Re: [PATCH] ALSA: seq: fix use-after-free of borrowed substream in snd-seq-midi

[Takashi Iwai <tiwai@suse.de>]

On Wed, 10 Jun 2026 22:03:36 +0200,
Doruk Tan Ozturk wrote:
> 
> event_process_midi() borrows the rawmidi output substream via
> msynth->output_rfile.output and uses it through dump_midi() ->
> snd_rawmidi_kernel_write() -> snd_rawmidi_kernel_write1() without
> synchronizing against a concurrent port unsubscribe.
> 
> A concurrent UNSUBSCRIBE_PORT on the output connection runs the
> unuse callback midisynth_unuse() -> snd_rawmidi_kernel_release() ->
> close_substream() -> snd_rawmidi_runtime_free(), freeing
> substream->runtime while an in-flight event_input callback is still
> inside snd_rawmidi_kernel_write1(). The borrowed substream runtime is
> exposed to teardown before the write path takes its own buffer
> reference (snd_rawmidi_buffer_ref()), so the early derefs of
> substream->runtime / runtime->buffer read freed memory.
> 
> The buggy scenario involves two paths, with each column showing the
> order within that path:
> 
>   path A: event_input path                path B: last unuse path
>   1. event_process_midi() reads           1. midisynth_unuse() runs on the
>      msynth->output_rfile.output.            last UNSUBSCRIBE_PORT.
>   2. snd_rawmidi_kernel_write1() has      2. snd_rawmidi_kernel_release()
>      not yet pinned runtime.                 closes the output file.
>   3. The writer continues using           3. close_substream() frees
>      the borrowed substream.                 substream->runtime.
> 
> This is the snd-seq-midi sibling of the UMP-bridge race fixed by
> commit 60a1969fae62 ("ALSA: seq: Serialize UMP output teardown with
> event_input"); mirror its approach here.
> 
> Add a per-msynth rwlock for the event_input-visible output file.
> Publish a newly opened output file under the write side once it is
> fully set up, and hold the read side from the output lookup through
> snd_rawmidi_kernel_write() in event_process_midi(). The last unuse
> copies and clears the visible output file under the write side, then
> drops the lock and drains/releases the saved rawmidi file outside it
> (drain/release may sleep). Use IRQ-safe rwlock guards because
> event_input can be reached from atomic sequencer delivery.
> 
> Reproduced under KASAN: a single subscription to the midisynth output
> port (so the subscriber count oscillates 0<->1 and every unsubscribe
> frees the runtime via snd_rawmidi_kernel_release) is raced against a
> flood of events driving event_process_midi -> snd_rawmidi_kernel_write1.
> The narrow window (the runtime is read before snd_rawmidi_buffer_ref())
> was widened with an injected delay to land the race deterministically;
> the freed object is the kmalloc-192 snd_rawmidi_runtime. With this patch
> applied the same forced race shows no use-after-free.
> 
>   BUG: KASAN: slab-use-after-free in snd_rawmidi_kernel_write1+0x73e/0x800
>   Read of size 8 at addr ffff88800b04f310 by task seqmidi_uaf2/84
>   Call Trace:
>    snd_rawmidi_kernel_write1+0x73e/0x800
>    __dump_midi+0x70/0x100
>    dump_var_event+0x290/0x320
>    event_process_midi+0x1ff/0x310
>    snd_seq_deliver_single_event+0x1e6/0x670
>    snd_seq_deliver_event+0x323/0x5f0
>    snd_seq_client_enqueue_event.constprop.0+0x226/0x400
>    snd_seq_write+0x2f1/0x530
>    vfs_write+0x21e/0xd30
>    ksys_write+0x17c/0x1c0
>    do_syscall_64+0xf9/0x540
> 
>   Allocated by task 85:
>    open_substream+0xc7/0x7a0
>    rawmidi_open_priv+0x3df/0x660
>    snd_rawmidi_kernel_open+0x95/0x140
>    midisynth_use+0xda/0x1f0
>    check_and_subscribe_port+0x707/0xbd0
>    snd_seq_ioctl_subscribe_port+0x1f4/0x400
> 
>   Freed by task 85:
>    close_substream.part.0+0x1f9/0x790
>    rawmidi_release_priv+0x1b0/0x240
>    snd_rawmidi_kernel_release+0x2d/0xb0
>    __delete_and_unsubscribe_port+0x1b9/0x3c0
>    snd_seq_ioctl_unsubscribe_port+0x1ee/0x400
> 
>   The buggy address belongs to the object at ffff88800b04f300
>    which belongs to the cache kmalloc-192 of size 192
> 
> Found by 0sec.
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Cc: stable@vger.kernel.org
> Signed-off-by: Doruk Tan Ozturk <doruk@0sec.ai>

There has been already a similar fix queued in sound.git tree for-next
branch:
ef7607ab1c8adc6258fb1b27d08e26aecdc18a58
    ALSA: seq: midi: Serialize output teardown with event_input

I believe this already addressed the bug.

Let me know if there is still anything missing.


thanks,

Takashi