[PATCH 02/16] drm/amdkfd: Don't dereference kfd_process.mm

[Christian König]

Am 26.10.2017 um 18:47 schrieb Felix Kuehling:
> On 2017-10-26 03:33 AM, Christian König wrote:
>> Am 21.10.2017 um 02:23 schrieb Felix Kuehling:
>>> The kfd_process doesn't own a reference to the mm_struct, so it can
>>> disappear without warning even while the kfd_process still exists.
>>> In fact, the delayed kfd_process teardown is triggered by an MMU
>>> notifier when the mm_struct is destroyed. Permanently holding a
>>> reference to the mm_struct would prevent this from happening.
>>>
>>> Therefore, avoid dereferencing the kfd_process.mm pointer and make
>>> it opaque. Use get_task_mm to get a temporary reference to the mm
>>> when it's needed.
>> Actually that patch is unnecessary.
>>
>> Process tear down (and calling the MMU release callback) is triggered
>> when mm_struct->mm_users reaches zero.
>>
>> The mm_struct is freed up when mm_struct->mm_count becomes zero.
>>
>> So what you can do is grab a reference to the mm_struct with mmgrab()
>> and still be notified by process tear down.
> Hmm, the mm_struct has two different reference counters. So if I grab
> the right type of reference it would work. Either way, one of two
> changes is needed:
>
>    * Take a permanent reference to the mm-struct, or
>    * Don't dereference the mm pointer because I'm not holding a reference
>
> IMO, KFD doesn't need to hold a reference to the mm_struct permanently.

Ah! From the patch description I was assuming that you took a reference 
(but the wrong one) and tried to fix a memory leak with that approach.

> So I believe my change still makes sense. I should probably just remove
> the comment "Permanently holding a reference to the mm_struct would
> prevent ...". Like you say, that's not accurate.

Yeah, good idea.

But now reading the patch there is something else which I stumbled over:
> +    /*
> +     * This cast should be safe here because we grabbed a
> +     * reference to the mm in kfd_process_notifier_release
> +     */
> +    WARN_ON(atomic_read(&((struct mm_struct *)p->mm)->mm_count) <= 0);
>          mmdrop(p->mm);
Well that isn't good coding style. You shouldn't obfuscate what pointer 
it is by changing it to "void*", but rather set it to NULL as soon as 
you know that it is stale.

Additional to that it is certainly not job of the driver to warn on a 
run over mm_count.

Regards,
Christian.

>
> Regards,
>    Felix
>
>> Regards,
>> Christian.
>>
>>> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
>>> ---
>>>    drivers/gpu/drm/amd/amdkfd/kfd_events.c  | 19 +++++++++++++++----
>>>    drivers/gpu/drm/amd/amdkfd/kfd_priv.h    |  7 ++++++-
>>>    drivers/gpu/drm/amd/amdkfd/kfd_process.c |  6 +++++-
>>>    3 files changed, 26 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c
>>> b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
>>> index 944abfa..61ce547 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
>>> @@ -24,8 +24,8 @@
>>>    #include <linux/slab.h>
>>>    #include <linux/types.h>
>>>    #include <linux/sched/signal.h>
>>> +#include <linux/sched/mm.h>
>>>    #include <linux/uaccess.h>
>>> -#include <linux/mm.h>
>>>    #include <linux/mman.h>
>>>    #include <linux/memory.h>
>>>    #include "kfd_priv.h"
>>> @@ -904,14 +904,24 @@ void kfd_signal_iommu_event(struct kfd_dev
>>> *dev, unsigned int pasid,
>>>         * running so the lookup function returns a locked process.
>>>         */
>>>        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
>>> +    struct mm_struct *mm;
>>>          if (!p)
>>>            return; /* Presumably process exited. */
>>>    +    /* Take a safe reference to the mm_struct, which may otherwise
>>> +     * disappear even while the kfd_process is still referenced.
>>> +     */
>>> +    mm = get_task_mm(p->lead_thread);
>>> +    if (!mm) {
>>> +        mutex_unlock(&p->mutex);
>>> +        return; /* Process is exiting */
>>> +    }
>>> +
>>>        memset(&memory_exception_data, 0, sizeof(memory_exception_data));
>>>    -    down_read(&p->mm->mmap_sem);
>>> -    vma = find_vma(p->mm, address);
>>> +    down_read(&mm->mmap_sem);
>>> +    vma = find_vma(mm, address);
>>>          memory_exception_data.gpu_id = dev->id;
>>>        memory_exception_data.va = address;
>>> @@ -937,7 +947,8 @@ void kfd_signal_iommu_event(struct kfd_dev *dev,
>>> unsigned int pasid,
>>>            }
>>>        }
>>>    -    up_read(&p->mm->mmap_sem);
>>> +    up_read(&mm->mmap_sem);
>>> +    mmput(mm);
>>>          mutex_lock(&p->event_mutex);
>>>    diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> index 7d86ec9..1a483a7 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>> @@ -494,7 +494,12 @@ struct kfd_process {
>>>         */
>>>        struct hlist_node kfd_processes;
>>>    -    struct mm_struct *mm;
>>> +    /*
>>> +     * Opaque pointer to mm_struct. We don't hold a reference to
>>> +     * it so it should never be dereferenced from here. This is
>>> +     * only used for looking up processes by their mm.
>>> +     */
>>> +    void *mm;
>>>          struct mutex mutex;
>>>    diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> index 3ccb3b5..21d27e5 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
>>> @@ -200,7 +200,11 @@ static void kfd_process_destroy_delayed(struct
>>> rcu_head *rcu)
>>>        struct kfd_process *p;
>>>          p = container_of(rcu, struct kfd_process, rcu);
>>> -    WARN_ON(atomic_read(&p->mm->mm_count) <= 0);
>>> +    /*
>>> +     * This cast should be safe here because we grabbed a
>>> +     * reference to the mm in kfd_process_notifier_release
>>> +     */
>>> +    WARN_ON(atomic_read(&((struct mm_struct *)p->mm)->mm_count) <= 0);
>>>          mmdrop(p->mm);
>>>    
>>