* [RFC PATCH 1/2] net: add sysfs attributes for customized dim profile management
2024-03-14 13:09 [RFC PATCH 0/2] net: provides dim profile fine-tuning channels Heng Qi
@ 2024-03-14 13:09 ` Heng Qi
2024-03-14 13:09 ` [RFC PATCH 2/2] virtio-net: support net sysfs to fine-tune dim profile Heng Qi
2024-03-14 18:49 ` [RFC PATCH 0/2] net: provides dim profile fine-tuning channels Jakub Kicinski
2 siblings, 0 replies; 5+ messages in thread
From: Heng Qi @ 2024-03-14 13:09 UTC (permalink / raw)
To: David S. Miller, Tal Gilboa, Eric Dumazet, Paolo Abeni,
Jakub Kicinski, Jason Wang, Michael S. Tsirkin, Xuan Zhuo,
Jonathan Corbet, Randy Dunlap, Bjorn Helgaas, Breno Leitao,
Johannes Berg, netdev, linux-kernel, virtualization
The NetDIM library, currently leveraged by an array of NICs, delivers
excellent acceleration benefits. Nevertheless, NICs vary significantly
in their dim profile list prerequisites.
Specifically, virtio-net backends may present diverse sw or hw device
implementation, making a one-size-fits-all parameter list impractical.
On Alibaba Cloud, the virtio DPU's performance under the default DIM
profile falls short of expectations, partly due to a mismatch in
parameter configuration.
I also noticed that ice/idpf/ena and other NICs have customized
profilelist or placed some restrictions on dim capabilities.
Motivated by this, I tried adding new sysfs attributes that provides
a per-device control to modify and access a device's interrupt parameters.
Usage
========
1. Query the currently customized list of the device
$ cat dim_profs
The profiles of (RX, EQE):
{.usec = 1, .pkts = 256, .comps = 0,},
{.usec = 8, .pkts = 256, .comps = 0,},
{.usec = 64, .pkts = 256, .comps = 0,},
{.usec = 128, .pkts = 256, .comps = 0,},
{.usec = 256, .pkts = 256, .comps = 0,}
The profiles of (TX, EQE):
{.usec = 1, .pkts = 256, .comps = 0,},
{.usec = 2, .pkts = 256, .comps = 0,},
{.usec = 3, .pkts = 256, .comps = 0,},
{.usec = 4, .pkts = 256, .comps = 0,},
{.usec = 5, .pkts = 256, .comps = 0,}
2. Tune
$ echo "RX EQE 8,8,0 16,16,0 32,32,0 64,64,0 128,128,0" > dim_profs
$ echo " TX EQE 0,2,0 1,3,0 2,4,0 3,5,0 4,6,0 " > dim_profs
$ cat dim_profs
The profiles of (RX, EQE):
{.usec = 8, .pkts = 8, .comps = 0,},
{.usec = 16, .pkts = 16, .comps = 0,},
{.usec = 32, .pkts = 32, .comps = 0,},
{.usec = 64, .pkts = 64, .comps = 0,},
{.usec = 128, .pkts = 128, .comps = 0,}
The profiles of (TX, EQE):
{.usec = 0, .pkts = 2, .comps = 0,},
{.usec = 1, .pkts = 3, .comps = 0,},
{.usec = 2, .pkts = 4, .comps = 0,},
{.usec = 3, .pkts = 5, .comps = 0,},
{.usec = 4, .pkts = 6, .comps = 0,}
3. Warn
If the device does not support .ndo_dim_moder_{set, get},
the following warning will response:
"Profile is default and not customized by the device."
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
---
Documentation/ABI/testing/sysfs-class-net | 17 +++
include/linux/dim.h | 7 ++
include/linux/netdevice.h | 35 ++++++
lib/dim/net_dim.c | 6 --
net/core/net-sysfs.c | 172 ++++++++++++++++++++++++++++++
5 files changed, 231 insertions(+), 6 deletions(-)
diff --git a/Documentation/ABI/testing/sysfs-class-net b/Documentation/ABI/testing/sysfs-class-net
index ebf21be..1e4faa8 100644
--- a/Documentation/ABI/testing/sysfs-class-net
+++ b/Documentation/ABI/testing/sysfs-class-net
@@ -352,3 +352,20 @@ Description:
0 threaded mode disabled for this dev
1 threaded mode enabled for this dev
== ==================================
+
+What: /sys/class/net/<iface>/dim_profs
+Date: Mar 2024
+KernelVersion: 6.8
+Contact: netdev@vger.kernel.org
+Description:
+ String value to control the profile list of DIM per device. User could
+ set this value to tune the profile list for RX/TX direction and EQE/CQE
+ mode respectively.
