[PULL,1/2] seccomp: don't kill process for resource control syscalls

Commit Message

Eduardo Otubo March 27, 2019, 9:59 a.m. UTC

From: Daniel P. Berrangé <[e-mail]>

The Mesa library tries to set process affinity on some of its threads in
order to optimize its performance. Currently this results in QEMU being
immediately terminated when seccomp is enabled.

Mesa doesn't consider failure of the process affinity settings to be
fatal to its operation, but our seccomp policy gives it no choice in
gracefully handling this denial.

It is reasonable to consider that malicious code using the resource
control syscalls to be a less serious attack than if they were trying
to spawn processes or change UIDs and other such things. Generally
speaking changing the resource control setting will "merely" affect
quality of service of processes on the host. With this in mind, rather
than kill the process, we can relax the policy for these syscalls to
return the EPERM errno value. This allows callers to detect that QEMU
does not want them to change resource allocations, and apply some
reasonable fallback logic.

The main downside to this is for code which uses these syscalls but does
not check the return value, blindly assuming they will always
succeeed. Returning an errno could result in sub-optimal behaviour.
Arguably though such code is already broken & needs fixing regardless.

Signed-off-by: Daniel P. Berrangé <[e-mail]>
Reviewed-by: Marc-André Lureau <[e-mail]>
Acked-by: Eduardo Otubo <[e-mail]>
---
 qemu-seccomp.c | 32 +++++++++++++++++++++++++-------
 1 file changed, 25 insertions(+), 7 deletions(-)

Comments

Daniel P. Berrangé March 27, 2019, 10:03 a.m. UTC | #1

On Wed, Mar 27, 2019 at 10:59:03AM +0100, Eduardo Otubo wrote:
> From: Daniel P. Berrangé <[e-mail]>
> 
> The Mesa library tries to set process affinity on some of its threads in
> order to optimize its performance. Currently this results in QEMU being
> immediately terminated when seccomp is enabled.
> 
> Mesa doesn't consider failure of the process affinity settings to be
> fatal to its operation, but our seccomp policy gives it no choice in
> gracefully handling this denial.
> 
> It is reasonable to consider that malicious code using the resource
> control syscalls to be a less serious attack than if they were trying
> to spawn processes or change UIDs and other such things. Generally
> speaking changing the resource control setting will "merely" affect
> quality of service of processes on the host. With this in mind, rather
> than kill the process, we can relax the policy for these syscalls to
> return the EPERM errno value. This allows callers to detect that QEMU
> does not want them to change resource allocations, and apply some
> reasonable fallback logic.
> 
> The main downside to this is for code which uses these syscalls but does
> not check the return value, blindly assuming they will always
> succeeed. Returning an errno could result in sub-optimal behaviour.
> Arguably though such code is already broken & needs fixing regardless.
> 
> Signed-off-by: Daniel P. Berrangé <[e-mail]>
> Reviewed-by: Marc-André Lureau <[e-mail]>
> Acked-by: Eduardo Otubo <[e-mail]>

Normally the person sending the pull request should be adding a
Signed-off-by line, not an Acked-by line, as Acked-by doesn't
have any meaning wrt to the DCO.

IIUC, we don't really use Acked-by in QEMU. Only case I think
it would be used is where a maintainer is giving their approval
for a patch to be sent someone else's tree. eg if a seccomp patch
had to merge via a block maintainer tree for some reason, then
you could give an Acked-by to indicate you are ok with that going
via the different tree.

Regards,
Daniel

Patch

diff --git a/qemu-seccomp.c b/qemu-seccomp.c
index 36d5829831..cf520883c7 100644
--- a/qemu-seccomp.c
+++ b/qemu-seccomp.c
@@ -121,20 +121,37 @@  qemu_seccomp(unsigned int operation, unsigned int flags, void *args)
 #endif
 }
 
-static uint32_t qemu_seccomp_get_kill_action(void)
+static uint32_t qemu_seccomp_get_action(int set)
 {
+    switch (set) {
+    case QEMU_SECCOMP_SET_DEFAULT:
+    case QEMU_SECCOMP_SET_OBSOLETE:
+    case QEMU_SECCOMP_SET_PRIVILEGED:
+    case QEMU_SECCOMP_SET_SPAWN: {
 #if defined(SECCOMP_GET_ACTION_AVAIL) && defined(SCMP_ACT_KILL_PROCESS) && \
     defined(SECCOMP_RET_KILL_PROCESS)
-    {
-        uint32_t action = SECCOMP_RET_KILL_PROCESS;
+        static int kill_process = -1;
+        if (kill_process == -1) {
+            uint32_t action = SECCOMP_RET_KILL_PROCESS;
 
-        if (qemu_seccomp(SECCOMP_GET_ACTION_AVAIL, 0, &action) == 0) {
+            if (qemu_seccomp(SECCOMP_GET_ACTION_AVAIL, 0, &action) == 0) {
+                kill_process = 1;
+            }
+            kill_process = 0;
+        }
+        if (kill_process == 1) {
             return SCMP_ACT_KILL_PROCESS;
         }
-    }
 #endif
+        return SCMP_ACT_TRAP;
+    }
+
+    case QEMU_SECCOMP_SET_RESOURCECTL:
+        return SCMP_ACT_ERRNO(EPERM);
 
-    return SCMP_ACT_TRAP;
+    default:
+        g_assert_not_reached();
+    }
 }
 
 
@@ -143,7 +160,6 @@  static int seccomp_start(uint32_t seccomp_opts)
     int rc = 0;
     unsigned int i = 0;
     scmp_filter_ctx ctx;
-    uint32_t action = qemu_seccomp_get_kill_action();
 
     ctx = seccomp_init(SCMP_ACT_ALLOW);
     if (ctx == NULL) {
@@ -157,10 +173,12 @@  static int seccomp_start(uint32_t seccomp_opts)
     }
 
     for (i = 0; i < ARRAY_SIZE(blacklist); i++) {
+        uint32_t action;
         if (!(seccomp_opts & blacklist[i].set)) {
             continue;
         }
 
+        action = qemu_seccomp_get_action(blacklist[i].set);
         rc = seccomp_rule_add_array(ctx, action, blacklist[i].num,
                                     blacklist[i].narg, blacklist[i].arg_cmp);
         if (rc < 0) {