Hello,
[ Given the CC list and your mention of spectre_v2_user=prctl is spot
on to show the badness... I spawned a new thread to suggest another
thing related to seccomp that I've been intending to suggest for
a while ]
On Tue, Nov 03, 2020 at 04:29:38PM -0800, Kees Cook wrote:
> I assume this is from Indirect Branch Prediction Barrier (IBPB) and
> Single Threaded Indirect Branch Prediction (STIBP) (which get enabled
> for threads under seccomp by default).
>
> Try booting with "spectre_v2_user=prctl"
We need to change the kernel default to
"spec_store_bypass_disable=prctl spectre_v2_user=prctl".
I've been recommending to everyone to use
"spec_store_bypass_disable=prctl spectre_v2_user=prctl" for a while
now. I already recommend to Yifei too a few months ago when he first
found out of the huge seccomp regression when he upgraded his codebase
to the upstream kernel with both STIBP/SSBD enabled in seccomp jails.
Here's below a tentative RFC, the code is actually trivial, if you
could help reviewing/improving the commit header it would be great.
Thanks,
Andrea
From 3f7adb783262dc7f4e71cdbf07b4ef9f6b8d3ed9 Mon Sep 17 00:00:00 2001
From: Andrea Arcangeli <[email protected]>
Date: Wed, 4 Nov 2020 15:20:33 -0500
Subject: [PATCH 1/1] x86: change default to spec_store_bypass_disable=prctl
spectre_v2_user=prctl
Switch the kernel default of SSBD and STIBP to the ones with
CONFIG_SECCOMP=n (i.e. spec_store_bypass_disable=prctl
spectre_v2_user=prctl) even if CONFIG_SECCOMP=y.
Several motivations listed below:
- If SMT is enabled the seccomp jail can still attack the rest of the
system even with spectre_v2_user=seccomp by using MDS-HT (except on
XEON PHI where MDS can be tamed with SMT left enabled, but that's a
special case). Setting STIBP become a very expensive window dressing
after MDS-HT was discovered.
- The seccomp jail cannot attack the kernel with spectre-v2-HT
regardless (even if STIBP is not set), but with MDS-HT the seccomp
jail can attack the kernel too.
- With spec_store_bypass_disable=prctl the seccomp jail can attack the
other userland (guest or host mode) using spectre-v2-HT, but the
userland attack is already mitigated by both ASLR and pid namespaces
for host userland and through virt isolation with libkrun or
kata. (if something if somebody is worried about spectre-v2-HT it's
best to mount proc with hidepid=2,gid=proc on workstations where not
all apps may run under container runtimes, rather than slowing down
all seccomp jails, but the best is to add pid namespaces to the
seccomp jail). As opposed MDS-HT is not mitigated and the seccomp
jail can still attack all other host and guest userland if SMT is
enabled even with spec_store_bypass_disable=seccomp.
- If full security is required then MDS-HT must also be mitigated with
nosmt and then spectre_v2_user=prctl and spectre_v2_user=seccomp
would become identical.
- Setting spectre_v2_user=seccomp is overall lower priority than to
setting javascript.options.wasm false in about:config to protect
against remote wasm MDS-HT, instead of worrying about Spectre-v2-HT
and STIBP which again is already statistically well mitigated by
other means in userland and it's fully mitigated in kernel with
retpolines (unlike the wasm assist call with MDS-HT).
- SSBD is needed to prevent reading the JIT memory and the primary
user being the OpenJDK. However the primary user of SSBD wouldn't be
covered by spec_store_bypass_disable=seccomp because it doesn't use
seccomp and the primary user also explicitly declined to set
PR_SET_SPECULATION_CTRL+PR_SPEC_STORE_BYPASS despite it easily
could. In fact it would need to set it only when the sandboxing
mechanism is enabled for javaws applets, but it still declined it by
declaring security within the same user address space as an
untenable objective for their JIT, even in the sandboxing case where
performance would be a lesser concern (for the record: I kind of
disagree in not setting PR_SPEC_STORE_BYPASS in the sandbox case and
I prefer to run javaws through a wrapper that sets
PR_SPEC_STORE_BYPASS if I need). In turn it can be inferred that
even if the primary user of SSBD would use seccomp, they would
invoke it with SECCOMP_FILTER_FLAG_SPEC_ALLOW by now.
