eal: fix querying DPDK version at runtime

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}

05/02/2021 20:39, Bruce Richardson:

Quoted 8 lines

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

Quoted 6 lines

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}

Nice demonstration.

lib/librte_telemetry/rte_telemetry.h rte_telemetry_register_cmd(const char *cmd, telemetry_cb fn, const char *help); · line 315

Quoted 5 lines

__rte_experimental
int

lib/librte_telemetry/rte_telemetry.h +1 -1

-rte_telemetry_init(const char *runtime_dir, rte_cpuset_t *cpuset,
+rte_telemetry_init(const char *runtime_dir, const char *rte_version, rte_cpuset_t *cpuset,
 		const char **err_str);

It is changing the API.
As it is experimental, you just need to mention it in the release notes.

It is the fix. Do you think it should be merged quickly? Or wait for 21.05?

Quoted 18 lines

05/02/2021 20:39, Bruce Richardson:

Quoted 8 lines

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

Quoted 6 lines

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}

Nice demonstration.

Quoted 8 lines

lib/librte_telemetry/rte_telemetry.h rte_telemetry_register_cmd(const char *cmd, telemetry_cb fn, const char *help); · line 315

Quoted 5 lines

__rte_experimental
int

lib/librte_telemetry/rte_telemetry.h +1 -1

-rte_telemetry_init(const char *runtime_dir, rte_cpuset_t *cpuset,
+rte_telemetry_init(const char *runtime_dir, const char *rte_version, rte_cpuset_t *cpuset,
 		const char **err_str);

It is changing the API.
As it is experimental, you just need to mention it in the release notes.

I don't think I actually need to mention it there, because this API is more "INTERNAL" than "EXPERIMENTAL". It's called automatically from rte_eal_init(). I've done up patch http://patches.dpdk.org/patch/87806/
to correct this, including a RN addition. That should remove the need for a doc update for this patch.

Quoted 4 lines

It is the fix. Do you think it should be merged quickly? Or wait for 21.05?

I'm not sure either way to be honest. Given the bug has been around so long, it's not exactly urgent. On the other hand, to get the fix the user needs to rebuild their app, so having it sooner is nicer, and will mean it would make the next LTS point release. Overall, though, I'm fine whichever you decide.

05/02/2021 22:26, Bruce Richardson:

Quoted 39 lines · Thomas Monjalon

Quoted 18 lines

05/02/2021 20:39, Bruce Richardson:

Quoted 8 lines

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

Quoted 6 lines

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}

Nice demonstration.

Quoted 8 lines

lib/librte_telemetry/rte_telemetry.h rte_telemetry_register_cmd(const char *cmd, telemetry_cb fn, const char *help); · line 315

Quoted 5 lines

__rte_experimental
int

lib/librte_telemetry/rte_telemetry.h +1 -1

-rte_telemetry_init(const char *runtime_dir, rte_cpuset_t *cpuset,
+rte_telemetry_init(const char *runtime_dir, const char *rte_version, rte_cpuset_t *cpuset,
 		const char **err_str);

It is changing the API.
As it is experimental, you just need to mention it in the release notes.

I don't think I actually need to mention it there, because this API is more "INTERNAL" than "EXPERIMENTAL". It's called automatically from rte_eal_init(). I've done up patch http://patches.dpdk.org/patch/87806/
to correct this, including a RN addition. That should remove the need for a doc update for this patch.

Quoted 4 lines

It is the fix. Do you think it should be merged quickly? Or wait for 21.05?

Quoted 5 lines

I'm not sure either way to be honest. Given the bug has been around so long, it's not exactly urgent. On the other hand, to get the fix the user needs to rebuild their app, so having it sooner is nicer, and will mean it would make the next LTS point release. Overall, though, I'm fine whichever you decide.

There is not much help available to close 21.02, so I won't take any risk. I'll merge this fix in 21.05.

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}
  -->

To fix this, we need to convert the rte_version() function, and any other necessary parts of the rte_version.h, to be actual functions in EAL, not just inlines/macros. The only complication in doing so is that telemetry library cannot call rte_version() directly, and instead needs the version string passed in on init.

Fixes af75078fece3 ("first public release")

Cc [email protected]

Signed-off-by Bruce Richardson <[email protected]>

---
v2:
rebased on top of main for 21.05 consideration.

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}
  -->

To fix this, we need to convert the rte_version() function, and any other necessary parts of the rte_version.h, to be actual functions in EAL, not just inlines/macros. The only complication in doing so is that telemetry library cannot call rte_version() directly, and instead needs the version string passed in on init.

Fixes af75078fece3 ("first public release")

Cc [email protected]

lib/librte_eal/include/rte_version.h +0 -0

index f7a3a1ebc..2f3f727b4 100644
@@ -28,38 +28,47 @@ extern "C" {
  * All version numbers in one to compare with RTE_VERSION_NUM()

Quoted 40 lines · Bruce Richardson

For using a DPDK application, such as OVS, which is dynamically linked, the DPDK version in use should always report the actual version, not the version used at build time. This incorrect behaviour can be seen by building OVS against one version of DPDK and running it against a later one. Using "ovs-vsctl list Open_vSwitch" to query basic info, the dpdk_version returned will be the build version not the currently running one - which can be verified using the DPDK telemetry library client.

  $ sudo ovs-vsctl list Open_vSwitch | grep dpdk_version
  dpdk_version        : "DPDK 20.11.0-rc4"

  $ echo quit | sudo dpdk-telemetry.py
  Connecting to /var/run/dpdk/rte/dpdk_telemetry.v2
  {"version": "DPDK 21.02.0-rc2", "pid": 405659, "max_output_len": 16384}
  -->

To fix this, we need to convert the rte_version() function, and any other necessary parts of the rte_version.h, to be actual functions in EAL, not just inlines/macros. The only complication in doing so is that telemetry library cannot call rte_version() directly, and instead needs the version string passed in on init.

Fixes af75078fece3 ("first public release")

Cc [email protected]

lib/librte_eal/include/rte_version.h +1 -4

index f7a3a1ebc..2f3f727b4 100644
@@ -28,38 +28,47 @@ extern "C" {
  * All version numbers in one to compare with RTE_VERSION_NUM()
  */
 #define RTE_VERSION RTE_VERSION_NUM( \
-                       RTE_VER_YEAR, \
-                       RTE_VER_MONTH, \
-                       RTE_VER_MINOR, \
-                       RTE_VER_RELEASE)
+                       rte_version_year(), \

lib/librte_eal/include/rte_version.h extern "C" { · line 34

Quoted 3 lines

lib/librte_eal/include/rte_version.h +3 -0

+                       rte_version_month(), \
+                       rte_version_minor(), \
+                       rte_version_release())

It breaks SPDK and applications relying on RTE_VERSION in preprocessor directives.
RTE_VERSION* macros should be left alone, and applications that need the runtime value should call rte_version().

See logs in https://lab.dpdk.org/results/dashboard/results/results-uploads/test_runs/2f636aaf4ce244eba20844f9ff006033/log_upload_file/2021/3/dpdk_6857cb635821_2021-03-17_06-34-34_NA.zip