iommu: always enable the vfio backend

Even if we have a kernel with iommufd vfio features enabled, the kernel
may not be configured for modern iommufd (i.e., no device cdevs).

In that case we want to fall back to legacy vfio.

Signed-off-by: Klaus Jensen <[email protected]>

include/vfn/iommu/context.h

@@ -16,13 +16,13 @@
 #define __VFN_IOVA_MIN 0x10000
 
 /**
- * get_iommu_context - create a new iommu context
+ * iommu_get_context - create a new iommu context
  * @name: context identifier
  *
  * Create a new iommu context. The mechanism depends on the backend.
  *
  * Return: a new &struct iommu_ctx.
  */
-struct iommu_ctx *get_iommu_context(const char *name);
+struct iommu_ctx *iommu_get_context(const char *name);
 
 #endif /* LIBVFN_IOMMU_CONTEXT_H */

src/iommu/context.c

@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: LGPL-2.1-or-later or MIT
+
+/*
+ * This file is part of libvfn.
+ *
+ * Copyright (C) 2023 The libvfn Authors. All Rights Reserved.
+ *
+ * This library (libvfn) is dual licensed under the GNU Lesser General
+ * Public License version 2.1 or later or the MIT license. See the
+ * COPYING and LICENSE files for more information.
+ */
+
+#define log_fmt(fmt) "iommu/context: " fmt
+
+#include <stddef.h>
+#include <stdint.h>
+#include <pthread.h>
+
+#include <sys/stat.h>
+
+#include "vfn/support.h"
+
+#include "util/iova_map.h"
+#include "context.h"
+
+#ifdef HAVE_VFIO_DEVICE_BIND_IOMMUFD
+static bool __iommufd_broken;
+
+static void __attribute__((constructor)) __check_iommufd_broken(void)
+{
+	struct stat sb;
+
+	if (stat("/dev/vfio/devices", &sb) || (sb.st_mode & S_IFDIR)) {
+		log_info("iommufd broken; probably missing CONFIG_VFIO_DEVICE_CDEV=y\n");
+
+		__iommufd_broken = true;
+	}
+}
+#endif
+
+struct iommu_ctx *iommu_get_default_context(void)
+{
+#ifdef HAVE_VFIO_DEVICE_BIND_IOMMUFD
+	if (__iommufd_broken)
+		goto fallback;
+
+	return iommufd_get_default_iommu_context();
+
+fallback:
+#endif
+	return vfio_get_default_iommu_context();
+}
+
+struct iommu_ctx *iommu_get_context(const char *name)
+{
+#ifdef HAVE_VFIO_DEVICE_BIND_IOMMUFD
+	if (__iommufd_broken)
+		goto fallback;
+
+	return iommufd_get_iommu_context(name);
+
+fallback:
+#endif
+	return vfio_get_iommu_context(name);
+}

src/iommu/context.h

@@ -30,4 +30,12 @@ struct iommu_ctx {
 	unsigned int flags;
 };
 
-struct iommu_ctx *get_default_iommu_ctx(void);
+struct iommu_ctx *iommu_get_default_context(void);
+
+struct iommu_ctx *vfio_get_default_iommu_context(void);
+struct iommu_ctx *vfio_get_iommu_context(const char *name);
+
+#ifdef HAVE_VFIO_DEVICE_BIND_IOMMUFD
+struct iommu_ctx *iommufd_get_default_iommu_context(void);
+struct iommu_ctx *iommufd_get_iommu_context(const char *name);
+#endif

src/iommu/iommufd.c

@@ -55,11 +55,6 @@ static struct iommu_ioas iommu_default_ioas = {
 	.name = "default",
 };
 
-struct iommu_ctx *get_default_iommu_ctx(void)
-{
-	return &iommu_default_ioas.ctx;
-}
-
 static int iommu_ioas_update_iova_ranges(struct iommu_ioas *ioas)
 {
 	struct iommu_ioas_iova_ranges iova_ranges = {
@@ -231,7 +226,7 @@ static int iommu_ioas_do_dma_unmap(struct iommu_ctx *ctx, uint64_t iova, size_t
 	return 0;
 }
 
-struct iommu_ctx_ops iommufd_ops = {
+static const struct iommu_ctx_ops iommufd_ops = {
 	.get_device_fd = iommufd_get_device_fd,
 
 	.dma_map = iommu_ioas_do_dma_map,
@@ -259,10 +254,22 @@ static int iommu_ioas_init(struct iommu_ioas *ioas)
 	return 0;
 }
 
