On Fri, Dec 04, 2020 at 08:57:37AM +0100, Helge Deller wrote:
> On 12/4/20 4:48 AM, Matthew Wilcox wrote:
> > On Thu, Dec 03, 2020 at 04:33:10PM -0800, James Bottomley wrote:
> >> These platforms define PMD_ORDER in asm/pgtable.h
> >
> > I think that's the real problem, though.
> >
> > #define PGD_ORDER 1 /* Number of pages per pgd */
> > #define PMD_ORDER 1 /* Number of pages per pmd */
> > #define PGD_ALLOC_ORDER (2 + 1) /* first pgd contains pmd */
> > #else
> > #define PGD_ORDER 1 /* Number of pages per pgd */
> > #define PGD_ALLOC_ORDER (PGD_ORDER + 1)
> >
> > That should clearly be PMD_ALLOC_ORDER, not PMD_ORDER. Or even
> > PAGES_PER_PMD like the comment calls it, because I really think
> > that doing an order-3 (8 pages) allocation for the PGD is wrong.
>
> We need a spinlock to protect parallel accesses to the PGD,
> search for pgd_spinlock().
> This spinlock is stored behind the memory used for the PGD, which
> is why we allocate more memory (and waste 3 pages).
There are better places to store it than that! For example, storing it
in the struct page, like many architectures do for split ptlocks.
You'll have to skip the _pt_pad_1 so it doesn't get confused with
being a compound_head, but soemthing like this:
struct { /* PA-RISC PGD */
unsigned long _pa_pad_1; /* compound_head */
#if ALLOC_SPLIT_PA_PTLOCKS
spinlock_t *pa_ptl;
#else
spinlock_t pa_ptl;
#endif
};
inside struct page (linux/mm_types.h) should do the trick. You'll
still need to allocate them separately if various debugging options
are enabled (see the ALLOC_SPLIT_PTLOCKS for details), but usually
this will save you a lot of memory.
You could also fill in pt_mm like x86 does for its pgds, and then use
mm->page_table_lock to protect whatever the PGD lock currently protects.
Maybe page_table_lock is sometimes held when calling ptep_set_wrprotect()
and sometimes isn't; then this wouldn't work.
Also, could you fix the comments? They don't match the code:
#define PGD_ORDER 1 /* Number of pages per pgd */
should be
#define PGD_ALLOC_ORDER 1 /* 2 pages (8KiB) per pgd */
_______________________________________________
Linux-nvdimm mailing list -- [email protected]
To unsubscribe send an email to [email protected]
