Re: ZONE_PADDING wastes 4 bytes of the new cacheline

[Nick Piggin <nickpiggin@yahoo.com.au>]

Andrea Arcangeli wrote:
> On Sat, Oct 23, 2004 at 02:33:39PM +1000, Nick Piggin wrote:
> 
>>Andrea Arcangeli wrote:
>>
>>>I don't see any benefit in limiting the high order, infact it seems a
>>>bad bug. If something you should limit the _small_ order, so that the
>>>high order will have a slight chance to succeed. You're basically doing
>>>the opposite.
>>>
>>
>>You need the order there so someone can't allocate a huge amount
>>of memory and deplete all your reserves and crash the system.
> 
> 
> what? the point of alloc_pages is to allow people to allocate memory,
> what's the difference of 1 2M allocation and 512 4k allocations? No
> difference at all. Infact if something the 512 4k allocations hurts a
> lot more since they can fragment the memory, while the single 2M
> allocation will not fragment the memory. So if you really want to limit
> something you should do the exact opposite of what the allocator is
> doing.
> 
> 
>>For day to day running, it should barely make a difference because
>>the watermarks will be an order of magnitude larger.
> 
> 
> yes, it makes little difference, this is why it doesn't hurt that much.
> 

It is an unlikely scenario, but it is definitely good for robustness.
Especially on small memory systems where the amount allocated doesn't
have to be that large.

Let's say a 16MB system pages_low ~= 64K, so we'll also say we've
currently got 64K free. Someone then wants to do an order 4 allocation
OK they succeed (assuming memory isn't fragmented) and there's 0K free.

Which is bad because you can now get deadlocks when trying to free
memory.

> 
>>AFAIKS, pages_min, pages_low and pages_high are all required for
>>what we want to be doing. I don't see you you could remove any one
>>of them and still have everything functioning properly....
>>
>>I haven't really looked at 2.4 or your patches though. Maybe I
>>misunderstood you.
> 
> 
> 2.4 has everything functionally properly but it has no
> pages_min/low/high, it only has the watermarks. Infact the watermarks
> _are_ low/min/high. That's what I'm forward porting to 2.6 (besides
> fixing minor mostly not noticeable but harmful bits like the order
> nosense described above).
> 

Oh if you've still got the three watermarks then that may work -
I thought you meant getting rid of one of the *completely*.

But I'm still not sure what advantage you see in moving from
pages_xxx + protection to a single watermark.