Contributed by Peter N. M. Hansteen on from the where-would-sir-like-his-640kB? dept.
jsg@), but following this commit, OpenBSD/amd64 kernels now have a 512GB address space.
The commit message innocently reads,
List: openbsd-cvs Subject: CVS: cvs.openbsd.org: src From: Jonathan Gray <jsg () cvs ! openbsd ! org> Date: 2026-06-22 0:27:33 Message-ID: 162f075352ec66cd () cvs ! openbsd ! org CVSROOT: /cvs Module name: src Changes by: jsg@cvs.openbsd.org 2026/06/21 18:27:33 Modified files: sys/arch/amd64/include: vmparam.h Log message: raise the size of amd64 kernel virtual address space from 4G to 512G
This allows mapping all of VRAM on recent Radeon cards where the PCI BAR size is the full (> 4G) size. Reported by Petre Rodan, who has a Radeon RX 6800 with 16G of VRAM. feedback from claudio@ ok kettenis@
By Noah Altun (naltun) noah@altun.cc on
Good update to amd64 support. Great work devs!
Reply
By Amit Kulkarni (amitkulz) on
This commit is also important
https://marc.info/?l=openbsd-cvs&m=178222561369914&w=2
Allocate mbufs in high memory if only 64 bit DMA interfaces exist.
Mbufs on amd64 were allocated below 4 GB so that devices not capable
of 64 bit DMA can access the memory. Interface drivers use
BUS_DMA_64BIT to allow the DMA layer doing 64 bit transfers. Now
flag interfaces with IFXF_MBUF_64BIT that are capable of 64 bit DMA
on all their mbuf rings.
If only such interfaces exist in the system during amd64 boot,
allocate mbufs and mbuf clusters also in high memory. Other
architectures may be limited to 32 bit memory anyway or use an
IOMMU. On riscv64 or arm64 busses may exists that support less
than 64 bit, this will be handled later.
Hotplug devices that do not support 64 bit DMA will use bounce
buffering. By changing the flags in device drivers we can force
bounce buffering and find missing calls to bus_dmamap_sync().
OK kettenis@ deraadt@
Reply