_channel_struct *ltt_chan,
+               unsigned int cpu)
+{
+       struct ltt_trace_struct *trace = ltt_chan->trace;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_chan->buf, cpu);
+
+       if (local_read(&ltt_buf->events_lost))
+               printk(KERN_ALERT
+                       "LTT : %s : %ld events lost "
+                       "in %s channel (cpu %u).\n",
+                       ltt_chan->channel_name,
+                       local_read(&ltt_buf->events_lost),
+                       ltt_chan->channel_name, cpu);
+       if (local_read(&ltt_buf->corrupted_subbuffers))
+               printk(KERN_ALERT
+                       "LTT : %s : %ld corrupted subbuffers "
+                       "in %s channel (cpu %u).\n",
+                       ltt_chan->channel_name,
+                       local_read(&ltt_buf->corrupted_subbuffers),
+                       ltt_chan->channel_name, cpu);
+
+       ltt_relay_print_errors(trace, ltt_chan, cpu);
+}
+
+static void ltt_relay_remove_dirs(struct ltt_trace_struct *trace)
+{
+       debugfs_remove(trace->dentry.trace_root);
+}
+
+static void ltt_relay_release_channel(struct kref *kref)
+{
+       struct ltt_channel_struct *ltt_chan = container_of(kref,
+                       struct ltt_channel_struct, kref);
+       percpu_free(ltt_chan->buf);
+}
+
+/*
+ * Create ltt buffer.
+ */
+static int ltt_relay_create_buffer(struct ltt_trace_struct *trace,
+               struct ltt_channel_struct *ltt_chan, struct rchan_buf *buf,
+               unsigned int cpu, unsigned int n_subbufs)
+{
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_chan->buf, cpu);
+       unsigned int j;
+
+       ltt_buf->commit_count =
+               kzalloc_node(sizeof(ltt_buf->commit_count) * n_subbufs,
+                       GFP_KERNEL, cpu_to_node(cpu));
+       if (!ltt_buf->commit_count)
+               return -ENOMEM;
+       kref_get(&trace->kref);
+       kref_get(&trace->ltt_transport_kref);
+       kref_get(&ltt_chan->kref);
+       local_set(&ltt_buf->offset, ltt_subbuffer_header_size());
+       atomic_long_set(&ltt_buf->consumed, 0);
+       atomic_long_set(&ltt_buf->active_readers, 0);
+       for (j = 0; j < n_subbufs; j++)
+               local_set(&ltt_buf->commit_count[j], 0);
+       init_waitqueue_head(&ltt_buf->write_wait);
+       atomic_set(&ltt_buf->wakeup_readers, 0);
+       spin_lock_init(&ltt_buf->full_lock);
+
+       ltt_buffer_begin_callback(buf, trace->start_tsc, 0);
+       /* atomic_add made on local variable on data that belongs to
+        * various CPUs : ok because tracing not started (for this cpu). */
+       local_add(ltt_subbuffer_header_size(), &ltt_buf->commit_count[0]);
+
+       local_set(&ltt_buf->events_lost, 0);
+       local_set(&ltt_buf->corrupted_subbuffers, 0);
+
+       return 0;
+}
+
+static void ltt_relay_destroy_buffer(struct ltt_channel_struct *ltt_chan,
+               unsigned int cpu)
+{
+       struct ltt_trace_struct *trace = ltt_chan->trace;
+       struct ltt_channel_buf_struct *ltt_buf =
+               percpu_ptr(ltt_chan->buf, cpu);
+
+       kref_put(&ltt_chan->trace->ltt_transport_kref,
+               ltt_release_transport);
+       ltt_relay_print_buffer_errors(ltt_chan, cpu);
+       kfree(ltt_buf->commit_count);
+       ltt_buf->commit_count = NULL;
+       kref_put(&ltt_chan->kref, ltt_relay_release_channel);
+       kref_put(&trace->kref, ltt_release_trace);
+       wake_up_interruptible(&trace->kref_wq);
+}
+
+/*
+ * Create channel.
