On Fri, Sep 20, 2019 at 03:35:00PM +0200, Mark Kettenis wrote: > > Date: Fri, 20 Sep 2019 06:06:40 -0700 > > From: Mike Larkin <[email protected]> > > > > On Fri, Sep 20, 2019 at 02:22:13PM +0200, Mark Kettenis wrote: > > > > Date: Fri, 20 Sep 2019 02:55:27 -0700 > > > > From: Mike Larkin <[email protected]> > > > > > > > > On Fri, Sep 20, 2019 at 01:09:56AM +0900, YASUOKA Masahiko wrote: > > > > > Hi, > > > > > > > > > > I recently got a VAIO Pro PK. The diff below is required to boot. > > > > > Without the diff, it freezes during boot. > > > > > > > > > > > > > > Its EFI framebuffer is located 0x4000000000 (9 zeros). This is > 4GB > > > > > and higher than highest available memory of the machine. These > > > > > configuraions seem to cause the problem. > > > > > > > > > > * * * > > > > > > > > > > Call cninit() after pmap_bootstrap() is called. Since the EFI > > > > > framebuffer may be located > 4GB which is not initialized by locore, > > > > > but by pmap_bootstrap(). Also make the address parameter passed to > > > > > pmap_bootstrap() cover the framebuffer. Actually VAIO pro PK's > > > > > framebuffer is located higher than the highest available memory > > > > > region. > > > > > > > > > > ok? comments? > > > > > > > > > > > > > Hi, > > > > > > > > I have a few questions... > > > > > > > > 1. There seems to be no limit on the max PA that we extend to here. > > > > This means, for example, if EFI places the framebuffer past 2TB > > > > PA, we won't have enough direct map to cover the mapping. Plus > > > > I think this will end up extending the direct map to cover any hole > > > > between "end of phys mem" and "efi fb addr". At a minimum, I think > > > > we need some sort of max PA clamp here. I don't know what Sony's > > > > placement algorithm is, but 0x4000000000 is 256GB PA. > > > > > > A dmesg and pcidump output would be useful. > > > > > > I suspect that this is a discrete graphics card where the EFI frame > > > buffer resides in VRAM. Using the direct map in this case is probably > > > not the right thing to do. > > > > > > > 2. What does delaying cninit do for machines that have errors or > > > > printfs before this? Would those even print anymore? This would > > > > affect all machines, even those without efifb, correct? > > > > > > Yes and no. It doesn't affect the classic VGA glass console, but it > > > does mean serial output might disappear. That isn't acceptable I'd > > > say. > > > > > > > 3. I am not a big fan of placing device-specific quirks in > > > > init_x86_64. Could this not be done in the efifb specific console > > > > init code? You could pmap_enter whatever you wanted there, based on > > > > the PA EFI sent you. Or does efifb go through the direct map for > > > > all video output? If so, we may be stuck creating that big direct map > > > > range. If that's the case though, we should probably try to restrict > > > > the permissions in the unused holes. > > > > > > The direct map is only used early on in the boot process. The frame > > > buffer is remapped in mainbus_attach() such that we can use > > > write-combining. But that is done after we print copyright. I think > > > the remapping could be done a bit earlier, but not before uvm gets > > > initialized, which happens after we print the copyright message. > > > > > > We don't have to use the direct map during early boot. If you gave us > > > some other way to map the framebuffer before pmap_bootstrap() has been > > > called we could stick that into efifb_cnattach_common(). We'd need > > > your help with that though. Note that the framebuffer can be fairly > > > large though (but we can probably come up with a reasonable upper > > > limit). > > > > > > > What sort of function do you need? Map this PA range at X, but before > > pmap_bootstrap? > > Map this PA range and hand me back the VA where you mapped it, indeed > before pmap_bootstrap(). > > I reckon you'd need to reserve slots in the early page tables for > this. The mapping needs to be uncached (UC or WC). I suspect we'll > continue to remap the frame buffer later, UC is fine as we're not > going to produce massive amounts of output to the console before doing > so.
Likely doable, but it will be some time before I can get to it. -ml
