Hi Michal,
thank you very much for your questions, please see my notes below.
On 12/26/2018 11:35 AM, Michal Hocko wrote:
> On Tue 25-12-18 18:39:26, Konstantin Khorenko wrote:
>> Q: Why do we need to bother at all?
>> A: If a node is highly loaded and its memory is significantly fragmented
>> (unfortunately almost any node with serious load has highly fragmented memory)
>> then any high order memory allocation can trigger massive memory shrink and
>> result in quite a big allocation latency. And the node becomes less responsive
>> and users don't like it.
>> The ultimate solution here is to get rid of large allocations, but we need an
>> instrument to detect them.
>
> Can you point to an example of the problem you are referring here? At
> least for costly orders we do bail out early and try to not cause
> massive reclaim. So what is the order that you are concerned about?
Well, this is the most difficult question to answer.
Unfortunately i don't have a reproducer for that, usually we get into situation
when someone experiences significant node slowdown, nodes most often have a lot of RAM,
we check what is going on there and see the node is busy with reclaim.
And almost every time the reason was - fragmented memory and high order allocations.
Mostly of 2nd and 3rd (which is still considered not costly) order.
Recent related issues we faced were about FUSE dev pipe:
d6d931adce11 ("fuse: use kvmalloc to allocate array of pipe_buffer structs.")
and about bnx driver + mtu 9000 which for each packet required page of 2nd order
(and it even failed sometimes, though it was not the root cause):
kswapd0: page allocation failure: order:2, mode:0x4020
Call Trace:
dump_stack+0x19/0x1b
warn_alloc_failed+0x110/0x180
__alloc_pages_nodemask+0x7bf/0xc60
alloc_pages_current+0x98/0x110
kmalloc_order+0x18/0x40
kmalloc_order_trace+0x26/0xa0
__kmalloc+0x279/0x290
bnx2x_frag_alloc.isra.61+0x2a/0x40 [bnx2x]
bnx2x_rx_int+0x227/0x17c0 [bnx2x]
bnx2x_poll+0x1dd/0x260 [bnx2x]
net_rx_action+0x179/0x390
__do_softirq+0x10f/0x2aa
call_softirq+0x1c/0x30
do_softirq+0x65/0xa0
irq_exit+0x105/0x110
do_IRQ+0x56/0xe0
common_interrupt+0x6d/0x6d
And as both places were called very often - the system latency was high.
This warning can be also used to catch allocation of 4th order and higher which may
easily fail. Those places which are ready to get allocation errors and have
fallbacks are marked with __GFP_NOWARN.
>> Q: Why warning? Use tracepoints!
>> A: Well, this is a matter of magic defaults.
>> Yes, you can use tracepoints to catch large allocations, but you need to do this
>> on purpose and regularly and this is to be done by every developer which is
>> quite unreal.
>> On the other hand if you develop something and get a warning, you'll have to
>> think about the reason and either succeed with reworking the code to use
>> smaller allocation sizes (and thus decrease allocation latency!) or just use
>> kvmalloc() if you don't really need physically continuos chunk or come to the
>> conclusion you definitely need physically continuos memory and shut up the
>> warning.
>
> Well, not really. For one thing, there are systems to panic on warning
> and you really do not want to blow up just because somebody is doing a
> large order allocation.
Well, on one hand - yes, i agree with you. That's why i don't suggest to enable
the warning by default right now - until all (most) of large allocations are marked
properly.
But after it's done and there are no (almost) unmarked high order allocations -
why not? This will reveal new cases of high order allocations soon.
i think people who run systems with "kernel.panic_on_warn" enabled do care
about reporting issues.
>> Q: Why compile time config option?
>> A: In order not to decrease the performance even a bit in case someone does not
>> want to hunt for large allocations.
>> In an ideal life i'd prefer this check/warning is enabled by default and may be
>> even without a config option so it works on every node. Once we find and rework
>> or mark all large allocations that would be good by default. Until that though
>> it will be noisy.
>
> So who is going to enable this option?
At the beginning - people who want to debug kernel and verify their fallbacks on
memory allocations failures in the code or just speed up their code on nodes
with fragmented memory - for 2nd and 3rd orders.
mm performance issues are tough, you know, and this is just another way to
gain more performance. It won't avoid the necessity of digging mm for sure,
but might decrease the pressure level.
Later (i hope) i could be enabled by default so all big new allocations
are verified sooner and either reworked or marked with __GFP_NOWARN if the code is
ready.
--
Best regards,
Konstantin Khorenko,
Virtuozzo Linux Kernel Team
>> Another option is to rework the patch via static keys (having the warning
>> disabled by default surely). That makes it possible to turn on the feature
>> without recompiling the kernel - during testing period for example.
>>
>> If you prefer this way, i would be happy to rework the patch via static keys.
>
> I would rather go and chase the underlying issue. So can we get an
> actual data please?
