Skip to contentSkip to conversation

Add mdev (Mediated device) support in DPDK

RFC Superseded Tiwei Bie 8 participants 33 messages Apr 3, 2019 – Jul 3, 2023 Arm/Amlogic Meson SoC support

lore
Share & tools
lore

Tiwei Bie

author
#0 lore raw

cover Add mdev (Mediated device) support in DPDK

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) bus support in DPDK. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this RFC, one mdev bus is introduced to scan the mdev devices in the system and do the probe based on the device-api.

Take the mdev devices whose device-api is "vfio-pci" as an example, in this RFC, these devices will be probed by a mdev driver provided by PCI bus, which will plug them to the PCI bus. And they will be probed with the drivers registered on the PCI bus based on VendorID/ DeviceID/... then.

                     +----------+
                     | mdev bus |
                     +----+-----+
                          |
         +----------------+----+------+------+
         |                     |      |      |
   mdev_vfio_pci               ......

(device-api: vfio-pci)

There are also other ways to add mdev device support in DPDK (e.g. let PCI bus scan /sys/bus/mdev/devices directly). Comments would be appreciated!

[1] https://github.com/torvalds/linux/blob/master/Documentation/vfio-mediated-device.txt

Thanks,
Tiwei

Tiwei Bie (3):
eal: add a helper for reading string from sysfs
bus/mdev: add mdev bus support
bus/pci: add mdev support

 config/common_base                        |   5 +
 config/common_linux                       |   1 +
 drivers/bus/Makefile                      |   1 +
 drivers/bus/mdev/Makefile                 |  41 +++
 drivers/bus/mdev/linux/Makefile           |   6 +
 drivers/bus/mdev/linux/mdev.c             | 117 ++++++++
 drivers/bus/mdev/mdev.c                   | 310 ++++++++++++++++++++++
 drivers/bus/mdev/meson.build              |  15 ++
 drivers/bus/mdev/private.h                |  90 +++++++
 drivers/bus/mdev/rte_bus_mdev.h           | 141 ++++++++++
 drivers/bus/mdev/rte_bus_mdev_version.map |  12 +
 drivers/bus/meson.build                   |   2 +-
 drivers/bus/pci/Makefile                  |   3 +
 drivers/bus/pci/linux/Makefile            |   4 +
 drivers/bus/pci/linux/pci_vfio.c          |  35 ++-
 drivers/bus/pci/linux/pci_vfio_mdev.c     | 305 +++++++++++++++++++++
 drivers/bus/pci/meson.build               |   4 +-
 drivers/bus/pci/pci_common.c              |  17 +-
 drivers/bus/pci/private.h                 |   9 +
 drivers/bus/pci/rte_bus_pci.h             |  11 +-
 lib/librte_eal/common/eal_filesystem.h    |   7 +
 lib/librte_eal/freebsd/eal/eal.c          |  22 ++
 lib/librte_eal/linux/eal/eal.c            |  22 ++
 lib/librte_eal/rte_eal_version.map        |   1 +
 mk/rte.app.mk                             |   1 +
 25 files changed, 1163 insertions(+), 19 deletions(-)
 create mode 100644 drivers/bus/mdev/Makefile
 create mode 100644 drivers/bus/mdev/linux/Makefile
 create mode 100644 drivers/bus/mdev/linux/mdev.c
 create mode 100644 drivers/bus/mdev/mdev.c
 create mode 100644 drivers/bus/mdev/meson.build
 create mode 100644 drivers/bus/mdev/private.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev_version.map
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
Signature · 1 line
2.17.1
author
#1 lore raw

RFC 1/3 eal: add a helper for reading string from sysfs

This patch adds a helper for reading string from sysfs.

---
 lib/librte_eal/common/eal_filesystem.h |  7 +++++++
 lib/librte_eal/freebsd/eal/eal.c       | 22 ++++++++++++++++++++++
 lib/librte_eal/linux/eal/eal.c         | 22 ++++++++++++++++++++++
 lib/librte_eal/rte_eal_version.map     |  1 +
 4 files changed, 52 insertions(+)
lib/librte_eal/common/eal_filesystem.h +7 -0
index 89a3added..2c823b27d 100644
@@ -116,4 +116,11 @@ eal_get_hugefile_lock_path(char *buffer, size_t buflen, int f_id)
  * Used to read information from files on /sys */
 int eal_parse_sysfs_value(const char *filename, unsigned long *val);
 
+/**
+ * Function to read a line from a file on the filesystem.
+ * Used to read information from files on /sys
+ */
+int __rte_experimental
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz);
+
 #endif /* EAL_FILESYSTEM_H */
lib/librte_eal/freebsd/eal/eal.c +22 -0
index 4e86b10b1..816cb9b91 100644
@@ -208,6 +208,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/librte_eal/linux/eal/eal.c +22 -0
index 13f401684..865cb19d7 100644
@@ -293,6 +293,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/librte_eal/rte_eal_version.map +1 -0
index d6e375135..d16258ffc 100644
@@ -298,6 +298,7 @@ EXPERIMENTAL {
 	rte_devargs_remove;
 	rte_devargs_type_count;
 	rte_eal_cleanup;
+	rte_eal_parse_sysfs_str;
 	rte_extmem_attach;
 	rte_extmem_detach;
 	rte_extmem_register;
Signature · 1 line
2.17.1
author
#2 lore raw

RFC 2/3 bus/mdev: add mdev bus support

In reply to Tiwei Bie

This patch adds the mdev (Mediated device) bus support in DPDK. This bus driver will scan all the mdev devices in the system, and do the probe based on device API (mdev_type/device_api).

---
 config/common_base                        |   5 +
 config/common_linux                       |   1 +
 drivers/bus/Makefile                      |   1 +
 drivers/bus/mdev/Makefile                 |  41 +++
 drivers/bus/mdev/linux/Makefile           |   6 +
 drivers/bus/mdev/linux/mdev.c             | 117 ++++++++
 drivers/bus/mdev/mdev.c                   | 310 ++++++++++++++++++++++
 drivers/bus/mdev/meson.build              |  15 ++
 drivers/bus/mdev/private.h                |  90 +++++++
 drivers/bus/mdev/rte_bus_mdev.h           | 141 ++++++++++
 drivers/bus/mdev/rte_bus_mdev_version.map |  12 +
 drivers/bus/meson.build                   |   2 +-
 mk/rte.app.mk                             |   1 +
 13 files changed, 741 insertions(+), 1 deletion(-)
 create mode 100644 drivers/bus/mdev/Makefile
 create mode 100644 drivers/bus/mdev/linux/Makefile
 create mode 100644 drivers/bus/mdev/linux/mdev.c
 create mode 100644 drivers/bus/mdev/mdev.c
 create mode 100644 drivers/bus/mdev/meson.build
 create mode 100644 drivers/bus/mdev/private.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev_version.map
config/common_base +5 -0
index 6292bc4af..d29e9a089 100644
@@ -168,6 +168,11 @@ CONFIG_RTE_LIBRTE_COMMON_DPAAX=n
 #
 CONFIG_RTE_LIBRTE_IFPGA_BUS=y
 
+#
+# Compile the mdev bus
+#
+CONFIG_RTE_LIBRTE_MDEV_BUS=n
+
 #
 # Compile PCI bus driver
 #
config/common_linux +1 -0
index 75334273d..7de9624c0 100644
@@ -25,6 +25,7 @@ CONFIG_RTE_LIBRTE_AVP_PMD=y
 CONFIG_RTE_LIBRTE_VDEV_NETVSC_PMD=y
 CONFIG_RTE_LIBRTE_NFP_PMD=y
 CONFIG_RTE_LIBRTE_POWER=y
+CONFIG_RTE_LIBRTE_MDEV_BUS=y
 CONFIG_RTE_VIRTIO_USER=y
 CONFIG_RTE_PROC_INFO=y
 
drivers/bus/Makefile +1 -0
index cea3b55e6..b2144ee63 100644
@@ -8,6 +8,7 @@ ifeq ($(CONFIG_RTE_EAL_VFIO),y)
 DIRS-$(CONFIG_RTE_LIBRTE_FSLMC_BUS) += fslmc
 endif
 DIRS-$(CONFIG_RTE_LIBRTE_IFPGA_BUS) += ifpga
+DIRS-$(CONFIG_RTE_LIBRTE_MDEV_BUS) += mdev
 DIRS-$(CONFIG_RTE_LIBRTE_PCI_BUS) += pci
 DIRS-$(CONFIG_RTE_LIBRTE_VDEV_BUS) += vdev
 DIRS-$(CONFIG_RTE_LIBRTE_VMBUS) += vmbus
drivers/bus/mdev/Makefile new file +41 -0

41 lines not on this page · raw message

drivers/bus/mdev/linux/Makefile new file +6 -0
new file mode 100644
index 000000000..a777ad3d4
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright(c) 2019 Intel Corporation
+
+SRCS += mdev.c
+
+CFLAGS += -D_GNU_SOURCE
drivers/bus/mdev/linux/mdev.c new file +117 -0

117 lines not on this page · raw message

drivers/bus/mdev/mdev.c new file +310 -0

310 lines not on this page · raw message

drivers/bus/mdev/meson.build new file +15 -0
new file mode 100644
index 000000000..33c701cb9
@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright(c) 2019 Intel Corporation
+
+version = 1
+allow_experimental_apis = true
+install_headers('rte_bus_mdev.h')
+sources = files('mdev.c')
+
+if host_machine.system() == 'linux'
+	sources += files('linux/mdev.c')
+	includes += include_directories('linux')
+	cflags += ['-D_GNU_SOURCE']
+else
+	build = false
+endif
drivers/bus/mdev/private.h new file +90 -0

90 lines not on this page · raw message

drivers/bus/mdev/rte_bus_mdev.h new file +141 -0

141 lines not on this page · raw message

drivers/bus/mdev/rte_bus_mdev_version.map new file +12 -0
new file mode 100644
index 000000000..7f73bf96b
@@ -0,0 +1,12 @@
+DPDK_19.05 {
+
+	local: *;
+};
+
+EXPERIMENTAL {
+	global:
+
+	rte_mdev_get_sysfs_path;
+	rte_mdev_register;
+	rte_mdev_unregister;
+};
drivers/bus/meson.build +1 -1
index 80de2d91d..f0ab19a03 100644
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # Copyright(c) 2017 Intel Corporation
 
-drivers = ['dpaa', 'fslmc', 'ifpga', 'pci', 'vdev', 'vmbus']
+drivers = ['dpaa', 'fslmc', 'ifpga', 'mdev', 'pci', 'vdev', 'vmbus']
 std_deps = ['eal']
 config_flag_fmt = 'RTE_LIBRTE_@0@_BUS'
 driver_name_fmt = 'rte_bus_@0@'
mk/rte.app.mk +1 -0
index 262132fc6..f8abe8237 100644
@@ -123,6 +123,7 @@ ifeq ($(CONFIG_RTE_LIBRTE_FSLMC_BUS),y)
 _LDLIBS-$(CONFIG_RTE_LIBRTE_COMMON_DPAAX)   += -lrte_common_dpaax
 endif
 
+_LDLIBS-$(CONFIG_RTE_LIBRTE_MDEV_BUS)       += -lrte_bus_mdev
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PCI_BUS)        += -lrte_bus_pci
 _LDLIBS-$(CONFIG_RTE_LIBRTE_VDEV_BUS)       += -lrte_bus_vdev
 _LDLIBS-$(CONFIG_RTE_LIBRTE_DPAA_BUS)       += -lrte_bus_dpaa
Signature · 1 line
2.17.1
author
#3 lore raw

RFC 3/3 bus/pci: add mdev support

In reply to Tiwei Bie

This patch adds the mdev support in PCI bus driver. A mdev driver is introduced to probe the mdev devices whose device API is "vfio-pci" on the mdev bus.

PS. There are some hacks in this patch for now.

---
 drivers/bus/pci/Makefile              |   3 +
 drivers/bus/pci/linux/Makefile        |   4 +
 drivers/bus/pci/linux/pci_vfio.c      |  35 ++-
 drivers/bus/pci/linux/pci_vfio_mdev.c | 305 ++++++++++++++++++++++++++
 drivers/bus/pci/meson.build           |   4 +-
 drivers/bus/pci/pci_common.c          |  17 +-
 drivers/bus/pci/private.h             |   9 +
 drivers/bus/pci/rte_bus_pci.h         |  11 +-
 8 files changed, 370 insertions(+), 18 deletions(-)
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
drivers/bus/pci/Makefile +3 -0
index de53ce1bf..085ec9066 100644
@@ -27,6 +27,9 @@ CFLAGS += -DALLOW_EXPERIMENTAL_API
 
 LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring
 LDLIBS += -lrte_ethdev -lrte_pci -lrte_kvargs
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+LDLIBS += -lrte_bus_mdev
+endif
 
 include $(RTE_SDK)/drivers/bus/pci/$(SYSTEM)/Makefile
 SRCS-$(CONFIG_RTE_LIBRTE_PCI_BUS) := $(addprefix $(SYSTEM)/,$(SRCS))
drivers/bus/pci/linux/Makefile +4 -0
index 90404468b..88bbc2390 100644
@@ -4,3 +4,7 @@
 SRCS += pci.c
 SRCS += pci_uio.c
 SRCS += pci_vfio.c
+
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+	SRCS += pci_vfio_mdev.c
+endif
drivers/bus/pci/linux/pci_vfio.c +31 -4
index ebf6ccd3c..c2c4c6a50 100644
@@ -13,6 +13,9 @@
 
 #include <rte_log.h>
 #include <rte_pci.h>
+#ifdef RTE_LIBRTE_MDEV_BUS
+#include <rte_bus_mdev.h>
+#endif
 #include <rte_bus_pci.h>
 #include <rte_eal_memconfig.h>
 #include <rte_malloc.h>
@@ -20,6 +23,7 @@
 #include <rte_eal.h>
 #include <rte_bus.h>
 #include <rte_spinlock.h>
+#include <rte_uuid.h>
 
 #include "eal_filesystem.h"
 
@@ -648,6 +652,7 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 {
 	struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
 	char pci_addr[PATH_MAX] = {0};
+	const char *sysfs_path;
 	int vfio_dev_fd;
 	struct rte_pci_addr *loc = &dev->addr;
 	int i, ret;
@@ -663,10 +668,20 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 #endif
 
 	/* store PCI address string */
-	snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
+	if (dev->use_uuid) {
+#ifdef RTE_LIBRTE_MDEV_BUS
+		sysfs_path = rte_mdev_get_sysfs_path();
+		rte_uuid_unparse(dev->uuid, pci_addr, sizeof(pci_addr));
+#else
+		return -1;
+#endif
+	} else {
+		sysfs_path = rte_pci_get_sysfs_path();
+		snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
 			loc->domain, loc->bus, loc->devid, loc->function);
+	}
 
-	ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
+	ret = rte_vfio_setup_device(sysfs_path, pci_addr,
 					&vfio_dev_fd, &device_info);
 	if (ret)
 		return ret;
@@ -793,6 +808,7 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 {
 	struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
 	char pci_addr[PATH_MAX] = {0};
+	const char *sysfs_path;
 	int vfio_dev_fd;
 	struct rte_pci_addr *loc = &dev->addr;
 	int i, ret;
@@ -808,8 +824,19 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 #endif
 
 	/* store PCI address string */
-	snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
+	if (dev->use_uuid) {
+#ifdef RTE_LIBRTE_MDEV_BUS
+		sysfs_path = rte_mdev_get_sysfs_path();
+		rte_uuid_unparse(dev->uuid, pci_addr, sizeof(pci_addr));
+#else
+		return -1;
+#endif
+	} else {
+		sysfs_path = rte_pci_get_sysfs_path();
+		snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
 			loc->domain, loc->bus, loc->devid, loc->function);
+	}
+
 
 	/* if we're in a secondary process, just find our tailq entry */
 	TAILQ_FOREACH(vfio_res, vfio_res_list, next) {
@@ -825,7 +852,7 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 		return -1;
 	}
 
-	ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
+	ret = rte_vfio_setup_device(sysfs_path, pci_addr,
 					&vfio_dev_fd, &device_info);
 	if (ret)
 		return ret;
drivers/bus/pci/linux/pci_vfio_mdev.c new file +305 -0

305 lines not on this page · raw message

drivers/bus/pci/meson.build +3 -1
index a3140ff97..c3e884657 100644
@@ -11,8 +11,10 @@ sources = files('pci_common.c',
 if host_machine.system() == 'linux'
 	sources += files('linux/pci.c',
 			'linux/pci_uio.c',
-			'linux/pci_vfio.c')
+			'linux/pci_vfio.c',
+			'linux/pci_vfio_mdev.c')
 	includes += include_directories('linux')
+	deps += ['bus_mdev']
 else
 	sources += files('bsd/pci.c')
 	includes += include_directories('bsd')
drivers/bus/pci/pci_common.c +5 -12
index 704b9d71a..6b47333e6 100644
@@ -124,21 +124,17 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 {
 	int ret;
 	bool already_probed;
-	struct rte_pci_addr *loc;
 
 	if ((dr == NULL) || (dev == NULL))
 		return -EINVAL;
 
-	loc = &dev->addr;
-
 	/* The device is not blacklisted; Check if driver supports it */
 	if (!rte_pci_match(dr, dev))
 		/* Match of device and driver failed */
 		return 1;
 
-	RTE_LOG(INFO, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid, loc->function,
-			dev->device.numa_node);
+	RTE_LOG(INFO, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	/* no initialization when blacklisted, return without error */
 	if (dev->device.devargs != NULL &&
@@ -208,7 +204,6 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 static int
 rte_pci_detach_dev(struct rte_pci_device *dev)
 {
-	struct rte_pci_addr *loc;
 	struct rte_pci_driver *dr;
 	int ret = 0;
 
@@ -216,11 +211,9 @@ rte_pci_detach_dev(struct rte_pci_device *dev)
 		return -EINVAL;
 
 	dr = dev->driver;
-	loc = &dev->addr;
 
-	RTE_LOG(DEBUG, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid,
-			loc->function, dev->device.numa_node);
+	RTE_LOG(DEBUG, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	RTE_LOG(DEBUG, EAL, "  remove driver: %x:%x %s\n", dev->id.vendor_id,
 			dev->id.device_id, dr->driver.name);
@@ -387,7 +380,7 @@ rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 }
 
 /* Remove a device from PCI bus */
-static void
+void
 rte_pci_remove_device(struct rte_pci_device *pci_dev)
 {
 	TAILQ_REMOVE(&rte_pci_bus.device_list, pci_dev, next);
drivers/bus/pci/private.h +9 -0
index 13c3324bb..d5815ee44 100644
@@ -67,6 +67,15 @@ void rte_pci_add_device(struct rte_pci_device *pci_dev);
 void rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 		struct rte_pci_device *new_pci_dev);
 
+/**
+ * Remove a PCI device from the PCI Bus.
+ *
+ * @param pci_dev
+ *	PCI device to remove
+ * @return void
+ */
+void rte_pci_remove_device(struct rte_pci_device *pci_dev);
+
 /**
  * Update a pci device object by asking the kernel for the latest information.
  *
drivers/bus/pci/rte_bus_pci.h +10 -1
index 06e004cd3..465a44935 100644
@@ -51,6 +51,13 @@ TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver);
 
 struct rte_devargs;
 
+/* It's RTE_UUID_STRLEN, which is bigger than PCI_PRI_STR_SIZE. */
+#define RTE_PCI_NAME_LEN		(36 + 1)
+
+// XXX: we can't include rte_uuid.h directly due to the conflicts
+//      introduced by stdbool.h
+typedef unsigned char rte_uuid_t[16];
+
 /**
  * A structure describing a PCI device.
  */
@@ -58,6 +65,8 @@ struct rte_pci_device {
 	TAILQ_ENTRY(rte_pci_device) next;   /**< Next probed PCI device. */
 	struct rte_device device;           /**< Inherit core device */
 	struct rte_pci_addr addr;           /**< PCI location. */
+	rte_uuid_t uuid;                    /**< Mdev location. */
+	uint8_t use_uuid;                   /**< True if uuid field valid. */
 	struct rte_pci_id id;               /**< PCI ID. */
 	struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
 					    /**< PCI Memory Resource */
@@ -65,7 +74,7 @@ struct rte_pci_device {
 	struct rte_pci_driver *driver;      /**< PCI driver used in probing */
 	uint16_t max_vfs;                   /**< sriov enable if not zero */
 	enum rte_kernel_driver kdrv;        /**< Kernel driver passthrough */
-	char name[PCI_PRI_STR_SIZE+1];      /**< PCI location (ASCII) */
+	char name[RTE_PCI_NAME_LEN];        /**< PCI/Mdev location (ASCII) */
 	struct rte_intr_handle vfio_req_intr_handle;
 				/**< Handler of VFIO request interrupt */
 };
Signature · 1 line
2.17.1
#4 lore raw

Re: [RFC 3/3] bus/pci: add mdev support

Some minor nits.

drivers/bus/pci/Makefile CFLAGS += -DALLOW_EXPERIMENTAL_API

Quoted 27 lines · Tiwei Bie

This patch adds the mdev support in PCI bus driver. A mdev driver is introduced to probe the mdev devices whose device API is "vfio-pci" on the mdev bus.

PS. There are some hacks in this patch for now.

---

drivers/bus/pci/Makefile              |   3 +
drivers/bus/pci/linux/Makefile        |   4 +
drivers/bus/pci/linux/pci_vfio.c      |  35 ++-
drivers/bus/pci/linux/pci_vfio_mdev.c | 305 ++++++++++++++++++++++++++
drivers/bus/pci/meson.build           |   4 +-
drivers/bus/pci/pci_common.c          |  17 +-
drivers/bus/pci/private.h             |   9 +
drivers/bus/pci/rte_bus_pci.h         |  11 +-
8 files changed, 370 insertions(+), 18 deletions(-)

create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c

drivers/bus/pci/Makefile +0 -0
index de53ce1bf..085ec9066 100644
@@ -27,6 +27,9 @@ CFLAGS += -DALLOW_EXPERIMENTAL_API

This define is enabled in 50-70 Makefiles, we can leave this here, but we should refactor this to a common place in the future.

drivers/bus/pci/Makefile CFLAGS += -DALLOW_EXPERIMENTAL_API · line 32

Quoted 6 lines

LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring LDLIBS += -lrte_ethdev -lrte_pci -lrte_kvargs

drivers/bus/pci/Makefile +3 -0
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+LDLIBS += -lrte_bus_mdev
+endif

See comment below.

Quoted 11 lines

include $(RTE_SDK)/drivers/bus/pci/$(SYSTEM)/Makefile SRCS-$(CONFIG_RTE_LIBRTE_PCI_BUS) := $(addprefix $(SYSTEM)/,$(SRCS))

drivers/bus/pci/linux/Makefile +0 -0
index 90404468b..88bbc2390 100644
@@ -4,3 +4,7 @@

SRCS += pci.c
SRCS += pci_uio.c
SRCS += pci_vfio.c

drivers/bus/pci/linux/Makefile line 10

Quoted 4 lines
drivers/bus/pci/linux/Makefile +4 -0
+
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+	SRCS += pci_vfio_mdev.c
+endif

Do we need a configuration option for MDEV?
Can it be enabled for all builds or reuse a current configuration if only for some OS or arch?

