Quoting Chris Wilson (2018-10-16 19:31:06)
> Fwiw, the shmem_unlock_mapping() call feels quite expensive, almost
> nullifying the advantage gained from not walking the lists in reclaim.
> I'll have better numbers in a couple of days.
Using a test ("igt/benchmarks/gem_syslatency -t 120 -b -m" on kbl)
consisting of cycletest with a background load of trying to allocate +
populate 2MiB (to hit thp) while catting all files to /dev/null, the
result of using mapping_set_unevictable is mixed.
Each test run consists of running cycletest for 120s measuring the mean
and maximum wakeup latency and then repeating that 120 times.
x baseline-mean.txt # no i915 activity
+ tip-mean.txt # current stock i915 with a continuous load
+------------------------------------------------------------------------+
| x + |
| x + |
|xx + |
|xx + |
|xx + |
|xx ++ |
|xx +++ |
|xx +++ |
|xx +++ |
|xx +++ |
|xx +++ |
|xx ++++ |
|xx +++++ |
|xx ++++++ |
|xx ++++++ |
|xx ++++++ |
|xx ++++++ |
|xx ++++++ |
|xx +++++++ + |
|xx ++++++++ + |
|xx ++++++++++ |
|xx+++++++++++ + + |
|xx+++++++++++ + + + + + ++ +|
| A |
||______M_A_________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 359.153 876.915 863.548 778.80319 186.15875
+ 120 2475.318 73172.303 7666.812 9579.4671 9552.865
Our target then is 863us, but currently i915 adds 7ms of uninterruptable
delay on hitting the shrinker.
x baseline-mean.txt
+ mapping-mean.txt # applying the mapping_set_evictable patch
* tip-mean.txt
+------------------------------------------------------------------------+
| x * + |
| x * + |
|xx * + |
|xx * + |
|xx * + |
|xx ** + |
|xx *** ++ |
|xx *** ++ |
|xx *** ++ |
|xx *** ++ |
|xx *** ++ |
|xx **** + ++ |
|xx *****+ ++ ++ |
|xx ******+ ++ ++ |
|xx ******+ ++ + ++ |
|xx ******+ ++ + ++ |
|xx ******+ ++ ++++ |
|xx ******+ ++ ++++ |
|xx ******* *+ ++++ |
|xx ******** *+ +++++ |
|xx **********+ +++++ |
|xx***********+*+++++* |
|xx***********+*+++++* * + * * ** *|
| A |
| |___AM___| |
||______M_A_________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 359.153 876.915 863.548 778.80319 186.15875
+ 120 3291.633 26644.894 15829.186 14654.781 4466.6997
* 120 2475.318 73172.303 7666.812 9579.4671 9552.865
Shows that if we use the mapping_set_evictable() +
shmem_unlock_mapping() we add a further 8ms uninterruptable delay to the
system... That's the opposite of our goal! ;)
x baseline-mean.txt
+ lock_vma-mean.txt # the old approach of pinning each page
* tip-mean.txt
+------------------------------------------------------------------------+
| *+ * |
| *+ * * |
| *+ * * |
| *+ * * |
| *+ *** |
| *+ *** |
| *+ *** |
| *+ *** |
| *+ *** |
| *+ *** |
| *+ *** |
| *+ **** |
| *+ ***** |
| *+ ****** |
| *+ ****** * |
| *+ ****** * |
| *+ ******* * |
| *+******** * |
| *+******** * |
| *+******** * |
| *+******** * * * |
| *+******** * * + * * * * * * *|
| A |
||MA| |
||_______M_A________| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 359.153 876.915 863.548 778.80319 186.15875
+ 120 511.415 18757.367 1276.302 1416.0016 1679.3965
* 120 2475.318 73172.303 7666.812 9579.4671 9552.865
By contrast, the previous approach of using mlock_page_vma() does
dramatically reduce the uninterruptable delay -- which suggests that the
mapping_set_evictable() isn't keeping our unshrinkable pages off the
shrinker lru.
However, if instead of looking at the average uninterruptable delay
during the 120s of cycletest, but look at the worst case, things get a
little more interesting. Currently i915 is terrible.
x baseline-max.txt
+ tip-max.txt
+------------------------------------------------------------------------+
| * |
[snip 100 lines]
| * |
| * |
| * |
| * |
| * |
| * |
| * |
| * |
| * +++ ++ + + + + +|
| A |
||_____M_A_______| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 7391 58543 51953 51564.033 5044.6375
+ 120 2284928 6.752085e+08 3385097 20825362 80352645
Worst case with no i915 is 52ms, but as soon as we load up i915 with
some work, the worst case uninterruptable delay is on average 20s!!! As
suggested by the median, the data is severely skewed by a few outliers.
(Worst worst case is so bad khungtaskd often makes an appearance.)
x baseline-max.txt
+ mapping-max.txt
* tip-max.txt
+------------------------------------------------------------------------+
| * |
[snip 100 lines]
| * |
| * |
| * |
| * |
| * |
| * |
| * |
| *+ |
| *+*** ** * * +* * *|
| A |
| |_A__| |
||_____M_A_______| |
+------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 120 7391 58543 51953 51564.033 5044.6375
+ 120 3088140 2.9181602e+08 4022581 6528993.3 26278426
* 120 2284928 6.752085e+08 3385097 20825362 80352645
So while the mapping_set_evictable patch did reduce the maximum observed
delay within the 4 hour sample, on average (median, to exclude those worst
worst case outliers) it still fares worse than stock i915. The
mlock_page_vma() has no impact on worst case wrt stock.
My conclusion is that the mapping_set_evictable patch makes both the
average and worst case uninterruptable latency (as observed by other
users of the system) significantly worse. (Although the maximum latency
is not stable enough to draw a real conclusion other than i915 is
shockingly terrible.)
-Chris