I'm sitting here next to Farid Zakaria at Tacosprint where we looked at the stat storm that has been haunting nixpkgs for a decade.
devenv auto activation runs ``` devenv hook-should-activate
on every shell prompt to decide whether you've stepped
into a project directory. It does almost nothing: discover the project, check
the trust database, print a path. So its runtime is pure startup overhead, and
it runs on every single prompt redraw.
``` bash
$ time devenv hook-should-activate
/home/domen/dev/myproject
real 0m0.070s
...
70ms before a prompt, every prompt.
And this isn't devenv's tax to pay, it's nixpkgs'. Every program pays it before it runs a line of its own code: the dynamic has to find each shared library, and the way Nix scatters packages across the store makes that search slow. This is not news. The cost has been measured, written up, and partly fixed more than once, and yet it has sat in limbo for the better part of a decade with no general fix merged into nixpkgs.
Most of that is the dynamic looking for a shared object that is sitting
right there in the store, just not in the first directory it tried. The
knocks on 486 wrong doors before it finds the right ones, and almost all of it
happens before main
even starts.
That number is the whole game. Above ~30ms you have to bolt a caching layer on top of the hook; in single digit milliseconds you just run it on every prompt and throw the cache away.
And it scales with the closure: imagemagick
's magick --version
makes 1225 failing opens:
$ strace -f -e openat magick --version 2>&1 >/dev/null | grep '\.so' | grep -c ENOENT
1225
The community has been circling a real fix for years. This post walks through the problem, the approaches people have tried with their tradeoffs, and a more radical one we spiked for devenv to see if it was even possible: deleting the dynamic altogether by linking the whole program into one static binary.
The umbrella tracking issue for the general problem is NixOS/nixpkgs#481620.
Why Nix makes the work so hard #
On a traditional distribution every shared library lives in a handful of global
directories such as /usr/lib
. The dynamic has a short, mostly cached
search path, and ld.so.cache
(built by ldconfig
) turns soname lookups into a hash table hit.
Nix is different by design. Every package lives in its own
/nix/store/<hash>-name/lib
directory, and there is no global ld.so.cache
for
store libraries. To make a binary find its dependencies, Nix records a
DT_RUNPATH
in the ELF header that lists one directory per dependency. A
program linked against fifty libraries gets a DT_RUNPATH
with dozens of entries.
Now recall how glibc resolves a DT_NEEDED
soname with DT_RUNPATH
present: it
walks every DT_RUNPATH
directory in order, trying to open dir/soname
in
each, until one succeeds. So resolving N libraries against a path of M
directories costs on the order of N times M openat()
attempts, almost all of which fail. That is the stat storm.
It gets worse. For every directory it searches, glibc first probes the
glibc-hwcaps
subdirectories for your CPU (x86-64-v3
, x86-64-v2
, and so on), which adds roughly three more failing opens per directory on a modern machine. On a fast SSD with a warm cache none of this is noticeable. On a slow disk, a network filesystem, a cold cache, or a low power ARM board, it is the difference between snappy and sluggish, and it multiplies across every process a shell script spawns.
Concretely, the two workloads we traced most closely:
| Workload | Loaded libraries | DT_RUNPATH dirs |
Failing .so opens |
|---|---|---|---|
devenv version |
83 | 12 (leaf binary) | ~486 |
imagemagick magick --version |
91 | 35 | ~1225 |
The wider a binary's own DT_RUNPATH
and the deeper its transitive graph, the worse the storm.
