Cell: shutdown method for spu_sysdev_class
Subsystems (2)
Quoted 15 lines · Geoff Levand
arch/powerpc/platforms/cell/spu_base.c +9
-0@@ -463,8 +463,21 @@ void spu_free(struct spu *spu) }EXPORT_SYMBOL_GPL (spu_free); +static int spu_shutdown (struct sys_device *sysdev) +{ + struct spu *spu = container_of(sysdev, struct spu, sysdev); + +// what else here??? + +spu_free_irqs (spu); +spu_destroy_spu (spu); +kfree (spu);
arch/powerpc/platforms/cell/spu_base.c void spu_free(struct spu *spu) · line 483
Quoted 10 lines
arch/powerpc/platforms/cell/spu_base.c +5
-1+return 0 ; +} +struct sysdev_class spu_sysdev_class = { -set_kset_name ("spu" ) +set_kset_name ("spu" ), + .shutdown = spu_shutdown, };int spu_add_sysdev_attr (struct sysdev_attribute *attr)
After some debugging, I found that this patch creates an oops when slab debugging is enabled, the reason for that being that sysdev_shutdown() iterates over all system devices using list_for_each_entry(), not list_for_each_entry_safe().
There are two ways of fixing this:
- use list_for_each_entry_safe() to go over all devices so they can be freed in their ->shutdown method.
- not free the device, because we know the system is going down anyway.
Is it documented or implied somewhere that ->shutdown must not free the device? If not, the first option is probably the safer choice.
Arnd <><
Quoted 34 lines · Arnd Bergmann
Quoted 15 lines · Geoff Levand
arch/powerpc/platforms/cell/spu_base.c +9
-0@@ -463,8 +463,21 @@ void spu_free(struct spu *spu) }EXPORT_SYMBOL_GPL (spu_free); +static int spu_shutdown (struct sys_device *sysdev) +{ + struct spu *spu = container_of(sysdev, struct spu, sysdev); + +// what else here??? + +spu_free_irqs (spu); +spu_destroy_spu (spu); +kfree (spu);arch/powerpc/platforms/cell/spu_base.c void spu_free(struct spu *spu) · line 483
Quoted 10 lines
arch/powerpc/platforms/cell/spu_base.c +5
-1+return 0 ; +} +struct sysdev_class spu_sysdev_class = { -set_kset_name ("spu" ) +set_kset_name ("spu" ), + .shutdown = spu_shutdown, };int spu_add_sysdev_attr (struct sysdev_attribute *attr)After some debugging, I found that this patch creates an oops when slab debugging is enabled, the reason for that being that
sysdev_shutdown()iterates over all system devices usinglist_for_each_entry(), notlist_for_each_entry_safe().There are two ways of fixing this:
Quoted 7 lines
- use
list_for_each_entry_safe()to go over all devices so they can be freed in their ->shutdown method.
- not free the device, because we know the system is going down anyway.Is it documented or implied somewhere that ->shutdown must not free the device? If not, the first option is probably the safer choice.
It's not documented anywhere, but implicitly assumed. I don't know of any shutdown implementation that frees the device.
Quoted 6 lines · Christoph Hellwig
Quoted 3 lines
Is it documented or implied somewhere that ->shutdown must not free the device? If not, the first option is probably the safer choice.
It's not documented anywhere, but implicitly assumed. I don't know of any shutdown implementation that frees the device.
Right, I found the explanation now myself:
The ->shutdown method is called for system devices that are still part of a linked linked list. Freeing the object would destroy that list.
Geoff, please merge the patch below in your tree.
Arnd <><
---
Subject: cell: don't free spu objects in sysdev shutdown
From: Arnd Bergmann <[email protected]> System devices are accessed after they are shut down, so we must not free the data structures.
arch/powerpc/platforms/cell/spu_base.c +0 -1
@@ -517,7 +517,6 @@ static int spu_shutdown(struct sys_devicspu_free_irqs (spu);spu_destroy_spu (spu); -kfree (spu);return 0 ; }