line6_buffer_release() frees the stream buffer after line6_wait_clear_audio_urbs() gave up waiting for the URBs that are still active. The buffer is still referenced by those URBs, and the completion callback of such an URB resubmits it from submit_audio_out_urb() with line6pcm->out.buffer already set to NULL, which ends up in a NULL pointer dereference in the memset() in submit_audio_out_urb(). While the URBs remain queued, the host controller can also write the received isochronous data into the freed buffer. Such a write does not go through the KASAN checks, so it silently corrupts the freed slab object; the corruption was observed as a bogus KASAN quarantine entry that later led to a general protection fault in qlist_free_all(). The active_urbs bitmap cannot prove that the URBs are gone: line6_unlink_audio_urbs() is called outside the stream lock and only marks the URBs that are active at that moment, so an URB resubmitted by a completion callback right after the scan is never unlinked and keeps being resubmitted after the stream was stopped. Set the unlink bit for every currently active URB under the stream lock, so that the completion callbacks cannot resubmit them, and then kill all URBs synchronously. usb_kill_urb() waits for the giveback, hence the buffer is no longer referenced by any URB when it is freed. Drop the unlink marks afterwards: line6_unlink_audio_urbs() can leave a stale mark on an URB that completed and was not resubmitted, and such a mark makes line6_stream_start() skip submitting the URBs. Fixes: 63e20df1e5b2 ("ALSA: line6: Reorganize PCM stream handling") Cc: stable@vger.kernel.org Co-developed-by: Bin Yu Signed-off-by: Bin Yu Co-developed-by: Cheng Wen Signed-off-by: Cheng Wen Signed-off-by: YuanZhe Liu <25031212351@stu.xidian.edu.cn> --- v3: - Kill the URBs unconditionally instead of rechecking active_urbs outside the stream lock, and clear the unlink marks after the kills. v2: - Rebased on sound/for-next. v1 was generated against v7.3.1 and no longer applied, because line6_buffer_release() gained the prev_fbuf clearing block in the meantime. No code change. --- sound/usb/line6/pcm.c | 32 +++++++++++++++++++++++++++++++- 1 file changed, 31 insertions(+), 1 deletion(-) diff --git a/sound/usb/line6/pcm.c b/sound/usb/line6/pcm.c index d17baf7761c6..4be3b41a98a2 100644 --- a/sound/usb/line6/pcm.c +++ b/sound/usb/line6/pcm.c @@ -134,6 +134,36 @@ static void line6_wait_clear_audio_urbs(struct snd_line6_pcm *line6pcm, "timeout: still %d active urbs..\n", alive); } +/* + * Mark all currently active URBs for unlinking and kill them. + * + * The unlink bit is set under the stream lock, so the completion + * callbacks cannot resubmit these URBs any longer. usb_kill_urb() then + * waits for the giveback of every URB, hence the stream buffer is no + * longer referenced by the host controller. The unlink marks are + * cleared afterwards, so that a later stream start can submit the URBs + * again. This function must be called from a sleepable context. + */ +static void line6_kill_audio_urbs(struct snd_line6_pcm *line6pcm, + struct line6_pcm_stream *pcms) +{ + int i; + + scoped_guard(spinlock_irqsave, &pcms->lock) { + for (i = 0; i < line6pcm->line6->iso_buffers; i++) + if (test_bit(i, &pcms->active_urbs)) + set_bit(i, &pcms->unlink_urbs); + } + + for (i = 0; i < line6pcm->line6->iso_buffers; i++) + if (pcms->urbs[i]) + usb_kill_urb(pcms->urbs[i]); + + scoped_guard(spinlock_irqsave, &pcms->lock) { + pcms->unlink_urbs = 0; + } +} + static inline struct line6_pcm_stream * get_stream(struct snd_line6_pcm *line6pcm, int direction) { @@ -172,7 +202,7 @@ static void line6_buffer_release(struct snd_line6_pcm *line6pcm, { clear_bit(type, &pstr->opened); if (!pstr->opened) { - line6_wait_clear_audio_urbs(line6pcm, pstr); + line6_kill_audio_urbs(line6pcm, pstr); if (pstr == &line6pcm->in) { /* prev_fbuf points into the capture buffer we are about * to free; drop it while no capture URB is in flight. -- 2.34.1