Rename struct device to struct _device

[markj]

Summary

Right now types.h has device_t as a typedef of struct device *. struct
device is defined in subr_bus.c. This means that subsystems that define
a struct device can easily introduce type confusion. The LinuxKPI does
exactly this and there are a couple of bugs as a result.

The type confusion also causes problems for debuggers, since device_t
can refer to different structures depending on how the debugger tries to
resolve it to a type. It also results in type duplication in the CTF
type graph.

Rename struct device to struct _device to avoid these problems.

Almost all of the kernel refers to devices with device_t. A small
handful of drivers referenced struct device * instead, and I have
separate commits to fix those. This diff contains the main name change
and updates headers that shouldn't depend on types.h.

[markj]

Use struct device_list.

[markj]

Update a couple of man9 pages that referenced struct device.

[gbe]

LGTM from manpages

[imp]

The root cause here is that when linking different compilation units together, struct foo needs to be globally the same for strict compliance with the C standard.
In the long run, the Linux KPI should define strict linux_kpi_device and #define compatibility from there for the user visible parts of the KPI. I imagine this will be quite hard, in practice, given the nature of the linuxkpi code though.

In the mean time, changing this to _device is fine with me. It resolves the conflict in a way that's adequate.

[markj]

In D29676#665663, @imp wrote:

The root cause here is that when linking different compilation units together, struct foo needs to be globally the same for strict compliance with the C standard.
In the long run, the Linux KPI should define strict linux_kpi_device and #define compatibility from there for the user visible parts of the KPI. I imagine this will be quite hard, in practice, given the nature of the linuxkpi code though.

"device" is a common identifier so I'm not sure how easy it would be to use #define to fix the problem. I remember that this topic came up with the DRM drivers and there weren't any attractive solutions.

[imp]

In D29676#665678, @markj wrote:

In D29676#665663, @imp wrote:

The root cause here is that when linking different compilation units together, struct foo needs to be globally the same for strict compliance with the C standard.
In the long run, the Linux KPI should define strict linux_kpi_device and #define compatibility from there for the user visible parts of the KPI. I imagine this will be quite hard, in practice, given the nature of the linuxkpi code though.

"device" is a common identifier so I'm not sure how easy it would be to use #define to fix the problem. I remember that this topic came up with the DRM drivers and there weren't any attractive solutions.

True. Which is why _device is good for now, but the issues of namespace collision will continue to vex us in general... I too see issues with a simple #define because it is an imperfect match to c's multiple namespaces...

[rpokala]

sys/kern/subr_bus.c

118

* The structure is named "_device" instead of "device" to avoid type
* confusion caused by other subsystems defining a (struct device).

For readability.

sys/powerpc/include/bus_dma.h

36

Why not

int bus_dma_tag_set_iommu(bus_dma_tag_t, device_t iommu, void *cookie);

?

sys/sys/pcpu.h

190

Similarly,

device_t pc_device; /* CPU device handle */

sys/sys/systm.h

630

And

void _gone_in_dev(device_t dev, int major, const char *msg);

[markj]

Address feedback.

[rpokala]

LGTM, FWIW.