Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1400115 > unrolled thread

[PATCH] arm64: Implement optimised IP checksum helpers

Back to article view | Back to linux.kernel

Contents

  [PATCH] arm64: Implement optimised IP checksum helpers Robin Murphy <robin.murphy@arm.com> - 2016-05-12 16:30 +0200
    Re: [PATCH] arm64: Implement optimised IP checksum helpers Luke Starrett <luke.starrett@broadcom.com> - 2016-05-12 17:40 +0200
      Re: [PATCH] arm64: Implement optimised IP checksum helpers Robin Murphy <robin.murphy@arm.com> - 2016-05-12 19:10 +0200
        Re: [PATCH] arm64: Implement optimised IP checksum helpers Luke Starrett <luke.starrett@broadcom.com> - 2016-05-12 20:40 +0200

#1400115 — [PATCH] arm64: Implement optimised IP checksum helpers

AArch64 is capable of 128-bit memory accesses without alignment
restrictions, which makes it both possible and highly practical to slurp
up a typical 20-byte IP header in just 2 loads. Implement our own
version of ip_fast_checksum() to take advantage of that, resulting in
considerably fewer instructions and memory accesses than the generic
version. We can also get more optimal code generation for csum_fold() by
defining it a slightly different way round from the generic version, so
throw that into the mix too.

Suggested-by: Luke Starrett <luke.starrett@broadcom.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---

Having spent an evening poring over disassembler output, it turns out
that there's really no need to drop into asm for this, hurrah! I've
taken it for a spin on both big and little-endian, with no ill effects.

 arch/arm64/include/asm/checksum.h | 49 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 arch/arm64/include/asm/checksum.h

diff --git a/arch/arm64/include/asm/checksum.h b/arch/arm64/include/asm/checksum.h
new file mode 100644
index 000000000000..7e975ffce837
--- /dev/null
+++ b/arch/arm64/include/asm/checksum.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (C) 2016 ARM Ltd.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+#ifndef __ASM_CHECKSUM_H
+#define __ASM_CHECKSUM_H
+
+static inline __sum16 csum_fold(__wsum csum)
+{
+	u32 sum = (__force u32)csum;
+	sum += (sum >> 16) | (sum << 16);
+	return ~(__force __sum16)(sum >> 16);
+}
+#define csum_fold csum_fold
+
+static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
+{
+	__uint128_t tmp;
+	u64 sum;
+
+	tmp = *(const __uint128_t *)iph;
+	iph += 16;
+	ihl -= 4;
+	tmp += ((tmp >> 64) | (tmp << 64));
+	sum = tmp >> 64;
+	do {
+		sum += *(const u32 *)iph;
+		iph += 4;
+	} while (--ihl);
+
+	sum += ((sum >> 32) | (sum << 32));
+	return csum_fold(sum >> 32);
+}
+#define ip_fast_csum ip_fast_csum
+
+#include <asm-generic/checksum.h>
+
+#endif	/* __ASM_CHECKSUM_H */
-- 
2.8.1.dirty

[toc] | [next] | [standalone]

#1400161

Hi Robin,

I pulled this in to a userspace test app expecting that the __uint128_t 
type might cause GCC to emit 'ldp'.  Seems like that was that your 
intent based on your commit note.  Instead I see two 64b loads (ldr Xn), 
and a single 32b load (ldr Wn) for the trailing 4B.  This was with 
Linaro GCC 4.9-2015.06.

Otherwise, the C cycle count looks good enough compared to the asm version.

