Skip to contentSkip to conversation

qemu/main-loop: Disable block backend global state assertion on Cocoa

PATCH Superseded Philippe Mathieu-Daudé 4 participants 8 messages Mar 22, 2022 – Mar 25, 2022

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 qemu/main-loop: Disable block backend global state assertion on Cocoa

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

  $ qemu-system-i386
  Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
  Abort trap: 6

Looking with lldb:

Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552. Process 76914 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert

     frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
  at block-backend.c:552:5 [opt]
      549    */
      550   BlockBackend *blk_all_next(BlockBackend *blk)
      551   {
  --> 552       GLOBAL_STATE_CODE();
      553       return blk ? QTAILQ_NEXT(blk, link)
      554                  : QTAILQ_FIRST(&block_backends);
      555   }

Target 1: (qemu-system-i386) stopped.

(lldb) bt
* thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert

     frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
     frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
     frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
     frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
   * frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1 at block-backend.c:552:5 [opt]
     frame #5: 0x00000001003c00b4 qemu-system-i386`blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt]
     frame #6: 0x00000001003d8f04 qemu-system-i386`qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt]
     frame #7: 0x000000010003ab0c qemu-system-i386`main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt]
     frame #8: 0x000000010003ab04 qemu-system-i386`main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt]
     frame #9: 0x00000001012690f4 dyld`start + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>

---
 include/qemu/main-loop.h | 12 ++++++++++++
 1 file changed, 12 insertions(+)
include/qemu/main-loop.h +12 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
 bool qemu_in_main_thread(void);
 
 /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926
+ */
+#define GLOBAL_STATE_CODE()
+#else
 #define GLOBAL_STATE_CODE()                                         \
     do {                                                            \
         assert(qemu_in_main_thread());                              \
     } while (0)
+#endif /* CONFIG_DARWIN */
 
 /* Mark and check that the function is part of the I/O API. */
 #define IO_CODE()                                                   \
Signature · 1 line
2.35.1
#1 lore raw
Quoted 63 lines · Philippe Mathieu-Daudé

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

   $ qemu-system-i386
   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Abort trap: 6

Looking with lldb:

   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Process 76914 stopped
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
   at block-backend.c:552:5 [opt]
       549    */
       550   BlockBackend *blk_all_next(BlockBackend *blk)
       551   {
   --> 552       GLOBAL_STATE_CODE();
       553       return blk ? QTAILQ_NEXT(blk, link)
       554                  : QTAILQ_FIRST(&block_backends);
       555   }
   Target 1: (qemu-system-i386) stopped.

   (lldb) bt
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
      frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
      frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
      frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
    * frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1 at block-backend.c:552:5 [opt]
      frame #5: 0x00000001003c00b4 qemu-system-i386`blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt]
      frame #6: 0x00000001003d8f04 qemu-system-i386`qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt]
      frame #7: 0x000000010003ab0c qemu-system-i386`main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt]
      frame #8: 0x000000010003ab04 qemu-system-i386`main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt]
      frame #9: 0x00000001012690f4 dyld`start + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>

---
  include/qemu/main-loop.h | 12 ++++++++++++
  1 file changed, 12 insertions(+)
include/qemu/main-loop.h +0 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
  bool qemu_in_main_thread(void);

include/qemu/main-loop.h bool qemu_mutex_iothread_locked(void); · line 280

Quoted 10 lines
include/qemu/main-loop.h +8 -0
 
  /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926

It is better to state that addRemovableDevicesMenuItems() is called from a thread different from the QEMU main thread and the main thread is waiting for that.

include/qemu/main-loop.h bool qemu_mutex_iothread_locked(void); · line 288

Quoted 8 lines
include/qemu/main-loop.h +4 -0
+ */
+#define GLOBAL_STATE_CODE()
+#else
  #define GLOBAL_STATE_CODE()                                         \
      do {                                                            \
          assert(qemu_in_main_thread());                              \
      } while (0)
+#endif /* CONFIG_DARWIN */

It still says CONFIG_DARWIN ;)

Regards,
Akihiko Odaki

Quoted 3 lines
include/qemu/main-loop.h +0 -0
  /* Mark and check that the function is part of the I/O API. */
  #define IO_CODE()                                                   \
#2 lore raw
In reply to Philippe Mathieu-Daudé
Quoted 74 lines · Philippe Mathieu-Daudé

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

   $ qemu-system-i386
   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Abort trap: 6

Looking with lldb:

   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Process 76914 stopped
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
   at block-backend.c:552:5 [opt]
       549    */
       550   BlockBackend *blk_all_next(BlockBackend *blk)
       551   {
   --> 552       GLOBAL_STATE_CODE();
       553       return blk ? QTAILQ_NEXT(blk, link)
       554                  : QTAILQ_FIRST(&block_backends);
       555   }
   Target 1: (qemu-system-i386) stopped.

