current linux-2.6.git: cpusets completely broken
2008/7/12 Dmitry Adamushko <[email protected]>:
Quoted 32 lines
2008/7/12 Linus Torvalds <[email protected]>:
Quoted 11 lines · Vegard Nossum
Quoted 3 lines
Can somebody else please test/ack/review it too? This should eventually go into 2.6.26 if it doesn't break anything else.
And Dmitry, _please_ also explain what was going on. Why did things break from calling
common_cpu_mem_hotplug_unplug()too much? That function is called pretty randomly anyway (for just about any random CPU event), so why did it fail in some circumstances?Upon
CPU_DOWN_PREPARE,update_sched_domains()->detach_destroy_domains(&cpu_online_map);
does the following:/*
* Force a reinitialization of the sched domains hierarchy. The domains
* and groups cannot be updated in place without racing with the balancing
* code, so we temporarily attach all running cpus to the NULL domain
* which will prevent rebalancing while the sched domains are recalculated. */The sched-domains should be rebuilt when a
CPU_DOWNops. is completed, effectivelly either upon CPU_DEAD{_FROZEN} (upon success) or CPU_DOWN_FAILED{_FROZEN} (upon failure -- restore the things to their initial state). That's whatupdate_sched_domains()also does but only for !CPUSETS case.With Max's patch, sched-domains' reinitialization is delegated to CPUSETS code:
Quoted 7 lines
cpuset_handle_cpuhp()->common_cpu_mem_hotplug_unplug()->rebuild_sched_domains()which as you've said "called pretty randomly anyway", e.g. for
CPU_UP_PREPARE.[ ah, then
rebuild_sched_domains()should not be there. It should be nop for MEMPLUG events I presume - should make another patch. ]
I had in mind something like this:
[ yes, probably the patch makes things somewhat uglier. I tried to bring a minimal amount of changes so far, just to emulate the 'old' behavior of update_sched_domains(). I guess, common_cpu_mem_hotplug_unplug() needs to be split up into cpu- and mem-hotplug parts to make it cleaner ]
(not tested yet)
---
kernel/cpuset.c +18 -6
@@ -1882,7 +1882,7 @@ static void scan_for_empty_cpusets(const struct cpuset *root)* in order to minimize text size. */ -static void common_cpu_mem_hotplug_unplug (void ) +static void common_cpu_mem_hotplug_unplug (int rebuild_sd) {cgroup_lock ();
@@ -1894,7 +1894,8 @@ static void common_cpu_mem_hotplug_unplug(void)* Scheduler destroys domains on hotplug events. * Rebuild them based on the current settings. */ -rebuild_sched_domains (); +if (rebuild_sd) +rebuild_sched_domains ();cgroup_unlock (); }
@@ -1912,11 +1913,22 @@ static void common_cpu_mem_hotplug_unplug(void)static int cpuset_handle_cpuhp (struct notifier_block *unused_nb,unsigned long phase,void *unused_cpu) { -if (phase ==CPU_DYING || phase ==CPU_DYING_FROZEN ) +swicth (phase) { +case CPU_UP_CANCELED : +case CPU_UP_CANCELED_FROZEN : +case CPU_DOWN_FAILED : +case CPU_DOWN_FAILED_FROZEN : +case CPU_ONLINE : +case CPU_ONLINE_FROZEN : +case CPU_DEAD : +case CPU_DEAD_FROZEN : +common_cpu_mem_hotplug_unplug (1 ); +break ; +default :return NOTIFY_DONE ; + } -common_cpu_mem_hotplug_unplug (); -return 0 ; +return NOTIFY_OK ; }#ifdef CONFIG_MEMORY_HOTPLUG
@@ -1929,7 +1941,7 @@ static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,void cpuset_track_online_nodes (void ) { -common_cpu_mem_hotplug_unplug (); +common_cpu_mem_hotplug_unplug (0 ); }#endif
Quoted 41 lines · Dmitry Adamushko
2008/7/12 Dmitry Adamushko <[email protected]>:
Quoted 32 lines
2008/7/12 Linus Torvalds <[email protected]>:
Quoted 11 lines · Vegard Nossum
Quoted 3 lines
Can somebody else please test/ack/review it too? This should eventually go into 2.6.26 if it doesn't break anything else.
