On 7/15/20 6:23 AM, Hans de Goede wrote:
> When we receive a PD data packet which ends up being for the alt-mode
> driver we have the following lock order:
>
> 1. tcpm_pd_rx_handler take the tcpm-port lock
> 2. We call into the alt-mode driver which takes the alt-mode's lock
>
> And when the alt-mode driver initiates communication we have the following
> lock order:
>
> 3. alt-mode driver takes the alt-mode's lock
> 4. alt-mode driver calls tcpm_altmode_enter which takes the tcpm-port lock
>
> This is a classic AB BA lock inversion issue.
>
> With the refactoring of tcpm_handle_vdm_request() done before this patch,
> we don't rely on, or need to make changes to the tcpm-port data by the
> time we make call 2. from above. All data to be passed to the alt-mode
> driver sits on our stack at this point, and thus does not need locking.
>
> So after the refactoring we can simply fix this by releasing the
> tcpm-port lock before calling into the alt-mode driver.
>
> This fixes the following lockdep warning:
>
> [ 191.454238] ======================================================
> [ 191.454240] WARNING: possible circular locking dependency detected
> [ 191.454244] 5.8.0-rc5+ #1 Not tainted
> [ 191.454246] ------------------------------------------------------
> [ 191.454248] kworker/u8:5/794 is trying to acquire lock:
> [ 191.454251] ffff9bac8e30d4a8 (&dp->lock){+.+.}-{3:3}, at: dp_altmode_vdm+0x30/0xf0 [typec_displayport]
> [ 191.454263]
> but task is already holding lock:
> [ 191.454264] ffff9bac9dc240a0 (&port->lock#2){+.+.}-{3:3}, at: tcpm_pd_rx_handler+0x43/0x12c0 [tcpm]
> [ 191.454273]
> which lock already depends on the new lock.
>
> [ 191.454275]
> the existing dependency chain (in reverse order) is:
> [ 191.454277]
> -> #1 (&port->lock#2){+.+.}-{3:3}:
> [ 191.454286] __mutex_lock+0x7b/0x820
> [ 191.454290] tcpm_altmode_enter+0x23/0x90 [tcpm]
> [ 191.454293] dp_altmode_work+0xca/0xe0 [typec_displayport]
> [ 191.454299] process_one_work+0x23f/0x570
> [ 191.454302] worker_thread+0x55/0x3c0
> [ 191.454305] kthread+0x138/0x160
> [ 191.454309] ret_from_fork+0x22/0x30
> [ 191.454311]
> -> #0 (&dp->lock){+.+.}-{3:3}:
> [ 191.454317] __lock_acquire+0x1241/0x2090
> [ 191.454320] lock_acquire+0xa4/0x3d0
> [ 191.454323] __mutex_lock+0x7b/0x820
> [ 191.454326] dp_altmode_vdm+0x30/0xf0 [typec_displayport]
> [ 191.454330] tcpm_pd_rx_handler+0x11ae/0x12c0 [tcpm]
> [ 191.454333] process_one_work+0x23f/0x570
> [ 191.454336] worker_thread+0x55/0x3c0
> [ 191.454338] kthread+0x138/0x160
> [ 191.454341] ret_from_fork+0x22/0x30
> [ 191.454343]
> other info that might help us debug this:
>
> [ 191.454345] Possible unsafe locking scenario:
>
> [ 191.454347] CPU0 CPU1
> [ 191.454348] ---- ----
> [ 191.454350] lock(&port->lock#2);
> [ 191.454353] lock(&dp->lock);
> [ 191.454355] lock(&port->lock#2);
> [ 191.454357] lock(&dp->lock);
> [ 191.454360]
> *** DEADLOCK ***
>
> Signed-off-by: Hans de Goede <hdegoede@xxxxxxxxxx>
> ---
> drivers/usb/typec/tcpm/tcpm.c | 31 ++++++++++++++++++++++++++++++-
> 1 file changed, 30 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/usb/typec/tcpm/tcpm.c b/drivers/usb/typec/tcpm/tcpm.c
> index 4745b4062000..ea14240423d1 100644
> --- a/drivers/usb/typec/tcpm/tcpm.c
> +++ b/drivers/usb/typec/tcpm/tcpm.c
> @@ -1247,6 +1247,27 @@ static void tcpm_handle_vdm_request(struct tcpm_port *port,
> if (PD_VDO_SVDM(p[0]))
> rlen = tcpm_pd_svdm(port, p, cnt, response, &adev_action);
>
> + /*
> + * We are done with any state stored in the port struct now, except
> + * for any port struct changes done by the tcpm_queue_vdm() call
> + * below, which is a separate operation.
> + *
> + * So we can safely release the lock here; and we MUST release the
> + * lock here to avoid an AB BA lock inversion:
> + *
> + * If we keep the lock here then the lock ordering in this path is:
> + * 1. tcpm_pd_rx_handler take the tcpm port lock
> + * 2. One of the typec_altmode_* calls below takes the alt-mode's lock
> + *
> + * And we also have this ordering:
> + * 1. alt-mode driver takes the alt-mode's lock
> + * 2. alt-mode driver calls tcpm_altmode_enter which takes the
> + * tcpm port lock
> + *
> + * Dropping our lock here avoids this.
> + */
> + mutex_unlock(&port->lock);
> +
> if (adev) {
> switch (adev_action) {
> case ADEV_NONE:
> @@ -1272,7 +1293,15 @@ static void tcpm_handle_vdm_request(struct tcpm_port *port,
> }
>
> if (rlen > 0)
> - tcpm_queue_vdm_unlocked(port, response[0], &response[1], rlen - 1);
> + tcpm_queue_vdm(port, response[0], &response[1], rlen - 1);
> +
> + /*
> + * We must re-take the lock here to balance the unlock in
> + * tcpm_pd_rx_handler, note that no changes are made while the lock
> + * is held again. All that is done is unwinding the call stack until
> + * we return to tcpm_pd_rx_handler and do the unlock there.
> + */
> + mutex_lock(&port->lock);
Unless I am missing something, tcpm_queue_vdm() now also acquires the lock
and releases it. Why not move this further up and keep tcpm_queue_vdm_unlocked() ?
This would avoid one set of lock/unlock calls.
Thanks,
Guenter
> }
>
> static void tcpm_send_vdm(struct tcpm_port *port, u32 vid, int cmd,
>
