On Mon, Sep 30, 2024 at 2:34 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 29.09.24 05:27, Xiang Gao wrote:
> > From: gaoxiang17 <gaoxiang17@xxxxxxxxxx>
> >
>
> We should add here
>
> "To debug CMA allocations (especially failing ones), it is valuable to
> know the state of CMA: how many pages out of the total ones are
> allocated, and how many were requested to be allocated. Let's print
> some more information."
>
> I assume Andrew can fix that up when applying.
>
> > before:
> > [ 24.407814] cma: cma_alloc(cma (____ptrval____), name: reserved, count 1, align 0)
> > [ 24.413397] cma: cma_alloc(cma (____ptrval____), name: reserved, count 1, align 0)
> > [ 24.415886] cma: cma_alloc(cma (____ptrval____), name: reserved, count 1, align 0)
> >
> > after:
> > [ 24.069738] cma: cma_alloc(cma (____ptrval____), name: reserved, total pages: 16384, used pages: 64, request pages: 1, align 0)
> > [ 24.075317] cma: cma_alloc(cma (____ptrval____), name: reserved, total pages: 16384, used pages: 65, request pages: 1, align 0)
> > [ 24.078455] cma: cma_alloc(cma (____ptrval____), name: reserved, total pages: 16384, used pages: 66, request pages: 1, align 0)
> >
> > Signed-off-by: gaoxiang17 <gaoxiang17@xxxxxxxxxx>
> > ---
> > mm/cma.c | 15 +++++++++++++--
> > 1 file changed, 13 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/cma.c b/mm/cma.c
> > index 2d9fae939283..90b3fdbac19c 100644
> > --- a/mm/cma.c
> > +++ b/mm/cma.c
> > @@ -403,6 +403,17 @@ static void cma_debug_show_areas(struct cma *cma)
> > spin_unlock_irq(&cma->lock);
> > }
> >
> > +static unsigned long cma_get_used_pages(struct cma *cma)
> > +{
> > + unsigned long used;
> > +
> > + spin_lock_irq(&cma->lock);
> > + used = bitmap_weight(cma->bitmap, (int)cma_bitmap_maxno(cma));
> > + spin_unlock_irq(&cma->lock);
>
> This adds overhead to each allocation, even if debug outputs are ignored
> I assume?
>
> I wonder if we'd want to print these details only when our allocation
> failed?
>
> Alternatively, we could actually track how many pages are allocated in
> the cma, so we don't have to traverse the complete bitmap on every
> allocation.
>
Yep, that's what I did as part of
https://lore.kernel.org/all/20240724124845.614c03ad39f8af3729cebee6@xxxxxxxxxxxxxxxxxxxx/T/
That patch didn't make it in (yet). I'm happy for it to be combined
with this one if that's easier.
- Frank
