2.6.4-mm1
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 10 lines · Andrew Morton
- The CPU scheduler changes in -mm (sched-domains) have been hanging about for too long. I had been hoping that the people who care about SMT and NUMA performance would have some results by now but all seems to be silent.
I do not wish to merge these up until the big-iron guys can say that they suit their requirements, with a reasonable expectation that we will not need to churn this code later in the 2.6 series.
So. If you have been testing, please speak up. If you have not been testing, please do so.
I tested them on Opteron NUMA systems and they are worse on simple tests than the stock scheduler (e.g. the parallelized STREAM test, which is a bit silly, but still fairly important)
For SMT there is a patch from Intel pending that teaches x86-64 to set up the SMT scheduler. They said they got slightly better benchmark results. The SMT setup seems to be racy though.
Some kind of SMT scheduler is definitely needed, we have a serious regression compared to 2.4 here right now. I'm not sure this is the right approach though, it seems to be far too complex.
-Andi
Quoted 11 lines · Andi Kleen
Quoted 8 lines · Andrew Morton
- The CPU scheduler changes in -mm (sched-domains) have been hanging about for too long. I had been hoping that the people who care about SMT and NUMA performance would have some results by now but all seems to be silent.
I do not wish to merge these up until the big-iron guys can say that they suit their requirements, with a reasonable expectation that we will not need to churn this code later in the 2.6 series.
Quoted 6 lines
Quoted 2 lines
So. If you have been testing, please speak up. If you have not been testing, please do so.
I tested them on Opteron NUMA systems and they are worse on simple tests than the stock scheduler (e.g. the parallelized STREAM test, which is a bit silly, but still fairly important)
OK, thanks.
Quoted 3 lines
For SMT there is a patch from Intel pending that teaches x86-64 to set up the SMT scheduler. They said they got slightly better benchmark results. The SMT setup seems to be racy though.
Am I correct in thinking that this patch provides the necessary hooks to integrate x86_4 into the new functionality which sched-domains provides, or is the Intel patch independent of sched-domains?
Quoted 3 lines
Some kind of SMT scheduler is definitely needed, we have a serious regression compared to 2.4 here right now. I'm not sure this is the right approach though, it seems to be far too complex.
Well that's discouraging. I really do want to push this thing along a bit.
Yours is the only report of regression of which I am aware. Is the reason understood?
And is anyone developing alternative SMT enhancements?
Quoted 7 lines
Quoted 3 lines
For SMT there is a patch from Intel pending that teaches x86-64 to set up the SMT scheduler. They said they got slightly better benchmark results. The SMT setup seems to be racy though.
Am I correct in thinking that this patch provides the necessary hooks to integrate
x86_4into the new functionality which sched-domains provides, or is the Intel patch independent of sched-domains?
It sets up the sched-domains code to know about HyperThreading CPUs on x86-64 too (basically same thing as the i386 code does with a few minor tweaks)
So it's dependent on that.
I will send it to you in separate mail.
Quoted 8 lines
Quoted 3 lines
Some kind of SMT scheduler is definitely needed, we have a serious regression compared to 2.4 here right now. I'm not sure this is the right approach though, it seems to be far too complex.
Well that's discouraging. I really do want to push this thing along a bit.
Yours is the only report of regression of which I am aware. Is the reason understood?
I think the reason is that it doesn't do balance on clone/fork. The normal scheduler also doesn't do that, but for some reason it still does better on the benchmarks (but worse than the old 2.4 -aa/Intel O(1) HT scheduler)
Quoted 1 line
And is anyone developing alternative SMT enhancements?
I thought there was a patch from Ingo Molnar? ("shared runqueue") I must admit I never tried it, just remember seeing the patches.
Also I've been playing with the entitlement scheduler to fix some of the interactivity problems I have on UP, but it also seems to still have problems.
-Andi
Quoted 4 lines · Andi Kleen
Also I've been playing with the entitlement scheduler to fix some of the interactivity problems I have on UP, but it also seems to still have problems.
You may find that nicksched fixes interactivity problems. It's a fairly fundamental rethink of the relationship between priorities and timeslices but back when I understood it I thought it made sense.
Quoted 3 lines · Andi Kleen
Quoted 3 lines
Quoted 3 lines
Some kind of SMT scheduler is definitely needed, we have a serious
regression compared to 2.4 here right now. I'm not sure this is the right approach though, it seems to be far too complex.
