RSS is a receive steering algorithm that can be negotiated to use with
virtio_net. Conventionally the hash calculation was done by the VMM.
However, computing the hash after the queue was chosen defeats the
purpose of RSS.
Another approach is to use eBPF steering program. This approach has
another downside: it cannot report the calculated hash due to the
restrictive nature of eBPF steering program.
Introduce the code to perform RSS to the kernel in order to overcome
thse challenges. An alternative solution is to extend the eBPF steering
program so that it will be able to report to the userspace, but I didn't
opt for it because extending the current mechanism of eBPF steering
program as is because it relies on legacy context rewriting, and
introducing kfunc-based eBPF will result in non-UAPI dependency while
the other relevant virtualization APIs such as KVM and vhost_net are
UAPIs.
Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
---
drivers/net/tun.c | 158 ++++++++++++++++++++++++++++++++++++++------
include/uapi/linux/if_tun.h | 27 ++++++++
2 files changed, 163 insertions(+), 22 deletions(-)
diff --git a/drivers/net/tun.c b/drivers/net/tun.c
index 986e4a5bf04d..680eb4561a7f 100644
--- a/drivers/net/tun.c
+++ b/drivers/net/tun.c
@@ -175,6 +175,9 @@ struct tun_prog {
struct tun_vnet_hash_container {
struct tun_vnet_hash common;
+ struct tun_vnet_hash_rss rss;
+ __be32 rss_key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
+ u16 rss_indirection_table[];
};
/* Since the socket were moved to tun_file, to preserve the behavior of persist
@@ -214,7 +217,7 @@ struct tun_struct {
struct bpf_prog __rcu *xdp_prog;
struct tun_prog __rcu *steering_prog;
struct tun_prog __rcu *filter_prog;
- struct tun_vnet_hash vnet_hash;
+ struct tun_vnet_hash_container __rcu *vnet_hash;
struct ethtool_link_ksettings link_ksettings;
/* init args */
struct file *file;
@@ -227,7 +230,7 @@ struct veth {
};
static const struct tun_vnet_hash tun_vnet_hash_cap = {
- .flags = TUN_VNET_HASH_REPORT,
+ .flags = TUN_VNET_HASH_REPORT | TUN_VNET_HASH_RSS,
.types = VIRTIO_NET_SUPPORTED_HASH_TYPES
};
@@ -333,8 +336,10 @@ static long tun_set_vnet_be(struct tun_struct *tun, int __user *argp)
return -EFAULT;
if (be) {
+ struct tun_vnet_hash_container *vnet_hash = rtnl_dereference(tun->vnet_hash);
+
if (!(tun->flags & TUN_VNET_LE) &&
- (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT))
+ vnet_hash && (vnet_hash->flags & TUN_VNET_HASH_REPORT))
return -EBUSY;
tun->flags |= TUN_VNET_BE;
@@ -537,7 +542,8 @@ static inline void tun_flow_save_rps_rxhash(struct tun_flow_entry *e, u32 hash)
* the userspace application move between processors, we may get a
* different rxq no. here.
