UBUNTU-CVE-2026-99184: Medium severity vulnerability
In the Linux kernel, the following vulnerability has been resolved: accel/rocket: search every core slot when a core is removed rocketremove() decrements rdev->numcores for each core it removes, while findcorefordev() searches slots 0 to numcores - 1. Unbinding the cores in the order they were bound therefore loses the last one: by the time it is removed the search range has already shrunk past its slot, so findcorefordev() returns -1 and rocketremove() gives up without doing anything. numcores never reaches zero, rocketdevicefini() never runs, and the file-scoped rdev keeps pointing at a device that is going away. Binding the cores again starts from that stale count, because rocketprobe() takes rdev->numcores as the slot to fill. On an RK3588, which describes three cores, the second round lands on slots 1, 2 and 3 while rdev->cores was allocated with room for three: rocket fdab0000.npu: Rockchip NPU core 1 version: 1179210309 rocket fdac0000.npu: Rockchip NPU core 2 version: 1179210309 rocket fdad0000.npu: Rockchip NPU core 3 version: 1179210309 The write to rdev->cores[3] is past the end of the array. Nothing in tree reads the core array often enough to notice, so the overrun is silent today. It turned up while testing a devfreq series on top of this, where a worker walks every core a few times a second, and UBSAN caught the first bool it read out of the overrun entry: UBSAN: invalid-load in drivers/accel/rocket/rocketdevfreq.c:47:10 load of value 5 is not a valid value for type 'Bool' Workqueue: devfreqwq devfreqmonitor Record how many slots were allocated and search all of them. Every core is then found on removal, numcores reaches zero, the device is torn down and a later bind starts from a clean rdev. This does not make unbinding a single core out of several work. probe still takes numcores as the slot to fill, so rebinding one core while its siblings stay bound would write over a slot that is already in use, and rocketopen() still reaches for cores[0] whether or not anything is there. Both of those want more thought than a fix should carry. Found by unbinding and rebinding all three cores on an Orange Pi 5 Plus. With this applied, 25 unbind/rebind rounds and 5 module unload/reload rounds run clean there: the cores land in slots 0, 1 and 2 every time, whichever order they are bound in, and the shared supply goes back to a single user after each round.
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Compensating control
Record how many core slots were allocated and search all allocated slots when removing a core, rather than limiting the search to slots 0 through num_cores - 1.
Event History
Frequently Asked Questions
What systems are exposed to this issue?
Systems using the Linux kernel rocket accelerator driver are exposed when Rocket NPU cores are removed and later bound again. The reported example is an RK3588 system describing three cores.
What sequence triggers the out-of-bounds write?
Removing cores in the same order they were bound can leave the driver with a stale core count and prevent device cleanup. Binding the cores again then uses that stale count as the next slot, which can write beyond the allocated core array.
What can be done if an updated kernel is not immediately available?
Avoid unbinding and rebinding Rocket NPU cores, particularly repeated core removal and reprobe cycles. A restart that prevents reuse of the stale driver state may avoid the described second binding sequence, but the provided information does not specify a formal workaround.
How can administrators recognize the affected rebind behavior?
On the reported three-core RK3588 configuration, kernel messages during a second bind showed cores numbered 1, 2, and 3 rather than starting at 0. The core numbered 3 corresponds to a write past an array allocated for three entries.