Andi, I'll agree that the way domains currently get set up is pretty ugly. Maybe some additional functions or macros could be used to make this process a bit clearer.
The actual kernel/sched.c code is really not that complex. In some ways it is *less* complicated than the old numa scheduler because it all goes through one code path.
It also handles SMT, which is where a bit of complexity is coming from. The other alternative is shared runqueues which is uglier and less flexible.
Quoted 9 lines
Quoted 5 lines
Well that's discouraging. I really do want to push this thing along a bit.
Yours is the only report of regression of which I am aware. Is the reason understood?
I think the reason is that it doesn't do balance on clone/fork. The normal scheduler also doesn't do that, but for some reason it still does better on the benchmarks (but worse than the old 2.4 -aa/Intel O(1) HT scheduler)
There have been a few changes and bug fixes since you last tested. Maybe that would help.
Quoted 5 lines
Quoted 2 lines
And is anyone developing alternative SMT enhancements?
I thought there was a patch from Ingo Molnar? ("shared runqueue") I must admit I never tried it, just remember seeing the patches.
Yep shared runqueues. Ingo and Rusty both had implementations but they both agreed sched-domains was a better alternative.
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 8 lines · Mickael Marchand
[snip]
Quoted 7 lines
Quoted 6 lines
while I am at it, I am running a 64 bits kernel with 32 bits debian testing and it seems some ioctl conversion fails
that happened with all 2.6 I tried.
here is the relevant kernel messages part :
ioctl32(dmsetup:26199): Unknown cmd fd(3) cmd(c134fd00){01} arg(0804c0b0) on/dev/mapper/control
Quoted 4 lines
Quoted 2 lines
The device mapper version 1 ioctl interface was removed. Perhaps you need to update your dm tools?
the debian tools are built with ioctlv4 (and compat for v1) I also tried with my own compiled dm tools from source without success
If it just uses them for compatibility probes then the ioctl handler can be silenced.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
ioctl32(fsck.reiserfs:201): Unknown cmd fd(4) cmd(80081272){00} arg(ffffdab8) on
/dev/ide/host0/bus0/target0/lun0/part4Is this something which 2.6 has always done, or is it new behaviour?
always since 2.6 IIRC
Quoted 1 line
reiserfs ioctl translation appears to be incomplete...
ha :)
I will take a look at it.
-Andi
Le jeudi 18 Mars 2004 00:25, Andi Kleen a écrit :
Quoted 10 lines
Quoted 9 lines · Mickael Marchand
[snip]
Quoted 7 lines
Quoted 6 lines
while I am at it, I am running a 64 bits kernel with 32 bits debian testing and it seems some ioctl conversion fails
that happened with all 2.6 I tried.
here is the relevant kernel messages part :
ioctl32(dmsetup:26199): Unknown cmd fd(3) cmd(c134fd00){01} arg(0804c0b0) on/dev/mapper/control
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
The device mapper version 1 ioctl interface was removed. Perhaps you need to update your dm tools?
the debian tools are built with ioctlv4 (and compat for v1) I also tried with my own compiled dm tools from source without success
If it just uses them for compatibility probes then the ioctl handler can be silenced.
hmm right now, dm/lvm absolutely does not work on amd64/32 bits. all ioctls calls are failling...
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
reiserfs ioctl translation appears to be incomplete...
ha :)
I will take a look at it.
thanks
Mik
Quoted 2 lines
hmm right now, dm/lvm absolutely does not work on amd64/32 bits. all ioctls calls are failling...
With no messages in the log?
Maybe they have broken data structures again, most likely because of different long long alignment. A lot of people who attempt to design data structures that don't need translation get that wrong unfortunately.
Emulating that stuff would be hard unfortunately because it has an rather over complicated ioctl structure that would be hard to write sane emulation code for.
Complain to the DM maintainers.
-Andi
just the ioctl cmd failed I reported in my first mail. then dmsetup just stops...
Cheers,
Mik
Le jeudi 11 Mars 2004 15:48, Andi Kleen a écrit :
Quoted 10 lines
Quoted 2 lines
hmm right now, dm/lvm absolutely does not work on amd64/32 bits. all ioctls calls are failling...
