Skip to contentSkip to conversation

uio: ensure class is registered before devices

PATCH In review Alexandre Belloni 2 participants 6 messages Aug 16, 2018 Userspace I/O (UIO)

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 uio: ensure class is registered before devices

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

This may result in a NULL pointer dereference later on in get_device_parent() when accessing the class glue_dirs spinlock.

The trace looks like that:

Unable to handle kernel NULL pointer dereference at virtual address 00000140
[...]
[<ffff0000089cc234>] _raw_spin_lock+0x14/0x48
[<ffff0000084f56bc>] device_add+0x154/0x6a0
[<ffff0000084f5e48>] device_create_groups_vargs+0x120/0x128
[<ffff0000084f5edc>] device_create+0x54/0x60
[<ffff0000086e72c0>] __uio_register_device+0x120/0x4a8
[<ffff000008528b7c>] jaguar2_pci_probe+0x2d4/0x558
[<ffff0000083fc18c>] local_pci_probe+0x3c/0xb8
[<ffff0000083fd81c>] pci_device_probe+0x11c/0x180
[<ffff0000084f88bc>] driver_probe_device+0x22c/0x2d8
[<ffff0000084f8a24>] __driver_attach+0xbc/0xc0
[<ffff0000084f69fc>] bus_for_each_dev+0x4c/0x98
[<ffff0000084f81b8>] driver_attach+0x20/0x28
[<ffff0000084f7d08>] bus_add_driver+0x1b8/0x228
[<ffff0000084f93c0>] driver_register+0x60/0xf8
[<ffff0000083fb918>] __pci_register_driver+0x40/0x48

Return EPROBE_DEFER in that case so the driver can register the device later.

---
 drivers/uio/uio.c | 9 +++++++++
 1 file changed, 9 insertions(+)
drivers/uio/uio.c +9 -0
index 5d421d7e8904..55b523f7499b 100644
@@ -275,6 +275,8 @@ static struct class uio_class = {
 	.dev_groups = uio_groups,
 };
 
+bool uio_class_registered;
+
 /*
  * device functions
  */
@@ -884,6 +886,9 @@ static int init_uio_class(void)
 		printk(KERN_ERR "class_register failed for uio\n");
 		goto err_class_register;
 	}
+
+	uio_class_registered = true;
+
 	return 0;
 
 err_class_register:
@@ -894,6 +899,7 @@ static int init_uio_class(void)
 
 static void release_uio_class(void)
 {
+	uio_class_registered = false;
 	class_unregister(&uio_class);
 	uio_major_cleanup();
 }
@@ -920,6 +926,9 @@ int __uio_register_device(struct module *owner,
 	struct uio_device *idev;
 	int ret = 0;
 
+	if (!uio_class_registered)
+		return -EPROBE_DEFER;
+
 	if (!parent || !info || !info->name || !info->version)
 		return -EINVAL;
 
Signature · 1 line
2.18.0
#1 lore raw
Quoted 2 lines · Alexandre Belloni

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

How does this happen? The link order should solve this issue, right?

thanks,

greg k-h

#2 lore raw
Quoted 6 lines · Greg Kroah-Hartman
Quoted 2 lines · Alexandre Belloni

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

How does this happen? The link order should solve this issue, right?

Sure, if we can ensure uio_init() is called before any driver calls uio_register_device() then this would not happen. However, I'm not sure how you would want to achieve that.

Signature · 3 lines
Alexandre Belloni, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#3 lore raw
Quoted 11 lines · Alexandre Belloni
Quoted 6 lines · Greg Kroah-Hartman
Quoted 2 lines · Alexandre Belloni

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

How does this happen? The link order should solve this issue, right?

Sure, if we can ensure uio_init() is called before any driver calls uio_register_device() then this would not happen. However, I'm not sure how you would want to achieve that.

That is the job of the link order, does this not work properly today? How have you triggered this so that you could test your patch?

thanks,

greg k-h

#4 lore raw
Quoted 9 lines · Greg Kroah-Hartman
Quoted 8 lines · Alexandre Belloni
Quoted 6 lines · Greg Kroah-Hartman
Quoted 2 lines · Alexandre Belloni

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

How does this happen? The link order should solve this issue, right?

Quoted 7 lines
Quoted 3 lines

Sure, if we can ensure uio_init() is called before any driver calls uio_register_device() then this would not happen. However, I'm not sure how you would want to achieve that.

That is the job of the link order, does this not work properly today? How have you triggered this so that you could test your patch?

I have a (not yet upstreamed) MFD driver in drivers/mfd that is registering the uio device. It mostly look like uio_pci_generic.c.

When compiling both as builtin, it will crash that way.

There are no link dependency between uio_init and uio_register_device calls so I guess the linker can't reorder that properly.

Signature · 3 lines
Alexandre Belloni, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#5 lore raw
Quoted 18 lines · Alexandre Belloni
Quoted 9 lines · Greg Kroah-Hartman
Quoted 8 lines · Alexandre Belloni
Quoted 6 lines · Greg Kroah-Hartman
Quoted 2 lines · Alexandre Belloni

When both uio and the uio drivers are built in the kernel, it is possible for a driver to register devices before the uio class is registered.

How does this happen? The link order should solve this issue, right?

Quoted 7 lines
Quoted 3 lines

Sure, if we can ensure uio_init() is called before any driver calls uio_register_device() then this would not happen. However, I'm not sure how you would want to achieve that.

That is the job of the link order, does this not work properly today? How have you triggered this so that you could test your patch?

Quoted 7 lines

I have a (not yet upstreamed) MFD driver in drivers/mfd that is registering the uio device. It mostly look like uio_pci_generic.c.

When compiling both as builtin, it will crash that way.

There are no link dependency between uio_init and uio_register_device calls so I guess the linker can't reorder that properly.

Ah, ok, yes, that will break, I was looking at the kernel tree _today_ :)

I'll queue this up after 4.19-rc1 is out, thanks.

greg k-h