On 29.10.2020 13.38, Sebastian Andrzej Siewior wrote:
> On 2020-10-29 13:22:20 [+0200], Mathias Nyman wrote:
>> On 29.10.2020 13.11, Mike Galbraith wrote:
>>> On Thu, 2020-10-29 at 11:41 +0200, Mathias Nyman wrote:
>>>> Can you check if this can be reproduced with 5.9 kernel?
>>>
>>> Nope, 5.9.2 didn't reproduce.
>>>
>>
>> Very odd.
>> It might be hard to reproduce as it requires xhci ring expansion to trigger it,
>> meaning at some point there is so much data queued to a device the current ring buffer
>> can't fit it.
>>
>> Other possibility is that there were some radix tree changes in 5.10-rc1, haven't looked into those.
>
> This came with commit
> 673d74683627b ("usb: xhci: add debugfs support for ep with stream")
>
> which is appeared in v5.10-rc1. This hunk below works around it:
>
> diff --git a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
> index 482fe8c5e3b47..699777fb523b6 100644
> --- a/drivers/usb/host/xhci.c
> +++ b/drivers/usb/host/xhci.c
> @@ -3533,11 +3533,13 @@ static int xhci_alloc_streams(struct usb_hcd *hcd, struct usb_device *udev,
> xhci_dbg(xhci, "Slot %u ep ctx %u now has streams.\n",
> udev->slot_id, ep_index);
> vdev->eps[ep_index].ep_state |= EP_HAS_STREAMS;
> - xhci_debugfs_create_stream_files(xhci, vdev, ep_index);
> }
> xhci_free_command(xhci, config_cmd);
> spin_unlock_irqrestore(&xhci->lock, flags);
>
> + for (i = 0; i < num_eps; i++)
> + xhci_debugfs_create_stream_files(xhci, vdev, ep_index);
> +
> /* Subtract 1 for stream 0, which drivers can't use */
> return num_streams - 1;
>
Sebastian, I think you might be missing the first mail in this thread.
https://lore.kernel.org/linux-usb/20201028203124.375344-4-mathias.nyman@xxxxxxxxxxxxxxx/
Let me do a quick recap to avoid confusion here. So far we have 2 issues.
1. Lockdep issue #1, stream debugfs entry created with local interrupts disabled and xhci->lock held
Cause: 673d74683627b ("usb: xhci: add debugfs support for ep with stream") in v5.10-rc1
Fix: [PATCH 3/3] xhci: Don't create stream debugfs files with spinlock held. (i.e. the mail we are replying to)
Comment: easily reproduced, enable lockdep and uas, mount debugfs and connect a UAS device.
discovered by both Hans and Mike
.
2. Lockdep issue #2, adding entries to radix tree during (stream) ring expansion with interrupts disabled and xhci->lock held.
Cause: unknown, probably a patch since we started using radix trees for finding streams
Fix: unknown.
Comment: Discovered by Mike when testing fix for issue#1. I suspect it can be reproduced on 5.9 but is
probably really hard as it involves ring expansion.
-Mathias
