31 Mar
2011
31 Mar
'11
7:52 p.m.
Hi all,
I'm running 3.1-branch git head currently under valgrind. And I've seen
several invalid reads and writes (apparently most are off by one).
==22274== Invalid read of size 1
==22274== at 0x6DA2C0A: pv_set_ruri_user (pv_core.c:1755)
==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x13A86F: run_actions_safe (action.c:1607)
==22274== by 0x1E2666: rval_get_int (rvalue.c:904)
==22274== by 0x1E5252: rval_expr_eval_int (rvalue.c:1866)
==22274== by 0x131B8A: do_action (action.c:1071)
==22274== by 0x13A6F1: run_actions (action.c:1555)
Is fairly obvious. modules_k/pv/pv_core.c has several places which take
backup of "val->rs.s[val->rs.len]" and replaces it with zero for string
termination. It's than later on set back to the original value. However,
it would appear that the value passed does not always point to a memory
area which is large enough. This results in multiple invalid reads and
writes of one.
The remaining reads are not so clear to me.
==22274== Invalid read of size 1
==22274== at 0x4811A69: strncpy (mc_replace_strmem.c:339)
==22274== by 0x6DA77FF: set_var_value (pv_svar.c:122)
==22274== by 0x6DA1EDB: pv_set_scriptvar (pv_core.c:1636)
==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x131E05: do_action (action.c:1086)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x131E05: do_action (action.c:1086)
==22274== Invalid read of size 2
==22274== at 0x4811E1F: memcpy (mc_replace_strmem.c:523)
==22274== by 0x1E134D: rval_new_str (rvalue.c:269)
==22274== by 0x1E394E: rval_convert (rvalue.c:1301)
==22274== by 0x1E4780: rval_str_add2 (rvalue.c:1610)
==22274== by 0x1E8681: rval_expr_eval (rvalue.c:2399)
==22274== by 0x13BBD9: lval_assign (lvalue.c:389)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x139229: do_action (action.c:1342)
==22274== by 0x13A6F1: run_actions (action.c:1555)
Any ideas how to fix these? Or any instructions what to provide more in
the bug report?
Cheers,
Timo
1 Apr
1 Apr
10:29 a.m.
On 03/31/2011 07:52 PM, Timo Teräs wrote:
Hi all,
I'm running 3.1-branch git head currently under valgrind. And I've seen
several invalid reads and writes (apparently most are off by one).
Hello,
I haven't used Valgrind memcheck tool for Kamailio , but from what I
know it might not be that useful. What Valgrind does it hooks into the
memory allocation functions like malloc/calloc/realloc/alloca and keeps
track of said allocation. Because, by default, Kamailio uses it's own
memory manager (both for private memory or shared), this messes up on
how Valgrind can keep track of its allocation/deallocation.
I don't know if there is a way to make Valgrind to work reliable with K
memory managers, but I will take a look on the cases you provide just to
be sure.
As a side note, I had uses static analysis tools quite succesfully on
3.1 . I had use clang analyzer from the llvm suite. There are still some
outstanding bugs which I didn't had time to look at.
Marius
==22274== Invalid read of size 1
==22274== at 0x6DA2C0A: pv_set_ruri_user (pv_core.c:1755)
==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x13A86F: run_actions_safe (action.c:1607)
==22274== by 0x1E2666: rval_get_int (rvalue.c:904)
==22274== by 0x1E5252: rval_expr_eval_int (rvalue.c:1866)
==22274== by 0x131B8A: do_action (action.c:1071)
==22274== by 0x13A6F1: run_actions (action.c:1555)
Is fairly obvious. modules_k/pv/pv_core.c has several places which take
backup of "val->rs.s[val->rs.len]" and replaces it with zero for string
termination. It's than later on set back to the original value. However,
it would appear that the value passed does not always point to a memory
area which is large enough. This results in multiple invalid reads and
writes of one.
The remaining reads are not so clear to me.