- runc/crun already set SECCOMP_FILTER_FLAG_SPEC_ALLOW by default, k8s
and podman have a default json seccomp allowlist that cannot be
slowed down, so for the #1 seccomp user this change is already a
noop.
- systemd/sshd or other apps that use seccomp, if they really need
STIBP or SSBD, they need to explicitly set the
PR_SET_SPECULATION_CTRL by now. The stibp/ssbd seccomp blind
catch-all approach was done probably initially with a wishful
thinking objective to pretend to have a peace of mind that it could
magically fix it all. That was wishful thinking before MDS-HT was
discovered, but after MDS-HT has been discovered it become just
window dressing.
- For qemu "-sandbox" seccomp jail it wouldn't make sense to set STIBP
or SSBD. SSBD doesn't help with KVM because there's no JIT (if it's
needed with TCG it should be an opt-in with
PR_SET_SPECULATION_CTRL+PR_SPEC_STORE_BYPASS and it shouldn't
slowdown KVM for nothing). For qemu+KVM STIBP would be even more
window dressing than it is for all other apps, because in the
qemu+KVM case there's not only the MDS attack to worry about with
SMT enabled. Even after disabling SMT, there's still a theoretical
spectre-v2 attack possible within the same thread context from guest
mode to host ring3 that the host kernel retpoline mitigation has no
theoretical chance to mitigate. On some kernels a
ibrs-always/ibrs-retpoline opt-in model is provided that will
enabled IBRS in the qemu host ring3 userland which fixes this
theoretical concern. Only after enabling IBRS in the host userland
it would then make sense to proceed and worry about STIBP and an
attack on the other host userland, but then again SMT would need to
be disabled for full security anyway, so that would render STIBP
again a noop.
- last but not the least: the lack of "spec_store_bypass_disable=prctl
spectre_v2_user=prctl" means the moment a guest boots and
sshd/systemd runs, the guest kernel will write to SPEC_CTRL MSR
which will make the guest vmexit forever slower, forcing KVM to
issue a very slow rdmsr instruction at every vmexit. So the end
result is that SPEC_CTRL MSR is only available in GCE. Most other
public cloud providers don't expose SPEC_CTRL, which means that not
only STIBP/SSBD isn't available, but IBPB isn't available either
(which would cause no overhead to the guest or the hypervisor
because it's write only and requires no reading during vmexit). So
the current default already net loss in security (missing IBPB)
which means most public cloud providers cannot achieve a fully
secure guest with nosmt (and nosmt is enough to fully mitigate
MDS-HT). It also means GCE and is unfairly penalized in performance
because it provides the option to enable full security in the guest
as an opt-in (i.e. nosmt and IBPB). So this change will allow all
cloud providers to expose SPEC_CTRL without incurring into any
hypervisor slowdown and at the same time it will remove the unfair
penalization of GCE performance for doing the right thing and it'll
allow to get full security with nosmt with IBPB being available (and
STIBP becoming meaningless).
Example to put things in prospective: the STIBP enabled in seccomp has
never been about protecting apps using seccomp like sshd from an
attack from a malicious userland, but to the contrary it has always
been about protecting the system from an attack from sshd, after a
successful remote network exploit against sshd. In fact initially it
wasn't obvious STIBP would work both ways (STIBP was about preventing
the task that runs with STIBP to be attacked with spectre-v2-HT, but
accidentally in the STIBP case it also prevents the attack in the
other direction). In the hypothetical case that sshd has been remotely
exploited the last concern should be STIBP being set, because it'll be
still possible to obtain info even from the kernel by using MDS if
nosmt wasn't set (and if it was set, STIBP is a noop in the first
place). As opposed kernel cannot leak anything with spectre-v2 HT
because of retpolines and the userland is mitigated by ASLR already
and ideally PID namespaces too. If something it'd be worth checking if
sshd run the seccomp thread under pid namespaces too if available in
the running kernel. SSBD also would be a noop for sshd, since sshd
uses no JIT. If sshd prefers to keep doing the STIBP window dressing
exercise, it still can even after this change of defaults by opting-in
with PR_SPEC_STORE_BYPASS.