+
+ Possible values:
+ ================================================ ==========================
+ RX EQE 1,1,0 2,2,0 3,3,0 4,4,0 5,5,0 tune RX + EQE profile list
+ RX CQE 8,8,0 16,16,0 32,32,0 64,64,0 128,128,0 tune RX + CQE profile list
+ TX EQE 16,8,0 2,16,0 16,8,0 32,64,0 128,64,0 tune TX + EQE profile list
+ TX CQE 8,5,0 8,16,0 32,12,0 128,64,0 256,128,0 tune TX + CQE profile list
+ ================================================ ==========================
diff --git a/include/linux/dim.h b/include/linux/dim.h
index f343bc9..43398f5 100644
--- a/include/linux/dim.h
+++ b/include/linux/dim.h
@@ -10,6 +10,13 @@
#include <linux/types.h>
#include <linux/workqueue.h>
+/* Number of DIM profiles and period mode. */
+#define NET_DIM_PARAMS_NUM_PROFILES 5
+#define NET_DIM_DEFAULT_RX_CQ_PKTS_FROM_EQE 256
+#define NET_DIM_DEFAULT_TX_CQ_PKTS_FROM_EQE 128
+#define NET_DIM_DEF_PROFILE_CQE 1
+#define NET_DIM_DEF_PROFILE_EQE 1
+
/*
* Number of events between DIM iterations.
* Causes a moderation of the algorithm run.
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index c6f6ac7..bc2f3ac 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -49,6 +49,7 @@
#include <uapi/linux/netdev.h>
#include <linux/hashtable.h>
#include <linux/rbtree.h>
+#include <linux/dim.h>
#include <net/net_trackers.h>
#include <net/net_debug.h>
#include <net/dropreason-core.h>
@@ -998,6 +999,27 @@ struct netdev_net_notifier {
struct notifier_block *nb;
};
+enum dim_direction {
+ DIM_RX_DIRECTION = 0x0,
+ DIM_TX_DIRECTION = 0x1,
+ DIM_NUM_DIRECTIONS
+};
+/**
+ * struct dim_profs_list - Structure for dim sysfs configuration.
+ * Used to exchange profile list between the sysfs and the driver.
+ *
+ * @direction: RX or TX dim information
+ * @mode: CQ period count mode (from CQE/EQE)
+ * @num: the number of profs array
+ * @profs: dim profile list
+ */
+struct dim_profs_list {
+ u8 direction;
+ u8 mode;
+ u8 num;
+ struct dim_cq_moder profs[];
+};
+
/*
* This structure defines the management hooks for network devices.
* The following hooks can be defined; unless noted otherwise, they are
@@ -1351,6 +1373,14 @@ struct netdev_net_notifier {
* struct kernel_hwtstamp_config *kernel_config,
* struct netlink_ext_ack *extack);
* Change the hardware timestamping parameters for NIC device.
+ *
+ * int (*ndo_dim_moder_get)(struct net_device *dev,
+ * struct dim_profs_list *list);
+ * Get dim profiles list from the NIC device.
+ *
+ * int (*ndo_dim_moder_set)(struct net_device *dev,
+ * struct dim_profs_list *list);
+ * Configure dim profiles list for the NIC device.
*/
struct net_device_ops {
int (*ndo_init)(struct net_device *dev);
@@ -1595,6 +1625,11 @@ struct net_device_ops {
int (*ndo_hwtstamp_set)(struct net_device *dev,
struct kernel_hwtstamp_config *kernel_config,
struct netlink_ext_ack *extack);
+ int (*ndo_dim_moder_get)(struct net_device *dev,
+ struct dim_profs_list *list);
+
+ int (*ndo_dim_moder_set)(struct net_device *dev,
+ struct dim_profs_list *list);
};
/**
diff --git a/lib/dim/net_dim.c b/lib/dim/net_dim.c
index 4e32f7a..67d5beb 100644
--- a/lib/dim/net_dim.c
+++ b/lib/dim/net_dim.c
@@ -11,12 +11,6 @@
* There are different set of profiles for RX/TX CQs.
