Re: [PATCH qemu v20] spapr: Implement Open Firmware client interface

[David Gibson]

On Tue, May 25, 2021 at 12:08:45PM +0200, BALATON Zoltan wrote:
> On Tue, 25 May 2021, David Gibson wrote:
> > On Mon, May 24, 2021 at 12:55:07PM +0200, BALATON Zoltan wrote:
> > > On Mon, 24 May 2021, David Gibson wrote:
> > > > On Sun, May 23, 2021 at 07:09:26PM +0200, BALATON Zoltan wrote:
> > > > > On Sun, 23 May 2021, BALATON Zoltan wrote:
> > > > > > On Sun, 23 May 2021, Alexey Kardashevskiy wrote:
> > > > > > > One thing to note about PCI is that normally I think the client
> > > > > > > expects the firmware to do PCI probing and SLOF does it. But VOF
> > > > > > > does not and Linux scans PCI bus(es) itself. Might be a problem 
> > > > > > > for
> > > > > > > you kernel.
> > > > > > 
> > > > > > I'm not sure what info does MorphOS get from the device tree and 
> > > > > > what it
> > > > > > probes itself but I think it may at least need device ids and info 
> > > > > > about
> > > > > > the PCI bus to be able to access the config regs, after that it 
> > > > > > should
> > > > > > set the devices up hopefully. I could add these from the board code 
> > > > > > to
> > > > > > device tree so VOF does not need to do anything about it. However 
> > > > > > I'm
> > > > > > not getting to that point yet because it crashes on something that 
> > > > > > it's
> > > > > > missing and couldn't yet find out what is that.
> > > > > > 
> > > > > > I'd like to get Linux working now as that would be enough to test 
> > > > > > this
> > > > > > and then if for MorphOS we still need a ROM it's not a problem if at
> > > > > > least we can boot Linux without the original firmware. But I can't 
> > > > > > make
> > > > > > Linux open a serial console and I don't know what it needs for 
> > > > > > that. Do
> > > > > > you happen to know? I've looked at the sources in 
> > > > > > Linux/arch/powerpc but
> > > > > > not sure how it would find and open a serial port on pegasos2. It 
> > > > > > seems
> > > > > > to work with the board firmware and now I can get it to boot with 
> > > > > > VOF
> > > > > > but then it does not open serial so it probably needs something in 
> > > > > > the
> > > > > > device tree or expects the firmware to set something up that we 
> > > > > > should
> > > > > > add in pegasos2.c when using VOF.
> > > > > 
> > > > > I've now found that Linux uses rtas methods read-pci-config and
> > > > > write-pci-config for PCI access on pegasos2 so this means that we'll
> > > > > probably need rtas too (I hoped we could get away without it if it 
> > > > > were only
> > > > > used for shutdown/reboot or so but seems Linux needs it for PCI as 
> > > > > well and
> > > > > does not scan the bus and won't find some devices without it).
> > > > 
> > > > Yes, definitely sounds like you'll need an RTAS implementation.
> > > 
> > > I plan to fix that after managed to get serial working as that seems to 
> > > not
> > > need it. If I delete the rtas-size property from /rtas on the original
> > > firmware that makes Linux skip instantiating rtas, but I still get serial
> > > output just not accessing PCI devices. So I think it should work and keeps
> > > things simpler at first. Then I'll try rtas later.
> > > 
> > > > > While VOF can do rtas, this causes a problem with the hypercall 
> > > > > method using
> > > > > sc 1 that goes through vhyp but trips the assert in ppc_store_sdr1() 
> > > > > so
> > > > > cannot work after guest is past quiesce.
> > > > 
> > > > > So the question is why is that
> > > > > assert there
> > > > 
> > > > Ah.. right.  So, vhyp was designed for the PAPR use case, where we
> > > > want to model the CPU when it's in supervisor and user mode, but not
> > > > when it's in hypervisor mode.  We want qemu to mimic the behaviour of
> > > > the hypervisor, rather than attempting to actually execute hypervisor
> > > > code in the virtual CPU.
> > > > 
> > > > On systems that have a hypervisor mode, SDR1 is hypervisor privileged,
> > > > so it makes no sense for the guest to attempt to set it.  That should
> > > > be caught by the general SPR code and turned into a 0x700, hence the
> > > > assert() if we somehow reach ppc_store_sdr1().
> > > > 
> > > > So, we are seeing a problem here because you want the 'sc 1'
> > > > interception of vhyp, but not the rest of the stuff that goes with it.
> > > > 
> > > > > and would using sc 1 for hypercalls on pegasos2 cause other
> > > > > problems later even if the assert could be removed?
> > > > 
> > > > At least in the short term, I think you probably can remove the
> > > > assert.  In your case the 'sc 1' calls aren't truly to a hypervisor,
> > > > but a special case escape to qemu for the firmware emulation.  I think
> > > > it's unlikely to cause problems later, because nothing on a 32-bit
> > > > system should be attempting an 'sc 1'.  The only thing I can think of
> > > > that would fail is some test case which explicitly verified that 'sc
> > > > 1' triggered a 0x700 (SIGILL from userspace).
> > > 
> > > OK so the assert should check if the CPU has an HV bit. I think there was 
> > > a
> > > #detine for that somewhere that I can add to the assert then I can try 
> > > that.
> > > What I wasn't sure about is that sc 1 would conflict with the guest's 
> > > usage
> > > of normal sc calls or are these going through different paths and only sc 
> > > 1
> > > will trigger vhyp callback not affecting notmal sc calls?
> > 
> > The vhyp shouldn't affect normal system calls, 'sc 1' is specifically
> > for hypercalls, as opposed to normal 'sc' (a.k.a. 'sc 0'), and the
> > vhyp only intercepts the hypercall version (after all Linux on PAPR
> > certainly uses its own system calls, and hypercalls are active for the
> > lifetime of the guest there).
> > 
> > > (Or if this causes
> > > an otherwise unnecessary VM exit on KVM even when it works then maybe
> > > looking for a different way in the future might be needed.
> > 
> > What you're doing here won't work with KVM as it stands.  There are
> > basically two paths into the vhyp hypercall path: 1) from TCG, if we
> > interpret an 'sc 1' instruction we enter vhyp, 2) from KVM, if we get
> > a KVM_EXIT_PAPR_HCALL KVM exit then we also go to the vhyp path.
> > 
> > The second path is specific to the PAPR (ppc64) implementation of KVM,
> > and will not work for a non-PAPR platform without substantial
> > modification of the KVM code.
> 
> OK so then at that point when we try KVM we'll need to look at alternative
> ways, I think MOL OSI worked with KVM at least in MOL but will probably make
> all syscalls exit KVM but since we'll probably need to use KVM PR it will
> exit anyway. For now I keep this vhyp as it does not run with KVM for other
> reasons yet so that's another area to clean up so as a proof of concept
> first version of using VOF vhyp will do.