With no messages in the log?
Maybe they have broken data structures again, most likely because of different long long alignment. A lot of people who attempt to design data structures that don't need translation get that wrong unfortunately.
Quoted 5 lines
Emulating that stuff would be hard unfortunately because it has an rather over complicated ioctl structure that would be hard to write sane emulation code for.
Complain to the DM maintainers.
already did.
Quoted 2 lines · Mickael Marchand
just the ioctl cmd failed I reported in my first mail. then dmsetup just stops...
Either it doesn't handle the fallback correctly or the ioctls are not compatible.
Quoted 1 line
From a quick look at
dm-ioctl.hI found some suspicious cases,
but no clear failures.
-Andi
Quoted 4 lines · Andi Kleen
Maybe they have broken data structures again, most likely because of different long long alignment. A lot of people who attempt to design data structures that don't need translation get that wrong unfortunately.
I'd thought we'd been careful about this. You're suggesting that the size of this structure has changed between kernel versions ?!
struct dm_ioctl {
uint32_t version[3];
uint32_t data_size;
uint32_t data_start;
uint32_t target_count;
int32_t open_count;
uint32_t flags;
uint32_t event_nr;
uint32_t padding;
uint64_t dev;
char name[DM_NAME_LEN];
char uuid[DM_UUID_LEN];
};- Joe
Quoted 2 lines · Joe Thornber
I'd thought we'd been careful about this. You're suggesting that the size of this structure has changed between kernel versions ?!
Ignore me, I'm being an idiot.
- Joe
Quoted 11 lines · Joe Thornber
struct
dm_ioctl{ 0uint32_t version[3]; uint32_t data_size; 4 uint32_t data_start; uint32_t target_count; int32_t open_count; uint32_t flags; 8 uint32_t event_nr; uint32_t padding; 10 ***
Here's probably the problem. Many 64bit arches align 64bit
numbers on a 64bit boundary. So it is adding 2 more words of padding to start the u64 at offset 12.
Quoted 5 lines
uint64_t dev;
char name[DM_NAME_LEN];
char uuid[DM_UUID_LEN];
};Joel
Signature · 9 lines
Life's Little Instruction Book #313 "Never underestimate the power of love." Joel Becker Senior Member of Technical Staff Oracle Corporation E-mail: [email protected] Phone: (650) 506-8127
Quoted 2 lines · Joel Becker
Quoted 1 line · Joe Thornber
struct
dm_ioctl{ 0
Don't mind me, I'm an idiot.
Joel
Signature · 10 lines
"All alone at the end of the evening When the bright lights have faded to blue. I was thinking about a woman who had loved me And I never knew" Joel Becker Senior Member of Technical Staff Oracle Corporation E-mail: [email protected] Phone: (650) 506-8127
Quoted 15 lines · Joel Becker
Quoted 11 lines · Joe Thornber
struct
dm_ioctl{ 0uint32_t version[3]; uint32_t data_size; 4 uint32_t data_start; uint32_t target_count; int32_t open_count; uint32_t flags; 8 uint32_t event_nr; uint32_t padding; 10 ***
Quoted 3 lines
Here's probably the problem. Many 64bit arches align 64bit
numbers on a 64bit boundary. So it is adding 2 more words of padding to start the u64 at offset 12.
But wouldn't this be applied across the board and therefore still work?
Or is it defined as "packed" somewhere?
Chris
Signature · 4 lines
Chris Friesen | MailStop: 043/33/F10 Nortel Networks | work: (613) 765-0557 3500 Carling Avenue | fax: (613) 765-2986 Nepean, ON K2H 8E9 Canada | email: [email protected]
Quoted 9 lines · Joe Thornber
Quoted 4 lines · Andi Kleen
Maybe they have broken data structures again, most likely because of different long long alignment. A lot of people who attempt to design data structures that don't need translation get that wrong unfortunately.
I'd thought we'd been careful about this. You're suggesting that the size of this structure has changed between kernel versions ?!
Quoted 16 lines
struct dm_ioctl {
uint32_t version[3];
uint32_t data_size;
uint32_t data_start;
uint32_t target_count;
int32_t open_count;
uint32_t flags;
uint32_t event_nr;
uint32_t padding;
uint64_t dev;
char name[DM_NAME_LEN];
char uuid[DM_UUID_LEN];Are DM_NAME_LEN and DM_UUID_LEN not both a multiple of 8?
