Implement assembly routine for csum_partial for 64 bit x86. This
primarily speeds up checksum calculation for smaller lengths such as
those that are present when doing skb_postpull_rcsum when getting
CHECKSUM_COMPLETE from device or after CHECKSUM_UNNECESSARY
conversion.
CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is checked to determine whether
we need to avoid unaligned accesses. Efficient unaligned accesses
offer a nice additional speedup as demonstrated in the results
provided below.
This implementation is similar to csum_partial implemented in
checksum_32.S, however since we are dealing with 8 bytes at a time
there are more cases for small lengths (and alignments)-- for that
we employ a jump table.
Testing:
Correctness:
Verified correctness by testing arbitrary length buffer filled with
random data. For each buffer I compared the computed checksum
using the original algorithm for each possible alignment (0-7 bytes).
Checksum performance:
Isolating old and new implementation for some common cases:
Old NewA NewA % NewNoA NewNoA %
Len/Aln nsec nsec Improv nsecs Improve
--------+-------+--------+-------+-------+---------------------
1400/0 192.9 175.1 10% 174.9 10% (Big packet)
40/0 13.8 7.7 44% 5.7 58% (Ipv6 hdr cmn case)
8/4 8.4 6.9 18% 2.8 67% (UDP, VXLAN in IPv4)
14/0 10.5 7.3 30% 5.4 48% (Eth hdr)
14/4 10.8 8.7 19% 5.4 50% (Eth hdr in IPv4)
14/3 11.0 9.8 11% 5.6 49% (Eth with odd align)
7/1 10.0 5.8 42% 4.8 52% (buffer in one quad)
NewA=>CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS not set
NewNoA=>CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is set
Results from: Intel(R) Xeon(R) CPU X5650 @ 2.67GHz
Also test on these with similar results:
Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz
Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
Branch prediction:
To test the effects of poor branch prediction in the jump tables I
tested checksum performance with runs for two combinations of length
and alignment. As the baseline I performed the test by doing half of
calls with the first combination, followed by using the second
combination for the second half. In the test case, I interleave the
two combinations so that in every call the length and alignment are
different to defeat the effects of branch prediction. Running several
cases, I did not see any material performance difference between
the baseline and the interleaving test case.
Signed-off-by: Tom Herbert <t...@herbertland.com>
---
arch/x86/include/asm/checksum_64.h | 5 +
arch/x86/lib/csum-partial_64.S | 277 +++++++++++++++++++++++++++++++++++++
arch/x86/lib/csum-partial_64.c | 148 --------------------
3 files changed, 282 insertions(+), 148 deletions(-)
create mode 100644 arch/x86/lib/csum-partial_64.S
delete mode 100644 arch/x86/lib/csum-partial_64.c
diff --git a/arch/x86/include/asm/checksum_64.h
b/arch/x86/include/asm/checksum_64.h
index cd00e17..a888f65 100644
--- a/arch/x86/include/asm/checksum_64.h
+++ b/arch/x86/include/asm/checksum_64.h
@@ -128,6 +128,11 @@ static inline __sum16 csum_tcpudp_magic(__be32 saddr,
__be32 daddr,
*/
extern __wsum csum_partial(const void *buff, int len, __wsum sum);
+static inline __sum16 ip_compute_csum(const void *buff, int len)
+{
+ return csum_fold(csum_partial(buff, len, 0));
+}
+
#define _HAVE_ARCH_COPY_AND_CSUM_FROM_USER 1
#define HAVE_CSUM_COPY_USER 1
diff --git a/arch/x86/lib/csum-partial_64.S b/arch/x86/lib/csum-partial_64.S
new file mode 100644
index 0000000..520b400
--- /dev/null
+++ b/arch/x86/lib/csum-partial_64.S
@@ -0,0 +1,277 @@
+/* Copyright 2016 Tom Herbert <t...@herbertland.com>
+ *
+ * Checksum partial calculation
+ *
+ * __wsum csum_partial(const void *buff, int len, __wsum sum)
+ *
+ * Computes the checksum of a memory block at buff, length len,
+ * and adds in "sum" (32-bit)
+ *
+ * Returns a 32-bit number suitable for feeding into itself
+ * or csum_tcpudp_magic
+ *
+ * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS determines whether alignment of the
+ * buffer must be dealt with.