Ultimately setting SSBD and STIBP by default for all seccomp jails is
a bad sweet spot and bad default with more cons than pros that end up
reducing security in the public cloud (by giving an huge incentive to
not expose SPEC_CTRL which would be needed to get full security with
IBPB after setting nosmt in the guest) and by excessively hurting
performance to more secure apps using seccomp that end up having to
opt out with SECCOMP_FILTER_FLAG_SPEC_ALLOW.
The following is the verified result of the new default with SMT
enabled:
(gdb) print spectre_v2_user_stibp
$1 = SPECTRE_V2_USER_PRCTL
(gdb) print spectre_v2_user_ibpb
$2 = SPECTRE_V2_USER_PRCTL
(gdb) print ssb_mode
$3 = SPEC_STORE_BYPASS_PRCTL
Signed-off-by: Andrea Arcangeli <[email protected]>
---
Documentation/admin-guide/kernel-parameters.txt | 5 ++---
arch/x86/kernel/cpu/bugs.c | 4 ++--
2 files changed, 4 insertions(+), 5 deletions(-)
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index 526d65d8573a..105401a3582f 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -4980,8 +4980,7 @@
auto - Kernel selects the mitigation depending on
the available CPU features and vulnerability.
- Default mitigation:
- If CONFIG_SECCOMP=y then "seccomp", otherwise "prctl"
+ Default mitigation: "prctl"
Not specifying this option is equivalent to
spectre_v2_user=auto.
@@ -5025,7 +5024,7 @@
will disable SSB unless they explicitly opt out.
Default mitigations:
- X86: If CONFIG_SECCOMP=y "seccomp", otherwise "prctl"
+ X86: "prctl"
On powerpc the options are:
diff --git a/arch/x86/kernel/cpu/bugs.c b/arch/x86/kernel/cpu/bugs.c
index d3f0db463f96..5ec39397fe9c 100644
--- a/arch/x86/kernel/cpu/bugs.c
+++ b/arch/x86/kernel/cpu/bugs.c
@@ -721,11 +721,11 @@ spectre_v2_user_select_mitigation(enum spectre_v2_mitigation_cmd v2_cmd)
case SPECTRE_V2_USER_CMD_FORCE:
mode = SPECTRE_V2_USER_STRICT;
break;
+ case SPECTRE_V2_USER_CMD_AUTO:
case SPECTRE_V2_USER_CMD_PRCTL:
case SPECTRE_V2_USER_CMD_PRCTL_IBPB:
mode = SPECTRE_V2_USER_PRCTL;
break;
- case SPECTRE_V2_USER_CMD_AUTO:
case SPECTRE_V2_USER_CMD_SECCOMP:
case SPECTRE_V2_USER_CMD_SECCOMP_IBPB:
if (IS_ENABLED(CONFIG_SECCOMP))
@@ -1132,7 +1132,6 @@ static enum ssb_mitigation __init __ssb_select_mitigation(void)
return mode;
switch (cmd) {
- case SPEC_STORE_BYPASS_CMD_AUTO:
case SPEC_STORE_BYPASS_CMD_SECCOMP:
/*
* Choose prctl+seccomp as the default mode if seccomp is
@@ -1146,6 +1145,7 @@ static enum ssb_mitigation __init __ssb_select_mitigation(void)
case SPEC_STORE_BYPASS_CMD_ON:
mode = SPEC_STORE_BYPASS_DISABLE;
break;
+ case SPEC_STORE_BYPASS_CMD_AUTO:
case SPEC_STORE_BYPASS_CMD_PRCTL:
mode = SPEC_STORE_BYPASS_PRCTL;
break;
On Wed, Nov 04, 2020 at 04:57:02PM -0500, Andrea Arcangeli wrote:
> Switch the kernel default of SSBD and STIBP to the ones with
> CONFIG_SECCOMP=n (i.e. spec_store_bypass_disable=prctl
> spectre_v2_user=prctl) even if CONFIG_SECCOMP=y.
