[dpdk-dev] [PATCH v3 4/4] net/mlx5: spawn rdma-core dependency plug-in

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

When mlx5 is not compiled directly as an independent shared object (e.g.
CONFIG_RTE_BUILD_SHARED_LIB not enabled for performance reasons), DPDK
applications inherit its dependencies on libibverbs and libmlx5 through
rte.app.mk.

This is an issue both when DPDK is delivered as a binary package (Linux
distributions) and for end users because rdma-core then propagates as a
mandatory dependency for everything.

Application writers relying on binary DPDK packages are not necessarily
aware of this fact and may end up delivering packages with broken
dependencies.

This patch therefore introduces an intermediate internal plug-in
hard-linked with rdma-core (to preserve symbol versioning) loaded by the
PMD through dlopen(), so that a missing rdma-core does not cause unresolved
symbols, allowing applications to start normally.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
---
 config/common_base        |  1 +
 doc/guides/nics/mlx5.rst  | 13 ++++++++++++
 drivers/net/mlx5/Makefile | 31 ++++++++++++++++++++++++++-
 drivers/net/mlx5/mlx5.c   | 48 ++++++++++++++++++++++++++++++++++++++++++
 mk/rte.app.mk             |  4 ++++
 5 files changed, 96 insertions(+), 1 deletion(-)

diff --git a/config/common_base b/config/common_base
index 6df20be0e..b9b421db7 100644
--- a/config/common_base
+++ b/config/common_base
@@ -306,6 +306,7 @@ CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE=8
 #
 CONFIG_RTE_LIBRTE_MLX5_PMD=n
 CONFIG_RTE_LIBRTE_MLX5_DEBUG=n
+CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS=n
 CONFIG_RTE_LIBRTE_MLX5_TX_MP_CACHE=8
 
 #
diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
index 2e6d1e45a..a9e4bf51a 100644
--- a/doc/guides/nics/mlx5.rst
+++ b/doc/guides/nics/mlx5.rst
@@ -159,6 +159,19 @@ These options can be modified in the ``.config`` file.
 
   Toggle compilation of librte_pmd_mlx5 itself.
 
+- ``CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS`` (default **n**)
+
+  Build PMD with additional code to make it loadable without hard
+  dependencies on **libibverbs** nor **libmlx5**, which may not be installed
+  on the target system.
+
+  In this mode, their presence is still required for it to run properly,
+  however their absence won't prevent a DPDK application from starting (with
+  ``CONFIG_RTE_BUILD_SHARED_LIB`` disabled) and they won't show up as
+  missing with ``ldd(1)``.
+
+  This option has no performance impact.
+
 - ``CONFIG_RTE_LIBRTE_MLX5_DEBUG`` (default **n**)
 
   Toggle debugging code and stricter compilation flags. Enabling this option
diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
index bdec30692..4b20d718b 100644
--- a/drivers/net/mlx5/Makefile
+++ b/drivers/net/mlx5/Makefile
@@ -33,9 +33,13 @@ include $(RTE_SDK)/mk/rte.vars.mk
 
 # Library name.
 LIB = librte_pmd_mlx5.a
+LIB_GLUE = librte_pmd_mlx5_glue.so
 
 # Sources.
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5.c
+ifneq ($(CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS),y)
+SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_glue.c
+endif
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_txq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rxtx.c
@@ -53,7 +57,10 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_rss.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_mr.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_flow.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_socket.c
-SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5_glue.c
+
+ifeq ($(CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS),y)
+INSTALL-$(CONFIG_RTE_LIBRTE_MLX5_PMD)-lib += $(LIB_GLUE)
+endif
 
 # Basic CFLAGS.
 CFLAGS += -O3
@@ -65,7 +72,13 @@ CFLAGS += -D_DEFAULT_SOURCE
 CFLAGS += -D_XOPEN_SOURCE=600
 CFLAGS += $(WERROR_FLAGS)
 CFLAGS += -Wno-strict-prototypes
+ifeq ($(CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS),y)
+CFLAGS += -DMLX5_GLUE='"$(LIB_GLUE)"'
+CFLAGS_mlx5_glue.o += -fPIC
+LDLIBS += -ldl
+else
 LDLIBS += -libverbs -lmlx5
+endif
 LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring
 LDLIBS += -lrte_ethdev -lrte_net -lrte_kvargs
 LDLIBS += -lrte_bus_pci
@@ -158,7 +171,23 @@ mlx5_autoconf.h: mlx5_autoconf.h.new
 
