ring-buffer: Get timestamp after event is allocated
authorSteven Rostedt (Red Hat) <rostedt@goodmis.org>
Fri, 29 May 2015 13:40:18 +0000 (09:40 -0400)
committerSteven Rostedt <rostedt@goodmis.org>
Tue, 21 Jul 2015 02:30:48 +0000 (22:30 -0400)
Move the capturing of the timestamp to after an event is allocated.
If the event is not a commit (where it is an event that preempted
another event), then no timestamp is needed, because the delta of
nested events is always zero.

If the event starts on a new page, no delta needs to be calculated
as the full timestamp will be added to the page header, and the
event will have a delta of zero.

Now if the event requires a time extend (the delta does not fit
in the 27 bit delta slot in the header), then the event is discarded,
the length is extended to hold the TIME_EXTEND event that allows for
a 59 bit delta, and the commit is tried again.

If the event can't be discarded (another event came in after it),
then the TIME_EXTEND is added directly to the allocated event and
the rest of the event is given padding.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
kernel/trace/ring_buffer.c

index a78d4ee..b5ed553 100644 (file)
@@ -2009,7 +2009,7 @@ rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
  * and with this, we can determine what to place into the
  * data field.
  */
-static void
+static void __always_inline
 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
                struct ring_buffer_event *event,
                struct rb_event_info *info)
@@ -2017,10 +2017,6 @@ rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
        unsigned length = info->length;
        u64 delta = info->delta;
 
-       /* Only a commit updates the timestamp */
-       if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
-               delta = 0;
-
        /*
         * If we need to add a timestamp, then we
         * add it to the start of the resevered space.
@@ -2286,6 +2282,8 @@ rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
        local_sub(length, &tail_page->write);
 }
 
+static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
+
 /*
  * This is the slow path, force gcc not to inline it.
  */
@@ -2300,6 +2298,16 @@ rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
        int ret;
        u64 ts;
 
+       /*
+        * If the event had a timestamp attached to it, remove it.
+        * The first event on a page (nested or not) always uses
+        * the full timestamp of the new page.
+        */
+       if (info->add_timestamp) {
+               info->add_timestamp = 0;
+               info->length -= RB_LEN_TIME_EXTEND;
+       }
+
        next_page = tail_page;
 
        rb_inc_page(cpu_buffer, &next_page);
@@ -2386,6 +2394,11 @@ rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
 
        rb_reset_tail(cpu_buffer, tail, info);
 
+       /* Commit what we have for now to update timestamps */
+       rb_end_commit(cpu_buffer);
+       /* rb_end_commit() decs committing */
+       local_inc(&cpu_buffer->committing);
+
        /* fail and let the caller try again */
        return ERR_PTR(-EAGAIN);
 
@@ -2403,10 +2416,23 @@ static inline bool sched_clock_stable(void)
 }
 #endif
 
+static inline int
+rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
+                 struct ring_buffer_event *event);
+static inline void rb_event_discard(struct ring_buffer_event *event);
+static void
+rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
+                     struct ring_buffer_event *event);
+
 static noinline void
 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
+                   struct ring_buffer_event *event,
                    struct rb_event_info *info)
 {
+       struct ring_buffer_event *padding;
+       int length;
+       int size;
+
        WARN_ONCE(info->delta > (1ULL << 59),
                  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
                  (unsigned long long)info->delta,
@@ -2416,7 +2442,61 @@ rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
                  "If you just came from a suspend/resume,\n"
                  "please switch to the trace global clock:\n"
                  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
-       info->add_timestamp = 1;
+
+       /*
+        * Discarding this event to add a timestamp in front, but
+        * we still need to update the length of it to perform the discard.
+        */
+       rb_update_event(cpu_buffer, event, info);
+
+       if (rb_try_to_discard(cpu_buffer, event)) {
+               info->add_timestamp = 1;
+               /*
+                * The time delta since the last event is too big to
+                * hold in the time field of the event, then we append a
+                * TIME EXTEND event ahead of the data event.
+                */
+               info->length += RB_LEN_TIME_EXTEND;
+               return;
+       }
+
+       /*
+        * Humpf! An event came in after this one, and because it is not a
+        * commit, it will have a delta of zero, thus, it will take on
+        * the timestamp of the previous commit, which happened a long time
+        * ago (we need to add a timestamp, remember?).
+        * We need to add the timestamp here. A timestamp is a fixed size
+        * of 8 bytes. That means the rest of the event needs to be
+        * padding.
+        */
+       size = info->length - RB_LEN_TIME_EXTEND;
+
+       /* The padding will have a delta of 1 */
+       if (size)
+               info->delta--;
+
+       padding = rb_add_time_stamp(event, info->delta);
+
+       if (size) {
+               length = info->length;
+               info->delta = 0;
+               info->length = size;
+               rb_update_event(cpu_buffer, padding, info);
+
+               rb_event_discard(padding);
+
+               /* Still visible, need to update write_stamp */
+               rb_update_write_stamp(cpu_buffer, event);
+
+               /* Still need to commit the padding. */
+               rb_end_commit(cpu_buffer);
+
+               /* rb_end_commit() decs committing */
+               local_inc(&cpu_buffer->committing);
+
+               /* The next iteration still uses the original length */
+               info->length = length;
+       }
 }
 
