subprocess can block all Python threads when using vfork()
until the child process exec()
succeeds or fails.
#104372
Labels
extension-modules
C modules in the Modules dir
stdlib
Python modules in the Lib dir
type-bug
An unexpected behavior, bug, or error
Background
When using the
subprocess
module on Linux,vfork()
rather thanfork()
has been preferred when possible for a few Python releases as is generally offers MUCH higher performance, the larger the parent process, the more cpu time it saves on page table copying.Problem
vfork()
- by design - does not return control to parent process until after the child process has successfully performed anexec()
system call or has died.Python
subprocess
uses_posixsubprocess.fork_exec()
extension module implementation to do handle async-signal-safe fork/vfork+exec code path (as that cannot safely be implemented in Python). This function does not release the GIL. Thus the GIL is held across thevfork()
call. Meaning the parent process does not resume execution until the exec has succeeded, or failed and returned an error which our child writes to it'serrpipe_write
fd before exiting.This can result in all Python threads hanging (ie: the GIL is held by a blocked thread) when the child
exec
system call takes a long time. Such as the binary existing on a potentially slow or high latency network filesystem. (We witnessed this on a FUSE filesystem backed by remote (high latency) cloud storage with a large executable... but any slow or high latency filesystem containing an executable or anything interfering in the exec system call can cause the same problem).Workarounds
Building without
vfork
support in_posixsubprocess
or otherwise passing an arg to subprocess API that happens to disable vfork because it is incompatible with the vfork-concept are workarounds. (I won't recommend any of those because none of them are a good idea to pass when not needed for their primary purpose).Another workaround viable in our specific case is to pre-read the executable before asking subprocess to launch it as that moves the high IO latency from the
exec
system call into a read without the GIL held beforehand as it pulls the executable into a fast-enough local storage cache. (That being useful at all depends on your particular filesystem latency implementation situation)Potential Solutions
If we could simply release the GIL around the
vfork()
call that would seem to be ideal... It's not clear to me that we can actually do that, I'm working out what a PR would look like or what blocks us from doing so safely.Offering yet another keyword argument to subprocess APIs to disable the use of APIs that could block the entire process upon
exec()
such asvfork
is viable - but doesn't feel great given our existing long list of subprocess args.Linked PRs
make_inheritable
for async signal safety #104518The text was updated successfully, but these errors were encountered: