| Message ID | 20200715132301.99816-6-hdegoede@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | usb: typec: tcpm: Fix AB BA lock inversion between tcpm and alt-mode drivers | expand |
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@redhat.com> > --- > 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, >
Hi, On 7/15/20 6:05 PM, Guenter Roeck wrote: > 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@redhat.com> >> --- >> 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. You're right, I've changed this for v2 of the patch-set. Regards, Hans
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); } static void tcpm_send_vdm(struct tcpm_port *port, u32 vid, int cmd,
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@redhat.com> --- drivers/usb/typec/tcpm/tcpm.c | 31 ++++++++++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-)