Date:	Thu, 4 Sep 2014 16:24:54 +0200
From:	Oleg Nesterov <oleg@...hat.com>
To:	Denys Vlasenko <dvlasenk@...hat.com>
Cc:	linux-kernel@...r.kernel.org, Jim Keniston <jkenisto@...ibm.com>,
	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>,
	Srikar Dronamraju <srikar@...ux.vnet.ibm.com>,
	Ingo Molnar <mingo@...nel.org>
Subject: Re: [PATCH 1/3] uprobes: add comment with insn opcodes, mnemonics
	and why we dont support them

Confused...

Does this replace the previous series (which at first glance looks
pretty much the same) ?

I already applied the previous one in my local tree, but didn't ask
to pull yet.

On 09/04, Denys Vlasenko wrote:
>
> After adding these, it's clear we have some awkward choices there.
> Some valid instructions are prohibited from uprobing while
> several invalid ones are allowed.
> 
> Hopefully future edits to the good-opcode tables will fix wrong bits
> or explain why those bits are not wrong.
> 
> No actual code changes.
> 
> Signed-off-by: Denys Vlasenko <dvlasenk@...hat.com>
> CC: Jim Keniston <jkenisto@...ibm.com>
> CC: Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>
> CC: Srikar Dronamraju <srikar@...ux.vnet.ibm.com>
> CC: Ingo Molnar <mingo@...nel.org>
> CC: Oleg Nesterov <oleg@...hat.com>
> ---
>  arch/x86/kernel/uprobes.c | 153 ++++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 134 insertions(+), 19 deletions(-)
> 
> diff --git a/arch/x86/kernel/uprobes.c b/arch/x86/kernel/uprobes.c
> index 5d1cbfe..9ba2fb2 100644
> --- a/arch/x86/kernel/uprobes.c
> +++ b/arch/x86/kernel/uprobes.c
> @@ -66,6 +66,49 @@
>   * Good-instruction tables for 32-bit apps.  This is non-const and volatile
>   * to keep gcc from statically optimizing it out, as variable_test_bit makes
>   * some versions of gcc to think only *(unsigned long*) is used.
> + *
> + * Prefixes. Most marked as "bad", but it doesn't matter, since insn decoder
> + * won't report *prefixes* as OPCODE1(insn).
> + * 0f - 2-byte opcode prefix
> + * 26,2e,36,3e - es:/cs:/ss:/ds:
> + * 64 - fs: (marked as "good", why?)
> + * 65 - gs: (marked as "good", why?)
> + * 66 - operand-size prefix
> + * 67 - address-size prefix
> + * f0 - lock prefix
> + * f2 - repnz    (marked as "good", why?)
> + * f3 - rep/repz (marked as "good", why?)
> + *
> + * Opcodes we'll probably never support:
> + * 6c-6f - ins,outs. SEGVs if used in userspace
> + * e4-e7 - in,out imm. SEGVs if used in userspace
> + * ec-ef - in,out acc. SEGVs if used in userspace
> + * cc - int3. SIGTRAP if used in userspace
> + * ce - into. Not used in userspace - no kernel support to make it useful. SEGVs
> + *	(why we support bound (62) then? it's similar, and similarly unused...)
> + * f1 - int1. SIGTRAP if used in userspace
> + * f4 - hlt. SEGVs if used in userspace
> + * fa - cli. SEGVs if used in userspace
> + * fb - sti. SEGVs if used in userspace
> + *
> + * Opcodes which need some work to be supported:
> + * 07,17,1f - pop es/ss/ds
> + *	Normally not used in userspace, but would execute if used.
> + *	Can cause GP or stack exception if tries to load wrong segment descriptor.
> + *	We hesitate to run them under single step since kernel's handling
> + *	of userspace single-stepping (TF flag) is fragile.
> + *	We can easily refuse to support push es/cs/ss/ds (06/0e/16/1e)
> + *	on the same grounds that they are never used.
> + * cd - int N.
> + *	Used by userspace for "int 80" syscall entry. (Other "int N"
> + *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
> + *	Not supported since kernel's handling of userspace single-stepping
> + *	(TF flag) is fragile.
> + * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
> + *
> + * Opcodes which can be enabled right away:
> + * 63 - arpl. This insn has no unusual exceptions (it's basically an arith op).
> + * d6 - salc. Undocumented "sign-extend carry flag to AL" insn
>   */
>  #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
>  static volatile u32 good_insns_32[256 / 32] = {
> @@ -94,7 +137,55 @@ static volatile u32 good_insns_32[256 / 32] = {
>  #define good_insns_32	NULL
>  #endif
>  
> -/* Good-instruction tables for 64-bit apps */
> +/* Good-instruction tables for 64-bit apps.
> + *
> + * Prefixes. Most marked as "bad", but it doesn't matter, since insn decoder
> + * won't report *prefixes* as OPCODE1(insn).
> + * 0f - 2-byte opcode prefix
> + * 26,2e,36,3e - es:/cs:/ss:/ds:
> + * 40-4f - rex prefixes
> + * 64 - fs: (marked as "good", why?)
> + * 65 - gs: (marked as "good", why?)
