Skip to contentSkip to conversation

Get device's parent from parent field, fix sleeping IRQs disabled

PATCH In review Sakari Ailus 5 participants 18 messages Nov 3, 2021 – Nov 4, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Get device's parent from parent field, fix sleeping IRQs disabled

Hi all,

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f
 __dynamic_pr_debug+0xe2/0x100
 __v4l2_fwnode_endpoint_parse+0x10c/0x450 [v4l2_fwnode]
 v4l2_fwnode_endpoint_parse+0x11/0x40 [v4l2_fwnode]
 cio2_pci_probe.cold+0x7fb/0x969 [ipu3_cio2]
 ? _raw_spin_unlock_irqrestore+0x42/0x70
 pci_device_probe+0xb6/0x140
 really_probe+0x1f5/0x3f0
 __driver_probe_device+0xfe/0x180
 driver_probe_device+0x1e/0x90
 __driver_attach+0xc4/0x1d0
 ? __device_attach_driver+0xe0/0xe0
 ? __device_attach_driver+0xe0/0xe0
 bus_for_each_dev+0x7b/0xc0
 bus_add_driver+0x14c/0x1f0
 driver_register+0x8b/0xe0
 ? 0xffffffffa0428000
 do_one_initcall+0x5b/0x300
 ? rcu_read_lock_sched_held+0x12/0x70
 ? trace_kmalloc+0x29/0xd0
 ? kmem_cache_alloc_trace+0x11d/0x630
 do_init_module+0x5c/0x260
 __do_sys_finit_module+0xae/0x110
 do_syscall_64+0x3b/0x90
 entry_SYSCALL_64_after_hwframe+0x44/0xae

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

Sakari Ailus (3):
ACPI: Make acpi_fwnode_handle safer
ACPI: Get acpi_device's parent from the parent field
ACPI: Make acpi_node_get_parent() local

 drivers/acpi/property.c | 15 ++++++---------
 include/acpi/acpi_bus.h |  2 +-
 include/linux/acpi.h    |  7 -------
 3 files changed, 7 insertions(+), 17 deletions(-)
Signature · 1 line
2.30.2
#1 lore raw

PATCH 1/3 ACPI: Make acpi_fwnode_handle safer

Check that the fwnode argument passed to acpi_fwnode_handle is non-NULL, and return NULL if it is, otherwise the fwnode. Thus the caller doesn't have to ensure the argument is a valid non-NULL fwnode.

---
 include/acpi/acpi_bus.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/acpi/acpi_bus.h +1 -1
index 53b6e9f9de7b4..c34d94521d40c 100644
@@ -445,7 +445,7 @@ static inline bool acpi_data_node_match(const struct fwnode_handle *fwnode,
 
 static inline struct fwnode_handle *acpi_fwnode_handle(struct acpi_device *adev)
 {
-	return &adev->fwnode;
+	return adev ? &adev->fwnode : NULL;
 }
 
 static inline void *acpi_driver_data(struct acpi_device *d)
Signature · 1 line
2.30.2
Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] ACPI: Make acpi_fwnode_handle safer

Quoted 1 line · Sakari Ailus

Check that the fwnode argument passed to acpi_fwnode_handle is non-NULL,

acpi_fwnode_handle()

Quoted 2 lines

and return NULL if it is, otherwise the fwnode. Thus the caller doesn't have to ensure the argument is a valid non-NULL fwnode.

Quoted 23 lines
---
 include/acpi/acpi_bus.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/acpi/acpi_bus.h +1 -1
index 53b6e9f9de7b4..c34d94521d40c 100644
@@ -445,7 +445,7 @@ static inline bool acpi_data_node_match(const struct fwnode_handle *fwnode,
 
 static inline struct fwnode_handle *acpi_fwnode_handle(struct acpi_device *adev)
 {
-	return &adev->fwnode;
+	return adev ? &adev->fwnode : NULL;
 }
 
 static inline void *acpi_driver_data(struct acpi_device *d)
Signature · 1 line
2.30.2
Signature · 2 lines
With Best Regards,
Andy Shevchenko
#3 lore raw

Re: [PATCH 1/3] ACPI: Make acpi_fwnode_handle safer

In reply to Sakari Ailus
Quoted 6 lines · Sakari Ailus

Check that the fwnode argument passed to acpi_fwnode_handle is non-NULL, and return NULL if it is, otherwise the fwnode. Thus the caller doesn't have to ensure the argument is a valid non-NULL fwnode.

