Skip to contentSkip to conversation

d451564 breakage

DISCUSSION Russell King 4 participants 10 messages Nov 13, 2009 – Nov 14, 2009 Arm (32-bit)

lore
Share & tools
lore
#0 lore raw

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

mm/highmem.c: In function ‘debug_kmap_atomic’:
mm/highmem.c:436: error: ‘KM_NMI’ undeclared (first use in this function) mm/highmem.c:436: error: (Each undeclared identifier is reported only once mm/highmem.c:436: error: for each function it appears in.) mm/highmem.c:436: error: ‘KM_NMI_PTE’ undeclared (first use in this function) mm/highmem.c:443: error: ‘KM_IRQ_PTE’ undeclared (first use in this function) make[2]: *** [mm/highmem.o] Error 1
make[1]: *** [mm] Error 2

I'd prefer not to have to add these definitions just so that highmem debugging can work for two reasons:

1. either we allocate mappings for these which will never be used, which unnecessarily wastes precious virtual memory space.

2. we define them to be some other KM_* value and hope that they never get used. If they do get used, we'll never know by way of compiler error but could result in silent data corruption.

The only sane alternative I can see would be to define these as KM_TYPE_NR and either ensure that kmap_atomic() always fails for out-of-bounds kmap types (more preferable to catch problems but has a performance impact) or we have the kmap debugging detect this as well.

Any preferences?

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#1 lore raw
Quoted 25 lines · On Fri, 13 Nov 2009 15:11:19 +0000 Russell King <[email protected]> wrote:

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

mm/highmem.c: In function ___debug_kmap_atomic___:
mm/highmem.c:436: error: ___KM_NMI___ undeclared (first use in this function) mm/highmem.c:436: error: (Each undeclared identifier is reported only once mm/highmem.c:436: error: for each function it appears in.) mm/highmem.c:436: error: ___KM_NMI_PTE___ undeclared (first use in this function) mm/highmem.c:443: error: ___KM_IRQ_PTE___ undeclared (first use in this function) make[2]: *** [mm/highmem.o] Error 1
make[1]: *** [mm] Error 2

I'd prefer not to have to add these definitions just so that highmem debugging can work for two reasons:

1. either we allocate mappings for these which will never be used, which unnecessarily wastes precious virtual memory space.

2. we define them to be some other KM_* value and hope that they never get used. If they do get used, we'll never know by way of compiler error but could result in silent data corruption.

Quoted 6 lines

The only sane alternative I can see would be to define these as KM_TYPE_NR and either ensure that kmap_atomic() always fails for out-of-bounds kmap types (more preferable to catch problems but has a performance impact) or we have the kmap debugging detect this as well.

Any preferences?

Could we do something nasty with ifdefs?

In arch header:
#define KM_NMI KM_NMI

In generic code:
#ifdef KM_NMI
<stuff which references KM_NMI>
#endif

or similar?

#2 lore raw
Quoted 41 lines · Andrew Morton
Quoted 25 lines · On Fri, 13 Nov 2009 15:11:19 +0000 Russell King <[email protected]> wrote:

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

mm/highmem.c: In function ___debug_kmap_atomic___:
mm/highmem.c:436: error: ___KM_NMI___ undeclared (first use in this function) mm/highmem.c:436: error: (Each undeclared identifier is reported only once mm/highmem.c:436: error: for each function it appears in.) mm/highmem.c:436: error: ___KM_NMI_PTE___ undeclared (first use in this function) mm/highmem.c:443: error: ___KM_IRQ_PTE___ undeclared (first use in this function) make[2]: *** [mm/highmem.o] Error 1
make[1]: *** [mm] Error 2

I'd prefer not to have to add these definitions just so that highmem debugging can work for two reasons:

1. either we allocate mappings for these which will never be used, which unnecessarily wastes precious virtual memory space.

2. we define them to be some other KM_* value and hope that they never get used. If they do get used, we'll never know by way of compiler error but could result in silent data corruption.

