On Thu, 2023-07-13 at 09:23 +0200, David Hildenbrand wrote:
> On 13.07.23 08:45, Verma, Vishal L wrote:
> >
> > I'm taking a shot at implementing the splitting internally in
> > memory_hotplug.c. The caller (kmem) side does become trivial with this
> > approach, but there's a slight complication if I don't have the module
> > param override (patch 1 of this series).
> >
> > The kmem diff now looks like:
> >
> > diff --git a/drivers/dax/kmem.c b/drivers/dax/kmem.c
> > index 898ca9505754..8be932f63f90 100644
> > --- a/drivers/dax/kmem.c
> > +++ b/drivers/dax/kmem.c
> > @@ -105,6 +105,8 @@ static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
> > data->mgid = rc;
> >
> > for (i = 0; i < dev_dax->nr_range; i++) {
> > + mhp_t mhp_flags = MHP_NID_IS_MGID | MHP_MEMMAP_ON_MEMORY |
> > + MHP_SPLIT_MEMBLOCKS;
> > struct resource *res;
> > struct range range;
> >
> > @@ -141,7 +143,7 @@ static int dev_dax_kmem_probe(struct dev_dax *dev_dax)
> > * this as RAM automatically.
> > */
> > rc = add_memory_driver_managed(data->mgid, range.start,
> > - range_len(&range), kmem_name, MHP_NID_IS_MGID);
> > + range_len(&range), kmem_name, mhp_flags);
> >
> > if (rc) {
> > dev_warn(dev, "mapping%d: %#llx-%#llx memory add failed\n",
> >
> >
>
> Why do we need the MHP_SPLIT_MEMBLOCKS?
I thought we still wanted either an opt-in or opt-out for the kmem
driver to be able to do memmap_on_memory, in case there were
performance implications or the lack of 1GiB PUDs. I haven't
implemented that yet, but I was thinking along the lines of a sysfs
knob exposed by kmem, that controls setting of this new
MHP_SPLIT_MEMBLOCKS flag.
>
> In add_memory_driver_managed(), if memmap_on_memory = 1 AND is effective for a
> single memory block, you can simply split up internally, no?
>
> Essentially in add_memory_resource() something like
>
> if (mhp_flags & MHP_MEMMAP_ON_MEMORY &&
> mhp_supports_memmap_on_memory(memory_block_size_bytes())) {
> for (cur_start = start, cur_start < start + size;
> cur_start += memory_block_size_bytes()) {
> mhp_altmap.free = PHYS_PFN(memory_block_size_bytes());
> mhp_altmap.base_pfn = PHYS_PFN(start);
> params.altmap = &mhp_altmap;
>
> ret = arch_add_memory(nid, start,
> memory_block_size_bytes(), ¶ms);
> if (ret < 0) ...
>
> ret = create_memory_block_devices(start, memory_block_size_bytes(),
> mhp_altmap.alloc, group);
> if (ret) ...
>
> }
> } else {
> /* old boring stuff */
> }
>
> Of course, doing it a bit cleaner, factoring out adding of mem+creating devices into
> a helper so we can use it on the other path, avoiding repeating memory_block_size_bytes()
> ...
My current approach was looping a level higher, on the call to
add_memory_resource, but this looks reasonable too, and I can switch to
this. In fact it is better as in my case I had to loop twice, once for
the regular add_memory() path and once for the _driver_managed() path.
Yours should avoid that.
>
> If any adding of memory failed, we remove what we already added. That works, because as
> long as we're holding the relevant locks, memory cannot get onlined in the meantime.
>
> Then we only have to teach remove_memory() to deal with individual blocks when finding
> blocks that have an altmap.
>
