Re: [Qemu-devel] [PULL 48/49] i386: populate floppy drive information in DSDT

[Michael S. Tsirkin]

On Wed, Feb 10, 2016 at 11:14:30AM -0500, John Snow wrote:
> 
> 
> On 02/09/2016 01:48 PM, Michael S. Tsirkin wrote:
> > On Tue, Feb 09, 2016 at 07:36:12PM +0100, Laszlo Ersek wrote:
> >> On 02/09/16 17:22, John Snow wrote:
> >>>
> >>>
> >>> On 02/09/2016 10:52 AM, Roman Kagan wrote:
> >>>> On Mon, Feb 08, 2016 at 03:20:47PM -0500, John Snow wrote:
> >>>>> On 02/08/2016 08:14 AM, Roman Kagan wrote:
> >>>>>> On Fri, Feb 05, 2016 at 07:25:07PM +0100, Igor Mammedov wrote:
> >>>>>>>> +    aml_append(fdi,
> >>>>>>>> +        aml_int(cylinders - 1));  /* Maximum Cylinder Number */
> >>>>>>> this puts uint64_t(-1) in AML i.e. cylinders == 0 and overflow 
> >>>>>>> happens here
> >>>>>>>
> >>>>>>> CCing Jon
> >>>>>>
> >>>>>> I guess this is the effect of John's fdc rework.  I used to think zero
> >>>>>> geometry was impossible at the time this patch was developed.
> >>>>>>
> >>>>>> I wonder if it hasn't been fixed already by
> >>>>>>
> >>>>>>   commit fd9bdbd3459e5b9d51534f0747049bc5b6145e07
> >>>>>>   Author: John Snow <address@hidden>
> >>>>>>   Date:   Wed Feb 3 11:28:55 2016 -0500
> >>>>>>
> >>>>>>       fdc: fix detection under Linux
> >>>>>
> >>>>> Yes, hopefully solved on my end. The geometry values for an empty disk
> >>>>> are not well defined (they certainly don't have any *meaning*) so if you
> >>>>> are populating tables based on an empty drive, I just hope you also have
> >>>>> the mechanisms needed to update said tables when the media changes.
> >>>>
> >>>> I don't.  At the time the patch was developed there basically were no
> >>>> mechanisms to update the geometry at all (and this was what you patchset
> >>>> addressed, in particular, wasn't it?) so I didn't care.
> >>>>
> >>>
> >>> That's not true.
> >>>
> >>> You could swap different 1.44MB-class diskettes for other geometries,
> >>> check this out:
> >>>
> >>> static const FDFormat fd_formats[] = {
> >>>     /* First entry is default format */
> >>>     /* 1.44 MB 3"1/2 floppy disks */
> >>>     { FDRIVE_DRV_144, 18, 80, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 20, 80, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 21, 80, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 21, 82, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 21, 83, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 22, 80, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 23, 80, 1, FDRIVE_RATE_500K, },
> >>>     { FDRIVE_DRV_144, 24, 80, 1, FDRIVE_RATE_500K, },
> >>> ...
> >>>
> >>> You absolutely could get different sector and track counts before my
> >>> patchset.
> >>>
> >>>
> >>>> Now if it actually has to be fully dynamic it's gonna be more
> >>>> involved...
> >>>>
> >>>>> What do the guests use these values for? Are they fixed at boot?
> >>>>
> >>>> Only Windows guests use it so it's hard to tell.  I can only claim that
> >>>> if I stick bogus values into that ACPI object the guest fails to read
> >>>> the floppy.
> >>
> >> We discussed this with John a bit on IRC.
> >>
> >> In my opinion, the real mess in this case is in the ACPI spec itself. If
> >> you re-read the _FDI control method's description, the Package that it
> >> returns contains *dynamic* geometry data, about the *disk* (not *drive*):
> >>
> >> - Maximum Cylinder Number // Integer (WORD)
> >> - Maximum Sector Number   // Integer (WORD)
> >> - Maximum Head Number     // Integer (WORD)
> >>
> >> What this seems to require is: the firmware developer should write ACPI
> >> code that
> >> - dynamically accesses the floppy drive / controller (using MMIO or IO
> >>   port registers),
> >> - retrieves the geometry of the disk actually inserted,
> >> - and returns the data nicely packaged.
> >>
> >> In effect, an ACPI-level driver for the disk.
> >>
> >> Now, John explained to me (and please keep in mind that this is my
> >> personal account of his explanation, not a factual rendering thereof),
> >> that there used to be no *standard* way to interrogate the current
> >> disk's geometry, other than trial and error with seeking.
> >>
> 
> At the very least, the Intel 82078 does not appear to be capable of
> replying to any query with a CHS triplet. It *can* report back the total
