Re: [PATCH bpf-next 2/3] bpf: Add bpf_perf_event_read_sample() helper

[Yonghong Song]

On 11/1/22 3:02 AM, Jiri Olsa wrote:
> On Mon, Oct 31, 2022 at 10:23:39PM -0700, Namhyung Kim wrote:
> > The bpf_perf_event_read_sample() helper is to get the specified sample
> > data (by using PERF_SAMPLE_* flag in the argument) from BPF to make a
> > decision for filtering on samples.  Currently PERF_SAMPLE_IP and
> > PERF_SAMPLE_DATA flags are supported only.
> >
> > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > ---
> >   include/uapi/linux/bpf.h       | 23 ++++++++++++++++
> >   kernel/trace/bpf_trace.c       | 49 ++++++++++++++++++++++++++++++++++
> >   tools/include/uapi/linux/bpf.h | 23 ++++++++++++++++
> >   3 files changed, 95 insertions(+)
> >
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index 94659f6b3395..cba501de9373 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -5481,6 +5481,28 @@ union bpf_attr {
> >    *		0 on success.
> >    *
> >    *		**-ENOENT** if the bpf_local_storage cannot be found.
> > + *
> > + * long bpf_perf_event_read_sample(struct bpf_perf_event_data *ctx, void *buf, u32 size, u64 sample_flags)
> > + *	Description
> > + *		For an eBPF program attached to a perf event, retrieve the
> > + *		sample data associated to *ctx*	and store it in the buffer
> > + *		pointed by *buf* up to size *size* bytes.
> > + *
> > + *		The *sample_flags* should contain a single value in the
> > + *		**enum perf_event_sample_format**.
> > + *	Return
> > + *		On success, number of bytes written to *buf*. On error, a
> > + *		negative value.
> > + *
> > + *		The *buf* can be set to **NULL** to return the number of bytes
> > + *		required to store the requested sample data.
> > + *
> > + *		**-EINVAL** if *sample_flags* is not a PERF_SAMPLE_* flag.
> > + *
> > + *		**-ENOENT** if the associated perf event doesn't have the data.
> > + *
> > + *		**-ENOSYS** if system doesn't support the sample data to be
> > + *		retrieved.
> >    */
> >   #define ___BPF_FUNC_MAPPER(FN, ctx...)			\
> >   	FN(unspec, 0, ##ctx)				\
> > @@ -5695,6 +5717,7 @@ union bpf_attr {
> >   	FN(user_ringbuf_drain, 209, ##ctx)		\
> >   	FN(cgrp_storage_get, 210, ##ctx)		\
> >   	FN(cgrp_storage_delete, 211, ##ctx)		\
> > +	FN(perf_event_read_sample, 212, ##ctx)		\
> >   	/* */
> >   
> >   /* backwards-compatibility macros for users of __BPF_FUNC_MAPPER that don't
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index ce0228c72a93..befd937afa3c 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -28,6 +28,7 @@
> >   
> >   #include <uapi/linux/bpf.h>
> >   #include <uapi/linux/btf.h>
> > +#include <uapi/linux/perf_event.h>
> >   
> >   #include <asm/tlb.h>
> >   
> > @@ -1743,6 +1744,52 @@ static const struct bpf_func_proto bpf_read_branch_records_proto = {
> >   	.arg4_type      = ARG_ANYTHING,
> >   };
> >   
> > +BPF_CALL_4(bpf_perf_event_read_sample, struct bpf_perf_event_data_kern *, ctx,
> > +	   void *, buf, u32, size, u64, flags)
> > +{
>
> I wonder we could add perf_btf (like we have tp_btf) program type that
> could access ctx->data directly without helpers

Martin and I have discussed an idea to introduce a generic helper like
    bpf_get_kern_ctx(void *ctx)
Given a context, the helper will return a PTR_TO_BTF_ID representing the
corresponding kernel ctx. So in the above example, user could call

    struct bpf_perf_event_data_kern *kctx = bpf_get_kern_ctx(ctx);
    ...

To implement bpf_get_kern_ctx helper, the verifier can find the type
of the context and provide a hidden btf_id as the second parameter of
the actual kernel helper function like
    bpf_get_kern_ctx(ctx) {
       return ctx;
    }
    /* based on ctx_btf_id, find kctx_btf_id and return it to verifier */

    The bpf_get_kern_ctx helper can be inlined as well.

> > +	struct perf_sample_data *sd = ctx->data;
> > +	void *data;
> > +	u32 to_copy = sizeof(u64);
> > +
> > +	/* only allow a single sample flag */
> > +	if (!is_power_of_2(flags))
> > +		return -EINVAL;
> > +
> > +	/* support reading only already populated info */
> > +	if (flags & ~sd->sample_flags)
> > +		return -ENOENT;
> > +
> > +	switch (flags) {
> > +	case PERF_SAMPLE_IP:
> > +		data = &sd->ip;
> > +		break;
> > +	case PERF_SAMPLE_ADDR:
> > +		data = &sd->addr;
> > +		break;
>
> AFAICS from pe_prog_convert_ctx_access you should be able to read addr
> directly from context right? same as sample_period.. so I think if this
> will be generic way to read sample data, should we add sample_period
> as well?
>
>
> > +	default:
> > +		return -ENOSYS;
> > +	}
> > +
> > +	if (!buf)
> > +		return to_copy;
> > +
> > +	if (size < to_copy)
> > +		to_copy = size;
>
> should we fail in here instead? is there any point in returning
> not complete data?
>
> jirka
>
>
> > +
> > +	memcpy(buf, data, to_copy);
> > +	return to_copy;
> > +}
> > +
> > +static const struct bpf_func_proto bpf_perf_event_read_sample_proto = {
> > +	.func           = bpf_perf_event_read_sample,
> > +	.gpl_only       = true,
> > +	.ret_type       = RET_INTEGER,
> > +	.arg1_type      = ARG_PTR_TO_CTX,
> > +	.arg2_type      = ARG_PTR_TO_MEM_OR_NULL,
> > +	.arg3_type      = ARG_CONST_SIZE_OR_ZERO,
> > +	.arg4_type      = ARG_ANYTHING,
> > +};
> > +
> [...]