   (lldb) bt
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
      frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
      frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
      frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
    * frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1 at block-backend.c:552:5 [opt]
      frame #5: 0x00000001003c00b4 qemu-system-i386`blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt]
      frame #6: 0x00000001003d8f04 qemu-system-i386`qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt]
      frame #7: 0x000000010003ab0c qemu-system-i386`main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt]
      frame #8: 0x000000010003ab04 qemu-system-i386`main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt]
      frame #9: 0x00000001012690f4 dyld`start + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>

---
  include/qemu/main-loop.h | 12 ++++++++++++
  1 file changed, 12 insertions(+)
include/qemu/main-loop.h +9 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
  bool qemu_in_main_thread(void);
 
  /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926
+ */

include/qemu/main-loop.h bool qemu_mutex_iothread_locked(void); · line 291

Quoted 10 lines
include/qemu/main-loop.h +3 -0
+#define GLOBAL_STATE_CODE()
+#else
  #define GLOBAL_STATE_CODE()                                         \
      do {                                                            \
          assert(qemu_in_main_thread());                              \
      } while (0)
+#endif /* CONFIG_DARWIN */
 
  /* Mark and check that the function is part of the I/O API. */
  #define IO_CODE()                                                   \

I don't know, it seems to me that the reorganized initialization code had only advantages.

For now, it fixes the regression and makes the Cocoa build much more similar to the others. There is an easy way to fix the -runas regression, by moving the code up to the call of -sharedApplication in cocoa_display_init.

Later, in 7.1, you have time to move more code out of -applicationDidFinishLaunching:, remove the bools that are copied out of the display options, etc.

Paolo

#3 lore raw
Quoted 82 lines · Paolo Bonzini
Quoted 81 lines · Philippe Mathieu-Daudé

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

   $ qemu-system-i386
   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Abort trap: 6

Looking with lldb:

   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552. Process 76914 stopped
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
   at block-backend.c:552:5 [opt]
       549    */
       550   BlockBackend *blk_all_next(BlockBackend *blk)
       551   {
   --> 552       GLOBAL_STATE_CODE();
       553       return blk ? QTAILQ_NEXT(blk, link)
       554                  : QTAILQ_FIRST(&block_backends);
       555   }
   Target 1: (qemu-system-i386) stopped.

   (lldb) bt
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
      frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
      frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
      frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
    * frame #4: 0x000000010057c2d4 qemu-system-i386blk_all_next.cold.1 at block-backend.c:552:5 [opt] frame #5: 0x00000001003c00b4 qemu-system-i386blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt] frame #6: 0x00000001003d8f04 qemu-system-i386qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt] frame #7: 0x000000010003ab0c qemu-system-i386main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt] frame #8: 0x000000010003ab04 qemu-system-i386main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt] frame #9: 0x00000001012690f4 dyldstart + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>
---
  include/qemu/main-loop.h | 12 ++++++++++++

  1 file changed, 12 insertions(+)
include/qemu/main-loop.h +4 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
   bool qemu_in_main_thread(void);
   /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls 
 qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering 

this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926
+ */

include/qemu/main-loop.h +2 -0
+#define GLOBAL_STATE_CODE()
+#else

  #define GLOBAL_STATE_CODE()                                         \

Quoted 8 lines
Quoted 6 lines

      do {                                                            \
          assert(qemu_in_main_thread());                              \
      } while (0)
+#endif /* CONFIG_DARWIN */
  /* Mark and check that the function is part of the I/O API. */ #define IO_CODE()                                                   \

I don't know, it seems to me that the reorganized initialization code had only advantages.

Yeah, I just don't want to break the "no new code after soft-freeze" rule.

Quoted 1 line

For now, it fixes the regression and makes the Cocoa build much more similar to the others.  There is an easy way to fix the -runas regression, by moving the code up to the call of -sharedApplication in cocoa_display_init.

That worked for me, I'll respin and wait to see if Akihiko is happy with the change and Peter won't object to the patchset past hard-freeze.

Quoted 5 lines

Later, in 7.1, you have time to move more code out of -applicationDidFinishLaunching:, remove the bools that are copied out of the display options, etc.

