feat: pv-iommu + tpu fixes

arch/x86/include/asm/xen/hypercall.h

@@ -497,6 +497,12 @@ HYPERVISOR_xenpmu_op(unsigned int op, void *arg)
 	return _hypercall2(int, xenpmu_op, op, arg);
 }
 
+static inline long
+HYPERVISOR_iommu_op(unsigned int subop, void *arg)
+{
+	return _hypercall2(int, iommu_op, subop, arg);
+}
+
 static inline int
 HYPERVISOR_dm_op(
 	domid_t dom, unsigned int nr_bufs, struct xen_dm_op_buf *bufs)

arch/x86/pci/xen.c

@@ -401,7 +401,24 @@ static void xen_teardown_msi_irqs(struct pci_dev *dev)
 
 static void xen_pv_teardown_msi_irqs(struct pci_dev *dev)
 {
-	if (dev->msix_enabled)
+	struct msi_desc *desc;
+	bool msix = false;
+
+	/*
+	 * pci_disable_msix() clears msix_enabled in pci_msix_shutdown() before
+	 * freeing the irqs brings us here, so it cannot say which of the two
+	 * the device was using. Ask a descriptor instead: getting this wrong
+	 * leaves the backend with MSI-X still enabled on the real device, and
+	 * the next attempt to enable it fails with -EALREADY.
+	 *
+	 * The descriptor lock is already held by pci_disable_msix().
+	 */
+	msi_for_each_desc(desc, &dev->dev, MSI_DESC_ALL) {
+		msix = desc->pci.msi_attrib.is_msix;
+		break;
+	}
+
+	if (msix)
 		xen_pci_frontend_disable_msix(dev);
 	else
 		xen_pci_frontend_disable_msi(dev);

drivers/iommu/Kconfig

@@ -363,6 +363,18 @@ config VIRTIO_IOMMU
 
 	  Say Y here if you intend to run this kernel as a guest.
 
+config XEN_IOMMU
+	bool "Xen IOMMU driver"
+	depends on XEN
+	select IOMMU_API
+	select IOMMU_IO_PGTABLE
+	help
+		Xen PV-IOMMU driver.
+
+		Say Y here if you intend to run this kernel under Xen, either as
+		Dom0 or as a guest whose passed-through devices come through
+		vPCI.
+
 config SPRD_IOMMU
 	tristate "Unisoc IOMMU Support"
 	depends on ARCH_SPRD || COMPILE_TEST

drivers/iommu/Makefile

@@ -16,6 +16,7 @@ obj-$(CONFIG_IOMMU_IO_PGTABLE_ARMV7S) += io-pgtable-arm-v7s.o
 obj-$(CONFIG_IOMMU_IO_PGTABLE_LPAE) += io-pgtable-arm.o
 obj-$(CONFIG_IOMMU_IO_PGTABLE_LPAE_KUNIT_TEST) += io-pgtable-arm-selftests.o
 obj-$(CONFIG_IOMMU_IO_PGTABLE_DART) += io-pgtable-dart.o
+obj-$(CONFIG_XEN_IOMMU) += io-pgtable-xen.o
 obj-$(CONFIG_IOMMU_IOVA) += iova.o
 obj-$(CONFIG_OF_IOMMU)	+= of_iommu.o
 obj-$(CONFIG_MSM_IOMMU) += msm_iommu.o
@@ -38,3 +39,4 @@ obj-$(CONFIG_SPRD_IOMMU) += sprd-iommu.o
 obj-$(CONFIG_APPLE_DART) += apple-dart.o
 obj-$(CONFIG_VSI_IOMMU) += vsi-iommu.o
 obj-$(CONFIG_IOMMU_DEBUG_PAGEALLOC) += iommu-debug-pagealloc.o
+obj-$(CONFIG_XEN_IOMMU) += xen-iommu.o

drivers/iommu/dma-iommu.c

@@ -29,6 +29,8 @@
 #include <linux/of_iommu.h>
 #include <linux/pci.h>
 #include <linux/pci-p2pdma.h>
+
+#include "iommu-priv.h"
 #include <linux/scatterlist.h>
 #include <linux/spinlock.h>
 #include <linux/swiotlb.h>
@@ -559,7 +561,17 @@ static int iova_reserve_iommu_regions(struct device *dev,
 	LIST_HEAD(resv_regions);
 	int ret = 0;
 
-	if (dev_is_pci(dev)) {
+	/*
+	 * A PCI host bridge's windows are reserved so that an IOVA cannot
+	 * collide with an address the bridge would route to MMIO. Under a
+	 * paravirtual IOMMU the IOVA space belongs to the hypervisor and is not
+	 * the bridge's address space at all, and the windows a Xen guest sees
+	 * are invented -- pcifront hands out iomem_resource itself, and a PVH
+	 * guest with no host bridge _CRS gets Linux's catch-all default. There
+	 * is nothing meaningful to reserve, and reserving it leaves no usable
+	 * IOVA space whatsoever.
+	 */
+	if (dev_is_pci(dev) && !xen_iommu_manages_iova(dev)) {
 		ret = iova_reserve_pci_windows(to_pci_dev(dev), iovad);
 		if (ret)
 			return ret;