Skip to contentSkip to conversation

misc: ocxl: fix possible double free in ocxl_file_register_afu

PATCH Applied Hangyu Hua 4 participants 8 messages Apr 18, 2022 – May 15, 2022

Subsystems (2)
lore
Share & tools
lore

Hangyu Hua

author
#0 lore raw

PATCH 1/1 misc: ocxl: fix possible double free in ocxl_file_register_afu

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")
---
 drivers/misc/ocxl/file.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +2 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)
 
 err_unregister:
 	ocxl_sysfs_unregister_afu(info); // safe to call even if register failed
+	free_minor(info);
 	device_unregister(&info->dev);
+	return rc;
 err_put:
 	ocxl_afu_put(afu);
 	free_minor(info);
Signature · 1 line
2.25.1
Acked-by on v1 1/1
#1 lore raw
Quoted 9 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---

Thanks for fixing that error path!
I'm now thinking it would be cleaner to have the call to free_minor() in the info_release() callback but that would be another patch. In any case:

Fred

Quoted 17 lines
  drivers/misc/ocxl/file.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +2 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)
 
  err_unregister:
  	ocxl_sysfs_unregister_afu(info); // safe to call even if register failed
+	free_minor(info);
  	device_unregister(&info->dev);
+	return rc;
  err_put:
  	ocxl_afu_put(afu);
  	free_minor(info);
author
#2 lore raw
Quoted 14 lines · Frederic Barrat
Quoted 9 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")

---

Quoted 7 lines

Thanks for fixing that error path!
I'm now thinking it would be cleaner to have the call to free_minor() in the info_release() callback but that would be another patch. In any case:

  Fred

I think it is a good idea to use callbacks to handle all garbage collections. And free_minor is used only in ocxl_file_register_afu() andocxl_file_unregister_afu(). So this should only require minor changes.

Thanks.

Quoted 17 lines
Quoted 17 lines

  drivers/misc/ocxl/file.c | 2 ++

  1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +0 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)

  err_unregister:
      ocxl_sysfs_unregister_afu(info); // safe to call even if register failed +    free_minor(info);
      device_unregister(&info->dev);
+    return rc;
  err_put:
      ocxl_afu_put(afu);
      free_minor(info);

#3 lore raw
In reply to Hangyu Hua
Quoted 3 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Double frees are often exploitable. But it looks to me like this error path is not easily reachable by an attacker.

ocxl_file_register_afu() is only called from ocxl_probe(), and we only go to err_unregister if the sysfs or cdev initialisation fails, which should only happen if we hit ENOMEM, or we have a duplicate device which would be a device-tree/hardware error. But maybe Fred can check more closely, I don't know the driver that well.

cheers

Quoted 24 lines

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")
---
 drivers/misc/ocxl/file.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +2 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)
 
 err_unregister:
 	ocxl_sysfs_unregister_afu(info); // safe to call even if register failed
+	free_minor(info);
 	device_unregister(&info->dev);
+	return rc;
 err_put:
 	ocxl_afu_put(afu);
 	free_minor(info);
Signature · 1 line
2.25.1
author
#4 lore raw
Quoted 8 lines · Michael Ellerman
Quoted 3 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Double frees are often exploitable. But it looks to me like this error path is not easily reachable by an attacker.

Quoted 8 lines

ocxl_file_register_afu() is only called from ocxl_probe(), and we only go to err_unregister if the sysfs or cdev initialisation fails, which should only happen if we hit ENOMEM, or we have a duplicate device which would be a device-tree/hardware error. But maybe Fred can check more closely, I don't know the driver that well.

cheers

Hi Michael,

You are right. It is hard to exploit at least in my understanding. That's why I didn't give this to security@. But it still need to be fix out. By the way, if you are interesting in this kind of bug, you can check out the other three patches I submitted recently.

rpmsg: virtio: fix possible double free in rpmsg_virtio_add_ctrl_dev()
https://lore.kernel.org/all/[email protected]/

rpmsg: virtio: fix possible double free in rpmsg_probe()
https://lore.kernel.org/all/[email protected]/

hwtracing: stm: fix possible double free in stm_register_device()
https://lore.kernel.org/all/[email protected]/

They are all the similar bugs i could find in the kernel.

