On Thu, Jan 27, 2022 at 12:59:52AM +0900, Masami Hiramatsu wrote: > On Wed, 26 Jan 2022 11:50:22 +0900 > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote: > > > > one more question.. > > > > > > I'm adding support for user to pass function symbols to bpf fprobe link > > > and I thought I'd pass symbols array to register_fprobe, but I'd need to > > > copy the whole array of strings from user space first, which could take > > > lot of memory considering attachment of 10k+ functions > > > > > > so I'm thinking better way is to resolve symbols already in bpf fprobe > > > link code and pass just addresses to register_fprobe > > > > That is OK. Fprobe accepts either ::syms or ::addrs. > > > > > > > > I assume you want to keep symbol interface, right? could we have some > > > flag ensuring the conversion code is skipped, so we don't go through > > > it twice? > > > > Yeah, we still have many unused bits in fprobe::flags. :) > > Instead of that, according to Steve's comment, I would like to introduce > 3 registration APIs. > > int register_fprobe(struct fprobe *fp, const char *filter, const char *notrace); > int register_fprobe_ips(struct fprobe *fp, unsigned long *addrs, int num); > int register_fprobe_syms(struct fprobe *fp, const char **syms, int num); > > The register_fprobe_ips() will not touch the @addrs. You have to set the > correct ftrace location address in the @addrs. ok, sounds good thanks, jirka