+ */
+static int ltt_relay_create_channel(const char *trace_name,
+               struct ltt_trace_struct *trace, struct dentry *dir,
+               const char *channel_name, struct ltt_channel_struct *ltt_chan,
+               unsigned int subbuf_size, unsigned int n_subbufs,
+               int overwrite)
+{
+       char *tmpname;
+       unsigned int tmpname_len;
+       int err = 0;
+
+       tmpname = kmalloc(PATH_MAX, GFP_KERNEL);
+       if (!tmpname)
+               return EPERM;
+       if (overwrite) {
+               strncpy(tmpname, LTT_FLIGHT_PREFIX, PATH_MAX-1);
+               strncat(tmpname, channel_name,
+                       PATH_MAX-1-sizeof(LTT_FLIGHT_PREFIX));
+       } else {
+               strncpy(tmpname, channel_name, PATH_MAX-1);
+       }
+       strncat(tmpname, "_", PATH_MAX-1-strlen(tmpname));
+
+       kref_init(&ltt_chan->kref);
+
+       ltt_chan->trace = trace;
+       ltt_chan->buffer_begin = ltt_buffer_begin_callback;
+       ltt_chan->buffer_end = ltt_buffer_end_callback;
+       ltt_chan->overwrite = overwrite;
+       ltt_chan->n_subbufs_order = get_count_order(n_subbufs);
+       ltt_chan->commit_count_mask = (~0UL >> ltt_chan->n_subbufs_order);
+       ltt_chan->buf = percpu_alloc_mask(sizeof(struct ltt_channel_buf_struct),
+                                         GFP_KERNEL, cpu_possible_map);
+       if (!ltt_chan->buf)
+               goto ltt_percpu_alloc_error;
+       ltt_chan->trans_channel_data = ltt_relay_open(tmpname,
+                       dir,
+                       subbuf_size,
+                       n_subbufs,
+                       &trace->callbacks,
+                       ltt_chan);
+       tmpname_len = strlen(tmpname);
+       if (tmpname_len > 0) {
+               /* Remove final _ for pretty printing */
+               tmpname[tmpname_len-1] = '\0';
+       }
+       if (ltt_chan->trans_channel_data == NULL) {
+               printk(KERN_ERR "LTT : Can't open %s channel for trace %s\n",
+                               tmpname, trace_name);
+               goto relay_open_error;
+       }
+
+       err = 0;
+       goto end;
+
+relay_open_error:
+       percpu_free(ltt_chan->buf);
+ltt_percpu_alloc_error:
+       err = EPERM;
+end:
+       kfree(tmpname);
+       return err;
+}
+
+static int ltt_relay_create_dirs(struct ltt_trace_struct *new_trace)
+{
+       new_trace->dentry.trace_root = debugfs_create_dir(new_trace->trace_name,
+                       get_ltt_root());
+       if (new_trace->dentry.trace_root == NULL) {
+               printk(KERN_ERR "LTT : Trace directory name %s already taken\n",
+                               new_trace->trace_name);
+               return EEXIST;
+       }
+
+       new_trace->callbacks.create_buf_file = ltt_create_buf_file_callback;
+       new_trace->callbacks.remove_buf_file = ltt_remove_buf_file_callback;
+
+       return 0;
+}
+
+/*
+ * LTTng channel flush function.
+ *
+ * Must be called when no tracing is active in the channel, because of
+ * accesses across CPUs.
+ */
+static notrace void ltt_relay_buffer_flush(struct rchan_buf *buf)
+{
+       buf->finalized = 1;
+       ltt_force_switch(buf, FORCE_FLUSH);
+}
+
+static void ltt_relay_async_wakeup_chan(struct ltt_channel_struct *ltt_channel)
+{
+       unsigned int i;
+       struct rchan *rchan = ltt_channel->trans_channel_data;
+
+       for_each_possible_cpu(i) {
+               struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, i);
+
+               if (atomic_read(&ltt_buf->wakeup_readers) == 1) {
+                       atomic_set(&ltt_buf->wakeup_readers, 0);
+                       wake_up_interruptible(&rchan->buf[i]->read_wait);
+               }
+       }
+}
+
+static void ltt_relay_finish_buffer(struct ltt_channel_struct *ltt_channel,
+               unsigned int cpu)
+{
+       struct rchan *rchan = ltt_channel->trans_channel_data;
+
+       if (rchan->buf[cpu]) {
+               struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, cpu);
+               ltt_relay_buffer_flush(rchan->buf[cpu]);
+               ltt_relay_wake_writers(ltt_buf);
+       }
+}
+
+
+static void ltt_relay_finish_channel(struct ltt_channel_struct *ltt_channel)
+{
+       unsigned int i;
+
+       for_each_possible_cpu(i)
+               ltt_relay_finish_buffer(ltt_channel, i);
+}
+
+static void ltt_relay_remove_channel(struct ltt_channel_struct *channel)
+{
+       struct rchan *rchan = channel->trans_channel_data;
+
+       ltt_relay_close(rchan);
+       kref_put(&channel->kref, ltt_relay_release_channel);
+}
+
+struct ltt_reserve_switch_offsets {
+       long begin, end, old;
+       long begin_switch, end_switch_current, end_switch_old;
+       long commit_count, reserve_commit_diff;
+       size_t before_hdr_pad, size;
+};
+
+/*
+ * Returns :
+ * 0 if ok
+ * !0 if execution must be aborted.
