Re: [PATCH] mm/damon/vaddr: Safely walk page table

[David Hildenbrand]

On 31.08.21 12:49, SeongJae Park wrote:
> From: SeongJae Park <sjpark@xxxxxxxxx>
>
> On Tue, 31 Aug 2021 11:53:05 +0200 David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> > On 27.08.21 17:04, SeongJae Park wrote:
> > > From: SeongJae Park <sjpark@xxxxxxxxx>
> > >
> > > Commit d7f647622761 ("mm/damon: implement primitives for the virtual
> > > memory address spaces") of linux-mm[1] tries to find PTE or PMD for
> > > arbitrary virtual address using 'follow_invalidate_pte()' without proper
> > > locking[2].  This commit fixes the issue by using another page table
> > > walk function for more general use case under proper locking.
> > >
> > > [1] https://github.com/hnaz/linux-mm/commit/d7f647622761
> > > [2] https://lore.kernel.org/linux-mm/3b094493-9c1e-6024-bfd5-7eca66399b7e@xxxxxxxxxx
> > >
> > > Fixes: d7f647622761 ("mm/damon: implement primitives for the virtual memory address spaces")
> > > Reported-by: David Hildenbrand <david@xxxxxxxxxx>
> > > Signed-off-by: SeongJae Park <sjpark@xxxxxxxxx>
> > > ---
> > >    mm/damon/vaddr.c | 81 +++++++++++++++++++++++++++++++++++++++++++-----
> > >    1 file changed, 74 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
> > > index 230db7413278..b3677f2ef54b 100644
> > > --- a/mm/damon/vaddr.c
> > > +++ b/mm/damon/vaddr.c
> > > @@ -8,10 +8,12 @@
> > >    #define pr_fmt(fmt) "damon-va: " fmt
> > >    
> > >    #include <linux/damon.h>
> > > +#include <linux/hugetlb.h>
> > >    #include <linux/mm.h>
> > >    #include <linux/mmu_notifier.h>
> > >    #include <linux/highmem.h>
> > >    #include <linux/page_idle.h>
> > > +#include <linux/pagewalk.h>
> > >    #include <linux/random.h>
> > >    #include <linux/sched/mm.h>
> > >    #include <linux/slab.h>
> > > @@ -446,14 +448,69 @@ static void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm,
> > >    #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
> > >    }
> > >    
> > > +struct damon_walk_private {
> > > +	pmd_t *pmd;
> > > +	pte_t *pte;
> > > +	spinlock_t *ptl;
> > > +};
> > > +
> > > +static int damon_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next,
> > > +		struct mm_walk *walk)
> > > +{
> > > +	struct damon_walk_private *priv = walk->private;
> > > +
> > > +	if (pmd_huge(*pmd)) {
> > > +		priv->ptl = pmd_lock(walk->mm, pmd);
> > > +		if (pmd_huge(*pmd)) {
> > > +			priv->pmd = pmd;
> > > +			return 0;
> > > +		}
> > > +		spin_unlock(priv->ptl);
> > > +	}
> > > +
> > > +	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
> > > +		return -EINVAL;
> > > +	priv->pte = pte_offset_map_lock(walk->mm, pmd, addr, &priv->ptl);
> > > +	if (!pte_present(*priv->pte)) {
> > > +		pte_unmap_unlock(priv->pte, priv->ptl);
> > > +		priv->pte = NULL;
> > > +		return -EINVAL;
> > > +	}
> > > +	return 0;
> > > +}
> > > +
> > > +static struct mm_walk_ops damon_walk_ops = {
> > > +	.pmd_entry = damon_pmd_entry,
> > > +};
> > > +
> > > +int damon_follow_pte_pmd(struct mm_struct *mm, unsigned long addr,
> > > +		struct damon_walk_private *private)
> > > +{
> > > +	int rc;
> > > +
> > > +	private->pte = NULL;
> > > +	private->pmd = NULL;
> > > +	rc = walk_page_range(mm, addr, addr + 1, &damon_walk_ops, private);
> > > +	if (!rc && !private->pte && !private->pmd)
> > > +		return -EINVAL;
> > > +	return rc;
> > > +}
> > > +
> > >    static void damon_va_mkold(struct mm_struct *mm, unsigned long addr)
> > >    {
> > > -	pte_t *pte = NULL;
> > > -	pmd_t *pmd = NULL;
> > > +	struct damon_walk_private walk_result;
> > > +	pte_t *pte;
> > > +	pmd_t *pmd;
> > >    	spinlock_t *ptl;
> > >    
> > > -	if (follow_invalidate_pte(mm, addr, NULL, &pte, &pmd, &ptl))
> > > +	mmap_write_lock(mm);
> >
> > Can you elaborate why mmap_read_lock() isn't sufficient for your use
> > case? The write mode might heavily affect damon performance and workload
> > impact.
>
> Because as you also mentioned in the previous mail, 'we can walk page tables
> ignoring VMAs with the mmap semaphore held in write mode', and in this case we
> don't know to which VMA the address is belong.  I thought the link to the mail
> can help people understanding the reason.  But, as you are suggesting, I now
> think putting an elaborated explanation here would be much better.  I will also
> put a warning for the possible performance impact.

walk_page_range() make sure to skip any VMA holes and only walks ranges 
within VMAs. With the mmap sem in read mode, the VMA layout (mostly) 
cannot change, so calling walk_page_range() is fine. So pagewalk.c 
properly takes care of VMAs.

As an example, take a look at MADV_COLD handling in mm/madvise.c.

madvise_need_mmap_write() returns "0", and we end up calling 
madvise_cold()->...->walk_page_range() with mmap_lock_read().

You can exclude any VMAs you don't care about in the test_walk() 
callback, if required.

--
Thanks,

David / dhildenb