linux-kernel - Re: [PATCH -next 1/1] mm: hugetlb_vmemmap: Fix WARN_ON in vmemmap_remap

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <Y1v64lILNN2woIcu@arm.com>
Date:   Fri, 28 Oct 2022 16:53:06 +0100
From:   Catalin Marinas <catalin.marinas@....com>
To:     Muchun Song <muchun.song@...ux.dev>
Cc:     Anshuman Khandual <anshuman.khandual@....com>,
        Wupeng Ma <mawupeng1@...wei.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Mike Kravetz <mike.kravetz@...cle.com>,
        Muchun Song <songmuchun@...edance.com>,
        Michal Hocko <mhocko@...e.com>,
        Oscar Salvador <osalvador@...e.de>,
        Linux Memory Management List <linux-mm@...ck.org>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH -next 1/1] mm: hugetlb_vmemmap: Fix WARN_ON in
 vmemmap_remap_pte

On Fri, Oct 28, 2022 at 10:45:09AM +0800, Muchun Song wrote:
> On Oct 27, 2022, at 18:50, Catalin Marinas <catalin.marinas@....com> wrote:
> > On Wed, Oct 26, 2022 at 02:06:00PM +0530, Anshuman Khandual wrote:
> >> On 10/26/22 12:31, Muchun Song wrote:
> >>>> On 10/25/22 12:06, Muchun Song wrote:
> >>>>>> On Oct 25, 2022, at 09:42, Wupeng Ma <mawupeng1@...wei.com> wrote:
> >>>>>> From: Ma Wupeng <mawupeng1@...wei.com>
> >>>>>> 
> >>>>>> Commit f41f2ed43ca5 ("mm: hugetlb: free the vmemmap pages associated with
> >>>>>> each HugeTLB page") add vmemmap_remap_pte to remap the tail pages as
> >>>>>> read-only to catch illegal write operation to the tail page.
> >>>>>> 
> >>>>>> However this will lead to WARN_ON in arm64 in __check_racy_pte_update()
> >>>>> 
> >>>>> Thanks for your finding this issue.
> >>>>> 
> >>>>>> since this may lead to dirty state cleaned. This check is introduced by
> >>>>>> commit 2f4b829c625e ("arm64: Add support for hardware updates of the
> >>>>>> access and dirty pte bits") and the initial check is as follow:
> >>>>>> 
> >>>>>> BUG_ON(pte_write(*ptep) && !pte_dirty(pte));
> >>>>>> 
> >>>>>> Since we do need to mark this pte as read-only to catch illegal write
> >>>>>> operation to the tail pages, use set_pte  to replace set_pte_at to bypass
> >>>>>> this check.
> >>>>> 
> >>>>> In theory, the waring does not affect anything since the tail vmemmap
> >>>>> pages are supposed to be read-only. So, skipping this check for vmemmap
> >>>> 
> >>>> Tails vmemmap pages are supposed to be read-only, in practice but their
> >>>> backing pages do have pte_write() enabled. Otherwise the VM_WARN_ONCE()
> >>>> warning would not have triggered.
> >>> 
> >>> Right.
> >>> 
> >>>> 
> >>>>       VM_WARN_ONCE(pte_write(old_pte) && !pte_dirty(pte),
> >>>>                    "%s: racy dirty state clearing: 0x%016llx -> 0x%016llx",
> >>>>                    __func__, pte_val(old_pte), pte_val(pte));
> >>>> 
> >>>> Also, is not it true that the pte being remapped into a different page
> >>>> as read only, than what it had originally (which will be freed up) i.e 
> >>>> the PFN in 'old_pte' and 'pte' will be different. Hence is there still
> >>> 
> >>> Right.
> >>> 
> >>>> a possibility for a race condition even when the PFN changes ?
> >>> 
> >>> Sorry, I didn't get this question. Did you mean the PTE is changed from
> >>> new (pte) to the old one (old_pte) by the hardware because of the update
> >>> of dirty bit when a concurrent write operation to the tail vmemmap page?
> >> 
> >> No, but is not vmemmap_remap_pte() reuses walk->reuse_page for all remaining
> >> tails pages ? Is not there a PFN change, along with access permission change
> >> involved in this remapping process ?
> > 
> > For the record, as we discussed offline, changing the output address
> > (pfn) of a pte is not safe without break-before-make if at least one of
> > the mappings was writeable. The caller (vmemmap_remap_pte()) would need
> > to be fixed to first invalidate the pte and then write the new pte. I
> 
> Could you expose more details about what issue it will be caused? I am
> not familiar with arm64.

Well, it's not allowed by the architecture, so some CPU implementations
may do weird things like accessing incorrect memory or triggering TLB
conflict aborts if, for some reason, they end up with two entries in
the TLB for the same VA but pointing to different pfns. The hardware
expects an invalid PTE and TLB invalidation between such changes. In
practice most likely nothing happens and this works fine but we need to
stick to the architecture requirements in case some CPUs take advantage
of this requirement.

> > assume no other CPU accesses this part of the vmemmap while the pte is
> > being remapped.
> 
> However, there is no guarantee that no other CPU accesses this pte.
> E.g. memory failure or memory compaction, both can obtain head page
> from any tail struct pages (only read) anytime.

Oh, so we cannot safely go through a break-before-make sequence here
(zero the PTE, flush the TLB, write the new PTE) as some CPU may access
this pte.

-- 
Catalin