Re: [PATCH] rcu/tree: Add a trace event for RCU stall warnings

[Uladzislau Rezki <urezki@gmail.com>]

On Thu, Feb 18, 2021 at 12:24:14PM -0500, Mathieu Desnoyers wrote:
> ----- On Feb 18, 2021, at 11:58 AM, Uladzislau Rezki urezki@gmail.com wrote:
> 
> > On Thu, Feb 18, 2021 at 10:50:53AM -0500, Mathieu Desnoyers wrote:
> >> ----- On Feb 18, 2021, at 10:06 AM, paulmck paulmck@kernel.org wrote:
> >> 
> >> > On Thu, Feb 18, 2021 at 09:17:01AM -0500, Mathieu Desnoyers wrote:
> >> >> ----- On Feb 18, 2021, at 7:52 AM, Sangmoon Kim sangmoon.kim@samsung.com wrote:
> >> >> 
> >> >> >> -----Original Message-----
> >> >> >> From: neeraju=codeaurora.org@mg.codeaurora.org
> >> >> >> <neeraju=codeaurora.org@mg.codeaurora.org>
> >> >> >> Sent: Thursday, February 18, 2021 3:18 AM
> >> >> >> 
> >> >> >> Hi Sangmoon,
> >> >> >> 
> >> >> >> On 2/17/2021 7:19 PM, Sangmoon Kim wrote:
> >> >> >> >> -----Original Message-----
> >> >> >> >> From: Paul E. McKenney <paulmck@kernel.org>
> >> >> >> >> Sent: Wednesday, February 17, 2021 2:50 AM
> >> >> >> >>
> >> >> >> >> On Mon, Feb 15, 2021 at 05:53:25PM +0900, Sangmoon Kim wrote:
> >> >> >> >>> The event allows us to trace the RCU stall when
> >> >> >> >>> sysctl_panic_on_rcu_stall is disabled.
> >> >> >> >>>
> >> >> >> >>> The first parameter is the name of RCU flavour like other trace
> >> >> >> >>> events. The second one shows us which function detected stalls.
> >> >> >> >>>
> >> >> >> >>> The RCU stall is mainly caused by external factors such as interrupt
> >> >> >> >>> handling or task scheduling or something else. Therefore, this event
> >> >> >> >>> uses TRACE_EVENT macro, not dedicated one, so that someone interested
> >> >> >> >>> in the RCU stall can use it without CONFIG_RCU_TRACE.
> >> >> >> >>>
> >> >> >> >>> Signed-off-by: Sangmoon Kim <sangmoon.kim@samsung.com>
> >> >> >> >>
> >> >> >> >> The patch looks plausible, but I have to ask...  Why not instead just
> >> >> >> >> get the existing information out of the console log?
> >> >> >> >>
> >> >> >> >> 							Thanx, Paul
> >> >> >> >
> >> >> >> > This can provide a trigger point for the RCU stall warning.
> >> >> >> > If a module in the kernel wants to trace the stall for debugging purposes,
> >> >> >> > there is a cost of continuing to parse the console log.
> >> >> >> > This tracepoint is useful because it is hard to pay these costs
> >> >> >> > especially on mobile devices.
> >> >> >> >
> >> >> >> > Thanks,
> >> >> >> > Sangmoon
> >> >> >> >
> >> >> >> 
> >> >> >> So, the idea here is to register to these trace events from kernel
> >> >> >> module and use that for debugging? Just curious what debugging action
> >> >> >> module does on these traces, as they have limited information
> >> >> >> about the stall, compared to console stall warnings, which gives a
> >> >> >> much more detailed information about stall.
> >> >> >> 
> >> >> >> 
> >> >> >> Thanks
> >> >> >> Neeraj
> >> >> > 
> >> >> > Hi Neeraj,
> >> >> > 
> >> >> > Yes, a module can log the stall occurence using the trace, although
> >> >> > there is no detailed information. If the kernel panic occurs for some
> >> >> > reasons, the debugging report generated by the module can include that
> >> >> > RCU stall warning has occurred before.
> >> >> > 
> >> >> > In addition, it's just an idea now, when a trace event happens, the
> >> >> > module can store the console log including detailed information, or may
> >> >> > also obtain CPU/task information by parsing the console log.
> >> >> 
> >> >> Adding a new tracepoint is not just about what is extracted by this specific
> >> >> tracepoint, but rather how it can be analyzed when combined with all other