==22274== Invalid read of size 1
==22274== at 0x4811A69: strncpy (mc_replace_strmem.c:339)
==22274== by 0x6DA77FF: set_var_value (pv_svar.c:122)
==22274== by 0x6DA1EDB: pv_set_scriptvar (pv_core.c:1636)
==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x131E05: do_action (action.c:1086)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x131E05: do_action (action.c:1086)
==22274== Invalid read of size 2
==22274== at 0x4811E1F: memcpy (mc_replace_strmem.c:523)
==22274== by 0x1E134D: rval_new_str (rvalue.c:269)
==22274== by 0x1E394E: rval_convert (rvalue.c:1301)
==22274== by 0x1E4780: rval_str_add2 (rvalue.c:1610)
==22274== by 0x1E8681: rval_expr_eval (rvalue.c:2399)
==22274== by 0x13BBD9: lval_assign (lvalue.c:389)
==22274== by 0x139E4D: do_action (action.c:1472)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x12FDEB: do_action (action.c:711)
==22274== by 0x13A6F1: run_actions (action.c:1555)
==22274== by 0x139229: do_action (action.c:1342)
==22274== by 0x13A6F1: run_actions (action.c:1555)
Any ideas how to fix these? Or any instructions what to provide more in
the bug report?
Cheers,
Timo
_______________________________________________
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list
sr-users(a)lists.sip-router.org
http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
10:38 a.m.
On 04/01/2011 10:29 AM, marius zbihlei wrote:
On 03/31/2011 07:52 PM, Timo Teräs wrote:
It does not make valgrind useless. Basically, it just makes valgrind not
detect all errors. Any errors it does detect, should be valid.
Hi all,
I'm running 3.1-branch git head currently under valgrind. And I've seen
several invalid reads and writes (apparently most are off by one).
Hello,
I haven't used Valgrind memcheck tool for Kamailio , but from what I
know it might not be that useful. What Valgrind does it hooks into the
memory allocation functions like malloc/calloc/realloc/alloca and keeps
track of said allocation. Because, by default, Kamailio uses it's own
memory manager (both for private memory or shared), this messes up on
how Valgrind can keep track of its allocation/deallocation. I don't know if there is a way to make Valgrind to
work reliable with K
memory managers, but I will take a look on the cases you provide just to
be sure.
There are valgrind hooks for memory managers. Should probably add them,
so it becomes aware of the things. However, those should be only needed
for memory pools. Ultimately kamailio uses malloc or mmap to allocate
the memory and valgrind knows both.
As a side note, I had uses static analysis tools quite
succesfully on
3.1 . I had use clang analyzer from the llvm suite. There are still some
outstanding bugs which I didn't had time to look at.
Looks like all my bugs are related to pvars and kamailio modules.
The first really evil programming mistake I found is at:
modules_k/textops/txtvar.c:tr_txt_eval_re()
In line 86, it will fills the transformed variable with tr_txt_buf and
returns that. However, that is a stack variable, which will get
corrupted pretty soon after function return. I'm not sure how to fix
this properly.
Marius
> ==22274== Invalid read of size 1
> ==22274== at 0x6DA2C0A: pv_set_ruri_user (pv_core.c:1755)
> ==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
> ==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
> ==22274== by 0x139E4D: do_action (action.c:1472)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x12FDEB: do_action (action.c:711)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x13A86F: run_actions_safe (action.c:1607)
> ==22274== by 0x1E2666: rval_get_int (rvalue.c:904)
> ==22274== by 0x1E5252: rval_expr_eval_int (rvalue.c:1866)
> ==22274== by 0x131B8A: do_action (action.c:1071)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
>
> Is fairly obvious. modules_k/pv/pv_core.c has several places which take
> backup of "val->rs.s[val->rs.len]" and replaces it with zero for
string
> termination. It's than later on set back to the original value. However,
> it would appear that the value passed does not always point to a memory
> area which is large enough. This results in multiple invalid reads and
> writes of one.
I fixed this by making a temporary copy of the string data to local
stack variable and null terminating it there. That made valgrind happy.
So obviously pv_set_ruri_user and others, are called with strings that
are allocated for the exact length, and the hacky zero-termination
cannot be used in this module.