+ */
+static inline int ltt_relay_try_reserve(
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf, struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets, size_t data_size,
+               u64 *tsc, unsigned int *rflags, int largest_align)
+{
+       offsets->begin = local_read(&ltt_buf->offset);
+       offsets->old = offsets->begin;
+       offsets->begin_switch = 0;
+       offsets->end_switch_current = 0;
+       offsets->end_switch_old = 0;
+
+       *tsc = trace_clock_read64();
+       if (last_tsc_overflow(ltt_buf, *tsc))
+               *rflags = LTT_RFLAG_ID_SIZE_TSC;
+
+       if (SUBBUF_OFFSET(offsets->begin, buf->chan) == 0) {
+               offsets->begin_switch = 1;              /* For offsets->begin */
+       } else {
+               offsets->size = ltt_get_header_size(ltt_channel,
+                                       offsets->begin, data_size,
+                                       &offsets->before_hdr_pad, *rflags);
+               offsets->size += ltt_align(offsets->begin + offsets->size,
+                                          largest_align)
+                                + data_size;
+               if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size)
+                               > buf->chan->subbuf_size) {
+                       offsets->end_switch_old = 1;    /* For offsets->old */
+                       offsets->begin_switch = 1;      /* For offsets->begin */
+               }
+       }
+       if (offsets->begin_switch) {
+               long subbuf_index;
+
+               if (offsets->end_switch_old)
+                       offsets->begin = SUBBUF_ALIGN(offsets->begin,
+                                                     buf->chan);
+               offsets->begin = offsets->begin + ltt_subbuffer_header_size();
+               /* Test new buffer integrity */
+               subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
+               offsets->reserve_commit_diff =
+                       (BUFFER_TRUNC(offsets->begin, buf->chan)
+                        >> ltt_channel->n_subbufs_order)
+                       - (local_read(&ltt_buf->commit_count[subbuf_index])
+                               & ltt_channel->commit_count_mask);
+               if (offsets->reserve_commit_diff == 0) {
+                       /* Next buffer not corrupted. */
+                       if (!ltt_channel->overwrite &&
+                               (SUBBUF_TRUNC(offsets->begin, buf->chan)
+                                - SUBBUF_TRUNC(atomic_long_read(
+                                                       &ltt_buf->consumed),
+                                               buf->chan))
+                               >= rchan->alloc_size) {
+                               /*
+                                * We do not overwrite non consumed buffers
+                                * and we are full : event is lost.
+                                */
+                               local_inc(&ltt_buf->events_lost);
+                               return -1;
+                       } else {
+                               /*
+                                * next buffer not corrupted, we are either in
+                                * overwrite mode or the buffer is not full.
+                                * It's safe to write in this new subbuffer.
+                                */
+                       }
+               } else {
+                       /*
+                        * Next subbuffer corrupted. Force pushing reader even
+                        * in normal mode. It's safe to write in this new
+                        * subbuffer.
+                        */
+               }
+               offsets->size = ltt_get_header_size(ltt_channel,
+                                       offsets->begin, data_size,
+                                       &offsets->before_hdr_pad, *rflags);
+               offsets->size += ltt_align(offsets->begin + offsets->size,
+                                          largest_align)
+                                + data_size;
+               if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size)
+                               > buf->chan->subbuf_size) {
+                       /*
+                        * Event too big for subbuffers, report error, don't
+                        * complete the sub-buffer switch.
+                        */
+                       local_inc(&ltt_buf->events_lost);
+                       return -1;
+               } else {
+                       /*
+                        * We just made a successful buffer switch and the event
+                        * fits in the new subbuffer. Let's write.
+                        */
+               }
+       } else {
+               /*
+                * Event fits in the current buffer and we are not on a switch
+                * boundary. It's safe to write.
+                */
+       }
+       offsets->end = offsets->begin + offsets->size;
+
+       if ((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0) {
+               /*
+                * The offset_end will fall at the very beginning of the next
+                * subbuffer.
+                */
+               offsets->end_switch_current = 1;        /* For offsets->begin */
+       }
+       return 0;
+}
+
+/*
+ * Returns :
+ * 0 if ok
+ * !0 if execution must be aborted.
+ */
+static inline int ltt_relay_try_switch(
+               enum force_switch_mode mode,
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf, struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets,
+               u64 *tsc)
+{
+       long subbuf_index;
+
+       offsets->begin = local_read(&ltt_buf->offset);
+       offsets->old = offsets->begin;
+       offsets->begin_switch = 0;
+       offsets->end_switch_old = 0;
+
+       *tsc = trace_clock_read64();
+
+       if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) {
+               offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan);
+               offsets->end_switch_old = 1;
+       } else {
+               /* we do not have to switch : buffer is empty */
+               return -1;
+       }
+       if (mode == FORCE_ACTIVE)
+               offsets->begin += ltt_subbuffer_header_size();
+       /*
+        * Always begin_switch in FORCE_ACTIVE mode.
