mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device
@ 2026-11-11 11:56 Slark Xiao
  2026-11-13 18:42 ` Jeff Hugo
  0 siblings, 1 reply; 4+ messages in thread
From: Slark Xiao @ 2026-11-11 11:56 UTC (permalink / raw)
  To: mani, jeff.hugo, bbhatt
  Cc: mhi, linux-arm-msm, linux-kernel, Slark Xiao, stable

The serial_number, oem_pk_hash and soc_reset attributes are registered
through mhi_bus_type.dev_groups, so the driver core creates them for
every device on the MHI bus.  MHI client devices (the per-channel
devices such as mhi0_DUN or mhi0_MBIM) are also MHI bus devices, so
each of them ends up with its own copies of these files even though
they all operate on the same mhi_controller:

  $ ls /sys/bus/mhi/devices/mhi0_DUN/
  driver  oem_pk_hash  power  serial_number  soc_reset  subsystem  uevent

This is wrong for several reasons:

  - The attributes describe the controller, not the channel.  Their
    callbacks all dereference mhi_dev->mhi_cntrl and read BHI registers
    or issue a controller-wide SoC reset, so writing soc_reset through
    mhi0_DUN resets the whole device while looking like a per-channel
    operation.
  - oem_pk_hash reports failures with dev_err(dev, ...), so a single
    failing controller produces one log line per registered channel
    device instead of one.
  - Userspace has no way to tell which copy is authoritative.

Move the attributes off the bus and attach them to the controller
device only.  mhi_register_controller() now installs the group via
mhi_dev->dev.groups before device_add(), while the channel devices
created by mhi_create_devices() are left untouched.  No functional
change for the controller device; the files stay at the paths listed
in Documentation/ABI/stable/sysfs-bus-mhi.

Note that trigger_edl was already handled this way, through an explicit
sysfs_create_file() on the controller kobject, which is why it never
appeared on the channel devices.

Fixes: 21cb9b18edea ("bus: mhi: core: Introduce sysfs entries for MHI")
Cc: stable@vger.kernel.org
Signed-off-by: Slark Xiao <slark_xiao@163.com>
---
 drivers/bus/mhi/host/init.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/drivers/bus/mhi/host/init.c b/drivers/bus/mhi/host/init.c
index fd3050889412..b4b254d1c6b9 100644
--- a/drivers/bus/mhi/host/init.c
+++ b/drivers/bus/mhi/host/init.c
@@ -150,13 +150,21 @@ static ssize_t trigger_edl_store(struct device *dev,
 }
 static DEVICE_ATTR_WO(trigger_edl);
 
-static struct attribute *mhi_dev_attrs[] = {
+static struct attribute *mhi_cntrl_dev_attrs[] = {
 	&dev_attr_serial_number.attr,
 	&dev_attr_oem_pk_hash.attr,
 	&dev_attr_soc_reset.attr,
 	NULL,
 };
-ATTRIBUTE_GROUPS(mhi_dev);
+
+static const struct attribute_group mhi_cntrl_dev_group = {
+	.attrs = mhi_cntrl_dev_attrs,
+};
+
+static const struct attribute_group *mhi_cntrl_dev_groups[] = {
+	&mhi_cntrl_dev_group,
+	NULL,
+};
 
 /* MHI protocol requires the transfer ring to be aligned with ring length */
 static int mhi_alloc_aligned_ring(struct mhi_controller *mhi_cntrl,
@@ -1019,6 +1027,8 @@ int mhi_register_controller(struct mhi_controller *mhi_cntrl,
 	dev_set_name(&mhi_dev->dev, "mhi%d", mhi_cntrl->index);
 	mhi_dev->name = dev_name(&mhi_dev->dev);
 
+	mhi_dev->dev.groups = mhi_cntrl_dev_groups;
+
 	/* Init wakeup source */
 	device_init_wakeup(&mhi_dev->dev, true);
 
@@ -1455,7 +1465,6 @@ const struct bus_type mhi_bus_type = {
 	.uevent = mhi_uevent,
 	.probe = mhi_probe,
 	.remove = mhi_remove,
-	.dev_groups = mhi_dev_groups,
 };
 
 static int __init mhi_init(void)
-- 
2.53.0


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device
  2026-11-11 11:56 [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device Slark Xiao
@ 2026-11-13 18:42 ` Jeff Hugo
  2026-11-18  7:05   ` Slark Xiao
  0 siblings, 1 reply; 4+ messages in thread
From: Jeff Hugo @ 2026-11-13 18:42 UTC (permalink / raw)
  To: Slark Xiao, mani, bbhatt; +Cc: mhi, linux-arm-msm, linux-kernel, stable

On 11/11/2026 4:56 AM, Slark Xiao wrote:
> The serial_number, oem_pk_hash and soc_reset attributes are registered
> through mhi_bus_type.dev_groups, so the driver core creates them for
> every device on the MHI bus.  MHI client devices (the per-channel
> devices such as mhi0_DUN or mhi0_MBIM) are also MHI bus devices, so
> each of them ends up with its own copies of these files even though
> they all operate on the same mhi_controller:
> 
>    $ ls /sys/bus/mhi/devices/mhi0_DUN/
>    driver  oem_pk_hash  power  serial_number  soc_reset  subsystem  uevent
> 
> This is wrong for several reasons:
> 
>    - The attributes describe the controller, not the channel.  Their
>      callbacks all dereference mhi_dev->mhi_cntrl and read BHI registers
>      or issue a controller-wide SoC reset, so writing soc_reset through
>      mhi0_DUN resets the whole device while looking like a per-channel
>      operation.
>    - oem_pk_hash reports failures with dev_err(dev, ...), so a single
>      failing controller produces one log line per registered channel
>      device instead of one.
>    - Userspace has no way to tell which copy is authoritative.
> 
> Move the attributes off the bus and attach them to the controller
> device only.  mhi_register_controller() now installs the group via
> mhi_dev->dev.groups before device_add(), while the channel devices
> created by mhi_create_devices() are left untouched.  No functional
> change for the controller device; the files stay at the paths listed
> in Documentation/ABI/stable/sysfs-bus-mhi.
> 
> Note that trigger_edl was already handled this way, through an explicit
> sysfs_create_file() on the controller kobject, which is why it never
> appeared on the channel devices.
I agree with the rationale, but this does appear to change uAPI. Are you 
sure no one is using these on the channel devices?

-Jeff

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re:Re: [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device
  2026-11-13 18:42 ` Jeff Hugo
@ 2026-11-18  7:05   ` Slark Xiao
  2026-11-20 16:45     ` Jeff Hugo
  0 siblings, 1 reply; 4+ messages in thread
From: Slark Xiao @ 2026-11-18  7:05 UTC (permalink / raw)
  To: Jeff Hugo; +Cc: mani, bbhatt, mhi, linux-arm-msm, linux-kernel, stable


At 2026-11-14 02:42:46, "Jeff Hugo" <jeff.hugo@oss.qualcomm.com> wrote:
>On 11/11/2026 4:56 AM, Slark Xiao wrote:
>> The serial_number, oem_pk_hash and soc_reset attributes are registered
>> through mhi_bus_type.dev_groups, so the driver core creates them for
>> every device on the MHI bus.  MHI client devices (the per-channel
>> devices such as mhi0_DUN or mhi0_MBIM) are also MHI bus devices, so
>> each of them ends up with its own copies of these files even though
>> they all operate on the same mhi_controller:
>> 
>>    $ ls /sys/bus/mhi/devices/mhi0_DUN/
>>    driver  oem_pk_hash  power  serial_number  soc_reset  subsystem  uevent
>> 
>> This is wrong for several reasons:
>> 
>>    - The attributes describe the controller, not the channel.  Their
>>      callbacks all dereference mhi_dev->mhi_cntrl and read BHI registers
>>      or issue a controller-wide SoC reset, so writing soc_reset through
>>      mhi0_DUN resets the whole device while looking like a per-channel
>>      operation.
>>    - oem_pk_hash reports failures with dev_err(dev, ...), so a single
>>      failing controller produces one log line per registered channel
>>      device instead of one.
>>    - Userspace has no way to tell which copy is authoritative.
>> 
>> Move the attributes off the bus and attach them to the controller
>> device only.  mhi_register_controller() now installs the group via
>> mhi_dev->dev.groups before device_add(), while the channel devices
>> created by mhi_create_devices() are left untouched.  No functional
>> change for the controller device; the files stay at the paths listed
>> in Documentation/ABI/stable/sysfs-bus-mhi.
>> 
>> Note that trigger_edl was already handled this way, through an explicit
>> sysfs_create_file() on the controller kobject, which is why it never
>> appeared on the channel devices.
>I agree with the rationale, but this does appear to change uAPI. Are you 
>sure no one is using these on the channel devices?
>
>-Jeff
Hi Jeff,
For oem_pk_hash and serial_number, all interface returns same value.
I think there is no need to keep these interfaces under the channel 
devices. 

