hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/IC5EHB ----------------------------------------- Adapt 910b npu driver for xsched Signed-off-by: Hui Tang <tanghui20@huawei.com> Signed-off-by: Konstantin Meskhidze <konstantin.meskhidze@huawei.com> Signed-off-by: Liu Kai <liukai284@huawei.com> Signed-off-by: Xia Fukun <xiafukun@huawei.com> --- .../0001-Adapt-910b-npu-driver-for-xsched.txt | 917 ++++++++++++++++++ 1 file changed, 917 insertions(+) create mode 100644 drivers/xcu/0001-Adapt-910b-npu-driver-for-xsched.txt diff --git a/drivers/xcu/0001-Adapt-910b-npu-driver-for-xsched.txt b/drivers/xcu/0001-Adapt-910b-npu-driver-for-xsched.txt new file mode 100644 index 000000000000..1ed6726dffec --- /dev/null +++ b/drivers/xcu/0001-Adapt-910b-npu-driver-for-xsched.txt @@ -0,0 +1,917 @@ +From d9c4001d6b8ac3b19a8e24d4fafb5fa3285733f5 Mon Sep 17 00:00:00 2001 +From: Hui Tang <tanghui20@huawei.com> +Date: Tue, 25 Feb 2025 10:18:24 +0000 +Subject: [PATCH openEuler-25.03] Adapt 910b npu driver for xsched + +hulk inclusion +category: feature +bugzilla: https://gitee.com/openeuler/kernel/issues/IC5EHB + +----------------------------------------- + +Adapt 910b npu driver for xsched + +Signed-off-by: Hui Tang <tanghui20@huawei.com> +Signed-off-by: Konstantin Meskhidze <konstantin.meskhidze@huawei.com> +Signed-off-by: Liu Kai <liukai284@huawei.com> +Signed-off-by: Xia Fukun <xiafukun@huawei.com> +--- + .../depends/inc_driver/ascend_hal_define.h | 2 +- + rms/trs_drv/chan/chan_rxtx.c | 2 +- + .../lba/near/comm/adapt/trs_near_adapt_init.h | 2 + + rms/trs_drv/lba/near/sia/adapt/Makefile | 7 +- + .../lba/near/sia/adapt/trs_host_init.c | 3 + + .../lba/near/sia/adapt/xsched_xpu_interface.c | 262 ++++++++++++++++++ + rms/trs_drv/trs_core/Makefile | 1 + + rms/trs_drv/trs_core/trs_fops.c | 143 +++++++--- + rms/trs_drv/trs_core/trs_hw_sqcq.c | 3 +- + rms/trs_drv/trs_core/trs_hw_sqcq.h | 3 +- + rms/trs_drv/trs_core/trs_logic_cq.c | 100 ++++--- + rms/trs_drv/trs_core/trs_logic_cq.h | 3 +- + rms/trs_drv/trs_core/trs_sqcq_map.c | 4 + + ts_agent/src/ts_agent_update_sqe.c | 6 + + 14 files changed, 465 insertions(+), 76 deletions(-) + create mode 100755 rms/trs_drv/lba/near/sia/adapt/xsched_xpu_interface.c + +diff --git a/dev_inc_open/inc/depends/inc_driver/ascend_hal_define.h b/dev_inc_open/inc/depends/inc_driver/ascend_hal_define.h +index a76efda..cc51c4d 100644 +--- a/dev_inc_open/inc/depends/inc_driver/ascend_hal_define.h ++++ b/dev_inc_open/inc/depends/inc_driver/ascend_hal_define.h +@@ -893,7 +893,7 @@ typedef enum tagDrvSqCqType { + } drvSqCqType_t; + + struct halSqCqInputInfo { +- drvSqCqType_t type; // normal : 0, callback : 1 ++ drvSqCqType_t type; // normal : 0, callback : 1, logic : 2 + uint32_t tsId; + /* The size and depth of each cqsq can be configured in normal mode, but this function is not yet supported */ + uint32_t sqeSize; // normal : 64Byte +diff --git a/rms/trs_drv/chan/chan_rxtx.c b/rms/trs_drv/chan/chan_rxtx.c +index 1fc72da..1e4ef38 100755 +--- a/rms/trs_drv/chan/chan_rxtx.c ++++ b/rms/trs_drv/chan/chan_rxtx.c +@@ -156,7 +156,7 @@ static int trs_chan_fill_sqe(struct trs_chan *chan, u8 *sqe, int timeout, int ad + /* if using bar to r/w sqe, it should use stack value to store sqe to avoid waster time */ + sqe_addr = trs_chan_mem_is_local_mem(&sq->mem_attr) ? dst_addr : sqe_tmp; + +- if (addr_domain == CHAN_ADDR_DOMAIN_KERNEL) { ++ if (addr_domain == CHAN_ADDR_DOMAIN_KERNEL || !access_ok(sqe, sq->para.sqe_size)) { + memcpy_s(sqe_addr, sq->para.sqe_size, sqe, sq->para.sqe_size); + } else { + ret_cpy = copy_from_user(sqe_addr, sqe, sq->para.sqe_size); +diff --git a/rms/trs_drv/lba/near/comm/adapt/trs_near_adapt_init.h b/rms/trs_drv/lba/near/comm/adapt/trs_near_adapt_init.h +index 3a60d1d..7a240a6 100755 +--- a/rms/trs_drv/lba/near/comm/adapt/trs_near_adapt_init.h ++++ b/rms/trs_drv/lba/near/comm/adapt/trs_near_adapt_init.h +@@ -21,4 +21,6 @@ + void trs_ts_adapt_init(struct trs_id_inst *inst); + void trs_ts_adapt_uninit(struct trs_id_inst *inst); + ++int xsched_xcu_group_init(u32 dev_id, u32 ts_num); ++ + #endif /* TRS_NEAR_ADAPT_INIT_H */ +diff --git a/rms/trs_drv/lba/near/sia/adapt/Makefile b/rms/trs_drv/lba/near/sia/adapt/Makefile +index 16a3f05..2cbdd43 100755 +--- a/rms/trs_drv/lba/near/sia/adapt/Makefile ++++ b/rms/trs_drv/lba/near/sia/adapt/Makefile +@@ -59,8 +59,13 @@ EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/rms/trs_drv/lba/near/sia/adapt/comm/tsc + EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/dms/include/ + EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/dms/config/ + EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/dbl/dev_urd/ ++EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/rms/trs_drv/trs_core ++EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/dev/inc/ ++EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/drv_davinci_intf_host ++EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/tsch/ ++ + obj-m += ascend_trs_pm_adapt.o +-ascend_trs_pm_adapt-objs := trs_host_init.o near_comm/trs_host_comm.o near_comm/trs_host_db.o near_comm/trs_host_id.o near_comm/trs_host_mbox.o near_comm/trs_host_msg.o near_comm/trs_near_adapt_init.o ++ascend_trs_pm_adapt-objs := trs_host_init.o near_comm/trs_host_comm.o near_comm/trs_host_db.o near_comm/trs_host_id.o near_comm/trs_host_mbox.o near_comm/trs_host_msg.o near_comm/trs_near_adapt_init.o xsched_xpu_interface.o + + ascend_trs_pm_adapt-objs += near_comm/soc_adapt/soc_adapt.o trs_host_init/trs_host.o trs_host_chan/trs_host_chan.o trs_host_chan/trs_sqe_update.o trs_host_core/trs_host_core.o + ascend_trs_pm_adapt-objs += near_comm/trs_host_chan/stars_v1/trs_chan_stars_v1_ops.o near_comm/trs_host_chan/stars_v1/trs_chan_stars_v1_ops_stars.o +diff --git a/rms/trs_drv/lba/near/sia/adapt/trs_host_init.c b/rms/trs_drv/lba/near/sia/adapt/trs_host_init.c +index abdabc6..4c270f5 100755 +--- a/rms/trs_drv/lba/near/sia/adapt/trs_host_init.c ++++ b/rms/trs_drv/lba/near/sia/adapt/trs_host_init.c +@@ -17,6 +17,7 @@ + #include <linux/types.h> + #include <linux/init.h> + #include <linux/module.h> ++#include <linux/xcu_group.h> + + #include "soc_res.h" + #include "trs_pub_def.h" +@@ -162,6 +163,8 @@ int trs_host_init(u32 phy_devid) + } + } + ++ xsched_xcu_group_init(phy_devid, ts_num); ++ + return 0; + } + EXPORT_SYMBOL(trs_host_init); +diff --git a/rms/trs_drv/lba/near/sia/adapt/xsched_xpu_interface.c b/rms/trs_drv/lba/near/sia/adapt/xsched_xpu_interface.c +new file mode 100755 +index 0000000..df7167d +--- /dev/null ++++ b/rms/trs_drv/lba/near/sia/adapt/xsched_xpu_interface.c +@@ -0,0 +1,262 @@ ++/* ++ * Copyright (c) Huawei Technologies Co., Ltd. 2019-2020. All rights reserved. ++ * ++ * This program is free software; you can redistribute it and/or modify ++ * it under the terms of the GNU General Public License version 2 and ++ * only 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. ++ * ++ * Description: ++ * Author: Huawei ++ * Create: 2024-06-17 ++ */ ++ ++#ifndef TSDRV_KERNEL_UT ++#include <linux/cdev.h> ++#include <linux/slab.h> ++#include <linux/spinlock.h> ++#include <linux/errno.h> ++#include <linux/list.h> ++#include <linux/platform_device.h> ++#include <linux/vmalloc.h> ++#include <linux/xcu_group.h> ++#include <linux/file.h> ++ ++#include "securec.h" ++#include "devdrv_manager_comm.h" ++#include "ascend_hal_define.h" ++#include "trs_pub_def.h" ++#include "trs_res_id_def.h" ++#include "trs_proc.h" ++#include "trs_cmd.h" ++#include "davinci_api.h" ++#include "davinci_interface.h" ++#include "davinci_intf_init.h" ++#include "task_struct.h" ++ ++int ioctl_trs_sqcq_send(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg); ++int ioctl_trs_sqcq_alloc(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg); ++int ioctl_trs_sqcq_free(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg); ++int ioctl_trs_sqcq_recv(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg); ++ ++extern struct xcu_group *xcu_group_root; ++extern int xsched_xcu_register(struct xcu_group *group, uint32_t phys_id); ++ ++/* Gets device driver TS context from a file descriptor of opened device. */ ++static void *get_tsdrv_ctx(int fd) ++{ ++ struct davinci_intf_private_stru *file_private_data; ++ void *ctx = NULL; ++ struct fd f; ++ ++ f = fdget(fd); ++ if (!f.file) ++ goto out; ++ ++ file_private_data = f.file->private_data; ++ if (!file_private_data) ++ goto out; ++ ++ ctx = file_private_data->priv_filep.private_data; ++ ++out: ++ fdput(f); ++ return ctx; ++} ++ ++int trs_xsched_ctx_run(struct xcu_op_handler_params *params) ++{ ++ uint32_t sq_id = *(uint32_t *)params->param_1; ++ uint32_t tsId = *(uint32_t *)params->param_2; ++ uint8_t *sqe_addr = params->param_3; ++ uint32_t sqe_num = *(uint32_t *)params->param_4; ++ int32_t timeout = *(int32_t *)params->param_5; ++ int32_t type = *(int32_t *)params->param_6; ++ struct halTaskSendInfo input = {0}; ++ struct trs_proc_ctx *ctx = params->param_7; ++ uint32_t logic_cqId = *(uint32_t *)params->param_8; ++ ++ input.tsId = tsId; ++ input.sqId = sq_id; ++ input.timeout = timeout; ++ input.sqe_addr = sqe_addr; ++ input.sqe_num = sqe_num; ++ input.type = type; ++ ++ trs_debug("%s %d: tsId %u sqId %u timeout %d num %u\n", ++ __FUNCTION__, __LINE__, tsId, sq_id, timeout, sqe_num); ++ ++ /* Send SQ tail to a doorbel. */ ++ return ioctl_trs_sqcq_send(ctx, logic_cqId, (unsigned long)&input);; ++} ++ ++int trs_xsched_ctx_free(struct xcu_op_handler_params *params) ++{ ++ struct trs_proc_ctx *ctx; ++ ++ ctx = get_tsdrv_ctx(params->fd); ++ if (!ctx) ++ return -ENOENT; ++ ++ return ioctl_trs_sqcq_free(ctx, 0, (unsigned long)params->payload); ++} ++ ++int trs_xsched_ctx_wait(struct xcu_op_handler_params *params) ++{ ++ uint32_t tsId = *(uint32_t *)params->param_1; ++ uint32_t cqId = *(uint32_t *)params->param_2; ++ uint32_t streamId = *(uint32_t *)params->param_3; ++ struct ts_stars_sqe_header *sqe = params->param_4; ++ uint8_t *cqe_addr = params->param_5; ++ struct trs_proc_ctx *ctx = params->param_6; ++ int32_t timeout = *(uint32_t *)params->param_7; ++ int32_t cqe_num = 1; ++ struct halReportRecvInfo input = {0}; ++ uint32_t task_id = sqe->task_id; ++ ++ input.type = DRV_LOGIC_TYPE; ++ input.tsId = tsId; ++ input.cqId = cqId; ++ input.timeout = timeout; ++ input.cqe_num = cqe_num; ++ input.cqe_addr = cqe_addr; ++ input.stream_id = streamId; ++ input.task_id = task_id; ++ input.res[0] = 1; /* version 1 for new runtime. */ ++ ++ trs_debug("%s %d: tdId %u logic_cqId %u streamid %u task_id %d timeout %d \n", ++ __FUNCTION__, __LINE__, tsId, cqId, streamId, task_id, timeout); ++ ++ /* Wait for cq irq and read result. */ ++ return ioctl_trs_sqcq_recv(ctx, 0, (unsigned long)&input); ++} ++ ++int trs_xsched_ctx_complete(struct xcu_op_handler_params *params) ++{ ++ return 0; ++} ++ ++int trs_xsched_ctx_alloc(struct xcu_op_handler_params *params) ++{ ++ struct halSqCqInputInfo *input_info = params->payload; ++ uint32_t *tgid = (uint32_t *)params->param_1; ++ uint32_t *sq_id = (uint32_t *)params->param_2; ++ uint32_t *cq_id = (uint32_t *)params->param_3; ++ uint32_t *user_stream_id = (uint32_t *)params->param_4; ++ struct trs_proc_ctx *ctx; ++ int ret = 0; ++ ++ trs_debug("%s %d, input_info %lx, type: %d\n", ++ __FUNCTION__, __LINE__, (unsigned long)input_info, input_info->type); ++ ++ ctx = get_tsdrv_ctx(params->fd); ++ if (!ctx) ++ return -ENOENT; ++ ++ trs_debug("%s %d, pid %d, task_id %d, size %ld\n", ++ __FUNCTION__, __LINE__, ctx->pid, ctx->task_id, sizeof(*ctx)); ++ ret = ioctl_trs_sqcq_alloc(ctx, 0, (unsigned long)input_info); ++ if (ret != 0) ++ return ret; ++ ++ *tgid = ctx->pid; ++ *sq_id = input_info->sqId; ++ *cq_id = input_info->cqId; ++ *user_stream_id = input_info->info[0]; ++ params->param_5 = ctx; ++ return 0; ++} ++ ++int trs_xsched_ctx_logic_alloc(struct xcu_op_handler_params *params) ++{ ++ struct halSqCqInputInfo *input_info = params->payload; ++ uint32_t *logic_cq_id = (uint32_t *)params->param_1; ++ struct trs_proc_ctx *ctx; ++ int ret = 0; ++ ++ trs_debug("%s %d, type: %d\n", __FUNCTION__, __LINE__, input_info->type); ++ ++ ctx = get_tsdrv_ctx(params->fd); ++ if (!ctx) ++ return -ENOENT; ++ ++ trs_debug("%s %d, pid %d, task_id %d, size %ld\n", ++ __FUNCTION__, __LINE__, ctx->pid, ctx->task_id, sizeof(*ctx)); ++ ++ ret = ioctl_trs_sqcq_alloc(ctx, 0, (unsigned long)input_info); ++ if (ret != 0) ++ return ret; ++ ++ *logic_cq_id = input_info->cqId; ++ trs_debug("%s %d, type: %d, cq_id: %u\n", ++ __FUNCTION__, __LINE__, input_info->type, *logic_cq_id); ++ return 0; ++} ++ ++int trs_xsched_ctx_sqe_op(struct xcu_op_handler_params *params) ++{ ++ struct ts_stars_sqe_header *sqe = params->param_2; ++ int op_type = *(int *)(params->param_1); ++ ++ switch (op_type) { ++ case SQE_IS_NOTIFY: ++ return (sqe->type == 0) && (sqe->wr_cqe == 1); ++ case SQE_SET_NOTIFY: ++ if (sqe->type == 0) ++ sqe->wr_cqe = 1; ++ break; ++ default: ++ break; ++ } ++ ++ return 0; ++} ++ ++static struct xcu_operation trs_xsched_ctx_xcu_ops = { ++ .run = trs_xsched_ctx_run, ++ .finish = trs_xsched_ctx_free, ++ .wait = trs_xsched_ctx_wait, ++ .complete = trs_xsched_ctx_complete, ++ .alloc = trs_xsched_ctx_alloc, ++ .logic_alloc = trs_xsched_ctx_logic_alloc, ++ .sqe_op = trs_xsched_ctx_sqe_op, ++}; ++ ++int xsched_xcu_group_init(u32 dev_id, u32 ts_num) ++{ ++ struct xcu_group *type_group; ++ struct xcu_group *dev_group; ++ struct xcu_group *ts_group; ++ int tsid; ++ ++ trs_debug("dev_id %u ts_num %u\n", dev_id, ts_num); ++ type_group = xcu_group_find(xcu_group_root, XCU_TYPE_XPU); ++ ++ if (!type_group) { ++ type_group = xcu_group_init(XCU_TYPE_XPU); ++ xcu_group_attach(type_group, xcu_group_root); ++ } ++ ++ dev_group = xcu_group_init(dev_id); ++ ++ trs_debug("%s %d deviceid is %d\n", __FUNCTION__, __LINE__, dev_id); ++ dev_group->id = dev_id; ++ xcu_group_attach(dev_group, type_group); ++ ++ for (tsid = 0; tsid < ts_num; tsid++) { ++ ts_group = xcu_group_init(tsid); ++ ts_group->opt = &trs_xsched_ctx_xcu_ops; ++ ++ xcu_group_attach(ts_group, dev_group); ++ xsched_xcu_register(ts_group, dev_id); ++ ++ cond_resched(); ++ } ++ ++ return 0; ++} ++#endif +diff --git a/rms/trs_drv/trs_core/Makefile b/rms/trs_drv/trs_core/Makefile +index e0a6a55..8d27ad9 100755 +--- a/rms/trs_drv/trs_core/Makefile ++++ b/rms/trs_drv/trs_core/Makefile +@@ -41,6 +41,7 @@ endif + + EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/rms/trs_drv/inc + EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/rms/trs_drv/trs_core ++EXTRA_CFLAGS += -I$(DRIVER_SRC_BASE_DIR)/tsch/ + + obj-m += ascend_trs_core.o + ascend_trs_core-objs := trs_fops.o trs_ts_inst.o trs_proc.o trs_res_mng.o trs_sqcq_map.o trs_hw_sqcq.o trs_sw_sqcq.o trs_logic_cq.o trs_cb_sqcq.o trs_shm_sqcq.o trs_proc_fs.o +diff --git a/rms/trs_drv/trs_core/trs_fops.c b/rms/trs_drv/trs_core/trs_fops.c +index e5702d2..1a9b3c7 100755 +--- a/rms/trs_drv/trs_core/trs_fops.c ++++ b/rms/trs_drv/trs_core/trs_fops.c +@@ -21,6 +21,7 @@ + #include <linux/mm.h> + #include <linux/sched/mm.h> + #include <linux/version.h> ++#include <linux/vstream.h> + + #include "ascend_hal_define.h" + +@@ -33,6 +34,8 @@ + #include "trs_ts_inst.h" + #include "trs_cmd.h" + #include "trs_fops.h" ++#include "trs_logic_cq.h" ++#include "task_struct.h" + + static int (*const trs_res_id_handles[TRS_MAX_CMD])(struct trs_proc_ctx *proc_ctx, + struct trs_core_ts_inst *ts_inst, struct trs_res_id_para *para) = { +@@ -195,8 +198,17 @@ static int (*const trs_sqcq_alloc_handles[DRV_INVALID_TYPE])(struct trs_proc_ctx + [DRV_CTRL_TYPE] = trs_sw_sqcq_alloc + }; + +-static int ioctl_trs_sqcq_alloc(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) ++static bool is_xsched_used(void __user *ptr, int size) + { ++ if (access_ok(ptr, size)) ++ return false; ++ ++ return true; ++} ++ ++int ioctl_trs_sqcq_alloc(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) ++{ ++ bool xsched_used = is_xsched_used((void __user *)arg, sizeof(struct halSqCqInputInfo)); + struct trs_core_ts_inst *ts_inst = NULL; + struct halSqCqInputInfo para; + struct trs_alloc_para *alloc_para = NULL; +@@ -204,10 +216,14 @@ static int ioctl_trs_sqcq_alloc(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + struct trs_uio_info uio_info; + int ret; + +- ret = copy_from_user(¶, (struct halSqCqInputInfo __user *)arg, sizeof(para)); +- if (ret != 0) { +- trs_err("Copy from user failed. (ret=%d)\n", ret); +- return ret; ++ if (xsched_used) { ++ memcpy(¶, (struct halSqCqInputInfo *)arg, sizeof(para)); ++ } else { ++ ret = copy_from_user(¶, (struct halSqCqInputInfo __user *)arg, sizeof(para)); ++ if (ret != 0) { ++ trs_err("Copy from user failed. (ret=%d)\n", ret); ++ return ret; ++ } + } + + alloc_para = get_alloc_para_addr(¶); +@@ -238,15 +254,24 @@ static int ioctl_trs_sqcq_alloc(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + trs_core_inst_put(ts_inst); + + if (ret == 0) { +- ret = copy_to_user((struct halSqCqInputInfo __user *)arg, ¶, sizeof(para)); +- ret |= copy_to_user((struct trs_uio_info __user *)user_uio_info, &uio_info, sizeof(uio_info)); +- if (ret != 0) { +- trs_err("Copy to user failed. (ret=%d)\n", ret); ++ if (xsched_used) { ++ memcpy((struct halSqCqInputInfo *)arg, ¶, sizeof(para)); ++ ret = copy_to_user((struct trs_uio_info __user *)user_uio_info, &uio_info, sizeof(uio_info)); ++ if (ret != 0) { ++ trs_err("Copy to user failed. (ret=%d)\n", ret); ++ } ++ } else { ++ ret = copy_to_user((struct halSqCqInputInfo __user *)arg, ¶, sizeof(para)); ++ ret |= copy_to_user((struct trs_uio_info __user *)user_uio_info, &uio_info, sizeof(uio_info)); ++ if (ret != 0) { ++ trs_err("Copy