[fix]日志整改-hcomm-AtomGit

Mexperimental/base_comm/channel/host_cpu_roce_channel.cc+3-2

	@@ -193,7 +193,8 @@ HcclResult HostCpuRoceChannel::BuildSocket()
193	uint16_t port = channelDesc_.port;	193	uint16_t port = channelDesc_.port;
194	if (port == 0) {	194	if (port == 0) {
195	port = DEFAULT_LISTENING_PORT;	195	port = DEFAULT_LISTENING_PORT;
196	- HCCL_INFO("[HostCpuRoceChannel::%s] channelDesc port is 0, use default port [%u]", __func__, port);	196	+ HCCL_INFO(
		197	+ "[HostCpuRoceChannel::%s] channelDesc port is 0, use default port [%u] for BuildSocket", __func__, port);
197	}	198	}
198	std::string socketTag = "AUTOMATIC_SOCKET_TAG";	199	std::string socketTag = "AUTOMATIC_SOCKET_TAG";
199	Hccl::SocketConfig socketConfig	200	Hccl::SocketConfig socketConfig
	@@ -201,7 +202,7 @@ HcclResult HostCpuRoceChannel::BuildSocket()
201	Hccl::SocketConfig(linkData, port, socketTag, channelDesc_.role == HCOMM_SOCKET_ROLE_SERVER) :	202	Hccl::SocketConfig(linkData, port, socketTag, channelDesc_.role == HCOMM_SOCKET_ROLE_SERVER) :
202	Hccl::SocketConfig(linkData, port, socketTag);	203	Hccl::SocketConfig(linkData, port, socketTag);
203	CHK_RET(hcomm::SocketMgr::GetInstance(devicePhyId_).GetSocket(socketConfig, socket_));	204	CHK_RET(hcomm::SocketMgr::GetInstance(devicePhyId_).GetSocket(socketConfig, socket_));
204	- HCCL_INFO("[HostCpuRoceChannel::%s] SUCCESS. port[%u].", __func__, port);	205	+ HCCL_INFO("[HostCpuRoceChannel::%s] BuildSocket SUCCESS. port[%u].", __func__, port);
205	return HCCL_SUCCESS;	206	return HCCL_SUCCESS;
206	}	207	}
207		208

Mexperimental/cluster_link_diag/dispatcher/exec_realtime_cmd.py+13-11

	@@ -29,21 +29,21 @@ _executor_instance = None
29		29
30		30
31	def signal_handler(sig, frame):	31	def signal_handler(sig, frame):
32	- """处理Ctrl+C信号的函数"""	32	+ """Handle Ctrl+C signal."""
33	- common.log_info("\n\033[33m接收到Ctrl+C，正在关闭所有连接...\033[0m")	33	+ common.log_info("\n\033[33mReceived Ctrl+C, closing all connections...\033[0m")
34	if _executor_instance:	34	if _executor_instance:
35	_executor_instance.force_close()	35	_executor_instance.force_close()
36		36
37		37
38	def setup_signal_handler(executor):	38	def setup_signal_handler(executor):
39	- """设置信号处理器"""	39	+ """Set up signal handler."""
40	global _executor_instance	40	global _executor_instance
41	_executor_instance = executor	41	_executor_instance = executor
42	signal.signal(signal.SIGINT, signal_handler)	42	signal.signal(signal.SIGINT, signal_handler)
43		43
44		44
45	def setup_logger(host_path, host_name):	45	def setup_logger(host_path, host_name):
46	- """为每个主机设置独立的logger"""	46	+ """Set up an independent logger for each host."""
47	return common.setup_file_logger(host_path, log_dir, _enable_logger_ref["value"])	47	return common.setup_file_logger(host_path, log_dir, _enable_logger_ref["value"])
48		48
49		49
	@@ -69,9 +69,9 @@ class ParallelSSHExecutor:
69	context.logger.info(f"command: {original_cmd}")	69	context.logger.info(f"command: {original_cmd}")
70		70
71	def force_close(self):	71	def force_close(self):
72	- """强制关闭所有连接"""	72	+ """Forcefully close all connections."""
73	self.force_closing = True	73	self.force_closing = True
74	- common.log_info("\033[33m正在强制关闭所有SSH连接...\033[0m")	74	+ common.log_info("\033[33mForcefully closing all SSH connections...\033[0m")
75		75
76	# 首先尝试正常关闭所有客户端	76	# 首先尝试正常关闭所有客户端
77	for client in self.clients[::-1]:	77	for client in self.clients[::-1]:
	@@ -94,7 +94,8 @@ class ParallelSSHExecutor:
94	if alive_threads:	94	if alive_threads:
95	self.completed = False	95	self.completed = False
96	common.log_error(	96	common.log_error(
97	- "\033[31m有些线程无法正常结束，请确认远端命令是否仍在运行。\033[0m"	97	+ "\033[31mSome threads cannot terminate normally, please check if remote "
		98	+ "commands are still running.\033[0m"
98	)	99	)
99		100
100	def exec_realtime_commands(self, context):	101	def exec_realtime_commands(self, context):
	@@ -155,10 +156,11 @@ class ParallelSSHExecutor:
155	if self.force_closing:	156	if self.force_closing:
156	return	157	return
157	common.log_error(	158	common.log_error(
158	- f"\033[31m[{context.host:<{self._ipv4_print_len}}] 执行命令时发生错误: {str(err)}\033[0m"	159	+ f"\033[31m[{context.host:<{self._ipv4_print_len}}] error occurred while executing command: "
		160	+ f"{str(err)}\033[0m"
159	)	161	)
160	if context.logger:	162	if context.logger:
161	- context.logger.error(f"执行命令时发生错误: {str(err)}")	163	+ context.logger.error(f"error occurred while executing command: {str(err)}")
162		164
163	def _log_realtime_line(self, context, line):	165	def _log_realtime_line(self, context, line):
164	common.log_info(f"[{context.host:<{self._ipv4_print_len}}] {line.rstrip()}")	166	common.log_info(f"[{context.host:<{self._ipv4_print_len}}] {line.rstrip()}")
	@@ -250,7 +252,7 @@ class ParallelSSHExecutor:
250	self._recurse_realtime_child(target_cmd, client, task)	252	self._recurse_realtime_child(target_cmd, client, task)
251	except (paramiko.SSHException, OSError, RuntimeError, ValueError) as err:	253	except (paramiko.SSHException, OSError, RuntimeError, ValueError) as err:
252	if not self.force_closing:	254	if not self.force_closing:
253	- common.log_error(f"\033[31m连接 {task.host_name} 失败: {str(err)}\033[0m")	255	+ common.log_error(f"\033[31mconnect to {task.host_name} failed: {str(err)}\033[0m")
254		256
255	def _realtime_exec_host(self, target_cmd, task):	257	def _realtime_exec_host(self, target_cmd, task):
256	host_name = task.host_name	258	host_name = task.host_name
	@@ -298,5 +300,5 @@ if __name__ == "__main__":
298	executor.re_realtime_exec(commands)	300	executor.re_realtime_exec(commands)
299	executor.wait_and_close()	301	executor.wait_and_close()
300	except (paramiko.SSHException, OSError, RuntimeError, ValueError) as err:	302	except (paramiko.SSHException, OSError, RuntimeError, ValueError) as err:
301	- common.log_error(f"\033[31m程序执行出错: {str(err)}\033[0m")	303	+ common.log_error(f"\033[31mprogram execution error: {str(err)}\033[0m")
302	executor.force_close()	304	executor.force_close()