What a good fix has to preserve #
The reason this problem has stayed open so long is that the obvious fixes break things people rely on. Any serious solution is judged against a checklist:
NixOS injects the GPU driver by puttingLD_LIBRARY_PATH
override./run/opengl-driver/lib
onLD_LIBRARY_PATH
. If a fix stops that from winning, graphics break.Interposers and shims must still load first.LD_PRELOAD
.The libGL / glvnd runtime swap. A program built against Mesa must be able to pick up the vendor driver at runtime.Two libraries with the same soname. This is the heart of the Nix model: different parts of one closure can legitimately depend on different builds of the same soname, and resolution must stay per object.Plugins loaded at runtime are a related but separate problem.dlopen
.Cross compilation. A fix that has to run the target cannot cross compile cleanly.Disk and closure size. Whatever metadata you add ships in every NAR.Maintenance burden. A glibc or patch has to be rebased onto every new glibc release, and patching glibc rebuilds the world.
No approach so far ticks every box. The interesting part is how each one chooses which boxes to give up.
Approach 1: freeze the resolution with absolute paths #
The simplest idea: rewrite every DT_NEEDED
entry from a bare soname like
libfoo.so.1
to the absolute store path of the library it resolves to. glibc
has a "slash short circuit": a DT_NEEDED
containing a /
is opened directly,
skipping all search. No search means no storm, and not even the glibc-hwcaps
probes happen.
This is well trodden ground:
Farid Zakaria'sshrinkwrap and thenix-harden-needed tool do exactly this as external post processing. Shrinkwrap is described in the paper(Zakaria, Scogland, Gamblin, Maltzahn, 2022; arXiv:2211.05118), which measures the storm directly: an Emacs launch drops from 1823Mapping Out the HPC Dependency Chaosstat
/openat
syscalls to 104, a 36 times speedup, and a 900 library MPI application starting across 2048 processes on NFS goes from 344.6s to 47.8s, 7.2 times faster. Those NFS numbers are the clearest evidence that this overhead, invisible on a warm local cache, becomes brutal on a network or cold filesystem.- patchelf
PR #357(--shrink-wrap
, open since 2021) pulls all transitiveDT_NEEDED
up onto the top binary and rewrites them to absolute paths. - Spack has a similar
bind
feature in the HPC world. - Inside nixpkgs, this mechanism is already used ad hoc in dozens of packages.
The cost is steep on the checklist. Absolute paths lose the LD_LIBRARY_PATH override, so the glvnd driver swap stops working, and you need an exemption list for libc, the itself, the GL stack, and initrd. There is also no runtime fallback: if the pinned path is gone, the program does not start.
There is also a build time fork in the road here. To rewrite a soname to an
absolute path you first have to resolve it, and there are two ways to do that:
**run the binary's own ** and record what glibc actually picks, or walk
DT_RUNPATH
statically and resolve it yourself. The first is exact but executes target code, so it cannot cross compile; the second cross compiles cleanly. The absolute path tooling only ever did the first, which is why it stays a manual, per package tool rather than a default. The static walk is the same technique the ELF note cache (approach 3) later builds on.
So absolute paths are the zero disk, maximum speed option, attractive for self contained leaf applications, but wrong as a default because of the override semantics.
Approach 2: the RUNPATH symlink farm #
If the problem is that the searches many directories, give it one. The
farm idea, floated early on by Linus Heckemann
(see #24844), is: for each ELF,
create
a single directory of symlinks pointing at exactly the libraries that ELF needs,
and set its DT_RUNPATH
to that one directory.
The crucial detail is that the sonames in DT_NEEDED
stay short. The farm only
changes where they are found, not how. Because the farm lives in DT_RUNPATH
,
which the consults after LD_LIBRARY_PATH
, every override keeps working.
And it builds with nothing but stock patchelf --set-rpath
and symlinks, with no glibc or patchelf fork, and never executes the target binary, so it cross compiles.
But keeping the sonames short is also where it breaks the Nix model. A farm
directory is a flat namespace keyed by soname, so it can hold exactly one
libfoo.so.1
. When a closure legitimately pulls two different builds of the same soname (the case Nix exists to allow), the farm cannot represent both, and glibc's soname based dedup collapses them to whichever loads first. Absolute paths (approach 1) sidestep this because the store path becomes the key; the farm, which deliberately keeps the bare soname, cannot.
