[U-Boot] netconsole: USB Ethernet connection dropping with ping or tftpboot

["Jörg Krause" <jkrause@posteo.de>]

On Do, 2015-02-05 at 15:23 -0700, Stephen Warren wrote:
> On 02/05/2015 03:10 PM, J?rg Krause wrote:
> > On Do, 2015-02-05 at 08:33 -0700, Stephen Warren wrote:
> >> On 02/05/2015 04:21 AM, J?rg Krause wrote:
> ...
> >>> This reminded me about an issue I had some months ago:
> >>> http://lists.denx.de/pipermail/u-boot/2014-July/182885.html
> >>>
> >>> I enabled debug output in arch/arm/cpu/arm926ejs/cache.c and I get
> >>> error:
> >>> => tftp u-boot.sb
> >>> using ci_udc, OUT ep- IN ep- STATUS ep-
> >>> CACHE: Misaligned operation at range [43f7b010, 43f7b070]
> >>>
> >>> I removed the flush_cache() call in cmd_net.c:netboot_common() as
> >>> suggested by Marek in the thread. But the error message is still there.
> >>
> >> Perhaps make flush_cache() a macro that also passes in the file/line
> >> number where it's called from, and print those out along with he
> >> "misaligned operation" error message?
> >>
> >> (or find some other way to perform a stack dump from within that function).
> >
> > I've added in each function in ci_udc a printf statement before a cache
> > function is called, eg:
> >
> >          static void ci_flush_qh(int ep_num)
> >          {
> >          	[..]
> >          	printf("CACHE: flush_dcache_range %s %d
> >          \n",__FUNCTION__,__LINE__);
> >          	flush_dcache_range(start, end);
> >          }
> >
> > I've also looked at all calling functions of flush_cache or
> > flush_dcache_range, but I think there is nothing of relevance.
> >
> > This is a snippet of the output:
> >
> ...
> >          CACHE: flush_dcache_range ci_bounce 356
> >          CACHE: Misaligned operation at range [43f7b010, 43f7b070]
> >          timeout sending packets to usb ethernet
> >          ping failed; host 10.0.0.1 is not alive
> 
> Which git commit did you build, and which board?

Building tag v2015.01 for a custom i.MX28 board which is not upstream.
It's configuration mainly based on mx28evk and m28evk.

> 
> I'm curious what values you have for ARCH_DMA_MINALIGN and 
> CONFIG_SYS_CACHELINE_SIZE.

I defined CONFIG_SYS_CACHELINE_SIZE 32 in the config. ARCH_DMA_MINALIGN
is also 32.

> 
> Are you sure flush_dcache_range() is the code printing the alignment 
> error, not e.g. invalidate_dcache_range()?

I've also added a printf statement to all functions in ci_udc which
calls invalidate_dcache_range(), but it is not logged when the alignment
error occurs. e.g.

        using ci_udc, OUT ep- IN ep- STATUS ep-
        MAC 00:19:b8:00:00:02
        HOST MAC 00:19:b8:00:00:01
        CACHE: flush_dcache_range ci_flush_qh 166 
        CACHE: invalidate_dcache_range ci_invalidate_qh 182 
        CACHE: flush_dcache_range ci_bounce 356 
        CACHE: flush_dcache_range ci_flush_qtd 199 
        CACHE: flush_dcache_range ci_flush_qh 166 

> 
> The reason I ask most of these questions is that line 356 mentioned in 
> your debug spew is in function ci_debounce() in the code I have; no 
> doubt I have a different git commit than you have checked out, and the 
> debug printfs you added changed the line numbers too.

You're right, the debug printfs changed the line number. It's the
flush_dcache_range() call at line 348 in ci_bounce(). I checked this
after removing all printfs.

> About the only thing I can think of is that:
> 
> a) You're not using an up-to-date ci_udc.c; IIRC I fixed quite a few 
> cache issues when I reworked it a while back.

Yes, we had a troubleshooting about this last summer. I had problems
with timeouts using tftp. Marek and you worked on this issue.

> 
> b) In ci_bounce(), the bounce buffer is only allocated if the 
> user-buffer is already aligned, and if a large-enough bounce buffer 
> wasn't previously allocated. If ci_req->b_buf was uninitialized it could 
> be non-zero (thus preventing the expected aligned allocation) yet not 
> actually aligned enough.

Maybe, we should work on this?

> 
> c) Perhaps ARCH_DMA_MINALIGN or CONFIG_SYS_CACHELINE_SIZE aren't correct 
> or are inconsistent.
> 
> Ah. I note that check_cache_range() in either arch/arm/cpu/arm1136/cpu.c 
> or arch/arm/cpu/arm926ejs/cache.c uses CONFIG_SYS_CACHELINE_SIZE to 
> check for alignment, whereas ci_udc.c uses ARCH_DMA_MINALIGN. 
> Inconsistency between those two could be at fault.

Both are set to 32, so this should not be the problem.