Re: Linux 2.6.24-rc7 Build-Failure at __you_cannot_kmalloc_that_much

[Jean Delvare <khali@linux-fr.org>]

Hi Andrew, hi Chritoph,

On Mon, 7 Jan 2008 11:38:31 -0800, Andrew Morton wrote:
> On Mon, 7 Jan 2008 10:31:53 -0800 (PST)
> Christoph Lameter <clameter@sgi.com> wrote:
> 
> > On Mon, 7 Jan 2008, Andrew Morton wrote:
> > 
> > > > : undefined reference to `__you_cannot_kmalloc_that_much'
> > 
> > There is also a kernel.org bugzilla for this at
> > 
> > http://bugzilla.kernel.org/show_bug.cgi?id=9669
> > For some reason my adds to this do not show up.
> > 
> > In both cases we have a 
> > 
> > k(z/m)alloc(sizeof(*pointer), ...)
> > 
> > that is for some reason failing. I guess what happens is that the function 
> > in which this occurs is too complex for gcc 3.2. Thus it stops constant 
> > folding the sizeof(*pointer) in the complex inline-if-cascade that SLAB 
> > needs to determine the cache and does not eliminate the 
> > __you_cannot_kmalloc_that_much branch().

Interesting theory... So I tried to split half of the code of
dmi_id_init() to a subfunction and bingo! gcc 3.2.3 is now able to
build it properly. Thanks for the hint!

> > SLUB in that case just puts a series of if comparisions in the code. This 
> > means compilation does not fail but a large amount of code is generated.
> 
> ug.  Silent and nasty.
> 
> > We could replace the __you_cannot_kmalloc_that_much() with a BUG() 
> > statement so we have the same effect in SLAB?
> 
> I think it'd be better to just put suitable workarounds at the offending
> callsites.  We've only seen three or four of them in several months.

Here's a workaround for dmi-id.

Subject: Fix for __you_cannot_kmalloc_that_much failure in dmi-id

gcc 3.2 has a hard time coping with the code in dmi_id_init():

drivers/built-in.o(.init.text+0x789e): In function `dmi_id_init':
: undefined reference to `__you_cannot_kmalloc_that_much'
make: *** [.tmp_vmlinux1] Error 1

Moving half of the code to a separate function seems to help. This is
a no-op for gcc 4.1 which will successfully inline the code anyway.

Signed-off-by: Jean Delvare <khali@linux-fr.org>
---
 drivers/firmware/dmi-id.c |   19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

--- linux-2.6.24-rc7.orig/drivers/firmware/dmi-id.c	2007-10-24 09:59:28.000000000 +0200
+++ linux-2.6.24-rc7/drivers/firmware/dmi-id.c	2008-01-08 10:32:00.000000000 +0100
@@ -175,12 +175,11 @@ static struct device *dmi_dev;
 
 extern int dmi_available;
 
-static int __init dmi_id_init(void)
+/* In a separate function to keep gcc 3.2 happy - do NOT merge this in
+   dmi_id_init! */
+static void __init dmi_id_init_attr_table(void)
 {
-	int ret, i;
-
-	if (!dmi_available)
-		return -ENODEV;
+	int i;
 
 	/* Not necessarily all DMI fields are available on all
 	 * systems, hence let's built an attribute table of just
@@ -205,6 +204,16 @@ static int __init dmi_id_init(void)
 	ADD_DMI_ATTR(chassis_serial,    DMI_CHASSIS_SERIAL);
 	ADD_DMI_ATTR(chassis_asset_tag, DMI_CHASSIS_ASSET_TAG);
 	sys_dmi_attributes[i++] = &sys_dmi_modalias_attr.attr;
+}
+
+static int __init dmi_id_init(void)
+{
+	int ret;
+
+	if (!dmi_available)
+		return -ENODEV;
+
+	dmi_id_init_attr_table();
 
 	ret = class_register(&dmi_class);
 	if (ret)

I'll now check if I can do something similar for snd-mixer-oss.

-- 
Jean Delvare