Quoted 128 lines
drivers/bus/pci/linux/pci_vfio.c +31 -4
index ebf6ccd3c..c2c4c6a50 100644
@@ -13,6 +13,9 @@
 
 #include <rte_log.h>
 #include <rte_pci.h>
+#ifdef RTE_LIBRTE_MDEV_BUS
+#include <rte_bus_mdev.h>
+#endif
 #include <rte_bus_pci.h>
 #include <rte_eal_memconfig.h>
 #include <rte_malloc.h>
@@ -20,6 +23,7 @@
 #include <rte_eal.h>
 #include <rte_bus.h>
 #include <rte_spinlock.h>
+#include <rte_uuid.h>
 
 #include "eal_filesystem.h"
 
@@ -648,6 +652,7 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 {
 	struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
 	char pci_addr[PATH_MAX] = {0};
+	const char *sysfs_path;
 	int vfio_dev_fd;
 	struct rte_pci_addr *loc = &dev->addr;
 	int i, ret;
@@ -663,10 +668,20 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 #endif
 
 	/* store PCI address string */
-	snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
+	if (dev->use_uuid) {
+#ifdef RTE_LIBRTE_MDEV_BUS
+		sysfs_path = rte_mdev_get_sysfs_path();
+		rte_uuid_unparse(dev->uuid, pci_addr, sizeof(pci_addr));
+#else
+		return -1;
+#endif
+	} else {
+		sysfs_path = rte_pci_get_sysfs_path();
+		snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
 			loc->domain, loc->bus, loc->devid, loc->function);
+	}
 
-	ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
+	ret = rte_vfio_setup_device(sysfs_path, pci_addr,
 					&vfio_dev_fd, &device_info);
 	if (ret)
 		return ret;
@@ -793,6 +808,7 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 {
 	struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
 	char pci_addr[PATH_MAX] = {0};
+	const char *sysfs_path;
 	int vfio_dev_fd;
 	struct rte_pci_addr *loc = &dev->addr;
 	int i, ret;
@@ -808,8 +824,19 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 #endif
 
 	/* store PCI address string */
-	snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
+	if (dev->use_uuid) {
+#ifdef RTE_LIBRTE_MDEV_BUS
+		sysfs_path = rte_mdev_get_sysfs_path();
+		rte_uuid_unparse(dev->uuid, pci_addr, sizeof(pci_addr));
+#else
+		return -1;
+#endif
+	} else {
+		sysfs_path = rte_pci_get_sysfs_path();
+		snprintf(pci_addr, sizeof(pci_addr), PCI_PRI_FMT,
 			loc->domain, loc->bus, loc->devid, loc->function);
+	}
+
 
 	/* if we're in a secondary process, just find our tailq entry */
 	TAILQ_FOREACH(vfio_res, vfio_res_list, next) {
@@ -825,7 +852,7 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 		return -1;
 	}
 
-	ret = rte_vfio_setup_device(rte_pci_get_sysfs_path(), pci_addr,
+	ret = rte_vfio_setup_device(sysfs_path, pci_addr,
 					&vfio_dev_fd, &device_info);
 	if (ret)
 		return ret;
drivers/bus/pci/linux/pci_vfio_mdev.c new file +31 -0
new file mode 100644
index 000000000..92498c2fe
@@ -0,0 +1,305 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Intel Corporation
+ */
+
+#include <string.h>
+#include <dirent.h>
+#include <fcntl.h>
+#include <sys/ioctl.h>
+#include <linux/pci_regs.h>
+
+#include <rte_log.h>
+#include <rte_pci.h>
+#include <rte_eal_memconfig.h>
+#include <rte_malloc.h>
+#include <rte_devargs.h>
+#include <rte_memcpy.h>
+#include <rte_vfio.h>
+#include <rte_bus_mdev.h>
+
+#include "eal_private.h"
+#include "eal_filesystem.h"
+
+#include "private.h"
+
+extern struct rte_pci_bus rte_pci_bus;
+
+static int
+get_pci_id(const char *sysfs_base, const char *dev_addr,
+	   struct rte_pci_id *pci_id)
+{
+	int ret = 0;

drivers/bus/pci/linux/pci_vfio_mdev.c line 41

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+	int iommu_group_num;
+	int vfio_group_fd;
+	int vfio_dev_fd;
+	int container;
+	int class;
+	char name[PATH_MAX];
+	struct vfio_group_status group_status = {
+		.argsz = sizeof(group_status) };
+
+	container = open("/dev/vfio/vfio", O_RDWR);

Should this one use the VFIO_CONTAINER_PATH define in rte_vfio.h? The define is gated by VFIO_PRESENT in that header.

Quoted 12 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +12 -0
+	if (container < 0) {
+		RTE_LOG(WARNING, EAL, "Failed to open VFIO container\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (ioctl(container, VFIO_GET_API_VERSION) != VFIO_API_VERSION) {
+		/* Unknown API version */
+		RTE_LOG(WARNING, EAL, "Unknown VFIO API version\n");
+		ret = -1;
+		goto close_container;
+	}

drivers/bus/pci/linux/pci_vfio_mdev.c line 63

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+
+	if (rte_vfio_get_group_num(sysfs_base, dev_addr,
+				   &iommu_group_num) <= 0) {
+		RTE_LOG(WARNING, EAL, "%s not managed by VFIO driver\n",
+			dev_addr);
+		ret = -1;
+		goto close_container;
+	}
+
+	snprintf(name, sizeof(name), "/dev/vfio/%d", iommu_group_num);

We should be testing the return value from snprintf, but it is not done anyplace else in the code? We need to look at fixing this in a different patch, but not here.

Quoted 87 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +87 -0
+
+	vfio_group_fd = open(name, O_RDWR);
+	if (vfio_group_fd < 0) {
+		ret = -1;
+		goto close_container;
+	}
+
+	/* if group_fd == 0, that means the device isn't managed by VFIO */
+	if (vfio_group_fd == 0) {
+		RTE_LOG(WARNING, EAL, "%s not managed by VFIO driver\n",
+			dev_addr);
+		ret = -1;
+		goto close_group;
+	}
+
+	if (ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status)) {
+		RTE_LOG(ERR, EAL, "%s cannot get group status, error %i (%s)\n",
+			dev_addr, errno, strerror(errno));
+		ret = -1;
+		goto close_group;
+	}
+
+	if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
+		RTE_LOG(ERR, EAL, "%s VFIO group is not viable!\n", dev_addr);
+		ret = -1;
+		goto close_group;
+	}
+
+	if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
+		if (ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
+			    &container)) {
+			RTE_LOG(ERR, EAL, "%s cannot add VFIO group to container, error %i (%s)\n",
+				dev_addr, errno, strerror(errno));
+			ret = -1;
+			goto close_group;
+		}
+	}
+
+	if (ioctl(container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU)) {
+		RTE_LOG(ERR, EAL, "%s cannot set iommu, error %i (%s)\n",
+			dev_addr, errno, strerror(errno));
+		ret = -1;
+		goto close_group;
+	}
+
+	vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
+	if (vfio_dev_fd < 0) {
+		/* if we cannot get a device fd, this implies a problem with
+		 * the VFIO group or the container not having IOMMU configured.
+		 */
+		RTE_LOG(ERR, EAL, "Getting a vfio_dev_fd for %s failed errno %d\n",
+			dev_addr, errno);
+		ret = -1;
+		goto close_group;
+	}
+
+	/* vendor_id */
+	if (pread64(vfio_dev_fd, &pci_id->vendor_id, sizeof(uint16_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_VENDOR_ID) != sizeof(uint16_t)) {
+		RTE_LOG(ERR, EAL, "Cannot read VendorID from PCI config space\n");
+		ret = -1;
+		goto close_device;
+	}
+
+	/* device_id */
+	if (pread64(vfio_dev_fd, &pci_id->device_id, sizeof(uint16_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_DEVICE_ID) != sizeof(uint16_t)) {
+		RTE_LOG(ERR, EAL, "Cannot read DeviceID from PCI config space\n");
+		ret = -1;
+		goto close_device;
+	}
+
+	/* subsystem_vendor_id */
+	if (pread64(vfio_dev_fd, &pci_id->subsystem_vendor_id, sizeof(uint16_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_SUBSYSTEM_VENDOR_ID) != sizeof(uint16_t)) {
+		RTE_LOG(ERR, EAL, "Cannot read SubVendorID from PCI config space\n");
+		ret = -1;
+		goto close_device;
+	}
+
+	/* subsystem_device_id */
+	if (pread64(vfio_dev_fd, &pci_id->subsystem_device_id, sizeof(uint16_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_SUBSYSTEM_ID) != sizeof(uint16_t)) {

drivers/bus/pci/linux/pci_vfio_mdev.c

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+		RTE_LOG(ERR, EAL, "Cannot read SubDeviceID from PCI config space\n");
+		ret = -1;
+		goto close_device;
+	}
+
+	/* class_id */
+	if (pread64(vfio_dev_fd, &class, sizeof(uint32_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_CLASS_REVISION) != sizeof(uint32_t)) {
+		RTE_LOG(ERR, EAL, "Cannot read ClassID from PCI config space\n”);

These should possible be DEBUG messages, but ERR is ok I guess. To me filling up the log with a bunch of messages when it is also flagged and log at a higher layer to many log messages. It would require us to look and make a cleaner

drivers/bus/pci/linux/pci_vfio_mdev.c line 169

Quoted 9 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +9 -0
+		ret = -1;
+		goto close_device;
+	}
+	pci_id->class_id = class >> 8;
+
+close_device:
+	if (close(vfio_dev_fd) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO device for %s\n",
+			dev_addr);

These should be ERR, DEBUG or WARN not INFO IMO or no log message at all.

Quoted 7 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +7 -0
+		ret = -1;
+	}
+
+close_group:
+	if (close(vfio_group_fd) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO group for %s\n",
+			dev_addr);

drivers/bus/pci/linux/pci_vfio_mdev.c line 186

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+		ret = -1;
+	}
+
+close_container:
+	if (close(container) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO container\n");
+		ret = -1;
+	}
+
+out:

Jumping to 4 different exit points makes this function complex, would it not be better to have one error exit point and test if the fds need to be closed e.g.

	if (pread64(...)) {
		RTE_LOG(ERR, EAL, “Error message”);
		goto err_exit;
	}

	return 0;
err_exit:
	if (vfio_dev_fd && close(vfio_dev_fd) < 0) {
		
	}
	if (…) {
	}
	return -1;

This should eliminate the variable ret and reduce the lines of code.

Quoted 16 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +16 -0
+	return ret;
+}
+
+static int vfio_pci_probe(struct rte_mdev_driver *mdev_drv __rte_unused,
+			  struct rte_mdev_device *mdev_dev)
+{
+	char name[RTE_UUID_STRLEN];
+	struct rte_pci_device *dev;
+	struct rte_bus *bus;
+	int ret;
+
+	bus = rte_bus_find_by_name("pci");
+	if (bus == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot find bus pci\n");
+		return -ENOENT;
+	}

drivers/bus/pci/linux/pci_vfio_mdev.c line 212

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+
+	if (bus->plug == NULL) {
+		RTE_LOG(ERR, EAL, "Function plug not supported by bus (%s)\n",
+			bus->name);
+		return -ENOTSUP;
+	}
+
+	dev = malloc(sizeof(*dev));
+	if (dev == NULL)
+		return -ENOMEM;

If going to add error logs for the above tests, why does this one not get one?
Should we just remove them and check in the calling function instead? Then convert these to DEBUG logs or remove them.

drivers/bus/pci/linux/pci_vfio_mdev.c line 223

Quoted 11 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +11 -0
+
+	memset(dev, 0, sizeof(*dev));
+	dev->device.bus = &rte_pci_bus.bus;
+	rte_uuid_unparse(mdev_dev->addr, name, sizeof(name));
+
+	if (get_pci_id(rte_mdev_get_sysfs_path(), name, &dev->id)) {
+		free(dev);
+		return -1;
+	}
+
+	snprintf(dev->name, sizeof(dev->name), "%s", name);

This should be strlcpy()

drivers/bus/pci/linux/pci_vfio_mdev.c line 231

Quoted 8 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +8 -0
+	dev->device.name = dev->name;
+	dev->kdrv = RTE_KDRV_VFIO;
+	dev->use_uuid = 1;
+	rte_uuid_copy(dev->uuid, mdev_dev->addr);
+
+	// TODO: dev->device.devargs, etc
+
+	memset(&dev->addr, -1, sizeof(dev->addr)); // XXX: TODO

I have seen in the past that TODO or FIXME is not something that should be in the code. The TODO items should be removed and tracked outside the code if needed to be done later.

drivers/bus/pci/linux/pci_vfio_mdev.c line 241

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+
+	/* device is valid, add to the list (sorted) */
+	if (TAILQ_EMPTY(&rte_pci_bus.device_list)) {
+		rte_pci_add_device(dev);
+	} else {
+		struct rte_pci_device *dev2;
+		int ret;
+
+		TAILQ_FOREACH(dev2, &rte_pci_bus.device_list, next) {
+			// XXX

What does this comment mean? remove it or explain it.

Quoted 16 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +16 -0
+			ret = rte_pci_addr_cmp(&dev->addr, &dev2->addr);
+			if (ret == 0)
+				ret = strncmp(dev->name, dev2->name,
+					      sizeof(dev->name));
+			if (ret > 0)
+				continue;
+			if (ret < 0) {
+				rte_pci_insert_device(dev2, dev);
+				goto plug;
+			}
+			/* already registered */
+			free(dev);
+			return 0;
+		}
+
+		rte_pci_add_device(dev);

drivers/bus/pci/linux/pci_vfio_mdev.c line 267

Quoted 10 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +10 -0
+	}
+
+plug:
+	ret = bus->plug(&dev->device);
+	if (ret != 0) {
+		rte_pci_remove_device(dev);
+		free(dev);
+	} else {
+		mdev_dev->private = dev;
+	}

The coding guide states we remove {} around single line statements.

Quoted 46 lines
drivers/bus/pci/linux/pci_vfio_mdev.c new file +38 -0
+	return ret;
+}
+
+static int vfio_pci_remove(struct rte_mdev_device *mdev_dev)
+{
+	struct rte_pci_device *dev = mdev_dev->private;
+	struct rte_bus *bus;
+	int ret;
+
+	if (dev == NULL)
+		return 0;
+
+	bus = rte_bus_find_by_name("pci");
+	if (bus == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot find bus pci\n");
+		return -ENOENT;
+	}
+
+	if (bus->unplug == NULL) {
+		RTE_LOG(ERR, EAL, "Function unplug not supported by bus (%s)\n",
+			bus->name);
+		return -ENOTSUP;
+	}
+
+	ret = bus->unplug(&dev->device);
+	if (ret == 0)
+		mdev_dev->private = NULL;
+
+	return ret;
+}
+
+static struct rte_mdev_driver vfio_pci_drv = {
+	.dev_api = RTE_MDEV_DEV_API_VFIO_PCI,
+	.probe = vfio_pci_probe,
+	.remove = vfio_pci_remove
+};
+
+RTE_MDEV_REGISTER_DRIVER(mdev_vfio_pci, vfio_pci_drv);
drivers/bus/pci/meson.build +0 -0
index a3140ff97..c3e884657 100644
@@ -11,8 +11,10 @@ sources = files('pci_common.c',

if host_machine.system() == 'linux'

	sources += files('linux/pci.c',
			'linux/pci_uio.c',

drivers/bus/pci/meson.build sources = files('pci_common.c',

Quoted 3 lines
drivers/bus/pci/meson.build +2 -1
-			'linux/pci_vfio.c')
+			'linux/pci_vfio.c',
+			'linux/pci_vfio_mdev.c’)

If you need the RTE_LIBRTE_MDEV define then pci_vfio_mdev.c needs to be built conditionally?

drivers/bus/pci/meson.build

Quoted 2 lines
	includes += include_directories('linux')
drivers/bus/pci/meson.build +1 -0
+	deps += ['bus_mdev’]

If this was added form dev then is too should be conditional.

Quoted 54 lines

else

	sources += files('bsd/pci.c')
	includes += include_directories('bsd')
drivers/bus/pci/pci_common.c +2 -6
index 704b9d71a..6b47333e6 100644
@@ -124,21 +124,17 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 {
 	int ret;
 	bool already_probed;
-	struct rte_pci_addr *loc;
 
 	if ((dr == NULL) || (dev == NULL))
 		return -EINVAL;
 
-	loc = &dev->addr;
-
 	/* The device is not blacklisted; Check if driver supports it */
 	if (!rte_pci_match(dr, dev))
 		/* Match of device and driver failed */
 		return 1;
 
-	RTE_LOG(INFO, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid, loc->function,
-			dev->device.numa_node);
+	RTE_LOG(INFO, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	/* no initialization when blacklisted, return without error */
 	if (dev->device.devargs != NULL &&
@@ -208,7 +204,6 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,

static int
rte_pci_detach_dev(struct rte_pci_device *dev)
{

drivers/bus/pci/pci_common.c +0 -1
-	struct rte_pci_addr *loc;
	struct rte_pci_driver *dr;
	int ret = 0;
drivers/bus/pci/pci_common.c +2 -4
@@ -216,11 +211,9 @@ rte_pci_detach_dev(struct rte_pci_device *dev)
 		return -EINVAL;
 
 	dr = dev->driver;
-	loc = &dev->addr;
 
-	RTE_LOG(DEBUG, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid,
-			loc->function, dev->device.numa_node);
+	RTE_LOG(DEBUG, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	RTE_LOG(DEBUG, EAL, "  remove driver: %x:%x %s\n", dev->id.vendor_id,
 			dev->id.device_id, dr->driver.name);

drivers/bus/pci/pci_common.c rte_pci_insert_device(struct rte_pci_device *exist_pci_dev, · line 384

Quoted 6 lines
drivers/bus/pci/pci_common.c +1 -1
@@ -387,7 +380,7 @@ rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 }
 
 /* Remove a device from PCI bus */
-static void
+void
 rte_pci_remove_device(struct rte_pci_device *pci_dev)

Have not looked yet, but did this function get added to the version.map file? Does converting a function to public function require experimental tag too, maybe not?

Quoted 27 lines
drivers/bus/pci/pci_common.c +0 -0
 {
 	TAILQ_REMOVE(&rte_pci_bus.device_list, pci_dev, next);
drivers/bus/pci/private.h +9 -0
index 13c3324bb..d5815ee44 100644
@@ -67,6 +67,15 @@ void rte_pci_add_device(struct rte_pci_device *pci_dev);
 void rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 		struct rte_pci_device *new_pci_dev);
 
+/**
+ * Remove a PCI device from the PCI Bus.
+ *
+ * @param pci_dev
+ *	PCI device to remove
+ * @return void
+ */
+void rte_pci_remove_device(struct rte_pci_device *pci_dev);
+
 /**
 * Update a pci device object by asking the kernel for the latest information.
 *
drivers/bus/pci/rte_bus_pci.h +0 -0
index 06e004cd3..465a44935 100644

drivers/bus/pci/rte_bus_pci.h TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver); · line 59

Quoted 9 lines
drivers/bus/pci/rte_bus_pci.h +6 -0
@@ -51,6 +51,13 @@ TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver);
 
 struct rte_devargs;
 
+/* It's RTE_UUID_STRLEN, which is bigger than PCI_PRI_STR_SIZE. */
+#define RTE_PCI_NAME_LEN		(36 + 1)
+
+// XXX: we can't include rte_uuid.h directly due to the conflicts
+//      introduced by stdbool.h
+typedef unsigned char rte_uuid_t[16];

Does this need to have a the string ‘XXX’ in the comment? Note maybe a better word.

Quoted 25 lines
drivers/bus/pci/rte_bus_pci.h +4 -1
+
 /**
 * A structure describing a PCI device.
 */
@@ -58,6 +65,8 @@ struct rte_pci_device {
 	TAILQ_ENTRY(rte_pci_device) next;   /**< Next probed PCI device. */
 	struct rte_device device;           /**< Inherit core device */
 	struct rte_pci_addr addr;           /**< PCI location. */
+	rte_uuid_t uuid;                    /**< Mdev location. */
+	uint8_t use_uuid;                   /**< True if uuid field valid. */
 	struct rte_pci_id id;               /**< PCI ID. */
 	struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
 					    /**< PCI Memory Resource */
@@ -65,7 +74,7 @@ struct rte_pci_device {
 	struct rte_pci_driver *driver;      /**< PCI driver used in probing */
 	uint16_t max_vfs;                   /**< sriov enable if not zero */
 	enum rte_kernel_driver kdrv;        /**< Kernel driver passthrough */
-	char name[PCI_PRI_STR_SIZE+1];      /**< PCI location (ASCII) */
+	char name[RTE_PCI_NAME_LEN];        /**< PCI/Mdev location (ASCII) */
 	struct rte_intr_handle vfio_req_intr_handle;
 				/**< Handler of VFIO request interrupt */
 };
Signature · 1 line
2.17.1

Regards,
Keith

author
#5 lore raw

Re: [RFC 3/3] bus/pci: add mdev support

Quoted 49 lines · Keith

Some minor nits.

drivers/bus/pci/Makefile CFLAGS += -DALLOW_EXPERIMENTAL_API

Quoted 27 lines · Tiwei Bie

This patch adds the mdev support in PCI bus driver. A mdev driver is introduced to probe the mdev devices whose device API is "vfio-pci" on the mdev bus.

PS. There are some hacks in this patch for now.

---

drivers/bus/pci/Makefile              |   3 +
drivers/bus/pci/linux/Makefile        |   4 +
drivers/bus/pci/linux/pci_vfio.c      |  35 ++-
drivers/bus/pci/linux/pci_vfio_mdev.c | 305 ++++++++++++++++++++++++++
drivers/bus/pci/meson.build           |   4 +-
drivers/bus/pci/pci_common.c          |  17 +-
drivers/bus/pci/private.h             |   9 +
drivers/bus/pci/rte_bus_pci.h         |  11 +-
8 files changed, 370 insertions(+), 18 deletions(-)

create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c

drivers/bus/pci/Makefile +0 -0
index de53ce1bf..085ec9066 100644
@@ -27,6 +27,9 @@ CFLAGS += -DALLOW_EXPERIMENTAL_API

This define is enabled in 50-70 Makefiles, we can leave this here, but we should refactor this to a common place in the future.

drivers/bus/pci/Makefile CFLAGS += -DALLOW_EXPERIMENTAL_API · line 32

Quoted 6 lines

LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring LDLIBS += -lrte_ethdev -lrte_pci -lrte_kvargs

drivers/bus/pci/Makefile +3 -0
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+LDLIBS += -lrte_bus_mdev
+endif

See comment below.

