re: kern/54486: athn driver panics on arm

[matthew green <mrg%eterna.com.au@localhost>]

>  Looks like gcc is optimizing the memcpy() call on line 2528 of
>  src/sys/dev/usb/if_athn_usb.c into an unaligned 32-bit store:
>  
>                  txf = (struct ar_tx_frame *)&htc[1];
>                  memset(txf, 0, sizeof(*txf));
>                  txf->data_type = AR_HTC_NORMAL;
>                  txf->node_idx = sta_index;
>                  txf->vif_idx = 0;
>                  txf->tid = tid;
>  
>  The fields at the beginning and end of *txf are being assigned to, so
>  only the middle part really needs the memset()ting, and gcc apparently
>  figures the way to do that is to store a 32-bit word of zeros there,
>  at an unaligned address.
>  
>  I would not be surprised if this optimizaion is also breaking other
>  parts of the arm kernel.  Does anyone know the right way to fix this?

this looks similar to the problem worked around in if_axe.c:

   http://mail-index.netbsd.org/source-changes/2019/01/06/msg102106.html

i forget the status of a real fix.  there's some issue with
unaligned on v6/v7, IIRC.

but i think this one is a code bug.

the axe case has a structure with two uint16_t's in it, and
assumes they will be 32 bit aligned.  that's wrong.

however, in this case, struct ar_tx_frame has 4 x uint8_ts
and then a uint32_t.

that means the compiler is probably right to assume that a
pointer to this structure will be aligned for uint32_t, so
in this case, the compiler seems to be not wrong, and we 
should fix the code.


.mrg.