x86 merge - a little feedback
Hi Thomas et al.
After spending several hours fiddeling with and improving the current Makefile for x86_64 I decided to take a closer look at the x86 merge og i386 and x86_64.
I took a closer look at x86/pci. There are 16 C files.
Quoted 1 line
From the mails and discussions I expected it be be
obvious what was i386 only, what was shared and what was x86_64 only.
But see following table
Filename i386 x86_64acpi.c X Xcommon.c X Xdirect.c X Xearly.c X Xfixup.c X Xi386.c X Xinit.c X Xirq.c Xk8-bus.c Xlegacy.c X Xmmconfig_32.c Xmmconfig_64.c Xmmconfig-shared.c X Xnuma.c Xpcbios.c Xvisws.c X
In the filename there is NOTHING for most of
the non-shared code that tell that this file is
used by only i386 or x86_64.
The exception is mmconfig that is prefixed with _32 versus _64. But as I have understood the mails floating around using _32,_64 is a way to say here are a potential candidate for futher merging.
In a meged x86 tree it would be very beneficial to either include in the filename that a specific file is i386 or x86_64 specific or stuff them in a separate subdirectory.
If legacy.c numa.c, pcibios.c and visws.c placed in a directory named i386 then it would be obvious that this is i386 only.
Or they could be named filename_32 (or the uglier filename_i386). As it stands out today the filename are kept but thier relationship are lost.
I dunno if this will address the concern of Andi about mixing i386 and x86_64 but to me at least things would be so much more obvious if the original relationship are spelled out.
Sam
Sam,
Quoted 31 lines · Sam Ravnborg
Hi Thomas et al.
After spending several hours fiddeling with and improving the current
Makefileforx86_64I decided to take a closer look at the x86 merge og i386 andx86_64.I took a closer look at x86/pci. There are 16 C files.
From the mails and discussions I expected it be be
obvious what was i386 only, what was shared and what wasx86_64only.But see following table
Filename i386
x86_64acpi.cX Xcommon.cX Xdirect.cX Xearly.cX Xfixup.cX Xi386.cX Xinit.cX Xirq.cXk8-bus.cXlegacy.cX Xmmconfig_32.cXmmconfig_64.cXmmconfig-shared.cX Xnuma.cXpcbios.cXvisws.cX
Quoted 3 lines
In the filename there is NOTHING for most of
the non-shared code that tell that this file is
used by only i386 orx86_64.
True. I tried to unify the Makefile by using
obj-$(CONFIG_X86_32) += ....
and
obj-$(CONFIG_X86_64) += ....
but I did fail due to link order problems in that code. I had not yet time to go down and figure that out yet, but it is on my todo list.
Quoted 3 lines
The exception is mmconfig that is prefixed with _32 versus _64. But as I have understood the mails floating around using _32,_64 is a way to say here are a potential candidate for futher merging.
Yes, if it is possible and makes sense.
Quoted 8 lines
In a meged x86 tree it would be very beneficial to either include in the filename that a specific file is i386 or
x86_64specific or stuff them in a separate subdirectory.If
legacy.cnuma.c,pcibios.candvisws.cplaced in a directory named i386 then it would be obvious that this is i386 only.
Or they could be namedfilename_32(or the uglierfilename_i386). As it stands out today the filename are kept but thier relationship are lost.
We concentrated first on the move to make it simple and binary equivalent. The cleanup of code (merging, location, makefile updates) are definitely on our todo list.
Quoted 3 lines
I dunno if this will address the concern of Andi about mixing i386 and
x86_64but to me at least things would be so much more obvious if the original relationship are spelled out.
Good point.
tglx
Quoted 6 lines · Thomas Gleixner
I tried to unify the
Makefileby usingobj-$(CONFIG_X86_32) += ....
and
obj-$(CONFIG_X86_64) += ....
Don't do that.
I think it would be much better to instead do something like
obj-y += mmconfig_$(CONFIG_WORD_SIZE).o
to make it clear when we have a file that is conceptually the same, but has different implementations.
That also makes the unification (assuming/hoping it gets done) of such files much cleaner - you just merge them, and the obj-y line can just drop the $(CONFIG_WORD_SIZE) thing. Very logical.
Quoted 1 line
but I did fail due to link order problems in that code.
