[PATCH 0/2] selftests/userfaultfd: fix build

These two patches fix building the userfaultfd test tool for me.
The first patch fixes a build error I experienced by removing a kernel
header from an uapi header file.
The second fix adds the syscall numbers for other architectures,
fixes a typo and makes the whole syscall numbering more future-proof.

Tested on x86, arm and arm64.

Andre Przywara (2):
  userfaultfd: remove kernel header include from uapi header
  selftests/userfaultfd: improve syscall number definition

 include/uapi/linux/userfaultfd.h         |  2 --
 tools/testing/selftests/vm/userfaultfd.c | 13 ++++++++++++-
 2 files changed, 12 insertions(+), 3 deletions(-)

-- 
2.5.1

At the moment the userfaultfd test program only supports x86 and an
architecture called "powewrpc" ;-)
Fix that typo and add the syscall numbers for other architectures as
well.
Also as in an ideal world a syscall number should come from the system
header file <asm/unistd.h>, include that header and guard the explicit
syscall number section below to avoid redefinitions.

Signed-off-by: Andre Przywara <redacted>
---
 tools/testing/selftests/vm/userfaultfd.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index 2c7cca6..63be27f 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c

@@ -65,16 +65,27 @@
 #include <sys/ioctl.h>
 #include <pthread.h>
 #include "../../../../include/uapi/linux/userfaultfd.h"
+#include <asm/unistd.h>
 
+/* ideally the above user header provides that number, but ... */
+#ifndef __NR_userfaultfd
 #ifdef __x86_64__
 #define __NR_userfaultfd 323
 #elif defined(__i386__)
 #define __NR_userfaultfd 374
-#elif defined(__powewrpc__)
+#elif defined(__powerpc__)
 #define __NR_userfaultfd 364
+#elif defined(__ia64__)
+#define __NR_userfaultfd 1343
+#elif defined(__arm__)
+#define __NR_userfaultfd 388
+#elif defined(__aarch64__)
+/* this is from the generic syscall table, used also on other architectures */
+#define __NR_userfaultfd 283
 #else
 #error "missing __NR_userfaultfd definition"
 #endif
+#endif /* !__NR_userfaultfd */
 
 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
 

-- 
2.5.1

As include/uapi/linux/userfaultfd.h is a user visible header file,
it should not include kernel-exclusive header files.
So trying to build the userfaultfd test program from the selftests
directory fails, since it contains a reference to linux/compiler.h.
As it turns out, that header is not really needed there, so we can
simply remove it to fix that issue.

Signed-off-by: Andre Przywara <redacted>
---
 include/uapi/linux/userfaultfd.h | 2 --
 1 file changed, 2 deletions(-)

diff --git a/include/uapi/linux/userfaultfd.h b/include/uapi/linux/userfaultfd.h
index df0e09b..9057d7a 100644
--- a/include/uapi/linux/userfaultfd.h
+++ b/include/uapi/linux/userfaultfd.h

@@ -11,8 +11,6 @@
 
 #include <linux/types.h>
 
-#include <linux/compiler.h>
-
 #define UFFD_API ((__u64)0xAA)
 /*
  * After implementing the respective features it will become:

-- 
2.5.1

On 09/22/2015 04:45 AM, Andre Przywara wrote:

This is not okay. User-space shouldn't be (re)defining/duplicating
syscall numbers. I can't take this patch.

thanks,
-- Shuah

-- 
Shuah Khan
Sr. Linux Kernel Developer
Open Source Innovation Group
Samsung Research America (Silicon Valley)
shuahkh-JPH+aEBZ4P+UEJcrhfAQsw@public.gmane.org | (970) 217-8978

On Tue, Sep 22, 2015 at 07:49:13AM -0600, Shuah Khan wrote:

On 09/22/2015 04:45 AM, Andre Przywara wrote:

quoted

At the moment the userfaultfd test program only supports x86 and an
architecture called "powewrpc" ;-)
Fix that typo and add the syscall numbers for other architectures as
well.
Also as in an ideal world a syscall number should come from the system
header file <asm/unistd.h>, include that header and guard the explicit
syscall number section below to avoid redefinitions.