Quoted 10 lines

include $(RTE_SDK)/drivers/bus/pci/$(SYSTEM)/Makefile SRCS-$(CONFIG_RTE_LIBRTE_PCI_BUS) := $(addprefix $(SYSTEM)/,$(SRCS))

drivers/bus/pci/linux/Makefile +0 -0
index 90404468b..88bbc2390 100644
@@ -4,3 +4,7 @@

SRCS += pci.c
SRCS += pci_uio.c

Quoted 8 lines

drivers/bus/pci/linux/Makefile line 10

Quoted 5 lines

SRCS += pci_vfio.c

drivers/bus/pci/linux/Makefile +4 -0
+
+ifeq ($(CONFIG_RTE_LIBRTE_MDEV_BUS),y)
+	SRCS += pci_vfio_mdev.c
+endif

Do we need a configuration option for MDEV?
Can it be enabled for all builds or reuse a current configuration if only for some OS or arch?

I think it's possible.

Quoted 1 line

[...]

Quoted 10 lines
Quoted 10 lines
drivers/bus/pci/linux/Makefile +10 -0
+static int
+get_pci_id(const char *sysfs_base, const char *dev_addr,
+	   struct rte_pci_id *pci_id)
+{
+	int ret = 0;
+	int iommu_group_num;
+	int vfio_group_fd;
+	int vfio_dev_fd;
+	int container;
+	int class;
Quoted 8 lines

drivers/bus/pci/linux/Makefile

Quoted 5 lines
drivers/bus/pci/linux/Makefile +5 -0
+	char name[PATH_MAX];
+	struct vfio_group_status group_status = {
+		.argsz = sizeof(group_status) };
+
+	container = open("/dev/vfio/vfio", O_RDWR);

Should this one use the VFIO_CONTAINER_PATH define in rte_vfio.h? The define is gated by VFIO_PRESENT in that header.

Yeah! And all the code in this file should be gated by VFIO_PRESENT as well.

Quoted 31 lines
Quoted 12 lines
drivers/bus/pci/linux/Makefile +12 -0
+	if (container < 0) {
+		RTE_LOG(WARNING, EAL, "Failed to open VFIO container\n");
+		ret = -1;
+		goto out;
+	}
+
+	if (ioctl(container, VFIO_GET_API_VERSION) != VFIO_API_VERSION) {
+		/* Unknown API version */
+		RTE_LOG(WARNING, EAL, "Unknown VFIO API version\n");
+		ret = -1;
+		goto close_container;
+	}

drivers/bus/pci/linux/Makefile

Quoted 10 lines
drivers/bus/pci/linux/Makefile +10 -0
+
+	if (rte_vfio_get_group_num(sysfs_base, dev_addr,
+				   &iommu_group_num) <= 0) {
+		RTE_LOG(WARNING, EAL, "%s not managed by VFIO driver\n",
+			dev_addr);
+		ret = -1;
+		goto close_container;
+	}
+
+	snprintf(name, sizeof(name), "/dev/vfio/%d", iommu_group_num);

We should be testing the return value from snprintf, but it is not done anyplace else in the code? We need to look at fixing this in a different patch, but not here.

Quoted 6 lines
drivers/bus/pci/linux/Makefile +6 -0
+
+	vfio_group_fd = open(name, O_RDWR);
+	if (vfio_group_fd < 0) {
+		ret = -1;
+		goto close_container;
+	}

[...]

Quoted 7 lines

drivers/bus/pci/linux/Makefile

Quoted 5 lines
drivers/bus/pci/linux/Makefile +5 -0
+	/* class_id */
+	if (pread64(vfio_dev_fd, &class, sizeof(uint32_t),
+		      VFIO_GET_REGION_ADDR(VFIO_PCI_CONFIG_REGION_INDEX) +
+		      PCI_CLASS_REVISION) != sizeof(uint32_t)) {
+		RTE_LOG(ERR, EAL, "Cannot read ClassID from PCI config space\n”);

These should possible be DEBUG messages, but ERR is ok I guess. To me filling up the log with a bunch of messages when it is also flagged and log at a higher layer to many log messages. It would require us to look and make a cleaner

Quoted 11 lines

drivers/bus/pci/linux/Makefile

Quoted 9 lines
drivers/bus/pci/linux/Makefile +9 -0
+		ret = -1;
+		goto close_device;
+	}
+	pci_id->class_id = class >> 8;
+
+close_device:
+	if (close(vfio_dev_fd) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO device for %s\n",
+			dev_addr);

These should be ERR, DEBUG or WARN not INFO IMO or no log message at all.

You are right. I missed this one.

Quoted 34 lines
Quoted 7 lines
drivers/bus/pci/linux/Makefile +7 -0
+		ret = -1;
+	}
+
+close_group:
+	if (close(vfio_group_fd) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO group for %s\n",
+			dev_addr);

drivers/bus/pci/linux/Makefile

Quoted 10 lines
drivers/bus/pci/linux/Makefile +10 -0
+		ret = -1;
+	}
+
+close_container:
+	if (close(container) < 0) {
+		RTE_LOG(INFO, EAL, "Error when closing VFIO container\n");
+		ret = -1;
+	}
+
+out:

Jumping to 4 different exit points makes this function complex, would it not be better to have one error exit point and test if the fds need to be closed e.g.

	if (pread64(...)) {
		RTE_LOG(ERR, EAL, “Error message”);
		goto err_exit;
	}

	return 0;
err_exit:
	if (vfio_dev_fd && close(vfio_dev_fd) < 0) {
		
	}
	if (…) {
	}
	return -1;
Quoted 1 line

This should eliminate the variable ret and reduce the lines of code.

Thanks for the suggestion. I can do that.

Quoted 21 lines
Quoted 21 lines
drivers/bus/pci/linux/Makefile +21 -0
+	return ret;
+}
+
+static int vfio_pci_probe(struct rte_mdev_driver *mdev_drv __rte_unused,
+			  struct rte_mdev_device *mdev_dev)
+{
+	char name[RTE_UUID_STRLEN];
+	struct rte_pci_device *dev;
+	struct rte_bus *bus;
+	int ret;
+
+	bus = rte_bus_find_by_name("pci");
+	if (bus == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot find bus pci\n");
+		return -ENOENT;
+	}
+
+	if (bus->plug == NULL) {
+		RTE_LOG(ERR, EAL, "Function plug not supported by bus (%s)\n",
+			bus->name);
+		return -ENOTSUP;
Quoted 8 lines

drivers/bus/pci/linux/Makefile

Quoted 5 lines
drivers/bus/pci/linux/Makefile +5 -0
+	}
+
+	dev = malloc(sizeof(*dev));
+	if (dev == NULL)
+		return -ENOMEM;

If going to add error logs for the above tests, why does this one not get one?
Should we just remove them and check in the calling function instead? Then convert these to DEBUG logs or remove them.

Thanks for the suggestion. Will improve the logs.

Quoted 13 lines

drivers/bus/pci/linux/Makefile

Quoted 11 lines
drivers/bus/pci/linux/Makefile +11 -0
+
+	memset(dev, 0, sizeof(*dev));
+	dev->device.bus = &rte_pci_bus.bus;
+	rte_uuid_unparse(mdev_dev->addr, name, sizeof(name));
+
+	if (get_pci_id(rte_mdev_get_sysfs_path(), name, &dev->id)) {
+		free(dev);
+		return -1;
+	}
+
+	snprintf(dev->name, sizeof(dev->name), "%s", name);

This should be strlcpy()

Quoted 10 lines

drivers/bus/pci/linux/Makefile

Quoted 8 lines
drivers/bus/pci/linux/Makefile +8 -0
+	dev->device.name = dev->name;
+	dev->kdrv = RTE_KDRV_VFIO;
+	dev->use_uuid = 1;
+	rte_uuid_copy(dev->uuid, mdev_dev->addr);
+
+	// TODO: dev->device.devargs, etc
+
+	memset(&dev->addr, -1, sizeof(dev->addr)); // XXX: TODO

I have seen in the past that TODO or FIXME is not something that should be in the code. The TODO items should be removed and tracked outside the code if needed to be done later.

Sorry for the confusion. There are some quick hacks in this RFC (especially in this function). I highlighted them with XXX or TODO. I didn't get rid of them for now, because there are different possible ways to add the mdev support in DPDK, and this RFC is just to demonstrate one possible way that we can do and to hear people's thoughts/opinions.

PS. All the hacks (including comments starting with //) in this RFC are temporary. They will be fixed or removed in the formal patch.

Quoted 4 lines
Quoted 4 lines
drivers/bus/pci/linux/Makefile +4 -0
+
+	/* device is valid, add to the list (sorted) */
+	if (TAILQ_EMPTY(&rte_pci_bus.device_list)) {
+		rte_pci_add_device(dev);
Quoted 8 lines

drivers/bus/pci/linux/Makefile

Quoted 6 lines
drivers/bus/pci/linux/Makefile +6 -0
+	} else {
+		struct rte_pci_device *dev2;
+		int ret;
+
+		TAILQ_FOREACH(dev2, &rte_pci_bus.device_list, next) {
+			// XXX

What does this comment mean? remove it or explain it.

It's to indicate that there is a quick hack here. It won't exist in the formal patch.

Quoted 79 lines
Quoted 16 lines
drivers/bus/pci/linux/Makefile +16 -0
+			ret = rte_pci_addr_cmp(&dev->addr, &dev2->addr);
+			if (ret == 0)
+				ret = strncmp(dev->name, dev2->name,
+					      sizeof(dev->name));
+			if (ret > 0)
+				continue;
+			if (ret < 0) {
+				rte_pci_insert_device(dev2, dev);
+				goto plug;
+			}
+			/* already registered */
+			free(dev);
+			return 0;
+		}
+
+		rte_pci_add_device(dev);

drivers/bus/pci/linux/Makefile

Quoted 10 lines
drivers/bus/pci/linux/Makefile +10 -0
+	}
+
+plug:
+	ret = bus->plug(&dev->device);
+	if (ret != 0) {
+		rte_pci_remove_device(dev);
+		free(dev);
+	} else {
+		mdev_dev->private = dev;
+	}

The coding guide states we remove {} around single line statements.

Quoted 46 lines
drivers/bus/pci/linux/Makefile +38 -0
+	return ret;
+}
+
+static int vfio_pci_remove(struct rte_mdev_device *mdev_dev)
+{
+	struct rte_pci_device *dev = mdev_dev->private;
+	struct rte_bus *bus;
+	int ret;
+
+	if (dev == NULL)
+		return 0;
+
+	bus = rte_bus_find_by_name("pci");
+	if (bus == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot find bus pci\n");
+		return -ENOENT;
+	}
+
+	if (bus->unplug == NULL) {
+		RTE_LOG(ERR, EAL, "Function unplug not supported by bus (%s)\n",
+			bus->name);
+		return -ENOTSUP;
+	}
+
+	ret = bus->unplug(&dev->device);
+	if (ret == 0)
+		mdev_dev->private = NULL;
+
+	return ret;
+}
+
+static struct rte_mdev_driver vfio_pci_drv = {
+	.dev_api = RTE_MDEV_DEV_API_VFIO_PCI,
+	.probe = vfio_pci_probe,
+	.remove = vfio_pci_remove
+};
+
+RTE_MDEV_REGISTER_DRIVER(mdev_vfio_pci, vfio_pci_drv);
drivers/bus/pci/meson.build +0 -0
index a3140ff97..c3e884657 100644
@@ -11,8 +11,10 @@ sources = files('pci_common.c',

if host_machine.system() == 'linux'

	sources += files('linux/pci.c',
			'linux/pci_uio.c',

drivers/bus/pci/meson.build sources = files('pci_common.c',

Quoted 3 lines
drivers/bus/pci/meson.build +2 -1
-			'linux/pci_vfio.c')
+			'linux/pci_vfio.c',
+			'linux/pci_vfio_mdev.c’)

If you need the RTE_LIBRTE_MDEV define then pci_vfio_mdev.c needs to be built conditionally?

Quoted 4 lines

drivers/bus/pci/meson.build

Quoted 2 lines
	includes += include_directories('linux')
drivers/bus/pci/meson.build +1 -0
+	deps += ['bus_mdev’]

If this was added form dev then is too should be conditional.

Yeah. There should be a check of dpdk_conf.has('...')

Quoted 56 lines
Quoted 56 lines

else

	sources += files('bsd/pci.c')
	includes += include_directories('bsd')
drivers/bus/pci/pci_common.c +2 -6
index 704b9d71a..6b47333e6 100644
@@ -124,21 +124,17 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 {
 	int ret;
 	bool already_probed;
-	struct rte_pci_addr *loc;
 
 	if ((dr == NULL) || (dev == NULL))
 		return -EINVAL;
 
-	loc = &dev->addr;
-
 	/* The device is not blacklisted; Check if driver supports it */
 	if (!rte_pci_match(dr, dev))
 		/* Match of device and driver failed */
 		return 1;
 
-	RTE_LOG(INFO, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid, loc->function,
-			dev->device.numa_node);
+	RTE_LOG(INFO, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	/* no initialization when blacklisted, return without error */
 	if (dev->device.devargs != NULL &&
@@ -208,7 +204,6 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,

static int
rte_pci_detach_dev(struct rte_pci_device *dev)
{

drivers/bus/pci/pci_common.c +0 -1
-	struct rte_pci_addr *loc;
	struct rte_pci_driver *dr;
	int ret = 0;
drivers/bus/pci/pci_common.c +2 -4
@@ -216,11 +211,9 @@ rte_pci_detach_dev(struct rte_pci_device *dev)
 		return -EINVAL;
 
 	dr = dev->driver;
-	loc = &dev->addr;
 
-	RTE_LOG(DEBUG, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid,
-			loc->function, dev->device.numa_node);
+	RTE_LOG(DEBUG, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	RTE_LOG(DEBUG, EAL, "  remove driver: %x:%x %s\n", dev->id.vendor_id,
 			dev->id.device_id, dr->driver.name);
@@ -387,7 +380,7 @@ rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 }
 
Quoted 7 lines

drivers/bus/pci/pci_common.c rte_pci_insert_device(struct rte_pci_device *exist_pci_dev, · line 384

Quoted 4 lines
drivers/bus/pci/pci_common.c +1 -1
 /* Remove a device from PCI bus */
-static void
+void
 rte_pci_remove_device(struct rte_pci_device *pci_dev)

Have not looked yet, but did this function get added to the version.map file? Does converting a function to public function require experimental tag too, maybe not?

This is just to make it a global function declared in private.h, so that we can call it from other C files inside PCI bus.

Quoted 30 lines
Quoted 30 lines
drivers/bus/pci/pci_common.c +0 -0
 {
 	TAILQ_REMOVE(&rte_pci_bus.device_list, pci_dev, next);
drivers/bus/pci/private.h +9 -0
index 13c3324bb..d5815ee44 100644
@@ -67,6 +67,15 @@ void rte_pci_add_device(struct rte_pci_device *pci_dev);
 void rte_pci_insert_device(struct rte_pci_device *exist_pci_dev,
 		struct rte_pci_device *new_pci_dev);
 
+/**
+ * Remove a PCI device from the PCI Bus.
+ *
+ * @param pci_dev
+ *	PCI device to remove
+ * @return void
+ */
+void rte_pci_remove_device(struct rte_pci_device *pci_dev);
+
 /**
 * Update a pci device object by asking the kernel for the latest information.
 *
drivers/bus/pci/rte_bus_pci.h +0 -0
index 06e004cd3..465a44935 100644
@@ -51,6 +51,13 @@ TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver);
 
 struct rte_devargs;
 
Quoted 8 lines

drivers/bus/pci/rte_bus_pci.h TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver); · line 59

Quoted 6 lines
drivers/bus/pci/rte_bus_pci.h +6 -0
+/* It's RTE_UUID_STRLEN, which is bigger than PCI_PRI_STR_SIZE. */
+#define RTE_PCI_NAME_LEN		(36 + 1)
+
+// XXX: we can't include rte_uuid.h directly due to the conflicts
+//      introduced by stdbool.h
+typedef unsigned char rte_uuid_t[16];

Does this need to have a the string ‘XXX’ in the comment? Note maybe a better word.

OK.

Thanks for the reviews/suggestions! Do appreciate it!

Regards,
Tiwei

#6 lore raw
In reply to Tiwei Bie
Quoted 22 lines · Tiwei Bie

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) bus support in DPDK. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this RFC, one mdev bus is introduced to scan the mdev devices in the system and do the probe based on the device-api.

Take the mdev devices whose device-api is "vfio-pci" as an example, in this RFC, these devices will be probed by a mdev driver provided by PCI bus, which will plug them to the PCI bus. And they will be probed with the drivers registered on the PCI bus based on VendorID/ DeviceID/... then.

                     +----------+
                     | mdev bus |
                     +----+-----+
                          |
         +----------------+----+------+------+
         |                     |      |      |
   mdev_vfio_pci               ......
Quoted 7 lines

(device-api: vfio-pci)

There are also other ways to add mdev device support in DPDK (e.g. let PCI bus scan /sys/bus/mdev/devices directly). Comments would be appreciated!

Hi Tiwei,

Thanks for the patchset. I was close to send a patchset with the same mdev support, but I'm glad to see your patchset first because I think it is interesting to see another view of how to implemented this.

After going through your patch I was a bit confused about how the mdev device to mdev driver match was done. But then I realized the approach you are following is different to my implementation, likely due to having different purposes. If I understand the idea behind, you want to have same PCI PMD drivers working with devices, PCI devices, created from mediated devices. That is the reason there is just one mdev driver, the one for vfio-pci mediated devices type.

My approach was different and I though having specific PMD mdev support was necessary, with the PMD requiring to register a mdev driver. I can see, after reading your patch, it can be perfectly possible to have the same PMDs for "pure" PCI devices and PCI devices made from mediated devices, and if the PMD requires to do something different due to the mediated devices intrinsics, then explicitly supporting that per PMD. I got specific ioctl calls between the PMD and the mediating driver but this can also be done with your approach.

I'm working on having a mediated PF, what is a different purpose than the Intel scalable I/O idea, so I will merge this patchset with my code and see if it works.

Thanks!

Quoted 51 lines

[1]
https://github.com/torvalds/linux/blob/master/Documentation/vfio-mediated-device.txt

Thanks,
Tiwei

Tiwei Bie (3):
eal: add a helper for reading string from sysfs
bus/mdev: add mdev bus support
bus/pci: add mdev support

 config/common_base                        |   5 +
 config/common_linux                       |   1 +
 drivers/bus/Makefile                      |   1 +
 drivers/bus/mdev/Makefile                 |  41 +++
 drivers/bus/mdev/linux/Makefile           |   6 +
 drivers/bus/mdev/linux/mdev.c             | 117 ++++++++
 drivers/bus/mdev/mdev.c                   | 310 ++++++++++++++++++++++
 drivers/bus/mdev/meson.build              |  15 ++
 drivers/bus/mdev/private.h                |  90 +++++++
 drivers/bus/mdev/rte_bus_mdev.h           | 141 ++++++++++
 drivers/bus/mdev/rte_bus_mdev_version.map |  12 +
 drivers/bus/meson.build                   |   2 +-
 drivers/bus/pci/Makefile                  |   3 +
 drivers/bus/pci/linux/Makefile            |   4 +
 drivers/bus/pci/linux/pci_vfio.c          |  35 ++-
 drivers/bus/pci/linux/pci_vfio_mdev.c     | 305 +++++++++++++++++++++
 drivers/bus/pci/meson.build               |   4 +-
 drivers/bus/pci/pci_common.c              |  17 +-
 drivers/bus/pci/private.h                 |   9 +
 drivers/bus/pci/rte_bus_pci.h             |  11 +-
 lib/librte_eal/common/eal_filesystem.h    |   7 +
 lib/librte_eal/freebsd/eal/eal.c          |  22 ++
 lib/librte_eal/linux/eal/eal.c            |  22 ++
 lib/librte_eal/rte_eal_version.map        |   1 +
 mk/rte.app.mk                             |   1 +
 25 files changed, 1163 insertions(+), 19 deletions(-)
 create mode 100644 drivers/bus/mdev/Makefile
 create mode 100644 drivers/bus/mdev/linux/Makefile
 create mode 100644 drivers/bus/mdev/linux/mdev.c
 create mode 100644 drivers/bus/mdev/mdev.c
 create mode 100644 drivers/bus/mdev/meson.build
 create mode 100644 drivers/bus/mdev/private.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev_version.map
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c

--
2.17.1

author
#7 lore raw
Quoted 35 lines · Alejandro Lucero
Quoted 22 lines · Tiwei Bie

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) bus support in DPDK. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this RFC, one mdev bus is introduced to scan the mdev devices in the system and do the probe based on the device-api.

Take the mdev devices whose device-api is "vfio-pci" as an example, in this RFC, these devices will be probed by a mdev driver provided by PCI bus, which will plug them to the PCI bus. And they will be probed with the drivers registered on the PCI bus based on VendorID/ DeviceID/... then.

                     +----------+
                     | mdev bus |
                     +----+-----+
                          |
         +----------------+----+------+------+
         |                     |      |      |
   mdev_vfio_pci               ......
Quoted 5 lines

(device-api: vfio-pci)

There are also other ways to add mdev device support in DPDK (e.g. let PCI bus scan /sys/bus/mdev/devices directly). Comments would be appreciated!

Hi Tiwei,

Thanks for the patchset. I was close to send a patchset with the same mdev support, but I'm glad to see your patchset first because I think it is interesting to see another view of how to implemented this.

Quoted 5 lines

After going through your patch I was a bit confused about how the mdev device to mdev driver match was done. But then I realized the approach you are following is different to my implementation, likely due to having different purposes. If I understand the idea behind, you want to have same PCI PMD drivers working with devices, PCI devices, created from mediated devices.

Exactly!

Quoted 12 lines

That
is the reason there is just one mdev driver, the one for vfio-pci mediated devices type.

My approach was different and I though having specific PMD mdev support was necessary, with the PMD requiring to register a mdev driver. I can see, after reading your patch, it can be perfectly possible to have the same PMDs for "pure" PCI devices and PCI devices made from mediated devices, and if the PMD requires to do something different due to the mediated devices intrinsics, then explicitly supporting that per PMD. I got specific ioctl calls between the PMD and the mediating driver but this can also be done with your approach.

Quoted 3 lines

I'm working on having a mediated PF, what is a different purpose than the Intel scalable I/O idea, so I will merge this patchset with my code and see if it works.

Cool! Thanks!

Quoted 54 lines

Thanks!

