[v1.22.x] prov/psm3: update provider to sync with IEFS 11.7.0.0.110

Updates: - Improved auto-tuning features for PSM3, including dynamic Credit Flows and detecting the presence of the rv kernel module. - Improved PSM3 intra-node performance for large message sizes. Signed-off-by: Scott Breyer <[email protected]> (cherry picked from commit 386f574)
ofiwg · Jul 20, 2024 · 19d2bf9 · 19d2bf9
1 parent bdaa18b
commit 19d2bf9
Show file tree

Hide file tree

Showing 72 changed files with 1,300 additions and 1,297 deletions.
diff --git a/prov/psm3/Makefile.include b/prov/psm3/Makefile.include
@@ -101,10 +101,6 @@ prov_psm3_psm3_libptl_ips_la_SOURCES = \
 	prov/psm3/psm3/ptl_ips/ips_recvq.h \
 	prov/psm3/psm3/ptl_ips/ips_scb.c \
 	prov/psm3/psm3/ptl_ips/ips_scb.h \
-	prov/psm3/psm3/ptl_ips/ips_tid.c \
-	prov/psm3/psm3/ptl_ips/ips_tid.h \
-	prov/psm3/psm3/ptl_ips/ips_tidcache.c \
-	prov/psm3/psm3/ptl_ips/ips_tidcache.h \
 	prov/psm3/psm3/ptl_ips/ips_tidflow.c \
 	prov/psm3/psm3/ptl_ips/ips_tidflow.h \
 	prov/psm3/psm3/ptl_ips/ptl.c \

diff --git a/prov/psm3/VERSION b/prov/psm3/VERSION
@@ -1 +1 @@
-3_6_0_1
+3_7_0_0
diff --git a/prov/psm3/autogen.sh b/prov/psm3/autogen.sh
@@ -5,14 +5,6 @@ if test ! -f src/psmx3.h; then
 	exit 1
 fi
 
-if [ -f psm3/Makefile.include.base ]
-then
-	make -f - <<EOF
-psm3/Makefile.include: psm3/Makefile.include.base
-	cp psm3/Makefile.include.base psm3/Makefile.include
-EOF
-fi
-
 set -x
 
 autoreconf -ivf
diff --git a/prov/psm3/configure.ac b/prov/psm3/configure.ac
@@ -129,7 +129,7 @@ AC_ARG_ENABLE([psm3-udp],
     ],[],[enable_psm3_udp=no])
 AC_ARG_ENABLE([psm3-rc],
     [AS_HELP_STRING([--enable-psm3-rc],
-        [EXPERIMENTAL: Enable User Space RC QPs on applicable HALs @<:@default=[Verbs HAL]@:>@])
+        [EXPERIMENTAL: Enable User Space RC QPs on applicable HALs @<:@default=check [check means match --enable-psm3-verbs option]@:>@])
     ],[],[enable_psm3_rc=check])
 AS_IF([test "x$enable_psm3_udp" = "xyes"],
       [
@@ -429,6 +429,10 @@ AS_IF([test x"$enable_atomics" != x"no"],
 ])
 unset LIBS_save
 
+dnl Check for 128-bit integer support
+AC_CHECK_TYPE([__int128],
+	[AC_DEFINE(HAVE___INT128, 1, [Set to 1 to use 128-bit ints])])
+
 dnl Check for gcc cpuid intrinsics
 AC_MSG_CHECKING(compiler support for cpuid)
 AC_LINK_IFELSE([AC_LANG_PROGRAM([[

diff --git a/prov/psm3/configure.m4 b/prov/psm3/configure.m4
@@ -456,7 +456,7 @@ AC_ARG_ENABLE([psm3-udp],
     [enable_psm3_udp=no])
 AC_ARG_ENABLE([psm3-rc],
     [AS_HELP_STRING([--enable-psm3-rc],
-        [EXPERIMENTAL: Enable User Space RC QPs on applicable HALs @<:@default=[Verbs HAL]@:>@])],
+        [EXPERIMENTAL: Enable User Space RC QPs on applicable HALs @<:@default=check [check means match --enable-psm3-verbs option]@:>@])],
     [],
     [enable_psm3_rc=check])
 dnl ------------- Extra Features

diff --git a/prov/psm3/debian/changelog b/prov/psm3/debian/changelog
@@ -1,4 +1,4 @@
-libpsm3-fi (11.6.0.0-231) unstable; urgency=medium
+libpsm3-fi (11.7.0.0-110) unstable; urgency=medium
 
