Using markers w/ preemptible RCU in early code
Hello,
I'm working on the kmemtrace GSoC project and I'm having a little problem.
My code inserts probes during early code (just after kmem_cache_init()). This works on the classic RCU. But on the preemptible RCU, this triggers BUG()s (supposedly on each probe, there are two), although the kernel runs fine:
Quoted 14 lines
BUG: scheduling while atomic: swapper/0/0x00000002Pid: 0, comm: swapper Not tainted 2.6.25-00002-gf80e324-dirty #16
Call Trace: [<ffffffff802365f5>] __schedule_bug+0x65/0x70 [<ffffffff804f8bb0>] thread_return+0x346/0x566 [<ffffffff802734ba>] ? get_marker+0x23a/0x260 [<ffffffff8025fbc6>] ? put_online_cpus+0x46/0x70 [<ffffffff80270c88>] __synchronize_sched+0x48/0x80 [<ffffffff802741a2>] marker_probe_register+0x152/0x660 [<ffffffff8029af10>] ? kmemtrace_probe_alloc+0x0/0x1b0 [<ffffffff8029ad6d>] kmemtrace_init+0x4d/0xd0 [<ffffffff80693c05>] start_kernel+0x205/0x300 [<ffffffff806931b2>] _sinittext+0x1b2/0x200
The is triggered by the following code in marker_probe_register():
#ifdef CONFIG_PREEMPT_RCU synchronize_sched(); /* Until we have the call_rcu_sched() */
#endif
Since preemption and SMP are disabled during early code, we could do something like:
kernel/marker.c +4 -1
@@ -23,6 +23,7 @@#include <linux/rcupdate.h> #include <linux/marker.h> #include <linux/err.h> +#include <linux/hardirq.h> extern struct marker __start___markers[];extern struct marker __stop___markers[];
@@ -672,7 +673,9 @@ int marker_probe_register(const char *name, const char *format,/* write rcu_pending before calling the RCU callback */ smp_wmb ();#ifdef CONFIG_PREEMPT_RCU -synchronize_sched ();/* Until we have the call_rcu_sched() */ +/* We are not preemptible when registering probes in early code */ +if (likely (preemptible ())) +synchronize_sched ();/* Until we have the call_rcu_sched() */ #endif call_rcu (&entry->rcu, free_old_closure); end:
Is this the best approach? marker_probe_register() isn't a hot path, so that check won't mess up the performance. Anyway, this works for me.
Eduard
(forgot LKML, added now.)
Quoted 45 lines · * Eduard - Gabriel Munteanu ([email protected]) wrote:
Hello,
I'm working on the kmemtrace GSoC project and I'm having a little problem.
My code inserts probes during early code (just after
kmem_cache_init()). This works on the classic RCU. But on the preemptible RCU, this triggers BUG()s (supposedly on each probe, there are two), although the kernel runs fine:Quoted 14 lines
BUG: scheduling while atomic: swapper/0/0x00000002Pid: 0, comm: swapper Not tainted 2.6.25-00002-gf80e324-dirty #16
Call Trace: [<ffffffff802365f5>] __schedule_bug+0x65/0x70 [<ffffffff804f8bb0>] thread_return+0x346/0x566 [<ffffffff802734ba>] ? get_marker+0x23a/0x260 [<ffffffff8025fbc6>] ? put_online_cpus+0x46/0x70 [<ffffffff80270c88>] __synchronize_sched+0x48/0x80 [<ffffffff802741a2>] marker_probe_register+0x152/0x660 [<ffffffff8029af10>] ? kmemtrace_probe_alloc+0x0/0x1b0 [<ffffffff8029ad6d>] kmemtrace_init+0x4d/0xd0 [<ffffffff80693c05>] start_kernel+0x205/0x300 [<ffffffff806931b2>] _sinittext+0x1b2/0x200The is triggered by the following code in
marker_probe_register():#ifdef CONFIG_PREEMPT_RCU synchronize_sched(); /* Until we have the call_rcu_sched() */#endif
Since preemption and SMP are disabled during early code, we could do something like:
kernel/marker.c +1
-0@@ -23,6 +23,7 @@#include <linux/rcupdate.h> #include <linux/marker.h> #include <linux/err.h> +#include <linux/hardirq.h> extern struct marker __start___markers[];extern struct marker __stop___markers[];
kernel/marker.c int marker_probe_register(const char *name, const char *format, · line 680
Quoted 9 lines
kernel/marker.c +3
-1@@ -672,7 +673,9 @@ int marker_probe_register(const char *name, const char *format,/* write rcu_pending before calling the RCU callback */ smp_wmb ();#ifdef CONFIG_PREEMPT_RCU -synchronize_sched ();/* Until we have the call_rcu_sched() */ +/* We are not preemptible when registering probes in early code */ +if (likely (preemptible ())) +synchronize_sched ();/* Until we have the call_rcu_sched() */ #endif call_rcu (&entry->rcu, free_old_closure);
I think call_rcu can become a problem too in that case. It is responsible for freeing the old closure and I don't think it will be executed if rcu or the scheduler is not active. Therefore, we should also do something like :
if (likely(preemptible()))
call_rcu(&entry->rcu, free_old_closure);
else
free_old_closure(&entry->rcu); (modulo passing the right
parameters..)
All the synchronize_sched, cal_rcu and rcu_barriers should be changed. I think the rcu_barriers should also have this kind of test to check for early boot use.
Mathieu
Quoted 10 lines
kernel/marker.c +0
-0end:Is this the best approach?
marker_probe_register()isn't a hot path, so that check won't mess up the performance. Anyway, this works for me.Eduard
(forgot LKML, added now.)
Signature · 2 lines
Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
A message that is not in the archive (sent off-list or not yet received): OFFCD1EB75.A69A5A3C-ON88257445.007BA01C-88257446.0002C7EF@us.ibm.com
Quoted 5 lines · On Sat, 10 May 2008 17:30:23 -0700 Paul McKenney <[email protected]>…
There might be an easy and harmless fix for early-boot use of
synchronize_sched()-- I will look into this.As noted above,
call_rcu()is not a problem.
Quoted 8 lines
The rcu_barrier*() primitives must never be invoked at early boot time. These primitives absolutely must block until all of the corresponding flavor of callback has been invoked (and completed), but it is not possible to block during early boot. Perhaps they should do a
WARN_ON()orBUG_ON()in this case?But why are you calling
rcu_barrier()during early boot in the first place?
Hi,
Mathieu's code does not call rcu_barrier() explicitely during early boot. It just happens that I need to use his code (i.e. markers) at that moment.
My understanding is that we could skip any synchronization primitives during early code, since there is no SMP and no preemption (and no IRQs popping around, maybe?). RCU stuff still needs to get registered for later on, though.
I guess we should write a marker_probe_register() variant for early code, with no rcu_barrier() and no synchronize_sched() around.
Mathieu, does this sound okay?
(Paul, I removed your alias from the Cc list, so you don't get duplicates.)
Thanks for your input,
Eduard