Comment by scottlamb | Hacker News Reader

Comment by scottlamb | Hacker News Reader

Parent Full thread

Are the authors here? Thanks for this! I'm always thrilled to see advances in profiling tools.

I'm curious what they have to say about complexity/necessity of interpreting all of DWARF. cargo-trace (an neat and conceptually similar but abandoned project, I think) [1] says:

> It can be empirically determined that almost all dwarf programs consist of a single instruction and use only three different instructions. rip+offset, rsp+offset or *cfa+offset, where cfa is the rsp value of the previous frame. The result of the unwinding is an array of instruction pointers.

Do you find this to be true? Is more complex interpreting of DWARF necessary?

And in the lkml thread linked from the article, Linus is extremely pessimistic about DWARF unwinding, [2] I'm sure not without justification. He's talking about kernel stacks, and I think the trade-off is different when you're trying to profile existing userspace applications and libraries compiled and implemented however, but nonetheless I'm curious to hear the authors say how applicable they think his points are.

The authors' pointer to .ctf_frame is also a very interesting development in this space. Not sure how long that will take to be widely usable (assuming it goes through at all).

[1] https://github.com/dvc94ch/cargo-trace

[2] https://lkml.org/lkml/2012/2/10/356

Author here!

This project looks interesting and as you mention it seems conceptually similar! I would agree with that quote, generally speaking, but I think it really depends on a binary to binary basis.

Some binaries outsource most of the work to dynamic libraries. Unfortunately, DWARF expressions are typically emitted for these program counter ranges, so it's desirable to at least implement a subset of expressions [0].

Even if that's not the case, we want to produce profiles that are as accurate as possible :)

You are totally right, they were discussing kernel stacks where stakes are higher as it needs to work perfectly, otherwise kernel live patching would not reliably work, among others. The kernel has now some unwind table format, that can be used in x86_64, called ORC [1].

That being said the parser for DWARF would still have to live in the kernel, and I am not sure if kernel devs would like to accept such a patch.

Ideally we would transition to an unwind-specific format for user-space (something like ours, for example) and perhaps have a suitable unwinder in the kernel, rather than having to implement it in a BPF program. This is something we are considering for the future, but it's not free from problems (increased executable size, redundant unwind information, etc). But this is exactly why we wanted to have a conversation with the communities interested in this work!

[0]: https://github.com/parca-dev/parca-agent/pull/1058/commits

[1]: https://lwn.net/Articles/728339/

> Ideally we would transition to an unwind-specific format for user-space (something like ours, for example) and perhaps have a suitable unwinder in the kernel, rather than having to implement it in a BPF program. This is something we are considering for the future, but it's not free from problems (increased executable size, redundant unwind information, etc). But this is exactly why we wanted to have a conversation with the communities interested in this work!

I'm just an interested user rather than someone who has an ownership stake in any major distribution, language, library, or application, but fwiw:

I'd love for every binary/library to embed enough information to cheaply unwind during profiling [edit: and symbolize shortly afterward], in one simple format, so that this sort of profiling tool Just Works, as well as backtraces generated by the program itself (e.g. Rust's panics). I'd prefer it support inlining well (so ideally not just turning on frame pointers everywhere but that'd still be a huge improvement over the status quo). As you know, this is all a complete mess now, and I hate that. I think a reasonable amount of space overhead for this is fine. I want to be able to profile pre-packaged applications and my own Rust applications that use some C libraries without hassle. And have profiles written in a format that I can ship to another machine without trouble reading the symbols there or concern about excessive sensitive information. All stuff you've touched on. I would consider it a genuine miracle if the parties involved could all get behind this.

Full DWARF information, e.g. Rust's debug = 2 [1] (to be clear: my understanding is this is more or less to allow debuggers to print variable contents reasonably well?) is another matter. Seems like that adds a huge amount of bloat and is more rarely used. I like split debuginfo. As long as there's an easy way to get it on demand via a debuginfod service and/or installing an additional package, it's fine.

