Updates the PSM2 library to match release 11.2.173

Fixes include:

Improved handling of the CUDA memory handle cache.

Improved handling of the case where the libpsm2 library is opened
multiple times by a single process. (This can happen when, for example,
multiple Open MPI transport layers all access libpsm2.)

Signed-off-by: Michael Heinz <[email protected]>

COMMIT

@@ -1 +1 @@
-bc17e0522f6b64e8e054d3cfea4506ac155724c2
+61553edd6b9fefce4a246b4701abc213e7c36b2b

Makefile

@@ -483,7 +483,7 @@ dist: distclean
 	PRUNE_LIST="";										\
 	for pd in ".git" "cscope*" "$(shell realpath --relative-to=${top_srcdir} ${OUTDIR})"	\
 		"*.orig" "*~" "#*" ".gitignore" "doc" "libcm" "psm.supp" "test" "psm_hal_MOCK"	\
-		 "tools" "artifacts" "*.rej.patch"; do			\
+		 "psm_test" "tools" "artifacts" "*.rej.patch"; do			\
 		PRUNE_LIST="$$PRUNE_LIST -name $$pd -prune -o";					\
 	done;											\
 	for hid in psm_hal_* ; do								\

README

@@ -220,6 +220,13 @@ Note: It is also possible to use rpm command to install rpm's, but it is recomme
 that one use yum/dnf as rpm tool has issues with name changes and obsoletes tags.
 yum or dnf should be better able to resolve dependency issues.
 
+TESTING
+=======
+
+Please see the subdirectory psm_test, starting with the file: psm_test/README for
+code and instructions on testing the psm2 library.  Additionally, the
+directory: psm_test/samples contains sample code to test.
+
 RELATED SOFTWARE TO PSM2
 ========================
 

include/rbtree.c

@@ -85,13 +85,22 @@
 
 #include <string.h> /* for memset declaration */
 
-#if !defined ( RBTREE_GET_LEFTMOST )       || \
+// RBTREE_CMP should be a comparator, i.e. RBTREE_CMP(a, b) should evaluate to
+// -1, 0, or 1 depending on if a < b, a == b, or a > b, respectively.
+#ifdef RBTREE_CMP
+
+#if defined(RBTREE_GET_LEFTMOST) || defined(RBTREE_GET_RIGHTMOST)
+#error Cannot define both RBTREE_CMP and RBTREE_GET_(LEFT|RIGHT)MOST
+#endif
+
+#elif !defined ( RBTREE_GET_LEFTMOST )       || \
 	! defined ( RBTREE_GET_RIGHTMOST ) || \
 	! defined ( RBTREE_MAP_COUNT )     || \
 	! defined ( RBTREE_ASSERT )
 #error "You must define RBTREE_GET_LEFTMOST and RBTREE_GET_RIGHTMOST and \
         RBTREE_MAP_COUNT and RBTREE_ASSERT before including rbtree.c"
-#endif
+
+#endif /* RBTREE_CMP */
 
 #define IN /* nothing */
 
@@ -117,13 +126,24 @@ static void ips_cl_qmap_remove_item(
 static cl_map_item_t* ips_cl_qmap_successor(
 				IN	cl_qmap_t* const	p_map,
 				IN	const cl_map_item_t*	p_item);
+
+
+#ifndef RBTREE_NO_EMIT_IPS_CL_QMAP_PREDECESSOR
 static cl_map_item_t* ips_cl_qmap_predecessor(
 				IN	cl_qmap_t* const	p_map,
 				IN	const cl_map_item_t*	p_item);
+#endif
+
+#if defined(RBTREE_GET_LEFTMOST)
 static cl_map_item_t* ips_cl_qmap_search(
 				IN	cl_qmap_t* const	p_map,
 				IN	unsigned long		start,
 				IN	unsigned long		end);
+#else
+static cl_map_item_t* ips_cl_qmap_searchv(
+				cl_qmap_t* const	p_map,
+				const RBTREE_MI_PL *key);
+#endif
 
 /*
  * Get the root.
@@ -380,7 +400,11 @@ ips_cl_qmap_insert_item(
 		p_insert_at = p_comp_item;
 
 		/* Traverse the tree until the correct insertion point is found. */
+#ifdef RBTREE_GET_LEFTMOST
 		if( RBTREE_GET_LEFTMOST(&p_item->payload) < RBTREE_GET_LEFTMOST(&p_insert_at->payload) )
+#else
+		if(RBTREE_CMP(&p_item->payload, &p_insert_at->payload) < 0)
+#endif
 		{
 			p_comp_item = p_insert_at->p_left;
 			compare_res = 1;
@@ -604,6 +628,11 @@ ips_cl_qmap_successor(
 	}
 }
 
+// When includer defines RBTREE_CMP, ips_cl_qmap_search() is not emitted.
+// When this happens, ips_cl_qmap_predecessor() may not be called.
+// Combined with -Werror -Wunused-function, libpsm2 fails to build.
+// So provide macro to control emitting this function
+#ifndef RBTREE_NO_EMIT_IPS_CL_QMAP_PREDECESSOR
 static cl_map_item_t *
 ips_cl_qmap_predecessor(
 	IN	cl_qmap_t* const		p_map,
@@ -627,7 +656,9 @@ ips_cl_qmap_predecessor(
 		return p_tmp;
 	}
 }
+#endif /* RBTREE_NO_EMIT_IPS_CL_QMAP_PREDECESSOR */
 
