Skip to contentSkip to conversation

Compile failure fix for ppc on 2.6.17-rc4-mm3 (2nd attempt)

PATCH In review Mel Gorman 4 participants 9 messages May 26, 2006 – Jun 13, 2006 PowerPC

lore
Share & tools
lore

Mel Gorman

author
#0 lore raw

PATCH 1/1 Compile failure fix for ppc on 2.6.17-rc4-mm3 (2nd attempt)

(Resending with Andrew's email address correct this time)

For the last few -mm releases, kernels built for an old RS6000 failed to compile with the message;

arch/powerpc/kernel/built-in.o(.init.text+0x77b4): In function `vrsqrtefp':
: undefined reference to __udivdi3' arch/powerpc/kernel/built-in.o(.init.text+0x7800): In function vrsqrtefp':
: undefined reference to `__udivdi3'
make: *** [.tmp_vmlinux1] Error 1

2.6.17-rc5 is not affected but I didn't search for the culprit patch in -mm. The following patch adds an implementation of __udivdi3 for plain old ppc32. This may not be the correct fix as Google tells me that __udivdi3 has been replaced by calls to do_div() in a number of cases. There was no obvious way to do that for vrsqrtefp, hence this workaround. The patch should be acked, rejected or replaced by a ppc expert.

arch/powerpc/lib/Makefile +1 -1
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.17-rc5-clean/arch/powerpc/lib/Makefile linux-2.6.17-rc5-udivdi3_ppc/arch/powerpc/lib/Makefile
@@ -4,7 +4,7 @@
 
 ifeq ($(CONFIG_PPC_MERGE),y)
 obj-y			:= string.o strcase.o
-obj-$(CONFIG_PPC32)	+= div64.o copy_32.o checksum_32.o
+obj-$(CONFIG_PPC32)	+= div64.o copy_32.o checksum_32.o udivdi3.o
 endif
 
 obj-y			+= bitops.o
arch/powerpc/lib/udivdi3.c +15 -0
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.17-rc5-clean/arch/powerpc/lib/udivdi3.c linux-2.6.17-rc5-udivdi3_ppc/arch/powerpc/lib/udivdi3.c
@@ -0,0 +1,15 @@
+/*
+ * Simple __udivdi3 function which doesn't use FPU.
+ */
+
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <asm-generic/div64.h>
+
+u64 __udivdi3(u64 n, u64 d)
+{
+	if (d & ~0xffffffff)
+		panic("Need true 64-bit/64-bit division");
+	return __div64_32(&n, (u32)d);
+}
+
Signature · 3 lines
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
#1 lore raw
Quoted 11 lines · [email protected] (Mel Gorman) wrote:

(Resending with Andrew's email address correct this time)

For the last few -mm releases, kernels built for an old RS6000 failed to compile with the message;

arch/powerpc/kernel/built-in.o(.init.text+0x77b4): In function `vrsqrtefp':
: undefined reference to __udivdi3' arch/powerpc/kernel/built-in.o(.init.text+0x7800): In function vrsqrtefp':
: undefined reference to `__udivdi3'
make: *** [.tmp_vmlinux1] Error 1

A function with a name like that doesn't _deserve_ to compile.

But actually vrsqrtefp() doesn't call __udivdi3 - the error lies somewhere else in the kernel and the toolchain gets it wrong, so we don't know where.

The way I usually hunt this problem down is to build the .s files (make arch/powerpc/kernel/foo.s) and then grep around, find the offending C function.

If the problem is specific to powerpc then a

	diffstat 2.6.17.rc4-mm3 | grep powerpc

will narrow down the number of files to be searched by rather a lot.

Quoted 6 lines

2.6.17-rc5 is not affected but I didn't search for the culprit patch in -mm. The following patch adds an implementation of __udivdi3 for plain old ppc32. This may not be the correct fix as Google tells me that __udivdi3 has been replaced by calls to do_div() in a number of cases. There was no obvious way to do that for vrsqrtefp, hence this workaround. The patch should be acked, rejected or replaced by a ppc expert.

Yes, we've traditionally avoided adding the 64-bit divide library functions.

author
#2 lore raw

On (26/05/06 09:49), Andrew Morton didst pronounce:

Quoted 7 lines
Quoted 6 lines · [email protected] (Mel Gorman) wrote:

(Resending with Andrew's email address correct this time)

For the last few -mm releases, kernels built for an old RS6000 failed to compile with the message;

Quoted 8 lines
Quoted 5 lines

arch/powerpc/kernel/built-in.o(.init.text+0x77b4): In function `vrsqrtefp':
: undefined reference to __udivdi3' arch/powerpc/kernel/built-in.o(.init.text+0x7800): In function vrsqrtefp':
: undefined reference to `__udivdi3'
make: *** [.tmp_vmlinux1] Error 1

A function with a name like that doesn't _deserve_ to compile.

heh

Quoted 3 lines

But actually vrsqrtefp() doesn't call __udivdi3 - the error lies somewhere else in the kernel and the toolchain gets it wrong, so we don't know where.

It explains why I couldn't find where vrsqrtefp() was.

Quoted 4 lines

The way I usually hunt this problem down is to build the .s files (make arch/powerpc/kernel/foo.s) and then grep around, find the offending C function.

That was a good idea, thanks. It showed that check_for_io_childs() in arch/powerpc/kernel/pci_32.c was the real problem. A quick look showed that is was doing divisions on resource_type_t which was 64 bits with my .config. Selecting CONFIG_RESOURCES_32BIT made the problem go away. This option is introduced by the kconfigurable-resources-* patches so I've cc'd Vivek Goyal to comment on the potential fixes below.

Quoted 8 lines
Quoted 6 lines

2.6.17-rc5 is not affected but I didn't search for the culprit patch in -mm. The following patch adds an implementation of __udivdi3 for plain old ppc32. This may not be the correct fix as Google tells me that __udivdi3 has been replaced by calls to do_div() in a number of cases. There was no obvious way to do that for vrsqrtefp, hence this workaround. The patch should be acked, rejected or replaced by a ppc expert.

Yes, we've traditionally avoided adding the 64-bit divide library functions.

Understandable. The fixes are obvious in this case. Easiest is for users to set CONFIG_RESOURCES_32BIT manually although the kernel build error is not very obvious. A fairly easy bodge is to default CONFIG_RESOURCES_32BIT to Y when building PPC32 but a determined user may still fail to build a kernel. A slightly riskier option is this patch that replaces a potential 64 bit divide with a do_div call. The patch has been successfully boot-tested on a RS6000.

arch/powerpc/kernel/pci_32.c +3 -1
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.17-rc4-mm3-clean/arch/powerpc/kernel/pci_32.c linux-2.6.17-rc4-mm3-resources-32/arch/powerpc/kernel/pci_32.c
@@ -1115,7 +1115,9 @@ check_for_io_childs(struct pci_bus *bus,
 	int	rc = 0;
 
 #define push_end(res, size) do { unsigned long __sz = (size) ; \
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	resource_size_t farEnd = (res->end + __sz); \
+	do_div(farEnd, (__sz + 1)); \
+	res->end = farEnd * (__sz + 1) + __sz; \
     } while (0)
 
 	list_for_each_entry(dev, &bus->devices, bus_list) {
#3 lore raw
Quoted 9 lines
Quoted 9 lines
Quoted 7 lines

arch/powerpc/kernel/built-in.o(.init.text+0x77b4): In function
`vrsqrtefp':
: undefined reference to `__udivdi3' arch/powerpc/kernel/built-in.o(.init.text+0x7800): In function
`vrsqrtefp':
: undefined reference to `__udivdi3'
make: *** [.tmp_vmlinux1] Error 1

A function with a name like that doesn't _deserve_ to compile.

Would vector_reciprocal_square_root_estimate_floating_point() be any better... Anyway, this is just a machine insn mnemonic, so the function name is fine I believe.

arch/powerpc/kernel/pci_32.c check_for_io_childs(struct pci_bus *bus, · line 1121

Quoted 6 lines

#define push_end(res, size) do { unsigned long __sz = (size) ; \

arch/powerpc/kernel/pci_32.c +3 -1
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	resource_size_t farEnd = (res->end + __sz); \
+	do_div(farEnd, (__sz + 1)); \
+	res->end = farEnd * (__sz + 1) + __sz; \
     } while (0)

Size here is a) a misnomer (size + 1 is the actual size) and b) always a power of two minus one. So instead, do

#define push_end(res, mask) res->end = -(-res->end & ~(unsigned long)mask)

(with a do { } while(0) armour if you prefer).