Quoted 48 lines

[1] https://github.com/torvalds/linux/blob/master/Documentation/vfio-mediated-device.txt

Thanks,
Tiwei

Tiwei Bie (3):
eal: add a helper for reading string from sysfs
bus/mdev: add mdev bus support
bus/pci: add mdev support

 config/common_base                        |   5 +
 config/common_linux                       |   1 +
 drivers/bus/Makefile                      |   1 +
 drivers/bus/mdev/Makefile                 |  41 +++
 drivers/bus/mdev/linux/Makefile           |   6 +
 drivers/bus/mdev/linux/mdev.c             | 117 ++++++++
 drivers/bus/mdev/mdev.c                   | 310 ++++++++++++++++++++++
 drivers/bus/mdev/meson.build              |  15 ++
 drivers/bus/mdev/private.h                |  90 +++++++
 drivers/bus/mdev/rte_bus_mdev.h           | 141 ++++++++++
 drivers/bus/mdev/rte_bus_mdev_version.map |  12 +
 drivers/bus/meson.build                   |   2 +-
 drivers/bus/pci/Makefile                  |   3 +
 drivers/bus/pci/linux/Makefile            |   4 +
 drivers/bus/pci/linux/pci_vfio.c          |  35 ++-
 drivers/bus/pci/linux/pci_vfio_mdev.c     | 305 +++++++++++++++++++++
 drivers/bus/pci/meson.build               |   4 +-
 drivers/bus/pci/pci_common.c              |  17 +-
 drivers/bus/pci/private.h                 |   9 +
 drivers/bus/pci/rte_bus_pci.h             |  11 +-
 lib/librte_eal/common/eal_filesystem.h    |   7 +
 lib/librte_eal/freebsd/eal/eal.c          |  22 ++
 lib/librte_eal/linux/eal/eal.c            |  22 ++
 lib/librte_eal/rte_eal_version.map        |   1 +
 mk/rte.app.mk                             |   1 +
 25 files changed, 1163 insertions(+), 19 deletions(-)
 create mode 100644 drivers/bus/mdev/Makefile
 create mode 100644 drivers/bus/mdev/linux/Makefile
 create mode 100644 drivers/bus/mdev/linux/mdev.c
 create mode 100644 drivers/bus/mdev/mdev.c
 create mode 100644 drivers/bus/mdev/meson.build
 create mode 100644 drivers/bus/mdev/private.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev_version.map
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c

--
2.17.1

#8 lore raw

Hi all,

I presented an approach in Fosdem
(https://archive.fosdem.org/2018/schedule/event/netmdev/) and feel happy someone is picking up.

If we step back a little, the mdev concept is to allow userland to be given a direct control over the hardware data path on a device still controlled by the kernel.
From a code base perspective, this can shrink down PMD code size b y a significant size: only 10% of the PMD code is actual data path, the rest being device control!
The concept is perfect for DPDK, SPDK and many other scenarios (AI accelerators).
Should the work be triggered by DPDK community, it should be applicable to a broader set of communities: SPDK, VPP, ODP, AF_XDP....

We bumped into many sharing (between kernel and userland) complexities particularly when a single PCI device controls two ports. So let's assume we try to solve a subset of the cases: coherent IO memory and a dedicated PCI space (by whatever mechanism) per port.

What are the "things to solve"?

1) enumeration: enumerating and "capturing" an mdev device (the patch I assume)
2) bifurcation: designating the queues to capture in userland (may be all) with a hardware driven rule (flow director or more generic)
3) memory management: dealing with rings and buffer management on rx and tx paths

The bifurcation can be as simple as : all queues in userland, or quite
rich: TCP port 80 goes to userland while the rest (ICMP...) go to kernel. If the kernel gets some of the traffic there will be a routing information sharing problem to solve. We had a few experiments here. Conclusion is its doable but many corner cases make it a big work. And it would be nice if the queue selection can be made very generic (and not tied to flow director).
Let's state this is for further study for now.

Lets focus on memory management of VFIO exposed devices.
I haven't refreshed my knowledge of the VFIO framework so you may want to correct a few points...
First of all, DPDK is made to switch packets and particularly between ports. With VFIO, this means all devices are in the same virtual IOVA which is tricky to implement in the kernel.
There are a few strategies to do that all requiring significant mdev extensions and more probably a kernel infrastructure change. The good news is it can be made in such a way that selected drivers implement the change, not requiring all the drivers to be touched.
Another big question is: is the kernel allocating the memory then the userland gets a map to it, or does the userland allocates the memory and the kernel just maintains the IOVA mapping.
I would favor kernel allocation and userland gets a map to it (in the unified IOVA). One reason being that memory allocation strategy can be very different from hardware to hardware:
- driver allocates packet buffers and populate a single ring of packet per queue
- driver allocates packet buffers of different sizes and populate multiple rings per queue (for instance rings of 128, 256, 1024, 2048 byte arrays per queue)
- driver allocates an unstructured memory area (say 32MB) and give it to hardware (no prepopulation of rings).
So the userland framework (DPDK, SPDK, ODP, VPP, AF_XDP, proprietary...) can just query for queues and rings to the kernel driver that knows what has to be done for the driver. The userland framework just has to create the relevant objects (queues, rings, packet buffers) to the provided kernel information.

Exposing VFIO devices to DPDK and other frameworks is a major topic, and I suggest that at the same time enumeration is done, a broader discussion on the data path itself happens.
Data path discussion is about memory management (above) and packet descriptors. Exposing hardware dependent structures in the userland is not the most widely accepted wisdom.
So I would rather assume hardware natively produce hardware, vendor, OS independent descriptors. Candidates can be: DPDK mbuf, VPP vlib_buf or virtio 1.1. I would favor a packet descriptor that supports a combination of inline offloads (VxLAN + IPSec + TSO...) : if virtio 1.1 could be extended with some DPDK mbuf fields that would be perfect ;-) That looks science fiction but I know that some smartNICs and other hardware, the hardware produced packet descriptor format can be flexible....

Cheers

FF

Quoted 57 lines · Tiwei Bie
Quoted 35 lines · Alejandro Lucero
Quoted 22 lines · Tiwei Bie

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) bus support in DPDK. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this RFC, one mdev bus is introduced to scan the mdev devices in the system and do the probe based on the device-api.

Take the mdev devices whose device-api is "vfio-pci" as an example, in this RFC, these devices will be probed by a mdev driver provided by PCI bus, which will plug them to the PCI bus. And they will be probed with the drivers registered on the PCI bus based on VendorID/ DeviceID/... then.

                     +----------+
                     | mdev bus |
                     +----+-----+
                          |
         +----------------+----+------+------+
         |                     |      |      |
   mdev_vfio_pci               ......
Quoted 5 lines

(device-api: vfio-pci)

There are also other ways to add mdev device support in DPDK (e.g. let PCI bus scan /sys/bus/mdev/devices directly). Comments would be appreciated!

Hi Tiwei,

Thanks for the patchset. I was close to send a patchset with the same mdev support, but I'm glad to see your patchset first because I think it is interesting to see another view of how to implemented this.

Quoted 5 lines

After going through your patch I was a bit confused about how the mdev device to mdev driver match was done. But then I realized the approach you are following is different to my implementation, likely due to having different purposes. If I understand the idea behind, you want to have same PCI PMD drivers working with devices, PCI devices, created from mediated devices.

Exactly!

Quoted 12 lines

That
is the reason there is just one mdev driver, the one for vfio-pci mediated devices type.

My approach was different and I though having specific PMD mdev support was necessary, with the PMD requiring to register a mdev driver. I can see, after reading your patch, it can be perfectly possible to have the same PMDs for "pure" PCI devices and PCI devices made from mediated devices, and if the PMD requires to do something different due to the mediated devices intrinsics, then explicitly supporting that per PMD. I got specific ioctl calls between the PMD and the mediating driver but this can also be done with your approach.

Quoted 60 lines
Quoted 3 lines

I'm working on having a mediated PF, what is a different purpose than the Intel scalable I/O idea, so I will merge this patchset with my code and see if it works.

Cool! Thanks!

Quoted 54 lines

Thanks!

Quoted 48 lines

[1] https://github.com/torvalds/linux/blob/master/Documentation/vfio-mediated-device.txt

Thanks,
Tiwei

Tiwei Bie (3):
eal: add a helper for reading string from sysfs
bus/mdev: add mdev bus support
bus/pci: add mdev support

 config/common_base                        |   5 +
 config/common_linux                       |   1 +
 drivers/bus/Makefile                      |   1 +
 drivers/bus/mdev/Makefile                 |  41 +++
 drivers/bus/mdev/linux/Makefile           |   6 +
 drivers/bus/mdev/linux/mdev.c             | 117 ++++++++
 drivers/bus/mdev/mdev.c                   | 310 ++++++++++++++++++++++
 drivers/bus/mdev/meson.build              |  15 ++
 drivers/bus/mdev/private.h                |  90 +++++++
 drivers/bus/mdev/rte_bus_mdev.h           | 141 ++++++++++
 drivers/bus/mdev/rte_bus_mdev_version.map |  12 +
 drivers/bus/meson.build                   |   2 +-
 drivers/bus/pci/Makefile                  |   3 +
 drivers/bus/pci/linux/Makefile            |   4 +
 drivers/bus/pci/linux/pci_vfio.c          |  35 ++-
 drivers/bus/pci/linux/pci_vfio_mdev.c     | 305 +++++++++++++++++++++
 drivers/bus/pci/meson.build               |   4 +-
 drivers/bus/pci/pci_common.c              |  17 +-
 drivers/bus/pci/private.h                 |   9 +
 drivers/bus/pci/rte_bus_pci.h             |  11 +-
 lib/librte_eal/common/eal_filesystem.h    |   7 +
 lib/librte_eal/freebsd/eal/eal.c          |  22 ++
 lib/librte_eal/linux/eal/eal.c            |  22 ++
 lib/librte_eal/rte_eal_version.map        |   1 +
 mk/rte.app.mk                             |   1 +
 25 files changed, 1163 insertions(+), 19 deletions(-)
 create mode 100644 drivers/bus/mdev/Makefile
 create mode 100644 drivers/bus/mdev/linux/Makefile
 create mode 100644 drivers/bus/mdev/linux/mdev.c
 create mode 100644 drivers/bus/mdev/mdev.c
 create mode 100644 drivers/bus/mdev/meson.build
 create mode 100644 drivers/bus/mdev/private.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev.h
 create mode 100644 drivers/bus/mdev/rte_bus_mdev_version.map
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c

--
2.17.1

--
François-Frédéric Ozog | Director Linaro Edge & Fog Computing Group
T: +33.67221.6485
[email protected] | Skype: ffozog

author
#10 lore raw

v2 cover Add mdev (Mediated device) support in DPDK

In reply to Tiwei Bie

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this RFCv2, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

RFCv2:
- Let PCI bus scan mediated PCI devices directly
- Address Keith's comments
- Merge below patch into this series (David)
http://patches.dpdk.org/patch/55927/
- Add internal representation of PCI device (David)
- Minor fixes and improvements

[1] https://github.com/torvalds/linux/blob/master/Documentation/vfio-mediated-device.txt

Thanks,
Tiwei

Tiwei Bie (5):
bus/pci: introduce an internal representation of PCI device bus/pci: avoid depending on private value in kernel source bus/pci: introduce helper for MMIO read and write
eal: add a helper for reading string from sysfs bus/pci: add mdev support

 drivers/bus/pci/bsd/pci.c               |  36 ++-
 drivers/bus/pci/linux/Makefile          |   1 +
 drivers/bus/pci/linux/pci.c             | 105 ++++++--
 drivers/bus/pci/linux/pci_init.h        |  29 ++-
 drivers/bus/pci/linux/pci_uio.c         |  22 ++
 drivers/bus/pci/linux/pci_vfio.c        | 314 ++++++++++++++++++++----
 drivers/bus/pci/linux/pci_vfio_mdev.c   | 236 ++++++++++++++++++
 drivers/bus/pci/meson.build             |   3 +-
 drivers/bus/pci/pci_common.c            |  31 +--
 drivers/bus/pci/private.h               |  22 ++
 drivers/bus/pci/rte_bus_pci.h           |  65 ++++-
 drivers/bus/pci/rte_bus_pci_version.map |   7 +
 lib/librte_eal/common/eal_filesystem.h  |  10 +
 lib/librte_eal/freebsd/eal/eal.c        |  22 ++
 lib/librte_eal/linux/eal/eal.c          |  39 ++-
 lib/librte_eal/rte_eal_version.map      |   1 +
 16 files changed, 838 insertions(+), 105 deletions(-)
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
Signature · 1 line
2.17.1
author
#11 lore raw

RFC v2 1/5 bus/pci: introduce an internal representation of PCI device

This patch introduces an internal representation of the PCI device which will be used to store the internal information that don't have to be exposed, e.g. the VFIO region sizes/offsets.

In this patch, the internal structure is simply a wrapper of the rte_pci_device structure. More fields will be added in the coming patches.

---
 drivers/bus/pci/bsd/pci.c    | 14 ++++++++------
 drivers/bus/pci/linux/pci.c  | 25 ++++++++++++++-----------
 drivers/bus/pci/pci_common.c |  2 +-
 drivers/bus/pci/private.h    | 12 ++++++++++++
 4 files changed, 35 insertions(+), 18 deletions(-)
drivers/bus/pci/bsd/pci.c +8 -6
index a2de70910..636868f38 100644
@@ -213,16 +213,18 @@ pci_uio_map_resource_by_index(struct rte_pci_device *dev, int res_idx,
 static int
 pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 {
+	struct rte_pci_device_internal *pdev;
 	struct rte_pci_device *dev;
 	struct pci_bar_io bar;
 	unsigned i, max;
 
-	dev = malloc(sizeof(*dev));
-	if (dev == NULL) {
+	pdev = malloc(sizeof(*pdev));
+	if (pdev == NULL)
 		return -1;
-	}
 
-	memset(dev, 0, sizeof(*dev));
+	memset(pdev, 0, sizeof(*pdev));
+
+	dev = &pdev->device;
 	dev->device.bus = &rte_pci_bus.bus;
 
 	dev->addr.domain = conf->pc_sel.pc_domain;
@@ -308,7 +310,7 @@ pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 				memmove(dev2->mem_resource,
 					dev->mem_resource,
 					sizeof(dev->mem_resource));
-				free(dev);
+				free(pdev);
 			}
 			return 0;
 		}
@@ -318,7 +320,7 @@ pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 	return 0;
 
 skipdev:
-	free(dev);
+	free(pdev);
 	return 0;
 }
 
drivers/bus/pci/linux/pci.c +14 -11
index 33c8ea7e9..dfab7b81b 100644
@@ -219,22 +219,25 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 {
 	char filename[PATH_MAX];
 	unsigned long tmp;
+	struct rte_pci_device_internal *pdev;
 	struct rte_pci_device *dev;
 	char driver[PATH_MAX];
 	int ret;
 
-	dev = malloc(sizeof(*dev));
-	if (dev == NULL)
+	pdev = malloc(sizeof(*pdev));
+	if (pdev == NULL)
 		return -1;
 
-	memset(dev, 0, sizeof(*dev));
+	memset(pdev, 0, sizeof(*pdev));
+
+	dev = &pdev->device;
 	dev->device.bus = &rte_pci_bus.bus;
 	dev->addr = *addr;
 
 	/* get vendor id */
 	snprintf(filename, sizeof(filename), "%s/vendor", dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.vendor_id = (uint16_t)tmp;
@@ -242,7 +245,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	/* get device id */
 	snprintf(filename, sizeof(filename), "%s/device", dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.device_id = (uint16_t)tmp;
@@ -251,7 +254,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/subsystem_vendor",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.subsystem_vendor_id = (uint16_t)tmp;
@@ -260,7 +263,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/subsystem_device",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.subsystem_device_id = (uint16_t)tmp;
@@ -269,7 +272,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/class",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	/* the least 24 bits are valid: class, subclass, program interface */
@@ -309,7 +312,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/resource", dirname);
 	if (pci_parse_sysfs_resource(filename, dev) < 0) {
 		RTE_LOG(ERR, EAL, "%s(): cannot parse resource\n", __func__);
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 
@@ -318,7 +321,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	ret = pci_get_kernel_driver_by_path(filename, driver, sizeof(driver));
 	if (ret < 0) {
 		RTE_LOG(ERR, EAL, "Fail to get kernel driver\n");
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 
@@ -382,7 +385,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 						RTE_LOG(ERR, EAL, "Unexpected device scan at %s!\n",
 							filename);
 				}
-				free(dev);
+				free(pdev);
 			}
 			return 0;
 		}
drivers/bus/pci/pci_common.c +1 -1
index d2af472ef..8b9deca8b 100644
@@ -523,7 +523,7 @@ pci_unplug(struct rte_device *dev)
 	if (ret == 0) {
 		rte_pci_remove_device(pdev);
 		rte_devargs_remove(dev->devargs);
-		free(pdev);
+		free(RTE_PCI_DEVICE_INTERNAL(pdev));
 	}
 	return ret;
 }
drivers/bus/pci/private.h +12 -0
index 8a5524052..3e2abd818 100644
@@ -10,6 +10,14 @@
 #include <rte_pci.h>
 #include <rte_bus_pci.h>
 
+/*
+ * Convert struct rte_pci_device to struct rte_pci_device_internal
+ */
+#define RTE_PCI_DEVICE_INTERNAL(ptr) \
+	container_of(ptr, struct rte_pci_device_internal, device)
+#define RTE_PCI_DEVICE_INTERNAL_CONST(ptr) \
+	container_of(ptr, const struct rte_pci_device_internal, device)
+
 extern struct rte_pci_bus rte_pci_bus;
 
 struct rte_pci_driver;
@@ -17,6 +25,10 @@ struct rte_pci_device;
 
 extern struct rte_pci_bus rte_pci_bus;
 
+struct rte_pci_device_internal {
+	struct rte_pci_device device;
+};
+
 /**
  * Probe the PCI bus
  *
Signature · 1 line
2.17.1
author
#12 lore raw

RFC v2 2/5 bus/pci: avoid depending on private value in kernel source

In reply to Tiwei Bie

The value 40 used in VFIO_GET_REGION_ADDR() is a private value (VFIO_PCI_OFFSET_SHIFT) defined in Linux kernel source [1]. It is not part of VFIO API, and we should not depend on it.

[1] https://github.com/torvalds/linux/blob/6fbc7275c7a9/drivers/vfio/pci/vfio_pci_private.h#L19

---
 drivers/bus/pci/linux/pci.c      |   4 +-
 drivers/bus/pci/linux/pci_init.h |   4 +-
 drivers/bus/pci/linux/pci_vfio.c | 176 ++++++++++++++++++++++++-------
 drivers/bus/pci/private.h        |  10 ++
 4 files changed, 154 insertions(+), 40 deletions(-)
drivers/bus/pci/linux/pci.c +2 -2
index dfab7b81b..00bfbb301 100644
@@ -639,7 +639,7 @@ int rte_pci_read_config(const struct rte_pci_device *device,
 		return pci_uio_read_config(intr_handle, buf, len, offset);
 #ifdef VFIO_PRESENT
 	case RTE_KDRV_VFIO:
-		return pci_vfio_read_config(intr_handle, buf, len, offset);
+		return pci_vfio_read_config(device, buf, len, offset);
 #endif
 	default:
 		rte_pci_device_name(&device->addr, devname,
@@ -663,7 +663,7 @@ int rte_pci_write_config(const struct rte_pci_device *device,
 		return pci_uio_write_config(intr_handle, buf, len, offset);
 #ifdef VFIO_PRESENT
 	case RTE_KDRV_VFIO:
-		return pci_vfio_write_config(intr_handle, buf, len, offset);
+		return pci_vfio_write_config(device, buf, len, offset);
 #endif
 	default:
 		rte_pci_device_name(&device->addr, devname,
drivers/bus/pci/linux/pci_init.h +2 -2
index c2e603a37..c6542a8f9 100644
@@ -64,9 +64,9 @@ int pci_uio_ioport_unmap(struct rte_pci_ioport *p);
 #endif
 
 /* access config space */
-int pci_vfio_read_config(const struct rte_intr_handle *intr_handle,
+int pci_vfio_read_config(const struct rte_pci_device *dev,
 			 void *buf, size_t len, off_t offs);
-int pci_vfio_write_config(const struct rte_intr_handle *intr_handle,
+int pci_vfio_write_config(const struct rte_pci_device *dev,
 			  const void *buf, size_t len, off_t offs);
 
 int pci_vfio_ioport_map(struct rte_pci_device *dev, int bar,
drivers/bus/pci/linux/pci_vfio.c +140 -36

317 lines not on this page · raw message

drivers/bus/pci/private.h +10 -0
index 3e2abd818..c09185b86 100644
@@ -10,6 +10,8 @@
 #include <rte_pci.h>
 #include <rte_bus_pci.h>
 
+#define RTE_MAX_PCI_REGIONS	9
+
 /*
  * Convert struct rte_pci_device to struct rte_pci_device_internal
  */
@@ -25,8 +27,16 @@ struct rte_pci_device;
 
 extern struct rte_pci_bus rte_pci_bus;
 
+struct rte_pci_region {
+	uint64_t size;
+	uint64_t offset;
+};
+
 struct rte_pci_device_internal {
 	struct rte_pci_device device;
+
+	/* PCI regions provided by e.g. VFIO. */
+	struct rte_pci_region region[RTE_MAX_PCI_REGIONS];
 };
 
 /**
Signature · 1 line
2.17.1
author
#13 lore raw

RFC v2 3/5 bus/pci: introduce helper for MMIO read and write

In reply to Tiwei Bie

The MMIO regions may not be mmap-able for mediated PCI device. In this case, the application should explicitly do read and write to access these regions.