Quoted 1 line
};
There are more structures here, right?
If yes, that's the problem.
-Andi
Quoted 1 line · Andi Kleen
Are
DM_NAME_LENandDM_UUID_LENnot both a multiple of 8?
name len == 128, uuid_len == 129, so is the uuid_len being rounded up to the nearest 64bit boundary on x86-64 and only 32bit boundary on x86-32 ? (Sounds likely)
Quoted 1 line
There are more structures here, right?
Not that effect this problem, we're getting unknown ioctl, not an oops.
- Joe
Quoted 3 lines · Joe Thornber
name len == 128,
uuid_len== 129, so is theuuid_lenbeing rounded up to the nearest 64bit boundary on x86-64 and only 32bit boundary on x86-32 ? (Sounds likely)
In which case the following ugly patch should fix things. Mickael, any chance you could test this please ?
- Joe
Fix ioctl breakage on x86-64.
include/linux/dm-ioctl.h +29 -15
@@ -187,23 +187,37 @@ enum { DM_TABLE_STATUS_CMD, }; +/* +* The dm_ioctl struct passed into the ioctl is just the header +* on a larger chunk of memory. On x86-64 the dm-ioctl struct +* will be padded to an 8 byte boundary so the size will be +* different, which would change the ioctl code - yes I really +* messed up. This hack forces x86-64 to have the correct ioctl +* code. +*/ +#ifdef CONFIG_X86_64 +typedef char ioctl_struct [308 ]; +#else +typedef struct dm_ioctl ioctl_struct ; +#endif +#define DM_IOCTL 0xfd -#define DM_VERSION _IOWR(DM_IOCTL, DM_VERSION_CMD, struct dm_ioctl) -#define DM_REMOVE_ALL _IOWR(DM_IOCTL, DM_REMOVE_ALL_CMD, struct dm_ioctl) -#define DM_LIST_DEVICES _IOWR(DM_IOCTL, DM_LIST_DEVICES_CMD, struct dm_ioctl) - -#define DM_DEV_CREATE _IOWR(DM_IOCTL, DM_DEV_CREATE_CMD, struct dm_ioctl) -#define DM_DEV_REMOVE _IOWR(DM_IOCTL, DM_DEV_REMOVE_CMD, struct dm_ioctl) -#define DM_DEV_RENAME _IOWR(DM_IOCTL, DM_DEV_RENAME_CMD, struct dm_ioctl) -#define DM_DEV_SUSPEND _IOWR(DM_IOCTL, DM_DEV_SUSPEND_CMD, struct dm_ioctl) -#define DM_DEV_STATUS _IOWR(DM_IOCTL, DM_DEV_STATUS_CMD, struct dm_ioctl) -#define DM_DEV_WAIT _IOWR(DM_IOCTL, DM_DEV_WAIT_CMD, struct dm_ioctl) - -#define DM_TABLE_LOAD _IOWR(DM_IOCTL, DM_TABLE_LOAD_CMD, struct dm_ioctl) -#define DM_TABLE_CLEAR _IOWR(DM_IOCTL, DM_TABLE_CLEAR_CMD, struct dm_ioctl) -#define DM_TABLE_DEPS _IOWR(DM_IOCTL, DM_TABLE_DEPS_CMD, struct dm_ioctl) -#define DM_TABLE_STATUS _IOWR(DM_IOCTL, DM_TABLE_STATUS_CMD, struct dm_ioctl) +#define DM_VERSION _IOWR(DM_IOCTL, DM_VERSION_CMD, ioctl_struct) +#define DM_REMOVE_ALL _IOWR(DM_IOCTL, DM_REMOVE_ALL_CMD, ioctl_struct) +#define DM_LIST_DEVICES _IOWR(DM_IOCTL, DM_LIST_DEVICES_CMD, ioctl_struct) + +#define DM_DEV_CREATE _IOWR(DM_IOCTL, DM_DEV_CREATE_CMD, ioctl_struct) +#define DM_DEV_REMOVE _IOWR(DM_IOCTL, DM_DEV_REMOVE_CMD, ioctl_struct) +#define DM_DEV_RENAME _IOWR(DM_IOCTL, DM_DEV_RENAME_CMD, ioctl_struct) +#define DM_DEV_SUSPEND _IOWR(DM_IOCTL, DM_DEV_SUSPEND_CMD, ioctl_struct) +#define DM_DEV_STATUS _IOWR(DM_IOCTL, DM_DEV_STATUS_CMD, ioctl_struct) +#define