Re: [RFC PATCH 0/2] Switch ARM to generic find_bit() API

[Yury Norov]

On Thu, Oct 20, 2022 at 05:51:34PM +0100, Russell King (Oracle) wrote:
> On Wed, Oct 19, 2022 at 08:20:22PM -0700, Yury Norov wrote:
> > Hi Russell, all,
> > 
> > I'd like to respin a patch that switches ARM to generic find_bit()
> > functions.
> > 
> > Generic code works on par with arch or better, according to my
> > testing [1], and with recent improvements merged in v6.1, it should
> > be even faster.
> > 
> > ARM already uses many generic find_bit() functions - those that it
> > doesn't implement. So we are talking about migrating a subset of the
> > API; most of find_bit() family has only generic implementation on ARM.
> > 
> > The only concern about this migration is that ARM code supports
> > byte-aligned bitmap addresses, while generic code is optimized for
> > word-aligned bitmaps.
> > 
> > In my practice, I've never seen unaligned bitmaps. But to check that on
> > ARM, I added a run-time check for bitmap alignment. I gave it run on
> > several architectures and found nothing.
> > 
> > Can you please check that on your hardware and compare performance of
> > generic vs arch code for you? If everything is OK, I suggest switching
> > ARM to generic find_bit() completely.
> > 
> > Thanks,
> > Yury
> > 
> > [1] https://lore.kernel.org/all/YuWk3titnOiQACzC@yury-laptop/
> 
> I _really_ don't want to play around with this stuff right now... 6.0
> appears to have a regression on arm32 early on during boot:
> 
> [    1.410115] EXT4-fs error (device sda1): htree_dirblock_to_tree:1093: inode #256: block 8797: comm systemd: bad entry in directory: rec_len % 4 != 0 - offset=0, inode=33188, rec_len=35097, size=4096 fake=0
> 
> Booting 5.19 with the same filesystem works without issue and without
> even a fsck, but booting 6.0 always results in some problem that
> prevents it booting.
> 
> Debugging this is not easy, because there also seems to be something
> up with the bloody serial console - sometimes I get nothing, other
> times I get nothing more than:
> 
> [    2.929502] EXT4-fs error (de
> 
> and then the output stops. Is the console no longer synchronous? If it
> isn't, that's a huge mistake which can be seen right here with the
> partial message output... so I also need to work out how to make the
> console output synchronous again.

Got it.

I you think that EXT4 problems are due to unaligned bitmaps, you can take
1st patch from this series to check.

Thanks,
Yury