RE: [PATCH] mm/vmalloc: Replace the ternary conditional operator with min()

[David Laight]

From: Lorenzo Stoakes
> Sent: 10 June 2023 22:07
...
> > > OK, as per the pedantic test bot, you'll need to change this to:-
> > >
> > > num = min_t(size_t, remains, PAGE_SIZE);
> >
> 
> Ordinarily I wouldn't respond to this (I go into why I feel this is not
> useful commentary below) but I am concerned Lu will take you seriously.
> 
> > There has to be a valid reason why min/max have strong type checks.
> 
> I really don't know what you mean by this? Yes there is a reason, I imagine
> it's to avoid unfortunate and invalid type comparisons.

Indeed, the 'unfortunate conversion' is a negative value
being converted to a large positive one.
That doesn't require anything like the type checking that min/max do.

> This is not applicable here (explained below...)
> 
> > Using min_t() all the time is just subverting them and means that
> > bugs are more likely than if the extra tests in min() were absent.
> 
> 'All the time' - are you just having a general whine + moan about perceived
> kernel practices? Can you please keep it focused on the actual issues at
> hand? I am not Linus and therefore not responsible for the entirety of the
> kernel.

I see a general problem (that Linus ought to worried about)
is that whenever min() reports a type error the answer is
do immediately drop in a min_t() instead of looking at the
type of the values and fixing them to that min() doesn't complain.
(Or fixing min() so it doesn't object to a much larger class
of comparisons.0

...
> > A 'safe' change is min(remains + 0ULL, PAGE_SIZE).
> 
> So now we're promoting an unsigned int (and sometimes unsigned long of
> course) to an unsigned long long (for reasons unknown) and comparing it
> with an unsigned long? Wouldn't this trigger the sensitive type check
> anyway?

PAGE size is defined to be 'long long' - so min() will be happy.
The compiler will just DTRT, even if 'remains' is 32bit it will
optimise away all the implied 64-bit extension.

Almost all the min_t() are min_t((some unsigned type),a,b).
If the values are known to be positive then:
#define min_unsigned(a, b) min((a) + 0u + 0ul + 0ull, (b) + 0u + 0ul + 0ull))
will zero extend whatever type is supplied before the comparison.
The compiler will just discard zero extensions.

> To be clear, I'd nack any such ridiculous change unless a hugely compelling
> reason is given (you've not given any). That's horrific. And again, you've
> not provided one single example of an _actual_ bug or situation where the
> 'problem' you tiresomely raise would occur.

The (size_t) cast isn't in itself a problem - provided you've
checked that it is larger than the types of both sides.
But search the kernel and you'll find places when min_t((u8),a,b)
is used.
This all follows the same pattern of min() gave a warning so
so use min_t().

...
> > But, in reality, min/max should always be valid when one
> > value is a constant between 0 and MAX_INT.
> 
> This is getting at a signed/unsigned comparison issue here afaict which is
> not the one we're dealing with here.

But it is exactly what min() is checking for and almost why min()
exists.
A min_unsafe() that didn't try to do any checks would be better
than train wreck that min_t() can create.

...
> Now since you kicked off this 'all the time' stuff I feel like I have been
> given permission to make some broad comments myself...
> 
> David, I am not one to commit-shame being a minor contributor myself buuuut
> I see 7,610 messages from you on lore and 4 commits, all from 4 years ago
> (please correct me if I'm wrong).

I don't work for google, intel, aws (etc).
Getting patches accepted is surprisingly hard.

I've been writing device driver and comms protocol stack code
for best part of 40 years.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)