On 2016年06月29日 20:43, Byungchul Park wrote:
On Mon, Jun 20, 2016 at 04:50:37PM +0900, byungchul.park wrote:
-----Original Message-----
From: xinhui [mailto:xinhui....@linux.vnet.ibm.com]
Sent: Monday, June 20, 2016 4:29 PM
To: Byungchul Park; pet...@infradead.org; mi...@kernel.org
Cc: linux-kernel@vger.kernel.org; npig...@suse.de; wal...@google.com;
a...@suse.de; t...@inhelltoy.tec.linutronix.de
Subject: Re: [RFC 12/12] x86/dumpstack: Optimize save_stack_trace
On 2016年06月20日 12:55, Byungchul Park wrote:
Currently, x86 implementation of save_stack_trace() is walking all stack
region word by word regardless of what the trace->max_entries is.
However, it's unnecessary to walk after already fulfilling caller's
requirement, say, if trace->nr_entries >= trace->max_entries is true.
For example, CONFIG_LOCKDEP_CROSSRELEASE implementation calls
save_stack_trace() with max_entries = 5 frequently. I measured its
overhead and printed its difference of sched_clock() with my QEMU x86
machine.
The latency was improved over 70% when trace->max_entries = 5.
[snip]
+static int save_stack_end(void *data)
+{
+ struct stack_trace *trace = data;
+ return trace->nr_entries >= trace->max_entries;
+}
+
static const struct stacktrace_ops save_stack_ops = {
.stack = save_stack_stack,
.address = save_stack_address,
then why not check the return value of ->address(), -1 indicate there is
no room to store any pointer.
Hello,
Indeed. It also looks good to me even though it has to propagate the condition
between callback functions. I will modify it if it's better.
Do you also think it would be better to make it propagate the result of
->address() rather than add a new callback, say, end_walk?
It's up to you. In my opinion, end_walk is better for reading.
Thank you.
Byungchul
.walk_stack = print_context_stack,
+ .end_walk = save_stack_end,
};
static const struct stacktrace_ops save_stack_ops_nosched = {