*/
-static u16 tun_automq_select_queue(struct tun_struct *tun, struct sk_buff *skb)
+static u16 tun_automq_select_queue(struct tun_struct *tun, struct sk_buff *skb,
+ const struct tun_vnet_hash_container *vnet_hash)
{
struct tun_vnet_hash_ext *ext;
struct flow_keys keys;
@@ -558,10 +564,10 @@ static u16 tun_automq_select_queue(struct tun_struct *tun, struct sk_buff *skb)
txq = reciprocal_scale(txq, numqueues);
}
- if (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT) {
+ if (vnet_hash && (vnet_hash->common.flags & TUN_VNET_HASH_REPORT)) {
ext = skb_ext_add(skb, SKB_EXT_TUN_VNET_HASH);
if (ext) {
- u32 types = tun->vnet_hash.types;
+ u32 types = vnet_hash->common.types;
ext->report = virtio_net_hash_report(types, keys.basic);
ext->value = skb->l4_hash ? skb->hash : txq;
@@ -588,6 +594,37 @@ static u16 tun_ebpf_select_queue(struct tun_struct *tun, struct sk_buff *skb)
return ret % numqueues;
}
+static u16 tun_vnet_rss_select_queue(struct tun_struct *tun,
+ struct sk_buff *skb,
+ const struct tun_vnet_hash_container *vnet_hash)
+{
+ struct tun_vnet_hash_ext *ext;
+ struct virtio_net_hash hash;
+ u32 numqueues = READ_ONCE(tun->numqueues);
+ u16 txq, index;
+
+ if (!numqueues)
+ return 0;
+
+ virtio_net_hash_rss(skb, vnet_hash->common.types, vnet_hash->rss_key, &hash);
+
+ if (!hash.report)
+ return vnet_hash->rss.unclassified_queue % numqueues;
+
+ if (vnet_hash->common.flags & TUN_VNET_HASH_REPORT) {
+ ext = skb_ext_add(skb, SKB_EXT_TUN_VNET_HASH);
+ if (ext) {
+ ext->value = hash.value;
+ ext->report = hash.report;
+ }
+ }
+
+ index = hash.value & vnet_hash->rss.indirection_table_mask;
+ txq = READ_ONCE(vnet_hash->rss_indirection_table[index]);
+
+ return txq % numqueues;
+}
+
static u16 tun_select_queue(struct net_device *dev, struct sk_buff *skb,
struct net_device *sb_dev)
{
@@ -595,10 +632,16 @@ static u16 tun_select_queue(struct net_device *dev, struct sk_buff *skb,
u16 ret;
rcu_read_lock();
- if (rcu_dereference(tun->steering_prog))
+ if (rcu_dereference(tun->steering_prog)) {
ret = tun_ebpf_select_queue(tun, skb);
- else
- ret = tun_automq_select_queue(tun, skb);
+ } else {
+ struct tun_vnet_hash_container *vnet_hash = rcu_dereference(tun->vnet_hash);
+
+ if (vnet_hash && (vnet_hash->common.flags & TUN_VNET_HASH_RSS))
+ ret = tun_vnet_rss_select_queue(tun, skb, vnet_hash);
+ else
+ ret = tun_automq_select_queue(tun, skb, vnet_hash);
+ }
rcu_read_unlock();
return ret;
@@ -2376,6 +2419,9 @@ static void tun_free_netdev(struct net_device *dev)
security_tun_dev_free_security(tun->security);
__tun_set_ebpf(tun, &tun->steering_prog, NULL);
__tun_set_ebpf(tun, &tun->filter_prog, NULL);
+ rtnl_lock();
+ kfree_rcu_mightsleep(rtnl_dereference(tun->vnet_hash));
+ rtnl_unlock();
}
static void tun_setup(struct net_device *dev)
@@ -3074,13 +3120,9 @@ static int tun_set_queue(struct file *file, struct ifreq *ifr)
}
static int tun_set_ebpf(struct tun_struct *tun, struct tun_prog __rcu **prog_p,
- void __user *data)
+ int fd)
{
struct bpf_prog *prog;
- int fd;
-
- if (copy_from_user(&fd, data, sizeof(fd)))
- return -EFAULT;
if (fd == -1) {
prog = NULL;
@@ -3146,10 +3188,12 @@ static long __tun_chr_ioctl(struct file *file, unsigned int cmd,