Segher

author
#4 lore raw
Quoted 24 lines · Segher Boessenkool
Quoted 9 lines
Quoted 9 lines
Quoted 7 lines

arch/powerpc/kernel/built-in.o(.init.text+0x77b4): In function
`vrsqrtefp':
: undefined reference to `__udivdi3' arch/powerpc/kernel/built-in.o(.init.text+0x7800): In function
`vrsqrtefp':
: undefined reference to `__udivdi3'
make: *** [.tmp_vmlinux1] Error 1

A function with a name like that doesn't _deserve_ to compile.

Would vector_reciprocal_square_root_estimate_floating_point() be any better... Anyway, this is just a machine insn mnemonic, so the function name is fine I believe.

arch/powerpc/kernel/pci_32.c check_for_io_childs(struct pci_bus *bus, · line 1121

Quoted 6 lines

#define push_end(res, size) do { unsigned long __sz = (size) ; \

arch/powerpc/kernel/pci_32.c +3 -1
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	resource_size_t farEnd = (res->end + __sz); \
+	do_div(farEnd, (__sz + 1)); \
+	res->end = farEnd * (__sz + 1) + __sz; \
     } while (0)

Size here is a) a misnomer (size + 1 is the actual size) and b) always a power of two minus one. So instead, do

Quoted 4 lines

#define push_end(res, mask) res->end = -(-res->end & ~(unsigned long)mask)

(with a do { } while(0) armour if you prefer).

It's not doing the same as the old code so is your suggested fix a correct replacement?

For example, given 0xfff for size the current code rounds res->end to the next 0x1000 boundary and adds 0xfff. Your propose fix just rounds it to the next 0x1000 if I'm reading it correctly but is what the code was meant to do in the first place? Using masks, the equivilant of the current code is something like;

#define push_end(res, mask) do { \
	res->end = -(-res->end & ~(unsigned long)mask); \
	res->end += mask; \
} while (0)
Signature · 3 lines
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
#5 lore raw
Quoted 15 lines
Quoted 10 lines

arch/powerpc/kernel/pci_32.c check_for_io_childs(struct pci_bus *bus, · line 1121

Quoted 6 lines

#define push_end(res, size) do { unsigned long __sz = (size) ; \

arch/powerpc/kernel/pci_32.c +3 -1
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	resource_size_t farEnd = (res->end + __sz); \
+	do_div(farEnd, (__sz + 1)); \
+	res->end = farEnd * (__sz + 1) + __sz; \
     } while (0)

Size here is a) a misnomer (size + 1 is the actual size) and b) always a power of two minus one. So instead, do

Quoted 4 lines

#define push_end(res, mask) res->end = -(-res->end & ~(unsigned long)mask)

(with a do { } while(0) armour if you prefer).

Quoted 8 lines

It's not doing the same as the old code so is your suggested fix a correct replacement?

For example, given 0xfff for size the current code rounds res->end to the next 0x1000 boundary and adds 0xfff. Your propose fix just rounds it to the next 0x1000 if I'm reading it correctly but is what the code was meant to do in the first place? Using masks, the equivilant of the current code is something like;

#define push_end(res, mask) do { \
	res->end = -(-res->end & ~(unsigned long)mask); \
	res->end += mask; \
} while (0)

Yeah forgot a bit, this looks fine.

Segher

author
#6 lore raw

On (29/05/06 19:56), Segher Boessenkool didst pronounce:

Quoted 26 lines
Quoted 26 lines
Quoted 11 lines

Review

Quoted 6 lines

#define push_end(res, size) do { unsigned long __sz = (size) ; \

(quoted diff) +3 -1
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	resource_size_t farEnd = (res->end + __sz); \
+	do_div(farEnd, (__sz + 1)); \
+	res->end = farEnd * (__sz + 1) + __sz; \
    } while (0)

Size here is a) a misnomer (size + 1 is the actual size) and b) always a power
of two minus one. So instead, do

Quoted 5 lines

#define push_end(res, mask) res->end = -(-res->end & ~(unsigned long)mask)

(with a do { } while(0) armour if you prefer).

