Re: [PATCH v2 09/17] powerpc/mm: Add new hash_page_mm()

From: Michael Neuling
Date: Thu Oct 02 2014 - 03:39:32 EST


On Thu, 2014-10-02 at 13:48 +1000, Michael Ellerman wrote:
> On Tue, 2014-30-09 at 10:34:58 UTC, Michael Neuling wrote:
> > From: Ian Munsie <imunsie@xxxxxxxxxxx>
> >
> > This adds a new function hash_page_mm() based on the existing hash_page().
> > This version allows any struct mm to be passed in, rather than assuming
> > current. This is useful for servicing co-processor faults which are not in the
> > context of the current running process.
>
> I'm not a big fan. hash_page() is already a train wreck, and this doesn't make
> it any better.

I can document it to make the situation a bit better. It's certainly
not clear which one to use here and under what circumstances. It's
basically ask benh territory.

> > diff --git a/arch/powerpc/mm/hash_utils_64.c b/arch/powerpc/mm/hash_utils_64.c
> > index bbdb054..0a5c8c0 100644
> > --- a/arch/powerpc/mm/hash_utils_64.c
> > +++ b/arch/powerpc/mm/hash_utils_64.c
> > @@ -904,7 +904,7 @@ void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
> > return;
> > slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
> > copro_flush_all_slbs(mm);
> > - if (get_paca_psize(addr) != MMU_PAGE_4K) {
> > + if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
> > get_paca()->context = mm->context;
> > slb_flush_and_rebolt();
>
> This is a bit fishy.
>
> If that mm is currently running on another cpu you just failed to update it's
> paca. But I think the call to check_paca_psize() in hash_page() will save you
> on that cpu.
>
> In fact we might be able to remove that synchronisation from
> demote_segment_4k() and always leave it up to check_paca_psize()?

Aneesh asked the same thing for v1 and we convinced ourselves it was ok.
I said this at the time...

I had a chat to benh offline about this and he thinks it's fine. A
running process in the same mm context will either have hit this mapping
or not. If it's hit it, the page will be invalidated and it'll come in
via hash_page and have it's segment demoted also (and paca updated). If
it hasn't hit, again it'll come into hash_page() and get demoted also.

> > @@ -989,26 +989,24 @@ static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
> > * -1 - critical hash insertion error
> > * -2 - access not permitted by subpage protection mechanism
> > */
> > -int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
> > +int hash_page_mm(struct mm_struct *mm, unsigned long ea, unsigned long access, unsigned long trap)
> > {
> > enum ctx_state prev_state = exception_enter();
> > pgd_t *pgdir;
> > unsigned long vsid;
> > - struct mm_struct *mm;
> > pte_t *ptep;
> > unsigned hugeshift;
> > const struct cpumask *tmp;
> > int rc, user_region = 0, local = 0;
> > int psize, ssize;
> >
> > - DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
> > - ea, access, trap);
> > + DBG_LOW("%s(ea=%016lx, access=%lx, trap=%lx\n",
> > + __func__, ea, access, trap);
> >
> > /* Get region & vsid */
> > switch (REGION_ID(ea)) {
> > case USER_REGION_ID:
> > user_region = 1;
> > - mm = current->mm;
> > if (! mm) {
> > DBG_LOW(" user region with no mm !\n");
> > rc = 1;
>
> What about the VMALLOC case where we do:
> mm = &init_mm;
>
> Is that what you want? It seems odd that you pass an mm to the routine, but
> then potentially it ends up using a different mm after all depending on the
> address.

Good point. We have hash_page() still. I can make that check in there
and decide which mm to use and pass that to hash_page_mm(). Then we
always use mm in hash_page_mm(). hash_page() will then look like this:

int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
{
struct mm_struct *mm = current->mm;

if (REGION_ID(ea) == VMALLOC_REGION_ID)
mm = &init_mm;

return hash_page_mm(mm, ea, access, trap);
}

Mikey
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/