Message-ID: <20251026203611.1608903-8-surenb@google.com>
Date: Sun, 26 Oct 2025 13:36:10 -0700
From: Suren Baghdasaryan <surenb@...gle.com>
To: akpm@...ux-foundation.org
Cc: david@...hat.com, lorenzo.stoakes@...cle.com, Liam.Howlett@...cle.com, 
	vbabka@...e.cz, alexandru.elisei@....com, peterx@...hat.com, sj@...nel.org, 
	rppt@...nel.org, mhocko@...e.com, corbet@....net, axboe@...nel.dk, 
	viro@...iv.linux.org.uk, brauner@...nel.org, hch@...radead.org, jack@...e.cz, 
	willy@...radead.org, m.szyprowski@...sung.com, robin.murphy@....com, 
	hannes@...xchg.org, zhengqi.arch@...edance.com, shakeel.butt@...ux.dev, 
	axelrasmussen@...gle.com, yuanchu@...gle.com, weixugc@...gle.com, 
	minchan@...nel.org, surenb@...gle.com, linux-mm@...ck.org, 
	linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org, 
	linux-block@...r.kernel.org, linux-fsdevel@...r.kernel.org, 
	iommu@...ts.linux.dev, Minchan Kim <minchan@...gle.com>
Subject: [PATCH v2 7/8] mm: introduce GCMA

From: Minchan Kim <minchan@...gle.com>

This patch introduces GCMA (Guaranteed Contiguous Memory Allocator)
cleacache backend which reserves some amount of memory at the boot
and then donates it to store clean file-backed pages in the cleancache.
GCMA aims to guarantee contiguous memory allocation success as well as
low and deterministic allocation latency.

Notes:
Originally, the idea was posted by SeongJae Park and Minchan Kim [1].
Later Minchan reworked it to be used in Android as a reference for
Android vendors to use [2].

[1] https://lwn.net/Articles/619865/
[2] https://android-review.googlesource.com/q/topic:%22gcma_6.12%22

Signed-off-by: Minchan Kim <minchan@...gle.com>
Signed-off-by: Suren Baghdasaryan <surenb@...gle.com>
---
 MAINTAINERS          |   2 +
 include/linux/gcma.h |  36 +++++++
 mm/Kconfig           |  15 +++
 mm/Makefile          |   1 +
 mm/gcma.c            | 244 +++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 298 insertions(+)
 create mode 100644 include/linux/gcma.h
 create mode 100644 mm/gcma.c

diff --git a/MAINTAINERS b/MAINTAINERS
index 3aabed281b71..40de200d1124 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -16384,6 +16384,7 @@ F:	Documentation/admin-guide/mm/
 F:	Documentation/mm/
 F:	include/linux/cma.h
 F:	include/linux/dmapool.h
+F:	include/linux/gcma.h
 F:	include/linux/ioremap.h
 F:	include/linux/memory-tiers.h
 F:	include/linux/page_idle.h
@@ -16395,6 +16396,7 @@ F:	mm/dmapool.c
 F:	mm/dmapool_test.c
 F:	mm/early_ioremap.c
 F:	mm/fadvise.c
+F:	mm/gcma.c
 F:	mm/ioremap.c
 F:	mm/mapping_dirty_helpers.c
 F:	mm/memory-tiers.c
diff --git a/include/linux/gcma.h b/include/linux/gcma.h
new file mode 100644
index 000000000000..20b2c85de87b
--- /dev/null
+++ b/include/linux/gcma.h
@@ -0,0 +1,36 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __GCMA_H__
+#define __GCMA_H__
+
+#include <linux/types.h>
+
+#ifdef CONFIG_GCMA
+
+int gcma_register_area(const char *name,
+		       unsigned long start_pfn, unsigned long count);
+
+/*
+ * NOTE: allocated pages are still marked reserved and when freeing them
+ * the caller should ensure they are isolated and not referenced by anyone
+ * other than the caller.
+ */
+int gcma_alloc_range(unsigned long start_pfn, unsigned long count, gfp_t gfp);
+int gcma_free_range(unsigned long start_pfn, unsigned long count);
+
+#else /* CONFIG_GCMA */
+
+static inline int gcma_register_area(const char *name,
+				     unsigned long start_pfn,
+				     unsigned long count)
+		{ return -EOPNOTSUPP; }
+static inline int gcma_alloc_range(unsigned long start_pfn,
+				   unsigned long count, gfp_t gfp)
+		{ return -EOPNOTSUPP; }
+
+static inline int gcma_free_range(unsigned long start_pfn,
+				   unsigned long count)
+		{ return -EOPNOTSUPP; }
+
+#endif /* CONFIG_GCMA */
+
+#endif /* __GCMA_H__ */
diff --git a/mm/Kconfig b/mm/Kconfig
index e1a169d5e5de..3166fde83340 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -1097,6 +1097,21 @@ config CMA_AREAS
 
 	  If unsure, leave the default value "8" in UMA and "20" in NUMA.
 
