Re: [RFC V1 2/4] iommufd: Lock all objects

[Jason Gunthorpe <jgg@nvidia.com>]

On Sat, Jul 20, 2024 at 11:56:43AM -0700, Steve Sistare wrote:
> +static int iommufd_take_all_iova_rwsem(struct iommufd_ctx *ictx,
> +				       struct xarray *ioas_list)
> +{
> +	struct iommufd_object *obj;
> +	unsigned long index;
> +	int rc;
> +
> +	/*
> +	 * This is very ugly, it is done instead of adding a lock around
> +	 * pages->source_mm, which is a performance path for mdev, we just
> +	 * obtain the write side of all the iova_rwsems which also protects the
> +	 * pages->source_*. Due to copies we can't know which IOAS could read
> +	 * from the pages, so we just lock everything. This is the only place
> +	 * locks are nested and they are uniformly taken in ID order.
> +	 *
> +	 * ioas_creation_lock prevents new IOAS from being installed in the
> +	 * xarray while we do this, and also prevents more than one thread from
> +	 * holding nested locks.
> +	 */
> +	down_write(&ictx->ioas_creation_lock);
> +	xa_lock(&ictx->objects);
> +	lockdep_off();

IIRC we used the nested annotation on the down_write instead of the
global lockdep_off? See how mm_take_all_locks() did a similar madness

Jason