Re: [PATCH 3/5] mm/mremap: use pmd_addr_end to calculate next in move_page_tables()

[Russell King - ARM Linux admin]

On Thu, Jan 30, 2020 at 09:30:00AM +0800, Wei Yang wrote:
> On Wed, Jan 29, 2020 at 11:24:41PM +0000, Russell King - ARM Linux admin wrote:
> >On Thu, Jan 30, 2020 at 05:57:45AM +0800, Wei Yang wrote:
> >> On Wed, Jan 29, 2020 at 09:47:38AM +0000, Russell King - ARM Linux admin wrote:
> >> >On Sun, Jan 26, 2020 at 05:47:57PM +0300, Dmitry Osipenko wrote:
> >> >> 18.01.2020 02:22, Wei Yang ÐÐÑÐÑ:
> >> >> > Use the general helper instead of do it by hand.
> >> >> > 
> >> >> > Signed-off-by: Wei Yang <richardw.yang@xxxxxxxxxxxxxxx>
> >> >> > ---
> >> >> >  mm/mremap.c | 7 ++-----
> >> >> >  1 file changed, 2 insertions(+), 5 deletions(-)
> >> >> > 
> >> >> > diff --git a/mm/mremap.c b/mm/mremap.c
> >> >> > index c2af8ba4ba43..a258914f3ee1 100644
> >> >> > --- a/mm/mremap.c
> >> >> > +++ b/mm/mremap.c
> >> >> > @@ -253,11 +253,8 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >> >> >  
> >> >> >  	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
> >> >> >  		cond_resched();
> >> >> > -		next = (old_addr + PMD_SIZE) & PMD_MASK;
> >> >> > -		/* even if next overflowed, extent below will be ok */
> >> >> > +		next = pmd_addr_end(old_addr, old_end);
> >> >> >  		extent = next - old_addr;
> >> >> > -		if (extent > old_end - old_addr)
> >> >> > -			extent = old_end - old_addr;
> >> >> >  		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
> >> >> >  		if (!old_pmd)
> >> >> >  			continue;
> >> >> > @@ -301,7 +298,7 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >> >> >  
> >> >> >  		if (pte_alloc(new_vma->vm_mm, new_pmd))
> >> >> >  			break;
> >> >> > -		next = (new_addr + PMD_SIZE) & PMD_MASK;
> >> >> > +		next = pmd_addr_end(new_addr, new_addr + len);
> >> >> >  		if (extent > next - new_addr)
> >> >> >  			extent = next - new_addr;
> >> >> >  		move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
> >> >> > 
> >> >> 
> >> >> Hello Wei,
> >> >> 
> >> >> Starting with next-20200122, I'm seeing the following in KMSG on NVIDIA
> >> >> Tegra (ARM32):
> >> >> 
> >> >>   BUG: Bad rss-counter state mm:(ptrval) type:MM_ANONPAGES val:190
> >> >> 
> >> >> and eventually kernel hangs.
> >> >> 
> >> >> Git's bisection points to this patch and reverting it helps. Please fix,
> >> >> thanks in advance.
> >> >
> >> >The above is definitely wrong - pXX_addr_end() are designed to be used
> >> >with an address index within the pXX table table and the address index
> >> >of either the last entry in the same pXX table or the beginning of the
> >> >_next_ pXX table.  Arbitary end address indicies are not allowed.
> >> >
> >> 
> >> #define pmd_addr_end(addr, end)						\
> >> ({	unsigned long __boundary = ((addr) + PMD_SIZE) & PMD_MASK;	\
> >> 	(__boundary - 1 < (end) - 1)? __boundary: (end);		\
> >> })
> >> 
> >> If my understanding is correct, the definition here align the addr to next PMD
> >> boundary or end.
> >> 
> >> I don't see the possibility to across another PMD. Do I miss something?
> >
> >Look at the definition of p*_addr_end() that are used when page tables
> >are rolled up.
> >
> 
> Sorry, I don't get your point.
> 
> What's the meaning of "roll up" here?
> 
> Would you mind giving me an example? I see pmd_addr_end() is not used in many
> places in core kernel. By glancing those usages, all the places use it like
> pmd_addr_end(addr, end). Seems no specially handing on the end address.
> 
> Or you mean the case when pmd_addr_end() is defined to return "end" directly? 

Not all hardware has five levels of page tables.  When hardware does not
have five levels, it is common to "roll up" some of the page tables into
others.

There are generic ways to implement this, which include using:

include/asm-generic/pgtable-nop4d.h
include/asm-generic/pgtable-nopud.h
include/asm-generic/pgtable-nopmd.h

and then there's architecture ways to implement this.  32-bit ARM takes
its implementation for PMD not from the generic version, which
post-dates 32-bit ARM, but from how page table roll-up was implemented
back at the time when the current ARM scheme was devised.  The generic
scheme is unsuitable for 32-bit ARM since we do more than just roll-up
page tables, but this is irrelevent for this discussion.

All three of the generic implementations, and 32-bit ARM, define the
pXX_addr_end() macros thusly:

include/asm-generic/pgtable-nop4d.h:#define p4d_addr_end(addr, end) (end)
include/asm-generic/pgtable-nopmd.h:#define pmd_addr_end(addr, end) (end)
include/asm-generic/pgtable-nopud.h:#define pud_addr_end(addr, end) (end)
arch/arm/include/asm/pgtable-2level.h:#define pmd_addr_end(addr,end) (end)

since, as I stated, pXX_addr_end() expects its "end" argument to be
the address index of the next entry in the immediately upper page
table level, or the address index of the last entry we wish to
process, which ever is smaller.

If it's larger than the address index of the next entry in the
immediately upper page table level, then the effect of all these
macros will be to walk off the end of the current level of page
table.

To see how they _should_ be used, see the loops in free_pgd_range()
and the free_pXX_range() functions called from there and below.

In all cases when the pXX_addr_end() macro was introduced, what I state
above holds true - and I believe still holds true today, until this
patch that has reportedly caused issues.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up