DPDK patches and discussions
 help / color / mirror / Atom feed
From: Jerin Jacob <jerin.jacob@caviumnetworks.com>
To: <dev@dpdk.org>
Subject: [dpdk-dev] [PATCH 2/3] arm64: acl: add neon based acl implementation
Date: Mon, 2 Nov 2015 19:48:40 +0530	[thread overview]
Message-ID: <1446473921-12706-3-git-send-email-jerin.jacob@caviumnetworks.com> (raw)
In-Reply-To: <1446473921-12706-2-git-send-email-jerin.jacob@caviumnetworks.com>

Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
---
 app/test-acl/main.c           |   4 +
 lib/librte_acl/Makefile       |   5 +
 lib/librte_acl/acl.h          |   4 +
 lib/librte_acl/acl_run_neon.c |  46 +++++++
 lib/librte_acl/acl_run_neon.h | 290 ++++++++++++++++++++++++++++++++++++++++++
 lib/librte_acl/rte_acl.c      |  25 ++++
 lib/librte_acl/rte_acl.h      |   1 +
 7 files changed, 375 insertions(+)
 create mode 100644 lib/librte_acl/acl_run_neon.c
 create mode 100644 lib/librte_acl/acl_run_neon.h

diff --git a/app/test-acl/main.c b/app/test-acl/main.c
index 72ce83c..0b0c093 100644
--- a/app/test-acl/main.c
+++ b/app/test-acl/main.c
@@ -101,6 +101,10 @@ static const struct acl_alg acl_alg[] = {
 		.name = "avx2",
 		.alg = RTE_ACL_CLASSIFY_AVX2,
 	},
+	{
+		.name = "neon",
+		.alg = RTE_ACL_CLASSIFY_NEON,
+	},
 };
 
 static struct {
diff --git a/lib/librte_acl/Makefile b/lib/librte_acl/Makefile
index 7a1cf8a..27f91d5 100644
--- a/lib/librte_acl/Makefile
+++ b/lib/librte_acl/Makefile
@@ -48,9 +48,14 @@ SRCS-$(CONFIG_RTE_LIBRTE_ACL) += rte_acl.c
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_bld.c
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_gen.c
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_scalar.c
+ifeq ($(CONFIG_RTE_ARCH_ARM64),y)
+SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_neon.c
+else
 SRCS-$(CONFIG_RTE_LIBRTE_ACL) += acl_run_sse.c
+endif
 
 CFLAGS_acl_run_sse.o += -msse4.1
+CFLAGS_acl_run_neon.o += -flax-vector-conversions -Wno-maybe-uninitialized
 
 #
 # If the compiler supports AVX2 instructions,
diff --git a/lib/librte_acl/acl.h b/lib/librte_acl/acl.h
index eb4930c..09d6784 100644
--- a/lib/librte_acl/acl.h
+++ b/lib/librte_acl/acl.h
@@ -230,6 +230,10 @@ int
 rte_acl_classify_avx2(const struct rte_acl_ctx *ctx, const uint8_t **data,
 	uint32_t *results, uint32_t num, uint32_t categories);
 
