Date: Tue, 13 Jan 2015 23:14:49 +0000 From: bugzilla-noreply@freebsd.org To: freebsd-dtrace@FreeBSD.org Subject: [Bug 195222] dtrace kernel sdt probes must always be anchored Message-ID: <bug-195222-32976-gDUE3Nju8o@https.bugs.freebsd.org/bugzilla/> In-Reply-To: <bug-195222-32976@https.bugs.freebsd.org/bugzilla/> References: <bug-195222-32976@https.bugs.freebsd.org/bugzilla/>
next in thread | previous in thread | raw e-mail | index | archive | help
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=195222 Mark Johnston <markj@FreeBSD.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |markj@FreeBSD.org Status|New |In Progress --- Comment #1 from Mark Johnston <markj@FreeBSD.org> --- I've done some work towards solving this problem, together with implementing Solaris-style SDT probes. By the latter I'm referring to the technique of overwriting SDT probe sites with NOPs until the probes are enabled. The core of the change is the move from #define SDT_PROBE(...) if (sdt_<probe identifier> != NULL) sdt_<probe identifier>(<args>); to #define SDT_PROBE(...) __dtrace_probe_<probe identifier>(<args>); In the new implementation, we call a function directly instead of through a pointer, and the call is unconditional. The main challenge is in identifying probe sites (represented by struct sdt_instance in my patch). We need to define the __dtrace_probe_* stub functions so that the kernel will link, but we also want to be able to find all of the calls to __dtrace_probe_* so that we know where the probe sites are actually located. Trying to do this after the kernel (or a kld) has been linked seems challenging - the only way to find __dtrace_probe_* calls would be to disassemble the text section and look for calls/jmps to the address of a __dtrace_probe_* stub. What Solaris does instead is look for relocations against the __dtrace_probe_* stubs, and handle them specially: instead of processing the relocation, the target offset is overwritten with NOPs, and the location is recorded for use by the SDT kernel code. I've written a userland program, sdtconvert (better name suggestions are welcome), which does exactly this. It is intended to be run against all object files in a kld immediately prior to linking. It looks for the relocations described above, overwrites the probe sites with NOPs, and records each probe site in a linker set, sdt_instances_set. Each struct sdt_instance in the set contains a pointer to the corresponding sdt_probe (filled in with a relocation emitted by sdtconvert), and the offset of the probe site within the kld. I've also changed sdt.ko to examine the sdt_instances set when it processes a kld. For each instance, we use the kernel linker to look up the name of the function containing the probe site, and create a probe with the correct function name. sdtconvert has already overwritten the probe sites with NOPs, so sdt.ko's responsibility is to insert/remove a breakpoint at the probe site whenever the probe is enabled/disabled. As a result I now get the following: markj@charmander: /usr/home/markj $ sudo dtrace -l -P tcp Password: ID PROVIDER MODULE FUNCTION NAME 36965 tcp kernel tcp_do_segment accept-established 36966 tcp kernel tcp_do_segment connect-established 36967 tcp kernel tcp_do_segment connect-refused 36968 tcp kernel tcp_input receive 36969 tcp kernel tcp_output send 36970 tcp kernel tcp_output connect-request 36971 tcp kernel tcp_state_change state-change 36972 tcp kernel tcp_drop state-change 36973 tcp kernel tcp_respond send 36974 tcp kernel tcp_respond accept-refused In particular, the function names are filled in automatically. Probes that specify their own function name are left alone. sdtconvert.c: https://github.com/markjdb/sdtconvert/blob/master/sdtconvert.c current FreeBSD changes (includes sdtconvert integration): http://people.freebsd.org/~markj/patches/sdt-proper/sdt-proper-1.diff The FreeBSD changes are on branch 20141214-sdt-proper in github.com/markjdb/freebsd. Note that this is still a WIP; I haven't yet written the kernel code to enable/disable probes by overwriting the probe site. But that's pretty straightforward. One thing I haven't figured out is how to handle compatibility. There are a number of SDT probes that specify the module and/or function, even though they're not supposed to. I'd like to rename them to avoid abusing the function and module components, but that would obviously break any script which uses these probes (though they aren't documented and mostly likely aren't widely used). It's preferable to just leave them alone, but I'd also like to have a new KPI for SDT probes that doesn't allow the module or function name to be specified, and convert "correct" SDT providers (io, ip, tcp, udp, proc, sched, ...) to use them. Then the old SDT(9) KPI would be deprecated. Note that the work done so far only supports amd64. Any thoughts? -- You are receiving this mail because: You are the assignee for the bug.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?bug-195222-32976-gDUE3Nju8o>