lkml.org 
[lkml]   [2022]   [Jul]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] efi/x86: use naked RET on mixed mode call wrapper
On Mon, Jul 18, 2022 at 01:41:37PM +0200, Peter Zijlstra wrote:
> On Fri, Jul 15, 2022 at 04:45:50PM -0300, Thadeu Lima de Souza Cascardo wrote:
> > When running with return thunks enabled under 32-bit EFI, the system
> > crashes with:
> >
> > [ 0.137688] kernel tried to execute NX-protected page - exploit attempt? (uid: 0)
> > [ 0.138136] BUG: unable to handle page fault for address: 000000005bc02900
> > [ 0.138136] #PF: supervisor instruction fetch in kernel mode
> > [ 0.138136] #PF: error_code(0x0011) - permissions violation
> > [ 0.138136] PGD 18f7063 P4D 18f7063 PUD 18ff063 PMD 190e063 PTE 800000005bc02063
> > [ 0.138136] Oops: 0011 [#1] PREEMPT SMP PTI
> > [ 0.138136] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.19.0-rc6+ #166
> > [ 0.138136] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
> > [ 0.138136] RIP: 0010:0x5bc02900
> > [ 0.138136] Code: Unable to access opcode bytes at RIP 0x5bc028d6.
> > [ 0.138136] RSP: 0018:ffffffffb3203e10 EFLAGS: 00010046
> > [ 0.138136] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000048
> > [ 0.138136] RDX: 000000000190dfac RSI: 0000000000001710 RDI: 000000007eae823b
> > [ 0.138136] RBP: ffffffffb3203e70 R08: 0000000001970000 R09: ffffffffb3203e28
> > [ 0.138136] R10: 747563657865206c R11: 6c6977203a696665 R12: 0000000000001710
> > [ 0.138136] R13: 0000000000000030 R14: 0000000001970000 R15: 0000000000000001
> > [ 0.138136] FS: 0000000000000000(0000) GS:ffff8e013ca00000(0000) knlGS:0000000000000000
> > [ 0.138136] CS: 0010 DS: 0018 ES: 0018 CR0: 0000000080050033
> > [ 0.138136] CR2: 000000005bc02900 CR3: 0000000001930000 CR4: 00000000000006f0
> > [ 0.138136] Call Trace:
> > [ 0.138136] <TASK>
> > [ 0.138136] ? efi_set_virtual_address_map+0x9c/0x175
> > [ 0.138136] efi_enter_virtual_mode+0x4a6/0x53e
> > [ 0.138136] start_kernel+0x67c/0x71e
> > [ 0.138136] x86_64_start_reservations+0x24/0x2a
> > [ 0.138136] x86_64_start_kernel+0xe9/0xf4
> > [ 0.138136] secondary_startup_64_no_verify+0xe5/0xeb
> > [ 0.138136] </TASK>
> >
> > That's because it cannot jump to the return thunk from the 32-bit code.
> > Using a naked RET and marking it as safe allows the system to proceed
> > booting.
> >
> > Fixes: aa3d480315ba ("x86: Use return-thunk in asm code")
> > Reported-by: Guenter Roeck <linux@roeck-us.net>
> > Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
> > Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
> > Cc: Borislav Petkov <bp@suse.de>
> > Cc: Josh Poimboeuf <jpoimboe@kernel.org>
> > Cc: <stable@vger.kernel.org>
> > ---
> >
> > Does this leave one potential attack vector open? Perhaps, since this is
> > running under a different mapping (AFAIU), the risk is reduced? Or rather, the
> > attacker could attack using the firmware RETs anyway?
> >
> > Alternatively, we could use IBPB when available when using the wrapper.
> >
> > Thoughts?
>
> What actual uarch are you running this on? Is this AMD hardware?
>
> For Intel we'll enable IBRS for firmware if it is not otherwise enabled
> (upstream will always enable IBRS for the SKL family chips, but Thomas
> just posted the retbleed=stuff approach yesterday that will not)
>
> On AMD I think you're stuck with IBPB, but that has to be issued
> *before* calling the firmware muck.
>
> In either case, I think the patch as proposed is fine. But perhaps we
> want something like the below on top.
>

I was testing this on Intel, because Guenter Roeck had reported such failures
when booting with 32-bit EFI. My patch fixes the boot problem, but then I asked
whether it would leave us vulnerable and, then, I was thinking about AMD
mostly, as you pointed out.