> number of sectors and the size of each sector, but that still requires
> geometry guesswork outside of the drive.
> 
> >> Apparently in UEFI Windows, Microsoft didn't want to implement this
> >> trial and error seeking, so -- given there was also no portable
> >> *hardware spec* to retrieve the same data, directly from the disk drive
> >> or controller -- they punted the entire question to ACPI. That is, to
> >> the firmware implementor.
> >>
> >> This is entirely bogus. For one, it ties the platform firmware (the UEFI
> >> binary in the flash chip on your motherboard) to your specific floppy
> >> drive / controller. Say good-bye to any independently bought & installed
> >> floppy drives.
> >>
> >> In theory, a floppy controller that comes with a separate PCI card could
> >> offer an option ROM with a UEFI driver in it, and that UEFI driver could
> >> install a separate SSDT with the hardware-matching _FDI in it. Still an
> >> unreasonable requirement, given that the *only* way Windows can be
> >> installed unattended (with external device drivers) is to provide those
> >> drivers on a floppy. Because, end-to-end,
> >>
> >>   do you want unattended UEFI installation with 3rd party drivers?
> >>
> >> translates to
> >>
> >>   better have a PCI-based floppy controller such that its oprom
> >>   installs an _FDI with dynamic hardware access,
> >>   *or* have your platform firmware match your floppy hardware
> >>
> >> Implementing this in QEMU would require:
> >> - inventing virt-only registers for the FDC that provide the current
> >> disk geometry,
> 
> We do secretly have these registers, but of course there's no spec to
> interpreting what they mean. For instance, what do they read when the
> drive is empty? I am not sure that information is codified in the ACPI spec.
> 
> Could be wrong, you seemed to indicate that the ACPI spec said that the
> info matches what you get from a certain legacy bios routine, but I
> don't know the specifics of that routine, either.
> 
> >> - and generating AML code that reads those registers
> >>
> >> *or*
> >>
> >> - implementing the trial and error seeking in AML
> >>
> >> Waste of time, don't do it. Microsoft have never documented their usage
> >> of _FDI. (Their forums are full of confused users whose physical floppy
> >> drives don't work under UEFI Windows!)
> >>
> >> I'm quite sure the _FDI addition in the ACPI spec is actually from
> >> Microsoft as well, but of course the *reasoning* / background for _FDI
> >> is also not public. Microsoft seem to push stuff into the ACPI spec that
> >> serves them, while conveniently ignoring non-optional parts of the spec
> >> that they don't feel like supporting (I'm looking at you,
> >> DataTableRegion). And their level of support is not public.
> >>
> >> So, the last paragraph of Roman's email
> >> <http://thread.gmane.org/gmane.comp.emulators.qemu.block/7978/focus=8081> 
> >> remains
> >> relevant -- do whatever ugly static hack is necessary in QEMU's AML
> >> generator to restore the one use case to working state that matters:
> >> unattended installation to a virtio disk.
> >>
> >> Noone in their right mind uses floppy in the guest interactively, and
> >> even for the unattended installation, floppy is used only because
> >> Windows is too stupid to work off a CD-ROM fully automatically. (Where
> >> everything one needs would be interrogable from on hardware / ATAPI / 
> >> SCSI.)
> >>
> >> IMHO, do the *absolute minimum* to adapt this AML generation patch to
> >> John's FDC rework, and ignore all dynamic aspects (like media change).
> >>
> >> Thanks
> >> Laszlo
> > 
> > I'm fine with reporting some static values that make windows work fine.
> > What I think is wrong is reporting geometry that happens to
> > match current windows.
> > 
> 
> Can you rephrase that last sentence? I don't follow, I'm sorry.
> 
> 
> Roman, does the 0xFF "empty disk geometry" hack appear to work for
> Windows 10?
> 
> Maybe it's fine enough as-is, as per Laszlo's good synopsis here.
> 
> --js

Right. What I would like to avoid, if possible, is reporting the current
data for whatever diskette happens to be used at machine_done
time.  This just makes things fragile if you switch
to a different one.

-- 
MST