The sync strategy in remove_all_compat_devs() can improved
by adopting that of rdma_dev_exit_net() which releases devices_rwsem
before calling remove_one_compat_dev()/device_del().

Also fixes a comment typo in rdma_dev_exit_net().

Fixes: 2b34c5580226 ("RDMA/core: Add command to set ib_core device net namspace sharing mode")
Signed-off-by: Sharath Srinivasan <sharath.srinivasan@oracle.com>
---
 drivers/infiniband/core/device.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
index 1174ab7da629..81689924fab8 100644
--- a/drivers/infiniband/core/device.c
+++ b/drivers/infiniband/core/device.c
@@ -1078,6 +1078,13 @@ static void remove_all_compat_devs(void)
        xa_for_each (&devices, index, dev) {
                unsigned long c_index = 0;

+               get_device(&dev->dev);
+               /*
+                * Release the devices_rwsem so that potentially blocking
+                * device_del doesn't hold the devices_rwsem for too long.
+                */
+               up_read(&devices_rwsem);
+
                /* Hold nets_rwsem so that any other thread modifying this
                 * system param can sync with this thread.
                 */
@@ -1085,6 +1092,9 @@ static void remove_all_compat_devs(void)
                xa_for_each (&dev->compat_devs, c_index, cdev)
                        remove_one_compat_dev(dev, c_index);
                up_read(&rdma_nets_rwsem);
+
+               put_device(&dev->dev);
+               down_read(&devices_rwsem);
        }
        up_read(&devices_rwsem);
 }
@@ -1168,8 +1178,8 @@ static void rdma_dev_exit_net(struct net *net)
        xa_for_each (&devices, index, dev) {
                get_device(&dev->dev);
                /*
-                * Release the devices_rwsem so that pontentially blocking
-                * device_del, doesn't hold the devices_rwsem for too long.
+                * Release the devices_rwsem so that potentially blocking
+                * device_del doesn't hold the devices_rwsem for too long.
                 */
                up_read(&devices_rwsem);

--
2.39.5 (Apple Git-154)

On Mon, Jan 19, 2026 at 11:43:52AM -0800, Sharath Srinivasan wrote:
> The sync strategy in remove_all_compat_devs() can improved
> by adopting that of rdma_dev_exit_net() which releases devices_rwsem
> before calling remove_one_compat_dev()/device_del().
> 
> Also fixes a comment typo in rdma_dev_exit_net().

You cannot change this locking without writing a huge commit message
explaining in detail the reason why any change like this is safe..

Jason

On Mon, Jan 19, 2026 at 03:53:29PM -0400, Jason Gunthorpe wrote:
> On Mon, Jan 19, 2026 at 11:43:52AM -0800, Sharath Srinivasan wrote:
> > The sync strategy in remove_all_compat_devs() can improved
> > by adopting that of rdma_dev_exit_net() which releases devices_rwsem
> > before calling remove_one_compat_dev()/device_del().
> > 
> > Also fixes a comment typo in rdma_dev_exit_net().
> 
> You cannot change this locking without writing a huge commit message
> explaining in detail the reason why any change like this is safe..

We can drop this patch, it doesn't even apply.

Thanks

> 
> Jason
>

On 2026-01-25 5:47 a.m., Leon Romanovsky wrote:
> On Mon, Jan 19, 2026 at 03:53:29PM -0400, Jason Gunthorpe wrote:
>> On Mon, Jan 19, 2026 at 11:43:52AM -0800, Sharath Srinivasan wrote:
>>> The sync strategy in remove_all_compat_devs() can improved
>>> by adopting that of rdma_dev_exit_net() which releases devices_rwsem
>>> before calling remove_one_compat_dev()/device_del().
>>>
>>> Also fixes a comment typo in rdma_dev_exit_net().
>>
>> You cannot change this locking without writing a huge commit message
>> explaining in detail the reason why any change like this is safe..
> 
> We can drop this patch, it doesn't even apply.
> 
> Thanks
> 

Thanks for trying. I'll send a rebased-v2 with an expanded commit message as Jason suggested.

Regards,
Sharath

>>
>> Jason
>>
>