Libfabric tracer

Makefile.am

@@ -3,8 +3,9 @@ pkgconfigdir = $(libdir)/pkgconfig
 pkgconfig_DATA = thapi.pc
 
 ACLOCAL_AMFLAGS = -I m4
-SUBDIRS = utils xprof sampling backends/opencl backends/ze backends/cuda backends/omp backends/hip backends/mpi backends/cxi backends/itt backends/pytorch
-
+SUBDIRS = utils xprof sampling backends/opencl backends/ze backends/cuda backends/omp backends/hip backends/mpi backends/cxi backends/itt backends/pytorch backends/libfabric
 EXTRA_DIST = autogen.sh README.md .valgrind/dlopen.supp thapi.pc.in
 
 CLEANFILES = thapi.pc
+
+DIST_SUBDIRS = utils xprof sampling backends/opencl backends/ze backends/cuda backends/omp backends/hip backends/mpi backends/cxi backends/itt backends/pytorch backends/libfabric

backends/libfabric/Makefile.am

@@ -0,0 +1,169 @@
+if STRICT
+  WERROR = -Werror
+else
+  WERROR =
+endif
+
+#
+# ----------------------------------------------------------------------
+# Tracepoint provider
+# ----------------------------------------------------------------------
+#
+# libfabric_events.yaml -> (gen_custom_probes.rb) -> .tp -> (lttng-gen-tp)
+#   -> libfabric_tracepoints.{c,h}
+#
+# No h2yaml and no <name>_model.rb anywhere in this backend: the tracer is
+# hand-written, so there is no header to parse.  backends/pytorch and
+# backends/cxi are the two other backends built this way.
+#
+
+FI_STATIC_PROBES = libfabric_tracepoints
+
+FI_STATIC_PROBES_TP = $(FI_STATIC_PROBES:=.tp)
+
+FI_STATIC_PROBES_INCL = $(FI_STATIC_PROBES:=.h)
+
+FI_STATIC_PROBES_SRC = $(FI_STATIC_PROBES:=.c)
+
+$(FI_STATIC_PROBES_TP): %.tp: $(top_srcdir)/utils/gen_custom_probes.rb $(srcdir)/libfabric_events.yaml $(top_srcdir)/utils/gen_probe_base.rb
+	$(RUBY) $< $(srcdir)/libfabric_events.yaml lttng_ust_fi > $@
+
+%.h %.c: %.tp
+	$(LTTNG_GEN_TP) $< -o $*.c -o $*.h
+
+noinst_LTLIBRARIES = libfitracepoints.la
+
+nodist_libfitracepoints_la_SOURCES = \
+	$(FI_STATIC_PROBES_INCL) \
+	$(FI_STATIC_PROBES_SRC)
+
+libfitracepoints_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I./
+libfitracepoints_la_CFLAGS = -fPIC -Wall -Wextra -Wno-unused-parameter -Wno-type-limits -Wno-sign-compare $(WERROR) $(LTTNG_UST_CFLAGS)
+libfitracepoints_la_LDFLAGS = $(LTTNG_UST_LIBS)
+
+EXTRA_DIST = \
+	libfabric_events.yaml \
+	$(top_srcdir)/utils/gen_custom_probes.rb \
+	include
+
+CLEANFILES = \
+	$(FI_STATIC_PROBES_TP) \
+	$(FI_STATIC_PROBES_INCL) \
+	$(FI_STATIC_PROBES_SRC)
+
+BUILT_SOURCES = \
+	$(FI_STATIC_PROBES_INCL)
+
+#
+# ----------------------------------------------------------------------
+# Tracer
+# ----------------------------------------------------------------------
+#
+# Note what is absent compared to backends/pytorch: there is no dummy_libs
+# stanza, because we never link against libfabric.  Every provider entry point
+# is reached either through dlvsym(RTLD_NEXT, "fi_fabric", ...) or through a
+# function pointer in a vtable the provider handed us, so the only thing
+# libfabric contributes at build time is its headers.
+#
+# The headers are vendored under include/rdma/ so the backend builds on a
+# machine with no libfabric installed -- same reason backends/mpi vendors
+# mpi.h.  Update them with:
+#     cp -r <libfabric>/include/rdma backends/libfabric/include/
+#
+
+bin_SCRIPTS = \
+	tracer_libfabric.sh
+
+lib_LTLIBRARIES = libTracerLibfabric.la
+
+nodist_libTracerLibfabric_la_SOURCES = \
+	$(FI_STATIC_PROBES_INCL)
+
+libTracerLibfabric_la_SOURCES = \
