archive/tar: do not use ustar long filename encoding and binary number encoding together

[lyonel]

Go version: go1.3.3 linux/amd64 on Fedora 20

When creating an archive (with tar.NewWriter), long names (> 100 chars) that don't fit into standard tar headers need to be encoded differently.
An optimisation in archive/tar/writer.go: writeHeader() tries to use a ustar header when only the name is too long creates files that are misinterpretated by other tar implementations (but read correctly by archive/tar)

For example, /home/support/.openoffice.org/3/user/uno_packages/cache/registry/com.sun.star.comp.deployment.executable.PackageRegistryBackend becomes com.sun.star.comp.deployment.executable.PackageRegistryBackend for external tar commands (tested with GNU tar, BSD tar and star)

Modifying archive/tar/writer.go and forcing preferPax to true in NewWriter fixes the issue

[bradfitz]

I see there were archive/tar changes between Go 1.3 and Go 1.4. Please confirm with this is a problem with Go 1.4 and we can re-open this bug.

[lyonel]

Actually there has been no code change affecting archive creation since 2013, please re-open.

[minux]

Could you please provide a short program that creates a tar file that
common tar(1) commands couldn't handle?

I tried but failed to produce such an example.
http://play.golang.org/p/atdcetMiDV

[lyonel]

it's actually quite tricky: the name has to be long enough (>100 chars), but not too long (<250 or so) and must contain slashes

here is an example that will untar a 'readme.txt' file instead of /path/to/readme/readme/.../readme.txt:
http://play.golang.org/p/sok5w4xal0

[minux]

I tried your example on both Linux and OS X, but the results are both
correct.

PS: if recreating the bug requires a trick test case, please provide the
example program in the original report, that will save us a lot of time.
Thanks.

[lyonel]

you're correct, I'll need to extract the exact code that exhibits the issue... stay tuned.

[lyonel]

It's indeed tricky: it only happens when archive/tar omits the ustar version (instead of setting it to '00'); this is triggered by "big" user ids (10000000)

cf. http://play.golang.org/p/pTFShmA0bZ

package main

import (
    "archive/tar"
    "log"
    "os"
    "strings"
)

func main() {
    // Create a buffer to write our archive to.
    buf, _ := os.Create("broken.tar")

    // Create a new tar archive.
    tw := tar.NewWriter(buf)

    // Add some files to the archive.
    var files = []struct {
        Name, Body string
    }{
        {"shortname.txt", "short file content"},
        {"path/to/" + strings.Repeat("readme/", 15) + "readme.txt", "This archive contains some text files."},
        {"b/shortname.txt", "short file content"},
    }
    for _, file := range files {
        hdr := &tar.Header{
            Name:       file.Name,
            Uid:        1000000000,
            Uname:      "root",
            Gname:      "root",
            Linkname:   ".",
            Mode:       0660,
            Size:       int64(len(file.Body)),
        }
        if err := tw.WriteHeader(hdr); err != nil {
            log.Fatalln(err)
        }
        if _, err := tw.Write([]byte(file.Body)); err != nil {
        log.Fatalln(err)
        }
    }
    // Make sure to check the error on Close.
    if err := tw.Close(); err != nil {
        log.Fatalln(err)
    }
    buf.Close()
}

[minux]

The problem is that if the uid/gid is bigger than 2^21-1 (that is, can't be
epresented in 8 octal digits), archive/tar will use binary representation
in the tar header.

When using that extension, it uses "ushar\x20\x20" as header magic,
instead of the normal "ushar". However, according to GNU tar source,
the long file name will only be considered if it's TMAGIC ("ushar").

By always using "ushar" fixed the problem, and both GNU tar and BSD
tar read the tar file correctly. I'm not sure why we're using
"ushar\x20\x20"
(OLDGNU_MAGIC) in the first place. It seems all tar readers can handle
binary-encoded uid/gid just fine with TMAGIC.

[minux]

According to libarchive source code,
https://github.com/libarchive/libarchive/blob/master/libarchive/archive_write_set_format_ustar.c
http://nxr.netbsd.org/xref/src-freebsd/contrib/libarchive/libarchive/archive_write_set_format_gnutar.c

Non-strict mode ustar support binary (base-256) uid/gid.
However, gnu mode (actually OLDGNU_MAGIC) doesn't support base-256 uid/gid.

According to GNU tar source code, once uid/gid is bigger than 2^21-1, it
will use pax header.

Thus we have at least two possible fixes:

1. if uid/gid is too large to be written in octal, always use PAX.
2. if uid/gid is too large, and there is nothing that prevents ushar format
   otherwise, use
   "ushar". This will generate a file not strictly compliant to the ushar
   format, but the majority
   of tar implementions (GNU tar and those based on libarchive) will accept it
   without issue.

the current way of using OLDGNU_MAGIC doesn't look right, as GNU tar itself
doesn't
use base-256 encoding when generating OLDGNU_MAGIC files.

/cc @dsymonds

[lyonel]

Thanks @minux for your analysis!
Just 2 remarks:

• any "binary-encoded" value will make the archive unreadable (especially the size, that's actually what happened in my initial bug report)
• I concur with you, PAX headers are probably the best option (forcing preferPax to true did the trick in my case)