+        * Test new buffer integrity
+        */
+       subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
+       offsets->reserve_commit_diff =
+               (BUFFER_TRUNC(offsets->begin, buf->chan)
+                >> ltt_channel->n_subbufs_order)
+               - (local_read(&ltt_buf->commit_count[subbuf_index])
+                       & ltt_channel->commit_count_mask);
+       if (offsets->reserve_commit_diff == 0) {
+               /* Next buffer not corrupted. */
+               if (mode == FORCE_ACTIVE
+                   && !ltt_channel->overwrite
+                   && offsets->begin - atomic_long_read(&ltt_buf->consumed)
+                      >= rchan->alloc_size) {
+                       /*
+                        * We do not overwrite non consumed buffers and we are
+                        * full : ignore switch while tracing is active.
+                        */
+                       return -1;
+               }
+       } else {
+               /*
+                * Next subbuffer corrupted. Force pushing reader even in normal
+                * mode
+                */
+       }
+       offsets->end = offsets->begin;
+       return 0;
+}
+
+static inline void ltt_reserve_push_reader(
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf,
+               struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets)
+{
+       long consumed_old, consumed_new;
+
+       do {
+               consumed_old = atomic_long_read(&ltt_buf->consumed);
+               /*
+                * If buffer is in overwrite mode, push the reader consumed
+                * count if the write position has reached it and we are not
+                * at the first iteration (don't push the reader farther than
+                * the writer). This operation can be done concurrently by many
+                * writers in the same buffer, the writer being at the farthest
+                * write position sub-buffer index in the buffer being the one
+                * which will win this loop.
+                * If the buffer is not in overwrite mode, pushing the reader
+                * only happens if a sub-buffer is corrupted.
+                */
+               if ((SUBBUF_TRUNC(offsets->end-1, buf->chan)
+                  - SUBBUF_TRUNC(consumed_old, buf->chan))
+                  >= rchan->alloc_size)
+                       consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan);
+               else {
+                       consumed_new = consumed_old;
+                       break;
+               }
+       } while (atomic_long_cmpxchg(&ltt_buf->consumed, consumed_old,
+                       consumed_new) != consumed_old);
+
+       if (consumed_old != consumed_new) {
+               /*
+                * Reader pushed : we are the winner of the push, we can
+                * therefore reequilibrate reserve and commit. Atomic increment
+                * of the commit count permits other writers to play around
+                * with this variable before us. We keep track of
+                * corrupted_subbuffers even in overwrite mode :
+                * we never want to write over a non completely committed
+                * sub-buffer : possible causes : the buffer size is too low
+                * compared to the unordered data input, or there is a writer
+                * that died between the reserve and the commit.
+                */
+               if (offsets->reserve_commit_diff) {
+                       /*
+                        * We have to alter the sub-buffer commit count.
+                        * We do not deliver the previous subbuffer, given it
+                        * was either corrupted or not consumed (overwrite
+                        * mode).
+                        */
+                       local_add(offsets->reserve_commit_diff,
+                                 &ltt_buf->commit_count[
+                                       SUBBUF_INDEX(offsets->begin,
+                                                    buf->chan)]);
+                       if (!ltt_channel->overwrite
+                           || offsets->reserve_commit_diff
+                              != rchan->subbuf_size) {
+                               /*
+                                * The reserve commit diff was not subbuf_size :
+                                * it means the subbuffer was partly written to
+                                * and is therefore corrupted. If it is multiple
+                                * of subbuffer size and we are in flight
+                                * recorder mode, we are skipping over a whole
+                                * subbuffer.
+                                */
+                               local_inc(&ltt_buf->corrupted_subbuffers);
+                       }
+               }
+       }
+}
+
+
+/*
+ * ltt_reserve_switch_old_subbuf: switch old subbuffer
+ *
+ * Concurrency safe because we are the last and only thread to alter this
+ * sub-buffer. As long as it is not delivered and read, no other thread can
+ * alter the offset, alter the reserve_count or call the
+ * client_buffer_end_callback on this sub-buffer.
+ *
+ * The only remaining threads could be the ones with pending commits. They will
+ * have to do the deliver themselves.  Not concurrency safe in overwrite mode.
+ * We detect corrupted subbuffers with commit and reserve counts. We keep a
+ * corrupted sub-buffers count and push the readers across these sub-buffers.
+ *
+ * Not concurrency safe if a writer is stalled in a subbuffer and another writer
+ * switches in, finding out it's corrupted.  The result will be than the old
+ * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
+ * will be declared corrupted too because of the commit count adjustment.
