pmap problem in FreeBSD current

[Mark Tinguely]

>  Hi Mark,
>  We've diagnosed the problem and it was related to arm_nocache_remap().
>  When the page was remaped into non-cacheable virtual memory region
>  there was no dcache write-back operation, so count of free items in
>  slab had differ value then should had.
>  This write-back operation should have been done in pmap_set_cache_entry()
>  but PVF_REF flag was missing so this part of code was omitted.
>
>  I think that we should enable PVF_REF flag in pmap_enter_pv().
>
>  ================================================
>
>  diff --git a/sys/arm/arm/pmap.c b/sys/arm/arm/pmap.c
>  index 3cdab65..133dc6d 100644
>  --- a/sys/arm/arm/pmap.c
>  +++ b/sys/arm/arm/pmap.c
>  @@ -1683,7 +1683,7 @@ pmap_enter_pv(struct vm_page *pg, struct pv_entry 
>  *pve, pmap_t pm,
>  /* PMAP_ASSERT_LOCKED(pmap_kernel()); */
>  pve->pv_pmap = pmap_kernel();
>  pve->pv_va = pg->md.pv_kva;
>  - pve->pv_flags = PVF_WRITE | PVF_UNMAN;
>  + pve->pv_flags = PVF_WRITE | PVF_UNMAN | PVF_REF;
>  pg->md.pv_kva = 0;
>
>  TAILQ_INSERT_HEAD(&pg->md.pv_list, pve, pv_list);
>
>  =================================================
>
>  With this patch I've had no panics.
>
>  Many thanks,
>  Michał Hajduk
>

Good job. This strange, your inclusion writes back the whole page of the
original buffer. This writeback of just the original buffer size should be
done in bus_dmamem_alloc() in busdma_machdep.c after the arm_remap_nocache()
with the line "cpu_idcache_wbinv_range(vm_offset_t)*vaddr, dmat->maxsize)".

I wonder if your cpu_idcache_wbinv_range() implementation correctly
wb/invalidates when the buffer is not cache line aligned or is smaller
than a cache-line? The other situation that would do this is if your
cache line larger than you think. I know the ARMv7 use 64 byte caches
instead of the more traditional 32 bytes. This should not be your case
because last week I looked up your processor information and thought it
said it used a 32 byte cache line.

Did http://www.casselton.net/~tinguely/arm_pmap_unmanaged.diff fix the lock
problems?

Thank-you for the test information and the code scouting.

--Mark Tinguely.