Fix cur_aux()->nospec_result test after do_check_insn() referring to the to-be-analyzed (potentially unsafe) instruction, not the already-analyzed (safe) instruction. This might allow a unsafe insn to slip through on a speculative path. Create some tests from the reproducer [1].
Commit d6f1c85f2253 ("bpf: Fall back to nospec for Spectre v1") should not be in any stable kernel yet, therefore bpf-next should suffice.
[1] https://lore.kernel.org/bpf/685b3c1b.050a0220.2303ee.0010.GAE@google.com/
Changes since v2: - Use insn_aux variable instead of introducing prev_aux() as suggested by Eduard (and therefore also drop patch 1) - v2: https://lore.kernel.org/bpf/20250628145016.784256-1-luis.gerhorst@fau.de/
Changes since v1: - Fix compiler error due to missed rename of prev_insn_idx in first patch - v1: https://lore.kernel.org/bpf/20250628125927.763088-1-luis.gerhorst@fau.de/
Changes since RFC: - Introduce prev_aux() as suggested by Alexei. For this, we must move the env->prev_insn_idx assignment to happen directly after do_check_insn(), for which I have created a separate commit. This patch could be simplified by using a local prev_aux variable as sugested by Eduard, but I figured one might find the new assignment-strategy easier to understand (before, prev_insn_idx and env->prev_insn_idx were out-of-sync for the latter part of the loop). Also, like this we do not have an additional prev_* variable that must be kept in-sync and the local variable's usage (old prev_insn_idx, new tmp) is much more local. If you think it would be better to not take the risk and keep the fix simple by just introducing the prev_aux variable, let me know. - Change WARN_ON_ONCE() to verifier_bug_if() as suggested by Alexei - Change assertion to check instruction is BPF_JMP[32] as suggested by Eduard - RFC: https://lore.kernel.org/bpf/8734bmoemx.fsf@fau.de/
Luis Gerhorst (2): bpf: Fix aux usage after do_check_insn() selftests/bpf: Add Spectre v4 tests
kernel/bpf/verifier.c | 19 ++- tools/testing/selftests/bpf/progs/bpf_misc.h | 4 + .../selftests/bpf/progs/verifier_unpriv.c | 149 ++++++++++++++++++ 3 files changed, 167 insertions(+), 5 deletions(-)
base-commit: 03fe01ddd1d8be7799419ea5e5f228a0186ae8c2
We must terminate the speculative analysis if the just-analyzed insn had nospec_result set. Using cur_aux() here is wrong because insn_idx might have been incremented by do_check_insn(). Therefore, introduce and use insn_aux variable.
Also change cur_aux(env)->nospec in case do_check_insn() ever manages to increment insn_idx but still fail.
Change the warning to check the insn class (which prevents it from triggering for ldimm64, for which nospec_result would not be problematic) and use verifier_bug_if().
In line with Eduard's suggestion, do not introduce prev_aux() because that requires one to understand that after do_check_insn() call what was current became previous. This would at-least require a comment.