> The remaining reads are not so clear to me.
I'm analysing the rest problems. Seems that they are almost always
related to using re.subst, which has the problem mentioned in the
beginning of the mail.
- Timo
11:17 a.m.
On 04/01/2011 10:38 AM, Timo Teräs wrote:
Hello,
Comments inline.
The allocation inside the memory pools, that is
what valgrind must be
aware of.
Well, indeed that's a ugly hack. As str's are not 0 terminated, you
either copy to a buffer length+1 or you overwrite the memory location
next to the end and revert the change afterwards. Not the cleanest way
to do it, but from a performance point of view, it make sense. If should
be taken extra care to ensure that on every code path the backup is
restored.
Marius
On 04/01/2011 10:29 AM, marius zbihlei wrote:
That is good to know.
On 03/31/2011 07:52 PM, Timo Teräs wrote:
It does not make valgrind useless. Basically, it just makes valgrind not
detect all errors. Any errors it does detect, should be valid.
Hi all,
I'm running 3.1-branch git head currently under valgrind. And I've seen
several invalid reads and writes (apparently most are off by one).
Hello,
I haven't used Valgrind memcheck tool for Kamailio , but from what I
know it might not be that useful. What Valgrind does it hooks into the
memory allocation functions like malloc/calloc/realloc/alloca and keeps
track of said allocation. Because, by default, Kamailio uses it's own
memory manager (both for private memory or shared), this messes up on
how Valgrind can keep track of its allocation/deallocation.
I don't
know if there is a way to make Valgrind to work reliable with K
memory managers, but I will take a look on the cases you provide just to
be sure.
There are valgrind hooks for memory managers. Should probably add them,
so it becomes aware of the things. However, those should be only needed
for memory pools. Ultimately kamailio uses malloc or mmap to allocate
the memory and valgrind knows both. The first really evil programming mistake I found is
at:
modules_k/textops/txtvar.c:tr_txt_eval_re()
In line 86, it will fills the transformed variable with tr_txt_buf and
returns that. However, that is a stack variable, which will get
corrupted pretty soon after function return. I'm not sure how to fix
this properly.
Good call. I truly belive that should a "static" keyword in front of
the
buffer declaration. Can you test it. (because in the cfg you usually
assign that transformation result to another pvar, the buffer will be
deep copied soon enough).
diff --git a/modules_k/textops/txt_var.c b/modules_k/textops/txt_var.c
index ae1fce6..a13e5e1 100644
--- a/modules_k/textops/txt_var.c
+++ b/modules_k/textops/txt_var.c
@@ -41,7 +41,7 @@ int tr_txt_eval_re(struct sip_msg *msg, tr_param_t
*tp, int subtype,
int nmatches;
str* result;
#define TR_TXT_BUF_SIZE 2048
- char tr_txt_buf[TR_TXT_BUF_SIZE];
+ static char tr_txt_buf[TR_TXT_BUF_SIZE];
if(val==NULL || (!(val->flags&PV_VAL_STR)) || val->rs.len<=0)
return -1;
Marius
> ==22274== Invalid read of size 1
> ==22274== at 0x6DA2C0A: pv_set_ruri_user (pv_core.c:1755)
> ==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
> ==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
> ==22274== by 0x139E4D: do_action (action.c:1472)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x12FDEB: do_action (action.c:711)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x13A86F: run_actions_safe (action.c:1607)
> ==22274== by 0x1E2666: rval_get_int (rvalue.c:904)
> ==22274== by 0x1E5252: rval_expr_eval_int (rvalue.c:1866)
> ==22274== by 0x131B8A: do_action (action.c:1071)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
>
> Is fairly obvious. modules_k/pv/pv_core.c has several places which take
> backup of "val->rs.s[val->rs.len]" and replaces it with zero for
string
> termination. It's than later on set back to the original value. However,
> it would appear that the value passed does not always point to a memory
> area which is large enough. This results in multiple invalid reads and
> writes of one.
>
I fixed this by making a temporary copy of the string data to local
stack variable and null terminating it there. That made valgrind happy.
