[PATCH] arm64: Fix the ptep_set_wrprotect() to set PTE_DIRTY if (PTE_DBM && !PTE_RDONLY)

[Catalin Marinas]

On Wed, Mar 09, 2016 at 05:17:39PM +0530, Ganapatrao Kulkarni wrote:
> On Wed, Mar 9, 2016 at 3:36 PM, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > On Wed, Mar 09, 2016 at 10:32:48AM +0530, Ganapatrao Kulkarni wrote:
> >> Commit 2f4b829c625e ("arm64: Add support for hardware updates of the
> >> access and dirty pte bits") introduced support for handling hardware
> >> updates of the access flag and dirty status.
> >>
> >> ptep_set_wrprotect is setting PTR_DIRTY if !PTE_RDONLY,
> >> however by design it suppose to set PTE_DIRTY
> >> only if (PTE_DBM && !PTE_RDONLY). This patch addes code to
> >> test and set accordingly.
> >
> > The reasoning behind the original code is that if !PTE_RDONLY, you have
> > no way to tell whether the page was written or not since it is already
> > writable, independent of the DBM. So by clearing the DBM bit (making the
> > page read-only), we need to ensure that a potential dirty state is
> > transferred to the software PTE_DIRTY bit.
> >
> > By checking PTE_DBM && !PTE_RDONLY, you kind of imply that you can have
> > a page with !PTE_DBM && !PTE_RDONLY. Given that PTE_DBM is actually
> > PTE_WRITE, PTE_RDONLY must always be set when !PTE_DBM. The bug may be
> > elsewhere not setting these bits correctly.
> 
> but i do see this macro,
> #define pte_hw_dirty(pte)       (pte_write(pte) && !(pte_val(pte) & PTE_RDONLY))

This was added in commit b847415ce96e ("arm64: Fix the pte_hw_dirty()
check when AF/DBM is enabled") for the pte_modify() case which is not
called on the actual PTE but a local variable. A pte passed to this
function as !PTE_DBM && !PTE_RDONLY should not be assumed dirty since
PTE_RDONLY will be set later by set_pte_at() when the actual page table
write occurs.

ptep_set_wrprotect() is run directly on the actual PTE, so here a
!PTE_RDONLY only means potentially dirty, independent of the PTE_DBM
bit. I consider the additional PTE_DBM check superfluous in this case
but we need to understand when we would actually get a pte with both
PTE_DBM and PTE_RDONLY cleared.

The only way I see this happening is if the pte doesn't have PTE_VALID
set, IOW it probably has PTE_PROT_NONE set which is used by the NUMA
balancing. So calling set_pte_at() on a !PTE_VALID && !PTE_DBM pte does
not currently set PTE_RDONLY and ptep_set_wrprotect() wrongly assumes it
is dirty.

> i dont see this issue, if i comment out arm64 implementation of
> ptep_set_wrprotect()

Because the default implementation discards any existing hw dirty
information by clearing the PTE_DBM bit and setting PTE_RDONLY via the
set_pte_at (of course, apart from the atomicity issues).

> >> This patch fixes BUG,
> >> kernel BUG at /build/linux-StrpB2/linux-4.4.0/fs/ext4/inode.c:2394!
> >> Internal error: Oops - BUG: 0 [#1] SMP
> >
> > Which bug is this? It's a PageWriteback() check in the for-next/core
> > branch. What kernel version are you using?
> 
> i am using 4.4.0

I guess with additional NUMA patches since it only fails when you enable
the NUMA_BALANCING configuration.

> > BTW, in 4.5-rc2 we pushed commit ac15bd63bbb2 ("arm64: Honour !PTE_WRITE
> > in set_pte_at() for kernel mappings"), though not sure that's what you
> > are hitting.
> 
> i have tried this patch, but issue still exist. crash log below
> 
> root at ubuntu:/home/ganapat/test# [  733.853009] kernel BUG at
> fs/ext4/inode.c:2394!

Is this the BUG_ON in page_buffers(!PagePrivate(page))? I can see in the
code above this that wrongly marking a page as dirty could have some
side effects.

Can you give this patch a try, on top of commit ac15bd63bbb2?

-------------8<----------------------

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index 7c73b365fcfa..b409a983f870 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -201,7 +201,7 @@ extern void __sync_icache_dcache(pte_t pteval, unsigned long addr);
 static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
 			      pte_t *ptep, pte_t pte)
 {
-	if (pte_valid(pte)) {
+	if (pte_present(pte)) {
 		if (pte_sw_dirty(pte) && pte_write(pte))
 			pte_val(pte) &= ~PTE_RDONLY;
 		else