   * Initial release 
 

diff --git a/prov/psm3/psm3/Makefile.include b/prov/psm3/psm3/Makefile.include
@@ -66,10 +66,6 @@ psm3_libptl_ips_la_SOURCES = \
 	psm3/ptl_ips/ips_recvq.h \
 	psm3/ptl_ips/ips_scb.c \
 	psm3/ptl_ips/ips_scb.h \
-	psm3/ptl_ips/ips_tid.c \
-	psm3/ptl_ips/ips_tid.h \
-	psm3/ptl_ips/ips_tidcache.c \
-	psm3/ptl_ips/ips_tidcache.h \
 	psm3/ptl_ips/ips_tidflow.c \
 	psm3/ptl_ips/ips_tidflow.h \
 	psm3/ptl_ips/ptl.c \

diff --git a/prov/psm3/psm3/hal_sockets/sockets_ep.c b/prov/psm3/psm3/hal_sockets/sockets_ep.c
@@ -1125,7 +1125,7 @@ psm3_sockets_ips_proto_init(struct ips_proto *proto, uint32_t cksum_sz)
 		// unfortunately default TCP max_buffering (16K) is too small
 		// so flow_credit_bytes < 16K would prevent getting a good pipeline of
 		// packets/ACKs going
-		proto->flow_credit_bytes = ep->mtu * proto->flow_credits;
+		proto->flow_credit_bytes = ep->mtu * proto->max_credits;
 	} else {
 		// sockets buffering needs to place an upper bound on bytes
 		// while flow_credits places an upper bound on pkts
@@ -1229,7 +1229,7 @@ psm3_sockets_ips_proto_init(struct ips_proto *proto, uint32_t cksum_sz)
 }
 
 // Fetch current link state to update linkinfo fields in ips_proto:
-// 	ep_base_lid, ep_lmc, ep_link_rate, QoS tables, CCA tables
+// 	ep_base_lid, ep_lmc, ep_link_rate
 // These are all fields which can change during a link bounce.
 // Note "active" state is not adjusted as on link down PSM will wait for
 // the link to become usable again so it's always a viable/active device

diff --git a/prov/psm3/psm3/hal_sockets/sockets_gdrcpy.c b/prov/psm3/psm3/hal_sockets/sockets_gdrcpy.c
@@ -58,7 +58,6 @@
 #include <sys/ioctl.h>
 #include <sys/types.h>
 #include "ips_proto.h"
-#include "ptl_ips/ips_tid.h"
 #include "ptl_ips/ips_expected_proto.h"
 
 // flags=0 for send, 1 for recv

diff --git a/prov/psm3/psm3/hal_sockets/sockets_hal.c b/prov/psm3/psm3/hal_sockets/sockets_hal.c
@@ -73,7 +73,7 @@ static int psm3_hfp_sockets_initialize(psmi_hal_instance_t *phi,
 #if defined(PSM_CUDA) || defined(PSM_ONEAPI)
 	// testing on HED-2629 suggests turning off RNDV can help
 	// latency for messages in size 8-256 KB
-	gpu_thresh_rndv = SOCKET_GPU_THRESH_RNDV;
+	psm3_gpu_thresh_rndv = SOCKET_GPU_THRESH_RNDV;
 #endif
 	/* we initialize a few HAL software specific capabilities which
 	 * are known before context_open can open RV or parse HAL specific
@@ -175,11 +175,11 @@ static void psm3_hfp_sockets_mq_init_defaults(struct psm2_mq *mq)
 	 * corresponding PSM3_* env variables.
 	 * Otherwise these defaults are used.
 	 */
-	mq->hfi_thresh_rv = PSM_MQ_NIC_RNDV_THRESH;
+	mq->rndv_nic_thresh = PSM3_MQ_RNDV_NIC_THRESH;
 	mq->ips_cpu_window_rv_str = PSM_CPU_NIC_RNDV_WINDOW_STR;
 	// Even without RDMA do we want to disable rendezvous?
 	// even without RDMA, the receiver controlled pacing helps scalability
-	mq->hfi_thresh_rv = (~(uint32_t)0); // disable rendezvous
+	mq->rndv_nic_thresh = (~(uint32_t)0); // disable rendezvous
 	mq->hfi_thresh_tiny = PSM_MQ_NIC_MAX_TINY;
 #if defined(PSM_CUDA) || defined(PSM_ONEAPI)
 	if (PSMI_IS_GPU_ENABLED)
@@ -220,16 +220,6 @@ static int psm3_hfp_sockets_get_unit_active(int unit)
 	return psm3_sockets_get_unit_active(unit, SIMS_FILTER);
 }
 
