From: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > Sent: 16 January 2020 07:40 > Quoting Daniel Vetter (2020-01-16 06:52:42) > > On Wed, Jan 15, 2020 at 08:52:45PM +0000, Chris Wilson wrote: > > > Since we may try and flush the cachelines associated with large buffers > > > (an 8K framebuffer is about 128MiB, even before we try HDR), this leads > > > to unacceptably long latencies (when using a voluntary CONFIG_PREEMPT). > > > If we call cond_resched() between each sg chunk, that it about every 128 > > > pages, we have a natural break point in which to check if the process > > > needs to be rescheduled. Naturally, this means that drm_clflush_sg() can > > > only be called from process context -- which is true at the moment. The > > > other clflush routines remain usable from atomic context. > > > > > > Even though flushing large objects takes a demonstrable amount to time > > > to flush all the cachelines, clflush is still preferred over a > > > system-wide wbinvd as the latter has unpredictable latencies affecting > > > the whole system not just the local task. > > > > > > Reported-by: David Laight <David.Laight@xxxxxxxxxx> > > > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > > > Cc: David Laight <David.Laight@xxxxxxxxxx> > > > > The original bug report is complaining about latencies for SCHED_RT > > threads, on a system that doesn't even use CONFIG_PREEMPT. I'm not sure > > it's terribly valid to cater to that use-case - all the desktop distros > > seem a lot more reasonable. So firmly *shrug* from my side ... > > Yeah, I had the same immediate response to the complaint), but otoh we've > inserted cond_resched() before when it looks like may consume entire > jiffies inside a loop. At the very minimum, we should have a > might_sleep() here and a reminder that this can be very slow (remember > byt?). I'm using RT to get more deterministic scheduling to look for long scheduling delays, rather than because we need very tight scheduling. Delays of several 100us aren't a real problem. The problem with CONFIG_PREEMPT is that the distros don't enable it and it isn't a command line option. So it is really useless unless you are able/willing to build your own kernel. I could run the code under the normal scheduler with 'nice -19'. I stlll wouldn't expect to have all but one cpu idle when I've just done a cv_broadcast() to wake up a lot of threads. I've added 'if (!(++i & 31)) cond_resched();' after the drm_clfulsh_page() call in drm_cflush_sg(). In my case that it 3600/32 reschedules in 3.3ms - plenty. However there is a call from __i915_gem_objet_set_pages() that is preceded by a lockdep_assert_held() check - so mustn't sleep. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales) _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
