THP Shrinker

From: Alexander Zhu <[email protected]>

Currently, when /sys/kernel/mm/transparent_hugepage/enabled=always is set there are a large number of transparent hugepages that are almost entirely zero filled. This is mentioned in a number of previous patchsets including:
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/all/
[email protected]/

Currently, split_huge_page() does not have a way to identify zero filled pages within the THP. Thus these zero pages get remapped and continue to create memory waste. In this patch, we identify and free tail pages that are zero filled in split_huge_page(). In this way, we avoid mapping these pages back into page table entries and can free up unused memory within THPs. This is based off the previously mentioned patchset by Yu Zhao. However, we chose to free zero tail pages whenever they are encountered instead of only on reclaim or migration. We also add a self test to verify the RssAnon value to make sure zero pages are not remapped.

!-------------------------------------------------------------------|
This Message Is From an External Sender

|-------------------------------------------------------------------!

Quoted 10 lines · On 25.08.22 23:30, [email protected] wrote:

From: Alexander Zhu <[email protected]>

Currently, when /sys/kernel/mm/transparent_hugepage/enabled=always is set there are a large number of transparent hugepages that are almost entirely zero filled. This is mentioned in a number of previous patchsets including:
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/all/
[email protected]/

Quoted 10 lines

Currently, split_huge_page() does not have a way to identify zero filled pages within the THP. Thus these zero pages get remapped and continue to create memory waste. In this patch, we identify and free tail pages that are zero filled in split_huge_page(). In this way, we avoid mapping these pages back into page table entries and can free up unused memory within THPs. This is based off the previously mentioned patchset by Yu Zhao. However, we chose to free zero tail pages whenever they are encountered instead of only on reclaim or migration. We also add a self test to verify the RssAnon value to make sure zero pages are not remapped.

Isn't this to some degree splitting the THP (PMDs->PTEs + dissolve compound page) and then letting KSM replace the zero-filled page by the shared zeropage?

Signature · 3 lines

Thanks,

David / dhildenb

Quoted 14 lines · On 25.08.22 23:30, [email protected] wrote:

From: Alexander Zhu <[email protected]>

Currently, split_huge_page() does not have a way to identify zero filled
pages within the THP. Thus these zero pages get remapped and continue to
create memory waste. In this patch, we identify and free tail pages that
are zero filled in split_huge_page(). In this way, we avoid mapping these
pages back into page table entries and can free up unused memory within
THPs.

Isn't this to some degree splitting the THP (PMDs->PTEs + dissolve compound page) and then letting KSM replace the zero-filled page by the
shared zeropage?

From: Alexander Zhu <[email protected]>

This change introduces a tool that scans through all of physical memory for anonymous THPs and groups them into buckets based on utilization. It also includes an interface under /sys/kernel/debug/thp_utilization.

Sample Output:

Utilized[0-50]: 1331 680884
Utilized[51-101]: 9 3983
Utilized[102-152]: 3 1187
Utilized[153-203]: 0 0
Utilized[204-255]: 2 539
Utilized[256-306]: 5 1135
Utilized[307-357]: 1 192
Utilized[358-408]: 0 0
Utilized[409-459]: 1 57
Utilized[460-512]: 400 13
Last Scan Time: 223.98
Last Scan Duration: 70.65

This indicates that there are 1331 THPs that have between 0 and 50 utilized (non zero) pages. In total there are 680884 zero pages in this utilization bucket. THPs in the [0-50] bucket compose 76% of total THPs, and are responsible for 99% of total zero pages across all THPs. In other words, the least utilized THPs are responsible for almost all of the memory waste when THP is always enabled. Similar results have been observed across production workloads.

The last two lines indicate the timestamp and duration of the most recent scan through all of physical memory. Here we see that the last scan occurred 223.98 seconds after boot time and took 70.65 seconds.

Utilization of a THP is defined as the percentage of nonzero pages in the THP. The worker thread will scan through all of physical memory and obtain utilization of all anonymous THPs. It will gather this information by periodically scanning through all of physical memory for anonymous THPs, group them into buckets based on utilization, and report utilization information through debugfs under /sys/kernel/debug/thp_utilization.

