Re: [PATCH v2 11/11] powerpc/mm/book3s64/pgtable: Uses counting method to skip serializing

From: John Hubbard <jhubbard@nvidia.com>
To: Leonardo Bras <leonardo@linux.ibm.com>,
	<linuxppc-dev@lists.ozlabs.org>, <linux-kernel@vger.kernel.org>,
	Linux-MM <linux-mm@kvack.org>
Cc: Jason Gunthorpe <jgg@ziepe.ca>,
	Thomas Gleixner <tglx@linutronix.de>,
	"Arnd Bergmann" <arnd@arndb.de>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	YueHaibing <yuehaibing@huawei.com>,
	Keith Busch <keith.busch@intel.com>,
	Nicholas Piggin <npiggin@gmail.com>,
	Mike Rapoport <rppt@linux.ibm.com>,
	Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>,
	Richard Fontana <rfontana@redhat.com>,
	Paul Mackerras <paulus@samba.org>,
	Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>,
	Ganesh Goudar <ganeshgr@linux.ibm.com>,
	"Andrew Morton" <akpm@linux-foundation.org>,
	Ira Weiny <ira.weiny@intel.com>,
	"Dan Williams" <dan.j.williams@intel.com>,
	Allison Randal <allison@lohutok.net>
Subject: Re: [PATCH v2 11/11] powerpc/mm/book3s64/pgtable: Uses counting method to skip serializing
Date: Mon, 23 Sep 2019 11:14:39 -0700	[thread overview]
Message-ID: <8706a1f1-0c5e-d152-938b-f355b9a5aaa8@nvidia.com> (raw)
In-Reply-To: <18c5c378db98f223a0663034baa9fd6ce42f1ec7.camel@linux.ibm.com>

On 9/23/19 10:25 AM, Leonardo Bras wrote:
> On Fri, 2019-09-20 at 17:48 -0700, John Hubbard wrote:
>>
> [...]
>> So it seems that full memory barriers (not just compiler barriers) are required.
>> If the irq enable/disable somehow provides that, then your new code just goes
>> along for the ride and Just Works. (You don't have any memory barriers in
>> start_lockless_pgtbl_walk() / end_lockless_pgtbl_walk(), just the compiler
>> barriers provided by the atomic inc/dec.)
>>
>> So it's really a pre-existing question about the correctness of the gup_fast()
>> irq disabling approach.
> 
> I am not experienced in other archs, and I am still pretty new to
> Power, but by what I could understand, this behavior is better
> explained in serialize_against_pte_lookup. 
> 
> What happens here is that, before doing a THP split/collapse, the
> function does a update of the pmd and a serialize_against_pte_lookup,
> in order do avoid a invalid output on a lockless pagetable walk.
> 
> Serialize basically runs a do_nothing in every cpu related to the
> process, and wait for it to return. 
> 
> This running depends on interrupt being enabled, so disabling it before
> gup_pgd_range() and re-enabling after the end, makes the THP
> split/collapse wait for gup_pgd_range() completion in every cpu before
> continuing. (here happens the lock)
> 

That part is all fine, but there are no run-time memory barriers in the 
atomic_inc() and atomic_dec() additions, which means that this is not
safe, because memory operations on CPU 1 can be reordered. It's safe
as shown *if* there are memory barriers to keep the order as shown:

CPU 0                            CPU 1
------                         --------------
                               atomic_inc(val) (no run-time memory barrier!)
pmd_clear(pte)
if (val)
    run_on_all_cpus(): IPI
                               local_irq_disable() (also not a mem barrier)

                               READ(pte)
                               if(pte)
                                  walk page tables

                               local_irq_enable() (still not a barrier)
                               atomic_dec(val)

free(pte)

thanks,
-- 
John Hubbard
NVIDIA

> (As told before, every gup_pgd_range() that occurs after it uses a
> updated pmd, so no problem.)
> 
> I am sure other archs may have a similar mechanism using
> local_irq_{disable,enable}.
> 
> Did it answer your questions?
> 
> Best regards,
> 
> Leonardo Bras
> 