 $(SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD):.c=.o): mlx5_autoconf.h
 
+# Generate dependency plug-in for rdma-core when the PMD must not be linked
+# directly, so that applications do not inherit this dependency.
+
+ifeq ($(CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS),y)
+
+$(LIB): $(LIB_GLUE)
+
+$(LIB_GLUE): mlx5_glue.o
+	$Q $(LD) $(LDFLAGS) $(EXTRA_LDFLAGS) \
+		-s -shared -o $@ $< -libverbs -lmlx5
+
+mlx5_glue.o: mlx5_autoconf.h
+
+endif
+
 clean_mlx5: FORCE
 	$Q rm -f -- mlx5_autoconf.h mlx5_autoconf.h.new
+	$Q rm -f -- mlx5_glue.o $(LIB_GLUE)
 
 clean: clean_mlx5
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index b911d0ad4..466a8b0a2 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -35,6 +35,7 @@
 #include <unistd.h>
 #include <string.h>
 #include <assert.h>
+#include <dlfcn.h>
 #include <stdint.h>
 #include <stdlib.h>
 #include <errno.h>
@@ -57,6 +58,7 @@
 #include <rte_pci.h>
 #include <rte_bus_pci.h>
 #include <rte_common.h>
+#include <rte_config.h>
 #include <rte_eal_memconfig.h>
 #include <rte_kvargs.h>
 
@@ -1075,6 +1077,47 @@ static struct rte_pci_driver mlx5_driver = {
 	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
 };
 
+#ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
+
+/**
+ * Initialization routine for run-time dependency on rdma-core.
+ */
+static int
+mlx5_glue_init(void)
+{
+	void *handle = NULL;
+	void **sym;
+	const char *dlmsg;
+
+	handle = dlopen(MLX5_GLUE, RTLD_LAZY);
+	if (!handle) {
+		rte_errno = EINVAL;
+		dlmsg = dlerror();
+		if (dlmsg)
+			ERROR("cannot load glue library: %s", dlmsg);
+		goto glue_error;
+	}
+	sym = dlsym(handle, "mlx5_glue");
+	if (!sym || !*sym) {
+		rte_errno = EINVAL;
+		dlmsg = dlerror();
+		if (dlmsg)
+			ERROR("cannot resolve glue symbol: %s", dlmsg);
+		goto glue_error;
+	}
+	mlx5_glue = *sym;
+	return 0;
+glue_error:
+	if (handle)
+		dlclose(handle);
+	ERROR("cannot initialize PMD due to missing run-time"
+	      " dependency on rdma-core libraries (libibverbs,"
+	      " libmlx5)");
+	return -rte_errno;
+}
+
+#endif
+
 /**
  * Driver initialization routine.
  */
@@ -1094,6 +1137,11 @@ rte_mlx5_pmd_init(void)
 	/* Match the size of Rx completion entry to the size of a cacheline. */
 	if (RTE_CACHE_LINE_SIZE == 128)
 		setenv("MLX5_CQE_SIZE", "128", 0);
+#ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
+	if (mlx5_glue_init())
+		return;
+	assert(mlx5_glue);
+#endif
 	mlx5_glue->fork_init();
 	rte_pci_register(&mlx5_driver);
 }
diff --git a/mk/rte.app.mk b/mk/rte.app.mk
index fdbe36630..88f626049 100644
--- a/mk/rte.app.mk
+++ b/mk/rte.app.mk
@@ -148,7 +148,11 @@ _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -ldl
 else
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -libverbs -lmlx4
 endif
+ifeq ($(CONFIG_RTE_LIBRTE_MLX5_DLOPEN_DEPS),y)
+_LDLIBS-$(CONFIG_RTE_LIBRTE_MLX5_PMD)       += -lrte_pmd_mlx5 -ldl
+else
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX5_PMD)       += -lrte_pmd_mlx5 -libverbs -lmlx5
+endif
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MRVL_PMD)       += -lrte_pmd_mrvl -L$(LIBMUSDK_PATH)/lib -lmusdk
 _LDLIBS-$(CONFIG_RTE_LIBRTE_NFP_PMD)        += -lrte_pmd_nfp
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_NULL)       += -lrte_pmd_null
-- 
2.11.0