Signed-off-by Alexander Zhu <[email protected]>

---
 Documentation/admin-guide/mm/transhuge.rst |   9 +
 include/linux/huge_mm.h                    |   2 +
 mm/huge_memory.c                           | 198 +++++++++++++++++++++
 3 files changed, 209 insertions(+)

Documentation/admin-guide/mm/transhuge.rst +9 -0

index c9c37f16eef8..d883ff9fddc7 100644
@@ -297,6 +297,15 @@ To identify what applications are mapping file transparent huge pages, it
 is necessary to read ``/proc/PID/smaps`` and count the FileHugeMapped fields
 for each mapping.
 
+The utilization of transparent hugepages can be viewed by reading
+``/sys/kernel/debug/thp_utilization``. The utilization of a THP is defined
+as the ratio of non zero filled 4kb pages to the total number of pages in a
+THP. The buckets are labelled by the range of total utilized 4kb pages with
+one line per utilization bucket. Each line contains the total number of
+THPs in that bucket and the total number of zero filled 4kb pages summed
+over all THPs in that bucket. The last two lines show the timestamp and
+duration respectively of the most recent scan over all of physical memory.
+
 Note that reading the smaps file is expensive and reading it
 frequently will incur overhead.
 

include/linux/huge_mm.h +2 -0

index 768e5261fdae..c9086239deb7 100644
@@ -179,6 +179,8 @@ bool hugepage_vma_check(struct vm_area_struct *vma,
 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 		unsigned long len, unsigned long pgoff, unsigned long flags);
 
+int thp_number_utilized_pages(struct page *page);
+
 void prep_transhuge_page(struct page *page);
 void free_transhuge_page(struct page *page);
 

mm/huge_memory.c +49 -0

index 8a7c1b344abe..8be1e320e70c 100644
@@ -45,6 +45,21 @@
 #define CREATE_TRACE_POINTS
 #include <trace/events/thp.h>
 
+/*
+ * The number of utilization buckets THPs will be grouped in
+ * under /sys/kernel/debug/thp_utilization.
+ */
+#define THP_UTIL_BUCKET_NR 10
+/*
+ * The maximum number of hugepages to scan through on each periodic
+ * run of the scanner that generates /sys/kernel/debug/thp_utilization.
+ * We scan through physical memory in chunks of size PMD_SIZE and
+ * record the timestamp and duration of each scan. In practice we have
+ * found that scanning THP_UTIL_SCAN_SIZE hugepages per second is sufficient
+ * for obtaining useful utilization metrics and does not have a noticeable
+ * impact on CPU.
+ */
+#define THP_UTIL_SCAN_SIZE 256
 /*
  * By default, transparent hugepage support is disabled in order to avoid
  * risking an increased memory footprint for applications that are not

@@ -70,6 +85,25 @@ static atomic_t huge_zero_refcount;
 struct page *huge_zero_page __read_mostly;
 unsigned long huge_zero_pfn __read_mostly = ~0UL;
 
+static void thp_utilization_workfn(struct work_struct *work);
+static DECLARE_DELAYED_WORK(thp_utilization_work, thp_utilization_workfn);
+
+struct thp_scan_info_bucket {
+	int nr_thps;
+	int nr_zero_pages;
+};
+
+struct thp_scan_info {
+	struct thp_scan_info_bucket buckets[THP_UTIL_BUCKET_NR];
+	struct zone *scan_zone;
+	struct timespec64 last_scan_duration;
+	struct timespec64 last_scan_time;
+	unsigned long pfn;
+};
+
+static struct thp_scan_info thp_scan_debugfs;
+static struct thp_scan_info thp_scan;
+
 bool hugepage_vma_check(struct vm_area_struct *vma,
 			unsigned long vm_flags,
 			bool smaps, bool in_pf)

@@ -486,6 +520,7 @@ static int __init hugepage_init(void)
 	if (err)
 		goto err_slab;
 
+	schedule_delayed_work(&thp_utilization_work, HZ);
 	err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 	if (err)
 		goto err_hzp_shrinker;

@@ -600,6 +635,11 @@ static inline bool is_transparent_hugepage(struct page *page)
 	       page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
 }
 
+static inline bool is_anon_transparent_hugepage(struct page *page)
+{
+	return PageAnon(page) && is_transparent_hugepage(page);
+}
+
 static unsigned long __thp_get_unmapped_area(struct file *filp,
 		unsigned long addr, unsigned long len,
 		loff_t off, unsigned long flags, unsigned long size)

