Skip to contentSkip to conversation

missing includes from infiniband merge

PATCH Applied Al Viro 3 participants 10 messages Sep 23, 2006 – Sep 24, 2006 InfiniBand

lore
Share & tools
lore

Al Viro

author
#0 lore raw

PATCH 1/1 missing includes from infiniband merge

indirect chains of includes are arch-specific and can't be relied upon... (hell, even attempt to build it for itanic would trigger vmalloc.h ones; err.h triggers on e.g. alpha).

---
 drivers/infiniband/core/mad_priv.h           |    1 +
 drivers/infiniband/hw/amso1100/c2_provider.c |    1 +
 drivers/infiniband/hw/amso1100/c2_rnic.c     |    1 +
 drivers/infiniband/hw/ipath/ipath_diag.c     |    1 +
 4 files changed, 4 insertions(+), 0 deletions(-)
drivers/infiniband/core/mad_priv.h +1 -0
index 1da9adb..d06b590 100644
@@ -38,6 +38,7 @@ #ifndef __IB_MAD_PRIV_H__
 #define __IB_MAD_PRIV_H__
 
 #include <linux/completion.h>
+#include <linux/err.h>
 #include <linux/pci.h>
 #include <linux/workqueue.h>
 #include <rdma/ib_mad.h>
drivers/infiniband/hw/amso1100/c2_provider.c +1 -0
index 8fddc8c..dd6af55 100644
@@ -49,6 +49,7 @@ #include <linux/tcp.h>
 #include <linux/init.h>
 #include <linux/dma-mapping.h>
 #include <linux/if_arp.h>
+#include <linux/vmalloc.h>
 
 #include <asm/io.h>
 #include <asm/irq.h>
drivers/infiniband/hw/amso1100/c2_rnic.c +1 -0
index 1c3c9d6..f49a32b 100644
@@ -50,6 +50,7 @@ #include <linux/init.h>
 #include <linux/dma-mapping.h>
 #include <linux/mm.h>
 #include <linux/inet.h>
+#include <linux/vmalloc.h>
 
 #include <linux/route.h>
 
drivers/infiniband/hw/ipath/ipath_diag.c +1 -0
index 28b6b46..29958b6 100644
@@ -43,6 +43,7 @@
 
 #include <linux/io.h>
 #include <linux/pci.h>
+#include <linux/vmalloc.h>
 #include <asm/uaccess.h>
 
 #include "ipath_kernel.h"
Signature · 1 line
1.4.2.GIT
#1 lore raw

Thanks, applied to my tree.

Sorry about that -- my cross-compilation testbed broke at an inopportune moment.

- R.

#2 lore raw
In reply to Al Viro
Quoted 5 lines · Al Viro

indirect chains of includes are arch-specific and can't be relied upon... (hell, even attempt to build it for itanic would trigger vmalloc.h ones; err.h triggers on e.g. alpha).

Quoted 7 lines
---
 drivers/infiniband/core/mad_priv.h           |    1 +
 drivers/infiniband/hw/amso1100/c2_provider.c |    1 +
 drivers/infiniband/hw/amso1100/c2_rnic.c     |    1 +
 drivers/infiniband/hw/ipath/ipath_diag.c     |    1 +
 4 files changed, 4 insertions(+), 0 deletions(-)

A better fix would be to avoid the arch dependency in the non-arch .h files so that in most cases it just works??

That will result in a few files being included twice or more but does that matter with current gcc - I doubt.

Sam

author
#3 lore raw
Quoted 13 lines · Sam Ravnborg
Quoted 5 lines · Al Viro

indirect chains of includes are arch-specific and can't be relied upon... (hell, even attempt to build it for itanic would trigger vmalloc.h ones; err.h triggers on e.g. alpha).

Quoted 7 lines
---
 drivers/infiniband/core/mad_priv.h           |    1 +
 drivers/infiniband/hw/amso1100/c2_provider.c |    1 +
 drivers/infiniband/hw/amso1100/c2_rnic.c     |    1 +
 drivers/infiniband/hw/ipath/ipath_diag.c     |    1 +
 4 files changed, 4 insertions(+), 0 deletions(-)
Quoted 2 lines

A better fix would be to avoid the arch dependency in the non-arch .h files so that in most cases it just works??

What "it"? Use of vmalloc() without including vmalloc.h since on i386 it just happens to be pulled via the
linux/pci.h -> linux/dmapool.h -> asm-i386/io.h -> linux/vmalloc.h chain?

