Re: dpkg fails on sparc64 (was: [PATCH v4 4/7] mm/thp: Carry over dirty bit when thp splits on pmd)

From: Peter Xu <peterx@redhat.com>
To: Anatoly Pugachev <matorola@gmail.com>
Cc: David Miller <davem@davemloft.net>,
	linux-mm@kvack.org, linux-kernel@vger.kernel.org,
	Hugh Dickins <hughd@google.com>,
	"Kirill A . Shutemov" <kirill@shutemov.name>,
	Alistair Popple <apopple@nvidia.com>,
	Andrea Arcangeli <aarcange@redhat.com>,
	Minchan Kim <minchan@kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	David Hildenbrand <david@redhat.com>,
	Andi Kleen <andi.kleen@intel.com>,
	Nadav Amit <nadav.amit@gmail.com>,
	Huang Ying <ying.huang@intel.com>,
	Vlastimil Babka <vbabka@suse.cz>,
	sparclinux@vger.kernel.org
Subject: Re: dpkg fails on sparc64 (was: [PATCH v4 4/7] mm/thp: Carry over dirty bit when thp splits on pmd)
Date: Tue, 25 Oct 2022 10:43:45 -0400	[thread overview]
Message-ID: <Y1f2IR+h4i2+/swj@x1n> (raw)
In-Reply-To: <CADxRZqy+cMHN4FjtDr7-LOyVf0y+G8MPiBoGiTEsSj48jBfVnw@mail.gmail.com>

On Tue, Oct 25, 2022 at 01:22:45PM +0300, Anatoly Pugachev wrote:
> On Sun, Oct 23, 2022 at 10:53 PM Peter Xu <peterx@redhat.com> wrote:
> > On Fri, Oct 21, 2022 at 07:06:03PM +0300, Anatoly Pugachev wrote:
> > >
> > >     Link: https://lkml.kernel.org/r/20220811161331.37055-5-peterx@redhat.com
> > >
> > > So, v6.0-rc3-176-g0d206b5d2e0d) does not segfault dpkg,
> > > v6.0-rc3-177-g0ccf7f168e17 segfaults it on package install.
> > >
> > > dpkg test was (apt) install/remove some packages, segfaults only on install
> > > (not remove).
> > >
> > > Reverted 0ccf7f168e17bb7eb5a322397ba5a841f4fbaccb from top of v6.1-rc1 and
> > > tried to compile kernel, but got error
> > >
> > > mm/huge_memory.c: In function ‘__split_huge_pmd_locked’:
> > > mm/huge_memory.c:2129:17: error: ‘dirty’ undeclared (first use in this function)
> > >  2129 |                 dirty = is_migration_entry_dirty(entry);
> > >       |                 ^~~~~
> > > mm/huge_memory.c:2129:17: note: each undeclared identifier is reported only once for each function it appears in
> > > make[2]: *** [scripts/Makefile.build:250: mm/huge_memory.o] Error 1
> > >
> > > So can't test v6.1-rc1 with patch reverted...
> >
> > Sorry to know this, and thanks for the report and debugging.  The revert
> > won't work because dirty variable is used in later patch for the swap path
> > too.  I've attached a partial (and minimum) revert, feel free to try.
> 
> Peter,
> 
> tested again with 6.1.0-rc2 already, non patched kernel segfaulting
> dpkg, using your patch makes dpkg
> (or kernel) to behave properly.
> Thanks!

Thanks for the quick feedback.

> 
> > I had a feeling that it's somehow related to the special impl of sparc64
> > pte_mkdirty() where a kernel patching mechanism is used to share code
> > between sun4[uv].  I'd assume your machine is sun4v?  As that's the one
> > that needs the patching, iiuc.
> 
> kernel boot log reports
> ARCH: SUN4V

Then it's expected but unfortunate too, as QEMU doesn't seem to have
support on sun4v so I cannot even try that out with a VM.

https://wiki.qemu.org/Documentation/Platforms/SPARC

I'd also expect there's nothing useful in either dmesg or relevant logs
because it's segv, but please share if you find anything that may be
helpful.

Maybe we need to have the minimum revert for v6.1 before we have more
clues.

Thanks,

-- 
Peter Xu