@@ -650,6 +690,38 @@ unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 }
 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 
+int thp_number_utilized_pages(struct page *page)
+{
+	struct folio *folio;
+	unsigned long page_offset, value;
+	int thp_nr_utilized_pages = HPAGE_PMD_NR;
+	int step_size = sizeof(unsigned long);
+	bool is_all_zeroes;
+	void *kaddr;
+	int i;

mm/huge_memory.c +139 -0

+			if (value != 0) {
+				is_all_zeroes = false;
+				break;
+			}
+		}
+		if (is_all_zeroes)
+			thp_nr_utilized_pages--;
+
+		kunmap_local(kaddr);
+	}
+	return thp_nr_utilized_pages;
+}
+
 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 			struct page *page, gfp_t gfp)
 {

@@ -3135,6 +3207,42 @@ static int __init split_huge_pages_debugfs(void)
 	return 0;
 }
 late_initcall(split_huge_pages_debugfs);
+
+static int thp_utilization_show(struct seq_file *seqf, void *pos)
+{
+	int i;
+	int start;
+	int end;
+
+	for (i = 0; i < THP_UTIL_BUCKET_NR; i++) {
+		start = i * HPAGE_PMD_NR / THP_UTIL_BUCKET_NR;
+		end = (i + 1 == THP_UTIL_BUCKET_NR)
+			   ? HPAGE_PMD_NR
+			   : ((i + 1) * HPAGE_PMD_NR / THP_UTIL_BUCKET_NR - 1);
+		/* The last bucket will need to contain 100 */
+		seq_printf(seqf, "Utilized[%d-%d]: %d %d\n", start, end,
+			   thp_scan_debugfs.buckets[i].nr_thps,
+			   thp_scan_debugfs.buckets[i].nr_zero_pages);
+	}
+	seq_printf(seqf, "Last Scan Time: %lu.%02lu\n",
+		   (unsigned long)thp_scan_debugfs.last_scan_time.tv_sec,
+		   (thp_scan_debugfs.last_scan_time.tv_nsec / (NSEC_PER_SEC / 100)));
+
+	seq_printf(seqf, "Last Scan Duration: %lu.%02lu\n",
+		   (unsigned long)thp_scan_debugfs.last_scan_duration.tv_sec,
+		   (thp_scan_debugfs.last_scan_duration.tv_nsec / (NSEC_PER_SEC / 100)));
+
+	return 0;
+}
+DEFINE_SHOW_ATTRIBUTE(thp_utilization);
+
+static int __init thp_utilization_debugfs(void)
+{
+	debugfs_create_file("thp_utilization", 0200, NULL, NULL,
+			    &thp_utilization_fops);
+	return 0;
+}
+late_initcall(thp_utilization_debugfs);
 #endif
 