The remaining costs are store pollution and the hwcaps floor. Every ELF gains
its own extra directory of symlinks, so the store fills up with farm directories
that shadow the real libraries. And the farm collapses the per directory
multiplier but not the per hwcaps multiplier: the still probes
glibc-hwcaps
inside the one farm directory. So it is a large constant factor win, not an asymptotic one.
How large depends entirely on how much of the graph you farm:
| Farmed scope | Failing opens | Reduction |
|---|---|---|
imagemagick , binary only (wide 35 dir DT_RUNPATH ) |
||
| 1225 → ~213 | 83% | |
devenv , leaf binary only (narrow 12 dir DT_RUNPATH ) |
||
| 486 → 392 | 19% | |
devenv , whole graph (every dep built with the hook) |
||
| 486 → 88 | 82% |
The two devenv rows are the lesson. Farming the leaf alone barely moves the
needle because the storm there is dominated by the 83 libraries resolving each
other, which a leaf only farm never touches. Only whole graph adoption reaches
82%, and the residual 88 are irreducible hwcaps probes rather than real library
searches. So the farm pays off immediately when a package's own binary has a
wide DT_RUNPATH
, but needs whole graph adoption for closure heavy applications.
Approach 3: a per DSO resolution cache in an ELF note #
This is the most ambitious approach and, on the checklist, the best. The idea,
designed by pennae in
#207893: have patchelf
write a
small PT_NOTE
into each library that records, for each DT_NEEDED
soname,
where the should find it. A patched glibc reads that note during ,
between the LD_LIBRARY_PATH
step and the DT_RUNPATH
walk, and resolves the dependency straight from it.
Placing the read after LD_LIBRARY_PATH
is what makes it safe: overrides,
LD_PRELOAD
, and the glvnd swap all keep winning, and soname based dedup is
unchanged because the sonames stay short. Each cache entry is either an exact
path, which is opened directly with no search and therefore no hwcaps probing,
or a directory hint for the rare cases that cannot be resolved at build time
($ORIGIN
relative entries, or directories that themselves contain a
glibc-hwcaps
tree).
This is the only approach that preserves every semantic, adds zero closure
references, and eliminates the hwcaps floor as well. pennae's original benchmark
showed an armv7 workload dropping from 44s to 29s (seconds, not ms, measured
under strace -cf
) with about 24000 fewer syscalls. In our own end to end test of a revived, cleaned up version, a note
bearing binary resolved its dependency with zero failing search probes,
versus the full storm for the same binary without the note, while the
LD_LIBRARY_PATH
override still took precedence.
The price is the heaviest of any approach. It needs two source changes: a glibc patch so the understands the note, and a patchelf change to write it. It is a staging mass rebuild, because patching glibc rebuilds the world. pennae's draft was closed for lack of a go or no go decision rather than any technical failure; the main worry raised was the long term maintenance of a glibc patch.
Approach 4: a Guix style per package ld.so.cache #
Guix solves the same problem in production by shipping a per package
ld.so.cache
, the same binary format ldconfig
produces, and having a patched consult it (written up in their Taming the 'stat' storm with a cache;
#207061proposed it for nixpkgs). It preserves
LD_LIBRARY_PATH
and is proven at scale,
but building the cache needs ldconfig
/ldd
for the target architecture, which
breaks cross compilation, and it hits buildEnv
collisions and dlmopen
namespace issues. The ELF note (approach 3) was in part a response: it reads
DT_NEEDED
and DT_RUNPATH
statically and never runs a foreign binary, so it
keeps the same LD_LIBRARY_PATH
guarantee without those costs.## Approach 5: delete the with static linking
The four approaches above all make the 's job easier. Static linking
removes the instead. For devenv, a self contained CLI, we spiked it:
building the whole closure through pkgsStatic
(which means musl, since glibc
doesn't support a complete static link) drops devenv version
and
hook-should-activate
from about 70ms to about 16ms.
| Build | Loaded libraries | startup |
|---|---|---|
| Baseline (all dynamic, glibc) | 83 | ~70ms |
| Fully static (musl) | 0 | ~16ms |
This is not a nixpkgs fix and was never meant to be. Deleting the also deletes everything the does at runtime: plugins on demand, honouring driver and interposer overrides, swapping in the GPU vendor's GL stack. A lot of nixpkgs depends on that, so static linking can never be a general default. It works for devenv only because devenv is a self contained CLI that talks to Nix through its own linked in C API and needs none of it.
