lkl: add PCI device interface and a vfio backend driver

LKL is now able to manage PCI devices by itself like DPDK!
This commit implements a PCI bus driver and PCI device interface.
The commit also includes a vfio-pci backend driver as a reference
implementation, thus no extra kernel modules is needed if
the PCI device is assigned to VFIO.

I believe the vfio backend driver fulfills the need of most cases,
but users can inject wrapper functions for another userspace PCI
framework such as uio-pci-generic or a handmade kernel module.
In either case, the framework should provide physically contiguous memory
for DMA, because the kernel and some drivers (e.g. NVMe) assume its memory
as physically contiguous.

Signed-off-by: Shinichi Awamoto <shinichi.awamoto@gmail.com>

.circleci/config.yml

@@ -56,8 +56,8 @@ do_steps: &do_steps
       command: |
         if [[ $CROSS_COMPILE == *android* ]]; then
           emulator -avd Nexus5_API24 -no-window -no-audio -no-boot-anim;
-        elif [[ $CROSS_COMPILE == *freebsd* ]]; then
+        elif [[ $CROSS_COMPILE == *freebsd* ]] || [[ -n "$LKL_QEMU_TEST" ]]; then
-          cd /home/ubuntu && $QEMU
+          cd /home/ubuntu && eval $QEMU
         fi
       background: true
   - run: cd tools/lkl && make -j8 ${MKARG}
@@ -71,7 +71,7 @@ do_steps: &do_steps
       command: |
         if [[ $CROSS_COMPILE == *android* ]]; then
           /home/ubuntu/circle-android.sh wait-for-boot;
-        elif [[ $CROSS_COMPILE == *freebsd* ]]; then
+        elif [[ $CROSS_COMPILE == *freebsd* ]] || [[ -n "$LKL_QEMU_TEST" ]]; then
           while ! $MYSSH -o ConnectTimeout=1 exit 2> /dev/null
           do
              sleep 5
@@ -147,6 +147,15 @@ jobs:
      VALGRIND: 1
    <<: *do_steps
 
+  x86_64_qemu:
+   docker:
+     - image: lkldocker/circleci-qemu-x86_64:v1.1
+   environment:
+     CROSS_COMPILE: ""
+     MKARG: "dpdk=no"
+     LKL_QEMU_TEST: 1
+   <<: *do_steps
+
   checkpatch:
    docker:
      - image: lkldocker/circleci:0.5
@@ -167,6 +176,7 @@ workflows:
   build:
     jobs:
      - x86_64
+     - x86_64_qemu
      - mingw32
      - android-arm32
      - android-aarch64

arch/lkl/Kconfig

@@ -36,6 +36,7 @@ config LKL
        select ARCH_NO_COHERENT_DMA_MMAP
        select HAVE_MEMBLOCK
        select NO_BOOTMEM
+       select BLK_DEV_NVME
 
 config OUTPUT_FORMAT
        string "Output format"
@@ -93,4 +94,8 @@ config CONSOLE_LOGLEVEL_QUIET
 	  will be used as the loglevel. IOW passing "quiet" will be the
 	  equivalent of passing "loglevel=<CONSOLE_LOGLEVEL_QUIET>"
 
-
+config PCI
+	bool "PCI support"
+	select NO_GENERIC_PCI_IOPORT_MAP
+	select GENERIC_PCI_IOMAP
+	default y