For soc_reset, it write the register directly and it works for the entire
device rather than a single channel.

So I think there is no problem to remove them.
Additionally, I reported this to Mani a long time ago, and he confirmed
that it was indeed a bug.

2024/7/24:
Slark: Hi Mani, do you think if it's a bug? The soc_reset and other
attr would be enumerated both the mhi0 and other mhi0_PORT
Mani: Gosh, yes 
They belong to the controller
Can you send a patch? Else I'll fix it

Thanks 

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device
  2026-11-18  7:05   ` Slark Xiao
@ 2026-11-20 16:45     ` Jeff Hugo
  0 siblings, 0 replies; 4+ messages in thread
From: Jeff Hugo @ 2026-11-20 16:45 UTC (permalink / raw)
  To: Slark Xiao; +Cc: mani, mhi, linux-arm-msm, linux-kernel, stable

On 11/18/2026 12:05 AM, Slark Xiao wrote:
> 
> At 2026-11-14 02:42:46, "Jeff Hugo" <jeff.hugo@oss.qualcomm.com> wrote:
>> On 11/11/2026 4:56 AM, Slark Xiao wrote:
>>> The serial_number, oem_pk_hash and soc_reset attributes are registered
>>> through mhi_bus_type.dev_groups, so the driver core creates them for
>>> every device on the MHI bus.  MHI client devices (the per-channel
>>> devices such as mhi0_DUN or mhi0_MBIM) are also MHI bus devices, so
>>> each of them ends up with its own copies of these files even though
>>> they all operate on the same mhi_controller:
>>>
>>>     $ ls /sys/bus/mhi/devices/mhi0_DUN/
>>>     driver  oem_pk_hash  power  serial_number  soc_reset  subsystem  uevent
>>>
>>> This is wrong for several reasons:
>>>
>>>     - The attributes describe the controller, not the channel.  Their
>>>       callbacks all dereference mhi_dev->mhi_cntrl and read BHI registers
>>>       or issue a controller-wide SoC reset, so writing soc_reset through
>>>       mhi0_DUN resets the whole device while looking like a per-channel
>>>       operation.
>>>     - oem_pk_hash reports failures with dev_err(dev, ...), so a single
>>>       failing controller produces one log line per registered channel
>>>       device instead of one.
>>>     - Userspace has no way to tell which copy is authoritative.
>>>
>>> Move the attributes off the bus and attach them to the controller
>>> device only.  mhi_register_controller() now installs the group via
>>> mhi_dev->dev.groups before device_add(), while the channel devices
>>> created by mhi_create_devices() are left untouched.  No functional
>>> change for the controller device; the files stay at the paths listed
>>> in Documentation/ABI/stable/sysfs-bus-mhi.
>>>
>>> Note that trigger_edl was already handled this way, through an explicit
>>> sysfs_create_file() on the controller kobject, which is why it never
>>> appeared on the channel devices.
>> I agree with the rationale, but this does appear to change uAPI. Are you
>> sure no one is using these on the channel devices?
>>
>> -Jeff
> Hi Jeff,
> For oem_pk_hash and serial_number, all interface returns same value.
> I think there is no need to keep these interfaces under the channel
> devices.
> 
> For soc_reset, it write the register directly and it works for the entire
> device rather than a single channel.
> 
> So I think there is no problem to remove them.

These are uAPI. The golden rule of the kernel is that we do not break 
userspace. It does not matter that their existence is a "bug", if there 
is userspace using them, then we must keep them.

I could agree with you assessment that having these items on the channel 
device is a mistake (and I do), but that does not change the rule.

So far you have not provided any data to say that userspace is in fact 
not using these, and therefore removing them would have no impact.

Without that analysis, I am sorry to say I am uncomfortable proceeding 
with this change.  Mani is free to ignore my concerns and apply this 
patch at his own risk.

-Jeff

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2026-11-20 16:45 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-11 11:56 [PATCH] bus: mhi: host: Only expose controller sysfs attributes on the controller device Slark Xiao
2026-11-13 18:42 ` Jeff Hugo
2026-11-18  7:05   ` Slark Xiao
2026-11-20 16:45     ` Jeff Hugo

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®