Re: [Qemu-devel] [PATCH v2] map 64-bit PCI BARs at location provided by emulator

[Michael S. Tsirkin]

On Mon, Oct 14, 2013 at 12:27:26PM +0200, Igor Mammedov wrote:
> On Sun, 13 Oct 2013 23:28:47 +0300
> "Michael S. Tsirkin" <address@hidden> wrote:
> 
> > On Sun, Oct 13, 2013 at 07:33:19PM +0200, Igor Mammedov wrote:
> > > On Sun, 13 Oct 2013 19:46:09 +0300
> > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > 
> > > > On Sun, Oct 13, 2013 at 06:23:28PM +0200, Igor Mammedov wrote:
> > > > > On Sun, 13 Oct 2013 18:59:20 +0300
> > > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > > 
> > > > > > On Sun, Oct 13, 2013 at 05:11:54PM +0200, Igor Mammedov wrote:
> > > > > > > On Sun, 13 Oct 2013 15:31:23 +0300
> > > > > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > > > > 
> > > > > > > > On Sun, Oct 13, 2013 at 02:13:44PM +0200, Igor Mammedov wrote:
> > > > > > > > > Currently 64-bit PCI BARs are unconditionally mapped by BIOS 
> > > > > > > > > right
> > > > > > > > > over 4G + RamSizeOver4G location, which doesn't allow to 
> > > > > > > > > reserve
> > > > > > > > > extra space before 64-bit PCI window. For memory hotplug an 
> > > > > > > > > extra
> > > > > > > > > RAM space might be reserved after present 64-bit RAM end and 
> > > > > > > > > BIOS
> > > > > > > > > should map 64-bit PCI BARs after it.
> > > > > > > > > 
> > > > > > > > > Introduce "etc/pcimem64-start" romfile to provide BIOS a hint
> > > > > > > > > where it should start mapping of 64-bit PCI BARs. If romfile 
> > > > > > > > > is
> > > > > > > > > missing, BIOS reverts to legacy behavior and starts mapping 
> > > > > > > > > right
> > > > > > > > > after high memory.
> > > > > > > > > 
> > > > > > > > > Signed-off-by: Igor Mammedov <address@hidden>
> > > > > > > > > ---
> > > > > > > > > v2:
> > > > > > > > >   * place 64-bit window behind high RAM end if 
> > > > > > > > > "etc/pcimem64-start"
> > > > > > > > >     points below it.
> > > > > > > > 
> > > > > > > > Hmm I had an alternative suggestion of passing smbios
> > > > > > > > tables in from QEMU (seabios already has a mechanism for this).
> > > > > > > > We could then simply increase the existing RamSize variable.
> > > > > > > > 
> > > > > > > > What do you think about this idea?
> > > > > > > There is several issues with RamSizeOver4G I see:
> > > > > > > 1. As Gerd pointed out it is also used to setup a part of e820 
> > > > > > > tables.
> > > > > > >    It possibly could be also changed see "[Qemu-devel] [RfC 
> > > > > > > PATCH] e820: pass
> > > > > > >    high memory too" but it will be more complicated to keep 
> > > > > > > seabios
> > > > > > >    backward/forward compatible if we use RamSizeOver4G for 
> > > > > > > passing.
> > > > > > 
> > > > > > What are the compatibility worries, exactly?
> > > > > > 
> > > > > > The way I see it, if QEMU wants to be compatible it can't
> > > > > > use memory hotplug. In that case RamSizeOver4G will have
> > > > > > the old meaning which should work without issues.
> > > > > QEMU can't know what BIOS version is used and error out in wrong 
> > > > > case, and that
> > > > > could cause hard to guess what going wrong bugs in the guest.
> > > > > 
> > > > > > 
> > > > > > > 2. It doesn't scale if we are going to use more than 1Tb future.
> > > > > > >    Extending RamSizeOver4G value via additional CMOS port wasn't 
> > > > > > > welcomed
> > > > > > >    either, see
> > > > > > > http://www.seabios.org/pipermail/seabios/2010-September/000911.html
> > > > > > > 
> > > > > > > It looks like fw_cfg interface is preferred one and 
> > > > > > > "etc/pcimem64-start"
> > > > > > > approach doesn't have backward/forward compatibility issues.
> > > > > > 
> > > > > > Hmm but you aren't passing in memory size - pcimem64-start
> > > > > > merely tells bios where to put PCI devices.
> > > > > > So how would it help?
> > > > > If a lot of memory is reserved for memory hotplug (terabytes), that 
> > > > > could
> > > > > put window behind what RamSizeOver4G could address it.
> > > > 
> > > > All we care about for a compatible bios is that it
> > > > keep working in old hardware configurations.
> > > > IMO there's no real need for old BIOS to work
> > > > with e.g. memory hotplug - it's reasonable
> > > > to ask users to update fitmware if they want
> > > > to use this feature.
> > > Fine with me if we don't care about strange hangs if something misused.
> > > 
> > > But RamSizeOver4G still doesn't allow correctly place PCI window if large
> > > amount of memory present/reserved.
> > 
> > I'm fine with adding a new fw cfg really.
> > I am however not sure we should have a single new base64 flag.
> > Hardware has a set of ranges where guest can
> > put PCI devices - would we be better off with
> > an interface that lists all these ranges?
> >
> > I'm not saying we should but this needs a bit of thought.
> I've offered to reuse your "etc/pci-info" fw_cfg that were rejected earlier,
> but it still seems no go: 
> http://permalink.gmane.org/gmane.comp.emulators.qemu/237307
> 
> And there is slight difference between PCI holes and PCI address space 
> mappings
> represented by MemoryRegion-s in QEMU.
> 
> Basically we only need to inform BIOS where to PCI address spaces start and
> simple "etc/pcimem64-start" + "etc/pcimem32-start" are just fine for that.
> And for now (memory hotplug) we need only the first one, the second one is
> very well hardcoded in Seabios/QEMU.