arch/lkl/Makefile

@@ -36,6 +36,7 @@ endif
 
 core-y += arch/lkl/kernel/
 core-y += arch/lkl/mm/
+core-y += arch/lkl/drivers/
 
 all: lkl.o
 

arch/lkl/drivers/Makefile

@@ -0,0 +1,2 @@
+
+obj-y = pci.o

arch/lkl/drivers/pci.c

@@ -0,0 +1,271 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include <linux/interrupt.h>
+#include <linux/types.h>
+#include <linux/io.h>
+#include <linux/platform_device.h>
+#include <linux/dma-mapping.h>
+#include <linux/scatterlist.h>
+#include <linux/mm.h>
+#include <asm/host_ops.h>
+
+static int lkl_pci_generic_read(struct pci_bus *bus, unsigned int devfn,
+				int where, int size, u32 *val)
+{
+	if (devfn == 0 &&
+	    lkl_ops->pci_ops->read(bus->sysdata, where, size, val) == size)
+		return PCIBIOS_SUCCESSFUL;
+	else
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+}
+
+static int lkl_pci_generic_write(struct pci_bus *bus, unsigned int devfn,
+				 int where, int size, u32 val)
+{
+	if (devfn == 0 &&
+	    lkl_ops->pci_ops->write(bus->sysdata, where, size, &val) == size)
+		return PCIBIOS_SUCCESSFUL;
+	else
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+}
+
+void __iomem *__pci_ioport_map(struct pci_dev *dev, unsigned long port,
+			       unsigned int nr)
+{
+	panic("%s is not supported\n", __func__);
+	return NULL;
+}
+
+static int lkl_pci_override_resource(struct pci_dev *dev, void *data)
+{
+	int i;
+	struct resource *r;
+	resource_size_t start, size;
+	void *remapped_start = NULL;
+
+	if (dev->devfn != 0)
+		return 0;
+
+	for (i = 0; i < PCI_NUM_RESOURCES; i++) {
+		r = &dev->resource[i];
+
+		if (!r->parent && r->start && r->flags) {
+			dev_info(&dev->dev, "claiming resource %s/%d\n",
+				 pci_name(dev), i);
+			if (pci_claim_resource(dev, i)) {
+				dev_err(&dev->dev,
+					"Could not claim resource %s/%d!",
+					pci_name(dev), i);
+			}
+
+			size = pci_resource_len(dev, i);
+
+			if (pci_resource_flags(dev, i) & IORESOURCE_MEM) {
+				remapped_start =
+					lkl_ops->pci_ops->resource_alloc(
+						dev->sysdata, size, i);
+			}
+
+			if (remapped_start) {
+				/* override values */
+				start = (resource_size_t)remapped_start;
+				pci_resource_start(dev, i) = start;
+				pci_resource_end(dev, i) = start + size - 1;
+			} else {
+				/*
+				 * A host library or the application could
+				 * not handle the resource. Disable it
+				 * not to be touched by drivers.
+				 */
+				pci_resource_flags(dev, i) |=
+					IORESOURCE_DISABLED;
+			}
+		}
+	}
+
+	dev->irq = lkl_get_free_irq("pci");
+
+	if (lkl_ops->pci_ops->irq_init(dev->sysdata, dev->irq) < 0)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static int lkl_pci_remove_devices(struct pci_dev *dev, void *data)
+{
+	lkl_ops->pci_ops->remove(dev->sysdata);
+	return 0;
+}
+
+static struct pci_ops lkl_pci_root_ops = {
+	.read = lkl_pci_generic_read,
+	.write = lkl_pci_generic_write,
+};
+
+static void *lkl_dma_alloc(struct device *dev, size_t size,
+			   dma_addr_t *dma_handle, gfp_t gfp,
+			   unsigned long attrs)
+{
+	void *vaddr = page_to_virt(alloc_pages(gfp, get_order(size)));
+	*dma_handle = (dma_addr_t)lkl_ops->pci_ops->map_page(
+		to_pci_dev(dev)->sysdata, vaddr, size);
+	return vaddr;
+}
+
+static void lkl_dma_free(struct device *dev, size_t size, void *cpu_addr,
+			 dma_addr_t dma_addr, unsigned long attrs)
+{
+	lkl_ops->pci_ops->unmap_page(to_pci_dev(dev)->sysdata, dma_addr, size);
+	__free_pages(cpu_addr, get_order(size));
+}
+
+static dma_addr_t lkl_dma_map_page(struct device *dev, struct page *page,
+				   unsigned long offset, size_t size,
+				   enum dma_data_direction dir,
+				   unsigned long attrs)
+{
+	dma_addr_t dma_handle = (dma_addr_t)lkl_ops->pci_ops->map_page(
+		to_pci_dev(dev)->sysdata, page_to_virt(page) + offset, size);
+	if (dma_handle == 0)
+		return DMA_MAPPING_ERROR;
+
+	return dma_handle;
+}
+
+static void lkl_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
+			       size_t size, enum dma_data_direction dir,
+			       unsigned long attrs)
+{
+	lkl_ops->pci_ops->unmap_page(to_pci_dev(dev)->sysdata, dma_addr, size);
+}
+
+static int lkl_dma_map_sg(struct device *dev, struct scatterlist *sgl,
+			  int nents, enum dma_data_direction dir,
+			  unsigned long attrs)
+{
+	int i;
+	struct scatterlist *sg;
+
+	for_each_sg(sgl, sg, nents, i) {
+		void *va;
+
+		WARN_ON(!sg_page(sg));
+		va = sg_virt(sg);
+		sg_dma_address(sg) = (dma_addr_t)lkl_dma_map_page(
+			dev, sg_page(sg), sg->offset, sg->length, dir, attrs);
+		sg_dma_len(sg) = sg->length;
+	}
+	return nents;
+}
+
+static void lkl_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
+			     int nents, enum dma_data_direction dir,
+			     unsigned long attrs)
+{
+	int i;
+	struct scatterlist *sg;
+
+	for_each_sg(sgl, sg, nents, i)
+		lkl_dma_unmap_page(dev, sg_dma_address(sg), sg_dma_len(sg), dir,
+				   attrs);
+}
+
+static int lkl_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static char *pcidev_name;
+
+static int __init setup_pci_device(char *str)
+{
+	if (pcidev_name) {
+		pr_info("The PCI driver supports only one PCI device.");
+		pr_info("'%s' will be discarded.", str);
+		return -1;
+	}
+	pcidev_name = str;
+	return 0;
+}
+
+early_param("lkl_pci", setup_pci_device);
+
+const struct dma_map_ops lkl_dma_ops = {
+	.alloc = lkl_dma_alloc,
+	.free = lkl_dma_free,
+	.map_sg = lkl_dma_map_sg,
+	.unmap_sg = lkl_dma_unmap_sg,
+	.map_page = lkl_dma_map_page,
+	.unmap_page = lkl_dma_unmap_page,
+	.dma_supported = lkl_dma_supported,
+};
+
+static int lkl_pci_probe(struct platform_device *pdev)
+{
+	struct lkl_pci_dev *dev;
+	struct pci_bus *bus;
+
+	if (!lkl_ops->pci_ops || !pcidev_name)
+		return -1;
+
+	dev = lkl_ops->pci_ops->add(pcidev_name, (void *)memory_start,
+				    memory_end - memory_start);
+	if (!dev)
+		return -1;
+
+	bus = pci_scan_bus(0, &lkl_pci_root_ops, (void *)dev);
+	if (!bus) {
+		lkl_ops->pci_ops->remove(dev);
+		return -1;
+	}
+	pci_walk_bus(bus, lkl_pci_override_resource, NULL);
+	pci_bus_add_devices(bus);
+	dev_set_drvdata(&pdev->dev, bus);
+
+	return 0;
+}
+
+static void lkl_pci_shutdown(struct platform_device *pdev)
+{
+	struct pci_bus *bus = (struct pci_bus *)dev_get_drvdata(&pdev->dev);
+
+	if (bus)
+		pci_walk_bus(bus, lkl_pci_remove_devices, NULL);
+}
+
+static struct platform_driver lkl_pci_driver = {
+	.driver = {
+		.name = "lkl_pci",
+	},
+	.probe = lkl_pci_probe,
+	.shutdown = lkl_pci_shutdown,
+};
+
+static int __init lkl_pci_init(void)
+{
+	int ret;
+	struct platform_device *dev;
+
+	/*register a platform driver*/
+	ret = platform_driver_register(&lkl_pci_driver);
+	if (ret != 0)
+		return ret;
+
+	dev = platform_device_alloc("lkl_pci", -1);
+	if (!dev)
+		return -ENOMEM;
+
+	ret = platform_device_add(dev);
+	if (ret != 0)
+		goto error;
+
+	return 0;
+error:
+	platform_device_put(dev);
+	return ret;
+}
+
+subsys_initcall(lkl_pci_init);

arch/lkl/include/asm/Kbuild

@@ -14,8 +14,6 @@ generic-y += current.h
 generic-y += delay.h
 generic-y += device.h
 generic-y += div64.h
-generic-y += dma.h
-generic-y += dma-mapping.h
 generic-y += emergency-restart.h
 generic-y += errno.h
 generic-y += extable.h
@@ -42,7 +40,7 @@ generic-y += module.h
 generic-y += msgbuf.h
 generic-y += param.h
 generic-y += parport.h
-generic-y += pci.h
+generic-y += pci_iomap.h
 generic-y += percpu.h
 generic-y += pgalloc.h
 generic-y += poll.h
@@ -75,5 +73,6 @@ generic-y += topology.h
 generic-y += trace_clock.h
 generic-y += unaligned.h
 generic-y += user.h
+generic-y += vga.h
 generic-y += word-at-a-time.h
 generic-y += kprobes.h

arch/lkl/include/asm/dma-mapping.h

@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_LKL_DMA_MAPPING_H
+#define _ASM_LKL_DMA_MAPPING_H
+
+extern const struct dma_map_ops lkl_dma_ops;
+
+static inline const struct dma_map_ops *get_arch_dma_ops(struct bus_type *bus)
+{
+	return &lkl_dma_ops;
+}
+
+#endif

arch/lkl/include/asm/dma.h

@@ -0,0 +1,13 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_LKL_DMA_H
+#define _ASM_LKL_DMA_H
+
+#include <asm-generic/dma.h>
+
+#ifdef CONFIG_PCI
+extern int isa_dma_bridge_buggy;
+#else
+#define isa_dma_bridge_buggy (0)
+#endif
+
+#endif /* _ASM_LKL_DMA_H */

arch/lkl/include/asm/pci.h

@@ -0,0 +1,11 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_LKL_PCI_H
+#define _ASM_LKL_PCI_H
+
+#include <asm-generic/pci.h>
+
+#define pcibios_assign_all_busses() 0
+#define PCIBIOS_MIN_IO 0x1000
+#define PCIBIOS_MIN_MEM 0x10000000
+
+#endif /* _ASM_LKL_PCI_H */

arch/lkl/include/uapi/asm/host_ops.h

@@ -9,6 +9,40 @@ typedef unsigned long lkl_thread_t;
 struct lkl_jmp_buf {
 	unsigned long buf[128];
 };
+struct lkl_pci_dev;
+
+/**
+ * lkl_dev_pci_ops - PCI host operations
+ *
+ * These operations would be a wrapper of userspace PCI drvier and
+ * must be provided by a host library or by the application.
+ *
+ * @add - add a new PCI device; returns a handler or NULL if fails
+ * @remove - release resources
+ * @init_irq - allocate resources for interrupts
+ * @read - read the PCI Configuration Space
+ * @write - write the PCI Configuration Space
+ * @resource_alloc - map BARx and return the mapped address. x is resource_index
+ *
+ * @map_page - return the DMA address of pages; vaddr might not be page-aligned
+ * @unmap_page - cleanup DMA region if needed
+ *
+ */
+struct lkl_dev_pci_ops {
+	struct lkl_pci_dev *(*add)(const char *name, void *kernel_ram,
+				   unsigned long ram_size);
+	void (*remove)(struct lkl_pci_dev *dev);
+	int (*irq_init)(struct lkl_pci_dev *dev, int irq);
+	int (*read)(struct lkl_pci_dev *dev, int where, int size, void *val);
+	int (*write)(struct lkl_pci_dev *dev, int where, int size, void *val);
+	void *(*resource_alloc)(struct lkl_pci_dev *dev,
+				unsigned long resource_size,
+				int resource_index);
+	unsigned long long (*map_page)(struct lkl_pci_dev *dev, void *vaddr,
+				       unsigned long size);
+	void (*unmap_page)(struct lkl_pci_dev *dev,
+			   unsigned long long dma_handle, unsigned long size);
+};
 
 /**
  * lkl_host_operations - host operations used by the Linux kernel
@@ -54,6 +88,8 @@ struct lkl_jmp_buf {
  *
  * @mem_alloc - allocate memory
  * @mem_free - free memory
+ * @page_alloc - allocate page aligned memory
+ * @page_free - free memory allocated by page_alloc
  *
  * @timer_create - allocate a host timer that runs fn(arg) when the timer
  * fires.
@@ -83,6 +119,7 @@ struct lkl_jmp_buf {
  * @jmp_buf_longjmp - perform a jump back to the saved jump buffer
  *
  * @memcpy - copy memory
+ * @pci_ops - pointer to PCI host operations
  */
 struct lkl_host_operations {
 	const char *virtio_devices;
@@ -114,6 +151,8 @@ struct lkl_host_operations {
 
 	void* (*mem_alloc)(unsigned long);
 	void (*mem_free)(void *);
+	void* (*page_alloc)(unsigned long size);
+	void (*page_free)(void *addr, unsigned long size);
 
 	unsigned long long (*time)(void);
 
@@ -131,6 +170,7 @@ struct lkl_host_operations {
 	void (*jmp_buf_longjmp)(struct lkl_jmp_buf *jmpb, int val);
 
 	void* (*memcpy)(void *dest, const void *src, unsigned long count);
+	struct lkl_dev_pci_ops *pci_ops;
 };
 
 /**

arch/lkl/mm/bootmem.c

@@ -12,7 +12,13 @@ void __init bootmem_init(unsigned long mem_sz)
 {
 	mem_size = mem_sz;
 