Msrc/base_comm/common/debug/plf_debug_config.cc+1-4

	@@ -78,9 +78,6 @@ u64 GetPlfDebugConfigValue()
78	return g_plfDebugConfig.load(std::memory_order_relaxed);	78	return g_plfDebugConfig.load(std::memory_order_relaxed);
79	}	79	}
80		80
81	-void SetPlfDebugConfigValue(u64 value)	81	+void SetPlfDebugConfigValue(u64 value) { g_plfDebugConfig.store(value, std::memory_order_relaxed); }
82	-{
83	- g_plfDebugConfig.store(value, std::memory_order_relaxed);
84	-}
85		82
86	} // namespace Hccl	83	} // namespace Hccl

Msrc/base_comm/resources/hccp/rdma_agent/peer/ra_peer_ctx.c+4-2

	@@ -313,7 +313,9 @@ int RaPeerCtxCqCreate(struct RaCtxHandle ctxHandle, struct CqInfoT info, struc
313	int ret = 0;	313	int ret = 0;
314		314
315	CHK_PRT_RETURN(info->in.ub.mode != JFC_MODE_NORMAL,	315	CHK_PRT_RETURN(info->in.ub.mode != JFC_MODE_NORMAL,
316	- hccp_err("[init][ctx_cq]jfc_mode[%d] not support, phyId[%u]", info->in.ub.mode, phyId), -EINVAL);	316	+ hccp_err("[init][ctx_cq]jfc_mode[%d] not support, only JFC_MODE_NORMAL is supported, phyId[%u]",
		317	+ info->in.ub.mode, phyId),
		318	+ -EINVAL);
317		319
318	RaRsSetDevInfo(&devInfo, phyId, ctxHandle->devIndex);	320	RaRsSetDevInfo(&devInfo, phyId, ctxHandle->devIndex);
319	RaCtxPrepareCqCreate(info, &cqAttr);	321	RaCtxPrepareCqCreate(info, &cqAttr);
	@@ -358,7 +360,7 @@ int RaPeerCtxQpCreate(struct RaCtxHandle ctxHandle, struct QpCreateAttr qpAttr
358	int ret = 0;	360	int ret = 0;
359		361
360	CHK_PRT_RETURN(qpAttr->ub.mode != JETTY_MODE_URMA_NORMAL,	362	CHK_PRT_RETURN(qpAttr->ub.mode != JETTY_MODE_URMA_NORMAL,
361	- hccp_err("[init][ctx_cq]jetty_mode[%d] not support,"	363	+ hccp_err("[init][ctx_cq]jetty_mode[%d] not support, only JETTY_MODE_URMA_NORMAL is supported,"
362	" phyId[%u]",	364	" phyId[%u]",
363	qpAttr->ub.mode, phyId),	365	qpAttr->ub.mode, phyId),
364	-EINVAL);	366	-EINVAL);

Msrc/base_comm/resources/hccp/rdma_service/rs_drv_rdma.c+2-2

	@@ -343,7 +343,7 @@ int RsDrvCreateCqWithAttrs(struct RsQpCb qpCb, int isExt, struct CqExtAttr cqA
343	if (isExt == 1) {	343	if (isExt == 1) {
344	qpCb->ibSendCq = RsIbvExpCreateCq(qpCb->rdevCb->ibCtx, cqAttr->sendCqDepth, NULL, channel, sendEqNum, &attr,	344	qpCb->ibSendCq = RsIbvExpCreateCq(qpCb->rdevCb->ibCtx, cqAttr->sendCqDepth, NULL, channel, sendEqNum, &attr,
345	&qpCb->qpResp.sendCqData);	345	&qpCb->qpResp.sendCqData);
346	- hccp_info("rs_ibv_exp_create_cq");	346	+ hccp_info("rs_ibv_exp_create_cq ibSendCq");
347	} else {	347	} else {
348	qpCb->ibSendCq = RsIbvCreateCq(qpCb->rdevCb->ibCtx, cqAttr->sendCqDepth, NULL, channel, sendEqNum);	348	qpCb->ibSendCq = RsIbvCreateCq(qpCb->rdevCb->ibCtx, cqAttr->sendCqDepth, NULL, channel, sendEqNum);
349	}	349	}
	@@ -354,7 +354,7 @@ int RsDrvCreateCqWithAttrs(struct RsQpCb qpCb, int isExt, struct CqExtAttr cqA
354	if (isExt == 1) {	354	if (isExt == 1) {
355	qpCb->ibRecvCq = RsIbvExpCreateCq(qpCb->rdevCb->ibCtx, cqAttr->recvCqDepth, NULL, channel, recvEqNum, &attr,	355	qpCb->ibRecvCq = RsIbvExpCreateCq(qpCb->rdevCb->ibCtx, cqAttr->recvCqDepth, NULL, channel, recvEqNum, &attr,
356	&qpCb->qpResp.recvCqData);	356	&qpCb->qpResp.recvCqData);
357	- hccp_info("rs_ibv_exp_create_cq");	357	+ hccp_info("rs_ibv_exp_create_cq ibRecvCq");
358	} else {	358	} else {
359	qpCb->ibRecvCq = RsIbvCreateCq(qpCb->rdevCb->ibCtx, cqAttr->recvCqDepth, NULL, channel, recvEqNum);	359	qpCb->ibRecvCq = RsIbvCreateCq(qpCb->rdevCb->ibCtx, cqAttr->recvCqDepth, NULL, channel, recvEqNum);
360	}	360	}

