Re: linux-next: build failure after merge of the driver-core tree

From: Jeffrey Hugo
Date: Tue Apr 11 2023 - 13:19:25 EST


On 4/11/2023 10:31 AM, Daniel Vetter wrote:
On Tue, Apr 11, 2023 at 09:29:27AM -0600, Jeffrey Hugo wrote:
On 4/11/2023 9:26 AM, Jeffrey Hugo wrote:
On 4/11/2023 9:13 AM, Greg KH wrote:
On Tue, Apr 11, 2023 at 09:08:39AM -0600, Jeffrey Hugo wrote:
On 4/11/2023 9:01 AM, Daniel Vetter wrote:
On Tue, Apr 11, 2023 at 12:40:28PM +0200, Greg KH wrote:
On Tue, Apr 11, 2023 at 11:55:20AM +0200, Daniel Vetter wrote:
On Tue, Apr 11, 2023 at 02:38:12PM +1000, Stephen Rothwell wrote:
Hi all,

After merging the driver-core tree, today's linux-next build (x86_64
allmodconfig) failed like this:

In file included from include/linux/linkage.h:7,
                   from include/linux/kernel.h:17,
                   from drivers/accel/qaic/mhi_qaic_ctrl.c:4:
drivers/accel/qaic/mhi_qaic_ctrl.c: In function
'mhi_qaic_ctrl_init':
include/linux/export.h:27:22: error: passing
argument 1 of 'class_create' from incompatible
pointer type
[-Werror=incompatible-pointer-types]
     27 | #define THIS_MODULE (&__this_module)
        |                     ~^~~~~~~~~~~~~~~
        |                      |
        |                      struct module *
drivers/accel/qaic/mhi_qaic_ctrl.c:544:38: note:
in expansion of macro 'THIS_MODULE'
    544 |         mqc_dev_class =
class_create(THIS_MODULE,
MHI_QAIC_CTRL_DRIVER_NAME);
        |                                      ^~~~~~~~~~~
In file included from include/linux/device.h:31,
                   from include/linux/mhi.h:9,
                   from drivers/accel/qaic/mhi_qaic_ctrl.c:5:
include/linux/device/class.h:229:54: note:
expected 'const char *' but argument is of type
'struct module *'
    229 | struct class * __must_check
class_create(const char *name);
        |                                          ~~~~~~~~~~~~^~~~
drivers/accel/qaic/mhi_qaic_ctrl.c:544:25:
error: too many arguments to function
'class_create'
    544 |         mqc_dev_class =
class_create(THIS_MODULE,
MHI_QAIC_CTRL_DRIVER_NAME);
        |                         ^~~~~~~~~~~~
include/linux/device/class.h:229:29: note: declared here
    229 | struct class * __must_check
class_create(const char *name);
        |                             ^~~~~~~~~~~~

Caused by commit

    1aaba11da9aa ("driver core: class: remove
module * from class_create()")

interacting with commit

    566fc96198b4 ("accel/qaic: Add mhi_qaic_cntl")

from the drm tree.

I have applied the following merge fix patch for today.

From: Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx>
Date: Tue, 11 Apr 2023 14:16:57 +1000
Subject: [PATCH] fixup for "driver core: class:
remove module * from class_create()"

interacting with "accel/qaic: Add mhi_qaic_cntl"

Signed-off-by: Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx>

Thanks for the fixup. Since Dave is out I've made a
note about this in my
handover mail so it won't get lost in the drm-next
merge window pull. I
don't think we need any other coordination than
mention it in each pull to
Linus, topic tree seems overkill for this. Plus there's no way I can
untangle the drm tree anyway :-).

Want me to submit a patch for the drm tree that moves this to use
class_register() instead, which will make the
merge/build issue go away
for you?  That's my long-term goal here anyway, so converting this new
code to this api today would be something I have to do eventually :)

We kinda closed drm-next for feature work mostly already (just pulling
stuff in from subtrees), so won't really help for this merge window.

For everything else I think this is up to Oded, I had no
idea qaic needed
it's entire own dev class and I don't want to dig into this
for the risk I
might freak out :-)

Adding Oded.

