parca-dev
Rust symbols are inconveniently shortened
Here's how parca resolves a symbol
$ git diff
diff --git a/pkg/symbol/elfutils/debuginfofile.go b/pkg/symbol/elfutils/debuginfofile.go
index 4a9ab2cf2..1c5a81aa9 100644
--- a/pkg/symbol/elfutils/debuginfofile.go
+++ b/pkg/symbol/elfutils/debuginfofile.go
@@ -102,6 +102,7 @@ func (f *debugInfoFile) SourceLines(addr uint64) ([]profile.LocationLine, error)
}
file, line := findLineInfo(f.lineEntries[cu.Offset], tr.Ranges)
+ fmt.Printf("addr = 0x%x, name = %s, file = %s, line = %d\n", addr, name, file, line)
lines = append(lines, profile.LocationLine{
Line: line,
Function: f.demangler.Demangle(&pb.Function{
addr = 0x17e976, name = small_slot_len, file = ?, line = 0
Compare this to addr2line from elfutils:
$ eu-addr2line -e ./target/debug/weird-unwind -S 0x17e976
_ZN14regex_automata4util8captures14GroupInfoInner14small_slot_len17h3554d4e8237ec643E+0x16
/home/ivan/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/regex-automata-0.4.9/src/util/captures.rs:2323:9
You couldn't even tell it was something about regexps, but it is:
$ echo _ZN14regex_automata4util8captures14GroupInfoInner14small_slot_len17h3554d4e8237ec643E+0x16 | c++filt
regex_automata::util::captures::GroupInfoInner::small_slot_len::h3554d4e8237ec643+0x16
Compare the flamegraph view in Parca (new and build are very opaque!):
To what perf manages to do:
Interesting to note that symbols come demangled before demangler runs, for example:
addr = 0x3659e6, name = min<core::iter::adapters::map::Map<core::slice::iter::Iter<regex_syntax::hir::literal::Literal>, regex_syntax::hir::literal::{impl#4}::min_literal_len::{closure_env#0}>>, file = /rustc/48a426eca9df23b24b3559e545cf88dee61d4de9/library/core/src/iter/traits/iterator.rs, line = 3143
Looking through what the entry dwarf provides:
(dlv) p tr.Entry
github.com/go-delve/delve/pkg/dwarf/godwarf.Entry(github.com/go-delve/delve/pkg/dwarf/godwarf.compositeEntry) [
*{
Offset: 1051394,
Tag: TagSubprogram (46),
Children: true,
Field: []debug/dwarf.Field len: 4, cap: 4, [
(*"debug/dwarf.Field")(0x140014ce680),
(*"debug/dwarf.Field")(0x140014ce6a0),
(*"debug/dwarf.Field")(0x140014ce6c0),
(*"debug/dwarf.Field")(0x140014ce6e0),
],},
*{
Offset: 980869,
Tag: TagSubprogram (46),
Children: true,
Field: []debug/dwarf.Field len: 7, cap: 7, [
(*"debug/dwarf.Field")(0x140014b5340),
(*"debug/dwarf.Field")(0x140014b5360),
(*"debug/dwarf.Field")(0x140014b5380),
(*"debug/dwarf.Field")(0x140014b53a0),
(*"debug/dwarf.Field")(0x140014b53c0),
(*"debug/dwarf.Field")(0x140014b53e0),
(*"debug/dwarf.Field")(0x140014b5400),
],},
]
The first batch:
(dlv) print *(*"debug/dwarf.Field")0x140014ce680
Command failed: 1:24: expected 'EOF', found 0x140014ce680
(dlv) print *(*"debug/dwarf.Field")(0x140014ce680)
debug/dwarf.Field {
Attr: AttrLowpc (17),
Val: interface {}(uint64) 1567072,
Class: ClassAddress (1),}
(dlv) print *(*"debug/dwarf.Field")(0x140014ce6a0)
debug/dwarf.Field {
Attr: AttrHighpc (18),
Val: interface {}(int64) 36,
Class: ClassConstant (3),}
(dlv) print *(*"debug/dwarf.Field")(0x140014ce6c0)
debug/dwarf.Field {
Attr: AttrFrameBase (64),
Val: interface {}([]uint8) [87],
Class: ClassExprLoc (4),}
(dlv) print *(*"debug/dwarf.Field")(0x140014ce6e0)
debug/dwarf.Field {
Attr: AttrSpecification (71),
Val: interface {}(debug/dwarf.Offset) 980869,
