READ_DMA (0xC8) command does not work correctly

Bug #1608802 reported by Stefan Weil on 2016-08-02
6
This bug affects 1 person
Affects Status Importance Assigned to Milestone
QEMU
Undecided
Unassigned

Bug Description

The QEMU PC emulation of DMA does not behave like real hardware or other virtualization software.

From the original bug report (Benjamin David Lunt):

    Back to the READ_DMA command, it is my conclusion that the
    READ_DMA command, more precisely, the BUS Master part of QEMU is
    in error. The tests that people have done to see if it works, is
    probably the guest finding out that DMA doesn't work and defaulting
    to PIO, but since the read was successful visually to the user, the
    user assumed the READ_DMA command works, where the guest actually
    defaulted back to PIO transfers without notice.

    My code works on real hardware (numerous machines), Bochs, and Oracle's
    Virtual Box.

    ...

    I have a small test suite, zipped and included at:
    www.fysnet.net/temp/c8bug.zip

    Within this zip file is a.img. This is a freeDOS bootable
    floppy. Emulate it with QEMU and then at the DOS prompt, run
    c8bug.exe.

    It will say that the drive is not ready.
     "Drive never became ready"
    (along with a few other lines of text)

    The source for this test suite is also included.
     c8bug.c is the c source code
     c8bug.h is the header file
     ctype.h is a quick way to get 'bit8u' type defines
     timer.h is a delay routine from another project I have
    (The base IO addresses are assumed and set at the top of c8bug.c)
    (compiled with DJGPP for DPMI)

    q.bat is my command line for QEMU

    On all other machines and VirtualBox, the controller is ready
    for me to receive the sector data.
     "Drive is ready to transmit data..."

    However, in QEMU, the controller never becomes ready.
     "Drive never became ready"

The bug was reported for QEMU for Windows, but I can confirm that QEMU for Linux also shows that behaviour, while VirtualBox works.

description: updated
Changed in qemu:
status: New → Confirmed
Kevin Wolf (kwolf-redhat) wrote :

Okay, I think there are two parts to this, and both the test case and qemu are
buggy, though perhaps the test case more seriously than qemu.

The ATA spec describes DMA commands as having two phases, preparation and
transfer. During the first phase, BSY=1,DRQ=0 is required. In the second phase,
the device can use the same status (BSY=1,DRQ=0) or switch to BSY=0,DRQ=1.
I couldn't find clear information on which state the device should be in between
receiving the ATA command and enabling the bus master operation, which is what
the test case checks, but it must be one of the two.

So the reason that the test case is buggy is that it waits for DRQ to be set,
while the spec only guarantees that either DRQ or BSY is set during the
transfer. It doesn't work on hardware that sets BSY rather than DRQ. (It also
means that the test expects the device to be in the transfer state before the
bus master is enabled.)

The problem with qemu is that it sets _both_ bits, whereas the spec requires
that only one of them is set. Apparently most drivers don't care about this as
long as any of the bits is set, but technically it's wrong.

To post a comment you must log in.
This report contains Public information  Edit
Everyone can see this information.

Other bug subscribers