Re: [PATCH] mm/swap_slots.c: don't reset the cache slot after use

[Tim Chen]

On 3/10/20 3:20 PM, Wei Yang wrote:
> On Tue, Mar 10, 2020 at 11:13:13AM -0700, Tim Chen wrote:
>> On 3/9/20 5:48 PM, Andrew Morton wrote:
>>> On Mon,  9 Mar 2020 17:09:40 +0800 Wei Yang <richard.weiyang@xxxxxxxxxxxxxxxxx> wrote:
>>>
>>>> Currently we would clear the cache slot if it is used. While this is not
>>>> necessary, since this entry would not be used until refilled.
>>>>
>>>> Leave it untouched and assigned the value directly to entry which makes
>>>> the code little more neat.
>>>>
>>>> Also this patch merges the else and if, since this is the only case we
>>>> refill and repeat swap cache.
>>>
>>> cc Tim, who can hopefully remember how this code works ;)
>>>
>>>> --- a/mm/swap_slots.c
>>>> +++ b/mm/swap_slots.c
>>>> @@ -309,7 +309,7 @@ int free_swap_slot(swp_entry_t entry)
>>>>  
>>>>  swp_entry_t get_swap_page(struct page *page)
>>>>  {
>>>> -	swp_entry_t entry, *pentry;
>>>> +	swp_entry_t entry;
>>>>  	struct swap_slots_cache *cache;
>>>>  
>>>>  	entry.val = 0;
>>>> @@ -336,13 +336,10 @@ swp_entry_t get_swap_page(struct page *page)
>>>>  		if (cache->slots) {
>>>>  repeat:
>>>>  			if (cache->nr) {
>>>> -				pentry = &cache->slots[cache->cur++];
>>>> -				entry = *pentry;
>>>> -				pentry->val = 0;
>>
>> The cache entry was cleared after assignment for defensive programming,  So there's
>> little chance I will be using a slot that has been assigned to someone else.
>> When I wrote swap_slots.c, this code was new and I want to make sure
>> that if something went wrong, and I assigned a swap slot that I shouldn't,
>> I will be able to detect quickly as I will only be stepping on entry 0.
>>
>> Otherwise such bug will be harder to detect as we will have two users of some random
>> swap slot stepping on each other.
>>
>> I'm okay if we want to get rid of this logic, now that the code has been
>> working correctly long enough.  But I think is good hygiene to clear the
>> cached entry after it has been assigned. 
>>
> 
> This is fine to keep the logic, while I am wondering whether we need to do
> this through pointer. cache->slots[] contain the value, we can get and reset
> without pointer.
> 
> The following code looks more obvious about the logic.
> 
> 		entry = cache->slots[cache->cur];
> 		cache->slots[cache->cur++].val = 0;

Yes, this looks pretty good.

Thanks.

Tim