Skip to contentSkip to conversation

cocoa: run qemu_init in the main thread

RFC In review Philippe Mathieu-Daudé 3 participants 7 messages Mar 17, 2022 – Mar 21, 2022

lore
Share & tools
lore
#0 lore raw

v4 cover cocoa: run qemu_init in the main thread

From: Philippe Mathieu-Daudé <[email protected]>

Posting v4 in case someone want to iterate.

Pending issue raised by Akihiko Odaki:

* this actually breaks the "runas" option with ui/cocoa.

[+NSApplication sharedApplication] calls issetugid() to see if setgid() or setuid() is called before and calls exit() if it evaluates true. It does not evaluate true without this patch since setgid() and setuid() are called after [+NSApplication sharedApplication]. This patch, however, changes the order and triggers the check.

There are two options to solve the problem:
1. Move setgid and setuid calls after [+NSApplication sharedApplication] to let NSApplication initialize as the original user.
2. Do: [[NSUserDefaults standardUserDefaults] setBool:YES forKey:@"_NSAppAllowsNonTrustedUGID"]

Option 2 would be preferred in terms of practicality since nobody would want to initialize NSApplication as the original user (usually superuser). However, _NSAppAllowsNonTrustedUGID is not documented by Apple.

* Oudated comment in main():

1970 /*

 1971   * Create the menu entries which depend on QEMU state (for consoles
 1972   * and removeable devices). These make calls back into QEMU functions,
 1973   * which is OK because at this point we know that the second thread
 1974   * holds the iothread lock and is synchronously waiting for us to
 1975   * finish.
 1976   */

(https://marc.info/?l=qemu-devel&m=164752136410805)

Since v3:
- Move qemu_event_init before cbowner alloc
- Reduce main_thread scope to applicationDidFinishLaunching
- Updated updateUIInfo() comment (s/cocoa_display_init/applicationDidFinishLaunching)

Since v2:
- Extracted code movement in preliminary patch

v3: https://lore.kernel.org/qemu-devel/[email protected]/
v2: https://lore.kernel.org/qemu-devel/[email protected]/
v1: https://lore.kernel.org/qemu-devel/[email protected]/

Paolo Bonzini (1):
ui/cocoa: run qemu_init in the main thread

Philippe Mathieu-Daudé (1):
ui/cocoa: Code movement

 softmmu/main.c |  12 ++--
 ui/cocoa.m     | 161 ++++++++++++++++++++++---------------------------
 2 files changed, 79 insertions(+), 94 deletions(-)
Signature · 1 line
2.34.1
#1 lore raw

RFC v4 1/2 ui/cocoa: Code movement

From: Philippe Mathieu-Daudé <[email protected]>

Pure code movement in preparation for the next commit.

---
 ui/cocoa.m | 86 +++++++++++++++++++++++++++---------------------------
 1 file changed, 43 insertions(+), 43 deletions(-)
ui/cocoa.m +43 -43
index cb6e7c41dc..027c3053f7 100644
@@ -75,6 +75,9 @@ typedef struct {
     int height;
 } QEMUScreen;
 
+@interface QemuCocoaPasteboardTypeOwner : NSObject<NSPasteboardTypeOwner>
+@end
+
 static void cocoa_update(DisplayChangeListener *dcl,
                          int x, int y, int w, int h);
 
@@ -109,6 +112,8 @@ static bool allow_events;
 
 static NSInteger cbchangecount = -1;
 static QemuClipboardInfo *cbinfo;
+static QemuClipboardPeer cbpeer;
+static QemuCocoaPasteboardTypeOwner *cbowner;
 static QemuEvent cbevent;
 
 // Utility functions to run specified code block with iothread lock held
@@ -142,6 +147,44 @@ static bool bool_with_iothread_lock(BoolCodeBlock block)
     return val;
 }
 