to user failed. (ret=%d)\n", ret); ++ } + } + } + + return ret; + } ++EXPORT_SYMBOL(ioctl_trs_sqcq_alloc); + + static int (*const trs_sqcq_free_handles[DRV_INVALID_TYPE])(struct trs_proc_ctx *proc_ctx, + struct trs_core_ts_inst *ts_inst, struct halSqCqFreeInfo *para) = { +@@ -257,16 +282,20 @@ static int (*const trs_sqcq_free_handles[DRV_INVALID_TYPE])(struct trs_proc_ctx + [DRV_CTRL_TYPE] = trs_sw_sqcq_free + }; + +-static int ioctl_trs_sqcq_free(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) ++int ioctl_trs_sqcq_free(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) + { + struct trs_core_ts_inst *ts_inst = NULL; + struct halSqCqFreeInfo para; + int ret; + +- ret = copy_from_user(¶, (struct halSqCqFreeInfo __user *)arg, sizeof(para)); +- if (ret != 0) { +- trs_err("Copy from user failed. (ret=%d)\n", ret); +- return ret; ++ if (is_xsched_used((void *)arg, sizeof(struct halSqCqFreeInfo))) { ++ memcpy(¶, (struct halSqCqFreeInfo *)arg, sizeof(para)); ++ } else { ++ ret = copy_from_user(¶, (struct halSqCqFreeInfo __user *)arg, sizeof(para)); ++ if (ret != 0) { ++ trs_err("Copy from user failed. (ret=%d)\n", ret); ++ return ret; ++ } + } + + if ((para.type < 0) || (para.type >= DRV_INVALID_TYPE)) { +@@ -287,6 +316,7 @@ static int ioctl_trs_sqcq_free(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + trs_core_inst_put(ts_inst); + return ret; + } ++EXPORT_SYMBOL(ioctl_trs_sqcq_free); + + static int ioctl_trs_sqcq_config(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) + { +@@ -362,17 +392,26 @@ static int (*const trs_sqcq_send_handles[DRV_INVALID_TYPE])(struct trs_proc_ctx + [DRV_CALLBACK_TYPE] = trs_cb_sqcq_send, + }; + +-static int ioctl_trs_sqcq_send(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) ++int ioctl_trs_sqcq_send(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) + { + struct trs_core_ts_inst *ts_inst = NULL; ++ struct halTaskSendInfo *kern_para = (struct halTaskSendInfo *)arg; + struct halTaskSendInfo __user *usr_para = (struct halTaskSendInfo __user *)arg; + struct halTaskSendInfo para; ++ struct trs_logic_cq *logic_cq = NULL; ++ struct ts_stars_sqe_header *sqe_header = NULL; ++ uint32_t logic_cqId = cmd; ++ bool xsched_used = is_xsched_used((void *)arg, sizeof(struct halTaskSendInfo)); + int ret; + +- ret = copy_from_user(¶, usr_para, sizeof(para)); +- if (ret != 0) { +- trs_err("Copy from user failed. (ret=%d)\n", ret); +- return ret; ++ if (xsched_used) { ++ memcpy(¶, (struct halTaskSendInfo *)arg, sizeof(para)); ++ } else { ++ ret = copy_from_user(¶, usr_para, sizeof(para)); ++ if (ret != 0) { ++ trs_err("Copy from user failed. (ret=%d)\n", ret); ++ return ret; ++ } + } + + if ((para.type < 0) || (para.type >= DRV_INVALID_TYPE) || (trs_sqcq_send_handles[para.type] == NULL) || +@@ -387,37 +426,69 @@ static int ioctl_trs_sqcq_send(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + return -EINVAL; + } + ++ if (xsched_used) { ++ logic_cq = &ts_inst->logic_cq_ctx.cq[logic_cqId]; ++ if (logic_cq == NULL) { ++ trs_err("Invalid para. (logic_cqId=%u)\n", logic_cqId); ++ return -EINVAL; ++ } ++ ++ sqe_header = (struct ts_stars_sqe_header *)para.sqe_addr; ++ trs_debug("sqe_header->type=%u logic_cqId=%u stream_id=%u task_id=%u\n", ++ sqe_header->type, logic_cqId, sqe_header->rt_stream_id, sqe_header->task_id); ++ ++ if ((sqe_header->type == 0) && (sqe_header->wr_cqe == 1)) { ++ trs_debug("logic_cq->wakeup_num=%u\n", atomic_read(&logic_cq->wakeup_num)); ++ ++ if (atomic_read(&logic_cq->wakeup_num) > 0) { ++ atomic_dec(&logic_cq->wakeup_num); ++ trs_debug("logic_cq->wakeup_num=%u\n", atomic_read(&logic_cq->wakeup_num)); ++ } ++ } ++ } ++ + ret = trs_sqcq_send_handles[para.type](proc_ctx, ts_inst, ¶); + + trs_core_inst_put(ts_inst); + + if ((ret == 0) && (para.type == DRV_NORMAL_TYPE)) { +- ret = put_user(para.pos, &usr_para->pos); +- if (ret != 0) { +- trs_err("Put to user fail. (devid=%u; tsid=%u; sqId=%u)\n", proc_ctx->devid, para.tsId, para.sqId); ++ if (xsched_used) { ++ kern_para->pos = para.pos; ++ } else { ++ ret = put_user(para.pos, &usr_para->pos); ++ if (ret != 0) { ++ trs_err("Put to user fail. (devid=%u; tsid=%u; sqId=%u)\n", proc_ctx->devid, para.tsId, para.sqId); ++ } + } + } + + return ret; + } ++EXPORT_SYMBOL(ioctl_trs_sqcq_send); + + static int (*const trs_sqcq_recv_handles[DRV_INVALID_TYPE])(struct trs_proc_ctx *proc_ctx, +- struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para) = { ++ struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para, bool is_xsched) = { + [DRV_NORMAL_TYPE] = trs_hw_sqcq_recv, + [DRV_LOGIC_TYPE] = trs_logic_cq_recv, + }; + +-static int ioctl_trs_sqcq_recv(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) ++int ioctl_trs_sqcq_recv(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) + { + struct trs_core_ts_inst *ts_inst = NULL; +- struct halReportRecvInfo *usr_para = (struct halReportRecvInfo __user *)arg; ++ struct halReportRecvInfo *kern_para = (struct halReportRecvInfo *)arg; ++ struct halReportRecvInfo __user *usr_para = (struct halReportRecvInfo __user *)arg; + struct halReportRecvInfo para; + int ret; ++ bool xsched_used = is_xsched_used((void *)arg, sizeof(struct halReportRecvInfo)); + +- ret = copy_from_user(¶, usr_para, sizeof(para)); +- if (ret != 0) { +- trs_err("Copy from user failed. (ret=%d)\n", ret); +- return ret; ++ if (xsched_used) { ++ memcpy(¶, (struct halReportRecvInfo *)arg, sizeof(para)); ++ } else { ++ ret = copy_from_user(¶, usr_para, sizeof(para)); ++ if (ret != 0) { ++ trs_err("Copy from user failed. (ret=%d)\n", ret); ++ return ret; ++ } + } + + if ((para.type < 0) || (para.type >= DRV_INVALID_TYPE) || (trs_sqcq_recv_handles[para.type] == NULL) || +@@ -432,11 +503,16 @@ static int ioctl_trs_sqcq_recv(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + return -EINVAL; + } + +- ret = trs_sqcq_recv_handles[para.type](proc_ctx, ts_inst, ¶); ++ ret = trs_sqcq_recv_handles[para.type](proc_ctx, ts_inst, ¶, xsched_used); ++ + if (ret == 0) { +- ret = put_user(para.report_cqe_num, &usr_para->report_cqe_num); +- if (ret != 0) { +- trs_err("Put to user fail. (devid=%u; tsid=%u; cqId=%u)\n", proc_ctx->devid, para.tsId, para.cqId); ++ if (xsched_used) { ++ kern_para->report_cqe_num = para.report_cqe_num; ++ } else { ++ ret = put_user(para.report_cqe_num, &usr_para->report_cqe_num); ++ if (ret != 0) { ++ trs_err("Put to user fail. (devid=%u; tsid=%u; cqId=%u)\n", proc_ctx->devid, para.tsId, para.cqId); ++ } + } + } else { + u32 ts_status; +@@ -449,6 +525,7 @@ static int ioctl_trs_sqcq_recv(struct trs_proc_ctx *proc_ctx, unsigned int cmd, + + return ret; + } ++EXPORT_SYMBOL(ioctl_trs_sqcq_recv); + + int ioctl_trs_stl_bind(struct trs_proc_ctx *proc_ctx, unsigned int cmd, unsigned long arg) + { +diff --git a/rms/trs_drv/trs_core/trs_hw_sqcq.c b/rms/trs_drv/trs_core/trs_hw_sqcq.c +index 825d603..10f3903 100755 +--- a/rms/trs_drv/trs_core/trs_hw_sqcq.c ++++ b/rms/trs_drv/trs_core/trs_hw_sqcq.c +@@ -1160,7 +1160,8 @@ int trs_hw_sqcq_send(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_ + return ret; + } + +-int trs_hw_sqcq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para) ++int trs_hw_sqcq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para, ++ bool is_xsched) + { + struct trs_id_inst *inst = &ts_inst->inst; + struct trs_chan_recv_para recv_para; +diff --git a/rms/trs_drv/trs_core/trs_hw_sqcq.h b/rms/trs_drv/trs_core/trs_hw_sqcq.h +index b32cd64..b6affdf 100755 +--- a/rms/trs_drv/trs_core/trs_hw_sqcq.h ++++ b/rms/trs_drv/trs_core/trs_hw_sqcq.h +@@ -32,7 +32,8 @@ int trs_sqcq_config(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_i + int trs_sqcq_query(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halSqCqQueryInfo *para); + + int trs_hw_sqcq_send(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halTaskSendInfo *para); +-int trs_hw_sqcq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para); ++int trs_hw_sqcq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para, ++ bool is_xsched); + void trs_proc_diable_sq_status(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, + int res_type, u32 res_id); + void trs_hw_sqcq_recycle(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, int res_type, u32 res_id); +diff --git a/rms/trs_drv/trs_core/trs_logic_cq.c b/rms/trs_drv/trs_core/trs_logic_cq.c +index d35b8d5..72cf64a 100755 +--- a/rms/trs_drv/trs_core/trs_logic_cq.c ++++ b/rms/trs_drv/trs_core/trs_logic_cq.c +@@ -265,13 +265,15 @@ static bool trs_logic_is_cqe_match(struct trs_logic_cq *logic_cq, void *cqe, u32 + } + + static int trs_logic_cq_recv_para_check(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, +- struct halReportRecvInfo *para) ++ struct halReportRecvInfo *para, bool is_xsched) + { + struct trs_id_inst *inst = &ts_inst->inst; + +- if (!trs_proc_has_res(proc_ctx, ts_inst, TRS_LOGIC_CQ, para->cqId)) { +- trs_err("Not proc owner cq. (devid=%u; tsid=%u; logic_cqid=%u)\n", inst->devid, inst->tsid, para->cqId); +- return -EINVAL; ++ if (!is_xsched) { ++ if (!trs_proc_has_res(proc_ctx, ts_inst, TRS_LOGIC_CQ, para->cqId)) { ++ trs_err("Not proc owner cq. (devid=%u; tsid=%u; logic_cqid=%u)\n", inst->devid, inst->tsid, para->cqId); ++ return -EINVAL; ++ } + } + + if (((para->timeout < 0) && (para->timeout != -1)) || (para->cqe_num == 0) || (para->cqe_addr == NULL)) { +@@ -441,7 +443,7 @@ static void trs_logic_cq_eliminate_holes(struct trs_logic_cq *logic_cq, u32 star + } + + static int trs_logic_cq_match_copy(struct trs_core_ts_inst *ts_inst, struct trs_logic_cq *logic_cq, +- struct halReportRecvInfo *para) ++ struct halReportRecvInfo *para, bool is_xsched) + { + u32 start, report_cnt, tail; + u32 rollback = 0; +@@ -463,11 +465,17 @@ static int trs_logic_cq_match_copy(struct trs_core_ts_inst *ts_inst, struct trs_ + } + + trs_logic_cq_copy_trace("Logic Cq Recv Match", ts_inst, logic_cq, start, report_cnt); +- ret = copy_to_user((void __user *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), +- (unsigned long)report_cnt * logic_cq->cqe_size); +- if (ret != 0) { +- trs_err("copy to user fail, cqid=%u report_cnt=%u\n", logic_cq->cqid, report_cnt); +- return ret; ++ ++ if (is_xsched) { ++ memcpy((void *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), ++ (unsigned long)report_cnt * logic_cq->cqe_size); ++ } else { ++ ret = copy_to_user((void __user *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), ++ (unsigned long)report_cnt * logic_cq->cqe_size); ++ if (ret != 0) { ++ trs_err("copy to user fail, cqid=%u report_cnt=%u\n", logic_cq->cqid, report_cnt); ++ return ret; ++ } + } + + para->report_cqe_num = report_cnt; +@@ -480,7 +488,7 @@ static int trs_logic_cq_match_copy(struct trs_core_ts_inst *ts_inst, struct trs_ + } + + static int trs_logic_cq_non_match_copy(struct trs_core_ts_inst *ts_inst, struct trs_logic_cq *logic_cq, +- struct halReportRecvInfo *para) ++ struct halReportRecvInfo *para, bool is_xsched) + { + u32 start, report_cnt, tail; + int ret; +@@ -490,11 +498,17 @@ static int trs_logic_cq_non_match_copy(struct trs_core_ts_inst *ts_inst, struct + report_cnt = (tail > start) ? tail - start : logic_cq->cq_depth - start; + + trs_logic_cq_copy_trace("Logic Cq Recv NoMatch", ts_inst, logic_cq, start, report_cnt); +- ret = copy_to_user((void __user *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), +- (unsigned long)report_cnt * logic_cq->cqe_size); +- if (ret != 0) { +- trs_err("copy to user fail, cqid=%u report_cnt=%u\n", logic_cq->cqid, report_cnt); +- return ret; ++ ++ if (is_xsched) { ++ memcpy((void *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), ++ (unsigned long)report_cnt * logic_cq->cqe_size); ++ } else { ++ ret = copy_to_user((void __user *)para->cqe_addr, logic_cq->addr + ((unsigned long)start * logic_cq->cqe_size), ++ (unsigned long)report_cnt * logic_cq->cqe_size); ++ if (ret != 0) { ++ trs_err("copy to user fail, cqid=%u report_cnt=%u\n", logic_cq->cqid, report_cnt); ++ return ret; ++ } + } + + para->report_cqe_num = report_cnt; +@@ -503,7 +517,7 @@ static int trs_logic_cq_non_match_copy(struct trs_core_ts_inst *ts_inst, struct + } + + static int trs_logic_cq_copy_report(struct trs_core_ts_inst *ts_inst, +- struct trs_logic_cq *logic_cq, struct halReportRecvInfo *para) ++ struct trs_logic_cq *logic_cq, struct halReportRecvInfo *para, bool is_xsched) + { + u32 version = para->res[0]; + int full_flag = 0; +@@ -522,9 +536,9 @@ static int trs_logic_cq_copy_report(struct trs_core_ts_inst *ts_inst, + } + + if (version == 1) { +- ret = trs_logic_cq_match_copy(ts_inst, logic_cq, para); // runtime new version ++ ret = trs_logic_cq_match_copy(ts_inst, logic_cq, para, is_xsched); // runtime new version + } else { +- ret = trs_logic_cq_non_match_copy(ts_inst, logic_cq, para); ++ ret = trs_logic_cq_non_match_copy(ts_inst, logic_cq, para, is_xsched); + } + if (ret != 0) { + return ret; +@@ -553,8 +567,8 @@ static int trs_logic_cq_wait_event(struct trs_logic_cq *logic_cq, int timeout) + long ret, tm; + + atomic_inc(&logic_cq->wait_thread_num); +- trs_debug("Wake wait start. (logic_cqid=%u; timeout=%d; wait_thread_num=%d)\n", +- logic_cq->cqid, timeout, atomic_read(&logic_cq->wait_thread_num)); ++ trs_debug("Wake wait