On 2020-11-26 13:55, stf_xl@xxxxx wrote:
> From: Stanislaw Gruszka <stf_xl@xxxxx>
>
> Currently 'while (q->queued > 0)' loop was removed from mt76u_stop_tx()
> code. This causes crash on device removal as we try to cleanup empty
> queue:
>
> [ 96.495571] kernel BUG at include/linux/skbuff.h:2297!
> [ 96.498983] invalid opcode: 0000 [#1] SMP PTI
> [ 96.501162] CPU: 3 PID: 27 Comm: kworker/3:0 Not tainted 5.10.0-rc5+ #11
> [ 96.502754] Hardware name: LENOVO 20DGS08H00/20DGS08H00, BIOS J5ET48WW (1.19 ) 08/27/2015
> [ 96.504378] Workqueue: usb_hub_wq hub_event
> [ 96.505983] RIP: 0010:skb_pull+0x2d/0x30
> [ 96.507576] Code: 00 00 8b 47 70 39 c6 77 1e 29 f0 89 47 70 3b 47 74 72 17 48 8b 87 c8 00 00 00 89 f6 48 01 f0 48 89 87 c8 00 00 00 c3 31 c0 c3 <0f> 0b 90 0f 1f 44 00 00 53 48 89 fb 48 8b bf c8 00 00 00 8b 43 70
> [ 96.509296] RSP: 0018:ffffb11b801639b8 EFLAGS: 00010287
> [ 96.511038] RAX: 000000001c6939ed RBX: ffffb11b801639f8 RCX: 0000000000000000
> [ 96.512964] RDX: ffffb11b801639f8 RSI: 0000000000000018 RDI: ffff90c64e4fb800
> [ 96.514710] RBP: ffff90c654551ee0 R08: ffff90c652bce7a8 R09: ffffb11b80163728
> [ 96.516450] R10: 0000000000000001 R11: 0000000000000001 R12: ffff90c64e4fb800
> [ 96.519749] R13: 0000000000000010 R14: 0000000000000020 R15: ffff90c64e352ce8
> [ 96.523455] FS: 0000000000000000(0000) GS:ffff90c96eec0000(0000) knlGS:0000000000000000
> [ 96.527171] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 96.530900] CR2: 0000242556f18288 CR3: 0000000146a10002 CR4: 00000000003706e0
> [ 96.534678] Call Trace:
> [ 96.538418] mt76x02u_tx_complete_skb+0x1f/0x50 [mt76x02_usb]
> [ 96.542231] mt76_queue_tx_complete+0x23/0x50 [mt76]
> [ 96.546028] mt76u_stop_tx.cold+0x71/0xa2 [mt76_usb]
> [ 96.549797] mt76x0u_stop+0x2f/0x90 [mt76x0u]
> [ 96.553638] drv_stop+0x33/0xd0 [mac80211]
> [ 96.557449] ieee80211_do_stop+0x558/0x860 [mac80211]
> [ 96.561262] ? dev_deactivate_many+0x298/0x2d0
> [ 96.565101] ieee80211_stop+0x16/0x20 [mac80211]
>
> Fix that by adding while loop again. We need loop, not just single
> check, to clean all pending entries.
>
> Additionally move mt76_worker_disable/enable after !mt76_has_tx_pending()
> as we want to tx_worker to run to process tx queues, while we wait for
> exactly that.
>
> I was a bit worried about accessing q->queued without lock, but
> mt76_worker_disable() -> kthread_park() should assure this value will
> be seen updated on other cpus.
>
> Fixes: fe5b5ab52e9d ("mt76: unify queue tx cleanup code")
> Signed-off-by: Stanislaw Gruszka <stf_xl@xxxxx>
Acked-by: Felix Fietkau <nbd@xxxxxxxx>
Thanks,
- Felix
