From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ard Biesheuvel Subject: Re: [RFC PATCH] skb: Define NET_IP_ALIGN based on CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS Date: Thu, 4 Oct 2018 19:43:59 +0200 Message-ID: References: <20181004173631.3nchegr6rm3jgz24@xylophone.i.decadent.org.uk> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Cc: "" , linux-kernel@lists.codethink.co.uk, linux-s390 , Ben Dooks , linux-arm-kernel To: Ben Hutchings , Russell King , Catalin Marinas , Will Deacon Return-path: Received: from mail-io1-f68.google.com ([209.85.166.68]:36609 "EHLO mail-io1-f68.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1727489AbeJEAiT (ORCPT ); Thu, 4 Oct 2018 20:38:19 -0400 Received: by mail-io1-f68.google.com with SMTP id p4-v6so8601584iom.3 for ; Thu, 04 Oct 2018 10:44:00 -0700 (PDT) In-Reply-To: <20181004173631.3nchegr6rm3jgz24@xylophone.i.decadent.org.uk> Sender: netdev-owner@vger.kernel.org List-ID: (+ Arnd, Russell, Catalin, Will) On 4 October 2018 at 19:36, Ben Hutchings wrote: > NET_IP_ALIGN is supposed to be defined as 0 if DMA writes to an > unaligned buffer would be more expensive than CPU access to unaligned > header fields, and otherwise defined as 2. > > Currently only ppc64 and x86 configurations define it to be 0. > However several other architectures (conditionally) define > CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, which seems to imply that > NET_IP_ALIGN should be 0. > > Remove the overriding definitions for ppc64 and x86 and define > NET_IP_ALIGN solely based on CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS. > > Signed-off-by: Ben Hutchings While this makes sense for arm64, I don't think it is appropriate for ARM per se. The unusual thing about ARM is that some instructions require 32-bit alignment even when CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is set, (i.e., load/store multiple, load/store double), and we rely on alignment fixups done by the kernel to deal with the fallout if such instructions happen to be used on unaligned quantities (Russell, please correct me if this is inaccurate) > --- > arch/powerpc/include/asm/processor.h | 11 ----------- > arch/x86/include/asm/processor.h | 8 -------- > include/linux/skbuff.h | 7 +++---- > 3 files changed, 3 insertions(+), 23 deletions(-) > > diff --git a/arch/powerpc/include/asm/processor.h b/arch/powerpc/include/asm/processor.h > index 52fadded5c1e..65c8210d2787 100644 > --- a/arch/powerpc/include/asm/processor.h > +++ b/arch/powerpc/include/asm/processor.h > @@ -525,17 +525,6 @@ extern void cvt_fd(float *from, double *to); > extern void cvt_df(double *from, float *to); > extern void _nmask_and_or_msr(unsigned long nmask, unsigned long or_val); > > -#ifdef CONFIG_PPC64 > -/* > - * We handle most unaligned accesses in hardware. On the other hand > - * unaligned DMA can be very expensive on some ppc64 IO chips (it does > - * powers of 2 writes until it reaches sufficient alignment). > - * > - * Based on this we disable the IP header alignment in network drivers. > - */ > -#define NET_IP_ALIGN 0 > -#endif > - > #endif /* __KERNEL__ */ > #endif /* __ASSEMBLY__ */ > #endif /* _ASM_POWERPC_PROCESSOR_H */ > diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h > index d53c54b842da..0108efc9726e 100644 > --- a/arch/x86/include/asm/processor.h > +++ b/arch/x86/include/asm/processor.h > @@ -33,14 +33,6 @@ struct vm86; > #include > #include > > -/* > - * We handle most unaligned accesses in hardware. On the other hand > - * unaligned DMA can be quite expensive on some Nehalem processors. > - * > - * Based on this we disable the IP header alignment in network drivers. > - */ > -#define NET_IP_ALIGN 0 > - > #define HBP_NUM 4 > /* > * Default implementation of macro that returns current > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h > index 17a13e4785fc..42467be8021f 100644 > --- a/include/linux/skbuff.h > +++ b/include/linux/skbuff.h > @@ -2435,11 +2435,10 @@ static inline int pskb_network_may_pull(struct sk_buff *skb, unsigned int len) > * The downside to this alignment of the IP header is that the DMA is now > * unaligned. On some architectures the cost of an unaligned DMA is high > * and this cost outweighs the gains made by aligning the IP header. > - * > - * Since this trade off varies between architectures, we allow NET_IP_ALIGN > - * to be overridden. > */ > -#ifndef NET_IP_ALIGN > +#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS > +#define NET_IP_ALIGN 0 > +#else > #define NET_IP_ALIGN 