start. (logic_cqid=%u; timeout=%d; wait_thread_num=%d, wakeup_num=%d)\n", ++ logic_cq->cqid, timeout, atomic_read(&logic_cq->wait_thread_num), atomic_read(&logic_cq->wakeup_num)); + + tm = (timeout == -1) ? MAX_SCHEDULE_TIMEOUT : msecs_to_jiffies((u32)timeout); + (void)prepare_to_wait_exclusive(&logic_cq->wait_queue, &wq_entry, TASK_INTERRUPTIBLE); +@@ -592,12 +606,13 @@ static int trs_logic_cq_wait_event(struct trs_logic_cq *logic_cq, int timeout) + return ret; + } + +-int trs_logic_cq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para) ++int trs_logic_cq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para, ++ bool is_xsched) + { + struct trs_logic_cq *logic_cq = NULL; + int ret; + +- ret = trs_logic_cq_recv_para_check(proc_ctx, ts_inst, para); ++ ret = trs_logic_cq_recv_para_check(proc_ctx, ts_inst, para, is_xsched); + if (ret != 0) { + return ret; + } +@@ -609,24 +624,35 @@ int trs_logic_cq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts + return trs_thread_bind_irq_wait(logic_cq, para->timeout); + } + trs_logic_cq_recv_trace("Recv start", ts_inst, para); +- do { +- mutex_lock(&logic_cq->mutex); +- ret = trs_logic_cq_copy_report(ts_inst, logic_cq, para); +- mutex_unlock(&logic_cq->mutex); +- if (ret == 0) { +- logic_cq->stat.recv++; +- trs_logic_cq_recv_trace("Recv finish", ts_inst, para); +- return ret; +- } + +- if (ret == -EAGAIN) { ++ if (is_xsched) { + if (para->timeout == 0) { +- para->report_cqe_num = 0; +- return 0; ++ para->report_cqe_num = 0; ++ return 0; + } + ret = trs_logic_cq_wait_event(logic_cq, para->timeout); +- } +- } while (ret >= 0); ++ trs_debug("Skip reading report for xsched, waiting for cq irq: logic_cqid=%u, timeout=%u, ret=%u)\n", ++ para->cqId, para->timeout, ret); ++ } else { ++ do { ++ mutex_lock(&logic_cq->mutex); ++ ret = trs_logic_cq_copy_report(ts_inst, logic_cq, para, is_xsched); ++ mutex_unlock(&logic_cq->mutex); ++ if (ret == 0) { ++ logic_cq->stat.recv++; ++ trs_logic_cq_recv_trace("Recv finish", ts_inst, para); ++ return ret; ++ } ++ ++ if (ret == -EAGAIN) { ++ if (para->timeout == 0) { ++ para->report_cqe_num = 0; ++ return 0; ++ } ++ ret = trs_logic_cq_wait_event(logic_cq, para->timeout); ++ } ++ } while (ret >= 0); ++ } + + return ret; + } +diff --git a/rms/trs_drv/trs_core/trs_logic_cq.h b/rms/trs_drv/trs_core/trs_logic_cq.h +index a45b110..b776b7f 100755 +--- a/rms/trs_drv/trs_core/trs_logic_cq.h ++++ b/rms/trs_drv/trs_core/trs_logic_cq.h +@@ -90,7 +90,8 @@ struct trs_core_ts_inst; + + int trs_logic_cq_alloc(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halSqCqInputInfo *para); + int trs_logic_cq_free(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halSqCqFreeInfo *para); +-int trs_logic_cq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para); ++int trs_logic_cq_recv(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst, struct halReportRecvInfo *para, ++ bool is_xsched); + + void trs_logic_set_cqe_version(struct trs_core_ts_inst *ts_inst, u32 logic_cqid, u32 cqe_verion); + int trs_logic_cq_enque(struct trs_core_ts_inst *ts_inst, u32 logic_cq_id, u32 stream_id, u32 task_id, void *cqe); +diff --git a/rms/trs_drv/trs_core/trs_sqcq_map.c b/rms/trs_drv/trs_core/trs_sqcq_map.c +index 8103d65..998ecfb 100755 +--- a/rms/trs_drv/trs_core/trs_sqcq_map.c ++++ b/rms/trs_drv/trs_core/trs_sqcq_map.c +@@ -305,6 +305,10 @@ int trs_sq_remap(struct trs_proc_ctx *proc_ctx, struct trs_core_ts_inst *ts_inst + int sq_reg_type = TRS_MAP_TYPE_REG; + int ret; + ++ ret = 0; ++ goto out; ++ ++ + if ((sq_info->sq_phy_addr == 0) || (sq_info->db_addr == 0) || (uio_info->sq_que_addr == 0)) { + ret = 0; + goto out; +diff --git a/ts_agent/src/ts_agent_update_sqe.c b/ts_agent/src/ts_agent_update_sqe.c +index bb4e3b2..01fe60c 100755 +--- a/ts_agent/src/ts_agent_update_sqe.c ++++ b/ts_agent/src/ts_agent_update_sqe.c +@@ -1146,6 +1146,12 @@ static void cqe_set_drop_flag(ts_stars_cqe_t *cqe) + if (cqe->warn || (cqe->sqe_type == TS_STARS_SQE_TYPE_PCIE_DMA)) { + /* cqe has been processed in ts_agent, no need to send to runtime */ + cqe->drop_flag = 1U; ++ ts_agent_debug("cqe has been processed in ts_agent, no need to send to runtime, drop_flag=%u\n", cqe->drop_flag); ++ ++ /* no drop, xsched needs to proc cqe */ ++ cqe->drop_flag = 0U; ++ ts_agent_debug("send cqe to runtime/xsched anyway, drop_flag=%u\n", cqe->drop_flag); ++ + return; + } + cqe->drop_flag = 0U; +-- +2.34.1 -- 2.34.1