Quoted 6 lines

The only sane alternative I can see would be to define these as KM_TYPE_NR and either ensure that kmap_atomic() always fails for out-of-bounds kmap types (more preferable to catch problems but has a performance impact) or we have the kmap debugging detect this as well.

Any preferences?

Could we do something nasty with ifdefs?

In arch header:
#define KM_NMI KM_NMI

Quoted 6 lines

In generic code:
#ifdef KM_NMI
<stuff which references KM_NMI>
#endif

or similar?

Yeah, that would be the simplest - and all of this will go away with Peter's stacked kmaps.

Ingo

#3 lore raw
In reply to Russell King
Quoted 8 lines · Russell King

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

mm/highmem.c: In function ???debug_kmap_atomic???:
mm/highmem.c:436: error: ???KM_NMI??? undeclared (first use in this function)

indeed - sorry.

Note that debug_kmap_atomic() will be removed in v2.6.33 so i'd suggest to just do the easy solution and add #ifndef dummy definitions to mm/highmem.c to cover ARM - we'll remove the whole cruft for good.

Btw., testing sidenote: i test the ARM defconfig and it didnt break there. Perhaps highmem is off in the ARM defconfig? It would be helpful if the ARM defconfig enabled highmem.

Thanks,

Ingo

#4 lore raw
Quoted 9 lines · Ingo Molnar
Quoted 6 lines · Russell King

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

Quoted 8 lines
Quoted 2 lines

mm/highmem.c: In function ???debug_kmap_atomic???:
mm/highmem.c:436: error: ???KM_NMI??? undeclared (first use in this function)

indeed - sorry.

Note that debug_kmap_atomic() will be removed in v2.6.33 so i'd suggest to just do the easy solution and add #ifndef dummy definitions to mm/highmem.c to cover ARM - we'll remove the whole cruft for good.

Actually, the following patch is probably the simplest solution.

Quoted 3 lines

Btw., testing sidenote: i test the ARM defconfig and it didnt break there. Perhaps highmem is off in the ARM defconfig? It would be helpful if the ARM defconfig enabled highmem.

Given that highmem on ARM is experimental, I'd rather not have it enabled in too many machine defconfigs as standard just yet.

However, enabling highmem on itself is not enough to show this breakage, you also need highmem debugging enabled.

 arch/arm/include/asm/kmap_types.h |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
arch/arm/include/asm/kmap_types.h +6 -0
index d16ec97..c019949 100644
@@ -22,4 +22,10 @@ enum km_type {
 	KM_TYPE_NR
 };
 
+#ifdef CONFIG_DEBUG_HIGHMEM
+#define KM_NMI		(-1)
+#define KM_NMI_PTE	(-1)
+#define KM_IRQ_PTE	(-1)
+#endif
+
 #endif
Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#5 lore raw
Quoted 39 lines · Russell King
Quoted 9 lines · Ingo Molnar
Quoted 6 lines · Russell King

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

Quoted 8 lines
Quoted 2 lines

mm/highmem.c: In function ???debug_kmap_atomic???:
mm/highmem.c:436: error: ???KM_NMI??? undeclared (first use in this function)

indeed - sorry.

Note that debug_kmap_atomic() will be removed in v2.6.33 so i'd suggest to just do the easy solution and add #ifndef dummy definitions to mm/highmem.c to cover ARM - we'll remove the whole cruft for good.

Actually, the following patch is probably the simplest solution.

Quoted 3 lines

Btw., testing sidenote: i test the ARM defconfig and it didnt break there. Perhaps highmem is off in the ARM defconfig? It would be helpful if the ARM defconfig enabled highmem.

Given that highmem on ARM is experimental, I'd rather not have it enabled in too many machine defconfigs as standard just yet.

However, enabling highmem on itself is not enough to show this breakage, you also need highmem debugging enabled.

 arch/arm/include/asm/kmap_types.h |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
arch/arm/include/asm/kmap_types.h +0 -0
index d16ec97..c019949 100644

arch/arm/include/asm/kmap_types.h enum km_type { · line 29

Quoted 8 lines
arch/arm/include/asm/kmap_types.h +5 -0
@@ -22,4 +22,10 @@ enum km_type {
 	KM_TYPE_NR
 };
 
+#ifdef CONFIG_DEBUG_HIGHMEM
+#define KM_NMI		(-1)
+#define KM_NMI_PTE	(-1)
+#define KM_IRQ_PTE	(-1)
+#endif

Please solve this in mm/highmem.c as Andrew suggested it - other architectures could be affected as well beyond ARM.