DM_DEV_WAIT _IOWR(DM_IOCTL, DM_DEV_WAIT_CMD, ioctl_struct) + +#define DM_TABLE_LOAD _IOWR(DM_IOCTL, DM_TABLE_LOAD_CMD, ioctl_struct) +#define DM_TABLE_CLEAR _IOWR(DM_IOCTL, DM_TABLE_CLEAR_CMD, ioctl_struct) +#define DM_TABLE_DEPS _IOWR(DM_IOCTL, DM_TABLE_DEPS_CMD, ioctl_struct) +#define DM_TABLE_STATUS _IOWR(DM_IOCTL, DM_TABLE_STATUS_CMD, ioctl_struct)#define DM_VERSION_MAJOR 4#define DM_VERSION_MINOR 0
Hi,
just tested, it works just fine :)
no more errors,
dmsetup version and dmsetup ls work nicely
I will try to make evms work now but I guess that should be okay. I will report if I have other troubles.
good candidate for next mm ?
thanks Joe,
Mik
Le vendredi 12 Mars 2004 10:49, Joe Thornber a écrit :
Quoted 73 lines · Joe Thornber
Quoted 3 lines
name len == 128,
uuid_len== 129, so is theuuid_lenbeing rounded up to the nearest 64bit boundary on x86-64 and only 32bit boundary on x86-32 ? (Sounds likely)In which case the following ugly patch should fix things. Mickael, any chance you could test this please ?
- Joe
Fix ioctl breakage on x86-64.
include/linux/dm-ioctl.h +14
-2@@ -187,23 +187,37 @@ enum { DM_TABLE_STATUS_CMD, }; +/* +* The dm_ioctl struct passed into the ioctl is just the header +* on a larger chunk of memory. On x86-64 the dm-ioctl struct +* will be padded to an 8 byte boundary so the size will be +* different, which would change the ioctl code - yes I really +* messed up. This hack forces x86-64 to have the correct ioctl +* code. +*/ +#ifdef CONFIG_X86_64 +typedef char ioctl_struct [308 ]; +#else +typedef struct dm_ioctl ioctl_struct ; +#endif +#define DM_IOCTL 0xfd -#define DM_VERSION _IOWR(DM_IOCTL, DM_VERSION_CMD, struct dm_ioctl) -#define DM_REMOVE_ALL _IOWR(DM_IOCTL, DM_REMOVE_ALL_CMD, struct dm_ioctl) -#define DM_LIST_DEVICES _IOWR(DM_IOCTL, DM_LIST_DEVICES_CMD,struct
dm_ioctl) -
-#defineDM_DEV_CREATE_IOWR(DM_IOCTL,DM_DEV_CREATE_CMD, structdm_ioctl) -#defineDM_DEV_REMOVE_IOWR(DM_IOCTL,DM_DEV_REMOVE_CMD,
structdm_ioctl) -#defineDM_DEV_RENAME_IOWR(DM_IOCTL,DM_DEV_RENAME_CMD, structdm_ioctl) -#defineDM_DEV_SUSPEND_IOWR(DM_IOCTL, DM_DEV_SUSPEND_CMD, struct dm_ioctl)-#defineDM_DEV_STATUS
_IOWR(DM_IOCTL, DM_DEV_STATUS_CMD, struct dm_ioctl)-#defineDM_DEV_WAIT
_IOWR(DM_IOCTL, DM_DEV_WAIT_CMD, struct dm_ioctl)-
-#defineDM_TABLE_LOAD_IOWR(DM_IOCTL,DM_TABLE_LOAD_CMD, structdm_ioctl) -#defineDM_TABLE_CLEAR_IOWR(DM_IOCTL,DM_TABLE_CLEAR_CMD,
structdm_ioctl) -#defineDM_TABLE_DEPS_IOWR(DM_IOCTL,DM_TABLE_DEPS_CMD, structdm_ioctl) -#defineDM_TABLE_STATUS_IOWR(DM_IOCTL, DM_TABLE_STATUS_CMD, struct dm_ioctl)+#defineDM_VERSION
_IOWR(DM_IOCTL, DM_VERSION_CMD, ioctl_struct)+#defineDM_REMOVE_ALL_IOWR(DM_IOCTL, DM_REMOVE_ALL_CMD, ioctl_struct)+#defineDM_LIST_DEVICES_IOWR(DM_IOCTL, DM_LIST_DEVICES_CMD, ioctl_struct)+include/linux/dm-ioctl.h +11