 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION

@@ -3220,3 +3328,93 @@ void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
 	trace_remove_migration_pmd(address, pmd_val(pmde));
 }
 #endif
+
+static void thp_scan_next_zone(void)
+{
+	struct timespec64 current_time;
+	int i;
+	bool update_debugfs;
+	/*
+	 * THP utilization worker thread has reached the end
+	 * of the memory zone. Proceed to the next zone.
+	 */
+	thp_scan.scan_zone = next_zone(thp_scan.scan_zone);
+	update_debugfs = !thp_scan.scan_zone;
+	thp_scan.scan_zone = update_debugfs ? (first_online_pgdat())->node_zones
+			: thp_scan.scan_zone;
+	thp_scan.pfn = (thp_scan.scan_zone->zone_start_pfn + HPAGE_PMD_NR - 1)
+			& ~(HPAGE_PMD_SIZE - 1);
+	if (!update_debugfs)
+		return;
+	/*
+	 * If the worker has scanned through all of physical
+	 * memory. Then update information displayed in /sys/kernel/debug/thp_utilization
+	 */
+	ktime_get_ts64(&current_time);
+	thp_scan_debugfs.last_scan_duration = timespec64_sub(current_time,
+							     thp_scan_debugfs.last_scan_time);
+	thp_scan_debugfs.last_scan_time = current_time;
+
+	for (i = 0; i < THP_UTIL_BUCKET_NR; i++) {
+		thp_scan_debugfs.buckets[i].nr_thps = thp_scan.buckets[i].nr_thps;
+		thp_scan_debugfs.buckets[i].nr_zero_pages = thp_scan.buckets[i].nr_zero_pages;
+		thp_scan.buckets[i].nr_thps = 0;
+		thp_scan.buckets[i].nr_zero_pages = 0;
+	}
+}
+
+static void thp_util_scan(unsigned long pfn_end)
+{
+	struct page *page = NULL;
+	int bucket, num_utilized_pages, current_pfn;
+	int i;
+	/*
+	 * Scan through each memory zone in chunks of up to THP_UTIL_SCAN_SIZE
+	 * hugepages every second looking for anonymous THPs.
+	 */
+	for (i = 0; i < THP_UTIL_SCAN_SIZE; i++) {
+		current_pfn = thp_scan.pfn;
+		thp_scan.pfn += HPAGE_PMD_NR;
+		if (current_pfn >= pfn_end)
+			return;
+
+		if (!pfn_valid(current_pfn))
+			continue;
+
+		page = pfn_to_page(current_pfn);
+		num_utilized_pages = thp_number_utilized_pages(page);
+		 /* Not a THP; skip it. */
+		if (num_utilized_pages < 0)
+			continue;
+		/* Group THPs into utilization buckets */
+		bucket = num_utilized_pages * THP_UTIL_BUCKET_NR / HPAGE_PMD_NR;
+		bucket = min(bucket, THP_UTIL_BUCKET_NR - 1);
+		thp_scan.buckets[bucket].nr_thps++;
+		thp_scan.buckets[bucket].nr_zero_pages += (HPAGE_PMD_NR - num_utilized_pages);
+	}
+}
+
+static void thp_utilization_workfn(struct work_struct *work)
+{
+	unsigned long pfn_end;
+
+	if (!thp_scan.scan_zone)
+		thp_scan.scan_zone = (first_online_pgdat())->node_zones;
+	/*
+	 * Worker function that scans through all of physical memory
+	 * for anonymous THPs.
+	 */
+	pfn_end = (thp_scan.scan_zone->zone_start_pfn +
+			thp_scan.scan_zone->spanned_pages + HPAGE_PMD_NR - 1)
+			& ~(HPAGE_PMD_SIZE - 1);
+	/* If we have reached the end of the zone or end of physical memory
+	 * move on to the next zone. Otherwise, scan the next PFNs in the
+	 * current zone.
+	 */
+	if (!populated_zone(thp_scan.scan_zone) || thp_scan.pfn >= pfn_end)
+		thp_scan_next_zone();
+	else
+		thp_util_scan(pfn_end);
+
+	schedule_delayed_work(&thp_utilization_work, HZ);
+}

Signature · 1 line

2.30.2

From: Alexander Zhu <[email protected]>

This patch introduces a shrinker that will remove THPs in the lowest utilization bucket. As previously mentioned, we have observed that almost all of the memory waste when THPs are always enabled is contained in the lowest utilization bucket. The shrinker will add these THPs to a list_lru and split anonymous THPs based off information from kswapd. It requires the changes from thp_utilization to identify the least utilized THPs, and the changes to split_huge_page to identify and free zero pages within THPs.

Signed-off-by Alexander Zhu <[email protected]>

---
 include/linux/huge_mm.h  |  7 +++
 include/linux/list_lru.h | 24 +++++++++++
 include/linux/mm_types.h |  5 +++
 mm/huge_memory.c         | 92 ++++++++++++++++++++++++++++++++++++++--
 mm/list_lru.c            | 49 +++++++++++++++++++++
 mm/page_alloc.c          |  6 +++
 6 files changed, 180 insertions(+), 3 deletions(-)

include/linux/huge_mm.h +7 -0

index c9086239deb7..13bd470173d2 100644
@@ -192,6 +192,8 @@ static inline int split_huge_page(struct page *page)
 }
 void deferred_split_huge_page(struct page *page);
 
+void add_underutilized_thp(struct page *page);
+
 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 		unsigned long address, bool freeze, struct folio *folio);
 

@@ -302,6 +304,11 @@ static inline struct list_head *page_deferred_list(struct page *page)
 	return &page[2].deferred_list;
 }
 