.. the above approach also gets rid of any link order problems.
Linus
Quoted 10 lines · Sam Ravnborg
...
In a meged x86 tree it would be very beneficial to either include in the filename that a specific file is i386 orx86_64specific or stuff them in a separate subdirectory.If
legacy.cnuma.c,pcibios.candvisws.cplaced in a directory named i386 then it would be obvious that this is i386 only.
Or they could be namedfilename_32(or the uglierfilename_i386). As it stands out today the filename are kept but thier relationship are lost.
...
I'm not agreeing with you on this.
It seems artificial to think 32bit<->64bit was the only interesting distinction on x86 machines.
You might as well create different directories for i386<->i486+ or pre-i686<->i686+.
As an example, visws.c is as much non-64bit as it is pre-i686.
Quoted 1 line
Sam
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
As an example,
visws.cis as much non-64bit as it is pre-i686.
Bad example...
visws is a visws specific file that should naver have allowed anywhere outside mach-visws.
Any link-order issues should have been dealt with in other ways.
Sam
Quoted 5 lines · Sam Ravnborg
Quoted 2 lines
As an example,
visws.cis as much non-64bit as it is pre-i686.Bad example...
visws is a visws specific file that should naver have allowed anywhere outside mach-visws.
Exactly, it's not about 64bit at all.
Similar, e.g. legacy.c is not about 32bit<->64bit.
Quoted 3 lines
Any link-order issues should have been dealt with in other ways.
Sam
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 1 line · Adrian Bunk
As an example,
visws.cis as much non-64bit as it is pre-i686.
Bullshit. visws were shipped with P3s.
Quoted 4 lines · Christoph Hellwig
Quoted 1 line · Adrian Bunk
As an example,
visws.cis as much non-64bit as it is pre-i686.Bullshit. visws were shipped with P3s.
Certainly true, but still not 64bit and never will be.
--
Len Sorensen
Quoted 4 lines
In the filename there is NOTHING for most of
the non-shared code that tell that this file is
used by only i386 orx86_64.
Exactly my point from KS. The big mash-up will not really make much difference in terms of Makefile clarity or whatever Thomas' point was. Apparently he wanted to eliminate a few lines of code from the Makefile and merge the header files completely?
Anyways, the end result will be roughly the same as it is now: i386 and x86-64 are shared in not completely obvious ways and if you change one you have to test compile the other too.
Same old, same old, as always.
Quoted 3 lines
I dunno if this will address the concern of Andi about mixing i386 and
x86_64but to me at least things would be so much more obvious if the original relationship are spelled out.
In the end it won't make much difference where the files are located (although I frankly don't see the advantage of this intrusive move).
You always have to at least compile test both if you change one and I doubt most people will be able to avoid this no matter how the Makefile looks or where the files are.
Even if everything was merged together and only ifdefs remained that fundamental fact would not change either.
A few more files could be also definitely shared, no argument. e.g. the time subsystem will likely to be shared soon anyways. And probably a few others. That should be better all done carefully step by step and properly reviewed though, not in some kind of brutal "rewrite the world" event.
My concern is mostly that he seems to want to merge some things between 32bit and 64bit (like the APIC drivers or the crappy i386 maze-of-inlines subarchitecture design) which ought not be together. I think I managed to keep x86-64 a relatively clean port over-all, but I see this now going down the drain :/
-Andi
Quoted 8 lines · Andi Kleen
...
In the end it won't make much difference where the files are located (although I frankly don't see the advantage of this intrusive move).You always have to at least compile test both if you change one and I doubt most people will be able to avoid this no matter how the
Makefilelooks or where the files are.
...
The important point is:
People do not expect code under arch/i386/ to be used by code under arch/x86_64/ and vice versa.
That regularly results in people sending patches that don't compile on the other architecture.
With one architecture it's much more obvious that the code is shared.
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 9 lines
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.That regularly results in people sending patches that don't compile on the other architecture.
With one architecture it's much more obvious that the code is shared.
Will that cause people to compile test both? I have my doubts that will really work.
e.g. a similar example would be CONFIG_MMU=n. The code is mostly shared and in the same directories, but people still break the MMUless architectures all the time.
I don't expect this to be different with 32bit/64bit.
