Revert "driver core: enforce device_lock for driver_match_device()"
Danilo Krummrich
maintainerPATCH 1/1 Revert "driver core: enforce device_lock for driver_match_device()"
This reverts commit dc23806a7c47 ("driver core: enforce device_lock for driver_match_device()") and commit 289b14592cef ("driver core: fix inverted "locked" suffix of driver_match_device()").
While technically correct, there is a major downside to this approach:
When a device is already present in the system and a driver is registered on the same bus, we iterate over all devices registered on this bus to see if one of them matches. If we come across an already bound one where the corresponding driver crashed while holding the device lock (e.g. in probe()) we can't make any progress anymore.
However, drivers are typically the least tested code in the kernel and hence it is a case that is likely to happen regularly. Besides hurting developer ergonomics, it potentially decreases chances of shutting things down cleanly and obtaining logs in production environments as well. [1]
This came up in the context of a firewire bug, which only in combination with the reverted commit, caused the machine to hang. [2]
Thus, revert commit dc23806a7c47 ("driver core: enforce device_lock for driver_match_device()") and add a brief note clarifying that an implementer of struct bus_type must not expect match() to be called with the device lock held.
--- drivers/base/base.h | 11 +---------- drivers/base/dd.c | 2 +- include/linux/device/bus.h | 2 ++ 3 files changed, 4 insertions(+), 11 deletions(-)
drivers/base/base.h +1 -10
@@ -179,19 +179,10 @@ void device_release_driver_internal(struct device *dev, const struct device_drivvoid driver_detach (const struct device_driver *drv);void driver_deferred_probe_del (struct device *dev);void device_set_deferred_probe_reason (const struct device *dev,struct va_format *vaf); -static inline int driver_match_device_locked (const struct device_driver *drv, -struct device *dev) -{ -device_lock_assert (dev); - -return drv->bus->match ? drv->bus->match (dev, drv) :1 ; -} -static inline int driver_match_device (const struct device_driver *drv,struct device *dev) { -guard (device)(dev); -return driver_match_device_locked (drv, dev); +return drv->bus->match ? drv->bus->match (dev, drv) :1 ; }static inline void dev_sync_state (struct device *dev)
drivers/base/dd.c +1 -1
@@ -928,7 +928,7 @@ static int __device_attach_driver(struct device_driver *drv, void *_data)bool async_allowed;int ret; - ret =driver_match_device_locked (drv, dev); + ret =driver_match_device (drv, dev);if (ret ==0 ) {/* no match */ return 0 ;
include/linux/device/bus.h +2 -0
@@ -35,6 +35,8 @@ struct fwnode_handle;* otherwise. It may also return error code if determining that * the driver supports the device is not possible. In case of * -EPROBE_DEFER it will queue the device for deferred probing. +* Note: This callback may be invoked with or without the device +* lock held. * @uevent: Called when a device is added, removed, or a few other things * that generate uevents to add the environment variables. * @probe: Called when a new device or driver add to this bus, and callback
base-commit: 78437ab3b769f80526416570f60173c89858dd84
Signature · 1 line
2.53.0
Danilo Krummrich
maintainerQuoted 26 lines · Danilo Krummrich
This reverts commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and commit289b14592cef("driver core: fix inverted "locked" suffix ofdriver_match_device()").While technically correct, there is a major downside to this approach:
When a device is already present in the system and a driver is registered on the same bus, we iterate over all devices registered on this bus to see if one of them matches. If we come across an already bound one where the corresponding driver crashed while holding the device lock (e.g. in
probe()) we can't make any progress anymore.However, drivers are typically the least tested code in the kernel and hence it is a case that is likely to happen regularly. Besides hurting developer ergonomics, it potentially decreases chances of shutting things down cleanly and obtaining logs in production environments as well. [1]
This came up in the context of a firewire bug, which only in combination with the reverted commit, caused the machine to hang. [2]
Thus, revert commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and add a brief note clarifying that an implementer of structbus_typemust not expectmatch()to be called with the device lock held.
Quoted 5 lines
Follow-up regarding driver_override: https://lore.kernel.org/driver-core/[email protected]/
Quoted 27 lines · Danilo Krummrich
This reverts commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and commit289b14592cef("driver core: fix inverted "locked" suffix ofdriver_match_device()").While technically correct, there is a major downside to this approach:
When a device is already present in the system and a driver is registered on the same bus, we iterate over all devices registered on this bus to see if one of them matches. If we come across an already bound one where the corresponding driver crashed while holding the device lock (e.g. in
probe()) we can't make any progress anymore.However, drivers are typically the least tested code in the kernel and hence it is a case that is likely to happen regularly. Besides hurting developer ergonomics, it potentially decreases chances of shutting things down cleanly and obtaining logs in production environments as well. [1]
This came up in the context of a firewire bug, which only in combination with the reverted commit, caused the machine to hang. [2]
Thus, revert commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and add a brief note clarifying that an implementer of structbus_typemust not expectmatch()to be called with the device lock held.
