On Tue, Aug 24, 2021 at 11:17 AM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Tue, Aug 24, 2021 at 11:02 AM Andrii Nakryiko > <andrii.nakryiko@xxxxxxxxx> wrote: > > > > On Tue, Aug 24, 2021 at 10:57 AM Alexei Starovoitov > > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > > > On Mon, Aug 23, 2021 at 9:50 PM Andrii Nakryiko > > > <andrii.nakryiko@xxxxxxxxx> wrote: > > > > > > > > On Fri, Aug 20, 2021 at 7:59 PM Dave Marchevsky <davemarchevsky@xxxxxx> wrote: > > > > > > > > > > This helper is meant to be "bpf_trace_printk, but with proper vararg > > > > > > > > We have bpf_snprintf() and bpf_seq_printf() names for other BPF > > > > helpers using the same approach. How about we call this one simply > > > > `bpf_printf`? It will be in line with other naming, it is logical BPF > > > > equivalent of user-space printf (which outputs to stderr, which in BPF > > > > land is /sys/kernel/debug/tracing/trace_pipe). And it will be logical > > > > to have a nice and short BPF_PRINTF() convenience macro provided by > > > > libbpf. > > > > > > > > > support". Follow bpf_snprintf's example and take a u64 pseudo-vararg > > > > > array. Write to dmesg using the same mechanism as bpf_trace_printk. > > > > > > > > Are you sure about the dmesg part?... bpf_trace_printk is outputting > > > > into /sys/kernel/debug/tracing/trace_pipe. > > > > > > Actually I like bpf_trace_vprintk() name, since it makes it obvious that > > > > It's the inconsistency with bpf_snprintf() and bpf_seq_printf() that's > > mildly annoying (it's f at the end, and no v- prefix). Maybe > > bpf_trace_printf() then? Or is it too close to bpf_trace_printk()? > > bpf_trace_printf could be ok, but see below. > > > But > > either way you would be using BPF_PRINTF() macro for this. And we can > > make that macro use bpf_trace_printk() transparently for <3 args, so > > that new macro works on old kernels. > > Cannot we change the existing bpf_printk() macro to work on old and new kernels? Only if we break backwards compatibility. And I only know how to detect the presence of new helper with CO-RE, which automatically makes any BPF program using this macro CO-RE-dependent, which might not be what users want (vmlinux BTF is still not universally available). If I could do something like that without breaking change and without CO-RE, I'd update bpf_printk() to use `const char *fmt` for format string a long time ago. But adding CO-RE dependency for bpf_printk() seems like a no-go. > So bpf_printk() would use bpf_trace_printf() on new and > bpf_trace_printk() on old? > I think bpf_trace_vprintk() looks cleaner in this context if we reuse > bpf_printk() macro.