Fixes: d6f1c85f2253 ("bpf: Fall back to nospec for Spectre v1") Reported-by: Paul Chaignon paul.chaignon@gmail.com Reported-by: Eduard Zingerman eddyz87@gmail.com Reported-by: syzbot+dc27c5fb8388e38d2d37@syzkaller.appspotmail.com Link: https://lore.kernel.org/bpf/685b3c1b.050a0220.2303ee.0010.GAE@google.com/ Link: https://lore.kernel.org/bpf/4266fd5de04092aa4971cbef14f1b4b96961f432.camel@g... Suggested-by: Eduard Zingerman eddyz87@gmail.com Signed-off-by: Luis Gerhorst luis.gerhorst@fau.de --- kernel/bpf/verifier.c | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 0f6cc2275695..96c737b41c3f 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -19923,6 +19923,7 @@ static int do_check(struct bpf_verifier_env *env)
for (;;) { struct bpf_insn *insn; + struct bpf_insn_aux_data *insn_aux; int err;
/* reset current history entry on each new instruction */ @@ -19936,6 +19937,7 @@ static int do_check(struct bpf_verifier_env *env) }
insn = &insns[env->insn_idx]; + insn_aux = &env->insn_aux_data[env->insn_idx];
if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) { verbose(env, @@ -20012,7 +20014,7 @@ static int do_check(struct bpf_verifier_env *env) /* Reduce verification complexity by stopping speculative path * verification when a nospec is encountered. */ - if (state->speculative && cur_aux(env)->nospec) + if (state->speculative && insn_aux->nospec) goto process_bpf_exit;
err = do_check_insn(env, &do_print_state); @@ -20020,11 +20022,11 @@ static int do_check(struct bpf_verifier_env *env) /* Prevent this speculative path from ever reaching the * insn that would have been unsafe to execute. */ - cur_aux(env)->nospec = true; + insn_aux->nospec = true; /* If it was an ADD/SUB insn, potentially remove any * markings for alu sanitization. */ - cur_aux(env)->alu_state = 0; + insn_aux->alu_state = 0; goto process_bpf_exit; } else if (err < 0) { return err; @@ -20033,7 +20035,7 @@ static int do_check(struct bpf_verifier_env *env) } WARN_ON_ONCE(err);
- if (state->speculative && cur_aux(env)->nospec_result) { + if (state->speculative && insn_aux->nospec_result) { /* If we are on a path that performed a jump-op, this * may skip a nospec patched-in after the jump. This can * currently never happen because nospec_result is only @@ -20042,8 +20044,15 @@ static int do_check(struct bpf_verifier_env *env) * never skip the following insn. Still, add a warning * to document this in case nospec_result is used * elsewhere in the future. + * + * All non-branch instructions have a single + * fall-through edge. For these, nospec_result should + * already work. */ - WARN_ON_ONCE(env->insn_idx != prev_insn_idx + 1); + if (verifier_bug_if(BPF_CLASS(insn->code) == BPF_JMP || + BPF_CLASS(insn->code) == BPF_JMP32, env, + "speculation barrier after jump instruction may not have the desired effect")) + return -EFAULT; process_bpf_exit: mark_verifier_state_scratched(env); err = update_branch_counts(env, env->cur_state);
On Sat, 2025-07-05 at 21:09 +0200, Luis Gerhorst wrote:
We must terminate the speculative analysis if the just-analyzed insn had nospec_result set. Using cur_aux() here is wrong because insn_idx might have been incremented by do_check_insn(). Therefore, introduce and use insn_aux variable.
Also change cur_aux(env)->nospec in case do_check_insn() ever manages to increment insn_idx but still fail.
Change the warning to check the insn class (which prevents it from triggering for ldimm64, for which nospec_result would not be problematic) and use verifier_bug_if().
In line with Eduard's suggestion, do not introduce prev_aux() because that requires one to understand that after do_check_insn() call what was current became previous. This would at-least require a comment.
Fixes: d6f1c85f2253 ("bpf: Fall back to nospec for Spectre v1") Reported-by: Paul Chaignon paul.chaignon@gmail.com Reported-by: Eduard Zingerman eddyz87@gmail.com Reported-by: syzbot+dc27c5fb8388e38d2d37@syzkaller.appspotmail.com Link: https://lore.kernel.org/bpf/685b3c1b.050a0220.2303ee.0010.GAE@google.com/ Link: https://lore.kernel.org/bpf/4266fd5de04092aa4971cbef14f1b4b96961f432.camel@g... Suggested-by: Eduard Zingerman eddyz87@gmail.com Signed-off-by: Luis Gerhorst luis.gerhorst@fau.de
Acked-by: Eduard Zingerman eddyz87@gmail.com
[...]
Add the following tests:
1. A test with an (unimportant) ldimm64 (16 byte insn) and a Spectre-v4--induced nospec that clarifies and serves as a basic Spectre v4 test.
2. Make sure a Spectre v4 nospec_result does not prevent a Spectre v1 nospec from being added before the dangerous instruction (tests that [1] is fixed).
3. Combine the two, which is the combination that triggers the warning in [2]. This is because the unanalyzed stack write has nospec_result set, but the ldimm64 (which was just analyzed) had incremented insn_idx by 2. That violates the assertion that nospec_result is only used after insns that increment insn_idx by 1 (i.e., stack writes).