Cheers, Daniel

Sorry for the mess.

I made a note to update to class_register() once my drm-misc access is
sorted out.  Looks like we'll address the conflict in the merge
window, and
catch the update to the new API in the following release.

Wait, I think the large question is, "why does this need a separate
class"?  Why are you not using the accel char device and class?  That is
what everything under accel/ should be using, otherwise why put it in
there?

And what exactly are you using that class for?  Just device nodes?  If
so, how many?

thanks,

greg k-h


Remember MHI_UCI that then evolved into the WWAN subsystem?  I pointed
out at the time that AIC100/QAIC would need the same functionality.
You/Jakub told myself/Mani/Loic that a combined implementation is not
acceptable, and every area needs to implement their own version of
MHI_UCI.

We took the WWAN subsystem and simplified it to meet our needs.

The functionality is QAIC specific, so wedging it into the Accel node
seems to be a poor fit as it would subject Habana and iVPU to the same.

Also, I forgot to mention. QAIC is sharing userspace components with WWAN,
so we really cannot diverge from what WWAN has done and define a new API
through the Accel node.

So there is an accel/drm_device in the qaic driver, but there's also this
different class thing, which I don't get.

And yeah if that's an entirely orthogonal thing then I guess that should
be in a different driver/subsystem, all supported with the aux bus to
multiplex the underlying device.

I haven't found any explanation for what MHI is (or any of the other
acrynoms), so I'm entirely lost.

MHI is documented at Documentation/mhi/
It is also referenced in the QAIC documentation - Documentation/accel/qaic/

It stands for "Modem Host Interface" (arguably a bad name now, but you can guess where it came from). It is a Qualcomm hardware block and associated software protocol that provides logical channels over a hardware link. Most commonly used for PCIe.

Pretty much any modern Qualcomm PCIe device implements it. 4G modems, 5G modems, Wifi adapters, AIC100, etc. Instead of talking "PCIe", the host talks "MHI" to the devices in most cases.

The core implementation for MHI exists in drivers/bus/mhi

MHI_UCI is the MHI Userspace Character Interface. It looked like most buses (eg USB) provide some direct device access to userspace. MHI_UCI was formulated along those same lines - provide direct userspace access to a whitelist of channels. Qualcomm provides some fairly extensive userspace utilities, and various communities have developed open source alternatives using this mechanism.

MHI_UCI was proposed to the community as the common driver (misc device) for all of the MHI devices. The Net folks came along, saw that it was used for 4G/5G modems (Wireless Wide Area Network devices or WWAN) and decided that they would not tolerate a common implementation. They NACK'd MHI_UCI and required that a WWAN specific subsystem be developed which would only service WWAN devices. The Net folks decreed that other subsystems which needed the same functionality need to have their own copy of the implementation.

QAIC devices expose Sahara (a boot time protocol) which has an existing userspace that is also used with Modems, although it looks like WWAN doesn't currently support those generations of products today. QAIC devices also support DIAG, which is currently supported in WWAN. The intent was to add the QAIC support for DIAG at a later time since it is not required for the bare minimum viable driver.

So, QAIC devices support the same services, would use the same userspace, but can't use a common implementation because Jakub(net) doesn't want to share and convinced Greg to go along. I'm not interested in pushing a cross tree fight (arguably already did that with MHI_UCI). If neither Greg nor Net will accept a common implementation that accelerators can use (QAIC), then the only place I can fit this is in the Accel area.

Using aux bus seems to make little difference if QAIC is the only consumer of this. I'm willing to refactor the implementation with some feedback and guidence, but the uAPI seems set in stone due to the existing userspace and WWAN (char devs with open/close/read/write/poll).

What would make you less unhappy?

-Daniel



We need (eventually) 128 device nodes.  We have systems with 32 QAIC
devices, and each QAIC device uses 4 device nodes (32 * 4 = 128).  WWAN
subsystem would be similar.  Looks like each 5G modem is 6 nodes per
device, so if you had 22 5G modems on a system, you'd have 132 device
nodes.  I'm not aware of any such system, but it could exist.

-Jeff