2008-10-24 05:00:19

by Hiroshi Shimamoto

Subject: [PATCH -tip/cpus4096-v2] cpumask: fix memory leak and cpumask corruption

From: Hiroshi Shimamoto <[email protected]>

On CONFIG_CPUMASK_OFFSTACK enabled kernel, when exiting
smp_call_function_many() without calling arch_send_call_function_ipi()
allbutself should be freed.

Also, the memory for allbutself shouldn't be freed when
arch_send_call_function_ipi() is called. The memory region will be
referenced in generic_smp_call_function_interrupt(). And the memory
will be freed at rcu_free_call_data().

Signed-off-by: Hiroshi Shimamoto <[email protected]>
---
kernel/smp.c | 8 ++++++--
1 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/kernel/smp.c b/kernel/smp.c
index da98191..96b39e3 100644
--- a/kernel/smp.c
+++ b/kernel/smp.c
@@ -88,6 +88,7 @@ static void rcu_free_call_data(struct rcu_head *head)
struct call_function_data *data;

data = container_of(head, struct call_function_data, rcu_head);
+ free_cpumask_var(data->cpumask);

kfree(data);
}
@@ -347,7 +348,7 @@ void smp_call_function_many(const struct cpumask *mask,
* into a targetted single call instead since it's faster.
*/
if (!num_cpus)
- return;
+ goto out;
else if (num_cpus == 1) {
cpu = cpumask_first(allbutself);
smp_call_function_single(cpu, func, info, wait);
@@ -382,9 +383,12 @@ void smp_call_function_many(const struct cpumask *mask,
/* optionally wait for the CPUs to complete */
if (wait) {
csd_flag_wait(&data->csd);
- if (unlikely(slowpath))
+ if (unlikely(slowpath)) {
smp_call_function_mask_quiesce_stack(allbutself);
+ free_cpumask_var(allbutself);
+ }
}
+ return;
out:
free_cpumask_var(allbutself);
}
--
1.5.6

2008-10-24 08:55:32

by Rusty Russell

Subject: Re: [PATCH -tip/cpus4096-v2] cpumask: fix memory leak and cpumask corruption

On Friday 24 October 2008 15:59:59 Hiroshi Shimamoto wrote:
> From: Hiroshi Shimamoto <[email protected]>

Hi Hiroshi,

Thanks for looking at this!

> On CONFIG_CPUMASK_OFFSTACK enabled kernel, when exiting
> smp_call_function_many() without calling arch_send_call_function_ipi()
> allbutself should be freed.

That bit's OK, changing the return to "goto out;".

> Also, the memory for allbutself shouldn't be freed when
> arch_send_call_function_ipi() is called. The memory region will be
> referenced in generic_smp_call_function_interrupt(). And the memory
> will be freed at rcu_free_call_data().

As your next patch shows, this should be solved a different way. I'll comment
there.

Cheers,
Rusty.