fs: non-bigint `stat` can give 0.20566138856% bigger results

[LiviaMedeiros]

Version

v24.2.0, v25.0.0-pre

Platform

Linux tumba 6.14.8-gentoo-yuran #2 SMP Tue May 27 17:51:23 +08 2025 x86_64 Intel(R) Core(TM)2 Quad CPU Q8200 @ 2.33GHz GenuineIntel GNU/Linux

Subsystem

fs

What steps will reproduce the bug?

import fs from 'node:fs/promises';
const f = '/tmp/deleteme-utimes-issue-file';
await fs.writeFile(f, '');

console.log('|', 'set', '|', 'got', '|', 'got/set', '|');
console.log('|', '---', '|', '---', '|', '---', '|');

for (let i = 1; i < Number.MAX_SAFE_INTEGER; i *= 3) {
await fs.utimes(f, i, i);
const { mtimeMs } = await fs.stat(f);
console.log('|', i, '|', mtimeMs, '|', mtimeMs / i, '|');
}

await fs.rm(f);

How often does it reproduce? Is there a required condition?

Reproduces on one specific PC all the time. Couldn't reproduce on any other machine so far.

What is the expected behavior? Why is that the expected behavior?

set	got	got/set
1	1000	1000
3	3000	1000
9	9000	1000
27	27000	1000
81	81000	1000
243	243000	1000
729	729000	1000
2187	2187000	1000
6561	6561000	1000
19683	19683000	1000
59049	59049000	1000
177147	177147000	1000
531441	531441000	1000
1594323	1594323000	1000
4782969	4782969000	1000
14348907	14348907000	1000
43046721	43046721000	1000
129140163	129140163000	1000
387420489	387420489000	1000
1162261467	1162261467000	1000
3486784401	3486784401000	1000
10460353203	10460353203000	1000
31381059609	15032385535000	479.0273407685939
94143178827	15032385535000	159.67578025619795
282429536481	15032385535000	53.22526008539932
847288609443	15032385535000	17.741753361799773
2541865828329	15032385535000	5.913917787266591
7625597484987	15032385535000	1.9713059290888637
22876792454961	15032385535000	0.6571019763629545
68630377364883	15032385535000	0.21903399212098484
205891132094649	15032385535000	0.07301133070699495
617673396283947	15032385535000	0.024337110235664983
1853020188851841	15032385535000	0.008112370078554995
5559060566555523	15032385535000	0.0027041233595183315

Expected because 1 second is 1000 milliseconds

What do you see instead?

set	got	got/set
1	1002.0566138856454	1002.0566138856454
3	3006.169841656936	1002.0566138856453
9	9018.509524970808	1002.0566138856453
27	27055.528574912427	1002.0566138856454
81	81166.58572473728	1002.0566138856454
243	243499.75717421185	1002.0566138856454
729	730499.2715226355	1002.0566138856453
2187	2191497.814567907	1002.0566138856456
6561	6574493.443703719	1002.0566138856454
19683	19723480.33111116	1002.0566138856454
59049	59170440.99333348	1002.0566138856454
177147	177511322.98000044	1002.0566138856454
531441	532533968.9400013	1002.0566138856454
1594323	1597601906.820004	1002.0566138856454
4782969	4792805720.4600115	1002.0566138856454
14348907	14378417161.380035	1002.0566138856454
43046721	43135251484.14011	1002.0566138856454
129140163	129405754452.42032	1002.0566138856454
387420489	388217263357.2609	1002.0566138856453
1162261467	1164651790071.783	1002.0566138856456
3486784401	3493955370215.3486	1002.0566138856454
10460353203	10481866110646.045	1002.0566138856453
31381059609	15063301347825.656	480.0125150492224
94143178827	15063301347825.656	160.0041716830741
282429536481	15063301347825.656	53.33472389435804
847288609443	15063301347825.656	17.778241298119347
2541865828329	15063301347825.656	5.926080432706449
7625597484987	15063301347825.656	1.975360144235483
22876792454961	15063301347825.656	0.6584533814118276
68630377364883	15063301347825.656	0.2194844604706092
205891132094649	15063301347825.656	0.0731614868235364
617673396283947	15063301347825.656	0.024387162274512136
1853020188851841	15063301347825.656	0.008129054091504046
5559060566555523	15063301347825.656	0.002709684697168015

Additional information

This issue causes test/parallel/test-fs-cp.mjs, test/parallel/test-fs-promises.js, test/parallel/test-fs-stat-bigint.js, and test/parallel/test-fs-utimes-y2K38.js fail for me locally every time. 😧

This issue persists for statSync(), too.
This issue persists between different filesystems.

This issue does not appear when using bigint option.
This issue does not appear when using touch/stat, or similar functions in other languages.

I suspect some weird bug with floating point.

[LiviaMedeiros]

After some investigation, it's not a Node.js/libuv/v8 bug but is still fixable on Node.js side.
The culprit was kMsPerSec coefficient and its accomplice was kNsPerMs.

The former had value of 10 ** 3 === 1002.05661388564544723 (powf64(0x4024000000000000, 0x4008000000000000) == 0x408f5073f1fb1ac6), the latter had value of 10 ** 6 === 998697.28497027803678066 (powf64(0x4024000000000000, 0x4018000000000000) == 0x412e7a5291e79fd1).

These results are not okay and normally nobody should get them, but we don't have to compute these values on every launch anyway, hence opened #58728 as fix.

The root cause was in glibc, somehow computing like this when using XMM registers and SSE instructions:

$ cat pew.c
#include <stdio.h>
#include <math.h>
int main(void) {
 double a = 10.0;
 double b = 1.0;
 double c = pow(a, b);
 printf("pow(%.1f, %.1f) == %.16f\n", a, b, c);
 return 0;
}

$ gcc -lm -o pew pew.c -O0 && ./pew
pow(10.0, 1.0) == 9.9887951003674988

...essentially corrupting any exponent that is not 2 and any base that is not power of 2.

Attempt to compile glibc with -mfpmath=387 failed:

In file included from ../sysdeps/x86/fpu/fenv_private.h:495,
 from ../sysdeps/ieee754/dbl-64/e_remainder.c:37:
../sysdeps/ieee754/dbl-64/e_remainder.c: In function ‘__ieee754_remainder’:
../sysdeps/generic/fenv_private.h:291:10: error: cleanup argument not a function
 291 | struct rm_ctx ctx __attribute__((cleanup (CLEANUPFUNC ## _ctx))); \
 | ^~~~~~
../sysdeps/generic/fenv_private.h:313:3: note: in expansion of macro ‘SET_RESTORE_ROUND_GENERIC’
 313 | SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_noex, \
 | ^~~~~~~~~~~~~~~~~~~~~~~~~
../sysdeps/ieee754/dbl-64/e_remainder.c:58:7: note: in expansion of macro ‘SET_RESTORE_ROUND_NOEX’
 58 | SET_RESTORE_ROUND_NOEX (FE_TONEAREST);
 | ^~~~~~~~~~~~~~~~~~~~~~

So can't confirm yet, but I'm positive that normally pow and by extension Math.pow and ** produce better results. 😅 In musl-based environments (SSE-capable Athlon XP and ARMv6l RPi) it works fine, with same versions of gcc and node.