+ *
+ * If CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is set then the steps are:
+ * 1) Initialize accumulator to initial sum
+ * 2) Sum 8 bytes at a time using adcq (unroll main loop
+ * to do 128 bytes at a time)
+ * 3) Sum remaining length (less than 8 bytes)
+ *
+ * If CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is not set then the steps are:
+ * 1) Handle buffer that is not aligned to 8 bytes, sum up to 8 byte
+ * alignment
+ * 2) Sum 8 bytes at a time using adcq (unroll main loop
+ * to do 128 bytes at a time)
+ * 3) Sum remaining length (less than 8 bytes)
+ * 4) Roll result if alignment is odd and add in initial sum argument
+ * 5) If buffer is not aligned to 8 bytes and length is less than
+ * or equal to 8 - alignment (whole buffer is in one quad), then
+ * treat that as a special case.
+ *
+ * Register usage:
+ * %rdi: argument #1, buff
+ * %rsi: argument #2, length
+ * %rdx: argument #3, add in value
+ * %rax,%eax: accumulator and return value
+ * %rcx,%ecx: counter and tmp
+ * %r11: tmp
+ * %r10: alignment (0-7) - when CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is set
+ */
+
+#include <linux/linkage.h>
+#include <asm/errno.h>
+#include <asm/asm.h>
+
+#define branch_tbl_len .L_branch_tbl_len
+
+#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
+
+/* Close the carry chain and return. */
+#define RETURN \
+ adcl $0, %eax; \
+ ret
+
+#else /* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
+
+/* Before returning need to roll the result if alignment was odd and then add
+ * in the initial sum.
+ */
+#define RETURN \
+ adcl $0, %eax; \
+ test $0x1, %r10d; \
+ jz 99f; \
+ roll $8, %eax; \
+99: addl %edx, %eax; \
+ adcl $0, %eax; \
+ ret
+
+#define branch_tbl_align .L_branch_tbl_align
+
+#endif /* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
+
+ENTRY(csum_partial)
+
+#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
+ movl %edx, %eax /* Initialize with initial sum argument */
+#else /* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
+ test %esi, %esi /* Zero length? */
+ jne 310f
+ movl %edx, %eax
+ ret
+
+310: xorl %eax, %eax
+
+ /* Determine alignment */
+ movl %edi, %r10d
+ andl $0x7, %r10d
+ jz 10f
+ movl $8, %ecx
+ subl %r10d, %ecx
+ cmpl %ecx, %esi
+ jle 320f
+ clc
+ jmpq *branch_tbl_align(, %r10, 8)
+
+ /* Whole buffer fits into one quad. Sum up to a four byte alignment
+ * and then call into the length table to finish.
+ */
+320: test $0x1, %r10d
+ jz 330f
+ movb (%rdi), %ah /* Align to two bytes */
+ decl %esi
+ lea 1(%rdi), %rdi
+330: cmpl $2, %esi
+ jl 340f
+ test $0x2, %r10d
+ jz 340f
+ addw (%rdi), %ax /* Align to four bytes */
+ adcl $0, %eax
+ lea 2(%rdi), %rdi
+ subl $2, %esi
+340:
+ clc
+ jmpq *branch_tbl_len(, %rsi, 8)
+
+/* Jumps table for alignments */
+
+201: /* Align 1 */
+ adcw 5(%rdi), %ax
+203: /* Align 3 */
+ adcw 3(%rdi), %ax
+205: /* Align 5 */
+ adcw 1(%rdi), %ax
+207: /* Align 7 */
+ adcl $0, %eax
+ addb (%rdi), %ah
+ jmp 222f
+202: /* Align 2 */
+ adcw 4(%rdi), %ax
+204: /* Align 4 */
+ adcw 2(%rdi), %ax
+206: /* Align 6 */
+ adcw (%rdi), %ax
+
+222: adcl $0, %eax
+ subl %ecx, %esi /* %rcx is 8 - alignment */
+ addq %rcx, %rdi
+200:
+ /* Fall through */
+
+#endif /* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
+
+ /* Check length */
+10: cmpl $8, %esi
+ jg 30f
+ jl 20f
+
+ /* Exactly 8 bytes length */
+ addl (%rdi), %eax
+ adcl 4(%rdi), %eax
+ RETURN
+
+ /* Less than 8 bytes length */
+20: clc
+ jmpq *branch_tbl_len(, %rsi, 8)
+
+ /* Greater than 8 bytes length. Determine number of quads (n). Sum
+ * over first n % 16 quads
+ */
+30: movl %esi, %ecx
+ shrl $3, %ecx
+ andl $0xf, %ecx
+ negq %rcx
+ lea 40f(, %rcx, 4), %r11
+ clc
+ jmp *%r11
+
+.align 8
+ adcq 14*8(%rdi),%rax
+ adcq 13*8(%rdi),%rax
+ adcq 12*8(%rdi),%rax
+ adcq 11*8(%rdi),%rax
+ adcq 10*8(%rdi),%rax
+ adcq 9*8(%rdi),%rax
+ adcq 8*8(%rdi),%rax
+ adcq 7*8(%rdi),%rax
+ adcq 6*8(%rdi),%rax
+ adcq 5*8(%rdi),%rax
+ adcq 4*8(%rdi),%rax
+ adcq 3*8(%rdi),%rax
+ adcq 2*8(%rdi),%rax
+ adcq 1*8(%rdi),%rax
+ adcq 0*8(%rdi),%rax
+ nop
+40: /* #quads % 16 jump table base */
+
+ adcq $0, %rax
+ shlq $3, %rcx
+ subq %rcx, %rdi /* %rcx is already negative length */
+
+ /* Now determine number of blocks of 8 quads. Sum 128 bytes at a time
+ * using unrolled loop.
+ */
+ movl %esi, %ecx
+ shrl $7, %ecx
+ jz 60f
+ clc
+
+ /* Main loop */
+50: adcq 0*8(%rdi),%rax
+ adcq 1*8(%rdi),%rax
+ adcq 2*8(%rdi),%rax
+ adcq 3*8(%rdi),%rax
+ adcq 4*8(%rdi),%rax
+ adcq 5*8(%rdi),%rax
+ adcq 6*8(%rdi),%rax
+ adcq 7*8(%rdi),%rax
+ adcq 8*8(%rdi),%rax
+ adcq 9*8(%rdi),%rax
+ adcq 10*8(%rdi),%rax
+ adcq 11*8(%rdi),%rax
+ adcq 12*8(%rdi),%rax
+ adcq 13*8(%rdi),%rax
+ adcq 14*8(%rdi),%rax
+ adcq 15*8(%rdi),%rax
+ lea 128(%rdi), %rdi
+ loop 50b
+
+ adcq $0, %rax
+
+ /* Handle remaining length which is <= 8 bytes */
+60: andl $0x7, %esi
+
+ /* Fold 64 bit sum to 32 bits */
+ movq %rax, %rcx
+ shrq $32, %rcx
+ addl %ecx, %eax
+
+ jmpq *branch_tbl_len(, %rsi, 8)
+
+/* Length table targets */
+
+107: /* Length 7 */
+ adcw 4(%rdi), %ax
+105: /* Length 5 */
+ adcw 2(%rdi), %ax
+103: /* Length 3 */
+ adcw (%rdi), %ax
+101: /* Length 1, grab the odd byte */
+ adcb -1(%rdi, %rsi), %al
+ adcb $0, %ah
+ RETURN
+106: /* Length 6 */
+ adcw 4(%rdi), %ax
+104: /* Length 4, optimized for double word access*/
+ adcl (%rdi), %eax
+ RETURN
+102: /* Length 2 */
+ adcw (%rdi), %ax
+100: /* Length 0 */
+ RETURN
+
+.section .rodata
+.align 64
+.L_branch_tbl_len:
+ .quad 100b
+ .quad 101b
+ .quad 102b
+ .quad 103b
+ .quad 104b
+ .quad 105b
+ .quad 106b
+ .quad 107b
+
+#ifndef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
+.L_branch_tbl_align:
+ .quad 200b
+ .quad 201b
+ .quad 202b
+ .quad 203b
+ .quad 204b
+ .quad 205b
+ .quad 206b
+ .quad 207b
+#endif
+
diff --git a/arch/x86/lib/csum-partial_64.c b/arch/x86/lib/csum-partial_64.c
deleted file mode 100644
index 9845371..0000000
--- a/arch/x86/lib/csum-partial_64.c
+++ /dev/null
@@ -1,148 +0,0 @@
-/*
- * arch/x86_64/lib/csum-partial.c
- *
- * This file contains network checksum routines that are better done
- * in an architecture-specific manner due to speed.
