[kernel] KVM: PPC: Fix hardware and emulated TCE tables matching

When attaching a hardware table to LIOBN in KVM, we match table parameters
such as page size, table offset and table size. However the tables are
created via very different paths - VFIO and KVM - and the VFIO path goes
through the platform code which has minimum TCE page size requirement
(which is 4K but since we allocate memory by pages and cannot avoid
alignment anyway, we align to 64k pages for powernv_defconfig).

So when we match the tables, one might be bigger that the other which
means the hardware table cannot get attached to LIOBN and DMA mapping
fails.

This removes the table size alignment from the guest visible table.
This does not affect the memory allocation which is still aligned -
kvmppc_tce_pages() takes care of this.

This relaxes the check we do when attaching tables to allow the hardware
table be bigger than the guest visible table.

Ideally we want the KVM table to cover the same space as the hardware
table does but since the hardware table may use multiple levels, and
all levels must use the same table size (IODA2 design), the area it can
actually cover might get very different from the window size which
the guest requested, even though the guest won't map it all.

Fixes: ca1fc489cf "KVM: PPC: Book3S: Allow backing bigger guest IOMMU pages with smaller physical pages"
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---
 arch/powerpc/kvm/book3s_64_vio.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

On Wed, Jun 20, 2018 at 06:42:58PM +1000, Alexey Kardashevskiy wrote:
> When attaching a hardware table to LIOBN in KVM, we match table parameters
> such as page size, table offset and table size. However the tables are
> created via very different paths - VFIO and KVM - and the VFIO path goes
> through the platform code which has minimum TCE page size requirement
> (which is 4K but since we allocate memory by pages and cannot avoid
> alignment anyway, we align to 64k pages for powernv_defconfig).
> 
> So when we match the tables, one might be bigger that the other which
> means the hardware table cannot get attached to LIOBN and DMA mapping
> fails.
> 
> This removes the table size alignment from the guest visible table.
> This does not affect the memory allocation which is still aligned -
> kvmppc_tce_pages() takes care of this.
> 
> This relaxes the check we do when attaching tables to allow the hardware
> table be bigger than the guest visible table.
> 
> Ideally we want the KVM table to cover the same space as the hardware
> table does but since the hardware table may use multiple levels, and
> all levels must use the same table size (IODA2 design), the area it can
> actually cover might get very different from the window size which
> the guest requested, even though the guest won't map it all.
> 
> Fixes: ca1fc489cf "KVM: PPC: Book3S: Allow backing bigger guest IOMMU pages with smaller physical pages"
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>

Reviewed-by: David Gibson <david@gibson.dropbear.id.au>

> ---
>  arch/powerpc/kvm/book3s_64_vio.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/powerpc/kvm/book3s_64_vio.c b/arch/powerpc/kvm/book3s_64_vio.c
> index 8c456fa..8167ce8 100644
> --- a/arch/powerpc/kvm/book3s_64_vio.c
> +++ b/arch/powerpc/kvm/book3s_64_vio.c
> @@ -180,7 +180,7 @@ extern long kvm_spapr_tce_attach_iommu_group(struct kvm *kvm, int tablefd,
>  		if ((tbltmp->it_page_shift <= stt->page_shift) &&
>  				(tbltmp->it_offset << tbltmp->it_page_shift ==
>  				 stt->offset << stt->page_shift) &&
> -				(tbltmp->it_size << tbltmp->it_page_shift ==
> +				(tbltmp->it_size << tbltmp->it_page_shift >=
>  				 stt->size << stt->page_shift)) {
>  			/*
>  			 * Reference the table to avoid races with
> @@ -296,7 +296,7 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
>  {
>  	struct kvmppc_spapr_tce_table *stt = NULL;
>  	struct kvmppc_spapr_tce_table *siter;
> -	unsigned long npages, size;
> +	unsigned long npages, size = args->size;
>  	int ret = -ENOMEM;
>  	int i;
>  
> @@ -304,7 +304,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
>  		(args->offset + args->size > (ULLONG_MAX >> args->page_shift)))
>  		return -EINVAL;
>  
> -	size = _ALIGN_UP(args->size, PAGE_SIZE >> 3);
>  	npages = kvmppc_tce_pages(size);
>  	ret = kvmppc_account_memlimit(kvmppc_stt_pages(npages), true);
>  	if (ret)

On Wed, Jun 20, 2018 at 06:42:58PM +1000, Alexey Kardashevskiy wrote:
> When attaching a hardware table to LIOBN in KVM, we match table parameters
> such as page size, table offset and table size. However the tables are
> created via very different paths - VFIO and KVM - and the VFIO path goes
> through the platform code which has minimum TCE page size requirement
> (which is 4K but since we allocate memory by pages and cannot avoid
> alignment anyway, we align to 64k pages for powernv_defconfig).
> 
> So when we match the tables, one might be bigger that the other which
> means the hardware table cannot get attached to LIOBN and DMA mapping
> fails.
> 
> This removes the table size alignment from the guest visible table.
> This does not affect the memory allocation which is still aligned -
> kvmppc_tce_pages() takes care of this.
> 
> This relaxes the check we do when attaching tables to allow the hardware
> table be bigger than the guest visible table.
> 
> Ideally we want the KVM table to cover the same space as the hardware
> table does but since the hardware table may use multiple levels, and
> all levels must use the same table size (IODA2 design), the area it can
> actually cover might get very different from the window size which
> the guest requested, even though the guest won't map it all.
> 
> Fixes: ca1fc489cf "KVM: PPC: Book3S: Allow backing bigger guest IOMMU pages with smaller physical pages"
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>

Thanks, patch applied to my kvm-ppc-next branch.

Paul.