+/*
+ * The startup process for the OSX/Cocoa UI is complicated, because
+ * OSX insists that the UI runs on the initial main thread, and so we
+ * need to start a second thread which runs the vl.c qemu_main():
+ *
+ * Initial thread:                    2nd thread:
+ * in main():
+ *  create qemu-main thread
+ *  wait on display_init semaphore
+ *                                    call qemu_main()
+ *                                    ...
+ *                                    in cocoa_display_init():
+ *                                     post the display_init semaphore
+ *                                     wait on app_started semaphore
+ *  create application, menus, etc
+ *  enter OSX run loop
+ * in applicationDidFinishLaunching:
+ *  post app_started semaphore
+ *                                     tell main thread to fullscreen if needed
+ *                                    [...]
+ *                                    run qemu main-loop
+ *
+ * We do this in two stages so that we don't do the creation of the
+ * GUI application menus and so on for command line options like --help
+ * where we want to just print text to stdout and exit immediately.
+ */
+
+static void *call_qemu_main(void *opaque)
+{
+    int status;
+
+    COCOA_DEBUG("Second thread: calling qemu_main()\n");
+    status = qemu_main(gArgc, gArgv, *_NSGetEnviron());
+    COCOA_DEBUG("Second thread: qemu_main() returned, exiting\n");
+    [cbowner release];
+    exit(status);
+}
+
 // Mac to QKeyCode conversion
 static const int mac_to_qkeycode_map[] = {
     [kVK_ANSI_A] = Q_KEY_CODE_A,
@@ -1804,9 +1847,6 @@ static void addRemovableDevicesMenuItems(void)
     qapi_free_BlockInfoList(pointerToFree);
 }
 
-@interface QemuCocoaPasteboardTypeOwner : NSObject<NSPasteboardTypeOwner>
-@end
-
 @implementation QemuCocoaPasteboardTypeOwner
 
 - (void)pasteboard:(NSPasteboard *)sender provideDataForType:(NSPasteboardType)type
@@ -1841,8 +1881,6 @@ static void addRemovableDevicesMenuItems(void)
 
 @end
 
-static QemuCocoaPasteboardTypeOwner *cbowner;
-
 static void cocoa_clipboard_notify(Notifier *notifier, void *data);
 static void cocoa_clipboard_request(QemuClipboardInfo *info,
                                     QemuClipboardType type);
@@ -1903,44 +1941,6 @@ static void cocoa_clipboard_request(QemuClipboardInfo *info,
     }
 }
 
-/*
- * The startup process for the OSX/Cocoa UI is complicated, because
- * OSX insists that the UI runs on the initial main thread, and so we
- * need to start a second thread which runs the vl.c qemu_main():
- *
- * Initial thread:                    2nd thread:
- * in main():
- *  create qemu-main thread
- *  wait on display_init semaphore
- *                                    call qemu_main()
- *                                    ...
- *                                    in cocoa_display_init():
- *                                     post the display_init semaphore
- *                                     wait on app_started semaphore
- *  create application, menus, etc
- *  enter OSX run loop
- * in applicationDidFinishLaunching:
- *  post app_started semaphore
- *                                     tell main thread to fullscreen if needed
- *                                    [...]
- *                                    run qemu main-loop
- *
- * We do this in two stages so that we don't do the creation of the
- * GUI application menus and so on for command line options like --help
- * where we want to just print text to stdout and exit immediately.
- */
-
-static void *call_qemu_main(void *opaque)
-{
-    int status;
-
-    COCOA_DEBUG("Second thread: calling qemu_main()\n");
-    status = qemu_main(gArgc, gArgv, *_NSGetEnviron());
-    COCOA_DEBUG("Second thread: qemu_main() returned, exiting\n");
-    [cbowner release];
-    exit(status);
-}
-
 int main (int argc, char **argv) {
     QemuThread thread;
 
Signature · 1 line
2.34.1
#2 lore raw

RFC v4 2/2 ui/cocoa: run qemu_init in the main thread

In reply to Philippe Mathieu-Daudé

From: Paolo Bonzini <[email protected]>

Simplify the initialization dance by running qemu_init() in the main thread before the Cocoa event loop starts. The cocoa_display_init() code that is post-applicationDidFinishLaunching: moves to the application delegate itself, and the secondary thread only runs the rest of qemu_main(), namely qemu_main_loop() and qemu_cleanup().

This fixes a case where addRemovableDevicesMenuItems() calls qmp_query_block() while expecting the main thread to still hold the BQL. The newly-introduced assertions in the block layer complain about this.

