| Message ID | 20181008233404.1909.37302.stgit@localhost.localdomain (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [mm] memremap: Fix reference count for pgmap in devm_memremap_pages | expand |
On Mon, Oct 8, 2018 at 4:34 PM Alexander Duyck <alexander.h.duyck@linux.intel.com> wrote: > > In the earlier patch "mm: defer ZONE_DEVICE page initialization to the > point where we init pgmap" I had overlooked the reference count that was > being held per page on the pgmap. As a result on running the ndctl test > "create.sh" we would call into devm_memremap_pages_release and encounter > the following percpu reference count error and hang: > WARNING: CPU: 30 PID: 0 at lib/percpu-refcount.c:155 > percpu_ref_switch_to_atomic_rcu+0xf3/0x120 > > This patch addresses that by performing an update for all of the device > PFNs in a single call. In my testing this seems to resolve the issue while > still allowing us to retain the improvements seen in memory initialization. > > Reported-by: Dan Williams <dan.j.williams@intel.com> Tested-by: Dan Williams <dan.j.williams@intel.com> Thanks Alex!
diff --git a/kernel/memremap.c b/kernel/memremap.c index 6ec81e0d7a33..9eced2cc9f94 100644 --- a/kernel/memremap.c +++ b/kernel/memremap.c @@ -218,6 +218,7 @@ void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap) memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE], align_start >> PAGE_SHIFT, align_size >> PAGE_SHIFT, pgmap); + percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap)); devm_add_action(dev, devm_memremap_pages_release, pgmap);
In the earlier patch "mm: defer ZONE_DEVICE page initialization to the point where we init pgmap" I had overlooked the reference count that was being held per page on the pgmap. As a result on running the ndctl test "create.sh" we would call into devm_memremap_pages_release and encounter the following percpu reference count error and hang: WARNING: CPU: 30 PID: 0 at lib/percpu-refcount.c:155 percpu_ref_switch_to_atomic_rcu+0xf3/0x120 This patch addresses that by performing an update for all of the device PFNs in a single call. In my testing this seems to resolve the issue while still allowing us to retain the improvements seen in memory initialization. Reported-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@linux.intel.com> --- kernel/memremap.c | 1 + 1 file changed, 1 insertion(+)