| Message ID | 5CA35D130200007800223E53@prv1-mh.provo.novell.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v3] x86emul/fuzz: add a state sanity checking function | expand |
On 4/2/19 2:01 PM, Jan Beulich wrote: > This is to accompany sanitize_input(). Just like for initial state we > want to have state between two emulated insns sane, at least as far as > assumptions in the main emulator go. Do minimal checking after segment > register, CR, and MSR writes, and roll back to the old value in case of > failure (raising #GP(0) at the same time). > > In the particular case observed, a CR0 write clearing CR0.PE was > followed by a VEX-encoded insn, which the decoder accepts based on > guest address size, restricting things just outside of the 64-bit case > (real and virtual modes don't allow VEX-encoded insns). Subsequently > _get_fpu() would then assert that CR0.PE must be set (and EFLAGS.VM > clear) when trying to invoke YMM, ZMM, or OPMASK state. > > Signed-off-by: Jan Beulich <jbeulich@suse.com> Reviewed-by: George Dunlap <george.dunlap@citrix.com> That said, I wonder if there's a way to avoid the duplication between sanitize_input() and check_state(). Another option would be to rework sanitize_input() (perhaps as sanizite_state()): * Accept a parameter saying whether to do optional changes (like CANONICALIZE_MAYBE) * Return a boolean saying whether any state was in fact sanitized. Then the current callers of check_state() could instead call sanitize_state(), and throw an exception if it returns 1. (Or some variation thereof.) Just a thought; I'm OK with checking this in as it is. -George
>>> On 27.05.19 at 12:51, <george.dunlap@citrix.com> wrote: > On 4/2/19 2:01 PM, Jan Beulich wrote: >> This is to accompany sanitize_input(). Just like for initial state we >> want to have state between two emulated insns sane, at least as far as >> assumptions in the main emulator go. Do minimal checking after segment >> register, CR, and MSR writes, and roll back to the old value in case of >> failure (raising #GP(0) at the same time). >> >> In the particular case observed, a CR0 write clearing CR0.PE was >> followed by a VEX-encoded insn, which the decoder accepts based on >> guest address size, restricting things just outside of the 64-bit case >> (real and virtual modes don't allow VEX-encoded insns). Subsequently >> _get_fpu() would then assert that CR0.PE must be set (and EFLAGS.VM >> clear) when trying to invoke YMM, ZMM, or OPMASK state. >> >> Signed-off-by: Jan Beulich <jbeulich@suse.com> > > Reviewed-by: George Dunlap <george.dunlap@citrix.com> Thanks. > That said, I wonder if there's a way to avoid the duplication between > sanitize_input() and check_state(). Another option would be to rework > sanitize_input() (perhaps as sanizite_state()): > * Accept a parameter saying whether to do optional changes (like > CANONICALIZE_MAYBE) > * Return a boolean saying whether any state was in fact sanitized. > > Then the current callers of check_state() could instead call > sanitize_state(), and throw an exception if it returns 1. (Or some > variation thereof.) I did consider this at the time, but the two functions aren't doing exactly the same validation. For example this /* EFLAGS.VM not available in long mode */ if ( long_mode_active(ctxt) ) regs->rflags &= ~X86_EFLAGS_VM; has no equivalent in check_state(), for it being an emulator bug to ever set EFLAGS.VM in long mode. I therefore thought it would be better to keep them separate despite there being partial redundancy. If the set of checks grows, we could consider factoring out the common subset into a helper function. Jan
--- a/tools/fuzz/x86_instruction_emulator/fuzz-emul.c +++ b/tools/fuzz/x86_instruction_emulator/fuzz-emul.c @@ -76,6 +76,8 @@ static inline bool input_read(struct fuz return true; } +static bool check_state(struct x86_emulate_ctxt *ctxt); + static const char* const x86emul_return_string[] = { [X86EMUL_OKAY] = "X86EMUL_OKAY", [X86EMUL_UNHANDLEABLE] = "X86EMUL_UNHANDLEABLE", @@ -424,8 +426,19 @@ static int fuzz_write_segment( rc = maybe_fail(ctxt, "write_segment", true); if ( rc == X86EMUL_OKAY ) + { + struct segment_register old = c->segments[seg]; + c->segments[seg] = *reg; + if ( !check_state(ctxt) ) + { + c->segments[seg] = old; + x86_emul_hw_exception(13 /* #GP */, 0, ctxt); + rc = X86EMUL_EXCEPTION; + } + } + return rc; } @@ -452,6 +465,7 @@ static int fuzz_write_cr( { struct fuzz_state *s = ctxt->data; struct fuzz_corpus *c = s->corpus; + unsigned long old; int rc; if ( reg >= ARRAY_SIZE(c->cr) ) @@ -461,9 +475,17 @@ static int fuzz_write_cr( if ( rc != X86EMUL_OKAY ) return rc; + old = c->cr[reg]; c->cr[reg] = val; - return X86EMUL_OKAY; + if ( !check_state(ctxt) ) + { + c->cr[reg] = old; + x86_emul_hw_exception(13 /* #GP */, 0, ctxt); + rc = X86EMUL_EXCEPTION; + } + + return rc; } #define fuzz_read_xcr emul_test_read_xcr @@ -561,7 +583,16 @@ static int fuzz_write_msr( { if ( msr_index[idx] == reg ) { + uint64_t old = c->msr[idx]; + c->msr[idx] = val; + + if ( !check_state(ctxt) ) + { + c->msr[idx] = old; + break; + } + return X86EMUL_OKAY; } } @@ -811,6 +842,30 @@ static void sanitize_input(struct x86_em } } +/* + * Call this function from hooks potentially altering machine state into + * something that's not architecturally valid, yet which - as per above - + * the emulator relies on. + */ +static bool check_state(struct x86_emulate_ctxt *ctxt) +{ + const struct fuzz_state *s = ctxt->data; + const struct fuzz_corpus *c = s->corpus; + const struct cpu_user_regs *regs = &c->regs; + + if ( long_mode_active(ctxt) && !(c->cr[0] & X86_CR0_PG) ) + return false; + + if ( (c->cr[0] & X86_CR0_PG) && !(c->cr[0] & X86_CR0_PE) ) + return false; + + if ( (regs->rflags & X86_EFLAGS_VM) && + (c->segments[x86_seg_cs].db || c->segments[x86_seg_ss].db) ) + return false; + + return true; +} + int LLVMFuzzerInitialize(int *argc, char ***argv) { if ( !emul_test_init() )
This is to accompany sanitize_input(). Just like for initial state we want to have state between two emulated insns sane, at least as far as assumptions in the main emulator go. Do minimal checking after segment register, CR, and MSR writes, and roll back to the old value in case of failure (raising #GP(0) at the same time). In the particular case observed, a CR0 write clearing CR0.PE was followed by a VEX-encoded insn, which the decoder accepts based on guest address size, restricting things just outside of the 64-bit case (real and virtual modes don't allow VEX-encoded insns). Subsequently _get_fpu() would then assert that CR0.PE must be set (and EFLAGS.VM clear) when trying to invoke YMM, ZMM, or OPMASK state. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- v3: Rename helper function to check_state(). v2: Correct placement of new declaration in fuzz_write_segment().