fix deadlock in shmget

[dead10ck]

shmget deadlocks when called more than once in the same process. It locks the mutex, then calls ashv_read_remote_segment, which asks the background thread in the same process to find a shmid, which in turn tries to lock the mutex which is already locked. I included a simple trivial test case that hits the deadlock.

I actually found this in Postgres, noticing that a CREATE TABLE AS ... query was never returning, and the process that was executing the query was stuck doing nothing. strace showed one thread stuck reading from a socket, and another stuck waiting on a lock.

Nevermind, misunderstood the logic

[Grimler91]

Nicely done debugging this!

It should not be necessary to bump all packages depending on libandroid-shmem after this, right? (or did you have to rebuild postgres to link it against the new library?)

[dead10ck]

Thanks! It was actually kind of fun.

As for patching, I guess it depends on each package, and whether it links statically or dynamically. If dynamically (which I imagine is most), it should not be necessary to do anything to those packages; I did not need to rebuild Postgres, for example. But it might be a good idea to track down if any packages link statically, or do something else (a quick grep shows libcairo looks like it vendors a copy, for example, which probably does require a patch to get this fix applied).