> >> >> relevant
> >> >> tracepoints.
> >> >> 
> >> >> For instance, if we have this added RCU stall warning added, here is how it can
> >> >> be
> >> >> used with the upcoming LTTng 2.13, which implements the "event notification"
> >> >> triggers
> >> >> feature:
> >> >> 
> >> >> 1) Setup "flight recorder" (snapshot) tracing to trace into a circular ring
> >> >> buffer,
> >> >>    enabling the following tracepoints:
> >> >>    - kernel activity (meaning all other RCU event, scheduling, irq, workqueues,
> >> >>    ...),
> >> >>    - this new RCU stall warning event.
> >> >> 
> >> >> 2) Add a "callstack-kernel" context to the RCU stall warning event. This will
> >> >> sample
> >> >>    the kernel stack when the event is hit. This will provide information similar to
> >> >>    the stack trace gathered into the console log on OOPS.
> >> >> 
> >> >> 3) Enable a trigger waiting on the RCU stall warning tracepoint to be hit. On
> >> >> this
> >> >>    trigger, actions can be associated, such as capturing a snapshot or waking up
> >> >>    an external user-space process to perform specific actions.
> >> >> 
> >> >> So you end up with a snapshot containing the sequence of events leading to the
> >> >> RCU stall warning, with a kernel stack trace of the context causing the stall
> >> >> warning to be emitted.
> >> >> 
> >> >> I would argue that this information is more complete than just the stack trace
> >> >> extracted through the console log.
> >> > 
> >> > I am not so sure about that.  RCU CPU stall warnings dump quite a bit more
> >> > than a stack trace to the console.  Which is why I am concerned about the
> >> > proverbial camel's nose in the tent.  ;-)
> >> > 
> >> > So Sangmoon, what is it that you really need for this to be useful to you?
> >> > 
> >> > Or am I missing your point?  (Either Mathieu's or Sangmoon's.)
> >> 
> >> Well there is a tracepoint to dump the console's content into the tracing
> >> buffers
> >> as well, so technically this existing RCU stall warning information could be
> >> extracted
> >> into a trace as well.
> >> 
> >> AFAIU, what the new "rcu stall warning" tracepoint provides is an easy way to
> >> hook on a specific
> >> event to trigger trace capture, without requiring to parse the console log
> >> continuously,
> >> and a way to know when the stall warning happens in time within the trace
> >> time-line.
> >> 
> >> That being said, there may be use-cases for extracting more details about the
> >> RCU stall
> >> as tracepoint event fields to make it more convenient, but it does not appear to
> >> be
> >> necessary considered that the console can be saved into trace buffers as well.
> >> 
> > Could you please clarify how the kernel ring-buffer can also be routed over
> > trace buffer?
> > Probably i am not aware of if it is possible on latest kernel.
> 
> This can be done by enabling the "printk console" tracepoint. See the call to
> trace_console_rcuidle() in kernel/printk/printk.c. The tracepoint is defined in
> include/trace/events/printk.h.
> 
OK, i see. At least "error" messages get routed:

<snip>
root@seldlx26095:/home/CORPUSERS/23060734# cat /sys/kernel/debug/tracing/trace_pipe 
     kworker/4:0-9693    [004] d..1 627080.409625: console: [627080.402633] cp210x ttyUSB0: failed set request 0x7 status: -19
     kworker/4:0-9693    [004] d..1 627080.409633: console: [627080.402642] cp210x ttyUSB0: failed set request 0x12 status: -19
     kworker/4:0-9693    [004] d..1 627080.409639: console: [627080.402647] cp210x ttyUSB0: failed set request 0x0 status: -19
<snip>

Thanks!

--
Vlad Rezki