On 7/29/22 18:53, Sean Christopherson wrote:

On Fri, Jul 29, 2022, Michal Luczaj wrote:

...

• • To trigger the emulator and feed it with LEA, we VM-exit on IO (with a
• • single OUTS), then race decoder's instruction fetch - hoping to replace the
• • initial IO op with an illegal LEA.

Rather than play games with memory, can't we just require and use force_emulation_prefix to force KVM to emulate a bogus LEA encoding? emulator.c in KVM-unit-tests already has most of what you need, e.g. I believe it's just a matter of implementing test_illegal_lea(). That test already has test_smsw_reg(), which is darn near the same thing, it just expects a different result (success instead of #UD).

diff --git a/x86/emulator.c b/x86/emulator.c index cd78e3cb..dd50578d 100644 --- a/x86/emulator.c +++ b/x86/emulator.c @@ -1193,6 +1193,7 @@ int main(void) test_smsw_reg(mem); test_nop(mem); test_mov_dr(mem);

•           test_illegal_lea();
    } else {
            report_skip("skipping register-only tests, "
                        "use kvm.force_emulation_prefix=1 to enable");
  
  

Ahh, right. Using force_emulation_prefix seems way better. Thanks!

I'll add "kvm-unit-tests" in the subject and send as v2 PATCH.

Michal

On Sun, Jul 31, 2022, Michal Luczaj wrote:

Check if the emulator throws #UD on illegal LEA.

Please explicitly state exactly what illegal LEA is being generated. Requiring readers to connect the dots of the LEA opcode and ModR/M encoding is unnecessarily mean :-)

Suggested-by: Sean Christopherson seanjc@google.com Signed-off-by: Michal Luczaj mhal@rbox.co

---

v1 -> v2: Instead of racing decoder make use of force_emulation_prefix

x86/emulator.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)

diff --git a/x86/emulator.c b/x86/emulator.c index cd78e3c..c3898f2 100644 --- a/x86/emulator.c +++ b/x86/emulator.c @@ -895,6 +895,24 @@ static void test_mov_dr(uint64_t *mem) report(rax == DR6_ACTIVE_LOW, "mov_dr6"); } +static void illegal_lea_handler(struct ex_regs *regs) +{

• extern char illegal_lea_cont;
• ++exceptions;
• regs->rip = (ulong)&illegal_lea_cont;

+}

+static void test_illegal_lea(uint64_t *mem)

@mem isn't needed.

+{

• exceptions = 0;
• handle_exception(UD_VECTOR, illegal_lea_handler);

No need to use a custom handler (ignore any patterns in emulator.c that suggest it's "mandatory", emulator is one of the oldest test). ASM_TRY() can handle all of this without any globals.

• asm(KVM_FEP ".byte 0x48; .byte 0x8d; .byte 0xc0\n\t"

•    "illegal_lea_cont:" : : : "rax");
  

"emulator" is compatible with 32-bit KUT, drop the REX prefix and clobber "eax" instead of "xax".

• report(exceptions == 1, "illegal lea");

Be nice to future debuggers. Call out what is illegal about LEA, capitalize LEA to make it more obvious that its an instruction, and print the expected versus actual.

• handle_exception(UD_VECTOR, 0);

+}

So this?

static void test_illegal_lea(void) { unsigned int vector;

asm volatile (ASM_TRY("1f") KVM_FEP ".byte 0x8d; .byte 0xc0\n\t" "1:" : : : "memory", "eax");

vector = exception_vector(); report(vector == UD_VECTOR, "Wanted #UD on LEA with /reg, got vector = %d", vector); }

static void test_push16(uint64_t *mem) { uint64_t rsp1, rsp2; @@ -1193,6 +1211,7 @@ int main(void) test_smsw_reg(mem); test_nop(mem); test_mov_dr(mem);

• test_illegal_lea(mem);
  

  } else { report_skip("skipping register-only tests, " "use kvm.force_emulation_prefix=1 to enable");

-- 2.32.0

Users of handle_exception() should always save and restore the handlers.

