On 24/03/16 10:56, Chris Wilson wrote:
On Wed, Mar 23, 2016 at 02:57:36PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Doing a lot of work in the interrupt handler introduces huge
latencies to the system as a whole.
Most dramatic effect can be seen by running an all engine
stress test like igt/gem_exec_nop/all where, when the kernel
config is lean enough, the whole system can be brought into
multi-second periods of complete non-interactivty. That can
look for example like this:
NMI watchdog: BUG: soft lockup - CPU#0 stuck for 23s! [kworker/u8:3:143]
Modules linked in: [redacted for brevity]
CPU: 0 PID: 143 Comm: kworker/u8:3 Tainted: G U L 4.5.0-160321+ #183
Hardware name: Intel Corporation Broadwell Client platform/WhiteTip Mountain 1
Workqueue: i915 gen6_pm_rps_work [i915]
task: ffff8800aae88000 ti: ffff8800aae90000 task.ti: ffff8800aae90000
RIP: 0010:[<ffffffff8104a3c2>] [<ffffffff8104a3c2>] __do_softirq+0x72/0x1d0
RSP: 0000:ffff88014f403f38 EFLAGS: 00000206
RAX: ffff8800aae94000 RBX: 0000000000000000 RCX: 00000000000006e0
RDX: 0000000000000020 RSI: 0000000004208060 RDI: 0000000000215d80
RBP: ffff88014f403f80 R08: 0000000b1b42c180 R09: 0000000000000022
R10: 0000000000000004 R11: 00000000ffffffff R12: 000000000000a030
R13: 0000000000000082 R14: ffff8800aa4d0080 R15: 0000000000000082
FS: 0000000000000000(0000) GS:ffff88014f400000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fa53b90c000 CR3: 0000000001a0a000 CR4: 00000000001406f0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
042080601b33869f ffff8800aae94000 00000000fffc2678 ffff88010000000a
0000000000000000 000000000000a030 0000000000005302 ffff8800aa4d0080
0000000000000206 ffff88014f403f90 ffffffff8104a716 ffff88014f403fa8
Call Trace:
<IRQ>
[<ffffffff8104a716>] irq_exit+0x86/0x90
[<ffffffff81031e7d>] smp_apic_timer_interrupt+0x3d/0x50
[<ffffffff814f3eac>] apic_timer_interrupt+0x7c/0x90
<EOI>
[<ffffffffa01c5b40>] ? gen8_write64+0x1a0/0x1a0 [i915]
[<ffffffff814f2b39>] ? _raw_spin_unlock_irqrestore+0x9/0x20
[<ffffffffa01c5c44>] gen8_write32+0x104/0x1a0 [i915]
[<ffffffff8132c6a2>] ? n_tty_receive_buf_common+0x372/0xae0
[<ffffffffa017cc9e>] gen6_set_rps_thresholds+0x1be/0x330 [i915]
[<ffffffffa017eaf0>] gen6_set_rps+0x70/0x200 [i915]
[<ffffffffa0185375>] intel_set_rps+0x25/0x30 [i915]
[<ffffffffa01768fd>] gen6_pm_rps_work+0x10d/0x2e0 [i915]
[<ffffffff81063852>] ? finish_task_switch+0x72/0x1c0
[<ffffffff8105ab29>] process_one_work+0x139/0x350
[<ffffffff8105b186>] worker_thread+0x126/0x490
[<ffffffff8105b060>] ? rescuer_thread+0x320/0x320
[<ffffffff8105fa64>] kthread+0xc4/0xe0
[<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170
[<ffffffff814f351f>] ret_from_fork+0x3f/0x70
[<ffffffff8105f9a0>] ? kthread_create_on_node+0x170/0x170
I could not explain, or find a code path, which would explain
a +20 second lockup, but from some instrumentation it was
apparent the interrupts off proportion of time was between
10-25% under heavy load which is quite bad.
By moving the GT interrupt handling to a tasklet in a most
simple way, the problem above disappears completely.
Perfect segue into gem_syslatency. I think gem_syslatency is the better
tool to correlate disruptive system behaviour. And then continue on with
gem_latency to demonstrate that is doesn't adversely affect our
performance.
Will do.
Also, gem_latency -n 100 shows 25% better throughput and CPU
usage, and 14% better latencies.
Mention the benefits of parallelising dispatch.
Hm, actually this should be the same as before I think.
As fairly hit-and-miss as perf testing is on these machines, it is
looking in favour of using tasklets vs the rt kthread. The numbers swing
between 2-10%, but consistently improves in the nop sync latencies.
There's still several hours to go in this run before we cover the
dispatch latenies, but so far reasonable.
(Hmm, looks like there may be a possible degredation on the single nop
dispatch but an improvement on the continuous nop dispatch.)
We can add all the numbers you get to the commit message as well.
I did not find any gains or regressions with Synmark2 or
GLbench under light testing. More benchmarking is certainly
required.
v2:
* execlists_lock should be taken as spin_lock_bh when
queuing work from userspace now. (Chris Wilson)
* uncore.lock must be taken with spin_lock_irq when
submitting requests since that now runs from either
softirq or process context.
There are a couple of execlist_lock usage outside of intel_lrc that may
or may not be useful to convert (low frequency reset / debug paths, so
way off the critical paths, but consistency in locking is invaluable).
Oh right, I've missed those.
+ tasklet_init(&engine->irq_tasklet, intel_lrc_irq_handler,
+ (unsigned long)engine);
I like trying to split lines to cluster arguments if possible. Here I
think intel_lrc_irq_handler pairs with engine,
tasklet_init(&engine->irq_tasklet,
intel_lrc_irq_handler, (unsigned long)engine);
*shrug*
Yeah it is nicer.
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index 221a94627aab..29810cba8a8c 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -266,6 +266,7 @@ struct intel_engine_cs {
} semaphore;
/* Execlists */
+ struct tasklet_struct irq_tasklet;
spinlock_t execlist_lock;
spinlock_t execlist_lock; /* used inside tasklet, use spin_lock_bh */
Will do.
It's looking good, but once this run completes, I'm going to repeat it
just to confirm how stable my numbers are.
Critical bugfix, improvements, simpler patch than my kthread
implementation,
Reviewed-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Okay I will respin with the above and we'll see.
Unfortunately my test platform just died so there will be a delay.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
