On 2023/6/13 10:46, Zhang Zekun wrote:
hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I7ASVH CVE: NA
The logic of cpu rcaches could cause severe performance problem. It is because that the logic of iova rcache would create percpu rcache which will be used only by local cpu, free iovas in cpu_rcache won't be shared among different cpus. For example, platform which have 256 cpus with 256 threads will have 417792 entiries( 6 * (256 * 2 + 32) * 128) to cache iova, but every thread running on a cpu can only use 1/256 of total cache size, and this could cause serve performance problem. We limit the max cpu size to 64 to fix this problem, and each iovad could save about 2MB memory.
In our test FIO with 256 threads:
Jobs: 12 (f=2): [/(1),R(2),/(9)][16.7%][r=2KiB/s][r=0 IOPS][eta 00m:30s] Jobs: 12 (f=12): [R(12)][20.0%][r=1091MiB/s][r=279k IOPS][eta 00m:28s] Jobs: 12 (f=12): [R(12)][22.2%][r=1426MiB/s][r=365k IOPS][eta 00m:28s] Jobs: 12 (f=12): [R(12)][25.0%][r=1607MiB/s][r=411k IOPS][eta 00m:27s] Jobs: 12 (f=12): [R(12)][27.8%][r=1501MiB/s][r=384k IOPS][eta 00m:26s] Jobs: 12 (f=12): [R(12)][30.6%][r=1486MiB/s][r=380k IOPS][eta 00m:25s] Jobs: 12 (f=12): [R(12)][33.3%][r=1393MiB/s][r=357k IOPS][eta 00m:24s] Jobs: 12 (f=12): [R(12)][36.1%][r=1550MiB/s][r=397k IOPS][eta 00m:23s] Jobs: 12 (f=12): [R(12)][38.9%][r=1485MiB/s][r=380k IOPS][eta 00m:22s]
After this patch:
bs: 10 (f=10): [R(10)][98.4%][r=7414MiB/s][r=1898k IOPS][eta 00m:15s] Jobs: 10 (f=10): [R(10)][98.5%][r=7495MiB/s][r=1919k IOPS][eta 00m:14s] Jobs: 10 (f=10): [R(10)][98.6%][r=7497MiB/s][r=1919k IOPS][eta 00m:13s] Jobs: 10 (f=10): [R(10)][98.7%][r=7497MiB/s][r=1919k IOPS][eta 00m:12s] Jobs: 10 (f=10): [R(10)][98.8%][r=7471MiB/s][r=1913k IOPS][eta 00m:11s] Jobs: 10 (f=10): [R(10)][98.9%][r=7483MiB/s][r=1916k IOPS][eta 00m:10s] Jobs: 10 (f=10): [R(10)][99.0%][r=7491MiB/s][r=1918k IOPS][eta 00m:09s] Jobs: 10 (f=10): [R(10)][99.1%][r=7436MiB/s][r=1904k IOPS][eta 00m:08s] Jobs: 10 (f=10): [R(10)][99.2%][r=7462MiB/s][r=1910k IOPS][eta 00m:07s]
can increase about 500% of IOPS.
Signed-off-by: Zhang Zekun zhangzekun11@huawei.com Reviewed-by: Weilong Chen chenweilong@huawei.com
v2: -add hulk inclusion
drivers/iommu/Kconfig | 10 ++++++++++ drivers/iommu/iova.c | 33 +++++++++++++++++++++++++++++++++ include/linux/iova.h | 5 +++++ 3 files changed, 48 insertions(+)
diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig index f04a2bde0018..c71eb20bc702 100644 --- a/drivers/iommu/Kconfig +++ b/drivers/iommu/Kconfig @@ -343,6 +343,16 @@ config ARM_SMMU_V3_PM help Add support for suspend and resume support for arm smmu v3.
+config ARM_SMMU_V3_REDUCE_CPURCACHE
- bool "Add support for cpu rcache"
- depends on ARM_SMMU_V3
- default n
- help
Add support for reducing the total amount of cpu rcache. Whennum of cpus is large, iova rcache will have serve performanceproblem. Reduce the amount of cpu rcaches to improve cache hitrate.- config S390_IOMMU def_bool y if S390 && PCI depends on S390 && PCI
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c index 1246e8f8bf08..ed86bb09db4f 100644 --- a/drivers/iommu/iova.c +++ b/drivers/iommu/iova.c @@ -881,11 +881,19 @@ static void init_iova_rcaches(struct iova_domain *iovad) rcache = &iovad->rcaches[i]; spin_lock_init(&rcache->lock); rcache->depot_size = 0; +#ifdef CONFIG_ARM_SMMU_V3_REDUCE_CPURCACHE
for (cpu = 0; cpu < MAX_CPU_SIZE; cpu++) {rcache->cpu_rcaches[cpu] = kmalloc(sizeof(*cpu_rcache), GFP_KERNEL);if (WARN_ON(!rcache->cpu_rcaches[cpu]))continue;cpu_rcache = rcache->cpu_rcaches[cpu];+#else rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size()); if (WARN_ON(!rcache->cpu_rcaches)) continue; for_each_possible_cpu(cpu) { cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu); +#endif
Too hack for the IOMMU core code,
NACK.
Thanks Hanjun