fix up generic csum_ipv6_magic function prototype
Subsystems (3)
PATCH 1/1 fix up generic csum_ipv6_magic function prototype
The generic version of csum_ipv6_magic has the len argument declared as __u16, while most arch dependent version declare it as __u32. After looking at the call site of this function, I come up to a conclusion that __u32 is a better match with the actual usage.
Hence, patch to change argument type for greater consistency.
---
asm-m32r and asm-parisc both have it declared as __u16. I think it is a copy-n-paste error and needs to be fixed by arch maintainer, I hope.
include/net/ip6_checksum.h renamed from include/net/ip6_checksum.h.orig +1 -1
@@ -36,7 +36,7 @@static __inline__unsigned short int csum_ipv6_magic (struct in6_addr *saddr,struct in6_addr *daddr, -__u16 len, +__u32 len,unsigned short proto,unsigned int csum) {
David Miller
maintainerFrom: "Chen, Kenneth W" <[email protected]>
Date: Wed, 8 Nov 2006 18:02:06 -0800
Quoted 8 lines
The generic version of
csum_ipv6_magichas the len argument declared as__u16, while most arch dependent version declare it as__u32. After looking at the call site of this function, I come up to a conclusion that__u32is a better match with the actual usage.Hence, patch to change argument type for greater consistency.
Signed-off-by Ken Chen <[email protected]>
Architecture implementations such as the ones for m32r and parisc have the same problem, so "for consistency" please fix them up as well.
Thanks a lot.
Quoted 8 lines · David Miller
From: "Chen, Kenneth W" <[email protected]>
Date: Wed, 8 Nov 2006 18:02:06 -0800Quoted 5 lines
The generic version of
csum_ipv6_magichas the len argument declared as__u16, while most arch dependent version declare it as__u32. After looking at the call site of this function, I come up to a conclusion that__u32is a better match with the actual usage.
Quoted 8 lines
Quoted 3 lines
Hence, patch to change argument type for greater consistency.
Signed-off-by Ken Chen <[email protected]>Architecture implementations such as the ones for m32r and parisc have the same problem, so "for consistency" please fix them up as well.
Thanks a lot.
Please, hold. One of the patches in my queue gets sanitized prototypes for all that stuff and it'll conflict like crazy.
I haven't touch that argument yet; if there's an agreement as to what should we switch to, I'll do that. So... does everyone agree that u32 is the way to go?
David Miller
maintainerFrom: Al Viro <[email protected]>
Date: Thu, 9 Nov 2006 07:22:16 +0000
Quoted 3 lines
I haven't touch that argument yet; if there's an agreement as to what should we switch to, I'll do that. So... does everyone agree that u32 is the way to go?
I definitely do.
Quoted 8 lines · David Miller
From: Al Viro <[email protected]>
Date: Thu, 9 Nov 2006 07:22:16 +0000Quoted 3 lines
I haven't touch that argument yet; if there's an agreement as to what should we switch to, I'll do that. So... does everyone agree that u32 is the way to go?
I definitely do.
OK. While we are at it, let's really sanitize the prototypes for that zoo.
General notes:
* I've added two types for checksums (__sum16 == checksum, __wsum == wide unfolded checksum). Those are fairly obvious.
* We have a bloody mess of what we do with pointers to data being checksumed - char *, unsigned char *, __user and const used inconsistently. I'm converting them all to pointers to (qualified) void, with const and __user where needed.
* struct in6_addr * in csum_ipv6_magic() should be pointer to const.
* IPv4 addresses should be passed as __be32.
* length in csum_ipv6_magic() should be u32.
After doing the above we have the following:
Consistent on all platforms:__sum16 ip_fast_csum(const void *, unsigned int);__sum16 ip_compute_csum(const void *, int);__sum16 csum_fold(__wsum);__wsum csum_partial_copy_from_user(const void __user *, void *, int, __wsum, int *); __wsum csum_partial_copy_nocheck(const void *, void *, int, __wsum);__wsum csum_and_copy_to_user(const void *, void *, int, __wsum, int *); __sum16 csum_ipv6_magic(struct in6_addr *, struct in6_addr *, __u32, unsigned short, __wsum);
Platform-dependent:__wsum csum_partial(const void *, T, __wsum);
On amd64 and uml/amd64 we have T = unsigned int, everywhere else it's int. __wsum csum_tcpudp_nofold(__be32, __be32, T1, T2, __wsum);
On arm/arm26: T1 = unsigned short, T2 = unsigned int.
On sparc/sparc64: T1 = unsigned int, T2 = unsigned short. Everywhere else: T1 = T2 = unsigned short.__sum16 csum_tcpudp_magic(__be32, __be32, unsigned short, T, __wsum); On arm/arm26 T is unsigned int, elsewhere it's unsigned short.
