Re: [dpdk-dev] [PATCH 3/3] app/test/meson: auto detect number of cores

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, Apr 12, 2019 at 09:46:17AM +0200, David Marchand wrote:
>    On Thu, Apr 11, 2019 at 9:52 PM Aaron Conole <[1]aconole@redhat.com>
>    wrote:
> 
>      The arguments being passed will cause failures on laptops that have,
>      for instance, 2 cores only.  Most of the tests don't require more
>      than a single core.  Some require multiple cores (but those tests
>      should be modified to 'SKIP' when the correct number of cores
>      aren't available).
>      The unit test results shouldn't be impacted by this change, but it
>      allows for a future enhancement to pass flags such as '--no-huge'.
>      Also include a fix to a reported issue with running on FreeBSD.
>      Signed-off-by: Aaron Conole <[2]aconole@redhat.com>
>      ---
>      Conflicts with [3]http://patches.dpdk.org/patch/50850/
>       app/test/meson.build | 24 +++++++++++++++++++++---
>       1 file changed, 21 insertions(+), 3 deletions(-)
>      diff --git a/app/test/meson.build b/app/test/meson.build
>      index 867cc5863..1010bfbc8 100644
>      --- a/app/test/meson.build
>      +++ b/app/test/meson.build
>      @@ -344,17 +344,32 @@ if get_option('tests')
>              timeout_seconds = 600
>              timeout_seconds_fast = 10
>      +       # Retreive the number of CPU cores
> 
>    nit: Retrieve
>    Little concern here on the approach of getting the max available cpu
>    index.
>    If we have non contiguous cpus (let's say hyper threading is disabled),
>    this won't work.
>    But we can just assume this won't happen for non regression setups
>    (vms).
> 
>      +       num_cores = run_command('lscpu',
>      '-p=cpu').stdout().strip().split('\n')[-1]
> 
>    lscpu is a linux command afaik.
>    Maybe for FreeBSD:
>    root@freebsd-10:~ # sysctl dev.cpu |cut -d . -f 3 |sort |tail -1
>    3
>    Not sure if FreeBSD ensures that the keys names/objects won't change
>    accross the versions.
> 

Very similar thoughs/concerns here. I think doing this per OS is probably
safer in the long term - even though lscpu is available for FreeBSD as an
extra package.

My suggestion: for FreeBSD, get the number of CPUs using
	"/sbin/sysctl -n hw.ncpu"
For Linux, rather than using a "0-N" range, use output from 
"cat /sys/devices/system/cpu/present", which will ensure that it works
even in the non-contiguous CPU numbering case. [Though I suspect we get
non-contiguous numbers only in the case a core or two has been hotplugged
out, I think - disabling hyperthreading won't leave gaps, just fewer cores]

/Bruce