Skip to contentSkip to conversation

ACPI: utils: Fix reference counting in for_each_acpi_dev_match()

PATCH Superseded Andy Shevchenko 3 participants 5 messages May 19, 2021 – May 26, 2021

Subsystems (4)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ACPI: utils: Fix reference counting in for_each_acpi_dev_match()

Currently it's possible to iterate over the dangling pointer in case the device suddenly disappears. This may happen becase callers put it at the end of a loop.

Instead, let's move that call inside acpi_dev_get_next_match_dev().

Fixes 803abec64ef9 ("media: ipu3-cio2: Add cio2-bridge to ipu3-cio2 driver")
Fixes bf263f64e804 ("media: ACPI / bus: Add acpi_dev_get_next_match_dev() and helper macro")
Cc Daniel Scally <djrscally(a)gmail.com>
Cc Sakari Ailus <sakari.ailus(a)linux.intel.com>
Signed-off-by Andy Shevchenko <andy.shevchenko(a)gmail.com>
---
 drivers/acpi/utils.c                       | 5 +----
 drivers/media/pci/intel/ipu3/cio2-bridge.c | 8 +++-----
 include/acpi/acpi_bus.h                    | 5 -----
 3 files changed, 4 insertions(+), 14 deletions(-)
drivers/acpi/utils.c +1 -4
index 3b54b8fd7396..ccfc484dbffd 100644
@@ -846,10 +846,6 @@ EXPORT_SYMBOL(acpi_dev_present);
  * Return the next match of ACPI device if another matching device was present
  * at the moment of invocation, or NULL otherwise.
  *
- * FIXME: The function does not tolerate the sudden disappearance of @adev, e.g.
- * in the case of a hotplug event. That said, the caller should ensure that
- * this will never happen.
- *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
  *
  * See additional information in acpi_dev_present() as well.
@@ -866,6 +862,7 @@ acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha
 	match.hrv = hrv;
 
 	dev = bus_find_device(&acpi_bus_type, start, &match, acpi_dev_match_cb);
+	acpi_dev_put(adev);
 	return dev ? to_acpi_device(dev) : NULL;
 }
 EXPORT_SYMBOL(acpi_dev_get_next_match_dev);
drivers/media/pci/intel/ipu3/cio2-bridge.c +3 -5
index e8511787c1e4..477417261b6e 100644
@@ -178,13 +178,11 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
 
 		if (bridge->n_sensors >= CIO2_NUM_PORTS) {
 			dev_err(&cio2->dev, "Exceeded available CIO2 ports\n");
-			cio2_bridge_unregister_sensors(bridge);
 			ret = -EINVAL;
-			goto err_out;
+			goto err_put_adev;
 		}
 
 		sensor = &bridge->sensors[bridge->n_sensors];
-		sensor->adev = adev;
 		strscpy(sensor->name, cfg->hid, sizeof(sensor->name));
 
 		ret = cio2_bridge_read_acpi_buffer(adev, "SSDB",
@@ -214,6 +212,7 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
 			goto err_free_swnodes;
 		}
 
+		sensor->adev = acpi_dev_get(adev);
 		adev->fwnode.secondary = fwnode;
 
 		dev_info(&cio2->dev, "Found supported sensor %s\n",
@@ -227,8 +226,7 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
 err_free_swnodes:
 	software_node_unregister_nodes(sensor->swnodes);
 err_put_adev:
-	acpi_dev_put(sensor->adev);
-err_out:
+	acpi_dev_put(adev);
 	return ret;
 }
 
include/acpi/acpi_bus.h +0 -5
index 3a82faac5767..bff6a11bb21f 100644
@@ -698,11 +698,6 @@ acpi_dev_get_first_match_dev(const char *hid, const char *uid, s64 hrv);
  * @hrv: Hardware Revision of the device, pass -1 to not check _HRV
  *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
- *
- * FIXME: Due to above requirement there is a window that may invalidate @adev
- * and next iteration will use a dangling pointer, e.g. in the case of a
- * hotplug event. That said, the caller should ensure that this will never
- * happen.
  */
 #define for_each_acpi_dev_match(adev, hid, uid, hrv)			\
 	for (adev = acpi_dev_get_first_match_dev(hid, uid, hrv);	\
Signature · 1 line
2.31.1
#1 lore raw
Quoted 5 lines · Andy Shevchenko

Currently it's possible to iterate over the dangling pointer in case the device suddenly disappears. This may happen becase callers put it at the end of a loop.

Instead, let's move that call inside acpi_dev_get_next_match_dev().

Not really.

