linux-kernel - Re: [GIT PULL] x86/mm for 6.2

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] [day] [month] [year] [list]

Date:   Sat, 17 Dec 2022 19:43:38 +0300
From:   "Kirill A. Shutemov" <kirill@...temov.name>
To:     Linus Torvalds <torvalds@...ux-foundation.org>
Cc:     kirill.shutemov@...ux.intel.com,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        linux-kernel@...r.kernel.org, x86@...nel.org, jejb@...ux.ibm.com,
        martin.petersen@...cle.com
Subject: Re: [GIT PULL] x86/mm for 6.2

On Fri, Dec 16, 2022 at 09:43:11AM -0600, Linus Torvalds wrote:
> On Fri, Dec 16, 2022 at 9:05 AM Kirill A. Shutemov <kirill@...temov.name> wrote:
> >
> > Below is preliminary fixup that suppose to address the issue. It does not
> > include change to untagged_addr() interface to avoid the clutter.
> 
> Looks like the right direction.
> 
> And once you change untagged_addr() to take 'tsk', you should then be
> able to cache all the information in the thread struct, and avoid the
> 'tsk->mm' dereference entirely.

Making untagged_addr() take task as an argument does not work well. We
don't have relevant task on hands in some places. The most notably GUP
only has 'mm'.

It is also not clear what untagging rules has to be applied: whoever runs
untagged_addr() or the target task/mm.

Up until now I choose the target mm rules. It makes things easier for
GDB, but it is not strong enough justification. Maybe always use rules of
'current' is the right way?

In other thread you suggested to make untagging in untagged_addr()
unconditional and I see how it solves the problem, but I don't think it is
good idea.

The current LAM patchset only enables LAM_U57 mode that doesn't compete
for real virtual address bits, but hardware also support LAM_U48 which
provides more tagging bits, but limits available address space for the
process. Unconditionally untag according to LAM_U48 rules is obviously
broken.

Although, x86 maintainers rejected LAM_U48 upstreaming as of now, I want
to keep door open in case a strong use case for it comes. I have code that
enables the mode stashed, just in case.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov