Patch Detail
get:
Show a patch.
patch:
Update a patch.
put:
Update a patch.
GET /api/patches/77498/?format=api
{ "id": 77498, "url": "http://patches.dpdk.org/api/patches/77498/?format=api", "web_url": "http://patches.dpdk.org/project/dpdk/patch/1599851920-16802-17-git-send-email-timothy.mcdaniel@intel.com/", "project": { "id": 1, "url": "http://patches.dpdk.org/api/projects/1/?format=api", "name": "DPDK", "link_name": "dpdk", "list_id": "dev.dpdk.org", "list_email": "dev@dpdk.org", "web_url": "http://core.dpdk.org", "scm_url": "git://dpdk.org/dpdk", "webscm_url": "http://git.dpdk.org/dpdk", "list_archive_url": "https://inbox.dpdk.org/dev", "list_archive_url_format": "https://inbox.dpdk.org/dev/{}", "commit_url_format": "" }, "msgid": "<1599851920-16802-17-git-send-email-timothy.mcdaniel@intel.com>", "list_archive_url": "https://inbox.dpdk.org/dev/1599851920-16802-17-git-send-email-timothy.mcdaniel@intel.com", "date": "2020-09-11T19:18:34", "name": "[v4,16/22] event/dlb: add dequeue and its burst variants", "commit_ref": null, "pull_url": null, "state": "superseded", "archived": true, "hash": "a7c9988dbde9f173227af0e61d5efbe90586390e", "submitter": { "id": 826, "url": "http://patches.dpdk.org/api/people/826/?format=api", "name": "Timothy McDaniel", "email": "timothy.mcdaniel@intel.com" }, "delegate": { "id": 310, "url": "http://patches.dpdk.org/api/users/310/?format=api", "username": "jerin", "first_name": "Jerin", "last_name": "Jacob", "email": "jerinj@marvell.com" }, "mbox": "http://patches.dpdk.org/project/dpdk/patch/1599851920-16802-17-git-send-email-timothy.mcdaniel@intel.com/mbox/", "series": [ { "id": 12163, "url": "http://patches.dpdk.org/api/series/12163/?format=api", "web_url": "http://patches.dpdk.org/project/dpdk/list/?series=12163", "date": "2020-09-11T19:18:18", "name": "Add DLB PMD", "version": 4, "mbox": "http://patches.dpdk.org/series/12163/mbox/" } ], "comments": "http://patches.dpdk.org/api/patches/77498/comments/", "check": "success", "checks": "http://patches.dpdk.org/api/patches/77498/checks/", "tags": {}, "related": [], "headers": { "Return-Path": "<dev-bounces@dpdk.org>", "X-Original-To": "patchwork@inbox.dpdk.org", "Delivered-To": "patchwork@inbox.dpdk.org", "Received": [ "from dpdk.org (dpdk.org [92.243.14.124])\n\tby inbox.dpdk.org (Postfix) with ESMTP id 02917A04C1;\n\tFri, 11 Sep 2020 21:24:36 +0200 (CEST)", "from [92.243.14.124] (localhost [127.0.0.1])\n\tby dpdk.org (Postfix) with ESMTP id 4370C1C20B;\n\tFri, 11 Sep 2020 21:22:28 +0200 (CEST)", "from mga18.intel.com (mga18.intel.com [134.134.136.126])\n by dpdk.org (Postfix) with ESMTP id 468F41C1AB\n for <dev@dpdk.org>; Fri, 11 Sep 2020 21:22:14 +0200 (CEST)", "from fmsmga005.fm.intel.com ([10.253.24.32])\n by orsmga106.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384;\n 11 Sep 2020 12:22:13 -0700", "from txasoft-yocto.an.intel.com ([10.123.72.192])\n by fmsmga005.fm.intel.com with ESMTP; 11 Sep 2020 12:22:12 -0700" ], "IronPort-SDR": [ "\n PLh7+J8XI4Kk8tFZ2kvGLFil+u2+QaQSHOTcJ0xQNRZv5lUkriVWhrrFf0uSz1FEelFWwE3Rsr\n KAr2+0tRNF+g==", "\n R6gEHlxL32lpzGykrbAM4OXMFwlrnAzobypbRPvZ9iXeulzO8F3AW0n5/xfIxhtTMSfdgXSmIR\n LkVD+FdBmvbQ==" ], "X-IronPort-AV": [ "E=McAfee;i=\"6000,8403,9741\"; a=\"146570557\"", "E=Sophos;i=\"5.76,416,1592895600\"; d=\"scan'208\";a=\"146570557\"", "E=Sophos;i=\"5.76,416,1592895600\"; d=\"scan'208\";a=\"506375699\"" ], "X-Amp-Result": "SKIPPED(no attachment in message)", "X-Amp-File-Uploaded": "False", "X-ExtLoop1": "1", "From": "Timothy McDaniel <timothy.mcdaniel@intel.com>", "To": "", "Cc": "dev@dpdk.org, erik.g.carrillo@intel.com, gage.eads@intel.com,\n harry.van.haaren@intel.com, jerinj@marvell.com", "Date": "Fri, 11 Sep 2020 14:18:34 -0500", "Message-Id": "<1599851920-16802-17-git-send-email-timothy.mcdaniel@intel.com>", "X-Mailer": "git-send-email 1.7.10", "In-Reply-To": "<1599851920-16802-1-git-send-email-timothy.mcdaniel@intel.com>", "References": "<1599851920-16802-1-git-send-email-timothy.mcdaniel@intel.com>", "Subject": "[dpdk-dev] [PATCH v4 16/22] event/dlb: add dequeue and its burst\n\tvariants", "X-BeenThere": "dev@dpdk.org", "X-Mailman-Version": "2.1.15", "Precedence": "list", "List-Id": "DPDK patches and discussions <dev.dpdk.org>", "List-Unsubscribe": "<https://mails.dpdk.org/options/dev>,\n <mailto:dev-request@dpdk.org?subject=unsubscribe>", "List-Archive": "<http://mails.dpdk.org/archives/dev/>", "List-Post": "<mailto:dev@dpdk.org>", "List-Help": "<mailto:dev-request@dpdk.org?subject=help>", "List-Subscribe": "<https://mails.dpdk.org/listinfo/dev>,\n <mailto:dev-request@dpdk.org?subject=subscribe>", "Errors-To": "dev-bounces@dpdk.org", "Sender": "\"dev\" <dev-bounces@dpdk.org>" }, "content": "Add support for dequeue, dequeue_burst, ...\n\nSigned-off-by: Timothy McDaniel <timothy.mcdaniel@intel.com>\n---\n drivers/event/dlb/dlb.c | 763 ++++++++++++++++++++++++++++++++++++++++++++++++\n 1 file changed, 763 insertions(+)", "diff": "diff --git a/drivers/event/dlb/dlb.c b/drivers/event/dlb/dlb.c\nindex 8fbd25a..ba2323b 100644\n--- a/drivers/event/dlb/dlb.c\n+++ b/drivers/event/dlb/dlb.c\n@@ -2887,9 +2887,764 @@ dlb_event_enqueue_forward_burst_delayed(void *event_port,\n \treturn __dlb_event_enqueue_burst(event_port, events, num, true);\n }\n \n+static __rte_always_inline int\n+dlb_recv_qe(struct dlb_port *qm_port, struct dlb_dequeue_qe *qe,\n+\t uint8_t *offset)\n+{\n+\tuint8_t xor_mask[2][4] = { {0x0F, 0x0E, 0x0C, 0x08},\n+\t\t\t\t {0x00, 0x01, 0x03, 0x07} };\n+\tuint8_t and_mask[4] = {0x0F, 0x0E, 0x0C, 0x08};\n+\tvolatile struct dlb_dequeue_qe *cq_addr;\n+\t__m128i *qes = (__m128i *)qe;\n+\tuint64_t *cache_line_base;\n+\tuint8_t gen_bits;\n+\n+\tcq_addr = dlb_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;\n+\tcq_addr = &cq_addr[qm_port->cq_idx];\n+\n+\tcache_line_base = (void *)(((uintptr_t)cq_addr) & ~0x3F);\n+\t*offset = ((uintptr_t)cq_addr & 0x30) >> 4;\n+\n+\t/* Load the next CQ cache line from memory. Pack these reads as tight\n+\t * as possible to reduce the chance that DLB invalidates the line while\n+\t * the CPU is reading it. Read the cache line backwards to ensure that\n+\t * if QE[N] (N > 0) is valid, then QEs[0:N-1] are too.\n+\t *\n+\t * (Valid QEs start at &qe[offset])\n+\t */\n+\tqes[3] = _mm_load_si128((__m128i *)&cache_line_base[6]);\n+\tqes[2] = _mm_load_si128((__m128i *)&cache_line_base[4]);\n+\tqes[1] = _mm_load_si128((__m128i *)&cache_line_base[2]);\n+\tqes[0] = _mm_load_si128((__m128i *)&cache_line_base[0]);\n+\n+\t/* Evict the cache line ASAP */\n+\tdlb_cldemote(cache_line_base);\n+\n+\t/* Extract and combine the gen bits */\n+\tgen_bits = ((_mm_extract_epi8(qes[0], 15) & 0x1) << 0) |\n+\t\t ((_mm_extract_epi8(qes[1], 15) & 0x1) << 1) |\n+\t\t ((_mm_extract_epi8(qes[2], 15) & 0x1) << 2) |\n+\t\t ((_mm_extract_epi8(qes[3], 15) & 0x1) << 3);\n+\n+\t/* XOR the combined bits such that a 1 represents a valid QE */\n+\tgen_bits ^= xor_mask[qm_port->gen_bit][*offset];\n+\n+\t/* Mask off gen bits we don't care about */\n+\tgen_bits &= and_mask[*offset];\n+\n+\treturn __builtin_popcount(gen_bits);\n+}\n+\n+static inline void\n+dlb_inc_cq_idx(struct dlb_port *qm_port, int cnt)\n+{\n+\tuint16_t idx = qm_port->cq_idx_unmasked + cnt;\n+\n+\tqm_port->cq_idx_unmasked = idx;\n+\tqm_port->cq_idx = idx & qm_port->cq_depth_mask;\n+\tqm_port->gen_bit = (~(idx >> qm_port->gen_bit_shift)) & 0x1;\n+}\n+\n+static inline int\n+dlb_process_dequeue_qes(struct dlb_eventdev_port *ev_port,\n+\t\t\tstruct dlb_port *qm_port,\n+\t\t\tstruct rte_event *events,\n+\t\t\tstruct dlb_dequeue_qe *qes,\n+\t\t\tint cnt)\n+{\n+\tuint8_t *qid_mappings = qm_port->qid_mappings;\n+\tint i, num;\n+\n+\tRTE_SET_USED(ev_port); /* avoids unused variable error */\n+\n+\tfor (i = 0, num = 0; i < cnt; i++) {\n+\t\tstruct dlb_dequeue_qe *qe = &qes[i];\n+\t\tint sched_type_map[4] = {\n+\t\t\t[DLB_SCHED_ATOMIC] = RTE_SCHED_TYPE_ATOMIC,\n+\t\t\t[DLB_SCHED_UNORDERED] = RTE_SCHED_TYPE_PARALLEL,\n+\t\t\t[DLB_SCHED_ORDERED] = RTE_SCHED_TYPE_ORDERED,\n+\t\t\t[DLB_SCHED_DIRECTED] = RTE_SCHED_TYPE_ATOMIC,\n+\t\t};\n+\n+\t\tDLB_LOG_DBG(\"dequeue success, data = 0x%llx, qid=%d, event_type=%d, subevent=%d\\npp_id = %d, sched_type = %d, qid = %d, err=%d\\n\",\n+\t\t\t (long long)qe->data, qe->qid,\n+\t\t\t qe->u.event_type.major,\n+\t\t\t qe->u.event_type.sub,\n+\t\t\t qe->pp_id, qe->sched_type, qe->qid, qe->error);\n+\n+\t\t/* Fill in event information.\n+\t\t * Note that flow_id must be embedded in the data by\n+\t\t * the app, such as the mbuf RSS hash field if the data\n+\t\t * buffer is a mbuf.\n+\t\t */\n+\t\tif (unlikely(qe->error)) {\n+\t\t\tDLB_LOG_ERR(\"QE error bit ON\\n\");\n+\t\t\tDLB_INC_STAT(ev_port->stats.traffic.rx_drop, 1);\n+\t\t\tdlb_consume_qe_immediate(qm_port, 1);\n+\t\t\tcontinue; /* Ignore */\n+\t\t}\n+\n+\t\tevents[num].u64 = qe->data;\n+\t\tevents[num].queue_id = qid_mappings[qe->qid];\n+\t\tevents[num].priority = DLB_TO_EV_PRIO((uint8_t)qe->priority);\n+\t\tevents[num].event_type = qe->u.event_type.major;\n+\t\tevents[num].sub_event_type = qe->u.event_type.sub;\n+\t\tevents[num].sched_type = sched_type_map[qe->sched_type];\n+\t\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qe->sched_type], 1);\n+\n+\t\tDLB_INC_STAT(ev_port->stats.traffic.rx_ok, 1);\n+\n+\t\tnum++;\n+\t}\n+\n+\treturn num;\n+}\n+\n+static inline int\n+dlb_process_dequeue_four_qes(struct dlb_eventdev_port *ev_port,\n+\t\t\t struct dlb_port *qm_port,\n+\t\t\t struct rte_event *events,\n+\t\t\t struct dlb_dequeue_qe *qes)\n+{\n+\tint sched_type_map[] = {\n+\t\t[DLB_SCHED_ATOMIC] = RTE_SCHED_TYPE_ATOMIC,\n+\t\t[DLB_SCHED_UNORDERED] = RTE_SCHED_TYPE_PARALLEL,\n+\t\t[DLB_SCHED_ORDERED] = RTE_SCHED_TYPE_ORDERED,\n+\t\t[DLB_SCHED_DIRECTED] = RTE_SCHED_TYPE_ATOMIC,\n+\t};\n+\tconst int num_events = DLB_NUM_QES_PER_CACHE_LINE;\n+\tuint8_t *qid_mappings = qm_port->qid_mappings;\n+\t__m128i sse_evt[2];\n+\tint i;\n+\n+\t/* In the unlikely case that any of the QE error bits are set, process\n+\t * them one at a time.\n+\t */\n+\tif (unlikely(qes[0].error || qes[1].error ||\n+\t\t qes[2].error || qes[3].error))\n+\t\treturn dlb_process_dequeue_qes(ev_port, qm_port, events,\n+\t\t\t\t\t qes, num_events);\n+\n+\tfor (i = 0; i < DLB_NUM_QES_PER_CACHE_LINE; i++) {\n+\t\tDLB_LOG_DBG(\"dequeue success, data = 0x%llx, qid=%d, event_type=%d, subevent=%d\\npp_id = %d, sched_type = %d, qid = %d, err=%d\\n\",\n+\t\t\t (long long)qes[i].data, qes[i].qid,\n+\t\t\t qes[i].u.event_type.major,\n+\t\t\t qes[i].u.event_type.sub,\n+\t\t\t qes[i].pp_id, qes[i].sched_type, qes[i].qid,\n+\t\t\t qes[i].error);\n+\t}\n+\n+\tevents[0].u64 = qes[0].data;\n+\tevents[1].u64 = qes[1].data;\n+\tevents[2].u64 = qes[2].data;\n+\tevents[3].u64 = qes[3].data;\n+\n+\t/* Construct the metadata portion of two struct rte_events\n+\t * in one 128b SSE register. Event metadata is constructed in the SSE\n+\t * registers like so:\n+\t * sse_evt[0][63:0]: event[0]'s metadata\n+\t * sse_evt[0][127:64]: event[1]'s metadata\n+\t * sse_evt[1][63:0]: event[2]'s metadata\n+\t * sse_evt[1][127:64]: event[3]'s metadata\n+\t */\n+\tsse_evt[0] = _mm_setzero_si128();\n+\tsse_evt[1] = _mm_setzero_si128();\n+\n+\t/* Convert the hardware queue ID to an event queue ID and store it in\n+\t * the metadata:\n+\t * sse_evt[0][47:40] = qid_mappings[qes[0].qid]\n+\t * sse_evt[0][111:104] = qid_mappings[qes[1].qid]\n+\t * sse_evt[1][47:40] = qid_mappings[qes[2].qid]\n+\t * sse_evt[1][111:104] = qid_mappings[qes[3].qid]\n+\t */\n+#define RTE_EVENT_QUEUE_ID_BYTE 5\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\t\t\t qid_mappings[qes[0].qid],\n+\t\t\t\t RTE_EVENT_QUEUE_ID_BYTE);\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\t\t\t qid_mappings[qes[1].qid],\n+\t\t\t\t RTE_EVENT_QUEUE_ID_BYTE + 8);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\t\t\t qid_mappings[qes[2].qid],\n+\t\t\t\t RTE_EVENT_QUEUE_ID_BYTE);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\t\t\t qid_mappings[qes[3].qid],\n+\t\t\t\t