+static inline struct list_head *page_underutilized_thp_list(struct page *page)
+{
+	return &page[3].underutilized_thp_list;
+}
+
 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })

include/linux/list_lru.h +24 -0

index b35968ee9fb5..c2cf146ea880 100644
@@ -89,6 +89,18 @@ void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *paren
  */
 bool list_lru_add(struct list_lru *lru, struct list_head *item);
 
+/**
+ * list_lru_add_page: add an element to the lru list's tail
+ * @list_lru: the lru pointer
+ * @page: the page containing the item
+ * @item: the item to be deleted.
+ *
+ * This function works the same as list_lru_add in terms of list
+ * manipulation. Used for non slab objects contained in the page.
+ *
+ * Return value: true if the list was updated, false otherwise
+ */
+bool list_lru_add_page(struct list_lru *lru, struct page *page, struct list_head *item);
 /**
  * list_lru_del: delete an element to the lru list
  * @list_lru: the lru pointer

@@ -102,6 +114,18 @@ bool list_lru_add(struct list_lru *lru, struct list_head *item);
  */
 bool list_lru_del(struct list_lru *lru, struct list_head *item);
 
+/**
+ * list_lru_del_page: delete an element to the lru list
+ * @list_lru: the lru pointer
+ * @page: the page containing the item
+ * @item: the item to be deleted.
+ *
+ * This function works the same as list_lru_del in terms of list
+ * manipulation. Used for non slab objects contained in the page.
+ *
+ * Return value: true if the list was updated, false otherwise
+ */
+bool list_lru_del_page(struct list_lru *lru, struct page *page, struct list_head *item);
 /**
  * list_lru_count_one: return the number of objects currently held by @lru
  * @lru: the lru pointer.

include/linux/mm_types.h +5 -0

index cf97f3884fda..05667a2030c0 100644
@@ -151,6 +151,11 @@ struct page {
 			/* For both global and memcg */
 			struct list_head deferred_list;
 		};
+		struct { /* Third tail page of compound page */
+			unsigned long _compound_pad_3; /* compound_head */
+			unsigned long _compound_pad_4;
+			struct list_head underutilized_thp_list;
+		};
 		struct {	/* Page table pages */
 			unsigned long _pt_pad_1;	/* compound_head */
 			pgtable_t pmd_huge_pte; /* protected by page->ptl */

mm/huge_memory.c +89 -3

index 0f774a7c0727..03dc42eba0ba 100644
@@ -8,6 +8,7 @@
 #include <linux/mm.h>
 #include <linux/sched.h>
 #include <linux/sched/mm.h>
+#include <linux/sched/clock.h>
 #include <linux/sched/coredump.h>
 #include <linux/sched/numa_balancing.h>
 #include <linux/highmem.h>

@@ -85,6 +86,8 @@ static atomic_t huge_zero_refcount;
 struct page *huge_zero_page __read_mostly;
 unsigned long huge_zero_pfn __read_mostly = ~0UL;
 
+struct list_lru huge_low_util_page_lru;
+
 static void thp_utilization_workfn(struct work_struct *work);
 static DECLARE_DELAYED_WORK(thp_utilization_work, thp_utilization_workfn);
 

@@ -269,6 +272,46 @@ static struct shrinker huge_zero_page_shrinker = {
 	.seeks = DEFAULT_SEEKS,
 };
 
+static enum lru_status low_util_free_page(struct list_head *item,
+					  struct list_lru_one *lru,
+					  spinlock_t *lock,
+					  void *cb_arg)
+{
+	struct page *head = compound_head(list_entry(item,
+									struct page,
+									underutilized_thp_list));
+
+	if (get_page_unless_zero(head)) {
+		lock_page(head);
+		list_lru_isolate(lru, item);
+		split_huge_page(head);
+		unlock_page(head);
+		put_page(head);
+	}
+
+	return LRU_REMOVED_RETRY;
+}
+
+static unsigned long shrink_huge_low_util_page_count(struct shrinker *shrink,
+						     struct shrink_control *sc)
+{
+	return list_lru_shrink_count(&huge_low_util_page_lru, sc);
+}
+
+static unsigned long shrink_huge_low_util_page_scan(struct shrinker *shrink,
+						    struct shrink_control *sc)
+{
+	return list_lru_shrink_walk(&huge_low_util_page_lru, sc, low_util_free_page, NULL);
+}
+
+static struct shrinker huge_low_util_page_shrinker = {
+	.count_objects = shrink_huge_low_util_page_count,
+	.scan_objects = shrink_huge_low_util_page_scan,
+	.seeks = DEFAULT_SEEKS,
+	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
+		SHRINKER_NONSLAB,
+};
+
 #ifdef CONFIG_SYSFS
 static ssize_t enabled_show(struct kobject *kobj,
 			    struct kobj_attribute *attr, char *buf)

