Bug #20863: `zlib.c` calls `rb_str_set_len` and `rb_str_modify_expand`(and others) without holding the GVL.

• Status: Open
• Assignee: -
• Target version: -
• Backport: 3.1: UNKNOWN, 3.2: UNKNOWN, 3.3: UNKNOWN

[ioquatix (Samuel Williams)]

Background

I was working on https://bugs.ruby-lang.org/issues/20876 and was investigating some problems with zlib.c and GVL, and noticed that zstream_run_func is executed without the GVL, but can invoke various rb_ string functions. Those functions in turn can invoke rb_raise and generally look problematic. However, maybe by luck, such code path does not appear to be invoked in typical usage.

However, even so, it is possible to cause zstream_run_func to segfault by a carefully crafted program which causes the internal buffer to be resized while the GVL is released: https://github.com/ruby/zlib/pull/88#issuecomment-2455772054

Proposal

I would like to modify zlib.c to only release the GVL around the CPU intensive compression/decompression operation: https://github.com/ruby/zlib/pull/88

In addition, I identified several more improvements to prevent segfaults and other related failures:

• Use rb_str_locktemp to prevent the z->buf changing size while in use by the rb_nogvl code.
• Expand the mutex to protect #deflate and #inflate completely, not just the internal operation.

In order to catch these issues earlier and find other bugs like this, I recommend we introduce additional checks: https://bugs.ruby-lang.org/issues/20877

[ioquatix (Samuel Williams)]

• Description updated (diff)

[ioquatix (Samuel Williams)]

• Description updated (diff)

[ioquatix (Samuel Williams)]

• Description updated (diff)

[byroot (Jean Boussier)]

@ko1 (Koichi Sasada) Do we have a proper description of what is safe and what it unsafe to do with the GVL released?

Because obviously it's OK to use ruby_xmalloc / ruby_xfree with the GVL released, so methods which allocate aren't necessarily problematic?\

In this case I'm unclear on why rb_str_set_len / rb_str_modify_expand shouldn't be called with the GVL released, assuming the objects on which they operate aren't visible to any other thread.

I think it would be helpful to have more clear guidelines on these things (unless of course I missed some existing documentation).

[ko1 (Koichi Sasada)]

Quoted from rb_thread_call_without_gvl doc:

 * NOTE: You can not execute most of Ruby C API and touch Ruby
 *       objects in `func()' and `ubf()', including raising an
 *       exception, because current thread doesn't acquire GVL
 *       (it causes synchronization problems).  If you need to
 *       call ruby functions either use rb_thread_call_with_gvl()
 *       or read source code of C APIs and confirm safety by
 *       yourself.
 *
 * NOTE: In short, this API is difficult to use safely.  I recommend you
 *       use other ways if you have.  We lack experiences to use this API.
 *       Please report your problem related on it.
 *
 * NOTE: Releasing GVL and re-acquiring GVL may be expensive operations
 *       for a short running `func()'. Be sure to benchmark and use this
 *       mechanism when `func()' consumes enough time.
 *
 * Safe C API:
 * * rb_thread_interrupted() - check interrupt flag
 * * ruby_xmalloc(), ruby_xrealloc(), ruby_xfree() -
 *   they will work without GVL, and may acquire GVL when GC is needed.

Again:

 * NOTE: In short, this API is difficult to use safely.  I recommend you
 *       use other ways if you have.  We lack experiences to use this API.
 *       Please report your problem related on it.

[byroot (Jean Boussier)]

@ko1 (Koichi Sasada) Not sure how I didn't think to check that, thank you. So indeed allocations are fine. From what I understand, the issue is mostly exceptions and of course using an object concurrently.

[ioquatix (Samuel Williams)]

I think the issue is, those methods from a public interface POV, are not allowed to be called without the GVL.

Even if today the implementation follows a "safe" code path, in the future, it may not.

Adding these annotations will help to clarify that "this method is not safe to call without the GVL" - a form of internal and run-time documentation.

[Eregon (Benoit Daloze)]

ioquatix (Samuel Williams) wrote in #note-7:

Even if today the implementation follows a "safe" code path, in the future, it may not.

This is a good point.
I think we should consider all C API functions unsafe to be called without the GVL, except the functions listed in Safe C API.
So I think we should update the docs to remove or read source code of C APIs and confirm safety by yourself. as it's not a good idea as it may change and it's very hard to assess if safe.

[byroot (Jean Boussier)]

There would be quite a lot of value in having some nogvl save APIs though. e.g. if database clients could allocate Hash/Array/String to build the response while the GVL is still released, it could really help with throughput of threaded servers like Puma.

[ioquatix (Samuel Williams)]

There would be quite a lot of value in having some nogvl save APIs though. e.g. if database clients could allocate Hash/Array/String to build the response while the GVL is still released, it could really help with throughput of threaded servers like Puma.

I think it's a great idea (seriously great), but out of scope for this issue. Do you want to create a new issue to start that discussion?