Exported symbols removed in 2.5.2

Exported symbols removed in 2.5.2

[Laurent Bigonville]

Hello,

In debian, during the build  of a package, we have a tool checking if 
symbols are removed from shared librearies.

With the 2.5.2 release, I get the following output:

--- debian/libauparse0.symbols (libauparse0_1:2.5.2-1_amd64)
+++ dpkg-gensymbolsmB_9P2    2016-05-03 15:01:10.606713310 +0000
@@ -1,7 +1,7 @@
  libauparse.so.0 libauparse0 #MINVER#
   au_terminate_all_events@Base 1:2.5.1
- audit_strsplit@Base 1:2.4.2
- audit_strsplit_r@Base 1:2.4.2
+#MISSING: 1:2.5.2-1# audit_strsplit@Base 1:2.4.2
+#MISSING: 1:2.5.2-1# audit_strsplit_r@Base 1:2.4.2
   auparse_add_callback@Base 1:2.2.1
   auparse_destroy@Base 1:2.2.1
   auparse_do_interpretation@Base 1:2.3.1
@@ -49,4 +49,4 @@
   ausearch_clear@Base 1:2.2.1
   ausearch_next_event@Base 1:2.2.1
   ausearch_set_stop@Base 1:2.2.1
- set_escape_mode@Base 1:2.4.4
+#MISSING: 1:2.5.2-1# set_escape_mode@Base 1:2.4.4
dpkg-gensymbols: warning: some symbols or patterns disappeared in the 
symbols file: see diff output below
dpkg-gensymbols: warning: debian/libaudit1/DEBIAN/symbols doesn't match 
completely debian/libaudit1.symbols
--- debian/libaudit1.symbols (libaudit1_1:2.5.2-1_amd64)
+++ dpkg-gensymbolsQATT_C    2016-05-03 15:01:10.802717308 +0000
@@ -56,7 +56,7 @@
   audit_rule_syscall_data@Base 1:2.2.1
   audit_rule_syscallbyname_data@Base 1:2.2.1
   audit_send@Base 1:2.2.1
- audit_send_user_message@Base 1:2.2.1
+#MISSING: 1:2.5.2-1# audit_send_user_message@Base 1:2.2.1
   audit_set_backlog_limit@Base 1:2.2.1
   audit_set_backlog_wait_time@Base 1:2.4.2
   audit_set_enabled@Base 1:2.2.1

Is that expected that these 4 symbols have been removed?

Cheers,

Laurent Bigonville

Re: Exported symbols removed in 2.5.2

[Steve Grubb]

On Tuesday, May 03, 2016 05:04:04 PM Laurent Bigonville wrote:
> Hello,
> 
> In debian, during the build  of a package, we have a tool checking if
> symbols are removed from shared librearies.
> 
> With the 2.5.2 release, I get the following output:
> 
> --- debian/libauparse0.symbols (libauparse0_1:2.5.2-1_amd64)
> +++ dpkg-gensymbolsmB_9P2    2016-05-03 15:01:10.606713310 +0000
> @@ -1,7 +1,7 @@
>   libauparse.so.0 libauparse0 #MINVER#
>    au_terminate_all_events@Base 1:2.5.1
> - audit_strsplit@Base 1:2.4.2
> - audit_strsplit_r@Base 1:2.4.2
> +#MISSING: 1:2.5.2-1# audit_strsplit@Base 1:2.4.2
> +#MISSING: 1:2.5.2-1# audit_strsplit_r@Base 1:2.4.2
>    auparse_add_callback@Base 1:2.2.1
>    auparse_destroy@Base 1:2.2.1
>    auparse_do_interpretation@Base 1:2.3.1
> @@ -49,4 +49,4 @@
>    ausearch_clear@Base 1:2.2.1
>    ausearch_next_event@Base 1:2.2.1
>    ausearch_set_stop@Base 1:2.2.1
> - set_escape_mode@Base 1:2.4.4
> +#MISSING: 1:2.5.2-1# set_escape_mode@Base 1:2.4.4
> dpkg-gensymbols: warning: some symbols or patterns disappeared in the
> symbols file: see diff output below
> dpkg-gensymbols: warning: debian/libaudit1/DEBIAN/symbols doesn't match
> completely debian/libaudit1.symbols
> --- debian/libaudit1.symbols (libaudit1_1:2.5.2-1_amd64)
> +++ dpkg-gensymbolsQATT_C    2016-05-03 15:01:10.802717308 +0000
> @@ -56,7 +56,7 @@
>    audit_rule_syscall_data@Base 1:2.2.1
>    audit_rule_syscallbyname_data@Base 1:2.2.1
>    audit_send@Base 1:2.2.1
> - audit_send_user_message@Base 1:2.2.1
> +#MISSING: 1:2.5.2-1# audit_send_user_message@Base 1:2.2.1
>    audit_set_backlog_limit@Base 1:2.2.1
>    audit_set_backlog_wait_time@Base 1:2.4.2
>    audit_set_enabled@Base 1:2.2.1
> 
> Is that expected that these 4 symbols have been removed?

Yes. This corresponds to the changelog entry:

- Revise function hiding technique to better protect audit ABI

All functions missing are internal to the audit libraries and could cause 
symbols collisions or worse if people start using them even though they are 
not declared in the library headers.