@@ -521,13 +564,18 @@ static int __init hugepage_init(void)
 		goto err_slab;
 
 	schedule_delayed_work(&thp_utilization_work, HZ);
+	err = register_shrinker(&huge_low_util_page_shrinker, "thp-low-util");
+	if (err)
+		goto err_low_util_shrinker;
 	err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
 	if (err)
 		goto err_hzp_shrinker;
 	err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
 	if (err)
 		goto err_split_shrinker;
-
+	err = list_lru_init_memcg(&huge_low_util_page_lru, &huge_low_util_page_shrinker);
+	if (err)
+		goto err_low_util_list_lru;
 	/*
 	 * By default disable transparent hugepages on smaller systems,
 	 * where the extra memory used could hurt more than TLB overhead

@@ -543,11 +591,16 @@ static int __init hugepage_init(void)
 		goto err_khugepaged;
 
 	return 0;
+
 err_khugepaged:
+	list_lru_destroy(&huge_low_util_page_lru);
+err_low_util_list_lru:
 	unregister_shrinker(&deferred_split_shrinker);
 err_split_shrinker:
 	unregister_shrinker(&huge_zero_page_shrinker);
 err_hzp_shrinker:
+	unregister_shrinker(&huge_low_util_page_shrinker);
+err_low_util_shrinker:
 	khugepaged_destroy();
 err_slab:
 	hugepage_exit_sysfs(hugepage_kobj);

@@ -622,6 +675,7 @@ void prep_transhuge_page(struct page *page)
 	 */
 
 	INIT_LIST_HEAD(page_deferred_list(page));
+	INIT_LIST_HEAD(page_underutilized_thp_list(page));
 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 }
 

@@ -2491,8 +2545,7 @@ static void __split_huge_page_tail(struct page *head, int tail,
 			 (1L << PG_dirty)));
 
 	/* ->mapping in first tail page is compound_mapcount */
-	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
-			page_tail);
+	VM_BUG_ON_PAGE(tail > 3 && page_tail->mapping != TAIL_MAPPING, page_tail);
 	page_tail->mapping = head->mapping;
 	page_tail->index = head->index + tail;
 	page_tail->private = 0;

@@ -2698,6 +2751,7 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 	struct folio *folio = page_folio(page);
 	struct page *head = &folio->page;
 	struct deferred_split *ds_queue = get_deferred_split_queue(head);
+	struct list_head *underutilized_thp_list = page_underutilized_thp_list(head);
 	XA_STATE(xas, &head->mapping->i_pages, head->index);
 	struct anon_vma *anon_vma = NULL;
 	struct address_space *mapping = NULL;

@@ -2796,6 +2850,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 			list_del(page_deferred_list(head));
 		}
 		spin_unlock(&ds_queue->split_queue_lock);
+		if (!list_empty(underutilized_thp_list))
+			list_lru_del_page(&huge_low_util_page_lru, head, underutilized_thp_list);
 		if (mapping) {
 			int nr = thp_nr_pages(head);
 

@@ -2838,6 +2894,7 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
 void free_transhuge_page(struct page *page)
 {
 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
+	struct list_head *underutilized_thp_list = page_underutilized_thp_list(page);
 	unsigned long flags;
 
 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);

@@ -2846,6 +2903,12 @@ void free_transhuge_page(struct page *page)
 		list_del(page_deferred_list(page));
 	}
 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
+	if (!list_empty(underutilized_thp_list))
+		list_lru_del_page(&huge_low_util_page_lru, page, underutilized_thp_list);
+
+	if (PageLRU(page))
+		__clear_page_lru_flags(page);
+
 	free_compound_page(page);
 }
 