-static int psm3_hfp_sockets_get_num_contexts(int unit)
-{
-	return 1024;
-}
-
-static int psm3_hfp_sockets_get_num_free_contexts(int unit)
-{
-	return 1024;
-}
-
 static int psm3_hfp_sockets_get_default_pkey(void)
 {
 	return 0;	/* use slot 0 as default */
@@ -305,8 +295,6 @@ static hfp_sockets_t psm3_sockets_hi = {
 		.hfp_get_num_ports			  = psm3_hfp_sockets_get_num_ports,
 		.hfp_get_unit_active			  = psm3_hfp_sockets_get_unit_active,
 		.hfp_get_port_active			  = psm3_hfp_sockets_get_port_active,
-		.hfp_get_num_contexts			  = psm3_hfp_sockets_get_num_contexts,
-		.hfp_get_num_free_contexts		  = psm3_hfp_sockets_get_num_free_contexts,
 		.hfp_get_default_pkey			  = psm3_hfp_sockets_get_default_pkey,
 		.hfp_get_port_subnet			  = psm3_hfp_sockets_get_port_subnet,
 		.hfp_get_unit_pci_bus			  = psm3_hfp_sockets_get_unit_pci_bus,

diff --git a/prov/psm3/psm3/hal_sockets/sockets_hal_inline_i.h b/prov/psm3/psm3/hal_sockets/sockets_hal_inline_i.h
@@ -189,7 +189,7 @@ static PSMI_HAL_INLINE psm2_error_t psm3_hfp_sockets_ips_proto_init(
 }
 
 // Fetch current link state to update linkinfo fields in ips_proto:
-// 	ep_base_lid, ep_lmc, ep_link_rate, QoS tables, CCA tables
+// 	ep_base_lid, ep_lmc, ep_link_rate
 // These are all fields which can change during a link bounce.
 // Note "active" state is not adjusted as on link down PSM will wait for
 // the link to become usable again so it's always a viable/active device
@@ -409,7 +409,7 @@ static PSMI_HAL_INLINE void psm3_hfp_sockets_ips_ipsaddr_disconnect(
 {
 }
 
-/* Handle HAL specific initialization of ibta path record query, CCA
+/* Handle HAL specific initialization of ibta path record query
  * and dispersive routing
  */
 static PSMI_HAL_INLINE psm2_error_t psm3_hfp_sockets_ips_ibta_init(

diff --git a/prov/psm3/psm3/hal_sockets/sockets_proto.c b/prov/psm3/psm3/hal_sockets/sockets_proto.c
@@ -74,7 +74,7 @@ psm3_tcp_proto_local_ack(struct ips_proto *proto, struct ips_flow *flow)
 
 	psmi_seqnum_t last_seq_num = STAILQ_LAST(unackedq, ips_scb, nextq)->seq_num;
 	while (between((scb = STAILQ_FIRST(unackedq))->seq_num.psn_num,
-		       last_seq_num.psn_num, flow->xmit_ack_num.psn_num-1)
+		       last_seq_num.psn_num, (flow->xmit_ack_num.psn_num-1) & proto->psn_mask)
 	    ) {
 		STAILQ_REMOVE_HEAD(unackedq, nextq);
 #ifdef PSM_DEBUG
@@ -151,6 +151,7 @@ psm3_tcp_proto_flow_flush_pio(struct ips_flow *flow, int *nflushed)
 		// local ack
 		if (scb) { // this check is unnecessary, but can make KW happy
 			flow->xmit_ack_num.psn_num = 1 + (__be32_to_cpu(scb->ips_lrh.bth[2]) & proto->psn_mask);
+			flow->xmit_ack_num.psn_num &= proto->psn_mask;
 		}
 		psm3_tcp_proto_local_ack(proto, flow);
 	}

