linux-kernel - Re: [PATCH] powerpc: optimise WARN

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]

Message-ID: <20190819075748.GY31406@gate.crashing.org>
Date:   Mon, 19 Aug 2019 02:57:48 -0500
From:   Segher Boessenkool <segher@...nel.crashing.org>
To:     Christophe Leroy <christophe.leroy@....fr>
Cc:     Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Paul Mackerras <paulus@...ba.org>,
        Michael Ellerman <mpe@...erman.id.au>,
        linux-kernel@...r.kernel.org, linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH] powerpc: optimise WARN_ON()

On Mon, Aug 19, 2019 at 07:40:42AM +0200, Christophe Leroy wrote:
> Le 18/08/2019 à 14:01, Segher Boessenkool a écrit :
> >On Sat, Aug 17, 2019 at 09:04:42AM +0000, Christophe Leroy wrote:
> >>Unlike BUG_ON(x), WARN_ON(x) uses !!(x) as the trigger
> >>of the t(d/w)nei instruction instead of using directly the
> >>value of x.
> >>
> >>This leads to GCC adding unnecessary pair of addic/subfe.
> >
> >And it has to, it is passed as an "r" to an asm, GCC has to put the "!!"
> >value into a register.
> >
> >>By using (x) instead of !!(x) like BUG_ON() does, the additional
> >>instructions go away:
> >
> >But is it correct?  What happens if you pass an int to WARN_ON, on a
> >64-bit kernel?
> 
> On a 64-bit kernel, an int is still in a 64-bit register, so there would 
> be no problem with tdnei, would it ? an int 0 is the same as an long 0, 
> right ?

The top 32 bits of a 64-bit register holding an int are undefined.  Take
as example

  int x = 42;
  x = ~x;

which may put ffff_ffff_ffff_ffd5 into the reg, not 0000_0000_ffff_ffd5
as you might expect or want.  For tw instructions this makes no difference
(they only look at the low 32 bits anyway); for td insns, it does.

> It is on 32-bit kernel that I see a problem, if one passes a long long 
> to WARN_ON(), the forced cast to long will just drop the upper size of 
> it. So as of today, BUG_ON() is buggy for that.

Sure, it isn't defined what types you can pass to that macro.  Another
thing that makes inline functions much saner to use.

> >(You might want to have 64-bit generate either tw or td.  But, with
> >your __builtin_trap patch, all that will be automatic).
> 
> Yes I'll discard this patch and focus on the __builtin_trap() one which 
> should solve most issues.

But see my comment there about the compiler knowing all code after an
unconditional trap is dead.


Segher