bus_dmamap_load_mbuf_sg failures

[thomas-bsd at skibo.net]

On 2020-05-24 14:41, Thomas Skibo wrote:
> Hello.
> 
> I just tracked down a bug in a network driver on an arm64 system I am
> tinkering with.  It seems when bus_dmamap_load_mbuf_sg() fails with 
> EFBIG
> due to too many segments, things are not cleaned up and it causes the
> subsequent attempt at bus_dmamap_load_mbuf_sg() with the defragged mbuf 
> to
> fail too (EFBIG again).
> 
> I noticed that in the arm version of busdma_machdep.c, when
> _bus_dmamap_load_buffer() fails with EFBIG, it calls 
> bus_dmamap_unload() on
> exit which cleans up the dma map.  In the arm64 version of 
> busdma_bounce.c,
> when bounce_bus_dmamap_load_buffer() fails, it does not.  Is this a bug 
> or
> are we expected to call bus_dmamap_unload() on a failed 
> bus_dmamap_load* call?
> 
> I looked around at other architectures and it is inconsistent: arm and 
> mips
> do the unload but x86, arm64, and riscv do not.
> 
> --Thomas
> 
> =====
> Thomas Skibo
> thomas-bsd at skibo.net

I have filed a bug for this.  It seems only arm64 and riscv have this 
issue.
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=246852