Thanks,

Luke


On 5/12/2016 10:26 AM, Robin Murphy wrote:
> AArch64 is capable of 128-bit memory accesses without alignment
> restrictions, which makes it both possible and highly practical to slurp
> up a typical 20-byte IP header in just 2 loads. Implement our own
> version of ip_fast_checksum() to take advantage of that, resulting in
> considerably fewer instructions and memory accesses than the generic
> version. We can also get more optimal code generation for csum_fold() by
> defining it a slightly different way round from the generic version, so
> throw that into the mix too.
>
> Suggested-by: Luke Starrett <luke.starrett@broadcom.com>
> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
> ---
>
> Having spent an evening poring over disassembler output, it turns out
> that there's really no need to drop into asm for this, hurrah! I've
> taken it for a spin on both big and little-endian, with no ill effects.
>
>   arch/arm64/include/asm/checksum.h | 49 +++++++++++++++++++++++++++++++++++++++
>   1 file changed, 49 insertions(+)
>   create mode 100644 arch/arm64/include/asm/checksum.h
>
> diff --git a/arch/arm64/include/asm/checksum.h b/arch/arm64/include/asm/checksum.h
> new file mode 100644
> index 000000000000..7e975ffce837
> --- /dev/null
> +++ b/arch/arm64/include/asm/checksum.h
> @@ -0,0 +1,49 @@
> +/*
> + * Copyright (C) 2016 ARM Ltd.
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2 as
> + * published by the Free Software Foundation.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> + * GNU General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public License
> + * along with this program.  If not, see <http://www.gnu.org/licenses/>.
> + */
> +#ifndef __ASM_CHECKSUM_H
> +#define __ASM_CHECKSUM_H
> +
> +static inline __sum16 csum_fold(__wsum csum)
> +{
> +	u32 sum = (__force u32)csum;
> +	sum += (sum >> 16) | (sum << 16);
> +	return ~(__force __sum16)(sum >> 16);
> +}
> +#define csum_fold csum_fold
> +
> +static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
> +{
> +	__uint128_t tmp;
> +	u64 sum;
> +
> +	tmp = *(const __uint128_t *)iph;
> +	iph += 16;
> +	ihl -= 4;
> +	tmp += ((tmp >> 64) | (tmp << 64));
> +	sum = tmp >> 64;
> +	do {
> +		sum += *(const u32 *)iph;
> +		iph += 4;
> +	} while (--ihl);
> +
> +	sum += ((sum >> 32) | (sum << 32));
> +	return csum_fold(sum >> 32);
> +}
> +#define ip_fast_csum ip_fast_csum
> +
> +#include <asm-generic/checksum.h>
> +
> +#endif	/* __ASM_CHECKSUM_H */

[toc] | [prev] | [next] | [standalone]

#1400246

Hi Luke,

On 12/05/16 16:34, Luke Starrett wrote:
> Hi Robin,
>
> I pulled this in to a userspace test app expecting that the __uint128_t
> type might cause GCC to emit 'ldp'.  Seems like that was that your
> intent based on your commit note.  Instead I see two 64b loads (ldr Xn),
> and a single 32b load (ldr Wn) for the trailing 4B.  This was with
> Linaro GCC 4.9-2015.06.

GCC 5 happily emits ldp there, but indeed I couldn't figure out how to 
convince GCC 4 to do so. From a quick ferret around in the GCC Git, it 
looks like the relevant optimisations may have only gone in post-4.9.

> Otherwise, the C cycle count looks good enough compared to the asm version.

Yeah, compiling as standalone functions with GCC 5 I get 19 instructions 
vs. 17 for the asm, but the loop logic gets optimised out completely 
when ihl is a compile-time constant (e.g. inet_gro_receive())

Cheers,
Robin.