Message-Id: <[email protected]>

[PMD: Fixed trivial build failures & rebased]
---
 softmmu/main.c |  12 +++--
 ui/cocoa.m     | 123 +++++++++++++++++++++----------------------------
 2 files changed, 60 insertions(+), 75 deletions(-)
softmmu/main.c +7 -5
index 639c67ff48..0c4384e980 100644
@@ -39,16 +39,18 @@ int main(int argc, char **argv)
 #endif
 #endif /* CONFIG_SDL */
 
-#ifdef CONFIG_COCOA
-#undef main
-#define main qemu_main
-#endif /* CONFIG_COCOA */
-
+#ifndef CONFIG_COCOA
 int main(int argc, char **argv, char **envp)
 {
+    /*
+     * ui/cocoa.m relies on this being the exact content of main(),
+     * because it has to run everything after qemu_init in a secondary
+     * thread.
+     */
     qemu_init(argc, argv, envp);
     qemu_main_loop();
     qemu_cleanup();
 
     return 0;
 }
+#endif /* CONFIG_COCOA */
ui/cocoa.m +53 -70

199 lines not on this page · raw message

Signature · 1 line
2.34.1
#3 lore raw
In reply to Philippe Mathieu-Daudé

Hi Akihiko, Paolo, Peter.

Quoted 14 lines · Philippe Mathieu-Daudé

From: Philippe Mathieu-Daudé <[email protected]>

Posting v4 in case someone want to iterate.

Pending issue raised by Akihiko Odaki:

* this actually breaks the "runas" option with ui/cocoa.

[+NSApplication sharedApplication] calls issetugid() to see if setgid() or setuid() is called before and calls exit() if it evaluates true. It does not evaluate true without this patch since setgid() and setuid() are called after [+NSApplication sharedApplication]. This patch, however, changes the order and triggers the check.

Quoted 4 lines

There are two options to solve the problem:
1. Move setgid and setuid calls after [+NSApplication sharedApplication] to let NSApplication initialize as the original user.

Akihiko, could you send a patch?

Quoted 7 lines

2. Do: [[NSUserDefaults standardUserDefaults] setBool:YES forKey:@"_NSAppAllowsNonTrustedUGID"]

Option 2 would be preferred in terms of practicality since nobody would want to initialize NSApplication as the original user (usually superuser). However, _NSAppAllowsNonTrustedUGID is not documented by Apple.

What are your views on this problem for 7.0-rc1? Keep modifying cocoa UI? Disable block layer assertions? Only disable them for Darwin?

Quoted 35 lines

* Oudated comment in main():

1970 /*

  1971   * Create the menu entries which depend on QEMU state (for consoles
  1972   * and removeable devices). These make calls back into QEMU functions,
  1973   * which is OK because at this point we know that the second thread
  1974   * holds the iothread lock and is synchronously waiting for us to
  1975   * finish.
  1976   */

(https://marc.info/?l=qemu-devel&m=164752136410805)

Since v3:
- Move qemu_event_init before cbowner alloc
- Reduce main_thread scope to applicationDidFinishLaunching
- Updated updateUIInfo() comment (s/cocoa_display_init/applicationDidFinishLaunching)

Since v2:
- Extracted code movement in preliminary patch

v3: https://lore.kernel.org/qemu-devel/[email protected]/
v2: https://lore.kernel.org/qemu-devel/[email protected]/
v1: https://lore.kernel.org/qemu-devel/[email protected]/

Paolo Bonzini (1):
ui/cocoa: run qemu_init in the main thread

Philippe Mathieu-Daudé (1):
ui/cocoa: Code movement

  softmmu/main.c |  12 ++--
  ui/cocoa.m     | 161 ++++++++++++++++++++++---------------------------
  2 files changed, 79 insertions(+), 94 deletions(-)
#4 lore raw
Quoted 17 lines · Philippe Mathieu-Daudé

Hi Akihiko, Paolo, Peter.

Quoted 14 lines · Philippe Mathieu-Daudé

From: Philippe Mathieu-Daudé <[email protected]>

Posting v4 in case someone want to iterate.

Pending issue raised by Akihiko Odaki:

* this actually breaks the "runas" option with ui/cocoa.

   [+NSApplication sharedApplication] calls issetugid() to see if setgid() or setuid() is called before and calls exit() if it evaluates true. It does not evaluate true without this patch since setgid() and setuid() are called after [+NSApplication sharedApplication]. This patch, however, changes the order and triggers the check.

Quoted 6 lines
Quoted 4 lines

   There are two options to solve the problem:
   1. Move setgid and setuid calls after [+NSApplication sharedApplication] to let NSApplication initialize as the original user.