@@ -2886,6 +2949,26 @@ void deferred_split_huge_page(struct page *page)
 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
 }
 
+void add_underutilized_thp(struct page *page)
+{
+	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
+
+	if (PageSwapCache(page))
+		return;
+
+	/*
+	 * Need to take a reference on the page to prevent the page from getting free'd from
+	 * under us while we are adding the THP to the shrinker.
+	 */
+	if (!get_page_unless_zero(page))
+		return;
+
+	if (!is_huge_zero_page(page) && is_anon_transparent_hugepage(page))
+		list_lru_add_page(&huge_low_util_page_lru, page, page_underutilized_thp_list(page));
+
+	put_page(page);
+}
+
 static unsigned long deferred_split_count(struct shrinker *shrink,
 		struct shrink_control *sc)
 {

@@ -3424,6 +3507,9 @@ static void thp_util_scan(unsigned long pfn_end)
 		/* Group THPs into utilization buckets */
 		bucket = num_utilized_pages * THP_UTIL_BUCKET_NR / HPAGE_PMD_NR;
 		bucket = min(bucket, THP_UTIL_BUCKET_NR - 1);
+		if (bucket == 0)
+			add_underutilized_thp(page);
+
 		thp_scan.buckets[bucket].nr_thps++;
 		thp_scan.buckets[bucket].nr_zero_pages += (HPAGE_PMD_NR - num_utilized_pages);
 	}

mm/list_lru.c +49 -0

index a05e5bef3b40..7e8b324cc840 100644
@@ -140,6 +140,32 @@ bool list_lru_add(struct list_lru *lru, struct list_head *item)
 }
 EXPORT_SYMBOL_GPL(list_lru_add);
 
+bool list_lru_add_page(struct list_lru *lru, struct page *page, struct list_head *item)
+{
+	int nid = page_to_nid(page);
+	struct list_lru_node *nlru = &lru->node[nid];
+	struct list_lru_one *l;
+	struct mem_cgroup *memcg;
+
+	spin_lock(&nlru->lock);
+	if (list_empty(item)) {
+		memcg = page_memcg(page);
+		memcg_list_lru_alloc(memcg, lru, GFP_KERNEL);
+		l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
+		list_add_tail(item, &l->list);
+		/* Set shrinker bit if the first element was added */
+		if (!l->nr_items++)
+			set_shrinker_bit(memcg, nid,
+					 lru_shrinker_id(lru));
+		nlru->nr_items++;
+		spin_unlock(&nlru->lock);
+		return true;
+	}
+	spin_unlock(&nlru->lock);
+	return false;
+}
+EXPORT_SYMBOL_GPL(list_lru_add_page);
+
 bool list_lru_del(struct list_lru *lru, struct list_head *item)
 {
 	int nid = page_to_nid(virt_to_page(item));

@@ -160,6 +186,29 @@ bool list_lru_del(struct list_lru *lru, struct list_head *item)
 }
 EXPORT_SYMBOL_GPL(list_lru_del);
 
+bool list_lru_del_page(struct list_lru *lru, struct page *page, struct list_head *item)
+{
+	int nid = page_to_nid(page);
+	struct list_lru_node *nlru = &lru->node[nid];
+	struct list_lru_one *l;
+	struct mem_cgroup *memcg;
+
+	spin_lock(&nlru->lock);
+	if (!list_empty(item)) {
+		memcg = page_memcg(page);
+		memcg_list_lru_alloc(memcg, lru, GFP_KERNEL);
+		l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
+		list_del_init(item);
+		l->nr_items--;
+		nlru->nr_items--;
+		spin_unlock(&nlru->lock);
+		return true;
+	}
+	spin_unlock(&nlru->lock);
+	return false;
+}
+EXPORT_SYMBOL_GPL(list_lru_del_page);
+
 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
 {
 	list_del_init(item);

mm/page_alloc.c +6 -0

index e5486d47406e..a2a33b4d71db 100644
@@ -1327,6 +1327,12 @@ static int free_tail_pages_check(struct page *head_page, struct page *page)
 		 * deferred_list.next -- ignore value.
 		 */
 		break;
+	case 3:
+		/*
+		 * the third tail page: ->mapping is
+		 * underutilized_thp_list.next -- ignore value.
+		 */
+		break;
 	default:
 		if (page->mapping != TAIL_MAPPING) {
 			bad_page(page, "corrupted mapping in tail page");