[1] https://lore.kernel.org/bpf/4266fd5de04092aa4971cbef14f1b4b96961f432.camel@g... [2] https://lore.kernel.org/bpf/685b3c1b.050a0220.2303ee.0010.GAE@google.com/
Signed-off-by: Luis Gerhorst luis.gerhorst@fau.de --- tools/testing/selftests/bpf/progs/bpf_misc.h | 4 + .../selftests/bpf/progs/verifier_unpriv.c | 149 ++++++++++++++++++ 2 files changed, 153 insertions(+)
diff --git a/tools/testing/selftests/bpf/progs/bpf_misc.h b/tools/testing/selftests/bpf/progs/bpf_misc.h index 20dce508d8e0..530752ddde8e 100644 --- a/tools/testing/selftests/bpf/progs/bpf_misc.h +++ b/tools/testing/selftests/bpf/progs/bpf_misc.h @@ -237,4 +237,8 @@ #define SPEC_V1 #endif
+#if defined(__TARGET_ARCH_x86) +#define SPEC_V4 +#endif + #endif diff --git a/tools/testing/selftests/bpf/progs/verifier_unpriv.c b/tools/testing/selftests/bpf/progs/verifier_unpriv.c index 4470541b5e71..28b4f7035ceb 100644 --- a/tools/testing/selftests/bpf/progs/verifier_unpriv.c +++ b/tools/testing/selftests/bpf/progs/verifier_unpriv.c @@ -801,4 +801,153 @@ l2_%=: \ : __clobber_all); }
+SEC("socket") +__description("unpriv: ldimm64 before Spectre v4 barrier") +__success __success_unpriv +__retval(0) +#ifdef SPEC_V4 +__xlated_unpriv("r1 = 0x2020200005642020") /* should not matter */ +__xlated_unpriv("*(u64 *)(r10 -8) = r1") +__xlated_unpriv("nospec") +#endif +__naked void unpriv_ldimm64_spectre_v4(void) +{ + asm volatile (" \ + r1 = 0x2020200005642020 ll; \ + *(u64 *)(r10 -8) = r1; \ + r0 = 0; \ + exit; \ +" ::: __clobber_all); +} + +SEC("socket") +__description("unpriv: Spectre v1 and v4 barrier") +__success __success_unpriv +__retval(0) +#ifdef SPEC_V1 +#ifdef SPEC_V4 +/* starts with r0 == r8 == r9 == 0 */ +__xlated_unpriv("if r8 != 0x0 goto pc+1") +__xlated_unpriv("goto pc+2") +__xlated_unpriv("if r9 == 0x0 goto pc+4") +__xlated_unpriv("r2 = r0") +/* Following nospec required to prevent following dangerous `*(u64 *)(NOT_FP -64) + * = r1` iff `if r9 == 0 goto pc+4` was mispredicted because of Spectre v1. The + * test therefore ensures the Spectre-v4--induced nospec does not prevent the + * Spectre-v1--induced speculative path from being fully analyzed. + */ +__xlated_unpriv("nospec") /* Spectre v1 */ +__xlated_unpriv("*(u64 *)(r2 -64) = r1") /* could be used to leak r2 */ +__xlated_unpriv("nospec") /* Spectre v4 */ +#endif +#endif +__naked void unpriv_spectre_v1_and_v4(void) +{ + asm volatile (" \ + r1 = 0; \ + *(u64*)(r10 - 8) = r1; \ + r2 = r10; \ + r2 += -8; \ + r1 = %[map_hash_8b] ll; \ + call %[bpf_map_lookup_elem]; \ + r8 = r0; \ + r2 = r10; \ + r2 += -8; \ + r1 = %[map_hash_8b] ll; \ + call %[bpf_map_lookup_elem]; \ + r9 = r0; \ + r0 = r10; \ + r1 = 0; \ + r2 = r10; \ + if r8 != 0 goto l0_%=; \ + if r9 != 0 goto l0_%=; \ + r0 = 0; \ +l0_%=: if r8 != 0 goto l1_%=; \ + goto l2_%=; \ +l1_%=: if r9 == 0 goto l3_%=; \ + r2 = r0; \ +l2_%=: *(u64 *)(r2 -64) = r1; \ +l3_%=: r0 = 0; \ + exit; \ +" : + : __imm(bpf_map_lookup_elem), + __imm_addr(map_hash_8b) + : __clobber_all); +} + +SEC("socket") +__description("unpriv: Spectre v1 and v4 barrier (simple)") +__success __success_unpriv +__retval(0) +#ifdef SPEC_V1 +#ifdef SPEC_V4 +__xlated_unpriv("if r8 != 0x0 goto pc+1") +__xlated_unpriv("goto pc+2") +__xlated_unpriv("goto pc-1") /* if r9 == 0 goto l3_%= */ +__xlated_unpriv("goto pc-1") /* r2 = r0 */ +__xlated_unpriv("nospec") +__xlated_unpriv("*(u64 *)(r2 -64) = r1") +__xlated_unpriv("nospec") +#endif +#endif +__naked void unpriv_spectre_v1_and_v4_simple(void) +{ + asm volatile (" \ + r8 = 0; \ + r9 = 0; \ + r0 = r10; \ + r1 = 0; \ + r2 = r10; \ + if r8 != 0 goto l0_%=; \ + if r9 != 0 goto l0_%=; \ + r0 = 0; \ +l0_%=: if r8 != 0 goto l1_%=; \ + goto l2_%=; \ +l1_%=: if r9 == 0 goto l3_%=; \ + r2 = r0; \ +l2_%=: *(u64 *)(r2 -64) = r1; \ +l3_%=: r0 = 0; \ + exit; \ +" ::: __clobber_all); +} + +SEC("socket") +__description("unpriv: ldimm64 before Spectre v1 and v4 barrier (simple)") +__success __success_unpriv +__retval(0) +#ifdef SPEC_V1 +#ifdef SPEC_V4 +__xlated_unpriv("if r8 != 0x0 goto pc+1") +__xlated_unpriv("goto pc+4") +__xlated_unpriv("goto pc-1") /* if r9 == 0 goto l3_%= */ +__xlated_unpriv("goto pc-1") /* r2 = r0 */ +__xlated_unpriv("goto pc-1") /* r1 = 0x2020200005642020 ll */ +__xlated_unpriv("goto pc-1") /* second part of ldimm64 */ +__xlated_unpriv("nospec") +__xlated_unpriv("*(u64 *)(r2 -64) = r1") +__xlated_unpriv("nospec") +#endif +#endif +__naked void unpriv_ldimm64_spectre_v1_and_v4_simple(void) +{ + asm volatile (" \ + r8 = 0; \ + r9 = 0; \ + r0 = r10; \ + r1 = 0; \ + r2 = r10; \ + if r8 != 0 goto l0_%=; \ + if r9 != 0 goto l0_%=; \ + r0 = 0; \ +l0_%=: if r8 != 0 goto l1_%=; \ + goto l2_%=; \ +l1_%=: if r9 == 0 goto l3_%=; \ + r2 = r0; \ + r1 = 0x2020200005642020 ll; \ +l2_%=: *(u64 *)(r2 -64) = r1; \ +l3_%=: r0 = 0; \ + exit; \ +" ::: __clobber_all); +} + char _license[] SEC("license") = "GPL";
Hello:
This series was applied to bpf/bpf-next.git (master) by Alexei Starovoitov ast@kernel.org:
On Sat, 5 Jul 2025 21:09:06 +0200 you wrote:
Fix cur_aux()->nospec_result test after do_check_insn() referring to the to-be-analyzed (potentially unsafe) instruction, not the already-analyzed (safe) instruction. This might allow a unsafe insn to slip through on a speculative path. Create some tests from the reproducer [1].
Commit d6f1c85f2253 ("bpf: Fall back to nospec for Spectre v1") should not be in any stable kernel yet, therefore bpf-next should suffice.
[...]
Here is the summary with links: - [bpf-next,v3,1/2] bpf: Fix aux usage after do_check_insn() https://git.kernel.org/bpf/bpf-next/c/dadb59104c64 - [bpf-next,v3,2/2] selftests/bpf: Add Spectre v4 tests https://git.kernel.org/bpf/bpf-next/c/92974cef83b5
You are awesome, thank you!
linux-kselftest-mirror@lists.linaro.org
-
Eduard Zingerman -
Luis Gerhorst -
patchwork-bot+netdevbpf＠kernel.org