-Steve

Re: Exported symbols removed in 2.5.2

[Laurent Bigonville]

Le 03/05/16 à 17:10, Steve Grubb a écrit :
> On Tuesday, May 03, 2016 05:04:04 PM Laurent Bigonville wrote:
>> Hello,
>>
>> In debian, during the build  of a package, we have a tool checking if
>> symbols are removed from shared librearies.
>>
>> With the 2.5.2 release, I get the following output:
>>
>> --- debian/libauparse0.symbols (libauparse0_1:2.5.2-1_amd64)
>> +++ dpkg-gensymbolsmB_9P2    2016-05-03 15:01:10.606713310 +0000
>> @@ -1,7 +1,7 @@
>>    libauparse.so.0 libauparse0 #MINVER#
>>     au_terminate_all_events@Base 1:2.5.1
>> - audit_strsplit@Base 1:2.4.2
>> - audit_strsplit_r@Base 1:2.4.2
>> +#MISSING: 1:2.5.2-1# audit_strsplit@Base 1:2.4.2
>> +#MISSING: 1:2.5.2-1# audit_strsplit_r@Base 1:2.4.2
>>     auparse_add_callback@Base 1:2.2.1
>>     auparse_destroy@Base 1:2.2.1
>>     auparse_do_interpretation@Base 1:2.3.1
>> @@ -49,4 +49,4 @@
>>     ausearch_clear@Base 1:2.2.1
>>     ausearch_next_event@Base 1:2.2.1
>>     ausearch_set_stop@Base 1:2.2.1
>> - set_escape_mode@Base 1:2.4.4
>> +#MISSING: 1:2.5.2-1# set_escape_mode@Base 1:2.4.4
>> dpkg-gensymbols: warning: some symbols or patterns disappeared in the
>> symbols file: see diff output below
>> dpkg-gensymbols: warning: debian/libaudit1/DEBIAN/symbols doesn't match
>> completely debian/libaudit1.symbols
>> --- debian/libaudit1.symbols (libaudit1_1:2.5.2-1_amd64)
>> +++ dpkg-gensymbolsQATT_C    2016-05-03 15:01:10.802717308 +0000
>> @@ -56,7 +56,7 @@
>>     audit_rule_syscall_data@Base 1:2.2.1
>>     audit_rule_syscallbyname_data@Base 1:2.2.1
>>     audit_send@Base 1:2.2.1
>> - audit_send_user_message@Base 1:2.2.1
>> +#MISSING: 1:2.5.2-1# audit_send_user_message@Base 1:2.2.1
>>     audit_set_backlog_limit@Base 1:2.2.1
>>     audit_set_backlog_wait_time@Base 1:2.4.2
>>     audit_set_enabled@Base 1:2.2.1
>>
>> Is that expected that these 4 symbols have been removed?
> Yes. This corresponds to the changelog entry:
>
> - Revise function hiding technique to better protect audit ABI
>
> All functions missing are internal to the audit libraries and could cause
> symbols collisions or worse if people start using them even though they are
> not declared in the library headers.

In the private.h header file, I can see the following comment:

// This is the main messaging function used internally
// Don't hide it, it used to be a part of the public API!
extern int audit_send_user_message(int fd, int type, hide_t hide_err,
         const char *message);

So doesn't this warrant a soname bump then?

Re: Exported symbols removed in 2.5.2

[Steve Grubb]

On Tuesday, May 03, 2016 05:15:01 PM Laurent Bigonville wrote:
> >> +#MISSING: 1:2.5.2-1# audit_send_user_message@Base 1:2.2.1
> >>
> >>     audit_set_backlog_limit@Base 1:2.2.1
> >>     audit_set_backlog_wait_time@Base 1:2.4.2
> >>     audit_set_enabled@Base 1:2.2.1
> >>
> >> Is that expected that these 4 symbols have been removed?
> > 
> > Yes. This corresponds to the changelog entry:
> > 
> > - Revise function hiding technique to better protect audit ABI
> > 
> > All functions missing are internal to the audit libraries and could cause
> > symbols collisions or worse if people start using them even though they
> > are
> > not declared in the library headers.
> 
> In the private.h header file, I can see the following comment:
> 
> // This is the main messaging function used internally
> // Don't hide it, it used to be a part of the public API!
> extern int audit_send_user_message(int fd, int type, hide_t hide_err,
>          const char *message);
> 
> So doesn't this warrant a soname bump then?

The answer is not simple. It was a hidden symbol:

hidden_proto(audit_send_user_message);

But I noticed that this broke at some point because it was hidden in old 
releases but then suddenly started being visible. There has been no changes in 
the hiding technique since the 1.2 release. My guess is that something changed 
in gcc somewhere along the way that broke the hiding technique from Ulrich 
Drepper's DSO programming guidelines.

The function was part of the public API in the 1.0.16 release. It was 
deprecated/hidden in the 1.2 release which dates to 7-Apr 2006. Its been about 
10 years that the function prototype has not been in libaudit.h. I would hope 
that a missing prototype message would have been reported and fixed in the last 
10 years. I have personally fixed use of the symbol in everything I know of 10 
years ago.

The only problem people would have is in very old utilities they wrote a long 
time ago, or very old versions of shadow-utils/pam. I wrote a script that 
looks for that symbol in all elf files. I have to test on RHEL 4 to find the 
symbol in any programs. So, I think you have a valid concern, but its been so 
long that in practice it has worked itself out.

-Steve