* Each profile size must be of NET_DIM_PARAMS_NUM_PROFILES
*/
-#define NET_DIM_PARAMS_NUM_PROFILES 5
-#define NET_DIM_DEFAULT_RX_CQ_PKTS_FROM_EQE 256
-#define NET_DIM_DEFAULT_TX_CQ_PKTS_FROM_EQE 128
-#define NET_DIM_DEF_PROFILE_CQE 1
-#define NET_DIM_DEF_PROFILE_EQE 1
-
#define NET_DIM_RX_EQE_PROFILES { \
{.usec = 1, .pkts = NET_DIM_DEFAULT_RX_CQ_PKTS_FROM_EQE,}, \
{.usec = 8, .pkts = NET_DIM_DEFAULT_RX_CQ_PKTS_FROM_EQE,}, \
diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index e3d7a8c..801cb07 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -23,6 +23,7 @@
#include <linux/of.h>
#include <linux/of_net.h>
#include <linux/cpu.h>
+#include <linux/dim.h>
#include <net/netdev_rx_queue.h>
#include <net/rps.h>
@@ -638,6 +639,176 @@ static ssize_t threaded_store(struct device *dev,
}
static DEVICE_ATTR_RW(threaded);
+static struct dim_profs_list *parse_dim_profs(const char *buf, ssize_t len)
+{
+ int i, ret, size, totlen = 0, retlen = 0;
+ char direction[3], period_mode[4];
+ struct dim_profs_list *list;
+
+ size = sizeof(*list) + NET_DIM_PARAMS_NUM_PROFILES * sizeof(struct dim_cq_moder);
+ list = kzalloc(size, GFP_KERNEL);
+ if (!list)
+ goto err_list;
+
+ list->num = NET_DIM_PARAMS_NUM_PROFILES;
+
+ ret = sscanf(buf, "%2s %3s%n", direction, period_mode, &retlen);
+ if (ret != 2)
+ goto err_parse;
+
+ if (!strcasecmp(direction, "RX"))
+ list->direction = DIM_RX_DIRECTION;
+ else if (!strcasecmp(direction, "TX"))
+ list->direction = DIM_TX_DIRECTION;
+ else
+ goto err_parse;
+
+ if (!strcasecmp(period_mode, "EQE"))
+ list->mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
+ else if (!strcasecmp(period_mode, "CQE"))
+ list->mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
+ else
+ goto err_parse;
+
+ totlen += retlen;
+ if (totlen > len)
+ goto err_parse;
+
+ buf += retlen;
+ if (!buf)
+ goto err_parse;
+
+ for (i = 0; i < NET_DIM_PARAMS_NUM_PROFILES; i++) {
+ ret = sscanf(buf, "%hu,%hu,%hu%n", &list->profs[i].usec,
+ &list->profs[i].pkts, &list->profs[i].comps, &retlen);
+ if (ret != 3)
+ goto err_parse;
+
+ totlen += retlen;
+ if (totlen > len)
+ goto err_parse;
+
+ buf += retlen;
+ if (i == NET_DIM_PARAMS_NUM_PROFILES - 1)
+ break;
+ }
+
+ return list;
+
+err_parse:
+ kfree(list);
+err_list:
+ return NULL;
+}
+
+static ssize_t dim_profs_store(struct device *dev,
+ struct device_attribute *attr,
+ const char *buf, size_t len)
+{
+ struct net_device *netdev = to_net_dev(dev);
+ const struct net_device_ops *ops = netdev->netdev_ops;
+ struct net *net = dev_net(netdev);
+ struct dim_profs_list *list;
+ int ret = 0;
+
+ if (!ns_capable(net->user_ns, CAP_NET_ADMIN))
+ return -EPERM;
+
+ list = parse_dim_profs(buf, len);
+ if (!list)
+ return -EINVAL;
+
+ if (!rtnl_trylock())
+ return restart_syscall();
+
+ if (dev_isalive(netdev)) {
+ if (!ops->ndo_dim_moder_set)
+ ret = -EINVAL;
+ else
+ ret = ops->ndo_dim_moder_set(netdev, list) ? : len;
+ }
+
+ kfree(list);
+ rtnl_unlock();
+
+ return ret;
+}
+
+static ssize_t dim_profs_show_one(struct device *dev,
+ struct device_attribute *attr,
+ char *buf, u8 direct, u8 mode,
+ size_t *len_)
+{
+ static const char fmt_body[] = "{.usec = %3hu, .pkts = %3hu, .comps = %3hu,}%s";