RTE_EVENT_QUEUE_ID_BYTE + 8);\n+\n+\t/* Convert the hardware priority to an event priority and store it in\n+\t * the metadata:\n+\t * sse_evt[0][55:48] = DLB_TO_EV_PRIO(qes[0].priority)\n+\t * sse_evt[0][119:112] = DLB_TO_EV_PRIO(qes[1].priority)\n+\t * sse_evt[1][55:48] = DLB_TO_EV_PRIO(qes[2].priority)\n+\t * sse_evt[1][119:112] = DLB_TO_EV_PRIO(qes[3].priority)\n+\t */\n+#define RTE_EVENT_PRIO_BYTE 6\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\t\t\t DLB_TO_EV_PRIO((uint8_t)qes[0].priority),\n+\t\t\t\t RTE_EVENT_PRIO_BYTE);\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\t\t\t DLB_TO_EV_PRIO((uint8_t)qes[1].priority),\n+\t\t\t\t RTE_EVENT_PRIO_BYTE + 8);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\t\t\t DLB_TO_EV_PRIO((uint8_t)qes[2].priority),\n+\t\t\t\t RTE_EVENT_PRIO_BYTE);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\t\t\t DLB_TO_EV_PRIO((uint8_t)qes[3].priority),\n+\t\t\t\t RTE_EVENT_PRIO_BYTE + 8);\n+\n+\t/* Write the event type and sub event type to the event metadata. Leave\n+\t * flow ID unspecified, since the hardware does not maintain it during\n+\t * scheduling:\n+\t * sse_evt[0][31:0] = qes[0].u.event_type.major << 28 |\n+\t *\t\t\tqes[0].u.event_type.sub << 20;\n+\t * sse_evt[0][95:64] = qes[1].u.event_type.major << 28 |\n+\t *\t\t\tqes[1].u.event_type.sub << 20;\n+\t * sse_evt[1][31:0] = qes[2].u.event_type.major << 28 |\n+\t *\t\t\tqes[2].u.event_type.sub << 20;\n+\t * sse_evt[1][95:64] = qes[3].u.event_type.major << 28 |\n+\t *\t\t\tqes[3].u.event_type.sub << 20;\n+\t */\n+#define RTE_EVENT_EV_TYPE_DW 0\n+#define RTE_EVENT_EV_TYPE_SHIFT 28\n+#define RTE_EVENT_SUB_EV_TYPE_SHIFT 20\n+\tsse_evt[0] = _mm_insert_epi32(sse_evt[0],\n+\t\t\tqes[0].u.event_type.major << RTE_EVENT_EV_TYPE_SHIFT |\n+\t\t\tqes[0].u.event_type.sub << RTE_EVENT_SUB_EV_TYPE_SHIFT,\n+\t\t\tRTE_EVENT_EV_TYPE_DW);\n+\tsse_evt[0] = _mm_insert_epi32(sse_evt[0],\n+\t\t\tqes[1].u.event_type.major << RTE_EVENT_EV_TYPE_SHIFT |\n+\t\t\tqes[1].u.event_type.sub << RTE_EVENT_SUB_EV_TYPE_SHIFT,\n+\t\t\tRTE_EVENT_EV_TYPE_DW + 2);\n+\tsse_evt[1] = _mm_insert_epi32(sse_evt[1],\n+\t\t\tqes[2].u.event_type.major << RTE_EVENT_EV_TYPE_SHIFT |\n+\t\t\tqes[2].u.event_type.sub << RTE_EVENT_SUB_EV_TYPE_SHIFT,\n+\t\t\tRTE_EVENT_EV_TYPE_DW);\n+\tsse_evt[1] = _mm_insert_epi32(sse_evt[1],\n+\t\t\tqes[3].u.event_type.major << RTE_EVENT_EV_TYPE_SHIFT |\n+\t\t\tqes[3].u.event_type.sub << RTE_EVENT_SUB_EV_TYPE_SHIFT,\n+\t\t\tRTE_EVENT_EV_TYPE_DW + 2);\n+\n+\t/* Write the sched type to the event metadata. 'op' and 'rsvd' are not\n+\t * set:\n+\t * sse_evt[0][39:32] = sched_type_map[qes[0].sched_type] << 6\n+\t * sse_evt[0][103:96] = sched_type_map[qes[1].sched_type] << 6\n+\t * sse_evt[1][39:32] = sched_type_map[qes[2].sched_type] << 6\n+\t * sse_evt[1][103:96] = sched_type_map[qes[3].sched_type] << 6\n+\t */\n+#define RTE_EVENT_SCHED_TYPE_BYTE 4\n+#define RTE_EVENT_SCHED_TYPE_SHIFT 6\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\tsched_type_map[qes[0].sched_type] << RTE_EVENT_SCHED_TYPE_SHIFT,\n+\t\tRTE_EVENT_SCHED_TYPE_BYTE);\n+\tsse_evt[0] = _mm_insert_epi8(sse_evt[0],\n+\t\tsched_type_map[qes[1].sched_type] << RTE_EVENT_SCHED_TYPE_SHIFT,\n+\t\tRTE_EVENT_SCHED_TYPE_BYTE + 8);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\tsched_type_map[qes[2].sched_type] << RTE_EVENT_SCHED_TYPE_SHIFT,\n+\t\tRTE_EVENT_SCHED_TYPE_BYTE);\n+\tsse_evt[1] = _mm_insert_epi8(sse_evt[1],\n+\t\tsched_type_map[qes[3].sched_type] << RTE_EVENT_SCHED_TYPE_SHIFT,\n+\t\tRTE_EVENT_SCHED_TYPE_BYTE + 8);\n+\n+\t/* Store the metadata to the event (use the double-precision\n+\t * _mm_storeh_pd because there is no integer function for storing the\n+\t * upper 64b):\n+\t * events[0].event = sse_evt[0][63:0]\n+\t * events[1].event = sse_evt[0][127:64]\n+\t * events[2].event = sse_evt[1][63:0]\n+\t * events[3].event = sse_evt[1][127:64]\n+\t */\n+\t_mm_storel_epi64((__m128i *)&events[0].event, sse_evt[0]);\n+\t_mm_storeh_pd((double *)&events[1].event, (__m128d) sse_evt[0]);\n+\t_mm_storel_epi64((__m128i *)&events[2].event, sse_evt[1]);\n+\t_mm_storeh_pd((double *)&events[3].event, (__m128d) sse_evt[1]);\n+\n+\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qes[0].sched_type], 1);\n+\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qes[1].sched_type], 1);\n+\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qes[2].sched_type], 1);\n+\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qes[3].sched_type], 1);\n+\n+\tDLB_INC_STAT(ev_port->stats.traffic.rx_ok, num_events);\n+\n+\treturn num_events;\n+}\n+\n+static inline bool\n+dlb_cq_is_empty(struct dlb_port *qm_port)\n+{\n+\tvolatile struct dlb_dequeue_qe *qe_ptr;\n+\tstruct dlb_dequeue_qe qe;\n+\n+\tqe_ptr = dlb_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;\n+\tqe = qe_ptr[qm_port->cq_idx];\n+\n+\treturn (qe.cq_gen != qm_port->gen_bit);\n+}\n+\n+static inline int\n+dlb_dequeue_wait(struct dlb_eventdev *dlb,\n+\t\t struct dlb_eventdev_port *ev_port,\n+\t\t struct dlb_port *qm_port,\n+\t\t uint64_t timeout,\n+\t\t uint64_t start_ticks)\n+{\n+\tstruct process_local_port_data *port_data;\n+\tuint64_t elapsed_ticks;\n+\n+\tport_data = &dlb_port[qm_port->id][PORT_TYPE(qm_port)];\n+\n+\telapsed_ticks = rte_get_timer_cycles() - start_ticks;\n+\n+\t/* Wait/poll time expired */\n+\tif (elapsed_ticks >= timeout) {\n+\t\t/* Interrupts not supported by PF PMD */\n+\t\treturn 1;\n+\t} else if (dlb->umwait_allowed) {\n+\t\tvolatile struct dlb_dequeue_qe *cq_base;\n+\n+\t\tcq_base = port_data->cq_base;\n+\n+\t\t/* Block on cache line write to CQ. Note: it's\n+\t\t * safe to access the per-process cq_base\n+\t\t * address here, since the PMD has already\n+\t\t * attempted at least one CQ dequeue.\n+\t\t */\n+\t\tdlb_umonitor(&cq_base[qm_port->cq_idx]);\n+\n+\t\t/* Avoid race condition. Check if still empty */\n+\t\tif (dlb_cq_is_empty(qm_port)) {\n+\t\t\tdlb_umwait(RTE_LIBRTE_PMD_DLB_UMWAIT_CTL_STATE,\n+\t\t\t\t timeout + start_ticks);\n+\t\t\tDLB_INC_STAT(ev_port->stats.traffic.rx_umonitor_umwait,\n+\t\t\t\t 1);\n+\t\t}\n+\t} else {\n+\t\tuint64_t poll_interval = RTE_LIBRTE_PMD_DLB_POLL_INTERVAL;\n+\t\tuint64_t curr_ticks = rte_get_timer_cycles();\n+\t\tuint64_t init_ticks = curr_ticks;\n+\n+\t\twhile ((curr_ticks - start_ticks < timeout) &&\n+\t\t (curr_ticks - init_ticks < poll_interval))\n+\t\t\tcurr_ticks = rte_get_timer_cycles();\n+\t}\n+\n+\treturn 0;\n+}\n+\n+static inline int16_t\n+dlb_hw_dequeue(struct dlb_eventdev *dlb,\n+\t struct dlb_eventdev_port *ev_port,\n+\t struct rte_event *events,\n+\t uint16_t max_num,\n+\t uint64_t dequeue_timeout_ticks)\n+{\n+\tuint64_t timeout;\n+\tuint64_t start_ticks = 0ULL;\n+\tstruct dlb_port *qm_port;\n+\tint num = 0;\n+\n+\tqm_port = &ev_port->qm_port;\n+\n+\t/* If configured for per dequeue wait, then use wait value provided\n+\t * to this API. Otherwise we must use the global\n+\t * value from eventdev config time.\n+\t */\n+\tif (!dlb->global_dequeue_wait)\n+\t\ttimeout = dequeue_timeout_ticks;\n+\telse\n+\t\ttimeout = dlb->global_dequeue_wait_ticks;\n+\n+\tif (timeout)\n+\t\tstart_ticks = rte_get_timer_cycles();\n+\n+\twhile (num < max_num) {\n+\t\tstruct dlb_dequeue_qe qes[DLB_NUM_QES_PER_CACHE_LINE];\n+\t\tuint8_t offset;\n+\t\tint num_avail;\n+\n+\t\t/* Copy up to 4 QEs from the current cache line into qes */\n+\t\tnum_avail = dlb_recv_qe(qm_port, qes, &offset);\n+\n+\t\t/* But don't process more than the user requested */\n+\t\tnum_avail = RTE_MIN(num_avail, max_num - num);\n+\n+\t\tdlb_inc_cq_idx(qm_port, num_avail);\n+\n+\t\tif (num_avail == DLB_NUM_QES_PER_CACHE_LINE)\n+\t\t\tnum += dlb_process_dequeue_four_qes(ev_port,\n+\t\t\t\t\t\t\t qm_port,\n+\t\t\t\t\t\t\t &events[num],\n+\t\t\t\t\t\t\t &qes[offset]);\n+\t\telse if (num_avail)\n+\t\t\tnum += dlb_process_dequeue_qes(ev_port,\n+\t\t\t\t\t\t\tqm_port,\n+\t\t\t\t\t\t\t&events[num],\n+\t\t\t\t\t\t\t&qes[offset],\n+\t\t\t\t\t\t\tnum_avail);\n+\t\telse if ((timeout == 0) || (num > 0))\n+\t\t\t/* Not waiting in any form, or 1+ events received? */\n+\t\t\tbreak;\n+\t\telse if (dlb_dequeue_wait(dlb, ev_port, qm_port,\n+\t\t\t\t\t timeout, start_ticks))\n+\t\t\tbreak;\n+\t}\n+\n+\tqm_port->owed_tokens += num;\n+\n+\tif (num && qm_port->token_pop_mode == AUTO_POP)\n+\t\tdlb_consume_qe_immediate(qm_port, num);\n+\n+\tev_port->outstanding_releases += num;\n+\n+\treturn num;\n+}\n+\n+static inline int\n+dlb_process_dequeue_qe(struct dlb_eventdev_port *ev_port __rte_unused,\n+\t\t struct dlb_port *qm_port,\n+\t\t struct rte_event *event,\n+\t\t struct dlb_dequeue_qe *qe)\n+{\n+\tint sched_type_map[4] = {\n+\t\t[DLB_SCHED_ATOMIC] = RTE_SCHED_TYPE_ATOMIC,\n+\t\t[DLB_SCHED_UNORDERED] = RTE_SCHED_TYPE_PARALLEL,\n+\t\t[DLB_SCHED_ORDERED] = RTE_SCHED_TYPE_ORDERED,\n+\t\t[DLB_SCHED_DIRECTED] = RTE_SCHED_TYPE_ATOMIC,\n+\t};\n+\tuint8_t *qid_mappings = qm_port->qid_mappings;\n+\n+\tDLB_LOG_DBG(\"dequeue success, data = 0x%llx, qid=%d, event_type=%d, subevent=%d\\npp_id = %d, sched_type = %d, qid = %d, err=%d\\n\",\n+\t\t (long long)qe->data, qe->qid,\n+\t\t qe->u.event_type.major,\n+\t\t qe->u.event_type.sub,\n+\t\t qe->pp_id, qe->sched_type, qe->qid, qe->error);\n+\n+\t/* Fill in event information.\n+\t * Note that flow_id must be embedded in the data by\n+\t * the app, such as the mbuf RSS hash field if the data\n+\t * buffer is a mbuf.\n+\t */\n+\tif (unlikely(qe->error)) {\n+\t\tDLB_LOG_ERR(\"QE error bit ON\\n\");\n+\t\tDLB_INC_STAT(ev_port->stats.traffic.rx_drop, 1);\n+\t\tdlb_consume_qe_immediate(qm_port, 1);\n+\t\treturn 0;\n+\t}\n+\n+\tevent->u64 = qe->data;\n+\tevent->queue_id = qid_mappings[qe->qid];\n+\tevent->priority = DLB_TO_EV_PRIO((uint8_t)qe->priority);\n+\tevent->event_type = qe->u.event_type.major;\n+\tevent->sub_event_type = qe->u.event_type.sub;\n+\tevent->sched_type = sched_type_map[qe->sched_type];\n+\tDLB_INC_STAT(ev_port->stats.rx_sched_cnt[qe->sched_type], 1);\n+\n+\tDLB_INC_STAT(ev_port->stats.traffic.rx_ok, 1);\n+\n+\treturn 1;\n+}\n+\n+static __rte_always_inline int\n+dlb_recv_qe_sparse(struct dlb_port *qm_port, struct dlb_dequeue_qe *qe)\n+{\n+\tvolatile struct dlb_dequeue_qe *cq_addr;\n+\tuint8_t xor_mask[2] = {0x0F, 0x00};\n+\tconst uint8_t and_mask = 0x0F;\n+\t__m128i *qes = (__m128i *)qe;\n+\tuint8_t gen_bits, gen_bit;\n+\tuintptr_t addr[4];\n+\tuint16_t idx;\n+\n+\tcq_addr = dlb_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;\n+\n+\tidx = qm_port->cq_idx;\n+\n+\t/* Load the next 4 QEs */\n+\taddr[0] = (uintptr_t)&cq_addr[idx];\n+\taddr[1] = (uintptr_t)&cq_addr[(idx + 4) & qm_port->cq_depth_mask];\n+\taddr[2] = (uintptr_t)&cq_addr[(idx + 8) & qm_port->cq_depth_mask];\n+\taddr[3] = (uintptr_t)&cq_addr[(idx + 12) & qm_port->cq_depth_mask];\n+\n+\t/* Prefetch next batch of QEs (all CQs occupy minimum 8 cache lines) */\n+\trte_prefetch0(&cq_addr[(idx + 16) & qm_port->cq_depth_mask]);\n+\trte_prefetch0(&cq_addr[(idx + 20) & qm_port->cq_depth_mask]);\n+\trte_prefetch0(&cq_addr[(idx + 24) & qm_port->cq_depth_mask]);\n+\trte_prefetch0(&cq_addr[(idx + 28) & qm_port->cq_depth_mask]);\n+\n+\t/* Correct the xor_mask for wrap-around QEs */\n+\tgen_bit = qm_port->gen_bit;\n+\txor_mask[gen_bit] ^= !!((idx + 4) > qm_port->cq_depth_mask) << 1;\n+\txor_mask[gen_bit] ^= !!((idx + 8) > qm_port->cq_depth_mask) << 2;\n+\txor_mask[gen_bit] ^= !!((idx + 12) > qm_port->cq_depth_mask) << 3;\n+\n+\t/* Read the cache lines backwards to ensure that if QE[N] (N > 0) is\n+\t * valid, then QEs[0:N-1] are too.