Msrc/base_comm/resources/hccp/rdma_service/rs_rdma.c+9-12

	@@ -63,7 +63,7 @@ STATIC int RsGetQpcb(struct RsRdevCb rdevCb, uint32_t qpn, struct RsQpCb *qpCb
63	}	63	}
64		64
65	*qpCb = NULL;	65	*qpCb = NULL;
66	- hccp_err("qp_cb for qp %u is not available!", qpn);	66	+ hccp_err("qp_cb for qp %u is not available", qpn);
67		67
68	return -ENODEV;	68	return -ENODEV;
69	}	69	}
	@@ -474,15 +474,13 @@ STATIC int RsMrInfoSync(struct RsMrCb *mrCb)
474	CHK_PRT_RETURN(mrCb->qpCb->connInfo == NULL, hccp_warn("no conn available !"), 0);	474	CHK_PRT_RETURN(mrCb->qpCb->connInfo == NULL, hccp_warn("no conn available !"), 0);
475		475
476	CHK_PRT_RETURN(mrCb->qpCb->state == RS_QP_STATUS_REM_FD_CLOSE,	476	CHK_PRT_RETURN(mrCb->qpCb->state == RS_QP_STATUS_REM_FD_CLOSE,
477	- hccp_warn("remote qp fd closed,"	477	+ hccp_warn("remote qp fd closed, "
478	"can not use it anymore! status[%d](RS_QP_STATUS_REM_FD_CLOSE)",	478	"can not use it anymore! status[%d](RS_QP_STATUS_REM_FD_CLOSE)",
479	mrCb->qpCb->state),	479	mrCb->qpCb->state),
480	-EFAULT);	480	-EFAULT);
481		481
482	CHK_PRT_RETURN(mrCb->qpCb->connInfo->connfd == RS_FD_INVALID,	482	CHK_PRT_RETURN(mrCb->qpCb->connInfo->connfd == RS_FD_INVALID,
483	- hccp_warn("mr info sync failed! fd not ready!"	483	+ hccp_warn("mr info sync failed! fd not ready! connfd[%d](RS_FD_INVALID)", mrCb->qpCb->connInfo->connfd),
484	- "connfd[%d](RS_FD_INVALID)",
485	- mrCb->qpCb->connInfo->connfd),
486	-ENETUNREACH);	484	-ENETUNREACH);
487		485
488	mrCb->mrInfo.cmd = (unsigned int)RS_CMD_MR_INFO;	486	mrCb->mrInfo.cmd = (unsigned int)RS_CMD_MR_INFO;
	@@ -638,10 +636,7 @@ RS_ATTRI_VISI_DEF int RsMrDereg(unsigned int phyId, unsigned int rdevIndex, unsi
638	CHK_PRT_RETURN(ret, hccp_err("rs_qpn2qpcb failed ret[%d]", ret), ret);	636	CHK_PRT_RETURN(ret, hccp_err("rs_qpn2qpcb failed ret[%d]", ret), ret);
639		637
640	CHK_PRT_RETURN(RsGetMrcb(qpCb, (uintptr_t)addr, &mrCb, &qpCb->mrList),	638	CHK_PRT_RETURN(RsGetMrcb(qpCb, (uintptr_t)addr, &mrCb, &qpCb->mrList),
641	- hccp_err("rs_get_mrcb failed "	639	+ hccp_err("rs_get_mrcb failed, qpn[%u] addr[0x%llx]", qpn, (unsigned long long)(uintptr_t)addr), -EFAULT);
642	- "g_rs_send_wr_num[%u]",
643	- gRsSendWrNum),
644	- -EFAULT);
645		640
646	ret = RsDrvMrDereg(mrCb->ibMr);	641	ret = RsDrvMrDereg(mrCb->ibMr);
647	CHK_PRT_RETURN(ret, hccp_err("rs_drv_mr_dereg failed ret[%d] ", ret), -EACCES);	642	CHK_PRT_RETURN(ret, hccp_err("rs_drv_mr_dereg failed ret[%d] ", ret), -EACCES);
	@@ -769,12 +764,14 @@ STATIC int RsInitTypicalMrCb(unsigned int phyId, struct RdmaMrRegInfo *mrRegInfo
769		764
770	if (devCb->rsCb->hccpMode == NETWORK_PEER_ONLINE \|\| devCb->rsCb->hccpMode == NETWORK_ONLINE) {	765	if (devCb->rsCb->hccpMode == NETWORK_PEER_ONLINE \|\| devCb->rsCb->hccpMode == NETWORK_ONLINE) {
771	mrCb->ibMr = RsDrvMrReg(devCb->ibPd, addr, len, access);	766	mrCb->ibMr = RsDrvMrReg(devCb->ibPd, addr, len, access);
772	- CHK_PRT_RETURN(mrCb->ibMr == NULL, hccp_err("rs_drv_mr_reg addr is NULL len[%lld] failed", len), -EACCES);	767	+ CHK_PRT_RETURN(mrCb->ibMr == NULL,
		768	+ hccp_err("rs_drv_mr_reg ibMr is NULL len[%lld] failed, errno:%d", len, errno), -EACCES);
773	} else {	769	} else {
774	ret = RsMrPrepareRoceSign(phyId, devCb, &roceSign);	770	ret = RsMrPrepareRoceSign(phyId, devCb, &roceSign);
775	CHK_PRT_RETURN(ret != 0, hccp_err("RsMrPrepareRoceSign failed, ret(%d)", ret), ret);	771	CHK_PRT_RETURN(ret != 0, hccp_err("RsMrPrepareRoceSign failed, ret(%d)", ret), ret);
776	mrCb->ibMr = RsDrvExpMrReg(devCb->ibPd, addr, len, access, roceSign);	772	mrCb->ibMr = RsDrvExpMrReg(devCb->ibPd, addr, len, access, roceSign);
777	- CHK_PRT_RETURN(mrCb->ibMr == NULL, hccp_err("rs_drv_exp_mr_reg addr is NULL len[%lld] failed", len), -EACCES);	773	+ CHK_PRT_RETURN(mrCb->ibMr == NULL,
		774	+ hccp_err("rs_drv_exp_mr_reg ibMr is NULL len[%lld] failed, errno:%d", len, errno), -EACCES);
778	}	775	}
779		776
780	mrCb->mrInfo.addr = (uintptr_t)addr;	777	mrCb->mrInfo.addr = (uintptr_t)addr;
	@@ -2287,7 +2284,7 @@ static int RsQpConnectAsyncInitPara(struct RsQpConnPara qpConnPara, int fd, stru
2287		2284
2288	CHK_PRT_RETURN(qpConnPara.phyId >= RS_MAX_DEV_NUM, hccp_err("param error ! phyId:%u", qpConnPara.phyId), -EINVAL);	2285	CHK_PRT_RETURN(qpConnPara.phyId >= RS_MAX_DEV_NUM, hccp_err("param error ! phyId:%u", qpConnPara.phyId), -EINVAL);
2289		2286
2290	- CHK_PRT_RETURN(fd < 0, hccp_err("param error ! fd:%d must be greater than or equal to 0", fd), -EINVAL);	2287	+ CHK_PRT_RETURN(fd < 0, hccp_err("param error ! fd:%d must be non-negative", fd), -EINVAL);
2291		2288
2292	ret = RsQpn2qpcb(qpConnPara.phyId, qpConnPara.rdevIndex, qpConnPara.qpn, qpCb);	2289	ret = RsQpn2qpcb(qpConnPara.phyId, qpConnPara.rdevIndex, qpConnPara.qpn, qpCb);
2293	CHK_PRT_RETURN(ret, hccp_err("get qpcb failed, qpn %u, ret %d", qpConnPara.qpn, ret), ret);	2290	CHK_PRT_RETURN(ret, hccp_err("get qpcb failed, qpn %u, ret %d", qpConnPara.qpn, ret), ret);

