Message ID | 20220222170645.860661-2-peter.maydell@linaro.org |
---|---|
State | Superseded |
Headers | show |
Series | ui/cocoa: updateUIInfo threading, autorelease pools | expand |
On Wed, Feb 23, 2022 at 2:06 AM Peter Maydell <peter.maydell@linaro.org> wrote: > > The updateUIInfo method makes Cocoa API calls. It also calls back > into QEMU functions like dpy_set_ui_info(). To do this safely, we > need to follow two rules: > * Cocoa API calls are made on the Cocoa UI thread > * When calling back into QEMU we must hold the iothread lock > > Fix the places where we got this wrong, by taking the iothread lock > while executing updateUIInfo, and moving the call in cocoa_switch() > inside the dispatch_async block. > > Some of the Cocoa UI methods which call updateUIInfo are invoked as > part of the initial application startup, while we're still doing the > little cross-thread dance described in the comment just above > call_qemu_main(). This meant they were calling back into the QEMU UI > layer before we'd actually finished initializing our display and > registered the DisplayChangeListener, which isn't really valid. Once > updateUIInfo takes the iothread lock, we no longer get away with > this, because during this startup phase the iothread lock is held by > the QEMU main-loop thread which is waiting for us to finish our > display initialization. So we must suppress updateUIInfo until > applicationDidFinishLaunching allows the QEMU main-loop thread to > continue, and instead defer telling the UI layer about our initial > window size until later. > > Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Thanks for fixing this bug. > --- > ui/cocoa.m | 27 ++++++++++++++++++++++++--- > 1 file changed, 24 insertions(+), 3 deletions(-) > > diff --git a/ui/cocoa.m b/ui/cocoa.m > index a8f1cdaf926..f8d3d8ad7a6 100644 > --- a/ui/cocoa.m > +++ b/ui/cocoa.m > @@ -522,8 +522,9 @@ QemuCocoaView *cocoaView; > } > } > > -- (void) updateUIInfo > +- (void) doUpdateUIInfo You may add the suffix "Locked" to align with handleEventLocked and make the intention more explicit. > { > + /* Must be called with the iothread lock, i.e. via updateUIInfo */ > NSSize frameSize; > QemuUIInfo info; > > @@ -554,6 +555,22 @@ QemuCocoaView *cocoaView; > dpy_set_ui_info(dcl.con, &info, TRUE); > } > > +- (void) updateUIInfo > +{ > + if (!allow_events) { > + /* > + * Don't try to tell QEMU about UI information in the application > + * startup phase -- we haven't yet registered dcl with the QEMU UI > + * layer, and also trying to take the iothread lock would deadlock. > + */ > + return; > + } > + > + with_iothread_lock(^{ > + [self doUpdateUIInfo]; > + }); > +} > + > - (void)viewDidMoveToWindow > { > [self updateUIInfo]; > @@ -1985,8 +2002,6 @@ static void cocoa_switch(DisplayChangeListener *dcl, > > COCOA_DEBUG("qemu_cocoa: cocoa_switch\n"); > > - [cocoaView updateUIInfo]; > - > // The DisplaySurface will be freed as soon as this callback returns. > // We take a reference to the underlying pixman image here so it does > // not disappear from under our feet; the switchSurface method will > @@ -1994,6 +2009,7 @@ static void cocoa_switch(DisplayChangeListener *dcl, > pixman_image_ref(image); > > dispatch_async(dispatch_get_main_queue(), ^{ > + [cocoaView updateUIInfo]; > [cocoaView switchSurface:image]; > }); > [pool release]; > @@ -2057,6 +2073,11 @@ static void cocoa_display_init(DisplayState *ds, DisplayOptions *opts) > qemu_event_init(&cbevent, false); > cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init]; > qemu_clipboard_peer_register(&cbpeer); > + > + /* Now we're set up, tell the UI layer our initial window size */ > + dispatch_async(dispatch_get_main_queue(), ^{ > + [cocoaView updateUIInfo]; > + }); > } This is not necessary since register_displaychangelistener calls dpy_gfx_switch. > > static QemuDisplay qemu_display_cocoa = { > -- > 2.25.1 >
