[RFC] munmap(64k) is much slower with patch set reduce synchronous TLB invalidation on ARM64 merged

[RFC] munmap(64k) is much slower with patch set reduce synchronous TLB invalidation on ARM64 merged

[Hanjun Guo]

Hi Will,

With patch set "Avoid synchronous TLB invalidation for intermediate
page-table entries on arm64" was merged, munmap() for hugepage such as
2M will be much faster, but with following case, munmap(64k) will take
much longer time than before:

#define MEM_LEN_TEST			(0x10000UL)
#define MEM_ADD_START			(0x8000fffef000UL)
#define MEM1_ADD_START			(0x800000090000UL)

mem = (unsigned char *)mmap((void *)MEM_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);

mem1 = (unsigned char *)mmap((void *)MEM1_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);

munmap(mem, MEM_LEN_TEST); // will take much longer time (from 50us to 230us on A55) than before

munmap(mem1, MEM_LEN_TEST);

munmap(mem, MEM_LEN_TEST) will finally call unmap_region() in the kernel,
and in free_pgtables() will adjust the start and end in mmu_gather, so when
running to __flush_tlb_range(), the (end - start) will more that 1G but less
than 2G in this case, on the other hand, (MAX_TLBI_OPS * stride) will be 2G
as the stride is 2M.

So in the end if ((end - start) > (MAX_TLBI_OPS * stride)) will not valid,
and will take loops to invalidate the TLB for many times which is time consuming.

Can we make the (MAX_TLBI_OPS * stride) as constant such as 4M (which is the
value before "Avoid synchronous TLB invalidation for intermediate page-table
entries on arm64" merged)? or a smaller value for MAX_TLBI_OPS?

Looking forward your insight on this.

Best regards
Hanjun


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [RFC] munmap(64k) is much slower with patch set reduce synchronous TLB invalidation on ARM64 merged

[Will Deacon]

On Sat, May 04, 2019 at 07:44:59PM +0800, Hanjun Guo wrote:
> With patch set "Avoid synchronous TLB invalidation for intermediate
> page-table entries on arm64" was merged, munmap() for hugepage such as
> 2M will be much faster, but with following case, munmap(64k) will take
> much longer time than before:

I didn't find anything grepping for that subject. Could you provide a commit
ID instead, please?

> #define MEM_LEN_TEST			(0x10000UL)
> #define MEM_ADD_START			(0x8000fffef000UL)
> #define MEM1_ADD_START			(0x800000090000UL)
> 
> mem = (unsigned char *)mmap((void *)MEM_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
> 
> mem1 = (unsigned char *)mmap((void *)MEM1_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
> 
> munmap(mem, MEM_LEN_TEST); // will take much longer time (from 50us to 230us on A55) than before
> 
> munmap(mem1, MEM_LEN_TEST);
> 
> munmap(mem, MEM_LEN_TEST) will finally call unmap_region() in the kernel,
> and in free_pgtables() will adjust the start and end in mmu_gather, so when
> running to __flush_tlb_range(), the (end - start) will more that 1G but less
> than 2G in this case, on the other hand, (MAX_TLBI_OPS * stride) will be 2G
> as the stride is 2M.
> 
> So in the end if ((end - start) > (MAX_TLBI_OPS * stride)) will not valid,
> and will take loops to invalidate the TLB for many times which is time consuming.
> 
> Can we make the (MAX_TLBI_OPS * stride) as constant such as 4M (which is the
> value before "Avoid synchronous TLB invalidation for intermediate page-table
> entries on arm64" merged)? or a smaller value for MAX_TLBI_OPS?

Changing MAX_TLBI_OPS will hurt the common case when tearing down PMD tables
with 4k pages, so I'm hesitant to change that again.

Will

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Re: [RFC] munmap(64k) is much slower with patch set reduce synchronous TLB invalidation on ARM64 merged

[Hanjun Guo]

On 2019/5/7 2:47, Will Deacon wrote:
> On Sat, May 04, 2019 at 07:44:59PM +0800, Hanjun Guo wrote:
>> With patch set "Avoid synchronous TLB invalidation for intermediate
>> page-table entries on arm64" was merged, munmap() for hugepage such as
>> 2M will be much faster, but with following case, munmap(64k) will take
>> much longer time than before:
> 
> I didn't find anything grepping for that subject. Could you provide a commit
> ID instead, please?

Sorry for not clear, this is a patch set:

https://lkml.org/lkml/2018/8/30/751

And for specific commit IDs,
f270ab8 arm64: tlb: Adjust stride and type of TLBI according to mmu_gather
67a902a arm64: tlbflush: Allow stride to be specified for __flush_tlb_range()

> 
>> #define MEM_LEN_TEST			(0x10000UL)
>> #define MEM_ADD_START			(0x8000fffef000UL)
>> #define MEM1_ADD_START			(0x800000090000UL)
>>
>> mem = (unsigned char *)mmap((void *)MEM_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
>>
>> mem1 = (unsigned char *)mmap((void *)MEM1_ADD_START, MEM_LEN_TEST, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
>>
>> munmap(mem, MEM_LEN_TEST); // will take much longer time (from 50us to 230us on A55) than before
>>
>> munmap(mem1, MEM_LEN_TEST);
>>
>> munmap(mem, MEM_LEN_TEST) will finally call unmap_region() in the kernel,
>> and in free_pgtables() will adjust the start and end in mmu_gather, so when
>> running to __flush_tlb_range(), the (end - start) will more that 1G but less
>> than 2G in this case, on the other hand, (MAX_TLBI_OPS * stride) will be 2G
>> as the stride is 2M.
>>
>> So in the end if ((end - start) > (MAX_TLBI_OPS * stride)) will not valid,
>> and will take loops to invalidate the TLB for many times which is time consuming.
>>
>> Can we make the (MAX_TLBI_OPS * stride) as constant such as 4M (which is the
>> value before "Avoid synchronous TLB invalidation for intermediate page-table
>> entries on arm64" merged)? or a smaller value for MAX_TLBI_OPS?
> 
> Changing MAX_TLBI_OPS will hurt the common case when tearing down PMD tables
> with 4k pages, so I'm hesitant to change that again.

It's true, but MAX_TLBI_OPS(1024) seems not an objective value fits all needs.
Is it reasonable to update the algorithm of adjusting the start and end in mmu_gather
in function free_pgtables()?

Thanks
Hanjun


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel