Stefan Hajnoczi <[email protected]> writes: > On Tue, Nov 01, 2016 at 04:21:36PM +0000, Alex Bennée wrote: >> >> Paolo Bonzini <[email protected]> writes: >> >> > On 21/10/2016 13:54, Alex Bennée wrote: >> >> There is a slight wart when checking for the state of the BQL when using >> >> GThread base co-routines (which we keep for ThreadSanitizer runs). While >> >> the main-loop holds the BQL it is suspended until the co-routine >> >> completes however the co-routines run in a separate thread so checking >> >> the TLS variable could be wrong. >> >> >> >> We fix this by expanding the check to include qemu_in_coroutine() for >> >> GThread based builds. As it is not used for production builds I'm not >> >> overly worried about any performance impact which should be negligible >> >> anyway. >> >> >> >> Signed-off-by: Alex Bennée <[email protected]> >> >> Cc: Stefan Hajnoczi <[email protected]> >> > >> > This is wrong unfortunately. It is possible to run coroutines outside >> > the BQL (e.g. with -device virtio-blk,iothread=foo). >> > >> > Do you know exactly why TSAN has no love for coroutines? >> >> The current production stuff is due to missing support for new stacks >> with setcontext. However I have built the latest tsan support library >> and that seems happy without the gthread co-routines. >> >> Currently I'm dealing with glib's racy gthread support however. > > I think Paolo suggested we drop the GThread backend on IRC. I agree > that we should do that since GThread co-routines break code that uses > thread-local variables and have never truly worked.
Indeed I have pulled your patch into my current series of sanitizer fixes. Once I can fix the setcontext/sigjmp confusion for a __SANITIZER__ builds I'll post the series. -- Alex Bennée
