Re: [PATCH v2 9/9] devm_memremap_pages: protect against pmem device unbind

[Williams, Dan J]

On Wed, 2015-08-26 at 14:46 +0200, Christoph Hellwig wrote:
> On Tue, Aug 25, 2015 at 09:28:13PM -0400, Dan Williams wrote:
> > Given that:
> > 
> > 1/ device ->remove() can not be failed
> > 
> > 2/ a pmem device may be unbound at any time
> > 
> > 3/ we do not know what other parts of the kernel are actively using a
> >    'struct page' from devm_memremap_pages()
> > 
> > ...provide a facility for active usages of device memory to block pmem
> > device unbind.  With a percpu_ref it should be feasible to take a
> > reference on a per-I/O or other high frequency basis.
> 
> Without a caller of get_page_map this is just adding dead code.  I'd
> suggest to group it in a series with that caller.
> 

Agreed, we can drop this until the first user arrives.

> Also if the page_map gets exposed in a header the name is a bit too generic.
> memremap_map maybe?

Done, and in the patch below I hide the internal implementation details
of page_map in kernel/memremap.c and only expose the percpu_ref in the
public memremap_map.

8<----
Subject: devm_memremap_pages: protect against pmem device unbind

From: Dan Williams <dan.j.williams@xxxxxxxxx>

Given that:

1/ device ->remove() can not be failed

2/ a pmem device may be unbound at any time

3/ we do not know what other parts of the kernel are actively using a
   'struct page' from devm_memremap_pages()

...provide a facility for active usages of device memory to block pmem
device unbind.  With a percpu_ref it should be feasible to take a
reference on a per-I/O or other high frequency basis.

Cc: Christoph Hellwig <hch@xxxxxx>
Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
---
 include/linux/io.h |   30 ++++++++++++++++
 kernel/memremap.c  |   96 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 125 insertions(+), 1 deletion(-)

diff --git a/include/linux/io.h b/include/linux/io.h
index de64c1e53612..9e696b114c6d 100644
--- a/include/linux/io.h
+++ b/include/linux/io.h
@@ -90,8 +90,25 @@ void devm_memunmap(struct device *dev, void *addr);
 void *__devm_memremap_pages(struct device *dev, struct resource *res);
 
 #ifdef CONFIG_ZONE_DEVICE
+#include <linux/percpu-refcount.h>
+
+struct memremap_map {
+	struct percpu_ref percpu_ref;
+};
+
 void *devm_memremap_pages(struct device *dev, struct resource *res);
+struct memremap_map * __must_check get_memremap_map(resource_size_t addr);
+static inline void ref_memremap_map(struct memremap_map *memremap_map)
+{
+	percpu_ref_get(&memremap_map->percpu_ref);
+}
+
+static inline void put_memremap_map(struct memremap_map *memremap_map)
+{
+	percpu_ref_put(&memremap_map->percpu_ref);
+}
 #else
+struct memremap_map;
 static inline void *devm_memremap_pages(struct device *dev, struct resource *res)
 {
 	/*
@@ -102,6 +119,19 @@ static inline void *devm_memremap_pages(struct device *dev, struct resource *res
 	WARN_ON_ONCE(1);
 	return ERR_PTR(-ENXIO);
 }
+
+static inline __must_check struct memremap_map *get_memremap_map(resource_size_t addr)
+{
+	return NULL;
+}
+
+static inline void ref_memremap_map(struct memremap_map *memremap_map)
+{
+}
+
+static inline void put_memremap_map(struct memremap_map *memremap_map)
+{
+}
 #endif
 
 /*
diff --git a/kernel/memremap.c b/kernel/memremap.c
index 72b0c66628b6..5b9f04789b96 100644
--- a/kernel/memremap.c
+++ b/kernel/memremap.c
@@ -11,7 +11,11 @@
  * General Public License for more details.
  */
 #include <linux/device.h>
+#include <linux/ioport.h>
 #include <linux/types.h>
+#include <linux/sched.h>
+#include <linux/wait.h>
+#include <linux/list.h>
 #include <linux/io.h>
 #include <linux/mm.h>
 #include <linux/memory_hotplug.h>
@@ -138,14 +142,74 @@ void devm_memunmap(struct device *dev, void *addr)
 EXPORT_SYMBOL(devm_memunmap);
 
 #ifdef CONFIG_ZONE_DEVICE
+static DEFINE_MUTEX(page_map_lock);
+static DECLARE_WAIT_QUEUE_HEAD(page_map_wait);
+static LIST_HEAD(page_maps);
+
+enum {
+	PAGE_MAP_LIVE,
+	PAGE_MAP_CONFIRM,
+};
+
 struct page_map {
 	struct resource res;
+	struct list_head list;
+	unsigned long flags;
+	struct memremap_map map;
+	struct device *dev;
 };
 
