Skip to contentSkip to conversation

eal: fix detection of static or shared DPDK builds

PATCH Applied Bruce Richardson 3 participants 4 messages Feb 8, 2021 – Feb 10, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 eal: fix detection of static or shared DPDK builds

When checking the loading of EAL shared lib to see if we have a shared DPDK build, we only want to include part of the ABI version in the check rather than the whole thing. For example, with ABI version 21.1 for DPDK release 21.02, the linker links the binary against librte_eal.so.21, without the ".1".

To avoid any further brittleness in this area, we can check for multiple versions when doing the check, since just about any version of EAL implies a shared build. Therefore we check for presence of librte_eal.so with full ABI_VERSION extension, and then repeatedly remove the end part of the filename after the last dot, checking each time. For example (debug log output for static build):

EAL: Checking presence of .so 'librte_eal.so.21.1'
EAL: Checking presence of .so 'librte_eal.so.21'
EAL: Checking presence of .so 'librte_eal.so'
EAL: Detected static linkage of DPDK

Fixes 7781950f4d38 ("eal: fix shared lib mode detection")
---
 lib/librte_eal/common/eal_common_options.c | 35 +++++++++++++++++++++-
 1 file changed, 34 insertions(+), 1 deletion(-)
lib/librte_eal/common/eal_common_options.c +34 -1
index 6b3707725f..94029bf7f1 100644
@@ -494,6 +494,39 @@ eal_dlopen(const char *pathname)
 	return retval;
 }
 
+static int
+is_shared_build(void)
+{
+#define EAL_SO "librte_eal.so"
+	char soname[32];
+	size_t len, minlen = strlen(EAL_SO);
+
+	len = strlcpy(soname, EAL_SO"."ABI_VERSION, sizeof(soname));
+	if (len > sizeof(soname)) {
+		RTE_LOG(ERR, EAL, "Shared lib name too long in shared build check\n");
+		len = sizeof(soname) - 1;
+	}
+
+	while (len >= minlen) {
+		/* check if we have this .so loaded, if so - shared build */
+		RTE_LOG(DEBUG, EAL, "Checking presence of .so '%s'\n", soname);
+		if (dlopen(soname, RTLD_LAZY | RTLD_NOLOAD) != NULL) {
+			RTE_LOG(INFO, EAL, "Detected shared linkage of DPDK\n");
+			return 1;
+		}
+
+		/* remove any version numbers off the end to retry */
+		while (len-- > 0)
+			if (soname[len] == '.') {
+				soname[len] = '\0';
+				break;
+			}
+	}
+
+	RTE_LOG(INFO, EAL, "Detected static linkage of DPDK\n");
+	return 0;
+}
+
 int
 eal_plugins_init(void)
 {
@@ -505,7 +538,7 @@ eal_plugins_init(void)
 	 * (Using dlopen with NOLOAD flag on EAL, will return NULL if the EAL
 	 * shared library is not already loaded i.e. it's statically linked.)
 	 */
-	if (dlopen("librte_eal.so."ABI_VERSION, RTLD_LAZY | RTLD_NOLOAD) != NULL &&
+	if (is_shared_build() &&
 			*default_solib_dir != '\0' &&
 			stat(default_solib_dir, &sb) == 0 &&
 			S_ISDIR(sb.st_mode))

--
2.27.0

#1 lore raw
Quoted 18 lines · Bruce Richardson

When checking the loading of EAL shared lib to see if we have a shared DPDK build, we only want to include part of the ABI version in the check rather than the whole thing. For example, with ABI version 21.1 for DPDK release 21.02, the linker links the binary against librte_eal.so.21, without the ".1".

To avoid any further brittleness in this area, we can check for multiple versions when doing the check, since just about any version of EAL implies a shared build. Therefore we check for presence of librte_eal.so with full ABI_VERSION extension, and then repeatedly remove the end part of the filename after the last dot, checking each time. For example (debug log output for static build):

EAL: Checking presence of .so 'librte_eal.so.21.1'
EAL: Checking presence of .so 'librte_eal.so.21'
EAL: Checking presence of .so 'librte_eal.so'
EAL: Detected static linkage of DPDK

Quoted 6 lines
Fixes 7781950f4d38 ("eal: fix shared lib mode detection")

---

I saw this issue with OVS, where I was getting weird failures about ports not being bound (in case of physical ports) or not being created (in case of virtio ports), when using a shared build. Since it's potentially serious, I'd appreciate if someone can reproduce the issue and verify the fix so we can consider it for 21.02 inclusion.

To demonstrate this with regular DPDK, do a usual build of DPDK and then do "ninja install" to install system-wide. Then build an example app, e.g. l2fwd, using "make" from the examples/l2fwd directory. Running the example normally, e.g. ./build/l2fwd -c F00, leads to no drivers being loaded or ports being found. Adding "-d /path/to/drivers" e.g. "/usr/local/lib/x86_64-linux-gnu/dpdk/pmds-21.1" on my system works as expected. This shows the driver loading is not correct.

After applying this patch and re-running "ninja install", l2fwd should run the same with and without the "-d" flag.

/Bruce

Tested-by
#2 lore raw

Hi Bruce,

Thanks for the fix.
I do see the issue mentioned when using DPDK shared libs with OVS and this patch fixes it.

However, I saw the issue only for system installed DPDK but not for directory installed DPDK.

Quoted 14 lines

I saw this issue with OVS, where I was getting weird failures about ports not being bound (in case of physical ports) or not being created (in case of virtio ports), when using a shared build. Since it's potentially serious, I'd appreciate if someone can reproduce the issue and verify the fix so we can consider it for 21.02 inclusion.

To demonstrate this with regular DPDK, do a usual build of DPDK and then do "ninja install" to install system-wide. Then build an example app, e.g. l2fwd, using "make" from the examples/l2fwd directory. Running the example normally, e.g. ./build/l2fwd -c F00, leads to no drivers being loaded or ports being found. Adding "-d /path/to/drivers" e.g. "/usr/local/lib/x86_64-linux-gnu/dpdk/pmds-21.1" on my system works as expected. This shows the driver loading is not correct.

Quoted 4 lines

After applying this patch and re-running "ninja install", l2fwd should run the same with and without the "-d" flag.

/Bruce

#3 lore raw

09/02/2021 21:10, Pai G, Sunil:

Quoted 22 lines

Hi Bruce,

Thanks for the fix.
I do see the issue mentioned when using DPDK shared libs with OVS and this patch fixes it.

However, I saw the issue only for system installed DPDK but not for directory installed DPDK.

Quoted 14 lines

I saw this issue with OVS, where I was getting weird failures about ports not being bound (in case of physical ports) or not being created (in case of virtio ports), when using a shared build. Since it's potentially serious, I'd appreciate if someone can reproduce the issue and verify the fix so we can consider it for 21.02 inclusion.

To demonstrate this with regular DPDK, do a usual build of DPDK and then do "ninja install" to install system-wide. Then build an example app, e.g. l2fwd, using "make" from the examples/l2fwd directory. Running the example normally, e.g. ./build/l2fwd -c F00, leads to no drivers being loaded or ports being found. Adding "-d /path/to/drivers" e.g. "/usr/local/lib/x86_64-linux-gnu/dpdk/pmds-21.1" on my system works as expected. This shows the driver loading is not correct.

Quoted 6 lines
Quoted 4 lines

After applying this patch and re-running "ninja install", l2fwd should run the same with and without the "-d" flag.

/Bruce

Applied, thanks