Comment 5 for bug 856072

(In reply to Patrick McManus from comment #2)
> (In reply to Doug Turner (:dougt) from comment #1)
> > lets add locking around the domain resolution and ifdef it for #android.
> > This should only be done for pre-honeycomb (i think). jim, want to throw up
> > a patch?
>
> wow. That could have such a huge negative impact. We need parallelism on
> high latency operations.
>
> Do you have a pointer to the android getaddrinfo() implementation in
> question? I can't find it because android.git.kernel.org is still offline.
> Is the io conditional on anything - a lazy init perhaps?
>
Here's a mirror (code is forked from netbsd): https://github.com/android/platform_bionic/blob/master/libc/netbsd/net/getaddrinfo.c

fopen() is called inside _sethtent() and fclose() in _endhtent(). Both are called by _files_getaddrinfo() which is responsible for reading /etc/hosts. I don't see a way to skip that (unless you count the workaround for our testing infra which is to delete /etc/hosts and make fopen() fail; but obviously a dirty hack used internally)

(In reply to Doug Turner (:dougt) from comment #3)
> slow is better than crashy. You're right, we also could just move the
> bionic code into nspr (or necko), but I am not sure how much that code fans
> out. That would be better than locking for sure.

Yeah. Also setting affinity will keep parallelism while reducing chance of race condition to pre-dual-core levels, which wasn't really an issue back then (I don't think).