It's not doing the same as the old code so is your suggested fix a correct replacement?

For example, given 0xfff for size the current code rounds res->end to the next 0x1000 boundary and adds 0xfff. Your propose fix just rounds it to the next 0x1000 if I'm reading it correctly but is what the code was meant to do in the first place? Using masks, the equivilant of the current code is something like;

Quoted 7 lines
Quoted 4 lines
#define push_end(res, mask) do { \
	res->end = -(-res->end & ~(unsigned long)mask); \
	res->end += mask; \
} while (0)

Yeah forgot a bit, this looks fine.

Ok. The following patch has been successfully boot tested on the same machine with 64 bit resource_size_t. Because the mask is assumed to be a power-of-two - 1, I added a comment and a BUG_ON() to be sure.

arch/powerpc/kernel/pci_32.c +10 -2
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.17-rc4-mm3-clean/arch/powerpc/kernel/pci_32.c linux-2.6.17-rc4-mm3-resources-32/arch/powerpc/kernel/pci_32.c
@@ -1114,8 +1114,16 @@ check_for_io_childs(struct pci_bus *bus,
 	int	i;
 	int	rc = 0;
 
-#define push_end(res, size) do { unsigned long __sz = (size) ; \
-	res->end = ((res->end + __sz) / (__sz + 1)) * (__sz + 1) + __sz; \
+	/*
+	 * Assuming mask is a power of two - 1, push_end
+	 * moves res->end to the end of the next
+	 * mask-aligned boundary.
+	 * e.g. res->end of 0x1fff moves to 0x2fff
+	 */
+#define push_end(res, mask) do { \
+	BUG_ON( ((mask+1) & mask) != 0); \
+	res->end = -(-res->end & ~(unsigned long)mask); \
+	res->end += mask; \
     } while (0)
 
 	list_for_each_entry(dev, &bus->devices, bus_list) {
#7 lore raw

arch/powerpc/kernel/pci_32.c check_for_io_childs(struct pci_bus *bus, · line 1126

Quoted 2 lines · Mel Gorman
arch/powerpc/kernel/pci_32.c +2 -0
+	res->end = -(-res->end & ~(unsigned long)mask); \
+	res->end += mask; \

I think this is equivalent to

	res->end = (res->end + mask) | mask;

and I have to say the latter seems more understandable to me (and doesn't need a cast) ...

Regards,
Paul.

author
#8 lore raw

On (09/06/06 19:36), Paul Mackerras didst pronounce:

Quoted 7 lines

Review

Quoted 2 lines · Mel Gorman
(quoted diff) +2 -0
+	res->end = -(-res->end & ~(unsigned long)mask); \
+	res->end += mask; \

I think this is equivalent to

Quoted 5 lines
	res->end = (res->end + mask) | mask;

and I have to say the latter seems more understandable to me (and doesn't need a cast) ...

Makes sense. The patch on top of the latest -mm is below if you want to push it. It has not been boot tested as building with the latest -mm is broken at the moment for older powerpcs because of the git-powerpc.patch patch . That patch assumes that mm_context_t is a struct with a vdso_base which is not always the case (it's an unsigned long in include/asm-ppc/mmu.h).

arch/powerpc/kernel/pci_32.c +2 -3
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.17-rc6-mm2-clean/arch/powerpc/kernel/pci_32.c linux-2.6.17-rc6-mm2-resources/arch/powerpc/kernel/pci_32.c
@@ -1121,9 +1121,8 @@ check_for_io_childs(struct pci_bus *bus,
 	 * e.g. res->end of 0x1fff moves to 0x2fff
 	 */
 #define push_end(res, mask) do {				\
-	BUG_ON(((mask+1) & mask) != 0);			\
-	res->end = -(-res->end & ~(unsigned long)mask);		\
-	res->end += mask;					\
+	BUG_ON(((mask+1) & mask) != 0);				\
+	res->end = (res->end + mask) | mask;			\
     } while (0)
 
 	list_for_each_entry(dev, &bus->devices, bus_list) {
Signature · 3 lines
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab