[53/75] patch: Adapt struct pcpu_sw_netstats change to u64_stats_t

Commit Message

Hauke Mehrtens June 27, 2024, 11:47 p.m. UTC

In kernel 6.0 the struct pcpu_sw_netstats which contains the counters
for network devices changes from attributes of type u64 to u64_stats_t.
The new types need special functions, use the old arithmetic operations
on older kernel versions.

Signed-off-by: Hauke Mehrtens <[e-mail]>
---
 patches/0111-pcpu_sw_netstats.patch | 30 +++++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)
 create mode 100644 patches/0111-pcpu_sw_netstats.patch

Comments

Johannes Berg June 28, 2024, 10:47 a.m. UTC | #1

On Fri, 2024-06-28 at 01:47 +0200, Hauke Mehrtens wrote:
> ++#if LINUX_VERSION_IS_GEQ(6,0,0)
> + 	u64_stats_inc(&stats64->rx_packets);
> + 	u64_stats_add(&stats64->rx_bytes, skb->len);
> ++#else
> ++	stats64->rx_packets++;
> ++	stats64->rx_bytes += skb->len;
> ++#endif

I suspect we _could_ do this with an auto-detect macro?

#if ... /* < 5.5 */
#define u64_stats_inc(v) (*v)++
#define u64_stats_add(v, a) (*v)+=(a)
#else
#define u64_stats_inc(v) \
 do {
	if (__builtin_types_compatible(typeof(v), *u64_stats_t)
		u64_stats_inc(v);
	else
		(*v)++;
 } while(0)


or so?

But then it's only one driver here ...

johannes

Hauke Mehrtens June 30, 2024, 8:20 p.m. UTC | #2

On 6/28/24 12:47, Johannes Berg wrote:
> On Fri, 2024-06-28 at 01:47 +0200, Hauke Mehrtens wrote:
>> ++#if LINUX_VERSION_IS_GEQ(6,0,0)
>> + 	u64_stats_inc(&stats64->rx_packets);
>> + 	u64_stats_add(&stats64->rx_bytes, skb->len);
>> ++#else
>> ++	stats64->rx_packets++;
>> ++	stats64->rx_bytes += skb->len;
>> ++#endif
> 
> I suspect we _could_ do this with an auto-detect macro?
> 
> #if ... /* < 5.5 */
> #define u64_stats_inc(v) (*v)++
> #define u64_stats_add(v, a) (*v)+=(a)
> #else
> #define u64_stats_inc(v) \
>   do {
> 	if (__builtin_types_compatible(typeof(v), *u64_stats_t)
> 		u64_stats_inc(v);
> 	else
> 		(*v)++;
>   } while(0)
> 
> 
> or so?
> 
> But then it's only one driver here ...
> 
> johannes

I think that should work because u64_stats_inc() and u64_stats_add() are 
not used in other places for now. This patch is needed because of this 
upstream change:
https://git.kernel.org/linus/9962acefbcb92736c268aafe5f52200948f60f3e
It changes the type from u64 to u64_stats_t. The functions are already 
in the kernel for much longer. Even kernel 5.10 uses them on some net 
counters. If we replace all u64_stats_inc() and u64_stats_add() usage 
with the ++ and similar operations this could also affect other places 
where we do not want to do this.

I would prefer to stay with this patch. The code in this area is also 
not changing often.

Hauke

Patch

diff --git a/patches/0111-pcpu_sw_netstats.patch b/patches/0111-pcpu_sw_netstats.patch
new file mode 100644
index 00000000..70f3fad6
--- /dev/null
+++ b/patches/0111-pcpu_sw_netstats.patch
@@ -0,0 +1,30 @@ 
+--- a/drivers/net/usb/usbnet.c
++++ b/drivers/net/usb/usbnet.c
+@@ -334,8 +334,13 @@ void usbnet_skb_return (struct usbnet *d
+ 		skb->protocol = eth_type_trans (skb, dev->net);
+ 
+ 	flags = u64_stats_update_begin_irqsave(&stats64->syncp);
++#if LINUX_VERSION_IS_GEQ(6,0,0)
+ 	u64_stats_inc(&stats64->rx_packets);
+ 	u64_stats_add(&stats64->rx_bytes, skb->len);
++#else
++	stats64->rx_packets++;
++	stats64->rx_bytes += skb->len;
++#endif
+ 	u64_stats_update_end_irqrestore(&stats64->syncp, flags);
+ 
+ 	netif_dbg(dev, rx_status, dev->net, "< rx, len %zu, type 0x%x\n",
+@@ -1253,8 +1258,13 @@ static void tx_complete (struct urb *urb
+ 		unsigned long flags;
+ 
+ 		flags = u64_stats_update_begin_irqsave(&stats64->syncp);
++#if LINUX_VERSION_IS_GEQ(6,0,0)
+ 		u64_stats_add(&stats64->tx_packets, entry->packets);
+ 		u64_stats_add(&stats64->tx_bytes, entry->length);
++#else
++		stats64->tx_packets += entry->packets;
++		stats64->tx_bytes += entry->length;
++#endif
+ 		u64_stats_update_end_irqrestore(&stats64->syncp, flags);
+ 	} else {
+ 		dev->net->stats.tx_errors++;