Tomáš Mráz
Tomáš Mráz
IMO if EVP_PKEY_KEYPAIR is specified - you should import at least public or private key part. Basically the selection should be the maximum that is possibly being imported. However importing...
Yeah this semantics also would make sense. It does not matter much which one is chosen but it should be documented.
Fix in #17408 however do we really want the change and if so, also on 3.0 branch?
I think we should consider doing a docfix for this. It is fairly surprising behavior so it should be at least mentioned in the documentation.
@nhorman with 1024 neighborhoods the collision limit was hit. I used 4096 as there will be external names added with 3rd party providers. This makes for 256KB hashtable.
Hmm... actually running the evp_fetch test directly on the work machine gives some better looking numbers: Old code: ``` tomas@work:~/tools/perf$ for i in 1 10 50 100 500 1000 ;...
Here is the current comparison of this branch with master branch: ### evp_fetch |thread count| number of iterations |openssl master per operation avg usec | master std dev |openssl namemap-update...
Rebased and squashed fixups to resolve conflict.
I think I've responded to your comments @nhorman. It has to be kept as is. I am still concerned whether always having a 256KB hash table per-libctx isn't a problem....
I'll try adding a rcu read lock and measure the performance in such case to see what is the performance penalty.