Re: [PATCH] char_dev: replace cdev_map with an xarray

[Greg KH]

On Mon, Jan 11, 2021 at 08:58:18PM +0000, Matthew Wilcox wrote:
> On Mon, Jan 11, 2021 at 06:05:13PM +0100, Christoph Hellwig wrote:
> > @@ -486,14 +491,22 @@ int cdev_add(struct cdev *p, dev_t dev, unsigned count)
> >  	if (WARN_ON(dev == WHITEOUT_DEV))
> >  		return -EBUSY;
> >  
> > -	error = kobj_map(cdev_map, dev, count, NULL,
> > -			 exact_match, exact_lock, p);
> > -	if (error)
> > -		return error;
> > +	mutex_lock(&chrdevs_lock);
> > +	for (i = 0; i < count; i++) {
> > +		error = xa_insert(&cdev_map, dev + i, p, GFP_KERNEL);
> > +		if (error)
> > +			goto out_unwind;
> > +	}
> > +	mutex_unlock(&chrdevs_lock);
> 
> Looking at some of the users ...
> 
> #define BSG_MAX_DEVS            32768
> ...
>         ret = cdev_add(&bsg_cdev, MKDEV(bsg_major, 0), BSG_MAX_DEVS);
> 
> So this is going to allocate 32768 entries; at 8 bytes each, that's 256kB.
> With XArray overhead, it works out to 73 pages or 292kB.  While I don't
> have bsg loaded on my laptop, I imagine a lot of machines do.
> 
> drivers/net/tap.c:#define TAP_NUM_DEVS (1U << MINORBITS)
> include/linux/kdev_t.h:#define MINORBITS        20
> drivers/net/tap.c:      err = cdev_add(tap_cdev, *tap_major, TAP_NUM_DEVS);
> 
> That's going to be even worse -- 8MB plus the overhead to be closer to 9MB.
> 
> I think we do need to implement the 'store a range' option ;-(

Looks like it will be needed here.

Wow that's a lot of tap devices allocated all at once, odd...

Note, we will get some additional savings when this goes in as I can
delete the kobject from the cdev (after cleaning up the misguided
drivers that tried to set it thinking it was used), which will help out
a lot for the individual structures being created, but not for these
ranges.

thanks,

greg k-h