Message-ID: <20170124225030.GC19920@node.shutemov.name>
Date:   Wed, 25 Jan 2017 01:50:30 +0300
From:   "Kirill A. Shutemov" <kirill@...temov.name>
To:     Andrew Morton <akpm@...ux-foundation.org>
Cc:     "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Andrea Arcangeli <aarcange@...hat.com>,
        Hugh Dickins <hughd@...gle.com>,
        Rik van Riel <riel@...hat.com>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 02/12] mm: introduce page_check_walk()

On Tue, Jan 24, 2017 at 01:41:22PM -0800, Andrew Morton wrote:
> On Tue, 24 Jan 2017 19:28:14 +0300 "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com> wrote:
> 
> > The patch introduce new interface to check if a page is mapped into a vma.
> > It aims to address shortcomings of page_check_address{,_transhuge}.
> > 
> > Existing interface is not able to handle PTE-mapped THPs: it only finds
> > the first PTE. The rest lefted unnoticed.
> > 
> > page_check_walk() iterates over all possible mapping of the page in the
> > vma.
> 
> I really don't like the name page_check_walk().  "check" could mean any
> damn thing.  Something like page_vma_mapped_walk() has meaning.  We
> could omit the "_walk" for brevity.

page_vma_mapped() would sound as predicate.
I'll rename to page_vma_mapped_walk().

> > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@...ux.intel.com>
> > ---
> >  include/linux/rmap.h |  65 ++++++++++++++++++++++
> >  mm/Makefile          |   6 ++-
> >  mm/huge_memory.c     |   9 ++--
> >  mm/page_check.c      | 148 +++++++++++++++++++++++++++++++++++++++++++++++++++
> >  4 files changed, 223 insertions(+), 5 deletions(-)
> >  create mode 100644 mm/page_check.c
> > 
> > diff --git a/include/linux/rmap.h b/include/linux/rmap.h
> > index 15321fb1df6b..474279810742 100644
> > --- a/include/linux/rmap.h
> > +++ b/include/linux/rmap.h
> > @@ -232,6 +232,71 @@ static inline bool page_check_address_transhuge(struct page *page,
> >  }
> >  #endif
> >  
> > +/* Avoid racy checks */
> > +#define PAGE_CHECK_WALK_SYNC		(1 << 0)
> > +/* Look for migarion entries rather than present ptes */
> > +#define PAGE_CHECK_WALK_MIGRATION	(1 << 1)
> > +
> > +struct page_check_walk {
> > +	struct page *page;
> > +	struct vm_area_struct *vma;
> > +	unsigned long address;
> > +	pmd_t *pmd;
> > +	pte_t *pte;
> > +	spinlock_t *ptl;
> > +	unsigned int flags;
> > +};
> 
> One thing which I don't think was documented is that it is the caller's
> responsibility to initialize this appropriately before calling
> page_check_walk().  At least, .pte and .ptl must be NULL, for
> page_check_walk_done().

Okay, I'll write it down.

> > +static inline void page_check_walk_done(struct page_check_walk *pcw)
> > +{
> > +	if (pcw->pte)
> > +		pte_unmap(pcw->pte);
> > +	if (pcw->ptl)
> > +		spin_unlock(pcw->ptl);
> > +}
> > +
> > +bool __page_check_walk(struct page_check_walk *pcw);
> > +
> > +/**
> > + * page_check_walk - check if @pcw->page is mapped in @pcw->vma at @pcw->address
> > + * @pcw: pointer to struce page_check_walk. page, vma and address must be set.
> 
> "struct"
> 
> > + *
> > + * Returns true, if the page is mapped in the vma. @pcw->pmd and @pcw->pte point
> 
> "Returns true if"
> 
> > + * to relevant page table entries. @pcw->ptl is locked. @pcw->address is
> > + * adjusted if needed (for PTE-mapped THPs).
> > + *
> > + * If @pcw->pmd is set, but @pcw->pte is not, you have found PMD-mapped page
> 
> "is set but"
> 
> > + * (usually THP). For PTE-mapped THP, you should run page_check_walk() in 
> > + * a loop to find all PTEs that maps the THP.
> 
> "that map"
> 
> > + *
> > + * For HugeTLB pages, @pcw->pte is set to relevant page table entry regardless
> 
> "set to the relevant", "regardless of"
> 
> > + * which page table level the page mapped at. @pcw->pmd is NULL.
> 
> "the page is"
> 
> > + *
> > + * Retruns false, if there's no more page table entries for the page in the vma.
> 
> "Returns false if there are"

Thanks for corrections. Applied.

