fanotify_mark()

fanotify_mark()

[Jeroen Roovers]

[-- Attachment #1: Type: text/plain, Size: 346 bytes --]

Trying out systemd I found that libc 2.17 does not build
fanotify_mark() support[1]. When I got it to do that (see attached
patch, which is probably incorrect), calls to fanotify_mark() still
failed on a C8000 (64-bit kernel). I am sure I am missing something
obvious.


Kind regards,
     jer


[1] https://bugs.gentoo.org/show_bug.cgi?id=480268

[-- Attachment #2: glibc-2.17-hppa-fanotify_mark.patch --]
[-- Type: text/plain, Size: 383 bytes --]

--- a/ports/sysdeps/unix/sysv/linux/hppa/syscalls.list
+++ b/ports/sysdeps/unix/sysv/linux/hppa/syscalls.list
@@ -36,3 +36,4 @@
 setrlimit	-	setrlimit	i:ip	__setrlimit	setrlimit	
 getrlimit	-	getrlimit	i:ip	__getrlimit	getrlimit	
 prlimit64	EXTRA	prlimit64	i:iipp	__prlimit64	prlimit64@@GLIBC_2.17
+fanotify_mark	EXTRA	fanotify_mark	i:iiiis	__fanotify_mark fanotify_mark@@GLIBC_2.17

Re: fanotify_mark()

[Helge Deller]

On 08/13/2013 03:49 PM, Jeroen Roovers wrote:
> Trying out systemd I found that libc 2.17 does not build
> fanotify_mark() support[1]. When I got it to do that (see attached
> patch, which is probably incorrect), calls to fanotify_mark() still
> failed on a C8000 (64-bit kernel). I am sure I am missing something
> obvious.
> 
> [1] https://bugs.gentoo.org/show_bug.cgi?id=480268

In the Linux kernel we use the compat layer for 32bit syscalls on 64bit kernel:
arch/parisc/kernel/syscall_table.S:     ENTRY_COMP(fanotify_mark)
which means we call (with 32bit userspace on 64bit kernel):
fs/notify/fanotify/fanotify_user.c:COMPAT_SYSCALL_DEFINE6(fanotify_mark, int, fanotify_fd, unsigned int, flags,
                __u32, mask0, __u32, mask1, int, dfd, const char  __user *, pathname)
which has 6 parameters.

So, I assume you need one more "i" in this line here (change i:iiiis -> i:iiiiis {to 5 i's}):
> +fanotify_mark	EXTRA	fanotify_mark	i:iiiis	__fanotify_mark fanotify_mark@@GLIBC_2.17

Just an assumption...

Helge

Re: fanotify_mark()

[Jeroen Roovers]

On Tue, 13 Aug 2013 23:29:42 +0200
Helge Deller <deller@gmx.de> wrote:

> > [1] https://bugs.gentoo.org/show_bug.cgi?id=480268
> 
> In the Linux kernel we use the compat layer for 32bit syscalls on
> 64bit kernel: arch/parisc/kernel/syscall_table.S:
> ENTRY_COMP(fanotify_mark) which means we call (with 32bit userspace
> on 64bit kernel):
> fs/notify/fanotify/fanotify_user.c:COMPAT_SYSCALL_DEFINE6(fanotify_mark,
> int, fanotify_fd, unsigned int, flags, __u32, mask0, __u32, mask1,
> int, dfd, const char  __user *, pathname) which has 6 parameters.
> 
> So, I assume you need one more "i" in this line here (change i:iiiis
> -> i:iiiiis {to 5 i's}):

I tried that too with a similarly bad result (I forgot to keep a
binary tarball of the packages glibc for easy switching, and I didn't
keep the strace output of the first attempt either).

As noted on the bug report. The bug report lists the relevant strace
output of the delightfully simple implementation of fanotify_mark in
the fatrace[1] tool as well:

# grep fanotif fatrace.strace
361   fanotify_init(0, 0)               = 3
361   fanotify_mark(0x3, 0x11, 0x4800003b, 0, 0xffffff9c) = -1 EINVAL
(Invalid argument)

I have since tried with a 32-bit kernel on a C3600, and there I get a
different one:

# grep fanotif fatrace.strace
2810  fanotify_init(0, 0)               = 3
2810  fanotify_mark(0x3, 0x11, 0x4800003b, 0, 0xffffff9c) = -1 EBADF
(Bad file descriptor)


     jer


[1] https://launchpad.net/fatrace

Re: fanotify_mark()

[Jeroen Roovers]

On Wed, 14 Aug 2013 00:54:00 +0200
Jeroen Roovers <jer@gentoo.org> wrote:

