btl/uct: complete re-work of the BTL

This commit is large and contains the following changes:

 - Disconnect the connection memory domain from the communication domain. This
   allows any memory domain to be used for connections. The default is to use
   tcp but it can be disabled which will allow UD and others to be used.

 - Move tl attributes off od the tl context structure.  In theory tl
   attributes do not differ betweeen contexts so query them once when the tl
   is created not once per context. This removes the need to allocate the
   first context so that code has also been removed.

 - Change mca_btl_uct_tl_t uct_dev_contexts member to be an array. The btl
   always allocates the maximum number of contexts. This is not a significant
   amount of memory. Rather than reduce it to be based on the configured
   maximum number of contexts it makes sense to just make it an array and
   remove the extra indirection when accessing the contexts.

 - Do not call mca_btl_uct_endpoint_set_flag before sending a message on the
   connection endpoint. This method may cause the release of the connection
   endpoint (cached on the BTL endpoint). If this happens it would lead to a
   SEGV.

 - Flush the endpoint only when it is being released. There is no need to do so
   on every send. Releasing the endpoint without flushing it may lead to it
   being destroyed while still processing data.

 - Downgrade endpoint lock from recursive. Recursive locks are not needed for
   the endpoint lock.

 - Move the async context from the module to the tl. There is no real benefit
   from sharing the async context between tls. Given this and some other changes
   that will be made it makes sense to move it from the module to the tl.

 - Connection TLs are only used to form connections for connect-to-endpoint TLs.
   They do not need to belong to the same memory domain as the one they are used
   with so there is no need to rely on a BTL module. Moved the
   pending_connection_reqs to the tl and changes the code to support a NULL
   module for the connection tl.

 - Put active tls in a list on the mca_btl_uct_md_t structure This simplifies
   the code a bit by moving mca_btl_uct_tl_t ownership to the mca_btl_uct_md_t
   class.

 - There is an issue with btl/uct which prevents the usage of the standard
   btl_uct_ MCA variables (eager limit, flags, etc). Because of the way the btl
   was written these values are all determined directly from UCT and can not be
   changed using the MCA variable interface. To address this issue this commit
   breaks apart the initialization code and separates out the pieces that are
   necessary for discovery only. The discovery pieces now use a new set of
   variables that include the memory domain name and directly control the
   behavior for BTLs on that memory domain as well as enabling the usage of the
   btl_uct variable to control the defaults for these variables.

   Example, using memory domain irdma0 will create variables:
     btl_uct_irdma0_eager_limit, btl_uct_irdma0_max_send_size, etc.

   The defaults will be based on what is reported by UCT and the user can set
   the values to a subset of what UCT reports. For example, if the max send size
   for the hardware is 8192B then it can be set to anything up to and including
   that value. The same is true for feature flags, if the hardware supports only
   some btl atomics or operations the user can specify a subset of them (others
   will be ignored).

 - Move device context code to a new file. There is a specific header for device
   contexts so it makes sense to move the context-specific code to a matching C
   file. No changes in this other than moving code around.

 - Use uct_ep_am_short_iov for short messages. The uct_ep_am_short_iov method
   should allow for faster short messages than uct_ep_am_short (which can only
   take a single buffer). This commit moves btl/uct to the newer method which
   breaks compatibility with some version of UCT. Since we already no longer
   support those versions this change is safe.

Signed-off-by: Nathan Hjelm <hjelmn@google.com>

Author: hjelmn

opal/mca/btl/uct/Makefile.am

@@ -13,6 +13,7 @@
 # Copyright (c) 2017      IBM Corporation.  All rights reserved.
 # Copyright (c) 2017-2018 Los Alamos National Security, LLC. All rights
 #                         reserved.
+# Copyright (c) 2025      Google, LLC. All rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -24,22 +25,31 @@ AM_CPPFLAGS = $(btl_uct_CPPFLAGS)
 
 amca_paramdir = $(AMCA_PARAM_SETS_DIR)
 
-sources = \
+headers = \
     btl_uct.h \
+    btl_uct_rdma.h \
+    btl_uct_endpoint.h \
+    btl_uct_am.h \
+    btl_uct_frag.h \
+    btl_uct_types.h \
+    btl_uct_device_context.h \
+    btl_uct_discover.h \
+    btl_uct_modex.h \
+    btl_uct_include_list.h
+
+sources = \
     btl_uct_module.c \
     btl_uct_component.c \
-    btl_uct_rdma.h \
     btl_uct_rdma.c \
-    btl_uct_endpoint.h \
     btl_uct_endpoint.c \
     btl_uct_amo.c \
-    btl_uct_am.h \
     btl_uct_am.c \
-    btl_uct_frag.h \
     btl_uct_frag.c \
     btl_uct_tl.c \
-    btl_uct_types.h \
-    btl_uct_device_context.h
+    btl_uct_discover.c \
+    btl_uct_modex.c \
+    btl_uct_include_list.c \
+    btl_uct_device_context.c
 