[1] https://doc.rust-lang.org/cargo/reference/profiles.html

> I'd love for every binary/library to embed enough information to cheaply unwind during profiling

Definitely! We hope to be able to change this. There's the .ctf_frame format [0] that could be used too. While its design goals are very much aligned to ours, implementing a reader for this format in BPF will be complicated.

> and symbolize shortly afterward

Definitely, I think that C Type Format (CTF) [1] could be a candidate here. Another possibility would be to get BTF (BPF Type Format) [2] working in userspace. While it wouldn't cover everything that DWARF deals with, it might be a reasonable tradeoff for symbolization.

FWIW both in Polar Signals and Parca (our OSS offering) we resolve to async symbolization done in the server, as it's very expensive to do and we want to reduce load where the Agent runs.

[0]: https://gcc.gnu.org/wiki/cauldron2022talks

[1]: https://lwn.net/Articles/795384/#:~:text=The%20Compact%20C%2....

[2]: https://docs.kernel.org/bpf/btf.html

I wish you success!

AFAICT from some searching this morning, CTF is implemented in the GNU toolchain (compiler -> debugger) but not at all in LLVM? so seems like it'd be a while before this really happens even if everyone gets behind that format.

> FWIW both in Polar Signals and Parca (our OSS offering) we resolve to async symbolization done in the server, as it's very expensive to do and we want to reduce load where the Agent runs.

Makes total sense for your cluster-wide always-on profiling. I've worked some with GWP [1] and Google Cloud Profiler, and they similarly defer/offload the symbolization to a dedicated service. But in a small-scale setting for e.g. a Rust std::backtrace::Backtrace with an Error or panic, I generally want to print a nice symbolized version in-process shortly after capture (or never). For small-scale on-demand profiling, similarly I likely want to symbolize soon after, although probably out-of-process.

[1] https://research.google/pubs/pub36575/

No idea about LLVM's support for it, but I know Apple's XNU kernel, built with LLVM, uses it (check for the __CTF,__ctf section). I think it uses ctf_insert(1), so it probably doesn't need the toolchain itself to support it.

Thanks so much!

You are right! Hopefully these formats will become standard and implemented in more toolchains. My worry is that there seem to be several "competing" formats for a more compact DWARF debug (as opposed to unwind) information. I hope this doesn't lead to more fragmentation of the ecosystem.

Definitely! I assume you work at Google, so most of your binaries probably have some crash handlers set up already, right? I find this very useful, and wish it were more commonplace. I really like Folly's implementation of this idea [0] that just requires a single line of code: `folly::symbolizer::installFatalSignalHandler()`

> I assume you work at Google

I did for a really long time. At a startup now.

> so most of your binaries probably have some crash handlers set up already, right?

Yes, there was nice instrumentation there too. All google3 binaries did something like InitGoogle() that set off commandline flag parsing, signal handler installation, log hooks, mlock/huge page setup, etc.

I've written a DWARF expression parser and stack unwinder and its not true that those three use-cases are enough to be robust [0]. Based on my own observations, you need to implement:

- undefined - same value - offset(N) - val_offset(N) - register(N)

I've never come across expression(E), val_expression(E), or architectural. But you definitely do need the above in the common cases.

More generally, I'm kinda confused by that quote, I think the author may be conflating DWARF expressions and stack unwinding a bit?

What I think they might have been getting at is that with DWARF expression evaluation [1] in particular (i.e. not just stack unwinding), I've never even seen +80% of the opcodes I've implemented in the wild across several programming languages and compilers. I also left about 1/4 of them unimplemented and I've never had an issue. The intention of the DWARF spec writers was to provide folks with a stack machine that was Turing-complete, but in practice, yeah, nobody really takes advantage of it because there's not really a need.

