An interesting discussion regarding "hybrid interrupt polling" for NVMe came to the conclusion that the ideal busyspin before sleeping was half of the expected request latency (and better if it was already halfway through that request). This suggested that we too should look again at our tradeoff between spinning and waiting. Currently, our spin simply tries to hide the cost of enabling the interrupt, which is good to avoid penalising nop requests (i.e. test throughput) and not much else. Studying real world workloads suggests that a spin of upto 500us can dramatically boost performance, but the suggestion is that this is not from avoiding interrupt latency per-se, but from secondary effects of sleeping such as allowing the CPU reduce cstate and context switch away. In a truly hybrid interrupt polling scheme, we would aim to sleep until just before the request completed and then wake up in advance of the interrupt and do a quick poll to handle completion. This is tricky for ourselves at the moment as we are not recording request times, and since we allow preemption, our requests are not on as a nicely ordered timeline as IO. However, the idea is interesting, for it will certainly help us decide when busyspinning is worthwhile. v2: Expose the spin setting via Kconfig options for easier adjustment and testing. v3: Don't get caught sneaking in a change to the busyspin parameters. v4: Explain more about the "hybrid interrupt polling" scheme that we want to migrate towards. Suggested-by: Sagar Kamble <sagar.a.kamble@xxxxxxxxx> References: http://events.linuxfoundation.org/sites/events/files/slides/lemoal-nvme-polling-vault-2017-final_0.pdf Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> Cc: Sagar Kamble <sagar.a.kamble@xxxxxxxxx> Cc: Eero Tamminen <eero.t.tamminen@xxxxxxxxx> Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> Cc: Ben Widawsky <ben@xxxxxxxxxxxx> Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx> Cc: Michał Winiarski <michal.winiarski@xxxxxxxxx> Reviewed-by: Sagar Kamble <sagar.a.kamble@xxxxxxxxx> --- drivers/gpu/drm/i915/Kconfig | 6 ++++++ drivers/gpu/drm/i915/Kconfig.profile | 13 +++++++++++++ drivers/gpu/drm/i915/i915_request.c | 27 +++++++++++++++++++++++++-- 3 files changed, 44 insertions(+), 2 deletions(-) create mode 100644 drivers/gpu/drm/i915/Kconfig.profile diff --git a/drivers/gpu/drm/i915/Kconfig b/drivers/gpu/drm/i915/Kconfig index 148be8e1a090..f0556310b851 100644 --- a/drivers/gpu/drm/i915/Kconfig +++ b/drivers/gpu/drm/i915/Kconfig @@ -133,3 +133,9 @@ depends on DRM_I915 depends on EXPERT source "drivers/gpu/drm/i915/Kconfig.debug" endmenu + +menu "drm/i915 Profile Guided Optimisation" + visible if EXPERT + depends on DRM_I915 + source "drivers/gpu/drm/i915/Kconfig.profile" +endmenu diff --git a/drivers/gpu/drm/i915/Kconfig.profile b/drivers/gpu/drm/i915/Kconfig.profile new file mode 100644 index 000000000000..0e5db98da8f3 --- /dev/null +++ b/drivers/gpu/drm/i915/Kconfig.profile @@ -0,0 +1,13 @@ +config DRM_I915_SPIN_REQUEST + int + default 5 # microseconds + help + Before sleeping waiting for a request (GPU operation) to complete, + we may spend some time polling for its completion. As the IRQ may + take a non-negligible time to setup, we do a short spin first to + check if the request will complete in the time it would have taken + us to enable the interrupt. + + May be 0 to disable the initial spin. In practice, we estimate + the cost of enabling the interrupt (if currently disabled) to be + a few microseconds. diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index b836721d3b13..b1f00b59bb95 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -1340,8 +1340,31 @@ long i915_request_wait(struct i915_request *rq, trace_i915_request_wait_begin(rq, flags); - /* Optimistic short spin before touching IRQs */ - if (__i915_spin_request(rq, state, 5)) + /* + * Optimistic spin before touching IRQs. + * + * We may use a rather large value here to offset the penalty of + * switching away from the active task. Frequently, the client will + * wait upon an old swapbuffer to throttle itself to remain within a + * frame of the gpu. If the client is running in lockstep with the gpu, + * then it should not be waiting long at all, and a sleep now will incur + * extra scheduler latency in producing the next frame. To try to + * avoid adding the cost of enabling/disabling the interrupt to the + * short wait, we first spin to see if the request would have completed + * in the time taken to setup the interrupt. + * + * We need upto 5us to enable the irq, and upto 20us to hide the + * scheduler latency of a context switch, ignoring the secondary + * impacts from a context switch such as cache eviction. + * + * The scheme used for low-latency IO is called "hybrid interrupt + * polling". The suggestion there is to sleep until just before you + * expect to be woken by the device interrupt and then poll for its + * completion. That requires having a good predictor for the request + * duration, which we currently lack. + */ + if (CONFIG_DRM_I915_SPIN_REQUEST && + __i915_spin_request(rq, state, CONFIG_DRM_I915_SPIN_REQUEST)) goto out; /* -- 2.20.1 _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