And Dmitry, _please_ also explain what was going on. Why did things break from calling
common_cpu_mem_hotplug_unplug()too much? That function is called pretty randomly anyway (for just about any random CPU event), so why did it fail in some circumstances?Upon
CPU_DOWN_PREPARE,update_sched_domains()->detach_destroy_domains(&cpu_online_map);
does the following:/*
* Force a reinitialization of the sched domains hierarchy. The domains
* and groups cannot be updated in place without racing with the balancing
* code, so we temporarily attach all running cpus to the NULL domain
* which will prevent rebalancing while the sched domains are recalculated. */The sched-domains should be rebuilt when a
CPU_DOWNops. is completed, effectivelly either upon CPU_DEAD{_FROZEN} (upon success) or CPU_DOWN_FAILED{_FROZEN} (upon failure -- restore the things to their initial state). That's whatupdate_sched_domains()also does but only for !CPUSETS case.With Max's patch, sched-domains' reinitialization is delegated to CPUSETS code:
Quoted 7 lines
cpuset_handle_cpuhp()->common_cpu_mem_hotplug_unplug()->rebuild_sched_domains()which as you've said "called pretty randomly anyway", e.g. for
CPU_UP_PREPARE.[ ah, then
rebuild_sched_domains()should not be there. It should be nop for MEMPLUG events I presume - should make another patch. ]
Quoted 8 lines
I had in mind something like this:
[ yes, probably the patch makes things somewhat uglier. I tried to bring a minimal amount of changes so far, just to emulate the 'old' behavior of
update_sched_domains(). I guess,common_cpu_mem_hotplug_unplug()needs to be split up into cpu- and mem-hotplug parts to make it cleaner ](not tested yet)
---
argh, this one compiles (will test shortly).
kernel/cpuset.c +0 -0
@@ -1882,7 +1882,7 @@ static void scan_for_empty_cpusets(const struct
cpuset *root)
* in order to minimize text size.
*/
-static void common_cpu_mem_hotplug_unplug(void) +static void common_cpu_mem_hotplug_unplug(int rebuild_sd)
{
cgroup_lock();
@@ -1894,7 +1894,8 @@ static void common_cpu_mem_hotplug_unplug(void)
* Scheduler destroys domains on hotplug events.
* Rebuild them based on the current settings.
*/
- rebuild_sched_domains();
+ if (rebuild_sd)
+ rebuild_sched_domains();
cgroup_unlock();
}
@@ -1912,11 +1913,22 @@ static void common_cpu_mem_hotplug_unplug(void)
static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
unsigned long phase, void *unused_cpu)
{
- if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
+ switch (phase) {
+ case CPU_UP_CANCELED:
+ case CPU_UP_CANCELED_FROZEN:
+ case CPU_DOWN_FAILED:
+ case CPU_DOWN_FAILED_FROZEN:
+ case CPU_ONLINE:
+ case CPU_ONLINE_FROZEN:
+ case CPU_DEAD:
+ case CPU_DEAD_FROZEN:
+ common_cpu_mem_hotplug_unplug(1);
+ break;
+ default:
return NOTIFY_DONE;
+ }
- common_cpu_mem_hotplug_unplug();
- return 0;
+ return NOTIFY_OK;
}
#ifdef CONFIG_MEMORY_HOTPLUG
@@ -1929,7 +1941,7 @@ static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
void cpuset_track_online_nodes(void)
{
- common_cpu_mem_hotplug_unplug();
+ common_cpu_mem_hotplug_unplug(0);
}
#endif
Linus,
(just that we have it all together in one place, ready for testing and further consideration).
below is the patch and explanation.
Basically the fix below just emulates the 'old' behavior
of update_sched_domains(). We call rebuild_sched_domains() for the same hotplug-events as it was called (and is still called for !CPUSETS case) in update_sched_domains(). The aim is to keep sched-domain consistent wrt cpu-down/up.
