Skip to contentSkip to conversation

Fix NULL dereference

PATCH New Zdenek Kabelac 4 participants 17 messages Oct 26, 2010 – Oct 27, 2010

lore
Share & tools
lore
#0 lore raw

cover Fix NULL dereference

Updated patchset for NULL pointer dereferences issues reported by clang.

Unlike the first version - this time less aggresive solution is used. INTERNAL_ERRORs are reported in these moments (if they would ever happen), and the execution path aborts when such conditions are met. Previous version was rather ignoring these paths and could lead to unwanted execution of other code parts.

Zdenek Kabelac (5):
Check type is not NULL before access Ensure seg is nonnull
Ensure first is not NULL before dereference Fix theoretical usage of NULL pointer dereference Check for NULL pointer

 lib/activate/dev_manager.c |    5 +++++
 lib/format1/format1.c      |    5 +++++
 libdm/libdm-report.c       |    6 ++++++
 libdm/regex/ttree.c        |    6 ++++++
 tools/reporter.c           |    4 ++++
 5 files changed, 26 insertions(+), 0 deletions(-)
Signature · 1 line
1.7.3.2
#1 lore raw

PATCH 1/5 Check type is not NULL before access

clang Logic error Dereference of null pointer

Check type before dereference and report internal error in case it's undefined and return from the function.

---
 libdm/libdm-report.c |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
libdm/libdm-report.c +6 -0
index 7631e21..db24fed 100644
@@ -273,6 +273,12 @@ static void _display_fields(struct dm_report *rh)
 			log_warn("%*.*s", (int) strlen(desc) + 7,
 				 (int) strlen(desc) + 7,
 				 "-------------------------------------------------------------------------------");
+
+			if (!type) {
+				log_error(INTERNAL_ERROR "Type is not defined.");
+				return;
+			}
+
 			log_warn("  %sall%-*s - %s", type->prefix,
 				 (int) (id_len - 3 - strlen(type->prefix)), "",
 				 "All fields in this section.");
Signature · 1 line
1.7.3.2
#2 lore raw

PATCH 1/5 Check type is not NULL before access

libdm/libdm-report.c static void _display_fields(struct dm_report *rh) · line 280

Quoted 7 lines · Zdenek Kabelac

Check type before dereference and report internal error in case it's undefined and return from the function.

libdm/libdm-report.c +4 -0
+			if (!type) {
+				log_error(INTERNAL_ERROR "Type is not defined.");
+				return;
+			}

Is this one really an internal library error?

Can it be triggered if someone usees the API incorrectly?

If so, it should be treated like validation.

Alasdair

#3 lore raw

PATCH 2/5 Ensure seg is nonnull

In reply to Zdenek Kabelac

clang Logic error Dereference of null pointer
Make sure we do not try to use NULL seg pointer.

---
 lib/activate/dev_manager.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
lib/activate/dev_manager.c +5 -0
index cbb5378..cc67a60 100644
@@ -626,6 +626,11 @@ int dev_manager_transient(struct dev_manager *dm, struct logical_volume *lv)
 		if (!type || !params)
 			continue;
 
+		if (!seg) {
+			log_error(INTERNAL_ERROR "Segment is not selected.");
+			goto out;
+		}
+
 		if (seg->segtype->ops->check_transient_status &&
 		    !seg->segtype->ops->check_transient_status(seg, params))
 			goto_out;
Signature · 1 line
1.7.3.2
#4 lore raw

PATCH 2/5 Ensure seg is nonnull

lib/activate/dev_manager.c int dev_manager_transient(struct dev_manager *dm, struct logical_volume *lv) · line 630

Quoted 2 lines · Zdenek Kabelac
lib/activate/dev_manager.c +2 -0
+		if (!seg) {
+			log_error(INTERNAL_ERROR "Segment is not selected.");

I don't mind assertions like this, but please describe the problem more accurately in the message. If this triggers, doesn't it mean we have a corrupt segment list? So say that.
"LV %s segment list corrupted: NULL struct lv_segment detected."

But isn't there a missing 'if (lv)' ?

Alasdair

#5 lore raw

PATCH 3/5 Ensure first is not NULL before dereference

In reply to Zdenek Kabelac

clang Logic error Dereference of null pointer

For empty lists make sure the NULL first pointer is not dereferenced and report error.

---
 lib/format1/format1.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
lib/format1/format1.c +5 -0
index 0dcec05..c5685c9 100644
@@ -102,6 +102,11 @@ static int _check_vgs(struct dm_list *pvs, struct volume_group *vg)
 		pv_count++;
 	}
 
+	if (!first) {
+		log_error(INTERNAL_ERROR "Unexpected empty PV list.");
+		return 0;
+	}
+
 	/* On entry to fn, list known to be non-empty */
 	if (pv_count != first->vgd.pv_cur) {
 		log_error("%d PV(s) found for VG %s: expected %d",
Signature · 1 line
1.7.3.2
#6 lore raw

PATCH 3/5 Ensure first is not NULL before dereference

lib/format1/format1.c static int _check_vgs(struct dm_list *pvs, struct volume_group *vg) · line 107

Quoted 3 lines · Zdenek Kabelac
lib/format1/format1.c +3 -0
+	if (!first) {
+		log_error(INTERNAL_ERROR "Unexpected empty PV list.");
+		return 0;

We already have a 'list empty' check prior to calling this function.

I think it's more readable with the check in the caller and the comment inside the function, instead of moving the check inside the function.

No need to change this that I can see.

What is the code path that could trigger this condition? (It's a static function only called from one place.)

What am I missing?

Alasdair

#7 lore raw

PATCH 4/5 Fix theoretical usage of NULL pointer dereference

In reply to Zdenek Kabelac

Make sure we have *c defined.
clang seems to be happier with this check.

---
 libdm/regex/ttree.c |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
libdm/regex/ttree.c +6 -0
index ec97c98..a384c59 100644
@@ -97,6 +97,12 @@ int ttree_insert(struct ttree *tt, unsigned int *key, void *data)
 			}
 		}
 	}
+
+	if (!*c) {
+		log_error(INTERNAL_ERROR "Insert failed.");
+		return 0;
+	}
+
 	(*c)->data = data;
 
 	return 1;
Signature · 1 line
1.7.3.2
#8 lore raw

PATCH 4/5 Fix theoretical usage of NULL pointer dereference

libdm/regex/ttree.c int ttree_insert(struct ttree *tt, unsigned int *key, void *data) · line 99

Quoted 4 lines · Zdenek Kabelac
libdm/regex/ttree.c +2 -0
@@ -97,6 +97,12 @@ int ttree_insert(struct ttree *tt, unsigned int *key, void *data)
 
+	if (!*c) {
+		log_error(INTERNAL_ERROR "Insert failed.");

What am I missing here?
Isn't that condition you are proposing to add logically impossible to trigger?!

It's not our job to add code just because a particular static analyser isn't clever enough to understand the logic.

Nack.

Alasdair

#9 lore raw

PATCH 4/5 Fix theoretical usage of NULL pointer dereference

At Tue, 26 Oct 2010 14:37:43 +0100,

Quoted 10 lines · Alasdair G Kergon

libdm/regex/ttree.c int ttree_insert(struct ttree *tt, unsigned int *key, void *data) · line 99

Quoted 4 lines · Zdenek Kabelac
libdm/regex/ttree.c +2 -0
@@ -97,6 +97,12 @@ int ttree_insert(struct ttree *tt, unsigned int *key, void *data)
 
+	if (!*c) {
+		log_error(INTERNAL_ERROR "Insert failed.");

What am I missing here?
Isn't that condition you are proposing to add logically impossible to trigger?!

Yes, it can't happen, as I said the first time this patch went round.

One thing I'm not clear on is how much benefit we're seeing from the CLang build? Kabi, how many genuine bugs did you find when you went through this process? If the benefits are real, then we can live with check like these. It would be nice if it was clearer that they are only there to pacify clang, maybe put a conditional compile in so they're only included with the clang build?

- Joe

#10 lore raw

PATCH 4/5 Fix theoretical usage of NULL pointer dereference

Dne 27.10.2010 12:19, ejt at redhat.com napsal(a):

Quoted 15 lines

At Tue, 26 Oct 2010 14:37:43 +0100,

Quoted 10 lines · Alasdair G Kergon

libdm/regex/ttree.c int ttree_insert(struct ttree *tt, unsigned int *key, void *data) · line 99

Quoted 4 lines · Zdenek Kabelac
libdm/regex/ttree.c +2 -0
@@ -97,6 +97,12 @@ int ttree_insert(struct ttree *tt, unsigned int *key, void *data)
 
+	if (!*c) {
+		log_error(INTERNAL_ERROR "Insert failed.");

What am I missing here?
Isn't that condition you are proposing to add logically impossible to trigger?!

Yes, it can't happen, as I said the first time this patch went round.

Quoted 8 lines

One thing I'm not clear on is how much benefit we're seeing from the CLang build? Kabi, how many genuine bugs did you find when you went through this process? If the benefits are real, then we can live with check like these. It would be nice if it was clearer that they are only there to pacify clang, maybe put a conditional compile in so they're only included with the clang build?

From the first original patchset real bugs are already commited.

I think patches 4, 5, 10, 14, 15. (Some of them are clearly problems of missing deep unit tests probably - but running scan-build - or spending months writing tests for every single error path in the code - I think we do not have manpower for this at this moment...)

The problem here is not what would happen if everything goes 'right', but what could happen if something goes 'wrong' - i.e. we may overwrite some bytes in memory by some other errors, we may leave some structure in wrong state, because of some unchecked error path - we could misuse or wrongly reuse something- obviously we will need to find the real cause of such memory overwrite - but the question is - is it the best thing to generate a coredump
- or should we nicely bailout from such case and give user some error report?

I didn't want to spend too much time with this thing in first place - so I've chosen easiest path here - and as mention in the patchset header - under normal circumstances lots of those (sometimes really crazy code paths) are not reachable - but if these checks are so cheap - why not add them - we do not mask the bug - we just avoid coredump for this case.