Signed-off-by: Andre Przywara <redacted>
---
 tools/testing/selftests/vm/userfaultfd.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index 2c7cca6..63be27f 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c

@@ -65,16 +65,27 @@
 #include <sys/ioctl.h>
 #include <pthread.h>
 #include "../../../../include/uapi/linux/userfaultfd.h"
+#include <asm/unistd.h>
 
+/* ideally the above user header provides that number, but ... */
+#ifndef __NR_userfaultfd
 #ifdef __x86_64__
 #define __NR_userfaultfd 323
 #elif defined(__i386__)
 #define __NR_userfaultfd 374
-#elif defined(__powewrpc__)
+#elif defined(__powerpc__)
 #define __NR_userfaultfd 364
+#elif defined(__ia64__)
+#define __NR_userfaultfd 1343
+#elif defined(__arm__)
+#define __NR_userfaultfd 388
+#elif defined(__aarch64__)
+/* this is from the generic syscall table, used also on other architectures */
+#define __NR_userfaultfd 283
 #else
 #error "missing __NR_userfaultfd definition"
 #endif
+#endif /* !__NR_userfaultfd */
 
 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
 

This is not okay. User-space shouldn't be (re)defining/duplicating
syscall numbers. I can't take this patch.

-mm has already been updated to do exactly that. Syscall numbers end
up hardcoded into userland binaries/libs somewhere, so it's not a
bugfix but certainly it's a nice cleanup to remove the whole #ifdef block.

Andre, could you see if linux-next (which includes -mm) works for you
by just running "cd tools/testing/selftests/vm/ && make"? If there's
any further change required could you diff it against linux-next?

Thanks!
Andrea

Hi Shuah, Andrea,

On 22/09/15 15:06, Andrea Arcangeli wrote:

On Tue, Sep 22, 2015 at 07:49:13AM -0600, Shuah Khan wrote:

quoted

On 09/22/2015 04:45 AM, Andre Przywara wrote:

quoted

At the moment the userfaultfd test program only supports x86 and an
architecture called "powewrpc" ;-)
Fix that typo and add the syscall numbers for other architectures as
well.
Also as in an ideal world a syscall number should come from the system
header file <asm/unistd.h>, include that header and guard the explicit
syscall number section below to avoid redefinitions.

Signed-off-by: Andre Przywara <andre.przywara@arm.com>
---
 tools/testing/selftests/vm/userfaultfd.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/tools/testing/selftests/vm/userfaultfd.c b/tools/testing/selftests/vm/userfaultfd.c
index 2c7cca6..63be27f 100644
--- a/tools/testing/selftests/vm/userfaultfd.c
+++ b/tools/testing/selftests/vm/userfaultfd.c

@@ -65,16 +65,27 @@
 #include <sys/ioctl.h>
 #include <pthread.h>
 #include "../../../../include/uapi/linux/userfaultfd.h"
+#include <asm/unistd.h>
 
+/* ideally the above user header provides that number, but ... */
+#ifndef __NR_userfaultfd
 #ifdef __x86_64__
 #define __NR_userfaultfd 323
 #elif defined(__i386__)
 #define __NR_userfaultfd 374
-#elif defined(__powewrpc__)
+#elif defined(__powerpc__)
 #define __NR_userfaultfd 364
+#elif defined(__ia64__)
+#define __NR_userfaultfd 1343
+#elif defined(__arm__)
+#define __NR_userfaultfd 388
+#elif defined(__aarch64__)
+/* this is from the generic syscall table, used also on other architectures */
+#define __NR_userfaultfd 283
 #else
 #error "missing __NR_userfaultfd definition"
 #endif
+#endif /* !__NR_userfaultfd */
 
 static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
 

This is not okay. User-space shouldn't be (re)defining/duplicating
syscall numbers. I can't take this patch.

While I agree that this isn't the right approach from a userland point
of view, I wonder how this is supposed to work for the next few months?
Is everybody required to overwrite their distribution-provided kernel
headers just for compiling this test program?

-mm has already been updated to do exactly that. Syscall numbers end
up hardcoded into userland binaries/libs somewhere, so it's not a
bugfix but certainly it's a nice cleanup to remove the whole #ifdef block.

