drm_cflush_sg() loops for over 3ms

From: David Laight
Date: Mon Jan 13 2020 - 09:34:47 EST


I've been looking at why some RT processes don't get scheduled promptly.
In my test the RT process's affinity ties it to a single cpu (this may not be such
a good idea as it seems).

What I've found is that the Intel i915 graphics driver uses the 'events_unbound'
kernel worker thread to periodically execute drm_cflush_sg().
(see https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/drm_cache.c)

I'm guessing this is to ensure that any writes to graphics memory become
visible is a semi-timely manner.

This loop takes about 1us per iteration split fairly evenly between whatever is in
for_each_sg_page() and drm_cflush_page().
With a 2560x1440 display the loop count is 3600 (4 bytes/pixel) and the whole
function takes around 3.3ms.

Since the kernel isn't pre-emptive (I though that wasn't much harder than SMP)
nothing else can run on that cpu until the loop finishes.

Adding a cond_resched() to the loop (maybe every 64 iterations) will
allow higher priority processes to run.
But really the code needs to be a lot faster.

I actually suspect that the (I assume IPI based) wbinv_on_all_cpus() would be
a lot faster - especially done by a per-cpu work queue?

I had moderate difficulty getting from the process (kworker/u8:3) to the
name of the worker thread pool, never mind the actual work.
Fortunately it runs so long that some of the output from 'echo t >/proc/sysrq-trigger'
still linked the pid (which I knew from ftrace scheduler events (and schedviz))
to the actual work item name.
(Oh, after I'd written a program to tidy up the raw ftrace output so schedviz
didn't barf on a trace that had wrapped.)

Is there anything in /proc (etc) that shows all the work queues and their current
work?

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)