trace: Add ring buffer stats to measure rate of events
authorVaibhav Nagarnaik <vnagarnaik@google.com>
Tue, 16 Aug 2011 21:46:16 +0000 (14:46 -0700)
committerSteven Rostedt <rostedt@goodmis.org>
Tue, 30 Aug 2011 16:27:45 +0000 (12:27 -0400)
The stats file under per_cpu folder provides the number of entries,
overruns and other statistics about the CPU ring buffer. However, the
numbers do not provide any indication of how full the ring buffer is in
bytes compared to the overall size in bytes. Also, it is helpful to know
the rate at which the cpu buffer is filling up.

This patch adds an entry "bytes: " in printed stats for per_cpu ring
buffer which provides the actual bytes consumed in the ring buffer. This
field includes the number of bytes used by recorded events and the
padding bytes added when moving the tail pointer to next page.

It also adds the following time stamps:
"oldest event ts:" - the oldest timestamp in the ring buffer
"now ts:"  - the timestamp at the time of reading

The field "now ts" provides a consistent time snapshot to the userspace
when being read. This is read from the same trace clock used by tracing
event timestamps.

Together, these values provide the rate at which the buffer is filling
up, from the formula:
bytes / (now_ts - oldest_event_ts)

Signed-off-by: Vaibhav Nagarnaik <vnagarnaik@google.com>
Cc: Michael Rubin <mrubin@google.com>
Cc: David Sharp <dhsharp@google.com>
Link: http://lkml.kernel.org/r/1313531179-9323-3-git-send-email-vnagarnaik@google.com
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
include/linux/ring_buffer.h
kernel/trace/ring_buffer.c
kernel/trace/trace.c

index b891de96000f35e3eaad5ecd7a6b5128fa6c65ff..67be0376d8e35e0593945dd9df507ed8528d02ce 100644 (file)
@@ -154,6 +154,8 @@ void ring_buffer_record_enable(struct ring_buffer *buffer);
 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu);
 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu);
 
+unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu);
+unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu);
 unsigned long ring_buffer_entries(struct ring_buffer *buffer);
 unsigned long ring_buffer_overruns(struct ring_buffer *buffer);
 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu);
index 731201bf4acc6af1efd7617bdc533ca1efaacabf..acf6b68dc4a8c2f9fa4e8b15313e3c4fc66b8b73 100644 (file)
@@ -488,12 +488,14 @@ struct ring_buffer_per_cpu {
        struct buffer_page              *reader_page;
        unsigned long                   lost_events;
        unsigned long                   last_overrun;
+       local_t                         entries_bytes;
        local_t                         commit_overrun;
        local_t                         overrun;
        local_t                         entries;
        local_t                         committing;
        local_t                         commits;
        unsigned long                   read;
+       unsigned long                   read_bytes;
        u64                             write_stamp;
        u64                             read_stamp;
 };
@@ -1708,6 +1710,7 @@ rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
                 * the counters.
                 */
                local_add(entries, &cpu_buffer->overrun);
+               local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
 
                /*
                 * The entries will be zeroed out when we move the
@@ -1863,6 +1866,9 @@ rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
        event = __rb_page_index(tail_page, tail);
        kmemcheck_annotate_bitfield(event, bitfield);
 
+       /* account for padding bytes */
+       local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
+
        /*
         * Save the original length to the meta data.
         * This will be used by the reader to add lost event
@@ -2054,6 +2060,9 @@ __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
        if (!tail)
                tail_page->page->time_stamp = ts;
 
+       /* account for these added bytes */
+       local_add(length, &cpu_buffer->entries_bytes);
+
        return event;
 }
 
@@ -2076,6 +2085,7 @@ rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
        if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
                unsigned long write_mask =
                        local_read(&bpage->write) & ~RB_WRITE_MASK;
+               unsigned long event_length = rb_event_length(event);
                /*
                 * This is on the tail page. It is possible that
                 * a write could come in and move the tail page
@@ -2085,8 +2095,11 @@ rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
                old_index += write_mask;
                new_index += write_mask;
                index = local_cmpxchg(&bpage->write, old_index, new_index);
-               if (index == old_index)
+               if (index == old_index) {
+                       /* update counters */
+                       local_sub(event_length, &cpu_buffer->entries_bytes);
                        return 1;
+               }
        }
 
