TFTP hangs with fragmented IP packets

[Tim Harvey]

On Tue, Apr 5, 2022 at 4:11 PM Laurent Pinchart
<laurent.pinchart at ideasonboard.com> wrote:
>
> Hi Ramon,
>
> On Sun, Apr 03, 2022 at 02:17:26AM +0300, Ramon Fried wrote:
> > On Thu, Mar 31, 2022 at 8:43 AM Ramon Fried wrote:
> > > On Tue, Mar 29, 2022 at 11:28 PM Laurent Pinchart wrote:
> > > >
> > > > Hello,
> > > >
> > > > I've banged my head a few days ago trying to debug an issue with a TFTP
> > > > transfer hanging in the middle.
> > > >
> > > > I'm testing U-Boot 2022-rc5 on a Toradex Verdin i.MX8MP module (using
> > > > the verdin-imx8mp defconfig). My local network MTU is 1500 bytes, and
> > > > the board uses the EQoS ethernet controller.
> > > >
> > > > The problem started occurring after rebuilding a kernel image. U-Boot
> > > > started transferring the image, and stopped in the middle, eventually
> > > > timing out. Capture network traffic showed that U-Boot was continuously
> > > > asking for retransmit of the same block, and eventually timed out.
> > > >
> > > > U-Boot is configured with a default TFTP block size of 4096 bytes, which
> > > > results in the TFTP blocks being sent in one UDP packet split in three
> > > > IP packets. U-Boot is configured with IP fragmentation supprot enabled.
> > > > This works fine for all TFTP blocks until a paticular one in the middle
> > > > of the kernel image.
> > > >
> > > > I've narrowed it down to a file of 1472 that can't be transferred at
> > > > all (I have attached the binary to this e-mail). Changing the value of
> > > > any of the last two bytes of the file allows transferring it correctly,
> > > > so I suspect a CRC issue, likely related to IP fragmentation. Lowering
> > > > the TFTP block size to avoid fragmentation works around the problem.
> > > >
> > > > Arguably a TFTP block size of 4096 bytes should probably not be used
> > > > with a 1500 bytes MTU network, but I thought it would be useful to fix
> > > > the issue nonetheless.
> > > >
> > > > I can test patches.
> > > >
> > > Interesting.
> > > I will try to reproduce.
> > > Thanks for reporting.
> >
> > Hi.
> > I couldn't reproduce the issue with the file you provided.
> > I used sandbox64, and changed the TFTP BLOCK size to 4096.
> > I managed to TFTP the file successfully.
> >
> > => setenv autoload no
> > => setenv ethrotate no
> > => setenv ethact eth0
> > => setenv ipaddr
> > => setenv serverip 172.23.1.137
> > => env set ipaddr 172.23.1.100
> > => tftpboot 0x100000 test.bin
> > Using host_eno1 device
> > TFTP from server 172.23.1.137; our IP address is 172.23.1.100
> > Filename 'test.bin'.
> > Load address: 0x100000
> > Loading: #
> >          718.8 KiB/s
> > done
> > Bytes transferred = 1472 (5c0 hex)
> > =>
>
> Hmmmm... I wonder if it could be specific to the EQoS ethernet
> controller in the i.MX8, maybe caused by issues in hardware CRC
> calculation ? I don't have any other board I can use for testing this at
> the moment as I'm travelling.
>
> > Can you perhaps send me a network PCAP of your device failing to
> > digest the file ?
>
> I'll try to do that tomorrow.
>
> --
> Regards,
>
> Laurent Pinchart

Laurent,

I'm working with a imx8mp-venice-gw74xx [1] board with EQoS<->RGMII
PHY and am able to transfer your test file with
CONFIG_TFTP_BLOCKSIZE=4096 fine. I haven't had any issues transferring
files in general and routinely transfer files up to 513MB.

Are you running into an issue perhaps with your loadaddr position? I'm
using CONFIG_SYS_LOAD_ADDR=0x40480000.

Best Regards,

Tim
[1] https://patchwork.ozlabs.org/project/uboot/list/?series=294275