Should we replicate it on every platform? Along with all kinds of fun that's going to cause wrt ordering, BTW...

author
#4 lore raw
Quoted 14 lines · Al Viro
Quoted 13 lines · Sam Ravnborg
Quoted 12 lines · Al Viro

indirect chains of includes are arch-specific and can't be relied upon... (hell, even attempt to build it for itanic would trigger vmalloc.h ones; err.h triggers on e.g. alpha).

---
 drivers/infiniband/core/mad_priv.h           |    1 +
 drivers/infiniband/hw/amso1100/c2_provider.c |    1 +
 drivers/infiniband/hw/amso1100/c2_rnic.c     |    1 +
 drivers/infiniband/hw/ipath/ipath_diag.c     |    1 +
 4 files changed, 4 insertions(+), 0 deletions(-)
Quoted 7 lines
Quoted 2 lines

A better fix would be to avoid the arch dependency in the non-arch .h files so that in most cases it just works??

What "it"? Use of vmalloc() without including vmalloc.h since on i386 it just happens to be pulled via the
linux/pci.h -> linux/dmapool.h -> asm-i386/io.h -> linux/vmalloc.h chain?

BTW, to do what you suggest we would mean the following:

let H(arch,foo) = {bar | asm-arch/foo.h pulls asm-arch/bar.h via chain that doesn't include linux/*}
let M(arch,foo) = {baz | asm-arch/bar.h includes linux/baz.h for some bar in H(arch,foo)}
let L(foo) = union of M(arch,foo) by all arch.
make sure that for each arch and for each bar in L(foo) we have asm-arch/foo.h pulling linux/bar.h

Have fun doing that and maintaining the results. Note that addition of include between asm-weird-crap/foo.h and asm-weird-crap/bar.h will require changes all over the place in asm-*/* - not only for asm-*/foo.h.

It won't be pretty, it will be hell to even try to clean dependencies _and_ it will break all the time due to ordering requirements.

#5 lore raw
In reply to Al Viro
Quoted 7 lines · Al Viro
Quoted 2 lines

A better fix would be to avoid the arch dependency in the non-arch .h files so that in most cases it just works??

What "it"? Use of vmalloc() without including vmalloc.h since on i386 it just happens to be pulled via the
linux/pci.h -> linux/dmapool.h -> asm-i386/io.h -> linux/vmalloc.h chain?

The other way around. Try to get rid of the evil includes in arch-$(ASM) that is just sitting there for no other purpose than to let a developer skip a single include.
In this case the right fix IMO would have been to kill the include of linux/vmalloc.h from asm-i386/io.h and let all users that previously failed to include vmalloc.h now do so themself.

Looking through asm-i386/io.h at fist look there is zero use of linux/vmalloc.h so the include has no business there.

With this your patch would obviously be needed and on top of this we would have to fix other places that 'forget' to include vmalloc.h but the good thing would be that this is now a bit more consistent across architectures.

Sam

author
#6 lore raw
Quoted 15 lines · Sam Ravnborg
Quoted 7 lines · Al Viro
Quoted 2 lines

A better fix would be to avoid the arch dependency in the non-arch .h files so that in most cases it just works??

What "it"? Use of vmalloc() without including vmalloc.h since on i386 it just happens to be pulled via the
linux/pci.h -> linux/dmapool.h -> asm-i386/io.h -> linux/vmalloc.h chain?

The other way around. Try to get rid of the evil includes in arch-$(ASM) that is just sitting there for no other purpose than to let a developer skip a single include.
In this case the right fix IMO would have been to kill the include of linux/vmalloc.h from asm-i386/io.h and let all users that previously failed to include vmalloc.h now do so themself.

Quoted 2 lines

Looking through asm-i386/io.h at fist look there is zero use of linux/vmalloc.h so the include has no business there.

There are obvious asm/page.h uses, so just ripping it out won't be enough. Even for that particular case. And we have shitloads of places were asm-foo/bar.h genuinely needs linux/baz.h for e.g. implementation of an inlined helper. With other targets not needing it at all. Would you mandate including it from every user of asm/foo.h? And maintain such rules afterwards ("asm/foo.h needs linux/baz.h included before it since on $WEIRD_TARGET we include asm/unique_turd.h that won't compile unless linux/baz.h will be aready there").

#7 lore raw
Quoted 4 lines · Al Viro
Quoted 3 lines

Looking through asm-i386/io.h at fist look there is zero use of linux/vmalloc.h so the include has no business there.