Msrc/coll_communicator_mgr/communicator/group_schedule_mgr/group_schedule_mgr.cc+3-1

	@@ -109,7 +109,9 @@ HcclResult GroupScheduleMgr::GetCurLocalRank(uint32_t& localRank)
109	}	109	}
110		110
111	if (curLocalRank >= this->serverToRankSize_.at(curServerIdx)) {	111	if (curLocalRank >= this->serverToRankSize_.at(curServerIdx)) {
112	- HCCL_ERROR("[getCurLocalRank] is invalid:%u", curLocalRank);	112	+ HCCL_ERROR(
		113	+ "[getCurLocalRank] curLocalRank[%u] is invalid, valid range [0, %u)", curLocalRank,
		114	+ this->serverToRankSize_.at(curServerIdx));
113	return HCCL_E_INTERNAL;	115	return HCCL_E_INTERNAL;
114	}	116	}
115		117

Msrc/coll_communicator_mgr/dfx/cluster_monitor/cluster_monitor.cc+6-2

	@@ -133,7 +133,9 @@ HcclResult ClusterMonitor::GetSocketDescFromRankInfo(
133	CHK_PTR_NULL(rankGraph);	133	CHK_PTR_NULL(rankGraph);
134	CHK_RET(rankGraph->GetDevicePort(remoteRank, &rmtPort));	134	CHK_RET(rankGraph->GetDevicePort(remoteRank, &rmtPort));
135	if (rmtPort > Hccl::MAX_VALUE_TCPPORT) {	135	if (rmtPort > Hccl::MAX_VALUE_TCPPORT) {
136	- HCCL_ERROR("[%s] Invalid port[%u] of Rank[%u]", __func__, rmtPort, remoteRank);	136	+ HCCL_ERROR(
		137	+ "[%s] Invalid port[%u] of Rank[%u], max valid port is %u", __func__, rmtPort, remoteRank,
		138	+ Hccl::MAX_VALUE_TCPPORT);
137	return HCCL_E_PARA;	139	return HCCL_E_PARA;
138	}	140	}
139	CommLink* links = nullptr;	141	CommLink* links = nullptr;
	@@ -161,7 +163,9 @@ HcclResult ClusterMonitor::GetSocketDescFromRankInfo(
161	CHK_RET(rankGraph->GetDevicePort(myRankId, &listenPort));	163	CHK_RET(rankGraph->GetDevicePort(myRankId, &listenPort));
162	socketDesc.role = HcommSocketRole::HCOMM_SOCKET_ROLE_SERVER;	164	socketDesc.role = HcommSocketRole::HCOMM_SOCKET_ROLE_SERVER;
163	if (listenPort > Hccl::MAX_VALUE_TCPPORT) {	165	if (listenPort > Hccl::MAX_VALUE_TCPPORT) {
164	- HCCL_ERROR("[%s] Invalid port[%u] of Rank[%u]", __func__, listenPort, myRankId);	166	+ HCCL_ERROR(
		167	+ "[%s] Invalid port[%u] of Rank[%u], max valid port is %u", __func__, listenPort, myRankId,
		168	+ Hccl::MAX_VALUE_TCPPORT);
165	return HCCL_E_PARA;	169	return HCCL_E_PARA;
166	}	170	}
167	socketDesc.listenPort = static_cast<uint16_t>(listenPort); // socketDesc.port中填监听端口号	171	socketDesc.listenPort = static_cast<uint16_t>(listenPort); // socketDesc.port中填监听端口号

