lkml.org 
[lkml]   [2018]   [Apr]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH v2] mm: vmalloc: Clean up vunmap to avoid pgtable ops twice
On Mon, 16 Apr 2018 16:29:02 +0530 Chintan Pandya <cpandya@codeaurora.org> wrote:

> vunmap does page table clear operations twice in the
> case when DEBUG_PAGEALLOC_ENABLE_DEFAULT is enabled.
>
> So, clean up the code as that is unintended.
>
> As a perf gain, we save few us. Below ftrace data was
> obtained while doing 1 MB of vmalloc/vfree on ARM64
> based SoC *without* this patch applied. After this
> patch, we can save ~3 us (on 1 extra vunmap_page_range).
>
> CPU DURATION FUNCTION CALLS
> | | | | | | |
> 6) | __vunmap() {
> 6) | vmap_debug_free_range() {
> 6) 3.281 us | vunmap_page_range();
> 6) + 45.468 us | }
> 6) 2.760 us | vunmap_page_range();
> 6) ! 505.105 us | }

It's been a long time since I looked at the vmap code :(

> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -603,26 +603,6 @@ static void unmap_vmap_area(struct vmap_area *va)
> vunmap_page_range(va->va_start, va->va_end);
> }
>
> -static void vmap_debug_free_range(unsigned long start, unsigned long end)
> -{
> - /*
> - * Unmap page tables and force a TLB flush immediately if pagealloc
> - * debugging is enabled. This catches use after free bugs similarly to
> - * those in linear kernel virtual address space after a page has been
> - * freed.
> - *
> - * All the lazy freeing logic is still retained, in order to minimise
> - * intrusiveness of this debugging feature.
> - *
> - * This is going to be *slow* (linear kernel virtual address debugging
> - * doesn't do a broadcast TLB flush so it is a lot faster).
> - */
> - if (debug_pagealloc_enabled()) {
> - vunmap_page_range(start, end);
> - flush_tlb_kernel_range(start, end);
> - }
> -}
> -
> /*
> * lazy_max_pages is the maximum amount of virtual address space we gather up
> * before attempting to purge with a TLB flush.
> @@ -756,6 +736,9 @@ static void free_unmap_vmap_area(struct vmap_area *va)
> {
> flush_cache_vunmap(va->va_start, va->va_end);
> unmap_vmap_area(va);
> + if (debug_pagealloc_enabled())
> + flush_tlb_kernel_range(va->va_start, va->va_end);
> +
> free_vmap_area_noflush(va);
> }
>
> @@ -1142,7 +1125,6 @@ void vm_unmap_ram(const void *mem, unsigned int count)
> BUG_ON(!PAGE_ALIGNED(addr));
>
> debug_check_no_locks_freed(mem, size);
> - vmap_debug_free_range(addr, addr+size);

This appears to be a functional change: if (count <= VMAP_MAX_ALLOC)
and we're in debug mode then the
vunmap_page_range/flush_tlb_kernel_range will no longer be performed.
Why is this ok?

> if (likely(count <= VMAP_MAX_ALLOC)) {
> vb_free(mem, size);
> @@ -1499,7 +1481,6 @@ struct vm_struct *remove_vm_area(const void *addr)
> va->flags |= VM_LAZY_FREE;
> spin_unlock(&vmap_area_lock);
>
> - vmap_debug_free_range(va->va_start, va->va_end);
> kasan_free_shadow(vm);
> free_unmap_vmap_area(va);
>

\
 
 \ /
  Last update: 2018-05-01 00:52    [W:0.054 / U:0.748 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site