Re: [dpdk-dev] [PATCH] eal/freebsd: lock memory device to prevent conflicts

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

On Mon, Sep 13, 2021 at 02:14:55PM +0100, Burakov, Anatoly wrote:
> On 13-Sep-21 12:06 PM, Bruce Richardson wrote:
> > Only a single DPDK process on the system can be using the /dev/contigmem
> > mappings at a time, but this was never explicitly enforced, e.g. when
> > using --in-memory flag on two processes. To prevent possible conflict
> > issues, we lock the dev node when it's in use, preventing other DPDK
> > processes from starting up and causing problems for us.
> > 
> > Fixes: 764bf26873b9 ("add FreeBSD support")
> > Cc: stable@dpdk.org
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> >   lib/eal/freebsd/eal_hugepage_info.c | 4 ++++
> >   1 file changed, 4 insertions(+)
> > 
> > diff --git a/lib/eal/freebsd/eal_hugepage_info.c b/lib/eal/freebsd/eal_hugepage_info.c
> > index 408f054f7a..4a8d87c23e 100644
> > --- a/lib/eal/freebsd/eal_hugepage_info.c
> > +++ b/lib/eal/freebsd/eal_hugepage_info.c
> > @@ -90,6 +90,10 @@ eal_hugepage_info_init(void)
> >   		RTE_LOG(ERR, EAL, "could not open "CONTIGMEM_DEV"\n");
> >   		return -1;
> >   	}
> > +	if (flock(fd, LOCK_EX) < 0) {
> > +		RTE_LOG(ERR, EAL, "could not lock memory. Is another DPDK process running?\n");
> > +		return -1;
> > +	}
> >   	if (buffer_size >= 1<<30)
> >   		RTE_LOG(INFO, EAL, "Contigmem driver has %d buffers, each of size %dGB\n",
> > 
> 
> This only gets triggered when regular init path is chosen, i.e. --no-huge
> still works.

Yes, but that is ok, I think, since no-huge doesn't use these resources or
suffer from this problem. On the other hand, except for running unit tests,
no-huge mode is pretty useless on FreeBSD as we don't have any
vfio-equivalent support, so all HW access has to use physical addresses
which can only be got using contigmem.

> I'm a bit uneasy with --in-memory mode pretending to work on
> FreeBSD and Windows, but that's a separate problem :)

Yes, it is, though one that does belong is the same area as this one. The
"fix" is probably to just print a warning when --in-memory is used,
informing the user that the flag is ignored and then continue.
Alternatively we can error out, but I think the warn+continue is better,
myself.

> As far as the patch
> goes, the problem it addresses does get fixed.
> 
> Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
> 
Thanks.

/Bruce