grub-probe && statfs(2)

grub-probe && statfs(2)

[Francis Gendreau]

[-- Attachment #1: Type: text/plain, Size: 3236 bytes --]

Hello everyone,

I'd like to know if someone here conceived using the statfs() system
call in find_root_device() (or grub_guess_root_device()) instead of
doing a linear look-up into the /dev directory.

My concern here is, that on FreeBSD system using devfs or not, the
approach employed by find_root_device() of getting the dev_t from struct
stat st_dev is not functional. The reason is simple, the function assume
that the st_dev value would be identical for an inode of a particular
device and for the device entry of /dev.

I wrote a short program (in attachment), that provided me the proof of
what I am reporting below.

FreeBSD-4.11 > ./test $HOME /dev/ad0s1f /tmp /dev/ad0s1h / /dev/ad0s1a
$HOME       : dev_t=160773, mntfrom="/dev/ad0s1f", mnton="/home"
/dev/ad0s1f : dev_t=160768, mntfrom="/dev/ad0s1a", mnton="/"
/tmp        : dev_t=160775, mntfrom="/dev/ad0s1h", mnton="/tmp"
/dev/ad0s1h : dev_t=160768, mntfrom="/dev/ad0s1a", mnton="/"
/           : dev_t=160768, mntfrom="/dev/ad0s1a", mnton="/"
/dev/ad0s1a : dev_t=160768, mntfrom="/dev/ad0s1a", mnton="/"

Freebsd-6.2_devfs> ./test . /dev/ad4s1h /tmp /dev/ad4s1e / /dev/ad4s1a /dev /mnt
.           : dev_t=117,      mntfrom="/dev/ad4s1h", mnton="/home"
/dev/ad4s1h : dev_t=16842496, mntfrom="devfs",       mnton="/dev"
/tmp        : dev_t=114,      mntfrom="/dev/ad4s1e", mnton="/tmp"
/dev/ad4s1e : dev_t=16842496, mntfrom="devfs",       mnton="/dev"
/           : dev_t=110,      mntfrom="/dev/ad4s1a", mnton="/"
/dev/ad4s1a : dev_t=16842496, mntfrom="devfs",       mnton="/dev"
/dev        : dev_t=16842496, mntfrom="devfs",       mnton="/dev"
/mnt        : dev_t=138,      mntfrom="/dev/ad0s1a", mnton="/mnt"

OpenBSD-3.2 > ./test . /dev/wd0a /mnt/cdrom/I386 /dev/cd0a
.               : dev_t=0,   mntfrom="/dev/wd0a", mnton="/"
/dev/wd0a       : dev_t=0,   mntfrom="/dev/wd0a", mnton="/"
/mnt/cdrom/I386 : dev_t=768, mntfrom="/dev/cd0a", mnton="/mnt/cdrom"
/dev/cd0a       : dev_t=0,   mntfrom="/dev/wd0a", mnton="/"

Speculating, it seems there is a difference between 44BSD (?) and Linux
in the assigned value of st_dev. While Linux seems to set the same value
for the inodes on the device and the device entry in /dev, BSDs tends to
treat entries in /dev as inodes of the device there are stored on.

My recommendation would be relatively simple. grub_guess_root_device()
could simply call statfs() and extract the device name from struct
statfs f_mntfromname, maybe as simply as the following, but I may lack
of knowledge about some technologies on which this might not work (like
RAID ...).

size_t strnlen (const char *s, size_t max)
{
  size_t l = 0;
  if ( s )
    for (; l < max && *s ; ++l );
  return l;
}

char *
grub_guess_root_device (const char *path)
{
  char *ret = NULL;
  struct statfs sfsb;

  if ( statfs (path, &sfsb) )
    {
      ret = xmalloc (strnlen (sfsb.f_mntfromname, MNAMELEN));
      if ( ret )
          strncpy (ret, sfsb.f_mntfromname, MNAMELEN);
      else
        {
          perror ("Root device string allocation failed!");
          exit (ENOMEM);
        }
    }
  else
    {
      grub_util_error ("unable to stat `%s'", path);
    }
  return ret;
}

I would like to know what you guys think about this.

Francis Gendreau


[-- Attachment #2: stat_vs_statfs.c --]
[-- Type: text/x-csrc, Size: 775 bytes --]

#include <stdio.h>
#include <sys/types.h>
#include <sys/stat.h>

#include <sys/param.h>
#include <sys/mount.h>

void do_stat( const char *path);
int main( int argc, char **argv)
{
  int i;

  if ( argc > 1 ) 
    for ( i = 1; i < argc; ++i)
      do_stat( argv[i]);
  else
    printf( "usage: %s <path> [<path> ...]\n", argv[0]);
  return 0;
}

void do_stat( const char *path)
{
  struct stat sb;
  struct statfs sfsb;

  if ( !stat( path, &sb) )
  {
    if ( !statfs( path, &sfsb) )
    {
      printf(
      "%s: dev_t=%d, mntfrom=\"%s\", mnton=\"%s\"\n",
      path,
      sb.st_dev,
      sfsb.f_mntfromname,
      sfsb.f_mntonname
      );
      return;
    }
    printf( "cannot statfs() %s\n", path); // should never print
  }
  printf( "cannot stat() %s\n", path);
}

Re: grub-probe && statfs(2)

[Yoshinori K. Okuji]

Hello,

On Saturday 29 December 2007 13:36, Francis Gendreau wrote:
> My recommendation would be relatively simple. grub_guess_root_device()
> could simply call statfs() and extract the device name from struct
> statfs f_mntfromname, maybe as simply as the following, but I may lack
> of knowledge about some technologies on which this might not work (like
> RAID ...).

Unfortunately, statfs in Linux does not return a filename. From statfs(2):

       The Linux statfs was inspired by the 4.4BSD one (but they  do  not  use
       the same structure).

In fact, struct statfs does not contain f_mntfromname or anything equivallent 
on Linux.

So I think it would be necessary to use statfs, only on *BSD (or if struct 
statfs contains f_mntfromname).

Thanks,
Okuji

Re: grub-probe && statfs(2)

[Francis Gendreau]

On Sun, 2007-12-30 at 13:38 +0100, Yoshinori K. Okuji wrote:
> Hello,
> 
> On Saturday 29 December 2007 13:36, Francis Gendreau wrote:
> > My recommendation would be relatively simple. grub_guess_root_device()
> > could simply call statfs() and extract the device name from struct
> > statfs f_mntfromname, maybe as simply as the following, but I may lack
> > of knowledge about some technologies on which this might not work (like
> > RAID ...).
> 
> Unfortunately, statfs in Linux does not return a filename. From statfs(2):
> 
>        The Linux statfs was inspired by the 4.4BSD one (but they  do  not  use
>        the same structure).
> 
> In fact, struct statfs does not contain f_mntfromname or anything equivallent 
> on Linux.
> 
> So I think it would be necessary to use statfs, only on *BSD (or if struct 
> statfs contains f_mntfromname).
> 
> Thanks,
> Okuji
> 

Well, after some googling, which I should have done before submitting, I
saw that Linux does not contain any informations required for the
operation. OpenBSD and FreeBSD does, but NetBSD do not offer this system
call.

I already wrote a patch for FreeBSD but lots of other errors makes it
almost useless. I'm currently tring to elaborate the necessary
alterations to the source so grub2 can be installed from a FreeBSD
system.

Francis Gendreau