Suggested-by: Sean Christopherson seanjc@google.com Signed-off-by: Michal Luczaj mhal@rbox.co --- I took the liberty of fixing the indentation of functions I've touched.

x86/emulator.c | 78 ++++++++++++++++++++++++++------------------------ 1 file changed, 41 insertions(+), 37 deletions(-)

diff --git a/x86/emulator.c b/x86/emulator.c index cd78e3c..769a049 100644 --- a/x86/emulator.c +++ b/x86/emulator.c @@ -710,6 +710,7 @@ static __attribute__((target("sse2"))) void test_sse_exceptions(void *cross_mem) void *page2 = (void *)(&bytes[4096]); struct pte_search search; pteval_t orig_pte; + handler old;

// setup memory for unaligned access mem = (uint32_t *)(&bytes[8]); @@ -725,10 +726,10 @@ static __attribute__((target("sse2"))) void test_sse_exceptions(void *cross_mem) asm("movupd %1, %0" : "=m"(*mem) : "x"(vv) : "memory"); report(sseeq(v, mem), "movupd unaligned"); exceptions = 0; - handle_exception(GP_VECTOR, unaligned_movaps_handler); + old = handle_exception(GP_VECTOR, unaligned_movaps_handler); asm("movaps %1, %0\n\t unaligned_movaps_cont:" : "=m"(*mem) : "x"(vv)); - handle_exception(GP_VECTOR, 0); + handle_exception(GP_VECTOR, old); report(exceptions == 1, "unaligned movaps exception");

// setup memory for cross page access @@ -746,10 +747,10 @@ static __attribute__((target("sse2"))) void test_sse_exceptions(void *cross_mem) invlpg(page2);

exceptions = 0; - handle_exception(PF_VECTOR, cross_movups_handler); + old = handle_exception(PF_VECTOR, cross_movups_handler); asm("movups %1, %0\n\t cross_movups_cont:" : "=m"(*mem) : "x"(vv) : "memory"); - handle_exception(PF_VECTOR, 0); + handle_exception(PF_VECTOR, old); report(exceptions == 1, "movups crosspage exception");

// restore invalidated page @@ -817,36 +818,38 @@ static void advance_rip_and_note_exception(struct ex_regs *regs)

static void test_mmx_movq_mf(uint64_t *mem) { - /* movq %mm0, (%rax) */ - extern char movq_start, movq_end; - - uint16_t fcw = 0; /* all exceptions unmasked */ - write_cr0(read_cr0() & ~6); /* TS, EM */ - exceptions = 0; - handle_exception(MF_VECTOR, advance_rip_and_note_exception); - asm volatile("fninit; fldcw %0" : : "m"(fcw)); - asm volatile("fldz; fldz; fdivp"); /* generate exception */ - - rip_advance = &movq_end - &movq_start; - asm(KVM_FEP "movq_start: movq %mm0, (%rax); movq_end:"); - /* exit MMX mode */ - asm volatile("fnclex; emms"); - report(exceptions == 1, "movq mmx generates #MF"); - handle_exception(MF_VECTOR, 0); + /* movq %mm0, (%rax) */ + extern char movq_start, movq_end; + handler old; + + uint16_t fcw = 0; /* all exceptions unmasked */ + write_cr0(read_cr0() & ~6); /* TS, EM */ + exceptions = 0; + old = handle_exception(MF_VECTOR, advance_rip_and_note_exception); + asm volatile("fninit; fldcw %0" : : "m"(fcw)); + asm volatile("fldz; fldz; fdivp"); /* generate exception */ + + rip_advance = &movq_end - &movq_start; + asm(KVM_FEP "movq_start: movq %mm0, (%rax); movq_end:"); + /* exit MMX mode */ + asm volatile("fnclex; emms"); + report(exceptions == 1, "movq mmx generates #MF"); + handle_exception(MF_VECTOR, old); }

static void test_jmp_noncanonical(uint64_t *mem) { extern char nc_jmp_start, nc_jmp_end; + handler old;

*mem = 0x1111111111111111ul;

exceptions = 0; rip_advance = &nc_jmp_end - &nc_jmp_start; - handle_exception(GP_VECTOR, advance_rip_and_note_exception); + old = handle_exception(GP_VECTOR, advance_rip_and_note_exception); asm volatile ("nc_jmp_start: jmp *%0; nc_jmp_end:" : : "m"(*mem)); report(exceptions == 1, "jump to non-canonical address"); - handle_exception(GP_VECTOR, 0); + handle_exception(GP_VECTOR, old); }

static void test_movabs(uint64_t *mem) @@ -979,22 +982,23 @@ static void ss_bad_rpl(struct ex_regs *regs)

static void test_sreg(volatile uint16_t *mem) { - u16 ss = read_ss(); + u16 ss = read_ss(); + handler old;

