[PATCH RESEND 1/3] uprobes: add comment with insn opcodes, mnemonics and why we dont support them

[Denys Vlasenko]

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@xxxxxxxxxx>
CC: linux-kernel@xxxxxxxxxxxxxxx
CC: X86 ML <x86@xxxxxxxxxx>
CC: Jim Keniston <jkenisto@xxxxxxxxxx>
CC: Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx>
CC: Srikar Dronamraju <srikar@xxxxxxxxxxxxxxxxxx>
CC: Ingo Molnar <mingo@xxxxxxxxxx>
CC: Oleg Nesterov <oleg@xxxxxxxxxx>
---

Last time these patches were sent, they were discussed a bit and then
tentatively okay'ed, but they aren't in the tree. So: ping...

 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@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/