-struct iommu_ctx *get_iommu_ctx(const char *name)
+static int iommufd_open(void)
+{
+	__iommufd = open("/dev/iommu", O_RDWR);
+	if (__iommufd < 0)
+		return -1;
+
+	return 0;
+}
+
+struct iommu_ctx *iommufd_get_iommu_context(const char *name)
 {
 	struct iommu_ioas *ioas = znew_t(struct iommu_ioas, 1);
 
+	if (__iommufd == -1)
+		log_fatal_if(iommufd_open(), "could not open /dev/iommu\n");
+
 	if (iommu_ioas_init(ioas) < 0) {
 		free(ioas);
 		return NULL;
@@ -273,12 +280,12 @@ struct iommu_ctx *get_iommu_ctx(const char *name)
 	return &ioas->ctx;
 }
 
-static void __attribute__((constructor)) init_iommufd(void)
+struct iommu_ctx *iommufd_get_default_iommu_context(void)
 {
-	__iommufd = open("/dev/iommu", O_RDWR);
-	if (__iommufd < 0)
-		log_fatal("could not open /dev/iommu\n");
+	if (__iommufd == -1) {
+		log_fatal_if(iommufd_open(), "could not open /dev/iommu\n");
+		log_fatal_if(iommu_ioas_init(&iommu_default_ioas), "iommu_ioas_init\n");
+	}
 
-	if (iommu_ioas_init(&iommu_default_ioas))
-		log_fatal("could not initialize default ioas: %s\n", strerror(errno));
+	return &iommu_default_ioas.ctx;
 }

src/iommu/meson.build

@@ -1,11 +1,11 @@
 iommu_sources = files(
+  'context.c',
   'dma.c',
+  'vfio.c',
 )
 
 if config_host.get('HAVE_VFIO_DEVICE_BIND_IOMMUFD', false)
   iommu_sources += files('iommufd.c',)
-else
-  iommu_sources += files('vfio.c')
 endif
 
 vfn_sources += iommu_sources

src/iommu/vfio.c

@@ -66,14 +66,10 @@ struct vfio_container {
 };
 
 static struct vfio_container vfio_default_container = {
+	.fd = -1,
 	.name = "default",
 };
 
-struct iommu_ctx *get_default_iommu_ctx(void)
-{
-	return &vfio_default_container.ctx;
-}
-
 #ifdef VFIO_IOMMU_INFO_CAPS
 # ifdef VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE
 static void vfio_iommu_type1_get_cap_iova_ranges(struct iova_map *map,
@@ -392,7 +388,7 @@ static int vfio_iommu_type1_do_dma_unmap(struct iommu_ctx *ctx, uint64_t iova, s
 	return 0;
 }
 
-struct iommu_ctx_ops vfio_ops = {
+static const struct iommu_ctx_ops vfio_ops = {
 	.get_device_fd = vfio_get_device_fd,
 
 	.dma_map = vfio_iommu_type1_do_dma_map,
@@ -422,7 +418,7 @@ static int vfio_init_container(struct vfio_container *vfio)
 	return 0;
 }
 
-struct iommu_ctx *get_iommu_context(const char *name)
+struct iommu_ctx *vfio_get_iommu_context(const char *name)
 {
 	struct vfio_container *vfio = znew_t(struct vfio_container, 1);
 
@@ -436,8 +432,11 @@ struct iommu_ctx *get_iommu_context(const char *name)
 	return &vfio->ctx;
 }
 
-static void __attribute__((constructor)) open_default_container(void)
+struct iommu_ctx *vfio_get_default_iommu_context(void)
 {
-	if (vfio_init_container(&vfio_default_container))
-		log_debug("default container not initialized\n");
+	if (vfio_default_container.fd == -1)
+		log_fatal_if(vfio_init_container(&vfio_default_container),
+			     "init default container\n");
+
+	return &vfio_default_container.ctx;
 }

src/vfio/pci.c

@@ -145,7 +145,7 @@ int vfio_pci_open(struct vfio_pci_device *pci, const char *bdf)
 	pci->bdf = bdf;
 
 	if (!pci->dev.ctx)
-		pci->dev.ctx = get_default_iommu_ctx();
+		pci->dev.ctx = iommu_get_default_context();
 
 	pci->dev.fd = pci->dev.ctx->ops.get_device_fd(pci->dev.ctx, bdf);
 	if (pci->dev.fd < 0) {