Re: [PATCH v5 2/3] media: rc: introduce BPF_PROG_LIRC_MODE2

[Matthias Reichl <hias@horus.com>]

Hi Sean,

I finally found the time to test your patch series and noticed
2 issues - comments are inline

On Sun, May 27, 2018 at 12:24:09PM +0100, Sean Young wrote:
> diff --git a/drivers/media/rc/Kconfig b/drivers/media/rc/Kconfig
> index eb2c3b6eca7f..d5b35a6ba899 100644
> --- a/drivers/media/rc/Kconfig
> +++ b/drivers/media/rc/Kconfig
> @@ -25,6 +25,19 @@ config LIRC
>  	   passes raw IR to and from userspace, which is needed for
>  	   IR transmitting (aka "blasting") and for the lirc daemon.
>  
> +config BPF_LIRC_MODE2
> +	bool "Support for eBPF programs attached to lirc devices"
> +	depends on BPF_SYSCALL
> +	depends on RC_CORE=y

Requiring rc-core to be built into the kernel could become
problematic in the future for people using media_build.

Currently the whole media tree (including rc-core) can be built
as modules so DVB and IR drivers can be replaced by newer versions.
But with rc-core in the kernel things could easily break if internal
data structures are changed.

Maybe we should add a small layer with a stable API/ABI between
bpf-lirc and rc-core to decouple them? Or would it be possible
to build rc-core with bpf support as a module?

> +	depends on LIRC
> +	help
> +	   Allow attaching eBPF programs to a lirc device using the bpf(2)
> +	   syscall command BPF_PROG_ATTACH. This is supported for raw IR
> +	   receivers.
> +
> +	   These eBPF programs can be used to decode IR into scancodes, for
> +	   IR protocols not supported by the kernel decoders.
> +
>  menuconfig RC_DECODERS
>  	bool "Remote controller decoders"
>  	depends on RC_CORE
> [...]
> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> index 388d4feda348..3c104113d040 100644
> --- a/kernel/bpf/syscall.c
> +++ b/kernel/bpf/syscall.c
> @@ -11,6 +11,7 @@
>   */
>  #include <linux/bpf.h>
>  #include <linux/bpf_trace.h>
> +#include <linux/bpf_lirc.h>
>  #include <linux/btf.h>
>  #include <linux/syscalls.h>
>  #include <linux/slab.h>
> @@ -1578,6 +1579,8 @@ static int bpf_prog_attach(const union bpf_attr *attr)
>  	case BPF_SK_SKB_STREAM_PARSER:
>  	case BPF_SK_SKB_STREAM_VERDICT:
>  		return sockmap_get_from_fd(attr, BPF_PROG_TYPE_SK_SKB, true);
> +	case BPF_LIRC_MODE2:
> +		return lirc_prog_attach(attr);
>  	default:
>  		return -EINVAL;
>  	}
> @@ -1648,6 +1651,8 @@ static int bpf_prog_detach(const union bpf_attr *attr)
>  	case BPF_SK_SKB_STREAM_PARSER:
>  	case BPF_SK_SKB_STREAM_VERDICT:
>  		return sockmap_get_from_fd(attr, BPF_PROG_TYPE_SK_SKB, false);
> +	case BPF_LIRC_MODE2:
> +		return lirc_prog_detach(attr);
>  	default:
>  		return -EINVAL;
>  	}
> @@ -1695,6 +1700,8 @@ static int bpf_prog_query(const union bpf_attr *attr,
>  	case BPF_CGROUP_SOCK_OPS:
>  	case BPF_CGROUP_DEVICE:
>  		break;
> +	case BPF_LIRC_MODE2:
> +		return lirc_prog_query(attr, uattr);

When testing this patch series I was wondering why I always got
-EINVAL when trying to query the registered programs.

Closer inspection revealed that bpf_prog_attach/detach/query and
calls to them in the bpf syscall are in "#ifdef CONFIG_CGROUP_BPF"
blocks - and as I built the kernel without CONFIG_CGROUP_BPF
BPF_PROG_ATTACH/DETACH/QUERY weren't handled in the syscall switch
and I got -EINVAL from the bpf syscall function.

I haven't checked in detail yet, but it looks to me like
bpf_prog_attach/detach/query could always be built (or when
either cgroup bpf or lirc bpf are enabled) and the #ifdefs moved
inside the switch(). So lirc bpf could be used without cgroup bpf.
Or am I missing something?

so long,

Hias
>  	default:
>  		return -EINVAL;
>  	}
> -- 
> 2.17.0
>