---
 drivers/bus/pci/bsd/pci.c               | 22 ++++++++++++
 drivers/bus/pci/linux/pci.c             | 46 ++++++++++++++++++++++++
 drivers/bus/pci/linux/pci_init.h        | 10 ++++++
 drivers/bus/pci/linux/pci_uio.c         | 22 ++++++++++++
 drivers/bus/pci/linux/pci_vfio.c        | 36 +++++++++++++++++++
 drivers/bus/pci/rte_bus_pci.h           | 48 +++++++++++++++++++++++++
 drivers/bus/pci/rte_bus_pci_version.map |  7 ++++
 7 files changed, 191 insertions(+)
drivers/bus/pci/bsd/pci.c +22 -0
index 636868f38..d4d2c9016 100644
@@ -527,6 +527,28 @@ int rte_pci_write_config(const struct rte_pci_device *dev,
 	return -1;
 }
 
+/* Read PCI MMIO space. */
+int rte_pci_mmio_read(const struct rte_pci_device *dev, int bar,
+		      void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy(buf, (uint8_t *)dev->mem_resource[bar].addr + offset, len);
+	return len;
+}
+
+/* Write PCI MMIO space. */
+int rte_pci_mmio_write(const struct rte_pci_device *dev, int bar,
+		       const void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy((uint8_t *)dev->mem_resource[bar].addr + offset, buf, len);
+	return len;
+}
+
 int
 rte_pci_ioport_map(struct rte_pci_device *dev, int bar,
 		struct rte_pci_ioport *p)
drivers/bus/pci/linux/pci.c +46 -0

52 lines not on this page · raw message

drivers/bus/pci/linux/pci_init.h +10 -0
index c6542a8f9..158a16977 100644
@@ -35,6 +35,11 @@ int pci_uio_read_config(const struct rte_intr_handle *intr_handle,
 int pci_uio_write_config(const struct rte_intr_handle *intr_handle,
 			 const void *buf, size_t len, off_t offs);
 
+int pci_uio_mmio_read(const struct rte_pci_device *dev, int bar,
+		      void *buf, size_t len, off_t offset);
+int pci_uio_mmio_write(const struct rte_pci_device *dev, int bar,
+		       const void *buf, size_t len, off_t offset);
+
 int pci_uio_ioport_map(struct rte_pci_device *dev, int bar,
 		       struct rte_pci_ioport *p);
 void pci_uio_ioport_read(struct rte_pci_ioport *p,
@@ -69,6 +74,11 @@ int pci_vfio_read_config(const struct rte_pci_device *dev,
 int pci_vfio_write_config(const struct rte_pci_device *dev,
 			  const void *buf, size_t len, off_t offs);
 
+int pci_vfio_mmio_read(const struct rte_pci_device *dev, int bar,
+		       void *buf, size_t len, off_t offset);
+int pci_vfio_mmio_write(const struct rte_pci_device *dev, int bar,
+			const void *buf, size_t len, off_t offset);
+
 int pci_vfio_ioport_map(struct rte_pci_device *dev, int bar,
 		        struct rte_pci_ioport *p);
 void pci_vfio_ioport_read(struct rte_pci_ioport *p,
drivers/bus/pci/linux/pci_uio.c +22 -0
index f240fe4f2..623273541 100644
@@ -45,6 +45,28 @@ pci_uio_write_config(const struct rte_intr_handle *intr_handle,
 	return pwrite(intr_handle->uio_cfg_fd, buf, len, offset);
 }
 
+int
+pci_uio_mmio_read(const struct rte_pci_device *dev, int bar,
+		  void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy(buf, (uint8_t *)dev->mem_resource[bar].addr + offset, len);
+	return len;
+}
+
+int
+pci_uio_mmio_write(const struct rte_pci_device *dev, int bar,
+		   const void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy((uint8_t *)dev->mem_resource[bar].addr + offset, buf, len);
+	return len;
+}
+
 static int
 pci_uio_set_bus_master(int dev_fd)
 {
drivers/bus/pci/linux/pci_vfio.c +36 -0
index 2dc4a9299..204698be0 100644
@@ -1164,6 +1164,42 @@ pci_vfio_ioport_unmap(struct rte_pci_ioport *p)
 	return -1;
 }
 
+int
+pci_vfio_mmio_read(const struct rte_pci_device *dev, int bar,
+		   void *buf, size_t len, off_t offs)
+{
+	uint64_t size, offset;
+	int fd;
+
+	fd = dev->intr_handle.vfio_dev_fd;
+
+	if (pci_vfio_get_region(dev, bar, &size, &offset) != 0)
+		return -1;
+
+	if ((uint64_t)len + offs > size)
+		return -1;
+
+	return pread64(fd, buf, len, offset + offs);
+}
+
+int
+pci_vfio_mmio_write(const struct rte_pci_device *dev, int bar,
+		    const void *buf, size_t len, off_t offs)
+{
+	uint64_t size, offset;
+	int fd;
+
+	fd = dev->intr_handle.vfio_dev_fd;
+
+	if (pci_vfio_get_region(dev, bar, &size, &offset) != 0)
+		return -1;
+
+	if ((uint64_t)len + offs > size)
+		return -1;
+
+	return pwrite64(fd, buf, len, offset + offs);
+}
+
 int
 pci_vfio_is_enabled(void)
 {
drivers/bus/pci/rte_bus_pci.h +48 -0

54 lines not on this page · raw message

drivers/bus/pci/rte_bus_pci_version.map +7 -0
index 27e9c4f10..141bdf48e 100644
@@ -16,3 +16,10 @@ DPDK_17.11 {
 
 	local: *;
 };
+
+EXPERIMENTAL {
+	global:
+
+	rte_pci_mmio_read;
+	rte_pci_mmio_write;
+};
Signature · 1 line
2.17.1
author
#14 lore raw

RFC v2 4/5 eal: add a helper for reading string from sysfs

In reply to Tiwei Bie

This patch adds a helper for reading string from sysfs.

---
 lib/librte_eal/common/eal_filesystem.h | 10 ++++++++++
 lib/librte_eal/freebsd/eal/eal.c       | 22 ++++++++++++++++++++++
 lib/librte_eal/linux/eal/eal.c         | 22 ++++++++++++++++++++++
 lib/librte_eal/rte_eal_version.map     |  1 +
 4 files changed, 55 insertions(+)
lib/librte_eal/common/eal_filesystem.h +10 -0
index 5d21f07c2..be4c51ebb 100644
@@ -104,4 +104,14 @@ eal_get_hugefile_path(char *buffer, size_t buflen, const char *hugedir, int f_id
  * Used to read information from files on /sys */
 int eal_parse_sysfs_value(const char *filename, unsigned long *val);
 
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice
+ *
+ * Function to read a line from a file on the filesystem.
+ * Used to read information from files on /sys
+ */
+__rte_experimental
+int rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz);
+
 #endif /* EAL_FILESYSTEM_H */
lib/librte_eal/freebsd/eal/eal.c +22 -0
index d53f0fe69..78720685f 100644
@@ -209,6 +209,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/librte_eal/linux/eal/eal.c +22 -0
index 2e5499f9b..44bad45d3 100644
@@ -295,6 +295,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/librte_eal/rte_eal_version.map +1 -0
index 1892d9ea9..a9559176b 100644
@@ -331,6 +331,7 @@ EXPERIMENTAL {
 	rte_dev_hotplug_handle_enable;
 	rte_dev_iterator_init;
 	rte_dev_iterator_next;
+	rte_eal_parse_sysfs_str;
 	rte_extmem_attach;
 	rte_extmem_detach;
 	rte_extmem_register;
Signature · 1 line
2.17.1
author
#15 lore raw

RFC v2 5/5 bus/pci: add mdev support

In reply to Tiwei Bie

This patch adds the mdev (Mediated device) support in PCI bus driver. With this patch, the PCI bus driver will be able to scan and probe the mediated PCI devices (i.e. the Mediated devices whose device API is "vfio-pci") in the system.

There are several things different between physical PCI devices and mediated PCI devices:

- Mediated PCI devices have to be accessed through VFIO API;
- The regions in mediated PCI devices may not be mmap-able, and drivers need to call read/write function to access them in this case;
- Mediated PCI devices use UUID as device address;

---
 drivers/bus/pci/linux/Makefile        |   1 +
 drivers/bus/pci/linux/pci.c           |  30 +++-
 drivers/bus/pci/linux/pci_init.h      |  15 +-
 drivers/bus/pci/linux/pci_vfio.c      | 104 ++++++++++--
 drivers/bus/pci/linux/pci_vfio_mdev.c | 236 ++++++++++++++++++++++++++
 drivers/bus/pci/meson.build           |   3 +-
 drivers/bus/pci/pci_common.c          |  29 ++--
 drivers/bus/pci/rte_bus_pci.h         |  17 +-
 lib/librte_eal/linux/eal/eal.c        |  17 +-
 9 files changed, 404 insertions(+), 48 deletions(-)
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
drivers/bus/pci/linux/Makefile +1 -0
index 90404468b..c17ab2484 100644
@@ -4,3 +4,4 @@
 SRCS += pci.c
 SRCS += pci_uio.c
 SRCS += pci_vfio.c
+SRCS += pci_vfio_mdev.c
drivers/bus/pci/linux/pci.c +22 -8
index bdfc8c5ff..5be898803 100644
@@ -31,7 +31,7 @@
 
 extern struct rte_pci_bus rte_pci_bus;
 
-static int
+int
 pci_get_kernel_driver_by_path(const char *filename, char *dri_name,
 			      size_t len)
 {
@@ -71,7 +71,7 @@ rte_pci_map_device(struct rte_pci_device *dev)
 	switch (dev->kdrv) {
 	case RTE_KDRV_VFIO:
 #ifdef VFIO_PRESENT
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			ret = pci_vfio_map_resource(dev);
 #endif
 		break;
@@ -100,7 +100,7 @@ rte_pci_unmap_device(struct rte_pci_device *dev)
 	switch (dev->kdrv) {
 	case RTE_KDRV_VFIO:
 #ifdef VFIO_PRESENT
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			pci_vfio_unmap_resource(dev);
 #endif
 		break;
@@ -348,6 +348,15 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 		int ret;
 
 		TAILQ_FOREACH(dev2, &rte_pci_bus.device_list, next) {
+			/*
+			 * Insert physical PCI devices before all mediated
+			 * PCI devices.
+			 */
+			if (dev2->is_mdev) {
+				rte_pci_insert_device(dev2, dev);
+				return 0;
+			}
+
 			ret = rte_pci_addr_cmp(&dev->addr, &dev2->addr);
 			if (ret > 0)
 				continue;
@@ -471,8 +480,14 @@ rte_pci_scan(void)
 		return 0;
 
 #ifdef VFIO_PRESENT
-	if (!pci_vfio_is_enabled())
-		RTE_LOG(DEBUG, EAL, "VFIO PCI modules not loaded\n");
+	if (!rte_vfio_is_enabled("vfio_pci"))
+		RTE_LOG(DEBUG, EAL, "VFIO PCI module not loaded\n");
+
+	if (!rte_vfio_is_enabled("vfio_mdev"))
+		RTE_LOG(DEBUG, EAL, "VFIO MDEV module not loaded\n");
+
+	if (pci_scan_mdev() != 0)
+		return -1;
 #endif
 
 	dir = opendir(rte_pci_get_sysfs_path());
@@ -788,7 +803,7 @@ rte_pci_ioport_map(struct rte_pci_device *dev, int bar,
 	switch (dev->kdrv) {
 #ifdef VFIO_PRESENT
 	case RTE_KDRV_VFIO:
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			ret = pci_vfio_ioport_map(dev, bar, p);
 		break;
 #endif
@@ -877,8 +892,7 @@ rte_pci_ioport_unmap(struct rte_pci_ioport *p)
 	switch (p->dev->kdrv) {
 #ifdef VFIO_PRESENT
 	case RTE_KDRV_VFIO:
-		if (pci_vfio_is_enabled())
-			ret = pci_vfio_ioport_unmap(p);
+		ret = -1;
 		break;
 #endif
 	case RTE_KDRV_IGB_UIO:
drivers/bus/pci/linux/pci_init.h +14 -1
index 158a16977..12739ba51 100644
@@ -17,6 +17,9 @@
 extern void *pci_map_addr;
 void *pci_find_max_end_va(void);
 
+int pci_get_kernel_driver_by_path(const char *filename, char *dri_name,
+				  size_t len);
+
 /* parse one line of the "resource" sysfs file (note that the 'line'
  * string is modified)
  */
@@ -91,7 +94,17 @@ int pci_vfio_ioport_unmap(struct rte_pci_ioport *p);
 int pci_vfio_map_resource(struct rte_pci_device *dev);
 int pci_vfio_unmap_resource(struct rte_pci_device *dev);
 
-int pci_vfio_is_enabled(void);
+int pci_vfio_is_enabled(struct rte_pci_device *dev);
+
+int pci_vfio_fill_regions(struct rte_pci_device *dev, int vfio_dev_fd,
+			  struct vfio_device_info *device_info);
+
+int pci_vfio_get_pci_id(struct rte_pci_device *dev, int vfio_dev_fd,
+			struct rte_pci_id *pci_id);
+
+const char *pci_mdev_get_sysfs_path(void);
+
+int pci_scan_mdev(void);
 
 #endif
 
drivers/bus/pci/linux/pci_vfio.c +93 -11

188 lines not on this page · raw message

drivers/bus/pci/linux/pci_vfio_mdev.c new file +236 -0

236 lines not on this page · raw message

drivers/bus/pci/meson.build +2 -1
index a312ecc03..890b7bda0 100644
@@ -11,7 +11,8 @@ sources = files('pci_common.c',
 if is_linux
 	sources += files('linux/pci.c',
 			'linux/pci_uio.c',
-			'linux/pci_vfio.c')
+			'linux/pci_vfio.c',
+			'linux/pci_vfio_mdev.c')
 	includes += include_directories('linux')
 else
 	sources += files('bsd/pci.c')
drivers/bus/pci/pci_common.c +12 -17
index 8b9deca8b..ec314cb07 100644
@@ -25,6 +25,7 @@
 #include <rte_common.h>
 #include <rte_devargs.h>
 #include <rte_vfio.h>
+#include <rte_uuid.h>
 
 #include "private.h"
 
@@ -61,8 +62,10 @@ pci_name_set(struct rte_pci_device *dev)
 	struct rte_devargs *devargs;
 
 	/* Each device has its internal, canonical name set. */
-	rte_pci_device_name(&dev->addr,
-			dev->name, sizeof(dev->name));
+	if (dev->is_mdev)
+		rte_uuid_unparse(dev->uuid, dev->name, sizeof(dev->name));
+	else
+		rte_pci_device_name(&dev->addr, dev->name, sizeof(dev->name));
 	devargs = pci_devargs_lookup(dev);
 	dev->device.devargs = devargs;
 	/* In blacklist mode, if the device is not blacklisted, no
@@ -124,21 +127,17 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 {
 	int ret;
 	bool already_probed;
-	struct rte_pci_addr *loc;
 
 	if ((dr == NULL) || (dev == NULL))
 		return -EINVAL;
 
-	loc = &dev->addr;
-
 	/* The device is not blacklisted; Check if driver supports it */
 	if (!rte_pci_match(dr, dev))
 		/* Match of device and driver failed */
 		return 1;
 
-	RTE_LOG(INFO, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid, loc->function,
-			dev->device.numa_node);
+	RTE_LOG(INFO, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	/* no initialization when blacklisted, return without error */
 	if (dev->device.devargs != NULL &&
@@ -208,7 +207,6 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
 static int
 rte_pci_detach_dev(struct rte_pci_device *dev)
 {
-	struct rte_pci_addr *loc;
 	struct rte_pci_driver *dr;
 	int ret = 0;
 
@@ -216,11 +214,9 @@ rte_pci_detach_dev(struct rte_pci_device *dev)
 		return -EINVAL;
 
 	dr = dev->driver;
-	loc = &dev->addr;
 
-	RTE_LOG(DEBUG, EAL, "PCI device "PCI_PRI_FMT" on NUMA socket %i\n",
-			loc->domain, loc->bus, loc->devid,
-			loc->function, dev->device.numa_node);
+	RTE_LOG(DEBUG, EAL, "PCI device %s on NUMA socket %i\n",
+		dev->name, dev->device.numa_node);
 
 	RTE_LOG(DEBUG, EAL, "  remove driver: %x:%x %s\n", dev->id.vendor_id,
 			dev->id.device_id, dr->driver.name);
@@ -297,10 +293,9 @@ rte_pci_probe(void)
 			ret = pci_probe_all_drivers(dev);
 		if (ret < 0) {
 			if (ret != -EEXIST) {
-				RTE_LOG(ERR, EAL, "Requested device "
-					PCI_PRI_FMT " cannot be used\n",
-					dev->addr.domain, dev->addr.bus,
-					dev->addr.devid, dev->addr.function);
+				RTE_LOG(ERR, EAL,
+					"Requested device %s cannot be used\n",
+					dev->name);
 				rte_errno = errno;
 				failed++;
 			}
drivers/bus/pci/rte_bus_pci.h +15 -2
index 86527b421..47e669e9c 100644
@@ -51,13 +51,26 @@ TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver);
 
 struct rte_devargs;
 
+/*
+ * NOTE: we can't include rte_uuid.h directly due to the conflicts
+ *      introduced by stdbool.h
+ */
+typedef unsigned char rte_uuid_t[16];
+
+/* It's RTE_UUID_STRLEN, which is bigger than PCI_PRI_STR_SIZE. */
+#define RTE_PCI_NAME_LEN		(36 + 1)
+
 /**
  * A structure describing a PCI device.
  */
 struct rte_pci_device {
 	TAILQ_ENTRY(rte_pci_device) next;   /**< Next probed PCI device. */
 	struct rte_device device;           /**< Inherit core device */
-	struct rte_pci_addr addr;           /**< PCI location. */
+	union {
+		struct rte_pci_addr addr;   /**< PCI location. */
+		rte_uuid_t uuid;            /**< Mdev location. */
+	};
+	uint8_t is_mdev;                    /**< True for mediated PCI device */
 	struct rte_pci_id id;               /**< PCI ID. */
 	struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
 					    /**< PCI Memory Resource */
@@ -65,7 +78,7 @@ struct rte_pci_device {
 	struct rte_pci_driver *driver;      /**< PCI driver used in probing */
 	uint16_t max_vfs;                   /**< sriov enable if not zero */
 	enum rte_kernel_driver kdrv;        /**< Kernel driver passthrough */
-	char name[PCI_PRI_STR_SIZE+1];      /**< PCI location (ASCII) */
+	char name[RTE_PCI_NAME_LEN];        /**< PCI/Mdev location (ASCII) */
 	struct rte_intr_handle vfio_req_intr_handle;
 				/**< Handler of VFIO request interrupt */
 };
lib/librte_eal/linux/eal/eal.c +9 -8
index 44bad45d3..942148180 100644
@@ -1068,6 +1068,15 @@ rte_eal_init(int argc, char **argv)
 		return -1;
 	}
 
+#ifdef VFIO_PRESENT
+	if (rte_eal_vfio_setup() < 0) {
+		rte_eal_init_alert("Cannot init VFIO");
+		rte_errno = EAGAIN;
+		rte_atomic32_clear(&run_once);
+		return -1;
+	}
+#endif
+
 	if (rte_bus_scan()) {
 		rte_eal_init_alert("Cannot scan the buses for devices");
 		rte_errno = ENODEV;
@@ -1151,14 +1160,6 @@ rte_eal_init(int argc, char **argv)
 		return -1;
 	}
 
-#ifdef VFIO_PRESENT
-	if (rte_eal_vfio_setup() < 0) {
-		rte_eal_init_alert("Cannot init VFIO");
-		rte_errno = EAGAIN;
-		rte_atomic32_clear(&run_once);
-		return -1;
-	}
-#endif
 	/* in secondary processes, memory init may allocate additional fbarrays
 	 * not present in primary processes, so to avoid any potential issues,
 	 * initialize memzones first.
Signature · 1 line
2.17.1
#16 lore raw

v3 cover Add mdev (Mediated device) support in DPDK

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci). In this patchset, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

The first four patches in this patchset are mainly preparation of mdev bus support. The left two patches are the key implementation of mdev bus.

The implementation of mdev bus in DPDK has several options:

1: Embed mdev bus in current pci bus

This patchset takes this option for an example. Mdev has several device types: pci/platform/amba/ccw/ap. DPDK currently only cares pci devices in all mdev device types so we could embed the mdev bus into current pci bus. Then pci bus with mdev support will scan/plug/ unplug/.. not only normal pci devices but also mediated pci devices.

2: A new mdev bus that scans mediated pci devices and probes mdev driver to
plug-in pci devices to pci bus

If we took this option, a new mdev bus will be implemented to scan mediated pci devices and a new mdev driver for pci devices will be implemented in pci bus to plug-in mediated pci devices to pci bus.

Our RFC v1 takes this option:
http://patchwork.dpdk.org/project/dpdk/cover/[email protected]/

Note that: for either option 1 or 2, device drivers do not know the implementation difference but only use structs/functions exposed by pci bus. Mediated pci devices are different from normal pci devices
on: 1. Mediated pci devices use UUID as address but normal ones use BDF.
2. Mediated pci devices may have some capabilities that normal pci devices do not have. For example, mediated pci devices could have regions that have sparse mmap capability, which allows a region to have multiple mmap areas. Another example is mediated pci devices may have regions/part of regions not mmaped but need to access them. Above difference will change the current ABI (i.e., struct rte_pci_device). Please check 5th and 6th patch for details.

3. A brand new mdev bus that does everything

This option will implement a new and standalone mdev bus. This option does not need any changes in current pci bus but only needs some shared code (linux vfio part) in pci bus. Drivers of devices that support mdev will register itself as a mdev driver and do not rely on pci bus anymore. This option, IMHO, will make the code clean. The only potential problem may be code duplication, which could be solved by making code of linux vfio part of pci bus common and shared.

Your comments on above three options are welcomed and appreciated!

Thanks!
Chenbo

----------------------------------------------------------------------------
RFC v3:
- Add sparse mmap support
- Minor fixes and improvements

RFC v2:
- Let PCI bus scan mediated PCI devices directly
- Address Keith's comments
- Merge below patch into this series (David)
http://patches.dpdk.org/patch/55927/
- Add internal representation of PCI device (David)
- Minor fixes and improvements

[1] https://github.com/torvalds/linux/blob/master/Documentation/driver-api/vfio-mediated-device.rst

Chenbo Xia (1):
bus/pci: add sparse mmap support for mediated PCI devices

Tiwei Bie (5):
bus/pci: introduce an internal representation of PCI device bus/pci: avoid depending on private value in kernel source bus/pci: introduce helper for MMIO read and write
eal: add a helper for reading string from sysfs bus/pci: add mdev support

 drivers/bus/pci/bsd/pci.c             |  36 +-
 drivers/bus/pci/linux/pci.c           | 107 ++++-
 drivers/bus/pci/linux/pci_init.h      |  29 +-
 drivers/bus/pci/linux/pci_uio.c       |  22 +
 drivers/bus/pci/linux/pci_vfio.c      | 586 ++++++++++++++++++++++----
 drivers/bus/pci/linux/pci_vfio_mdev.c | 277 ++++++++++++
 drivers/bus/pci/meson.build           |   1 +
 drivers/bus/pci/pci_common.c          |  86 ++--
 drivers/bus/pci/pci_params.c          |  36 +-
 drivers/bus/pci/private.h             |  40 ++
 drivers/bus/pci/rte_bus_pci.h         |  83 +++-
 drivers/bus/pci/version.map           |   4 +
 lib/eal/common/eal_filesystem.h       |  10 +
 lib/eal/freebsd/eal.c                 |  22 +
 lib/eal/linux/eal.c                   |  39 +-
 lib/eal/version.map                   |   3 +
 16 files changed, 1224 insertions(+), 157 deletions(-)
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
Signature · 1 line
2.17.1
#17 lore raw

RFC v3 1/6 bus/pci: introduce an internal representation of PCI device

From: Tiwei Bie <[email protected]>

This patch introduces an internal representation of the PCI device which will be used to store the internal information that don't have to be exposed, e.g. the VFIO region sizes/offsets.

In this patch, the internal structure is simply a wrapper of the rte_pci_device structure. More fields will be added in the coming patches.