+config GCMA
+       bool "GCMA (Guaranteed Contiguous Memory Allocator)"
+       depends on CLEANCACHE
+	help
+	  This enables the Guaranteed Contiguous Memory Allocator to allow
+	  low latency guaranteed contiguous memory allocations. Memory
+	  reserved by GCMA is donated to cleancache to be used as pagecache
+	  extension. Once GCMA allocation is requested, necessary pages are
+	  taken back from the cleancache and used to satisfy the request.
+	  Cleancache guarantees low latency successful allocation as long
+	  as the total size of GCMA allocations does not exceed the size of
+	  the memory donated to the cleancache.
+
+	  If unsure, say "N".
+
 #
 # Select this config option from the architecture Kconfig, if available, to set
 # the max page order for physically contiguous allocations.
diff --git a/mm/Makefile b/mm/Makefile
index 845841a140e3..05aee66a8b07 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -149,3 +149,4 @@ obj-$(CONFIG_TMPFS_QUOTA) += shmem_quota.o
 obj-$(CONFIG_PT_RECLAIM) += pt_reclaim.o
 obj-$(CONFIG_CLEANCACHE) += cleancache.o
 obj-$(CONFIG_CLEANCACHE_SYSFS)	+= cleancache_sysfs.o
+obj-$(CONFIG_GCMA)	+= gcma.o
diff --git a/mm/gcma.c b/mm/gcma.c
new file mode 100644
index 000000000000..b86f82b8fe9d
--- /dev/null
+++ b/mm/gcma.c
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * GCMA (Guaranteed Contiguous Memory Allocator)
+ *
+ */
+
+#define pr_fmt(fmt) "gcma: " fmt
+
+#include <linux/cleancache.h>
+#include <linux/gcma.h>
+#include <linux/hashtable.h>
+#include <linux/highmem.h>
+#include <linux/idr.h>
+#include <linux/slab.h>
+#include <linux/xarray.h>
+#include "internal.h"
+
+#define MAX_GCMA_AREAS		64
+#define GCMA_AREA_NAME_MAX_LEN	32
+
+struct gcma_area {
+	int pool_id;
+	unsigned long start_pfn;
+	unsigned long end_pfn;
+	char name[GCMA_AREA_NAME_MAX_LEN];
+};
+
+static struct gcma_area areas[MAX_GCMA_AREAS];
+static atomic_t nr_gcma_area = ATOMIC_INIT(0);
+static DEFINE_SPINLOCK(gcma_area_lock);
+
+static int free_folio_range(struct gcma_area *area,
+			     unsigned long start_pfn, unsigned long end_pfn)
+{
+	unsigned long scanned = 0;
+	unsigned long pfn;
+
+	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
+		int err;
+
+		if (!(++scanned % XA_CHECK_SCHED))
+			cond_resched();
+
+		err = cleancache_backend_put_folio(area->pool_id, pfn_folio(pfn));
+		if (err) {
+			pr_warn("PFN %lu: folio is still in use\n", pfn);
+			return err;
+		}
+	}
+
+	return 0;
+}
+
+static int alloc_folio_range(struct gcma_area *area,
+			      unsigned long start_pfn, unsigned long end_pfn,
+			      gfp_t gfp)
+{
+	unsigned long scanned = 0;
+	unsigned long pfn;
+
+	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
+		int err;
+
+		if (!(++scanned % XA_CHECK_SCHED))
+			cond_resched();
+
+		err = cleancache_backend_get_folio(area->pool_id, pfn_folio(pfn));
+		if (err) {
+			free_folio_range(area, start_pfn, pfn);
+			return err;
+		}
+	}
+
+	return 0;
+}
+
+static struct gcma_area *find_area(unsigned long start_pfn, unsigned long end_pfn)
+{
+	int nr_area = atomic_read_acquire(&nr_gcma_area);
+	int i;
+
+	for (i = 0; i < nr_area; i++) {
+		struct gcma_area *area = &areas[i];
+
+		if (area->end_pfn <= start_pfn)
+			continue;
+
+		if (area->start_pfn > end_pfn)
+			continue;
+
+		/* The entire range should belong to a single area */
+		if (start_pfn < area->start_pfn || end_pfn > area->end_pfn)
+			break;
+
+		/* Found the area containing the entire range */
+		return area;