int ifindex;
int sndbuf;
int vnet_hdr_sz;
+ int fd;
int le;
int ret;
bool do_notify = false;
- struct tun_vnet_hash vnet_hash;
+ struct tun_vnet_hash vnet_hash_common;
+ struct tun_vnet_hash_container *vnet_hash;
if (cmd == TUNSETIFF || cmd == TUNSETQUEUE ||
(_IOC_TYPE(cmd) == SOCK_IOC_TYPE && cmd != SIOCGSKNS)) {
@@ -3373,7 +3417,8 @@ static long __tun_chr_ioctl(struct file *file, unsigned int cmd,
break;
}
- if (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT &&
+ vnet_hash = rtnl_dereference(tun->vnet_hash);
+ if (vnet_hash && (vnet_hash->common.flags & TUN_VNET_HASH_REPORT) &&
vnet_hdr_sz < (int)sizeof(struct virtio_net_hdr_v1_hash)) {
ret = -EBUSY;
break;
@@ -3396,7 +3441,8 @@ static long __tun_chr_ioctl(struct file *file, unsigned int cmd,
if (le) {
tun->flags |= TUN_VNET_LE;
} else {
- if (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT &&
+ vnet_hash = rtnl_dereference(tun->vnet_hash);
+ if (vnet_hash && (vnet_hash->common.flags & TUN_VNET_HASH_REPORT) &&
!tun_legacy_is_little_endian(tun)) {
ret = -EBUSY;
break;
@@ -3446,11 +3492,27 @@ static long __tun_chr_ioctl(struct file *file, unsigned int cmd,
break;
case TUNSETSTEERINGEBPF:
- ret = tun_set_ebpf(tun, &tun->steering_prog, argp);
+ if (get_user(fd, (int __user *)argp)) {
+ ret = -EFAULT;
+ break;
+ }
+
+ vnet_hash = rtnl_dereference(tun->vnet_hash);
+ if (fd != -1 && vnet_hash && (vnet_hash->common.flags & TUN_VNET_HASH_RSS)) {
+ ret = -EBUSY;
+ break;
+ }
+
+ ret = tun_set_ebpf(tun, &tun->steering_prog, fd);
break;
case TUNSETFILTEREBPF:
- ret = tun_set_ebpf(tun, &tun->filter_prog, argp);
+ if (get_user(fd, (int __user *)argp)) {
+ ret = -EFAULT;
+ break;
+ }
+
+ ret = tun_set_ebpf(tun, &tun->filter_prog, fd);
break;
case TUNSETCARRIER:
@@ -3469,20 +3531,72 @@ static long __tun_chr_ioctl(struct file *file, unsigned int cmd,
break;
case TUNSETVNETHASH:
- if (copy_from_user(&vnet_hash, argp, sizeof(vnet_hash))) {
+ if (copy_from_user(&vnet_hash_common, argp, sizeof(vnet_hash_common))) {
ret = -EFAULT;
break;
}
argp = (struct tun_vnet_hash __user *)argp + 1;
- if ((vnet_hash.flags & TUN_VNET_HASH_REPORT) &&
+ if ((vnet_hash_common.flags & TUN_VNET_HASH_REPORT) &&
(tun->vnet_hdr_sz < sizeof(struct virtio_net_hdr_v1_hash) ||
!tun_is_little_endian(tun))) {
ret = -EBUSY;
break;
}
- tun->vnet_hash = vnet_hash;
+ if (vnet_hash_common.flags & TUN_VNET_HASH_RSS) {
+ struct tun_vnet_hash_rss rss;
+ size_t indirection_table_size;
+ size_t key_size;
+ size_t size;
+
+ if (tun->steering_prog) {
+ ret = -EBUSY;
+ break;
+ }
+
+ if (copy_from_user(&rss, argp, sizeof(rss))) {
+ ret = -EFAULT;
+ break;
+ }
+ argp = (struct tun_vnet_hash_rss __user *)argp + 1;
+
+ indirection_table_size = ((size_t)rss.indirection_table_mask + 1) * 2;
+ key_size = virtio_net_hash_key_length(vnet_hash_common.types);
+ size = struct_size(vnet_hash, rss_indirection_table,
+ (size_t)rss.indirection_table_mask + 1);
+
+ vnet_hash = kmalloc(size, GFP_KERNEL);
+ if (!vnet_hash) {
+ ret = -ENOMEM;
+ break;
+ }
+
+ if (copy_from_user(vnet_hash->rss_indirection_table,
+ argp, indirection_table_size)) {
+ kfree(vnet_hash);
+ ret = -EFAULT;
+ break;
+ }
+ argp = (u16 __user *)argp + rss.indirection_table_mask + 1;