+ *
+ * Note : offset_old should never be 0 here.
+ */
+static inline void ltt_reserve_switch_old_subbuf(
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf, struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
+{
+       long oldidx = SUBBUF_INDEX(offsets->old - 1, rchan);
+
+       ltt_channel->buffer_end(buf, *tsc, offsets->old, oldidx);
+       /* Must write buffer end before incrementing commit count */
+       smp_wmb();
+       offsets->commit_count =
+               local_add_return(rchan->subbuf_size
+                                - (SUBBUF_OFFSET(offsets->old - 1, rchan)
+                                + 1),
+                                &ltt_buf->commit_count[oldidx]);
+       if ((BUFFER_TRUNC(offsets->old - 1, rchan)
+                       >> ltt_channel->n_subbufs_order)
+                       - ((offsets->commit_count - rchan->subbuf_size)
+                               & ltt_channel->commit_count_mask) == 0)
+               ltt_deliver(buf, oldidx, NULL);
+}
+
+/*
+ * ltt_reserve_switch_new_subbuf: Populate new subbuffer.
+ *
+ * This code can be executed unordered : writers may already have written to the
+ * sub-buffer before this code gets executed, caution.  The commit makes sure
+ * that this code is executed before the deliver of this sub-buffer.
+ */
+static inline void ltt_reserve_switch_new_subbuf(
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf, struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
+{
+       long beginidx = SUBBUF_INDEX(offsets->begin, rchan);
+
+       ltt_channel->buffer_begin(buf, *tsc, beginidx);
+       /* Must write buffer end before incrementing commit count */
+       smp_wmb();
+       offsets->commit_count = local_add_return(ltt_subbuffer_header_size(),
+                       &ltt_buf->commit_count[beginidx]);
+       /* Check if the written buffer has to be delivered */
+       if ((BUFFER_TRUNC(offsets->begin, rchan)
+                       >> ltt_channel->n_subbufs_order)
+                       - ((offsets->commit_count - rchan->subbuf_size)
+                               & ltt_channel->commit_count_mask) == 0)
+               ltt_deliver(buf, beginidx, NULL);
+}
+
+
+/*
+ * ltt_reserve_end_switch_current: finish switching current subbuffer
+ *
+ * Concurrency safe because we are the last and only thread to alter this
+ * sub-buffer. As long as it is not delivered and read, no other thread can
+ * alter the offset, alter the reserve_count or call the
+ * client_buffer_end_callback on this sub-buffer.
+ *
+ * The only remaining threads could be the ones with pending commits. They will
+ * have to do the deliver themselves.  Not concurrency safe in overwrite mode.
+ * We detect corrupted subbuffers with commit and reserve counts. We keep a
+ * corrupted sub-buffers count and push the readers across these sub-buffers.
+ *
+ * Not concurrency safe if a writer is stalled in a subbuffer and another writer
+ * switches in, finding out it's corrupted.  The result will be than the old
+ * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
+ * will be declared corrupted too because of the commit count adjustment.
+ */
+static inline void ltt_reserve_end_switch_current(
+               struct ltt_channel_struct *ltt_channel,
+               struct ltt_channel_buf_struct *ltt_buf, struct rchan *rchan,
+               struct rchan_buf *buf,
+               struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
+{
+       long endidx = SUBBUF_INDEX(offsets->end - 1, rchan);
+
+       ltt_channel->buffer_end(buf, *tsc, offsets->end, endidx);
+       /* Must write buffer begin before incrementing commit count */
+       smp_wmb();
+       offsets->commit_count =
+               local_add_return(rchan->subbuf_size
+                                - (SUBBUF_OFFSET(offsets->end - 1, rchan)
+                                + 1),
+                                &ltt_buf->commit_count[endidx]);
+       if ((BUFFER_TRUNC(offsets->end - 1, rchan)
+                       >> ltt_channel->n_subbufs_order)
+                       - ((offsets->commit_count - rchan->subbuf_size)
+                               & ltt_channel->commit_count_mask) == 0)
+               ltt_deliver(buf, endidx, NULL);
+}
+
+/**
+ * ltt_relay_reserve_slot - Atomic slot reservation in a LTTng buffer.
+ * @trace: the trace structure to log to.
+ * @ltt_channel: channel structure
+ * @transport_data: data structure specific to ltt relay
+ * @data_size: size of the variable length data to log.
+ * @slot_size: pointer to total size of the slot (out)
+ * @buf_offset : pointer to reserved buffer offset (out)
+ * @tsc: pointer to the tsc at the slot reservation (out)
+ * @cpu: cpuid
+ *
+ * Return : -ENOSPC if not enough space, else returns 0.
+ * It will take care of sub-buffer switching.
