linux-kernel - Re: [PATCH] xfs: fix diff_two_keys calculation for cnt btree

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <vx6bowvzlqixc4ap7vvj4mwarsuqm7y65cejg6yoc5wgpeh4j6@74rej3wf6uqq>
Date: Sun, 27 Apr 2025 14:56:39 +0300
From: Fedor Pchelkin <pchelkin@...ras.ru>
To: "Darrick J. Wong" <djwong@...nel.org>, 
	Kent Overstreet <kent.overstreet@...ux.dev>
Cc: Carlos Maiolino <cem@...nel.org>, 
	Chandan Babu R <chandanbabu@...nel.org>, Brian Foster <bfoster@...hat.com>, linux-xfs@...r.kernel.org, 
	linux-kernel@...r.kernel.org, lvc-project@...uxtesting.org, Alexey Nepomnyashih <sdl@...ct.ru>, 
	stable@...r.kernel.org, Christoph Hellwig <hch@....de>
Subject: Re: [PATCH] xfs: fix diff_two_keys calculation for cnt btree

Hi,

On Sat, 26. Apr 08:03, Darrick J. Wong wrote:
> On Sat, Apr 26, 2025 at 04:42:31PM +0300, Fedor Pchelkin wrote:
> > Currently the difference is computed on 32-bit unsigned values although
> > eventually it is stored in a variable of int64_t type. This gives awkward
> > results, e.g. when the diff _should_ be negative, it is represented as
> > some large positive int64_t value.
> > 
> > Perform the calculations directly in int64_t as all other diff_two_keys
> > routines actually do.
> > 
> > Found by Linux Verification Center (linuxtesting.org) with Svace static
> > analysis tool.
> > 
> > Fixes: 08438b1e386b ("xfs: plumb in needed functions for range querying of the freespace btrees")
> > Cc: stable@...r.kernel.org
> > Signed-off-by: Fedor Pchelkin <pchelkin@...ras.ru>
> > ---
> >  fs/xfs/libxfs/xfs_alloc_btree.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> > 
> > diff --git a/fs/xfs/libxfs/xfs_alloc_btree.c b/fs/xfs/libxfs/xfs_alloc_btree.c
> > index a4ac37ba5d51..b3c54ae90e25 100644
> > --- a/fs/xfs/libxfs/xfs_alloc_btree.c
> > +++ b/fs/xfs/libxfs/xfs_alloc_btree.c
> > @@ -238,13 +238,13 @@ xfs_cntbt_diff_two_keys(
> >  	ASSERT(!mask || (mask->alloc.ar_blockcount &&
> >  			 mask->alloc.ar_startblock));
> >  
> > -	diff =  be32_to_cpu(k1->alloc.ar_blockcount) -
> > -		be32_to_cpu(k2->alloc.ar_blockcount);
> > +	diff = (int64_t)be32_to_cpu(k1->alloc.ar_blockcount) -
> > +			be32_to_cpu(k2->alloc.ar_blockcount);
> 
> Perhaps it's time to hoist cmp_int to include/ and refactor all these
> things to use it?
> 
> #define cmp_int(l, r)          ((l > r) - (l < r))
> 
> --D
> 

Thanks, that would be worth it, I think. Though the current xfs
***diff_two_keys() implementations try to compute and return the actual
difference between two values, not the result of their comparison. Now
looking at diff_two_keys() use cases, I see only the latter one is needed
anyway so a good bit to refactor.


The thing I'm pondering over now is whether the macro in its current
form is okay to move up to include/. There is no argument restrictions and
typechecking intended to catch up obviously misleading usage patterns
though we'd need some if this is hoisted to a generic header and exported
for potential use by others?

There are four places where cmp_int is defined at the moment:
- bcachefs
- md/bcache
- xfs_zone_gc
- pipe.c

bcachefs is the largest user having all kinds of different arguments
providing to the macro, bitfields included. It also has several rather
generic wrappers, like u64_cmp, unsigned_cmp, u8_cmp, cmp_le32 and
others..

AF_UNIX code even has

	#define cmp_ptr(l, r)	(((l) > (r)) - ((l) < (r)))

for pointer comparisons.


So in my opinion we'd probably need to come up with something like a new
include/linux/cmp.h header where all this stuff will be gathered in a
generic way.

Any objections/suggestions on that? Or just moving

	#define cmp_int(l, r)          ((l > r) - (l < r))

to a generic header and dropping the corresponding defines from the
separate in-kernel users would be enough in this case?

--
Thanks,
Fedor

> >  	if (diff)
> >  		return diff;
> >  
> > -	return  be32_to_cpu(k1->alloc.ar_startblock) -
> > -		be32_to_cpu(k2->alloc.ar_startblock);
> > +	return (int64_t)be32_to_cpu(k1->alloc.ar_startblock) -
> > +			be32_to_cpu(k2->alloc.ar_startblock);
> >  }
> >  
> >  static xfs_failaddr_t
> > -- 
> > 2.49.0
> > 
> >