One thing surprised us: at 16ms, with the gone, devenv is still far above
the ~2ms a static musl hello world starts in, the rest being execve
mapping the image and devenv's own startup work. Even so, 16ms is fast enough for the shell hook to drop its per directory activation cache and just run the check every prompt.
What about macOS? #
macOS uses a different , dyld
, and the storm isn't there. Nix on Darwin
already ships approach 1: every Mach-O records its dependencies as absolute
store paths in LC_LOAD_DYLIB
rather than bare sonames, and carries no
LC_RPATH
. So dyld
opens each library directly on the first path it tries,
and system frameworks come straight from the in memory dyld shared cache without
touching disk. Where the glibc devenv
made ~486 failing opens, the macOS one makes essentially none.
The startup cost macOS does have is specific to Nix. To decide whether to
advertise x86_64-darwin
as an extra platform, libstore forked a child running
arch -arch x86_64 /usr/bin/true
on startup, costing ~13ms on every Nix process
on Apple silicon. The fix answers the same question with a stat
of Rosetta 2's fixed install path in ~0.01ms (NixOS/nix#16067).
Side by side #
Every column is framed as a property you want, so ✅ is always good and ❌ always a cost. Legend: ✅ yes · ⚠️ with caveats · ❌ no · ➖ not applicable. Caveats marked ⚠️ or worth a word are footnoted below.
| Approach | No glibc fork | No patchelf change | Cheap on disk | Keeps LD_LIBRARY_PATH / glvnd |
Keeps dup sonames | Kills hwcaps floor | Cross safe |
|---|---|---|---|---|---|---|---|
Absolute DT_NEEDED |
✅ | ✅ a |
✅ | ❌ | ⚠️ b |
✅ | ⚠️ c |
| RUNPATH symlink farm | ✅ | ✅ d |
❌ e |
✅ | ❌ | ❌ | ✅ |
| Per DSO ELF note | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Per package ld.so.cache | ❌ | ✅ | ✅ | ✅ | ⚠️ f |
✅ | ❌ |
| Static linking (musl) | ✅ g |
✅ | ❌ h |
❌ | ➖ | ✅ | ✅ i |
a stock patchelf --replace-needed
·
b breaks on the rare duplicate soname ·
c only the static-resolution variant cross compiles, and it is unbuilt ·
d stock patchelf --set-rpath
·
e every ELF gains its own symlink directory in the store ·
f buildEnv
collisions ·
g uses musl, not glibc, so no glibc fork to maintain ·
h ~82MB binary ·
i via pkgsStatic
The final approach #
Over the week at Tacosprint we revived the ELF note cache, cleaned it up,
and got it built and tested end to end. After a decade in limbo it now works:
the note writer, patchelf --build-resolution-cache
(#647), shipped in patchelf 0.19.0, the first patchelf release since 0.18.0 in April 2023.
The last thing to land is
nixpkgs#535735, which turns the note on
across the whole package set. Because it patches glibc it has to go through
staging
, which rebuilds the world, so every binary in nixpkgs comes out the other side resolving its libraries straight from the note. That is also where it gets exercised at scale, and we're committed to fixing whatever shakes out as we go.
Once it has proven itself there, the longer term goal is to upstream the patch into glibc itself, so the fix isn't a nixpkgs carry but something every store based, nix style package manager, guix included, can rely on.