-	_memory_start = (unsigned long)lkl_ops->mem_alloc(mem_size);
+	if (lkl_ops->page_alloc) {
+		mem_size = PAGE_ALIGN(mem_size);
+		_memory_start = (unsigned long)lkl_ops->page_alloc(mem_size);
+	} else {
+		_memory_start = (unsigned long)lkl_ops->mem_alloc(mem_size);
+	}
+
 	memory_start = _memory_start;
 	BUG_ON(!memory_start);
 	memory_end = memory_start + mem_size;
@@ -62,5 +68,8 @@ void free_initmem(void)
 
 void free_mem(void)
 {
-	lkl_ops->mem_free((void *)_memory_start);
+	if (lkl_ops->page_free)
+		lkl_ops->page_free((void *)_memory_start, mem_size);
+	else
+		lkl_ops->mem_free((void *)_memory_start);
 }

tools/lkl/.gitignore

@@ -8,6 +8,7 @@ tests/valgrind*.xml
 *.dll
 tests/net-test
 tests/disk
+tests/vfio-pci
 Makefile.conf
 include/lkl_autoconf.h
 tests/autoconf.sh

tools/lkl/Makefile.autoconf

@@ -55,6 +55,7 @@ endef
 define posix_host
   $(call set_autoconf_var,POSIX,y)
   $(call set_autoconf_var,VIRTIO_NET,y)
+  $(if $(strip $(call find_include,linux/vfio.h)),$(call set_autoconf_var,VFIO_PCI,y))
   LDFLAGS += -pie
   CFLAGS += -fPIC -pthread
   SOSUF := .so

tools/lkl/Targets

@@ -23,5 +23,6 @@ LDLIBS_cptofs-$(LKL_HOST_CONFIG_NEEDS_LARGP) += -largp
 
 progs-y += tests/boot
 progs-y += tests/disk
+progs-y += tests/disk-vfio-pci
 progs-y += tests/net-test
 

tools/lkl/include/lkl.h

@@ -434,6 +434,25 @@ long lkl_mount_dev(unsigned int disk_id, unsigned int part, const char *fs_type,
 		   int flags, const char *opts,
 		   char *mnt_str, unsigned int mnt_str_len);
 
+/**
+ * lkl_mount_blkdev - mount a block device
+ *
+ * Like lkl_mount_dev, but mounts the device specified by dev.
+ *
+ * @dev - the device id (can be generated by LKL_MKDEV()) identifying the device
+ * to be mounted
+ * @fs_type - filesystem type
+ * @flags - mount flags
+ * @opts - additional filesystem specific mount options
+ * @mnt_str - a string that will be filled by this function with the path where
+ * the filesystem has been mounted
+ * @mnt_str_len - size of mnt_str
+ * @returns - 0 on success, a negative value on error
+ */
+long lkl_mount_blkdev(unsigned int dev, const char *fs_type, int flags,
+		      const char *opts, char *mnt_str,
+		      unsigned int mnt_str_len);
+
 /**
  * lkl_umount_dev - umount a disk
  *
@@ -450,6 +469,19 @@ long lkl_mount_dev(unsigned int disk_id, unsigned int part, const char *fs_type,
 long lkl_umount_dev(unsigned int disk_id, unsigned int part, int flags,
 		    long timeout_ms);
 
+/**
+ * lkl_umount_blkdev - umount a block device
+ *
+ * Like lkl_umount_dev, but unmounts the device specified by dev.
+ *
+ * @dev - the device id identifying the device to be mounted
+ * @flags - umount flags
+ * @timeout_ms - timeout to wait for the kernel to flush closed files so that
+ * umount can succeed
+ * @returns - 0 on success, a negative value on error
+ */
+long lkl_umount_blkdev(unsigned int dev, int flags, long timeout_ms);
+
 /**
  * lkl_umount_timeout - umount filesystem with timeout
  *

tools/lkl/lib/Build

@@ -21,5 +21,6 @@ liblkl-$(LKL_HOST_CONFIG_VIRTIO_NET_MACVTAP) += virtio_net_macvtap.o
 liblkl-$(LKL_HOST_CONFIG_VIRTIO_NET_DPDK) += virtio_net_dpdk.o
 liblkl-$(LKL_HOST_CONFIG_VIRTIO_NET_VDE) += virtio_net_vde.o
 liblkl-$(LKL_HOST_CONFIG_VIRTIO_NET) += virtio_net_pipe.o
+liblkl-$(LKL_HOST_CONFIG_VFIO_PCI) += vfio_pci.o
 liblkl-y += ../../perf/pmu-events/jsmn.o
 liblkl-y += config.o

tools/lkl/lib/fs.c

@@ -200,22 +200,16 @@ int lkl_get_virtio_blkdev(int disk_id, unsigned int part, uint32_t *pdevid)
 	return lkl_encode_dev_from_sysfs(sysfs_path, pdevid);
 }
 
-long lkl_mount_dev(unsigned int disk_id, unsigned int part,
+long lkl_mount_blkdev(unsigned int dev, const char *fs_type, int flags,
-		   const char *fs_type, int flags,
+		      const char *data, char *mnt_str, unsigned int mnt_str_len)
-		   const char *data, char *mnt_str, unsigned int mnt_str_len)
 {
 	char dev_str[] = { "/dev/xxxxxxxx" };
-	unsigned int dev;
 	int err;
 	char _data[4096]; /* FIXME: PAGE_SIZE is not exported by LKL */
 
 	if (mnt_str_len < sizeof(dev_str))
 		return -LKL_ENOMEM;
 
-	err = lkl_get_virtio_blkdev(disk_id, part, &dev);
-	if (err < 0)
-		return err;
-
 	snprintf(dev_str, sizeof(dev_str), "/dev/%08x", dev);
 	snprintf(mnt_str, mnt_str_len, "/mnt/%08x", dev);
 
@@ -263,6 +257,21 @@ long lkl_mount_dev(unsigned int disk_id, unsigned int part,
 	return 0;
 }
 
