Re: [PATCH v4 3/7] iommu/mediatek: Use gather to achieve the tlb range flush

[Will Deacon]

On Wed, Oct 16, 2019 at 11:33:08AM +0800, Yong Wu wrote:
> Use the iommu_gather mechanism to achieve the tlb range flush.
> Gather the iova range in the "tlb_add_page", then flush the merged iova
> range in iotlb_sync.
> 
> Suggested-by: Tomasz Figa <tfiga@xxxxxxxxxxxx>
> Signed-off-by: Yong Wu <yong.wu@xxxxxxxxxxxx>
> ---
>  drivers/iommu/mtk_iommu.c | 12 ++++++++----
>  1 file changed, 8 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/iommu/mtk_iommu.c b/drivers/iommu/mtk_iommu.c
> index c2f6c78..81ac95f 100644
> --- a/drivers/iommu/mtk_iommu.c
> +++ b/drivers/iommu/mtk_iommu.c
> @@ -245,11 +245,9 @@ static void mtk_iommu_tlb_flush_page_nosync(struct iommu_iotlb_gather *gather,
>  					    void *cookie)
>  {
>  	struct mtk_iommu_data *data = cookie;
> -	unsigned long flags;
> +	struct iommu_domain *domain = &data->m4u_dom->domain;
>  
> -	spin_lock_irqsave(&data->tlb_lock, flags);
> -	mtk_iommu_tlb_add_flush_nosync(iova, granule, granule, true, cookie);
> -	spin_unlock_irqrestore(&data->tlb_lock, flags);
> +	iommu_iotlb_gather_add_page(domain, gather, iova, granule);

You need to be careful here, because iommu_iotlb_gather_add_page() can
call iommu_tlb_sync() in some situations and you don't hold the lock.

>  static const struct iommu_flush_ops mtk_iommu_flush_ops = {
> @@ -469,9 +467,15 @@ static void mtk_iommu_iotlb_sync(struct iommu_domain *domain,
>  				 struct iommu_iotlb_gather *gather)
>  {
>  	struct mtk_iommu_data *data = mtk_iommu_get_m4u_data();
> +	size_t length = gather->end - gather->start;
>  	unsigned long flags;
>  
> +	if (gather->start == ULONG_MAX)
> +		return;
> +
>  	spin_lock_irqsave(&data->tlb_lock, flags);
> +	mtk_iommu_tlb_add_flush_nosync(gather->start, length, gather->pgsize,
> +				       false, data);
>  	mtk_iommu_tlb_sync(data);
>  	spin_unlock_irqrestore(&data->tlb_lock, flags);

Modulo my comment above, this fixes my previous comment. Given that mainline
is already broken, I guess the runtime bisectability isn't a problem.

Will