Quoted 5 lines
This aligns with our previous analysis:
https://lore.kernel.org/lkml/CALbr=LZ4v7N=tO1vgOsyj9AS+XuNbn6kG-QcF+PacdMjSo0iyw@mail.gmail.com/
Enforcing the device_lock here amplifies the impact of an oops during probe, so moving to a separate lock for driver_override is definitely a better approach.
The revert looks good to me.
As a minor note, it might be worth adding a tag for the initial Tegra issue where this lock stall was first observed. No big deal if you decide to leave it as is, though.
Quoted 62 lines
--- drivers/base/base.h | 11 +---------- drivers/base/dd.c | 2 +- include/linux/device/bus.h | 2 ++ 3 files changed, 4 insertions(+), 11 deletions(-)drivers/base/base.h +1
-10@@ -179,19 +179,10 @@ void device_release_driver_internal(struct device *dev, const struct device_drivvoid driver_detach (const struct device_driver *drv);void driver_deferred_probe_del (struct device *dev);void device_set_deferred_probe_reason (const struct device *dev,struct va_format *vaf); -static inline int driver_match_device_locked (const struct device_driver *drv, -struct device *dev) -{ -device_lock_assert (dev); - -return drv->bus->match ? drv->bus->match (dev, drv) :1 ; -} -static inline int driver_match_device (const struct device_driver *drv,struct device *dev) { -guard (device)(dev); -return driver_match_device_locked (drv, dev); +return drv->bus->match ? drv->bus->match (dev, drv) :1 ; }static inline void dev_sync_state (struct device *dev)drivers/base/dd.c +1
-1@@ -928,7 +928,7 @@ static int __device_attach_driver(struct device_driver *drv, void *_data)bool async_allowed;int ret; - ret =driver_match_device_locked (drv, dev); + ret =driver_match_device (drv, dev);if (ret ==0 ) {/* no match */ return 0 ;include/linux/device/bus.h +2
-0@@ -35,6 +35,8 @@ struct fwnode_handle;* otherwise. It may also return error code if determining that * the driver supports the device is not possible. In case of * -EPROBE_DEFER it will queue the device for deferred probing. +* Note: This callback may be invoked with or without the device +* lock held. * @uevent: Called when a device is added, removed, or a few other things * that generate uevents to add the environment variables. * @probe: Called when a new device or driver add to this bus, and callback base-commit:
78437ab3b769f80526416570f60173c89858dd84--
2.53.0
Quoted 31 lines · Danilo Krummrich
This reverts commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and commit289b14592cef("driver core: fix inverted "locked" suffix ofdriver_match_device()").While technically correct, there is a major downside to this approach:
When a device is already present in the system and a driver is registered on the same bus, we iterate over all devices registered on this bus to see if one of them matches. If we come across an already bound one where the corresponding driver crashed while holding the device lock (e.g. in
probe()) we can't make any progress anymore.However, drivers are typically the least tested code in the kernel and hence it is a case that is likely to happen regularly. Besides hurting developer ergonomics, it potentially decreases chances of shutting things down cleanly and obtaining logs in production environments as well. [1]
This came up in the context of a firewire bug, which only in combination with the reverted commit, caused the machine to hang. [2]
Thus, revert commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and add a brief note clarifying that an implementer of structbus_typemust not expectmatch()to be called with the device lock held.
Quoted 5 lines
--- drivers/base/base.h | 11 +---------- drivers/base/dd.c | 2 +- include/linux/device/bus.h | 2 ++ 3 files changed, 4 insertions(+), 11 deletions(-)
Quoted 26 lines · Danilo Krummrich
This reverts commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and commit289b14592cef("driver core: fix inverted "locked" suffix ofdriver_match_device()").While technically correct, there is a major downside to this approach:
When a device is already present in the system and a driver is registered on the same bus, we iterate over all devices registered on this bus to see if one of them matches. If we come across an already bound one where the corresponding driver crashed while holding the device lock (e.g. in
probe()) we can't make any progress anymore.However, drivers are typically the least tested code in the kernel and hence it is a case that is likely to happen regularly. Besides hurting developer ergonomics, it potentially decreases chances of shutting things down cleanly and obtaining logs in production environments as well. [1]
This came up in the context of a firewire bug, which only in combination with the reverted commit, caused the machine to hang. [2]
Thus, revert commit
dc23806a7c47("driver core: enforcedevice_lockfordriver_match_device()") and add a brief note clarifying that an implementer of structbus_typemust not expectmatch()to be called with the device lock held.
Quoted 5 lines
Applied to driver-core-linus, thanks!
[ Add additional Link: reference. - Danilo ]