Message-ID: <87lg7mf30o.fsf@yhuang-dev.intel.com>
Date:   Fri, 28 Sep 2018 16:19:03 +0800
From:   "Huang\, Ying" <ying.huang@...el.com>
To:     Daniel Jordan <daniel.m.jordan@...cle.com>
Cc:     Andrew Morton <akpm@...ux-foundation.org>, <linux-mm@...ck.org>,
        <linux-kernel@...r.kernel.org>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>,
        Andrea Arcangeli <aarcange@...hat.com>,
        Michal Hocko <mhocko@...nel.org>,
        Johannes Weiner <hannes@...xchg.org>,
        Shaohua Li <shli@...nel.org>, Hugh Dickins <hughd@...gle.com>,
        Minchan Kim <minchan@...nel.org>,
        Rik van Riel <riel@...hat.com>,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        Naoya Horiguchi <n-horiguchi@...jp.nec.com>,
        Zi Yan <zi.yan@...rutgers.edu>
Subject: Re: [PATCH -V5 RESEND 03/21] swap: Support PMD swap mapping in swap_duplicate()

Daniel Jordan <daniel.m.jordan@...cle.com> writes:

> On Thu, Sep 27, 2018 at 09:34:36AM +0800, Huang, Ying wrote:
>> Daniel Jordan <daniel.m.jordan@...cle.com> writes:
>> > On Wed, Sep 26, 2018 at 08:55:59PM +0800, Huang, Ying wrote:
>> >> Daniel Jordan <daniel.m.jordan@...cle.com> writes:
>> >> > On Tue, Sep 25, 2018 at 03:13:30PM +0800, Huang Ying wrote:
>> >> >>  /*
>> >> >>   * Increase reference count of swap entry by 1.
>> >> >> - * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
>> >> >> - * but could not be atomically allocated.  Returns 0, just as if it succeeded,
>> >> >> - * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
>> >> >> - * might occur if a page table entry has got corrupted.
>> >> >> + *
>> >> >> + * Return error code in following case.
>> >> >> + * - success -> 0
>> >> >> + * - swap_count_continuation is required but could not be atomically allocated.
>> >> >> + *   *entry is used to return swap entry to call add_swap_count_continuation().
>> >> >> + *								      -> ENOMEM
>> >> >> + * - otherwise same as __swap_duplicate()
>> >> >>   */
>> >> >> -int swap_duplicate(swp_entry_t entry)
>> >> >> +int swap_duplicate(swp_entry_t *entry, int entry_size)
>> >> >>  {
>> >> >>  	int err = 0;
>> >> >>  
>> >> >> -	while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
>> >> >> -		err = add_swap_count_continuation(entry, GFP_ATOMIC);
>> >> >> +	while (!err &&
>> >> >> +	       (err = __swap_duplicate(entry, entry_size, 1)) == -ENOMEM)
>> >> >> +		err = add_swap_count_continuation(*entry, GFP_ATOMIC);
>> >> >>  	return err;
>> >> >
>> >> > Now we're returning any error we get from __swap_duplicate, apparently to
>> >> > accommodate ENOTDIR later in the series, which is a change from the behavior
>> >> > introduced in 570a335b8e22 ("swap_info: swap count continuations").  This might
>> >> > belong in a separate patch given its potential for side effects.
>> >> 
>> >> I have checked all the calls of the function and found there will be no
>> >> bad effect.  Do you have any side effect?
>> >
>> > Before I was just being vaguely concerned about any unintended side effects,
>> > but looking again, yes I do.
>> >
>> > Now when swap_duplicate returns an error in copy_one_pte, copy_one_pte returns
>> > a (potentially nonzero) entry.val, which copy_pte_range interprets
>> > unconditionally as 'try adding a swap count continuation.'  Not what we want
>> > for returns other than -ENOMEM.
>> 
>> Thanks for pointing this out!  Before the change in the patchset, the
>> behavior is,
>> 
>> Something wrong is detected in swap_duplicate(), but the error is
>> ignored.  Then copy_one_pte() will think everything is OK, so that it
>> can proceed to call set_pte_at().  The system will be in inconsistent
>> state and some data may be polluted!
>
> Yes, the part about page table corruption in the comment above swap_duplicate.
>
>> But this doesn't cause any problem in practical.  Per my understanding,
>> because if other part of the kernel works correctly, it's impossible for
>> swap_duplicate() return any error except -ENOMEM before the change in
>> this patchset.
>
> I agree with that, but it's not what I'm trying to explain.  I didn't go into
> enough detail, let me try again.  Hopefully I'm understanding this right.
>
> While running with these patches, say we're at
>
>   copy_pte_range
>    copy_one_pte
>     swap_duplicate
>      __swap_duplicate
>       __swap_duplicate_locked
>     
> And say __swap_duplicate_locked returns an error that isn't -ENOMEM, such as
> -EEXIST.  That means __swap_duplicate and swap_duplicate also return -EEXIST.
> copy_one_pte returns entry.val, which can be and usually is nonzero, so we
> break out of the loop in copy_pte_range and then--erroneously--call
> add_swap_count_continuation.
>
> The add_swap_count_continuation call was added in 570a335b8e22 and relies on
> the assumption that callers can only get -ENOMEM from swap_duplicate.  This
> patch changes that assumption.
>
> Not a big deal: the continuation call just returns early, no harm done, but it
> allocs and frees a page needlessly, so we should fix it.  One way is to change
> copy_one_pte's return to int so we can just pass the error code back to
> copy_pte_range so it knows whether to try adding the continuation.

There may be even more problems.  After add_swap_count_continuation(),
copy_one_pte() will be retried, and the CPU may hang with dead loop.

But before the changes in this patchset, the behavior is,
__swap_duplicate() return an error that isn't -ENOMEM, such as -EEXIST.
Then copy_one_pte() would thought the operation has been done
successfully, and go to call set_pte_at().  This will cause the system
state become inconsistent, and the system may panic or hang somewhere
later.

So per my understanding, if we thought page table corruption isn't a
real problem (that is, __swap_duplicate() will never return e.g. -EEXIST
if copied by copy_one_pte() indirectly), both the original and the new
code should be OK.

If we thought it is a real problem, we need to fix the original code and
keep it fixed in the new code.  Do you agree?

There's several ways to fix the problem.  But the page table shouldn't
be corrupted in practice, unless there's some programming error.  So I
suggest to make it as simple as possible via adding,

VM_BUG_ON(error != -ENOMEM);

in swap_duplicate().

Do you agree?

Best Regards,
Huang, Ying

> The other swap_duplicate caller, try_to_unmap_one, seems ok.
>
>> But the error may be possible during development, and it
>> may serve as some kind of document too.  So I suggest to add
>> 
>> VM_BUG_ON(error != -ENOMEM);
>> 
>> in swap_duplicate().  What do you think about that?
>
> That doesn't seem necessary.
>
>> > So it might make sense to have a separate patch that changes swap_duplicate's
>> > return and makes callers handle it.
>> 
>> Thanks for your help to take a deep look at this.  I want to try to fix
>> all potential problems firstly, because the number of the caller is
>> quite limited.  Do you agree?
>
> Yes, makes sense to me.
>
> Daniel

Powered by blists - more mailing lists