On Mon, Aug 25, 2025 at 4:13 PM Luck, Tony <tony.luck@xxxxxxxxx> wrote: > > On Mon, Aug 25, 2025 at 10:33:48PM +0000, Jiaqi Yan wrote: > > Legacy persistent memory, e.g. Device DAX configured like the following > > > > 440000000-303fffffff : Persistent Memory (legacy) > > 440000000-47fffffff : dax1.0 > > 480000000-4bfffffff : dax2.0 > > 4c0000000-4ffffffff : dax3.0 > > 500000000-53fffffff : dax4.0 > > ... > > > > can support recover from Machine Check Exception due to memory failure. > > Therefore there is need to test it by injecting memory error > > using EINJ to legacy persistent memory. > > > > However, current EINJ only check if physical address falls into the > > IORES_DESC_PERSISTENT_MEMORY_LEGACY region. So attempt to inject > > error to "Persistent Memory (legacy)" fails with -EINVAL. > > > > Allow EINJ to inject at physical address belongin to > > IORES_DESC_PERSISTENT_MEMORY_LEGACY. > > > > Tested on a machine configured with Device DAX: > > memmap=4G!12G nd_e820.pmem=12G,4G,mode=fsdax memmap=176G!17G > > nd_e820.pmem=17G,1G,mode=devdax,align=1G memmap=176G!209G > > nd_e820.pmem=209G,1G,mode=devdax,align=1G memmap=176G!401G > > nd_e820.pmem=401G,1G,mode=devdax,align=1G memmap=176G!593G > > nd_e820.pmem=593G,1G,mode=devdax,align=1G > > > > Injected error at 0x35238d2000, within Device DAX region and > > allocated to a userspace test process. EINJ driver now issue the > > injection request to firmware, and firmware logs shows injection > > at 0x35238d2000 succeeded. The userspace test process then > > accessed 0x35238d2000, caused a MCE, and killed by SIGBUS. > > > > Signed-off-by: Jiaqi Yan <jiaqiyan@xxxxxxxxxx> > > --- > > drivers/acpi/apei/einj-core.c | 2 ++ > > 1 file changed, 2 insertions(+) > > > > diff --git a/drivers/acpi/apei/einj-core.c b/drivers/acpi/apei/einj-core.c > > index 2561b045acc7b..e746fa66f92ff 100644 > > --- a/drivers/acpi/apei/einj-core.c > > +++ b/drivers/acpi/apei/einj-core.c > > @@ -712,6 +712,8 @@ int einj_error_inject(u32 type, u32 flags, u64 param1, u64 param2, u64 param3, > > != REGION_INTERSECTS) && > > (region_intersects(base_addr, size, IORESOURCE_MEM, IORES_DESC_PERSISTENT_MEMORY) > > != REGION_INTERSECTS) && > > + (region_intersects(base_addr, size, IORESOURCE_MEM, IORES_DESC_PERSISTENT_MEMORY_LEGACY) > > + != REGION_INTERSECTS) && > > I chatted offline with Dan Williams. He wondered whether this sanity > check should just be reduced to pass through any address except MMIO > and leave it to the BIOS to decide what is a legitimate injection > target. I guess that's fine, but wonder should the EINJ driver still exclude IORES_DESC_ACPI_TABLES and IORES_DESC_ACPI_NV_STORAGE? to prevent that, say, by coincidence or not an error "corrupted" EINJ table itself, or "corrupted" other ACPI data. > > > (region_intersects(base_addr, size, IORESOURCE_MEM, IORES_DESC_SOFT_RESERVED) > > != REGION_INTERSECTS) && > > !arch_is_platform_page(base_addr))) > > -- > > 2.51.0.268.g9569e192d0-goog > > -Tony