> On Tue, 13 Aug 2013 23:29:42 +0200
> Helge Deller <deller@gmx.de> wrote:
> 
> > > [1] https://bugs.gentoo.org/show_bug.cgi?id=480268
> > 
> > In the Linux kernel we use the compat layer for 32bit syscalls on
> > 64bit kernel: arch/parisc/kernel/syscall_table.S:
> > ENTRY_COMP(fanotify_mark) which means we call (with 32bit userspace
> > on 64bit kernel):
> > fs/notify/fanotify/fanotify_user.c:COMPAT_SYSCALL_DEFINE6(fanotify_mark,
> > int, fanotify_fd, unsigned int, flags, __u32, mask0, __u32, mask1,
> > int, dfd, const char  __user *, pathname) which has 6 parameters.
> > 
> > So, I assume you need one more "i" in this line here (change i:iiiis
> > -> i:iiiiis {to 5 i's}):
> 
> I tried that too with a similarly bad result

After the 5 i patch =and= recompiling fatrace, fanotify_mark() now works
properly. Which brings me no closer to systemd support on HPPA, I'm
afraid, but at least systemd now installs properly. I'm working on
writing up another e-mail about the next hurdle[1] systemd presents us.


     jer


[1] Sneak preview: https://bugs.gentoo.org/show_bug.cgi?id=482214

Re: fanotify_mark()

[Helge Deller]

On 08/26/2013 01:46 AM, Jeroen Roovers wrote:
> On Wed, 14 Aug 2013 00:54:00 +0200
> Jeroen Roovers <jer@gentoo.org> wrote:
> 
>> On Tue, 13 Aug 2013 23:29:42 +0200
>> Helge Deller <deller@gmx.de> wrote:
>>
>>>> [1] https://bugs.gentoo.org/show_bug.cgi?id=480268
>>>
>>> In the Linux kernel we use the compat layer for 32bit syscalls on
>>> 64bit kernel: arch/parisc/kernel/syscall_table.S:
>>> ENTRY_COMP(fanotify_mark) which means we call (with 32bit userspace
>>> on 64bit kernel):
>>> fs/notify/fanotify/fanotify_user.c:COMPAT_SYSCALL_DEFINE6(fanotify_mark,
>>> int, fanotify_fd, unsigned int, flags, __u32, mask0, __u32, mask1,
>>> int, dfd, const char  __user *, pathname) which has 6 parameters.
>>>
>>> So, I assume you need one more "i" in this line here (change i:iiiis
>>> -> i:iiiiis {to 5 i's}):
>>
>> I tried that too with a similarly bad result
> 
> After the 5 i patch =and= recompiling fatrace, fanotify_mark() now works
> properly.

Great!
I didn't checked, but will the "5 i" work on 32bit too? Don't think so...

> Which brings me no closer to systemd support on HPPA, I'm
> afraid, but at least systemd now installs properly. I'm working on
> writing up another e-mail about the next hurdle[1] systemd presents us.
> [1] Sneak preview: https://bugs.gentoo.org/show_bug.cgi?id=482214

Did you already filed this signal-problem upstream as suggested in comment #3 (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?

Helge

Re: fanotify_mark()

[John David Anglin]

On 26-Aug-13, at 5:27 PM, Helge Deller wrote:

>> Which brings me no closer to systemd support on HPPA, I'm
>> afraid, but at least systemd now installs properly. I'm working on
>> writing up another e-mail about the next hurdle[1] systemd presents  
>> us.
>> [1] Sneak preview: https://bugs.gentoo.org/show_bug.cgi?id=482214
>
> Did you already filed this signal-problem upstream as suggested in  
> comment #3 (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?


I believe two of the signal numbers come from HP-UX.

Dave
--
John David Anglin	dave.anglin@bell.net

Re: fanotify_mark()

[Jeroen Roovers]

On Mon, 26 Aug 2013 23:27:39 +0200
Helge Deller <deller@gmx.de> wrote:

> > After the 5 i patch =and= recompiling fatrace, fanotify_mark() now
> > works properly.
> 
> Great!
> I didn't checked, but will the "5 i" work on 32bit too? Don't think
> so...

Yes, that works too.


     jer

systemd real-time signals choices clash with Linux/PARISC available SIGRT range, WAS: fanotify_mark()

[Jeroen Roovers]

On Mon, 26 Aug 2013 18:37:54 -0400
John David Anglin <dave.anglin@bell.net> wrote:

> On 26-Aug-13, at 5:27 PM, Helge Deller wrote:
> >> [1] Sneak preview: https://bugs.gentoo.org/show_bug.cgi?id=482214
> >
> > Did you already filed this signal-problem upstream as suggested in  
> > comment #3 (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?

Well, there are two ways to resolve this problem, and seeing who is
developing systemd and seeing the generous way the "available" signal
range is used, I'm pretty doubtful about changes there.

As I said in comment #2, that range could be compacted (a lot) and then
fit easily on any future platform. Since it was a design choice even
reflected in man pages[1] ("[...], SIGRTMIN+29   Sets the log level to
[...]"), I'm very afraid they will not change it easily.