The first question is in the types we are using for length. OK, csum_tcpudp_...() is a special case; there we want u16 and unless there's a reason _not_ to do so on sparc, I'd rather convert it to the same thing.
Another special case is ip_fast_csum() (length in 32bit words, never more than 15, actually; existing code happily breaks for (never passed) large values).
The rest is a mess and should clearly be the same type. Do we want it to be signed? Several architectures do have checks for the bit 31 being set and demonstrate bloody odd behaviour in such cases.
Incidentally, WTF is
define SK_CS_CALCULATE_CHECKSUM
#ifndef CONFIG_X86_64
#define SkCsCalculateChecksum(p,l) ((~ip_compute_csum(p, l)) & 0xffff)
#else
#define SkCsCalculateChecksum(p,l) ((~ip_fast_csum(p, l)) & 0xffff)
#endif
in ./drivers/net/sk98lin/h/skdrv1st.h? I don't see any callers of that beast, anyway, but I'd really love to know who'd written it in the first place; the meaning of the second argument is different in those two...
The next question: csum_tcpudp_...() 'proto' argument. What do we want there? u8? u16? u32? It always fits in u8 and many architectures rely on it never exceeding 255. It's always constant, BTW...
csum_partial_copy_fromuser(): Can die, only 3 targets have its rudiment and nothing in the tree uses it. ACK?
csum_partial_copy(). Rare alias for csum_partial_copy_nocheck(). Can die; all instances simply should be renamed to csum_partial_copy_nocheck. ACK?
Quoted 7 lines · Al Viro
After doing the above we have the following:
Platform-dependent:
__wsumcsum_tcpudp_nofold(__be32, __be32, T1, T2, __wsum); On arm/arm26: T1 = unsigned short, T2 = unsigned int.__sum16csum_tcpudp_magic(__be32, __be32, unsigned short, T, __wsum); On arm/arm26 T is unsigned int, elsewhere it's unsigned short.
Could both become unsigned short or unsigned int. Would prefer unsigned int on ARM, since otherwise the compiler generate code to truncate any variable "int"s to an unsigned short.
Signature · 3 lines
Russell King Linux kernel 2.6 ARM Linux - http://www.arm.linux.org.uk/ maintainer of: 2.6 Serial core
Quoted 9 lines · Russell King
Quoted 7 lines · Al Viro
After doing the above we have the following:
Platform-dependent:
__wsumcsum_tcpudp_nofold(__be32, __be32, T1, T2, __wsum); On arm/arm26: T1 = unsigned short, T2 = unsigned int.__sum16csum_tcpudp_magic(__be32, __be32, unsigned short, T, __wsum); On arm/arm26 T is unsigned int, elsewhere it's unsigned short.
Quoted 3 lines
Could both become unsigned short or unsigned int. Would prefer unsigned int on ARM, since otherwise the compiler generate code to truncate any variable "int"s to an unsigned short.
Frankly, as for the generated code... I'd expect s/htons(len)/len * 256/ and the same for proto in there to have much more effect. And it does give the same checksum...
Quoted 9 lines · Al Viro
Incidentally, WTF is
defineSK_CS_CALCULATE_CHECKSUM
#ifndefCONFIG_X86_64
#defineSkCsCalculateChecksum(p,l)((~ip_compute_csum(p, l)) &0xffff)
#else
#defineSkCsCalculateChecksum(p,l)((~ip_fast_csum(p, l)) &0xffff)
#endif
in./drivers/net/sk98lin/h/skdrv1st.h?
Looks totally bogus. The x86-64 ip_fast_csum is practically identical to the i386 version. Someone must have been asleep while reviewing that code.
But AFAIK sk98lin is scheduled for deletion anyways, to be replaced by sky*. Probably for good reasons. Hopefully soon.
-Andi
David Miller
maintainerFrom: Al Viro <[email protected]>
Date: Mon, 13 Nov 2006 08:52:23 +0000
Quoted 4 lines
The first question is in the types we are using for length. OK, csum_tcpudp_...() is a special case; there we want u16 and unless there's a reason _not_ to do so on sparc, I'd rather convert it to the same thing.
That's fine.
Quoted 2 lines
csum_partial_copy_fromuser(): Can die, only 3 targets have its rudiment and nothing in the tree uses it. ACK?
ACK.
Quoted 2 lines
csum_partial_copy(). Rare alias forcsum_partial_copy_nocheck(). Can die; all instances simply should be renamed tocsum_partial_copy_nocheck. ACK?
ACK.