+
+ if (copy_from_user(vnet_hash->rss_key, argp, key_size)) {
+ kfree(vnet_hash);
+ ret = -EFAULT;
+ break;
+ }
+
+ vnet_hash->rss = rss;
+ } else {
+ vnet_hash = kmalloc(sizeof(vnet_hash->common), GFP_KERNEL);
+ if (!vnet_hash) {
+ ret = -ENOMEM;
+ break;
+ }
+ }
+
+ vnet_hash->common = vnet_hash_common;
+ kfree_rcu_mightsleep(rcu_replace_pointer_rtnl(tun->vnet_hash, vnet_hash));
break;
default:
diff --git a/include/uapi/linux/if_tun.h b/include/uapi/linux/if_tun.h
index 1561e8ce0a0a..1c130409db5d 100644
--- a/include/uapi/linux/if_tun.h
+++ b/include/uapi/linux/if_tun.h
@@ -75,6 +75,14 @@
*
* The argument is a pointer to &struct tun_vnet_hash.
*
+ * The argument is a pointer to the compound of the following in order if
+ * %TUN_VNET_HASH_RSS is set:
+ *
+ * 1. &struct tun_vnet_hash
+ * 2. &struct tun_vnet_hash_rss
+ * 3. Indirection table
+ * 4. Key
+ *
* %TUNSETVNETHDRSZ ioctl must be called with a number greater than or equal to
* the size of &struct virtio_net_hdr_v1_hash before calling this ioctl with
* %TUN_VNET_HASH_REPORT.
@@ -144,6 +152,13 @@ struct tun_filter {
*/
#define TUN_VNET_HASH_REPORT 0x0001
+/**
+ * define TUN_VNET_HASH_RSS - Request virtio_net RSS
+ *
+ * This is mutually exclusive with eBPF steering program.
+ */
+#define TUN_VNET_HASH_RSS 0x0002
+
/**
* struct tun_vnet_hash - virtio_net hashing configuration
* @flags:
@@ -159,4 +174,16 @@ struct tun_vnet_hash {
__u32 types;
};
+/**
+ * struct tun_vnet_hash_rss - virtio_net RSS configuration
+ * @indirection_table_mask:
+ * Bitmask to be applied to the indirection table index
+ * @unclassified_queue:
+ * The index of the queue to place unclassified packets in
+ */
+struct tun_vnet_hash_rss {
+ __u16 indirection_table_mask;
+ __u16 unclassified_queue;
+};
+
#endif /* _UAPI__IF_TUN_H */
--
2.46.0On Tue, Sep 24, 2024 at 5:01 PM Akihiko Odaki <akihiko.odaki@daynix.com> wrote: > > RSS is a receive steering algorithm that can be negotiated to use with > virtio_net. Conventionally the hash calculation was done by the VMM. > However, computing the hash after the queue was chosen defeats the > purpose of RSS. > > Another approach is to use eBPF steering program. This approach has > another downside: it cannot report the calculated hash due to the > restrictive nature of eBPF steering program. > > Introduce the code to perform RSS to the kernel in order to overcome > thse challenges. An alternative solution is to extend the eBPF steering > program so that it will be able to report to the userspace, but I didn't > opt for it because extending the current mechanism of eBPF steering > program as is because it relies on legacy context rewriting, and > introducing kfunc-based eBPF will result in non-UAPI dependency while > the other relevant virtualization APIs such as KVM and vhost_net are > UAPIs. > > Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com> If we decide to go this way, we need to make it reusable for macvtap as well. Thanks
On Tue, Sep 24, 2024 at 11:01:12AM +0200, Akihiko Odaki wrote:
> RSS is a receive steering algorithm that can be negotiated to use with
> virtio_net. Conventionally the hash calculation was done by the VMM.
