Re: [PATCH] Add initial support for pax extended attributes

[Junio C Hamano]

Jeff King  writes:

> I do think we'd fail to notice the truncation, which isn't ideal. But it
> looks like the rest of the script suffers from the same issue.
>
> If anybody cares, it might not be too hard to wrap all of the 512-byte
> read calls into a helper that dies on bogus input.

Perhaps.  In any case, the patch presented here is an improvement
over the status quo, so let's move the world forward by taking it
without any further "while at it" fixes, which can come later when
people feel inclined to do so.

Thanks, both, for writing and reviewing ;-)

Re: [PATCH] Add initial support for pax extended attributes

[Jeff King]

On Wed, May 23, 2018 at 11:34:52AM +0900, Junio C Hamano wrote:

> > @@ -90,13 +99,32 @@ foreach my $tar_file (@ARGV)
> > Z8 Z1 Z100 Z6
> > Z2 Z32 Z32 Z8 Z8 Z*', $_;
> > }
> > -   next if $name =~ m{/\z};
> > $mode = oct $mode;
> > $size = oct $size;
> > $mtime = oct $mtime;
> > next if $typeflag == 5; # directory
> >  
> > -   if ($typeflag != 1) { # handle hard links later
> > +   if ($typeflag eq 'x') { # extended header
> > +   # If extended header, check for path
> > +   my $pax_header = '';
> > +   while ($size > 0 && read(I, $_, 512) == 512) {
> 
> Would we ever get a short-read (i.e. we ask to read 512 bytes,
> syscall returns after reading only 256 bytes, even though next call
> to read would give the remaining 256 bytes and later ones)?

No, because perl's read() is buffered (you need sysread() to get a real
syscall read). We might read fewer than 512 if we hit EOF, but I think
that would be a truncated input, then, since ustar does everything in
512-byte records.

I do think we'd fail to notice the truncation, which isn't ideal. But it
looks like the rest of the script suffers from the same issue.

If anybody cares, it might not be too hard to wrap all of the 512-byte
read calls into a helper that dies on bogus input. I sort of assumed
this was mostly a proof of concept script and nobody used it, though. :)

It makes me wonder if there is a better-tested tar-reading module in
CPAN that could be used (though at the expense of requiring an extra
dependency).

-Peff

Re: [PATCH] Add initial support for pax extended attributes

[Junio C Hamano]

Pedro Alvarez  writes:

> From: Pedro Alvarez Piedehierro 
> Subject: [PATCH] Add initial support for pax extended attributes

Lead it with the name of the area you are adding support for pax ext
header, e.g.

Subject: [PATCH] import-tars: read overlong names from pax extended header

or something.

> Sometimes the tar files will contain pax extended attributes to deal
> with cases where the information needed doesn't fit in a standard
> ustar entry.
>
> One of these cases is when the path is larger than 100 characters. A
> pax entry will appear containing two standard ustart entries. The first

u-start? us-tart? sound yummy.  I think s/ustart entries/ustar headers/

> entry will have an 'x' typeflag, and contain the the extended attributes.

s/contain//

>
> The pax extended attributes contain one or multiple records constructed as
> follows:
>
> "%d %s=%s\n", , , 

> This commit makes sure that we always read the extended attibutes from

s/This commit makes sure/Make sure/;

> pax entries, and in the case of finding one, we parse its records
> looking for 'path' information. If this information is found, it's
> stored to be used in the next ustar entry.
>
> Information about the Pax Interchange Format can be found at:
>
> 
> https://www.freebsd.org/cgi/man.cgi?manpath=FreeBSD+8-current=tar=5.

> Before this change, importing gcc tarballs[1] would fail with the
> following error:
>
> fast-import crash report:
> fast-import process: 82899
> parent process : 82897
> at 2018-05-21 12:35:27 +
>
> fatal: Unsupported command: 29 atime=1516870168.93527949

Drop "Before this change, " and move the above to the very beginning
of the proposed log message.  The problem description is always
"without this patch applied, we have this problem that needs to be
fixed", so "Before this change" is an unnecessary thing to say.

The remainder of the crash log may or may not be in the problem
description, if we want to shoot for brevity.  If I were writing a
log message for this patch, I'd go for even shorter version, e.g.

Importing gcc tarballs[1] with import-tars script (in
contrib/) fails when hitting a pax extended header that
records a long pathname.

Teach the code to parse and grab information from pax
extended headers, and reconstruct a long pathname that is
split into multiple records, to correct this problem.

The code to parse pax extended headers were written,
consulting the Pax Interchange Format documentation [2].

[1] http://ftp.gnu.org/gnu/gcc/gcc-7.3.0/gcc-7.3.0.tar.xz
[2] 
https://www.freebsd.org/cgi/man.cgi?manpath=FreeBSD+8-current=tar=5

> index d60b4315ed..c2e54ec7a3 100755
> --- a/contrib/fast-import/import-tars.perl
> +++ b/contrib/fast-import/import-tars.perl
> @@ -63,6 +63,8 @@ foreach my $tar_file (@ARGV)
>   my $have_top_dir = 1;
>   my ($top_dir, %files);
>  
> + my $next_path = '';
> +
>   while (read(I, $_, 512) == 512) {
>   my ($name, $mode, $uid, $gid, $size, $mtime,
>   $chksum, $typeflag, $linkname, $magic,
> @@ -70,6 +72,13 @@ foreach my $tar_file (@ARGV)
>   $prefix) = unpack 'Z100 Z8 Z8 Z8 Z12 Z12
>   Z8 Z1 Z100 Z6
>   Z2 Z32 Z32 Z8 Z8 Z*', $_;
> +
> + unless ($next_path eq '') {
> + # Recover name from previous extended header
> + $name = $next_path;
> + $next_path = '';
> + }
> +
>   last unless length($name);
>   if ($name eq '././@LongLink') {
>   # GNU tar extension
> @@ -90,13 +99,32 @@ foreach my $tar_file (@ARGV)
>   Z8 Z1 Z100 Z6
>   Z2 Z32 Z32 Z8 Z8 Z*', $_;
>   }
> - next if $name =~ m{/\z};
>   $mode = oct $mode;
>   $size = oct $size;
>   $mtime = oct $mtime;
>   next if $typeflag == 5; # directory
>  
> - if ($typeflag != 1) { # handle hard links later
> + if ($typeflag eq 'x') { # extended header
> + # If extended header, check for path
> + my $pax_header = '';
> + while ($size > 0 && read(I, $_, 512) == 512) {

Would we ever get a short-read (i.e. we ask to read 512 bytes,
syscall returns after reading only 256 bytes, even though next call
to read would give the remaining 256 bytes and later ones)?

If we do get a short-read, would that be an error (in which case,
how are we handling it)?  If it is not an error, should we continue
reading, instead of leaving the loop?

> + $pax_header = $pax_header . substr($_, 0, 
> $size);
> + $size -= 512;
> + }
> +
> + my @lines =