[PATCH] net: core: alloc_netdev_mqs reduce alignment overhead

[PATCH] net: core: alloc_netdev_mqs reduce alignment overhead

[Chris Gregory]

I was reading over alloc_netdev_mqs and noticed that it's trying to
align the allocation by adding on 31 bytes (NETDEV_ALIGN-1).  I'm not
an expert on kmalloc, but AFAICT it will always give back a pointer
aligned to ARCH_KMALLOC_MINALIGN.  On my x86_64,
ARCH_KMALLOC_MINALIGN=8 so this saves 7 bytes which could mean an
8-byte allocation is freed up.

I think there's some additional potential optimization for the case
where ARCH_KMALLOC_MINALIGN >= NETDEV_ALIGN.  In this case we can
delete the padding member completely.  I don't know if this ever
happens in practice though.

If this patch looks good, we should probably do the same thing at
linux/drivers/net/ethernet/wiznet/w5100.c:1091.

Chris
----

diff --git a/net/core/dev.c b/net/core/dev.c
index e1bb6d7856d9..73daa1573554 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -10943,7 +10943,7 @@ struct net_device *alloc_netdev_mqs(int
sizeof_priv, const char *name,
         alloc_size += sizeof_priv;
     }
     /* ensure 32-byte alignment of whole construct */
-    alloc_size += NETDEV_ALIGN - 1;
+    alloc_size += NETDEV_ALIGN - min(NETDEV_ALIGN, ARCH_KMALLOC_MINALIGN);

     p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
     if (!p)

Re: [PATCH] net: core: alloc_netdev_mqs reduce alignment overhead

[Jakub Kicinski]

On Mon, 24 Jun 2024 03:58:04 -0400 Chris Gregory wrote:
> I was reading over alloc_netdev_mqs and noticed that it's trying to
> align the allocation by adding on 31 bytes (NETDEV_ALIGN-1).  I'm not
> an expert on kmalloc, but AFAICT it will always give back a pointer
> aligned to ARCH_KMALLOC_MINALIGN.  On my x86_64,
> ARCH_KMALLOC_MINALIGN=8 so this saves 7 bytes which could mean an
> 8-byte allocation is freed up.

We're reworking the priv handling to use a flexible array:
https://lore.kernel.org/all/20240229225910.79e224cf@kernel.org/
one more series of prep needs to get merged, and we'll convert.
-- 
pw-bot: reject