Enabling SR-IOV VFs will cause the PCI subsystem to schedule a
work and flush its workqueue. Since the nfp driver schedules its
own work we can't enable VFs while holding driver load. Commit
6d48ceb27af1 ("nfp: allocate a private workqueue for driver work")
tried to avoid this deadlock by creating a separate workqueue.
Unfortunately, due to the architecture of workqueue subsystem this
does not guarantee a separate thread of execution. Luckily
we can simply take pci_enable_sriov() from under the driver lock.
Take pci_disable_sriov() from under the lock too for symmetry.
Fixes:
6d48ceb27af1 ("nfp: allocate a private workqueue for driver work")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct nfp_pf *pf = pci_get_drvdata(pdev);
int err;
- mutex_lock(&pf->lock);
-
if (num_vfs > pf->limit_vfs) {
nfp_info(pf->cpp, "Firmware limits number of VFs to %u\n",
pf->limit_vfs);
- err = -EINVAL;
- goto err_unlock;
+ return -EINVAL;
}
err = pci_enable_sriov(pdev, num_vfs);
if (err) {
dev_warn(&pdev->dev, "Failed to enable PCI SR-IOV: %d\n", err);
- goto err_unlock;
+ return err;
}
+ mutex_lock(&pf->lock);
+
err = nfp_app_sriov_enable(pf->app, num_vfs);
if (err) {
dev_warn(&pdev->dev,
return num_vfs;
err_sriov_disable:
- pci_disable_sriov(pdev);
-err_unlock:
mutex_unlock(&pf->lock);
+ pci_disable_sriov(pdev);
return err;
#endif
return 0;
pf->num_vfs = 0;
+ mutex_unlock(&pf->lock);
+
pci_disable_sriov(pdev);
dev_dbg(&pdev->dev, "Removed VFs.\n");
-
- mutex_unlock(&pf->lock);
#endif
return 0;
}