Re: [PATCH v3 04/14] mm/memremap: add ZONE_DEVICE support for compound pages

[Joao Martins <joao.m.martins@oracle.com>]

On 7/15/21 2:08 AM, Dan Williams wrote:
> On Wed, Jul 14, 2021 at 12:36 PM Joao Martins <joao.m.martins@oracle.com> wrote:
>>
>> Add a new align property for struct dev_pagemap which specifies that a
> 
> s/align/@geometry/
> 
Yeap, updated.

>> pagemap is composed of a set of compound pages of size @align,
> 
> s/@align/@geometry/
> 
Yeap, updated.

>> instead of
>> base pages. When a compound page geometry is requested, all but the first
>> page are initialised as tail pages instead of order-0 pages.
>>
>> For certain ZONE_DEVICE users like device-dax which have a fixed page size,
>> this creates an opportunity to optimize GUP and GUP-fast walkers, treating
>> it the same way as THP or hugetlb pages.
>>
>> Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
>> ---
>>  include/linux/memremap.h | 17 +++++++++++++++++
>>  mm/memremap.c            |  8 ++++++--
>>  mm/page_alloc.c          | 34 +++++++++++++++++++++++++++++++++-
>>  3 files changed, 56 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/linux/memremap.h b/include/linux/memremap.h
>> index 119f130ef8f1..e5ab6d4525c1 100644
>> --- a/include/linux/memremap.h
>> +++ b/include/linux/memremap.h
>> @@ -99,6 +99,10 @@ struct dev_pagemap_ops {
>>   * @done: completion for @internal_ref
>>   * @type: memory type: see MEMORY_* in memory_hotplug.h
>>   * @flags: PGMAP_* flags to specify defailed behavior
>> + * @geometry: structural definition of how the vmemmap metadata is populated.
>> + *     A zero or PAGE_SIZE defaults to using base pages as the memmap metadata
>> + *     representation. A bigger value but also multiple of PAGE_SIZE will set
>> + *     up compound struct pages representative of the requested geometry size.
>>   * @ops: method table
>>   * @owner: an opaque pointer identifying the entity that manages this
>>   *     instance.  Used by various helpers to make sure that no
>> @@ -114,6 +118,7 @@ struct dev_pagemap {
>>         struct completion done;
>>         enum memory_type type;
>>         unsigned int flags;
>> +       unsigned long geometry;
>>         const struct dev_pagemap_ops *ops;
>>         void *owner;
>>         int nr_range;
>> @@ -130,6 +135,18 @@ static inline struct vmem_altmap *pgmap_altmap(struct dev_pagemap *pgmap)
>>         return NULL;
>>  }
>>
>> +static inline unsigned long pgmap_geometry(struct dev_pagemap *pgmap)
>> +{
>> +       if (!pgmap || !pgmap->geometry)
>> +               return PAGE_SIZE;
>> +       return pgmap->geometry;
>> +}
>> +
>> +static inline unsigned long pgmap_pfn_geometry(struct dev_pagemap *pgmap)
>> +{
>> +       return PHYS_PFN(pgmap_geometry(pgmap));
>> +}
> 
> Are both needed? Maybe just have ->geometry natively be in nr_pages
> units directly, because pgmap_pfn_geometry() makes it confusing
> whether it's a geometry of the pfn or the geometry of the pgmap.
> 
I use pgmap_geometry() largelly when we manipulate memmap in sparse-vmemmap code, as we
deal with addresses/offsets/subsection-size. While using pgmap_pfn_geometry for code that
deals with PFN initialization. For this patch I could remove the confusion.

And actually maybe I can just store the pgmap_geometry() value in bytes locally in
vmemmap_populate_compound_pages() and we can remove this extra helper.

>> +
>>  #ifdef CONFIG_ZONE_DEVICE
>>  bool pfn_zone_device_reserved(unsigned long pfn);
>>  void *memremap_pages(struct dev_pagemap *pgmap, int nid);
>> diff --git a/mm/memremap.c b/mm/memremap.c
>> index 805d761740c4..ffcb924eb6a5 100644
>> --- a/mm/memremap.c
>> +++ b/mm/memremap.c
>> @@ -318,8 +318,12 @@ static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
>>         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
>>                                 PHYS_PFN(range->start),
>>                                 PHYS_PFN(range_len(range)), pgmap);
>> -       percpu_ref_get_many(pgmap->ref, pfn_end(pgmap, range_id)
>> -                       - pfn_first(pgmap, range_id));
>> +       if (pgmap_geometry(pgmap) > PAGE_SIZE)
> 
> This would become
> 
> if (pgmap_geometry(pgmap) > 1)
> 
>> +               percpu_ref_get_many(pgmap->ref, (pfn_end(pgmap, range_id)
>> +                       - pfn_first(pgmap, range_id)) / pgmap_pfn_geometry(pgmap));
> 
> ...and this would be pgmap_geometry()
> 
>> +       else
>> +               percpu_ref_get_many(pgmap->ref, pfn_end(pgmap, range_id)
>> +                               - pfn_first(pgmap, range_id));
>>         return 0;
>>
Let me adjust accordingly.

