On Tue, Jul 28, 2020 at 07:30:36PM +0200, Mark Kettenis wrote: > > Date: Tue, 28 Jul 2020 21:42:46 +1000 > > From: Jonathan Matthew <jonat...@d14n.org> > > > > On Tue, Jul 28, 2020 at 11:12:21AM +0200, Mark Kettenis wrote: > > > > Date: Tue, 28 Jul 2020 13:46:34 +1000 > > > > From: Jonathan Matthew <jonat...@d14n.org> > > > > > > > > On Mon, Jul 27, 2020 at 05:16:47PM +0200, Mark Kettenis wrote: > > > > > > Date: Mon, 27 Jul 2020 17:02:41 +0200 (CEST) > > > > > > From: Mark Kettenis <mark.kette...@xs4all.nl> > > > > > > > > > > > > Recent ACPI versions have deprecated "Processor()" nodes in favout > > > > > > of > > > > > > "Device()" nodes with a _HID() method that returns "ACPI0007". This > > > > > > diff tries to support machines with firmware that implements this. > > > > > > If > > > > > > you see something like: > > > > > > > > > > > > "ACPI0007" at acpi0 not configured > > > > > > > > > > > > please try the following diff and report back with an updated dmesg. > > > > > > > > > > > > Cheers, > > > > > > > > > > > > Mark > > > > > > > > > > And now with the right diff... > > > > > > > > On a dell r6415, it looks like this: > > > > > > > > acpicpu0 at acpi0copyvalue: 6: C1(@1 halt!) > > > > all the way up to > > > > acpicpu127 at acpi0copyvalue: 6: no cpu matching ACPI ID 127 > > > > > > > > which I guess means aml_copyvalue() needs to learn how to copy > > > > AML_OBJTYPE_DEVICE. > > > > > > Yes. It is not immediately obvious how this should work. Do we need > > > to copy the aml_node pointer or not? We don't do that for > > > AML_OBJTYPE_PROCESSOR and AML_OBJTYPE_POWERRSRC types which are > > > similar to AML_OBJTYPE_DEVICE. But AML_OBJTYPE_DEVICE object don't > > > carry any additional information. So we end up with just an empty > > > case to avoid the warning. > > > > > > Does this work on the Dell machines? > > > > We've seen crashes in pool_cache_get() in various places after all the > > acpicpus > > attach, which we haven't seen before on these machines, so I think it's > > corrupting memory somehow. > > Does that happen with only the acpicpu(4) diff?
Yes. Looking at this a bit more, in the case where aml_evalnode() can't copy the result value, it leaves it uninitialised, which means we'll call aml_freevalue(&res) where res is stack junk. memset(&res, 0, sizeof(res)) seems to fix it. > > > With this addition, we get this for each cpu: > > acpicpu0 at acpi0: C1(@1 halt!) > > The exclamation mark indicates that this is the "fallback" C-state. > Is there a _CST method at all? > > Anyway, given that this is a server system, it isn't really surprising > that there isn't any fancy power saving stuff. Right, there doesn't seem to be any. The processor devices look like this in the aml: Scope (_SB) { Device (C000) { Name (_HID, "ACPI0007" /* Processor Device */) // _HID: Hardware ID Name (_UID, 0x00) // _UID: Unique ID } Device (C001) { Name (_HID, "ACPI0007" /* Processor Device */) // _HID: Hardware ID Name (_UID, 0x01) // _UID: Unique ID } .. and so on. > > > > Index: dev/acpi/dsdt.c > > > =================================================================== > > > RCS file: /cvs/src/sys/dev/acpi/dsdt.c,v > > > retrieving revision 1.252 > > > diff -u -p -r1.252 dsdt.c > > > --- dev/acpi/dsdt.c 21 Jul 2020 03:48:06 -0000 1.252 > > > +++ dev/acpi/dsdt.c 28 Jul 2020 09:04:15 -0000 > > > @@ -996,6 +996,8 @@ aml_copyvalue(struct aml_value *lhs, str > > > lhs->v_objref = rhs->v_objref; > > > aml_addref(lhs->v_objref.ref, ""); > > > break; > > > + case AML_OBJTYPE_DEVICE: > > > + break; > > > default: > > > printf("copyvalue: %x", rhs->type); > > > break; > > > >