+ */
+static notrace int ltt_relay_reserve_slot(struct ltt_trace_struct *trace,
+               struct ltt_channel_struct *ltt_channel, void **transport_data,
+               size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc,
+               unsigned int *rflags, int largest_align, int cpu)
+{
+       struct rchan *rchan = ltt_channel->trans_channel_data;
+       struct rchan_buf *buf = *transport_data =
+                       rchan->buf[cpu];
+       struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, buf->cpu);
+       struct ltt_reserve_switch_offsets offsets;
+
+       offsets.reserve_commit_diff = 0;
+       offsets.size = 0;
+
+       /*
+        * Perform retryable operations.
+        */
+       if (__get_cpu_var(ltt_nesting) > 4) {
+               local_inc(&ltt_buf->events_lost);
+               return -EPERM;
+       }
+       do {
+               if (ltt_relay_try_reserve(ltt_channel, ltt_buf,
+                               rchan, buf, &offsets, data_size, tsc, rflags,
+                               largest_align))
+                       return -ENOSPC;
+       } while (local_cmpxchg(&ltt_buf->offset, offsets.old,
+                       offsets.end) != offsets.old);
+
+       /*
+        * Atomically update last_tsc. This update races against concurrent
+        * atomic updates, but the race will always cause supplementary full TSC
+        * events, never the opposite (missing a full TSC event when it would be
+        * needed).
+        */
+       save_last_tsc(ltt_buf, *tsc);
+
+       /*
+        * Push the reader if necessary
+        */
+       ltt_reserve_push_reader(ltt_channel, ltt_buf, rchan, buf, &offsets);
+
+       /*
+        * Switch old subbuffer if needed.
+        */
+       if (offsets.end_switch_old)
+               ltt_reserve_switch_old_subbuf(ltt_channel, ltt_buf, rchan, buf,
+                       &offsets, tsc);
+
+       /*
+        * Populate new subbuffer.
+        */
+       if (offsets.begin_switch)
+               ltt_reserve_switch_new_subbuf(ltt_channel, ltt_buf, rchan,
+                       buf, &offsets, tsc);
+
+       if (offsets.end_switch_current)
+               ltt_reserve_end_switch_current(ltt_channel, ltt_buf, rchan,
+                       buf, &offsets, tsc);
+
+       *slot_size = offsets.size;
+       *buf_offset = offsets.begin + offsets.before_hdr_pad;
+       return 0;
+}
+
+/*
+ * Force a sub-buffer switch for a per-cpu buffer. This operation is
+ * completely reentrant : can be called while tracing is active with
+ * absolutely no lock held.
+ *
+ * Note, however, that as a local_cmpxchg is used for some atomic
+ * operations, this function must be called from the CPU which owns the buffer
+ * for a ACTIVE flush.
+ */
+static notrace void ltt_force_switch(struct rchan_buf *buf,
+               enum force_switch_mode mode)
+{
+       struct ltt_channel_struct *ltt_channel =
+                       (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, buf->cpu);
+       struct rchan *rchan = ltt_channel->trans_channel_data;
+       struct ltt_reserve_switch_offsets offsets;
+       u64 tsc;
+
+       offsets.reserve_commit_diff = 0;
+       offsets.size = 0;
+
+       /*
+        * Perform retryable operations.
+        */
+       do {
+               if (ltt_relay_try_switch(mode, ltt_channel, ltt_buf,
+                               rchan, buf, &offsets, &tsc))
+                       return;
+       } while (local_cmpxchg(&ltt_buf->offset, offsets.old,
+                       offsets.end) != offsets.old);
+
+       /*
+        * Atomically update last_tsc. This update races against concurrent
+        * atomic updates, but the race will always cause supplementary full TSC
+        * events, never the opposite (missing a full TSC event when it would be
+        * needed).
+        */
+       save_last_tsc(ltt_buf, tsc);
+
+       /*
+        * Push the reader if necessary
+        */
+       if (mode == FORCE_ACTIVE)
+               ltt_reserve_push_reader(ltt_channel, ltt_buf, rchan,
+                                       buf, &offsets);
+
+       /*
+        * Switch old subbuffer if needed.
+        */
+       if (offsets.end_switch_old)
+               ltt_reserve_switch_old_subbuf(ltt_channel, ltt_buf, rchan, buf,
+                       &offsets, &tsc);
+
+       /*
+        * Populate new subbuffer.
+        */
+       if (mode == FORCE_ACTIVE)
+               ltt_reserve_switch_new_subbuf(ltt_channel,
+                       ltt_buf, rchan, buf, &offsets, &tsc);
+}
+
+/*
+ * for flight recording. must be called after relay_commit.
+ * This function decrements de subbuffer's lost_size each time the commit count
+ * reaches back the reserve offset (module subbuffer size). It is useful for
+ * crash dump.
