Re: [RFC 2/3] mm/vmalloc: add support for __GFP_NOFAIL

[NeilBrown]

On Thu, 21 Oct 2021, Uladzislau Rezki wrote:
> On Wed, Oct 20, 2021 at 05:00:28PM +0200, Uladzislau Rezki wrote:
> > >
> > > On Wed 20-10-21 16:29:14, Uladzislau Rezki wrote:
> > > > On Wed, Oct 20, 2021 at 4:06 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > > [...]
> > > > > As I've said I am OK with either of the two. Do you or anybody have any
> > > > > preference? Without any explicit event to wake up for neither of the two
> > > > > is more than just an optimistic retry.
> > > > >
> > > > From power perspective it is better to have a delay, so i tend to say
> > > > that delay is better.
> > >
> > > I am a terrible random number generator. Can you give me a number
> > > please?
> > >
> > Well, we can start from one jiffy so it is one timer tick: schedule_timeout(1)
> > 
> A small nit, it is better to replace it by the simple msleep() call: msleep(jiffies_to_msecs(1));

I disagree.  I think schedule_timeout_uninterruptible(1) is the best
wait to sleep for 1 ticl

msleep() contains
  timeout = msecs_to_jiffies(msecs) + 1;
and both jiffies_to_msecs and msecs_to_jiffies might round up too.
So you will sleep for at least twice as long as you asked for, possible
more.

NeilBrown


> 
> --
> Vlad Rezki
> 
>