• Chris Wilson's avatar
    drm/i915/execlists: Direct submission of new requests (avoid tasklet/ksoftirqd) · 9512f985
    Chris Wilson authored
    Back in commit 27af5eea ("drm/i915: Move execlists irq handler to a
    bottom half"), we came to the conclusion that running our CSB processing
    and ELSP submission from inside the irq handler was a bad idea. A really
    bad idea as we could impose nearly 1s latency on other users of the
    system, on average! Deferring our work to a tasklet allowed us to do the
    processing with irqs enabled, reducing the impact to an average of about
    50us.
    
    We have since eradicated the use of forcewaked mmio from inside the CSB
    processing and ELSP submission, bringing the impact down to around 5us
    (on Kabylake); an order of magnitude better than our measurements 2
    years ago on Broadwell and only about 2x worse on average than the
    gem_syslatency on an unladen system.
    
    In this iteration of the tasklet-vs-direct submission debate, we seek a
    compromise where by we submit new requests immediately to the HW but
    defer processing the CS interrupt onto a tasklet. We gain the advantage
    of low-latency and ksoftirqd avoidance when waking up the HW, while
    avoiding the system-wide starvation of our CS irq-storms.
    
    Comparing the impact on the maximum latency observed (that is the time
    stolen from an RT process) over a 120s interval, repeated several times
    (using gem_syslatency, similar to RT's cyclictest) while the system is
    fully laden with i915 nops, we see that direct submission an actually
    improve the worse case.
    
    Maximum latency in microseconds of a third party RT thread
    (gem_syslatency -t 120 -f 2)
      x Always using tasklets (a couple of >1000us outliers removed)
      + Only using tasklets from CS irq, direct submission of requests
    +------------------------------------------------------------------------+
    |          +                                                             |
    |          +                                                             |
    |          +                                                             |
    |          +       +                                                     |
    |          + +     +                                                     |
    |       +  + +     +  x     x     x                                      |
    |      +++ + +     +  x  x  x  x  x  x                                   |
    |      +++ + ++  + +  *x x  x  x  x  x                                   |
    |      +++ + ++  + *  *x x  *  x  x  x                                   |
    |    + +++ + ++  * * +*xxx  *  x  x  xx                                  |
    |    * +++ + ++++* *x+**xx+ *  x  x xxxx x                               |
    |   **x++++*++**+*x*x****x+ * +x xx xxxx x          x                    |
    |x* ******+***************++*+***xxxxxx* xx*x     xxx +                x+|
    |             |__________MA___________|                                  |
    |      |______M__A________|                                              |
    +------------------------------------------------------------------------+
        N           Min           Max        Median           Avg        Stddev
    x 118            91           186           124     125.28814     16.279137
    + 120            92           187           109     112.00833     13.458617
    Difference at 95.0% confidence
    	-13.2798 +/- 3.79219
    	-10.5994% +/- 3.02677%
    	(Student's t, pooled s = 14.9237)
    
    However the mean latency is adversely affected:
    
    Mean latency in microseconds of a third party RT thread
    (gem_syslatency -t 120 -f 1)
      x Always using tasklets
      + Only using tasklets from CS irq, direct submission of requests
    +------------------------------------------------------------------------+
    |           xxxxxx                                        +   ++         |
    |           xxxxxx                                        +   ++         |
    |           xxxxxx                                      + +++ ++         |
    |           xxxxxxx                                     +++++ ++         |
    |           xxxxxxx                                     +++++ ++         |
    |           xxxxxxx                                     +++++ +++        |
    |           xxxxxxx                                   + ++++++++++       |
    |           xxxxxxxx                                 ++ ++++++++++       |
    |           xxxxxxxx                                 ++ ++++++++++       |
    |          xxxxxxxxxx                                +++++++++++++++     |
    |         xxxxxxxxxxx    x                           +++++++++++++++     |
    |x       xxxxxxxxxxxxx   x           +            + ++++++++++++++++++  +|
    |           |__A__|                                                      |
    |                                                      |____A___|        |
    +------------------------------------------------------------------------+
        N           Min           Max        Median           Avg        Stddev
    x 120         3.506         3.727         3.631     3.6321417    0.02773109
    + 120         3.834         4.149         4.039     4.0375167   0.041221676
    Difference at 95.0% confidence
    	0.405375 +/- 0.00888913
    	11.1608% +/- 0.244735%
    	(Student's t, pooled s = 0.03513)
    
    However, since the mean latency corresponds to the amount of irqsoff
    processing we have to do for a CS interrupt, we only need to speed that
    up to benefit not just system latency but our own throughput.
    
    v2: Remember to defer submissions when under reset.
    v4: Only use direct submission for new requests
    v5: Be aware that with mixing direct tasklet evaluation and deferred
    tasklets, we may end up idling before running the deferred tasklet.
    v6: Remove the redudant likely() from tasklet_is_enabled(), restrict the
    annotation to reset_in_progress().
    v7: Take the full timeline.lock when enabling perf_pmu stats as the
    tasklet is no longer a valid guard. A consequence is that the stats are
    now only valid for engines also using the timeline.lock to process
    state.
    
    Testcase: igt/gem_exec_latency/*rthog*
    References: 27af5eea ("drm/i915: Move execlists irq handler to a bottom half")
    Suggested-by: default avatarTvrtko Ursulin <tvrtko.ursulin@intel.com>
    Signed-off-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
    Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
    Reviewed-by: default avatarTvrtko Ursulin <tvrtko.ursulin@intel.com>
    Link: https://patchwork.freedesktop.org/patch/msgid/20180628201211.13837-9-chris@chris-wilson.co.uk
    9512f985
intel_engine_cs.c 45.1 KB