+#if defined(RBTREE_GET_LEFTMOST)
 /*
  * return the first node with buffer overlapping or zero.
  */
@@ -690,3 +721,23 @@ ips_cl_qmap_search(cl_qmap_t * const p_map,
 
 	return p_item;
 }
+#else /* defined(...LEFTMOST) || defined(...RIGHTMOST) */
+static cl_map_item_t *
+ips_cl_qmap_searchv(cl_qmap_t * const p_map, const RBTREE_MI_PL *key)
+{
+	RBTREE_ASSERT( p_map );
+	cl_map_item_t *p_item = __cl_map_root(p_map);
+
+	while (p_item != p_map->nil_item) {
+		if (RBTREE_CMP(key, &p_item->payload) > 0) {
+			p_item = p_item->p_right;
+		} else if (RBTREE_CMP(key, &p_item->payload) < 0) {
+			p_item = p_item->p_left;
+		} else {
+			break;
+		}
+	}
+
+	return p_item;
+}
+#endif /* defined(...LEFTMOST) || defined(...RIGHTMOST) */

psm.c

@@ -65,11 +65,14 @@ static int psmi_verno = PSMI_VERNO_MAKE(PSM2_VERNO_MAJOR, PSM2_VERNO_MINOR);
 static int psmi_verno_client_val;
 int psmi_epid_ver;
 
+// Special psmi_refcount values
 #define PSMI_NOT_INITIALIZED    0
-#define PSMI_INITIALIZED        1
-#define PSMI_FINALIZED         -1	/* Prevent the user from calling psm2_init
-					 * once psm_finalize has been called. */
-static int psmi_isinit = PSMI_NOT_INITIALIZED;
+#define PSMI_FINALIZED         -1
+
+// PSM2 doesn't support transitioning out of the PSMI_FINALIZED state
+// once psmi_refcount is set to PSMI_FINALIZED, any further attempts to change
+// psmi_refcount should be treated as an error
+static int psmi_refcount = PSMI_NOT_INITIALIZED;
 
 /* Global lock used for endpoint creation and destroy
  * (in functions psm2_ep_open and psm2_ep_close) and also
@@ -104,9 +107,8 @@ uint32_t gdr_copy_threshold_recv;
  * It is supposed to be filled with logical OR
  * on conditional compilation basis
  * along with future features/capabilities.
- * At the very beginning we start with Multi EPs.
  */
-uint64_t psm2_capabilities_bitset = PSM2_MULTI_EP_CAP;
+uint64_t psm2_capabilities_bitset = PSM2_MULTI_EP_CAP | PSM2_LIB_REFCOUNT_CAP;
 
 int psmi_verno_client()
 {
@@ -130,7 +132,7 @@ int psmi_verno_isinteroperable(uint16_t verno)
 
 int MOCKABLE(psmi_isinitialized)()
 {
-	return (psmi_isinit == PSMI_INITIALIZED);
+	return (psmi_refcount > 0);
 }
 MOCK_DEF_EPILOGUE(psmi_isinitialized);
 
@@ -356,10 +358,12 @@ psm2_error_t __psm2_init(int *major, int *minor)
 	GENERIC_PERF_SET_SLOT_NAME(PSM_TX_SPEEDPATH_CTR, "TX");
 	GENERIC_PERF_SET_SLOT_NAME(PSM_RX_SPEEDPATH_CTR, "RX");
 
-	if (psmi_isinit == PSMI_INITIALIZED)
+	if (psmi_refcount > 0) {
+		psmi_refcount++;
 		goto update;
+	}
 
-	if (psmi_isinit == PSMI_FINALIZED) {
+	if (psmi_refcount == PSMI_FINALIZED) {
 		err = PSM2_IS_FINALIZED;
 		goto fail;
 	}
@@ -435,7 +439,7 @@ psm2_error_t __psm2_init(int *major, int *minor)
 				((id.eax & CPUID_EXMODEL_MASK) >> 12);
 	}
 
-	psmi_isinit = PSMI_INITIALIZED;
+	psmi_refcount++;
 	/* hfi_debug lives in libhfi.so */
 	psmi_getenv("PSM2_TRACEMASK",
 		    "Mask flags for tracing",
@@ -520,7 +524,6 @@ psm2_error_t __psm2_init(int *major, int *minor)
 #endif
 
 update:
-
 	if (getenv("PSM2_IDENTIFY")) {
                 Dl_info info_psm;
 		char ofed_delta[100] = "";
@@ -557,6 +560,8 @@ psm2_error_t __psm2_init(int *major, int *minor)
 	*major = (int)psmi_verno_major;
 	*minor = (int)psmi_verno_minor;
 fail:
+	_HFI_DBG("psmi_refcount=%d,err=%u\n", psmi_refcount, err);
+
 	PSM2_LOG_MSG("leaving");
 	return err;
 }
@@ -779,7 +784,14 @@ psm2_error_t __psm2_finalize(void)
 
