Re: [PATCH v2 2/3] sched_ext: Track bits[] storage size in struct scx_cmask

[Andrea Righi]

Hi Tejun,

On Mon, May 18, 2026 at 12:53:06PM -1000, Tejun Heo wrote:
> On Tue, May 19, 2026 at 12:11:35AM +0200, Andrea Righi wrote:
> > > +/**
> > > + * scx_cmask_reframe - Reshape @m's active range without resizing storage
> > > + * @m: cmask to reframe
> > > + * @base: new active range base
> > > + * @nr_cids: new active range length, must fit within @m->alloc_words
> > > + *
> > > + * Body bits within the new range become garbage - only the head and tail
> > > + * words are zeroed to keep the padding invariant.
> > > + */
> > > +static inline void scx_cmask_reframe(struct scx_cmask *m, u32 base, u32 nr_cids)
> > > +{
> > > +	if (WARN_ON_ONCE(SCX_CMASK_NR_WORDS(nr_cids) > m->alloc_words))
> > > +		return;
> > 
> > Considering that:
> > 
> >  #define SCX_CMASK_NR_WORDS(nr_cids)    (((nr_cids) + 63) / 64 + 1)
> > 
> > If we pass nr_cids == UINT_MAX here, we have:
> > 
> >  CMASK_NR_WORDS(UINT_MAX) = (UINT_MAX + 63)/64 + 1 = 62/64 + 1 = 1 (wraps)
> > 
> > Should we simply reject if it's greater than a certain reasonable upper bound?
> 
> I'm not sure what we do matters. No matter what, this would be a clear bug
> and an unlikely one at that. As long as the backtrace is dumped, I think
> anything is fine.

Agreed that the bug is unlikely to happen, but the WARN_ON_ONCE() wouldn't fire
at all for nr_cids == UINT_MAX.

However, IIUC scx_cmask_reframe() is internal kernel code, with no callers yet
and the upcoming consumers will probably drive nr_cids from bounded sources,
likely num_possible_cpus() and such. So, the wrap shouldn't really reachable on
the kernel side. Therefore, I guess we can simply drop the WARN_ON_ONCE().

Instead, cmask_reframe() is the mirrored version that is called from BPF, so any
loaded BPF prog can potentially pass an arbitrary u32.

How about changing the CMASK_NR_WORDS() macro as following?

 #define CMASK_NR_WORDS(nr_cids) ((u32)(((u64)(nr_cids) + 63) / 64 + 1))

In this way CMASK_NR_WORDS(UINT_MAX) returns ~67M instead of 1 and in this way
we'd get a clear scx_bpf_error() backtrace if it wraps. WDYT?

Thanks,
-Andrea