On 12/27/2018 07:46 PM, Michal Hocko wrote:
> On Thu 27-12-18 15:18:54, Konstantin Khorenko wrote:
>> 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.
>
> This is not true in general, though.
Right now - yep, this is not true. Sorry, i was not clear enough here.
i meant after we catch all high order allocations, we review them and either
switch to kvmalloc() or mark with NOWARN if we are ready to handle allocation errors
in that particular place. So this is an ideal situation when we reviewed all the things.
> [...]
>> 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.
>
> There will always be legitimate high order allocations.
Sure. But after we review them we either switch them to kvmalloc() or mark them with
NOWARN. In both cases we won't get new warnings about that places.
> I believe that
> for your particular use case it is much better to simply enable reclaim
> and page allocator tracepoints which will give you not only the source
> of the allocation but also a much better picture
Tracepoints are much better for issues investigation, right. And we do so.
And warnings are intended not for investigation but for prevention of possible future issues.
If we spot a big allocation, we can review it in advance, before we face any problem,
and in most cases just switch it to use kvmalloc() in 90% cases and we never ever have
a problem with unexpected reclaim due to this allocation. Ever.
With any reclaim algorithm - the compaction just won't be triggered.
>> i think people who run systems with "kernel.panic_on_warn" enabled do care
>> about reporting issues.
>
> You surely do not want to put the system down just because of the high
> order allocation though, right?
Right, i do not. (And i also don't want to run a node with "kernel.panic_on_warn"
enabled in production :) )
But people who do run nodes with "kernel.panic_on_warn" enabled in production
may disable high allocation warning by increasing warning order level higher than
MAX_ORDER. Or just not enable kernel config option.
i do understand the warning will be noisy at the beginning thus i surely don't even
suggest to make it enable by default now.
>>>> 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.
>
> But the warning alone will not give us useful information I am afraid.
> It will only give us, there are warnings but not whether those are
> actually a problem or not.
Yes. And even more - a lot of high order allocations which cannot be
switched to kvmalloc() are in drivers - for DMA zones - so they are very
rare and most probably won't ever cause a problem.
But some of them can potentially cause a problem some day. And warning
does not provide info how to distinguish "bad" and "good" ones.
But if we switch both "bad" and "good" big allocations to kvmalloc(),
that won't hurt, right? But that way we ensure we won't get any problems
from "bad" cases, even if we don't know exactly which of them are potentially "bad".
> So I really believe that using existing
> tracepoints or add some that will fill missing gaps will be much more
> better long term. And we do not have to add another config and touch the
> code as a bonus.
And Michal, thank you very much once again for this conversation.
i appreciate it.
--
Best regards,
Konstantin Khorenko,
Virtuozzo Linux Kernel Team
