Re: [PATCH] x86/mm/tlb: Remove flush_tlb_info from the stack

[Nadav Amit <namit@vmware.com>]

> On Apr 23, 2019, at 10:23 AM, Andy Lutomirski <luto@kernel.org> wrote:
> 
> On Tue, Apr 23, 2019 at 9:56 AM Nadav Amit <namit@vmware.com> wrote:
>>> On Apr 23, 2019, at 9:50 AM, Andy Lutomirski <luto@kernel.org> wrote:
>>> 
>>> On Tue, Apr 23, 2019 at 12:12 AM Nadav Amit <namit@vmware.com> wrote:
>> https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.kernel.org%2Fpub%2Fscm%2Flinux%2Fkernel%2Fgit%2Fluto%2Flinux.git%2Flog%2F%3Fh%3Dx86%2Ffixes&amp;data=02%7C01%7Cnamit%40vmware.com%7Cb0ba13a059b54abcb1c008d6c8107a54%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C636916370442070349&amp;sdata=5uXX95CD%2FAI5ZSJON%2BjCAKQ88sAwZgme6Az0ZHpEGZo%3D&amp;reserved=0 >> Remove flush_tlb_info variables from the stack. This allows to align
>>>> flush_tlb_info to cache-line and avoid potentially unnecessary cache
>>>> line movements. It also allows to have a fixed virtual-to-physical
>>>> translation of the variables, which reduces TLB misses.
>>>> 
>>>> Use per-CPU struct for flush_tlb_mm_range() and
>>>> flush_tlb_kernel_range(). Add debug assertions to ensure there are
>>>> no nested TLB flushes that might overwrite the per-CPU data. For
>>>> arch_tlbbatch_flush(), use a const struct.
>>>> 
>>>> Results when running a microbenchmarks that performs 10^6 MADV_DONTEED
>>>> operations and touching a page, in which 3 additional threads run a
>>>> busy-wait loop (5 runs):
>>> 
>>> Can you add a memset(,,,. 0, sizeof(struct flush_tlb_info)) everywhere
>>> you grab it?  Or, even better, perhaps do something like:
>>> 
>>> static inline struct flush_tlb_info *get_flush_tlb_info(void)
>>> {
>>> /* check reentrancy, make sure that we use smp_processor_id() or
>>> otherwise assert that we're bound to a single CPU. */
>>> struct flush_tlb_info *ptr = this_cpu_ptr(...);
>>> memset(ptr, 0, sizeof(*ptr));
>>> return ptr;
>>> }
>>> 
>>> static inline void put_flush_tlb_info(void)
>>> {
>>> /* finish checking reentrancy. */
>>> }
>> 
>> I’ll check if the compiler is smart enough to avoid redundant assignments,
>> and if it is not, I’ll just give all the struct arguments to
>> get_flush_tlb_info() instead of memset() if you don’t mind.
> 
> Sounds good.
> 
>> I also want to give a try for parallelizing the remote and local
>> invocations, which really annoys me every time I look at the code.
> 
> Yes please!

I have written some patches and they do provide a considerable performance
improvement of (>10%) for remote TLB flushes. There are still some issues
that need to be resolved, specifically a small slowdown for local TLB
flushes (~15ns).

Anyhow, based on my past experience, I will do this change in a separate
patch-set after the flush_tlb_info off-stack patch makes it through.