Hmm okay but the only reason we want to control this is
because it conflicts with memory hotplug, no?
I vaguely recall we discussed just passing amount of
hot-pluggable memory in, in the past, but I don't
remember the reason we decided against it.
Could you refresh my memory?

> It's less complicated than maintaining structure interface (which would be
> difficult to keep compatible if it will be changing in future).
> 
> A more flexible approach would be a separate table (like e820 table, but only
> for PCI ranges), but it seems a bit of overkill for now (I can't picture
> a need for more than 2 PCI address space mappings).

It's not unthinkable.
Multiple ECAM regions (for multi-root systems) can make holes in the address 
space.
Also, we just ignore everything above the ioapic, but that's
not a must, we could maybe use address space above ioapic.

> > > > 
> > > > > > > > 
> > > > > > > > 
> > > > > > > > > ---
> > > > > > > > >  src/fw/pciinit.c | 13 ++++++++++++-
> > > > > > > > >  1 file changed, 12 insertions(+), 1 deletion(-)
> > > > > > > > > 
> > > > > > > > > diff --git a/src/fw/pciinit.c b/src/fw/pciinit.c
> > > > > > > > > index b29db99..798309b 100644
> > > > > > > > > --- a/src/fw/pciinit.c
> > > > > > > > > +++ b/src/fw/pciinit.c
> > > > > > > > > @@ -18,6 +18,8 @@
> > > > > > > > >  #include "paravirt.h" // RamSize
> > > > > > > > >  #include "string.h" // memset
> > > > > > > > >  #include "util.h" // pci_setup
> > > > > > > > > +#include "byteorder.h" // le64_to_cpu
> > > > > > > > > +#include "romfile.h" // romfile_loadint
> > > > > > > > >  
> > > > > > > > >  #define PCI_DEVICE_MEM_MIN     0x1000
> > > > > > > > >  #define PCI_BRIDGE_IO_MIN      0x1000
> > > > > > > > > @@ -764,6 +766,15 @@ static void pci_bios_map_devices(struct 
> > > > > > > > > pci_bus *busses)
> > > > > > > > >  {
> > > > > > > > >      if (pci_bios_init_root_regions(busses)) {
> > > > > > > > >          struct pci_region r64_mem, r64_pref;
> > > > > > > > > +        u64 ram64_end = 0x100000000ULL + RamSizeOver4G;
> > > > > > > > > +        u64 base64 = 
> > > > > > > > > le64_to_cpu(romfile_loadint("etc/pcimem64-start",
> > > > > > > > > +                                 ram64_end));
> > > > > > > > > +        if (base64 < ram64_end) {
> > > > > > > > > +            dprintf(1, "ignorig etc/pcimem64-start [0x%llx] 
> > > > > > > > > below present RAM, "
> > > > > > > > > +                       "placing 64-bit PCI window behind RAM 
> > > > > > > > > end: %llx",
> > > > > > > > > +                        base64, ram64_end);
> > > > > > > > > +            base64 = ram64_end;
> > > > > > > > > +        }
> > > > > > > > >          r64_mem.list.first = NULL;
> > > > > > > > >          r64_pref.list.first = NULL;
> > > > > > > > >          
> > > > > > > > > pci_region_migrate_64bit_entries(&busses[0].r[PCI_REGION_TYPE_MEM],
> > > > > > > > > @@ -779,7 +790,7 @@ static void pci_bios_map_devices(struct 
> > > > > > > > > pci_bus *busses)
> > > > > > > > >          u64 align_mem = pci_region_align(&r64_mem);
> > > > > > > > >          u64 align_pref = pci_region_align(&r64_pref);
> > > > > > > > >  
> > > > > > > > > -        r64_mem.base = ALIGN(0x100000000LL + RamSizeOver4G, 
> > > > > > > > > align_mem);
> > > > > > > > > +        r64_mem.base = ALIGN(base64, align_mem);
> > > > > > > > >          r64_pref.base = ALIGN(r64_mem.base + sum_mem, 
> > > > > > > > > align_pref);
> > > > > > > > >          pcimem64_start = r64_mem.base;
> > > > > > > > >          pcimem64_end = r64_pref.base + sum_pref;
> > > > > > > > > -- 
> > > > > > > > > 1.8.3.1
> > > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > -- 
> > > > > > > Regards,
> > > > > > >   Igor
> > > > > 
> > > > > 
> > > > > -- 
> > > > > Regards,
> > > > >   Igor
> > > 
> > > 
> > > -- 
> > > Regards,
> > >   Igor