>
> Thanks,
>
> Luke
>
>
> On 5/12/2016 10:26 AM, Robin Murphy wrote:
>> AArch64 is capable of 128-bit memory accesses without alignment
>> restrictions, which makes it both possible and highly practical to slurp
>> up a typical 20-byte IP header in just 2 loads. Implement our own
>> version of ip_fast_checksum() to take advantage of that, resulting in
>> considerably fewer instructions and memory accesses than the generic
>> version. We can also get more optimal code generation for csum_fold() by
>> defining it a slightly different way round from the generic version, so
>> throw that into the mix too.
>>
>> Suggested-by: Luke Starrett <luke.starrett@broadcom.com>
>> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
>> ---
>>
>> Having spent an evening poring over disassembler output, it turns out
>> that there's really no need to drop into asm for this, hurrah! I've
>> taken it for a spin on both big and little-endian, with no ill effects.
>>
>>   arch/arm64/include/asm/checksum.h | 49
>> +++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 49 insertions(+)
>>   create mode 100644 arch/arm64/include/asm/checksum.h
>>
>> diff --git a/arch/arm64/include/asm/checksum.h
>> b/arch/arm64/include/asm/checksum.h
>> new file mode 100644
>> index 000000000000..7e975ffce837
>> --- /dev/null
>> +++ b/arch/arm64/include/asm/checksum.h
>> @@ -0,0 +1,49 @@
>> +/*
>> + * Copyright (C) 2016 ARM Ltd.
>> + *
>> + * This program is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License version 2 as
>> + * published by the Free Software Foundation.
>> + *
>> + * This program is distributed in the hope that it will be useful,
>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>> + * GNU General Public License for more details.
>> + *
>> + * You should have received a copy of the GNU General Public License
>> + * along with this program.  If not, see <http://www.gnu.org/licenses/>.
>> + */
>> +#ifndef __ASM_CHECKSUM_H
>> +#define __ASM_CHECKSUM_H
>> +
>> +static inline __sum16 csum_fold(__wsum csum)
>> +{
>> +    u32 sum = (__force u32)csum;
>> +    sum += (sum >> 16) | (sum << 16);
>> +    return ~(__force __sum16)(sum >> 16);
>> +}
>> +#define csum_fold csum_fold
>> +
>> +static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
>> +{
>> +    __uint128_t tmp;
>> +    u64 sum;
>> +
>> +    tmp = *(const __uint128_t *)iph;
>> +    iph += 16;
>> +    ihl -= 4;
>> +    tmp += ((tmp >> 64) | (tmp << 64));
>> +    sum = tmp >> 64;
>> +    do {
>> +        sum += *(const u32 *)iph;
>> +        iph += 4;
>> +    } while (--ihl);
>> +
>> +    sum += ((sum >> 32) | (sum << 32));
>> +    return csum_fold(sum >> 32);
>> +}
>> +#define ip_fast_csum ip_fast_csum
>> +
>> +#include <asm-generic/checksum.h>
>> +
>> +#endif    /* __ASM_CHECKSUM_H */
>

[toc] | [prev] | [next] | [standalone]

#1400282

Hi Robin,

On 5/12/2016 1:08 PM, Robin Murphy wrote:
> Hi Luke,
>
> On 12/05/16 16:34, Luke Starrett wrote:
>> Hi Robin,
>>
>> I pulled this in to a userspace test app expecting that the __uint128_t
>> type might cause GCC to emit 'ldp'.  Seems like that was that your
>> intent based on your commit note.  Instead I see two 64b loads (ldr Xn),
>> and a single 32b load (ldr Wn) for the trailing 4B.  This was with
>> Linaro GCC 4.9-2015.06.
>
> GCC 5 happily emits ldp there, but indeed I couldn't figure out how to 
> convince GCC 4 to do so. From a quick ferret around in the GCC Git, it 
> looks like the relevant optimisations may have only gone in post-4.9.
>
Not a problem.  I was just curious for my own selfish reasons.

>> Otherwise, the C cycle count looks good enough compared to the asm 
>> version.
>
> Yeah, compiling as standalone functions with GCC 5 I get 19 
> instructions vs. 17 for the asm, but the loop logic gets optimised out 
> completely when ihl is a compile-time constant (e.g. inet_gro_receive())
>

I updated to Linaro GCC 5.3-2016.02, and saw what you described.  I ran 
some smoke testing against a random header generator.  LGTM.

Acked-by: Luke Starrett <luke.starrett@broadcom.com>

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web