Skip to contentSkip to conversation

Segmentation Fault with 'm' Dependencies

PATCH New Martin Walch 3 participants 5 messages Oct 28, 2013 – Nov 19, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hello,

this test case leads to a segmentation fault:

config A
tristate "A" if m

config MODULES
boolean "MODULES"
option modules

As you can see, the MODULES symbol with the option modules is declared after the first occurrence of an 'm' dependency. (Actually you can drop the MODULES section or use a different symbol name. It does not matter.) Internally 'm' gets converted into (symbol_mod && modules_sym), which adds a dependency on a bad symbol, finally leading to dereferencing a null pointer.

If you move the declaration of the MODULES symbol to the top, everything works fine.

The crash has been introduced last month with

Quoted 2 lines

6902dccfda005fa4c42410fa064fdd331ab42479
kconfig: do not special-case 'MODULES' symbol

However, things were probably broken before. The problem has only become visible.

The reason that configuring a Linux kernel does not crash the configuration system is that there is currently no architecture that has a symbol with a dependency on 'm' anywhere before the MODULES symbol.

Regards
Martin Walch

#1 lore raw

Martin, All,

On 2013-10-28 03:16 +0100, Martin Walch spake thusly:

Quoted 9 lines

this test case leads to a segmentation fault:

config A
tristate "A" if m

config MODULES
boolean "MODULES"
option modules

Quoted 5 lines

As you can see, the MODULES symbol with the option modules is declared after the first occurrence of an 'm' dependency. (Actually you can drop the MODULES section or use a different symbol name. It does not matter.) Internally 'm' gets converted into (symbol_mod && modules_sym), which adds a dependency on a bad symbol, finally leading to dereferencing a null pointer.

Indeed, reproduced here. I'll investigate further (although anyone is free to hack it, too! :-p)

Quoted 8 lines

If you move the declaration of the MODULES symbol to the top, everything works fine.

The crash has been introduced last month with

Quoted 2 lines

6902dccfda005fa4c42410fa064fdd331ab42479
kconfig: do not special-case 'MODULES' symbol

Quoted 6 lines

However, things were probably broken before. The problem has only become visible.

The reason that configuring a Linux kernel does not crash the configuration system is that there is currently no architecture that has a symbol with a dependency on 'm' anywhere before the MODULES symbol.

Regards,
Yann E. MORIN.

Signature · 6 lines
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'
#2 lore raw
Quoted 12 lines · Yann E. MORIN

Martin, All,

On 2013-10-28 03:16 +0100, Martin Walch spake thusly:

Quoted 9 lines

this test case leads to a segmentation fault:

config A
tristate "A" if m

config MODULES
boolean "MODULES"
option modules

Quoted 8 lines
Quoted 5 lines

As you can see, the MODULES symbol with the option modules is declared after the first occurrence of an 'm' dependency. (Actually you can drop the MODULES section or use a different symbol name. It does not matter.) Internally 'm' gets converted into (symbol_mod && modules_sym), which adds a dependency on a bad symbol, finally leading to dereferencing a null pointer.

Indeed, reproduced here. I'll investigate further (although anyone is free to hack it, too! :-p)

Hi Yann, all,

I had a look at the problem, Martin reported and found out that menu_check_dep() is causing the problem:

...

	case E_SYMBOL:
		/* change 'm' into 'm' && MODULES */
		if (e->left.sym == &symbol_mod)
			return expr_alloc_and(e, expr_alloc_symbol(modules_sym));
		break;

...

It generates an expression that uses modules_sym which is NULL at that time.

The problem seems to be that since commit 6902dccfda005fa modules_sym is NULL until an "option modules" is found or the default is set but part of the code needs a valid pointer at any time.

I played with other possible fixes but got the impression that these would add even more complicated code and I decided to propose the patch that I will send in a minute. That patch basically reverts commit 6902dccfda005fa keeping the changes that (as far as I understood) are the important part of that commit. The parser also needs to be regenerated but that should happen in a separate commit if I remember corretly. So, for now and until a review, I left it out.

I would be glad if you could have a look at the patch.

Dirk

Quoted 17 lines
Quoted 8 lines

If you move the declaration of the MODULES symbol to the top, everything works fine.

The crash has been introduced last month with

Quoted 2 lines

6902dccfda005fa4c42410fa064fdd331ab42479
kconfig: do not special-case 'MODULES' symbol

Quoted 6 lines

However, things were probably broken before. The problem has only become visible.

The reason that configuring a Linux kernel does not crash the configuration system is that there is currently no architecture that has a symbol with a dependency on 'm' anywhere before the MODULES symbol.

Regards,
Yann E. MORIN.

#3 lore raw

PATCH 1/1 scripts/kconfig: handle cases with 'option modules' defined after reference

Currently, the following test-case causes a segfault:

config A
tristate "A" if m

config MODULES
boolean "MODULES"
option modules

This is, because modules_sym is NULL until 'option modules' appears, since commit
6902dccfda (kconfig: do not special-case 'MODULES' symbol).

The above example causes an expression of type symbol being generated with NULL as the pointer to the symbol in menu_check_dep(), later causing a segfault when that pointer is dereferenced.

