[PATCH] drm/amdkfd: rework criu_restore_bos error handling

Felix Kuehling felix.kuehling at amd.com
Sat Feb 19 03:03:18 UTC 2022


Am 2022-02-18 um 21:34 schrieb Tom Rix:
>
> On 2/18/22 10:35 AM, Felix Kuehling wrote:
>> Am 2022-02-18 um 12:39 schrieb trix at redhat.com:
>>> From: Tom Rix <trix at redhat.com>
>>>
>>> Clang static analysis reports this problem
>>> kfd_chardev.c:2327:2: warning: 1st function call argument
>>>    is an uninitialized value
>>>    kvfree(bo_privs);
>>>    ^~~~~~~~~~~~~~~~
>>>
>>> If the copy_from_users(bo_buckets, ...) fails, there is a jump to
>>> the generic error handler at exit:.  The freeing of bo_privs and
>>> unwinding of the dmabuf_fd loop do not need to be done.
>>>
>>> Add some specific labels for the early failures.
>>> Reorder the frees to be the reverse of their allocs.
>>>
>>> Move the initialize of 'i' back to the loop.
>>> The problem with the early frees predates the loop
>>> unwinding problem.
>>
>> I think the existing error handling strategy in this function is 
>> fine. Having only one exit label avoids potential issues when using 
>> the wrong label. Freeing NULL pointers is not a problem. The loop 
>> becomes a noop if i==0 (this was fixed by you in a previous patch). 
>> The only real problem I see is that bo_privs is not initialized. So 
>> this should really be a one-line or maybe two-line fix:
>>
>>     struct kfd_criu_bo_bucket *bo_buckets = NULL;
>>     struct kfd_criu_bo_priv_data *bo_privs = NULL;
>
> This is the other way I considered to fix the problem. So it will work.

OK. I have already submitted this version to amd-staging-drm-next. Thank 
you for reporting the problem.

Regards,
   Felix


>
> Tom
>
>>
>> Regards,
>>   Felix
>>
>>
>>>
>>> Fixes: 73fa13b6a511 ("drm/amdkfd: CRIU Implement KFD restore ioctl")
>>> Signed-off-by: Tom Rix <trix at redhat.com>
>>> ---
>>>   drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 15 +++++++++------
>>>   1 file changed, 9 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c 
>>> b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> index 965af2a08bc0..1d5f41ac3832 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
>>> @@ -2102,7 +2102,7 @@ static int criu_restore_bos(struct kfd_process 
>>> *p,
>>>       const bool criu_resume = true;
>>>       bool flush_tlbs = false;
>>>       int ret = 0, j = 0;
>>> -    uint32_t i = 0;
>>> +    uint32_t i;
>>>         if (*priv_offset + (args->num_bos * sizeof(*bo_privs)) > 
>>> max_priv_data_size)
>>>           return -EINVAL;
>>> @@ -2119,13 +2119,13 @@ static int criu_restore_bos(struct 
>>> kfd_process *p,
>>>       if (ret) {
>>>           pr_err("Failed to copy BOs information from user\n");
>>>           ret = -EFAULT;
>>> -        goto exit;
>>> +        goto free_buckets;
>>>       }
>>>         bo_privs = kvmalloc_array(args->num_bos, sizeof(*bo_privs), 
>>> GFP_KERNEL);
>>>       if (!bo_privs) {
>>>           ret = -ENOMEM;
>>> -        goto exit;
>>> +        goto free_buckets;
>>>       }
>>>         ret = copy_from_user(bo_privs, (void __user 
>>> *)args->priv_data + *priv_offset,
>>> @@ -2133,12 +2133,12 @@ static int criu_restore_bos(struct 
>>> kfd_process *p,
>>>       if (ret) {
>>>           pr_err("Failed to copy BOs information from user\n");
>>>           ret = -EFAULT;
>>> -        goto exit;
>>> +        goto free_privs;
>>>       }
>>>       *priv_offset += args->num_bos * sizeof(*bo_privs);
>>>         /* Create and map new BOs */
>>> -    for (; i < args->num_bos; i++) {
>>> +    for (i = 0; i < args->num_bos; i++) {
>>>           struct kfd_criu_bo_bucket *bo_bucket;
>>>           struct kfd_criu_bo_priv_data *bo_priv;
>>>           struct kfd_dev *dev;
>>> @@ -2323,8 +2323,11 @@ static int criu_restore_bos(struct 
>>> kfd_process *p,
>>>           if (bo_buckets[i].alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
>>>               close_fd(bo_buckets[i].dmabuf_fd);
>>>       }
>>> -    kvfree(bo_buckets);
>>> +free_privs:
>>>       kvfree(bo_privs);
>>> +free_buckets:
>>> +    kvfree(bo_buckets);
>>> +
>>>       return ret;
>>>   }
>>
>


More information about the amd-gfx mailing list