Re: [RFC PATCHv5 1/6] zram: introduce writeback bio batching

[Sergey Senozhatsky]

On (25/11/21 08:40), Hannes Reinecke wrote:
> > +static int zram_complete_done_reqs(struct zram *zram,
> > +				   struct zram_wb_ctl *wb_ctl)
> > +{
> > +	struct zram_wb_req *req;
> > +	unsigned long flags;
> >   	int ret = 0, err;
> > -	u32 index;
> > -	page = alloc_page(GFP_KERNEL);
> > -	if (!page)
> > -		return -ENOMEM;
> > +	while (1) {
> > +		spin_lock_irqsave(&wb_ctl->done_lock, flags);
> > +		req = list_first_entry_or_null(&wb_ctl->done_reqs,
> > +					       struct zram_wb_req, entry);
> > +		if (req)
> > +			list_del(&req->entry);
> > +		spin_unlock_irqrestore(&wb_ctl->done_lock, flags);
> > +
> > +		if (!req)
> > +			break;
> > +
> > +		err = zram_writeback_complete(zram, req);
> > +		if (err)
> > +			ret = err;
> > +
> > +		atomic_dec(&wb_ctl->num_inflight);
> > +		release_pp_slot(zram, req->pps);
> > +		req->pps = NULL;
> > +
> > +		list_add(&req->entry, &wb_ctl->idle_reqs);
> 
> Shouldn't this be locked?

See below.

> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +static struct zram_wb_req *zram_select_idle_req(struct zram_wb_ctl *wb_ctl)
> > +{
> > +	struct zram_wb_req *req;
> > +
> > +	req = list_first_entry_or_null(&wb_ctl->idle_reqs,
> > +				       struct zram_wb_req, entry);
> > +	if (req)
> > +		list_del(&req->entry);
> 
> See above. I think you need to lock this to avoid someone stepping in
> here an modify the element under you.

->idle_reqs list is mutated by one and one task only: the one that
does writeback.  ->done_reqs list, on the other hand, is accessed
both from IRQ (bio completion) and from the writeback task.