Msrc/coll_communicator_mgr/resource_mgr/local/my_rank/my_rank.cc+3-1

	@@ -426,7 +426,9 @@ HcclResult MyRank::QueryListenPort(
426	uint32_t rmtPort = 0;	426	uint32_t rmtPort = 0;
427	CHK_RET(GetDevicePortInternal(remoteRank, &rmtPort, remoteEndpointDesc.loc.locType));	427	CHK_RET(GetDevicePortInternal(remoteRank, &rmtPort, remoteEndpointDesc.loc.locType));
428	if (rmtPort > Hccl::MAX_VALUE_TCPPORT) {	428	if (rmtPort > Hccl::MAX_VALUE_TCPPORT) {
429	- HCCL_ERROR("[%s] Invalid port[%u] of Rank[%u]", __func__, rmtPort, remoteRank);	429	+ HCCL_ERROR(
		430	+ "[%s] Invalid port[%u] of Rank[%u], max valid port is %u", __func__, rmtPort, remoteRank,
		431	+ Hccl::MAX_VALUE_TCPPORT);
430	return HCCL_E_PARA;	432	return HCCL_E_PARA;
431	}	433	}
432	// 查询该socket链接的server端监听的端口（监听方的选择策略需要跟SocketConfig中保持一致）	434	// 查询该socket链接的server端监听的端口（监听方的选择策略需要跟SocketConfig中保持一致）

Msrc/legacy/ascend910/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_mesh_aiv_executor.cc+2-2

	@@ -171,7 +171,7 @@ HcclResult CollAllReduceMeshAivExecutor::GetAivExecParam(
171	}	171	}
172		172
173	HCCL_INFO(	173	HCCL_INFO(
174	- "SPK, buffersIn [%p] [%p] [%p] [%p] "	174	+ "[CollAllReduceMeshAivExecutor] buffersIn [%p] [%p] [%p] [%p] "
175	"buffersOut [%p] [%p] [%p] [%p]",	175	"buffersOut [%p] [%p] [%p] [%p]",
176	args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3], args.buffersOut[0],	176	args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3], args.buffersOut[0],
177	args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);	177	args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
	@@ -183,7 +183,7 @@ HcclResult CollAllReduceMeshAivExecutor::GetAivExecParam(
183	args.reduceOp = param.reduceType;	183	args.reduceOp = param.reduceType;
184	args.devType = static_cast<u32>(topoAttr_.deviceType);	184	args.devType = static_cast<u32>(topoAttr_.deviceType);
185	HCCL_INFO(	185	HCCL_INFO(
186	- "SPK [CollAllReduceMeshAivExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], "	186	+ "[CollAllReduceMeshAivExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], len[%llu],datatype[%llu], "
187	"op[%llu]",	187	"op[%llu]",
188	args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);	188	args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
189		189

Msrc/legacy/ascend910/algorithm/impl/coll_executor/coll_comm_executor.cc+6-6

	@@ -2739,17 +2739,17 @@ bool CollCommExecutor::IsLevel0Neighbor(u32 remoteRank, u32 level0RankSize)
2739		2739
2740	HcclResult CollCommExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& adjInfo)	2740	HcclResult CollCommExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& adjInfo)
2741	{	2741	{
2742	- HCCL_INFO("[nslbdp] Entry GetAdjInfo.");	2742	+ HCCL_INFO("[GetAdjInfo] Entry GetAdjInfo.");
2743	algResResp_ = &algRes;	2743	algResResp_ = &algRes;
2744	SubCommInfo level1CommInfo = {};	2744	SubCommInfo level1CommInfo = {};
2745	AdjInfo nslbAdjInfo = {};	2745	AdjInfo nslbAdjInfo = {};
2746	if (Getlevel1CommRank(level1CommInfo) != HCCL_SUCCESS) {	2746	if (Getlevel1CommRank(level1CommInfo) != HCCL_SUCCESS) {
2747	- HCCL_INFO("[nslbdp-GetAdjInfo] Getlevel1CommRank is NULL.");	2747	+ HCCL_WARNING("[GetAdjInfo] Getlevel1CommRank is NULL.");
2748	return HCCL_SUCCESS;	2748	return HCCL_SUCCESS;
2749	}	2749	}
2750	u32 localRank = level1CommInfo.localRank;	2750	u32 localRank = level1CommInfo.localRank;
2751	u32 localRankSize = level1CommInfo.localRankSize;	2751	u32 localRankSize = level1CommInfo.localRankSize;
2752	- HCCL_INFO("[nslbdp-GetAdjInfo] level1CommInfo.localRank = [%u] localRankSize = [%u].", localRank, localRankSize);	2752	+ HCCL_INFO("[GetAdjInfo] level1CommInfo.localRank = [%u] localRankSize = [%u].", localRank, localRankSize);
2753		2753
2754	if (localRankSize == 1) {	2754	if (localRankSize == 1) {
2755	return HCCL_SUCCESS;	2755	return HCCL_SUCCESS;
	@@ -2761,7 +2761,7 @@ HcclResult CollCommExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& ad
2761		2761
2762	std::unique_ptr<AlgTemplateBase> nslbdp_levelTempAlg;	2762	std::unique_ptr<AlgTemplateBase> nslbdp_levelTempAlg;
2763	if (SelectTempAlg(nslbdp_levelTempAlg, localRankSize) != HCCL_SUCCESS) {	2763	if (SelectTempAlg(nslbdp_levelTempAlg, localRankSize) != HCCL_SUCCESS) {
2764	- HCCL_INFO("[nslbdp-GetAdjInfo] SelectTempAlg is unsuccessful.");	2764	+ HCCL_WARNING("[GetAdjInfo] SelectTempAlg is unsuccessful.");
2765	return HCCL_SUCCESS;	2765	return HCCL_SUCCESS;
2766	}	2766	}
2767	if (nslbdp_levelTempAlg == nullptr) {	2767	if (nslbdp_levelTempAlg == nullptr) {
	@@ -2770,7 +2770,7 @@ HcclResult CollCommExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& ad
2770	CHK_RET(nslbdp_levelTempAlg->GetNslbAdjInfo(localRank, localRankSize, level1CommInfo.links, nslbAdjInfo));	2770	CHK_RET(nslbdp_levelTempAlg->GetNslbAdjInfo(localRank, localRankSize, level1CommInfo.links, nslbAdjInfo));
2771		2771
2772	adjInfo.dstRankNum = nslbAdjInfo.dstRankNum;	2772	adjInfo.dstRankNum = nslbAdjInfo.dstRankNum;
2773	- HCCL_INFO("[nslbdp-GetAdjInfo] adjInfo.dstRankNum[%u].", adjInfo.dstRankNum);	2773	+ HCCL_INFO("[GetAdjInfo] adjInfo.dstRankNum[%u].", adjInfo.dstRankNum);
2774		2774
2775	for (size_t i = 0; i < nslbAdjInfo.nsAdjInfo.size(); i++) {	2775	for (size_t i = 0; i < nslbAdjInfo.nsAdjInfo.size(); i++) {
2776	NslbDpAdjInfo dpAdjInfo = {};	2776	NslbDpAdjInfo dpAdjInfo = {};
	@@ -2779,7 +2779,7 @@ HcclResult CollCommExecutor::GetAdjInfo(AlgResourceResponse& algRes, AdjInfo& ad
2779	dpAdjInfo.rev = 0;	2779	dpAdjInfo.rev = 0;
2780	adjInfo.nsAdjInfo.push_back(dpAdjInfo);	2780	adjInfo.nsAdjInfo.push_back(dpAdjInfo);
2781	HCCL_INFO(	2781	HCCL_INFO(
2782	- "[nslbdp]GetAdjInfo dstLocalRankId[%u], phaseId[%u].", nslbAdjInfo.nsAdjInfo[i].dstLocalRankId,	2782	+ "[GetAdjInfo] dstLocalRankId[%u], phaseId[%u].", nslbAdjInfo.nsAdjInfo[i].dstLocalRankId,
2783	nslbAdjInfo.nsAdjInfo[i].phaseId);	2783	nslbAdjInfo.nsAdjInfo[i].phaseId);
2784	}	2784	}
2785	return HCCL_SUCCESS;	2785	return HCCL_SUCCESS;

