Date:   Fri, 4 Dec 2020 12:06:44 +0100
From:   Borislav Petkov <bp@...en8.de>
To:     Masami Hiramatsu <mhiramat@...nel.org>
Cc:     Tom Lendacky <thomas.lendacky@....com>, x86@...nel.org,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...nel.org>,
        Kees Cook <keescook@...omium.org>,
        "H . Peter Anvin" <hpa@...or.com>, Joerg Roedel <jroedel@...e.de>,
        "Gustavo A . R . Silva" <gustavoars@...nel.org>,
        Jann Horn <jannh@...gle.com>,
        Srikar Dronamraju <srikar@...ux.vnet.ibm.com>,
        Ricardo Neri <ricardo.neri-calderon@...ux.intel.com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 1/3] x86/uprobes: Fix not using prefixes.nbytes for
 loop over prefixes.bytes

On Fri, Dec 04, 2020 at 09:56:53AM +0900, Masami Hiramatsu wrote:
> Hmm, there is a difference between Intel SDM and AMD APM.
> 
> Intel SDM vol.2
> 
> 2.1.1 Instruction Prefixes
> Instruction prefixes are divided into four groups, each with a set of allowable prefix codes. For each instruction, it
> is only useful to include up to one prefix code from each of the four groups (Groups 1, 2, 3, 4).
> 
> AMD APM vol.3
> 
> 1.2.1 Summary of Legacy Prefixes
> Table 1-1 on page 7 shows the legacy prefixes. The legacy prefixes are organized into five groups, as
> shown in the left-most column of Table 1-1. An instruction encoding may include a maximum of one
> prefix from each of the five groups.
> 
> So, Intel CPU doesn't accept LOCK-REP because those are in a same prefix
> group, but AMD says it is acceptable.

That would be a huge problem for code if both vendors would behave
differently wrt prefixes.

> Actually, insn.c only accepts the prefix up to 4, so if there is any
> instruction which has 5 prefixes, it will fail to parse.

Well, actually it looks more like a difference in how both vendors group
things:

AMD has 5 groups and Intel 4 by putting LOCK and REP together.

The most important aspect, however, is that you can have as many
prefixes as you want and there's no hardware limitation on the number -
I'm being told - just that you can overflow the instruction limit of 15
and then get a #GP for invalid insn. See here:

https://sandpile.org/x86/opc_enc.htm

note #1

with examples how you can overflow the 15 bytes limit even with a valid
insn.

> Note that anyway the same prefix can be repeated, we can see a good example
> in K8_NOP*.

Yap.

> In this case, insn.c just store the 1 osp in the prefixes.bytes[], and
> just increment prefixes.nbytes for the repeated prefixes.
> 
> Anyway, if there is LOCK-REP prefix combination, I have to introduce new
> insn_field for legacy prefix.

Well, the legacy prefixes field needs to be of 4 fields because REP and
LOCK really are two separate but mutually exclusive groups. Why?

They're used by a disjoint set of instructions, see the AMD doc for both
REP and LOCK prefixes.

Which means, you can either have a REP (exclusive or) LOCK but not both.

Which means, as a stable@ fix I can use Tom's ARRAY_SIZE() suggestion
and then later on we can make the legacy prefixes a separate struct.
Maybe even a struct with a bitfield:

struct legacy_prefixes {
        /* operand-size override: 0x66 */
        u8 os_over: 1,
        /* address-size override: 0x67 */
           as_over: 1,
        /*
         * segment override: 0x2e(CS), 0x3e(DS), 0x26(ES), 0x64(FS), 0x65(GS),
         * 0x36(SS)
         */
           s_over: 1,
        /* lock prefix: 0xf0 */
           lock:   1,
        /* repeat prefixes: 0xf2: REPNx, 0xf3: REPx */
           rep:    1,
	   __resv: 3;
};

or so which you can set to denote when you've seen the respective
prefixes.

But that we can discuss later.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Powered by blists - more mailing lists