+ static const char fmt_hdr[] = "The profiles of (%2s, %3s):\n";
+ const char *direction[2] = {"RX", "TX"}, *period_mode[2] = {"EQE", "CQE"};
+ struct net_device *netdev = to_net_dev(dev);
+ const struct net_device_ops *ops = netdev->netdev_ops;
+ struct dim_profs_list *list;
+ size_t size, len = *len_;
+ ssize_t i;
+
+ size = sizeof(*list) + NET_DIM_PARAMS_NUM_PROFILES * sizeof(struct dim_cq_moder);
+ list = kzalloc(size, GFP_KERNEL);
+ if (!list)
+ return -ENOMEM;
+
+ list->num = NET_DIM_PARAMS_NUM_PROFILES;
+ list->direction = direct;
+ list->mode = mode;
+ if (ops->ndo_dim_moder_get(netdev, list))
+ goto ret_;
+
+ len += scnprintf(buf + len, PAGE_SIZE - len,
+ fmt_hdr, direction[direct], period_mode[mode]);
+ for (i = 0; i < NET_DIM_PARAMS_NUM_PROFILES; i++) {
+ len += scnprintf(buf + len, PAGE_SIZE - len, fmt_body,
+ list->profs[i].usec, list->profs[i].pkts,
+ list->profs[i].comps,
+ (i == NET_DIM_PARAMS_NUM_PROFILES - 1) ? "\n" : ",\n");
+ }
+ *len_ = len;
+ret_:
+ kfree(list);
+ return 0;
+}
+
+static ssize_t dim_profs_show(struct device *dev,
+ struct device_attribute *attr,
+ char *buf)
+{
+ static const char out[] = "profile is default and not customized by the device.";
+ struct net_device *netdev = to_net_dev(dev);
+ const struct net_device_ops *ops = netdev->netdev_ops;
+ ssize_t i, j, ret = 0;
+ size_t len = 0;
+
+ if (!rtnl_trylock())
+ return restart_syscall();
+
+ if (!ops->ndo_dim_moder_get) {
+ ret = sysfs_emit(buf, "%s\n", out);
+ goto ret_;
+ }
+
+ for (i = 0; i < DIM_NUM_DIRECTIONS; i++) {
+ for (j = 0; j < DIM_CQ_PERIOD_NUM_MODES; j++) {
+ ret = dim_profs_show_one(dev, attr, buf, i, j, &len);
+ if (ret)
+ goto ret_;
+ }
+ }
+
+ rtnl_unlock();
+ return len;
+
+ret_:
+ rtnl_unlock();
+ return ret;
+}
+
+static DEVICE_ATTR_RW(dim_profs);
+
static struct attribute *net_class_attrs[] __ro_after_init = {
&dev_attr_netdev_group.attr,
&dev_attr_type.attr,
@@ -671,6 +842,7 @@ static ssize_t threaded_store(struct device *dev,
&dev_attr_carrier_up_count.attr,
&dev_attr_carrier_down_count.attr,
&dev_attr_threaded.attr,
+ &dev_attr_dim_profs.attr,
NULL,
};
ATTRIBUTE_GROUPS(net_class);
--
1.8.3.1
^ permalink raw reply related [flat|nested] 5+ messages in thread* [RFC PATCH 2/2] virtio-net: support net sysfs to fine-tune dim profile
2024-03-14 13:09 [RFC PATCH 0/2] net: provides dim profile fine-tuning channels Heng Qi
2024-03-14 13:09 ` [RFC PATCH 1/2] net: add sysfs attributes for customized dim profile management Heng Qi
@ 2024-03-14 13:09 ` Heng Qi
2024-03-14 18:49 ` [RFC PATCH 0/2] net: provides dim profile fine-tuning channels Jakub Kicinski
2 siblings, 0 replies; 5+ messages in thread
From: Heng Qi @ 2024-03-14 13:09 UTC (permalink / raw)
To: David S. Miller, Tal Gilboa, Eric Dumazet, Paolo Abeni,
Jakub Kicinski, Jason Wang, Michael S. Tsirkin, Xuan Zhuo,
Jonathan Corbet, Randy Dunlap, Bjorn Helgaas, Breno Leitao,
Johannes Berg, netdev, linux-kernel, virtualization
Virtio-net has different types of back-end device
implementations. In order to effectively optimize
the dim library's gains for different device
implementations, let's use the interface provided
by net-sysfs to fine-tune the profile list.