+static struct page_map *to_page_map(struct percpu_ref *ref)
+{
+	return container_of(ref, struct page_map, map.percpu_ref);
+}
+
+static void page_map_release(struct percpu_ref *ref)
+{
+	struct page_map *page_map = to_page_map(ref);
+
+	/* signal page_map is idle (no more refs) */
+	clear_bit(PAGE_MAP_LIVE, &page_map->flags);
+	wake_up_all(&page_map_wait);
+}
+
+static void page_map_confirm(struct percpu_ref *ref)
+{
+	struct page_map *page_map = to_page_map(ref);
+
+	/* signal page_map is confirmed dead (slow path ref mode) */
+	set_bit(PAGE_MAP_CONFIRM, &page_map->flags);
+	wake_up_all(&page_map_wait);
+}
+
+static void page_map_destroy(struct page_map *page_map)
+{
+	long tmo;
+
+	/* flush new lookups */
+	mutex_lock(&page_map_lock);
+	list_del_rcu(&page_map->list);
+	mutex_unlock(&page_map_lock);
+	synchronize_rcu();
+
+	percpu_ref_kill_and_confirm(&page_map->map.percpu_ref, page_map_confirm);
+	do {
+		tmo = wait_event_interruptible_timeout(page_map_wait,
+			!test_bit(PAGE_MAP_LIVE, &page_map->flags)
+			&& test_bit(PAGE_MAP_CONFIRM, &page_map->flags), 5*HZ);
+		if (tmo <= 0)
+			dev_dbg(page_map->dev,
+					"page map active, continuing to wait...\n");
+	} while (tmo <= 0);
+}
+
 static void devm_memremap_pages_release(struct device *dev, void *res)
 {
 	struct page_map *page_map = res;
 
+	if (test_bit(PAGE_MAP_LIVE, &page_map->flags))
+		page_map_destroy(page_map);
+
 	/* pages are dead and unused, undo the arch mapping */
 	arch_remove_memory(page_map->res.start, resource_size(&page_map->res));
 }
@@ -155,7 +219,7 @@ void *devm_memremap_pages(struct device *dev, struct resource *res)
 	int is_ram = region_intersects(res->start, resource_size(res),
 			"System RAM");
 	struct page_map *page_map;
-	int error, nid;
+	int error, nid, rc;
 
 	if (is_ram == REGION_MIXED) {
 		WARN_ONCE(1, "%s attempted on mixed region %pr\n",
@@ -172,6 +236,12 @@ void *devm_memremap_pages(struct device *dev, struct resource *res)
 		return ERR_PTR(-ENOMEM);
 
 	memcpy(&page_map->res, res, sizeof(*res));
+	INIT_LIST_HEAD(&page_map->list);
+	page_map->dev = dev;
+	rc = percpu_ref_init(&page_map->map.percpu_ref, page_map_release, 0,
+				GFP_KERNEL);
+	if (rc)
+		return ERR_PTR(rc);
 
 	nid = dev_to_node(dev);
 	if (nid < 0)
@@ -183,8 +253,32 @@ void *devm_memremap_pages(struct device *dev, struct resource *res)
 		return ERR_PTR(error);
 	}
 
+	set_bit(PAGE_MAP_LIVE, &page_map->flags);
+	mutex_lock(&page_map_lock);
+	list_add_rcu(&page_map->list, &page_maps);
+	mutex_unlock(&page_map_lock);
+
 	devres_add(dev, page_map);
 	return __va(res->start);
 }
 EXPORT_SYMBOL(devm_memremap_pages);
+
+struct memremap_map * __must_check get_memremap_map(resource_size_t addr)
+{
+	struct memremap_map *ret = NULL;
+	struct page_map *page_map;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(page_map, &page_maps, list) {
+		if (addr >= page_map->res.start && addr <= page_map->res.end) {
+			if (percpu_ref_tryget(&page_map->map.percpu_ref))
+				ret = &page_map->map;
+			break;
+		}
+	}
+	rcu_read_unlock();
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(get_memremap_map);
 #endif /* CONFIG_ZONE_DEVICE */

N‹§²æ¸›yú²X¬¶ÇvØ–)Þ{.nlj·¥Š{±‘êX§¶›¡Ü}©ž²ÆzÚj:+v‰¨¾«‘êZ+€Êzf£¢·hšˆ§~†­†Ûÿû®w¥¢¸?™¨è&¢)ßf”ùy§m…á«a¶Úÿ0¶ìå