Cc [email protected] # v5.15 and up

Why?

Quoted 1 line

That's because you want to avoid a NULL check in the second patch and it adds a ton of redundant NULL checks all over the place.

Like for example in include/acpi/acpi.h:

#define ACPI_COMPANION_SET(dev, adev) set_primary_fwnode(dev, (adev) ? \
acpi_fwnode_handle(adev) : NULL)

You should at least get rid of this one.

Quoted 18 lines
---
 include/acpi/acpi_bus.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/acpi/acpi_bus.h +1 -2
index 53b6e9f9de7b4..c34d94521d40c 100644
@@ -445,7 +445,7 @@ static inline bool acpi_data_node_match(const struct fwnode_handle *fwnode,
 
 static inline struct fwnode_handle *acpi_fwnode_handle(struct acpi_device *adev)
 {
-       return &adev->fwnode;
+       return adev ? &adev->fwnode : NULL;
 }
 
 static inline void *acpi_driver_data(struct acpi_device *d)
--
#4 lore raw

Re: [PATCH 1/3] ACPI: Make acpi_fwnode_handle safer

Hi Rafael,

Quoted 16 lines · Rafael J. Wysocki
Quoted 6 lines · Sakari Ailus

Check that the fwnode argument passed to acpi_fwnode_handle is non-NULL, and return NULL if it is, otherwise the fwnode. Thus the caller doesn't have to ensure the argument is a valid non-NULL fwnode.

Cc [email protected] # v5.15 and up

Why?

Quoted 1 line

That's because you want to avoid a NULL check in the second patch and it adds a ton of redundant NULL checks all over the place.

Quoted 6 lines

Like for example in include/acpi/acpi.h:

#define ACPI_COMPANION_SET(dev, adev) set_primary_fwnode(dev, (adev) ? \
acpi_fwnode_handle(adev) : NULL)

You should at least get rid of this one.

That's a fair point.

acpi_fwnode_handle() doesn't have a tonne of users so it's entirely feasible to change the users as necessary but in either case I guess I'll drop this from stable as it's likely to conflict with some of these.

Signature · 3 lines
Regards,

Sakari Ailus
#5 lore raw

Re: [PATCH 1/3] ACPI: Make acpi_fwnode_handle safer

In reply to Rafael J. Wysocki
Quoted 16 lines · Rafael J. Wysocki
Quoted 6 lines · Sakari Ailus

Check that the fwnode argument passed to acpi_fwnode_handle is non-NULL, and return NULL if it is, otherwise the fwnode. Thus the caller doesn't have to ensure the argument is a valid non-NULL fwnode.

Cc [email protected] # v5.15 and up

Why?

Quoted 1 line

That's because you want to avoid a NULL check in the second patch and it adds a ton of redundant NULL checks all over the place.

Quoted 6 lines

Like for example in include/acpi/acpi.h:

#define ACPI_COMPANION_SET(dev, adev) set_primary_fwnode(dev, (adev) ? \
acpi_fwnode_handle(adev) : NULL)

You should at least get rid of this one.

I went through the users and it seems they don't actually even benefit from this. So I'll drop the patch.

Most of these little macros are made safer that way but in this case it seems callers have already acpi_device around, or do checks for other purposes as well.

Signature · 1 line
Sakari Ailus
#6 lore raw

PATCH 2/3 ACPI: Get acpi_device's parent from the parent field

In reply to Sakari Ailus

Printk modifier %pfw is used to print the full path of the device name. This is obtained device by device until a device no longer has a parent.

On ACPI getting the parent fwnode is done by calling acpi_get_parent() which tries to down() a semaphore. But local IRQs are now disabled in vprintk_store() before the mutex is acquired. This is obviously a problem.