-0+#define DM_DEV_CREATE _IOWR(DM_IOCTL, DM_DEV_CREATE_CMD, ioctl_struct) +#define DM_DEV_REMOVE _IOWR(DM_IOCTL, DM_DEV_REMOVE_CMD, ioctl_struct) +#define DM_DEV_RENAME _IOWR(DM_IOCTL, DM_DEV_RENAME_CMD, ioctl_struct) +#define DM_DEV_SUSPEND _IOWR(DM_IOCTL, DM_DEV_SUSPEND_CMD, ioctl_struct) +#define DM_DEV_STATUS _IOWR(DM_IOCTL, DM_DEV_STATUS_CMD, ioctl_struct) +#define DM_DEV_WAIT _IOWR(DM_IOCTL, DM_DEV_WAIT_CMD, ioctl_struct) + +#define DM_TABLE_LOAD _IOWR(DM_IOCTL, DM_TABLE_LOAD_CMD, ioctl_struct) +#define DM_TABLE_CLEAR _IOWR(DM_IOCTL, DM_TABLE_CLEAR_CMD, ioctl_struct) +#define DM_TABLE_DEPS _IOWR(DM_IOCTL, DM_TABLE_DEPS_CMD, ioctl_struct) +#define DM_TABLE_STATUS _IOWR(DM_IOCTL, DM_TABLE_STATUS_CMD,
ioctl_struct)#define
DM_VERSION_MAJOR4
#defineDM_VERSION_MINOR0
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Quoted 3 lines · Mickael Marchand
Hi,
just tested, it works just fine :)
Great I'm glad it works. Sorry I was so slow about realising what was wrong.
Quoted 1 line
good candidate for next mm ?
Yep, I'll forward a patch to akpm now.
- Joe
Quoted 3 lines · Mickael Marchand
just tested, it works just fine :)
no more errors,
dmsetup version and dmsetup ls work nicely
I concur -- this seems to be working now. Many thanks.
Quoted 4 lines
Quoted 4 lines
In which case the following ugly patch should fix things. Mickael, any chance you could test this please ?
- Joe
Eww... that's a really horrid-looking patch. I'm not complaining
right now, though. :)
Hugo.
Signature · 3 lines
=== Hugo Mills: hugo@... carfax.org.uk | darksatanic.net | lug.org.uk === PGP key: 1C335860 from wwwkeys.eu.pgp.net or http://www.carfax.org.uk --- Hey, Virtual Memory! Now I can have a *really big* ramdisk! ---
Quoted 2 lines · Mickael Marchand
hmm right now, dm/lvm absolutely does not work on amd64/32 bits. all ioctls calls are failling...
This one has me stumped. I've tested on sparc64/debian and Kevin Corry has tested on PPC and neither of us have problems. So it looks like an amd64 only problem, does 2.6.4 vanilla work ? (I don't have access to one of these machines).
- Joe
Quoted 6 lines · Joe Thornber
Quoted 2 lines · Mickael Marchand
hmm right now, dm/lvm absolutely does not work on amd64/32 bits. all ioctls calls are failling...
This one has me stumped. I've tested on sparc64/debian and Kevin Corry has tested on PPC and neither of us have problems. So it looks
ppc and sparc64 are different from x86-64 and ia64.
The problem on i386 is that alignof(long long) is different between 32bit and 64bit. That's not the case on the riscs.
This causes problems either with moving fields around/after 64bit values and worse it changes the alignment of whole structures in arrays too (because alignof(struct) is the largest alignment needed by any members)
Quoted 2 lines
like an amd64 only problem, does 2.6.4 vanilla work ? (I don't have access to one of these machines).
Most likely it's one of your arrays. You pass arrays, right?
-Andi