GeezerVison

info contains mtime rounded to an int, but in the binary header, mtime can be a floatting point number?

Correct. The mtime is rounded in the info after calling create_pax_header, and appears in the overall header twice: unrounded in the pax extended header (formatted as a decimal number in a UTF-8 string) and rounded in the ustar header (octal 1750, seen in the next line).

Does PAX have a limit for the number of digits after the dot for a floating point number?

For example, mtime=1/7 is formatted as 0.14285714285714285. The number 0.142,857,142,857,142,85 has 17 digits after the dot. A resolution of 10^-17 seconds, whereas the most accurate clocks have a resolution of 1 ns (10^-9) on Linux.

$ echo Hello > document.txt
$ tar -H posix -cf archive.tar document.txt
$ hexdump -C archive.tar 
(...)00000200  33 30 20 6d 74 69 6d 65  3d 31 36 34 34 33 32 35  |30 mtime=1644325|
00000210  30 33 33 2e 31 38 32 37  34 34 32 32 38 0a 33 30  |033.182744228.30|
00000220  20 61 74 69 6d 65 3d 31  36 34 34 33 32 35 30 33  | atime=164432503|
00000230  33 2e 31 38 32 37 34 34  32 32 38 0a 33 30 20 63  |3.182744228.30 c|
00000240  74 69 6d 65 3d 31 36 34  34 33 32 35 30 33 33 2e  |time=1644325033.|
00000250  31 38 32 37 34 34 32 32  38 0a 00 00 00 00 00 00  |182744228.......|
00000260  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |................|
(...)

$ python3
>>> import os
>>> st=os.stat("document.txt")
>>> st.st_atime_ns
1644325037796763404
>>> st.st_mtime_ns
1644325033182744228
>>> st.st_ctime_ns
1644325033182744228

I get:

• mtime=1644325033.182744228
• atime=1644325033.182744228
• ctime=1644325033.182744228

9 digits after the dot for mtime, atime and ctime.

Maybe for these 9 header names, round to 9 digits if there are more than 9 digits?

Maybe it's ok to have more digits. I don't know.

The standard doesn't specify a maximum number of digits. It suggests to use enough digits to allow the timestamp to be reconstructed exactly (if extracting on a system with the same timestamp precision as the originating system). Most implementations use nanosecond precision.

Doing additional rounding of the value in the pax extended header would be a behavioural change beyond what I'm setting out to do here. I suspect any potential issues with using more digits than an extractor can handle would be rare anyway, since you would have to deliberately use a made-up number rather than just dividing st_mtime_ns by 10^9.

1000.1 cannot be represented exactly in binary IEEE 754:

>>> (1001.1).hex()
'0x1.f48cccccccccdp+9'

Maybe to avoid any rounding issue in the test, we could use 1000.5 number which can be represented exaclty?

>>> (1000.5).hex()
'0x1.f440000000000p+9'

I can make that change if you want, but there are going to be quite a few real-world timestamps that don't have an exact representation in floating point.

And come to think of it, 1000.5 has the disadvantage of being exactly halfway between two integers, so the result of rounding it changes depending on the prevailing rounding strategy. I think round() is supposed to guarantee round-to-even, but it still adds a little ambiguity for humans.

All in all, switching to fixed-point would probably be a good idea. But that's again rather out of scope for what I'm trying to accomplish here. :)