Paolo

#4 lore raw
Quoted 9 lines · Philippe Mathieu-Daudé
Quoted 4 lines · Paolo Bonzini

For now, it fixes the regression and makes the Cocoa build much more similar to the others. There is an easy way to fix the -runas regression, by moving the code up to the call of -sharedApplication in cocoa_display_init.

That worked for me, I'll respin and wait to see if Akihiko is happy with the change and Peter won't object to the patchset past hard-freeze.

Cocoa UI is currently completely broken, so we should have some fix to it. I'm happy to go with whatever you all think is a reasonable approach.

-- PMM

#5 lore raw
In reply to Paolo Bonzini
Quoted 91 lines · Paolo Bonzini
Quoted 80 lines · Philippe Mathieu-Daudé

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

   $ qemu-system-i386
   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Abort trap: 6

Looking with lldb:

   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552. Process 76914 stopped
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
   at block-backend.c:552:5 [opt]
       549    */
       550   BlockBackend *blk_all_next(BlockBackend *blk)
       551   {
   --> 552       GLOBAL_STATE_CODE();
       553       return blk ? QTAILQ_NEXT(blk, link)
       554                  : QTAILQ_FIRST(&block_backends);
       555   }
   Target 1: (qemu-system-i386) stopped.

   (lldb) bt
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
      frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
      frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
      frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
    * frame #4: 0x000000010057c2d4 qemu-system-i386blk_all_next.cold.1 at block-backend.c:552:5 [opt] frame #5: 0x00000001003c00b4 qemu-system-i386blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt] frame #6: 0x00000001003d8f04 qemu-system-i386qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt] frame #7: 0x000000010003ab0c qemu-system-i386main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt] frame #8: 0x000000010003ab04 qemu-system-i386main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt] frame #9: 0x00000001012690f4 dyldstart + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>
---
  include/qemu/main-loop.h | 12 ++++++++++++

  1 file changed, 12 insertions(+)
include/qemu/main-loop.h +4 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
   bool qemu_in_main_thread(void);
   /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls 
 qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering 

this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926
+ */

include/qemu/main-loop.h +2 -0
+#define GLOBAL_STATE_CODE()
+#else
Quoted 7 lines

  #define GLOBAL_STATE_CODE()                                         \
      do {                                                            \
          assert(qemu_in_main_thread());                              \
      } while (0)
+#endif /* CONFIG_DARWIN */
  /* Mark and check that the function is part of the I/O API. */ #define IO_CODE()                                                   \

I don't know, it seems to me that the reorganized initialization code had only advantages.

Quoted 1 line

For now, it fixes the regression and makes the Cocoa build much more similar to the others.  There is an easy way to fix the -runas regression, by moving the code up to the call of -sharedApplication in cocoa_display_init.

So the options are:

#1 disabling the assert for cocoa (this patch)

#2 run qemu_init() in the main thread​ from Paolo [*] with this (?)
patch on top:

-- >8 --

ui/cocoa.m +0 -1
index e69ce97f44..867c222e18 100644
@@ -1946,7 +1946,6 @@ static void cocoa_clipboard_request(QemuClipboardInfo *info,
 int main(int argc, char **argv, char **envp)
 {
     COCOA_DEBUG("Entered main()\n");
-    qemu_event_init(&cbevent, false);
 
     /* Takes iothread lock, released in 

applicationDidFinishLaunching:. */

     qemu_init(argc, argv, envp);

@@ -1958,13 +1957,6 @@ int main(int argc, char **argv, char **envp)

     NSAutoreleasePool * pool = [[NSAutoreleasePool alloc] init];
-    // Pull this console process up to being a fully-fledged graphical
-    // app with a menubar and Dock icon
-    ProcessSerialNumber psn = { 0, kCurrentProcess };
-    TransformProcessType(&psn, kProcessTransformToForegroundApplication);
-
-    [QemuApplication sharedApplication];
-
     create_initial_menus();

     /*
@@ -1976,7 +1968,6 @@ int main(int argc, char **argv, char **envp)
      */
     add_console_menu_entries();
     addRemovableDevicesMenuItems();

- cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init];

// Create an Application controller
QemuCocoaAppController *appController = [[QemuCocoaAppController
alloc] init];
@@ -2089,6 +2080,17 @@ static void cocoa_display_init(DisplayState *ds, DisplayOptions *opts)

     if (opts->u.cocoa.has_left_command_key && 
!opts->u.cocoa.left_command_key) {
         left_command_key_enabled = 0;
     }

+
+ qemu_event_init(&cbevent, false);
+
+ cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init];
+

+    // Pull this console process up to being a fully-fledged graphical
+    // app with a menubar and Dock icon
+    ProcessSerialNumber psn = { 0, kCurrentProcess };
+    TransformProcessType(&psn, kProcessTransformToForegroundApplication);

+
+ [QemuApplication sharedApplication];
}

---

# 3 "Create menus in iothread" approach from Akihiko:
https://lore.kernel.org/qemu-devel/[email protected]/

Is that correct? (#2 patch and the 3 different options)

What is preferred between #2 and #3? I don't have enough knowledge to take the decision, which is why I suggested the chicken-hearted "disable assert" option #1.

