runqueue: Don't defer setscene tasks behind already completed tasks
PATCH 1/1 runqueue: Don't defer setscene tasks behind already completed tasks
When a task completes, the first setscene task deferred against it is released and any others are re-deferred behind that task. If the released task had already completed (e.g. its setscene was skipped due to a valid stamp), it never completes again, so the re-deferred tasks are never released. Setscene then can't finish until the deadlock fallback forces each one, printing "Runqueue deadlocked on deferred tasks".
This shows up with three or more multiconfigs sharing a task hash. Skip already completed tasks when picking the new blocker.
AI-Generated: Uses GitHub Copilot
--- lib/bb/runqueue.py | 4 ++++ 1 file changed, 4 insertions(+)
lib/bb/runqueue.py +4 -0
@@ -2112,6 +2112,10 @@ class RunQueueExecute: bb.debug(1, "Deferred task %s now buildable" % t) del self.sq_deferred[t] update_scenequeue_data([t], self.sqdata, self.rqdata, self.rq, self.cooker, self.stampcache, self, summary=False) + # A task which already completed will never complete again so must + # not become the new blocker, else the remaining tasks never get released + if t in self.runq_complete: + continue found = t else: bb.debug(1, "Deferring %s after %s" % (t, found))
Signature · 1 line
2.34.1
Quoted 23 lines · Mark Hatle via bitbake-devel
When a task completes, the first setscene task deferred against it is released and any others are re-deferred behind that task. If the released task had already completed (e.g. its setscene was skipped due to a valid stamp), it never completes again, so the re-deferred tasks are never released. Setscene then can't finish until the deadlock fallback forces each one, printing "Runqueue deadlocked on deferred tasks".
This shows up with three or more multiconfigs sharing a task hash. Skip already completed tasks when picking the new blocker.
AI-Generated: Uses GitHub Copilot
Signed-off-by Mark Hatle <[email protected]>---
lib/bb/runqueue.py | 4 ++++1 file changed, 4 insertions(+)lib/bb/runqueue.py +0
-0@@ -2112,6 +2112,10 @@ class RunQueueExecute: bb.debug(1, "Deferred task %s now buildable" % t)del self.sq_deferred[t]
Quoted 8 lines
update_scenequeue_data([t], self.sqdata, self.rqdata, self.rq, self.cooker, self.stampcache, self, summary=False)+ # A task which already completed will never complete again so must
+ # not become the new blocker, else the remaining tasks never get released
+ if t in self.runq_complete:
+ continue
found = t
else:
bb.debug(1, "Deferring %s after %s" % (t, found))
This doesn't look quite right to me.
If you do that, you need to also then follow through and handle anything that has a dependency on t.
Also, the code above this is the only codepath where self.runq_complete.add(task) is called. For something to be in runq_complete, this function has to have been called for t, so that deferral should have already been handled too.
I have a different suspicion. I think the loop ordering is subtly wrong, which is leading to the deadlocking.
We need to process deferred tasks in multilib order. The main code already does this but this section just has a sorted() call. I think it may need to be:
lib/bb/runqueue.py +4 -1
@@ -2109,7 +2109,10 @@ class RunQueueExecute: logger.debug("Marking task %s as buildable", revdep) found = None - for t in sorted(self.sq_deferred.copy()): + for mc in sorted(self.sqdata.multiconfigs): + for t in sorted(self.sq_deferred.copy()): + if mc_from_tid(t) != mc: + continue if self.sq_deferred[t] == task: # Allow the next deferred task to run. Any other deferred tasks should be deferred after that task. # We shouldn't allow all to run at once as it is prone to races.
which then matches the other code. We probably need a small helper function to give that sorting. (yes, I know the indentation needs fixing, I left it like this for clarity on what the change is).
Can you see if that helps please?
Thanks,
Richard
On 11/16/26 7:32 AM, Richard Purdie wrote:
...
Quoted 35 lines
This doesn't look quite right to me.
If you do that, you need to also then follow through and handle anything that has a dependency on t.
Also, the code above this is the only codepath where
self.runq_complete.add(task)is called. For something to be inrunq_complete, this function has to have been called for t, so that deferral should have already been handled too.I have a different suspicion. I think the loop ordering is subtly wrong, which is leading to the deadlocking.
We need to process deferred tasks in multilib order. The main code already does this but this section just has a
sorted()call. I think it may need to be:lib/bb/runqueue.py +4
-1@@ -2109,7 +2109,10 @@ class RunQueueExecute: logger.debug("Marking task %s as buildable", revdep) found = None - for t in sorted(self.sq_deferred.copy()): + for mc in sorted(self.sqdata.multiconfigs): + for t in sorted(self.sq_deferred.copy()): + if mc_from_tid(t) != mc: + continue if self.sq_deferred[t] == task: # Allow the next deferred task to run. Any other deferred tasks should be deferred after that task. # We shouldn't allow all to run at once as it is prone to races.
Quoted 5 lines
which then matches the other code. We probably need a small helper function to give that sorting. (yes, I know the indentation needs fixing, I left it like this for clarity on what the change is).
Can you see if that helps please?
Assuming I applied this properly, I'm still getting the same results:
WARNING: Runqueue deadlocked on deferred tasks, forcing task mc:zynqmp-zcu102-sdt-full-microblaze-pmu:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_populate_lic blocked by mc:zynqmp-zcu102-sdt-full-cortexa53-fsbl:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_populate_lic
WARNING: Runqueue deadlocked on deferred tasks, forcing task mc:zynqmp-zcu102-sdt-full-microblaze-pmu:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_populate_sysroot blocked by mc:zynqmp-zcu102-sdt-full-cortexa53-fsbl:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_populate_sysroot
WARNING: Runqueue deadlocked on deferred tasks, forcing task mc:zynqmp-zcu102-sdt-full-microblaze-pmu:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_recipe_qa blocked by mc:zynqmp-zcu102-sdt-full-cortexa53-fsbl:virtual:native:/scratch2/mhatle/git/external/experimental/release/meta/recipes-support/sqlite/sqlite3_3.54.1.bb:do_recipe_qa
(and many more)
--Mark
Quoted 5 lines
Thanks,
Richard
Quoted 23 lines · Mark Hatle via bitbake-devel
When a task completes, the first setscene task deferred against it is released and any others are re-deferred behind that task. If the released task had already completed (e.g. its setscene was skipped due to a valid stamp), it never completes again, so the re-deferred tasks are never released. Setscene then can't finish until the deadlock fallback forces each one, printing "Runqueue deadlocked on deferred tasks".
This shows up with three or more multiconfigs sharing a task hash. Skip already completed tasks when picking the new blocker.
AI-Generated: Uses GitHub Copilot
Signed-off-by Mark Hatle <[email protected]>---
lib/bb/runqueue.py | 4 ++++1 file changed, 4 insertions(+)lib/bb/runqueue.py +0
-0@@ -2112,6 +2112,10 @@ class RunQueueExecute: bb.debug(1, "Deferred task %s now buildable" % t)del self.sq_deferred[t]
Quoted 8 lines
update_scenequeue_data([t], self.sqdata, self.rqdata, self.rq, self.cooker, self.stampcache, self, summary=False)+ # A task which already completed will never complete again so must
+ # not become the new blocker, else the remaining tasks never get released
+ if t in self.runq_complete:
+ continue
found = t
else:
bb.debug(1, "Deferring %s after %s" % (t, found))
Just to confirm, this is not quite the right fix. I did work out exactly why it was breaking and whilst this works around the problem, this breaks things. There is a better fix and I've sent patches.
Cheers,
Richard