A colleague here noticed that the pthreads-based locking loses the distinction between read and write locks. We've collected mutex contention data, and found that the CRYPTO_ERR lock, used while getting error info, is one of the biggest offenders.
It turns out that pthreads_locking_callback ignores the CRYPTO_READ/WRITE flag that is passed in. It seems fairly simple to update that function to use NPTL rwlock's. Any interest? We'll put out a diff and pull request soon. -- Principal Security Engineer Akamai Technologies, Cambridge, MA IM: [email protected]<mailto:[email protected]>; Twitter: RichSalz
