Re: [RESEND PATCH 2/2] iommu/iova: allocate iova_rcache->depot dynamicly

From: John Garry
Date: Tue Aug 15 2023 - 11:17:56 EST


On 12/08/2023 08:18, zhangzekun (A) wrote:


在 2023/8/11 22:14, John Garry 写道:
On 11/08/2023 14:02, Zhang Zekun wrote:
In fio test with 4k,read,and allowed cpus to 0-255, we observe a performance
decrease of IOPS.
The normal IOPS

What do you mean by normal IOPS? Describe this "normal" scenario.

Hi, John

The reason why I think 1980K is normal is that I have test the iova_rache
hit rate with all iova size, the average iova cache hit rate can reach up to

Sorry to say, but I still don't know what you mean by the terms "normal" and "abnormal" here. Is "normal" prior to the drop in IOPS which you mention, above? If so, at what time do this occur?

around 99% (varys from diffent work loads and iova size), and I think
iova_rcache behave well in our test work loads. Besides, the IOPS is
behaving as expect which is acked by our test group.

? can reach up to 1980k, but we can only
get about 1600k.

abormal IOPS:
Jobs: 12 (f=12): [R(12)][99.3%][r=6220MiB/s][r=1592k IOPS][eta 00m:12s]
Jobs: 12 (f=12): [R(12)][99.4%][r=6215MiB/s][r=1591k IOPS][eta 00m:11s]
Jobs: 12 (f=12): [R(12)][99.4%][r=6335MiB/s][r=1622k IOPS][eta 00m:10s]
Jobs: 12 (f=12): [R(12)][99.5%][r=6194MiB/s][r=1586k IOPS][eta 00m:09s]
Jobs: 12 (f=12): [R(12)][99.6%][r=6173MiB/s][r=1580k IOPS][eta 00m:08s]
Jobs: 12 (f=12): [R(12)][99.6%][r=5984MiB/s][r=1532k IOPS][eta 00m:07s]
Jobs: 12 (f=12): [R(12)][99.7%][r=6374MiB/s][r=1632k IOPS][eta 00m:06s]
Jobs: 12 (f=12): [R(12)][99.7%][r=6343MiB/s][r=1624k IOPS][eta 00m:05s]

normal IOPS:
Jobs: 12 (f=12): [R(12)][99.3%][r=7736MiB/s][r=1980k IOPS][eta 00m:12s]
Jobs: 12 (f=12): [R(12)][99.4%][r=7744MiB/s][r=1982k IOPS][eta 00m:11s]
Jobs: 12 (f=12): [R(12)][99.4%][r=7737MiB/s][r=1981k IOPS][eta 00m:10s]
Jobs: 12 (f=12): [R(12)][99.5%][r=7735MiB/s][r=1980k IOPS][eta 00m:09s]
Jobs: 12 (f=12): [R(12)][99.6%][r=7741MiB/s][r=1982k IOPS][eta 00m:08s]
Jobs: 12 (f=12): [R(12)][99.6%][r=7740MiB/s][r=1982k IOPS][eta 00m:07s]
Jobs: 12 (f=12): [R(12)][99.7%][r=7736MiB/s][r=1981k IOPS][eta 00m:06s]
Jobs: 12 (f=12): [R(12)][99.7%][r=7736MiB/s][r=1980k IOPS][eta 00m:05s]

The current struct of iova_rcache will have iova_cpu_rcache for every
cpu, and these iova_cpu_rcaches use a common buffer iova_rcache->depot to
exchange iovas among iova_cpu_rcaches. A machine with 256 cpus will have
256 iova_cpu_rcaches and 1 iova_rcache->depot per iova_domain. However,
the max size of iova_rcache->depot is fixed to MAX_GLOBAL_MAGS which equals
to 32, and can't grow with the number of cpus, and this can cause problem
in some condition.

Some drivers will only free iovas in their irq call back function. For
the driver in this case it has 16 thread irqs to free iova, but these
irq call back function will only free iovas on 16 certain cpus(cpu{0,16,
32...,240}). Thread irq which smp affinity is 0-15, will only free iova on
cpu 0. However, the driver will alloc iova on all cpus(cpu{0-255}), cpus
without free iova in local cpu_rcache need to get free iovas from
iova_rcache->depot. The current size of iova_rcache->depot max size is 32,
and it seems to be too small for 256 users (16 cpus will put iovas to
iova_rcache->depot and 240 cpus will try to get iova from it). Set
iova_rcache->depot grow with the num of possible cpus, and the decrease
of IOPS disappear.

Doesn't it take a long time for all the depots to fill for you? From the description, this sounds like the hisilicon SAS controller which you are experimenting with and I found there that it took a long time for the depots to fill for high throughput testing.

Thanks,
John

Yes, it will take more time to fill rcahe->depot, but we don't need to fill
all depots before using iova in them. We can use iova in rcache->depot
when local cpu_rcache is empty and there are iova magazines in
rcache->depot, which means "rcache->depot_size > 0". The larger depot
size will help caching more iova magazines freed in __iova_rcache_insert()
which means more potential memory cost for iova_rcache, and should not
introduce performance issues.

Thanks,
John