Class: ClassReference (10),}
The second batch:
(dlv) print *(*"debug/dwarf.Field")(0x140014b5340)
debug/dwarf.Field {
Attr: AttrLinkageName (110),
Val: interface {}(string) "_ZN14regex_automata4util8captures14GroupInfoInner14small_slot_le...+21 more",
Class: ClassString (12),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b5360)
debug/dwarf.Field {
Attr: AttrName (3),
Val: interface {}(string) "small_slot_len",
Class: ClassString (12),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b5380)
debug/dwarf.Field {
Attr: AttrDeclFile (58),
Val: interface {}(int64) 47,
Class: ClassConstant (3),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b53a0)
debug/dwarf.Field {
Attr: AttrDeclLine (59),
Val: interface {}(int64) 2316,
Class: ClassConstant (3),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b53c0)
debug/dwarf.Field {
Attr: AttrType (73),
Val: interface {}(debug/dwarf.Offset) 979443,
Class: ClassReference (10),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b53e0)
debug/dwarf.Field {
Attr: AttrDeclaration (60),
Val: interface {}(bool) true,
Class: ClassFlag (5),}
(dlv) print *(*"debug/dwarf.Field")(0x140014b5400)
debug/dwarf.Field {
Attr: AttrExternal (63),
Val: interface {}(bool) true,
Class: ClassFlag (5),}
It seems that AttrLinkageName is what we want here rather than AttrName.
Looking at elfutls code, that's exactly what's being used:
- https://sourceware.org/git/?p=elfutils.git;a=blob;f=src/addr2line.c;h=d87e5b45e0a97bc1bb1e7b52981b9424fc0997dc;hb=HEAD#l310
302 static const char *
303 get_diename (Dwarf_Die *die)
304 {
305 Dwarf_Attribute attr;
306 const char *name;
307
308 name = dwarf_formstring (dwarf_attr_integrate (die, DW_AT_MIPS_linkage_name,
309 &attr)
310 ?: dwarf_attr_integrate (die, DW_AT_linkage_name,
311 &attr));
312
313 if (name == NULL)
314 name = dwarf_diename (die) ?: "??";
315
316 return name;
317 }
Swapping this out in Parca itself gives me the following flamegraph:
The problem now is that symbols aren't demangled properly. Taking a random one:
addr = 0x117da0, name = _ZN3std2rt10lang_start28_$u7b$$u7b$closure$u7d$$u7d$17h16960a2bf2e077d9E, file = /rustc/48a426eca9df23b24b3559e545cf88dee61d4de9/library/std/src/rt.rs, line = 195
It matches eu-addr2line:
$ eu-addr2line -e ./target/debug/weird-unwind -S 0x117da0
_ZN3std2rt10lang_start28_$u7b$$u7b$closure$u7d$$u7d$17h16960a2bf2e077d9E+0x10
/rustc/48a426eca9df23b24b3559e545cf88dee61d4de9/library/std/src/rt.rs:195:18
And c++filt can demangle it with no issues:
$ echo '_ZN3std2rt10lang_start28_$u7b$$u7b$closure$u7d$$u7d$17h16960a2bf2e077d9E' | c++filt
std::rt::lang_start::{{closure}}::h16960a2bf2e077d9
It's probably worth a separate issue to figure this part out.
I'm happy to open a PR to start using AttrLinkageName if available and falling back to AttrName. Let me know what you think.
It's probably worth a separate issue to figure this part out.
It's because we feed Name instead of SystemName into demangler and it think that there's nothing to do:
- https://github.com/parca-dev/parca/blob/v0.22.0/pkg/symbol/elfutils/debuginfofile.go#L108
- https://github.com/parca-dev/parca/blob/v0.22.0/pkg/symbol/demangle/demangle.go#L79-L81
With that fixed, I finally have my expected flamegraph:
I'm also happy to open a PR to fix that bit as well if it makes sense.