Luckily struct acpi_device has a parent field already. Use that field to get the parent instead of relying on acpi_get_parent().

Fixes 002eb6ad0751 ("printk: track/limit recursion")
Cc [email protected] # v5.15 and up

Depends-on: ("ACPI: Make acpi_fwnode_handle safer")

---
 drivers/acpi/property.c | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)
drivers/acpi/property.c +4 -8
index e312ebaed8db4..7403ee2816eb8 100644
@@ -1089,16 +1089,12 @@ struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
 	if (is_acpi_data_node(fwnode)) {
 		/* All data nodes have parent pointer so just return that */
 		return to_acpi_data_node(fwnode)->parent;
-	} else if (is_acpi_device_node(fwnode)) {
-		acpi_handle handle, parent_handle;
+	}
 
-		handle = to_acpi_device_node(fwnode)->handle;
-		if (ACPI_SUCCESS(acpi_get_parent(handle, &parent_handle))) {
-			struct acpi_device *adev;
+	if (is_acpi_device_node(fwnode)) {
+		struct acpi_device *device = to_acpi_device_node(fwnode);
 
-			if (!acpi_bus_get_device(parent_handle, &adev))
-				return acpi_fwnode_handle(adev);
-		}
+		return acpi_fwnode_handle(device->parent);
 	}
 
 	return NULL;
Signature · 1 line
2.30.2
Reviewed-by on v1 2/3
#7 lore raw

Re: [PATCH 2/3] ACPI: Get acpi_device's parent from the parent field

Quoted 9 lines · Sakari Ailus

Printk modifier %pfw is used to print the full path of the device name. This is obtained device by device until a device no longer has a parent.

On ACPI getting the parent fwnode is done by calling acpi_get_parent() which tries to down() a semaphore. But local IRQs are now disabled in vprintk_store() before the mutex is acquired. This is obviously a problem.

Luckily struct acpi_device has a parent field already. Use that field to get the parent instead of relying on acpi_get_parent().

I think the best if Rafael can confirm that we may use it like this. If he approved, nothing would stop me to give you

Quoted 37 lines
Fixes 002eb6ad0751 ("printk: track/limit recursion")
Cc [email protected] # v5.15 and up

Depends-on: ("ACPI: Make acpi_fwnode_handle safer")

---
 drivers/acpi/property.c | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)
drivers/acpi/property.c +4 -8
index e312ebaed8db4..7403ee2816eb8 100644
@@ -1089,16 +1089,12 @@ struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
 	if (is_acpi_data_node(fwnode)) {
 		/* All data nodes have parent pointer so just return that */
 		return to_acpi_data_node(fwnode)->parent;
-	} else if (is_acpi_device_node(fwnode)) {
-		acpi_handle handle, parent_handle;
+	}
 
-		handle = to_acpi_device_node(fwnode)->handle;
-		if (ACPI_SUCCESS(acpi_get_parent(handle, &parent_handle))) {
-			struct acpi_device *adev;
+	if (is_acpi_device_node(fwnode)) {
+		struct acpi_device *device = to_acpi_device_node(fwnode);
 
-			if (!acpi_bus_get_device(parent_handle, &adev))
-				return acpi_fwnode_handle(adev);
-		}
+		return acpi_fwnode_handle(device->parent);
 	}
 
 	return NULL;
Signature · 1 line
2.30.2
Signature · 2 lines
With Best Regards,
Andy Shevchenko
#8 lore raw

Re: [PATCH 2/3] ACPI: Get acpi_device's parent from the parent field

Quoted 10 lines · Andy Shevchenko
Quoted 8 lines · Sakari Ailus

Printk modifier %pfw is used to print the full path of the device name. This is obtained device by device until a device no longer has a parent.

On ACPI getting the parent fwnode is done by calling acpi_get_parent() which tries to down() a semaphore. But local IRQs are now disabled in vprintk_store() before the mutex is acquired. This is obviously a problem.

Luckily struct acpi_device has a parent field already.

Which I'm going to eliminate, because it is redundant.