- */
-
-#include <linux/compiler.h>
-#include <linux/module.h>
-#include <asm/checksum.h>
-
-static inline unsigned short from32to16(unsigned a)
-{
- unsigned short b = a >> 16;
- asm("addw %w2,%w0\n\t"
- "adcw $0,%w0\n"
- : "=r" (b)
- : "0" (b), "r" (a));
- return b;
-}
-
-/*
- * Do a 64-bit checksum on an arbitrary memory area.
- * Returns a 32bit checksum.
- *
- * This isn't as time critical as it used to be because many NICs
- * do hardware checksumming these days.
- *
- * Things tried and found to not make it faster:
- * Manual Prefetching
- * Unrolling to an 128 bytes inner loop.
- * Using interleaving with more registers to break the carry chains.
- */
-static unsigned do_csum(const unsigned char *buff, unsigned len)
-{
- unsigned odd, count;
- unsigned long result = 0;
-
- if (unlikely(len == 0))
- return result;
- odd = 1 & (unsigned long) buff;
- if (unlikely(odd)) {
- result = *buff << 8;
- len--;
- buff++;
- }
- count = len >> 1; /* nr of 16-bit words.. */
- if (count) {
- if (2 & (unsigned long) buff) {
- result += *(unsigned short *)buff;
- count--;
- len -= 2;
- buff += 2;
- }
- count >>= 1; /* nr of 32-bit words.. */
- if (count) {
- unsigned long zero;
- unsigned count64;
- if (4 & (unsigned long) buff) {
- result += *(unsigned int *) buff;
- count--;
- len -= 4;
- buff += 4;
- }
- count >>= 1; /* nr of 64-bit words.. */
-
- /* main loop using 64byte blocks */
- zero = 0;
- count64 = count >> 3;
- while (count64) {
- asm("addq 0*8(%[src]),%[res]\n\t"
- "adcq 1*8(%[src]),%[res]\n\t"
- "adcq 2*8(%[src]),%[res]\n\t"
- "adcq 3*8(%[src]),%[res]\n\t"
- "adcq 4*8(%[src]),%[res]\n\t"
- "adcq 5*8(%[src]),%[res]\n\t"
- "adcq 6*8(%[src]),%[res]\n\t"
- "adcq 7*8(%[src]),%[res]\n\t"
- "adcq %[zero],%[res]"
- : [res] "=r" (result)
- : [src] "r" (buff), [zero] "r" (zero),
- "[res]" (result));
- buff += 64;
- count64--;
- }
-
- /* last up to 7 8byte blocks */
- count %= 8;
- while (count) {
- asm("addq %1,%0\n\t"
- "adcq %2,%0\n"
- : "=r" (result)
- : "m" (*(unsigned long *)buff),
- "r" (zero), "0" (result));
- --count;
- buff += 8;
- }
- result = add32_with_carry(result>>32,
- result&0xffffffff);
-
- if (len & 4) {
- result += *(unsigned int *) buff;
- buff += 4;
- }
- }
- if (len & 2) {
- result += *(unsigned short *) buff;
- buff += 2;
- }
- }
- if (len & 1)
- result += *buff;
- result = add32_with_carry(result>>32, result & 0xffffffff);
- if (unlikely(odd)) {
- result = from32to16(result);
- result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
- }
- return result;
-}
-
-/*
- * computes the checksum of a memory block at buff, length len,
- * and adds in "sum" (32-bit)
- *
- * returns a 32-bit number suitable for feeding into itself
- * or csum_tcpudp_magic
- *
- * this function must be called with even lengths, except
- * for the last fragment, which may be odd
- *
- * it's best to have buff aligned on a 64-bit boundary
- */
-__wsum csum_partial(const void *buff, int len, __wsum sum)
-{
- return (__force __wsum)add32_with_carry(do_csum(buff, len),
- (__force u32)sum);
-}
-
-/*
- * this routine is used for miscellaneous IP-like checksums, mainly
- * in icmp.c
- */
-__sum16 ip_compute_csum(const void *buff, int len)
-{
- return csum_fold(csum_partial(buff,len,0));
-}
-EXPORT_SYMBOL(ip_compute_csum);
-
--
2.4.6