-Andi
Quoted 5 lines · Andi Kleen
Quoted 5 lines
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.
Quoted 8 lines
Quoted 4 lines
That regularly results in people sending patches that don't compile on the other architecture.
With one architecture it's much more obvious that the code is shared.
Will that cause people to compile test both? I have my doubts that will really work.
...
You will see that it could be shared, and it'll be much easier to see all configurations it's used in.
Currently, there are 6 or 7 different ways how a function under arch/i386/ could be used by a function under arch/x86_64/ (and vice versa) and it's non-trivial to figure out all usages when grep'ing for users.
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 3 lines · Andi Kleen
Will that cause people to compile test both? I have my doubts that will really work.
If people don't compile-test both now, then why would they compile-test things when merged?
So no, that's not the point.
But at least things like "grep" will work sanely, and people will be *aware* that "Oh, this touches a file that may be used by the other word-size".
Right now, we have people changing "i386-only" files that turn out to be used by x86-64 too - through very subtle Makefile things that the person who only looks into the i386 Makefile will never even *see*.
THAT is the problem (well, at least part of it).
Linus
Quoted 5 lines · Linus Torvalds
Quoted 3 lines · Andi Kleen
Will that cause people to compile test both? I have my doubts that will really work.
Quoted 8 lines
If people don't compile-test both now, then why would they compile-test things when merged?
So no, that's not the point.
But at least things like "grep" will work sanely, and people will be *aware* that "Oh, this touches a file that may be used by the other word-size".
Speaking of which, how does sparc/sparc64 handle things?
Jan
Quoted 6 lines · Andi Kleen
Quoted 6 lines
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.That regularly results in people sending patches that don't compile on the other architecture.
Quoted 9 lines
Quoted 1 line
With one architecture it's much more obvious that the code is shared.
Will that cause people to compile test both? I have my doubts that will really work.
e.g. a similar example would be
CONFIG_MMU=n. The code is mostly shared and in the same directories, but people still break the MMUless architectures all the time.
As I was the first one to do CONFIG_MMU=y/n in the same arch directory, since 2.5, I can tell you that that's simply crap. The only reason CONFIG_MMU=n gets broken all the time is because people don't think about it in generic code, it's rarely broken in the architecture code, and even with the most occasional of build tests most of that gets caught in a hurry.
You do of course have to consider both cases when writing new code, but those things tend to be pretty obvious. It's a bit more work for the arch maintainer, but it's certainly far less confusing and problematic than separating things out.
In fact, going the opposite route is what leads to endless trouble in the long run, since you brought up the MMUless example, m68knommu is a good example. Rather than beginning life in arch/m68k, it was forked off, mostly to deal with the ColdFire CPUs that weren't planned to have MMUs. Now that the product line has moved along, adding an MMU to it is in the roadmap, which means that inevitably they're both going to have to be combined anyways. Simply dealing with the initial trouble of having them combined initially would have solved a lot of that mess.
You can ignore the added maintenance for as long as possible, but sooner or later it's going to be a problem. Procrastination is not something that bodes particularly well for divergent hardware support.
Quoted 6 lines · On Wed, 12 Sep 2007 09:29:53 +0900 Paul Mundt <[email protected]> wrote:
As I was the first one to do
CONFIG_MMU=y/nin the same arch directory, since 2.5, I can tell you that that's simply crap. The only reasonCONFIG_MMU=ngets broken all the time is because people don't think about it in generic code, it's rarely broken in the architecture code, and even with the most occasional of build tests most of that gets caught in a hurry.
oy. I do sh allmodconfig test builds fairly regularly.
<does one>
It all went OK, up until moddep:
ERROR: "_ebss" [drivers/mtd/maps/uclinux.ko] undefined!
Quoted 2 lines · On Tue, 11 Sep 2007 23:14:22 +0200 Adrian Bunk <[email protected]> wrote:
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.
[OT: it drives me batshit that we ended up including stuff in both directions]
Quoted 6 lines · Andrew Morton
Quoted 2 lines · On Tue, 11 Sep 2007 23:14:22 +0200 Adrian Bunk <[email protected]> wrote:
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.[OT: it drives me batshit that we ended up including stuff in both directions]
Why?