+long lkl_mount_dev(unsigned int disk_id, unsigned int part,
+		   const char *fs_type, int flags,
+		   const char *data, char *mnt_str, unsigned int mnt_str_len)
+{
+	unsigned int dev;
+	int err;
+
+	err = lkl_get_virtio_blkdev(disk_id, part, &dev);
+	if (err < 0)
+		return err;
+
+	return lkl_mount_blkdev(dev, fs_type, flags, data, mnt_str,
+				mnt_str_len);
+}
+
 long lkl_umount_timeout(char *path, int flags, long timeout_ms)
 {
 	long incr = 10000000; /* 10 ms */
@@ -284,18 +293,12 @@ long lkl_umount_timeout(char *path, int flags, long timeout_ms)
 	return err;
 }
 
-long lkl_umount_dev(unsigned int disk_id, unsigned int part, int flags,
+long lkl_umount_blkdev(unsigned int dev, int flags, long timeout_ms)
-		    long timeout_ms)
 {
 	char dev_str[] = { "/dev/xxxxxxxx" };
 	char mnt_str[] = { "/mnt/xxxxxxxx" };
-	unsigned int dev;
 	int err;
 
-	err = lkl_get_virtio_blkdev(disk_id, part, &dev);
-	if (err < 0)
-		return err;
-
 	snprintf(dev_str, sizeof(dev_str), "/dev/%08x", dev);
 	snprintf(mnt_str, sizeof(mnt_str), "/mnt/%08x", dev);
 
@@ -310,6 +313,19 @@ long lkl_umount_dev(unsigned int disk_id, unsigned int part, int flags,
 	return lkl_sys_rmdir(mnt_str);
 }
 
+long lkl_umount_dev(unsigned int disk_id, unsigned int part, int flags,
+		    long timeout_ms)
+{
+	unsigned int dev;
+	int err;
+
+	err = lkl_get_virtio_blkdev(disk_id, part, &dev);
+	if (err < 0)
+		return err;
+
+	return lkl_umount_blkdev(dev, flags, timeout_ms);
+}
+
 struct lkl_dir {
 	int fd;
 	char buf[1024];

tools/lkl/lib/posix-host.c

@@ -13,6 +13,7 @@
 #include <sys/types.h>
 #include <sys/socket.h>
 #include <sys/syscall.h>
+#include <sys/mman.h>
 #include <poll.h>
 #include <lkl_host.h>
 #include "iomem.h"
@@ -311,6 +312,28 @@ static long _gettid(void)
 #endif
 }
 
+static void *page_alloc(unsigned long size)
+{
+	void *addr;
+
+	addr = mmap(NULL, size, PROT_READ | PROT_WRITE,
+		     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+
+	if (addr == MAP_FAILED)
+		return NULL;
+
+	return addr;
+}
+
+static void page_free(void *addr, unsigned long size)
+{
+	munmap((void *)addr, size);
+}
+
+#ifdef LKL_HOST_CONFIG_VFIO_PCI
+extern struct lkl_dev_pci_ops vfio_pci_ops;
+#endif
+
 struct lkl_host_operations lkl_host_ops = {
 	.panic = panic,
 	.thread_create = thread_create,
@@ -338,6 +361,8 @@ struct lkl_host_operations lkl_host_ops = {
 	.print = print,
 	.mem_alloc = malloc,
 	.mem_free = free,
+	.page_alloc = page_alloc,
+	.page_free = page_free,
 	.ioremap = lkl_ioremap,
 	.iomem_access = lkl_iomem_access,
 	.virtio_devices = lkl_virtio_devs,
@@ -345,6 +370,9 @@ struct lkl_host_operations lkl_host_ops = {
 	.jmp_buf_set = jmp_buf_set,
 	.jmp_buf_longjmp = jmp_buf_longjmp,
 	.memcpy = memcpy,
+#ifdef LKL_HOST_CONFIG_VFIO_PCI
+	.pci_ops = &vfio_pci_ops,
+#endif
 };
 
 static int fd_get_capacity(struct lkl_disk disk, unsigned long long *res)

