On 07.01.2026 17:54, Teddy Astie wrote: > Current logic prefer using CFC/CF8 and fallbacks on mmcfg when accessing >> 255 registers or a non-zero segment. Change the logic to always rely
(Minor: Many mail programs, like mine, will mistake a > in the first column as being reply quoting.) > on mmcfg unless it is not available to avoid locking on pci_config_lock > if possible. > > Signed-off-by: Teddy Astie <[email protected]> > --- > Are there x86 platforms where MMCFG is the only way to access PCI > configuration space ? If there were, how would that fact be communicated? > --- a/xen/arch/x86/x86_64/pci.c > +++ b/xen/arch/x86/x86_64/pci.c > @@ -14,62 +14,56 @@ > uint8_t pci_conf_read8(pci_sbdf_t sbdf, unsigned int reg) > { > uint32_t value; > + int ret = pci_mmcfg_read(sbdf.seg, sbdf.bus, sbdf.devfn, reg, 1, &value); Along the lines of what in particular Roger said in reply to the cover letter, I'm unconvinced we want to slow down (even if just minimally) things by unconditionally making this call (and similar ones below). Jan
