On 02/22/2014 12:10 AM, Alexander Graf wrote: > > On 21.02.2014, at 14:04, Alexey Kardashevskiy <a...@ozlabs.ru> wrote: > >> On 02/22/2014 12:02 AM, Alexander Graf wrote: >>> >>> On 21.02.2014, at 13:56, Alexey Kardashevskiy <a...@ozlabs.ru> wrote: >>> >>>> On 02/21/2014 07:57 PM, Alexander Graf wrote: >>>>> >>>>> On 21.02.2014, at 05:57, Alexey Kardashevskiy <a...@ozlabs.ru> wrote: >>>>> >>>>>> On 02/10/2014 05:32 PM, Alexey Kardashevskiy wrote: >>>>>>> At the moment if the user asked for huge pages and there is no more huge >>>>>>> pages, QEMU prints warning and falls back to the anonymous memory >>>>>>> allocator which is quite easy not to notice. QEMU also does so even >>>>>>> if the user specified -mem-prealloc and it seems wrong as the user >>>>>>> specifically requested huge pages for the entire RAM but QEMU failed to >>>>>>> do >>>>>>> so and continued. On PPC64 this will produce a fragile guest as QEMU >>>>>>> tells the guest via device-tree that it can use huge pages when it >>>>>>> actually cannot. >>>>>>> >>>>>>> This adds message+exit if RAM cannot be preallocated from huge pages. >>>>>> >>>>>> >>>>>> Too bad? Should I increase my personal pinging timeout from 1 to 2 weeks >>>>>> to >>>>>> avoid annoying the community? :) Thanks! >>>>> >>>> >>>>> The patch changes the semantics of -mem-prealloc from "make sure all >>>>> RAM is mapped" to "make sure all RAM is mapped and is backed by huge >>>>> pages if we use huge pages" and thus is just plain wrong. >>>> >>>> ? I did actually expect it to alloc RAM from hugepages only. Otherwise >>>> there is no point in mem-prealloc. Yes, I am ignorant, I know. >>>> >>>>> The real question is why are we allowing sparsely mapped huge page >>>> backing at all? Should we change that? Do we need a new flag for this to >>>> specify "yes, I do want all my pages backed by -mem-path"? >>>> >>>> >>>> ? Add a switch to -mem-path saying "yes I really want -mem-path"? Sorry, I >>>> lost you here. -mem-path + -mem-prealloc - like this is not enough? Why >>>> would I specify -mem-path after all if I did not want RAM to backed by huge >>>> pages? >>> >> >>> I think it makes sense to disable any fallback for -mem-path, so that it >>> always only allocates RAM pages from the -mem-path pool. But this is a >>> big change from how it used to work before and thus needs to be properly >>> coordinated. >> >> ROMs, BARs - this all will stop working if I understand things right. And >> we (ozlabs) do not really want these things to be in huge pages. >
> Only if they're backed by virtual memory. And in that case why don't you > want them be huge pages? What qualifies a region to be huge vs > non-huge? This just adds complication for no reason. If we disable small pages with -mam-path, we'll have to teach SLOF and our PCI hotplug code to align BARs and for what? HV KVM does not need this to function. -- Alexey
