[PATCH 3/3] mm: make pfn walker support ZONE_DEVICE

[Toshiki Fukasawa <t-fukasawa@xxxxxxxxxxxxx>]

This patch allows pfn walker to read pages on ZONE_DEVICE.
There are the following notes:

	a) The reserved pages indicated by vmem_altmap->reserve
	   are uninitialized, so it must be skipped to read.
	b) To get vmem_altmap, we need to use get_dev_pagemap(),
	   but doing it for all pfns is too slow.

This patch solves both of them. Since vmem_altmap could reserve
only first few pages, we can reduce the number of checks by
counting sequential valid pages.

Signed-off-by: Toshiki Fukasawa <t-fukasawa@xxxxxxxxxxxxx>
---
 fs/proc/page.c           | 22 ++++++++++++++++++----
 include/linux/memremap.h |  6 ++++++
 mm/memremap.c            | 29 +++++++++++++++++++++++++++++
 3 files changed, 53 insertions(+), 4 deletions(-)

diff --git a/fs/proc/page.c b/fs/proc/page.c
index a49b638..b6241ea 100644
--- a/fs/proc/page.c
+++ b/fs/proc/page.c
@@ -33,6 +33,7 @@ static ssize_t kpage_common_read(struct file *file, char __user *buf,
 	struct page *ppage;
 	unsigned long src = *ppos;
 	unsigned long pfn;
+	unsigned long valid_pages = 0;
 	ssize_t ret = 0;
 
 	pfn = src / KPMSIZE;
@@ -41,11 +42,24 @@ static ssize_t kpage_common_read(struct file *file, char __user *buf,
 		return -EINVAL;
 
 	while (count > 0) {
-		/*
-		 * TODO: ZONE_DEVICE support requires to identify
-		 * memmaps that were actually initialized.
-		 */
 		ppage = pfn_to_online_page(pfn);
+		if (!ppage && pfn_zone_device(pfn)) {
+			/*
+			 * Skip to read first few uninitialized pages on
+			 * ZONE_DEVICE. And count valid pages starting
+			 * with the pfn so that minimize the number of
+			 * calls to nr_valid_pages_zone_device().
+			 */
+			if (!valid_pages)
+				valid_pages = nr_valid_pages_zone_device(pfn);
+			if (valid_pages) {
+				ppage = pfn_to_page(pfn);
+				valid_pages--;
+			}
+		} else if (valid_pages) {
+			/* ZONE_DEVICE has been hot removed */
+			valid_pages = 0;
+		}
 
 		if (put_user(read_fn(ppage), out)) {
 			ret = -EFAULT;
diff --git a/include/linux/memremap.h b/include/linux/memremap.h
index 6fefb09..d111ae3 100644
--- a/include/linux/memremap.h
+++ b/include/linux/memremap.h
@@ -123,6 +123,7 @@ static inline struct vmem_altmap *pgmap_altmap(struct dev_pagemap *pgmap)
 }
 
 #ifdef CONFIG_ZONE_DEVICE
+unsigned long nr_valid_pages_zone_device(unsigned long pfn);
 void *memremap_pages(struct dev_pagemap *pgmap, int nid);
 void memunmap_pages(struct dev_pagemap *pgmap);
 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap);
@@ -133,6 +134,11 @@ struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap);
 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns);
 #else
+static inline unsigned long nr_valid_pages_zone_device(unsigned long pfn)
+{
+	return 0;
+}
+
 static inline void *devm_memremap_pages(struct device *dev,
 		struct dev_pagemap *pgmap)
 {
diff --git a/mm/memremap.c b/mm/memremap.c
index 8a97fd4..307c73e 100644
--- a/mm/memremap.c
+++ b/mm/memremap.c
@@ -73,6 +73,35 @@ static unsigned long pfn_next(unsigned long pfn)
 	return pfn + 1;
 }
 
+/*
+ * This returns the number of sequential valid pages starting from @pfn
+ * on ZONE_DEVICE. The invalid pages reserved by driver is first few
+ * pages on ZONE_DEVICE.
+ */
+unsigned long nr_valid_pages_zone_device(unsigned long pfn)
+{
+	struct dev_pagemap *pgmap;
+	struct vmem_altmap *altmap;
+	unsigned long pages;
+
+	pgmap = get_dev_pagemap(pfn, NULL);
+	if (!pgmap)
+		return 0;
+	altmap = pgmap_altmap(pgmap);
+	if (altmap && pfn < (altmap->base_pfn + altmap->reserve))
+		pages = 0;
+	else
+		/*
+		 * PHYS_PFN(pgmap->res.end) is end pfn on pgmap
+		 * (not start pfn on next mapping).
+		 */
+		pages = PHYS_PFN(pgmap->res.end) - pfn + 1;
+
+	put_dev_pagemap(pgmap);
+
+	return pages;
+}
+
 #define for_each_device_pfn(pfn, map) \
 	for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
 
-- 
1.8.3.1