Eh, since you'll need to modify KVM anyway, it probably makes just as
much sense to modify it to catch the 'sc 1' as MoL's magic thingy.

> [...]
> > > > > I've tested that the missing rtas is not the reason for getting no 
> > > > > output
> > > > > via serial though, as even when disabling rtas on pegasos2.rom it 
> > > > > boots and
> > > > > I still get serial output just some PCI devices are not detected 
> > > > > (such as
> > > > > USB, the video card and the not emulated ethernet port but these are 
> > > > > not
> > > > > fatal so it might even work as a first try without rtas, just to boot 
> > > > > a
> > > > > Linux kernel for testing it would be enough if I can fix the serial 
> > > > > output).
> > > > > I still don't know why it's not finding serial but I think it may be 
> > > > > some
> > > > > missing or wrong info in the device tree I generat. I'll try to focus 
> > > > > on
> > > > > this for now and leave the above rtas question for later.
> > > > 
> > > > Oh.. another thought on that.  You have an ISA serial port on Pegasos,
> > > > I believe.  I wonder if the PCI->ISA bridge needs some configuration /
> > > > initialization that the firmware is expected to do.  If so you'll need
> > > > to mimic that setup in qemu for the VOF case.
> > > 
> > > That's what I begin to think because I've added everything to the device
> > > tree that I thought could be needed and I still don't get it working so it
> > > may need some config from the firmware. But how do I access device 
> > > registers
> > > from board code? I've tried adding a machine reset method and write to
> > > memory mapped device registers but all my attempts failed. I've tried
> > > cpu_stl_le_data and even memory_region_dispatch_write but these did not 
> > > get
> > > to the device. What's the way to access guest mmio regs from QEMU?
> > 
> > That's odd, cpu_stl() and memory_region_dispatch_write() should work
> > from board code (after the relevant memory regions are configured, of
> > course).  As an ISA serial port, it's probably accessed through IO
> > space, not memory space though, so you'd need &address_space_io.  And
> > if there is some bridge configuration then it's the bridge control
> > registers you need to look at not the serial registers - you'd have to
> > look at the bridge documentation for that.  Or, I guess the bridge
> > implementation in qemu, which you wrote part of.
> 
> I've found at last that stl_le_phys() works. There are so many of these that
> I never know when to use which.
> 
> I think the address_space_rw calls in vof_client_call() in vof.c could also
> use these for somewhat shorter code. I've ended up with
> stl_le_phys(CPU(cpu)->as, addr, val) in my machine reset methodbut I don't
> even need that now as it works without additional setup. Also VOF's memory
> access is basically the same as the already existing rtas_st() and co. so
> maybe that could be reused to make code smaller?

rtas_ld() and rtas_st() should only be used for reading/writing RTAS
parameters to and from memory.  Accessing IO shouldn't be done with
those.

For IO you probably want the cpu_st*() variants in most cases, since
you're trying to emulate an IO access from the virtual cpu.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature