Re: [Qemu-devel] [PATCH 2/2] qemu-thread: let cur_mon be per-thread

[Peter Xu <peterx@redhat.com>]

On Wed, Apr 11, 2018 at 09:45:32AM +0800, Stefan Hajnoczi wrote:
> On Tue, Apr 10, 2018 at 08:49:13PM +0800, Peter Xu wrote:
> > cur_mon was only used in main loop so we don't really need that to be
> > per-thread variable.  Now it's possible that we have more than one
> > thread to operate on it.  Let's start to let it be per-thread variable.
> 
> Trying to understand the reason for this patch:
> 
> Are there any users of per-thread cur_mon?

Currently no.  But if considering future OOB-capable commands, they
will modify cur_mon in monitor IOThread at least.

> 
> or
> 
> Does this patch fix a bug?

No; currently we have no bug.  But we have encounter the bug when we
start to add more OOB commands.

Here is the problem (Marc-Andre reported this, and I'll try to
summarize): after we have valid OOB commands,
monitor_qmp_dispatch_one() can be run not only in main thread, but
also in monitor iothread.  When that happens, both of them (main
thread, and monitor iothread) can be modifying the cur_mon variable at
the same time. [1]

Considering that cur_mon is only used "just like" a stack variable, it
should be perfectly fine we just make it as a per thread variable,
hence this patch.

> 
> > In case we'll create threads within a valid cur_mon setup, we'd better
> > let the child threads to inherit the cur_mon from parent thread too.  Do
> > that for both posix and win32 threads.
> 
> Without actual users I don't like this.  It sounds like "let's make it
> global just in case something needs it some day".
> 
> It's ugly for QEMU's thread API to know about the monitor - that's a
> layering violation.

Yes, I'm sorry about it.  Actually I don't like it too.  But that
seems to be an efficient and simple solution to me now.  The ideal
solution should be totally removing cur_mon, which is non-trivial.

And for sure we can try to avoid layer violation.  For example, we can
have something like qemu_thread_register_variable(pthread_key_t), then
monitor code can register the cur_mon pthread_key to the qemu thread
module.  That'll somehow achieve isolation between modules but I'm not
sure whether that would be necessary for now, hence I chose the simple.

> 
> If there's a legitimate need I think this patch might be necessary, but
> I don't see enough justification to do this yet.

The problem was described at [1].  Thanks,

-- 
Peter Xu