Full kbuild support (atop of https://github.com/minipli-oss/open-gpu-kernel-modules/tree/fixes)

[minipli-oss]

This set of changes makes it possible to compile the OS-agnostic part below src/ to be compiled with Linux's kbuild which brings a lot of benefits, as in automatically picking up the right compiler flags and, hence, code instrumentation to support features like RANDSTRUCT, kCFI and various grsecurity features.

The changes build atop of the fixes branch (#1068).

Most significant changes are to get rid of system headers completely, which was kinda challenging for the C++ parts of src/nvidia-modeset/, as well as actually supporting C++ source files via kbuild. I took the poor man's approach and bend the rules for C to get all *FLAGS* / *flags* handling for free instead of reimplementing it. A "proper" implementation may do something more sophisticated.

Another challenge was to work around a limitation of kbuild for modules with lots of object files. kbuild simply doesn't allow compound object files for modules and there is, apparently, no interest in fixing this. So I had to implement a hack similar to this using a fake module. A little ugly under the hood, but works quite well.

The final obstacle was C++'s COMDAT sections confusing objtool and the linker. After finally understanding how they work, I was able fix that by simply collapsing them prior to linking the final module.

All in all quite some pragmatic choices, therefore just a proposed as a draft. However, I was able to successfully build the modules and actually use them.

Building with kbuild also allows to enable a lot of debugging options. For example, one of my build configs had UBSAN enabled which triggered a few interesting splats:

root@gpu-dev:~# dmesg | grep UBSAN -A1
[    8.831486] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/mem_mgr/io_vaspace.c:444:36
[    8.831490] index 1 is out of range for type 'RmPhysAddr [1]'
--
[    9.119483] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/arch/nvalloc/unix/src/registry.c:564:60
[    9.119487] index 1 is out of range for type 'PACKED_REGISTRY_ENTRY [*]'
--
[    9.924916] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/gpu/fifo/arch/maxwell/kernel_fifo_gm107.c:776:45
[    9.924920] index 15 is out of range for type 'NvU32 [15]'
--
[    9.957044] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/gpu/mem_mgr/mem_desc.c:2496:28
[    9.957047] index 1 is out of range for type 'RmPhysAddr [1]'
--
[    9.969087] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/mem_mgr/io_vaspace.c:336:68
[    9.969090] index 1 is out of range for type 'RmPhysAddr [1]'
--
[    9.981216] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/mem_mgr/io_vaspace.c:336:35
[    9.981219] index 1 is out of range for type 'RmPhysAddr [1]'
--
[   10.139401] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/rmapi/rmapi_cache_handlers.c:107:91
[   10.139470] index 4294967295 is out of range for type '<unknown> [32]'
--
[   10.150784] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/rmapi/rmapi_cache_handlers.c:108:91
[   10.150787] index 4294967295 is out of range for type '<unknown> [32]'
--
[   12.455149] UBSAN: array-index-out-of-bounds in /home/minipli/src/open-gpu-kernel-modules/kernel-open/../src/nvidia/src/kernel/gpu/mem_mgr/mem_desc.c:2503:32
[   12.455152] index 1 is out of range for type 'RmPhysAddr [1]'

Lots of these are false positives, as the code is making use of old-style flexibe arrays. However, the reports for kernel_fifo_gm107.c (letting though an invalid index (15)) and rmapi_cache_handlers.c or (index -1) may deserve a closer look.

[CLAassistant]

All committers have signed the CLA.

[minipli-oss]

I just realized, I completely forget about mentioning how to actually build the modules with kbuild:

$ make modules -j $(nproc) USE_KBUILD=1
[...]
$ sudo make modules_install -j $(nproc) USE_KBUILD=1

The key is to mention USE_KBUILD=1 for both, building and installing the modules.