[dpdk-dev] [PATCH v4 2/4] net/mlx4: spawn rdma-core dependency plug-in

[Adrien Mazarguil]

When mlx4 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 libmlx4 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 at 6wind.com>
---
 config/common_base        |  1 +
 doc/guides/nics/mlx4.rst  | 13 +++++++++++
 drivers/net/mlx4/Makefile | 29 +++++++++++++++++++++++++
 drivers/net/mlx4/mlx4.c   | 49 ++++++++++++++++++++++++++++++++++++++++++
 mk/rte.app.mk             |  4 ++++
 5 files changed, 96 insertions(+)

diff --git a/config/common_base b/config/common_base
index 170a38939..6df20be0e 100644
--- a/config/common_base
+++ b/config/common_base
@@ -298,6 +298,7 @@ CONFIG_RTE_LIBRTE_AVF_16BYTE_RX_DESC=n
 #
 CONFIG_RTE_LIBRTE_MLX4_PMD=n
 CONFIG_RTE_LIBRTE_MLX4_DEBUG=n
+CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS=n
 CONFIG_RTE_LIBRTE_MLX4_TX_MP_CACHE=8
 
 #
diff --git a/doc/guides/nics/mlx4.rst b/doc/guides/nics/mlx4.rst
index cab45dfe8..88161781c 100644
--- a/doc/guides/nics/mlx4.rst
+++ b/doc/guides/nics/mlx4.rst
@@ -86,6 +86,19 @@ These options can be modified in the ``.config`` file.
 
   Toggle compilation of librte_pmd_mlx4 itself.
 
+- ``CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS`` (default **n**)
+
+  Build PMD with additional code to make it loadable without hard
+  dependencies on **libibverbs** nor **libmlx4**, 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_MLX4_DEBUG`` (default **n**)
 
   Toggle debugging code and stricter compilation flags. Enabling this option
diff --git a/drivers/net/mlx4/Makefile b/drivers/net/mlx4/Makefile
index 1d33c38ed..c004ac71c 100644
--- a/drivers/net/mlx4/Makefile
+++ b/drivers/net/mlx4/Makefile
@@ -33,12 +33,15 @@ include $(RTE_SDK)/mk/rte.vars.mk
 
 # Library name.
 LIB = librte_pmd_mlx4.a
+LIB_GLUE = librte_pmd_mlx4_glue.so
 
 # Sources.
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_ethdev.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_flow.c
+ifneq ($(CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS),y)
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_glue.c
+endif
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_intr.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_mr.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_rxq.c
@@ -46,6 +49,10 @@ SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_rxtx.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_txq.c
 SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD) += mlx4_utils.c
 
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS),y)
+INSTALL-$(CONFIG_RTE_LIBRTE_MLX4_PMD)-lib += $(LIB_GLUE)
+endif
+
 # Basic CFLAGS.
 CFLAGS += -O3
 CFLAGS += -std=c11 -Wall -Wextra
@@ -55,7 +62,13 @@ CFLAGS += -D_BSD_SOURCE
 CFLAGS += -D_DEFAULT_SOURCE
 CFLAGS += -D_XOPEN_SOURCE=600
 CFLAGS += $(WERROR_FLAGS)
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS),y)
+CFLAGS += -DMLX4_GLUE='"$(LIB_GLUE)"'
+CFLAGS_mlx4_glue.o += -fPIC
+LDLIBS += -ldl
+else
 LDLIBS += -libverbs -lmlx4
+endif
 LDLIBS += -lrte_eal -lrte_mbuf -lrte_mempool -lrte_ring
 LDLIBS += -lrte_ethdev -lrte_net -lrte_kvargs
 LDLIBS += -lrte_bus_pci
@@ -109,7 +122,23 @@ mlx4_autoconf.h: mlx4_autoconf.h.new
 