Zdenek

#11 lore raw

PATCH 5/5 Check for NULL pointer

In reply to Zdenek Kabelac

clangs is happier and check for non NULL options value. Report internal error in this case.

---
 tools/reporter.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)
tools/reporter.c +4 -0
index f39d23f..24a27ba 100644
@@ -324,6 +324,10 @@ static int _report(struct cmd_context *cmd, int argc, char **argv,
 			return EINVALID_CMD_LINE;
 		}
 		if (*opts == '+') {
+			if (!options) {
+				log_error(INTERNAL_ERROR "Missing options.");
+				return EINVALID_CMD_LINE;
+			}
 			if (!(str = dm_pool_alloc(cmd->mem,
 					 strlen(options) + strlen(opts) + 1))) {
 				log_error("options string allocation failed");
Signature · 1 line
1.7.3.2
#12 lore raw

PATCH 5/5 Check for NULL pointer

tools/reporter.c static int _report(struct cmd_context *cmd, int argc, char **argv, · line 330

Quoted 4 lines · Zdenek Kabelac
tools/reporter.c +4 -0
+			if (!options) {
+				log_error(INTERNAL_ERROR "Missing options.");
+				return EINVALID_CMD_LINE;
+			}

Again - what is the bug here?

options is always non-NULL there surely?

What code path causes that new check to fire?

Nack without a decent explanation in the error message of the *cause* not the symptom.

Alasdair

#13 lore raw

cover Fix NULL dereference

In reply to Zdenek Kabelac
Quoted 7 lines · Zdenek Kabelac

Updated patchset for NULL pointer dereferences issues reported by clang.

Unlike the first version - this time less aggresive solution is used. INTERNAL_ERRORs are reported in these moments (if they would ever happen), and the execution path aborts when such conditions are met. Previous version was rather ignoring these paths and could lead to unwanted execution of other code parts.

Well the ones I've looked at here seem to be more about dealing with shortcomings in the static analysis code rather than fixing real bugs.

Alasdair

#14 lore raw

cover Fix NULL dereference

Hi,

Quoted 2 lines · Alasdair G Kergon

Well the ones I've looked at here seem to be more about dealing with shortcomings in the static analysis code rather than fixing real bugs.

talking about static analysis. We used to have access to Coverity Prevent reports about LVM, right? If I try to go to scan.coverity.com and click "sign in" near the LVM2 project, I get a connection refused... the link is sending me to

http://scan2.coverity.com:7494/

Alasdair, I suppose you are the Coverity's "contact" -- could you maybe ask what's up with that, and maybe get me an account so I can see the report as well? I reckon it could be useful...

Yours,
Petr.

#15 lore raw

cover Fix NULL dereference

Dne 26.10.2010 16:05, Petr Rockai napsal(a):

Quoted 10 lines

Hi,

Quoted 2 lines · Alasdair G Kergon

Well the ones I've looked at here seem to be more about dealing with shortcomings in the static analysis code rather than fixing real bugs.

talking about static analysis. We used to have access to Coverity Prevent reports about LVM, right? If I try to go to scan.coverity.com and click "sign in" near the LVM2 project, I get a connection refused... the link is sending me to

Quoted 6 lines

http://scan2.coverity.com:7494/

Alasdair, I suppose you are the Coverity's "contact" -- could you maybe ask what's up with that, and maybe get me an account so I can see the report as well? I reckon it could be useful...

I've tried 4 times to get any answer from then - never got any responce. So I've considered they probably stopped to give access to their results.

Zdenek

#16 lore raw

cover Fix NULL dereference

In reply to Alasdair G Kergon

Dne 26.10.2010 15:57, Alasdair G Kergon napsal(a):

Quoted 12 lines · Zdenek Kabelac
Quoted 7 lines

Updated patchset for NULL pointer dereferences issues reported by clang.

Unlike the first version - this time less aggresive solution is used. INTERNAL_ERRORs are reported in these moments (if they would ever happen), and the execution path aborts when such conditions are met. Previous version was rather ignoring these paths and could lead to unwanted execution of other code parts.

Well the ones I've looked at here seem to be more about dealing with shortcomings in the static analysis code rather than fixing real bugs.

Some of them can never be triggered within current LVM code. Static analyzer is currently incapable to model data structure behavior to understand, that some settings can never happen and sometimes it creates very complex code path to model NULL pointer at the end.
(Also instrumentation nonnull would be handy here - but it's long term goal)

However my small patches here really just try to clean warning - the price for checks seems to be quite low and we do not need to look into analyzer output again and again.

We may also put them into

#ifdef __clang__
#endif

section to avoid any runtime overheads - but I don't like spreading such ifdefs everywhere.

I can also keep these patches in my private branch - to not be always bothered with same error.

For now I did not want to spend too much time on this so I've rather fixed easily and quickly what I've considered to be even worth to look at.

Of course deeper analysis here will require some time - so - placing them to my low-prio background queue....

Zdenek