Quoted 8 lines

There are obvious asm/page.h uses, so just ripping it out won't be enough. Even for that particular case. And we have shitloads of places were asm-foo/bar.h genuinely needs linux/baz.h for e.g. implementation of an inlined helper. With other targets not needing it at all. Would you mandate including it from every user of asm/foo.h? And maintain such rules afterwards ("asm/foo.h needs linux/baz.h included before it since on $WEIRD_TARGET we include asm/unique_turd.h that won't compile unless linux/baz.h will be aready there").

The only thing I like to see is minimal suprise. And minimal suprise in this case is to be considered as "works on almost all archs if not all". In practical terms it could be that users of asm/* had to include baz.h before bar.h. Or we could stick to current mess where one has to have a shitload of crosscompiles and CPU power to check even trivial changes to a few include files.

Partly this could be fixed by making header files in asm-$(ARCH) second class citizen - that always got included via their linux/ counterpart.

Take a look at uaccess.h for example.
A grep shows that most architectures include almost the same header files with a few that require string.h.
So it would be so simple to move the include of string.h to the linux/uaccess.h counter part and no arch specific dependencies.

Grepping the tree I only find one user of linux/uaccess.h : filemap.* But even though my point holds that there are easy way to deal with this without the maintenace nightmare you try to picture up.

That said this would maybe work in two third of the cases and is no bullet proof solution. But each time we can decrease the arch differences we are one step in the right direction.

Sam

author
#8 lore raw
Quoted 4 lines · Sam Ravnborg

The only thing I like to see is minimal suprise. And minimal suprise in this case is to be considered as "works on almost all archs if not all". In practical terms it could be that users of asm/* had to include baz.h before bar.h.

Even though dependency on baz.h is due to implementation of one of the helpers in bar.h on a few targets and makes no sense whatsoever for everything else?

Quoted 3 lines

Or we could stick to current mess where one has
to have a shitload of crosscompiles and CPU power to check even trivial changes to a few include files.

We _are_ stuck with it.

Quoted 3 lines

Partly this could be fixed by making header files in asm-$(ARCH) second class citizen - that always got included via their linux/ counterpart.

Which only makes dependency graph fatter... What's the difference between including asm/uaccess.h and linux/uaccess.h?

Basically, you pull tons of includes into linux/blah.h because it happens to include asm/foo.h and _that_ depends on having linux/barf.h for $WEIRD_TARGET. And guess what? You are back to the same cross-compiles, since attempt to remove blah.h -> barf.h will break $WEIRD_TARGET, but you won't notice that unless you cross-compile for it.

So all you get is a bunch of harder to explain includes between linux/*.h, _and_ extra dependencies that make no sense whatsoever.

#9 lore raw
Quoted 5 lines
Quoted 3 lines

Or we could stick to current mess where one has
to have a shitload of crosscompiles and CPU power to check even trivial changes to a few include files.

We _are_ stuck with it.

For old stuff - to a high degree yes.

Quoted 6 lines
Quoted 3 lines

Partly this could be fixed by making header files in asm-$(ARCH) second class citizen - that always got included via their linux/ counterpart.

Which only makes dependency graph fatter... What's the difference between including asm/uaccess.h and linux/uaccess.h?

It makes it fatter on the horizontal level - yes.
But then files do not rely on files being included by asm-* files anymore - so less difference between architecutes.

Quoted 3 lines

Basically, you pull tons of includes into linux/blah.h because it happens to include asm/foo.h and _that_ depends on having linux/barf.h for $WEIRD_TARGET.

That would be wrong. Any file in asm-* that needs a file from linux/ should pull that include to the linux/ counterpart and in this way we are on track again.

It boils down to something simple as:
Do we want to keep arch independence lower on the price of higher number includes in asm- counter part files in linux/?

Quoted 3 lines

And guess what? You are back to the same cross-compiles, since attempt to remove blah.h -> barf.h will break $WEIRD_TARGET, but you won't notice that unless you cross-compile for it.

No.
There is a huge difference modifying architecture independent code like infiniband or module.c and to try to clean up headers. Cleaning up headers really really require that one cross compile a lot no matter how files are organized.

Quoted 2 lines

So all you get is a bunch of harder to explain includes between linux/*.h, _and_ extra dependencies that make no sense whatsoever.

They are easy to explain: They are there to keep arch independence lower - and the cost is a number (tons?) of extra includes.

OK - this was a fun thread talking about arch specific includes. But since I do not plan to follow-up with patches I will drop of for now. bugzilla and others tell me I should concentrate on other things.

Sam