On 10/19/2021 10:10 PM, Jason Gunthorpe wrote:
On Tue, Oct 19, 2021 at 12:55:13PM -0600, Alex Williamson wrote:
+static void mlx5vf_reset_work_handler(struct work_struct *work)
+{
+ struct mlx5vf_pci_core_device *mvdev =
+ container_of(work, struct mlx5vf_pci_core_device, work);
+
+ mutex_lock(&mvdev->state_mutex);
+ mlx5vf_reset_mig_state(mvdev);
I see this calls mlx5vf_reset_vhca_state() but how does that unfreeze
and unquiesce the device as necessary to get back to _RUNNING?
FLR of the function does it.
Same flow as if userspace attaches the vfio migration, freezes the
device then closes the FD. The FLR puts everything in the device right
and the next open will see a functional, unfrozen, blank device.
Right
+ mvdev->vmig.vfio_dev_state = VFIO_DEVICE_STATE_RUNNING;
+ mutex_unlock(&mvdev->state_mutex);
+}
+
+static void mlx5vf_pci_aer_reset_done(struct pci_dev *pdev)
+{
+ struct mlx5vf_pci_core_device *mvdev = dev_get_drvdata(&pdev->dev);
+
+ if (!mvdev->migrate_cap)
+ return;
+
+ schedule_work(&mvdev->work);
This seems troublesome, how long does userspace poll the device state
after reset to get back to _RUNNING? Seems we at least need a
flush_work() call when userspace reads the device state. Thanks,
The locking is very troubled here because the higher VFIO layers are
holding locks across reset and using those same locks with the mm_lock
The delay is a good point :(
What is the expectation for a reasonable delay ? we may expect this
system WQ to run only short tasks and be very responsive.
See
https://elixir.bootlin.com/linux/v5.15-rc6/source/include/linux/workqueue.h#L355
The other algorithm that can rescue this is to defer the cleanup work
to the mutex unlock, which ever context happens to get to it:
reset_done:
spin_lock(spin)
defered_reset = true;
if (!mutex_trylock(&state_mutex))
spin_unlock(spin)
return
spin_unlock(spin)
state_mutex_unlock()
state_mutex_unlock:
again:
spin_lock(spin)
if (defered_reset)
spin_unlock()
do_post_reset;
goto again;
mutex_unlock(state_mutex);
spin_unlock()
and call state_mutex_unlock() in all unlock cases.
It is a much more complicated algorithm than the work.
Right, it seems much more complicated compared to current code..
In addition, need to get into the details of the above algorithm, not
sure that I got all of them ..
Yishai this should also have had a comment explaining why this is
needed as nobody is going to guess a ABBA deadlock on mm_lock is the
reason.
Sure, this can be done.
Are we fine to start/stay with current simple approach that I wrote and
tested so far ?
Yishai