+int
+rte_acl_classify_neon(const struct rte_acl_ctx *ctx, const uint8_t **data,
+	uint32_t *results, uint32_t num, uint32_t categories);
+
 #ifdef __cplusplus
 }
 #endif /* __cplusplus */
diff --git a/lib/librte_acl/acl_run_neon.c b/lib/librte_acl/acl_run_neon.c
new file mode 100644
index 0000000..b014451
--- /dev/null
+++ b/lib/librte_acl/acl_run_neon.c
@@ -0,0 +1,46 @@
+/*
+ *   BSD LICENSE
+ *
+ *   Copyright (C) Cavium networks Ltd. 2015.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Cavium networks nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include "acl_run_neon.h"
+
+int
+rte_acl_classify_neon(const struct rte_acl_ctx *ctx, const uint8_t **data,
+		      uint32_t *results, uint32_t num, uint32_t categories)
+{
+	if (likely(num >= 8))
+		return search_neon_8(ctx, data, results, num, categories);
+	else if (num >= 4)
+		return search_neon_4(ctx, data, results, num, categories);
+	else
+		return rte_acl_classify_scalar(ctx, data, results, num,
+			categories);
+}
diff --git a/lib/librte_acl/acl_run_neon.h b/lib/librte_acl/acl_run_neon.h
new file mode 100644
index 0000000..4579476
--- /dev/null
+++ b/lib/librte_acl/acl_run_neon.h
@@ -0,0 +1,290 @@
+/*
+ *   BSD LICENSE
+ *
+ *   Copyright (C) Cavium networks Ltd. 2015.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Cavium networks nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include "acl_run.h"
+#include "acl_vect.h"
+
+struct _neon_acl_const {
+	rte_xmm_t xmm_shuffle_input;
+	rte_xmm_t xmm_index_mask;
+	rte_xmm_t range_base;
+} neon_acl_const  __attribute__((aligned(RTE_CACHE_LINE_SIZE))) = {
+	{
+		.u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c}
+	},
+	{
+		.u32 = {RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX,
+		RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX}
+	},
+	{
+		.u32 = {0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c}
+	},
+};
+
+/*
+ * Resolve priority for multiple results (neon version).
+ * This consists comparing the priority of the current traversal with the
+ * running set of results for the packet.
+ * For each result, keep a running array of the result (rule number) and
+ * its priority for each category.
+ */
+static inline void
+resolve_priority_neon(uint64_t transition, int n, const struct rte_acl_ctx *ctx,
+		      struct parms *parms,
+		      const struct rte_acl_match_results *p,
+		      uint32_t categories)
+{
+	uint32_t x;
+	int32x4_t results, priority, results1, priority1;
+	uint32x4_t selector;
+	int32_t *saved_results, *saved_priority;
+
+	for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
+		saved_results = (int32_t *)(&parms[n].cmplt->results[x]);
+		saved_priority = (int32_t *)(&parms[n].cmplt->priority[x]);
+
+		/* get results and priorities for completed trie */
+		results = vld1q_s32(
+			(const int32_t *)&p[transition].results[x]);
+		priority = vld1q_s32(
+			(const int32_t *)&p[transition].priority[x]);
+
+		/* if this is not the first completed trie */
+		if (parms[n].cmplt->count != ctx->num_tries) {
+			/* get running best results and their priorities */
+			results1 = vld1q_s32(saved_results);
+			priority1 = vld1q_s32(saved_priority);
+
+			/* select results that are highest priority */
+			selector = vcgtq_s32(priority1, priority);
+			results = vbslq_s32(selector, results1, results);
+			priority = vbslq_s32(selector, priority1, priority);
+		}
+
+		/* save running best results and their priorities */
+		vst1q_s32(saved_results, results);
+		vst1q_s32(saved_priority, priority);
+	}
+}
+
+/*
+ * Check for any match in 4 transitions
+ */
+static inline __attribute__((always_inline)) uint32_t
+check_any_match_x4(uint64_t val[])
+{
+	return ((val[0] | val[1] | val[2] | val[3]) & RTE_ACL_NODE_MATCH);
+}
+
+static inline __attribute__((always_inline)) void
+acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
+		   struct acl_flow_data *flows, uint64_t transitions[])
+{
+	while (check_any_match_x4(transitions)) {
+		transitions[0] = acl_match_check(transitions[0], slot, ctx,
+			parms, flows, resolve_priority_neon);
+		transitions[1] = acl_match_check(transitions[1], slot + 1, ctx,
+			parms, flows, resolve_priority_neon);
+		transitions[2] = acl_match_check(transitions[2], slot + 2, ctx,
+			parms, flows, resolve_priority_neon);
+		transitions[3] = acl_match_check(transitions[3], slot + 3, ctx,
+			parms, flows, resolve_priority_neon);
+	}
+}
+
+/*
+ * Process 4 transitions (in 2 NEON Q registers) in parallel
+ */
+static inline __attribute__((always_inline)) int32x4_t
+transition4(int32x4_t next_input, const uint64_t *trans, uint64_t transitions[])
+{
+	int32x4x2_t tr_hi_lo;
+	int32x4_t t, in, r;
+	uint32x4_t index_msk, node_type, addr;
+	uint32x4_t dfa_msk, mask, quad_ofs, dfa_ofs;
+
+	/* Move low 32 into tr_hi_lo.val[0] and high 32 into tr_hi_lo.val[1] */
+	tr_hi_lo = vld2q_s32((const int32_t *)transitions);
+
+	/* Calculate the address (array index) for all 4 transitions. */
+
+	index_msk = vld1q_u32((const uint32_t *)&neon_acl_const.xmm_index_mask);
+
+	/* Calc node type and node addr */
+	node_type = vbicq_s32(tr_hi_lo.val[0], index_msk);
+	addr = vandq_s32(tr_hi_lo.val[0], index_msk);
