linux-kernel - Re: [RFC][PATCH 5/7] x86: convert arch_futex_atomic_op_inuser() to user_access_begin/user_access

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: <20200327024227.GT23230@ZenIV.linux.org.uk>
Date:   Fri, 27 Mar 2020 02:42:27 +0000
From:   Al Viro <viro@...iv.linux.org.uk>
To:     Linus Torvalds <torvalds@...ux-foundation.org>
Cc:     Thomas Gleixner <tglx@...utronix.de>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [RFC][PATCH 5/7] x86: convert arch_futex_atomic_op_inuser() to
 user_access_begin/user_access_end()

On Tue, Mar 24, 2020 at 01:57:19PM -0700, Linus Torvalds wrote:
> On Tue, Mar 24, 2020 at 1:45 PM Al Viro <viro@...iv.linux.org.uk> wrote:
> >
> > OK...  BTW, I'd been trying to recall the reasons for the way
> > __futex_atomic_op2() loop is done; ISTR some discussion along
> > the lines of cacheline ping-pong prevention, but I'd been unable
> > to reconstruct enough details to find it and I'm not sure it
> > hadn't been about some other code ;-/
> 
> No, that doesn't look like any cacheline advantage I can think of -
> quite the reverse.
> 
> I suspect it's just lazy code, with the reload being unnecessary. Or
> it might be written that way because you avoid an extra variable.
> 
> In fact, from a cacheline optimization standpoint, there are
> advantages to not doing the load even on the initial run: if you know
> a certain value is particularly likely, there are advantages to just
> _assuming_ that value, rather than loading it. So no initial load at
> all, and just initialize the first value to the likely case.
> 
> That can avoid an unnecessary "load for shared ownership" cacheline
> state transition (since the cmpxchg will want to turn it into an
> exclusive modified cacheline anyway).
> 
> But I don't think that optimization is likely the case here, and
> you're right, the loop would be better written with the initial load
> outside the loop.

OK, updated branch is in the same place; changes: __futex_atomic_op{1,2}
turned into unsafe_atomic_op{1,2}, with "goto on error" folded into those.
And pointless reload removed from cmpxchg loop in unsafe_atomic_op2().
Diffstat:
 arch/alpha/include/asm/futex.h      |  5 +-
 arch/arc/include/asm/futex.h        |  5 +-
 arch/arm/include/asm/futex.h        |  5 +-
 arch/arm64/include/asm/futex.h      |  5 +-
 arch/hexagon/include/asm/futex.h    |  5 +-
 arch/ia64/include/asm/futex.h       |  5 +-
 arch/microblaze/include/asm/futex.h |  5 +-
 arch/mips/include/asm/futex.h       |  5 +-
 arch/nds32/include/asm/futex.h      |  6 +--
 arch/openrisc/include/asm/futex.h   |  5 +-
 arch/parisc/include/asm/futex.h     |  2 -
 arch/powerpc/include/asm/futex.h    |  5 +-
 arch/riscv/include/asm/futex.h      |  5 +-
 arch/s390/include/asm/futex.h       |  2 -
 arch/sh/include/asm/futex.h         |  4 --
 arch/sparc/include/asm/futex_64.h   |  4 --
 arch/x86/include/asm/futex.h        | 97 ++++++++++++++++++++++++-------------
 arch/x86/include/asm/uaccess.h      | 93 -----------------------------------
 arch/xtensa/include/asm/futex.h     |  5 +-
 include/asm-generic/futex.h         |  2 -
 kernel/futex.c                      |  5 +-
 tools/objtool/check.c               |  1 +
 22 files changed, 93 insertions(+), 183 deletions(-)

Sorry about the fuckup when sending that patchset ;-/  It ended up cc'd to
x86 list instead of the futex one; Message-Id of the beginning of the
thread is <20200327022836.881203-1-viro@...IV.linux.org.uk>.