This patch introduces 'write_backward' bit to perf_event_attr, which
controls the direction of a ring buffer. After set, the corresponding
ring buffer is written from end to beginning. This feature is design to
support reading from overwritable ring buffer.
Ring buffer can be created by mapping a perf event fd. Kernel puts event
records into ring buffer, user tooling like perf fetch them from
address returned by mmap(). To prevent racing between kernel and tooling,
they communicate to each other through 'head' and 'tail' pointers.
Kernel maintains 'head' pointer, points it to the next free area (tail
of the last record). Tooling maintains 'tail' pointer, points it to the
tail of last consumed record (record has already been fetched). Kernel
determines the available space in a ring buffer using these two
pointers to avoid overwrite unfetched records.
By mapping without 'PROT_WRITE', an overwritable ring buffer is created.
Different from normal ring buffer, tooling is unable to maintain 'tail'
pointer because writing is forbidden. Therefore, for this type of ring
buffers, kernel overwrite old records unconditionally, works like flight
recorder. This feature would be useful if reading from overwritable ring
buffer were as easy as reading from normal ring buffer. However,
there's an obscure problem.
The following figure demonstrates a full overwritable ring buffer. In
this figure, the 'head' pointer points to the end of last record, and a
long record 'E' is pending. For a normal ring buffer, a 'tail' pointer
would have pointed to position (X), so kernel knows there's no more
space in the ring buffer. However, for an overwritable ring buffer,
kernel ignore the 'tail' pointer.
(X) head
. |
. V
+------+-------+----------+------+---+
|A....A|B.....B|C........C|D....D| |
+------+-------+----------+------+---+
Record 'A' is overwritten by event 'E':
head
|
V
+--+---+-------+----------+------+---+
|.E|..A|B.....B|C........C|D....D|E..|
+--+---+-------+----------+------+---+
Now tooling decides to read from this ring buffer. However, none of these
two natural positions, 'head' and the start of this ring buffer, are
pointing to the head of a record. Even the full ring buffer can be
accessed by tooling, it is unable to find a position to start decoding.
The first attempt tries to solve this problem AFAIK can be found from
[1]. It makes kernel to maintain 'tail' pointer: updates it when ring
buffer is half full. However, this approach introduces overhead to
fast path. Test result shows a 1% overhead [2]. In addition, this method
utilizes no more tham 50% records.
Another attempt can be found from [3], which allows putting the size of
an event at the end of each record. This approach allows tooling to find
records in a backward manner from 'head' pointer by reading size of a
record from its tail. However, because of alignment requirement, it
needs 8 bytes to record the size of a record, which is a huge waste. Its
performance is also not good, because more data need to be written.
This approach also introduces some extra branch instructions to fast
path.
'write_backward' is a better solution to this problem.
Following figure demonstrates the state of the overwritable ring buffer
when 'write_backward' is set before overwriting:
head
|
V
+---+------+----------+-------+------+
| |D....D|C........C|B.....B|A....A|
+---+------+----------+-------+------+
and after overwriting:
head
|
V
+---+------+----------+-------+---+--+
|..E|D....D|C........C|B.....B|A..|E.|
+---+------+----------+-------+---+--+
In each situation, 'head' points to the beginning of the newest record.
From this record, tooling can iterate over the full ring buffer and fetch
records one by one.
The only limitation that needs to be considered is back-to-back reading.
Due to the non-deterministic of user programs, it is impossible to ensure
the ring buffer keeps stable during reading. Consider an extreme situation:
tooling is scheduled out after reading record 'D', then a burst of events
come, eat up the whole ring buffer (one or multiple rounds). When the
tooling process comes back, reading after 'D' is incorrect now.
To prevent this problem, we need to find a way to ensure the ring buffer
is stable during reading. ioctl(PERF_EVENT_IOC_PAUSE_OUTPUT) is
suggested because its overhead is lower than
ioctl(PERF_EVENT_IOC_ENABLE).
By carefully verifying 'header' pointer, reader can avoid pausing the
ring-buffer. For example:
/* A union of all possible events */
union perf_event event;
p = head = perf_mmap__read_head();
while (true) {
/* copy header of next event */
fetch(&event.header, p, sizeof(event.header));
/* read 'head' pointer */
head = perf_mmap__read_head();
/* check overwritten: is the header good? */
if (!verify(sizeof(event.header), p, head))
break;
/* copy the whole event */
fetch(&event, p, event.header.size);
/* read 'head' pointer again */
head = perf_mmap__read_head();
/* is the whole event good? */
if (!verify(event.header.size, p, head))
break;
p += event.header.size;
}
However, the overhead is high because:
a) In-place decoding is not safe.
Copying-verifying-decoding is required.
b) Fetching 'head' pointer requires additional synchronization.
(From Alexei Starovoitov:
Even when this trick works, pause is needed for more than stability of
reading. When we collect the events into overwrite buffer we're waiting
for some other trigger (like all cpu utilization spike or just one cpu
running and all others are idle) and when it happens the buffer has
valuable info from the past. At this point new events are no longer
interesting and buffer should be paused, events read and unpaused until
next trigger comes.)
This patch utilizes event's default overflow_handler introduced
previously. perf_event_output_backward() is created as the default
overflow handler for backward ring buffers. To avoid extra overhead to
fast path, original perf_event_output() becomes __perf_event_output()
and marked '__always_inline'. In theory, there's no extra overhead
introduced to fast path.
Performance testing:
Calling 3000000 times of 'close(-1)', use gettimeofday() to check
duration. Use 'perf record -o /dev/null -e raw_syscalls:*' to capture
system calls. In ns.
Testing environment:
CPU : Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz
Kernel : v4.5.0
MEAN STDVAR
BASE 800214.950 2853.083
PRE1 2253846.700 9997.014
PRE2 2257495.540 8516.293
POST 2250896.100 8933.921
Where 'BASE' is pure performance without capturing. 'PRE1' is test
result of pure 'v4.5.0' kernel. 'PRE2' is test result before this
patch. 'POST' is test result after this patch. See [4] for the detailed
experimental setup.
Considering the stdvar, this patch doesn't introduce performance
overhead to the fast path.
[1] http://lkml.iu.edu/hypermail/linux/kernel/1304.1/04584.html
[2] http://lkml.iu.edu/hypermail/linux/kernel/1307.1/00535.html
[3] http://lkml.iu.edu/hypermail/linux/kernel/1512.0/01265.html
[4] http://lkml.kernel.org/g/56F89DCD.1040202@huawei.com
Signed-off-by: Wang Nan <wangnan0@huawei.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Cc: <acme@kernel.org>
Cc: <pi3orama@163.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: He Kuang <hekuang@huawei.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Zefan Li <lizefan@huawei.com>
Link: http://lkml.kernel.org/r/1459865478-53413-1-git-send-email-wangnan0@huawei.com
[ Fixed the changelog some more. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
9ecda41acb