Message-ID: <aQjxjlJvLnx_zRx8@smile.fi.intel.com>
Date: Mon, 3 Nov 2025 20:16:46 +0200
From: Andy Shevchenko <andriy.shevchenko@...el.com>
To: David Laight <david.laight.linux@...il.com>
Cc: Kuan-Wei Chiu <visitorckw@...il.com>,
	Guan-Chun Wu <409411716@....tku.edu.tw>,
	Andrew Morton <akpm@...ux-foundation.org>, ebiggers@...nel.org,
	tytso@....edu, jaegeuk@...nel.org, xiubli@...hat.com,
	idryomov@...il.com, kbusch@...nel.org, axboe@...nel.dk, hch@....de,
	sagi@...mberg.me, home7438072@...il.com,
	linux-nvme@...ts.infradead.org, linux-fscrypt@...r.kernel.org,
	ceph-devel@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v4 0/6] lib/base64: add generic encoder/decoder, migrate
 users

On Mon, Nov 03, 2025 at 04:41:41PM +0200, Andy Shevchenko wrote:
> On Mon, Nov 03, 2025 at 01:22:13PM +0000, David Laight wrote:
> > On Mon, 3 Nov 2025 19:07:24 +0800
> > Kuan-Wei Chiu <visitorckw@...il.com> wrote:
> > > On Mon, Nov 03, 2025 at 11:24:35AM +0100, Andy Shevchenko wrote:
> > > > On Fri, Oct 31, 2025 at 09:09:47PM -0700, Andrew Morton wrote:  
> > > > > On Wed, 29 Oct 2025 18:17:25 +0800 Guan-Chun Wu <409411716@....tku.edu.tw> wrote:

...

> > > > > Looks like wonderful work, thanks.  And it's good to gain a selftest
> > > > > for this code.
> > > > >   
> > > > > > This improves throughput by ~43-52x.  
> > > > > 
> > > > > Well that isn't a thing we see every day.  
> > > > 
> > > > I agree with the judgement, the problem is that this broke drastically a build:
> > > > 
> > > > lib/base64.c:35:17: error: initializer overrides prior initialization of this subobject [-Werror,-Winitializer-overrides]
> > > >    35 |         [BASE64_STD] = BASE64_REV_INIT('+', '/'),
> > > >       |                        ^~~~~~~~~~~~~~~~~~~~~~~~~
> > > > lib/base64.c:26:11: note: expanded from macro 'BASE64_REV_INIT'
> > > >    26 |         ['A'] =  0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, \
> > > >       |                  ^
> > > > lib/base64.c:35:17: note: previous initialization is here
> > > >    35 |         [BASE64_STD] = BASE64_REV_INIT('+', '/'),
> > > >       |                        ^~~~~~~~~~~~~~~~~~~~~~~~~
> > > > lib/base64.c:25:16: note: expanded from macro 'BASE64_REV_INIT'
> > > >    25 |         [0 ... 255] = -1, \
> > > >       |                       ^~
> > > > ...
> > > > fatal error: too many errors emitted, stopping now [-ferror-limit=]
> > > > 20 errors generated.
> > > >   
> > > Since I didn't notice this build failure, I guess this happens during a
> > > W=1 build? Sorry for that. Maybe I should add W=1 compilation testing
> > > to my checklist before sending patches in the future. I also got an
> > > email from the kernel test robot with a duplicate initialization
> > > warning from the sparse tool [1], pointing to the same code.
> > > 
> > > This implementation was based on David's previous suggestion [2] to
> > > first default all entries to -1 and then set the values for the 64
> > > character entries. This was to avoid expanding the large 256 * 3 table
> > > and improve code readability.
> > > 
> > > Since I believe many people test and care about W=1 builds,
> > 
> > Last time I tried a W=1 build it failed horribly because of 'type-limits'.
> > The kernel does that all the time - usually for its own error tests inside
> > #define and inline functions.
> > Certainly some of the changes I've seen to stop W=1 warnings are really
> > a bad idea - but that is a bit of a digression.
> > 
> > Warnings can be temporarily disabled using #pragma.
> > That might be the best thing to do here with this over-zealous warning.
> > 
> > This compiles on gcc and clang (even though the warnings have different names):
> > #pragma GCC diagnostic push
> > #pragma GCC diagnostic ignored "-Woverride-init"
> > int x[16] = { [0 ... 15] = -1, [5] = 5};
> > #pragma GCC diagnostic pop
> > 
> > > I think we need to find another way to avoid this warning?
> > > Perhaps we could consider what you suggested:
> > > 
> > > #define BASE64_REV_INIT(val_plus, val_comma, val_minus, val_slash, val_under) { \
> > > 	[ 0 ... '+'-1 ] = -1, \
> > > 	[ '+' ] = val_plus, val_comma, val_minus, -1, val_slash, \
> > > 	[ '0' ] = 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, \
> > > 	[ '9'+1 ... 'A'-1 ] = -1, \
> > > 	[ 'A' ] = 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, \
> > > 		  23, 24, 25, 26, 27, 28, 28, 30, 31, 32, 33, 34, 35, \
> > > 	[ 'Z'+1 ... '_'-1 ] = -1, \
> > > 	[ '_' ] = val_under, \
> > > 	[ '_'+1 ... 'a'-1 ] = -1, \
> > > 	[ 'a' ] = 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, \
> > > 		  49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, \
> > > 	[ 'z'+1 ... 255 ] = -1 \
> > > }
> > 
> > I just checked, neither gcc nor clang allow empty ranges (eg [ 6 ... 5 ] = -1).
> > Which means the coder has to know which characters are adjacent as well
> > as getting the order right.
> > Basically avoiding the warning sucks.
> > 
> > > Or should we just expand the 256 * 3 table as it was before?
> > 
> > That has much the same issue - IIRC it relies on three big sequential lists.
> > 
> > The #pragma may be best - but doesn't solve sparse (unless it processes
> > them as well).
> 
> Pragma will be hated. I believe there is a better way to do what you want. Let
> me cook a PoC.

I tried locally several approaches and the best I can come up with is the pre-generated
(via Python script) pieces of C code that we can copy'n'paste instead of that shortened
form. So basically having a full 256 tables in the code is my suggestion to fix the build
issue. Alternatively we can generate that at run-time (on the first run) in
the similar way how prime_numbers.c does. The downside of such an approach is loosing
the const specifier, which I consider kinda important.

Btw, in the future here might be also the side-channel attack concerns appear, which would
require to reconsider the whole algo to get it constant-time execution.

> > > [1]: https://lore.kernel.org/oe-kbuild-all/202511021343.107utehN-lkp@intel.com/
> > > [2]: https://lore.kernel.org/lkml/20250928195736.71bec9ae@pumpkin/

-- 
With Best Regards,
Andy Shevchenko

Powered by blists - more mailing lists