Thanks,

Phil.

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

#6 lore raw
Quoted 146 lines · Philippe Mathieu-Daudé
Quoted 91 lines · Paolo Bonzini
Quoted 80 lines · Philippe Mathieu-Daudé

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

Since commit 0439c5a462 ("block/block-backend.c: assertions for block-backend") QEMU crashes when using Cocoa on Darwin hosts.

Example on macOS:

   $ qemu-system-i386
   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552.
   Abort trap: 6

Looking with lldb:

   Assertion failed: (qemu_in_main_thread()), function blk_all_next, file block-backend.c, line 552. Process 76914 stopped
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert frame #4: 0x000000010057c2d4 qemu-system-i386`blk_all_next.cold.1
   at block-backend.c:552:5 [opt]
       549    */
       550   BlockBackend *blk_all_next(BlockBackend *blk)
       551   {
   --> 552       GLOBAL_STATE_CODE();
       553       return blk ? QTAILQ_NEXT(blk, link)
       554                  : QTAILQ_FIRST(&block_backends);
       555   }
   Target 1: (qemu-system-i386) stopped.

   (lldb) bt
   * thread #1, queue = 'com.apple.main-thread', stop reason = hit program assert
      frame #0: 0x00000001908c99b8 libsystem_kernel.dylib`__pthread_kill + 8
      frame #1: 0x00000001908fceb0 libsystem_pthread.dylib`pthread_kill + 288
      frame #2: 0x000000019083a314 libsystem_c.dylib`abort + 164
      frame #3: 0x000000019083972c libsystem_c.dylib`__assert_rtn + 300
    * frame #4: 0x000000010057c2d4 qemu-system-i386blk_all_next.cold.1 at block-backend.c:552:5 [opt] frame #5: 0x00000001003c00b4 qemu-system-i386blk_all_next(blk=<unavailable>) at block-backend.c:552:5 [opt] frame #6: 0x00000001003d8f04 qemu-system-i386qmp_query_block(errp=0x0000000000000000) at qapi.c:591:16 [opt] frame #7: 0x000000010003ab0c qemu-system-i386main [inlined] addRemovableDevicesMenuItems at cocoa.m:1756:21 [opt] frame #8: 0x000000010003ab04 qemu-system-i386main(argc=<unavailable>, argv=<unavailable>) at cocoa.m:1980:5 [opt] frame #9: 0x00000001012690f4 dyldstart + 520

As we are in passed release 7.0 hard freeze, disable the block backend assertion which, while being valuable during development, is not helpful to users. We'll restore this assertion immediately once 7.0 is released and work on a fix.

Cc Kevin Wolf <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Peter Maydell <[email protected]>
Cc Emanuele Giuseppe Esposito <[email protected]>

---
Supersedes: <[email protected]>
---
  include/qemu/main-loop.h | 12 ++++++++++++

  1 file changed, 12 insertions(+)
include/qemu/main-loop.h +4 -0
index 7a4d6a0920..48061f736b 100644
@@ -270,10 +270,22 @@ bool qemu_mutex_iothread_locked(void);
   bool qemu_in_main_thread(void);
   /* Mark and check that the function is part of the global state API. */
+#ifdef CONFIG_COCOA
+/*
+ * When using Cocoa ui, addRemovableDevicesMenuItems() calls 
 qmp_query_block()
+ * while expecting the main thread to still hold the BQL, triggering 

this
+ * assertions in the block layer (commit 0439c5a462). As the Cocoa fix is not
+ * trivial, disable this assertion for the v7.0.0 release when using Cocoa; it
+ * will be restored immediately after the release. This issue is tracked as
+ * https://gitlab.com/qemu-project/qemu/-/issues/926
+ */

include/qemu/main-loop.h +2 -0
+#define GLOBAL_STATE_CODE()
+#else
Quoted 7 lines

  #define GLOBAL_STATE_CODE()                                         \
      do {                                                            \
          assert(qemu_in_main_thread());                              \
      } while (0)
+#endif /* CONFIG_DARWIN */
  /* Mark and check that the function is part of the I/O API. */ #define IO_CODE()                                                   \

I don't know, it seems to me that the reorganized initialization code had only advantages.

