When we want to duplicate a new process, dup_task_struct() will undergo
a series of allocations. If alloc_thread_info_node() fails, we call
free_task_struct() and return.
This seems right, but it is not. free_task_struct() will not only free
the task struct from the kmem_cache, but will also call
arch_release_task_struct(). The problem is that this function is
supposed to undo whatever arch-specific work done by
arch_dup_task_struct(), that is not yet called at this point. The
particular problem I ran accross was that in x86, we will arrive at
fpu_free() without having ever allocated it.
This code is very ancient, and according to git, it is there since the
pre-git era. But forks don't fail that often, so that made it well
hidden.
Signed-off-by: Glauber Costa <[email protected]>
Reported-by: Frederic Weisbecker <[email protected]>
---
kernel/fork.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/fork.c b/kernel/fork.c
index 152d023..b397435 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
ti = alloc_thread_info_node(tsk, node);
if (!ti) {
- free_task_struct(tsk);
+ kmem_cache_free(task_struct_cachep, tsk);
return NULL;
}
--
1.7.11.4
On Thu 23-08-12 17:08:46, Glauber Costa wrote:
> When we want to duplicate a new process, dup_task_struct() will undergo
> a series of allocations. If alloc_thread_info_node() fails, we call
> free_task_struct() and return.
>
> This seems right, but it is not. free_task_struct() will not only free
> the task struct from the kmem_cache, but will also call
> arch_release_task_struct(). The problem is that this function is
> supposed to undo whatever arch-specific work done by
> arch_dup_task_struct(), that is not yet called at this point. The
> particular problem I ran accross was that in x86, we will arrive at
> fpu_free() without having ever allocated it.
>
> This code is very ancient, and according to git, it is there since the
> pre-git era. But forks don't fail that often, so that made it well
> hidden.
>
> Signed-off-by: Glauber Costa <[email protected]>
> Reported-by: Frederic Weisbecker <[email protected]>
> ---
> kernel/fork.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/fork.c b/kernel/fork.c
> index 152d023..b397435 100644
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
>
> ti = alloc_thread_info_node(tsk, node);
> if (!ti) {
> - free_task_struct(tsk);
> + kmem_cache_free(task_struct_cachep, tsk);
What about ia64 (or !CONFIG_ARCH_THREAD_INFO_ALLOCATOR in general) which
doesn't allocate thread_info at all?
> return NULL;
> }
>
> --
> 1.7.11.4
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--
Michal Hocko
SUSE Labs
On Thu 23-08-12 16:33:12, Michal Hocko wrote:
> On Thu 23-08-12 17:08:46, Glauber Costa wrote:
> > When we want to duplicate a new process, dup_task_struct() will undergo
> > a series of allocations. If alloc_thread_info_node() fails, we call
> > free_task_struct() and return.
> >
> > This seems right, but it is not. free_task_struct() will not only free
> > the task struct from the kmem_cache, but will also call
> > arch_release_task_struct(). The problem is that this function is
> > supposed to undo whatever arch-specific work done by
> > arch_dup_task_struct(), that is not yet called at this point. The
> > particular problem I ran accross was that in x86, we will arrive at
> > fpu_free() without having ever allocated it.
> >
> > This code is very ancient, and according to git, it is there since the
> > pre-git era. But forks don't fail that often, so that made it well
> > hidden.
> >
> > Signed-off-by: Glauber Costa <[email protected]>
> > Reported-by: Frederic Weisbecker <[email protected]>
> > ---
> > kernel/fork.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 152d023..b397435 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
> >
> > ti = alloc_thread_info_node(tsk, node);
> > if (!ti) {
> > - free_task_struct(tsk);
> > + kmem_cache_free(task_struct_cachep, tsk);
>
> What about ia64 (or !CONFIG_ARCH_THREAD_INFO_ALLOCATOR in general) which
> doesn't allocate thread_info at all?
Hit send button too fast. Should read (or CONFIG_ARCH_THREAD_INFO_ALLOCATOR)
ia64 will not fail obviously and there is no other arch which would
define own thread infor allocators but there might be some in future.
--
Michal Hocko
SUSE Labs
On 08/23/2012 06:33 PM, Michal Hocko wrote:
> On Thu 23-08-12 17:08:46, Glauber Costa wrote:
>> When we want to duplicate a new process, dup_task_struct() will undergo
>> a series of allocations. If alloc_thread_info_node() fails, we call
>> free_task_struct() and return.
>>
>> This seems right, but it is not. free_task_struct() will not only free
>> the task struct from the kmem_cache, but will also call
>> arch_release_task_struct(). The problem is that this function is
>> supposed to undo whatever arch-specific work done by
>> arch_dup_task_struct(), that is not yet called at this point. The
>> particular problem I ran accross was that in x86, we will arrive at
>> fpu_free() without having ever allocated it.
>>
>> This code is very ancient, and according to git, it is there since the
>> pre-git era. But forks don't fail that often, so that made it well
>> hidden.