This should be a minimal change. Effectively, the change is against f18f982abf183e91f435990d337164c7a43d1e6d. So the logic of this patch should be easily visible comparing it to what the aforementioned commit does.
Ingo, could also please comment on this issue? TIA.
Subject: fix cpuset_handle_cpuhp()
The following commit
---
commit f18f982abf183e91f435990d337164c7a43d1e6d
Author: Max Krasnyansky <[email protected]> Date: Thu May 29 11:17:01 2008 -0700
sched: CPU hotplug events must not destroy scheduler domains created by the cpusets
---
[ Note, with this commit arch_update_cpu_topology is not called any more for CPUSETS. But it's just a nop. The whole scheme should be probably reworked later. ]
introduced a hotplug-related problem as described below:
[ Basically the fix below just emulates the 'old' behavior of update_sched_domains(). We call rebuild_sched_domains() for the same hotplug-events as it was called (and is still called for !CPUSETS case) in update_sched_domains(). ]
Upon CPU_DOWN_PREPARE, update_sched_domains() -> detach_destroy_domains(&cpu_online_map) does the following:
/*
* Force a reinitialization of the sched domains hierarchy. The domains
* and groups cannot be updated in place without racing with the
balancing
* code, so we temporarily attach all running cpus to the NULL domain
* which will prevent rebalancing while the sched domains are
recalculated.
*/
The sched-domains should be rebuilt when a CPU_DOWN ops. has been completed, effectively either upon CPU_DEAD{_FROZEN} (upon success) or CPU_DOWN_FAILED{_FROZEN} (upon failure -- restore the things to their initial state). That's what update_sched_domains() also does but only for !CPUSETS case.
With Max's patch, sched-domains' reinitialization is delegated to CPUSETS code:
cpuset_handle_cpuhp() -> common_cpu_mem_hotplug_unplug() -> rebuild_sched_domains()
Being called for CPU_UP_PREPARE and if its callback is called after update_sched_domains()), it just negates all the work done by update_sched_domains() -- i.e. a soon-to-be-offline cpu is included in the sched-domains and that makes it visible for the load-balancer while the CPU_DOWN ops. is in progress.
__migrate_live_tasks() moves the tasks off a 'dead' cpu (it's already "offline" when this function is called).
try_to_wake_up() is called for one of these tasks from another CPU -> the load-balancer (wake_idle()) picks up a "dead" CPU and places the task on it. Then e.g. BUG_ON(rq->nr_running) detects this a bit later -> oops.
---
kernel/cpuset.c +18 -6
@@ -1882,7 +1882,7 @@ static void scan_for_empty_cpusets(const struct cpuset *root)* in order to minimize text size. */ -static void common_cpu_mem_hotplug_unplug (void ) +static void common_cpu_mem_hotplug_unplug (int rebuild_sd) {cgroup_lock ();
@@ -1894,7 +1894,8 @@ static void common_cpu_mem_hotplug_unplug(void)* Scheduler destroys domains on hotplug events. * Rebuild them based on the current settings. */ -rebuild_sched_domains (); +if (rebuild_sd) +rebuild_sched_domains ();cgroup_unlock (); }
@@ -1912,11 +1913,22 @@ static void common_cpu_mem_hotplug_unplug(void)static int cpuset_handle_cpuhp (struct notifier_block *unused_nb,unsigned long phase,void *unused_cpu) { -if (phase ==CPU_DYING || phase ==CPU_DYING_FROZEN ) +switch (phase) { +case CPU_UP_CANCELED : +case CPU_UP_CANCELED_FROZEN : +case CPU_DOWN_FAILED : +case CPU_DOWN_FAILED_FROZEN : +case CPU_ONLINE : +case CPU_ONLINE_FROZEN : +case CPU_DEAD : +case CPU_DEAD_FROZEN : +common_cpu_mem_hotplug_unplug (1 ); +break ; +default :return NOTIFY_DONE ; + } -common_cpu_mem_hotplug_unplug (); -return 0 ; +return NOTIFY_OK ; }#ifdef CONFIG_MEMORY_HOTPLUG
@@ -1929,7 +1941,7 @@ static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,void cpuset_track_online_nodes (void ) { -common_cpu_mem_hotplug_unplug (); +common_cpu_mem_hotplug_unplug (0 ); }#endif
