libnbd thread-local errors and dlclose (was: Re: [ANNOUNCE] libblkio v0.1.0 preview release)

From: "Richard W.M. Jones" <rjones@redhat.com>
To: "Daniel P. Berrangé" <berrange@redhat.com>
Cc: Kevin Wolf <kwolf@redhat.com>,
	pkrempa@redhat.com, Alberto Garcia <berto@igalia.com>,
	slp@redhat.com, qemu-block@nongnu.org,
	Markus Armbruster <armbru@redhat.com>,
	qemu-devel@nongnu.org, mreitz@redhat.com,
	Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>,
	Stefan Hajnoczi <stefanha@redhat.com>,
	Klaus Jensen <its@irrelevant.dk>,
	philmd@redhat.com, sgarzare@redhat.com
Subject: libnbd thread-local errors and dlclose (was: Re: [ANNOUNCE] libblkio v0.1.0 preview release)
Date: Thu, 29 Apr 2021 18:17:32 +0100	[thread overview]
Message-ID: <20210429171732.GX26415@redhat.com> (raw)
In-Reply-To: <YIrL5kE+0ULVN2lK@redhat.com>

On Thu, Apr 29, 2021 at 04:08:22PM +0100, Daniel P. Berrangé wrote:
> On Thu, Apr 29, 2021 at 04:00:38PM +0100, Richard W.M. Jones wrote:
> > On Thu, Apr 29, 2021 at 03:41:29PM +0100, Stefan Hajnoczi wrote:
> > > On Thu, Apr 29, 2021 at 03:22:59PM +0100, Richard W.M. Jones wrote:
> > > > libvirt originally, and now libnbd, keep a per-thread error message
> > > > (stored in thread-local storage).  It's a lot nicer than having to
> > > > pass &errmsg to every function.  You can just write:
> > > > 
> > > >  if (nbd_connect_tcp (nbd, "remote", "nbd") == -1) {
> > > >    fprintf (stderr,
> > > >             "failed to connect to remote server: %s (errno = %d)\n",
> > > >             nbd_get_error (), nbd_get_errno ());
> > > >    exit (EXIT_FAILURE);
> > > >  }
> > > > 
> > > > (https://libguestfs.org/libnbd.3.html#ERROR-HANDLING)
> > > > 
> > > > It means you can extend the range of error information available in
> > > > future.  Also you can return a 'const char *' and the application
> > > > doesn't have to worry about lifetimes, at least in the common case.
> > > 
> > > Thanks for sharing the idea, I think it would work well for libblkio
> > > too.
> > > 
> > > Do you ignore the dlclose(3) memory leak?
> > 
> > I believe this mechanism in libnbd ensures there is no leak in the
> > ordinary shared library (not dlopen/dlclose) case:
> > 
> > https://gitlab.com/nbdkit/libnbd/-/blob/f9257a9fdc68706a4079deb4ced61e1d468f28d6/lib/errors.c#L35
> > 
> > However I'm not sure what happens in the dlopen case, so I'm going to
> > test that out now ...
> 
> pthread_key destructors are a disaster waiting to happen with
> dlclose, if the dlclose happens while non-main threads are
> still running. When those threads exit, they'll run the
> destructor which points to a function that is no longer
> resident in memory.

While libnbd had a memory leak there, now fixed by:

https://gitlab.com/nbdkit/libnbd/-/commit/026d281c57dd95485cc9bf829918b5efd9e32ddb

I don't actually think we have the bug you're describing.  We have a
destructor (errors_free()) which runs on dlclose, which deletes the
thread-local storage key, so the destructor will not run after the
library has been unloaded.

I added a test for this which works fine for me:

https://gitlab.com/nbdkit/libnbd/-/commit/831d142787aba4f5b638418e02cf7e0f2a051565

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
libguestfs lets you edit virtual machines.  Supports shell scripting,
bindings from many languages.  http://libguestfs.org