+	}
+
+	return NULL;
+}
+
+int gcma_register_area(const char *name,
+		       unsigned long start_pfn, unsigned long count)
+{
+	LIST_HEAD(folios);
+	int i, pool_id;
+	int nr_area;
+	int ret = 0;
+
+	pool_id = cleancache_backend_register_pool(name);
+	if (pool_id < 0)
+		return pool_id;
+
+	for (i = 0; i < count; i++) {
+		struct folio *folio;
+
+		folio = pfn_folio(start_pfn + i);
+		folio_clear_reserved(folio);
+		folio_set_count(folio, 0);
+		list_add(&folio->lru, &folios);
+	}
+
+	cleancache_backend_put_folios(pool_id, &folios);
+
+	spin_lock(&gcma_area_lock);
+
+	nr_area = atomic_read(&nr_gcma_area);
+	if (nr_area < MAX_GCMA_AREAS) {
+		struct gcma_area *area = &areas[nr_area];
+
+		area->pool_id = pool_id;
+		area->start_pfn = start_pfn;
+		area->end_pfn = start_pfn + count;
+		strscpy(area->name, name);
+		/* Ensure above stores complete before we increase the count */
+		atomic_set_release(&nr_gcma_area, nr_area + 1);
+	} else {
+		ret = -ENOMEM;
+	}
+
+	spin_unlock(&gcma_area_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(gcma_register_area);
+
+int gcma_alloc_range(unsigned long start_pfn, unsigned long count, gfp_t gfp)
+{
+	unsigned long end_pfn = start_pfn + count;
+	struct gcma_area *area;
+	struct folio *folio;
+	int err, order = 0;
+
+	gfp = current_gfp_context(gfp);
+	if (gfp & __GFP_COMP) {
+		if (!is_power_of_2(count))
+			return -EINVAL;
+
+		order = ilog2(count);
+		if (order >= MAX_PAGE_ORDER)
+			return -EINVAL;
+	}
+
+	area = find_area(start_pfn, end_pfn);
+	if (!area)
+		return -EINVAL;
+
+	err = alloc_folio_range(area, start_pfn, end_pfn, gfp);
+	if (err)
+		return err;
+
+	/*
+	 * GCMA returns pages with refcount 1 and expects them to have
+	 * the same refcount 1 when they are freed.
+	 */
+	if (order) {
+		folio = pfn_folio(start_pfn);
+		post_alloc_hook(&folio->page, order, gfp);
+		set_page_refcounted(&folio->page);
+		prep_compound_page(&folio->page, order);
+	} else {
+		for (unsigned long pfn = start_pfn; pfn < end_pfn; pfn++) {
+			folio = pfn_folio(pfn);
+			post_alloc_hook(&folio->page, order, gfp);
+			set_page_refcounted(&folio->page);
+		}
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(gcma_alloc_range);
+
+int gcma_free_range(unsigned long start_pfn, unsigned long count)
+{
+	unsigned long end_pfn = start_pfn + count;
+	struct gcma_area *area;
+	unsigned long pfn;
+	int err = -EINVAL;
+
+	area = find_area(start_pfn, end_pfn);
+	if (!area)
+		return -EINVAL;
+
+	/* First pass checks and drops folio refcounts */
+	for (pfn = start_pfn; pfn < end_pfn;) {
+		struct folio *folio = pfn_folio(pfn);
+		unsigned long nr_pages = folio_nr_pages(folio);
+
+		if (pfn + nr_pages > end_pfn) {
+			end_pfn = pfn;
+			goto error;
+
+		}
+		if (!folio_ref_dec_and_test(folio)) {
+			end_pfn = pfn + nr_pages;
+			goto error;
+		}
+		pfn += nr_pages;
+	}
+
+	/* Second pass prepares the folios */
+	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
+		struct folio *folio = pfn_folio(pfn);
+
+		free_pages_prepare(&folio->page, folio_order(folio));
+		pfn += folio_nr_pages(folio);
+	}
+
+	err = free_folio_range(area, start_pfn, end_pfn);
+	if (!err)
+		return 0;
+
+error:
+	/* Restore folio refcounts */
+	for (pfn = start_pfn; pfn < end_pfn;) {
+		struct folio *folio = pfn_folio(pfn);
+
+		folio_ref_inc(folio);
+		pfn += folio_nr_pages(folio);
+	}
+
+	return err;
+}
+EXPORT_SYMBOL_GPL(gcma_free_range);
-- 
2.51.1.851.g4ebd6896fd-goog

Powered by blists - more mailing lists