 # Make the output library in this directory, and name it either
 # mca_<type>_<name>.la (for DSO builds) or libmca_<type>_<name>.la
@@ -50,20 +60,22 @@ lib =
 lib_sources =
 component = mca_btl_uct.la
 component_sources = $(sources)
+component_headers = $(headers)
 else
 lib = libmca_btl_uct.la
 lib_sources = $(sources)
+lib_headers = ${headers}
 component =
 component_sources =
 endif
 
 mcacomponentdir = $(opallibdir)
 mcacomponent_LTLIBRARIES = $(component)
-mca_btl_uct_la_SOURCES = $(component_sources)
+mca_btl_uct_la_SOURCES = $(component_sources) $(component_headers)
 mca_btl_uct_la_LDFLAGS = -module -avoid-version $(btl_uct_LDFLAGS)
 mca_btl_uct_la_LIBADD = $(top_builddir)/opal/lib@OPAL_LIB_NAME@.la $(btl_uct_LIBS)
 
 noinst_LTLIBRARIES = $(lib)
-libmca_btl_uct_la_SOURCES = $(lib_sources)
+libmca_btl_uct_la_SOURCES = $(lib_sources) $(lib_headers)
 libmca_btl_uct_la_LDFLAGS = -module -avoid-version $(btl_uct_LDFLAGS)
 libmca_btl_uct_la_LIBADD = $(btl_uct_LIBS)

opal/mca/btl/uct/btl_uct.h

@@ -64,6 +64,9 @@ struct mca_btl_uct_module_t {
     /** base BTL interface */
     mca_btl_base_module_t super;
 
+    /** module index in the component module array */
+    int module_index;
+
     /** whether the module has been fully initialized or not */
     bool initialized;
 
@@ -76,31 +79,15 @@ struct mca_btl_uct_module_t {
     /** mutex to protect the module */
     opal_recursive_mutex_t lock;
 
-    /** async context */
-    ucs_async_context_t *ucs_async;
-
     /** transport for active messaging */
     mca_btl_uct_tl_t *am_tl;
 
     /** transport for RDMA/AMOs */
     mca_btl_uct_tl_t *rdma_tl;
 
-    /** transport for forming connections (if needed) */
-    mca_btl_uct_tl_t *conn_tl;
-
-    /** array containing the am_tl and rdma_tl */
-    mca_btl_uct_tl_t *comm_tls[2];
-
-#if UCT_API >= UCT_VERSION(1, 7)
-    uct_component_h uct_component;
-#endif
-
     /** registration cache */
     mca_rcache_base_module_t *rcache;
 
-    /** name of the memory domain backing this module */
-    char *md_name;
-
     /** am and rdma share endpoints */
     bool shared_endpoints;
 
@@ -119,8 +106,9 @@ struct mca_btl_uct_module_t {
     /** frags that were waiting on connections that are now ready to send */
     opal_list_t pending_frags;
 
-    /** pending connection requests */
-    opal_fifo_t pending_connection_reqs;
+    /** allowed transports */
+    char *allowed_transports;
+    mca_btl_uct_include_list_t allowed_transport_list;
 };
 typedef struct mca_btl_uct_module_t mca_btl_uct_module_t;
 
@@ -133,6 +121,9 @@ struct mca_btl_uct_component_t {
     /** base BTL component */
     mca_btl_base_component_3_0_0_t super;
 
+    /** whether the component is initialized. controls cleanup. */
+    bool initialized;
+
     /** number of TL modules */
     int module_count;
 
@@ -141,10 +132,15 @@ struct mca_btl_uct_component_t {
 
     /** allowed UCT memory domains */
     char *memory_domains;
+    mca_btl_uct_include_list_t memory_domain_list;
 
     /** allowed transports */
     char *allowed_transports;
 
+    /** transports to consider for forming connections */
+    char *connection_domains;
+    mca_btl_uct_include_list_t connection_domain_list;
+
     /** number of worker contexts to create */
     int num_contexts_per_module;
 