---
 drivers/bus/pci/bsd/pci.c    | 14 +++++++++-----
 drivers/bus/pci/linux/pci.c  | 27 ++++++++++++++++-----------
 drivers/bus/pci/pci_common.c |  2 +-
 drivers/bus/pci/private.h    | 12 ++++++++++++
 4 files changed, 38 insertions(+), 17 deletions(-)
drivers/bus/pci/bsd/pci.c +9 -5
index 4b8a208781..20ce979f60 100644
@@ -212,16 +212,20 @@ pci_uio_map_resource_by_index(struct rte_pci_device *dev, int res_idx,
 static int
 pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 {
+	struct rte_pci_device_internal *pdev;
 	struct rte_pci_device *dev;
 	struct pci_bar_io bar;
 	unsigned i, max;
 
-	dev = malloc(sizeof(*dev));
-	if (dev == NULL) {
+	pdev = malloc(sizeof(*pdev));
+	if (pdev == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot allocate memory for internal pci device\n");
 		return -1;
 	}
 
-	memset(dev, 0, sizeof(*dev));
+	memset(pdev, 0, sizeof(*pdev));
+
+	dev = &pdev->device;
 	dev->device.bus = &rte_pci_bus.bus;
 
 	dev->addr.domain = conf->pc_sel.pc_domain;
@@ -307,7 +311,7 @@ pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 				memmove(dev2->mem_resource,
 					dev->mem_resource,
 					sizeof(dev->mem_resource));
-				free(dev);
+				free(pdev);
 			}
 			return 0;
 		}
@@ -317,7 +321,7 @@ pci_scan_one(int dev_pci_fd, struct pci_conf *conf)
 	return 0;
 
 skipdev:
-	free(dev);
+	free(pdev);
 	return 0;
 }
 
drivers/bus/pci/linux/pci.c +16 -11
index 0dc99e9cb2..6dbba10657 100644
@@ -218,22 +218,27 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 {
 	char filename[PATH_MAX];
 	unsigned long tmp;
+	struct rte_pci_device_internal *pdev;
 	struct rte_pci_device *dev;
 	char driver[PATH_MAX];
 	int ret;
 
-	dev = malloc(sizeof(*dev));
-	if (dev == NULL)
+	pdev = malloc(sizeof(*pdev));
+	if (pdev == NULL) {
+		RTE_LOG(ERR, EAL, "Cannot allocate memory for internal pci device\n");
 		return -1;
+	}
+
+	memset(pdev, 0, sizeof(*pdev));
 
-	memset(dev, 0, sizeof(*dev));
+	dev = &pdev->device;
 	dev->device.bus = &rte_pci_bus.bus;
 	dev->addr = *addr;
 
 	/* get vendor id */
 	snprintf(filename, sizeof(filename), "%s/vendor", dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.vendor_id = (uint16_t)tmp;
@@ -241,7 +246,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	/* get device id */
 	snprintf(filename, sizeof(filename), "%s/device", dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.device_id = (uint16_t)tmp;
@@ -250,7 +255,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/subsystem_vendor",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.subsystem_vendor_id = (uint16_t)tmp;
@@ -259,7 +264,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/subsystem_device",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	dev->id.subsystem_device_id = (uint16_t)tmp;
@@ -268,7 +273,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/class",
 		 dirname);
 	if (eal_parse_sysfs_value(filename, &tmp) < 0) {
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 	/* the least 24 bits are valid: class, subclass, program interface */
@@ -308,7 +313,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	snprintf(filename, sizeof(filename), "%s/resource", dirname);
 	if (pci_parse_sysfs_resource(filename, dev) < 0) {
 		RTE_LOG(ERR, EAL, "%s(): cannot parse resource\n", __func__);
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 
@@ -317,7 +322,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 	ret = pci_get_kernel_driver_by_path(filename, driver, sizeof(driver));
 	if (ret < 0) {
 		RTE_LOG(ERR, EAL, "Fail to get kernel driver\n");
-		free(dev);
+		free(pdev);
 		return -1;
 	}
 
@@ -386,7 +391,7 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 						pci_name_set(dev2);
 					}
 				}
-				free(dev);
+				free(pdev);
 			}
 			return 0;
 		}
drivers/bus/pci/pci_common.c +1 -1
index ee7f966358..1c368c254c 100644
@@ -571,7 +571,7 @@ pci_unplug(struct rte_device *dev)
 	if (ret == 0) {
 		rte_pci_remove_device(pdev);
 		rte_devargs_remove(dev->devargs);
-		free(pdev);
+		free(RTE_PCI_DEVICE_INTERNAL(pdev));
 	}
 	return ret;
 }
drivers/bus/pci/private.h +12 -0
index 4cd9d14ec7..49a29d45cf 100644
@@ -12,11 +12,23 @@
 #include <rte_os_shim.h>
 #include <rte_pci.h>
 
+/*
+ * Convert struct rte_pci_device to struct rte_pci_device_internal
+ */
+#define RTE_PCI_DEVICE_INTERNAL(ptr) \
+	container_of(ptr, struct rte_pci_device_internal, device)
+#define RTE_PCI_DEVICE_INTERNAL_CONST(ptr) \
+	container_of(ptr, const struct rte_pci_device_internal, device)
+
 extern struct rte_pci_bus rte_pci_bus;
 
 struct rte_pci_driver;
 struct rte_pci_device;
 
+struct rte_pci_device_internal {
+	struct rte_pci_device device;
+};
+
 /**
  * Scan the content of the PCI bus, and the devices in the devices
  * list
Signature · 1 line
2.17.1
#18 lore raw

RFC v3 2/6 bus/pci: avoid depending on private value in kernel source

In reply to Chenbo Xia

From: Tiwei Bie <[email protected]>

The value 40 used in VFIO_GET_REGION_ADDR() is a private value (VFIO_PCI_OFFSET_SHIFT) defined in Linux kernel source [1]. It is not part of VFIO API, and we should not depend on it.

[1] https://github.com/torvalds/linux/blob/v5.12/drivers/vfio/pci/vfio_pci_private.h

---
 drivers/bus/pci/linux/pci.c      |   4 +-
 drivers/bus/pci/linux/pci_init.h |   4 +-
 drivers/bus/pci/linux/pci_vfio.c | 176 ++++++++++++++++++++++++-------
 drivers/bus/pci/private.h        |   9 ++
 4 files changed, 153 insertions(+), 40 deletions(-)
drivers/bus/pci/linux/pci.c +2 -2
index 6dbba10657..8f1fddbf20 100644
@@ -647,7 +647,7 @@ int rte_pci_read_config(const struct rte_pci_device *device,
 		return pci_uio_read_config(intr_handle, buf, len, offset);
 #ifdef VFIO_PRESENT
 	case RTE_PCI_KDRV_VFIO:
-		return pci_vfio_read_config(intr_handle, buf, len, offset);
+		return pci_vfio_read_config(device, buf, len, offset);
 #endif
 	default:
 		rte_pci_device_name(&device->addr, devname,
@@ -671,7 +671,7 @@ int rte_pci_write_config(const struct rte_pci_device *device,
 		return pci_uio_write_config(intr_handle, buf, len, offset);
 #ifdef VFIO_PRESENT
 	case RTE_PCI_KDRV_VFIO:
-		return pci_vfio_write_config(intr_handle, buf, len, offset);
+		return pci_vfio_write_config(device, buf, len, offset);
 #endif
 	default:
 		rte_pci_device_name(&device->addr, devname,
drivers/bus/pci/linux/pci_init.h +2 -2
index dcea726186..9f6659ba6e 100644
@@ -66,9 +66,9 @@ int pci_uio_ioport_unmap(struct rte_pci_ioport *p);
 #endif
 
 /* access config space */
-int pci_vfio_read_config(const struct rte_intr_handle *intr_handle,
+int pci_vfio_read_config(const struct rte_pci_device *dev,
 			 void *buf, size_t len, off_t offs);
-int pci_vfio_write_config(const struct rte_intr_handle *intr_handle,
+int pci_vfio_write_config(const struct rte_pci_device *dev,
 			  const void *buf, size_t len, off_t offs);
 
 int pci_vfio_ioport_map(struct rte_pci_device *dev, int bar,
drivers/bus/pci/linux/pci_vfio.c +140 -36

317 lines not on this page · raw message

drivers/bus/pci/private.h +9 -0
index 49a29d45cf..8b5fa70641 100644
@@ -12,6 +12,8 @@
 #include <rte_os_shim.h>
 #include <rte_pci.h>
 
+#define RTE_MAX_PCI_REGIONS	9
+
 /*
  * Convert struct rte_pci_device to struct rte_pci_device_internal
  */
@@ -25,8 +27,15 @@ extern struct rte_pci_bus rte_pci_bus;
 struct rte_pci_driver;
 struct rte_pci_device;
 
+struct rte_pci_region {
+	uint64_t size;
+	uint64_t offset;
+};
+
 struct rte_pci_device_internal {
 	struct rte_pci_device device;
+	/* PCI regions provided by e.g. VFIO. */
+	struct rte_pci_region region[RTE_MAX_PCI_REGIONS];
 };
 
 /**
Signature · 1 line
2.17.1
#19 lore raw

RFC v3 3/6 bus/pci: introduce helper for MMIO read and write

In reply to Chenbo Xia

From: Tiwei Bie <[email protected]>

The MMIO regions may not be mmap-able for mediated PCI device. In this case, the application should explicitly do read and write to access these regions.

---
 drivers/bus/pci/bsd/pci.c        | 22 +++++++++++++++
 drivers/bus/pci/linux/pci.c      | 46 ++++++++++++++++++++++++++++++
 drivers/bus/pci/linux/pci_init.h | 10 +++++++
 drivers/bus/pci/linux/pci_uio.c  | 22 +++++++++++++++
 drivers/bus/pci/linux/pci_vfio.c | 36 ++++++++++++++++++++++++
 drivers/bus/pci/rte_bus_pci.h    | 48 ++++++++++++++++++++++++++++++++
 drivers/bus/pci/version.map      |  4 +++
 7 files changed, 188 insertions(+)
drivers/bus/pci/bsd/pci.c +22 -0
index 20ce979f60..781f65c637 100644
@@ -494,6 +494,28 @@ int rte_pci_write_config(const struct rte_pci_device *dev,
 	return -1;
 }
 
+/* Read PCI MMIO space. */
+int rte_pci_mmio_read(const struct rte_pci_device *dev, int bar,
+		      void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy(buf, (uint8_t *)dev->mem_resource[bar].addr + offset, len);
+	return len;
+}
+
+/* Write PCI MMIO space. */
+int rte_pci_mmio_write(const struct rte_pci_device *dev, int bar,
+		       const void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy((uint8_t *)dev->mem_resource[bar].addr + offset, buf, len);
+	return len;
+}
+
 int
 rte_pci_ioport_map(struct rte_pci_device *dev, int bar,
 		struct rte_pci_ioport *p)
drivers/bus/pci/linux/pci.c +46 -0

52 lines not on this page · raw message

drivers/bus/pci/linux/pci_init.h +10 -0
index 9f6659ba6e..6853fa88a3 100644
@@ -37,6 +37,11 @@ int pci_uio_read_config(const struct rte_intr_handle *intr_handle,
 int pci_uio_write_config(const struct rte_intr_handle *intr_handle,
 			 const void *buf, size_t len, off_t offs);
 
+int pci_uio_mmio_read(const struct rte_pci_device *dev, int bar,
+		      void *buf, size_t len, off_t offset);
+int pci_uio_mmio_write(const struct rte_pci_device *dev, int bar,
+		       const void *buf, size_t len, off_t offset);
+
 int pci_uio_ioport_map(struct rte_pci_device *dev, int bar,
 		       struct rte_pci_ioport *p);
 void pci_uio_ioport_read(struct rte_pci_ioport *p,
@@ -71,6 +76,11 @@ int pci_vfio_read_config(const struct rte_pci_device *dev,
 int pci_vfio_write_config(const struct rte_pci_device *dev,
 			  const void *buf, size_t len, off_t offs);
 
+int pci_vfio_mmio_read(const struct rte_pci_device *dev, int bar,
+		       void *buf, size_t len, off_t offset);
+int pci_vfio_mmio_write(const struct rte_pci_device *dev, int bar,
+			const void *buf, size_t len, off_t offset);
+
 int pci_vfio_ioport_map(struct rte_pci_device *dev, int bar,
 		        struct rte_pci_ioport *p);
 void pci_vfio_ioport_read(struct rte_pci_ioport *p,
drivers/bus/pci/linux/pci_uio.c +22 -0
index 39ebeac2a0..2482635058 100644
@@ -45,6 +45,28 @@ pci_uio_write_config(const struct rte_intr_handle *intr_handle,
 	return pwrite(intr_handle->uio_cfg_fd, buf, len, offset);
 }
 
+int
+pci_uio_mmio_read(const struct rte_pci_device *dev, int bar,
+		  void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy(buf, (uint8_t *)dev->mem_resource[bar].addr + offset, len);
+	return len;
+}
+
+int
+pci_uio_mmio_write(const struct rte_pci_device *dev, int bar,
+		   const void *buf, size_t len, off_t offset)
+{
+	if (bar >= PCI_MAX_RESOURCE || dev->mem_resource[bar].addr == NULL ||
+			(uint64_t)offset + len > dev->mem_resource[bar].len)
+		return -1;
+	memcpy((uint8_t *)dev->mem_resource[bar].addr + offset, buf, len);
+	return len;
+}
+
 static int
 pci_uio_set_bus_master(int dev_fd)
 {
drivers/bus/pci/linux/pci_vfio.c +36 -0
index 012e7f72c1..3ecd984215 100644
@@ -1212,6 +1212,42 @@ pci_vfio_ioport_unmap(struct rte_pci_ioport *p)
 	return -1;
 }
 
+int
+pci_vfio_mmio_read(const struct rte_pci_device *dev, int bar,
+		   void *buf, size_t len, off_t offs)
+{
+	uint64_t size, offset;
+	int fd;
+
+	fd = dev->intr_handle.vfio_dev_fd;
+
+	if (pci_vfio_get_region(dev, bar, &size, &offset) != 0)
+		return -1;
+
+	if ((uint64_t)len + offs > size)
+		return -1;
+
+	return pread64(fd, buf, len, offset + offs);
+}
+
+int
+pci_vfio_mmio_write(const struct rte_pci_device *dev, int bar,
+		    const void *buf, size_t len, off_t offs)
+{
+	uint64_t size, offset;
+	int fd;
+
+	fd = dev->intr_handle.vfio_dev_fd;
+
+	if (pci_vfio_get_region(dev, bar, &size, &offset) != 0)
+		return -1;
+
+	if ((uint64_t)len + offs > size)
+		return -1;
+
+	return pwrite64(fd, buf, len, offset + offs);
+}
+
 int
 pci_vfio_is_enabled(void)
 {
drivers/bus/pci/rte_bus_pci.h +48 -0

54 lines not on this page · raw message

drivers/bus/pci/version.map +4 -0
index f33ed0abd1..02e4219aab 100644
@@ -21,4 +21,8 @@ EXPERIMENTAL {
 	global:
 
 	rte_pci_find_ext_capability;
+
+	# added in 21.08
+	rte_pci_mmio_read;
+	rte_pci_mmio_write;
 };
Signature · 1 line
2.17.1
#20 lore raw

RFC v3 4/6 eal: add a helper for reading string from sysfs

In reply to Chenbo Xia

From: Tiwei Bie <[email protected]>

This patch adds a helper for reading string from sysfs.

---
 lib/eal/common/eal_filesystem.h | 10 ++++++++++
 lib/eal/freebsd/eal.c           | 22 ++++++++++++++++++++++
 lib/eal/linux/eal.c             | 22 ++++++++++++++++++++++
 lib/eal/version.map             |  3 +++
 4 files changed, 57 insertions(+)
lib/eal/common/eal_filesystem.h +10 -0
index 5d21f07c20..be4c51ebb2 100644
@@ -104,4 +104,14 @@ eal_get_hugefile_path(char *buffer, size_t buflen, const char *hugedir, int f_id
  * Used to read information from files on /sys */
 int eal_parse_sysfs_value(const char *filename, unsigned long *val);
 
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice
+ *
+ * Function to read a line from a file on the filesystem.
+ * Used to read information from files on /sys
+ */
+__rte_experimental
+int rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz);
+
 #endif /* EAL_FILESYSTEM_H */
lib/eal/freebsd/eal.c +22 -0
index f4d1676754..002f07f4da 100644
@@ -169,6 +169,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/eal/linux/eal.c +22 -0
index ba19fc6347..d5917a48ca 100644
@@ -260,6 +260,28 @@ eal_parse_sysfs_value(const char *filename, unsigned long *val)
 	return 0;
 }
 
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}
 
 /* create memory configuration in shared/mmap memory. Take out
  * a write lock on the memsegs, so we can auto-detect primary/secondary.
lib/eal/version.map +3 -0
index fe5c3dac98..3d7fce26a4 100644
@@ -423,6 +423,9 @@ EXPERIMENTAL {
 	rte_version_release; # WINDOWS_NO_EXPORT
 	rte_version_suffix; # WINDOWS_NO_EXPORT
 	rte_version_year; # WINDOWS_NO_EXPORT
+
+	# added in 21.08
+	rte_eal_parse_sysfs_str; # WINDOWS_NO_EXPORT
 };
 
 INTERNAL {
Signature · 1 line
2.17.1
#21 lore raw

Re: [dpdk-dev] [RFC v3 4/6] eal: add a helper for reading string from sysfs

Quoted 12 lines · On Tue, 1 Jun 2021 11:06:42 +0800 Chenbo Xia <[email protected]> wrote:
lib/eal/freebsd/eal.c +12 -0
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+

lib/eal/freebsd/eal.c eal_parse_sysfs_value(const char *filename, unsigned long *val) · line 193

Quoted 10 lines
lib/eal/freebsd/eal.c +10 -0
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}

It would be helpful if function removed trailing newline.

	strchrnul(buf, '\n') = '\0';
#22 lore raw

Re: [dpdk-dev] [RFC v3 4/6] eal: add a helper for reading string from sysfs

Hi Stephen,

Quoted 39 lines

-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Tuesday, June 1, 2021 1:38 PM
To: Xia, Chenbo <[email protected]>
Cc: [email protected]; [email protected]; Liang, Cunming <[email protected]>; Wu, Jingjing <[email protected]>; Burakov, Anatoly <[email protected]>; Yigit, Ferruh <[email protected]>; [email protected]; [email protected]; Richardson, Bruce <[email protected]>; [email protected]; Ananyev, Konstantin <[email protected]>; Tiwei Bie <[email protected]>
Subject: Re: [RFC v3 4/6] eal: add a helper for reading string from sysfs

Quoted 12 lines · On Tue, 1 Jun 2021 11:06:42 +0800 Chenbo Xia <[email protected]> wrote:
lib/eal/freebsd/eal.c +12 -0
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);
+		return -1;
+	}
+

lib/eal/freebsd/eal.c eal_parse_sysfs_value(const char *filename, unsigned long *val) · line 193

Quoted 10 lines
lib/eal/freebsd/eal.c +10 -0
+	if (fgets(buf, sz, f) == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot read sysfs file %s\n",
+			__func__, filename);
+		fclose(f);
+		return -1;
+	}
+
+	fclose(f);
+	return 0;
+}

It would be helpful if function removed trailing newline.

Quoted 1 line
	strchrnul(buf, '\n') = '\0';

Make sense. Will fix it.

Thanks,
Chenbo

#23 lore raw

Re: [dpdk-dev] [RFC v3 4/6] eal: add a helper for reading string from sysfs

In reply to Chenbo Xia

lib/eal/freebsd/eal.c eal_parse_sysfs_value(const char *filename, unsigned long *val) · line 180

Quoted 10 lines · On Tue, 1 Jun 2021 11:06:42 +0800 Chenbo Xia <[email protected]> wrote:
lib/eal/freebsd/eal.c +9 -0
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);

Helpful to decode errno.

		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s:%s\n",
			__func__, filename, strerror(errno));
#24 lore raw

Re: [dpdk-dev] [RFC v3 4/6] eal: add a helper for reading string from sysfs

Hi Stephen,

Quoted 27 lines

-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Tuesday, June 1, 2021 1:39 PM
To: Xia, Chenbo <[email protected]>
Cc: [email protected]; [email protected]; Liang, Cunming <[email protected]>; Wu, Jingjing <[email protected]>; Burakov, Anatoly <[email protected]>; Yigit, Ferruh <[email protected]>; [email protected]; [email protected]; Richardson, Bruce <[email protected]>; [email protected]; Ananyev, Konstantin <[email protected]>; Tiwei Bie <[email protected]>
Subject: Re: [RFC v3 4/6] eal: add a helper for reading string from sysfs

lib/eal/freebsd/eal.c eal_parse_sysfs_value(const char *filename, unsigned long *val) · line 180

Quoted 10 lines · On Tue, 1 Jun 2021 11:06:42 +0800 Chenbo Xia <[email protected]> wrote:
lib/eal/freebsd/eal.c +9 -0
+int
+rte_eal_parse_sysfs_str(const char *filename, char *buf, unsigned long sz)
+{
+	FILE *f;
+
+	f = fopen(filename, "r");
+	if (f == NULL) {
+		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s\n",
+			__func__, filename);

Helpful to decode errno.

Quoted 2 lines
		RTE_LOG(ERR, EAL, "%s(): cannot open sysfs file %s:%s\n",
			__func__, filename, strerror(errno));

Yes. Will fix.

Thanks,
Chenbo

#25 lore raw

Re: [dpdk-dev] [RFC v3 4/6] eal: add a helper for reading string from sysfs

In reply to Chenbo Xia

01/06/2021 05:06, Chenbo Xia:

Quoted 4 lines

From: Tiwei Bie <[email protected]>

This patch adds a helper for reading string from sysfs.

Quoted 8 lines
---
 lib/eal/common/eal_filesystem.h | 10 ++++++++++
 lib/eal/freebsd/eal.c           | 22 ++++++++++++++++++++++
 lib/eal/linux/eal.c             | 22 ++++++++++++++++++++++
 lib/eal/version.map             |  3 +++
 4 files changed, 57 insertions(+)

3 separate comments:

1/ How much code is portable between Linux and FreeBSD? I guess the path will be different?

2/ Please think about Windows stub.

3/ Instead of EAL, we should start lib/sysfs/ I have other ideas of sysfs functions for PMD use.

#26 lore raw

RFC v3 5/6 bus/pci: add mdev support

In reply to Chenbo Xia

From: Tiwei Bie <[email protected]>

This patch adds the mdev (Mediated device) support in PCI bus driver. With this patch, the PCI bus driver will be able to scan and probe the mediated PCI devices (i.e. the Mediated devices whose device API is "vfio-pci") in the system.

