On Wed, 2024-03-13 at 11:16 +1300, Hamish Martin wrote:
> There is an issue with ACPI overlay table removal specifically
> related
> to I2C multiplexers.
>
> Consider an ACPI SSDT Overlay that defines a PCA9548 I2C mux on an
> existing I2C bus. When this table is loaded we see the creation of a
> device for the overall PCA9548 chip and 8 further devices - one
> i2c_adapter each for the mux channels. These are all bound to their
> ACPI equivalents via an eventual invocation of acpi_bind_one().
>
> When we unload the SSDT overlay we run into the problem. The ACPI
> devices are deleted as normal via acpi_device_del_work_fn() and the
> acpi_device_del_list.
>
> However, the following warning and stack trace is output as the
> deletion does not go smoothly:
> ------------[ cut here ]------------
> kernfs: can not remove 'physical_node', no directory
> WARNING: CPU: 1 PID: 11 at fs/kernfs/dir.c:1674
> kernfs_remove_by_name_ns+0xb9/0xc0
> Modules linked in:
> CPU: 1 PID: 11 Comm: kworker/u128:0 Not tainted 6.8.0-rc6+ #1
> Hardware name: congatec AG conga-B7E3/conga-B7E3, BIOS 5.13
> 05/16/2023
> Workqueue: kacpi_hotplug acpi_device_del_work_fn
> RIP: 0010:kernfs_remove_by_name_ns+0xb9/0xc0
> Code: e4 00 48 89 ef e8 07 71 db ff 5b b8 fe ff ff ff 5d 41 5c 41 5d
> e9 a7 55 e4 00 0f 0b eb a6 48 c7 c7 f0 38 0d 9d e8 97 0a d5 ff <0f>
> 0b eb dc 0f 1f 00 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90
> RSP: 0018:ffff9f864008fb28 EFLAGS: 00010286
> RAX: 0000000000000000 RBX: ffff8ef90a8d4940 RCX: 0000000000000000
> RDX: ffff8f000e267d10 RSI: ffff8f000e25c780 RDI: ffff8f000e25c780
> RBP: ffff8ef9186f9870 R08: 0000000000013ffb R09: 00000000ffffbfff
> R10: 00000000ffffbfff R11: ffff8f000e0a0000 R12: ffff9f864008fb50
> R13: ffff8ef90c93dd60 R14: ffff8ef9010d0958 R15: ffff8ef9186f98c8
> FS: 0000000000000000(0000) GS:ffff8f000e240000(0000)
> knlGS:0000000000000000
> CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00007f48f5253a08 CR3: 00000003cb82e000 CR4: 00000000003506f0
> Call Trace:
> <TASK>
> ? kernfs_remove_by_name_ns+0xb9/0xc0
> ? __warn+0x7c/0x130
> ? kernfs_remove_by_name_ns+0xb9/0xc0
> ? report_bug+0x171/0x1a0
> ? handle_bug+0x3c/0x70
> ? exc_invalid_op+0x17/0x70
> ? asm_exc_invalid_op+0x1a/0x20
> ? kernfs_remove_by_name_ns+0xb9/0xc0
> ? kernfs_remove_by_name_ns+0xb9/0xc0
> acpi_unbind_one+0x108/0x180
> device_del+0x18b/0x490
> ? srso_return_thunk+0x5/0x5f
> ? srso_return_thunk+0x5/0x5f
> device_unregister+0xd/0x30
> i2c_del_adapter.part.0+0x1bf/0x250
> i2c_mux_del_adapters+0xa1/0xe0
> i2c_device_remove+0x1e/0x80
> device_release_driver_internal+0x19a/0x200
> bus_remove_device+0xbf/0x100
> device_del+0x157/0x490
> ? __pfx_device_match_fwnode+0x10/0x10
> ? srso_return_thunk+0x5/0x5f
> device_unregister+0xd/0x30
> i2c_acpi_notify+0x10f/0x140
> notifier_call_chain+0x58/0xd0
> blocking_notifier_call_chain+0x3a/0x60
> acpi_device_del_work_fn+0x85/0x1d0
> process_one_work+0x134/0x2f0
> worker_thread+0x2f0/0x410
> ? __pfx_worker_thread+0x10/0x10
> kthread+0xe3/0x110
> ? __pfx_kthread+0x10/0x10
> ret_from_fork+0x2f/0x50
> ? __pfx_kthread+0x10/0x10
> ret_from_fork_asm+0x1b/0x30
> </TASK>
> ---[ end trace 0000000000000000 ]---
> ...