Anyways, i wouldn't have a problem with putting the already shared files into a different directory or move it over to one of the architectures, although I must admit I personally wouldn't see a big benefit from it. But if it gives people a warm fuzzy feeling I'm all for it.
-Andi
Quoted 9 lines · On Sat, 15 Sep 2007 20:36:23 +0200 Andi Kleen <[email protected]> wrote:
Quoted 6 lines · Andrew Morton
Quoted 2 lines · On Tue, 11 Sep 2007 23:14:22 +0200 Adrian Bunk <[email protected]> wrote:
People do not expect code under
arch/i386/to be used by code underarch/x86_64/and vice versa.[OT: it drives me batshit that we ended up including stuff in both directions]
Why?
It's more complex, obviously. More surprising. It used to be the case that arch/x86^4 files were xx86_64 and arch/i386 files were i386 and possibly x86_64. Now it's the case that arch/x86_64 files are x86_64 and maybe i386 and arch/i386 files are i386 and maybe x86_64. Additional and quite unnecessary complexity.
I mean, how often do x86_64 changes in your tree break i386? Once every 3ish weeks would be my guess. Often this will be because the person making (and reviewing) the x86_64 change didn't know (or forgot) that the file is also used by x86_64.
Quoted 4 lines
Anyways, i wouldn't have a problem with putting the already shared files into a different directory or move it over to one of the architectures, although I must admit I personally wouldn't see a big benefit from it. But if it gives people a warm fuzzy feeling I'm all for it.
Doing something like that would reduce complexity, reduce surprise and increase maintainability. That's more than warm-and-fuzzies.
Quoted 3 lines · Sam Ravnborg
From the mails and discussions I expected it be be obvious what was i386 only, what was shared and what was
x86_64only.
The problem right now is the *reverse* - even though they are in different subdirectories (and thus *look* like they are all separate), they aren't.
So the merged end result is much better: at a first approximation everything is shared (largely true), and the ones that aren't shared can be made more obvious.
Quoted 12 lines
But see following table
Filename i386
x86_64acpi.cX Xcommon.cX Xdirect.cX Xearly.cX Xfixup.cX Xi386.cX Xinit.cX Xlegacy.cX Xmmconfig-shared.cX X
So sharing is the default, as it should be.
Quoted 4 lines
irq.cXk8-bus.cXnuma.cXpcbios.cX
And these are examples of things that likely *should* be shared, and have nothing what-so-ever to do with 32- vs 64-bit issues. For example, neither NUMA nor k8 is in *any* way a 32-bit vs 64-bit issue, they are issues with chips that can be used as either.
Quoted 2 lines
mmconfig_32.cXmmconfig_64.cX
And this is obvious, and correct.
Quoted 1 line
visws.cX
Now, arguably this should be entirely elsewhere (ie in the "mach-visw" directory).
Quoted 3 lines
In the filename there is NOTHING for most of
the non-shared code that tell that this file is
used by only i386 orx86_64.
.. and that's because of historical issues, and has nothing to do with the merge, and everything to do with all the problems that the merge is supposed to eventually help us FIX!
Quoted 3 lines
The exception is mmconfig that is prefixed with _32 versus _64. But as I have understood the mails floating around using _32,_64 is a way to say here are a potential candidate for futher merging.
Yes. But so are others.
Quoted 2 lines
If
legacy.cnuma.c,pcibios.candvisws.cplaced in a directory named i386 then it would be obvious that this is i386 only.
But none of it is "i386 only" and putting it in a directory of its own would be stupid and wrong. The visws.c thing is platform-specific thing, and the fact that the platform happens to be 32-bit is totally secondary to the much bigger issue of the *platform*, so again, it would be totally wrong to split it up by wordsize.
Quoted 3 lines
I dunno if this will address the concern of Andi about mixing i386 and
x86_64but to me at least things would be so much more obvious if the original relationship are spelled out.
I obviously disagree violently. We should absolutely *not* make any of this depend on word-size. Quite the reverse!
Linus
Quoted 9 lines
Quoted 3 lines
If
legacy.cnuma.c,pcibios.candvisws.cplaced in a directory named i386 then it would be obvious that this is i386 only.But none of it is "i386 only" and putting it in a directory of its own would be stupid and wrong. The
visws.cthing is platform-specific thing, and the fact that the platform happens to be 32-bit is totally secondary to the much bigger issue of the *platform*, so again, it would be totally wrong to split it up by wordsize.
