Re: darktable performance regression on AMD systems caused by "mm: align larger anonymous mappings on THP boundaries"

[Petr Tesarik <ptesarik@suse.com>]

On Thu, 24 Oct 2024 12:56:27 +0200
Vlastimil Babka <vbabka@suse.cz> wrote:

> On 10/24/24 12:49, Petr Tesarik wrote:
> > On Thu, 24 Oct 2024 12:23:48 +0200
> > Vlastimil Babka <vbabka@suse.cz> wrote:
> >   
> >> On 10/24/24 11:58, Vlastimil Babka wrote:  
> >> > On 10/24/24 09:45, Thorsten Leemhuis wrote:    
> >> >> Hi, Thorsten here, the Linux kernel's regression tracker.
> >> >> 
> >> >> Rik, I noticed a report about a regression in bugzilla.kernel.org that
> >> >> appears to be caused by the following change of yours:
> >> >> 
> >> >> efa7df3e3bb5da ("mm: align larger anonymous mappings on THP boundaries")
> >> >> [v6.7]
> >> >> 
> >> >> It might be one of those "some things got faster, a few things became
> >> >> slower" situations. Not sure. Felt odd that the reporter was able to
> >> >> reproduce it on two AMD systems, but not on a Intel system. Maybe there
> >> >> is a bug somewhere else that was exposed by this.    
> >> > 
> >> > It seems very similar to what we've seen with spec benchmarks such as cactus
> >> > and bisected to the same commit:
> >> > 
> >> > https://bugzilla.suse.com/show_bug.cgi?id=1229012
> >> > 
> >> > The exact regression varies per system. Intel regresses too but relatively
> >> > less. The theory is that there are many large-ish allocations that don't
> >> > have individual sizes aligned to 2MB and would have been merged, commit
> >> > efa7df3e3bb5da causes them to become separate areas where each aligns its
> >> > start at 2MB boundary and there are gaps between. This (gaps and vma
> >> > fragmentation) itself is not great, but most of the problem seemed to be
> >> > from the start alignment, which togethter with the access pattern causes
> >> > more TLB or cache missess due to limited associtativity.
> >> > 
> >> > So maybe darktable has a similar problem. A simple candidate fix could
> >> > change commit efa7df3e3bb5da so that the mapping size has to be a multiple
> >> > of THP size (2MB) in order to become aligned, right now it's enough if it's
> >> > THP sized or larger.    
> >> 
> >> Maybe this could be enough to fix the issue? (on 6.12-rc4)  
> > 
> > 
> > Yes, this should work. I was unsure if thp_get_unmapped_area_vmflags()
> > differs in other ways from mm_get_unmapped_area_vmflags(), which might
> > still be relevant. I mean, does mm_get_unmapped_area_vmflags() also
> > prefer to allocate THPs if the virtual memory block is large enough?  
> 
> Well any sufficiently large area spanning a PMD aligned/sized block (either
> a result of a single allocation or merging of several allocations) can
> become populated by THPs (at least in those aligned blocks), and the
> preference depends on system-wide THP settings and madvise(MADV_HUGEPAGE) or
> prctl.
> 
> But mm_get_unmapped_area_vmflags() will AFAIK not try to align the area to
> PMD size like the thp_ version would, even if the request is large enough.

Then it sounds like exactly what we want. But I prefer to move the
check down to __thp_get_unmapped_area() like this:

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 2fb328880b50..8d80f3af5525 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1082,6 +1082,9 @@ static unsigned long __thp_get_unmapped_area(struct file *filp,
 	if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())
 		return 0;
 
+	if (!IS_ALIGNED(len, size))
+		return 0;
+
 	if (off_end <= off_align || (off_end - off_align) < size)
 		return 0;