Re: [PATCH -mm -v4 05/21] mm, THP, swap: Support PMD swap mapping in free_swap_and_cache()/swap_free()

["Huang\, Ying" <ying.huang@intel.com>]

Dave Hansen <dave.hansen@linux.intel.com> writes:

> On 07/10/2018 12:13 AM, Huang, Ying wrote:
>> Dave Hansen <dave.hansen@linux.intel.com> writes:
>>> The code non-resuse was, and continues to be, IMNHO, one of the largest
>>> sources of bugs with the original THP implementation.  It might be
>>> infeasible to do here, but let's at least give it as much of a go as we can.
>> 
>> I totally agree that we should unify the code path for huge and normal
>> page/swap if possible.  One concern is code size for !CONFIG_THP_SWAP.
>
> I've honestly never heard that as an argument before.  In general, our
> .c files implement *full* functionality: the most complex case.  The
> headers #ifdef that functionality down because of our .config or
> architecture.
>
> The thing that matters here is debugging and reviewing the _complicated_
> case, IMNHO.

I agree with your point here.  I will try it and measure the code size
change too.

>> The original method is good for that.  The new method may introduce some
>> huge swap related code that is hard to be eliminated for
>> !CONFIG_THP_SWAP.  Andrew Morton pointed this out for the patchset of
>> the first step of the THP swap optimization.
>> 
>> This may be mitigated at least partly via,
>> 
>> `
>> #ifdef CONFIG_THP_SWAP
>> #define nr_swap_entries(nr)          (nr)
>> #else
>> #define nr_swap_entries(nr)          1
>> #endif
>> 
>> void do_something(swp_entry_t entry, int __nr_entries)
>> {
>>         int i, nr_entries = nr_swap_entries(__nr_entries);
>> 
>>         if (nr_entries = SWAPFILE_CLUSTER)
>>                 ; /* huge swap specific */
>>         else
>>                 ; /* normal swap specific */
>> 
>>         for (i = 0; i < nr_entries; i++) {
>>                 ; /* do something for each entry */
>>         }
>> 
>>         /* ... */
>> }
>> `
>
> While that isn't perfect, it's better than the current state of things.
>
> While you are refactoring things, I think you also need to take a good
> look at roughly chopping this series in half by finding another stopping
> point.  You've done a great job so far of trickling this functionality
> in so far, but 21 patches is quite a bit, and the set is only going to
> get larger.

Yes.  The patchset is too large.  I will try to reduce it if possible.
At least [21/21] can be separated.  [02/21] may be sent separately
too.  Other parts are hard, THP swapin and creating/supporting PMD swap
mapping need to be in one patchset.

Best Regards,
Huang, Ying