diff --git a/prov/psm3/psm3/hal_verbs/verbs_ep.c b/prov/psm3/psm3/hal_verbs/verbs_ep.c
@@ -397,8 +397,6 @@ psm3_verbs_parse_params(psm2_ep_t ep)
 	// min size is (HFI_TF_NFLOWS + ep->hfi_num_send_rdma) *
 	// chunk size (psm3_mq_max_window_rv(mq, 0) after
 	// psm3_mq_initialize_params)
-	// for OPA native, actual window_rv may be smaller, but for UD it
-	// is not reduced
 	psm3_getenv("PSM3_RV_MR_CACHE_SIZE",
 			"kernel space MR cache size"
 			" (MBs, 0 lets rv module decide) [0]",
@@ -550,7 +548,7 @@ psm3_verbs_ips_proto_init(struct ips_proto *proto, uint32_t cksum_sz)
 	ep->chunk_max_size = ep->mtu;
 #ifdef PSM_BYTE_FLOW_CREDITS
 	// let flow_credits be the control
-	proto->flow_credit_bytes = ep->mtu * proto->flow_credits;
+	proto->flow_credit_bytes = ep->mtu * proto->max_credits;
 	_HFI_DBG("initial flow_credits %d bytes %d\n",
 				proto->flow_credits, proto->flow_credit_bytes);
 #else
@@ -594,7 +592,7 @@ psm3_verbs_ips_proto_init(struct ips_proto *proto, uint32_t cksum_sz)
 }
 
 // Fetch current link state to update linkinfo fields in ips_proto:
-// 	ep_base_lid, ep_lmc, ep_link_rate, QoS tables, CCA tables
+// 	ep_base_lid, ep_lmc, ep_link_rate
 // These are all fields which can change during a link bounce.
 // Note "active" state is not adjusted as on link down PSM will wait for
 // the link to become usable again so it's always a viable/active device
@@ -2884,8 +2882,11 @@ unsigned psm3_verbs_parse_rdmamode(int reload)
 	// IPS_PROTOEXP_FLAGS_INTERLEAVE are N/A when RDMA not enabled
 
 	default_value = 0;
-#if defined(PSM_CUDA) || defined(PSM_ONEAPI)
 #ifdef RNDV_MOD
+	if (psm3_rv_available()) {
+		default_value = IPS_PROTOEXP_FLAG_RDMA_KERNEL;
+	}
+#if defined(PSM_CUDA) || defined(PSM_ONEAPI)
 	// GPUDIRECT causes default_value of RDMA=1
 	if (PSMI_IS_GPU_ENABLED && psmi_parse_gpudirect())
 		default_value = IPS_PROTOEXP_FLAG_RDMA_KERNEL;

diff --git a/prov/psm3/psm3/hal_verbs/verbs_gdrcpy.c b/prov/psm3/psm3/hal_verbs/verbs_gdrcpy.c
@@ -58,7 +58,6 @@
 #include <sys/ioctl.h>
 #include <sys/types.h>
 #include "ips_proto.h"
-#include "ptl_ips/ips_tid.h"
 #include "ptl_ips/ips_expected_proto.h"
 
 // flags=0 for send, 1 for recv

diff --git a/prov/psm3/psm3/hal_verbs/verbs_hal.c b/prov/psm3/psm3/hal_verbs/verbs_hal.c
@@ -166,12 +166,12 @@ static void psm3_hfp_verbs_mq_init_defaults(struct psm2_mq *mq)
 	 * Otherwise these defaults are used.
 	 */
 	unsigned rdmamode = psm3_verbs_parse_rdmamode(1);
-	mq->hfi_thresh_rv = PSM_MQ_NIC_RNDV_THRESH;
+	mq->rndv_nic_thresh = PSM3_MQ_RNDV_NIC_THRESH;
 	mq->ips_cpu_window_rv_str = PSM_CPU_NIC_RNDV_WINDOW_STR;
 	if (! (rdmamode & IPS_PROTOEXP_FLAG_ENABLED)) {
 		// TBD - when RDMA is disabled do we want to disable rendezvous?
 		// even without RDMA, the receiver controlled pacing helps scalability
-		mq->hfi_thresh_rv = (~(uint32_t)0); // disable rendezvous
+		mq->rndv_nic_thresh = (~(uint32_t)0); // disable rendezvous
 	}
 	mq->hfi_thresh_tiny = PSM_MQ_NIC_MAX_TINY;
 #if defined(PSM_CUDA) || defined(PSM_ONEAPI)
@@ -213,16 +213,6 @@ static int psm3_hfp_verbs_get_unit_active(int unit)
 	return psm3_verbs_get_unit_active(unit, VIMS_FILTER);
 }
 
