[sldev] Request for comments on SNOW-103

[Aleric Inglewood]

Hiya,

I found that this bug is being caused by what I consider to be a pretty
serious design flaw. I cannot fix it without significant changes to how
things are "working" now.

I'd like to get some Linden eyes on this at this point for comments :).
You probably want to start reading somewhere here:

https://jira.secondlife.com/browse/SNOW-103?focusedCommentId=122363&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#action_122363

The kernel of the problem, summarized, is as follows:

If two different threads use the same APR memory pool for any LLAPRFile,
then closing of a file by one thread causes the file descriptor of the other
to be closed as well (and other nasty things).

I think that it should just never happen that two different threads
use the same pool for LLAPRFile objects.

However, this is the case, EVEN though the code attempts to not
do that by calling LLThread::getLocalAPRFilePool(). Namely,
getLocalAPRFilePool() returns the same pool for two different
threads without any problem.

Can the original designer/coder of this part please comment?
Am I correct to think that the idea of getLocalAPRFilePool() is to
return a memory pool that is local the calling thread?

It is not, because all it does is return a pointer from LLThread;
when that LLThread object is accessible by any thread,
then it doesn't really work.

The problem LLThread is the base class of LLTextureCache
which is a singleton! Obviously it is going to be used by
more than one thread.