There are several things different between physical PCI devices and mediated PCI devices:

- Mediated PCI devices have to be accessed through VFIO API;
- The regions in mediated PCI devices may not be mmap-able, and drivers need to call read/write function to access them in this case;
- Mediated PCI devices use UUID as device address;

---
 drivers/bus/pci/linux/pci.c           |  30 ++-
 drivers/bus/pci/linux/pci_init.h      |  15 +-
 drivers/bus/pci/linux/pci_vfio.c      | 147 ++++++++++++--
 drivers/bus/pci/linux/pci_vfio_mdev.c | 277 ++++++++++++++++++++++++++
 drivers/bus/pci/meson.build           |   1 +
 drivers/bus/pci/pci_common.c          |  84 +++++---
 drivers/bus/pci/pci_params.c          |  36 +++-
 drivers/bus/pci/private.h             |  17 ++
 drivers/bus/pci/rte_bus_pci.h         |  17 +-
 lib/eal/linux/eal.c                   |  17 +-
 10 files changed, 571 insertions(+), 70 deletions(-)
 create mode 100644 drivers/bus/pci/linux/pci_vfio_mdev.c
drivers/bus/pci/linux/pci.c +22 -8
index 4805f277c5..29dd9ba26f 100644
@@ -30,7 +30,7 @@
 
 extern struct rte_pci_bus rte_pci_bus;
 
-static int
+int
 pci_get_kernel_driver_by_path(const char *filename, char *dri_name,
 			      size_t len)
 {
@@ -70,7 +70,7 @@ rte_pci_map_device(struct rte_pci_device *dev)
 	switch (dev->kdrv) {
 	case RTE_PCI_KDRV_VFIO:
 #ifdef VFIO_PRESENT
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			ret = pci_vfio_map_resource(dev);
 #endif
 		break;
@@ -99,7 +99,7 @@ rte_pci_unmap_device(struct rte_pci_device *dev)
 	switch (dev->kdrv) {
 	case RTE_PCI_KDRV_VFIO:
 #ifdef VFIO_PRESENT
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			pci_vfio_unmap_resource(dev);
 #endif
 		break;
@@ -347,6 +347,15 @@ pci_scan_one(const char *dirname, const struct rte_pci_addr *addr)
 		int ret;
 
 		TAILQ_FOREACH(dev2, &rte_pci_bus.device_list, next) {
+			/*
+			 * Insert physical PCI devices before all mediated
+			 * PCI devices.
+			 */
+			if (dev2->is_mdev) {
+				rte_pci_insert_device(dev2, dev);
+				return 0;
+			}
+
 			ret = rte_pci_addr_cmp(&dev->addr, &dev2->addr);
 			if (ret > 0)
 				continue;
@@ -465,8 +474,14 @@ rte_pci_scan(void)
 		return 0;
 
 #ifdef VFIO_PRESENT
-	if (!pci_vfio_is_enabled())
-		RTE_LOG(DEBUG, EAL, "VFIO PCI modules not loaded\n");
+	if (!rte_vfio_is_enabled("vfio_pci"))
+		RTE_LOG(DEBUG, EAL, "VFIO PCI module not loaded\n");
+
+	if (!rte_vfio_is_enabled("vfio_mdev"))
+		RTE_LOG(DEBUG, EAL, "VFIO MDEV module not loaded\n");
+
+	if (pci_scan_mdev() != 0)
+		return -1;
 #endif
 
 	dir = opendir(rte_pci_get_sysfs_path());
@@ -737,7 +752,7 @@ rte_pci_ioport_map(struct rte_pci_device *dev, int bar,
 	switch (dev->kdrv) {
 #ifdef VFIO_PRESENT
 	case RTE_PCI_KDRV_VFIO:
-		if (pci_vfio_is_enabled())
+		if (pci_vfio_is_enabled(dev))
 			ret = pci_vfio_ioport_map(dev, bar, p);
 		break;
 #endif
@@ -801,8 +816,7 @@ rte_pci_ioport_unmap(struct rte_pci_ioport *p)
 	switch (p->dev->kdrv) {
 #ifdef VFIO_PRESENT
 	case RTE_PCI_KDRV_VFIO:
-		if (pci_vfio_is_enabled())
-			ret = pci_vfio_ioport_unmap(p);
+		ret = -1;
 		break;
 #endif
 	case RTE_PCI_KDRV_IGB_UIO:
drivers/bus/pci/linux/pci_init.h +14 -1
index 6853fa88a3..0c0191b6d5 100644
@@ -19,6 +19,9 @@
 extern void *pci_map_addr;
 void *pci_find_max_end_va(void);
 
+int pci_get_kernel_driver_by_path(const char *filename, char *dri_name,
+				  size_t len);
+
 /* parse one line of the "resource" sysfs file (note that the 'line'
  * string is modified)
  */
@@ -93,7 +96,17 @@ int pci_vfio_ioport_unmap(struct rte_pci_ioport *p);
 int pci_vfio_map_resource(struct rte_pci_device *dev);
 int pci_vfio_unmap_resource(struct rte_pci_device *dev);
 
-int pci_vfio_is_enabled(void);
+int pci_vfio_is_enabled(struct rte_pci_device *dev);
+
+int pci_vfio_fill_regions(struct rte_pci_device *dev, int vfio_dev_fd,
+			  struct vfio_device_info *device_info);
+
+int pci_vfio_get_pci_id(struct rte_pci_device *dev, int vfio_dev_fd,
+			struct rte_pci_id *pci_id);
+
+const char *pci_mdev_get_sysfs_path(void);
+
+int pci_scan_mdev(void);
 
 #endif
 
drivers/bus/pci/linux/pci_vfio.c +126 -21

248 lines not on this page · raw message

drivers/bus/pci/linux/pci_vfio_mdev.c new file +277 -0

277 lines not on this page · raw message

drivers/bus/pci/meson.build +1 -0
index 81c7e94c00..fb7a9a1fa8 100644
@@ -11,6 +11,7 @@ if is_linux
             'linux/pci.c',
             'linux/pci_uio.c',
             'linux/pci_vfio.c',
+            'linux/pci_vfio_mdev.c',
     )
     includes += include_directories('linux')
 endif
drivers/bus/pci/pci_common.c +59 -25

167 lines not on this page · raw message

drivers/bus/pci/pci_params.c +31 -5
index 3192e9c967..231e57213e 100644
@@ -2,12 +2,15 @@
  * Copyright 2018 Gaëtan Rivet
  */
 
+#include <string.h>
+
 #include <rte_bus.h>
 #include <rte_bus_pci.h>
 #include <rte_dev.h>
 #include <rte_errno.h>
 #include <rte_kvargs.h>
 #include <rte_pci.h>
+#include <rte_uuid.h>
 
 #include "private.h"
 
@@ -35,6 +38,19 @@ pci_addr_kv_cmp(const char *key __rte_unused,
 	return -abs(rte_pci_addr_cmp(addr1, addr2));
 }
 
+static int
+mdev_addr_kv_cmp(const char *key __rte_unused,
+		const char *value,
+		void *_addr2)
+{
+	rte_uuid_t addr1;
+	unsigned char *addr2 = _addr2;
+
+	if (rte_uuid_parse(value, addr1))
+		return -1;
+	return -abs(rte_uuid_compare(addr1, addr2));
+}
+
 static int
 pci_dev_match(const struct rte_device *dev,
 	      const void *_kvlist)
@@ -47,11 +63,21 @@ pci_dev_match(const struct rte_device *dev,
 		return 0;
 	pdev = RTE_DEV_TO_PCI_CONST(dev);
 	/* if any field does not match. */
-	if (rte_kvargs_process(kvlist, pci_params_keys[RTE_PCI_PARAM_ADDR],
-			       &pci_addr_kv_cmp,
-			       (void *)(intptr_t)&pdev->addr))
-		return 1;
-	return 0;
+	if (!pdev->is_mdev) {
+		if (rte_kvargs_process(kvlist,
+			pci_params_keys[RTE_PCI_PARAM_ADDR], &pci_addr_kv_cmp,
+			(void *)(intptr_t)&pdev->addr))
+			return 1;
+		else
+			return 0;
+	} else {
+		if (rte_kvargs_process(kvlist,
+			pci_params_keys[RTE_PCI_PARAM_ADDR], &mdev_addr_kv_cmp,
+			(void *)(intptr_t)&pdev->uuid))
+			return 1;
+		else
+			return 0;
+	}
 }
 
 void *
drivers/bus/pci/private.h +17 -0
index 8b5fa70641..3515c086aa 100644
@@ -64,6 +64,18 @@ pci_name_set(struct rte_pci_device *dev);
  */
 bool rte_pci_ignore_device(const struct rte_pci_addr *pci_addr);
 
+/**
+ * Validate whether a mediated PCI device with given uuid should be
+ * ignored or not.
+ *
+ * @param mdev_addr
+ *	MDEV address of device to be validated
+ * @return
+ *	true: if device is to be ignored,
+ *	false: if device is to be scanned,
+ */
+bool rte_mdev_ignore_device(const rte_uuid_t mdev_addr);
+
 /**
  * Add a PCI device to the PCI Bus (append to PCI Device list). This function
  * also updates the bus references of the PCI Device (and the generic device
@@ -114,6 +126,11 @@ struct pci_msix_table {
 struct mapped_pci_resource {
 	TAILQ_ENTRY(mapped_pci_resource) next;
 
+	union {
+		struct rte_pci_addr addr;
+		rte_uuid_t uuid;
+	};
+	uint8_t is_mdev;
 	struct rte_pci_addr pci_addr;
 	char path[PATH_MAX];
 	int nb_maps;
drivers/bus/pci/rte_bus_pci.h +15 -2
index dc26811b0a..fb7d934bd0 100644
@@ -51,6 +51,15 @@ TAILQ_HEAD(rte_pci_driver_list, rte_pci_driver);
 
 struct rte_devargs;
 
+/*
+ * NOTE: we can't include rte_uuid.h directly due to the conflicts
+ *      introduced by stdbool.h
+ */
+typedef unsigned char rte_uuid_t[16];
+
+/* It's RTE_UUID_STRLEN, which is bigger than PCI_PRI_STR_SIZE. */
+#define RTE_PCI_NAME_LEN		(36 + 1)
+
 enum rte_pci_kernel_driver {
 	RTE_PCI_KDRV_UNKNOWN = 0,  /* may be misc UIO or bifurcated driver */
 	RTE_PCI_KDRV_IGB_UIO,      /* igb_uio for Linux */
@@ -67,7 +76,11 @@ enum rte_pci_kernel_driver {
 struct rte_pci_device {
 	TAILQ_ENTRY(rte_pci_device) next;   /**< Next probed PCI device. */
 	struct rte_device device;           /**< Inherit core device */
-	struct rte_pci_addr addr;           /**< PCI location. */
+	union {
+		struct rte_pci_addr addr;   /**< PCI location. */
+		rte_uuid_t uuid;            /**< Mdev location. */
+	};
+	uint8_t is_mdev;                    /**< True for mediated PCI device */
 	struct rte_pci_id id;               /**< PCI ID. */
 	struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
 					    /**< PCI Memory Resource */
@@ -75,7 +88,7 @@ struct rte_pci_device {
 	struct rte_pci_driver *driver;      /**< PCI driver used in probing */
 	uint16_t max_vfs;                   /**< sriov enable if not zero */
 	enum rte_pci_kernel_driver kdrv;    /**< Kernel driver passthrough */
-	char name[PCI_PRI_STR_SIZE+1];      /**< PCI location (ASCII) */
+	char name[RTE_PCI_NAME_LEN];        /**< PCI/Mdev location (ASCII) */
 	struct rte_intr_handle vfio_req_intr_handle;
 				/**< Handler of VFIO request interrupt */
 };
lib/eal/linux/eal.c +9 -8
index d5917a48ca..323f13107e 100644
@@ -1089,6 +1089,15 @@ rte_eal_init(int argc, char **argv)
 		return -1;
 	}
 
+#ifdef VFIO_PRESENT
+	if (rte_eal_vfio_setup() < 0) {
+		rte_eal_init_alert("Cannot init VFIO");
+		rte_errno = EAGAIN;
+		__atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
+		return -1;
+	}
+#endif
+
 	if (rte_bus_scan()) {
 		rte_eal_init_alert("Cannot scan the buses for devices");
 		rte_errno = ENODEV;
@@ -1194,14 +1203,6 @@ rte_eal_init(int argc, char **argv)
 		return -1;
 	}
 
-#ifdef VFIO_PRESENT
-	if (rte_eal_vfio_setup() < 0) {
-		rte_eal_init_alert("Cannot init VFIO");
-		rte_errno = EAGAIN;
-		__atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
-		return -1;
-	}
-#endif
 	/* in secondary processes, memory init may allocate additional fbarrays
 	 * not present in primary processes, so to avoid any potential issues,
 	 * initialize memzones first.
Signature · 1 line
2.17.1
#27 lore raw

RFC v3 6/6 bus/pci: add sparse mmap support for mediated PCI devices

In reply to Chenbo Xia

This patch adds sparse mmap support in PCI bus. Sparse mmap is a capability defined in VFIO which allows multiple mmap areas in one VFIO region. Mediated pci devices could use this capability to let mdev parent driver have control over access of non-mmapable part of regions.

---
 drivers/bus/pci/linux/pci_vfio.c | 229 +++++++++++++++++++++++++++----
 drivers/bus/pci/private.h        |   2 +
 drivers/bus/pci/rte_bus_pci.h    |  18 ++-
 3 files changed, 218 insertions(+), 31 deletions(-)
drivers/bus/pci/linux/pci_vfio.c +199 -30
index 00ba5db03a..e68eccb63f 100644
@@ -654,6 +654,82 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
 	return 0;
 }
 
+static int
+pci_vfio_sparse_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
+		struct vfio_region_sparse_mmap_area *vfio_areas,
+		uint32_t nr_areas, int bar_index, int additional_flags,
+		int numa_node)
+{
+	struct pci_map *map = &vfio_res->maps[bar_index];
+	struct rte_mem_map_area *area;
+	struct vfio_region_sparse_mmap_area *sparse;
+	void *bar_addr;
+	uint32_t i, j;
+
+	map->nr_areas = nr_areas;
+
+	if (map->size == 0) {
+		RTE_LOG(DEBUG, EAL, "Bar size is 0, skip BAR%d\n", bar_index);
+		return 0;
+	}
+
+	if (!map->nr_areas) {
+		RTE_LOG(DEBUG, EAL, "Skip bar %d with no sparse mmap areas\n",
+			bar_index);
+		map->areas = NULL;
+		return 0;
+	}
+
+	if (map->areas == NULL) {
+		map->areas = rte_zmalloc_socket(NULL,
+				sizeof(*map->areas) * nr_areas,
+				RTE_CACHE_LINE_SIZE, numa_node);
+		if (map->areas == NULL) {
+			RTE_LOG(ERR, EAL,
+				"Cannot alloc memory for sparse map areas\n");
+			return -1;
+		}
+	}
+
+	for (i = 0; i < map->nr_areas; i++) {
+		area = &map->areas[i];
+		sparse = &vfio_areas[i];
+
+		bar_addr = mmap(map->addr, sparse->size, 0, MAP_PRIVATE |
+				MAP_ANONYMOUS | additional_flags, -1, 0);
+		if (bar_addr != MAP_FAILED) {
+			area->addr = pci_map_resource(bar_addr, vfio_dev_fd,
+				map->offset + sparse->offset, sparse->size,
+				RTE_MAP_FORCE_ADDRESS);
+			if (area->addr == NULL) {
+				munmap(bar_addr, sparse->size);
+				RTE_LOG(ERR, EAL, "Failed to map pci BAR%d\n",
+					bar_index);
+				goto err_map;
+			}
+
+			area->offset = sparse->offset;
+			area->size = sparse->size;
+		} else {
+			RTE_LOG(ERR, EAL, "Failed to create inaccessible mapping for BAR%d\n",
+				bar_index);
+			goto err_map;
+		}
+	}
+
+	return 0;
+
+err_map:
+	for (j = 0; j < i; j++) {
+		pci_unmap_resource(map->areas[j].addr, map->areas[j].size);
+		map->areas[j].offset = 0;
+		map->areas[j].size = 0;
+	}
+	rte_free(map->areas);
+	map->nr_areas = 0;
+	return -1;
+}
+
 /*
  * region info may contain capability headers, so we need to keep reallocating
  * the memory until we match allocated memory size with argsz.
@@ -770,6 +846,31 @@ pci_vfio_fill_regions(struct rte_pci_device *dev, int vfio_dev_fd,
 	return 0;
 }
 
+static void
+clean_up_pci_resource(struct mapped_pci_resource *vfio_res)
+{
+	struct pci_map *map;
+	uint32_t i, j;
+
+	for (i = 0; i < PCI_MAX_RESOURCE; i++) {
+		map = &vfio_res->maps[i];
+		if (map->nr_areas > 1) {
+			for (j = 0; j < map->nr_areas; j++)
+				pci_unmap_resource(map->areas[j].addr,
+					map->areas[j].size);
+		} else {
+			/*
+			 * We do not need to be aware of MSI-X BAR mappings.
+			 * Using current maps array is enough.
+			 */
+			if (map->addr)
+				pci_unmap_resource(map->addr, map->size);
+		}
+	}
+
+	rte_free(map->areas);
+}
+
 static int
 pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 {
@@ -866,6 +967,8 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 
 	for (i = 0; i < vfio_res->nb_maps; i++) {
 		void *bar_addr;
+		struct vfio_info_cap_header *hdr;
+		struct vfio_region_info_cap_sparse_mmap *sparse;
 
 		ret = pci_vfio_get_region_info(vfio_dev_fd, &reg, i);
 		if (ret < 0) {
@@ -911,15 +1014,59 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 		maps[i].size = reg->size;
 		maps[i].path = NULL; /* vfio doesn't have per-resource paths */
 
-		ret = pci_vfio_mmap_bar(vfio_dev_fd, vfio_res, i, 0);
-		if (ret < 0) {
-			RTE_LOG(ERR, EAL, "%s mapping BAR%i failed: %s\n",
-					pci_addr, i, strerror(errno));
-			free(reg);
-			goto err_vfio_res;
-		}
+		hdr = pci_vfio_info_cap(reg, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
+
+		if (dev->is_mdev && hdr != NULL) {
+			sparse = container_of(hdr,
+				struct vfio_region_info_cap_sparse_mmap,
+				header);
+
+			ret = pci_vfio_sparse_mmap_bar(vfio_dev_fd, vfio_res,
+				sparse->areas, sparse->nr_areas, i, 0,
+				dev->device.numa_node);
+			if (ret < 0) {
+				RTE_LOG(ERR, EAL, "%s sparse mapping BAR%i failed: %s\n",
+						pci_addr, i, strerror(errno));
+				free(reg);
+				goto err_vfio_res;
+			}
 
-		dev->mem_resource[i].addr = maps[i].addr;
+			dev->sparse_mem[i].size = reg->size;
+			dev->sparse_mem[i].nr_maps = vfio_res->maps[i].nr_areas;
+			dev->sparse_mem[i].areas = vfio_res->maps[i].areas;
+		} else {
+			ret = pci_vfio_mmap_bar(vfio_dev_fd, vfio_res, i, 0);
+			if (ret < 0) {
+				RTE_LOG(ERR, EAL, "%s mapping BAR%i failed: %s\n",
+						pci_addr, i, strerror(errno));
+				free(reg);
+				goto err_vfio_res;
+			}
+
+			if (dev->is_mdev) {
+				struct pci_map *mdev_map = &maps[i];
+				mdev_map->nr_areas = 1;
+				mdev_map->areas = rte_zmalloc_socket(NULL,
+					sizeof(*mdev_map->areas),
+					RTE_CACHE_LINE_SIZE,
+					dev->device.numa_node);
+				if (maps[i].areas == NULL) {
+					RTE_LOG(ERR, EAL,
+						"Cannot allocate memory for sparse map areas\n");
+					goto err_vfio_res;
+				}
+				mdev_map->areas[0].addr = maps[i].addr;
+				mdev_map->areas[0].offset = 0;
+				mdev_map->areas[0].size = reg->size;
+				dev->sparse_mem[i].size = reg->size;
+				dev->sparse_mem[i].nr_maps = 1;
+				dev->sparse_mem[i].areas = mdev_map->areas;
+			} else {
+				maps[i].nr_areas = 0;
+				maps[i].areas = NULL;
+				dev->mem_resource[i].addr = maps[i].addr;
+			}
+		}
 
 		free(reg);
 	}
@@ -940,6 +1087,7 @@ pci_vfio_map_resource_primary(struct rte_pci_device *dev)
 
 	return 0;
 err_vfio_res:
+	clean_up_pci_resource(vfio_res);
 	rte_free(vfio_res);
 err_vfio_dev_fd:
 	rte_vfio_release_device(rte_pci_get_sysfs_path(),
@@ -960,7 +1108,7 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 	struct mapped_pci_res_list *vfio_res_list =
 		RTE_TAILQ_CAST(rte_vfio_tailq.head, mapped_pci_res_list);
 
-	struct pci_map *maps;
+	struct pci_map *maps, *cur;
 
 	dev->intr_handle.fd = -1;
 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
@@ -1012,14 +1160,49 @@ pci_vfio_map_resource_secondary(struct rte_pci_device *dev)
 	maps = vfio_res->maps;
 
 	for (i = 0; i < vfio_res->nb_maps; i++) {
-		ret = pci_vfio_mmap_bar(vfio_dev_fd, vfio_res, i, MAP_FIXED);
-		if (ret < 0) {
-			RTE_LOG(ERR, EAL, "%s mapping BAR%i failed: %s\n",
-					pci_addr, i, strerror(errno));
-			goto err_vfio_dev_fd;
+		cur = &maps[i];
+		if (cur->nr_areas > 1) {
+			struct vfio_region_sparse_mmap_area *areas;
+			uint32_t i;
+
+			areas = malloc(sizeof(*areas) * cur->nr_areas);
+			if (areas == NULL) {
+				RTE_LOG(ERR, EAL, "Failed to alloc vfio areas for %s\n",
+					pci_addr);
+				goto err_vfio_dev_fd;
+			}
+
+			for (i = 0; i < cur->nr_areas; i++) {
+				areas[i].offset = cur->areas[i].offset;
+				areas[i].size = cur->areas[i].size;
+			}
+
+			ret = pci_vfio_sparse_mmap_bar(vfio_dev_fd, vfio_res,
+				areas, cur->nr_areas, i, MAP_FIXED,
+				dev->device.numa_node);
+			if (ret < 0) {
+				RTE_LOG(ERR, EAL, "%s sparse mapping BAR%i failed: %s\n",
+						pci_addr, i, strerror(errno));
+				free(areas);
+				goto err_vfio_dev_fd;
+			}
+
+			free(areas);
+		} else {
+			ret = pci_vfio_mmap_bar(vfio_dev_fd, vfio_res,
+				i, MAP_FIXED);
+			if (ret < 0) {
+				RTE_LOG(ERR, EAL, "%s mapping BAR%i failed: %s\n",
+						pci_addr, i, strerror(errno));
+				goto err_vfio_dev_fd;
+			}
+
+			if (dev->is_mdev)
+				cur->areas[0].addr = cur->addr;
+			else
+				dev->mem_resource[i].addr = cur->addr;
 		}
 
-		dev->mem_resource[i].addr = maps[i].addr;
 	}
 
 	/* we need save vfio_dev_fd, so it can be used during release */
@@ -1054,8 +1237,6 @@ find_and_unmap_vfio_resource(struct mapped_pci_res_list *vfio_res_list,
 			const char *pci_addr)
 {
 	struct mapped_pci_resource *vfio_res = NULL;
-	struct pci_map *maps;
-	int i;
 
 	/* Get vfio_res */
 	TAILQ_FOREACH(vfio_res, vfio_res_list, next) {
@@ -1079,19 +1260,7 @@ find_and_unmap_vfio_resource(struct mapped_pci_res_list *vfio_res_list,
 	RTE_LOG(INFO, EAL, "Releasing PCI mapped resource for %s\n",
 		pci_addr);
 
-	maps = vfio_res->maps;
-	for (i = 0; i < vfio_res->nb_maps; i++) {
-
-		/*
-		 * We do not need to be aware of MSI-X table BAR mappings as
-		 * when mapping. Just using current maps array is enough
-		 */
-		if (maps[i].addr) {
-			RTE_LOG(INFO, EAL, "Calling pci_unmap_resource for %s at %p\n",
-				pci_addr, maps[i].addr);
-			pci_unmap_resource(maps[i].addr, maps[i].size);
-		}
-	}
+	clean_up_pci_resource(vfio_res);
 
 	return vfio_res;
 }
drivers/bus/pci/private.h +2 -0
index 3515c086aa..8d94d8acf8 100644
@@ -110,6 +110,8 @@ struct pci_map {
 	uint64_t offset;
 	uint64_t size;
 	uint64_t phaddr;
+	uint32_t nr_areas;
+	struct rte_mem_map_area *areas;
 };
 
 struct pci_msix_table {
drivers/bus/pci/rte_bus_pci.h +17 -1
index fb7d934bd0..ddc913f121 100644
@@ -70,6 +70,18 @@ enum rte_pci_kernel_driver {
 	RTE_PCI_KDRV_NET_UIO,      /* NetUIO for Windows */
 };
 
+struct rte_mem_map_area {
+	void *addr;
+	uint64_t offset;
+	uint64_t size;
+};
+
+struct rte_sparse_mem_map {
+	uint64_t size;
+	uint32_t nr_maps;
+	struct rte_mem_map_area *areas;
+};
+
 /**
  * A structure describing a PCI device.
  */
@@ -82,8 +94,12 @@ struct rte_pci_device {
 	};
 	uint8_t is_mdev;                    /**< True for mediated PCI device */
 	struct rte_pci_id id;               /**< PCI ID. */
-	struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
+	union {
+		struct rte_mem_resource mem_resource[PCI_MAX_RESOURCE];
 					    /**< PCI Memory Resource */
+		struct rte_sparse_mem_map sparse_mem[PCI_MAX_RESOURCE];
+					    /**< Sparse Memory Map for Mdev */
+	};
 	struct rte_intr_handle intr_handle; /**< Interrupt handle */
 	struct rte_pci_driver *driver;      /**< PCI driver used in probing */
 	uint16_t max_vfs;                   /**< sriov enable if not zero */
Signature · 1 line
2.17.1
#28 lore raw
In reply to Chenbo Xia

01/06/2021 05:06, Chenbo Xia:

Quoted 6 lines

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci).