This patch basically reverts commit 6902dccfda keeping only the change that modules_sym references symbol 'n' as a default when 'option modules' is not used in the config files.

---
 scripts/kconfig/menu.c  |  9 ++++++---
 scripts/kconfig/zconf.y | 11 +++++++++--
 2 files changed, 15 insertions(+), 5 deletions(-)
scripts/kconfig/menu.c +6 -3
index db1512a..c4ec1ff 100644
@@ -198,15 +198,18 @@ void menu_add_symbol(enum prop_type type, struct symbol *sym, struct expr *dep)
 
 void menu_add_option(int token, char *arg)
 {
+	struct property *prop;
+
 	switch (token) {
 	case T_OPT_MODULES:
-		if (modules_sym)
+		if (modules_sym->prop)
 			zconf_error("symbol '%s' redefines option 'modules'"
 				    " already defined by symbol '%s'",
 				    current_entry->sym->name,
-				    modules_sym->name
+				    prop_get_symbol(modules_sym->prop)->name
 				    );
-		modules_sym = current_entry->sym;
+		prop = prop_alloc(P_DEFAULT, modules_sym);
+		prop->expr = expr_alloc_symbol(current_entry->sym);
 		break;
 	case T_OPT_DEFCONFIG_LIST:
 		if (!sym_defconfig_list)
scripts/kconfig/zconf.y +9 -2
index 0653886..941ec32 100644
@@ -493,6 +493,9 @@ void conf_parse(const char *name)
 
 	sym_init();
 	_menu_init();
+	modules_sym = sym_lookup(NULL, 0);
+	modules_sym->type = S_BOOLEAN;
+	modules_sym->flags |= SYMBOL_AUTO;
 	rootmenu.prompt = menu_add_prompt(P_MENU, "Linux Kernel Configuration", NULL);
 
 	if (getenv("ZCONF_DEBUG"))
@@ -500,8 +503,12 @@ void conf_parse(const char *name)
 	zconfparse();
 	if (zconfnerrs)
 		exit(1);
-	if (!modules_sym)
-		modules_sym = sym_find( "n" );
+	if (!modules_sym->prop) {
+		struct property *prop;
+
+		prop = prop_alloc(P_DEFAULT, modules_sym);
+		prop->expr = expr_alloc_symbol(sym_find("n"));
+	}
 
 	rootmenu.prompt->text = _(rootmenu.prompt->text);
 	rootmenu.prompt->text = sym_expand_string_value(rootmenu.prompt->text);
Signature · 1 line
1.8.4
#4 lore raw
In reply to Dirk Gouders

Dirk, All,

On 2013-11-19 21:57 +0100, Dirk Gouders spake thusly:

Quoted 31 lines
Quoted 10 lines · Yann E. MORIN

On 2013-10-28 03:16 +0100, Martin Walch spake thusly:

Quoted 9 lines

this test case leads to a segmentation fault:

config A
tristate "A" if m

config MODULES
boolean "MODULES"
option modules

Quoted 8 lines
Quoted 5 lines

As you can see, the MODULES symbol with the option modules is declared after the first occurrence of an 'm' dependency. (Actually you can drop the MODULES section or use a different symbol name. It does not matter.) Internally 'm' gets converted into (symbol_mod && modules_sym), which adds a dependency on a bad symbol, finally leading to dereferencing a null pointer.

Indeed, reproduced here. I'll investigate further (although anyone is free to hack it, too! :-p)

Hi Yann, all,

I had a look at the problem, Martin reported and found out that menu_check_dep() is causing the problem:

...

	case E_SYMBOL:
		/* change 'm' into 'm' && MODULES */
		if (e->left.sym == &symbol_mod)
			return expr_alloc_and(e, expr_alloc_symbol(modules_sym));
		break;
Quoted 8 lines

...

It generates an expression that uses modules_sym which is NULL at that time.

The problem seems to be that since commit 6902dccfda005fa modules_sym is NULL until an "option modules" is found or the default is set but part of the code needs a valid pointer at any time.

Indeed. The idea behind that patch was that we would post-pone allocating modules_sym until we either see the actual 'modules' option, or until the end of the parsing, since checking the value of the symbol is not needed until we actually display the menu.

But that's wrong, since we do need that symbol at the time of parsing, to generate the dependency list, as Martin discovered.

I missed that since no architecture in the kernel has any option that depends on 'MODULES' before it is defined in Kconfig.

Quoted 7 lines

I played with other possible fixes but got the impression that these would add even more complicated code and I decided to propose the patch that I will send in a minute. That patch basically reverts commit 6902dccfda005fa keeping the changes that (as far as I understood) are the important part of that commit. The parser also needs to be regenerated but that should happen in a separate commit if I remember corretly. So, for now and until a review, I left it out.

Yes, for a first review, it is OK not to regenerate it.
But on final submission, you should regenerate the parser in the same patch. It is a single, self-contained and semantically-related change, so it makes sense to group them in a single patch.

Quoted 1 line

I would be glad if you could have a look at the patch.

I've had a quick look, and it looks OK at first sight. I'll look into it more in depth in a moment, and will comment directly as a reply to the patch.

Thank you! :-)

Regards,
Yann E. MORIN.

Signature · 6 lines
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'