[release/8.0-staging] Fix performance of System.Security. Cryptography.Tests #95310
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
These tests take a significant amount of time on Linux due to primality tests of RSA 16384.
|
Tagging subscribers to this area: @dotnet/area-system-security, @bartonjs, @vcsjones Issue DetailsBackport of #95284 to release/8.0-staging /cc @vcsjones Customer ImpactTestingRiskIMPORTANT: If this backport is for a servicing release, please verify that:
|
|
Thanks for tagging me. For future reference: Area owners can merge their own PRs to the Test-only changes are considered tell mode, so area owners can apply the @dotnet/area-system-security can you please take a look? |
| @@ -10,7 +10,8 @@ namespace System.Security.Cryptography.Rsa.Tests | |||
| [SkipOnPlatform(TestPlatforms.Browser, "Not supported on Browser")] | |||
| public partial class ImportExport | |||
| { | |||
| public static bool Supports16384 { get; } = TestRsa16384(); | |||
| private static readonly Lazy<bool> s_supports16384 = new Lazy<bool>(TestRsa16384); | |||
| public static bool Supports16384 => s_supports16384.Value; | |||
There was a problem hiding this comment.
TestRsa16384 itself takes a lot of time?
There was a problem hiding this comment.
Woh. Ok. That's not something we have any control over, right? That's all time spent in the OS implementation? Why does testing for even the support for this take so long?
There was a problem hiding this comment.
That's not something we have any control over, right? That's all time spent in the OS implementation?
Yeah.
Why does testing for even the support for this take so long?
It basically boils down to: This is the largest possible RSA key. No one uses them (or should), but it is a theoretical max, so we want to test our upper boundary condition. Using an RSA key that large is a guaranteed way to hit performance problems.
The performance hit is in OpenSSL's EVP_PKEY_check. We use this method on key inputs so we can validate the key's consistency (making sure public parameters match private ones, etc). OpenSSL does a rather "thorough" job of this by running Miller-Rabin on the key input. We have no control over that or any flags, and we can't really avoid the called to EVP_PKEY_check.
Woh. Ok.
This PR cut the total execution time of System.Security.Cryptography on Linux by more than half on my Linux box.
Backport of #95284 to release/8.0-staging
/cc @vcsjones @bartonjs
This is a test-only change that moves a few slow tests to outerloop to alleviate inner loop test timeouts.
Fixes #95094