+
+	/* t = 0 */
+	t = veorq_s32(node_type, node_type);
+
+	/* mask for DFA type(0) nodes */
+	dfa_msk = vceqq_u32(node_type, t);
+
+	mask = vld1q_s32((const int32_t *)&neon_acl_const.xmm_shuffle_input);
+	in = vqtbl1q_u8((uint8x16_t)next_input, (uint8x16_t)mask);
+
+	/* DFA calculations. */
+	r = vshrq_n_u32(in, 30); /* div by 64 */
+	mask = vld1q_s32((const int32_t *)&neon_acl_const.range_base);
+	r = vaddq_u8(r, mask);
+	t = vshrq_n_u32(in, 24);
+	r = vqtbl1q_u8((uint8x16_t)tr_hi_lo.val[1], (uint8x16_t)r);
+	dfa_ofs = vsubq_s32(t, r);
+
+	/* QUAD/SINGLE calculations. */
+	t = vcgtq_s8(in, tr_hi_lo.val[1]);
+	t = vabsq_s8(t);
+	t = vpaddlq_u8(t);
+	quad_ofs = vpaddlq_u16(t);
+
+	/* blend DFA and QUAD/SINGLE. */
+	t = vbslq_u8(dfa_msk, dfa_ofs, quad_ofs);
+
+	/* calculate address for next transitions */
+	addr = vaddq_u32(addr, t);
+
+	/* Fill next transitions */
+	transitions[0] = trans[vgetq_lane_u32(addr, 0)];
+	transitions[1] = trans[vgetq_lane_u32(addr, 1)];
+	transitions[2] = trans[vgetq_lane_u32(addr, 2)];
+	transitions[3] = trans[vgetq_lane_u32(addr, 3)];
+
+	return vshrq_n_u32(next_input, CHAR_BIT);
+}
+
+/*
+ * Execute trie traversal with 8 traversals in parallel
+ */
+static inline int
+search_neon_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
+	      uint32_t *results, uint32_t total_packets, uint32_t categories)
+{
+	int n;
+	struct acl_flow_data flows;
+	uint64_t index_array[8];
+	struct completion cmplt[8];
+	struct parms parms[8];
+	int32x4_t input0, input1;
+
+	acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
+		     total_packets, categories, ctx->trans_table);
+
+	for (n = 0; n < 8; n++) {
+		cmplt[n].count = 0;
+		index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
+	}
+
+	 /* Check for any matches. */
+	acl_match_check_x4(0, ctx, parms, &flows, &index_array[0]);
+	acl_match_check_x4(4, ctx, parms, &flows, &index_array[4]);
+
+	while (flows.started > 0) {
+		/* Gather 4 bytes of input data for each stream. */
+		input0 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 0), input0, 0);
+		input1 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 4), input1, 0);
+
+		input0 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 1), input0, 1);
+		input1 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 5), input1, 1);
+
+		input0 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 2), input0, 2);
+		input1 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 6), input1, 2);
+
+		input0 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 3), input0, 3);
+		input1 = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 7), input1, 3);
+
+		/* Process the 4 bytes of input on each stream. */
+
+		input0 = transition4(input0, flows.trans, &index_array[0]);
+		input1 = transition4(input1, flows.trans, &index_array[4]);
+
+		input0 = transition4(input0, flows.trans, &index_array[0]);
+		input1 = transition4(input1, flows.trans, &index_array[4]);
+
+		input0 = transition4(input0, flows.trans, &index_array[0]);
+		input1 = transition4(input1, flows.trans, &index_array[4]);
+
+		input0 = transition4(input0, flows.trans, &index_array[0]);
+		input1 = transition4(input1, flows.trans, &index_array[4]);
+
+		 /* Check for any matches. */
+		acl_match_check_x4(0, ctx, parms, &flows, &index_array[0]);
+		acl_match_check_x4(4, ctx, parms, &flows, &index_array[4]);
+	}
+
+	return 0;
+}
+
+/*
+ * Execute trie traversal with 4 traversals in parallel
+ */
+static inline int
+search_neon_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
+	      uint32_t *results, int total_packets, uint32_t categories)
+{
+	int n;
+	struct acl_flow_data flows;
+	uint64_t index_array[4];
+	struct completion cmplt[4];
+	struct parms parms[4];
+	int32x4_t input;
+
+	acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
+		     total_packets, categories, ctx->trans_table);
+
+	for (n = 0; n < 4; n++) {
+		cmplt[n].count = 0;
+		index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
+	}
+
+	/* Check for any matches. */
+	acl_match_check_x4(0, ctx, parms, &flows, index_array);
+
+	while (flows.started > 0) {
+		/* Gather 4 bytes of input data for each stream. */
+		input = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 0), input, 0);
+		input = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 1), input, 1);
+		input = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 2), input, 2);
+		input = vsetq_lane_s32(GET_NEXT_4BYTES(parms, 3), input, 3);
+
+		/* Process the 4 bytes of input on each stream. */
+		input = transition4(input, flows.trans, index_array);
+		input = transition4(input, flows.trans, index_array);
+		input = transition4(input, flows.trans, index_array);
+		input = transition4(input, flows.trans, index_array);
+
+		/* Check for any matches. */
+		acl_match_check_x4(0, ctx, parms, &flows, index_array);
+	}
+
+	return 0;
+}
+
diff --git a/lib/librte_acl/rte_acl.c b/lib/librte_acl/rte_acl.c
index d60219f..e2fdebd 100644
--- a/lib/librte_acl/rte_acl.c
+++ b/lib/librte_acl/rte_acl.c
@@ -55,11 +55,32 @@ rte_acl_classify_avx2(__rte_unused const struct rte_acl_ctx *ctx,
 	return -ENOTSUP;
 }
 