>>
>> Signed-off-by: Glauber Costa <[email protected]>
>> Reported-by: Frederic Weisbecker <[email protected]>
>> ---
>> kernel/fork.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/kernel/fork.c b/kernel/fork.c
>> index 152d023..b397435 100644
>> --- a/kernel/fork.c
>> +++ b/kernel/fork.c
>> @@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
>>
>> ti = alloc_thread_info_node(tsk, node);
>> if (!ti) {
>> - free_task_struct(tsk);
>> + kmem_cache_free(task_struct_cachep, tsk);
>
> What about ia64 (or !CONFIG_ARCH_THREAD_INFO_ALLOCATOR in general) which
> doesn't allocate thread_info at all?
They would return something meaningful here anyway, otherwise this would
already error out and exit.
But you actually have a point. Not all architectures (all but ia64) will
allocathe the task struct from the slab... sigh...
Sorry, I will come up with something for this.
On Thu 23-08-12 16:38:50, Michal Hocko wrote:
> On Thu 23-08-12 16:33:12, Michal Hocko wrote:
> > On Thu 23-08-12 17:08:46, Glauber Costa wrote:
> > > When we want to duplicate a new process, dup_task_struct() will undergo
> > > a series of allocations. If alloc_thread_info_node() fails, we call
> > > free_task_struct() and return.
> > >
> > > This seems right, but it is not. free_task_struct() will not only free
> > > the task struct from the kmem_cache, but will also call
> > > arch_release_task_struct(). The problem is that this function is
> > > supposed to undo whatever arch-specific work done by
> > > arch_dup_task_struct(), that is not yet called at this point. The
> > > particular problem I ran accross was that in x86, we will arrive at
> > > fpu_free() without having ever allocated it.
> > >
> > > This code is very ancient, and according to git, it is there since the
> > > pre-git era. But forks don't fail that often, so that made it well
> > > hidden.
> > >
> > > Signed-off-by: Glauber Costa <[email protected]>
> > > Reported-by: Frederic Weisbecker <[email protected]>
> > > ---
> > > kernel/fork.c | 2 +-
> > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > index 152d023..b397435 100644
> > > --- a/kernel/fork.c
> > > +++ b/kernel/fork.c
> > > @@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
> > >
> > > ti = alloc_thread_info_node(tsk, node);
> > > if (!ti) {
> > > - free_task_struct(tsk);
> > > + kmem_cache_free(task_struct_cachep, tsk);
> >
> > What about ia64 (or !CONFIG_ARCH_THREAD_INFO_ALLOCATOR in general) which
> > doesn't allocate thread_info at all?
>
> Hit send button too fast. Should read (or CONFIG_ARCH_THREAD_INFO_ALLOCATOR)
> ia64 will not fail obviously and there is no other arch which would
> define own thread infor allocators but there might be some in future.
Bahh, and I should have been looking at CONFIG_ARCH_TASK_STRUCT_ALLOCATOR
instead. Anyway ia64 uses page allocator directly so kmem_cache_free is
not appropriate.
--
Michal Hocko
SUSE Labs
On 08/23/2012 06:45 PM, Michal Hocko wrote:
> On Thu 23-08-12 16:38:50, Michal Hocko wrote:
>> On Thu 23-08-12 16:33:12, Michal Hocko wrote:
>>> On Thu 23-08-12 17:08:46, Glauber Costa wrote:
>>>> When we want to duplicate a new process, dup_task_struct() will undergo
>>>> a series of allocations. If alloc_thread_info_node() fails, we call
>>>> free_task_struct() and return.
>>>>
>>>> This seems right, but it is not. free_task_struct() will not only free
>>>> the task struct from the kmem_cache, but will also call
>>>> arch_release_task_struct(). The problem is that this function is
>>>> supposed to undo whatever arch-specific work done by
>>>> arch_dup_task_struct(), that is not yet called at this point. The
>>>> particular problem I ran accross was that in x86, we will arrive at
>>>> fpu_free() without having ever allocated it.
>>>>
>>>> This code is very ancient, and according to git, it is there since the
>>>> pre-git era. But forks don't fail that often, so that made it well
>>>> hidden.
>>>>
>>>> Signed-off-by: Glauber Costa <[email protected]>
>>>> Reported-by: Frederic Weisbecker <[email protected]>
>>>> ---
>>>> kernel/fork.c | 2 +-
>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/kernel/fork.c b/kernel/fork.c
>>>> index 152d023..b397435 100644
>>>> --- a/kernel/fork.c
>>>> +++ b/kernel/fork.c
>>>> @@ -299,7 +299,7 @@ static struct task_struct *dup_task_struct(struct task_struct *orig)
>>>>
>>>> ti = alloc_thread_info_node(tsk, node);
>>>> if (!ti) {
>>>> - free_task_struct(tsk);
>>>> + kmem_cache_free(task_struct_cachep, tsk);
>>>
>>> What about ia64 (or !CONFIG_ARCH_THREAD_INFO_ALLOCATOR in general) which
>>> doesn't allocate thread_info at all?
>>
>> Hit send button too fast. Should read (or CONFIG_ARCH_THREAD_INFO_ALLOCATOR)
>> ia64 will not fail obviously and there is no other arch which would
>> define own thread infor allocators but there might be some in future.
>
> Bahh, and I should have been looking at CONFIG_ARCH_TASK_STRUCT_ALLOCATOR
> instead. Anyway ia64 uses page allocator directly so kmem_cache_free is
> not appropriate.
>
Yes, you are right. Thanks for spotting this