CVE-2020-10708 kernel: race condition in kernel/audit.c may allow low privilege users trigger kernel panic

Hello,

Since this is public...no harm dropping a copy over here. My thoughts are that there is a race here. But since starting/stopping the audit daemon requires root privs and as root you can do worse things. I don't know if this is fixable or working per design.

---------- Forwarded Message ----------

Subject: [oss-security] CVE-2020-10708 kernel: race condition in kernel/ audit.c may allow low privilege users trigger kernel panic
Date: Friday, April 17, 2020, 12:40:10 AM EDT
From: 陈伟宸(田各) <[email protected]>
To: oss-security <[email protected]>

"A race condition was found in the Linux kernel audit subsystem. When the system is configured to panic on events being dropped, an attacker who is able to trigger an audit event that starts while auditd is in the process of starting may be able to cause the system to panic by exploiting a race condition in audit event handling. This creates a denial of service by causing a panic."

https://bugzilla.redhat.com/show_bug.cgi?id=1822593

Env:
Red Hat Enterprise Linux Server release 7.7 (Maipo) 3.10.0-1062.12.1.el7.x86_64

Details:
Function audit_log_end and audit_panic may have race conditions when auditd is restarting because audit_pid can be NULL in audit_log_end and then become not NULL in audit_panic, which may allow attackers to trigger kernel panic. Here is panic call stack:

void audit_log_end(struct audit_buffer *ab)

{
    if (!ab)
        return;
    if (!audit_rate_check()) {
        audit_log_lost("rate limit exceeded");
    } else {
        struct nlmsghdr *nlh = nlmsg_hdr(ab->skb);
        nlh->nlmsg_len = ab->skb->len - NLMSG_HDRLEN;

        if (audit_pid) {
            skb_queue_tail(&audit_skb_queue, ab->skb);
            wake_up_interruptible(&kauditd_wait);
        } else {
            audit_printk_skb(ab->skb); // <- audit_pid == NULL when auditd is
killed
        }
        ab->skb = NULL;
    }
    audit_buffer_free(ab);
}
-> audit_printk_skb -> audit_log_lost ->
void audit_panic(const char *message)
{
    switch (audit_failure)
    {
    case AUDIT_FAIL_SILENT:
        break;
    case AUDIT_FAIL_PRINTK:
        if (printk_ratelimit())
            printk(KERN_ERR "audit: %s\n", message);
        break;
    case AUDIT_FAIL_PANIC:
        /* test audit_pid since printk is always losey, why bother? */
        if (audit_pid) // <- audit_pid not NULL because auditd is restarting
            panic("audit: %s\n", message);
        break;
    }
}

How to reproduce：
1. set audit-failure to AUDIT_FAIL_PANIC(2) and add a random audit rule like:
[root@test ~]# cat /etc/audit/rules.d/audit.rules

(quoted diff) +0 -4

-D
-b 8192
-f 2
-w /etc/hosts -p rwa -k hosts

2. keep killing auditd and then starting auditd, for example:
while true; do ps aux | grep "/sbin/auditd" | grep -v "grep" | awk '{print $2}' | xargs kill; service auditd start; systemctl reset-failed auditd.service; done
3. log in a low privilege user and keep reading /etc/hosts, for example:
while true; do cat /etc/hosts > /dev/null; done
4. kernel panic will happen within several minutes

Thanks.

Quoted 6 lines · Steve Grubb

Hello,

Since this is public...no harm dropping a copy over here. My thoughts are that there is a race here. But since starting/stopping the audit daemon requires root privs and as root you can do worse things. I don't know if this is fixable or working per design.

The first thing to note is that the code snippets below are not from a modern upstream kernel; I'm guessing it's a RHEL-7.x based kernel given the problem report.

The relevant code changed substantially
when I reworked how the audit queue was managed; it's been a while now so I can't say with 100% confidence that this has been sorted out, but I believe the current upstream code should handle auditd restarts within the bounds of the record queue.

Ultimately if you configure the system to panic when you start dropping audit records, and you stop the audit daemon for a period of time such that the queues start to overflow, the system will panic. That's what you have configured it to do :)

It is also important to note that the panic-on-overflow behavior is not the default, the sysadmin must explicitly configure the system to behave in this way.

Quoted 84 lines

---------- Forwarded Message ----------

Subject: [oss-security] CVE-2020-10708 kernel: race condition in kernel/ audit.c may allow low privilege users trigger kernel panic
Date: Friday, April 17, 2020, 12:40:10 AM EDT
From: 陈伟宸(田各) <[email protected]>
To: oss-security <[email protected]>

"A race condition was found in the Linux kernel audit subsystem. When the system is configured to panic on events being dropped, an attacker who is able to trigger an audit event that starts while auditd is in the process of starting may be able to cause the system to panic by exploiting a race condition in audit event handling. This creates a denial of service by causing a panic."

https://bugzilla.redhat.com/show_bug.cgi?id=1822593

Env:
Red Hat Enterprise Linux Server release 7.7 (Maipo) 3.10.0-1062.12.1.el7.x86_64

Details:
Function audit_log_end and audit_panic may have race conditions when auditd is restarting because audit_pid can be NULL in audit_log_end and then become not NULL in audit_panic, which may allow attackers to trigger kernel panic. Here is panic call stack:

void audit_log_end(struct audit_buffer *ab)

{
    if (!ab)
        return;
    if (!audit_rate_check()) {
        audit_log_lost("rate limit exceeded");
    } else {
        struct nlmsghdr *nlh = nlmsg_hdr(ab->skb);
        nlh->nlmsg_len = ab->skb->len - NLMSG_HDRLEN;

        if (audit_pid) {
            skb_queue_tail(&audit_skb_queue, ab->skb);
            wake_up_interruptible(&kauditd_wait);
        } else {
            audit_printk_skb(ab->skb); // <- audit_pid == NULL when auditd is
killed
        }
        ab->skb = NULL;
    }
    audit_buffer_free(ab);
}
-> audit_printk_skb -> audit_log_lost ->
void audit_panic(const char *message)
{
    switch (audit_failure)
    {
    case AUDIT_FAIL_SILENT:
        break;
    case AUDIT_FAIL_PRINTK:
        if (printk_ratelimit())
            printk(KERN_ERR "audit: %s\n", message);
        break;
    case AUDIT_FAIL_PANIC:
        /* test audit_pid since printk is always losey, why bother? */
        if (audit_pid) // <- audit_pid not NULL because auditd is restarting
            panic("audit: %s\n", message);
        break;
    }
}

How to reproduce：
1. set audit-failure to AUDIT_FAIL_PANIC(2) and add a random audit rule like:
[root@test ~]# cat /etc/audit/rules.d/audit.rules

(quoted diff) +0 -4

-D
-b 8192
-f 2
-w /etc/hosts -p rwa -k hosts

2. keep killing auditd and then starting auditd, for example:
while true; do ps aux | grep "/sbin/auditd" | grep -v "grep" | awk '{print $2}' | xargs kill; service auditd start; systemctl reset-failed auditd.service; done
3. log in a low privilege user and keep reading /etc/hosts, for example:
while true; do cat /etc/hosts > /dev/null; done
4. kernel panic will happen within several minutes

Quoted 1 line

Thanks.

paul moore