tools/lkl/lib/vfio_pci.c

@@ -0,0 +1,401 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <pthread.h>
+#include <lkl_host.h>
+#include <stdio.h>
+#include <sys/types.h>
+#include <stdint.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <linux/vfio.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#include <linux/vfio.h>
+#include <sys/eventfd.h>
+#include "iomem.h"
+
+struct lkl_pci_dev {
+	struct lkl_sem *thread_init_sem;
+	int irq;
+	lkl_thread_t int_thread;
+	int quit;
+	int fd;
+	int irq_fd;
+	struct vfio_device_info device_info;
+	struct vfio_region_info config_reg;
+	struct vfio_iommu_type1_dma_map dma_map;
+};
+
+/**
+ * vfio_pci_add - Create a new pci device
+ *
+ * The device should be assigned to VFIO by the host in advance.
+ *
+ * @name - PCI device name (as %x:%x:%x.%x format)
+ * @kernel_ram - the start address of kernel memory needed to be mapped for DMA.
+ * The address must be aligned to the page size.
+ * @ram_size - the size of kernel memory, should be page-aligned as well.
+ */
+
+static struct lkl_pci_dev *vfio_pci_add(const char *name, void *kernel_ram,
+					unsigned long ram_size)
+{
+	struct lkl_pci_dev *dev;
+	char path[128];
+	int segn, busn, devn, funcn;
+	int i;
+	int container_fd = 0, group_fd = 0;
+	struct vfio_group_status group_status = { .argsz = sizeof(
+							  group_status) };
+	struct vfio_iommu_type1_info iommu_info = { .argsz = sizeof(
+							    iommu_info) };
+
+	dev = malloc(sizeof(*dev));
+	if (!dev)
+		return NULL;
+
+	memset(dev, 0, sizeof(*dev));
+
+	dev->device_info.argsz = sizeof(struct vfio_device_info);
+	dev->config_reg.argsz = sizeof(struct vfio_region_info);
+	dev->dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
+
+	container_fd = open("/dev/vfio/vfio", O_RDWR);
+	if (container_fd < 0)
+		goto error;
+
+	if (ioctl(container_fd, VFIO_GET_API_VERSION) != VFIO_API_VERSION ||
+	    ioctl(container_fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) == 0)
+		goto error;
+
+	if (sscanf(name, "vfio%x:%x:%x.%x", &segn, &busn, &devn, &funcn) != 4)
+		goto error;
+
+	snprintf(path, sizeof(path),
+		 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/iommu_group", segn,
+		 busn, devn, funcn);
+
+	i = readlink(path, path, sizeof(path));
+	if (i < 0)
+		goto error;
+
+	path[i] = '\0';
+	snprintf(path, sizeof(path), "/dev/vfio%s", strrchr(path, '/'));
+
+	group_fd = open(path, O_RDWR);
+	if (group_fd < 0)
+		goto error;
+
+	if (ioctl(group_fd, VFIO_GROUP_GET_STATUS, &group_status) < 0)
+		goto error;
+
+	if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE))
+		goto error;
+
+	if (ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container_fd) < 0)
+		goto error;
+
+	if (ioctl(container_fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU) < 0)
+		goto error;
+
+	if (ioctl(container_fd, VFIO_IOMMU_GET_INFO, &iommu_info) < 0)
+		goto error;
+
+	/* if kernel_ram is null, assume the memory is already initialized
+	 * by another device, and skip this step.
+	 */
+	if (kernel_ram) {
+		dev->dma_map.vaddr = (uint64_t)kernel_ram;
+		dev->dma_map.size = ram_size;
+		dev->dma_map.iova = 0;
+		dev->dma_map.flags =
+			VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
+		if (ioctl(container_fd, VFIO_IOMMU_MAP_DMA, &dev->dma_map) < 0)
+			goto error;
+	}
+
+	snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x", segn, busn, devn,
+		 funcn);
+	dev->fd = ioctl(group_fd, VFIO_GROUP_GET_DEVICE_FD, path);
+
+	if (dev->fd < 0)
+		goto error;
+
+	if (ioctl(dev->fd, VFIO_DEVICE_GET_INFO, &dev->device_info) < 0)
+		goto error;
+
+	if (dev->device_info.num_regions <= VFIO_PCI_CONFIG_REGION_INDEX)
+		goto error;
+
+	dev->config_reg.index = VFIO_PCI_CONFIG_REGION_INDEX;
+
+	if (ioctl(dev->fd, VFIO_DEVICE_GET_REGION_INFO, &dev->config_reg) < 0)
+		goto error;
+
+	return dev;
+
+error:
+	lkl_printf("lkl_vfio_pci: failed to create a PCI device for %s\n",
+		   name);
+	if (container_fd > 0)
+		close(container_fd);
+	if (group_fd > 0)
+		close(group_fd);
+	free(dev);
+	return NULL;
+}
+
+static void vfio_pci_remove(struct lkl_pci_dev *dev)
+{
+	dev->quit = 1;
+	lkl_host_ops.thread_join(dev->int_thread);
+	close(dev->fd);
+	free(dev);
+}
+
+static int check_irq_status(struct lkl_pci_dev *dev)
+{
+	unsigned short status;
+
+	if (pread(dev->fd, &status, 2, dev->config_reg.offset + 6) != 2)
+		return 0;
+	return (status & (1 << 3)) ? 1 : 0;
+}
+
+/* Currently, we only support INTx. */
+static void vfio_int_thread(void *_dev)
+{
+	eventfd_t icount;
+	struct lkl_pci_dev *dev = (struct lkl_pci_dev *)_dev;
+	struct timespec req = { 0, 1000 * 1000 };
+	struct vfio_irq_info irq = { .argsz = sizeof(irq) };
+	struct vfio_irq_set *irq_set;
+	char irq_set_buf[sizeof(struct vfio_irq_set) + sizeof(int)];
+	fd_set rfds;
+
+	if (dev->device_info.num_irqs <= VFIO_PCI_INTX_IRQ_INDEX)
+		goto init_error;
+
+	irq.index = VFIO_PCI_INTX_IRQ_INDEX;
+
+	if (ioctl(dev->fd, VFIO_DEVICE_GET_IRQ_INFO, &irq))
+		goto init_error;
+
+	if (irq.count != 1)
+		goto init_error;
+
+	irq_set = (struct vfio_irq_set *)irq_set_buf;
+	irq_set->argsz = sizeof(irq_set_buf);
+	irq_set->count = 1;
+	irq_set->flags =
+		VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
