On Fri, Nov 07, 2014 at 10:16:06AM -0200, Eduardo Habkost wrote: > On Fri, Nov 07, 2014 at 12:21:26PM +0100, Andrew Jones wrote: > > On Fri, Nov 07, 2014 at 10:52:31AM +0100, Andrew Jones wrote: > > > On Fri, Nov 07, 2014 at 10:40:14AM +0100, Paolo Bonzini wrote: > > > > > > > > > > > > On 07/11/2014 10:29, Andrew Jones wrote: > > > > >> > I think this would cause too many failures in the wild. Perhaps > > > > >> > error > > > > >> > out if it is lower, and warn if sockets * cores * threads > > > > > >> > max_cpus > > > > >> > since we actually allow hot-plug a thread at a time? > > > > > We'd still have more failures if we choose to error out when it's > > > > > lower, > > > > > since we currently silently adjust threads in some of those cases, or > > > > > just don't care that the topology doesn't support up to maxcpus in > > > > > other. > > > > > > > > So I guess we need a decent fallback if it doesn't match. Something > > > > like (based also on the reply from Eduardo): > > > > > > > > 1) always warn if max_cpus % (cores*threads) != 0 || smp_cpus % > > > > (cores*threads) != 0 > > > > > > > > 2) if sockets*cpus*threads < max_cpus, adjust sockets to > > > > DIV_ROUND_UP(max_cpus, cores*threads). If we didn't warn in step 1, do > > > > it now. Give a different, less harsh warning if the cmdline > > > > sockets*cpus*threads did match smp_cpus. In the latter case, the user > > > > _almost_ knows what he was doing. > > > > > > > > Not perfect, but it could be something to start from. Adjusting sockets > > > > is better than adjusting threads. > > > > > > OK. I can whip up a v2 that is less harsh (more warnings, less aborts). > > > I'll also address the other issue I mentioned in the bottom of my reply > > > to Eduardo, which is to make sure we consider machine_class->max_cpus. > > > > After talking with Igor, it seems like the better approach is to get > > smp parameters converted over to machine properties, allowing us to > > use the old parser for old machine types, and the new for new. I'll > > look into it. > > While we work on that, I think we could at least change the existing > code to abort if sockets*cores*threads < smp_cpus (when all 4 options > are present in the command-line), and never adjust any option silently.
But then we'll end up with 3 different behaviors. If we do anything for the short-term, then maybe it should only be to add warnings. The machine property version will then convert those warnings to aborts for new machines types. drew