> > + * @pcw->ptl is unlocked and @pcw->pte is unmapped.
> > + *
> > + * If you need to stop the walk before page_check_walk() returned false, use
> > + * page_check_walk_done(). It will do the housekeeping.
> > + */
> > +static inline bool page_check_walk(struct page_check_walk *pcw)
> > +{
> > +	/* The only possible pmd mapping has been handled on last iteration */
> > +	if (pcw->pmd && !pcw->pte) {
> > +		page_check_walk_done(pcw);
> > +		return false;
> > +	}
> > +
> > +	/* Only for THP, seek to next pte entry makes sense */
> > +	if (pcw->pte) {
> > +		if (!PageTransHuge(pcw->page) || PageHuge(pcw->page)) {
> > +			page_check_walk_done(pcw);
> > +			return false;
> > +		}
> > +	}
> > +
> > +	return __page_check_walk(pcw);
> > +}
> 
> Was the decision to inline this a correct one?

Well, my logic was that in most cases we would have exactly one iteration.
The only case when we need more than one iteration is PTE-mapped THP which
is rare.
I hoped to avoid additional function call. Not sure if it worth it.

Should I move it inside the function?

> > --- /dev/null
> > +++ b/mm/page_check.c
> > @@ -0,0 +1,148 @@
> > +#include <linux/mm.h>
> > +#include <linux/rmap.h>
> > +#include <linux/hugetlb.h>
> > +#include <linux/swap.h>
> > +#include <linux/swapops.h>
> > +
> > +#include "internal.h"
> > +
> > +static inline bool check_pmd(struct page_check_walk *pcw)
> > +{
> > +	pmd_t pmde = *pcw->pmd;
> > +	barrier();
> > +	return pmd_present(pmde) && !pmd_trans_huge(pmde);
> > +}
> 
> Can we please have a comment explaining what the barrier() does?

I copied it from page_check_address_transhuge().

I think we can get aways with READ_ONCE() instead.
I'll add a comment.

> > +static inline bool not_found(struct page_check_walk *pcw)
> > +{
> > +	page_check_walk_done(pcw);
> > +	return false;
> > +}
> > +
> > +static inline bool map_pte(struct page_check_walk *pcw)
> > +{
> > +	pcw->pte = pte_offset_map(pcw->pmd, pcw->address);
> > +	if (!(pcw->flags & PAGE_CHECK_WALK_SYNC)) {
> > +		if (pcw->flags & PAGE_CHECK_WALK_MIGRATION) {
> > +			if (!is_swap_pte(*pcw->pte))
> > +				return false;
> > +		} else {
> > +			if (!pte_present(*pcw->pte))
> > +				return false;
> > +		}
> > +	}
> > +	pcw->ptl = pte_lockptr(pcw->vma->vm_mm, pcw->pmd);
> > +	spin_lock(pcw->ptl);
> > +	return true;
> > +}
> 
> The compiler will just ignore all these "inline" statements.

These helpers grew during development. I forgot to drop "inline".

