On 7/5/22 04:40, Hillf Danton wrote:
If no compat devices can be added to ib_device with DEVICE_REGISTERED
cleared then they can be removed without ib_device's refcount dropping
to zero.
Even if that is not strictly true, a new flag that marks ib device
disabled and prevents new compact devices from being added can be added
in bid to cut the wait for completion.
Hillf
+++ b/drivers/infiniband/core/device.c
@@ -1265,6 +1265,7 @@ static void disable_device(struct ib_dev
down_write(&devices_rwsem);
xa_clear_mark(&devices, device->index, DEVICE_REGISTERED);
+ // device->disabled = true;
up_write(&devices_rwsem);
/*
@@ -1282,17 +1283,10 @@ static void disable_device(struct ib_dev
}
ib_cq_pool_cleanup(device);
+ remove_compat_devs(device);
/* Pairs with refcount_set in enable_device */
ib_device_put(device);
- wait_for_completion(&device->unreg_completion);
-
- /*
- * compat devices must be removed after device refcount drops to zero.
- * Otherwise init_net() may add more compatdevs after removing compat
- * devices and before device is disabled.
- */
- remove_compat_devs(device);
}
I'm not convinced the above patch is a step in the right direction nor
that it is correct. Anyway, since the RDMA maintainers know this code
better than I do I will let them comment on the above patch.
Bart.