Ingo

#6 lore raw
Quoted 45 lines · Ingo Molnar
Quoted 42 lines · Russell King
Quoted 9 lines · Ingo Molnar
Quoted 6 lines · Russell King

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

Quoted 8 lines
Quoted 2 lines

mm/highmem.c: In function ???debug_kmap_atomic???:
mm/highmem.c:436: error: ???KM_NMI??? undeclared (first use in this function)

indeed - sorry.

Note that debug_kmap_atomic() will be removed in v2.6.33 so i'd suggest to just do the easy solution and add #ifndef dummy definitions to mm/highmem.c to cover ARM - we'll remove the whole cruft for good.

Actually, the following patch is probably the simplest solution.

Quoted 3 lines

Btw., testing sidenote: i test the ARM defconfig and it didnt break there. Perhaps highmem is off in the ARM defconfig? It would be helpful if the ARM defconfig enabled highmem.

Given that highmem on ARM is experimental, I'd rather not have it enabled in too many machine defconfigs as standard just yet.

However, enabling highmem on itself is not enough to show this breakage, you also need highmem debugging enabled.

 arch/arm/include/asm/kmap_types.h |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
arch/arm/include/asm/kmap_types.h +0 -0
index d16ec97..c019949 100644
@@ -22,4 +22,10 @@ enum km_type {
 	KM_TYPE_NR
 };
 
Quoted 8 lines

arch/arm/include/asm/kmap_types.h enum km_type { · line 29

Quoted 5 lines
arch/arm/include/asm/kmap_types.h +5 -0
+#ifdef CONFIG_DEBUG_HIGHMEM
+#define KM_NMI		(-1)
+#define KM_NMI_PTE	(-1)
+#define KM_IRQ_PTE	(-1)
+#endif

Please solve this in mm/highmem.c as Andrew suggested it - other architectures could be affected as well beyond ARM.

Have you actually put any thought into that approach? What you're suggesting means:

- adding preprocessor definitions to all existing places where these symbols are defined
- adding multiple complex ifdefs to mm/highmem.c thusly:

		if (
#ifdef KM_NMI
			type != KM_NMI
#else
			1
#endif
			&&
#ifdef KM_NMI_PTE
			type != KM_NMI_PTE
#else
			1
#endif
		) {
			WARN_ON(1);
			warn_count--;
		}

and I really don't think this is a realistic solution, even for the interim.

The other solution is to determine if KM_NMI/KM_NMI_PTE/KM_IRQ_PTE are x86 only (which does seem to be the case). If that is the case, moving the new definitions in my patch into mm/highmem.c itself and ifdef out the ones in asm-generic/km_types.h would seem like another possible approach.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#7 lore raw
Quoted 55 lines · Russell King
Quoted 45 lines · Ingo Molnar
Quoted 42 lines · Russell King
Quoted 9 lines · Ingo Molnar
Quoted 6 lines · Russell King

Change:

highmem: Fix debug_kmap_atomic() to also handle KM_IRQ_PTE, KM_NMI, and KM_NMI_PTE

Appears to break ARM:

Quoted 8 lines
Quoted 2 lines

mm/highmem.c: In function ???debug_kmap_atomic???:
mm/highmem.c:436: error: ???KM_NMI??? undeclared (first use in this function)

indeed - sorry.

Note that debug_kmap_atomic() will be removed in v2.6.33 so i'd suggest to just do the easy solution and add #ifndef dummy definitions to mm/highmem.c to cover ARM - we'll remove the whole cruft for good.

Actually, the following patch is probably the simplest solution.

Quoted 3 lines

Btw., testing sidenote: i test the ARM defconfig and it didnt break there. Perhaps highmem is off in the ARM defconfig? It would be helpful if the ARM defconfig enabled highmem.

