Re: [RFC PATCH 1/2] lib, stackdepot: check stackdepot handle before accessing slabs.

Re: [RFC PATCH 1/2] lib, stackdepot: check stackdepot handle before accessing slabs.

[Vlastimil Babka]

+CC linux-mm, Alex

On 9/1/21 07:19, Imran Khan wrote:
> stack_depot_save allocates slabs that will be used for storing
> objects in future.If this slab allocation fails we may get to
> a situation where space allocation for a new stack_record fails,
> causing stack_depot_save to return 0 as handle.
> If user of this handle ends up invoking stack_depot_fetch with
> this handle value, current implementation of stack_depot_fetch
> will end up using slab from wrong index.
> To avoid this check handle value at the beginning.
> Also issue a warning for nil handle values and when slab allocation
> for stackdepot fails for the first time.
> 
> Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
> Suggested-by: Vlastimil Babka <vbabka@suse.cz>

Agree but without the warnings please, especially the "stack depot handle is
absent" one. It's just something that can happen e.g. in GFP_NOWAIT contexts
and no need to spam dmesg.

> ---
>  lib/stackdepot.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/lib/stackdepot.c b/lib/stackdepot.c
> index 0a2e417f83cb..1d42ef9ef766 100644
> --- a/lib/stackdepot.c
> +++ b/lib/stackdepot.c
> @@ -232,6 +232,10 @@ unsigned int stack_depot_fetch(depot_stack_handle_t handle,
>  	struct stack_record *stack;
>  
>  	*entries = NULL;
> +	if (!handle) {
> +		WARN(1, "stack depot handle is absent.\n");
> +		return 0;
> +	}
>  	if (parts.slabindex > depot_index) {
>  		WARN(1, "slab index %d out of bounds (%d) for stack id %08x\n",
>  			parts.slabindex, depot_index, handle);
> @@ -303,6 +307,8 @@ depot_stack_handle_t stack_depot_save(unsigned long *entries,
>  		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
>  		if (page)
>  			prealloc = page_address(page);
> +		else
> +			WARN_ONCE(1, "slab allocation for stack depot failed.\n");
>  	}
>  
>  	raw_spin_lock_irqsave(&depot_lock, flags);
> 

Re: [RFC PATCH 1/2] lib, stackdepot: check stackdepot handle before accessing slabs.

[imran.f.khan]

On 1/9/21 6:58 pm, Vlastimil Babka wrote:
> +CC linux-mm, Alex
> 
> On 9/1/21 07:19, Imran Khan wrote:
>> stack_depot_save allocates slabs that will be used for storing
>> objects in future.If this slab allocation fails we may get to
>> a situation where space allocation for a new stack_record fails,
>> causing stack_depot_save to return 0 as handle.
>> If user of this handle ends up invoking stack_depot_fetch with
>> this handle value, current implementation of stack_depot_fetch
>> will end up using slab from wrong index.
>> To avoid this check handle value at the beginning.
>> Also issue a warning for nil handle values and when slab allocation
>> for stackdepot fails for the first time.
>>
>> Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
>> Suggested-by: Vlastimil Babka <vbabka@suse.cz>
> 
> Agree but without the warnings please, especially the "stack depot handle is
> absent" one. It's just something that can happen e.g. in GFP_NOWAIT contexts
> and no need to spam dmesg.
> 

Okay. I have removed warnings in v2 of patch set.

>> ---
>>   lib/stackdepot.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/lib/stackdepot.c b/lib/stackdepot.c
>> index 0a2e417f83cb..1d42ef9ef766 100644
>> --- a/lib/stackdepot.c
>> +++ b/lib/stackdepot.c
>> @@ -232,6 +232,10 @@ unsigned int stack_depot_fetch(depot_stack_handle_t handle,
>>   	struct stack_record *stack;
>>   
>>   	*entries = NULL;
>> +	if (!handle) {
>> +		WARN(1, "stack depot handle is absent.\n");
>> +		return 0;
>> +	}
>>   	if (parts.slabindex > depot_index) {
>>   		WARN(1, "slab index %d out of bounds (%d) for stack id %08x\n",
>>   			parts.slabindex, depot_index, handle);
>> @@ -303,6 +307,8 @@ depot_stack_handle_t stack_depot_save(unsigned long *entries,
>>   		page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER);
>>   		if (page)
>>   			prealloc = page_address(page);
>> +		else
>> +			WARN_ONCE(1, "slab allocation for stack depot failed.\n");
>>   	}
>>   
>>   	raw_spin_lock_irqsave(&depot_lock, flags);
>>
> 
Thanks for reviewing this.

--Imran

Re: [RFC PATCH 2/2] lib, stackdepot: Add helper to print stack entries.

[Vlastimil Babka]

On 9/1/21 07:19, Imran Khan wrote:
> To print a stack entries, users of stackdepot, first
> use stack_depot_fetch to get a list of stack entries
> and then use stack_trace_print to print this list.
> Provide a helper in stackdepot to print stack entries
> based on stackdepot handle.
> 
> Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
> Suggested-by: Vlastimil Babka <vbabka@suse.cz>

You should convert existing users together with the patch that introduces
the helper. I think print_stack() in mm/kasan/report.c, and
__dump_page_owner() could use this.