+	irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
+	irq_set->start = 0;
+	dev->irq_fd = eventfd(0, EFD_CLOEXEC);
+	if (dev->irq_fd < 0)
+		goto init_error;
+	*(int *)&irq_set->data = dev->irq_fd;
+
+	if (ioctl(dev->fd, VFIO_DEVICE_SET_IRQS, irq_set))
+		goto init_error;
+
+	lkl_host_ops.sem_up(dev->thread_init_sem);
+
+	while (1) {
+		/* We should wait until the driver actually handles
+		 * an interrupt by monitoring the PCI interrupt status bit.
+		 */
+		while (check_irq_status(dev) && !dev->quit) {
+			lkl_trigger_irq(dev->irq);
+			nanosleep(&req, NULL);
+		}
+
+		if (dev->quit)
+			return;
+
+		/* unmask interrupts */
+		irq_set->argsz = sizeof(*irq_set);
+		irq_set->count = 1;
+		irq_set->flags =
+			VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
+		irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
+		irq_set->start = 0;
+		if (ioctl(dev->fd, VFIO_DEVICE_SET_IRQS, irq_set))
+			goto handling_error;
+
+		/* Wait for next interrupt. */
+		while (1) {
+			struct timeval tv;
+			int rc;
+
+			FD_ZERO(&rfds);
+			FD_SET(dev->irq_fd, &rfds);
+			tv.tv_sec = 0;
+			tv.tv_usec = 100 * 1000;
+			rc = select(dev->irq_fd + 1, &rfds, NULL, NULL, &tv);
+			if (rc == -1)
+				goto handling_error;
+			else if (rc)
+				if (read(dev->irq_fd, &icount, sizeof(icount)) <
+				    0)
+					goto handling_error;
+				else
+					break;
+			else if (dev->quit)
+				return;
+		}
+	}
+
+init_error:
+	lkl_printf("lkl_vfio_pci: failed to setup INTx for a device\n");
+	return;
+handling_error:
+	lkl_printf("lkl_vfio_pci: unknown error in the interrupt handler\n");
+}
+
+static int vfio_pci_irq_init(struct lkl_pci_dev *dev, int irq)
+{
+	dev->thread_init_sem = lkl_host_ops.sem_alloc(0);
+	if (!dev->thread_init_sem)
+		return -1;
+
+	dev->irq = irq;
+
+	dev->int_thread =
+		lkl_host_ops.thread_create(vfio_int_thread, (void *)dev);
+	if (!dev->int_thread) {
+		lkl_host_ops.sem_free(dev->thread_init_sem);
+		return -1;
+	}
+
+	/* wait until the interrupt handler thread is ready */
+	lkl_host_ops.sem_down(dev->thread_init_sem);
+	lkl_host_ops.sem_free(dev->thread_init_sem);
+	return 0;
+}
+
+static unsigned long long vfio_map_page(struct lkl_pci_dev *dev, void *vaddr,
+					unsigned long size)
+{
+	return (unsigned long long)vaddr - dev->dma_map.vaddr;
+}
+
+static void vfio_unmap_page(struct lkl_pci_dev *dev,
+			    unsigned long long dma_handle, unsigned long size)
+{
+}
+
+static int vfio_pci_read(struct lkl_pci_dev *dev, int where, int size,
+			 void *val)
+{
+	return pread(dev->fd, val, size, dev->config_reg.offset + where);
+}
+
+static int vfio_pci_write(struct lkl_pci_dev *dev, int where, int size,
+			  void *val)
+{
+	return pwrite(dev->fd, val, size, dev->config_reg.offset + where);
+}
+
+static int pci_resource_read(void *data, int offset, void *res, int size)
+{
+	void *addr = data + offset;
+
+	switch (size) {
+	case 8:
+		*(uint64_t *)res = *(uint64_t *)addr;
+		break;
+	case 4:
+		*(uint32_t *)res = *(uint32_t *)addr;
+		break;
+	case 2:
+		*(uint16_t *)res = *(uint16_t *)addr;
+		break;
+	case 1:
+		*(uint8_t *)res = *(uint8_t *)addr;
+		break;
+	default:
+		return -LKL_EOPNOTSUPP;
+	}
+	return 0;
+}
+
+static int pci_resource_write(void *data, int offset, void *res, int size)
+{
+	void *addr = data + offset;
+
+	switch (size) {
+	case 8:
+		*(uint64_t *)addr = *(uint64_t *)res;
+		break;
+	case 4:
+		*(uint32_t *)addr = *(uint32_t *)res;
+		break;
+	case 2:
+		*(uint16_t *)addr = *(uint16_t *)res;
+		break;
+	case 1:
+		*(uint8_t *)addr = *(uint8_t *)res;
+		break;
+	default:
+		return -LKL_EOPNOTSUPP;
+	}
+	return 0;
+}
+
+static const struct lkl_iomem_ops pci_resource_ops = {
+	.read = pci_resource_read,
+	.write = pci_resource_write,
+};
+
+#define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
+
+static void *vfio_resource_alloc(struct lkl_pci_dev *dev,
+				 unsigned long resource_size,
+				 int resource_index)
+{
+	unsigned int region_index_list[] = {
+		VFIO_PCI_BAR0_REGION_INDEX, VFIO_PCI_BAR1_REGION_INDEX,
+		VFIO_PCI_BAR2_REGION_INDEX, VFIO_PCI_BAR3_REGION_INDEX,
+		VFIO_PCI_BAR4_REGION_INDEX, VFIO_PCI_BAR5_REGION_INDEX,
+	};
+	struct vfio_region_info reg = { .argsz = sizeof(reg) };
+	void *mmio_addr;
+
+	if ((unsigned int)resource_index >= ARRAY_SIZE(region_index_list))
+		return NULL;
+
+	reg.index = region_index_list[resource_index];
+
+	if (dev->device_info.num_regions <= reg.index)
+		return NULL;
+
+	/* We assume the resource is a memory space. */
+
+	if (ioctl(dev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg) < 0)
+		return NULL;
+
+	if (reg.size < resource_size)
+		return NULL;
+
+	mmio_addr = mmap(NULL, resource_size, PROT_READ | PROT_WRITE,
+			 MAP_SHARED, dev->fd, reg.offset);
+
+	if (mmio_addr == MAP_FAILED)
+		return NULL;
+
+	return register_iomem(mmio_addr, resource_size, &pci_resource_ops);
+}
+
+struct lkl_dev_pci_ops vfio_pci_ops = {
+	.add = vfio_pci_add,
+	.remove = vfio_pci_remove,
+	.irq_init = vfio_pci_irq_init,
+	.read = vfio_pci_read,
+	.write = vfio_pci_write,
+	.resource_alloc = vfio_resource_alloc,
+	.map_page = vfio_map_page,
+	.unmap_page = vfio_unmap_page,
+};

