ERROR::SDT(7stap)

💖 error::sdt — <sys/sdt.h> marker failures 💖

okay bestie, buckle up, this is the 7stap page for error::sdt, and it's all about when your <sys/sdt.h> markers decide to be dramatic. 🎀

✨ NAME ✨

error::sdt — <sys/sdt.h> marker failures

🌸 DESCRIPTION 🌸

So here's the tea. Systemtap's <sys/sdt.h> probes are modeled after the dtrace USDT API, but they're implemented completely differently — like, same vibe, different girl. 💅

What they actually leave behind in your userspace program's text segment is just a NOP instruction. That's it. One lil' nothing-instruction. 🫶 And then they add an ELF note to the binary with metadata. That metadata is the important part — it describes the marker's name and its parameters.

The whole encoding is designed to be parseable by multiple tools — and not just systemtap, bestie. GDB, the GNU Debugger, also contains support. 👑 These tools can find parameters and their types, wherever they happen to reside, even without DWARF debuginfo. iconic behavior honestly.

Now, why is finding parameters so tricky? Because the STAP_PROBE / DTRACE_PROBE markers store an assembly language expression for each operand — that's just what happens when you use gcc inline-assembly directives. The compiler gets a broad gcc operand constraint string ("nor") for the operands, and usually that works well. Usually it does not force the compiler to load the parameters into or out of registers, which would slow down an instrumented program. Nobody wants that. 🐌

BUT. Some instrumentation sites with some parameters do not work well with the default "nor" constraint. And that's where the three little gremlins come in:

💀 unresolveable at run-time

GCC may emit strings that an assembler could resolve (from the context of compiling the original program), but a run-time tool cannot. For example, the operand string might refer to a label of a local symbol that is not emitted into the ELF object file at all — which leaves no trace for the run-time. 😭 Reference to such parameters from within systemtap can result in "SDT asm not understood" errors.

💀 too complicated expression

GCC might synthesize very complicated assembly addressing modes from complex C data types / pointer expressions. systemtap or gdb may not be able to parse some valid but complicated expressions. Reference to such parameters from within systemtap can result in "SDT asm not understood" errors. (yes, same error, the universe is not creative today 🙈)

💀 overly restrictive constraint

GCC might not be able to even compile the original program with the default "nor" constraint, due to shortage of registers or other reasons. A compile-time gcc error such as "asm operand has impossible constraints" may result. not slay. 💀

🧚 There are two general workarounds to this family of problems 🧚

✨ change the constraints

While compiling the original instrumented program, set the STAP_SDT_ARG_CONSTRAINT macro to different constraint strings. See the GCC manual about various options. For example, on many machine architectures, "r" forces operands into registers, and "g" leaves operands essentially unconstrained. pick your fighter 💅

✨ revert to debuginfo

As long as the instrumented program compiles, it may be fine simply to keep using <sys/sdt.h> but eschew extraction of a few individual parameters. In the worst case, disable <sys/sdt.h> macros entirely to eschew the compiled-in instrumentation. And if DWARF debuginfo was generated and preserved, a systemtap script could refer to the underlying source context variables instead of the positional STAP_PROBE parameters. 🦋

👀 SEE ALSO 👀

• stap(1)
• stapprobes(3stap)
• error::dwarf(7stap)
• http://gcc.gnu.org/onlinedocs/gcc/Constraints.html
• http://sourceware.org/systemtap/wiki/UserSpaceProbeImplementation
• error::reporting(7stap)

and that's the whole page, bestie 🫶 this was ERROR::SDT(7stap) — the 7stap manual, brought to you by the systemtap girlies. go forth and resolve your operands. ✨💖🎀