@@ -158,6 +154,17 @@ struct mca_btl_uct_component_t {
 
     /** connection retry timeout */
     unsigned int connection_retry_timeout;
+
+#if UCT_API >= UCT_VERSION(1, 7)
+    uct_component_h *uct_components;
+    unsigned num_uct_components;
+#endif
+
+    /** list of memory domains (btl_uct_md_t) */
+    opal_list_t md_list;
+
+    /** connection transport (if needed). reference is owned by conn_md */
+    mca_btl_uct_tl_t *conn_tl;
 };
 typedef struct mca_btl_uct_component_t mca_btl_uct_component_t;
 
@@ -293,19 +300,24 @@ ucs_status_t mca_btl_uct_am_handler(void *arg, void *data, size_t length, unsign
 struct mca_btl_base_endpoint_t *mca_btl_uct_get_ep(struct mca_btl_base_module_t *module,
                                                    opal_proc_t *proc);
 
-int mca_btl_uct_query_tls(mca_btl_uct_module_t *module, mca_btl_uct_md_t *md,
-                          uct_tl_resource_desc_t *tl_descs, unsigned tl_count);
+int mca_btl_uct_populate_tls(mca_btl_uct_md_t *md, uct_tl_resource_desc_t *tl_descs, unsigned tl_count);
 int mca_btl_uct_process_connection_request(mca_btl_uct_module_t *module,
                                            mca_btl_uct_conn_req_t *req);
 
+mca_btl_uct_module_t *mca_btl_uct_alloc_module(mca_btl_uct_md_t *md,
+                                               size_t registration_size);
+
+int mca_btl_uct_evaluate_tl(mca_btl_uct_module_t *module, mca_btl_uct_tl_t *tl);
+int mca_btl_uct_enable_tl_conn(mca_btl_uct_tl_t *tl);
+
 /**
  * @brief Checks if a tl is suitable for using for RDMA
  *
  * @param[in] tl  btl/uct tl pointer
  */
 static inline bool mca_btl_uct_tl_supports_rdma(mca_btl_uct_tl_t *tl)
 {
-    return (MCA_BTL_UCT_TL_ATTR(tl, 0).cap.flags
+    return (tl->uct_iface_attr.cap.flags
             & (UCT_IFACE_FLAG_PUT_ZCOPY | UCT_IFACE_FLAG_GET_ZCOPY))
            == (UCT_IFACE_FLAG_PUT_ZCOPY | UCT_IFACE_FLAG_GET_ZCOPY);
 }
@@ -315,7 +327,7 @@ static inline bool mca_btl_uct_tl_supports_rdma(mca_btl_uct_tl_t *tl)
  */
 static inline bool mca_btl_uct_tl_support_am(mca_btl_uct_tl_t *tl)
 {
-    return (MCA_BTL_UCT_TL_ATTR(tl, 0).cap.flags
+    return (tl->uct_iface_attr.cap.flags
             & (UCT_IFACE_FLAG_AM_SHORT | UCT_IFACE_FLAG_AM_BCOPY | UCT_IFACE_FLAG_AM_ZCOPY));
 }
 
@@ -326,7 +338,7 @@ static inline bool mca_btl_uct_tl_support_am(mca_btl_uct_tl_t *tl)
  */
 static inline bool mca_btl_uct_tl_supports_conn(mca_btl_uct_tl_t *tl)
 {
-    return (MCA_BTL_UCT_TL_ATTR(tl, 0).cap.flags
+    return (tl->uct_iface_attr.cap.flags
             & (UCT_IFACE_FLAG_AM_SHORT | UCT_IFACE_FLAG_CONNECT_TO_IFACE))
            == (UCT_IFACE_FLAG_AM_SHORT | UCT_IFACE_FLAG_CONNECT_TO_IFACE);
 }
@@ -338,7 +350,11 @@ static inline bool mca_btl_uct_tl_supports_conn(mca_btl_uct_tl_t *tl)
  */
 static inline bool mca_btl_uct_tl_requires_connection_tl(mca_btl_uct_tl_t *tl)
 {
-    return !(MCA_BTL_UCT_TL_ATTR(tl, 0).cap.flags & UCT_IFACE_FLAG_CONNECT_TO_IFACE);
+    if (NULL == tl) {
+        return false;
+    }
+
+    return !(tl->uct_iface_attr.cap.flags & UCT_IFACE_FLAG_CONNECT_TO_IFACE);
 }
 
 END_C_DECLS