Msrc/legacy/ascend910/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_aiv_deter_small_executor.cc+2-2

	@@ -146,7 +146,7 @@ HcclResult CollReduceScatterAivDeterSmallExecutor::GetAivExecParam(
146	}	146	}
147		147
148	HCCL_INFO(	148	HCCL_INFO(
149	- "SPK, buffersIn [%p] [%p] [%p] [%p] "	149	+ "[CollReduceScatterAivDeterSmallExecutor] buffersIn [%p] [%p] [%p] [%p] "
150	"buffersOut [%p] [%p] [%p] [%p]",	150	"buffersOut [%p] [%p] [%p] [%p]",
151	args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3], args.buffersOut[0],	151	args.buffersIn[0], args.buffersIn[1], args.buffersIn[2], args.buffersIn[3], args.buffersOut[0],
152	args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);	152	args.buffersOut[1], args.buffersOut[2], args.buffersOut[3]);
	@@ -158,7 +158,7 @@ HcclResult CollReduceScatterAivDeterSmallExecutor::GetAivExecParam(
158	args.reduceOp = param.reduceType;	158	args.reduceOp = param.reduceType;
159	args.devType = static_cast<u32>(topoAttr_.deviceType);	159	args.devType = static_cast<u32>(topoAttr_.deviceType);
160	HCCL_INFO(	160	HCCL_INFO(
161	- "SPK [CollReduceScatterAivDeterSmallExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], "	161	+ "[CollReduceScatterAivDeterSmallExecutor][GetAivExecParam], rank[%llu], rankSize[%llu], "
162	"len[%llu],datatype[%llu], op[%llu]",	162	"len[%llu],datatype[%llu], op[%llu]",
163	args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);	163	args.rank, args.rankSize, args.len, args.dataType, args.reduceOp);
164		164

Msrc/legacy/ascend910/framework/communicator/comm_config.cc+1-1

	@@ -240,7 +240,7 @@ HcclResult CommConfig::SetConfigByVersion(const CommConfigHandle& config)
240	// 版本大于等于11，支持配置通信域级别的sqDepth	240	// 版本大于等于11，支持配置通信域级别的sqDepth
241	sqDepth_ = config.sqDepth;	241	sqDepth_ = config.sqDepth;
242	}	242	}
243	- HCCL_INFO("NSLBDP-VERSION config.info.version = [%u] .", config.info.version);	243	+ HCCL_INFO("CommConfig version: config.info.version = [%u].", config.info.version);
244	return HCCL_SUCCESS;	244	return HCCL_SUCCESS;
245	}	245	}
246		246

Msrc/legacy/ascend910/framework/device/framework/aicpu_zero_copy_exchanger.cc+2-2

	@@ -333,7 +333,7 @@ HcclResult AicpuZeroCopyExchanger::BatchSetLocalAddrToRemote(void* in, void* out
333	CHK_PRT_RET(	333	CHK_PRT_RET(
334	ret != 0,	334	ret != 0,
335	HCCL_ERROR(	335	HCCL_ERROR(
336	- "[[AicpuZeroCopyExchanger][BatchSetLocalAddrToRemote] Batch get remote "	336	+ "[AicpuZeroCopyExchanger][BatchSetLocalAddrToRemote] sdma batch copy "
337	"failed, ret[%u]",	337	"failed, ret[%u]",
338	ret),	338	ret),
339	HCCL_E_INTERNAL);	339	HCCL_E_INTERNAL);
	@@ -368,7 +368,7 @@ HcclResult AicpuZeroCopyExchanger::BatchSetLocalAddrToRemote(void* in, void* out
368	CHK_PRT_RET(	368	CHK_PRT_RET(
369	ret != 0,	369	ret != 0,
370	HCCL_ERROR(	370	HCCL_ERROR(
371	- "[[AicpuZeroCopyExchanger][BatchSetLocalAddrToRemote] Batch get remote "	371	+ "[AicpuZeroCopyExchanger][BatchSetLocalAddrToRemote] sdma batch copy "
372	"failed, ret[%u]",	372	"failed, ret[%u]",
373	ret),	373	ret),
374	HCCL_E_INTERNAL);	374	HCCL_E_INTERNAL);

Msrc/legacy/ascend910/platform/common/adapter/adapter_rts.cc+1-1

	@@ -326,7 +326,7 @@ HcclResult __hrtGetDevice(s32* deviceLogicId)
326	return HCCL_SUCCESS;	326	return HCCL_SUCCESS;
327	}	327	}
328	*deviceLogicId = 0;	328	*deviceLogicId = 0;
329	- HCCL_WARNING("[hrtGetDevice]Does not support this interface.");	329	+ HCCL_WARNING("[hrtGetDevice] Does not support this interface.");
330	return HCCL_E_NOT_SUPPORT;	330	return HCCL_E_NOT_SUPPORT;
331	#endif	331	#endif
332	}	332	}

