On Mon, Jan 12, 2015 at 03:06:20PM +0800, Li, Zhen-Hua wrote:
+
+#ifdef CONFIG_CRASH_DUMP
+
+/*
+ * Fix Crashdump failure caused by leftover DMA through a hardware IOMMU
+ *
+ * Fixes the crashdump kernel to deal with an active iommu and legacy
+ * DMA from the (old) panicked kernel in a manner similar to how legacy
+ * DMA is handled when no hardware iommu was in use by the old kernel --
+ * allow the legacy DMA to continue into its current buffers.
+ *
+ * In the crashdump kernel, this code:
+ * 1. skips disabling the IOMMU's translating of IO Virtual Addresses (IOVA).
+ * 2. Do not re-enable IOMMU's translating.
+ * 3. In kdump kernel, use the old root entry table.
+ * 4. Leaves the current translations in-place so that legacy DMA will
+ * continue to use its current buffers.
+ * 5. Allocates to the device drivers in the crashdump kernel
+ * portions of the iova address ranges that are different
+ * from the iova address ranges that were being used by the old kernel
+ * at the time of the panic.
+ *
+ */
It looks like you are still copying the io-page-tables from the old
kernel into the kdump kernel, is that right? With the approach that was
proposed you only need to copy over the context entries 1-1. They are
still pointing to the page-tables in the old kernels memory (which is
just fine).
The root-entry of the old kernel is also re-used, and when the kdump
kernel starts to use a device, its context entry is updated to point to
a newly allocated page-table.