>>  err_add_memory:
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 79f3b38afeca..188cb5f8c308 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -6597,6 +6597,31 @@ static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
>>         }
>>  }
>>
>> +static void __ref memmap_init_compound(struct page *page, unsigned long pfn,
> 
> I'd feel better if @page was renamed @head... more below:
> 
Oh yeah -- definitely more readable.

>> +                                       unsigned long zone_idx, int nid,
>> +                                       struct dev_pagemap *pgmap,
>> +                                       unsigned long nr_pages)
>> +{
>> +       unsigned int order_align = order_base_2(nr_pages);
>> +       unsigned long i;
>> +
>> +       __SetPageHead(page);
>> +
>> +       for (i = 1; i < nr_pages; i++) {
> 
> The switch of loop styles is jarring. I.e. the switch from
> memmap_init_zone_device() that is using pfn, end_pfn, and a local
> 'struct page *' variable to this helper using pfn + i and a mix of
> helpers (__init_zone_device_page,  prep_compound_tail) that have
> different expectations of head page + tail_idx and current page.
> 
> I.e. this reads more obviously correct to me, but maybe I'm just in
> the wrong headspace:
> 
>         for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
>                 struct page *page = pfn_to_page(pfn);
> 
>                 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
>                 prep_compound_tail(head, pfn - head_pfn);
> 
Personally -- and I am dubious given I have been staring at this code -- I find that what
I wrote a little better as it follows more what compound page initialization does. Like
it's easier for me to read that I am initializing a number of tail pages and a head page
(for a known geometry size).

Additionally, it's unnecessary (and a tiny ineficient?) to keep doing pfn_to_page(pfn)
provided ZONE_DEVICE requires SPARSEMEM_VMEMMAP and so your page pointers are all
contiguous and so for any given PFN we can avoid having deref vmemmap vaddrs back and
forth. Which is the second reason I pass a page, and iterate over its tails based on a
head page pointer. But I was at too minds when writing this, so if the there's no added
inefficiency I can rewrite like the above.

>> +               __init_zone_device_page(page + i, pfn + i, zone_idx,
>> +                                       nid, pgmap);
>> +               prep_compound_tail(page, i);
>> +
>> +               /*
>> +                * The first and second tail pages need to
>> +                * initialized first, hence the head page is
>> +                * prepared last.
> 
> I'd change this comment to say why rather than restate what can be
> gleaned from the code. It's actually not clear to me why this order is
> necessary.
> 
So the first tail page stores mapcount_ptr and compound order, and the
second tail page stores pincount_ptr. prep_compound_head() does this:

	set_compound_order(page, order);
	atomic_set(compound_mapcount_ptr(page), -1);
	if (hpage_pincount_available(page))
		atomic_set(compound_pincount_ptr(page), 0);

So we need those tail pages initialized first prior to initializing the head.

I can expand the comment above to make it clear why we need first and second tail pages.

>> +                */
>> +               if (i == 2)
>> +                       prep_compound_head(page, order_align);
>> +       }
>> +}
>> +
>>  void __ref memmap_init_zone_device(struct zone *zone,
>>                                    unsigned long start_pfn,
>>                                    unsigned long nr_pages,
>> @@ -6605,6 +6630,7 @@ void __ref memmap_init_zone_device(struct zone *zone,
>>         unsigned long pfn, end_pfn = start_pfn + nr_pages;
>>         struct pglist_data *pgdat = zone->zone_pgdat;
>>         struct vmem_altmap *altmap = pgmap_altmap(pgmap);
>> +       unsigned int pfns_per_compound = pgmap_pfn_geometry(pgmap);
>>         unsigned long zone_idx = zone_idx(zone);
>>         unsigned long start = jiffies;
>>         int nid = pgdat->node_id;
>> @@ -6622,10 +6648,16 @@ void __ref memmap_init_zone_device(struct zone *zone,
>>                 nr_pages = end_pfn - start_pfn;
>>         }
>>
>> -       for (pfn = start_pfn; pfn < end_pfn; pfn++) {
>> +       for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
>>                 struct page *page = pfn_to_page(pfn);
>>
>>                 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
>> +
>> +               if (pfns_per_compound == 1)
>> +                       continue;
>> +
>> +               memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
>> +                                    pfns_per_compound);
> 
> I otherwise don't see anything broken with this patch, so feel free to include:
> 
> Reviewed-by: Dan Williams <dan.j.williams@intel.com>
> 
> ...on the resend with the fixups.
> 
Thanks.

I will wait whether you still want to retain the tag provided the implied changes
fixing the failure you reported.