Andre, could you see if linux-next (which includes -mm) works for you
by just running "cd tools/testing/selftests/vm/ && make"? If there's
any further change required could you diff it against linux-next?

This doesn't compile now for me, because it looks into
/usr/include/asm/unistd.h, which I keep to the distribution copy of it.
Also linux/userfaultfd.h is missing, because it's brand new.

If that tool lives in the kernel repo, it should be able to either use
the uapi headers directly or hardcode the syscall numbers - strictly
it's not a sane userland program anymore, but for that kind of tools I
deem it's totally acceptable. I think this is one rationale for keeping
it inside the linux.git repo.

Obviously you were facing the same problem in the beginning (looking at
the original code), so I was just extending the original solution to
cover more architectures and prepare for the time when those symbols
start to appear in distributions.

I guess the right solution would be to hack the Makefile to set the
include path to the kernel's copy of include/uapi, though I am not sure
this works cleanly for different architectures and separate build
directories. I will give this a try ...

Cheers,
Andre.

On Tue, 2015-09-22 at 18:15 +0100, Andre Przywara wrote:

On 22/09/15 15:06, Andrea Arcangeli wrote:

quoted

Andre, could you see if linux-next (which includes -mm) works for you
by just running "cd tools/testing/selftests/vm/ && make"? If there's
any further change required could you diff it against linux-next?

This doesn't compile now for me, because it looks into
/usr/include/asm/unistd.h, which I keep to the distribution copy of it.
Also linux/userfaultfd.h is missing, because it's brand new.

...

I guess the right solution would be to hack the Makefile to set the
include path to the kernel's copy of include/uapi, though I am not sure
this works cleanly for different architectures and separate build
directories. I will give this a try ...

Not that's not the right solution.

The right solution is to export the kernel headers, the Makefile will pick them
up (at least in linux-next):

  $ cd linux
  $ make headers_install
  $ ls usr/include/
  asm/  asm-generic/  drm/  linux/  misc/  mtd/  rdma/  scsi/  sound/  uapi/  video/  xen/
  $ make -C tools/testing/selftests TARGETS=vm
  ...
  $ ls tools/testing/selftests/vm/userfaultfd
  tools/testing/selftests/vm/userfaultfd*

cheers

Hi Michael,

On 23/09/15 10:55, Michael Ellerman wrote:

On Tue, 2015-09-22 at 18:15 +0100, Andre Przywara wrote:

quoted

On 22/09/15 15:06, Andrea Arcangeli wrote:

quoted

Andre, could you see if linux-next (which includes -mm) works for you
by just running "cd tools/testing/selftests/vm/ && make"? If there's
any further change required could you diff it against linux-next?

This doesn't compile now for me, because it looks into
/usr/include/asm/unistd.h, which I keep to the distribution copy of it.
Also linux/userfaultfd.h is missing, because it's brand new.

...

quoted

I guess the right solution would be to hack the Makefile to set the
include path to the kernel's copy of include/uapi, though I am not sure
this works cleanly for different architectures and separate build
directories. I will give this a try ...

Not that's not the right solution.

The right solution is to export the kernel headers, the Makefile will pick them
up (at least in linux-next):

Actually that was I was hoping for, I just missed the possibility of
headers_install to put the files back into the kernel tree (I usually
use it with INSTALL_HDR_PATH for packaging).

  $ cd linux
  $ make headers_install
  $ ls usr/include/
  asm/  asm-generic/  drm/  linux/  misc/  mtd/  rdma/  scsi/  sound/  uapi/  video/  xen/
  $ make -C tools/testing/selftests TARGETS=vm
  ...
  $ ls tools/testing/selftests/vm/userfaultfd
  tools/testing/selftests/vm/userfaultfd*

Yes, that works. Now I have just to figure out how to arrange this with
out-of-tree build directories.

Cheers,
Andre.

On Wed, 2015-09-23 at 14:11 +0100, Andre Przywara wrote:

Hi Michael,

On 23/09/15 10:55, Michael Ellerman wrote:

quoted

On Tue, 2015-09-22 at 18:15 +0100, Andre Przywara wrote:

quoted

On 22/09/15 15:06, Andrea Arcangeli wrote:

quoted

Andre, could you see if linux-next (which includes -mm) works for you
by just running "cd tools/testing/selftests/vm/ && make"? If there's
any further change required could you diff it against linux-next?

