Peter Xu <[email protected]> writes: > On Tue, Nov 25, 2025 at 12:46:01PM +0100, Markus Armbruster wrote: >> Daniel P. Berrangé <[email protected]> writes: >> >> > On Tue, Nov 25, 2025 at 08:40:07AM +0100, Markus Armbruster wrote: >> >> g_autoptr(T) is quite useful when the object's extent matches the >> >> function's. >> >> >> >> This isn't the case for an Error object the function propagates to its >> >> caller. It is the case for an Error object the function reports or >> >> handles itself. However, the functions to report Error also free it. > > I'd confess I didn't pay enough attention on how the error API was designed > deliberately to always free the Error objects before almost whenever > possible. But I see now, thanks for the write up.
You're welcome! >> >> >> >> Thus, g_autoptr(Error) is rarely applicable. We have just three >> >> instances out of >1100 local Error variables, all in migration code. >> >> >> >> Two want to move the error to the MigrationState for later handling / >> >> reporting. Since migrate_set_error() doesn't move, but stores a copy, >> >> the original needs to be freed, and g_autoptr() is correct there. We >> >> have 17 more that instead manually free with error_free() or >> >> error_report_err() right after migrate_set_error(). >> >> >> >> We recently discussed storing a copy vs. move the original: >> >> >> >> From: Peter Xu <[email protected]> >> >> Subject: Re: [PATCH 0/3] migration: Error fixes and improvements >> >> Date: Mon, 17 Nov 2025 11:03:37 -0500 >> >> Message-ID: <[email protected]> >> >> >> >> The two g_autoptr() gave me pause when I investigated this topic, simply >> >> because they deviate from the common pattern migrate_set_error(s, err) >> >> followed by error_free() or error_report_err(). >> >> >> >> The third one became wrong when I cleaned up the reporting (missed in >> >> the cleanup patch, fixed in the patch I'm replying to). I suspect my >> >> mistake escaped review for the same reason I made it: g_autoptr(Error) >> >> is unusual and not visible in the patch hunk. >> >> >> >> Would you like me to replace the two correct uses of g_autoptr(Error) by >> >> more common usage? > > Works for me. > > Now I also think it should be good migrate_set_error() follow QEMU's Error > API design if we decide to stick with it freeing errors in such APIs. > > Said that, I wonder if you think we could still consider passing Error** > into migrate_set_error(), though, which will be a merged solution of > current Error API and what Marc-Andre proposed on resetting pointers to > avoid any possible UAF, which I would still slightly prefer personally. > > If we rework migrate_set_error() to take ownership first, then we can > naturally drop the two use cases, and remove the cleanup function. > > Markus, please also let me know if you want me to do it. I think the first step should replace the two g_autoptr() by error_free(), then delete g_autoptr() support. A possible second step is to replace migrate_set_error() by a function that takes ownership. "Replace" because I think migrate_set_error() would be a bad name for such a function. What's a better name? Naming is hard... migrate_error_propagate_to_state()? Because there's similarity: error_propagate(errp, err); stores @err in @errp, or else frees it, and migrate_error_propagate_to_state(s, err) stores @err in @s, or else frees it. We could also forgo encapsulation and simply use error_propagate(&s->error, err); Matter of taste, which means migration maintainers decide. I can do just the first step, or both. Up to you. [...]