Signature · 1 line

2.30.2

Quoted 16 lines · David Hildenbrand

Quoted 14 lines · On 25.08.22 23:30, [email protected] wrote:

From: Alexander Zhu <[email protected]>

Currently, split_huge_page() does not have a way to identify zero filled
pages within the THP. Thus these zero pages get remapped and continue to
create memory waste. In this patch, we identify and free tail pages that
are zero filled in split_huge_page(). In this way, we avoid mapping these
pages back into page table entries and can free up unused memory within
THPs.

Quoted 5 lines

Isn't this to some degree splitting the THP (PMDs->PTEs + dissolve compound page) and then letting KSM replace the zero-filled page by the
shared zeropage?

Many systems do not run KSM, though, and even on the systems where it does, KSM only covers a subset of the memory in the system.

I could see wanting to maybe consolidate the scanning between KSM and this thing at some point, if it could be done without too much complexity, but keeping this change to split_huge_page looks like it might make sense even when KSM is enabled, since it will get rid of the unnecessary memory much faster than KSM could.

Keeping a hundred MB of unnecessary memory around for longer would simply result in more THPs getting split up, and more memory pressure for a longer time than we need.

Quoted 11 lines · Rik van Riel

Quoted 2 lines · David Hildenbrand

Quoted 1 line · On 25.08.22 23:30, [email protected] wrote:

From: Alexander Zhu <[email protected]>

I could see wanting to maybe consolidate the scanning between KSM and this thing at some point, if it could be done without too much complexity, but keeping this change to split_huge_page looks like it might make sense even when KSM is enabled, since it will get rid of the unnecessary memory much faster than KSM could.

Quoted 3 lines

Keeping a hundred MB of unnecessary memory around for longer would simply result in more THPs getting split up, and more memory pressure for a longer time than we need.

Right. I was wondering if we want to map the shared zeropage instead of
the "detected to be zero" page, similar to how KSM would do it. For example, with userfaultfd there would be an observable difference.

(maybe that's already done in this patch set)

On 25 Aug 2022, at 17:30, [email protected] wrote:

How large is the memory? Just wonder the scanning speed. Also, it might be better to explicitly add the time unit, second, in the output.

Is it possible to use cache-bypassing read to avoid cache pollution? You are scanning for 256*2M at a time. Wouldn’t that wipe out all the useful data in the cache?

How stale could the information in the utilization bucket be?

Is it possible that THP shrinker splits a THP used to have a lot of zero-filled subpages but now have all subpages filled with useful values?

In Patch 2, split_huge_page() only unmap zero-filled subpages, but for THP shrinker, should it verify the utilization before it splits the page?

Quoted 16 lines · David Hildenbrand

Quoted 14 lines · Rik van Riel

Quoted 2 lines · David Hildenbrand

Quoted 1 line · On 25.08.22 23:30, [email protected] wrote:

From: Alexander Zhu <[email protected]>

I could see wanting to maybe consolidate the scanning between KSM and this thing at some point, if it could be done without too much complexity, but keeping this change to split_huge_page looks like it might make sense even when KSM is enabled, since it will get rid of the unnecessary memory much faster than KSM could.

Keeping a hundred MB of unnecessary memory around for longer would simply result in more THPs getting split up, and more memory pressure for a longer time than we need.

Quoted 7 lines

Right. I was wondering if we want to map the shared zeropage instead of
the "detected to be zero" page, similar to how KSM would do it. For example, with userfaultfd there would be an observable difference.

(maybe that's already done in this patch set)

The patch does not currently do that, but I suppose it could?

What exactly are the userfaultfd differences here, and how does dropping 4kB pages break things vs. using the shared zeropage?

If you unmap something (resulting in pte_none()) where previously something used to be mapped in a page table, you might suddenly inform the user space fault handler about a page fault that it doesn't expect, because it previously placed a page and did not zap that page itself (MADV_DONTNEED).

So at least with userfaultfd I think we have to be careful. Not sure if there are other corner cases (again, KSM behavior is interesting)

Signature · 3 lines

Thanks,

David / dhildenb