Re: [PATCH v4 00/20] Add Allwinner H3 SoC and Orange Pi PC Machine

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 2/17/20 9:27 PM, Niek Linnenbank wrote:
> Hi Philippe,
> 
> On Wed, Feb 12, 2020 at 11:12 PM Philippe Mathieu-Daudé 
> <philmd@redhat.com <mailto:philmd@redhat.com>> wrote:
> 
>     On 2/12/20 10:47 PM, Niek Linnenbank wrote:
>      > Hi all,
>      >
>      > Short status update regarding this series.
>      >
>      > Currently I am debugging booting NetBSD 9.0-RC2, which is recently
>      > released [1]
>      > The problem is that, apparently, RC2 can't mount the rootfs properly:
>      >
>      > [   3.1871510] vfs_mountroot: can't open root device
>      > [   3.2141370] cannot mount root, error = 6
>      >
>      > This only happens using the RC2 filesystem image. However, the RC2
>      > kernel can
>      > properly load & boot the RC1 filesystem image, and the RC1 kernel
>     gives
>      > the same error
>      > on the RC2 filesystem.  But I dont think its a NetBSD-RC2 issue,
>     because
>      > on a real
>      > Orange Pi PC hardware board, this problem does not show.
> 
>     Linux shows:
>         console: mmc0: Problem switching card into high-speed mode!
> 
>     QEMU MMC support is not in good shape, and high-speed SD card
>     support is
>     limited. I keep procrastinating at reviewing your SDHCI patch hmmmmm.
> 
>     So you added:
> 
>        * DMA transfers
>        * Direct FIFO I/O
>        * Short/Long format command responses
> 
>     I'd start diffing the trace output of the following events:
> 
>     allwinner_sdhost*
>     sdcard*
> 
>     with both kernels.
> 
> Thanks for the kind suggestions Philippe.
> Indeed, comparing the trace files of both the RC1 and RC2 kernels is one 
> of the things I did, and many more.
> 
> After extensive low-level debugging, I discovered that the issue is much 
> more simple than I thought.
> In particular, when using -sd <file>, the emulated device gets the same 
> physical size as the file.
> Normally this is not a problem, but for the NetBSD RC2 image, the kernel 
> reads the MBR partition table,
> compares it with the device size and concludes that the NetBSD partition 
> in the image is larger than the actual device.
> 
> Unfortunately, this root cause is not printed on the NetBSD console, 
> only the final  'can't open root device' message.
> Also, when running on hardware, obviously the SD card will be larger, 
> e.g. 4GB or more, so that is why this issue does not show.
> 
> So the fix is to extend the input image by a few megabytes before 
> booting it.
> And on the positive side, with this search we now have more confidence 
> that the
> emulated SD/MMC device in Qemu works as expected.

Good news!

IIRC from the specs, cards are block devices and the only alignment 
required is the size of a block (512KiB for your 4GiB card).

That said I never saw a card not pow2 aligned, but the card firmware 
should be able to discard blocks and announce fewer. Maybe FreeBSD is 
incorrect assuming a pow2 alignment?

> 
> Kind regards,
> Niek
> 
>      >
>      > I'm comparing traces and adding more low-level debug output to
>     NetBSD
>      > and QEMU to find the issue.
>      >
>      > Also I processed the review remarks which were send so far.
>      >
>      > Kind regards,
>      > Niek
>      >
>      > [1] https://www.netbsd.org/releases/formal-9/NetBSD-9.0.html
> 
> 
> 
> -- 
> Niek Linnenbank
>