+ * We use slot_size - 1 to make sure we deal correctly with the case where we
+ * fill the subbuffer completely (so the subbuf index stays in the previous
+ * subbuffer).
+ */
+#ifdef CONFIG_LTT_VMCORE
+static inline void ltt_write_commit_counter(struct rchan_buf *buf,
+               long buf_offset, size_t slot_size)
+{
+       struct ltt_channel_struct *ltt_channel =
+               (struct ltt_channel_struct *)buf->chan->private_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, buf->cpu);
+       struct ltt_subbuffer_header *header;
+       long offset, subbuf_idx, commit_count;
+       uint32_t lost_old, lost_new;
+
+       subbuf_idx = SUBBUF_INDEX(buf_offset - 1, buf->chan);
+       offset = buf_offset + slot_size;
+       header = (struct ltt_subbuffer_header *)
+                       ltt_relay_offset_address(buf,
+                               subbuf_idx * buf->chan->subbuf_size);
+       for (;;) {
+               lost_old = header->lost_size;
+               commit_count =
+                       local_read(&ltt_buf->commit_count[subbuf_idx]);
+               /* SUBBUF_OFFSET includes commit_count_mask */
+               if (!SUBBUF_OFFSET(offset - commit_count, buf->chan)) {
+                       lost_new = (uint32_t)buf->chan->subbuf_size
+                                  - SUBBUF_OFFSET(commit_count, buf->chan);
+                       lost_old = cmpxchg_local(&header->lost_size, lost_old,
+                                               lost_new);
+                       if (lost_old <= lost_new)
+                               break;
+               } else {
+                       break;
+               }
+       }
+}
+#else
+static inline void ltt_write_commit_counter(struct rchan_buf *buf,
+               long buf_offset, size_t slot_size)
+{
+}
+#endif
+
+/*
+ * Atomic unordered slot commit. Increments the commit count in the
+ * specified sub-buffer, and delivers it if necessary.
+ *
+ * Parameters:
+ *
+ * @ltt_channel : channel structure
+ * @transport_data: transport-specific data
+ * @buf_offset : offset following the event header.
+ * @slot_size : size of the reserved slot.
+ */
+static notrace void ltt_relay_commit_slot(
+               struct ltt_channel_struct *ltt_channel,
+               void **transport_data, long buf_offset, size_t slot_size)
+{
+       struct rchan_buf *buf = *transport_data;
+       struct ltt_channel_buf_struct *ltt_buf =
+                       percpu_ptr(ltt_channel->buf, buf->cpu);
+       struct rchan *rchan = buf->chan;
+       long offset_end = buf_offset;
+       long endidx = SUBBUF_INDEX(offset_end - 1, rchan);
+       long commit_count;
+
+       /* Must write slot data before incrementing commit count */
+       smp_wmb();
+       commit_count = local_add_return(slot_size,
+               &ltt_buf->commit_count[endidx]);
+       /* Check if all commits have been done */
+       if ((BUFFER_TRUNC(offset_end - 1, rchan)
+                       >> ltt_channel->n_subbufs_order)
+                       - ((commit_count - rchan->subbuf_size)
+                          & ltt_channel->commit_count_mask) == 0)
+               ltt_deliver(buf, endidx, NULL);
+       /*
+        * Update lost_size for each commit. It's needed only for extracting
+        * ltt buffers from vmcore, after crash.
+        */
+       ltt_write_commit_counter(buf, buf_offset, slot_size);
+}
+
+/*
+ * This is called with preemption disabled when user space has requested
+ * blocking mode.  If one of the active traces has free space below a
+ * specific threshold value, we reenable preemption and block.
+ */
+static int ltt_relay_user_blocking(struct ltt_trace_struct *trace,
+               unsigned int chan_index, size_t data_size,
+               struct user_dbg_data *dbg)
+{
+       struct rchan *rchan;
+       struct ltt_channel_buf_struct *ltt_buf;
+       struct ltt_channel_struct *channel;
+       struct rchan_buf *relay_buf;
+       int cpu;
+       DECLARE_WAITQUEUE(wait, current);
+
+       channel = &trace->channels[chan_index];
+       rchan = channel->trans_channel_data;
+       cpu = smp_processor_id();
+       relay_buf = rchan->buf[cpu];
+       ltt_buf = percpu_ptr(channel->buf, cpu);
+
+       /*
+        * Check if data is too big for the channel : do not
+        * block for it.
+        */
+       if (LTT_RESERVE_CRITICAL + data_size > relay_buf->chan->subbuf_size)
+               return 0;
+
+       /*
+        * If free space too low, we block. We restart from the
+        * beginning after we resume (cpu id may have changed
+        * while preemption is active).
+        */
+       spin_lock(&ltt_buf->full_lock);
+       if (!channel->overwrite) {
+               dbg->write = local_read(&ltt_buf->offset);
+               dbg->read = atomic_long_read(&ltt_buf->consumed);
+               dbg->avail_size = dbg->write + LTT_RESERVE_CRITICAL + data_size
+                                 - SUBBUF_TRUNC(dbg->read,
+                                                relay_buf->chan);
+               if (dbg->avail_size > rchan->alloc_size) {
+                       __set_current_state(TASK_INTERRUPTIBLE);
+                       add_wait_queue(&ltt_buf->write_wait, &wait);
+                       spin_unlock(&ltt_buf->full_lock);
+                       preempt_enable();
+                       schedule();
+                       __set_current_state(TASK_RUNNING);
+                       remove_wait_queue(&ltt_buf->write_wait, &wait);
+                       if (signal_pending(current))
+                               return -ERESTARTSYS;
+                       preempt_disable();
+                       return 1;
+               }
+       }
+       spin_unlock(&ltt_buf->full_lock);
+       return 0;
+}
+
+static void ltt_relay_print_user_errors(struct ltt_trace_struct *trace,
+               unsigned int chan_index, size_t data_size,
+               struct user_dbg_data *dbg, int cpu)
+{
+       struct rchan *rchan;
+       struct ltt_channel_buf_struct *ltt_buf;
+       struct ltt_channel_struct *channel;
+       struct rchan_buf *relay_buf;
+
+       channel = &trace->channels[chan_index];
+       rchan = channel->trans_channel_data;
+       relay_buf = rchan->buf[cpu];
+       ltt_buf = percpu_ptr(channel->buf, cpu);
+
+       printk(KERN_ERR "Error in LTT usertrace : "
+       "buffer full : event lost in blocking "
+       "mode. Increase LTT_RESERVE_CRITICAL.\n");
+       printk(KERN_ERR "LTT nesting level is %u.\n",
+               per_cpu(ltt_nesting, cpu));
+       printk(KERN_ERR "LTT avail size %lu.\n",
+               dbg->avail_size);
+       printk(KERN_ERR "avai write : %lu, read : %lu\n",
+                       dbg->write, dbg->read);
+
+       dbg->write = local_read(&ltt_buf->offset);
+       dbg->read = atomic_long_read(&ltt_buf->consumed);
+
+       printk(KERN_ERR "LTT cur size %lu.\n",
+               dbg->write + LTT_RESERVE_CRITICAL + data_size
+               - SUBBUF_TRUNC(dbg->read, relay_buf->chan));
+       printk(KERN_ERR "cur write : %lu, read : %lu\n",
+                       dbg->write, dbg->read);
+}
+
+static struct ltt_transport ltt_relay_transport = {
+       .name = "relay",
+       .owner = THIS_MODULE,
+       .ops = {
+               .create_dirs = ltt_relay_create_dirs,
+               .remove_dirs = ltt_relay_remove_dirs,
+               .create_channel = ltt_relay_create_channel,
+               .finish_channel = ltt_relay_finish_channel,
+               .remove_channel = ltt_relay_remove_channel,
+               .wakeup_channel = ltt_relay_async_wakeup_chan,
+               .commit_slot = ltt_relay_commit_slot,
+               .reserve_slot = ltt_relay_reserve_slot,
+               .user_blocking = ltt_relay_user_blocking,
+               .user_errors = ltt_relay_print_user_errors,
+       },
+};
+
+static int __init ltt_relay_init(void)
+{
+       printk(KERN_INFO "LTT : ltt-relay init\n");
+
+       ltt_file_operations = ltt_relay_file_operations;
+       ltt_file_operations.owner = THIS_MODULE;
+       ltt_file_operations.open = ltt_open;
+       ltt_file_operations.release = ltt_release;
+       ltt_file_operations.poll = ltt_poll;
+       ltt_file_operations.splice_read = ltt_relay_file_splice_read,
+       ltt_file_operations.ioctl = ltt_ioctl;
+#ifdef CONFIG_COMPAT
+       ltt_file_operations.compat_ioctl = ltt_compat_ioctl;
+#endif
+
+       ltt_transport_register(&ltt_relay_transport);
+
+       return 0;
+}
+
+static void __exit ltt_relay_exit(void)
+{
+       printk(KERN_INFO "LTT : ltt-relay exit\n");
+
+       ltt_transport_unregister(&ltt_relay_transport);
+}
+
+module_init(ltt_relay_init);
+module_exit(ltt_relay_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Mathieu Desnoyers");
+MODULE_DESCRIPTION("Linux Trace Toolkit Next Generation Lockless Relay");
This page took 0.044735 seconds and 4 git commands to generate.