Given that highmem on ARM is experimental, I'd rather not have it enabled in too many machine defconfigs as standard just yet.

However, enabling highmem on itself is not enough to show this breakage, you also need highmem debugging enabled.

 arch/arm/include/asm/kmap_types.h |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
arch/arm/include/asm/kmap_types.h +0 -0
index d16ec97..c019949 100644
@@ -22,4 +22,10 @@ enum km_type {
 	KM_TYPE_NR
 };
 
Quoted 8 lines

arch/arm/include/asm/kmap_types.h enum km_type { · line 29

Quoted 5 lines
arch/arm/include/asm/kmap_types.h +5 -0
+#ifdef CONFIG_DEBUG_HIGHMEM
+#define KM_NMI		(-1)
+#define KM_NMI_PTE	(-1)
+#define KM_IRQ_PTE	(-1)
+#endif

Please solve this in mm/highmem.c as Andrew suggested it - other architectures could be affected as well beyond ARM.

Quoted 1 line

Have you actually put any thought into that approach? [...]

Yes.

Quoted 17 lines

[...] What you're suggesting means:

- adding preprocessor definitions to all existing places where these symbols are defined
- adding multiple complex ifdefs to mm/highmem.c thusly:

		if (
#ifdef KM_NMI
			type != KM_NMI
#else
			1
#endif
			&&
#ifdef KM_NMI_PTE
			type != KM_NMI_PTE
#else
			1
Quoted 1 line

#endif

I'd NAK such ugly code.

What i was thinking of was a really simple thing in highmem.c (or rather, in include/linux/highmem.h, after kmap_types.h gets included):

#ifndef KM_NMI

# define KM_NMI			-1
#endif
#ifndef KM_NMI_PTE
# define KM_NMI_PTE		-1
#endif

etc.

Thanks,

Ingo

#8 lore raw
Quoted 11 lines · Ingo Molnar

What i was thinking of was a really simple thing in highmem.c (or rather, in include/linux/highmem.h, after kmap_types.h gets included):

#ifndef KM_NMI

# define KM_NMI			-1
#endif
#ifndef KM_NMI_PTE
# define KM_NMI_PTE		-1
#endif

etc.

That means we need to ensure that kmap_atomic() fails when these are used, to avoid kmap_atomic corrupting virtual mappings below the kmap area should one of these be used. That means additional run-time code in all those implementations - I'd much rather leave that to someone else to do.

Alternatively, they could be wrapped with CONFIG_DEBUG_HIGHMEM so they aren't normally visible.

In any case, as I've already pointed out, we still need to add definitions to all places which define these constants. Having them exist as enum constants does not make them visible to the preprocessor.

I don't think it's as simple as you're trying to make it. However, since you seem to have a solid idea which you're trying to get me to implement, maybe it'd just be much simpler and faster if you provided that patch.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#9 lore raw
Quoted 27 lines · Russell King
Quoted 11 lines · Ingo Molnar

What i was thinking of was a really simple thing in highmem.c (or rather, in include/linux/highmem.h, after kmap_types.h gets included):

#ifndef KM_NMI

# define KM_NMI			-1
#endif
#ifndef KM_NMI_PTE
# define KM_NMI_PTE		-1
#endif

etc.

That means we need to ensure that kmap_atomic() fails when these are used, to avoid kmap_atomic corrupting virtual mappings below the kmap area should one of these be used. That means additional run-time code in all those implementations - I'd much rather leave that to someone else to do.

Alternatively, they could be wrapped with CONFIG_DEBUG_HIGHMEM so they aren't normally visible.

In any case, as I've already pointed out, we still need to add definitions to all places which define these constants. Having them exist as enum constants does not make them visible to the preprocessor.

Quoted 4 lines

I don't think it's as simple as you're trying to make it. However, since you seem to have a solid idea which you're trying to get me to implement, maybe it'd just be much simpler and faster if you provided that patch.

Since we are going to remove the generic code (the patch for that has been posted) it doesnt really matter and i'd suggest for you to commit the ARM patch.

Ingo