* Christian Brauner <brauner@xxxxxxxxxx> [241213 16:07]:
> On Fri, Dec 13, 2024 at 09:50:56PM +0100, Christian Brauner wrote:
> > On Fri, Dec 13, 2024 at 09:11:04PM +0100, Christian Brauner wrote:
> > > On Fri, Dec 13, 2024 at 08:25:21PM +0100, Christian Brauner wrote:
> > > > On Fri, Dec 13, 2024 at 08:01:30PM +0100, Christian Brauner wrote:
> > > > > On Fri, Dec 13, 2024 at 06:53:55PM +0000, Matthew Wilcox wrote:
> > > > > > On Fri, Dec 13, 2024 at 07:51:50PM +0100, Christian Brauner wrote:
> > > > > > > Yeah, it does. Did you see the patch that is included in the series?
> > > > > > > I've replaced the macro with always inline functions that select the
> > > > > > > lock based on the flag:
> > > > > > >
> > > > > > > static __always_inline void mtree_lock(struct maple_tree *mt)
> > > > > > > {
> > > > > > > if (mt->ma_flags & MT_FLAGS_LOCK_IRQ)
> > > > > > > spin_lock_irq(&mt->ma_lock);
> > > > > > > else
> > > > > > > spin_lock(&mt->ma_lock);
> > > > > > > }
> > > > > > > static __always_inline void mtree_unlock(struct maple_tree *mt)
> > > > > > > {
> > > > > > > if (mt->ma_flags & MT_FLAGS_LOCK_IRQ)
> > > > > > > spin_unlock_irq(&mt->ma_lock);
> > > > > > > else
> > > > > > > spin_unlock(&mt->ma_lock);
> > > > > > > }
> > > > > > >
> > > > > > > Does that work for you?
> > > > > >
> > > > > > See the way the XArray works; we're trying to keep the two APIs as
> > > > > > close as possible.
> > > > > >
> > > > > > The caller should use mtree_lock_irq() or mtree_lock_irqsave()
> > > > > > as appropriate.
> > > > >
> > > > > Say I need:
> > > > >
> > > > > spin_lock_irqsave(&mt->ma_lock, flags);
> > > > > mas_erase(...);
> > > > > -> mas_nomem()
> > > > > -> mtree_unlock() // uses spin_unlock();
> > > > > // allocate
> > > > > -> mtree_lock() // uses spin_lock();
> > > > > spin_lock_irqrestore(&mt->ma_lock, flags);
> > > > >
> > > > > So that doesn't work, right? IOW, the maple tree does internal drop and
> > > > > retake locks and they need to match the locks of the outer context.
> > > > >
> > > > > So, I think I need a way to communicate to mas_*() what type of lock to
> > > > > take, no? Any idea how you would like me to do this in case I'm not
> > > > > wrong?
> > > >
> > > > My first inclination has been to do it via MA_STATE() and the mas_flag
> > > > value but I'm open to any other ideas.
> > >
> > > Braino on my part as free_pid() can be called with write_lock_irq() held.
> >
> > I don't think I can use the maple tree because even an mas_erase()
> > operation may allocate memory and that just makes it rather unpleasant
> > to use in e.g., free_pid(). So I think I'm going to explore using the
> > xarray to get the benefits of ULONG_MAX indices and I see that btrfs is
> > using it already for similar purposes.
>
> Hm, __xa_alloc_cyclic() doesn't support ULONG_MAX indices. So any ideas
> how I can proceed? Can I use the maple tree to wipe an entry at a given
> index and have the guarantee it won't have to allocate memory?
There are two ways you can do this:
1. preallocate, then store in the locked area.
2. store XA_ZERO_ENTRY and translate that to NULL on reading.