Quoted 27 lines
Fixes 803abec64ef9 ("media: ipu3-cio2: Add cio2-bridge to ipu3-cio2 driver")
Fixes bf263f64e804 ("media: ACPI / bus: Add acpi_dev_get_next_match_dev() and helper macro")
Cc Daniel Scally <djrscally(a)gmail.com>
Cc Sakari Ailus <sakari.ailus(a)linux.intel.com>
Signed-off-by Andy Shevchenko <andy.shevchenko(a)gmail.com>
---
 drivers/acpi/utils.c                       | 5 +----
 drivers/media/pci/intel/ipu3/cio2-bridge.c | 8 +++-----
 include/acpi/acpi_bus.h                    | 5 -----
 3 files changed, 4 insertions(+), 14 deletions(-)
drivers/acpi/utils.c +0 -4
index 3b54b8fd7396..ccfc484dbffd 100644
@@ -846,10 +846,6 @@ EXPORT_SYMBOL(acpi_dev_present);
  * Return the next match of ACPI device if another matching device was present
  * at the moment of invocation, or NULL otherwise.
  *
- * FIXME: The function does not tolerate the sudden disappearance of @adev, e.g.
- * in the case of a hotplug event. That said, the caller should ensure that
- * this will never happen.
- *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
  *
  * See additional information in acpi_dev_present() as well.

drivers/acpi/utils.c acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha · line 865

Quoted 4 lines
drivers/acpi/utils.c +1 -0
@@ -866,6 +862,7 @@ acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha
        match.hrv = hrv;
 
        dev = bus_find_device(&acpi_bus_type, start, &match, acpi_dev_match_cb);
+       acpi_dev_put(adev);

What's the point?

The caller may as well put it after this returns, can't it?

And what if the caller still wants to do something with adev after this returns?

drivers/acpi/utils.c acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha · line 868

Quoted 3 lines
drivers/acpi/utils.c +0 -0
        return dev ? to_acpi_device(dev) : NULL;
 }
 EXPORT_SYMBOL(acpi_dev_get_next_match_dev);

The original changelog is somewhat unclear. It should say that it is required to reference-count the start-point device before passing it to this function, but beyond that I don't see a bug here.

Quoted 56 lines
drivers/media/pci/intel/ipu3/cio2-bridge.c +3 -5
index e8511787c1e4..477417261b6e 100644
@@ -178,13 +178,11 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
 
                if (bridge->n_sensors >= CIO2_NUM_PORTS) {
                        dev_err(&cio2->dev, "Exceeded available CIO2 ports\n");
-                       cio2_bridge_unregister_sensors(bridge);
                        ret = -EINVAL;
-                       goto err_out;
+                       goto err_put_adev;
                }
 
                sensor = &bridge->sensors[bridge->n_sensors];
-               sensor->adev = adev;
                strscpy(sensor->name, cfg->hid, sizeof(sensor->name));
 
                ret = cio2_bridge_read_acpi_buffer(adev, "SSDB",
@@ -214,6 +212,7 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
                        goto err_free_swnodes;
                }
 
+               sensor->adev = acpi_dev_get(adev);
                adev->fwnode.secondary = fwnode;
 
                dev_info(&cio2->dev, "Found supported sensor %s\n",
@@ -227,8 +226,7 @@ static int cio2_bridge_connect_sensor(const struct cio2_sensor_config *cfg,
 err_free_swnodes:
        software_node_unregister_nodes(sensor->swnodes);
 err_put_adev:
-       acpi_dev_put(sensor->adev);
-err_out:
+       acpi_dev_put(adev);
        return ret;
 }
 
include/acpi/acpi_bus.h +0 -6
index 3a82faac5767..bff6a11bb21f 100644
@@ -698,11 +698,6 @@ acpi_dev_get_first_match_dev(const char *hid, const char *uid, s64 hrv);
  * @hrv: Hardware Revision of the device, pass -1 to not check _HRV
  *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
- *
- * FIXME: Due to above requirement there is a window that may invalidate @adev
- * and next iteration will use a dangling pointer, e.g. in the case of a
- * hotplug event. That said, the caller should ensure that this will never
- * happen.
  */
 #define for_each_acpi_dev_match(adev, hid, uid, hrv)                   \
        for (adev = acpi_dev_get_first_match_dev(hid, uid, hrv);        \
--

2.31.1

#2 lore raw
Quoted 10 lines · Rafael J. Wysocki
Quoted 5 lines · Andy Shevchenko

Currently it's possible to iterate over the dangling pointer in case the device suddenly disappears. This may happen becase callers put it at the end of a loop.

Instead, let's move that call inside acpi_dev_get_next_match_dev().

Not really.

OK, I see what you want to achieve and the macro is actually buggy, because it leaves unbalanced references behind.

Quoted 11 lines
Quoted 11 lines
Fixes 803abec64ef9 ("media: ipu3-cio2: Add cio2-bridge to ipu3-cio2 driver")
Fixes bf263f64e804 ("media: ACPI / bus: Add acpi_dev_get_next_match_dev() and helper macro")
Cc Daniel Scally <djrscally(a)gmail.com>
Cc Sakari Ailus <sakari.ailus(a)linux.intel.com>
Signed-off-by Andy Shevchenko <andy.shevchenko(a)gmail.com>
---
 drivers/acpi/utils.c                       | 5 +----
 drivers/media/pci/intel/ipu3/cio2-bridge.c | 8 +++-----
 include/acpi/acpi_bus.h                    | 5 -----
 3 files changed, 4 insertions(+), 14 deletions(-)
Quoted 15 lines
Quoted 5 lines
drivers/acpi/utils.c +0 -0
index 3b54b8fd7396..ccfc484dbffd 100644

drivers/acpi/utils.c EXPORT_SYMBOL(acpi_dev_present); · line 851

Quoted 10 lines
drivers/acpi/utils.c +0 -4
@@ -846,10 +846,6 @@ EXPORT_SYMBOL(acpi_dev_present);
  * Return the next match of ACPI device if another matching device was present
  * at the moment of invocation, or NULL otherwise.
  *
- * FIXME: The function does not tolerate the sudden disappearance of @adev, e.g.
- * in the case of a hotplug event. That said, the caller should ensure that
- * this will never happen.
- *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
  *
  * See additional information in acpi_dev_present() as well.