-static int psm3_hfp_verbs_get_num_contexts(int unit)
-{
-	return 1024;
-}
-
-static int psm3_hfp_verbs_get_num_free_contexts(int unit)
-{
-	return 1024;
-}
-
 static int psm3_hfp_verbs_get_default_pkey(void)
 {
 	return 0;	/* use slot 0 as default */
@@ -293,8 +283,6 @@ static hfp_verbs_t psm3_verbs_hi = {
 		.hfp_get_num_ports			  = psm3_hfp_verbs_get_num_ports,
 		.hfp_get_unit_active			  = psm3_hfp_verbs_get_unit_active,
 		.hfp_get_port_active			  = psm3_hfp_verbs_get_port_active,
-		.hfp_get_num_contexts			  = psm3_hfp_verbs_get_num_contexts,
-		.hfp_get_num_free_contexts		  = psm3_hfp_verbs_get_num_free_contexts,
 		.hfp_get_default_pkey			  = psm3_hfp_verbs_get_default_pkey,
 		.hfp_get_port_subnet			  = psm3_hfp_verbs_get_port_subnet,
 		.hfp_get_unit_pci_bus			  = psm3_hfp_verbs_get_unit_pci_bus,

diff --git a/prov/psm3/psm3/hal_verbs/verbs_hal_inline_i.h b/prov/psm3/psm3/hal_verbs/verbs_hal_inline_i.h
@@ -181,7 +181,7 @@ static PSMI_HAL_INLINE psm2_error_t psm3_hfp_verbs_ips_proto_init(
 }
 
 // Fetch current link state to update linkinfo fields in ips_proto:
-// 	ep_base_lid, ep_lmc, ep_link_rate, QoS tables, CCA tables
+// 	ep_base_lid, ep_lmc, ep_link_rate
 // These are all fields which can change during a link bounce.
 // Note "active" state is not adjusted as on link down PSM will wait for
 // the link to become usable again so it's always a viable/active device
@@ -610,7 +610,7 @@ static PSMI_HAL_INLINE void psm3_hfp_verbs_ips_ipsaddr_disconnect(
 #endif
 }
 
-/* Handle HAL specific initialization of ibta path record query, CCA
+/* Handle HAL specific initialization of ibta path record query
  * and dispersive routing
  */
 static PSMI_HAL_INLINE psm2_error_t psm3_hfp_verbs_ips_ibta_init(

diff --git a/prov/psm3/psm3/hal_verbs/verbs_service.h b/prov/psm3/psm3/hal_verbs/verbs_service.h
@@ -112,7 +112,5 @@ int psm3_verbs_get_unit_active(int unit, enum verbs_init_max_speed init_max_spee
    returns <= 0 if no port on any of the units is active. */
 int psm3_hfp_verbs_have_active_unit(int num_units);
 
-/* get the number of contexts from the unit id. */
-int psm3_verbs_get_num_contexts(int unit);
 #endif /* PSM_HAL_VERBS_SERVICE_H */
 #endif /* PSM_VERBS */
diff --git a/prov/psm3/psm3/include/utils_env.h b/prov/psm3/psm3/include/utils_env.h
@@ -153,6 +153,8 @@ int MOCKABLE(psm3_getenv_range)(const char *name, const char *descr,
 		union psmi_envvar_val *newval);
 MOCK_DCL_EPILOGUE(psm3_getenv_range);
 
+int psm3_count_tuples(const char *str);
+
 /*
  * Parsing int, unsigned int and long parameters
  * 0 -> ok, *val updated