QEMU GPU vfio-pci pass through

@tlosm
Quote:May be a problem on a real PegasosII, but @balaton fixed all problems (by converting 64 bit BARs to 32 bit, interrupt related fixes, etc.) in his bboot.fth and BBoot, and probably in QEmu itself as well.
There should be no problems using the QEmu emulation of the PegasosII any more with PCIe gfx cards, at least not with AmigaOS 4.x, MorphOS may be different.
Using BBoot on a real PegasosII is probably not possible, but using bboot.fth should be, and it might even help getting PCIe gfx cards working with a PCI<->PCIe bridge.
The real AmigaOne SE/XE/µA1 systems have at least 100 times as much problems, hardware limits and bugs as the QEmu emulation of it, but QEmu doesn't emulate the hardware limits and bugs of the PegasosII and AmigaOne SE/XE/µA1

Several people got Radeon HD/RX PCIe cards working with vfio-pci with QEmu, and all used the PegasosII emulation since it used to be the best one for AmigaOS 4.x.
Now, with NVRAM support etc. added, the AmigaOne XE emulation is probably better than the PegasosII one for AmigaOS 4.x.
Sam4x0 emulation is too slow, one of the reasons is the missing "real" MMU, using TLB caches instead, which cause a lot of exceptions, and which probably can't be fixed/worked around in the QEmu emulation.
However, only @geenaam got vfio-pci working with usable speeds, for everyone else it was extremely slow, the real Radeon PCIe gfx card was even much slower than using SM501/502 software emulation.

What @white has to fix are the iommu groups, the Radeon gfx card has to be in a separate group with nothing else (required by the host) in it.
The easiest, and maybe only, option is trying to use a different PCIe slot for the Radeon gfx card.

@joerg
Quote:

Just to clarify a bit: I use bboot on real pegasos2 and it works fine. It speed up loading of os4 for about 10 seconds (30s vs 40s with os4s loader)

And, while maybe not very important at the moment because of "beta":

Quote:

In beta kernel for os4 all this also fixed already, and i can use real RadeonRx and RadeonHd via bridges, but.. that not released for now, sadly.

@joerg

Quote:

Swapping PCI slots is one way. The other is to use ACS-overwrite. This requires both a new kernel and some lines in the grub boot. I don't have the link handy that tells how to do that, butas for the ACS overwrite/patch, you just need ot install linux-zen as the kernel (I think I did that on arch linux using pacman -Sy linux-zen or similar... can be found online). Think it is similar on ubuntu-based systems.

After that, and reboot, etc, you will see only a single pci device per IOMMU group (you will have many groups). You can then just make a '-device ' entry in the qemu command line for just the video part and not bother with the audio if you don't want.

@themamboman

CachyOS based on Arch and its Linux kernels also have similar functionality.

https://wiki.cachyos.org/cachyos_basic/why_cachyos/
https://github.com/CachyOS/kernel-patches

The Linux kernel for the desktop in popular distributions is best recompiled yourself.
The versions included in the distributions are too (so to speak...) neutral.