Re: [PATCH] cxl/memdev: fix deadlock in cxl_memdev_autoremove() on attach failure

[dan.j.williams]

Gregory Price wrote:
> cxl_memdev_autoremove() takes device_lock(&cxlmd->dev) via guard(device)
> and then calls cxl_memdev_unregister() when the attach callback was
> provided but cxl_mem_probe() failed to bind.
> 
> cxl_memdev_unregister() calls
>   cdev_device_del()
>     device_del()
>       bus_remove_device()
>         device_release_driver()
> 
> which also takes device_lock(), deadlocking the calling thread.
> 
> This path is reached when a driver uses the @attach parameter to
> devm_cxl_add_memdev() and the CXL topology fails to enumerate (e.g.
> DVSEC range registers decode outside platform-defined CXL ranges,
> causing the endpoint port probe to fail).
> 
> Fix by using scoped_guard() and breaking out of the guard scope before
> calling cxl_memdev_unregister(), so device_lock() is released first.

Whoops, yes, good catch, but...

> 
> Fixes: 29317f8dc6ed ("cxl/mem: Introduce cxl_memdev_attach for CXL-dependent operation")
> Signed-off-by: Gregory Price <gourry@xxxxxxxxxx>
> ---
>  drivers/cxl/core/memdev.c | 25 ++++++++++++++-----------
>  1 file changed, 14 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/cxl/core/memdev.c b/drivers/cxl/core/memdev.c
> index af3d0cc65138..c0de767b24fb 100644
> --- a/drivers/cxl/core/memdev.c
> +++ b/drivers/cxl/core/memdev.c
> @@ -1098,19 +1098,22 @@ static struct cxl_memdev *cxl_memdev_autoremove(struct cxl_memdev *cxlmd)
>  	 * return. Note that failure here could be the result of a race to
>  	 * teardown the CXL port topology. I.e. cxl_mem_probe() could have
>  	 * succeeded and then cxl_mem unbound before the lock is acquired.
> +	 *
> +	 * Check under device_lock but unregister outside of it, as
> +	 * cxl_memdev_unregister() will also take the device lock.
>  	 */
> -	guard(device)(&cxlmd->dev);
> -	if (cxlmd->attach && !cxlmd->dev.driver) {
> -		cxl_memdev_unregister(cxlmd);
> -		return ERR_PTR(-ENXIO);

...please just make a helper cxl_memdev_did_attach(), and move the
existing comment there. No need for the addtional comment that just adds
narration.

> +	scoped_guard(device, &cxlmd->dev) {

In general, I feel that scoped_guard() is almost always the wrong choice
for existing code and a helper with plain guard() is preferred.

> +		if (cxlmd->attach && !cxlmd->dev.driver)
> +			break;
> +
> +		rc = devm_add_action_or_reset(cxlmd->cxlds->dev,
> +					      cxl_memdev_unregister, cxlmd);

Just for completeness, no need to hold the lock over this. The helper
can just check ->attach and ->dev.driver.