The dev object embedded in struct acpi_device has a parent field pointing to the same object and that one is used by the driver core.

Quoted 5 lines
Quoted 1 line

Use that field to get the parent instead of relying on acpi_get_parent().

I think the best if Rafael can confirm that we may use it like this. If he approved, nothing would stop me to give you

In fact, I would prefer the parent field of the dev object embedded in struct acpi_device to be used and for completeness it should be tested against NULL unless you know that the parent is not going away ATM.

Quoted 20 lines
Quoted 20 lines
Fixes 002eb6ad0751 ("printk: track/limit recursion")
Cc [email protected] # v5.15 and up

Depends-on: ("ACPI: Make acpi_fwnode_handle safer")

---
 drivers/acpi/property.c | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)
drivers/acpi/property.c +1 -2
index e312ebaed8db4..7403ee2816eb8 100644
@@ -1089,16 +1089,12 @@ struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
      if (is_acpi_data_node(fwnode)) {
              /* All data nodes have parent pointer so just return that */
              return to_acpi_data_node(fwnode)->parent;
-     } else if (is_acpi_device_node(fwnode)) {
-             acpi_handle handle, parent_handle;
+     }
 
Quoted 5 lines

drivers/acpi/property.c struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode) · line 1095

Quoted 5 lines
drivers/acpi/property.c +2 -3
-             handle = to_acpi_device_node(fwnode)->handle;
-             if (ACPI_SUCCESS(acpi_get_parent(handle, &parent_handle))) {
-                     struct acpi_device *adev;
+     if (is_acpi_device_node(fwnode)) {
+             struct acpi_device *device = to_acpi_device_node(fwnode);

Call this variable adev please.

Quoted 5 lines

drivers/acpi/property.c struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode) · line 1097

Quoted 5 lines
drivers/acpi/property.c +1 -3
-                     if (!acpi_bus_get_device(parent_handle, &adev))
-                             return acpi_fwnode_handle(adev);
-             }
+             return acpi_fwnode_handle(device->parent);

And then

             adev = to_acpi_device(&device->dev.parent);
             if (adev)
                          return acpi_fwnode_handle(adev);
Quoted 13 lines
Quoted 7 lines
drivers/acpi/property.c +0 -1
      }
 
      return NULL;
--

2.30.2

--
With Best Regards,
Andy Shevchenko

#9 lore raw

Re: [PATCH 2/3] ACPI: Get acpi_device's parent from the parent field

Hi Rafael,

Quoted 11 lines · Rafael J. Wysocki
Quoted 9 lines · Andy Shevchenko
Quoted 7 lines · Sakari Ailus

Printk modifier %pfw is used to print the full path of the device name. This is obtained device by device until a device no longer has a parent.

On ACPI getting the parent fwnode is done by calling acpi_get_parent() which tries to down() a semaphore. But local IRQs are now disabled in vprintk_store() before the mutex is acquired. This is obviously a problem.

Quoted 6 lines
Quoted 1 line
Quoted 1 line

Luckily struct acpi_device has a parent field already.

Which I'm going to eliminate, because it is redundant.

The dev object embedded in struct acpi_device has a parent field pointing to the same object and that one is used by the driver core.

Indeed, I'll do that in v2.

Quoted 45 lines
Quoted 6 lines
Quoted 2 lines

Use that field to get the parent instead of relying on acpi_get_parent().

I think the best if Rafael can confirm that we may use it like this. If he approved, nothing would stop me to give you

In fact, I would prefer the parent field of the dev object embedded in struct acpi_device to be used and for completeness it should be tested against NULL unless you know that the parent is not going away ATM.