> + * 66 - operand-size prefix
> + * 67 - address-size prefix
> + * f0 - lock prefix
> + * f2 - repnz    (marked as "good", why?)
> + * f3 - rep/repz (marked as "good", why?)
> + *
> + * Genuinely invalid opcodes:
> + * 06,07 - formerly push/pop es
> + * 0e - formerly push cs
> + * 16,17 - formerly push/pop ss
> + * 1e,1f - formerly push/pop ds
> + * 27,2f,37,3f - formerly daa/das/aaa/aas
> + * 60,61 - formerly pusha/popa
> + * 62 - formerly bound. EVEX prefix for AVX512
> + * 82 - formerly redundant encoding of Group1
> + * 9a - formerly call seg:ofs (marked as "supported"???)
> + * c4,c5 - formerly les/lds. VEX prefixes for AVX
> + * ce - formerly into
> + * d4,d5 - formerly aam/aad
> + * d6 - formerly undocumented salc
> + * ea - formerly jmp seg:ofs (marked as "supported"???)
> + *
> + * Opcodes we'll probably never support:
> + * 6c-6f - ins,outs. SEGVs if used in userspace
> + * e4-e7 - in,out imm. SEGVs if used in userspace
> + * ec-ef - in,out acc. SEGVs if used in userspace
> + * cc - int3. SIGTRAP if used in userspace
> + * f1 - int1. SIGTRAP if used in userspace
> + * f4 - hlt. SEGVs if used in userspace
> + * fa - cli. SEGVs if used in userspace
> + * fb - sti. SEGVs if used in userspace
> + *
> + * Opcodes which need some work to be supported:
> + * cd - int N.
> + *	Used by userspace for "int 80" syscall entry. (Other "int N"
> + *	cause GP -> SEGV since their IDT gates don't allow calls from CPL 3).
> + *	Not supported since kernel's handling of userspace single-stepping
> + *	(TF flag) is fragile.
> + * cf - iret. Normally not used in userspace. Doesn't SEGV unless arguments are bad
> + */
>  #if defined(CONFIG_X86_64)
>  static volatile u32 good_insns_64[256 / 32] = {
>  	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
> @@ -122,7 +213,48 @@ static volatile u32 good_insns_64[256 / 32] = {
>  #define good_insns_64	NULL
>  #endif
>  
> -/* Using this for both 64-bit and 32-bit apps */
> +/* Using this for both 64-bit and 32-bit apps.
> + * Opcodes we don't support:
> + * 0f 00 - SLDT/STR/LLDT/LTR/VERR/VERW/-/- group. System insns
> + * 0f 01 - SGDT/SIDT/LGDT/LIDT/SMSW/-/LMSW/INVLPG group.
> + *	Also encodes tons of other system insns if mod=11.
> + *	Some are in fact non-system: xend, xtest, rdtscp, maybe more
> + * 0f 02 - lar (why? should be safe, it throws no exceptipons)
> + * 0f 03 - lsl (why? should be safe, it throws no exceptipons)
> + * 0f 04 - undefined
> + * 0f 05 - syscall
> + * 0f 06 - clts (CPL0 insn)
> + * 0f 07 - sysret
> + * 0f 08 - invd (CPL0 insn)
> + * 0f 09 - wbinvd (CPL0 insn)
> + * 0f 0a - undefined
> + * 0f 0b - ud2
> + * 0f 0c - undefined
> + * 0f 0d - prefetchFOO (amd prefetch insns)
> + * 0f 18 - prefetchBAR (intel prefetch insns)
> + * 0f 24 - mov from test regs (perhaps entire 20-27 area can be disabled (special reg ops))
> + * 0f 25 - undefined
> + * 0f 26 - mov to test regs
> + * 0f 27 - undefined
> + * 0f 30 - wrmsr (CPL0 insn)
> + * 0f 34 - sysenter
> + * 0f 35 - sysexit
> + * 0f 36 - undefined
> + * 0f 37 - getsec
> + * 0f 38-3f - 3-byte opcodes (why?? all look safe)
> + * 0f 78 - vmread
> + * 0f 79 - vmwrite
> + * 0f 7a - undefined
> + * 0f 7b - undefined
> + * 0f 7c - undefined
> + * 0f 7d - undefined
> + * 0f a6 - undefined
> + * 0f a7 - undefined
> + * 0f b8 - popcnt (why?? it's an ordinary ALU op)
> + * 0f d0 - undefined
> + * 0f f0 - lddqu (why?? it's an ordinary vector load op)
> + * 0f ff - undefined
> + */
>  static volatile u32 good_2byte_insns[256 / 32] = {
>  	/*      0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f         */
>  	/*      ----------------------------------------------         */
> @@ -148,23 +280,6 @@ static volatile u32 good_2byte_insns[256 / 32] = {
>  #undef W
>  
>  /*
> - * opcodes we'll probably never support:
> - *
> - *  6c-6d, e4-e5, ec-ed - in
> - *  6e-6f, e6-e7, ee-ef - out
> - *  cc, cd - int3, int
> - *  cf - iret
> - *  d6 - illegal instruction
> - *  f1 - int1/icebp
> - *  f4 - hlt
> - *  fa, fb - cli, sti
> - *  0f - lar, lsl, syscall, clts, sysret, sysenter, sysexit, invd, wbinvd, ud2
> - *
> - * invalid opcodes in 64-bit mode:
> - *
> - *  06, 0e, 16, 1e, 27, 2f, 37, 3f, 60-62, 82, c4-c5, d4-d5
> - *  63 - we support this opcode in x86_64 but not in i386.
> - *
>   * opcodes we may need to refine support for:
>   *
>   *  0f - 2-byte instructions: For many of these instructions, the validity
> -- 
> 1.8.1.4
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists