Skip to contentSkip to conversation

[2.4] Multiple memleaks in IBM Hot Plug Controller Driver

PATCH Oleg Drokin 4 participants 7 messages Mar 13, 2003 – Mar 14, 2003

lore
Share & tools
lore
#0 lore raw

Hello!

There seem to be memleak convert_2digits_to_char() function that is triggered during normal operations.
Also I think there are some memleaks on error exit paths ebda_rsrc_controller()
All of this is addressed by below patch.
2.5 seems to have totally different version of this code (and no convert_2digits_to_char() function at all for example).
Found with help of smatch + enhanced unfree script.

Bye,
Oleg

===== drivers/hotplug/ibmphp_ebda.c 1.6 vs edited =====

drivers/hotplug/ibmphp_ebda.c +22 -3
@@ -597,8 +597,8 @@
 	char *str1;
 
 	str = (char *) kmalloc (3, GFP_KERNEL);
-	memset (str, 0, 3);
-	str1 = (char *) kmalloc (2, GFP_KERNEL);
+	if (!str)
+		return NULL;
 	memset (str, 0, 3);
 	bit = (int)(var / 10);
 	switch (bit) {
@@ -608,13 +608,20 @@
 		return str;
 	default: 	
 		//2 digits number
+		str1 = (char *) kmalloc (2, GFP_KERNEL);
+		if (!str1) {
+			break;
+		}
+		memset (str, 0, 3);
 		*str1 = (char)(bit + 48);
 		strncpy (str, str1, 1);
 		memset (str1, 0, 3);
 		*str1 = (char)((var % 10) + 48);
 		strcat (str, str1);
+		kfree(str1);
 		return str;
-	}	
+	}
+	kfree(str);
 	return NULL;	
 }
 
@@ -1022,6 +1029,10 @@
 			bus_info_ptr1 = ibmphp_find_same_bus_num (hpc_ptr->slots[index].slot_bus_num);
 			if (!bus_info_ptr1) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return -ENODEV;
 			}
 			((struct slot *) hp_slot_ptr->private)->bus_on = bus_info_ptr1;
@@ -1036,12 +1047,20 @@
 			rc = ibmphp_hpc_fillhpslotinfo (hp_slot_ptr);
 			if (rc) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return rc;
 			}
 
 			rc = ibmphp_init_devno ((struct slot **) &hp_slot_ptr->private);
 			if (rc) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return rc;
 			}
 			hp_slot_ptr->ops = &ibmphp_hotplug_slot_ops;
#1 lore raw
Quoted 9 lines · Oleg Drokin

Hello!

There seem to be memleak convert_2digits_to_char() function that is triggered during normal operations.
Also I think there are some memleaks on error exit paths ebda_rsrc_controller()
All of this is addressed by below patch.
2.5 seems to have totally different version of this code (and no convert_2digits_to_char() function at all for example).

Yes, the 2.5 version should be backported to 2.4. There have been a number of error patch fixes in the 2.5 version, care to make up a patch?

Quoted 1 line

Found with help of smatch + enhanced unfree script.

The 2.5 changes were found with smatch too :)

thanks,

greg k-h

#2 lore raw

Hello!

Quoted 9 lines · Greg KH
Quoted 7 lines

There seem to be memleak convert_2digits_to_char() function that is triggered during normal operations.
Also I think there are some memleaks on error exit paths ebda_rsrc_controller()
All of this is addressed by below patch.
2.5 seems to have totally different version of this code (and no convert_2digits_to_char() function at all for example).

Yes, the 2.5 version should be backported to 2.4. There have been a number of error patch fixes in the 2.5 version, care to make up a patch?

Ok, see below.
Note there is no credit for me. Your 2 patches from 2.5 applied to 2.4 just fine even without offsets, and stuff compiled (I have no hardware to test how it works, but I presume it will work ok).
Comments for patches were:
C IBM PCI Hotplug driver: Clean up the slot filename generation logic a lot. C IBM PCI Hotplug: fix a load of memory leak errors found by the checker project .

Bye,
Oleg
===== drivers/hotplug/ibmphp_ebda.c 1.7 vs 1.8 =====

drivers/hotplug/ibmphp_ebda.c +6 -71
@@ -65,8 +65,6 @@
 static LIST_HEAD (opt_lo_head);
 static void *io_mem;
 
-char *chassis_str, *rxe_str, *str;
-
 /* Local functions */
 static int ebda_rsrc_controller (void);
 static int ebda_rsrc_rsrc (void);
@@ -591,32 +589,6 @@
 	return 0;	
 }
 	
-static char *convert_2digits_to_char (int var)
-{
-	int bit;	
-	char *str1;
-
-	str = (char *) kmalloc (3, GFP_KERNEL);
-	memset (str, 0, 3);
-	str1 = (char *) kmalloc (2, GFP_KERNEL);
-	memset (str, 0, 3);
-	bit = (int)(var / 10);
-	switch (bit) {
-	case 0:
-		//one digit number
-		*str = (char)(var + 48);
-		return str;
-	default: 	
-		//2 digits number
-		*str1 = (char)(bit + 48);
-		strncpy (str, str1, 1);
-		memset (str1, 0, 3);
-		*str1 = (char)((var % 10) + 48);
-		strcat (str, str1);
-		return str;
-	}	
-	return NULL;	
-}
 
 /* Since we don't know the max slot number per each chassis, hence go
  * through the list of all chassis to find out the range
@@ -701,7 +673,7 @@
 {
 	struct opt_rio *opt_vg_ptr = NULL;
 	struct opt_rio_lo *opt_lo_ptr = NULL;
-	char *ptr_chassis_num, *ptr_rxe_num, *ptr_slot_num;
+	static char str[30];
 	int which = 0; /* rxe = 1, chassis = 0 */
 	u8 number = 1; /* either chassis or rxe # */
 	u8 first_slot = 1;
@@ -715,19 +687,7 @@
 	
 	slot_num = slot_cur->number;
 
-	chassis_str = (char *) kmalloc (30, GFP_KERNEL);
-	memset (chassis_str, 0, 30);
-	rxe_str = (char *) kmalloc (30, GFP_KERNEL);
-	memset (rxe_str, 0, 30);
-	ptr_chassis_num = (char *) kmalloc (3, GFP_KERNEL);
-	memset (ptr_chassis_num, 0, 3);
-	ptr_rxe_num = (char *) kmalloc (3, GFP_KERNEL);
-	memset (ptr_rxe_num, 0, 3);
-	ptr_slot_num = (char *) kmalloc (3, GFP_KERNEL);
-	memset (ptr_slot_num, 0, 3);
-	
-	strcpy (chassis_str, "chassis");
-	strcpy (rxe_str, "rxe");
+	memset (str, 0, sizeof(str));
 	
 	if (rio_table_ptr) {
 		if (rio_table_ptr->ver_num == 3) {
@@ -772,31 +732,10 @@
 		}
 	}
 
-	switch (which) {
-	case 0:
-		/* Chassis */
-		*ptr_chassis_num = (char)(number + 48);
-		strcat (chassis_str, ptr_chassis_num);
-		kfree (ptr_chassis_num);
-		strcat (chassis_str, "slot");
-		ptr_slot_num = convert_2digits_to_char (slot_num - first_slot + 1);
-		strcat (chassis_str, ptr_slot_num);
-		kfree (ptr_slot_num);
-		return chassis_str;
-		break;
-	case 1:
-		/* RXE */
-		*ptr_rxe_num = (char)(number + 48);
-		strcat (rxe_str, ptr_rxe_num);
-		kfree (ptr_rxe_num);
-		strcat (rxe_str, "slot");
-		ptr_slot_num = convert_2digits_to_char (slot_num - first_slot + 1);
-		strcat (rxe_str, ptr_slot_num);
-		kfree (ptr_slot_num);
-		return rxe_str;
-		break;
-	}	
-	return NULL;
+	sprintf(str, "%s%dslot%d",
+		which == 0 ? "chassis" : "rxe",
+		number, slot_num - first_slot + 1);
+	return str;
 }
 
 static struct pci_driver ibmphp_driver;
@@ -1060,10 +999,6 @@
 		slot_cur = list_entry (list, struct slot, ibm_slot_list);
 
 		snprintf (slot_cur->hotplug_slot->name, 30, "%s", create_file_name (slot_cur));
-		if (chassis_str) 
-			kfree (chassis_str);
-		if (rxe_str)
-			kfree (rxe_str);
 		pci_hp_register (slot_cur->hotplug_slot);
 	}
 
#3 lore raw
In reply to Oleg Drokin

On Thursday 13 March 2003 21.45, Oleg Drokin wrote:

Below, why allocating 2 bytes on heap (str1,) only to non-conditionally free it a few lines further down? Why not keep the two bytes on stack instead? It also seems like a bad idea to strncopy/strcat 1 byte long strings.
_
/Bjorn.

Quoted 62 lines

===== drivers/hotplug/ibmphp_ebda.c 1.6 vs edited =====

drivers/hotplug/ibmphp_ebda.c +20 -2
@@ -608,13 +608,20 @@
 		return str;
 	default:
 		//2 digits number
+		str1 = (char *) kmalloc (2, GFP_KERNEL);
+		if (!str1) {
+			break;
+		}
+		memset (str, 0, 3);
 		*str1 = (char)(bit + 48);
 		strncpy (str, str1, 1);
 		memset (str1, 0, 3);
 		*str1 = (char)((var % 10) + 48);
 		strcat (str, str1);
+		kfree(str1);
 		return str;
-	}
+	}
+	kfree(str);
 	return NULL;
 }
 
@@ -1022,6 +1029,10 @@
 			bus_info_ptr1 = ibmphp_find_same_bus_num
 (hpc_ptr->slots[index].slot_bus_num); if (!bus_info_ptr1) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return -ENODEV;
 			}
 			((struct slot *) hp_slot_ptr->private)->bus_on = bus_info_ptr1;
@@ -1036,12 +1047,20 @@
 			rc = ibmphp_hpc_fillhpslotinfo (hp_slot_ptr);
 			if (rc) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return rc;
 			}
 
 			rc = ibmphp_init_devno ((struct slot **) &hp_slot_ptr->private);
 			if (rc) {
 				iounmap (io_mem);
+				kfree (hp_slot_ptr->name);
+				kfree (hp_slot_ptr->info);
+				kfree (hp_slot_ptr->private);
+				kfree (hp_slot_ptr);
 				return rc;
 			}
 			hp_slot_ptr->ops = &ibmphp_hotplug_slot_ops;
-

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

#4 lore raw
Quoted 5 lines · Björn Fahller

On Thursday 13 March 2003 21.45, Oleg Drokin wrote:

Below, why allocating 2 bytes on heap (str1,) only to non-conditionally free it a few lines further down? Why not keep the two bytes on stack instead? It also seems like a bad idea to strncopy/strcat 1 byte long strings.

Like I previously said, this whole function is a bad dream. Look at what is now in 2.5, all of this nonsense is now gone.

thanks,

greg k-h

#5 lore raw
In reply to Oleg Drokin
Quoted 11 lines · Oleg Drokin

Hello!

There seem to be memleak convert_2digits_to_char() function that
is triggered during normal operations.
Also I think there are some memleaks on error exit paths ebda_rsrc_controller()
All of this is addressed by below patch.
2.5 seems to have totally different version of this code (and no convert_2digits_to_char() function at all for example). Found with help of smatch + enhanced unfree script.

Quoted 2 lines

Bye,
Oleg

Oleg, you are doing great job. Thanks!

Well.. I just looked into that function. Do we have a kernel obfuscated C contest? ;)

drivers/hotplug/ibmphp_ebda.c line 599

Quoted 7 lines

===== drivers/hotplug/ibmphp_ebda.c 1.6 vs edited =====

drivers/hotplug/ibmphp_ebda.c +0 -0
@@ -597,8 +597,8 @@
 	char *str1;
 
 	str = (char *) kmalloc (3, GFP_KERNEL);

BTW, that char* str is not local. It is not even static. It is global symbol: "char *chassis_str, *rxe_str, *str;"

drivers/hotplug/ibmphp_ebda.c line 601

Quoted 2 lines
drivers/hotplug/ibmphp_ebda.c +0 -2
-	memset (str, 0, 3);
-	str1 = (char *) kmalloc (2, GFP_KERNEL);

A bit weird to have 3 and 2 bytes allocated in two separate kmalloc()

drivers/hotplug/ibmphp_ebda.c line 602

Quoted 3 lines
drivers/hotplug/ibmphp_ebda.c +2 -0
+	if (!str)
+		return NULL;
 	memset (str, 0, 3);

This was fun, right? "Lets add second memset just in case first failed" ;) ;)

Quoted 5 lines
drivers/hotplug/ibmphp_ebda.c +0 -0
 	bit = (int)(var / 10);
 	switch (bit) {
@@ -608,13 +608,20 @@
 		return str;
 	default:

drivers/hotplug/ibmphp_ebda.c line 619

Quoted 10 lines
drivers/hotplug/ibmphp_ebda.c +6 -0
 		//2 digits number
+		str1 = (char *) kmalloc (2, GFP_KERNEL);
+		if (!str1) {
+			break;
+		}
+		memset (str, 0, 3);
+             memset (str, 0, 3);
              *str1 = (char)(bit + 48);
              strncpy (str, str1, 1);
              memset (str1, 0, 3);

Wow! *str1 is 2 bytes long, not 3!