 	PSM2_LOG_MSG("entering");
 
+	_HFI_DBG("psmi_refcount=%d\n", psmi_refcount);
 	PSMI_ERR_UNLESS_INITIALIZED(NULL);
+	psmi_assert(psmi_refcount > 0);
+	psmi_refcount--;
+
+	if (psmi_refcount > 0) {
+		return PSM2_OK;
+	}
 
 	/* When PSM_PERF is enabled, the following line causes the
 	   instruction cycles gathered in the current run to be dumped
@@ -856,7 +868,7 @@ psm2_error_t __psm2_finalize(void)
 	}
 #endif
 
-	psmi_isinit = PSMI_FINALIZED;
+	psmi_refcount = PSMI_FINALIZED;
 	PSM2_LOG_MSG("leaving");
 	psmi_log_fini();
 

psm2.h

@@ -488,6 +488,7 @@ psm2_error_t psm2_init(int *api_verno_major, int *api_verno_minor);
  * consecutive bits : 0x2, 0x4 ... and so on.
  */
 #define PSM2_MULTI_EP_CAP 0x1	/* Multiple Endpoints capability */
+#define PSM2_LIB_REFCOUNT_CAP 0x2	/* Library finalization is managed with reference count */
 
 /** @brief PSM2 capabilities provider
  *

psm2_hal.c

@@ -326,14 +326,10 @@ static struct _psmi_hal_instance *psmi_hal_get_pi_inst(void)
 				p->params.num_ports = nports;
 				p->params.default_pkey = dflt_pkey;
 				p->params.sw_status |= valid_flags;
-				p->params.unit_active = (uint8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits,
-										sizeof(uint8_t));
-				p->params.unit_active_valid = (uint8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits,
-										      sizeof(uint8_t));
-				p->params.port_active = (uint8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits*nports,
-										sizeof(uint8_t));
-				p->params.port_active_valid = (uint8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits*nports,
-										      sizeof(uint8_t));
+				p->params.unit_active = (int8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits, sizeof(int8_t));
+				p->params.unit_active_valid = (int8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits, sizeof(int8_t));
+				p->params.port_active = (int8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits*nports, sizeof(int8_t));
+				p->params.port_active_valid = (int8_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits*nports, sizeof(int8_t));
 				p->params.num_contexts = (uint16_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits,
 										  sizeof(uint16_t));
 				p->params.num_contexts_valid = (uint16_t *) psmi_calloc(PSMI_EP_NONE, UNDEFINED, nunits,

psm2_hal.h

@@ -183,8 +183,8 @@ typedef struct _psmi_hal_params
 	uint16_t   num_units;
 	uint16_t   num_ports;
 	uint16_t   default_pkey;
-	uint8_t    *unit_active,*unit_active_valid;
-	uint8_t    *port_active,*port_active_valid;
+	int8_t     *unit_active,*unit_active_valid;
+	int8_t     *port_active,*port_active_valid;
 	uint16_t   *num_contexts,*num_contexts_valid;
 	uint16_t   *num_free_contexts,*num_free_contexts_valid;
 } psmi_hal_params_t;

psm_user.h

@@ -362,6 +362,36 @@ CUresult (*psmi_cuDevicePrimaryCtxRelease)(CUdevice device);
 		}							\
 	} while (0)
 
+/**
+ * Similar to PSMI_CUDA_CALL() except does not error out
+ * if func(args) returns CUDA_SUCCESS or except_err
+ *
+ * Invoker must provide 'CUresult cudaerr' in invoked scope
+ * so invoker can inspect whether cudaerr == CUDA_SUCCESS or
+ * cudaerr == except_err after expanded code is executed.
+ */
+#define PSMI_CUDA_CALL_EXCEPT(except_err, func, args...) do { \
+		cudaerr = psmi_##func(args);				\
+		if (cudaerr != CUDA_SUCCESS && cudaerr != except_err) {	\
+			if (ctxt == NULL)				\
+				_HFI_ERROR(				\
+				"Check if CUDA is initialized"	\
+				"before psm2_ep_open call \n");		\
+			_HFI_ERROR(					\
+				"CUDA failure: %s() (at %s:%d)"		\
+				"returned %d\n",			\
+				#func, __FILE__, __LINE__, cudaerr);	\
+			psmi_handle_error(				\
+				PSMI_EP_NORETURN, PSM2_INTERNAL_ERR,	\
+				"Error returned from CUDA function.\n");\
+		} else if (cudaerr == except_err) { \
+			_HFI_INFO( \
+				"CUDA warning: %s() (at %s:%d)"		\
+				"returned %d\n",			\
+				#func, __FILE__, __LINE__, cudaerr);	\
+		} \
+	} while (0)
+
 #define PSMI_CUDA_CHECK_EVENT(event, cudaerr) do {			\
 		cudaerr = psmi_cuEventQuery(event);			\
 		if ((cudaerr != CUDA_SUCCESS) &&			\