And I think you nailed what I had in mind for using IBPB when doing firmware
calls, and perhaps this is wanted even when we ignore this naked RET here.

There is a typo on your patch below, but I will give it a try and see if it
doesn't blow up on AMD systems without IBPB (by way of emulation).

Thanks.
Cascardo.

> ---
> Subject: x86/amd: Use IBPB for firmware calls
>
> On AMD IBRS does not prevent Retbleed; as such use IBPB before a
> firmware call to flush the branch history state.
>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
> arch/x86/include/asm/cpufeatures.h | 1 +
> arch/x86/include/asm/nospec-branch.h | 2 ++
> arch/x86/kernel/cpu/bugs.c | 11 ++++++++++-
> 3 files changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
> index 00f5227c8459..a77b915d36a8 100644
> --- a/arch/x86/include/asm/cpufeatures.h
> +++ b/arch/x86/include/asm/cpufeatures.h
> @@ -302,6 +302,7 @@
> #define X86_FEATURE_RETPOLINE_LFENCE (11*32+13) /* "" Use LFENCE for Spectre variant 2 */
> #define X86_FEATURE_RETHUNK (11*32+14) /* "" Use REturn THUNK */
> #define X86_FEATURE_UNRET (11*32+15) /* "" AMD BTB untrain return */
> +#define X86_FEATURE_USE_IBPB_FW (11*32+16) /* "" Use IBPB during runtime firmware calls */
>
> /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */
> #define X86_FEATURE_AVX_VNNI (12*32+ 4) /* AVX VNNI instructions */
> diff --git a/arch/x86/include/asm/nospec-branch.h b/arch/x86/include/asm/nospec-branch.h
> index 10a3bfc1eb23..f934dcdb7c0d 100644
> --- a/arch/x86/include/asm/nospec-branch.h
> +++ b/arch/x86/include/asm/nospec-branch.h
> @@ -297,6 +297,8 @@ do { \
> alternative_msr_write(MSR_IA32_SPEC_CTRL, \
> spec_ctrl_current() | SPEC_CTRL_IBRS, \
> X86_FEATURE_USE_IBRS_FW); \
> + altnerative_msr_write(MSR_IA32_PRED_CMD, PRED_CMD_IBPB, \
> + X86_FEATURE_USE_IBPB_FW); \
> } while (0)
>
> #define firmware_restrict_branch_speculation_end() \
> diff --git a/arch/x86/kernel/cpu/bugs.c b/arch/x86/kernel/cpu/bugs.c
> index aa34f908c39f..78c9082242a9 100644
> --- a/arch/x86/kernel/cpu/bugs.c
> +++ b/arch/x86/kernel/cpu/bugs.c
> @@ -1516,7 +1516,16 @@ static void __init spectre_v2_select_mitigation(void)
> * the CPU supports Enhanced IBRS, kernel might un-intentionally not
> * enable IBRS around firmware calls.
> */
> - if (boot_cpu_has(X86_FEATURE_IBRS) && !spectre_v2_in_ibrs_mode(mode)) {
> + if (boot_cpu_has_bug(X86_BUG_RETBLEED) &&
> + (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
> + boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)) {
> +
> + if (retbleed_cmd != RETBLEED_CMD_IBPB) {
> + setup_force_cpu_cap(X86_FEATURE_USE_IBPB_FW);
> + pr_info("Enabling Speculation Barrier for firmware calls\n");
> + }
> +
> + } else if (boot_cpu_has(X86_FEATURE_IBRS) && !spectre_v2_in_ibrs_mode(mode)) {
> setup_force_cpu_cap(X86_FEATURE_USE_IBRS_FW);
> pr_info("Enabling Restricted Speculation for firmware calls\n");
> }

\
 
 \ /
  Last update: 2022-07-18 16:00    [W:0.081 / U:0.064 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site