On Thu, 6 Oct 2022 13:09:08 +0100
Alexandru Elisei <alexandru.elisei@xxxxxxx> wrote:
> Hi,
>
> On Thu, Oct 06, 2022 at 01:35:52PM +0200, Claudio Imbrenda wrote:
> > On Thu, 6 Oct 2022 12:12:39 +0100
> > Alexandru Elisei <alexandru.elisei@xxxxxxx> wrote:
> >
> > > All architectures that implements virt_to_pte_phys() (s390x, x86,
> > > arm and arm64) return a physical address from the function. Teach
> > > vmalloc to treat it as such, instead of confusing the return
> > > value with a page table entry.
> >
> > I'm not sure I understand what you mean
>
> I thought that vmalloc uses PAGE_MASK because it expects
> virt_to_pte_phys() to return a pteval (because of the "pte' part in
> the virt_to_pte_phys()
I agree that the name of the function is confusing; there are comments
in lib/vmalloc.h and for virt_to_pte_phys it says:
/* Walk the page table and resolve the virtual address to a physical
address */
> function name), which might have the [PAGE_SHIFT-1:0] bits used to store
> page metadata by an architecture (like permissions), but like you've
> explained below it uses PAGE_MASK to align the page address (which is
> identically mapped) before passing it to the page allocator to be freed.
>
> >
> > > Changing things the other way around (having the function return a page
> > > table entry instead) is not feasible, because it is possible for an
> > > architecture to use the upper bits of the table entry to store metadata
> > > about the page.
> > >
> > > Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > > Cc: Thomas Huth <thuth@xxxxxxxxxx>
> > > Cc: Andrew Jones <andrew.jones@xxxxxxxxx>
> > > Cc: Laurent Vivier <lvivier@xxxxxxxxxx>
> > > Cc: Janosch Frank <frankja@xxxxxxxxxxxxx>
> > > Cc: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx>
> > > Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
> > > ---
> > > lib/vmalloc.c | 4 ++--
> > > 1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/lib/vmalloc.c b/lib/vmalloc.c
> > > index 572682576cc3..0696b5da8190 100644
> > > --- a/lib/vmalloc.c
> > > +++ b/lib/vmalloc.c
> > > @@ -169,7 +169,7 @@ static void vm_free(void *mem)
> > > /* the pointer is not page-aligned, it was a single-page allocation */
> > > if (!IS_ALIGNED((uintptr_t)mem, PAGE_SIZE)) {
> > > assert(GET_MAGIC(mem) == VM_MAGIC);
> > > - page = virt_to_pte_phys(page_root, mem) & PAGE_MASK;
> > > + page = virt_to_pte_phys(page_root, mem);
> >
> > this will break things for small allocations, though. if the pointer is
> > not aligned, then the result of virt_to_pte_phys will also not be
> > aligned....
>
> I agree, I missed that part. Would be nice if it were written using
> PAGE_ALIGN to avoid mistakes like mine in the future, but that's
PAGE_ALIGN rounds UP, though, and we need to round down.
I think it's easier and more readable to & PAGE_MASK, instead of a more
cumbersome ALIGN_DOWN((thing), PAGE_SIZE)
> unimportant.
>
> >
> > > assert(page);
> > > free_page(phys_to_virt(page));
> >
> > ...and phys_to_virt will also return an unaligned address, and
> > free_page will complain about it.
> >
> > > return;
> > > @@ -183,7 +183,7 @@ static void vm_free(void *mem)
> > > /* free all the pages including the metadata page */
> > > ptr = (uintptr_t)m & PAGE_MASK;
> >
> > ptr gets page aligned here
> >
> > > for (i = 0 ; i < m->npages + 1; i++, ptr += PAGE_SIZE) {
> > > - page = virt_to_pte_phys(page_root, (void *)ptr) & PAGE_MASK;
> > > + page = virt_to_pte_phys(page_root, (void *)ptr);
> >
> > so virt_to_pte_phys will also return an aligned address;
> > I agree that & PAGE_MASK is redundant here
>
> You are correct, if we've ended up here it means that the pointer is
> already page aligned, and it will be incremented by PAGE_SIZE each
> iteration, hence the virt_to_pte_phys() will also be paged aligned.
>
> I don't see much point in writing a patch just to remove the unnecessary
> alignment here, so I'll drop this patch entirely.
>
> Thank you for the prompt explanation!
I'm glad things have been clarified :)
>
> Alex
>
> >
> > > assert(page);
> > > free_page(phys_to_virt(page));
> > > }
> >
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