In other words - of all directories I used the worst one to prove my point. I had no idea that legacy.c, numa.c and pcbios.c was candidates for x86_64 usage.
The point I try to make and which seems to have been lost in platform/wordsize inputs are that there is a reason for being able to see which of the two architectures a given file belong to.
Try to grep for csum_partial in x86/lib and you will get this:
checksum.S: * Changes: Ingo Molnar, converted csum_partial_copy() to 2.1 exception
checksum.S:unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum) checksum.S:ENTRY(csum_partial)
checksum.S:ENDPROC(csum_partial)
checksum.S:ENTRY(csum_partial)
checksum.S:ENDPROC(csum_partial)
checksum.S:unsigned int csum_partial_copy_generic (const char *src, char *dst, checksum.S: * Copy from ds while checksumming, otherwise like csum_partial checksum.S:ENTRY(csum_partial_copy_generic) checksum.S:ENDPROC(csum_partial_copy_generic) checksum.S:ENTRY(csum_partial_copy_generic) checksum.S:ENDPROC(csum_partial_copy_generic) csum-copy.S:ENTRY(csum_partial_copy_generic) csum-copy.S:ENDPROC(csum_partial_copy_generic)
csum-partial.c:__wsum csum_partial(const void *buff, int len, __wsum sum) csum-partial.c:EXPORT_SYMBOL(csum_partial);csum-partial.c: return csum_fold(csum_partial(buff,len,0));csum-wrappers.c: * csum_partial_copy_from_user - Copy and checksum from user space. csum-wrappers.c:csum_partial_copy_from_user(const void __user *src, void *dst,csum-wrappers.c: isum = csum_partial_copy_generic((__force const void *)src,
csum-wrappers.c:EXPORT_SYMBOL(csum_partial_copy_from_user);csum-wrappers.c: * csum_partial_copy_to_user - Copy and checksum to user space. csum-wrappers.c:csum_partial_copy_to_user(const void *src, void __user *dst,csum-wrappers.c: return csum_partial_copy_generic(src, (void __force *)dst,len,isum,NULL,errp);
csum-wrappers.c:EXPORT_SYMBOL(csum_partial_copy_to_user);csum-wrappers.c: * csum_partial_copy_nocheck - Copy and checksum. csum-wrappers.c:csum_partial_copy_nocheck(const void *src, void *dst, int len, __wsum sum)csum-wrappers.c: return csum_partial_copy_generic(src,dst,len,sum,NULL,NULL);
csum-wrappers.c:EXPORT_SYMBOL(csum_partial_copy_nocheck);
OK - maybe this is obvious for you and a few others. But for me I get utterly confused about where to look for the x86_64 version.
Diving into the Makefile I can figure it out.
But thats one indirection too much.
As an example where this plays out better are in x86/crypto. When grepping for aes_dec_blk I got following output:
aes_32.c:asmlinkage void aes_dec_blk(struct crypto_tfm *tfm, u8 *dst, const u8 *src);aes_32.c: aes_dec_blk(tfm, dst, src);
aes_64.c:asmlinkage void aes_dec_blk(struct crypto_tfm *tfm, u8 *out, const u8 *in);aes_64.c: aes_dec_blk(tfm, dst, src);
aes-i586-asm.S:/* void aes_dec_blk(struct crypto_tfm *tfm, u8 *out_blk, const u8 *in_blk) */ aes-i586-asm.S:.global aes_dec_blk
aes-i586-asm.S:aes_dec_blk:
aes-x86_64-asm.S:/* void aes_dec_blk(struct crypto_tfm *tfm, u8 *out, const u8 *in) */aes-x86_64-asm.S: entry(aes_dec_blk,240,dec128,dec192)
See how obvious it is what are x86_64 specific and what are i386 specific. And that despite the mixed naming convention used.
All files that _truely_ belongs to only one of the two architectures ought to be named such that this is obvious when grepping like the above examples shows.
Using the wordsize to distingush the filename seems to cause confusion since there are files that _truely_ only belongs to i386 but is not 32 bit specific because they exist only on i386 because they are not needed on any x86_64 system.
Sam