On Sun, Jul 03, 2022 at 06:20:53PM +0300, Andy Shevchenko wrote:
> On Sun, Jul 03, 2022 at 08:50:19AM +0200, Christophe JAILLET wrote:
> > Le 02/07/2022 à 23:09, Yury Norov a écrit :
> > > On Sat, Jul 02, 2022 at 08:29:36PM +0200, Christophe JAILLET wrote:
>
> ...
>
> > > This should be dropped, for sure, and kmalloc() at line 128 should be
> > > replaced with bitmap_alloc().
> >
> > This kmalloc() is for a structure and a flexible array.
> >
> > You mean re-arranging the code to allocate the structure alone at first,
> > then the bitmap?
We can change struct primes to:
struct primes {
struct rcu_head rcu;
unsigned long last, sz;
unsigned long *primes;
};
And then either allocate twice:
new = kmalloc(sizeof(struct primes);
new->primes = bitmap_alloc(sz);
Or keep the same struct primes for all expansions, and just allocate
new bitmap for ->primes when needed. This is what I meant.
This a bit deeper rework, but it addresses Andy's concern about excessive
fragmentation. (Did anyone before complain? Is it measurable?)
> It's one way, but it will increase fragmentation of memory. The other one
> as it seems to me is to name a new API properly, i.e. bitmap_size_to_bytes().
>
> In such case you won't need renames to begin with. And then would be able
> to convert driver-by-driver in cases of duplicated code.
>
> I think that's what confused Yuri and I kinda agree that bitmap_size() should
> return bits, and not bytes. Also argument for pure bitmap_size() would be
> bitmap itself, but we have no way to detect the length of bitmap because we
> are using POD and not a specific data structure for it.
bitmap_size_to_bytes() sounds better. How many places in the kernel
do we have where we can't simply use bitmap_alloc(), and need this
machinery? If this is the only one, I'd prefer to switch it to
bitmap_alloc() instead.
Thanks,
Yury