---
Vegard Nossum
Tested-byQuoted 57 lines · Dmitry Adamushko
Subject: fix
cpuset_handle_cpuhp()The following commit
---
commitf18f982abf183e91f435990d337164c7a43d1e6dAuthor: Max Krasnyansky <[email protected]> Date: Thu May 29 11:17:01 2008 -0700sched: CPU hotplug events must not destroy scheduler domains created by the cpusets
---[ Note, with this commit
arch_update_cpu_topologyis not called any more for CPUSETS. But it's just a nop. The whole scheme should be probably reworked later. ]introduced a hotplug-related problem as described below:
[ Basically the fix below just emulates the 'old' behavior of
update_sched_domains(). We callrebuild_sched_domains()for the same hotplug-events as it was called (and is still called for !CPUSETS case) inupdate_sched_domains(). ]Upon
CPU_DOWN_PREPARE,update_sched_domains()->detach_destroy_domains(&cpu_online_map)does the following:/*
* Force a reinitialization of the sched domains hierarchy. The domains
* and groups cannot be updated in place without racing with the
balancing
* code, so we temporarily attach all running cpus to the NULL domain
* which will prevent rebalancing while the sched domains are
recalculated.
*/The sched-domains should be rebuilt when a
CPU_DOWNops. has been completed, effectively either upon CPU_DEAD{_FROZEN} (upon success) or CPU_DOWN_FAILED{_FROZEN} (upon failure -- restore the things to their initial state). That's whatupdate_sched_domains()also does but only for !CPUSETS case.With Max's patch, sched-domains' reinitialization is delegated to CPUSETS code:
cpuset_handle_cpuhp()->common_cpu_mem_hotplug_unplug()->rebuild_sched_domains()Being called for
CPU_UP_PREPAREand if its callback is called afterupdate_sched_domains()), it just negates all the work done byupdate_sched_domains()-- i.e. a soon-to-be-offline cpu is included in the sched-domains and that makes it visible for the load-balancer while theCPU_DOWNops. is in progress.
__migrate_live_tasks()moves the tasks off a 'dead' cpu (it's already "offline" when this function is called).
Quoted 7 lines
try_to_wake_up()is called for one of these tasks from another CPU -> the load-balancer (wake_idle()) picks up a "dead" CPU and places the task on it. Then e.g.BUG_ON(rq->nr_running)detects this a bit later -> oops.Signed-off-by Dmitry Adamushko <[email protected]>
Works :-)
Vegard
Signature · 4 lines
"The animistic metaphor of the bug that maliciously sneaked in while the programmer was not looking is intellectually dishonest as it disguises that the error is the programmer's own creation." -- E. W. Dijkstra, EWD1036
Quoted 65 lines · Vegard Nossum
Quoted 63 lines · Dmitry Adamushko
Subject: fix
cpuset_handle_cpuhp()The following commit
---
commitf18f982abf183e91f435990d337164c7a43d1e6dAuthor: Max Krasnyansky <[email protected]> Date: Thu May 29 11:17:01 2008 -0700sched: CPU hotplug events must not destroy scheduler domains created by the cpusets
---[ Note, with this commit
arch_update_cpu_topologyis not called any more for CPUSETS. But it's just a nop. The whole scheme should be probably reworked later. ]introduced a hotplug-related problem as described below:
[ Basically the fix below just emulates the 'old' behavior of
update_sched_domains(). We callrebuild_sched_domains()for the same hotplug-events as it was called (and is still called for !CPUSETS case) inupdate_sched_domains(). ]Upon
CPU_DOWN_PREPARE,update_sched_domains()->detach_destroy_domains(&cpu_online_map)does the following:/*
* Force a reinitialization of the sched domains hierarchy. The domains
* and groups cannot be updated in place without racing with the
balancing
* code, so we temporarily attach all running cpus to the NULL domain
* which will prevent rebalancing while the sched domains are
recalculated.
