On Thu, Jun 18, 2015 at 11:55:42AM +0200, Vlastimil Babka wrote: > >>>If there are many mirror regions in one node, then it will be many holes in the > >>>normal zone, is this fine? > >> > >>Yeah, it doesn't matter how many holes there are. > > > >So mirror zone and normal zone will span each other, right? > > > >e.g. node 1: 4G-8G(normal), 8-12G(mirror), 12-16G(normal), 16-24G(mirror), 24-28G(normal) ... > >normal: start=4G, size=28-4=24G, > >mirror: start=8G, size=24-8=16G, > > Yes, that works. It's somewhat unfortunate wrt performance that the hardware > does it like this though. With current Xeon h/w you can have one mirrored range per memory controller ... and there are two memory controllers on a cpu socket, so two mirrored ranges per node. So a map might look like: SKT0: MC0: 0-2G Mirrored (but we may want to ignore mirror here to keep it for ZONE_DMA) SKT0: MC0: 2G-4G No memory ... I/O mapping area SKT0: MC0: 4G-34G Not mirrored SKT0: MC1: 34G-40G Mirrored SKT0: MC1: 40G-66G Not mirrored SKT1: MC0: 66G-70G Mirror SKT1: MC0: 70G-98G Not Mirrored SKT1: MC1: 98G-102G Mirror SKT1: MC1: 102G-130G Not Mirrored ... and so on. > >I think zone is defined according to the special address range, like 16M(DMA), 4G(DMA32), > > Traditionally yes. But then there is ZONE_MOVABLE, this year's LSF/MM we > discussed (and didn't outright deny) ZONE_CMA... > I'm not saying others will favour the new zone approach though, it's just my > opinion that it might be a better option than a new migratetype. If we are going to have lots of zones ... then perhaps we will need a fast way to look at a "struct page" and decide which zone it belongs to. Complicated math on the address deosn't sound ideal. If the complex zone model is just for 64-bit, are there enough bits available in page->flags (3 bits for 8 options ... which we are close to filling now ... 4 bits for future breathing room). > >and is it appropriate to add a new mirror zone with a volatile physical address? > > By "volatile" you mean what, that the example above would change > dynamically? That would be rather challenging... If we hot-add another cpu together with on die memory controllers connected to more memory ... then some of the new memory might be mirrored. Current h/w doesn't allow mirrored areas to grow/shrink (though if there are a lot of errors we may break a mirror so a whole range could lose the mirror attribute). -Tony -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>