Re: [PATCH] [RFC] arm64: mmu: use range based TLB flushing when hot unplugging memory

[Alistair Popple <apopple@nvidia.com>]

On 2026-05-21 at 21:24 +1000, Ryan Roberts <ryan.roberts@arm.com> wrote...
> On 21/05/2026 11:46, Balbir Singh wrote:
> > On Thu, May 21, 2026 at 09:50:04AM +0100, Ryan Roberts wrote:
> >> On 21/05/2026 05:24, Alistair Popple wrote:
> >>> Hot unplugging memory on ARM64 requires a TLB invalidate after unmapping
> >>> the page to be hot unplugged from the direct map. Currently that happens
> >>> one page at a time, meaning range based invalidates cannot be used. The
> >>> result of this is that removing large amounts of memory takes a long
> >>> time and in some cases can trigger an RCU stall warning.
> >>>
> >>> For example on one system hot unplugging 480GB of memory takes ~1
> >>> minute. With this change the same operation took ~1 second, a 60x
> >>> improvement.
> >>>
> >>> Signed-off-by: Alistair Popple <apopple@nvidia.com>
> >>>
> >>> ---
> >>>
> >>> This is an RFC, because I'm not sure the change is correct as it frees
> >>> the PTE page before flushing the TLB. I'm not familiar enough with ARM64
> >>> architecture to be sure this is safe, for example I don't know if HW
> >>> can update PTE bits such as access/dirty in the page through a stale
> >>> TLB entry.
> >>>
> >>> If so this would open a window during which the page is free but could
> >>> still be written to. Likely the safe option would be to collect all the
> >>> pages to be free on a list and free them after doing the range based TLB
> >>> flush, but wanted to get feedback on the approach before implementing it
> >>> which is the goal of this RFC.
> >>
> >> Hi Alistair,
> >>
> >> This patch doesn't apply on v7.1-rc4 because it conflicts with this patch:
> >>
> >> Commit 48478b9f79137 ("arm64/mm: Enable batched TLB flush in unmap_hotplug_range()")
> >>
> >> which has a very similar performance improvement, so hopefully it solves your
> >> problem?

Yes it does. This issue has been niggling me for a while, so how coincidental
that I finally try and do something about it just after someone has fixed it :-)

> >> There are two paths which use this logic; unmapping the linear map and unmapping
> >> the corresponding vmemmap. In the latter case, the memory is also freed, so we
> >> can't safely do the range optimizaiton there since the TLB needs to be flushed
> >> before freeing the memory. But the linear map is the big, slow bit so hopefully
> >> it's sufficent for you?
> >>
> > 
> > I assume vmemmap path is for tearing down the struct pages corresponding
> > to the physical memory and vmemmap teardowns taking a flush should be
> > OK. It is worth checking if the issue is already fixed.
> 
> Yes, exactly. Assuming you're using 64K pages, I think vmemmap is mapped using
> base pages. So you'd get 1024 struct pages on a page, so 1 TLBI per 64M of
> memory that is being unplugged on the vmemmap path.
> 
> On the linear map path, it's a single range tlbi operation to over the entire range.
> 
> Alistair's patch is doing it per 512M in both cases, but is unsafe for vmemmap
> as currently written.

Right - hence the RFC. This was somewhat sent in the spirit of it being wrong
in the hope that somebody could help me figure out how to do it properly. As
it turns out though it's already been fixed which is an even better outcome so
thanks for pointing that out.

So this patch can be ignored, I see a similar offlining performance improvement
on my system (~1m -> ~1s) with 48478b9f79137 applied.

Thanks.

 - Alistair

> > 
> > Balbir
> > <snip>
>