+int __attribute__ ((weak))
+rte_acl_classify_sse(__rte_unused const struct rte_acl_ctx *ctx,
+	__rte_unused const uint8_t **data,
+	__rte_unused uint32_t *results,
+	__rte_unused uint32_t num,
+	__rte_unused uint32_t categories)
+{
+	return -ENOTSUP;
+}
+
+int __attribute__ ((weak))
+rte_acl_classify_neon(__rte_unused const struct rte_acl_ctx *ctx,
+	__rte_unused const uint8_t **data,
+	__rte_unused uint32_t *results,
+	__rte_unused uint32_t num,
+	__rte_unused uint32_t categories)
+{
+	return -ENOTSUP;
+}
+
 static const rte_acl_classify_t classify_fns[] = {
 	[RTE_ACL_CLASSIFY_DEFAULT] = rte_acl_classify_scalar,
 	[RTE_ACL_CLASSIFY_SCALAR] = rte_acl_classify_scalar,
 	[RTE_ACL_CLASSIFY_SSE] = rte_acl_classify_sse,
 	[RTE_ACL_CLASSIFY_AVX2] = rte_acl_classify_avx2,
+	[RTE_ACL_CLASSIFY_NEON] = rte_acl_classify_neon,
 };
 
 /* by default, use always available scalar code path. */
@@ -93,6 +114,9 @@ rte_acl_init(void)
 {
 	enum rte_acl_classify_alg alg = RTE_ACL_CLASSIFY_DEFAULT;
 
+#ifdef RTE_ARCH_ARM64
+	alg =  RTE_ACL_CLASSIFY_NEON;
+#else
 #ifdef CC_AVX2_SUPPORT
 	if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
 		alg = RTE_ACL_CLASSIFY_AVX2;
@@ -102,6 +126,7 @@ rte_acl_init(void)
 #endif
 		alg = RTE_ACL_CLASSIFY_SSE;
 
+#endif
 	rte_acl_set_default_classify(alg);
 }
 
diff --git a/lib/librte_acl/rte_acl.h b/lib/librte_acl/rte_acl.h
index 98ef2fc..0979a09 100644
--- a/lib/librte_acl/rte_acl.h
+++ b/lib/librte_acl/rte_acl.h
@@ -270,6 +270,7 @@ enum rte_acl_classify_alg {
 	RTE_ACL_CLASSIFY_SCALAR = 1,  /**< generic implementation. */
 	RTE_ACL_CLASSIFY_SSE = 2,     /**< requires SSE4.1 support. */
 	RTE_ACL_CLASSIFY_AVX2 = 3,    /**< requires AVX2 support. */
+	RTE_ACL_CLASSIFY_NEON = 4,    /**< requires NEON support. */
 	RTE_ACL_CLASSIFY_NUM          /* should always be the last one. */
 };
 
-- 
2.1.0

  reply	other threads:[~2015-11-02 14:20 UTC|newest]

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2015-11-02 14:18 [dpdk-dev] [PATCH 0/3] ARM64: NEON ACL implementation Jerin Jacob
2015-11-02 14:18 ` [dpdk-dev] [PATCH 1/3] arm: ret_vector.h improvements Jerin Jacob
2015-11-02 14:18   ` Jerin Jacob [this message]
2015-11-02 14:18     ` [dpdk-dev] [PATCH 3/3] arm64: defconfig: enabled CONFIG_RTE_LIBRTE_ACL Jerin Jacob
2015-11-02 15:39     ` [dpdk-dev] [PATCH 2/3] arm64: acl: add neon based acl implementation Jan Viktorin
2015-11-02 16:19       ` Jerin Jacob
2015-11-02 17:31         ` Jan Viktorin
2015-11-02 16:54     ` Ananyev, Konstantin
2015-11-03  4:30       ` Jerin Jacob
2015-11-03 10:23         ` Ananyev, Konstantin
2015-11-03 10:35           ` Jan Viktorin
2015-11-03 13:20             ` Ananyev, Konstantin

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1446473921-12706-3-git-send-email-jerin.jacob@caviumnetworks.com \
    --to=jerin.jacob@caviumnetworks.com \
    --cc=dev@dpdk.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).