On Tue, 22 Feb 2022 at 17:13, Akihiko Odaki <akihiko.odaki@gmail.com> wrote: > > On Wed, Feb 23, 2022 at 2:06 AM Peter Maydell <peter.maydell@linaro.org> wrote: > > diff --git a/ui/cocoa.m b/ui/cocoa.m > > index a8f1cdaf926..f8d3d8ad7a6 100644 > > --- a/ui/cocoa.m > > +++ b/ui/cocoa.m > > @@ -522,8 +522,9 @@ QemuCocoaView *cocoaView; > > } > > } > > > > -- (void) updateUIInfo > > +- (void) doUpdateUIInfo > > You may add the suffix "Locked" to align with handleEventLocked and > make the intention more explicit. Ah, good idea. > > + /* Now we're set up, tell the UI layer our initial window size */ > > + dispatch_async(dispatch_get_main_queue(), ^{ > > + [cocoaView updateUIInfo]; > > + }); > > } > > This is not necessary since register_displaychangelistener calls dpy_gfx_switch. Cool, I'll drop this bit. -- PMM
diff --git a/ui/cocoa.m b/ui/cocoa.m index a8f1cdaf926..f8d3d8ad7a6 100644 --- a/ui/cocoa.m +++ b/ui/cocoa.m @@ -522,8 +522,9 @@ QemuCocoaView *cocoaView; } } -- (void) updateUIInfo +- (void) doUpdateUIInfo { + /* Must be called with the iothread lock, i.e. via updateUIInfo */ NSSize frameSize; QemuUIInfo info; @@ -554,6 +555,22 @@ QemuCocoaView *cocoaView; dpy_set_ui_info(dcl.con, &info, TRUE); } +- (void) updateUIInfo +{ + if (!allow_events) { + /* + * Don't try to tell QEMU about UI information in the application + * startup phase -- we haven't yet registered dcl with the QEMU UI + * layer, and also trying to take the iothread lock would deadlock. + */ + return; + } + + with_iothread_lock(^{ + [self doUpdateUIInfo]; + }); +} + - (void)viewDidMoveToWindow { [self updateUIInfo]; @@ -1985,8 +2002,6 @@ static void cocoa_switch(DisplayChangeListener *dcl, COCOA_DEBUG("qemu_cocoa: cocoa_switch\n"); - [cocoaView updateUIInfo]; - // The DisplaySurface will be freed as soon as this callback returns. // We take a reference to the underlying pixman image here so it does // not disappear from under our feet; the switchSurface method will @@ -1994,6 +2009,7 @@ static void cocoa_switch(DisplayChangeListener *dcl, pixman_image_ref(image); dispatch_async(dispatch_get_main_queue(), ^{ + [cocoaView updateUIInfo]; [cocoaView switchSurface:image]; }); [pool release]; @@ -2057,6 +2073,11 @@ static void cocoa_display_init(DisplayState *ds, DisplayOptions *opts) qemu_event_init(&cbevent, false); cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init]; qemu_clipboard_peer_register(&cbpeer); + + /* Now we're set up, tell the UI layer our initial window size */ + dispatch_async(dispatch_get_main_queue(), ^{ + [cocoaView updateUIInfo]; + }); } static QemuDisplay qemu_display_cocoa = {
The updateUIInfo method makes Cocoa API calls. It also calls back into QEMU functions like dpy_set_ui_info(). To do this safely, we need to follow two rules: * Cocoa API calls are made on the Cocoa UI thread * When calling back into QEMU we must hold the iothread lock Fix the places where we got this wrong, by taking the iothread lock while executing updateUIInfo, and moving the call in cocoa_switch() inside the dispatch_async block. Some of the Cocoa UI methods which call updateUIInfo are invoked as part of the initial application startup, while we're still doing the little cross-thread dance described in the comment just above call_qemu_main(). This meant they were calling back into the QEMU UI layer before we'd actually finished initializing our display and registered the DisplayChangeListener, which isn't really valid. Once updateUIInfo takes the iothread lock, we no longer get away with this, because during this startup phase the iothread lock is held by the QEMU main-loop thread which is waiting for us to finish our display initialization. So we must suppress updateUIInfo until applicationDidFinishLaunching allows the QEMU main-loop thread to continue, and instead defer telling the UI layer about our initial window size until later. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> --- ui/cocoa.m | 27 ++++++++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-)