But the kerneldoc really needs to say that the caller is required to obtain a reference on adev before passing it here, because that reference will be dropped and the object pointed to by adev may not be present any more after this returns.

Quoted 5 lines
Quoted 5 lines
drivers/acpi/utils.c +1 -0
@@ -866,6 +862,7 @@ acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha
        match.hrv = hrv;
 
        dev = bus_find_device(&acpi_bus_type, start, &match, acpi_dev_match_cb);
+       acpi_dev_put(adev);
#3 lore raw

Hi Rafael, Andy

Quoted 10 lines · Rafael J. Wysocki
Quoted 7 lines · Rafael J. Wysocki
Quoted 4 lines · Andy Shevchenko

Currently it's possible to iterate over the dangling pointer in case the device suddenly disappears. This may happen becase callers put it at the end of a loop.

Instead, let's move that call inside acpi_dev_get_next_match_dev().

Not really.

OK, I see what you want to achieve and the macro is actually buggy, because it leaves unbalanced references behind.

Yeah; I guess the originally intended use (which was "get all these devices") doesn't really tally with the naming or with the operation of similar functions in the kernel like the fwnode_handle ops; sorry about that. Anyway; I think Andy's fix is the right way to do it, so the calling code's responsible for holding onto a reference if it wants to keep it.

Quoted 37 lines
Quoted 12 lines
Quoted 12 lines
Fixes 803abec64ef9 ("media: ipu3-cio2: Add cio2-bridge to ipu3-cio2 driver")
Fixes bf263f64e804 ("media: ACPI / bus: Add acpi_dev_get_next_match_dev() and helper macro")
Cc Daniel Scally <[email protected]>
Cc Sakari Ailus <[email protected]>
---
 drivers/acpi/utils.c                       | 5 +----
 drivers/media/pci/intel/ipu3/cio2-bridge.c | 8 +++-----
 include/acpi/acpi_bus.h                    | 5 -----
 3 files changed, 4 insertions(+), 14 deletions(-)
Quoted 15 lines
Quoted 5 lines
drivers/acpi/utils.c +0 -0
index 3b54b8fd7396..ccfc484dbffd 100644

drivers/acpi/utils.c EXPORT_SYMBOL(acpi_dev_present); · line 851

Quoted 10 lines
drivers/acpi/utils.c +0 -4
@@ -846,10 +846,6 @@ EXPORT_SYMBOL(acpi_dev_present);
  * Return the next match of ACPI device if another matching device was present
  * at the moment of invocation, or NULL otherwise.
  *
- * FIXME: The function does not tolerate the sudden disappearance of @adev, e.g.
- * in the case of a hotplug event. That said, the caller should ensure that
- * this will never happen.
- *
  * The caller is responsible for invoking acpi_dev_put() on the returned device.
  *
  * See additional information in acpi_dev_present() as well.

But the kerneldoc really needs to say that the caller is required to obtain a reference on adev before passing it here, because that reference will be dropped and the object pointed to by adev may not be present any more after this returns.

Quoted 5 lines
Quoted 5 lines
drivers/acpi/utils.c +1 -0
@@ -866,6 +862,7 @@ acpi_dev_get_next_match_dev(struct acpi_device *adev, const char *hid, const cha
        match.hrv = hrv;
 
        dev = bus_find_device(&acpi_bus_type, start, &match, acpi_dev_match_cb);
+       acpi_dev_put(adev);
#4 lore raw
Quoted 13 lines · Daniel Scally
Quoted 10 lines · Rafael J. Wysocki
Quoted 7 lines · Rafael J. Wysocki
Quoted 4 lines · Andy Shevchenko

Currently it's possible to iterate over the dangling pointer in case the device suddenly disappears. This may happen becase callers put it at the end of a loop.

Instead, let's move that call inside acpi_dev_get_next_match_dev().

Not really.

OK, I see what you want to achieve and the macro is actually buggy, because it leaves unbalanced references behind.

Quoted 6 lines

Yeah; I guess the originally intended use (which was "get all these devices") doesn't really tally with the naming or with the operation of similar functions in the kernel like the fwnode_handle ops; sorry about that. Anyway; I think Andy's fix is the right way to do it, so the calling code's responsible for holding onto a reference if it wants to keep it.

I think we need to postpone the fix till v5.14-rc1 is out. It appears that some code has been changed in EFI and media subsystems so that patch will be in conflict with.

Signature · 2 lines
With Best Regards,
Andy Shevchenko