net/sched: taprio: avoid disabling offload when it was never enabled
Linux
Affected versions
| Product | Affected | Fixed in |
|---|---|---|
| Linux Linux |
| No fix yet |
Linux
| Product | Affected | Fixed in |
|---|---|---|
| Linux Linux |
| No fix yet |
| Linux Linux |
|
|
In the Linux kernel, the following vulnerability has been resolved:
net/sched: taprio: avoid disabling offload when it was never enabled
In an incredibly strange API design decision, qdisc->destroy() gets called even if qdisc->init() never succeeded, not exclusively since commit 87b60cfacf9f ("net_sched: fix error recovery at qdisc creation"), but apparently also earlier (in the case of qdisc_create_dflt()).
The taprio qdisc does not fully acknowledge this when it attempts full offload, because it starts off with q->flags = TAPRIO_FLAGS_INVALID in taprio_init(), then it replaces q->flags with TCA_TAPRIO_ATTR_FLAGS parsed from netlink (in taprio_change(), tail called from taprio_init()).
But in taprio_destroy(), we call taprio_disable_offload(), and this determines what to do based on FULL_OFFLOAD_IS_ENABLED(q->flags).
But looking at the implementation of FULL_OFFLOAD_IS_ENABLED() (a bitwise check of bit 1 in q->flags), it is invalid to call this macro on q->flags when it contains TAPRIO_FLAGS_INVALID, because that is set to U32_MAX, and therefore FULL_OFFLOAD_IS_ENABLED() will return true on an invalid set of flags.
As a result, it is possible to crash the kernel if user space forces an error between setting q->flags = TAPRIO_FLAGS_INVALID, and the calling of taprio_enable_offload(). This is because drivers do not expect the offload to be disabled when it was never enabled.
The error that we force here is to attach taprio as a non-root qdisc, but instead as child of an mqprio root qdisc:
From the CVE record · Linux
The Linux kernel's taprio scheduler can try to disable hardware offload even when offload was never enabled. Local user space that can configure qdiscs can trigger a kernel crash by causing taprio initialization to fail before offload is enabled. No user interaction is described.
Engineers running the Linux kernel should check their exposure; affected versions are not specified.