On Thu, 3 Dec 2020, Marek Szyprowski wrote:
> On 03.12.2020 16:46, Marek Szyprowski wrote:
> > On 25.11.2020 03:32, Matthew Wilcox wrote:
> >> On Tue, Nov 17, 2020 at 11:43:02PM +0000, Matthew Wilcox wrote:
> >>> On Tue, Nov 17, 2020 at 07:15:13PM +0000, Matthew Wilcox wrote:
> >>>> I find both of these functions exceptionally confusing.  Does this
> >>>> make it easier to understand?
> >>> Never mind, this is buggy.  I'll send something better tomorrow.
> >> That took a week, not a day.  *sigh*.  At least this is shorter.
> >>
> >> commit 1a02863ce04fd325922d6c3db6d01e18d55f966b
> >> Author: Matthew Wilcox (Oracle) <willy@infradead.org>
> >> Date:   Tue Nov 17 10:45:18 2020 -0500
> >>
> >>      fix mm-truncateshmem-handle-truncates-that-split-thps.patch
> >
> > This patch landed in todays linux-next (20201203) as commit 
> > 8678b27f4b8b ("8678b27f4b8bfc130a13eb9e9f27171bcd8c0b3b"). Sadly it 
> > breaks booting of ANY of my ARM 32bit test systems, which use initrd. 
> > ARM64bit based systems boot fine. Here is example of the crash:

kernel BUG at fs/inode.c:531
evict() hitting clear_inode()'s BUG_ON(inode->i_data.nr_pages)
Same here on i386 on mmotm (slightly different line number on mmotm).

> 
> One more thing. Reverting those two:
> 
> 1b1aa968b0b6 mm-truncateshmem-handle-truncates-that-split-thps-fix-fix
> 
> 8678b27f4b8b mm-truncateshmem-handle-truncates-that-split-thps-fix
> 
> on top of linux next-20201203 fixes the boot issues.

Thanks a lot for the report, Marek.  Yes, reverting those two
(of which "-fix" amounts to a rewrite, and "-fix-fix" is far from
complete) takes the linux-next tree back to how truncate was before it
took in yesterday's mmotm: not crashing on 32-bit, but still not good.

The 32-bit breakage may turn out to be a simple one-liner like a
missing cast, or overflow from 0 to -1, somewhere in the rewritten
truncate_inode_pages_range(); but it did not stand out to me, and
it does not immediately matter, since other fixes are needed to
that patch.  I'm afraid it's proving to be work in progress.

I did ask Andrew to revert these earlier in the thread, but it looks
like that got lost in the jungle of his inbox: I'll send a better
targetted mail, but what we need is to revert these *four* patches,
until we have a better tested and stable version.

mm-truncateshmem-handle-truncates-that-split-thps.patch
mm-truncateshmem-handle-truncates-that-split-thps-fix.patch
mm-truncateshmem-handle-truncates-that-split-thps-fix-fix.patch
mm-filemap-return-only-head-pages-from-find_get_entries.patch

Hugh