Anyway, here is the diff against some old 2.5 (sorry don't have latest tree here at the moment). Also here are old and new functions for easy visual diff. Completely untested:

static char *convert_2digits_to_char (int var)

{
        int bit;
        char *str1;

        str = (char *) kmalloc (3, GFP_KERNEL);
        memset (str, 0, 3);
        str1 = (char *) kmalloc (2, GFP_KERNEL);
        memset (str, 0, 3);
        bit = (int)(var / 10);
        switch (bit) {
        case 0:
                //one digit number
                *str = (char)(var + 48);
                return str;
        default:
                //2 digits number
                *str1 = (char)(bit + 48);
                strncpy (str, str1, 1);
                memset (str1, 0, 3);
                *str1 = (char)((var % 10) + 48);
                strcat (str, str1);
                return str;
        }
        return NULL;
}

static char *convert_2digits_to_char (int var)

{
        int bit;

        char *str = (char *) kmalloc (3, GFP_KERNEL);
        if (!str)
                return NULL;
        bit = (int)(var / 10);
        switch (bit) {
        case 0:
                //one digit number
                str[0] = (char)(var + '0');
                str[1] = 0;
                break;
        default:
                //2 digits number
                str[0] = (char)(bit + '0');
                str[1] = (char)((var % 10) + '0');
                str[2] = 0;
                break;
        }
        return str;
}

--
vda

ibmphp_ebda.new.c renamed from ibmphp_ebda.c +13 -16
@@ -65,7 +65,7 @@
 static LIST_HEAD (opt_lo_head);
 static void *io_mem;
 
-char *chassis_str, *rxe_str, *str;
+char *chassis_str, *rxe_str;
 
 /* Local functions */
 static int ebda_rsrc_controller (void);
@@ -594,28 +594,25 @@
 static char *convert_2digits_to_char (int var)
 {
 	int bit;	
-	char *str1;
 
-	str = (char *) kmalloc (3, GFP_KERNEL);
-	memset (str, 0, 3);
-	str1 = (char *) kmalloc (2, GFP_KERNEL);
-	memset (str, 0, 3);
+	char *str = (char *) kmalloc (3, GFP_KERNEL);
+	if (!str)
+		return NULL;	
 	bit = (int)(var / 10);
 	switch (bit) {
 	case 0:
 		//one digit number
-		*str = (char)(var + 48);
-		return str;
+		str[0] = (char)(var + '0');
+		str[1] = 0;
+		break;
 	default: 	
 		//2 digits number
-		*str1 = (char)(bit + 48);
-		strncpy (str, str1, 1);
-		memset (str1, 0, 3);
-		*str1 = (char)((var % 10) + 48);
-		strcat (str, str1);
-		return str;
-	}	
-	return NULL;	
+		str[0] = (char)(bit + '0');
+		str[1] = (char)((var % 10) + '0');
+		str[2] = 0;
+		break;
+	}
+	return str;
 }
 
 /* Since we don't know the max slot number per each chassis, hence go
#6 lore raw

Hello!

Quoted 5 lines · Denis Vlasenko

ibmphp_ebda.new.c

Quoted 3 lines
ibmphp_ebda.new.c renamed from ibmphp_ebda.c +2 -0
+	if (!str)
+		return NULL;
 	memset (str, 0, 3);

This was fun, right? "Lets add second memset just in case first failed" ;) ;)

Ah, I was half asleep when doing the change, so I missed
other obvious stuff it seems, like memset of three bytes to two byte variabe :)
memset on wrong thing (second memset should be applied to str1 of course.)

But as Greg KH said already, whis stuff should be replaced by version from 2.5 instead ;)

Quoted 8 lines
Quoted 8 lines
ibmphp_ebda.new.c renamed from ibmphp_ebda.c +2 -0
 	bit = (int)(var / 10);
 	switch (bit) {
@@ -608,13 +608,20 @@
 		return str;
 	default:
 		//2 digits number
+		str1 = (char *) kmalloc (2, GFP_KERNEL);
+		if (!str1) {
Quoted 8 lines

Review line 619

Quoted 7 lines
(quoted diff) +4 -0
+			break;
+		}
+		memset (str, 0, 3);
+             memset (str, 0, 3);
              *str1 = (char)(bit + 48);
              strncpy (str, str1, 1);
              memset (str1, 0, 3);

Wow! *str1 is 2 bytes long, not 3!

yup.

Quoted 3 lines

Anyway, here is the diff against some old 2.5 (sorry don't have latest tree here at the moment). Also here are old and new functions for easy visual diff. Completely untested:

New 2.5 code does not have this function at all.
I will look into porting 2.5 changes back to 2.4 tonight. This seems to be proper solution in this case.

Bye,
Oleg