2 > #endif > > -- > Ben Hutchings, Software Developer Codethink Ltd > https://www.codethink.co.uk/ Dale House, 35 Dale Street > Manchester, M1 2HF, United Kingdom > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel From mboxrd@z Thu Jan 1 00:00:00 1970 From: ard.biesheuvel@linaro.org (Ard Biesheuvel) Date: Thu, 4 Oct 2018 19:43:59 +0200 Subject: [RFC PATCH] skb: Define NET_IP_ALIGN based on CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS In-Reply-To: <20181004173631.3nchegr6rm3jgz24@xylophone.i.decadent.org.uk> References: <20181004173631.3nchegr6rm3jgz24@xylophone.i.decadent.org.uk> Message-ID: To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org (+ Arnd, Russell, Catalin, Will) On 4 October 2018 at 19:36, Ben Hutchings wrote: > NET_IP_ALIGN is supposed to be defined as 0 if DMA writes to an > unaligned buffer would be more expensive than CPU access to unaligned > header fields, and otherwise defined as 2. > > Currently only ppc64 and x86 configurations define it to be 0. > However several other architectures (conditionally) define > CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, which seems to imply that > NET_IP_ALIGN should be 0. > > Remove the overriding definitions for ppc64 and x86 and define > NET_IP_ALIGN solely based on CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS. > > Signed-off-by: Ben Hutchings While this makes sense for arm64, I don't think it is appropriate for ARM per se. The unusual thing about ARM is that some instructions require 32-bit alignment even when CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is set, (i.e., load/store multiple, load/store double), and we rely on alignment fixups done by the kernel to deal with the fallout if such instructions happen to be used on unaligned quantities (Russell, please correct me if this is inaccurate) > --- > arch/powerpc/include/asm/processor.h | 11 ----------- > arch/x86/include/asm/processor.h | 8 -------- > include/linux/skbuff.h | 7 +++---- > 3 files changed, 3 insertions(+), 23 deletions(-) > > diff --git a/arch/powerpc/include/asm/processor.h b/arch/powerpc/include/asm/processor.h > index 52fadded5c1e..65c8210d2787 100644 > --- a/arch/powerpc/include/asm/processor.h > +++ b/arch/powerpc/include/asm/processor.h > @@ -525,17 +525,6 @@ extern void cvt_fd(float *from, double *to); > extern void cvt_df(double *from, float *to); > extern void _nmask_and_or_msr(unsigned long nmask, unsigned long or_val); > > -#ifdef CONFIG_PPC64 > -/* > - * We handle most unaligned accesses in hardware. On the other hand > - * unaligned DMA can be very expensive on some ppc64 IO chips (it does > - * powers of 2 writes until it reaches sufficient alignment). > - * > - * Based on this we disable the IP header alignment in network drivers. > - */ > -#define NET_IP_ALIGN 0 > -#endif > - > #endif /* __KERNEL__ */ > #endif /* __ASSEMBLY__ */ > #endif /* _ASM_POWERPC_PROCESSOR_H */ > diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h > index d53c54b842da..0108efc9726e 100644 > --- a/arch/x86/include/asm/processor.h > +++ b/arch/x86/include/asm/processor.h > @@ -33,14 +33,6 @@ struct vm86; > #include > #include > > -/* > - * We handle most unaligned accesses in hardware. On the other hand > - * unaligned DMA can be quite expensive on some Nehalem processors. > - * > - * Based on this we disable the IP header alignment in network drivers. > - */ > -#define NET_IP_ALIGN 0 > - > #define HBP_NUM 4 > /* > * Default implementation of macro that returns current > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h > index 17a13e4785fc..42467be8021f 100644 > --- a/include/linux/skbuff.h > +++ b/include/linux/skbuff.h > @@ -2435,11 +2435,10 @@ static inline int pskb_network_may_pull(struct sk_buff *skb, unsigned int len) > * The downside to this alignment of the IP header is that the DMA is now > * unaligned. On some architectures the cost of an unaligned DMA is high > * and this cost outweighs the gains made by aligning the IP header. > - * > - * Since this trade off varies between architectures, we allow NET_IP_ALIGN > - * to be overridden. > */ > -#ifndef NET_IP_ALIGN > +#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS > +#define NET_IP_ALIGN 0 > +#else > #define NET_IP_ALIGN 2 > #endif > > -- > Ben Hutchings, Software Developer Codethink Ltd > https://www.codethink.co.uk/ Dale House, 35 Dale Street > Manchester, M1 2HF, United Kingdom > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel at lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel