RFC: let x86_64 no longer define X86
I'd like to send a patch after 2.6.10 that removes the following from arch/x86_64/Kconfig:
config X86
bool
default y
Additionally, I'll also check all current X86 uses to prevent breakages.
Why?
X86 is _the_ symbol to identify the i386 architecture, but the x86_64 port hijacked it. Kernel-wise, x86_64 is mostly simply a new port like e.g. ia64.
Where is the problem?
To say "X86", you currently have to write "(X86 && !X86_64)" in the Kconfig file. This is not intuitive.
Why is e.g. CONFIG_LBD available on x86_64 and even enabled in defconfig?
Isn't this an incompatible change?
Yes it is.
But according to the current development model, such changes are allowed in 2.6 .
And if you want to support both older and more recent kernels, the following dependencies will be correct both before and after this change:
- (X86 && !X86_64)
- (X86 && X86_64)
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 4 lines · Adrian Bunk
And if you want to support both older and more recent kernels, the following dependencies will be correct both before and after this change:
- (X86 && !X86_64)
- (X86 &&X86_64)
This last one surely can't be correct before *and* afterwards. But even in the current system, it is a pretty perverse thing to check for. I guess you meant:
(X86 || X86_64)
Quoted 9 lines · Nick Piggin
Quoted 6 lines · Adrian Bunk
And if you want to support both older and more recent kernels, the following dependencies will be correct both before and after this change:
- (X86 && !X86_64)
- (X86 &&X86_64)
Quoted 5 lines
This last one surely can't be correct before *and* afterwards. But even in the current system, it is a pretty perverse thing to check for. I guess you meant:
(X86 ||
X86_64)
Yes, thanks for the correcting my typo.
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 9 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:config X86
bool
default yAdditionally, I'll also check all current X86 uses to prevent breakages.
Or, you could define an X86_32 config symbol in i386. This seems a little more backward compatible, and means that you can continue to just test X86 for the rather large set of code that works fine on both 32-bit and 64-bit.
I guess it depends on whether you think there are more places in the generic code that the two architectures share code, vs places that are 32-bit only.
Paul
Quoted 10 lines · Paul Menage
Quoted 8 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:config X86
bool
default yAdditionally, I'll also check all current X86 uses to prevent breakages.
Quoted 8 lines
Or, you could define an
X86_32config symbol in i386. This seems a little more backward compatible, and means that you can continue to just test X86 for the rather large set of code that works fine on both 32-bit and 64-bit.I guess it depends on whether you think there are more places in the generic code that the two architectures share code, vs places that are 32-bit only.
We are not talking about thousands of places.
We are talking about less than hundred places.
And many people do currently get it wrong like with CONFIG_LBD.
The most important improvement would be to prevent such bugs and to have the X86_64 dependency explicitely stated.
The #ifdef CONFIG_X86 in init/main.c is an example where it currently takes some time to understand whether it's correct or a bug.
X86_32 would be a solution, but it would IMHO also create confusion since i386 and ia64 also have some things in common (e.g. ACPI support). The cleanest thing is simply, to state X86_64 dependencies explicitely.
Quoted 1 line
Paul
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 2 lines
The most important improvement would be to prevent such bugs and to have the
X86_64dependency explicitely stated.
This would just end up with me having to hunt through all the drivers all the time and enabling drivers that need to be enabled on x86-64 too.
It's much easier to disable the few drivers that are broken with !X86_64.
Again please don't do it. It will just cause more work long term.
-Andi
Quoted 8 lines · Andi Kleen
Quoted 2 lines
The most important improvement would be to prevent such bugs and to have the
X86_64dependency explicitely stated.This would just end up with me having to hunt through all the drivers all the time and enabling drivers that need to be enabled on x86-64 too.
It's much easier to disable the few drivers that are broken with !X86_64.
...
The issue you describe only occurs when a new dependency on X86 is added. It is not a problem for the common case that a driver is portable and therefore available on all architectures.
If the driver also works on X86_64, adding a " || X86_64" is trivial. In the worst case, a new driver is not available on X86_64 until this is added to the dependencies.
But the current setup might cause real bugs.
If one new user of CONFIG_LBD wouldn't additionally (and not strictly required) check BITS_PER_LONG, it might currently have unwanted effects on X86_64. Explicite annotations with X86_64 if it works on this architecture would prevent such bugs.
Quoted 1 line
-Andi
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 6 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:config X86
bool
default y
I'm against this. Please don't do this.
-Andi
Quoted 10 lines · Andi Kleen
Quoted 5 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:config X86
bool
default yI'm against this. Please don't do this.
An explanation would be nice.
Jeff
Quoted 6 lines · Jeff Garzik
Quoted 5 lines · Andi Kleen
Quoted 3 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
config X86
bool
default yI'm against this. Please don't do this.
An explanation would be nice.
Basically what Paul Menage said. There is a lot of common code, and you would end up writing X86 && X86_64 more often than X86 && !X86_64.
In addition such a change is quite intrusive and I don't think it's a good idea to do right now because it'll very likely introduce bugs.
If someone really thinks the X86 && !X86_64 is too ugly
(I personally don't think it is because it says clearly
what the matter is) then adding an additional X86_32 would be the right thing to do.
-Andi
Quoted 3 lines · Andi Kleen
Basically what Paul Menage said. There is a lot of common code, and you would end up writing X86 &&
X86_64more often than X86 && !X86_64.
(assuming you meant the first to be X86 || X86_64)
Can you show some examples? We don't have this for any other architecture.
Signature · 1 line
dwmw2
Quoted 2 lines · David Woodhouse
Can you show some examples? We don't have this for any other architecture.
Just grep for any use of X86 in Kconfig.
-Andi
Quoted 5 lines · Andi Kleen
Quoted 2 lines · David Woodhouse
Can you show some examples? We don't have this for any other architecture.
Just grep for any use of X86 in
Kconfig.
OK, I'll pick the first I find...
config ATP
tristate "AT-LAN-TEC/RealTek pocket adapter support" depends on NET_POCKET && ISA && X86
Why is that OK on x86_64 but not Alpha? Why isn't the use of X86 a bug in that case?
Signature · 1 line
dwmw2
Quoted 7 lines · David Woodhouse
Quoted 5 lines · Andi Kleen
Quoted 2 lines · David Woodhouse
Can you show some examples? We don't have this for any other architecture.
Just grep for any use of X86 in
Kconfig.
Quoted 8 lines
OK, I'll pick the first I find...
config ATP
tristate "AT-LAN-TEC/RealTek pocket adapter support" depends onNET_POCKET&& ISA && X86Why is that OK on
x86_64but not Alpha? Why isn't the use of X86 a bug in that case?
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
-Andi
Quoted 1 line · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
which I disagree with. CONFIG_ISA should include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.
Jeff
Quoted 8 lines · Jeff Garzik
Quoted 2 lines · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
which I disagree with.
CONFIG_ISAshould include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.
There is. It gets rid of many tens of drivers that are not and will never be 64bit clean and have a snowball in hell chances to work on x86-64.
In theory you could invent a new ISA_SLOT or ISA_BROKEN config for them, but since ISA does the job quite well for near everybody except for one or two corner cases I don't see any sense in changing it.
-Andi
Quoted 10 lines · Andi Kleen
Quoted 6 lines · Jeff Garzik
Quoted 1 line · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
which I disagree with.
CONFIG_ISAshould include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.
Quoted 6 lines
There is. It gets rid of many tens of drivers that are not and will never be 64bit clean and have a snowball in hell chances to work on x86-64.
In theory you could invent a new
ISA_SLOTorISA_BROKENconfig for them, but since ISA does the job quite well for near everybody except for one or two corner cases I don't see any sense in changing it.
The traditional legacy ISA devices -- floppy, serial, parallel, mouse, keyboard, IDE -- are still around. Yet now we need to invent a new name to classify ISA devices that have been with us for 20 years?
CONFIG_ISA_BROKEN is more appropriate than pretending devices we've called ISA since the 1980's do not imply/depend on CONFIG_ISA.
Jeff
Quoted 21 lines · Jeff Garzik
Quoted 13 lines · Andi Kleen
Quoted 9 lines · Jeff Garzik
Quoted 2 lines · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
which I disagree with.
CONFIG_ISAshould include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.Quoted 6 lines
There is. It gets rid of many tens of drivers that are not and will never be 64bit clean and have a snowball in hell chances to work on x86-64.
In theory you could invent a new
ISA_SLOTorISA_BROKENconfig for them, but since ISA does the job quite well for near everybody except for one or two corner cases I don't see any sense in changing it.
Quoted 6 lines
The traditional legacy ISA devices -- floppy, serial, parallel, mouse, keyboard, IDE -- are still around. Yet now we need to invent a new name to classify ISA devices that have been with us for 20 years?
CONFIG_ISA_BROKENis more appropriate than pretending devices we've called ISA since the 1980's do not imply/depend onCONFIG_ISA.
Silly question:
Why CONFIG_ISA_BROKEN?
A new CONFIG_ISA_SLOT might solve such cases, and otherwise the case Andi described would be perfectly covered by !CONFIG_64BIT .
Quoted 1 line
Jeff
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 18 lines · Jeff Garzik
Quoted 10 lines · Andi Kleen
Quoted 6 lines · Jeff Garzik
Quoted 1 line · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
which I disagree with.
CONFIG_ISAshould include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.Quoted 6 lines
There is. It gets rid of many tens of drivers that are not and will never be 64bit clean and have a snowball in hell chances to work on x86-64.
In theory you could invent a new
ISA_SLOTorISA_BROKENconfig for them, but since ISA does the job quite well for near everybody except for one or two corner cases I don't see any sense in changing it.
Quoted 6 lines
The traditional legacy ISA devices -- floppy, serial, parallel, mouse, keyboard, IDE -- are still around. Yet now we need to invent a new name to classify ISA devices that have been with us for 20 years?
CONFIG_ISA_BROKENis more appropriate than pretending devices we've called ISA since the 1980's do not imply/depend onCONFIG_ISA.Jeff
Why CONFIG_ISA_BROKEN. That implies (states) that its broken and it's not. All modern ix86's have such a bus, even though it doesn't go out to some connectors anymore. Just leave the 20-year-old local bus, then called ISA, alone. ISA means:
Industry Standard Architecture
It __is__! Don't muck with it.
Cheers,
Dick Johnson
Penguin : Linux version 2.6.9 on an i686 machine (5537.79 BogoMips).
Notice : All mail here is now cached for review by John Ashcroft.
98.36% of all statistics are fiction.
Quoted 1 line · linux-os
Why
CONFIG_ISA_BROKEN. That implies (states) that its broken and
The name is appropriate because the drivers in question _are_ broken.
Jeff
Quoted 2 lines
Quoted 1 line · linux-os
Why
CONFIG_ISA_BROKEN. That implies (states) that its broken and
Quoted 1 line · Jeff Garzik
The name is appropriate because the drivers in question _are_ broken.
On some architectures -- this is a porting issue, and not a clean binary distinction.
ASSUMES_32_BIT or some other "32 bit" name would probably better capture this particular issue.
Even better might be to get the compiler to catch the most obvious mistakes and use #define decorations to override the compiler's determination (you'd need two of these, because the compiler can get this wrong in two different ways).
Signature · 1 line
Raul
Quoted 6 lines · Andi Kleen
Quoted 5 lines · Jeff Garzik
Quoted 2 lines · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
Quoted 7 lines
Quoted 3 lines
which I disagree with.
CONFIG_ISAshould include southbridge devices behind a PCI<->ISA bridge. There is zero value to a more stricter "there is a physical ISA bus in this machine" definition.There is. It gets rid of many tens of drivers that are not and will never be 64bit clean and have a snowball in hell chances to work on x86-64.
...
If this was correctly annotated with (!64BIT || BROKEN), such a broken driver wasn't offered on Alpha, too.
Quoted 1 line
-Andi
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 2 lines · Andi Kleen
I don't know details about the driver, but it's not enabled on x86-64 because x86-64 doesn't have ISA set.
That looks bogus actually -- I think it should only depend on the existence of parport_pc style hardware. CONFIG_ISA is definitely a digression. But still, either way the example is wrong. It shouldn't be limited to X86 and X86_64.
I still haven't found good examples of cases where X86 is used and we would want to change that to X86 || X86_64. Could this be one?
config HW_RANDOM
tristate "Intel/AMD/VIA HW Random Number Generator support" depends on (X86 || IA64) && PCI
...or this?
config FTAPE
tristate "Ftape (QIC-80/Travan) support" depends on BROKEN_ON_SMP && (ALPHA || X86)
I also see some which already have it:
config NVRAM
tristate "/dev/nvram support"
depends on ATARI || X86 || X86_64 || ARM || GENERIC_NVRAM
config HANGCHECK_TIMER
tristate "Hangcheck timer"
depends on X86_64 || X86
And some which seem to be wrong because they want only X86 not X86_64:
config SONYPI
tristate "Sony Vaio Programmable I/O Control Device support (EXPERIMENTAL)" depends on EXPERIMENTAL && X86 && PCI && INPUT && !64BIT
config MWAVE
tristate "ACP Modem (Mwave) support"
depends on X86
select SERIAL_8250
Using X86 to include X86_64 is bizarre and inconsistent, and it's already leading to errors in Kconfig. Let's fix it.
Signature · 1 line
dwmw2
Quoted 5 lines · Andi Kleen
...
In addition such a change is quite intrusive and I don't think it's a good idea to do right now because it'll very likely introduce bugs.
...
I's say it will fix some bugs and make some future bugs less likely.
The CONFIG_LBD example in my initial mail is an example for such a bug.
Quoted 1 line
-Andi
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedQuoted 8 lines · Adrian Bunk
Quoted 5 lines · Andi Kleen
...
In addition such a change is quite intrusive and I don't think it's a good idea to do right now because it'll very likely introduce bugs.
...I's say it will fix some bugs and make some future bugs less likely.
I doubt it (and so far I did most of these changes)
-Andi
At Fri, 19 Nov 2004 01:51:17 +0100,
Quoted 8 lines · Adrian Bunk
I'd like to send a patch after 2.6.10 that removes the following from
arch/x86_64/Kconfig:config X86
bool
default y
Quoted 8 lines
Additionally, I'll also check all current X86 uses to prevent breakages.
Why?
X86 is _the_ symbol to identify the i386 architecture, but the
x86_64port hijacked it. Kernel-wise,x86_64is mostly simply a new port like e.g. ia64.
If we do this for x86-64, shouldn't we do the same for ppc64 and sh64, too? (only sparc64 seems not defining 32bit config.)
Takashi
Quoted 2 lines · Takashi Iwai
If we do this for x86-64, shouldn't we do the same for ppc64 and sh64, too? (only sparc64 seems not defining 32bit config.)
Both ppc32 and ppc64 define CONFIG_PPC already.
Paul.