Re: Re : breakage in emacsclient

[Samuel Thibault]

Andrew Eggenberger, le jeu. 29 oct. 2020 16:12:47 -0500, a ecrit:
>     I think the problem is with the pflocal/io.c implementation of S_io_stat.
[...]
>     calling fstat on the 
>     file descriptor of a socket created with cloexec_socket (AF_UNIX,
>     SOCK_STREAM, 0);. But S_io_stat doesn't set uid.

Indeed.  The more complete story is that

- on the application side, socket() calls _hurd_socket_server, which
  opens /servers/socket/1
- on the pflocal side, that translates to calling trivfs_S_dir_lookup,
  which does receive the user credentials in its cred parameter, and
  calls trivfs_open that creates a struct trivfs_peropen and struct
  trivfs_protid in which the user field points to the credentials.

- then on the application side, socket() calls __socket_create
- on the pflocal side, that translates to calling S_socket_create.
  There, the pf port is for now untranslated (showing up as a
  mach_port_t), but you can add to pflocal/mig-mutate.h the same
  PF_INTRAN, PF_INTRAN_PAYLOAD, and PF_DESTRUCTOR as in e.g.
  pfinet/mig-mutate.h, so that S_socket_create (and others in the same
  file) get the translated port (showing up as a struct trivfs_protid
  *master). Then you have access to the credential, and can record
  uid/gid in a new field of struct sock.

> The problem is that /hurd/pflocal is owned by and in
> the group root,

Well, that's unrelated. io_stat memsets the struct stat to 0 anyway.

Samuel