[PATCH] powerpc/iommu: TCEs are incorrectly manipulated with DLPAR add/remove of memory

[Gaurav Batra]

Hello Michael,

I found this bug while going though the code. This bug is exposed when 
DDW is smaller than the max memory of the LPAR. This will result in 
creating DDW which will have Dynamically mapped TCEs (no direct mapping).

I would like to stress that this  bug is exposed only in Upstream 
kernel. Current kernel level in Distros are not exposed to this since 
they don't have the  concept of "dynamically mapped" DDW.

I didn't have access to any of the P10 boxes with large amount of memory 
to  re-create the scenario. On P10 we have 2MB TCEs, which results in 
DDW large enough to be able to cover  max memory I could have for the 
LPAR. As a result,  IO Bus Addresses generated were always within DDW 
limits and no H_PARAMETER was returned by HCALL.

So, I hacked the kernel to force the use of 64K TCEs. This resulted in 
DDW smaller than max memory.

When I tried to DLPAR ADD memory, it failed with error code of -4 
(H_PARAMETER) from HCALL (H_PUT_TCE/H_PUT_TCE_INDIRECT), when 
iommu_mem_notifier() invoked tce_setrange_multi_pSeriesLP().

I didn't test the DLPAR REMOVE path, to verify if incorrect TCEs are 
removed by tce_clearrange_multi_pSeriesLP(), since I would need to hack 
kernel to force dynamically added TCEs to the high IO Bus Addresses. 
But, the concept is  same.

Thanks,

Gaurav

On 6/13/23 12:16 PM, Gaurav Batra wrote:
> When memory is dynamically added/removed, iommu_mem_notifier() is invoked. This
> routine traverses through all the DMA windows (DDW only, not default windows)
> to add/remove "direct" TCE mappings. The routines for this purpose are
> tce_clearrange_multi_pSeriesLP() and tce_clearrange_multi_pSeriesLP().
>
> Both these routines are designed for Direct mapped DMA windows only.
>
> The issue is that there could be some DMA windows in the list which are not
> "direct" mapped. Calling these routines will either,
>
> 1) remove some dynamically mapped TCEs, Or
> 2) try to add TCEs which are out of bounds and HCALL returns H_PARAMETER
>
> Here are the side affects when these routines are incorrectly invoked for
> "dynamically" mapped DMA windows.
>
> tce_setrange_multi_pSeriesLP()
>
> This adds direct mapped TCEs. Now, this could invoke HCALL to add TCEs with
> out-of-bound range. In this scenario, HCALL will return H_PARAMETER and DLAR
> ADD of memory will fail.
>
> tce_clearrange_multi_pSeriesLP()
>
> This will remove range of TCEs. The TCE range that is calculated, depending on
> the memory range being added, could infact be mapping some other memory
> address (for dynamic DMA window scenario). This will wipe out those TCEs.
>
> The solution is for iommu_mem_notifier() to only invoke these routines for
> "direct" mapped DMA windows.
>
> Signed-off-by: Gaurav Batra <gbatra at linux.vnet.ibm.com>
> Reviewed-by: Brian King <brking at linux.vnet.ibm.com>
> ---
>   arch/powerpc/platforms/pseries/iommu.c | 17 +++++++++++++----
>   1 file changed, 13 insertions(+), 4 deletions(-)
>
> diff --git a/arch/powerpc/platforms/pseries/iommu.c b/arch/powerpc/platforms/pseries/iommu.c
> index 918f511837db..24dd61636400 100644
> --- a/arch/powerpc/platforms/pseries/iommu.c
> +++ b/arch/powerpc/platforms/pseries/iommu.c
> @@ -363,6 +363,7 @@ struct dynamic_dma_window_prop {
>   struct dma_win {
>   	struct device_node *device;
>   	const struct dynamic_dma_window_prop *prop;
> +	bool    direct;
>   	struct list_head list;
>   };
>
> @@ -1409,6 +1410,8 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
>   		goto out_del_prop;
>
>   	if (direct_mapping) {
> +		window->direct = true;
> +
>   		/* DDW maps the whole partition, so enable direct DMA mapping */
>   		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
>   					    win64->value, tce_setrange_multi_pSeriesLP_walk);
> @@ -1425,6 +1428,8 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
>   		int i;
>   		unsigned long start = 0, end = 0;
>
> +		window->direct = false;
> +
>   		for (i = 0; i < ARRAY_SIZE(pci->phb->mem_resources); i++) {
>   			const unsigned long mask = IORESOURCE_MEM_64 | IORESOURCE_MEM;
>
> @@ -1587,8 +1592,10 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
>   	case MEM_GOING_ONLINE:
>   		spin_lock(&dma_win_list_lock);
>   		list_for_each_entry(window, &dma_win_list, list) {
> -			ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
> -					arg->nr_pages, window->prop);
> +			if (window->direct) {
> +				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
> +						arg->nr_pages, window->prop);
> +			}
>   			/* XXX log error */
>   		}
>   		spin_unlock(&dma_win_list_lock);
> @@ -1597,8 +1604,10 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
>   	case MEM_OFFLINE:
>   		spin_lock(&dma_win_list_lock);
>   		list_for_each_entry(window, &dma_win_list, list) {
> -			ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
> -					arg->nr_pages, window->prop);
> +			if (window->direct) {
> +				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
> +						arg->nr_pages, window->prop);
> +			}
>   			/* XXX log error */
>   		}
>   		spin_unlock(&dma_win_list_lock);