Msrc/legacy/ascend910/pub_inc/transport_pub.h+2-2

	@@ -420,10 +420,10 @@ struct TransportDeviceNormalData {
420	"remoteOutputMem: addr[%llu], size[%llu], key[%u]", remoteOutputMem.addr, remoteOutputMem.size,	420	"remoteOutputMem: addr[%llu], size[%llu], key[%u]", remoteOutputMem.addr, remoteOutputMem.size,
421	remoteOutputMem.key);	421	remoteOutputMem.key);
422	HCCL_DEBUG(	422	HCCL_DEBUG(
423	- "remoteInputMem: addr[%llu], size[%llu], key[%u]", localInputMem.addr, localInputMem.size,	423	+ "localInputMem: addr[%llu], size[%llu], key[%u]", localInputMem.addr, localInputMem.size,
424	localInputMem.key);	424	localInputMem.key);
425	HCCL_DEBUG(	425	HCCL_DEBUG(
426	- "remoteOutputMem: addr[%llu], size[%llu], key[%u]", localOutputMem.addr, localOutputMem.size,	426	+ "localOutputMem: addr[%llu], size[%llu], key[%u]", localOutputMem.addr, localOutputMem.size,
427	localOutputMem.key);	427	localOutputMem.key);
428	HCCL_DEBUG("qpInfo: qpPtr[%llu], sqIndex[%u], dbIndex[%u]", qpInfo.qpPtr, qpInfo.sqIndex, qpInfo.dbIndex);	428	HCCL_DEBUG("qpInfo: qpPtr[%llu], sqIndex[%u], dbIndex[%u]", qpInfo.qpPtr, qpInfo.sqIndex, qpInfo.dbIndex);
429	HCCL_DEBUG("qpMode[%d]", static_cast<int32_t>(qpMode));	429	HCCL_DEBUG("qpMode[%d]", static_cast<int32_t>(qpMode));

Msrc/legacy/ascend950/service/collective/alg/coll_alg_factory/alg_ccu_context/reduce/ccu_context_reduce_mesh1d_two_shot_mem2mem.cc+1-1

	@@ -358,7 +358,7 @@ void CcuContextReduceMeshTwoShotMem2Mem1D::BcastLocToRmt(
358	{	358	{
359	CHK_PRT_THROW(	359	CHK_PRT_THROW(
360	dstAddr.size() != transports.size() + 1,	360	dstAddr.size() != transports.size() + 1,
361	- HCCL_ERROR("[ReduceRmtToLoc] srcAddr.size[%zu] != transports size[%zu] + 1", dstAddr.size(), transports.size()),	361	+ HCCL_ERROR("[ReduceRmtToLoc] dstAddr.size[%zu] != transports size[%zu] + 1", dstAddr.size(), transports.size()),
362	InvalidParamsException, "Invalid srcAddr size");	362	InvalidParamsException, "Invalid srcAddr size");
363	gatherSrc_[rankId_].addr = srcAddr;	363	gatherSrc_[rankId_].addr = srcAddr;
364	gatherSrc_[rankId_].addr += sliceOffset_[rankId_];	364	gatherSrc_[rankId_].addr += sliceOffset_[rankId_];

Msrc/legacy/ascend950/service/collective/alg/coll_alg_factory/alg_template/ins_alg_template/ins_temp_reduce_aicpu_reduce.cc+1-1

	@@ -73,7 +73,7 @@ HcclResult InsTempReduceAicpuReduce::RunGatherMesh(
73	DataInfo recvData(neighborLinkData, rxSlicesList);	73	DataInfo recvData(neighborLinkData, rxSlicesList);
74	CHK_PRT_RET(	74	CHK_PRT_RET(
75	Recv(recvData, tempInsQues[queIdx], 0, true, dmaMode_),	75	Recv(recvData, tempInsQues[queIdx], 0, true, dmaMode_),
76	- HCCL_ERROR("[InsTempReduceAicpuReduce] BatchSend failed"), HcclResult::HCCL_E_INTERNAL);	76	+ HCCL_ERROR("[InsTempReduceAicpuReduce] BatchRecv failed"), HcclResult::HCCL_E_INTERNAL);
77	}	77	}
78	CHK_RET(PostSyncInterQueues(tempInsQues));	78	CHK_RET(PostSyncInterQueues(tempInsQues));
79	} else {	79	} else {

Mtest/ut/framework/next/coll_comms/communicator/group_schedule_mgr/ut_group_schedule_mgr.cc+14-1

	@@ -12,12 +12,12 @@
12	#include <mockcpp/mockcpp.hpp>	12	#include <mockcpp/mockcpp.hpp>
13	#include <vector>	13	#include <vector>
14	#include <string>	14	#include <string>
15	-#include "group_schedule_mgr.h"
16	#include "hccl/hccl_types.h"	15	#include "hccl/hccl_types.h"
17	#include "hccl/hccl_launch.h"	16	#include "hccl/hccl_launch.h"
18	#include "hccl/hccl_rank_graph.h"	17	#include "hccl/hccl_rank_graph.h"
19	#include "acl/acl_base_rt.h"	18	#include "acl/acl_base_rt.h"
20	#define private public	19	#define private public
		20	+#include "group_schedule_mgr.h"
21	#include "hccl_comm_pub.h"	21	#include "hccl_comm_pub.h"
22	#include "coll_comm.h"	22	#include "coll_comm.h"
23	#undef private	23	#undef private
	@@ -208,3 +208,16 @@ TEST_F(GroupScheduleMgrTest, Ut_pow2Up_When_InputPowerOfTwo_Expect_ReturnSameVal
208	}	208	}
209	EXPECT_EQ(power, expected);	209	EXPECT_EQ(power, expected);
210	}	210	}
		211	+
		212	+TEST_F(GroupScheduleMgrTest, Ut_GetCurLocalRank_When_RankNotInList_Expect_InternalError)
		213	+{
		214	+ GroupScheduleMgr mgr;
		215	+ mgr.userRank_ = 5;
		216	+ mgr.serverNum_ = 1;
		217	+ mgr.serverToRankSize_[0] = 2;
		218	+ mgr.serverToRankList_[0] = {0, 1};
		219	+
		220	+ uint32_t localRank = 0;
		221	+ HcclResult ret = mgr.GetCurLocalRank(localRank);
		222	+ EXPECT_EQ(ret, HCCL_E_INTERNAL);
		223	+}