So obviously pv_set_ruri_user and others, are called with strings that
are allocated for the exact length, and the hacky zero-termination
cannot be used in this module.
> The
remaining reads are not so clear to me.
>
I'm analysing the rest problems. Seems that they are almost
always
related to using re.subst, which has the problem mentioned in the
beginning of the mail.
- Timo
12:53 p.m.
On 04/01/2011 11:17 AM, marius zbihlei wrote:
Well, indeed that's a ugly hack. As str's are not 0 terminated, you
either copy to a buffer length+1 or you overwrite the memory location
next to the end and revert the change afterwards. Not the cleanest way
to do it, but from a performance point of view, it make sense. If should
be taken extra care to ensure that on every code path the backup is
restored.
Yes. My argument was that it absolutely does not work in two cases:
- the memory allocated for the string is exactly the strings length
(you'd be overwriting random places with the string terminator)
- the same memory area is used for two overlapping strings: changing
one byte would change all string references using that area
Though, in this instance the second item does not seem possible. I'm not
sure if the first item can happen or not.
On 04/01/2011 10:38 AM, Timo Teräs wrote:
static buffers are not nice solution and can break depending on the call
chain. However, since if it's going to be deep copied anyway, it'll
probably work here. I was not sure about the deep copying part, so I
hesitated to suggest it as possible fix.
On 04/01/2011 10:29 AM, marius zbihlei wrote:
Good call. I truly belive that should a "static" keyword in front of
the
buffer declaration. Can you test it. (because in the cfg you usually
assign that transformation result to another pvar, the buffer will be
deep copied soon enough). On 03/31/2011 07:52 PM, Timo Teräs wrote:
The first really evil programming mistake I found is at:
modules_k/textops/txtvar.c:tr_txt_eval_re()
In line 86, it will fills the transformed variable with tr_txt_buf and
returns that. However, that is a stack variable, which will get
corrupted pretty soon after function return. I'm not sure how to fix
this properly.
diff --git a/modules_k/textops/txt_var.c
b/modules_k/textops/txt_var.c
index ae1fce6..a13e5e1 100644
--- a/modules_k/textops/txt_var.c
+++ b/modules_k/textops/txt_var.c
@@ -41,7 +41,7 @@ int tr_txt_eval_re(struct sip_msg *msg, tr_param_t
*tp, int subtype,
int nmatches;
str* result;
#define TR_TXT_BUF_SIZE 2048
- char tr_txt_buf[TR_TXT_BUF_SIZE];
+ static char tr_txt_buf[TR_TXT_BUF_SIZE];
if(val==NULL || (!(val->flags&PV_VAL_STR)) || val->rs.len<=0)
return -1;
Yes, I tried this actually earlier right after sending my previous mail.
It seems that this cured all the remaining valgrind errors I was seeing.
So yes, this should definitely pushed to all applicable git repos/branches.
> ==22274== Invalid read of size 1
> ==22274== at 0x6DA2C0A: pv_set_ruri_user (pv_core.c:1755)
> ==22274== by 0x13BA15: lval_pvar_assign (lvalue.c:357)
> ==22274== by 0x13BF0F: lval_assign (lvalue.c:405)
> ==22274== by 0x139E4D: do_action (action.c:1472)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x12FDEB: do_action (action.c:711)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
> ==22274== by 0x13A86F: run_actions_safe (action.c:1607)
> ==22274== by 0x1E2666: rval_get_int (rvalue.c:904)
> ==22274== by 0x1E5252: rval_expr_eval_int (rvalue.c:1866)
> ==22274== by 0x131B8A: do_action (action.c:1071)
> ==22274== by 0x13A6F1: run_actions (action.c:1555)
>
> Is fairly obvious. modules_k/pv/pv_core.c has several places which take
> backup of "val->rs.s[val->rs.len]" and replaces it with zero for
string
> termination. It's than later on set back to the original value.
> However,
> it would appear that the value passed does not always point to a memory
> area which is large enough. This results in multiple invalid reads and
> writes of one.
>
I fixed this by making a temporary copy of the string data to local
stack variable and null terminating it there. That made valgrind happy.
