Re: [PATCH v4] drm/trace: Buffer DRM logs in a ringbuffer accessible via debugfs

[Sean Paul <sean@poorly.run>]

On Wed, Jan 15, 2020 at 02:01:19PM +0000, Chris Wilson wrote:
> Quoting Sean Paul (2020-01-15 13:41:58)
> > On Wed, Jan 15, 2020 at 10:36:36AM +0000, Chris Wilson wrote:
> > > Quoting Sean Paul (2020-01-14 17:21:43)
> > > > From: Sean Paul <seanpaul@chromium.org>
> > > > 
> > > > This patch uses a ring_buffer to keep a "flight recorder" (name credit Weston)
> > > > of DRM logs for a specified set of debug categories. The user writes a
> > > > bitmask of debug categories to the "trace_mask" node and can read log
> > > > messages from the "trace" node.
> > > > 
> > > > These nodes currently exist in debugfs under the dri directory. I
> > > > intended on exposing all of this through tracefs originally, but the
> > > > tracefs entry points are not exposed, so there's no way to create
> > > > tracefs files from drivers at the moment. I think it would be a
> > > > worthwhile endeavour, but one requiring more time and conversation to
> > > > ensure the drm traces fit somewhere sensible.
> > > 
> > > Fwiw, I have a need for client orientated debug message store, with
> > > the primary purpose of figuring out -EINVAL. We need per-client so we can
> > > put sensitive information about the potentially buggy client behaviour,
> > > and of course it needs to be accessible by the non-privileged client.
> > > 
> > > On the execution side, it's easy to keep track of the client so we could
> > > trace execution flow per client, within reason. And we could do
> > > similarly for kms clients.
> > 
> > Could you build such a thing with drm_trace underpinning it, just put the
> > pertinent information in the message?
> 
> Not as is. The global has to go, and there's no use for debugfs. So we
> are just left with a sprintf() around a ring_buffer. I am left in the
> same position as just wanting to generalise tracek to take the ringbuffer
> as a parameter.
> 

Ah, I think I see what you're getting at now. I think it would be reasonable to
split out a drm_trace_buffer from the current code for this purpose. We could
have an interface like:

struct drm_trace_buffer *drm_trace_buffer_init(unsigned int num_pages);
int drm_trace_buffer_resize(struct drm_trace_buffer *buf, unsigned int num_pages);
int drm_trace_buffer_printf(struct drm_trace_buffer *buf, const char *format, ...);
int drm_trace_buffer_output(struct seq_file *seq);
void drm_trace_buffer_cleanup(struct drm_trace_buffer *buf);

Then to Joonas' point, we could have drm_trace_log which uses this interface to
mirror the logs with a debugfs interface.

Would that work for your purpose?


> > > Just chiming to say, I don't think a duplicate of dmesg hidden inside
> > > debugfs achieves much. But a generic tracek-esque ringbuf would be very
> > > useful -- even if only so we can separate our GEM_TRACE from the global
> > > tracek.
> > 
> > I think that's essentially what we've got, I've just narrowly focused on
> > surfacing debug logs. If drm_trace_printf were exported, replacing
> > GEM_TRACE would be as simple as s/trace_printk/drm_trace_printf/. Initially I
> > thought exporting it to drivers would be a bad idea, but I'm open to changing my
> > mind on this as long as drivers are using it responsibly. 
> 
> I definitely can't make the mistake of flooding kms tracing with
> overwhelming execution traces -- we can't go back to mixing kms traces
> with execution traces.

Yeah, I assumed this wouldn't be enabled during normal operation, just for
debugging (as it is used now).

Sean

> -Chris

-- 
Sean Paul, Software Engineer, Google / Chromium OS