        /* could not discard */
@@ -2660,6 +2673,58 @@ rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
                (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
 }
 
+/**
+ * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
+ * @buffer: The ring buffer
+ * @cpu: The per CPU buffer to read from.
+ */
+unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
+{
+       unsigned long flags;
+       struct ring_buffer_per_cpu *cpu_buffer;
+       struct buffer_page *bpage;
+       unsigned long ret;
+
+       if (!cpumask_test_cpu(cpu, buffer->cpumask))
+               return 0;
+
+       cpu_buffer = buffer->buffers[cpu];
+       spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
+       /*
+        * if the tail is on reader_page, oldest time stamp is on the reader
+        * page
+        */
+       if (cpu_buffer->tail_page == cpu_buffer->reader_page)
+               bpage = cpu_buffer->reader_page;
+       else
+               bpage = rb_set_head_page(cpu_buffer);
+       ret = bpage->page->time_stamp;
+       spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
+
+/**
+ * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
+ * @buffer: The ring buffer
+ * @cpu: The per CPU buffer to read from.
+ */
+unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
+{
+       struct ring_buffer_per_cpu *cpu_buffer;
+       unsigned long ret;
+
+       if (!cpumask_test_cpu(cpu, buffer->cpumask))
+               return 0;
+
+       cpu_buffer = buffer->buffers[cpu];
+       ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
+
 /**
  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
  * @buffer: The ring buffer
@@ -3527,11 +3592,13 @@ rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
        cpu_buffer->reader_page->read = 0;
 
        local_set(&cpu_buffer->commit_overrun, 0);
+       local_set(&cpu_buffer->entries_bytes, 0);
        local_set(&cpu_buffer->overrun, 0);
        local_set(&cpu_buffer->entries, 0);
        local_set(&cpu_buffer->committing, 0);
        local_set(&cpu_buffer->commits, 0);
        cpu_buffer->read = 0;
+       cpu_buffer->read_bytes = 0;
 
        cpu_buffer->write_stamp = 0;
        cpu_buffer->read_stamp = 0;
@@ -3918,6 +3985,7 @@ int ring_buffer_read_page(struct ring_buffer *buffer,
        } else {
                /* update the entry counter */
                cpu_buffer->read += rb_page_entries(reader);
+               cpu_buffer->read_bytes += BUF_PAGE_SIZE;
 
                /* swap the pages */
                rb_init_page(bpage);
index 01176788c387e30d0b72dfc65e245161e901bd71..b41907000bc605e5a65d8e475a70ac422df5011c 100644 (file)
@@ -4056,6 +4056,8 @@ tracing_stats_read(struct file *filp, char __user *ubuf,
        struct trace_array *tr = &global_trace;
        struct trace_seq *s;
        unsigned long cnt;
+       unsigned long long t;
+       unsigned long usec_rem;
 
        s = kmalloc(sizeof(*s), GFP_KERNEL);
        if (!s)
@@ -4072,6 +4074,17 @@ tracing_stats_read(struct file *filp, char __user *ubuf,
        cnt = ring_buffer_commit_overrun_cpu(tr->buffer, cpu);
        trace_seq_printf(s, "commit overrun: %ld\n", cnt);
 
+       cnt = ring_buffer_bytes_cpu(tr->buffer, cpu);
+       trace_seq_printf(s, "bytes: %ld\n", cnt);
+
+       t = ns2usecs(ring_buffer_oldest_event_ts(tr->buffer, cpu));
+       usec_rem = do_div(t, USEC_PER_SEC);
+       trace_seq_printf(s, "oldest event ts: %5llu.%06lu\n", t, usec_rem);
+
+       t = ns2usecs(ring_buffer_time_stamp(tr->buffer, cpu));
+       usec_rem = do_div(t, USEC_PER_SEC);
+       trace_seq_printf(s, "now ts: %5llu.%06lu\n", t, usec_rem);
+
        count = simple_read_from_buffer(ubuf, count, ppos, s->buffer, s->len);
 
        kfree(s);