So obviously pv_set_ruri_user and others, are called with strings that
are allocated for the exact length, and the hacky zero-termination
cannot be used in this module.
>>> The remaining reads are not so clear to
me.
>>>
> I'm analysing the rest problems. Seems that they are almost always
> related to using re.subst, which has the problem mentioned in the
> beginning of the mail.
Yes. Looks like all the nags were related to usage re.subst results.
Thanks,
Timo
1:33 p.m.
On 04/01/2011 12:53 PM, Timo Teräs wrote:
On 04/01/2011 11:17 AM, marius zbihlei wrote:
static buffers are not nice solution and can break depending on the call
chain. However, since if it's going to be deep copied anyway, it'll
probably work here. I was not sure about the deep copying part, so I
hesitated to suggest it as possible fix.
Hello,
I pushed the patch to master and 3.1.
Yes, I tried this actually earlier right after sending my previous mail.
It seems that this cured all the remaining valgrind errors I was seeing.
So yes, this should definitely pushed to all applicable git repos/branches.
Yes. My argument was that it absolutely does not work in two cases:
- the memory allocated for the string is exactly the strings length
(you'd be overwriting random places with the string terminator)
- the same memory area is used for two overlapping strings: changing
one byte would change all string references using that area
Though, in this instance the second item does not seem possible. I'm not
sure if the first item can happen or not.
Well, I think the first case is what happens. it will not overwrite
random places, it will place a \0 at the end of the allocated
buffer(even the memory location is not part of that buffer). Because
buffer[length] points to some memory in pkg memory pool, accessing it
will not cause an access violation (it's not an address in .txt or .ro
segments) . The only problem I forsee is that during the set of \0 and
the revert to the original value, the address location would be accessed
in some other way, but I doubt it (keep in mind that locations have to
be properly aligned - so i presume in most cases a padding byte will be
overwritten). Not a good practice, but if no bugs are present I don't
recommend changing it.
Cheers,
Marius
1:48 p.m.
On 04/01/2011 01:33 PM, marius zbihlei wrote:
On 04/01/2011 12:53 PM, Timo Teräs wrote:
What if that byte is actually pkg memory pool management data?
That breaks memory allocation which is done when doing deep copy?
But yes, the re.subst fix alone seems to have removed all the valgrind
reported issues.
- Timo
Yes. My argument was that it absolutely does not
work in two cases:
- the memory allocated for the string is exactly the strings length
(you'd be overwriting random places with the string terminator)
- the same memory area is used for two overlapping strings: changing
one byte would change all string references using that area
Though, in this instance the second item does not seem possible. I'm not
sure if the first item can happen or not.
Well, I think the first case is what happens. it will not overwrite
random places, it will place a \0 at the end of the allocated
buffer(even the memory location is not part of that buffer). Because
buffer[length] points to some memory in pkg memory pool, accessing it
will not cause an access violation (it's not an address in .txt or .ro
segments) . The only problem I forsee is that during the set of \0 and
the revert to the original value, the address location would be accessed
in some other way, but I doubt it (keep in mind that locations have to
be properly aligned - so i presume in most cases a padding byte will be
overwritten). Not a good practice, but if no bugs are present I don't
recommend changing it. 
11:35 a.m.
On Friday 01 April 2011, Timo Teräs wrote:
[...]
There are valgrind hooks for memory managers. Should probably add them,
so it becomes aware of the things. However, those should be only needed
for memory pools. Ultimately kamailio uses malloc or mmap to allocate
the memory and valgrind knows both.
Hi Timo,
as a small side note: i think you could switch the PGK internal memory manager
to use system malloc instead of its pool - maybe this makes it easier for
valgrind. At least in 1.5 this was present, e.g.:
https://openser.svn.sourceforge.net/svnroot/openser/branches/1.5/mem/mem.h
(SYSTEM_MALLOC)
Cheers,
Henning
4984
days inactive
4985
days old
7 comments
3 participants
participants (3)
-
Henning Westerholt -
marius zbihlei -
Timo Teräs