[0] https://dwarfstd.org/doc/DWARF5.pdf (section 6.4) [1] https://dwarfstd.org/doc/DWARF5.pdf (section 7.7)

I've found two examples of DWARF expressions in the wild: 1. A handwritten expression in libpthread <= 2.19, which was removed in 2014 [1], and 2. the expression covering the PLT section of most binaries. [2]

[1] https://bugzilla.mozilla.org/show_bug.cgi?id=1157194#c14

[2] https://github.com/mstange/framehop/blob/6777199dd0c2c1b3768...

> More generally, I'm kinda confused by that quote, I think the author may be conflating DWARF expressions and stack unwinding a bit?

I suspect they only cared about / looked at the portion of DWARF information needed for stack unwinding. I'd read the quote with that implicit qualifier after "almost all dwarf programs".

Ah gotcha, in that case, as I said, I've never come across a DWARF expression while doing stack unwinding in the wild. Just registers and offsets :)

I find this surprising! Was this for off the shelf applications or some custom binaries?

We see DWARF expressions such as `DW_CFA_def_cfa_expression` on the regular. See the "Test Plan" section and commit messages of the PR that introduced support for this particular opcode, as well as the simple executable that shows this opcode [0] and its printed unwind table [1]

[0]: https://github.com/parca-dev/parca-agent/pull/1058

[1]: https://github.com/parca-dev/testdata/blob/7f9206570a743206f...

Yep! I've never seen anything besides DW_CFA_def_cfa, DW_CFA_offset, and DW_CFA_nop in my CIE definition instructions before. Not to say you shouldn't support them or that they don't exist - they just haven't come up for me yet on the compilers I've been digging in to. I've been examining programs I wrote myself, compiled on Linux using a variety of tools (gcc, g++, clang, clang++, rust, go, zig, and jai are the compilers I've been supporting). I've been writing a debugger from scratch for the past few months, so the set of programs I've been testing on is quite small compared to you I'm sure.

I've never encountered a DW_CFA_def_cfa_expression before in the wild, which compilers(s) are you seeing that in out of curiosity? I see Parca supports many more languages/compilers than I do. I do support that opcode (and all the other CIE opcodes) in my debugger, I just have yet to find a compiler that generates it.

Side note, just looked through the Parca marketing materials, it looks awesome! Excited to try it out on some of my stuff and see what I can find.

That's very curious! We definitely see way more DWARF opcodes in the wild, including `DW_CFA_remember_state`, `DW_CFA_restore_state`, `DW_CFA_advance_*`, among others.

Statistically speaking, a profiler has more chances of bumping into program counters (PCs) that fall in ranges defined by a larger set of DWARF opcodes. We run at 100Hz, which can quickly yield a large number of different program counters, while I assume that most debugging sessions won't hit that many different PCs.

I'm using Fedora 36 which ships `gcc (GCC) 12.2.1 20220819 (Red Hat 12.2.1-2)`

> I've been writing a debugger from scratch for the past few months

Sounds fun! Are you planning on writing about this? I am sure there would be many people interested in reading about this topic!

Either way feel free to reach out at javier @ my company's domain if you want to discuss the DWARF opcodes thing further!

Good point about profilers' stack unwinders hitting much more code than a traditional debugger does, I'm sure that's mostly to blame. I've only been debugging a few thousand lines of code with my debugger so far, so we can't really call it a serious tool yet! I also am using Fedora 36 with the gcc version that it ships with. :shrug:

I'm not sure if I plan to write about it, it doesn't feel like I'm uncovering anything that anyone can't go look up pretty easily. I've been building it mostly because I'm frustrated with all the existing tools out there, save for RemedyBG (which is awesome, but Windows-only, and I'm only on Windows every so often). However, I'm still quite a ways off from being a reasonable daily driver!

> The intention of the DWARF spec writers was to provide folks with a stack machine that was Turing-complete, but in practice, yeah, nobody really takes advantage of it because there's not really a need.

There definitely is a need for better DWARF information, be it with new opcodes or not. It's just that generating them has not been a priority for compilers.

Reply on news.ycombinator.com