On Thu, Sep 10, 2015 at 8:39 AM, Vlastimil Babka <vbabka@xxxxxxx> wrote:
[CC Joonsoo, Mel]
On 09/09/2015 08:31 PM, Vitaly Wool wrote:
> Hi Laura,
>
> On Wed, Sep 9, 2015 at 7:56 PM, Laura Abbott <labbott@xxxxxxxxxx> wrote:
>
>> (cc-ing linux-mm)
>> On 09/09/2015 07:44 AM, Vitaly Wool wrote:
>>
>>> __zone_watermark_ok() does not corrrectly take high-order
>>> CMA pageblocks into account: high-order CMA blocks are not
>>> removed from the watermark check. Moreover, CMA pageblocks
>>> may suddenly vanish through CMA allocation, so let's not
>>> regard these pages as free in __zone_watermark_ok().
>>>
>>> This patch also adds some primitive testing for the method
>>> implemented which has proven that it works as it should.
>>>
>>>
>> The choice to include CMA as part of watermarks was pretty deliberate.
>> Do you have a description of the problem you are facing with
>> the watermark code as is? Any performance numbers?
>>
>>
> let's start with facing the fact that the calculation in
> __zone_watermark_ok() is done incorrectly for the case when ALLOC_CMA is
> not set. While going through pages by order it is implicitly considered
You're not the first who tried to fix it, I think Joonsoo tried as well?
I think the main objection was against further polluting fastpaths due to CMA.
I believe Joonsoo was calculating free_pages incorrectly, too, but in a different way: he was subtracting CMA pages twice.
Note that Mel has a patchset removing high-order watermark checks (in the last
patch of https://lwn.net/Articles/655406/ ) so this will be moot afterwards.
I am not quite convinced that nested loops are a better solution than what I suggest.
> that CMA pages can be used and this impacts the result of the function.
>
> This can be solved in a slightly different way compared to what I proposed
> but it needs per-order CMA pages accounting anyway. Then it would have
> looked like:
>
> for (o = 0; o < order; o++) {
> /* At the next order, this order's pages become unavailable
> */
> free_pages -= z->free_area[o].nr_free << o;
> #ifdef CONFIG_CMA
> if (!(alloc_flags & ALLOC_CMA))
> free_pages -= z->free_area[o].nr_free_cma << o;
> /* Require fewer higher order pages to be free */
> min >>= 1;
> ...
>
> But what we have also seen is that CMA pages may suddenly disappear due to
> CMA allocator work so the whole watermark checking was still unreliable,
> causing compaction to not run when it ought to and thus leading to
Well, watermark checking is inherently racy. CMA pages disappearing is no
exception, non-CMA pages may disappear as well.
Right, that is why I decided to play on the safe side.
~vitaly