On 01/03/2021 15:48, John Garry wrote:
While max32_alloc_size indirectly tracks the largest*contiguous*
available space, one of the ideas from which it grew was to simply keep
count of the total number of free PFNs. If you're really spending
significant time determining that the tree is full, as opposed to just
taking longer to eventually succeed, then it might be relatively
innocuous to tack on that semi-redundant extra accounting as a
self-contained quick fix for that worst case.
Anyway, we see ~50% throughput regression, which is intolerable. As seen
in [0], I put this down to the fact that we have so many IOVA requests
which exceed the rcache size limit, which means many RB tree accesses
for non-cacheble IOVAs, which are now slower.
I will attempt to prove this by increasing RCACHE RANGE, such that all
IOVA sizes may be cached.
About this one, as expected, we restore performance by increasing the
RCACHE RANGE.
Some figures:
Baseline v5.12-rc1
strict mode:
600K IOPs
Revert "iommu/iova: Retry from last rb tree node if iova search fails":
1215K
Increase IOVA RCACHE range 6 -> 10 (All IOVAs size requests now
cacheable for this experiment):
1400K
Reduce LLDD max SGE count 124 -> 16:
1288K
non-strict mode
1650K
So ideally we can work towards something for which IOVAs of all size
could be cached.
Cheers,
John
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu
