Re: [PATCH v3] i2c: dev: don't allow user-space to deadlock the kernel

[Wolfram Sang]

On Wed, Jan 18, 2023 at 02:49:40PM +0100, Bartosz Golaszewski wrote:
> From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> 
> If we open an i2c character device and then unbind the underlying i2c
> adapter (either by unbinding it manually via sysfs or - for a real-life
> example - when unplugging a USB device with an i2c adaper), the kernel
> thread calling i2c_del_adapter() will become blocked waiting for the
> completion that only completes once all references to the character
> device get dropped.
> 
> In order to fix that, we introduce a couple changes. They need to be
> part of a single commit in order to preserve bisectability. First, drop
> the dev_release completion. That removes the risk of a deadlock but
> we now need to protect the character device structures against NULL
> pointer dereferences. To that end introduce an rw semaphore. It will
> protect the dummy i2c_client structure against dropping the adapter from
> under it. It will be taken for reading by all file_operations callbacks
> and for writing by the notifier's unbind handler. This way we don't
> prohibit the syscalls that don't get in each other's way from running
> concurrently but the adapter will not be unbound before all syscalls
> return.
> 
> Finally: upon being notified about an unbind event for the i2c adapter,
> we take the lock for writing and set the adapter pointer in the character
> device's structure to NULL. This "numbs down" the device - it still exists
> but is no longer functional. Meanwhile every syscall callback checks that
> pointer after taking the lock but before executing any code that requires
> it. If it's NULL, we return an error to user-space.
> 
> This way we can safely open an i2c device from user-space, unbind the
> device without triggering a deadlock and any subsequent system-call for
> the file descriptor associated with the removed adapter will gracefully
> fail.
> 
> Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> ---
> v1 -> v2:
> - keep the device release callback and use it to free the IDR number
> - rebase on top of v6.2-rc1
> 
> v2 -> v3:
> - make symbol names more descriptive
> - protect the name_show() sysfs callback too
> - zero the adapter's struct device on device release
> - make sure the code works nicely with CONFIG_DEBUG_KOBJECT_RELEASE enabled

So, this code handled all my stress-testing well so far. I'll try to
think of some more ideas until this evening, but likely I will apply it
later. Nonetheless, more review eyes are still welcome!

Attachment: signature.asc
Description: PGP signature