On Tue, Jul 08, 2025 at 09:26:49AM +0200, Peter Zijlstra wrote:
> On Mon, Jul 07, 2025 at 06:12:26PM +0300, Mike Rapoport wrote:
> > On Mon, Jul 07, 2025 at 11:06:25AM -0400, Liam R. Howlett wrote:
> > > * Mike Rapoport <r...@kernel.org> [250707 07:32]:
> > > > On Mon, Jul 07, 2025 at 01:11:02PM +0200, Peter Zijlstra wrote:
> > > > >
> > > > > err = __execmem_cache_free(&mas, ptr, GFP_KERNEL |
> > > > > __GFP_NORETRY);
> > > > > if (err) {
> > > > > mas_store_gfp(&mas, pending_free_set(ptr), GFP_KERNEL);
> > > > > execmem_cache.pending_free_cnt++;
> > > > > schedule_delayed_work(&execmem_cache_free_work,
> > > > > FREE_DELAY);
> > > > > return true;
> > > > > }
> > > > >
> > > > > schedule_work(&execmem_cache_clean_work);
> > > > > return true;
> > > > > }
> > > > >
> > > > > And now I have to ask what happens if mas_store_gfp() returns an
> > > > > error?
> > > >
> > > > AFAIU it won't. mas points to exact slot we've got the area from,
> > > > nothing else
> > > > can modify the tree because of the mutex, so that mas_store_gfp()
> > > > essentially updates the value at an existing entry.
> > > >
> > > > I'll add a comment about it.
> > > >
> > > > Added @Liam to make sure I'm not saying nonsense :)
> > > >
> > >
> > > Yes, if there is already a node with a value with the same range, there
> > > will be no allocations that will happen, so it'll just change the
> > > pointer for you. This is a slot store operation.
> > >
> > > But, if it's possible to have no entries (an empty tree, or a single
> > > value at 0), you will most likely allocate a node to store it, which is
> > > 256B.
> > >
> > > I don't think this is a concern in this particular case though as you
> > > are searching for an entry and storing, so it needs to exist. So
> > > really, the only scenario here is if you store 1 - ULONG_MAX (without
> > > having expanded a root node) or 0 - ULONG_MAX, and that seems invalid.
> >
> > Thanks for clarification, Liam!
> > The tree cannot be empty at that point and if it has a single value, it
> > won't be at 0, I'm quite sure no architecture has execmem areas at 0.
>
> Would it make sense to have something like GFP_NO_ALLOC to pass to
> functions like this where we know it won't actually allocate -- and
> which when it does reach the allocator generates a WARN and returns NULL
> ?
We can add a WARN at the caller as well, that won't require a new gfp flag.
The question is how to recover if such thing happen, I don't really see
what execmem can do here if mas_store_gfp() returns an error :/
--
Sincerely yours,
Mike.