- // check for null segment load - *mem = 0; - asm volatile("mov %0, %%ss" : : "m"(*mem)); - report(read_ss() == 0, "mov null, %%ss"); - - // check for exception when ss.rpl != cpl on null segment load - exceptions = 0; - handle_exception(GP_VECTOR, ss_bad_rpl); - *mem = 3; - asm volatile("mov %0, %%ss; ss_bad_rpl_cont:" : : "m"(*mem)); - report(exceptions == 1 && read_ss() == 0, - "mov null, %%ss (with ss.rpl != cpl)"); - handle_exception(GP_VECTOR, 0); - write_ss(ss); + // check for null segment load + *mem = 0; + asm volatile("mov %0, %%ss" : : "m"(*mem)); + report(read_ss() == 0, "mov null, %%ss"); + + // check for exception when ss.rpl != cpl on null segment load + exceptions = 0; + old = handle_exception(GP_VECTOR, ss_bad_rpl); + *mem = 3; + asm volatile("mov %0, %%ss; ss_bad_rpl_cont:" : : "m"(*mem)); + report(exceptions == 1 && read_ss() == 0, + "mov null, %%ss (with ss.rpl != cpl)"); + handle_exception(GP_VECTOR, old); + write_ss(ss); }

static uint64_t usr_gs_mov(void)

LEA with a register-direct source operand is illegal. Verify that the emulator raises #UD.

Suggested-by: Sean Christopherson seanjc@google.com Signed-off-by: Michal Luczaj mhal@rbox.co --- x86/emulator.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)

diff --git a/x86/emulator.c b/x86/emulator.c index d4488a7..df0bc49 100644 --- a/x86/emulator.c +++ b/x86/emulator.c @@ -896,6 +896,20 @@ static void test_mov_dr(uint64_t *mem) report(rax == DR6_ACTIVE_LOW, "mov_dr6"); }

+static void test_illegal_lea(void) +{ + unsigned int vector; + + asm volatile (ASM_TRY("1f") + KVM_FEP ".byte 0x8d; .byte 0xc0\n\t" + "1:" + : : : "memory", "eax"); + + vector = exception_vector(); + report(vector == UD_VECTOR, + "Wanted #UD on LEA with /reg, got vector = %u", vector); +} + static void test_push16(uint64_t *mem) { uint64_t rsp1, rsp2; @@ -1187,6 +1201,7 @@ int main(void) test_smsw_reg(mem); test_nop(mem); test_mov_dr(mem); + test_illegal_lea(); } else { report_skip("skipping register-only tests, " "use kvm.force_emulation_prefix=1 to enable");

On Wed, Aug 03, 2022, Michal Luczaj wrote:

TRY_ASM() mishandles #UD thrown by the forced-emulation-triggered emulator. While the faulting address stored in the exception table points at forced emulation prefix, when #UD comes, RIP is 5 bytes (size of KVM_FEP) ahead and the exception ends up unhandled.

Ah, but that's only the behavior if FEP is allowed. If FEP is disabled, then the #UD will be on the prefix.

Suggested-by: Sean Christopherson seanjc@google.com

Heh, I didn't really suggest this because I didn't even realize it was a problem :-)

Signed-off-by: Michal Luczaj mhal@rbox.co

While here, I've also took the opportunity to merge both 32 and 64-bit versions of ASM_TRY() (.dc.a for .long and .quad), but perhaps there were some reasons for not using .dc.a?

This should be a separate patch, and probably as the very last patch in case dc.a isn't viable for whatever reason. I've never seen/used dc.a so I really have no idea whether or not it's ok to use.

As a "safe" alternative that can be done before adding ASM_TRY_FEP(), we can steal the kernel's __ASM_SEL() approach so that you don't have to implement two versions of ASM_TRY_FEP(). That's worth doing because __ASM_SEL() can probably be used in other places too. Patch at the bottom.

lib/x86/desc.h | 11 +++++------ x86/emulator.c | 4 ++-- 2 files changed, 7 insertions(+), 8 deletions(-)

diff --git a/lib/x86/desc.h b/lib/x86/desc.h index 2a285eb..99cc224 100644 --- a/lib/x86/desc.h +++ b/lib/x86/desc.h @@ -80,21 +80,20 @@ typedef struct __attribute__((packed)) { u16 iomap_base; } tss64_t; -#ifdef __x86_64 #define ASM_TRY(catch) \ "movl $0, %%gs:4 \n\t" \ ".pushsection .data.ex \n\t" \

• ".quad 1111f, " catch "\n\t" \

• ".dc.a 1111f, " catch "\n\t" \ ".popsection \n\t" \ "1111:"

-#else -#define ASM_TRY(catch) \

+#define ASM_TRY_PREFIXED(prefix, catch) \

Rather than a generic ASM_TRY_PREFIXED, I think it makes sense to add an explicit ASM_TRY_FEP() that takes in only the label and hardcodes the FEP prefix. The "#UD skips the prefix" behavior is unique to "successful" forced emulation, so this is literally the only scenario where skipping a prefix is expected/correct.