Quoted 20 lines
Quoted 20 lines
Fixes 002eb6ad0751 ("printk: track/limit recursion")
Cc [email protected] # v5.15 and up

Depends-on: ("ACPI: Make acpi_fwnode_handle safer")

---
 drivers/acpi/property.c | 12 ++++--------
 1 file changed, 4 insertions(+), 8 deletions(-)
drivers/acpi/property.c +1 -2
index e312ebaed8db4..7403ee2816eb8 100644
@@ -1089,16 +1089,12 @@ struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
      if (is_acpi_data_node(fwnode)) {
              /* All data nodes have parent pointer so just return that */
              return to_acpi_data_node(fwnode)->parent;
-     } else if (is_acpi_device_node(fwnode)) {
-             acpi_handle handle, parent_handle;
+     }
 
Quoted 5 lines

drivers/acpi/property.c struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode) · line 1095

Quoted 5 lines
drivers/acpi/property.c +2 -3
-             handle = to_acpi_device_node(fwnode)->handle;
-             if (ACPI_SUCCESS(acpi_get_parent(handle, &parent_handle))) {
-                     struct acpi_device *adev;
+     if (is_acpi_device_node(fwnode)) {
+             struct acpi_device *device = to_acpi_device_node(fwnode);

Call this variable adev please.

Quoted 5 lines

drivers/acpi/property.c struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode) · line 1097

Quoted 5 lines
drivers/acpi/property.c +1 -3
-                     if (!acpi_bus_get_device(parent_handle, &adev))
-                             return acpi_fwnode_handle(adev);
-             }
+             return acpi_fwnode_handle(device->parent);
Quoted 5 lines

And then

             adev = to_acpi_device(&device->dev.parent);
             if (adev)
                          return acpi_fwnode_handle(adev);

Ack.

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
drivers/acpi/property.c +0 -0
      }
 
      return NULL;
Signature · 3 lines
Kind regards,

Sakari Ailus
#10 lore raw

PATCH 3/3 ACPI: Make acpi_node_get_parent() local

In reply to Sakari Ailus

acpi_node_get_parent() isn't used outside drivers/acpi/property.c. Make it local.

---
 drivers/acpi/property.c | 3 ++-
 include/linux/acpi.h    | 7 -------
 2 files changed, 2 insertions(+), 8 deletions(-)
drivers/acpi/property.c +2 -1
index 7403ee2816eb8..49301d1bba4ff 100644
@@ -1084,7 +1084,8 @@ struct fwnode_handle *acpi_get_next_subnode(const struct fwnode_handle *fwnode,
  * Returns parent node of an ACPI device or data firmware node or %NULL if
  * not available.
  */
-struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
+static struct fwnode_handle *
+acpi_node_get_parent(const struct fwnode_handle *fwnode)
 {
 	if (is_acpi_data_node(fwnode)) {
 		/* All data nodes have parent pointer so just return that */
include/linux/acpi.h +0 -7
index 375715b0535fb..c65a754b1db53 100644
@@ -1168,7 +1168,6 @@ int acpi_node_prop_get(const struct fwnode_handle *fwnode, const char *propname,
 
 struct fwnode_handle *acpi_get_next_subnode(const struct fwnode_handle *fwnode,
 					    struct fwnode_handle *child);
-struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode);
 
 struct acpi_probe_entry;
 typedef bool (*acpi_probe_entry_validate_subtbl)(struct acpi_subtable_header *,
@@ -1273,12 +1272,6 @@ acpi_get_next_subnode(const struct fwnode_handle *fwnode,
 	return NULL;
 }
 
-static inline struct fwnode_handle *
-acpi_node_get_parent(const struct fwnode_handle *fwnode)
-{
-	return NULL;
-}
-
 static inline struct fwnode_handle *
 acpi_graph_get_next_endpoint(const struct fwnode_handle *fwnode,
 			     struct fwnode_handle *prev)
Signature · 1 line
2.30.2
Reviewed-by on v1 3/3
#11 lore raw

Re: [PATCH 3/3] ACPI: Make acpi_node_get_parent() local

Quoted 2 lines · Sakari Ailus

acpi_node_get_parent() isn't used outside drivers/acpi/property.c. Make it local.

Always in favour for such patches!