This doesn't compile now for me, because it looks into
/usr/include/asm/unistd.h, which I keep to the distribution copy of it.
Also linux/userfaultfd.h is missing, because it's brand new.

...

quoted

I guess the right solution would be to hack the Makefile to set the
include path to the kernel's copy of include/uapi, though I am not sure
this works cleanly for different architectures and separate build
directories. I will give this a try ...

Not that's not the right solution.

The right solution is to export the kernel headers, the Makefile will pick them
up (at least in linux-next):

Actually that was I was hoping for, I just missed the possibility of
headers_install to put the files back into the kernel tree (I usually
use it with INSTALL_HDR_PATH for packaging).

OK great.

It does seem that folks are struggling to work out how to do this, so maybe we
can come up with a way to have it automated.

quoted

  $ cd linux
  $ make headers_install
  $ ls usr/include/
  asm/  asm-generic/  drm/  linux/  misc/  mtd/  rdma/  scsi/  sound/  uapi/  video/  xen/
  $ make -C tools/testing/selftests TARGETS=vm
  ...
  $ ls tools/testing/selftests/vm/userfaultfd
  tools/testing/selftests/vm/userfaultfd*

Yes, that works. Now I have just to figure out how to arrange this with
out-of-tree build directories.

Right, I think it doesn't work at the moment.

The makefiles that look for the kernel headers use ../../../../usr/include,
which probably isn't what you want for an out of tree build.

Maybe we should do something like the below.

It allows you to do:

  $ make O=~/work/build/ headers_install
  $ ls ~/work/build/usr/include/
  asm/  asm-generic/  drm/  linux/  misc/  mtd/  rdma/  scsi/  sound/  uapi/  video/  xen/
  $ ls usr/include
  ls: cannot access usr/include: No such file or directory
  $ make -C tools/testing/selftests KERNEL_HEADERS=~/work/build/usr/include/ TARGEST=vm
  $ ls tools/testing/selftests/vm/userfaultfd
  tools/testing/selftests/vm/userfaultfd*

cheers

diff --git a/tools/testing/selftests/Makefile b/tools/testing/selftests/Makefile
index 8922c2155a47..617a76be3889 100644
--- a/tools/testing/selftests/Makefile
+++ b/tools/testing/selftests/Makefile

@@ -41,6 +41,9 @@ override LDFLAGS =
 override MAKEFLAGS =
 endif
 
+KERNEL_HEADERS ?= $(abspath ../../../usr/include)
+export KERNEL_HEADERS
+
 all:
        for TARGET in $(TARGETS); do \
                make -C $$TARGET; \

diff --git a/tools/testing/selftests/vm/Makefile b/tools/testing/selftests/vm/Makefile
index e4bb1de1d526..4f45b9c8dd75 100644
--- a/tools/testing/selftests/vm/Makefile
+++ b/tools/testing/selftests/vm/Makefile

@@ -1,6 +1,6 @@
 # Makefile for vm selftests
 
-CFLAGS = -Wall -I ../../../../usr/include $(EXTRA_CFLAGS)
+CFLAGS = -Wall -I $(KERNEL_HEADERS) $(EXTRA_CFLAGS)
 BINARIES = compaction_test
 BINARIES += hugepage-mmap
 BINARIES += hugepage-shm

@@ -14,12 +14,9 @@ BINARIES += userfaultfd
 all: $(BINARIES)
 %: %.c
        $(CC) $(CFLAGS) -o $@ $^ -lrt
-userfaultfd: userfaultfd.c ../../../../usr/include/linux/kernel.h
+userfaultfd: userfaultfd.c $(KERNEL_HEADERS)/linux/kernel.h
        $(CC) $(CFLAGS) -O2 -o $@ $< -lpthread
 
-../../../../usr/include/linux/kernel.h:
-       make -C ../../../.. headers_install
-
 TEST_PROGS := run_vmtests
 TEST_FILES := $(BINARIES)