Re: [PATCH] iommu/iova: change IOVA_MAG_SIZE to 127 to save memory

[John Garry <john.garry@huawei.com>]

>>>
>>>>    [    4.319253] iommu: Adding device 0000:06:00.2 to group 5
>>>>    [    4.325869] iommu: Adding device 0000:20:01.0 to group 15
>>>>    [    4.332648] iommu: Adding device 0000:20:02.0 to group 16
>>>>    [    4.338946] swapper/0 invoked oom-killer: 
>>>> gfp_mask=0x6040c0(GFP_KERNEL|__GFP_COMP), nodemask=(null), order=0, 
>>>> oom_score_adj=0
>>>>    [    4.350251] swapper/0 cpuset=/ mems_allowed=0
>>>>    [    4.354618] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 
>>>> 4.19.57.mx64.282 #1
>>>>    [    4.355612] Hardware name: Dell Inc. PowerEdge R7425/08V001, 
>>>> BIOS 1.9.3 06/25/2019
>>>>    [    4.355612] Call Trace:
>>>>    [    4.355612]  dump_stack+0x46/0x5b
>>>>    [    4.355612]  dump_header+0x6b/0x289
>>>>    [    4.355612]  out_of_memory+0x470/0x4c0
>>>>    [    4.355612]  __alloc_pages_nodemask+0x970/0x1030
>>>>    [    4.355612]  cache_grow_begin+0x7d/0x520
>>>>    [    4.355612]  fallback_alloc+0x148/0x200
>>>>    [    4.355612]  kmem_cache_alloc_trace+0xac/0x1f0
>>>>    [    4.355612]  init_iova_domain+0x112/0x170

Note for Feng Tang: This callchain does not exist anymore since we 
separated out the rcache init from the IOVA domain init. Indeed, not so 
much memory is wasted for unused rcaches now.

My point really is that it would be nicer to see a modern callchain - 
but don't read that as me saying that the change is this patch is bad.

>>>>    [    4.355612]  amd_iommu_domain_alloc+0x138/0x1a0
>>>>    [    4.355612]  iommu_group_get_for_dev+0xc4/0x1a0
>>>>    [    4.355612]  amd_iommu_add_device+0x13a/0x610
>>>>    [    4.355612]  add_iommu_group+0x20/0x30
>>>>    [    4.355612]  bus_for_each_dev+0x76/0xc0
>>>>    [    4.355612]  bus_set_iommu+0xb6/0xf0
>>>>    [    4.355612]  amd_iommu_init_api+0x112/0x132
>>>>    [    4.355612]  state_next+0xfb1/0x1165
>>>>    [    4.355612]  amd_iommu_init+0x1f/0x67
>>>>    [    4.355612]  pci_iommu_init+0x16/0x3f
>>>>    ...
>>>>    [    4.670295] Unreclaimable slab info:
>>>>    ...
>>>>    [    4.857565] kmalloc-2048           59164KB      59164KB
>>>>
>>>> Change IOVA_MAG_SIZE from 128 to 127 to make size of 'iova_magazine'
>>>> 1024 bytes so that no memory will be wasted.
>>>>
>>>> [1]. https://lkml.org/lkml/2019/8/12/266
>>>>
>>>> Signed-off-by: Feng Tang <feng.tang@intel.com>
>>>> ---
>>>>   drivers/iommu/iova.c | 7 ++++++-
>>>>   1 file changed, 6 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
>>>> index db77aa675145b..27634ddd9b904 100644
>>>> --- a/drivers/iommu/iova.c
>>>> +++ b/drivers/iommu/iova.c
>>>> @@ -614,7 +614,12 @@ EXPORT_SYMBOL_GPL(reserve_iova);
>>>>    * dynamic size tuning described in the paper.
>>>>    */
>>>> -#define IOVA_MAG_SIZE 128
>>>> +/*
>>>> + * As kmalloc's buffer size is fixed to power of 2, 127 is chosen to
>>>> + * assure size of 'iova_magzine' to be 1024 bytes, so that no memory
>>>
>>> Typo: iova_magazine
>>>
>>>> + * will be wasted.
>>>> + */
>>>> +#define IOVA_MAG_SIZE 127
>>
>> I do wonder if we will see some strange new behaviour since 
>> IOVA_FQ_SIZE % IOVA_MAG_SIZE != 0 now...
> 
> I doubt it - even if a flush queue does happen to be entirely full of 
> equal-sized IOVAs, a CPU's loaded magazines also both being perfectly 
> empty when it comes to dump a full fq seem further unlikely, so in 
> practice I don't see this making any appreciable change to the 
> likelihood of spilling back to the depot or not. In fact the smaller the 
> magazines get, the less time would be spent flushing the depot back to 
> the rbtree, where your interesting workload falls off the cliff and 
> never catches back up with the fq timer, so at some point it might even 
> improve (unless it's also already close to the point where smaller 
> caches would bottleneck allocation)... might be interesting to 
> experiment with a wider range of magazine sizes if you had the time and 
> inclination.
> 

ok, what you are saying sounds reasonable. I just remember that when we 
analyzed the longterm aging issue that we concluded that the FQ size and 
its relation to the magazine size was a factor and this change makes me 
a little worried about new issues. Better the devil you know and all that...

Anyway, if I get some time I might do some testing to see if this change 
has any influence.

Another thought is if we need even store the size in the iova_magazine? 
mags in the depot are always full. As such, we only need worry about 
mags loaded in the cpu rcache and their sizes, so maybe we could have 
something like this:

struct iova_magazine {
-       unsigned long size;
        unsigned long pfns[IOVA_MAG_SIZE];
};

@@ -631,6 +630,8 @@ struct iova_cpu_rcache {
        spinlock_t lock;
        struct iova_magazine *loaded;
        struct iova_magazine *prev;
+       int loaded_size;
+       int prev_size;
};

I haven't tried to implement it though..

Thanks,
John