Quoted 49 lines
---
 drivers/acpi/property.c | 3 ++-
 include/linux/acpi.h    | 7 -------
 2 files changed, 2 insertions(+), 8 deletions(-)
drivers/acpi/property.c +2 -1
index 7403ee2816eb8..49301d1bba4ff 100644
@@ -1084,7 +1084,8 @@ struct fwnode_handle *acpi_get_next_subnode(const struct fwnode_handle *fwnode,
  * Returns parent node of an ACPI device or data firmware node or %NULL if
  * not available.
  */
-struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode)
+static struct fwnode_handle *
+acpi_node_get_parent(const struct fwnode_handle *fwnode)
 {
 	if (is_acpi_data_node(fwnode)) {
 		/* All data nodes have parent pointer so just return that */
include/linux/acpi.h +0 -7
index 375715b0535fb..c65a754b1db53 100644
@@ -1168,7 +1168,6 @@ int acpi_node_prop_get(const struct fwnode_handle *fwnode, const char *propname,
 
 struct fwnode_handle *acpi_get_next_subnode(const struct fwnode_handle *fwnode,
 					    struct fwnode_handle *child);
-struct fwnode_handle *acpi_node_get_parent(const struct fwnode_handle *fwnode);
 
 struct acpi_probe_entry;
 typedef bool (*acpi_probe_entry_validate_subtbl)(struct acpi_subtable_header *,
@@ -1273,12 +1272,6 @@ acpi_get_next_subnode(const struct fwnode_handle *fwnode,
 	return NULL;
 }
 
-static inline struct fwnode_handle *
-acpi_node_get_parent(const struct fwnode_handle *fwnode)
-{
-	return NULL;
-}
-
 static inline struct fwnode_handle *
 acpi_graph_get_next_endpoint(const struct fwnode_handle *fwnode,
 			     struct fwnode_handle *prev)
Signature · 1 line
2.30.2
Signature · 2 lines
With Best Regards,
Andy Shevchenko
#12 lore raw

Re: [PATCH 3/3] ACPI: Make acpi_node_get_parent() local

Quoted 6 lines · Andy Shevchenko
Quoted 2 lines · Sakari Ailus

acpi_node_get_parent() isn't used outside drivers/acpi/property.c. Make it local.

Always in favour for such patches!

Thanks, Andy!

Signature · 1 line
Sakari Ailus
#13 lore raw
In reply to Sakari Ailus

added CC: printk maintainer (Petr Mladek)

Quoted 32 lines · Sakari Ailus

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f

...

Quoted 5 lines

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

Well, before my work it was vprintk_emit() that was disabling local interrupts. So this has always been broken.

Really it should be:

Fixes 3bd32d6a2ee6 ("lib/vsprintf: Add %pfw conversion specifier for

printing fwnode names")

Regardless, the fix should go into 5.10 and 5.14 stables.

John Ogness

#14 lore raw
Quoted 35 lines · John Ogness

added CC: printk maintainer (Petr Mladek)

Quoted 24 lines · Sakari Ailus

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
Quoted 8 lines
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f
Quoted 6 lines

...

Quoted 5 lines

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

It looks like a generic problem.

If I get it properly, we should make sure that any struct fwnode_operations implementation will _not_ use sleeping locks in the .get_parent() callback. Or anything that is called indirectly from from vsprintf.

Adding Andy into Cc.

Quoted 7 lines

Well, before my work it was vprintk_emit() that was disabling local interrupts. So this has always been broken.

Really it should be:

Fixes 3bd32d6a2ee6 ("lib/vsprintf: Add %pfw conversion specifier for

printing fwnode names")

Yes, I consider this to be the culprit.

Quoted 1 line

Regardless, the fix should go into 5.10 and 5.14 stables.

Please add the Fixes tag and the following into the commit message:

Best Regards,
Petr

#15 lore raw
Quoted 41 lines · Petr Mladek
Quoted 33 lines · John Ogness
Quoted 24 lines · Sakari Ailus

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
Quoted 8 lines
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f
Quoted 6 lines

...

Quoted 5 lines

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

Quoted 6 lines

It looks like a generic problem.

If I get it properly, we should make sure that any struct fwnode_operations implementation will _not_ use sleeping locks in the .get_parent() callback. Or anything that is called indirectly from from vsprintf.

