Re: [PATCH] Reinstate "resource: avoid unnecessary lookups in find_next_iomem_res()"

["andriy.shevchenko@linux.intel.com" <andriy.shevchenko@linux.intel.com>]

On Mon, Nov 24, 2025 at 11:30:46PM +0000, Stamatis, Ilias wrote:
> On Mon, 2025-11-24 at 21:52 +0200, andriy.shevchenko@linux.intel.com wrote:
> > On Mon, Nov 24, 2025 at 07:35:31PM +0000, Stamatis, Ilias wrote:
> > > On Mon, 2025-11-24 at 20:55 +0200, andriy.shevchenko@linux.intel.com wrote:
> > > > On Mon, Nov 24, 2025 at 06:01:35PM +0000, Stamatis, Ilias wrote:
> > > > > On Mon, 2025-11-24 at 08:58 -0800, Andrew Morton wrote:
> > > > > > On Mon, 24 Nov 2025 16:53:49 +0000 Ilias Stamatis <ilstam@amazon.com> wrote:
> > > > > > 
> > > > > > > Commit 97523a4edb7b ("kernel/resource: remove first_lvl / siblings_only
> > > > > > > logic") removed an optimization introduced by commit 756398750e11
> > > > > > > ("resource: avoid unnecessary lookups in find_next_iomem_res()"). That
> > > > > > > was not called out in the message of the first commit explicitly so it's
> > > > > > > not entirely clear whether removing the optimization happened
> > > > > > > inadvertently or not.
> > > > > > > 
> > > > > > > As the original commit message of the optimization explains there is no
> > > > > > > point considering the children of a subtree in find_next_iomem_res() if
> > > > > > > the top level range does not match. Reinstating the optimization results
> > > > > > > in significant performance improvements in systems with very large iomem
> > > > > > > maps when mmaping /dev/mem.
> > > > > > 
> > > > > > It would be great if we could quantify "significant performance
> > > > > > improvements"?
> > > > > 
> > > > > Hi Andrew and Andy,
> > > > > 
> > > > > You are right to call that out and apologies for leaving it vague.
> > > > > 
> > > > > I've done my testing with older kernel versions in systems where `wc -l
> > > > > /proc/iomem` can return ~5k. In that environment I see mmaping parts of
> > > > > /dev/mem taking 700-1500μs without the optimisation and 10-50μs with the
> > > > > optimisation.
> > > > > 
> > > > > The real-world use case we care about is hypervisor live update where having to
> > > > > do lots of these mmaps() serially can significantly affect the guest downtime
> > > > > if the cost is 20-30x.
> > > > 
> > > > Thanks for providing this information.
> > > > 
> > > > > > It also would be good to know which exact function(s) is a bottleneck.
> > > > > 
> > > > > Perf tracing shows that ~95% of CPU time is spent in find_next_iomem_res(),
> > > > 
> > > > Have you investigated possibility to return that check directly into
> > > > the culprit?
> > > 
> > > I'm sorry, I don't understand this. Could you please clarify what you mean?
> > > What do you consider to be the culprit and which check do you refer to?
> > 
> > The mentioned patch removed the check for siblings from next_resource().
> > The function that your test case complains about is find_next_iomem_res().
> > Hence, have you tried to reinstantiate the (removed) check from next_resource()
> > in find_next_iomem_res() and see if it helps?
> 
> next_resource() does accept a 'skip_children' parameter in the latest kernel
> today which is equivalent to the 'sibling_only' parameter in the older
> kernels.

It used to be

       if (sibling_only)
               return p->sibling;

       if (p->child)
               return p->child;
       ...


and become (in the latest kernels)

       if (!skip_children && p->child)
               return p->child;
       ...


Can you elaborate how are they interoperable?

TL;DR: I don't think it's an equivalent.

> And the for_each_resource() macro currently used in find_next_iomem_res()
> calls next_resource().
> 
> Hope that makes sense.

-- 
With Best Regards,
Andy Shevchenko