> However, computing the hash after the queue was chosen defeats the
> purpose of RSS.
>
> Another approach is to use eBPF steering program. This approach has
> another downside: it cannot report the calculated hash due to the
> restrictive nature of eBPF steering program.
>
> Introduce the code to perform RSS to the kernel in order to overcome
> thse challenges. An alternative solution is to extend the eBPF steering
> program so that it will be able to report to the userspace, but I didn't
> opt for it because extending the current mechanism of eBPF steering
> program as is because it relies on legacy context rewriting, and
> introducing kfunc-based eBPF will result in non-UAPI dependency while
> the other relevant virtualization APIs such as KVM and vhost_net are
> UAPIs.
>
> Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
...
> diff --git a/drivers/net/tun.c b/drivers/net/tun.c
...
> @@ -333,8 +336,10 @@ static long tun_set_vnet_be(struct tun_struct *tun, int __user *argp)
> return -EFAULT;
>
> if (be) {
> + struct tun_vnet_hash_container *vnet_hash = rtnl_dereference(tun->vnet_hash);
> +
> if (!(tun->flags & TUN_VNET_LE) &&
> - (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT))
> + vnet_hash && (vnet_hash->flags & TUN_VNET_HASH_REPORT))
Hi Odaki-san,
I am wondering if the above should this be vnet_hash->common.flags?
I am seeing this:
../drivers/net/tun.c:342:44: error: ‘struct tun_vnet_hash_container’ has no member named ‘flags’
342 | vnet_hash && (vnet_hash->flags & TUN_VNET_HASH_REPORT))
...
On 2024/09/24 22:05, Simon Horman wrote:
> On Tue, Sep 24, 2024 at 11:01:12AM +0200, Akihiko Odaki wrote:
>> RSS is a receive steering algorithm that can be negotiated to use with
>> virtio_net. Conventionally the hash calculation was done by the VMM.
>> However, computing the hash after the queue was chosen defeats the
>> purpose of RSS.
>>
>> Another approach is to use eBPF steering program. This approach has
>> another downside: it cannot report the calculated hash due to the
>> restrictive nature of eBPF steering program.
>>
>> Introduce the code to perform RSS to the kernel in order to overcome
>> thse challenges. An alternative solution is to extend the eBPF steering
>> program so that it will be able to report to the userspace, but I didn't
>> opt for it because extending the current mechanism of eBPF steering
>> program as is because it relies on legacy context rewriting, and
>> introducing kfunc-based eBPF will result in non-UAPI dependency while
>> the other relevant virtualization APIs such as KVM and vhost_net are
>> UAPIs.
>>
>> Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
>
> ...
>
>> diff --git a/drivers/net/tun.c b/drivers/net/tun.c
>
> ...
>
>> @@ -333,8 +336,10 @@ static long tun_set_vnet_be(struct tun_struct *tun, int __user *argp)
>> return -EFAULT;
>>
>> if (be) {
>> + struct tun_vnet_hash_container *vnet_hash = rtnl_dereference(tun->vnet_hash);
>> +
>> if (!(tun->flags & TUN_VNET_LE) &&
>> - (tun->vnet_hash.flags & TUN_VNET_HASH_REPORT))
>> + vnet_hash && (vnet_hash->flags & TUN_VNET_HASH_REPORT))
>
> Hi Odaki-san,
>
> I am wondering if the above should this be vnet_hash->common.flags?
> I am seeing this:
>
> ../drivers/net/tun.c:342:44: error: ‘struct tun_vnet_hash_container’ has no member named ‘flags’
> 342 | vnet_hash && (vnet_hash->flags & TUN_VNET_HASH_REPORT))
>
> ...
You are right. I couldn't notice this error because I was testing
without CONFIG_TUN_VNET_CROSS_LE; I'll test with the configuration and
submit a new version with fix.
Regards,
Akihiko Odaki
© 2016 - 2024 Red Hat, Inc.