Re: [PATCH 1/2] x86/efi: Map EFI memmap entries in-order at runtime

[James Bottomley <jbottomley@odin.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset="utf-8", Size: 5044 bytes --]

On Wed, 2015-09-30 at 09:43 -0700, Andy Lutomirski wrote:
> On Wed, Sep 30, 2015 at 2:30 AM, Ard Biesheuvel
> <ard.biesheuvel@linaro.org> wrote:
> > On 29 September 2015 at 23:58, Laszlo Ersek <lersek@redhat.com> wrote:
> >> On 09/28/15 08:41, Matthew Garrett wrote:
> >>> On Mon, Sep 28, 2015 at 08:16:46AM +0200, Ingo Molnar wrote:
> >>>
> >>>> So the question is, what does Windows do?
> >>>
> >>> It's pretty trivial to hack OVMF to dump the SetVirtualAddressMap()
> >>> arguments to the qemu debug port. Unfortunately I'm about to drop
> >>> mostly  offline for a week, otherwise I'd give it a go...
> > [...]
> >> Then I booted my Windows Server 2012 R2, Windows 8.1, and Windows 10
> >> guests, with the properties table feature enabled vs. disabled in the
> >> firmware. (All three Windows guests were updated first though.)
> >>
> >> All three Windows OSes adapt their SetVirtualAddressMap() calls, when
> >> the feature is enabled in the firmware. However, Windows 8.1 crashes
> >> nonetheless (BSOD, I forget the fault details, sorry). Windows Server
> >> 2012 R2 and Windows 10 boot fine.
> >>
> >
> > Looking at the log, it seems the VA mapping strategy is actually the
> > same (i.e., bottom-up for Win10), and the difference can be explained
> > by the differences in the memory map provided by the firmware to the
> > OS. And indeed, the Win8.1 log shows the following:
> >
> >  # MemType Phys 0x  Virt 0x  Size 0x Attributes
> > -- ------- -------- -------- ------- -------------------------------
> >  0 RtData  7EC21000 FFBFA000 0006000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  1 RtCode  7EC27000 FFBF3000 0007000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> >  2 RtData  7EC2E000 FFBEC000 0007000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  3 RtData  7EC35000 FFBEB000 0001000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  4 RtCode  7EC36000 FFBE6000 0005000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> >  5 RtData  7EC3B000 FFBE4000 0002000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  6 RtData  7EC60000 FFBDE000 0006000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  7 RtCode  7EC66000 FFBD5000 0009000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> >  8 RtData  7EC6F000 FFBD3000 0002000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> >  9 RtData  7EC9E000 FFAFA000 00D9000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 10 RtCode  7ED77000 FFA63000 0097000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> > 11 RtData  7EE0E000 FFA58000 000B000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 12 RtData  7FE99000 FFA52000 0006000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 13 RtCode  7FE9F000 FFA4C000 0006000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> > 14 RtData  7FEA5000 FFA49000 0003000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 15 RtCode  7FEA8000 FFA42000 0007000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> > 16 RtData  7FEAF000 FFA3F000 0003000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 17 RtCode  7FEB2000 FFA36000 0009000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> > 18 RtData  7FEBB000 FFA33000 0003000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 19 RtCode  7FEBE000 FFA2A000 0009000 [UC|WC|WT|WB|  |  |RO|  |  |RT]
> > 20 RtData  7FEC7000 FFA04000 0026000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 21 RtData  7FFD0000 FF9E4000 0020000 [UC|WC|WT|WB|  |XP|  |  |  |RT]
> > 22 RtData  FFE00000 FF7E4000 0200000 [UC|  |  |  |  |XP|  |  |  |RT]
> >
> > I.e., the physical addresses increase while the virtual addresses
> > decrease, and since each consecutive RuntimeCode/RuntimeData pair
> > constitutes a PE/COFF image (.text and .data, respectively), the
> > PE/COFF images appear corrupted in the virtual space.
> 
> All of this garbage makes me want to ask a rhetorical question:
> 
> Why on Earth did anyone think it's a good idea to invoke EFI functions
> at CPL0 once the OS is booted?

I'm afraid the originators of EFI (Intel) look on it as a DOS
replacement ... with the same OS support.

> And a more practical question:
> 
> Do we actually have to invoke EFI functions at CPL0?
> 
> I really mean it.  Sure, for things like reboot where we give up
> control and don't get it back, we need to do that.  But for things
> like variable access, the EFI code should really only need access to
> EFI memor (with a known PA -> VA map) and the ability to trigger an
> SMI.  Doing it at CPL3 could require more fixups than would really
> make sense, but could we virtualize it instead?
> 
> Actually, CPL3 + IOPL3 just might work.
> 
> Heck, on mixed-mode, we're already invoke EFI functions in compat
> mode, and that seems okay, so those functions can't be poking at any
> CPU state that varies between long and 32-bit modes.

It's hard.  The EFI functions expect to interact directly with kernel
memory, which they can't at CPL3.  We could vector all that through a
CPL3 readable buffer but anything within EFI that uses privileged
instructions will fault and we'll have to handle it ... this really
sounds like a can of worms.  Especially as windows will be no help
testing all of this because it will call in at CPL0.

James

ÿôèº{.nÇ+‰·Ÿ®‰­†+%ŠËÿ±éݶ\x17¥Šwÿº{.nÇ+‰·¥Š{±þG«éÿŠ{ayº\x1dʇڙë,j\a­¢f£¢·hšïêÿ‘êçz_è®\x03(­éšŽŠÝ¢j"ú\x1a¶^[m§ÿÿ¾\a«þG«éÿ¢¸?™¨è­Ú&£ø§~á¶iO•æ¬z·švØ^\x14\x04\x1a¶^[m§ÿÿÃ\fÿ¶ìÿ¢¸?–I¥