\n+\t */\n+\tqes[3] = _mm_load_si128((__m128i *)(void *)addr[3]);\n+\trte_compiler_barrier();\n+\tqes[2] = _mm_load_si128((__m128i *)(void *)addr[2]);\n+\trte_compiler_barrier();\n+\tqes[1] = _mm_load_si128((__m128i *)(void *)addr[1]);\n+\trte_compiler_barrier();\n+\tqes[0] = _mm_load_si128((__m128i *)(void *)addr[0]);\n+\n+\t/* Extract and combine the gen bits */\n+\tgen_bits = ((_mm_extract_epi8(qes[0], 15) & 0x1) << 0) |\n+\t\t ((_mm_extract_epi8(qes[1], 15) & 0x1) << 1) |\n+\t\t ((_mm_extract_epi8(qes[2], 15) & 0x1) << 2) |\n+\t\t ((_mm_extract_epi8(qes[3], 15) & 0x1) << 3);\n+\n+\t/* XOR the combined bits such that a 1 represents a valid QE */\n+\tgen_bits ^= xor_mask[gen_bit];\n+\n+\t/* Mask off gen bits we don't care about */\n+\tgen_bits &= and_mask;\n+\n+\treturn __builtin_popcount(gen_bits);\n+}\n+\n+static inline int16_t\n+dlb_hw_dequeue_sparse(struct dlb_eventdev *dlb,\n+\t\t struct dlb_eventdev_port *ev_port,\n+\t\t struct rte_event *events,\n+\t\t uint16_t max_num,\n+\t\t uint64_t dequeue_timeout_ticks)\n+{\n+\tuint64_t timeout;\n+\tuint64_t start_ticks = 0ULL;\n+\tstruct dlb_port *qm_port;\n+\tint num = 0;\n+\n+\tqm_port = &ev_port->qm_port;\n+\n+\t/* If configured for per dequeue wait, then use wait value provided\n+\t * to this API. Otherwise we must use the global\n+\t * value from eventdev config time.\n+\t */\n+\tif (!dlb->global_dequeue_wait)\n+\t\ttimeout = dequeue_timeout_ticks;\n+\telse\n+\t\ttimeout = dlb->global_dequeue_wait_ticks;\n+\n+\tif (timeout)\n+\t\tstart_ticks = rte_get_timer_cycles();\n+\n+\twhile (num < max_num) {\n+\t\tstruct dlb_dequeue_qe qes[DLB_NUM_QES_PER_CACHE_LINE];\n+\t\tint num_avail;\n+\n+\t\t/* Copy up to 4 QEs from the current cache line into qes */\n+\t\tnum_avail = dlb_recv_qe_sparse(qm_port, qes);\n+\n+\t\t/* But don't process more than the user requested */\n+\t\tnum_avail = RTE_MIN(num_avail, max_num - num);\n+\n+\t\tdlb_inc_cq_idx(qm_port, num_avail << 2);\n+\n+\t\tif (num_avail == DLB_NUM_QES_PER_CACHE_LINE)\n+\t\t\tnum += dlb_process_dequeue_four_qes(ev_port,\n+\t\t\t\t\t\t\t qm_port,\n+\t\t\t\t\t\t\t &events[num],\n+\t\t\t\t\t\t\t &qes[0]);\n+\t\telse if (num_avail)\n+\t\t\tnum += dlb_process_dequeue_qes(ev_port,\n+\t\t\t\t\t\t\tqm_port,\n+\t\t\t\t\t\t\t&events[num],\n+\t\t\t\t\t\t\t&qes[0],\n+\t\t\t\t\t\t\tnum_avail);\n+\t\telse if ((timeout == 0) || (num > 0))\n+\t\t\t/* Not waiting in any form, or 1+ events received? */\n+\t\t\tbreak;\n+\t\telse if (dlb_dequeue_wait(dlb, ev_port, qm_port,\n+\t\t\t\t\t timeout, start_ticks))\n+\t\t\tbreak;\n+\t}\n+\n+\tqm_port->owed_tokens += num;\n+\n+\tif (num && qm_port->token_pop_mode == AUTO_POP)\n+\t\tdlb_consume_qe_immediate(qm_port, num);\n+\n+\tev_port->outstanding_releases += num;\n+\n+\treturn num;\n+}\n+\n+static int\n+dlb_event_release(struct dlb_eventdev *dlb, uint8_t port_id, int n)\n+{\n+\tstruct process_local_port_data *port_data;\n+\tstruct dlb_eventdev_port *ev_port;\n+\tstruct dlb_port *qm_port;\n+\tint i;\n+\n+\tif (port_id > dlb->num_ports) {\n+\t\tDLB_LOG_ERR(\"Invalid port id %d in dlb-event_release\\n\",\n+\t\t\t port_id);\n+\t\trte_errno = -EINVAL;\n+\t\treturn rte_errno;\n+\t}\n+\n+\tev_port = &dlb->ev_ports[port_id];\n+\tqm_port = &ev_port->qm_port;\n+\tport_data = &dlb_port[qm_port->id][PORT_TYPE(qm_port)];\n+\n+\ti = 0;\n+\n+\tif (qm_port->is_directed) {\n+\t\ti = n;\n+\t\tgoto sw_credit_update;\n+\t}\n+\n+\twhile (i < n) {\n+\t\tint pop_offs = 0;\n+\t\tint j = 0;\n+\n+\t\t/* Zero-out QEs */\n+\t\tqm_port->qe4[0].cmd_byte = 0;\n+\t\tqm_port->qe4[1].cmd_byte = 0;\n+\t\tqm_port->qe4[2].cmd_byte = 0;\n+\t\tqm_port->qe4[3].cmd_byte = 0;\n+\n+\t\tfor (; j < DLB_NUM_QES_PER_CACHE_LINE && (i + j) < n; j++) {\n+\t\t\tint16_t thresh = qm_port->token_pop_thresh;\n+\n+\t\t\tif (qm_port->token_pop_mode == DELAYED_POP &&\n+\t\t\t qm_port->issued_releases >= thresh - 1) {\n+\t\t\t\t/* Insert the token pop QE */\n+\t\t\t\tdlb_construct_token_pop_qe(qm_port, j);\n+\n+\t\t\t\t/* Reset the releases for the next QE batch */\n+\t\t\t\tqm_port->issued_releases -= thresh;\n+\n+\t\t\t\t/* When using delayed token pop mode, the\n+\t\t\t\t * initial token threshold is the full CQ\n+\t\t\t\t * depth. After the first token pop, we need to\n+\t\t\t\t * reset it to the dequeue_depth.