CVE-2026-74691: net: thunderbolt: Tear down DMA paths before stopping the rings

In the Linux kernel, the following vulnerability has been resolved:

net: thunderbolt: Tear down DMA paths before stopping the rings

tbnetteardown() stops both rings and frees their frame buffers before calling tbxdomaindisablepaths(). tbringstop() zeroes the ring's descriptor base and tbnetfreebuffers() unmaps and frees the pages the frames sit in, so by the time tbpathdeactivatehop() polls the hop's 'pending' bit, anything still in flight has nowhere to drain to.

The teardown sequence has been in this order since the driver was added. The setup path has not: commit ff7cd07f3064 ("net: thunderbolt: Enable DMA paths only after rings are enabled") moved the path enable to the end of tbnetconnectedwork() and documented why:

/ Both logins successful so enable the rings, high-speed DMA paths and start the network device queue. Note we enable the DMA paths last to make sure we have primed the Rx ring before any incoming packets are allowed to arrive. /

Teardown was never updated to match, so the rings and the paths now come down in the same order they go up instead of in reverse.

On an ASMedia ASM4242 host router the 'pending' bit then never clears: every teardown burns the full 500 ms timeout and tbpathdeactivatehop() returns -ETIMEDOUT. Raising the timeout to 5 s does not help, so the hop is not slow to drain, it never drains at all.

The failure is invisible above the thunderbolt core. tbpathdeactivatehops() is void and only calls tbportwarn(); tbpathdeactivate(), tbtunneldeactivate() and tbdisconnectxdomainpaths() are void as well, and tbdisconnectxdomainpaths() ends in an unconditional "return 0". So tbxdomaindisablepaths() reports success and the netdevwarn() below it never fires. Repeated teardowns eventually take the XDomain control channel down, after which the peer node is gone and only a power cycle brings the controller back.

Deactivating the paths first fixes it. Measured with kretprobes on a stock v6.17 tree with no other patches applied, on a link that was up and had just carried traffic:

before: tbpathdeactivatehop() returns 0 for the first hop, then -ETIMEDOUT for the second 500335 us later after: 0 for both, 525 us apart

Alternating the two orderings ABBA over three load levels, four teardowns per arm: every teardown failed before the change (21 of 21 that ran), none failed after (0 of 24). The before arms ran short because the link died partway through. The same split shows up when the interface is enslaved to a bond instead of just brought down, which is how I ran into this in the first place. Throughput and latency after the change are unchanged.

Hosts whose routers drain the hop despite the stale descriptor base see no functional difference, since the paths end up deactivated either way.