[PATCH] fsi/master: Clarify master lifetimes & fix use-after-free in hub master

Christopher Bostic cbostic at linux.vnet.ibm.com
Tue Oct 3 12:11:49 AEDT 2017


Tested-by: Christopher Bostic <cbostic at linux.vnet.ibm.com>


On 10/2/17 7:32 PM, Jeremy Kerr wrote:
> Once we call fsi_master_unregister, the core will put_device,
> potentially freeing the hub master. This change adds a comment
> explaining the lifetime of an allocated fsi_master.
>
> We then add a reference from the driver to the hub master, so it stays
> around until we've finished ->remove().
>
> Signed-off-by: Jeremy Kerr <jk at ozlabs.org>
> ---
>   drivers/fsi/fsi-master-hub.c | 19 ++++++++++++++++---
>   drivers/fsi/fsi-master.h     | 15 +++++++++++++++
>   2 files changed, 31 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/fsi/fsi-master-hub.c b/drivers/fsi/fsi-master-hub.c
> index 5e4cd31..614d5f5 100644
> --- a/drivers/fsi/fsi-master-hub.c
> +++ b/drivers/fsi/fsi-master-hub.c
> @@ -288,10 +288,19 @@ static int hub_master_probe(struct device *dev)
>   	hub_master_init(hub);
>
>   	rc = fsi_master_register(&hub->master);
> -	if (!rc)
> -		return 0;
> +	if (rc)
> +		goto err_release;
> +
> +	/* At this point, fsi_master_register performs the device_initialize(),
> +	 * and holds the sole reference on master.dev. This means the device
> +	 * will be freed (via ->release) during any subsequent call to
> +	 * fsi_master_unregister.  We add our own reference to it here, so we
> +	 * can perform cleanup (in _remove()) without it being freed before
> +	 * we're ready.
> +	 */
> +	get_device(&hub->master.dev);
> +	return 0;
>
> -	kfree(hub);
>   err_release:
>   	fsi_slave_release_range(fsi_dev->slave, FSI_HUB_LINK_OFFSET,
>   			FSI_HUB_LINK_SIZE * links);
> @@ -306,6 +315,10 @@ static int hub_master_remove(struct device *dev)
>   	fsi_slave_release_range(hub->upstream->slave, hub->addr, hub->size);
>   	of_node_put(hub->master.dev.of_node);
>
> +	/* master.dev will likely be ->release()ed after this, which free()s
> +	 * the hub */
> +	put_device(&hub->master.dev);
> +
>   	return 0;
>   }
>
> diff --git a/drivers/fsi/fsi-master.h b/drivers/fsi/fsi-master.h
> index 546257d..1530ab2 100644
> --- a/drivers/fsi/fsi-master.h
> +++ b/drivers/fsi/fsi-master.h
> @@ -37,6 +37,21 @@ struct fsi_master {
>
>   #define dev_to_fsi_master(d) container_of(d, struct fsi_master, dev)
>
> +/**
> + * fsi_master registration & lifetime: the fsi_master_register() and
> + * fsi_master_unregister() functions will take ownership of the master, and
> + * ->dev in particular. The registration path performs a get_device(), which
> + * takes the first reference on the device. Similarly, the unregistration path
> + * performs a put_device(), which may well drop the last reference.
> + *
> + * This means that master implementations *may* need to hold their own
> + * reference (via get_device()) on master->dev. In particular, if the device's
> + * ->release callback frees the fsi_master, then fsi_master_unregister will
> + * invoke this free if no other reference is held.
> + *
> + * The same applies for the error path of fsi_master_register; if the call
> + * fails, dev->release will have been invoked.
> + */
>   extern int fsi_master_register(struct fsi_master *master);
>   extern void fsi_master_unregister(struct fsi_master *master);
>



More information about the openbmc mailing list