Re: [PATCH v4 2/2] mm: Init page count in reserve_bootmem_region when MEMINIT_EARLY

From: David Hildenbrand
Date: Mon Oct 02 2023 - 04:31:16 EST


On 29.09.23 10:30, Mike Rapoport wrote:
On Thu, Sep 28, 2023 at 04:33:02PM +0800, Yajun Deng wrote:
memmap_init_range() would init page count of all pages, but the free
pages count would be reset in __free_pages_core(). There are opposite
operations. It's unnecessary and time-consuming when it's MEMINIT_EARLY
context.

Init page count in reserve_bootmem_region when in MEMINIT_EARLY context,
and check the page count before reset it.

At the same time, the INIT_LIST_HEAD in reserve_bootmem_region isn't
need, as it already done in __init_single_page.

The following data was tested on an x86 machine with 190GB of RAM.

before:
free_low_memory_core_early() 341ms

after:
free_low_memory_core_early() 285ms

Signed-off-by: Yajun Deng <yajun.deng@xxxxxxxxx>
---
v4: same with v2.
v3: same with v2.
v2: check page count instead of check context before reset it.
v1: https://lore.kernel.org/all/20230922070923.355656-1-yajun.deng@xxxxxxxxx/
---
mm/mm_init.c | 18 +++++++++++++-----
mm/page_alloc.c | 20 ++++++++++++--------
2 files changed, 25 insertions(+), 13 deletions(-)

diff --git a/mm/mm_init.c b/mm/mm_init.c
index 9716c8a7ade9..3ab8861e1ef3 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -718,7 +718,7 @@ static void __meminit init_reserved_page(unsigned long pfn, int nid)
if (zone_spans_pfn(zone, pfn))
break;
}
- __init_single_page(pfn_to_page(pfn), pfn, zid, nid, INIT_PAGE_COUNT);
+ __init_single_page(pfn_to_page(pfn), pfn, zid, nid, 0);
}
#else
static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
@@ -756,8 +756,8 @@ void __meminit reserve_bootmem_region(phys_addr_t start,
init_reserved_page(start_pfn, nid);
- /* Avoid false-positive PageTail() */
- INIT_LIST_HEAD(&page->lru);
+ /* Init page count for reserved region */

Please add a comment that describes _why_ we initialize the page count here.

+ init_page_count(page);
/*
* no need for atomic set_bit because the struct
@@ -888,9 +888,17 @@ void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone
}
page = pfn_to_page(pfn);
- __init_single_page(page, pfn, zone, nid, INIT_PAGE_COUNT);
- if (context == MEMINIT_HOTPLUG)
+
+ /* If the context is MEMINIT_EARLY, we will init page count and
+ * mark page reserved in reserve_bootmem_region, the free region
+ * wouldn't have page count and we will check the pages count
+ * in __free_pages_core.
+ */
+ __init_single_page(page, pfn, zone, nid, 0);
+ if (context == MEMINIT_HOTPLUG) {
+ init_page_count(page);
__SetPageReserved(page);

Rather than calling init_page_count() and __SetPageReserved() for
MEMINIT_HOTPLUG you can set flags to INIT_PAGE_COUNT | INIT_PAGE_RESERVED
an call __init_single_page() after the check for MEMINIT_HOTPLUG.

But more generally, I wonder if we have to differentiate HOTPLUG here at all.
@David, can you comment please?

There are a lot of details to that, and I'll share some I can briefly think of.

1) __SetPageReserved()

I tried removing that a while ago, but there was a blocker (IIRC something about
ZONE_DEVICE). I still have the patches at [1] and I could probably take a look
if that blocker still exists (I recall that something changed at some point, but
I never had the time to follow up).

But once we stop setting the pages reserved, we might run into issues with ...


2) init_page_count()

virtio-mem, XEN balloon and HV-balloon add memory blocks that can contain holes.
set_online_page_callback() is used to intercept memory onlining and to expose
only the pages that are not holes to the buddy: calling generic_online_page() on !hole.

Holes are PageReserved but with an initialized page count. Memory offlining will fail on
PageReserved pages -- has_unmovable_pages().


At least virtio-mem clears the PageReserved flag of holes when onlining memory,
and currently relies in the page count to be reasonable (so memory offlining can work).

static void virtio_mem_set_fake_offline(unsigned long pfn,
unsigned long nr_pages, bool onlined)
{
page_offline_begin();
for (; nr_pages--; pfn++) {
struct page *page = pfn_to_page(pfn);

__SetPageOffline(page);
if (!onlined) {
SetPageDirty(page);
/* FIXME: remove after cleanups */
ClearPageReserved(page);
}
}
page_offline_end();
}


For virtio-mem, we could initialize the page count there instead. The other PV drivers
might require a bit more thought.


[1] https://github.com/davidhildenbrand/linux/tree/online_reserved_cleanup


+ }
/*
* Usually, we want to mark the pageblock MIGRATE_MOVABLE,
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 06be8821d833..b868caabe8dc 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1285,18 +1285,22 @@ void __free_pages_core(struct page *page, unsigned int order)
unsigned int loop;
/*
- * When initializing the memmap, __init_single_page() sets the refcount
- * of all pages to 1 ("allocated"/"not free"). We have to set the
- * refcount of all involved pages to 0.
+ * When initializing the memmap, memmap_init_range sets the refcount
+ * of all pages to 1 ("reserved" and "free") in hotplug context. We
+ * have to set the refcount of all involved pages to 0. Otherwise,
+ * we don't do it, as reserve_bootmem_region only set the refcount on
+ * reserve region ("reserved") in early context.
*/

Again, why hotplug and early init should be different?

- prefetchw(p);
- for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
- prefetchw(p + 1);
+ if (page_count(page)) {
+ prefetchw(p);
+ for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
+ prefetchw(p + 1);
+ __ClearPageReserved(p);
+ set_page_count(p, 0);
+ }
__ClearPageReserved(p);
set_page_count(p, 0);

That looks wrong. if the page count would by pure luck be 0 already for hotplugged memory,
you wouldn't clear the reserved flag.

These changes make me a bit nervous.

--
Cheers,

David / dhildenb