+	src/fit_hook.c \
+	src/fit_addr.c \
+	src/fit_lttng.c
+
+libTracerLibfabric_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I$(srcdir)/src -I./
+libTracerLibfabric_la_CFLAGS = -Wall -Wextra -Wno-unused-parameter -Wno-missing-field-initializers $(WERROR) $(LTTNG_UST_CFLAGS)
+libTracerLibfabric_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -avoid-version -module
+libTracerLibfabric_la_LIBADD = libfitracepoints.la
+
+EXTRA_DIST += \
+	src/fi_trace.h
+
+#
+# ----------------------------------------------------------------------
+# Analysis: raw lttng_ust_fi trace -> interval stream
+# ----------------------------------------------------------------------
+#
+# Three names here are load-bearing and fail confusingly rather than loudly if
+# you change them:
+#
+#   -p fiinterval -c interval   utils/babeltrace_thapi.in:161 looks the
+#                               component up as "filter.#{b}interval.interval"
+#                               where b is the short backend name, "fi".
+#   libFIInterval.la            utils/test_wrapper_thapi_text_pretty.sh.in:69
+#                               builds lib"${be^^}"Interval.so.
+#   $(pkglibdir)/bt2            babeltrace_thapi.in:151 scans exactly this one
+#                               directory for plugins.
+#
+
+BTX_FI_GENERATED = \
+	btx_filter_fi/metababel/metababel.h \
+	btx_filter_fi/metababel/btx_component.h \
+	btx_filter_fi/metababel/btx_component.c \
+	btx_filter_fi/metababel/btx_upstream.h \
+	btx_filter_fi/metababel/btx_upstream.c \
+	btx_filter_fi/metababel/btx_downstream.h \
+	btx_filter_fi/metababel/btx_downstream.c \
+	btx_filter_fi/btx_main.c
+
+EXTRA_DIST += \
+	$(top_srcdir)/xprof/btx_interval_model.yaml \
+	$(srcdir)/btx_libfabric_model.yaml
+
+$(BTX_FI_GENERATED) &: $(top_srcdir)/xprof/btx_interval_model.yaml $(srcdir)/btx_libfabric_model.yaml
+	$(METABABEL) -u $(srcdir)/btx_libfabric_model.yaml -d $(top_srcdir)/xprof/btx_interval_model.yaml -t FILTER -o btx_filter_fi -p fiinterval -c interval
+
+CLEANFILES += \
+	$(BTX_FI_GENERATED)
+
+BUILT_SOURCES += \
+	$(BTX_FI_GENERATED)
+
+nodist_libFIInterval_la_SOURCES = \
+	$(BTX_FI_GENERATED)
+
+libFIInterval_la_SOURCES = \
+	btx_fiinterval_callbacks.cpp
+
+libFIInterval_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I./ -I./btx_filter_fi
+libFIInterval_la_CFLAGS = -Wall -Wextra -Wno-unused-parameter $(WERROR) $(BABELTRACE2_CFLAGS)
+libFIInterval_la_CXXFLAGS = -std=c++17 -Wall -Wextra -Wno-unused-parameter $(WERROR) $(BABELTRACE2_CFLAGS)
+libFIInterval_la_LDFLAGS = $(BABELTRACE2_LIBS) -avoid-version -module
+
+bt2dir = $(pkglibdir)/bt2
+bt2_LTLIBRARIES = libFIInterval.la
+
+#
+# ----------------------------------------------------------------------
+# Enable pretty.print in babeltrace_thapi
+# ----------------------------------------------------------------------
+#
+# The library name uses the *directory* name (libfabric), not the short
+# backend name (fi) -- babeltrace_thapi.in:167 does the same for OpenCL,
+# requiring babeltrace_opencl_lib for backend key "cl".
+#
+
+babeltrace_libfabric_lib.rb: $(srcdir)/gen_babeltrace_libfabric_lib.rb $(top_srcdir)/utils/gen_babeltrace_lib_helper.rb $(top_srcdir)/utils/gen_probe_base.rb $(srcdir)/btx_libfabric_model.yaml
+	SRC_DIR=$(srcdir) $(RUBY) $< > $@
+
+EXTRA_DIST += \
+	gen_babeltrace_libfabric_lib.rb
+
+CLEANFILES += \
+	babeltrace_libfabric_lib.rb
+
+data_DATA = \
+	babeltrace_libfabric_lib.rb