PT-2026-128176
CVE-2026-99183
·
Published
2026-10-28
·
Updated
2026-10-28
Description by
Score
N/A
Not availableNo severity ratings or metrics are available. When they are, we'll update the corresponding info on the page.
In the Linux kernel, the following vulnerability has been resolved:
accel/rocket: search every core slot when looking up a scheduler
sched to core() walks rdev->cores[] up to rdev->num cores, and
rocket remove() decrements num cores for every core it removes. Unbind a
core that is not the last one and the cores behind it fall outside the
search, so sched to core() returns NULL for a core that is still bound and
still running jobs. Neither caller checks the result:
rocket job run(): rocket fence create(core), core->dev
rocket job timedout(): dev err(core->dev, "NPU job timed out")
Unbinding the middle core of the three on an RK3588 while three clients are
submitting to all of them faults twice, once from the surviving core's
job queue and once from its reset work:
KASAN: null-ptr-deref in range [0x0000000000000220-0x0000000000000227]
Workqueue: fdad0000.npu drm sched run job work [gpu sched]
pc : rocket job run+0x234/0x838 [rocket]
Call trace:
rocket job run+0x234/0x838 [rocket]
drm sched run job work+0x2cc/0xad8 [gpu sched]
process one work+0x640/0x14f0
KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
Workqueue: rocket-reset-2 drm sched job timedout [gpu sched]
pc : rocket job timedout+0xf0/0x1e0 [rocket]
Call trace:
rocket job timedout+0xf0/0x1e0 [rocket]
drm sched job timedout+0x188/0x6a0 [gpu sched]
Both are the third core: the workqueue names are its device and its
core->index, and it was left at slot 2 while num cores had dropped to 2.
Search all the slots that were allocated, the way find core for dev() now
does. A core that is still bound is then found, and the two callers get
the pointer they already assume they have.
This does not make unbinding one core out of several safe. An open client
keeps an entity pointing at the scheduler of the core that went away:
drm sched reports it as not ready for every job that lands on it, and the
client waits in dma fence default wait for a fence that will never signal.
Stopping the NULL dereference is what belongs in a fix; the rest wants
more thought.
Found an issue in the description? Have something to add? Feel free to write us 👾