Date: Wed, 16 Jun 2021 18:19:49 +1000 From: Peter Grehan <grehan@freebsd.org> To: =?UTF-8?Q?Corvin_K=c3=b6hne?= <c.koehne@beckhoff.com> Cc: p.bruenn@beckhoff.com, freebsd-virtualization@freebsd.org Subject: Re: [PATCH 0/4] Prepare bhyve's OVMF for GPU-Passthrough Message-ID: <d93fe505-2818-2f00-fab4-b18fcf1f8a8d@freebsd.org> In-Reply-To: <20210615083000.1286-1-c.koehne@beckhoff.com> References: <20210615083000.1286-1-c.koehne@beckhoff.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Hi Corvin, Thanks for posting these here. > GPU-Passthrough for bhyve requires a few patches to work properly. > These patches will allow GPU-Passthrough for bhyve. > It will work for dedicated AMD GPUs and integrated Intel GPUs. > Nvidia GPUs are untested. I have no issue with patches 3/4 and 4/4. However, 1+2/4 switch over to having EFI do the PCI scan/address assignment rather than accepting bhyve's probe. This history here is that tychon@'s initial version of bhyve/EFI had the scan when it was mandatory in OVMF. That didn't work with passthru devices, requiring the "Duet" (EFI-over-BIOS) PCI code to be used. The recent upstreaming of Bhyve/EFI was done prior to the fix allowing dynamic relocation of pass-thru BARs. With Duet not being available, the flag allowing EFI to inherit PCI BAR assignment was set (PcdPciDisableBusEnumeration, as you've found). While it's now possible to allow EFI to scan/assign PCI BARs, I'd like to avoid it if possible for 2 reasons: - assignment policy can stay in bhyve, such as whether to locate 64-bit BARs in the 32-bit region which EFI didn't (doesn't?) allow. Bugs or corner-cases can be fixed in bhyve without requiring a modification to upstream EFI. - there is no need for EFI to perform a slow can via PCI bus operations, resulting in VM-exits, where bhyve can perform all this in memory, which can result in faster boot. Your patch description states: >For Linux guests, AMD GPUs require that their PCI ROM is processed by UEFI. Is it possible to fix this in bhyve ? Can pass-thru ROMs be mapped just like mmio BARs are ? later, Peter.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?d93fe505-2818-2f00-fab4-b18fcf1f8a8d>