Creation of sysfs entries is expensive, mainly for workloads that
constantly creates netdev and netns often.
Do not create BQL sysfs entries for devices that don't need,
basically those that do not have a real queue, i.e, devices that has
NETIF_F_LLTX and IFF_NO_QUEUE, such as `lo` interface.
This will remove the /sys/class/net/eth0/queues/tx-X/byte_queue_limits/
directory for these devices.
In the example below, eth0 has the `byte_queue_limits` directory but not
`lo`.
# ls /sys/class/net/lo/queues/tx-0/
traffic_class tx_maxrate tx_timeout xps_cpus xps_rxqs
# ls /sys/class/net/eth0/queues/tx-0/byte_queue_limits/
hold_time inflight limit limit_max limit_min
This also removes the #ifdefs, since we can also use netdev_uses_bql() to
check if the config is enabled. (as suggested by Jakub).
Suggested-by: Eric Dumazet <[email protected]>
Signed-off-by: Breno Leitao <[email protected]>
---
Changelog
v1:
* https://lore.kernel.org/all/[email protected]/
v2:
* Removed the ifdefs by adding them inside the netdev_uses_bql(), as
suggested by Jakub Kicinski.
v3:
* Added an extra parenthesis between or (||) clauses as suggested by
Stephen Hemminger
---
net/core/net-sysfs.c | 35 ++++++++++++++++++++++++-----------
1 file changed, 24 insertions(+), 11 deletions(-)
diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index 946caefdd959..e2a01b7b7992 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -1459,6 +1459,9 @@ static const struct attribute_group dql_group = {
.name = "byte_queue_limits",
.attrs = dql_attrs,
};
+#else
+/* Fake declaration, all the code using it should be dead */
+extern const struct attribute_group dql_group;
#endif /* CONFIG_BQL */
#ifdef CONFIG_XPS
@@ -1696,6 +1699,15 @@ static const struct kobj_type netdev_queue_ktype = {
.get_ownership = netdev_queue_get_ownership,
};
+static bool netdev_uses_bql(const struct net_device *dev)
+{
+ if ((dev->features & NETIF_F_LLTX) ||
+ (dev->priv_flags & IFF_NO_QUEUE))
+ return false;
+
+ return IS_ENABLED(CONFIG_BQL);
+}
+
static int netdev_queue_add_kobject(struct net_device *dev, int index)
{
struct netdev_queue *queue = dev->_tx + index;
@@ -1713,11 +1725,11 @@ static int netdev_queue_add_kobject(struct net_device *dev, int index)
if (error)
goto err;
-#ifdef CONFIG_BQL
- error = sysfs_create_group(kobj, &dql_group);
- if (error)
- goto err;
-#endif
+ if (netdev_uses_bql(dev)) {
+ error = sysfs_create_group(kobj, &dql_group);
+ if (error)
+ goto err;
+ }
kobject_uevent(kobj, KOBJ_ADD);
return 0;
@@ -1738,9 +1750,9 @@ static int tx_queue_change_owner(struct net_device *ndev, int index,
if (error)
return error;
-#ifdef CONFIG_BQL
- error = sysfs_group_change_owner(kobj, &dql_group, kuid, kgid);
-#endif
+ if (netdev_uses_bql(ndev))
+ error = sysfs_group_change_owner(kobj, &dql_group, kuid, kgid);
+
return error;
}
#endif /* CONFIG_SYSFS */
@@ -1772,9 +1784,10 @@ netdev_queue_update_kobjects(struct net_device *dev, int old_num, int new_num)
if (!refcount_read(&dev_net(dev)->ns.count))
queue->kobj.uevent_suppress = 1;
-#ifdef CONFIG_BQL
- sysfs_remove_group(&queue->kobj, &dql_group);
-#endif
+
+ if (netdev_uses_bql(dev))
+ sysfs_remove_group(&queue->kobj, &dql_group);
+
kobject_put(&queue->kobj);
}
--
2.39.3
On Mon, Feb 19, 2024 at 02:42:36AM -0800, Breno Leitao wrote:
> Creation of sysfs entries is expensive, mainly for workloads that
> constantly creates netdev and netns often.
>
> Do not create BQL sysfs entries for devices that don't need,
> basically those that do not have a real queue, i.e, devices that has
> NETIF_F_LLTX and IFF_NO_QUEUE, such as `lo` interface.
>
> This will remove the /sys/class/net/eth0/queues/tx-X/byte_queue_limits/
> directory for these devices.
>
> In the example below, eth0 has the `byte_queue_limits` directory but not
> `lo`.
>
> # ls /sys/class/net/lo/queues/tx-0/
> traffic_class tx_maxrate tx_timeout xps_cpus xps_rxqs
>
> # ls /sys/class/net/eth0/queues/tx-0/byte_queue_limits/
> hold_time inflight limit limit_max limit_min
>
> This also removes the #ifdefs, since we can also use netdev_uses_bql() to
> check if the config is enabled. (as suggested by Jakub).
>
> Suggested-by: Eric Dumazet <[email protected]>
> Signed-off-by: Breno Leitao <[email protected]>
Thanks Breno,
I like that this removes sysfs entries for devices for which they do not
act on. Although I understand it is not a complete solution - still not
all devices that have these entries use BQL - I think a key value there is
to alleviate potential confusion for users. And as such, this is a step in
the right direction.
I also like that some #ifdefs disappear.
Reviewed-by: Simon Horman <[email protected]>
..
On Mon, 19 Feb 2024 02:42:36 -0800 Breno Leitao wrote:
> Creation of sysfs entries is expensive, mainly for workloads that
> constantly creates netdev and netns often.
>
> Do not create BQL sysfs entries for devices that don't need,
> basically those that do not have a real queue, i.e, devices that has
> NETIF_F_LLTX and IFF_NO_QUEUE, such as `lo` interface.
FTR, I applied v2, thanks!