driver core: Attach devices on CPU local to device node
authorAlexander Duyck <>
Tue, 22 Jan 2019 18:39:37 +0000 (10:39 -0800)
committerGreg Kroah-Hartman <>
Thu, 31 Jan 2019 13:20:54 +0000 (14:20 +0100)
Call the asynchronous probe routines on a CPU local to the device node. By
doing this we should be able to improve our initialization time
significantly as we can avoid having to access the device from a remote
node which may introduce higher latency.

For example, in the case of initializing memory for NVDIMM this can have a
significant impact as initialing 3TB on remote node can take up to 39
seconds while initialing it on a local node only takes 23 seconds. It is
situations like this where we will see the biggest improvement.

Reviewed-by: Dan Williams <>
Reviewed-by: Bart Van Assche <>
Signed-off-by: Alexander Duyck <>
Reviewed-by: Rafael J. Wysocki <>
Signed-off-by: Greg Kroah-Hartman <>

index 627ad05064e07667590620630eefd03b3b779509..aa6a9c613595f67c01cae3efccc9a42d68906a90 100644 (file)
@@ -829,7 +829,7 @@ static int __device_attach(struct device *dev, bool allow_async)
                        dev_dbg(dev, "scheduling asynchronous probe\n");
-                       async_schedule(__device_attach_async_helper, dev);
+                       async_schedule_dev(__device_attach_async_helper, dev);
                } else {
@@ -989,7 +989,7 @@ static int __driver_attach(struct device *dev, void *data)
                if (!dev->driver) {
                        dev->p->async_driver = drv;
-                       async_schedule(__driver_attach_async_helper, dev);
+                       async_schedule_dev(__driver_attach_async_helper, dev);
                return 0;