Re: wininet(1/3): Only initialize SSL_CTX once

[Hans Leidekker]

Hi Juan,
> These patches collectively fix bug 18364, which was a crash in OpenSSL
> in a multithreaded app (utorrent.exe.)

The bug reporter sees long delays between torrents starting up after
applying these patches, which makes we wonder if the ssl lock effectively
serializes secure https requests?

Would patch 1 and 3 not be sufficient to fix the utorrent crash?

-Hans

wininet(1/3): Only initialize SSL_CTX once

[Juan Lang]

Hi Hans,
> The bug reporter sees long delays between torrents starting up after
> applying these patches, which makes we wonder if the ssl lock effectively
> serializes secure https requests?

There isn't a single SSL lock.  There's an interlocked exchange for
assigning SSL_CTX, which should only be created once except in a race
between threads.  There's a critical section to protect against two
threads initializing the SSL library at the same time.  Finally, there
are a number of SSL locks (CRYPTO_num_locks() many, specifically) that
OpenSSL uses.  When I was debugging it, there are at least 16 of them.
 So I don't think serialization of locks is the root cause of the
slowdown.
> Would patch 1 and 3 not be sufficient to fix the utorrent crash?

I don't think so.  Other apps have had similar crashes with the
equivalent of patch 2 not applied, and the OpenSSL man page warns of
random crashes due to failing to implement and set these callbacks.
--Juan