Mtest/ut/framework/next/coll_comms/dfx/ut_cluster_monitor.cc+61-0

	@@ -20,6 +20,7 @@
20	#include "dfx/cluster_monitor/cluster_monitor.h"	20	#include "dfx/cluster_monitor/cluster_monitor.h"
21	#include "hccl_comm_pub.h"	21	#include "hccl_comm_pub.h"
22	#include "rank_graph_v2.h"	22	#include "rank_graph_v2.h"
		23	+#include "new_rank_info.h"
23	#undef private	24	#undef private
24	#include "hccl_comm_socket_c_adpt.h"	25	#include "hccl_comm_socket_c_adpt.h"
25	#include "base_config_legacy.h"	26	#include "base_config_legacy.h"
	@@ -959,6 +960,66 @@ TEST_F(ClusterMonitorTest, Ut_FormatUID_When_NormalInput_Expect_ReturnUid)
959	EXPECT_TRUE(uid.id[0] != '\0');	960	EXPECT_TRUE(uid.id[0] != '\0');
960	}	961	}
961		962
		963	+TEST_F(ClusterMonitorTest, Ut_GetSocketDescFromRankInfo_When_RemotePortInvalid_Expect_ParaError)
		964	+{
		965	+ auto ctx = CreateHcclCommForInsertTest();
		966	+ ASSERT_NE(ctx.hcclCommPtr, nullptr);
		967	+ HcclComm comm = static_cast<HcclComm>(ctx.hcclCommPtr.get());
		968	+
		969	+ uint32_t invalidPort = Hccl::MAX_VALUE_TCPPORT + 1;
		970	+ MOCKER_CPP(&Hccl::IRankGraph::GetDevicePort)
		971	+ .stubs()
		972	+ .with(mockcpp::any(), outBoundP(&invalidPort, sizeof(invalidPort)))
		973	+ .will(returnValue(HCCL_SUCCESS));
		974	+
		975	+ SocketDesc socketDesc{};
		976	+ HcclResult result = g_monitor.GetSocketDescFromRankInfo(comm, 0, 0, ClusterUIDType{}, socketDesc);
		977	+ EXPECT_EQ(result, HCCL_E_PARA);
		978	+
		979	+ GlobalMockObject::verify();
		980	+}
		981	+
		982	+// 对端端口正常，本端作为server但监听端口非法时命中 GetSocketDescFromRankInfo 的校验分支
		983	+static HcclResult GetDevicePortStub(Hccl::IRankGraph* self, uint32_t rank, uint32_t* devPort)
		984	+{
		985	+ // remoteRank=0 返回合法端口，myRankId=1 返回非法端口（触发 listenPort 校验）
		986	+ *devPort = (rank == 0) ? 10000 : (Hccl::MAX_VALUE_TCPPORT + 1);
		987	+ return HCCL_SUCCESS;
		988	+}
		989	+
		990	+static HcclResult HcclRankGraphGetLinksDeviceStub(
		991	+ HcclComm comm, uint32_t netLayer, uint32_t srcRank, uint32_t dstRank, CommLink** links, uint32_t* linkNum)
		992	+{
		993	+ static CommLink link;
		994	+ (void)memset_s(&link, sizeof(link), 0, sizeof(link));
		995	+ // local(src)地址 < remote(dst)地址，使本端作为server监听
		996	+ link.srcEndpointDesc.commAddr.type = CommAddrType::COMM_ADDR_TYPE_IP_V4;
		997	+ inet_pton(AF_INET, "1.0.0.1", &link.srcEndpointDesc.commAddr.addr);
		998	+ link.srcEndpointDesc.loc.locType = EndpointLocType::ENDPOINT_LOC_TYPE_DEVICE;
		999	+ link.dstEndpointDesc.commAddr.type = CommAddrType::COMM_ADDR_TYPE_IP_V4;
		1000	+ inet_pton(AF_INET, "2.0.0.1", &link.dstEndpointDesc.commAddr.addr);
		1001	+ link.dstEndpointDesc.loc.locType = EndpointLocType::ENDPOINT_LOC_TYPE_DEVICE;
		1002	+ *links = &link;
		1003	+ *linkNum = 1;
		1004	+ return HCCL_SUCCESS;
		1005	+}
		1006	+
		1007	+TEST_F(ClusterMonitorTest, Ut_GetSocketDescFromRankInfo_When_ListenPortInvalid_Expect_ParaError)
		1008	+{
		1009	+ auto ctx = CreateHcclCommForInsertTest();
		1010	+ ASSERT_NE(ctx.hcclCommPtr, nullptr);
		1011	+ HcclComm comm = static_cast<HcclComm>(ctx.hcclCommPtr.get());
		1012	+
		1013	+ MOCKER_CPP(&Hccl::IRankGraph::GetDevicePort).stubs().will(invoke(GetDevicePortStub));
		1014	+ MOCKER(HcclRankGraphGetLinks).stubs().will(invoke(HcclRankGraphGetLinksDeviceStub));
		1015	+
		1016	+ SocketDesc socketDesc{};
		1017	+ HcclResult result = g_monitor.GetSocketDescFromRankInfo(comm, 0, 0, ClusterUIDType{}, socketDesc);
		1018	+ EXPECT_EQ(result, HCCL_E_PARA);
		1019	+
		1020	+ GlobalMockObject::verify();
		1021	+}
		1022	+
962	TEST_F(ClusterMonitorTest, Ut_GetUID_When_NormalInput_Expect_ReturnString)	1023	TEST_F(ClusterMonitorTest, Ut_GetUID_When_NormalInput_Expect_ReturnString)
963	{	1024	{
964	std::string netInstId = "testInstance";	1025	std::string netInstId = "testInstance";