Not sure how I may help here, but if there is a requirement to all *printf() to be non-sleepable, then yes, this should be done in that way. I'm not sure if every single %p extension does this...

It may require to update fwnode (source code) documentation to reflect the requirement. Also might_sleep() / migth_sleep_if() in the API calls may shed a light on (potential) issues.

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#16 lore raw
In reply to Petr Mladek

Hi Petr, John,

Quoted 43 lines · Petr Mladek
Quoted 35 lines · John Ogness

added CC: printk maintainer (Petr Mladek)

Quoted 24 lines · Sakari Ailus

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
Quoted 8 lines
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f
Quoted 6 lines

...

Quoted 5 lines

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

Quoted 8 lines

It looks like a generic problem.

If I get it properly, we should make sure that any struct fwnode_operations implementation will _not_ use sleeping locks in the .get_parent() callback. Or anything that is called indirectly from from vsprintf.

Adding Andy into Cc.

That holds for OF and swnode already, ACPI was an exception to this.

Quoted 3 lines
Quoted 3 lines

Well, before my work it was vprintk_emit() that was disabling local interrupts. So this has always been broken.

Fair enough.

Quoted 6 lines
Quoted 5 lines

Really it should be:

Fixes 3bd32d6a2ee6 ("lib/vsprintf: Add %pfw conversion specifier for

printing fwnode names")

Quoted 7 lines

Yes, I consider this to be the culprit.

Quoted 1 line

Regardless, the fix should go into 5.10 and 5.14 stables.

Please add the Fixes tag and the following into the commit message:

I'll use that in v2.

Signature · 3 lines
Regards,

Sakari Ailus
#17 lore raw
In reply to John Ogness

(and really add Andy)

Quoted 35 lines · John Ogness

added CC: printk maintainer (Petr Mladek)

Quoted 24 lines · Sakari Ailus

This set changes getting fwnode's parent on ACPI fwnode so it no longer needs a semaphore, using struct acpi_device->parent field instead of calling acpi_get_parent(). The semaphore is being acquired when the device's full path is printed which now takes place local IRQs disabled:

--------8<------------------------

BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163

...

Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 __might_resched.cold+0xf4/0x12f
 down_timeout+0x21/0x70
 acpi_os_wait_semaphore+0x63/0x180
 acpi_ut_acquire_mutex+0x123/0x1ba
 acpi_get_parent+0x30/0x71
 acpi_node_get_parent+0x64/0x90
 ? lock_acquire+0x1a0/0x300
 fwnode_count_parents+0x6d/0xb0
 fwnode_full_name_string+0x18/0x90
 fwnode_string+0xd7/0x140
 vsnprintf+0x1ec/0x4f0
Quoted 8 lines
 va_format.constprop.0+0x6a/0x130
 vsnprintf+0x1ec/0x4f0
 vprintk_store+0x271/0x5a0
 ? rcu_read_lock_sched_held+0x12/0x70
 ? lock_release+0x228/0x310
 ? acpi_initialize_hp_context+0x50/0x50
 vprintk_emit+0xd5/0x340
 _printk+0x58/0x6f
Quoted 6 lines

...

Quoted 5 lines

--------8<------------------------

I guess one could argue it wasn't great to begin with that getting fwnode's parent required a semaphore to begin with, nevertheless John's patch made it a concrete problem. Added Cc: stable, too.

It looks like a generic problem.

If I get it properly, we should make sure that any struct fwnode_operations implementation will _not_ use sleeping locks in the .get_parent() callback. Or anything that is called indirectly from from vsprintf.

Adding Andy into Cc.

Quoted 7 lines

Well, before my work it was vprintk_emit() that was disabling local interrupts. So this has always been broken.

Really it should be:

Fixes 3bd32d6a2ee6 ("lib/vsprintf: Add %pfw conversion specifier for

printing fwnode names")

Yes, I consider this to be the culprit.

Quoted 1 line

Regardless, the fix should go into 5.10 and 5.14 stables.

Please add the Fixes tag and the following into the commit message:

Best Regards,
Petr