tools/lkl/tests/Build

@@ -1,3 +1,4 @@
 boot-y += boot.o test.o
 disk-y += disk.o cla.o test.o
+disk-vfio-pci-y += disk-vfio-pci.o cla.o test.o
 net-test-y += net-test.o cla.o test.o

tools/lkl/tests/disk-vfio-pci.c

@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <stdio.h>
+#include <unistd.h>
+#include <string.h>
+#include <time.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <lkl.h>
+#include <lkl_host.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+
+#include "test.h"
+#include "cla.h"
+
+static struct {
+	int printk;
+	const char *fstype;
+	const char *pciname;
+} cla;
+
+struct cl_arg args[] = {
+	{ "type", 't', "filesystem type", 1, CL_ARG_STR, &cla.fstype },
+	{ "pciname", 'n', "PCI device name (as %x:%x:%x.%x format)", 1,
+	  CL_ARG_STR, &cla.pciname },
+	{ 0 },
+};
+
+static char mnt_point[32];
+static char bootparams[128];
+
+static int lkl_test_umount_dev(void)
+{
+	long ret, ret2;
+
+	ret = lkl_sys_chdir("/");
+
+	ret2 = lkl_umount_blkdev(LKL_MKDEV(259, 0), 0, 1000);
+
+	lkl_test_logf("%ld %ld", ret, ret2);
+
+	if (!ret && !ret2)
+		return TEST_SUCCESS;
+
+	return TEST_FAILURE;
+}
+
+struct lkl_dir *dir;
+
+static int lkl_test_opendir(void)
+{
+	int err;
+
+	dir = lkl_opendir(mnt_point, &err);
+
+	lkl_test_logf("lkl_opedir(%s) = %d %s\n", mnt_point, err,
+		      lkl_strerror(err));
+
+	if (err == 0)
+		return TEST_SUCCESS;
+
+	return TEST_FAILURE;
+}
+
+static int lkl_test_readdir(void)
+{
+	struct lkl_linux_dirent64 *de = lkl_readdir(dir);
+	int wr = 0;
+
+	while (de) {
+		wr += lkl_test_logf("%s ", de->d_name);
+		if (wr >= 70) {
+			lkl_test_logf("\n");
+			wr = 0;
+			break;
+		}
+		de = lkl_readdir(dir);
+	}
+
+	if (lkl_errdir(dir) == 0)
+		return TEST_SUCCESS;
+
+	return TEST_FAILURE;
+}
+
+LKL_TEST_CALL(mount_dev, lkl_mount_blkdev, 0, LKL_MKDEV(259, 0),
+	      cla.fstype, 0, NULL, mnt_point, sizeof(mnt_point))
+LKL_TEST_CALL(closedir, lkl_closedir, 0, dir);
+LKL_TEST_CALL(chdir_mnt_point, lkl_sys_chdir, 0, mnt_point);
+LKL_TEST_CALL(start_kernel, lkl_start_kernel, 0, &lkl_host_ops, bootparams);
+LKL_TEST_CALL(stop_kernel, lkl_sys_halt, 0);
+
+struct lkl_test tests[] = {
+	LKL_TEST(start_kernel),	   LKL_TEST(mount_dev),
+	LKL_TEST(chdir_mnt_point), LKL_TEST(opendir),
+	LKL_TEST(readdir),	   LKL_TEST(closedir),
+	LKL_TEST(umount_dev),	   LKL_TEST(stop_kernel),
+};
+
+int main(int argc, const char **argv)
+{
+	if (parse_args(argc, argv, args) < 0)
+		return -1;
+
+	snprintf(bootparams, sizeof(bootparams),
+		 "mem=16M loglevel=8 lkl_pci=vfio%s", cla.pciname);
+
+	lkl_host_ops.print = lkl_test_log;
+
+	return lkl_test_run(tests, sizeof(tests) / sizeof(struct lkl_test),
+			    "disk-vfio-pci %s", cla.fstype);
+}

tools/lkl/tests/disk-vfio-pci.sh

@@ -0,0 +1,69 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: GPL-2.0
+
+script_dir=$(cd $(dirname ${BASH_SOURCE:-$0}); pwd)
+source $script_dir/test.sh
+
+pciname="0000:00:03.0"
+nvme_id="8086 5845"
+bin_name="disk-vfio-pci"
+
+function wait_guest()
+{
+    for i in `seq 300`; do
+	if $MYSSH exit 2> /dev/null; then
+	    break
+	fi
+	sleep 1
+    done
+}
+
+function init()
+{
+    # initialize
+    dd if=/dev/zero of=/home/ubuntu/nvme.img bs=1024 count=102400
+    yes | sudo mkfs.$fstype /home/ubuntu/nvme.img
+    $MYSSH sudo modprobe vfio-pci
+    $MYSSH "sh -c 'echo $nvme_id |
+    	       	       sudo tee /sys/bus/pci/drivers/vfio-pci/new_id'"
+    $MYSSH "sh -c 'echo $pciname |
+    	       	       sudo tee /sys/bus/pci/drivers/nvme/unbind'"
+    $MYSSH "sh -c 'echo $pciname |
+    	       	       sudo tee /sys/bus/pci/drivers/vfio-pci/bind'"
+    $MYSSH sudo chown lkl:lkl /dev/vfio/3
+    $MYSCP $script_dir/$bin_name lkl@localhost:
+}
+
+function cleanup()
+{
+    $MYSSH "sh -c 'echo $pciname |
+    	       	       sudo tee /sys/bus/pci/drivers/vfio-pci/unbind'"
+    $MYSSH "sh -c 'echo $pciname |
+    	       	       sudo tee /sys/bus/pci/drivers/nvme/bind'"
+}
+
+function run()
+{
+    if [ -z "$LKL_QEMU_TEST" ]; then
+	lkl_test_plan 0 "disk-vfio-pci $fstype"
+	echo "vfio not supported"
+    else
+	lkl_test_plan 1 "disk-vfio-pci $fstype"
+	lkl_test_run 1 init
+	lkl_test_exec $MYSSH ./$bin_name -n 0000:00:03.0 -t $fstype
+	lkl_test_plan 1 "disk-vfio-pci $fstype"
+	lkl_test_run 1 cleanup
+    fi
+}
+
+if [ "$1" = "-t" ]; then
+    shift
+    fstype=$1
+    shift
+fi
+
+if [ -z "$fstype" ]; then
+    fstype="ext4"
+fi
+
+"$@"

tools/lkl/tests/run.py

@@ -62,7 +62,11 @@ tests = [
     'lklfuse.sh -t btrfs',
     'lklfuse.sh -t vfat',
     'lklfuse.sh -t xfs',
-    'hijack-test.sh'
+    'hijack-test.sh',
+    'disk-vfio-pci.sh -t ext4 run',
+    'disk-vfio-pci.sh -t btrfs run',
+    'disk-vfio-pci.sh -t vfat run',
+    'disk-vfio-pci.sh -t xfs run'
 ]
 
 parser = argparse.ArgumentParser(description='LKL test runner')