Please could you illustrate with an usage of mdev in DPDK? What does it enable which is not possible today?

Quoted 16 lines

In this patchset, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

The first four patches in this patchset are mainly preparation of mdev bus support. The left two patches are the key implementation of mdev bus.

The implementation of mdev bus in DPDK has several options:

Quoted 7 lines

1: Embed mdev bus in current pci bus

This patchset takes this option for an example. Mdev has several device types: pci/platform/amba/ccw/ap. DPDK currently only cares pci devices in all mdev device types so we could embed the mdev bus into current pci bus. Then pci bus with mdev support will scan/plug/ unplug/.. not only normal pci devices but also mediated pci devices.

I think it is a different bus.
It would be cleaner to not touch the PCI bus.
Having a separate bus will allow an easy way to identify a device with the new generic devargs syntax, example:
bus=mdev,uuid=XXX
or more complex:
bus=mdev,uuid=XXX/class=crypto/driver=qat,foo=bar

Quoted 24 lines

2: A new mdev bus that scans mediated pci devices and probes mdev driver to
plug-in pci devices to pci bus

If we took this option, a new mdev bus will be implemented to scan mediated pci devices and a new mdev driver for pci devices will be implemented in pci bus to plug-in mediated pci devices to pci bus.

Our RFC v1 takes this option:
http://patchwork.dpdk.org/project/dpdk/cover/[email protected]/

Note that: for either option 1 or 2, device drivers do not know the implementation difference but only use structs/functions exposed by pci bus. Mediated pci devices are different from normal pci devices
on: 1. Mediated pci devices use UUID as address but normal ones use BDF.
2. Mediated pci devices may have some capabilities that normal pci devices do not have. For example, mediated pci devices could have regions that have sparse mmap capability, which allows a region to have multiple mmap areas. Another example is mediated pci devices may have regions/part of regions not mmaped but need to access them. Above difference will change the current ABI (i.e., struct rte_pci_device). Please check 5th and 6th patch for details.

3. A brand new mdev bus that does everything

Quoted 7 lines

This option will implement a new and standalone mdev bus. This option does not need any changes in current pci bus but only needs some shared code (linux vfio part) in pci bus. Drivers of devices that support mdev will register itself as a mdev driver and do not rely on pci bus anymore. This option, IMHO, will make the code clean. The only potential problem may be code duplication, which could be solved by making code of linux vfio part of pci bus common and shared.

Yes I prefer this third option.
We can find an elegant way of sharing some VFIO code between buses.

#29 lore raw

Hi Thomas,

Quoted 15 lines

-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Friday, June 11, 2021 3:16 PM
To: Xia, Chenbo <[email protected]>
Cc: [email protected]; Liang, Cunming <[email protected]>; Wu, Jingjing <[email protected]>; Burakov, Anatoly <[email protected]>; Yigit, Ferruh <[email protected]>; [email protected]; [email protected]; Richardson, Bruce <[email protected]>; [email protected]; [email protected]; Ananyev, Konstantin <[email protected]>
Subject: Re: [dpdk-dev] [RFC v3 0/6] Add mdev (Mediated device) support in DPDK

01/06/2021 05:06, Chenbo Xia:

Quoted 2 lines

Hi everyone,

Quoted 7 lines
Quoted 4 lines

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci).

Please could you illustrate with an usage of mdev in DPDK? What does it enable which is not possible today?

The main purpose is for DPDK to drive mdev-based devices, which is not possible today.

I'd take PCI devices for an example. Currently DPDK can only drive devices of physical pci bus under /sys/bus/pci and kernel exposes the pci devices to APP in that way.

But there are PCI devices using vfio-mdev as a software framework to expose Mdev to APP under /sys/bus/mdev. Devices could choose this way of virtualizing itself to let multiple APPs share one physical device. For example, Intel Scalable IOV technology is known to use vfio-mdev as SW framework for Scalable IOV enabled devices (and Intel net/crypto/raw devices support this tech). For those mdev-based devices, DPDK needs support on the bus layer to scan/plug/probe/.. them, which is the main effort this patchset does. There are also other devices using the vfio-mdev framework, AFAIK, Nvidia's GPU is the first one using mdev and Intel's GPU virtualization also uses it.

Quoted 25 lines
Quoted 17 lines

In this patchset, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

The first four patches in this patchset are mainly preparation of mdev bus support. The left two patches are the key implementation of mdev bus.

The implementation of mdev bus in DPDK has several options:

Quoted 7 lines

1: Embed mdev bus in current pci bus

This patchset takes this option for an example. Mdev has several device types: pci/platform/amba/ccw/ap. DPDK currently only cares pci devices in all mdev device types so we could embed the mdev bus into current pci bus. Then pci bus with mdev support will scan/plug/ unplug/.. not only normal pci devices but also mediated pci devices.

Quoted 7 lines

I think it is a different bus.
It would be cleaner to not touch the PCI bus.
Having a separate bus will allow an easy way to identify a device with the new generic devargs syntax, example:
bus=mdev,uuid=XXX
or more complex:
bus=mdev,uuid=XXX/class=crypto/driver=qat,foo=bar

OK. Agree on cleaner to not touch PCI bus. And there may also be a 'type=pci' as mdev has several types in its definition (pci/ap/platform/ccw/...).

Quoted 34 lines
Quoted 10 lines

2: A new mdev bus that scans mediated pci devices and probes mdev driver to
plug-in pci devices to pci bus

If we took this option, a new mdev bus will be implemented to scan mediated pci devices and a new mdev driver for pci devices will be implemented in pci bus to plug-in mediated pci devices to pci bus.

Our RFC v1 takes this option:

   http://patchwork.dpdk.org/project/dpdk/cover/20190403071844.21126-1-

[email protected]/

Quoted 15 lines

Note that: for either option 1 or 2, device drivers do not know the implementation difference but only use structs/functions exposed by pci bus. Mediated pci devices are different from normal pci devices
on: 1. Mediated pci devices use UUID as address but normal ones use BDF.
2. Mediated pci devices may have some capabilities that normal pci devices do not have. For example, mediated pci devices could have regions that have sparse mmap capability, which allows a region to have multiple mmap areas. Another example is mediated pci devices may have regions/part of regions not mmaped but need to access them. Above difference will change the current ABI (i.e., struct rte_pci_device). Please check 5th and 6th patch for details.

3. A brand new mdev bus that does everything

Quoted 7 lines

This option will implement a new and standalone mdev bus. This option does not need any changes in current pci bus but only needs some shared code (linux vfio part) in pci bus. Drivers of devices that support mdev will register itself as a mdev driver and do not rely on pci bus anymore. This option, IMHO, will make the code clean. The only potential problem may be code duplication, which could be solved by making code of linux vfio part of pci bus common and shared.

Quoted 2 lines

Yes I prefer this third option.
We can find an elegant way of sharing some VFIO code between buses.

Yes, I have not thought about the details of the code sharing but will try to make it elegant.

Thanks,
Chenbo

Quoted 1 line
#30 lore raw

15/06/2021 04:49, Xia, Chenbo:

Quoted 19 lines

From: Thomas Monjalon <[email protected]>

Quoted 10 lines

01/06/2021 05:06, Chenbo Xia:

Quoted 6 lines

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci).

Please could you illustrate with an usage of mdev in DPDK? What does it enable which is not possible today?

The main purpose is for DPDK to drive mdev-based devices, which is not possible today.

I'd take PCI devices for an example. Currently DPDK can only drive devices of physical pci bus under /sys/bus/pci and kernel exposes the pci devices to APP in that way.

Quoted 9 lines

But there are PCI devices using vfio-mdev as a software framework to expose Mdev to APP under /sys/bus/mdev. Devices could choose this way of virtualizing itself to let multiple APPs share one physical device. For example, Intel Scalable IOV technology is known to use vfio-mdev as SW framework for Scalable IOV enabled devices (and Intel net/crypto/raw devices support this tech). For those mdev-based devices, DPDK needs support on the bus layer to scan/plug/probe/.. them, which is the main effort this patchset does. There are also other devices using the vfio-mdev framework, AFAIK, Nvidia's GPU is the first one using mdev and Intel's GPU virtualization also uses it.

Yes mdev was designed for virtualization I think.
The use of mdev for Scalable IOV without virtualization
may be seen as an abuse by Linux maintainers,
as they currently seem to prefer the auxiliary bus (which is a real bus).

Mellanox got a push back when trying to use mdev for the same purpose (Scalable Function, also called Sub-Function) in the kernel. The Linux community decided to use the auxiliary bus.

Any other feedback on the choice mdev vs aux?
Is there any kernel code supporting this mdev model for Intel devices?

Quoted 68 lines
Quoted 24 lines
Quoted 16 lines

In this patchset, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

The first four patches in this patchset are mainly preparation of mdev bus support. The left two patches are the key implementation of mdev bus.

The implementation of mdev bus in DPDK has several options:

Quoted 7 lines

1: Embed mdev bus in current pci bus

This patchset takes this option for an example. Mdev has several device types: pci/platform/amba/ccw/ap. DPDK currently only cares pci devices in all mdev device types so we could embed the mdev bus into current pci bus. Then pci bus with mdev support will scan/plug/ unplug/.. not only normal pci devices but also mediated pci devices.

Quoted 7 lines

I think it is a different bus.
It would be cleaner to not touch the PCI bus.
Having a separate bus will allow an easy way to identify a device with the new generic devargs syntax, example:
bus=mdev,uuid=XXX
or more complex:
bus=mdev,uuid=XXX/class=crypto/driver=qat,foo=bar

OK. Agree on cleaner to not touch PCI bus. And there may also be a 'type=pci' as mdev has several types in its definition (pci/ap/platform/ccw/...).

Quoted 33 lines
Quoted 9 lines

2: A new mdev bus that scans mediated pci devices and probes mdev driver to
plug-in pci devices to pci bus

If we took this option, a new mdev bus will be implemented to scan mediated pci devices and a new mdev driver for pci devices will be implemented in pci bus to plug-in mediated pci devices to pci bus.

Our RFC v1 takes this option:

   http://patchwork.dpdk.org/project/dpdk/cover/20190403071844.21126-1-

[email protected]/

Quoted 22 lines

Note that: for either option 1 or 2, device drivers do not know the implementation difference but only use structs/functions exposed by pci bus. Mediated pci devices are different from normal pci devices
on: 1. Mediated pci devices use UUID as address but normal ones use BDF.
2. Mediated pci devices may have some capabilities that normal pci devices do not have. For example, mediated pci devices could have regions that have sparse mmap capability, which allows a region to have multiple mmap areas. Another example is mediated pci devices may have regions/part of regions not mmaped but need to access them. Above difference will change the current ABI (i.e., struct rte_pci_device). Please check 5th and 6th patch for details.

3. A brand new mdev bus that does everything

This option will implement a new and standalone mdev bus. This option does not need any changes in current pci bus but only needs some shared code (linux vfio part) in pci bus. Drivers of devices that support mdev will register itself as a mdev driver and do not rely on pci bus anymore. This option, IMHO, will make the code clean. The only potential problem may be code duplication, which could be solved by making code of linux vfio part of pci bus common and shared.

Quoted 5 lines
Quoted 2 lines

Yes I prefer this third option.
We can find an elegant way of sharing some VFIO code between buses.

Yes, I have not thought about the details of the code sharing but will try to make it elegant.

Great, thanks.

#31 lore raw

Hi Thomas,

Quoted 54 lines

-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Tuesday, June 15, 2021 3:48 PM
To: Xia, Chenbo <[email protected]>
Cc: [email protected]; Liang, Cunming <[email protected]>; Wu, Jingjing <[email protected]>; Burakov, Anatoly <[email protected]>; Yigit, Ferruh <[email protected]>; [email protected]; [email protected]; Richardson, Bruce <[email protected]>; [email protected]; [email protected]; Ananyev, Konstantin <[email protected]>; [email protected]; [email protected]; [email protected]
Subject: Re: [dpdk-dev] [RFC v3 0/6] Add mdev (Mediated device) support in DPDK

15/06/2021 04:49, Xia, Chenbo:

Quoted 15 lines

From: Thomas Monjalon <[email protected]>

Quoted 10 lines

01/06/2021 05:06, Chenbo Xia:

Quoted 6 lines

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci).

Please could you illustrate with an usage of mdev in DPDK? What does it enable which is not possible today?

The main purpose is for DPDK to drive mdev-based devices, which is not possible today.

Quoted 6 lines

I'd take PCI devices for an example. Currently DPDK can only drive devices of physical pci bus under /sys/bus/pci and kernel exposes the pci devices to APP in that way.

But there are PCI devices using vfio-mdev as a software framework to expose Mdev to APP under /sys/bus/mdev. Devices could choose this way of

virtualizing

Quoted 2 lines

itself to let multiple APPs share one physical device. For example, Intel Scalable IOV technology is known to use vfio-mdev as SW framework for

Scalable

Quoted 1 line

IOV enabled devices (and Intel net/crypto/raw devices support this tech).

For

Quoted 1 line

those mdev-based devices, DPDK needs support on the bus layer to

scan/plug/probe/..

Quoted 1 line

them, which is the main effort this patchset does. There are also other

devices

Quoted 1 line

using the vfio-mdev framework, AFAIK, Nvidia's GPU is the first one using

mdev

Quoted 1 line

and Intel's GPU virtualization also uses it.

Yes mdev was designed for virtualization I think.
The use of mdev for Scalable IOV without virtualization
may be seen as an abuse by Linux maintainers,
as they currently seem to prefer the auxiliary bus (which is a real bus).

Quoted 5 lines

Mellanox got a push back when trying to use mdev for the same purpose (Scalable Function, also called Sub-Function) in the kernel. The Linux community decided to use the auxiliary bus.

Any other feedback on the choice mdev vs aux?

OK. Thanks for the info. Much appreciated.

I could investigate a bit about the choice and later come back to you.

Quoted 1 line

Is there any kernel code supporting this mdev model for Intel devices?

Now there's only intel GPU. But I think you care more about devices that DPDK could
drive: a dma device (DPDK's name ioat under raw/ioat) is on its way upstreaming
(https://www.spinics.net/lists/kvm/msg244417.html)

Thanks,
Chenbo

Quoted 87 lines
Quoted 33 lines
Quoted 25 lines
Quoted 17 lines

In this patchset, the PCI bus driver is extended to support scanning and probing the mdev devices whose device-api is "vfio-pci".

                     +---------+
                     | PCI bus |
                     +----+----+
                          |
         +--------+-------+-------+--------+
         |        |               |        |
  Physical PCI devices ...   Mediated PCI devices ...

The first four patches in this patchset are mainly preparation of mdev bus support. The left two patches are the key implementation of mdev bus.

The implementation of mdev bus in DPDK has several options:

Quoted 7 lines

1: Embed mdev bus in current pci bus

This patchset takes this option for an example. Mdev has several device types: pci/platform/amba/ccw/ap. DPDK currently only cares pci devices in all mdev device types so we could embed the mdev bus into current pci bus. Then pci bus with mdev support will scan/plug/ unplug/.. not only normal pci devices but also mediated pci devices.

Quoted 7 lines

I think it is a different bus.
It would be cleaner to not touch the PCI bus.
Having a separate bus will allow an easy way to identify a device with the new generic devargs syntax, example:
bus=mdev,uuid=XXX
or more complex:
bus=mdev,uuid=XXX/class=crypto/driver=qat,foo=bar

Quoted 1 line

OK. Agree on cleaner to not touch PCI bus. And there may also be a

'type=pci'

Quoted 3 lines

as mdev has several types in its definition (pci/ap/platform/ccw/...).

Quoted 1 line
Quoted 1 line

2: A new mdev bus that scans mediated pci devices and probes mdev driver

to

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

plug-in pci devices to pci bus

If we took this option, a new mdev bus will be implemented to scan mediated pci devices and a new mdev driver for pci devices will be implemented in pci bus to plug-in mediated pci devices to pci bus.

Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

Our RFC v1 takes this option:

   http://patchwork.dpdk.org/project/dpdk/cover/20190403071844.21126-1-

[email protected]/

Quoted 5 lines

Note that: for either option 1 or 2, device drivers do not know the implementation difference but only use structs/functions exposed by pci bus. Mediated pci devices are different from normal pci devices
on: 1. Mediated pci devices use UUID as address but normal ones use

BDF.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

2. Mediated pci devices may have some capabilities that normal pci devices do not have. For example, mediated pci devices could have regions that have sparse mmap capability, which allows a region to

have

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines

multiple mmap areas. Another example is mediated pci devices may have regions/part of regions not mmaped but need to access them. Above difference will change the current ABI (i.e., struct rte_pci_device). Please check 5th and 6th patch for details.

3. A brand new mdev bus that does everything

This option will implement a new and standalone mdev bus. This option does not need any changes in current pci bus but only needs some

shared

Quoted 1 line
Quoted 1 line
Quoted 1 line

code (linux vfio part) in pci bus. Drivers of devices that support

mdev

Quoted 1 line
Quoted 1 line
Quoted 1 line

will register itself as a mdev driver and do not rely on pci bus

anymore.

Quoted 1 line
Quoted 1 line
Quoted 1 line

This option, IMHO, will make the code clean. The only potential

problem

Quoted 1 line
Quoted 1 line
Quoted 1 line

may be code duplication, which could be solved by making code of

linux

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

vfio part of pci bus common and shared.

Yes I prefer this third option.
We can find an elegant way of sharing some VFIO code between buses.

Yes, I have not thought about the details of the code sharing but will try

to make

Quoted 1 line

it elegant.

Great, thanks.

#32 lore raw
In reply to Thomas Monjalon
Quoted 35 lines · Thomas Monjalon

15/06/2021 04:49, Xia, Chenbo:

Quoted 19 lines

From: Thomas Monjalon <[email protected]>

Quoted 10 lines

01/06/2021 05:06, Chenbo Xia:

Quoted 6 lines

Hi everyone,

This is a draft implementation of the mdev (Mediated device [1]) support in DPDK PCI bus driver. Mdev is a way to virtualize devices in Linux kernel. Based on the device-api (mdev_type/device_api), there could be different types of mdev devices (e.g. vfio-pci).

Please could you illustrate with an usage of mdev in DPDK? What does it enable which is not possible today?

The main purpose is for DPDK to drive mdev-based devices, which is not possible today.

I'd take PCI devices for an example. Currently DPDK can only drive devices of physical pci bus under /sys/bus/pci and kernel exposes the pci devices to APP in that way.

Quoted 9 lines

But there are PCI devices using vfio-mdev as a software framework to expose Mdev to APP under /sys/bus/mdev. Devices could choose this way of virtualizing itself to let multiple APPs share one physical device. For example, Intel Scalable IOV technology is known to use vfio-mdev as SW framework for Scalable IOV enabled devices (and Intel net/crypto/raw devices support this tech). For those mdev-based devices, DPDK needs support on the bus layer to scan/plug/probe/.. them, which is the main effort this patchset does. There are also other devices using the vfio-mdev framework, AFAIK, Nvidia's GPU is the first one using mdev and Intel's GPU virtualization also uses it.

Yes mdev was designed for virtualization I think.
The use of mdev for Scalable IOV without virtualization
may be seen as an abuse by Linux maintainers,
as they currently seem to prefer the auxiliary bus (which is a real bus).

Quoted 6 lines

Mellanox got a push back when trying to use mdev for the same purpose (Scalable Function, also called Sub-Function) in the kernel. The Linux community decided to use the auxiliary bus.

Any other feedback on the choice mdev vs aux?
Is there any kernel code supporting this mdev model for Intel devices?

IMHO until a kernel networking driver is accepted that uses mdev this is all just dead code in dpdk and shouldn't be merged.

I think it is unlikely that future networking drivers will use mdev.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

2: A new mdev bus that scans mediated pci devices and probes mdev driver to
plug-in pci devices to pci bus

And we are likely not doing 'mediated pci devices' at all..

Jason