*sigh*

That'll require moving the KVM_FEP definition, but that's a good change on its own. Probably throw it in lib/x86/processor.h?

"movl $0, %%gs:4 \n\t" \ ".pushsection .data.ex \n\t" \

• ".long 1111f, " catch "\n\t" \

• ".dc.a 1111f, " catch "\n\t" \ ".popsection \n\t" \
• prefix "\n\t" \ "1111:"

-#endif /*

• selector 32-bit 64-bit

diff --git a/x86/emulator.c b/x86/emulator.c index df0bc49..d2a5302 100644 --- a/x86/emulator.c +++ b/x86/emulator.c @@ -900,8 +900,8 @@ static void test_illegal_lea(void) { unsigned int vector;

• asm volatile (ASM_TRY("1f")

•       KVM_FEP ".byte 0x8d; .byte 0xc0\n\t"
  

• asm volatile (ASM_TRY_PREFIXED(KVM_FEP, "1f")

•       ".byte 0x8d; .byte 0xc0\n\t"
  

Put this patch earlier in the series, before test_illegal_lea() is introduced. Both so that there's no unnecessary churn, and also because running the illegal LEA testcase without this patch will fail.

      "1:"
      : : : "memory", "eax");

--- From: Sean Christopherson seanjc@google.com Date: Wed, 3 Aug 2022 11:09:41 -0700 Subject: [PATCH] x86: Dedup 32-bit vs. 64-bit ASM_TRY() by stealing kernel's __ASM_SEL()

Steal the kernel's __ASM_SEL() implementation and use it to consolidate ASM_TRY(). The only difference between the 32-bit and 64-bit versions is the size of the address stored in the table.

No functional change intended.

Signed-off-by: Sean Christopherson seanjc@google.com --- lib/x86/desc.h | 19 +++++-------------- lib/x86/processor.h | 12 ++++++++++++ 2 files changed, 17 insertions(+), 14 deletions(-)

diff --git a/lib/x86/desc.h b/lib/x86/desc.h index 2a285eb6..e670ebf2 100644 --- a/lib/x86/desc.h +++ b/lib/x86/desc.h @@ -80,21 +80,12 @@ typedef struct __attribute__((packed)) { u16 iomap_base; } tss64_t;

-#ifdef __x86_64 -#define ASM_TRY(catch) \ - "movl $0, %%gs:4 \n\t" \ - ".pushsection .data.ex \n\t" \ - ".quad 1111f, " catch "\n\t" \ - ".popsection \n\t" \ +#define ASM_TRY(catch) \ + "movl $0, %%gs:4 \n\t" \ + ".pushsection .data.ex \n\t" \ + __ASM_SEL(.long, .quad) " 1111f, " catch "\n\t" \ + ".popsection \n\t" \ "1111:" -#else -#define ASM_TRY(catch) \ - "movl $0, %%gs:4 \n\t" \ - ".pushsection .data.ex \n\t" \ - ".long 1111f, " catch "\n\t" \ - ".popsection \n\t" \ - "1111:" -#endif

/* * selector 32-bit 64-bit diff --git a/lib/x86/processor.h b/lib/x86/processor.h index 03242206..30e2de87 100644 --- a/lib/x86/processor.h +++ b/lib/x86/processor.h @@ -19,6 +19,18 @@ # define S "4" #endif

+#ifdef __ASSEMBLY__ +#define __ASM_FORM(x, ...) x,## __VA_ARGS__ +#else +#define __ASM_FORM(x, ...) " " xstr(x,##__VA_ARGS__) " " +#endif + +#ifndef __x86_64__ +#define __ASM_SEL(a,b) __ASM_FORM(a) +#else +#define __ASM_SEL(a,b) __ASM_FORM(b) +#endif + #define DB_VECTOR 1 #define BP_VECTOR 3 #define UD_VECTOR 6

base-commit: a106b30d39425b7afbaa3bbd4aab16fd26d333e7 --