Re: [Qemu-devel] [PATCH v2 2/2] RAM API: Make use of it for x86 PC

[Anthony Liguori]

On 11/18/2010 09:22 AM, Avi Kivity wrote:
> On 11/18/2010 01:42 AM, Anthony Liguori wrote:
> > > Gack.  For the benefit of those that want to join the fun without
> > > digging up the spec, these magic flippable segments the i440fx can
> > > toggle are 12 fixed 16k segments from 0xc0000 to 0xeffff and a single
> > > 64k segment from 0xf0000 to 0xfffff.  There are read-enable and
> > > write-enable bits for each, so the chipset can be configured to read
> > > from the bios and write to memory (to setup BIOS-RAM caching), and read
> > > from memory and write to the bios (to enable BIOS-RAM caching).  The
> > > other bit combinations are also available.
> >
> > Yup.  As Gleb mentions, there's the SDRAM register which controls 
> > whether 0xa0000 is mapped to PCI or whether it's mapped to RAM (but 
> > KVM explicitly disabled SMM support).
>
> KVM not supporting SMM is a bug (albeit one that is likely to remain 
> unresolved for a while).  Let's pretend that kvm smm support is not an 
> issue.
>
> IIUC, SMM means that there two memory maps when the cpu accesses 
> memory, one for SMM, one for non-SMM.

No.  That's not what it means.  With the i440fx, when the CPU accesses 
0xa0000, it gets forwarded to the PCI bus no different than an access to 
0xe0000.

If the CPU asserts the EXF4#/Ab7# signal, then the i440fx directs CPU 
accesses to 0xa0000 to RAM instead of the PCI bus.

Alternatively, if the SMRAM register is activated, then the i440fx will 
redirect 0xa0000 to RAM regardless of whether the CPU asserts that 
signal.  That means that even without KVM supporting SMM, this mode can 
happen.

In general, the memory controller can redirect IO accesses to RAM or to 
the PCI bus.  The PCI bus may redirect the access to the ISA bus.

> > > For my purpose in using this to program the IOMMU with guest 
> > > physical to
> > > host virtual addresses for device assignment, it doesn't really matter
> > > since there should never be a DMA in this range of memory.  But for a
> > > general RAM API, I'm not sure either.  I'm tempted to say that while
> > > this is in fact a use of RAM, the RAM is never presented to the 
> > > guest as
> > > usable system memory (E820_RAM for x86), and should therefore be
> > > excluded from the RAM API if we're using it only to track regions that
> > > are actual guest usable physical memory.
> > >
> > > We had talked on irc that pc.c should be registering 0x0 to
> > > below_4g_mem_size as ram, but now I tend to disagree with that.  The
> > > memory backing 0xa0000-0x100000 is present, but it's not presented to
> > > the guest as usable RAM.  What's your strict definition of what the RAM
> > > API includes?  Is it only what the guest could consider usable RAM or
> > > does it also include quirky chipset accelerator features like this
> > > (everything with a guest physical address)?  Thanks,
> >
> > Today we model on flat space that's a mixed of device memory, RAM, or 
> > ROM.  This is not how machines work and the limitations of this model 
> > is holding us back.
> >
> > IRL, there's a block of RAM that's connected to a memory controller.  
> > The CPU is also connected to the memory controller.  Devices are 
> > connected to another controller which is in turn connected to the 
> > memory controller.  There may, in fact, be more than one controller 
> > between a device and the memory controller.
> >
> > A controller may change the way a device sees memory in arbitrary 
> > ways.  In fact, two controllers accessing the same page might see 
> > something totally different.
> >
> > The idea behind the RAM API is to begin to establish this hierarchy.  
> > RAM is not what any particular device sees--it's actual RAM.  IOW, 
> > the RAM API should represent what address mapping I would get if I 
> > talked directly to DIMMs.
> >
> > This is not what RamBlock is even though the name would suggest 
> > otherwise.  RamBlocks are anything that qemu represents as cache 
> > consistency directly accessable memory.  Device ROMs and areas of 
> > device RAM are all allocated from the RamBlock space.
> >
> > So the very first task of a RAM API is to simplify differentiate 
> > these two things.  Once we have the base RAM API, we can start adding 
> > the proper APIs that sit on top of it (like a PCI memory API).
>
> Things aren't that bad - a ram_addr_t and a physical address are 
> already different things, so we already have one level of translation.

Yeah, but ram_addr_t doesn't model anything meaningful IRL.  It's an 
internal implementation detail.

Regards,

Anthony Liguori