 static struct ring_buffer_event *
@@ -2426,14 +2506,7 @@ __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
        struct ring_buffer_event *event;
        struct buffer_page *tail_page;
        unsigned long tail, write;
-
-       /*
-        * If the time delta since the last event is too big to
-        * hold in the time field of the event, then we append a
-        * TIME EXTEND event ahead of the data event.
-        */
-       if (unlikely(info->add_timestamp))
-               info->length += RB_LEN_TIME_EXTEND;
+       bool is_commit;
 
        tail_page = info->tail_page = cpu_buffer->tail_page;
        write = local_add_return(info->length, &tail_page->write);
@@ -2442,31 +2515,42 @@ __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
        write &= RB_WRITE_MASK;
        tail = write - info->length;
 
-       /*
-        * If this is the first commit on the page, then it has the same
-        * timestamp as the page itself.
-        */
-       if (!tail)
-               info->delta = 0;
-
        /* See if we shot pass the end of this buffer page */
        if (unlikely(write > BUF_PAGE_SIZE))
                return rb_move_tail(cpu_buffer, tail, info);
 
        /* We reserved something on the buffer */
-
        event = __rb_page_index(tail_page, tail);
-       kmemcheck_annotate_bitfield(event, bitfield);
-       rb_update_event(cpu_buffer, event, info);
-
-       local_inc(&tail_page->entries);
 
        /*
-        * If this is the first commit on the page, then update
-        * its timestamp.
+        * If this is the first commit on the page, then it has the same
+        * timestamp as the page itself, otherwise we need to figure out
+        * the delta.
         */
-       if (!tail)
+       info->ts = rb_time_stamp(cpu_buffer->buffer);
+       is_commit = rb_event_is_commit(cpu_buffer, event);
+
+       /* Commits are special (non nested events) */
+       info->delta = is_commit ? info->ts - cpu_buffer->write_stamp : 0;
+
+       if (!tail) {
+               /*
+                * If this is the first commit on the page, set the
+                * page to its timestamp.
+                */
                tail_page->page->time_stamp = info->ts;
+               info->delta = 0;
+
+       } else if (unlikely(test_time_stamp(info->delta)) &&
+                  !info->add_timestamp) {
+               rb_handle_timestamp(cpu_buffer, event, info);
+               return ERR_PTR(-EAGAIN);
+       }
+
+       kmemcheck_annotate_bitfield(event, bitfield);
+       rb_update_event(cpu_buffer, event, info);
+
+       local_inc(&tail_page->entries);
 
        /* account for these added bytes */
        local_add(info->length, &cpu_buffer->entries_bytes);
@@ -2560,7 +2644,6 @@ rb_reserve_next_event(struct ring_buffer *buffer,
        struct ring_buffer_event *event;
        struct rb_event_info info;
        int nr_loops = 0;
-       u64 diff;
 
        rb_start_commit(cpu_buffer);
 
@@ -2578,12 +2661,9 @@ rb_reserve_next_event(struct ring_buffer *buffer,
                return NULL;
        }
 #endif
-
        info.length = rb_calculate_event_length(length);
- again:
        info.add_timestamp = 0;
-       info.delta = 0;
-
+ again:
        /*
         * We allow for interrupts to reenter here and do a trace.
         * If one does, it will cause this original code to loop
@@ -2596,19 +2676,6 @@ rb_reserve_next_event(struct ring_buffer *buffer,
        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
                goto out_fail;
 
-       info.ts = rb_time_stamp(cpu_buffer->buffer);
-       diff = info.ts - cpu_buffer->write_stamp;
-
-       /* make sure this diff is calculated here */
-       barrier();
-
-       /* Did the write stamp get updated already? */
-       if (likely(info.ts >= cpu_buffer->write_stamp)) {
-               info.delta = diff;
-               if (unlikely(test_time_stamp(info.delta)))
-                       rb_handle_timestamp(cpu_buffer, &info);
-       }
-
        event = __rb_reserve_next(cpu_buffer, &info);
 
        if (unlikely(PTR_ERR(event) == -EAGAIN))