Akihiko, could you send a patch?

Do you mean a patch for option 1?

Quoted 11 lines
Quoted 8 lines

   2. Do: [[NSUserDefaults standardUserDefaults] setBool:YES forKey:@"_NSAppAllowsNonTrustedUGID"]

   Option 2 would be preferred in terms of practicality since nobody would want to initialize NSApplication as the original user (usually superuser). However, _NSAppAllowsNonTrustedUGID is not documented by Apple.

What are your views on this problem for 7.0-rc1? Keep modifying cocoa UI? Disable block layer assertions? Only disable them for Darwin?

I think we should disable block layer assertions. It is not a change visible to user and its value is more apparent in development. We can preserve most of its benefit if we restore the assertions immediately after 7.0 release and let them work during the next development cycle.

If it is not preferred, we can apply the following change as a less-intrusive alternative:
https://patchew.org/QEMU/[email protected]/

The change faced objections as it uses Cocoa APIs from iothread. It is still in accordance with Cocoa's API convention and the only negative effect is that it can confuse developers. It is not affecting users and we can also minimize the possibility of confusion by immediately following with this "qemu_init in the main thread" patch after 7.0 release.

Regards,
Akihiko Odaki

Quoted 40 lines
Quoted 39 lines

* Oudated comment in main():

  1970  /*
  1971   * Create the menu entries which depend on QEMU state (for consoles 1972   * and removeable devices). These make calls back into QEMU functions, 1973   * which is OK because at this point we know that the second thread 1974   * holds the iothread lock and is synchronously waiting for us to 1975   * finish.
  1976   */

(https://marc.info/?l=qemu-devel&m=164752136410805)

Since v3:
- Move qemu_event_init before cbowner alloc
- Reduce main_thread scope to applicationDidFinishLaunching
- Updated updateUIInfo() comment (s/cocoa_display_init/applicationDidFinishLaunching)

Since v2:
- Extracted code movement in preliminary patch

v3: https://lore.kernel.org/qemu-devel/[email protected]/

v2: https://lore.kernel.org/qemu-devel/[email protected]/

v1: https://lore.kernel.org/qemu-devel/[email protected]/

Paolo Bonzini (1):
   ui/cocoa: run qemu_init in the main thread

Philippe Mathieu-Daudé (1):
   ui/cocoa: Code movement

  softmmu/main.c |  12 ++--
  ui/cocoa.m     | 161 ++++++++++++++++++++++---------------------------

  2 files changed, 79 insertions(+), 94 deletions(-)
#5 lore raw
In reply to Philippe Mathieu-Daudé
Quoted 3 lines · Philippe Mathieu-Daudé
Quoted 3 lines

   1. Move setgid and setuid calls after [+NSApplication sharedApplication] to let NSApplication initialize as the original user.

Another possibility is to move the code up to "[QemuApplication sharedApplication]" from main() to cocoa_display_init().

Paolo

#6 lore raw
Quoted 8 lines · Paolo Bonzini
Quoted 3 lines · Philippe Mathieu-Daudé
Quoted 3 lines

   1. Move setgid and setuid calls after [+NSApplication sharedApplication] to let NSApplication initialize as the original user.

Another possibility is to move the code up to "[QemuApplication sharedApplication]" from main() to cocoa_display_init().

Paolo

I'm for moving everything except [NSApp run] to cocoa_display_init().

This series moved Cocoa display initialization code to [-QemuCocoaApplicatinController applicationDidFinishLaunching:] to keep the initialization order for Cocoa, but it is unlikely that doing initialization things before [-QemuCocoaApplicatinController applicationDidFinishLaunching:] would cause problems. (The only Apple thing affected by the change would be dispatch_sync, which is rarely used.) The code movement is rather intrusive for QEMU as found with this "runas" problem.

Doing Cocoa initialization things in [-QemuCocoaApplicatinController applicationDidFinishLaunching:] and after the method call can be even dangerous as it would modify the state of Cocoa, which is already running. "Create menus in iothread" patch series actually triggered such a problem and required careful coding:
https://patchew.org/QEMU/[email protected]/

The code movement also requires more code to keep the content of DisplayOptions and to have the definition of [-QemuCocoaApplicatinController applicationDidFinishLaunching:]. It would be simply nice if we could remove them.

Regards,
Akihiko Odaki