Fix segfault when using vgsplit in stacked environment
PATCH 1/1 Fix segfault when using vgsplit in stacked environment
Hi,
this is a small patch for segmentation fault in vgsplit while using stacked VGs (rh bz 481793). Segfault is caused by null dereference in 'pv_uses_vg' function that is called to check if a PV uses VG somewhere in its construction. In stacked environment, when the LV/VG from lower layer is disabled, the PV constructed above does not exist as well and so the device value in PV's structure is set to NULL. We should check for this NULL value in 'pv_uses_vg' and return immediately if the situation occurs.
Peter
lib/activate/activate.c +3 -0
@@ -1105,6 +1105,9 @@ int pv_uses_vg(struct physical_volume *pv,if (!activation ())return 0 ; +if (!pv->dev || !pv->dev->dev) +return 1 ; +if (!dm_is_dm_major (MAJOR (pv->dev->dev)))return 0 ;
PATCH 1/1 Fix segfault when using vgsplit in stacked environment
lib/activate/activate.c int pv_uses_vg(struct physical_volume *pv, · line 1110
Quoted 11 lines · Peter Rajnoha
lib/activate/activate.c +3
-0@@ -1105,6 +1105,9 @@ int pv_uses_vg(struct physical_volume *pv,if (!activation ())return 0 ; +if (!pv->dev || !pv->dev->dev) +return 1 ; +
(answering myself - why "return 1". the function name is misleading for me...) /*
* Does PV use VG somewhere in its construction?
* Returns 1 on failure.
*/
int pv_uses_vg(struct physical_volume *pv,
struct volume_group *vg)
Please add something like this to cover it by testsuite (best add it to already existing vgsplit tests):
. ./test-utils.sh
aux prepare_devs 3
pvcreate $devs
vgcreate $vg1 $dev1 $dev2
lvcreate -n $lv1 -l 100%FREE $vg1
#top VG
pvcreate $G_dev_/$vg1/$lv1
vgcreate $vg $G_dev_/$vg1/$lv1 $dev3
vgchange -a n $vg
vgchange -a n $vg1
# this should fail but not segfault not vgsplit $vg $vg1 $dev3
Milan
--
mbroz at redhat.com
PATCH 1/1 Fix segfault when using vgsplit in stacked environment
In reply to Peter RajnohaQuoted 5 lines · Peter Rajnoha
this is a small patch for segmentation fault in vgsplit while using stacked VGs (rh bz 481793). Segfault is caused by null dereference in '
pv_uses_vg' function that is called to check if a PV uses VG somewhere in its construction. In stacked environment, when the LV/VG from lower layer is disabled, the PV constructed above does not exist as well and
I don't think the code should be getting as far as it does in that situation.
It is meaningless to call pv_uses_vg() if the PV does not exist. And the function currently returns 0 if the answer is unknown - it'll get even more confusing if it returns 0 for some types of 'unknown' and 1 for other types!
Alasdair
Signature · 1 line
agk at redhat.com
PATCH 1/1 Fix segfault when using vgsplit in stacked environment
Quoted 1 line · Alasdair G Kergon
It is meaningless to call
pv_uses_vg()if the PV does not exist.
Well, it's some kind of an exception, so I thought it would be better to check it right here than all the other places outside the function.
Quoted 3 lines
And the function currently returns 0 if the answer is unknown - it'll get even more confusing if it returns 0 for some types of 'unknown' and 1 for other types!
Hmm, isn't it the opposite way? Looking at the dev_manager_device_uses_vg function that is called from pv_uses_vg and returs its value, dev_manager_device_uses_vg should return 1 if uncertain. So I relied on this at that time... I return 1 when checking for NULL value of pv->dev, because we can't decide if PV uses VG in this situation, too.
Peter
PATCH 1/1 Fix segfault when using vgsplit in stacked environment
In reply to Alasdair G KergonHi,
Quoted 2 lines · Alasdair G Kergon
I don't think the code should be getting as far as it does in that situation.
(explanation for the list) After an IRC discussion, we have agreed that a different approach would work here, specifically, that we should not allow tools to try tinkering with VGs that have PVs missing, unless they specifically know what they are doing. There's probably half dozen other places where we assume that pv->dev is valid. The attached patch changes the meaning of cmd->handles_missing_pvs somewhat: If a tool now opens a VG *for writing* but it does not set handles_missing_pvs, vg_read will fail.
This check was previously done only in vg_write, which led to situations like the above bug, where a little less vigorous code path trips a NULL pointer. This behaviour change of handles_missing_pvs affects these situations:
- lvchange -a, vgchange -a take a write lock, so they need to set, for the -a case, handles_missing_pvs. This is safe.
- vgreduce needs to set handles_missing_pvs, since it is supposed to work in that situation... it previously did not need the flag in general, since it most of the time writes out a VG that has no PVs missing in it; however, since it's now needed also for opening the VG for writing, this needs to be always set for --removemissing; there's a small risk of bugs associated, but it can be mitigated by adding appropriate check to the code...
Please note that this patch depends on the "vg_read" patchset found elsewhere on this list.
Yours,
Petr.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: lvm-vg_read-missing_pvs.diff
Type: text/x-diff
Size: 4043 bytes
Desc: lvm-vg_read-missing_pvs.diff
URL: <http://listman.redhat.com/archives/lvm-devel/attachments/20090216/47777170/attachment.bin>
-------------- next part --------------
PS: The patch passes the testsuite and also makes Milan's testcase pass (it is included with the patch).
Signature · 6 lines
Peter Rockai | me()mornfall!net | prockai()redhat!com
http://blog.mornfall.net | http://web.mornfall.net
"In My Egotistical Opinion, most people's C programs should be
indented six feet downward and covered with dirt."
-- Blair P. Houghton on the subject of C program indentation