> > +static inline bool check_pte(struct page_check_walk *pcw)
> > +{
> > +	if (pcw->flags & PAGE_CHECK_WALK_MIGRATION) {
> > +		swp_entry_t entry;
> > +		if (!is_swap_pte(*pcw->pte))
> > +			return false;
> > +		entry = pte_to_swp_entry(*pcw->pte);
> > +		if (!is_migration_entry(entry))
> > +			return false;
> > +		if (migration_entry_to_page(entry) - pcw->page >=
> > +				hpage_nr_pages(pcw->page)) {
> > +			return false;
> > +		}
> > +		if (migration_entry_to_page(entry) < pcw->page)
> > +			return false;
> > +	} else {
> > +		if (!pte_present(*pcw->pte))
> > +			return false;
> > +
> > +		/* THP can be referenced by any subpage */
> > +		if (pte_page(*pcw->pte) - pcw->page >=
> > +				hpage_nr_pages(pcw->page)) {
> > +			return false;
> > +		}
> > +		if (pte_page(*pcw->pte) < pcw->page)
> > +			return false;
> > +	}
> > +
> > +	return true;
> > +}
> 
> Thankfully, because inlining this one does seem inappropriate - it's
> enormous!
> 
> > +bool __page_check_walk(struct page_check_walk *pcw)
> > +{
> > +	struct mm_struct *mm = pcw->vma->vm_mm;
> > +	struct page *page = pcw->page;
> > +	pgd_t *pgd;
> > +	pud_t *pud;
> > +
> > +	/* For THP, seek to next pte entry */
> > +	if (pcw->pte)
> > +		goto next_pte;
> > +
> > +	if (unlikely(PageHuge(pcw->page))) {
> > +		/* when pud is not present, pte will be NULL */
> > +		pcw->pte = huge_pte_offset(mm, pcw->address);
> > +		if (!pcw->pte)
> > +			return false;
> > +
> > +		pcw->ptl = huge_pte_lockptr(page_hstate(page), mm, pcw->pte);
> > +		spin_lock(pcw->ptl);
> > +		if (!check_pte(pcw))
> > +			return not_found(pcw);
> > +		return true;
> > +	}
> > +restart:
> > +	pgd = pgd_offset(mm, pcw->address);
> > +	if (!pgd_present(*pgd))
> > +		return false;
> > +	pud = pud_offset(pgd, pcw->address);
> > +	if (!pud_present(*pud))
> > +		return false;
> > +	pcw->pmd = pmd_offset(pud, pcw->address);
> > +	if (pmd_trans_huge(*pcw->pmd)) {
> > +		pcw->ptl = pmd_lock(mm, pcw->pmd);
> > +		if (!pmd_present(*pcw->pmd))
> > +			return not_found(pcw);
> > +		if (likely(pmd_trans_huge(*pcw->pmd))) {
> > +			if (pcw->flags & PAGE_CHECK_WALK_MIGRATION)
> > +				return not_found(pcw);
> > +			if (pmd_page(*pcw->pmd) != page)
> > +				return not_found(pcw);
> > +			return true;
> > +		} else {
> > +			/* THP pmd was split under us: handle on pte level */
> > +			spin_unlock(pcw->ptl);
> > +			pcw->ptl = NULL;
> > +		}
> > +	} else {
> > +		if (!check_pmd(pcw))
> > +			return false;
> > +	}
> > +	if (!map_pte(pcw))
> > +		goto next_pte;
> > +	while (1) {
> > +		if (check_pte(pcw))
> > +			return true;
> > +next_pte:	do {
> > +			pcw->address += PAGE_SIZE;
> > +			if (pcw->address >= __vma_address(pcw->page, pcw->vma) +
> > +					hpage_nr_pages(pcw->page) * PAGE_SIZE)
> > +				return not_found(pcw);
> > +			/* Did we cross page table boundary? */
> > +			if (pcw->address % PMD_SIZE == 0) {
> > +				pte_unmap(pcw->pte);
> > +				if (pcw->ptl) {
> > +					spin_unlock(pcw->ptl);
> > +					pcw->ptl = NULL;
> > +				}
> > +				goto restart;
> > +			} else {
> > +				pcw->pte++;
> > +			}
> > +		} while (pte_none(*pcw->pte));
> > +
> > +		if (!pcw->ptl) {
> > +			pcw->ptl = pte_lockptr(mm, pcw->pmd);
> > +			spin_lock(pcw->ptl);
> > +		}
> > +	}
> > +}
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@...ck.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@...ck.org"> email@...ck.org </a>

-- 
 Kirill A. Shutemov

Powered by blists - more mailing lists