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
---
drivers/net/virtio_net.c | 64 +++++++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 63 insertions(+), 1 deletion(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index e709d44..7fae737 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -57,6 +57,16 @@
#define VIRTNET_DRIVER_VERSION "1.0.0"
+/* This is copied from NET_DIM_RX_EQE_PROFILES in DIM library */
+#define VIRTNET_DIM_RX_PKTS 256
+static struct dim_cq_moder rx_itr_conf[] = {
+ {.usec = 1, .pkts = VIRTNET_DIM_RX_PKTS,},
+ {.usec = 8, .pkts = VIRTNET_DIM_RX_PKTS,},
+ {.usec = 64, .pkts = VIRTNET_DIM_RX_PKTS,},
+ {.usec = 128, .pkts = VIRTNET_DIM_RX_PKTS,},
+ {.usec = 256, .pkts = VIRTNET_DIM_RX_PKTS,}
+};
+
static const unsigned long guest_offloads[] = {
VIRTIO_NET_F_GUEST_TSO4,
VIRTIO_NET_F_GUEST_TSO6,
@@ -3584,7 +3594,10 @@ static void virtnet_rx_dim_work(struct work_struct *work)
if (!rq->dim_enabled)
continue;
- update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
+ if (dim->profile_ix >= ARRAY_SIZE(rx_itr_conf))
+ dim->profile_ix = ARRAY_SIZE(rx_itr_conf) - 1;
+
+ update_moder = rx_itr_conf[dim->profile_ix];
if (update_moder.usec != rq->intr_coal.max_usecs ||
update_moder.pkts != rq->intr_coal.max_packets) {
err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
@@ -4170,6 +4183,53 @@ static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
}
+static int virtnet_dim_moder_valid(struct net_device *dev, struct dim_profs_list *list)
+{
+ struct virtnet_info *vi = netdev_priv(dev);
+
+ if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
+ return -EOPNOTSUPP;
+
+ if (!list || list->direction != DIM_RX_DIRECTION ||
+ list->num != NET_DIM_PARAMS_NUM_PROFILES ||
+ list->mode != DIM_CQ_PERIOD_MODE_START_FROM_EQE) {
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+static int virtnet_dim_moder_get(struct net_device *dev, struct dim_profs_list *list)
+{
+ int ret;
+
+ ret = virtnet_dim_moder_valid(dev, list);
+ if (ret)
+ return ret;
+
+ memcpy(list->profs, rx_itr_conf, sizeof(*list->profs) * list->num);
+
+ return 0;
+}
+
+static int virtnet_dim_moder_set(struct net_device *dev, struct dim_profs_list *list)
+{
+ int i, ret;
+
+ ret = virtnet_dim_moder_valid(dev, list);
+ if (ret)
+ return ret;
+
+ for (i = 0; i < list->num; i++) {
+ rx_itr_conf[i].usec = list->profs[i].usec;
+ rx_itr_conf[i].pkts = list->profs[i].pkts;
+ if (list->profs[i].comps)
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
static const struct net_device_ops virtnet_netdev = {
.ndo_open = virtnet_open,
.ndo_stop = virtnet_close,
@@ -4186,6 +4246,8 @@ static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
.ndo_get_phys_port_name = virtnet_get_phys_port_name,
.ndo_set_features = virtnet_set_features,
.ndo_tx_timeout = virtnet_tx_timeout,
+ .ndo_dim_moder_get = virtnet_dim_moder_get,
+ .ndo_dim_moder_set = virtnet_dim_moder_set,
};
static void virtnet_config_changed_work(struct work_struct *work)
--
1.8.3.1
^ permalink raw reply related [flat|nested] 5+ messages in thread