\n+\t\t\t\t */\n+\t\t\t\tqm_port->token_pop_thresh =\n+\t\t\t\t\tqm_port->dequeue_depth;\n+\n+\t\t\t\tpop_offs = 1;\n+\t\t\t\tj++;\n+\t\t\t\tbreak;\n+\t\t\t}\n+\n+\t\t\tqm_port->qe4[j].cmd_byte = DLB_COMP_CMD_BYTE;\n+\t\t\tqm_port->issued_releases++;\n+\t\t}\n+\n+\t\tdlb_hw_do_enqueue(qm_port, i == 0, port_data);\n+\n+\t\t/* Don't include the token pop QE in the release count */\n+\t\ti += j - pop_offs;\n+\t}\n+\n+sw_credit_update:\n+\t/* each release returns one credit */\n+\tif (!ev_port->outstanding_releases) {\n+\t\tDLB_LOG_ERR(\"Unrecoverable application error. Outstanding releases underflowed.\\n\");\n+\t\trte_errno = -ENOTRECOVERABLE;\n+\t\treturn rte_errno;\n+\t}\n+\n+\tev_port->outstanding_releases -= i;\n+\tev_port->inflight_credits += i;\n+\n+\t/* Replenish s/w credits if enough releases are performed */\n+\tdlb_replenish_sw_credits(dlb, ev_port);\n+\treturn 0;\n+}\n+\n+static uint16_t\n+dlb_event_dequeue_burst(void *event_port, struct rte_event *ev, uint16_t num,\n+\t\t\tuint64_t wait)\n+{\n+\tstruct dlb_eventdev_port *ev_port = event_port;\n+\tstruct dlb_port *qm_port = &ev_port->qm_port;\n+\tstruct dlb_eventdev *dlb = ev_port->dlb;\n+\tuint16_t cnt;\n+\tint ret;\n+\n+\trte_errno = 0;\n+\n+\tRTE_ASSERT(ev_port->setup_done);\n+\tRTE_ASSERT(ev != NULL);\n+\n+\tif (ev_port->implicit_release && ev_port->outstanding_releases > 0) {\n+\t\tuint16_t out_rels = ev_port->outstanding_releases;\n+\n+\t\tret = dlb_event_release(dlb, ev_port->id, out_rels);\n+\t\tif (ret)\n+\t\t\treturn(ret);\n+\n+\t\tDLB_INC_STAT(ev_port->stats.tx_implicit_rel, out_rels);\n+\t}\n+\n+\tif (qm_port->token_pop_mode == DEFERRED_POP &&\n+\t\t\tqm_port->owed_tokens)\n+\t\tdlb_consume_qe_immediate(qm_port, qm_port->owed_tokens);\n+\n+\tcnt = dlb_hw_dequeue(dlb, ev_port, ev, num, wait);\n+\n+\tDLB_INC_STAT(ev_port->stats.traffic.total_polls, 1);\n+\tDLB_INC_STAT(ev_port->stats.traffic.zero_polls, ((cnt == 0) ? 1 : 0));\n+\treturn cnt;\n+}\n+\n+static uint16_t\n+dlb_event_dequeue(void *event_port, struct rte_event *ev, uint64_t wait)\n+{\n+\treturn dlb_event_dequeue_burst(event_port, ev, 1, wait);\n+}\n+\n+static uint16_t\n+dlb_event_dequeue_burst_sparse(void *event_port, struct rte_event *ev,\n+\t\t\t uint16_t num, uint64_t wait)\n+{\n+\tstruct dlb_eventdev_port *ev_port = event_port;\n+\tstruct dlb_port *qm_port = &ev_port->qm_port;\n+\tstruct dlb_eventdev *dlb = ev_port->dlb;\n+\tuint16_t cnt;\n+\tint ret;\n+\n+\trte_errno = 0;\n+\n+\tRTE_ASSERT(ev_port->setup_done);\n+\tRTE_ASSERT(ev != NULL);\n+\n+\tif (ev_port->implicit_release && ev_port->outstanding_releases > 0) {\n+\t\tuint16_t out_rels = ev_port->outstanding_releases;\n+\n+\t\tret = dlb_event_release(dlb, ev_port->id, out_rels);\n+\t\tif (ret)\n+\t\t\treturn(ret);\n+\n+\t\tDLB_INC_STAT(ev_port->stats.tx_implicit_rel, out_rels);\n+\t}\n+\n+\tif (qm_port->token_pop_mode == DEFERRED_POP &&\n+\t qm_port->owed_tokens)\n+\t\tdlb_consume_qe_immediate(qm_port, qm_port->owed_tokens);\n+\n+\tcnt = dlb_hw_dequeue_sparse(dlb, ev_port, ev, num, wait);\n+\n+\tDLB_INC_STAT(ev_port->stats.traffic.total_polls, 1);\n+\tDLB_INC_STAT(ev_port->stats.traffic.zero_polls, ((cnt == 0) ? 1 : 0));\n+\treturn cnt;\n+}\n+\n+static uint16_t\n+dlb_event_dequeue_sparse(void *event_port, struct rte_event *ev, uint64_t wait)\n+{\n+\treturn dlb_event_dequeue_burst_sparse(event_port, ev, 1, wait);\n+}\n+\n void\n dlb_entry_points_init(struct rte_eventdev *dev)\n {\n+\tstruct dlb_eventdev *dlb;\n+\n \tstatic struct rte_eventdev_ops dlb_eventdev_entry_ops = {\n \t\t.dev_infos_get = dlb_eventdev_info_get,\n \t\t.dev_configure = dlb_eventdev_configure,\n@@ -2915,7 +3670,15 @@ dlb_entry_points_init(struct rte_eventdev *dev)\n \tdev->enqueue_burst = dlb_event_enqueue_burst;\n \tdev->enqueue_new_burst = dlb_event_enqueue_new_burst;\n \tdev->enqueue_forward_burst = dlb_event_enqueue_forward_burst;\n+\tdev->dequeue = dlb_event_dequeue;\n+\tdev->dequeue_burst = dlb_event_dequeue_burst;\n \n+\tdlb = dev->data->dev_private;\n+\n+\tif (dlb->poll_mode == DLB_CQ_POLL_MODE_SPARSE) {\n+\t\tdev->dequeue = dlb_event_dequeue_sparse;\n+\t\tdev->dequeue_burst = dlb_event_dequeue_burst_sparse;\n+\t}\n }\n \n int\n", "prefixes": [ "v4", "16/22" ] }