*/The sched-domains should be rebuilt when a
CPU_DOWNops. has been completed, effectively either upon CPU_DEAD{_FROZEN} (upon success) or CPU_DOWN_FAILED{_FROZEN} (upon failure -- restore the things to their initial state). That's whatupdate_sched_domains()also does but only for !CPUSETS case.With Max's patch, sched-domains' reinitialization is delegated to CPUSETS code:
cpuset_handle_cpuhp()->common_cpu_mem_hotplug_unplug()->rebuild_sched_domains()Being called for
CPU_UP_PREPAREand if its callback is called afterupdate_sched_domains()), it just negates all the work done byupdate_sched_domains()-- i.e. a soon-to-be-offline cpu is included in the sched-domains and that makes it visible for the load-balancer while theCPU_DOWNops. is in progress.
__migrate_live_tasks()moves the tasks off a 'dead' cpu (it's already "offline" when this function is called).
try_to_wake_up()is called for one of these tasks from another CPU -> the load-balancer (wake_idle()) picks up a "dead" CPU and places the task on it. Then e.g.BUG_ON(rq->nr_running)detects this a bit later -> oops.
Quoted 5 lines
Quoted 1 line
Signed-off-by Dmitry Adamushko <[email protected]>Tested-by Vegard Nossum <[email protected]>Works :-)
thanks! I've tidied up the changelog and queued it up into tip/sched/urgent. I'd prefer this more conservative patch so late in the cycle, but i'll also queue up the more intrusive real fix from Linus and Dmitry in sched/devel.
Linus, if you've not applied it already, you can pull Dmitry's fix from:
git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip.git sched-fixes-for-linus
shortlog, diffstat and diff below.
Thanks,
Ingo
------------------>
Dmitry Adamushko (1):
cpusets, hotplug, scheduler: fix scheduler domain breakage
kernel/cpuset.c | 24 ++++++++++++++++++------ 1 files changed, 18 insertions(+), 6 deletions(-)
kernel/cpuset.c +18 -6
@@ -1882,7 +1882,7 @@ static void scan_for_empty_cpusets(const struct cpuset *root)* in order to minimize text size. */ -static void common_cpu_mem_hotplug_unplug (void ) +static void common_cpu_mem_hotplug_unplug (int rebuild_sd) {cgroup_lock ();
@@ -1894,7 +1894,8 @@ static void common_cpu_mem_hotplug_unplug(void)* Scheduler destroys domains on hotplug events. * Rebuild them based on the current settings. */ -rebuild_sched_domains (); +if (rebuild_sd) +rebuild_sched_domains ();cgroup_unlock (); }
@@ -1912,11 +1913,22 @@ static void common_cpu_mem_hotplug_unplug(void)static int cpuset_handle_cpuhp (struct notifier_block *unused_nb,unsigned long phase,void *unused_cpu) { -if (phase ==CPU_DYING || phase ==CPU_DYING_FROZEN ) +switch (phase) { +case CPU_UP_CANCELED : +case CPU_UP_CANCELED_FROZEN : +case CPU_DOWN_FAILED : +case CPU_DOWN_FAILED_FROZEN : +case CPU_ONLINE : +case CPU_ONLINE_FROZEN : +case CPU_DEAD : +case CPU_DEAD_FROZEN : +common_cpu_mem_hotplug_unplug (1 ); +break ; +default :return NOTIFY_DONE ; + } -common_cpu_mem_hotplug_unplug (); -return 0 ; +return NOTIFY_OK ; }#ifdef CONFIG_MEMORY_HOTPLUG
@@ -1929,7 +1941,7 @@ static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,void cpuset_track_online_nodes (void ) { -common_cpu_mem_hotplug_unplug (); +common_cpu_mem_hotplug_unplug (0 ); }#endif