Quoted 1 line

For now, it fixes the regression and makes the Cocoa build much more similar to the others.  There is an easy way to fix the -runas regression, by moving the code up to the call of -sharedApplication in cocoa_display_init.

So the options are:

#1 disabling the assert for cocoa (this patch)

#2 run qemu_init() in the main thread​ from Paolo [*] with this (?)
   patch on top:

-- >8 --

ui/cocoa.m +0 -1
index e69ce97f44..867c222e18 100644
@@ -1946,7 +1946,6 @@ static void 
 cocoa_clipboard_request(QemuClipboardInfo *info,
  int main(int argc, char **argv, char **envp)
  {
      COCOA_DEBUG("Entered main()\n");
-    qemu_event_init(&cbevent, false);
 
      /* Takes iothread lock, released in 

applicationDidFinishLaunching:.  */
     qemu_init(argc, argv, envp);

ui/cocoa.m +0 -8
@@ -1958,13 +1957,6 @@ int main(int argc, char **argv, char **envp)
 
      NSAutoreleasePool * pool = [[NSAutoreleasePool alloc] init];
 
-    // Pull this console process up to being a fully-fledged graphical
-    // app with a menubar and Dock icon
-    ProcessSerialNumber psn = { 0, kCurrentProcess };
-    TransformProcessType(&psn, kProcessTransformToForegroundApplication);
-
-    [QemuApplication sharedApplication];
-
      create_initial_menus();
 
      /*
@@ -1976,7 +1968,6 @@ int main(int argc, char **argv, char **envp)
       */
      add_console_menu_entries();
      addRemovableDevicesMenuItems();
-    cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init];
 
      // Create an Application controller
      QemuCocoaAppController *appController = [[QemuCocoaAppController 

alloc] init];

ui/cocoa.m +1 -0
@@ -2089,6 +2080,17 @@ static void cocoa_display_init(DisplayState *ds, 
 DisplayOptions *opts)
      if (opts->u.cocoa.has_left_command_key && 
 !opts->u.cocoa.left_command_key) {
          left_command_key_enabled = 0;
      }
+

ui/cocoa.m static void cocoa_display_init(DisplayState *ds, · line 2095

Quoted 10 lines
ui/cocoa.m +10 -0
+    qemu_event_init(&cbevent, false);
+
+    cbowner = [[QemuCocoaPasteboardTypeOwner alloc] init];
+
+    // Pull this console process up to being a fully-fledged graphical
+    // app with a menubar and Dock icon
+    ProcessSerialNumber psn = { 0, kCurrentProcess };
+    TransformProcessType(&psn, kProcessTransformToForegroundApplication);
+
+    [QemuApplication sharedApplication];

I think the changes for cbevent and cbowner are irrelevant, but basically this should do.

Quoted 8 lines
ui/cocoa.m +0 -1
  }
 
---

# 3 "Create menus in iothread" approach from Akihiko:
https://lore.kernel.org/qemu-devel/[email protected]/

Quoted 5 lines

Is that correct? (#2 patch and the 3 different options)

What is preferred between #2 and #3? I don't have enough knowledge to take the decision, which is why I suggested the chicken-hearted "disable assert" option #1.

#2 can help allowing ui/cocoa to live in the same binary with ui/gtk and ui/sdl2 in the future and I believe it should be eventually (possible after 7.0) done, but requires relatively large code change.

#3 aims to solve only the problem with iothread dependency when creating menus and its scope of modification is minimal. It is concerned that #3 may confuse developers since it touches Cocoa interfaces from iothread although it is still according to its API convention.

I'm pretty sure both of #2 and #3 work, but I would like to rather insist #1 as the safest. As a developer, I don't want see such an ugly workaround like #1 of course, but such feeling does not matter much in practice. On the other hand, having the entire next release cycle for testing #2 is a real advantage even if the possibility of discovering a new problem is very tiny.

Regards,
Akihiko Odaki

Quoted 7 lines

Thanks,

Phil.

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

#7 lore raw
In reply to Philippe Mathieu-Daudé
Quoted 4 lines · Philippe Mathieu-Daudé

What is preferred between #2 and #3? I don't have enough knowledge to take the decision, which is why I suggested the chicken-hearted "disable assert" option #1.

I don't have a strong opinion, but I do strongly think we need to get a fix of some kind into the tree before rc2, and we are rapidly running out of time for that. Pick something, and send patches/pullreq as appropriate, please.

thanks
-- PMM