pci: rewrite pci_call_probe() to use workqueue instead of work_on_cpu()

pci_call_probe() uses work_on_cpu(), which creates and tears down a full kthread on each call, to invoke ->probe() on node local CPU for allocation affinity.

The same goal can easily be achieved using a work item. This patch rewrites pci_call_probe() so that it uses a work item instead of work_on_cpu().

Note that the function is restructured for simplicity. This adds get/put_online_cpus() pair for devices without node but the overhead of doing so isn't anything material at this level.

Signed-off-by Tejun Heo <[email protected]>

---
I tested both paths but had to simulate device on foreign node.

Thanks.

 drivers/pci/pci-driver.c |   48 +++++++++++++++++++++++++++--------------------
 1 file changed, 28 insertions(+), 20 deletions(-)

drivers/pci/pci-driver.c +28 -20

index 5270f1a..554c15a 100644
@@ -274,11 +274,14 @@ struct drv_dev_and_id {
        struct pci_driver *drv;
        struct pci_dev *dev;
        const struct pci_device_id *id;
+       struct work_struct work;
+       int error;
 };
 
-static long local_pci_probe(void *_ddi)
+static void pci_probe_fn(struct work_struct *work)
 {
-       struct drv_dev_and_id *ddi = _ddi;
+       struct drv_dev_and_id *ddi = container_of(work, struct drv_dev_and_id,
+                                                 work);
        struct device *dev = &ddi->dev->dev;
        int rc;
 

@@ -298,33 +301,38 @@ static long local_pci_probe(void *_ddi)
                pm_runtime_set_suspended(dev);
                pm_runtime_put_noidle(dev);
        }
-       return rc;
+
+       ddi->error = rc;
 }
 
 static int pci_call_probe(struct pci_driver *drv, struct pci_dev *dev,
                          const struct pci_device_id *id)
 {
-       int error, node;
+       int node;
+       int cpu = nr_cpu_ids;
        struct drv_dev_and_id ddi = { drv, dev, id };
 
-       /* Execute driver initialization on node where the device's
-          bus is attached to.  This way the driver likely allocates
-          its local memory on the right node without any need to
-          change it. */
-       node = dev_to_node(&dev->dev);
-       if (node >= 0) {
-               int cpu;
+       /*
+        * Execute driver initialization on node where the device's bus is
+        * attached to.  This way the driver likely allocates its local
+        * memory on the right node without any need to change it.
+        */
+       get_online_cpus();
 
-               get_online_cpus();
+       node = dev_to_node(&dev->dev);
+       if (node >= 0)
                cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
-               if (cpu < nr_cpu_ids)
-                       error = work_on_cpu(cpu, local_pci_probe, &ddi);
-               else
-                       error = local_pci_probe(&ddi);
-               put_online_cpus();
-       } else
-               error = local_pci_probe(&ddi);
-       return error;
+
+       if (cpu < nr_cpu_ids) {
+               INIT_WORK_ONSTACK(&ddi.work, pci_probe_fn);
+               schedule_work_on(cpu, &ddi.work);
+               flush_work(&ddi.work);
+       } else {
+               pci_probe_fn(&ddi.work);
+       }
+
+       put_online_cpus();
+       return ddi.error;
 }
 
 /**

Quoted 8 lines · Tejun Heo

pci_call_probe() uses work_on_cpu(), which creates and tears down a full kthread on each call, to invoke ->probe() on node local CPU for allocation affinity.

The same goal can easily be achieved using a work item. This patch rewrites pci_call_probe() so that it uses a work item instead of work_on_cpu().

Quoted 5 lines

Note that the function is restructured for simplicity. This adds get/put_online_cpus() pair for devices without node but the overhead of doing so isn't anything material at this level.

Signed-off-by Tejun Heo <[email protected]>

I applied this to my "next" branch and pushed it. Thanks!

Quoted 13 lines · Bjorn Helgaas

Quoted 12 lines · Tejun Heo

pci_call_probe() uses work_on_cpu(), which creates and tears down a full kthread on each call, to invoke ->probe() on node local CPU for allocation affinity.

The same goal can easily be achieved using a work item. This patch rewrites pci_call_probe() so that it uses a work item instead of work_on_cpu().

Note that the function is restructured for simplicity. This adds get/put_online_cpus() pair for devices without node but the overhead of doing so isn't anything material at this level.

Quoted 3 lines

Quoted 1 line

Signed-off-by Tejun Heo <[email protected]>

I applied this to my "next" branch and pushed it. Thanks!

Can you please drop / revert this one? It seems we need work_on_cpu() after all and I'm gonna reimplement it in sane way.