Re: [PATCH 0/2] fix devm_memremap_pages() mem hotplug locking

From: Masayoshi Mizuma
Date: Sun Feb 12 2017 - 20:39:14 EST

Hi Dan,

On Sun, 12 Feb 2017 14:34:11 -0800 Dan Williams wrote:
Ben notes that commit f931ab479dd2 "mm: fix devm_memremap_pages crash,
use mem_hotplug_{begin, done}" is incomplete and broken. Writes to
mem_hotplug.active_writer need to be coordinated under the device
hotplug lock. Otherwise, we can potentially corrupt mem_hotplug.refcount
leading to soft lockups.

I think mem_hotplug_{begin,done} is not suitable to exclude devm_memremap_pages()
because it seems that memory hotplug is not related to this context.
How about using pgmap_lock instead?

Like this:

diff --git a/kernel/memremap.c b/kernel/memremap.c
index 9ecedc2..e9b9cfa 100644
--- a/kernel/memremap.c
+++ b/kernel/memremap.c
@@ -246,9 +246,9 @@ static void devm_memremap_pages_release(struct device *dev, void *data)
/* pages are dead and unused, undo the arch mapping */
align_start = res->start & ~(SECTION_SIZE - 1);
align_size = ALIGN(resource_size(res), SECTION_SIZE);
- mem_hotplug_begin();
+ mutex_lock(&pgmap_lock);
arch_remove_memory(align_start, align_size);
- mem_hotplug_done();
+ mutex_unlock(&pgmap_lock);
untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
@@ -360,9 +360,9 @@ void *devm_memremap_pages(struct device *dev, struct resource *res,
if (error)
goto err_pfn_remap;
- mem_hotplug_begin();
+ mutex_lock(&pgmap_lock);
error = arch_add_memory(nid, align_start, align_size, true);
- mem_hotplug_done();
+ mutex_unlock(&pgmap_lock);
if (error)
goto err_add_memory;