Thanks.

Quoted 25 lines
Quoted 25 lines

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")
---
  drivers/misc/ocxl/file.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +2 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)
 
  err_unregister:
  	ocxl_sysfs_unregister_afu(info); // safe to call even if register failed
+	free_minor(info);
  	device_unregister(&info->dev);
+	return rc;
  err_put:
  	ocxl_afu_put(afu);
  	free_minor(info);
Signature · 1 line
2.25.1
#5 lore raw
In reply to Michael Ellerman
Quoted 5 lines · Michael Ellerman
Quoted 3 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Quoted 8 lines

Double frees are often exploitable. But it looks to me like this error path is not easily reachable by an attacker.

ocxl_file_register_afu() is only called from ocxl_probe(), and we only go to err_unregister if the sysfs or cdev initialisation fails, which should only happen if we hit ENOMEM, or we have a duplicate device which would be a device-tree/hardware error. But maybe Fred can check more closely, I don't know the driver that well.

The linux devices built here are based on what is parsed on the physical devices. Those could be FPGAs but updating the FPGA image requires root privilege. In any case, duplicate AFU names are possible, that's why the driver adds an index (the afu->config.idx part of the name) to the linux device name. So we would need to mess that up in the driver as well to have a duplicate device name. So I would agree the double free is hard to hit.

mpe: I think this patch can be taken as is. The "beautification" I talked about is just that and I don't intend to work on it except if something else shows up.

Fred

Quoted 27 lines

cheers

Quoted 24 lines

Fix this by adding free_minor() and return to err_unregister error path.

Fixes 75ca758adbaf ("ocxl: Create a clear delineation between ocxl backend & frontend")
---
  drivers/misc/ocxl/file.c | 2 ++
  1 file changed, 2 insertions(+)
drivers/misc/ocxl/file.c +2 -0
index d881f5e40ad9..6777c419a8da 100644
@@ -556,7 +556,9 @@ int ocxl_file_register_afu(struct ocxl_afu *afu)
 
  err_unregister:
  	ocxl_sysfs_unregister_afu(info); // safe to call even if register failed
+	free_minor(info);
  	device_unregister(&info->dev);
+	return rc;
  err_put:
  	ocxl_afu_put(afu);
  	free_minor(info);
Signature · 1 line
2.25.1
#6 lore raw
Quoted 15 lines · Frederic Barrat
Quoted 5 lines · Michael Ellerman
Quoted 3 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Quoted 8 lines

Double frees are often exploitable. But it looks to me like this error path is not easily reachable by an attacker.

ocxl_file_register_afu() is only called from ocxl_probe(), and we only go to err_unregister if the sysfs or cdev initialisation fails, which should only happen if we hit ENOMEM, or we have a duplicate device which would be a device-tree/hardware error. But maybe Fred can check more closely, I don't know the driver that well.

Quoted 7 lines

The linux devices built here are based on what is parsed on the physical devices. Those could be FPGAs but updating the FPGA image requires root privilege. In any case, duplicate AFU names are possible, that's why the driver adds an index (the afu->config.idx part of the name) to the linux device name. So we would need to mess that up in the driver as well to have a duplicate device name.
So I would agree the double free is hard to hit.

Thanks for confirming.

Quoted 3 lines

mpe: I think this patch can be taken as is. The "beautification" I talked about is just that and I don't intend to work on it except if something else shows up.

OK, will pick this up.

cheers

#7 lore raw
In reply to Hangyu Hua
Quoted 8 lines · Hangyu Hua

info_release() will be called in device_unregister() when info->dev's reference count is 0. So there is no need to call ocxl_afu_put() and kfree() again.

Fix this by adding free_minor() and return to err_unregister error path.

[...]

Applied to powerpc/next.

[1/1] misc: ocxl: fix possible double free in ocxl_file_register_afu

      https://git.kernel.org/powerpc/c/950cf957fe34d40d63dfa3bf3968210430b6491e

cheers