Arrrgh, after much fiddling with floppies, as I trundle patched code between my
laptop and my hosed machine, this looks partly like a bug in LVM.
Specifically, in tools/lib/pv_read_all_pv_of_vg.c, in this section of code:
/* make array contiguous again */
for ( i = 0; i < np - 2; i++) {
if ( pv_this[i] == NULL) {
/* ensure we don't have a sequence of NULLs */
if ( pv_this[i+1] == NULL) {
int j = i + 1;
while ( pv_this[j] == NULL && j < np - 1) j++;
if ( j < np - 1) pv_this[i+1] = pv_this[j];
}
pv_this[i] = pv_this[i+1];
pv_this[i+1] = NULL;
}
}
np=0;
while ( pv_this[np] != NULL) np++;
Let's step through this. Before multiple path/md handling, pv_this[] looked
like this:
[0] /dev/md0
[1] /dev/sda
[2] /dev/sde3
[3] /dev/sde4
pv_this[1] is set to NULL, because sda is part of md0, so we enter the
above section of code with pv_this[] looking like this:
[0] /dev/md0
[1] NULL
[2] /dev/sde3
[3] /dev/sde4
In the first iteration of the loop (i=0) nothing changes.
In the second iteration of this loop (i=1), we encounter a NULL value, so
we shift the list down one and set the following element to NULL, leaving us
with:
[0] /dev/md0
[1] /dev/sde3
[2] NULL
[3] /dev/sde4
And that's it. Due to the loop condition (i < np - 2), the loop exits,
since np=4 and i would be 2 in the next iteration. The following code:
while ( pv_this[np] != NULL) np++;
Will end up setting np to 2, since the NULL value in pv_this[2] will cause
the loop to exit.
I think this loop will work as intended if the loop condition were
i < np -1. In our situtation, this would cause one more iteration of the loop, and the NULL value in pv_this[2] would cause another shift to occur leaving us with:
[0] /dev/md0
[1] /dev/sde3
[2] /dev/sde4
[3] NULL
Which looks just great. I'm ignoring the sub-loop (...ensure we don't have
a sequence of NULLs...) for now, but it probably needs fixing.
I'm about to go check this out. I noticed that there is at least one other
place in this file that may have a similar problem. Results in a few minutes.
-- Lars
