Re: [PATCH 1/3] memory: extern memory_block_size_bytes and set_memory_block_size_order

[Gregory Price]

On Mon, Oct 14, 2024 at 10:32:36PM +0200, David Hildenbrand wrote:
> On 14.10.24 16:25, Gregory Price wrote:
> > On Mon, Oct 14, 2024 at 01:54:27PM +0200, David Hildenbrand wrote:
> > > On 08.10.24 17:21, Gregory Price wrote:
> > > > On Tue, Oct 08, 2024 at 05:02:33PM +0200, David Hildenbrand wrote:
> > > > > On 08.10.24 16:51, Gregory Price wrote:
> > > > > > > > +int __weak set_memory_block_size_order(unsigned int order)
> > > > > > > > +{
> > > > > > > > +	return -ENODEV;
> > > > > > > > +}
> > > > > > > > +EXPORT_SYMBOL_GPL(set_memory_block_size_order);
> > > > > > > 
> > > > > > > I can understand what you are trying to achieve, but letting arbitrary
> > > > > > > modules mess with this sounds like a bad idea.
> > > > > > > 
> > > > > > 
> > > > > > I suppose the alternative is trying to scan the CEDT from inside each
> > > > > > machine, rather than the ACPI driver?  Seems less maintainable.
> > > > > > 
> > > > > > I don't entirely disagree with your comment.  I hummed and hawwed over
> > > > > > externing this - hence the warning in the x86 machine.
> > > > > > 
> > > > > > Open to better answers.
> > > > > 
> > > > > Maybe an interface to add more restrictions on the maximum size might be
> > > > > better (instead of setting the size/order, you would impose another upper
> > > > > limit).
> > > > 
> > > > That is effectively what set_memory_block_size_order is, though.  Once
> > > > blocks are exposed to the allocators, its no longer safe to change the
> > > > size (in part because it was built assuming it wouldn't change, but I
> > > > imagine there are other dragons waiting in the shadows to bite me).
> > > 
> > > Yes, we must run very early.
> > > 
> > > How is this supposed to interact with code like
> > > 
> > > set_block_size()
> > > 
> > > that also calls set_memory_block_size_order() on UV systems (assuming there
> > > will be CXL support sooner or later?)?
> > > 
> > > 
> > 
> > Tying the other email to this one - just clarifying the way forward here.
> > 
> > It sounds like you're saying at a minimum drop EXPORT tags to prevent
> > modules from calling it - but it also sounds like built-ins need to be
> > prevented from touching it as well after a certain point in early boot.
> 
> Right, at least the EXPORT is not required.
> 
> > 
> > Do you think I should go down the advise() path as suggested by Ira,
> > just adding a arch_lock_blocksize() bit and have set_..._order check it,
> > or should we just move towards each architecture having to go through
> > the ACPI:CEDT itself?
> 
> Let's summarize what we currently have on x86 is:
> 
> 1) probe_memory_block_size()
> 
> Triggered on first memory_block_size_bytes() invocation. Makes a decision
> based on:
> 
> a) Already set size using set_memory_block_size_order()
> b) RAM size
> c) Bare metal vs. virt (bare metal -> use max)
> d) Virt: largest block size aligned to memory end
> 
> 
> 2) set_memory_block_size_order()
> 
> Triggered by set_block_size() on UV systems.
> 
> 
> I don't think set_memory_block_size_order() is the right tool to use. We
> just want to leave that alone I think -- it's a direct translation of a
> kernel cmdline parameter that should win.
> 
> You essentially want to tweak the b)->d) logic to take other alignment into
> consideration.
> 
> Maybe have some simple callback mechanism probe_memory_block_size() that can
> consult other sources for alignment requirements?
>

Thanks for this - I'll cobble something together.

Probably this ends up falling out similar to what Ira suggested. 

drivers/acpi/numa/srat.c
    acpi_numa_init():
        order = parse_cfwm(...)
        memblock_advise_size(order);

drivers/base/memory.c
    static int memblock_size_order = 0; /* let arch choose */

    int memblock_advise_size(order)
        int old_order;
        int new_order;
        if (order <= 0)
            return -EINVAL;

        do {
            old_order = memblock_size_order;
            new_order = MIN(old_order, order);
        } while (!atomic_cmpxchg(&memblock_size_order, old_order, new_order));

        /* memblock_size_order is now <= order, if -1 then the probe won */
        return new_order;

    int memblock_probe_size()
        return atomic_xchg(&memblock_size_order, -1);

drivers/base/memblock.h
    #ifdef HOTPLUG
        export memblock_advise_size()
        export memblock_probe_size()
    #else
        static memblock_advice_size() { return -ENODEV; } /* always fail */
        static memblock_probe_size() { return 0; } /* arch chooses */
    #endif

arch/*/mm/...
    probe_block_size():
        memblock_probe_size();
        /* select minimum across above suggested values */

> If that's not an option, then another way to set further min-alignment
> requirements (whereby we take MIN(old_align, new_align))?
> 
> -- 
> Cheers,
> 
> David / dhildenb
>