 $(SRCS-$(CONFIG_RTE_LIBRTE_MLX4_PMD):.c=.o): mlx4_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 ($(CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS),y)
+
+$(LIB): $(LIB_GLUE)
+
+$(LIB_GLUE): mlx4_glue.o
+	$Q $(LD) $(LDFLAGS) $(EXTRA_LDFLAGS) \
+		-s -shared -o $@ $< -libverbs -lmlx4
+
+mlx4_glue.o: mlx4_autoconf.h
+
+endif
+
 clean_mlx4: FORCE
 	$Q rm -f -- mlx4_autoconf.h mlx4_autoconf.h.new
+	$Q rm -f -- mlx4_glue.o $(LIB_GLUE)
 
 clean: clean_mlx4
diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
index f304359bb..12e6b431c 100644
--- a/drivers/net/mlx4/mlx4.c
+++ b/drivers/net/mlx4/mlx4.c
@@ -37,6 +37,7 @@
  */
 
 #include <assert.h>
+#include <dlfcn.h>
 #include <errno.h>
 #include <inttypes.h>
 #include <stddef.h>
@@ -44,6 +45,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <unistd.h>
 
 /* Verbs headers do not support -pedantic. */
 #ifdef PEDANTIC
@@ -55,6 +57,7 @@
 #endif
 
 #include <rte_common.h>
+#include <rte_config.h>
 #include <rte_dev.h>
 #include <rte_errno.h>
 #include <rte_ethdev_driver.h>
@@ -724,6 +727,47 @@ static struct rte_pci_driver mlx4_driver = {
 		     RTE_PCI_DRV_INTR_RMV,
 };
 
+#ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
+
+/**
+ * Initialization routine for run-time dependency on rdma-core.
+ */
+static int
+mlx4_glue_init(void)
+{
+	void *handle = NULL;
+	void **sym;
+	const char *dlmsg;
+
+	handle = dlopen(MLX4_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, "mlx4_glue");
+	if (!sym || !*sym) {
+		rte_errno = EINVAL;
+		dlmsg = dlerror();
+		if (dlmsg)
+			ERROR("cannot resolve glue symbol: %s", dlmsg);
+		goto glue_error;
+	}
+	mlx4_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,"
+	      " libmlx4)");
+	return -rte_errno;
+}
+
+#endif
+
 /**
  * Driver initialization routine.
  */
@@ -744,6 +788,11 @@ rte_mlx4_pmd_init(void)
 	 * using this PMD, which is not supported in forked processes.
 	 */
 	setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
+#ifdef RTE_LIBRTE_MLX4_DLOPEN_DEPS
+	if (mlx4_glue_init())
+		return;
+	assert(mlx4_glue);
+#endif
 	mlx4_glue->fork_init();
 	rte_pci_register(&mlx4_driver);
 }
diff --git a/mk/rte.app.mk b/mk/rte.app.mk
index 0169f3f5b..fdbe36630 100644
--- a/mk/rte.app.mk
+++ b/mk/rte.app.mk
@@ -143,7 +143,11 @@ ifeq ($(CONFIG_RTE_LIBRTE_KNI),y)
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_KNI)        += -lrte_pmd_kni
 endif
 _LDLIBS-$(CONFIG_RTE_LIBRTE_LIO_PMD)        += -lrte_pmd_lio
+ifeq ($(CONFIG_RTE_LIBRTE_MLX4_DLOPEN_DEPS),y)
+_LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -ldl
+else
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX4_PMD)       += -lrte_pmd_mlx4 -libverbs -lmlx4
+endif
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MLX5_PMD)       += -lrte_pmd_mlx5 -libverbs -lmlx5
 _LDLIBS-$(CONFIG_RTE_LIBRTE_MRVL_PMD)       += -lrte_pmd_mrvl -L$(LIBMUSDK_PATH)/lib -lmusdk
 _LDLIBS-$(CONFIG_RTE_LIBRTE_NFP_PMD)        += -lrte_pmd_nfp
-- 
2.11.0