> repeated 7 more times, 1 for each channel of the mux
> ...
>
> The issue is that the binding of the ACPI devices to their peer I2C
> adapters is not correctly cleaned up. Digging deeper into the issue
> we
> see that the deletion order is such that the ACPI devices matching
> the
> mux channel i2c adapters are deleted first during the SSDT overlay
> removal. For each of the channels we see a call to i2c_acpi_notify()
> with ACPI_RECONFIG_DEVICE_REMOVE but, because these devices are not
> actually i2c_clients, nothing is done for them.
>
> Later on, after each of the mux channels has been dealt with, we come
> to delete the i2c_client representing the PCA9548 device. This is the
> call stack we see above, whereby the kernel cleans up the i2c_client
> including destruction of the mux and its channel adapters. At this
> point we do attempt to unbind from the ACPI peers but those peers no
> longer exist and so we hit the kernfs errors.
>
> The fix is to augment i2c_acpi_notify() to handle i2c_adapters. But,
> given that the life cycle of the adapters is linked to the
> i2c_client,
> instead of deleting the i2c_adapters during the i2c_acpi_notify(), we
> just trigger unbinding of the ACPI device from the adapter device,
> and
> allow the clean up of the adapter to continue in the way it always
> has.
>
> Signed-off-by: Hamish Martin <hamish.martin@xxxxxxxxxxxxxxxxxxx>
> Reviewed-by: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
> Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxx>
> Fixes: 525e6fabeae2 ("i2c / ACPI: add support for ACPI reconfigure
> notifications")
> Cc: <stable@xxxxxxxxxxxxxxx> # v4.8+
> ---
>
> Notes:
> v4:
> Resolve Build failure noted by:
> Linux Kernel Functional Testing <lkft@xxxxxxxxxx>, and
> kernel test robot <lkp@xxxxxxxxx>
> These failures led to revert of the v3 version of this patch
> that had been accepted earlier.
> v3:
> Add reviewed by tags (Mika Westerberg and Andi Shyti) and Fixes
> tag.
> v2:
> Moved long problem description from cover letter to commit
> description at Mika's suggestion
>
> drivers/i2c/i2c-core-acpi.c | 19 +++++++++++++++----
> 1 file changed, 15 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/i2c/i2c-core-acpi.c b/drivers/i2c/i2c-core-
> acpi.c
> index d6037a328669..14ae0cfc325e 100644
> --- a/drivers/i2c/i2c-core-acpi.c
> +++ b/drivers/i2c/i2c-core-acpi.c
> @@ -445,6 +445,11 @@ static struct i2c_client
> *i2c_acpi_find_client_by_adev(struct acpi_device *adev)
> return i2c_find_device_by_fwnode(acpi_fwnode_handle(adev));
> }
>
> +static struct i2c_adapter *i2c_acpi_find_adapter_by_adev(struct
> acpi_device *adev)
> +{
> + return i2c_find_adapter_by_fwnode(acpi_fwnode_handle(adev));
> +}
> +
> static int i2c_acpi_notify(struct notifier_block *nb, unsigned long
> value,
> void *arg)
> {
> @@ -471,11 +476,17 @@ static int i2c_acpi_notify(struct
> notifier_block *nb, unsigned long value,
> break;
>
> client = i2c_acpi_find_client_by_adev(adev);
> - if (!client)
> - break;
> + if (client) {
> + i2c_unregister_device(client);
> + put_device(&client->dev);
> + }
> +
> + adapter = i2c_acpi_find_adapter_by_adev(adev);
> + if (adapter) {
> + acpi_unbind_one(&adapter->dev);
> + put_device(&adapter->dev);
> + }
>
> - i2c_unregister_device(client);
> - put_device(&client->dev);
> break;
> }
>
I wonder if this patch has slipped through the net due to a process
error on my part.
The v3 version was accepted by Wolfram into for-current on March 4th,
but a build issue was detected by a couple of test robots, leading to
it being reverted 3 days later.
I submitted this corrected v4 version on March 13th. Perhaps I should
have not submitted a v4 and that has screwed something up, or perhaps
folks are just busy ;-).
Let me know what further work is required to get this one accepted.
Thanks,
Hamish M.
