Re: [PATCH v3] EDAC, ghes: Fix locking and memory barrier issues

[Robert Richter <rrichter@marvell.com>]

On 05.11.19 20:07:51, Robert Richter wrote:
> The ghes registration and refcount is broken in several ways:
> 
>  * ghes_edac_register() returns with success for a 2nd instance even
>    if a first instance is still running. This is not correct as the
>    first instance may fail later. A subsequent registration may not
>    finish before the first. Parallel registrations must be avoided.
> 
>  * The refcount was increased even if a registration failed. This
>    leads to stale counters preventing the device from being released.
> 
>  * The ghes refcount may not be decremented properly on
>    unregistration. Always decrement the refcount once
>    ghes_edac_unregister() is called to keep the refcount sane.
> 
>  * The ghes_pvt pointer is handed to the irq handler before
>    registration finished.
> 
>  * The mci structure could be freed while the irq handler is running.
> 
> Fix this by adding a mutex to ghes_edac_register(). This mutex
> serializes instances to register and unregister. The refcount is only
> increased if the registration succeeded. This makes sure the refcount
> is in a consistent state after registering or unregistering a device.
> Note: A spinlock cannot be used here as the code section may sleep.
> 
> The ghes_pvt is protected by ghes_lock now. This ensures the pointer
> is not updated before registration was finished or while the irq
> handler is running. It is unset before unregistering the device
> including necessary (implicit) memory barriers making the changes
> visible to other cpus. Thus, the device can not be used anymore by an
> interrupt.
> 
> Also, rename ghes_init to ghes_refcount for better readability and
> switch to refcount API.
> 
> A refcount is needed. There can be multiple GHES structures being
> defined (see ACPI 6.3 specification, 18.3.2.7 Generic Hardware Error
> Source, "Some platforms may describe multiple Generic Hardware Error
> Source structures with different notification types, ...").
> 
> Another approach to use the mci's device refcount (get_device()) and
> have a release function does not work here. A release function will be
> called only for device_release() with the last put_device() call. The
> device must be deleted *before* that with device_del(). This is only
> possible by maintaining an own refcount.
> 
> Fixes: 0fe5f281f749 ("EDAC, ghes: Model a single, logical memory controller")
> Fixes: 1e72e673b9d1 ("EDAC/ghes: Fix Use after free in ghes_edac remove path")
> Co-developed-by: James Morse <james.morse@arm.com>
> Signed-off-by: James Morse <james.morse@arm.com>
> Co-developed-by: Borislav Petkov <bp@suse.de>
> Signed-off-by: Borislav Petkov <bp@suse.de>
> Signed-off-by: Robert Richter <rrichter@marvell.com>

I hope this SOB chain is correct now.

Thanks,

-Robert

> ---
> V3:
>  * fixed SOB chain again (added SOBs according to
>    submitting-patches.rst)
> 
> V2:
>  * fixed missing 'goto unlock' in error path
>  * fixed SOB chain
>  * added comment on how to protect ghes_pvt with ghes_lock
>  * renamed ghes_init to ghes_refcount
>  * switched to refcount API instead of atomic_*
> ---
>  drivers/edac/ghes_edac.c | 89 +++++++++++++++++++++++++++++-----------
>  1 file changed, 65 insertions(+), 24 deletions(-)