Re: [PATCH 4/7] rcu/tree: support reclaim for head-less object

[Uladzislau Rezki]

> 
> Very nice work, Vlad! And beautifully split! Makes it easy to review! One
> comment below but otherwise patches 1-4 look good to me, I will look at
> others as well now. I have some patches on top of the series, mostly little
> clean ups which I will send together with yours.
> 
Thanks, Joel!


> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> > ---
> >  kernel/rcu/tree.c | 94 +++++++++++++++++++++++++++++++++++++++++++----
> >  1 file changed, 86 insertions(+), 8 deletions(-)
> > 
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 869a72e25d38..5a64c92feafc 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -2892,19 +2892,34 @@ static void kfree_rcu_work(struct work_struct *work)
> >  	 * when we could not allocate a bulk array.
> >  	 *
> >  	 * Under that condition an object is queued to the
> > -	 * list instead.
> > +	 * list instead. Please note that head-less objects
> > +	 * have dynamically attached rcu_head, so they also
> > +	 * contain a back-pointer that has to be freed.
> >  	 */
> >  	for (; head; head = next) {
> >  		unsigned long offset = (unsigned long)head->func;
> > -		void *ptr = (void *)head - offset;
> > +		bool headless;
> > +		void *ptr;
> >  
> >  		next = head->next;
> > +
> > +		/* We tag the headless object, if so adjust offset. */
> > +		headless = (((unsigned long) head - offset) & BIT(0));
> > +		if (headless)
> > +			offset -= 1;
> 
> Just to be sure, can vmalloc() ever allocate an object at an odd valued
> memory address? I'm not fully sure looking at vmalloc code whether this is
> the case.
>
No. It must be PAGE_ALIGNED(addr). 

> 
> As per the tagging, allocated objects have to at least at a 2-byte boundary
> for the pointer's BIT(0) to be available. If that's not the case, we need to
> add a warning to the code at a bare minimum.
> 
> Another approach which is better I think is to add the tag to the offset
> itself. So if the offset is > LONG_MAX / 2 or something like that, then
> assume it is headless, and override offset to sizeof(unsigned long *) in that
> case. Then you would arrive at the correct pointer for the wrapper. That
> would take care of the case where in the future, either SLAB or vmalloc()
> ends up returning pointers that are only byte-aligned.
> 
> Thoughts?
> 
I saw your https://lkml.org/lkml/2020/3/29/480. I think it is OK,
and we can go your way. There is advantage that it is tolerate to
any alignment :)

Thanks!

--
Vlad Rezki