On Fri, Apr 02, 2021 at 06:30:48PM -0700, Josh Don wrote:
> Hi Peter,
>
> I walked through the reference counting, and it seems good to me
> (though it did take a few passes to fully digest the invariants for
> the fat cookie stuff).
>
> > +unsigned long sched_core_alloc_cookie(unsigned int type)
> > {
> > struct sched_core_cookie *ck = kmalloc(sizeof(*ck), GFP_KERNEL);
> > if (!ck)
> > return 0;
> > - refcount_set(&ck->refcnt, 1);
> > + WARN_ON_ONCE(type > GROUP_COOKIE);
> > + sched_core_init_cookie(ck, type);
> > sched_core_get();
> >
> > - return (unsigned long)ck;
> > + return (unsigned long)ck | type;
> > }
>
> This feels like it needs to be stronger than a WARN_ON_ONCE; could
> create a corrupted address that we later try to kfree().
The fattie is also released with kfree(), not a real problem that.
> Also, for my own edification, why will the bottom two bits here always be 0?
#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
> > -unsigned long sched_core_alloc_cookie(void)
> > +static inline void *cookie_ptr(unsigned long cookie)
> > +{
> > + return (void *)(cookie & ~3UL);
> > +}
> > +
> > +static inline int cookie_type(unsigned long cookie)
> > +{
> > + return cookie & 3;
> > +}
>
> s/3/FAT_COOKIE
COOKIE_MASK if anything..
> > +#define FAT_COOKIE 0x03
>
> Move to sched.h to group with TASK/GROUP_COOKIE?
Didn't want to expose FAT outside the ifdef
> > +static unsigned long __sched_core_fat_cookie(struct task_struct *p,
> > + void **spare_fat,
> > + unsigned long cookie)
> > +{
>
> This function looks good to me, but could use some more comments about
> the pre/post-condition assumptions. Ie. cookie already has a get()
> associated with it, caller is expected to kfree the spare_fat.
Fair enough. I'll go write up something. I had to fix a refcount issue
here, which should've been clue a comment it required.
There's actually a bug with the rb tree magic too, I'll go fix.
> > + raw_spin_lock_irqsave(&fat_lock, flags);
> > + n = rb_find_add(&fat->node, &fat_root, fat_cmp);
> > + raw_spin_unlock_irqrestore(&fat_lock, flags);
> > +
> > + if (n) {
> > + sched_core_put_fat(fat);
> > + fat = node_2_fat(n);
>
> This put() doesn't seem strictly necessary; caller will be
> unconditionally freeing the spare_fat. Keep anyway for completeness,
> but add a comment?
It needs to put the references we got on constructing the new fat.