Agreed. I think this is the right time to flip this switch. I agree with
the (very well described) rationales. :)
Fundamentally, likely everyone who is interested in manipulating the
mitigations are doing so now, and it doesn't make sense (on many fronts)
to tie some to seccomp mode any more (which was intended as a temporary
defense to gain coverage while sysadmins absorbed what the best
practices should be).
Thanks for sending this!
Acked-by: Kees Cook <[email protected]>
--
Kees Cook
On Thu, Nov 05, 2020 at 12:22:29AM +0100, Thomas Gleixner wrote:
> On Wed, Nov 04 2020 at 16:57, Andrea Arcangeli wrote:
> > ---
> > Documentation/admin-guide/kernel-parameters.txt | 5 ++---
>
> Is Documentation/admin-guide/hw-vuln/* still correct? If not, please
> fix that as well.
Right, I missed two seccomp mention that needed removing there too.
Also I noticed below I intended PR_SPEC_INDIRECT_BRANCH
(PR_SPEC_STORE_BYPASS there is no point to even mention it as a
possibility to be considered), so I corrected it.
==
uses no JIT. If sshd prefers to keep doing the STIBP window dressing
exercise, it still can even after this change of defaults by opting-in
with PR_SPEC_STORE_BYPASS.
==
> > >with PR_SPEC_INDIRECT_BRANCH.
> Aside of that please send patches in the proper format so they do not
> need manual interaction when picking them up.
This was a RFC per subject since I expected it wouldn't be final, but
I added Kees' Acked-by and I'll submit it now.
Thanks,
Andrea
On Wed, Nov 04 2020 at 16:57, Andrea Arcangeli wrote:
> ---
> Documentation/admin-guide/kernel-parameters.txt | 5 ++---
Is Documentation/admin-guide/hw-vuln/* still correct? If not, please
fix that as well.
Aside of that please send patches in the proper format so they do not
need manual interaction when picking them up.
Thanks,
tglx
Hello,
Could you help checking if this incremental doc cleanup is possible?
After the previous patch is applied, there's still a leftover mention
of seccomp that should be removed in a duped bit of documentation, so
I tentatively referred the original documentation already updated in
sync, instead of keeping the dup around and applying the same update
to the dup.
Note: as far as I can tell spec_store_bypass_disable= documentation is
not duplicated in spectre.rst, that's better in my view. The more dups
we have the more one goes out of sync..
Andrea Arcangeli (1):
x86: deduplicate the spectre_v2_user documentation
Documentation/admin-guide/hw-vuln/spectre.rst | 51 +------------------
1 file changed, 2 insertions(+), 49 deletions(-)
Is anything happening with this proposal? Is there anything I could do
to help it along?
My personal motivation is that I'm involved in developing and using the
[Shadow] simulator, which we use to run hours and days long simulations.
We're currently looking into running some simulations in gitlab CI
Docker runner to take advantage of shared hardware, but Docker currently
doesn't expose a way to opt out of these mitigations without turning off
seccomp altogether [Docker FR].
I've measured these mitigations to cause simulations to take 50% longer
[Overhead], so I'm pretty motivated to find a way to disable them :).
[Shadow]: https://shadow.github.io/
[Docker FR]: https://github.com/moby/moby/issues/42619
[Overhead]:
https://github.com/shadow/shadow/issues/1489#issuecomment-871445482
P.S. Attempting to respond to a thread without actually being subscribed
to the list; sorry if this ends up not threading correctly. The CC
header was truncated so also some original recipients have been dropped.
Original thread: https://lkml.org/lkml/2020/11/4/1135