> ---
>  lib/stackdepot.c | 17 +++++++++++++++++
>  1 file changed, 17 insertions(+)

It's missing a declaration in include/linux/stackdepot.h
Perhaps it could be all be a static inline there anyway.

> diff --git a/lib/stackdepot.c b/lib/stackdepot.c
> index 1d42ef9ef766..eab4476b060b 100644
> --- a/lib/stackdepot.c
> +++ b/lib/stackdepot.c
> @@ -214,6 +214,23 @@ static inline struct stack_record *find_stack(struct stack_record *bucket,
>  	return NULL;
>  }
>  
> +/**
> + * stack_depot_print - print stack entries from a depot
> + *
> + * @handle:		Stack depot handle which was returned from
> + *			stack_depot_save().
> + *
> + */
> +void stack_depot_print(depot_stack_handle_t stack)
> +{
> +	unsigned long *entries;
> +	unsigned int nr_entries;
> +
> +	nr_entries = stack_depot_fetch(stack, &entries);
> +	stack_trace_print(entries, nr_entries, 0);
> +}
> +EXPORT_SYMBOL_GPL(stack_depot_print);
> +
>  /**
>   * stack_depot_fetch - Fetch stack entries from a depot
>   *
> 

Re: [RFC PATCH 2/2] lib, stackdepot: Add helper to print stack entries.

[imran.f.khan]

Hi Vlastimil,

On 1/9/21 7:07 pm, Vlastimil Babka wrote:
> On 9/1/21 07:19, Imran Khan wrote:
>> To print a stack entries, users of stackdepot, first
>> use stack_depot_fetch to get a list of stack entries
>> and then use stack_trace_print to print this list.
>> Provide a helper in stackdepot to print stack entries
>> based on stackdepot handle.
>>
>> Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
>> Suggested-by: Vlastimil Babka <vbabka@suse.cz>
> 
> You should convert existing users together with the patch that introduces
> the helper. I think print_stack() in mm/kasan/report.c, and
> __dump_page_owner() could use this.
> 

Okay. I have done this in v2 of the patch set. BTW I also see some users 
(one place in page owner, rest all in some gpu drivers) of 
stack_depot_fetch + stack_trace_snprintf. Could you please let me know 
if it would be okay to add a helper corresponding to 
stack_trace_snprintf as well.

>> ---
>>   lib/stackdepot.c | 17 +++++++++++++++++
>>   1 file changed, 17 insertions(+)
> 
> It's missing a declaration in include/linux/stackdepot.h
> Perhaps it could be all be a static inline there anyway.
> 

Agree. I have added missing declaration now. I have not made it inline
though.
>> diff --git a/lib/stackdepot.c b/lib/stackdepot.c
>> index 1d42ef9ef766..eab4476b060b 100644
>> --- a/lib/stackdepot.c
>> +++ b/lib/stackdepot.c
>> @@ -214,6 +214,23 @@ static inline struct stack_record *find_stack(struct stack_record *bucket,
>>   	return NULL;
>>   }
>>   
>> +/**
>> + * stack_depot_print - print stack entries from a depot
>> + *
>> + * @handle:		Stack depot handle which was returned from
>> + *			stack_depot_save().
>> + *
>> + */
>> +void stack_depot_print(depot_stack_handle_t stack)
>> +{
>> +	unsigned long *entries;
>> +	unsigned int nr_entries;
>> +
>> +	nr_entries = stack_depot_fetch(stack, &entries);
>> +	stack_trace_print(entries, nr_entries, 0);
>> +}
>> +EXPORT_SYMBOL_GPL(stack_depot_print);
>> +
>>   /**
>>    * stack_depot_fetch - Fetch stack entries from a depot
>>    *
>>
> 

Thanks again for the review.

--Imran

Re: [RFC PATCH 2/2] lib, stackdepot: Add helper to print stack entries.

[Vlastimil Babka]

On 9/2/21 02:11, imran.f.khan@oracle.com wrote:
> Hi Vlastimil,
> 
> On 1/9/21 7:07 pm, Vlastimil Babka wrote:
>> On 9/1/21 07:19, Imran Khan wrote:
>>> To print a stack entries, users of stackdepot, first
>>> use stack_depot_fetch to get a list of stack entries
>>> and then use stack_trace_print to print this list.
>>> Provide a helper in stackdepot to print stack entries
>>> based on stackdepot handle.
>>>
>>> Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
>>> Suggested-by: Vlastimil Babka <vbabka@suse.cz>
>>
>> You should convert existing users together with the patch that introduces
>> the helper. I think print_stack() in mm/kasan/report.c, and
>> __dump_page_owner() could use this.
>>
> 
> Okay. I have done this in v2 of the patch set. BTW I also see some users
> (one place in page owner, rest all in some gpu drivers) of stack_depot_fetch
> + stack_trace_snprintf. Could you please let me know if it would be okay to
> add a helper corresponding to stack_trace_snprintf as well.

Yeah looks like that would make sense too, i915 even has this as
__print_depot_stack(). Thanks.