> I believe two of the signal numbers come from HP-UX.

#define SIGXCPU         33
#define SIGXFSZ         34
#define SIGSTKFLT       36

According to [2], SIGSTKFLT isn't used.

Do we still support HP-UX? I have never seen a binary for it I could
try with, but then maybe I never went looking for one, either. :)


     jer


[1] http://www.freedesktop.org/software/systemd/man/systemd.html
[2]
http://h21007.www2.hp.com/portal/download/files/unprot/STK/Linux_STK/impacts/i60.html

Re: systemd real-time signals choices clash with Linux/PARISC available SIGRT range, WAS: fanotify_mark()

[John David Anglin]

On 8/27/2013 10:46 AM, Jeroen Roovers wrote:
> On Mon, 26 Aug 2013 18:37:54 -0400
> John David Anglin <dave.anglin@bell.net> wrote:
>
>> On 26-Aug-13, at 5:27 PM, Helge Deller wrote:
>>>> [1] Sneak preview: https://bugs.gentoo.org/show_bug.cgi?id=482214
>>> Did you already filed this signal-problem upstream as suggested in
>>> comment #3 (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?
> Well, there are two ways to resolve this problem, and seeing who is
> developing systemd and seeing the generous way the "available" signal
> range is used, I'm pretty doubtful about changes there.
>
> As I said in comment #2, that range could be compacted (a lot) and then
> fit easily on any future platform. Since it was a design choice even
> reflected in man pages[1] ("[...], SIGRTMIN+29   Sets the log level to
> [...]"), I'm very afraid they will not change it easily.
>
>> I believe two of the signal numbers come from HP-UX.
> #define SIGXCPU         33
> #define SIGXFSZ         34
The signal numbers for these two signals come from HP-UX but the signals 
are used
by Linux, so I can't see how they can change.
> #define SIGSTKFLT       36
>
> According to [2], SIGSTKFLT isn't used.
This signal isn't used by HP-UX but it is used by Linux, so again this 
can't change.

Can we change _NSIG to 69 so there are 32 RT signals as on other arches?

>
> Do we still support HP-UX? I have never seen a binary for it I could
> try with, but then maybe I never went looking for one, either. :)
Doesn't really matter...
>
>
>       jer
>
>
> [1] http://www.freedesktop.org/software/systemd/man/systemd.html
> [2]
> http://h21007.www2.hp.com/portal/download/files/unprot/STK/Linux_STK/impacts/i60.html
> --
> To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
>


-- 
John David Anglin    dave.anglin@bell.net

Re: systemd real-time signals choices clash with Linux/PARISC available SIGRT range, WAS: fanotify_mark()

[John David Anglin]

On 8/27/2013 11:26 AM, John David Anglin wrote:
> Can we change _NSIG to 69 so there are 32 RT signals as on other arches?
It looks like it needs to be a power of 2.  MIPS uses 128.

Dave

-- 
John David Anglin    dave.anglin@bell.net

Re: systemd real-time signals choices clash with Linux/PARISC available SIGRT range, WAS: fanotify_mark()

[Helge Deller]

On 08/27/2013 05:26 PM, John David Anglin wrote:
> On 8/27/2013 10:46 AM, Jeroen Roovers wrote:
>> On Mon, 26 Aug 2013 18:37:54 -0400 John David Anglin
>> <dave.anglin@bell.net> wrote:
>> 
>>> On 26-Aug-13, at 5:27 PM, Helge Deller wrote:
>>>>> [1] Sneak preview:
>>>>> https://bugs.gentoo.org/show_bug.cgi?id=482214
>>>> Did you already filed this signal-problem upstream as suggested
>>>> in comment #3
>>>> (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?
>> Well, there are two ways to resolve this problem, and seeing who
>> is developing systemd and seeing the generous way the "available"
>> signal range is used, I'm pretty doubtful about changes there.

I can see why they wouldn't like to change this again.
It might break systemd binaries on other existing arches which already 
support systemd.

Anyway, has there been any further upstream discussion?

>> As I said in comment #2, that range could be compacted (a lot) and
>> then fit easily on any future platform. Since it was a design
>> choice even reflected in man pages[1] ("[...], SIGRTMIN+29   Sets
>> the log level to [...]"), I'm very afraid they will not change it
>> easily.
>> 
>>> I believe two of the signal numbers come from HP-UX.
>> #define SIGXCPU         33 #define SIGXFSZ         34
> The signal numbers for these two signals come from HP-UX but the
> signals are used by Linux, so I can't see how they can change.
>> #define SIGSTKFLT       36
>> 
>> According to [2], SIGSTKFLT isn't used.
> This signal isn't used by HP-UX but it is used by Linux, so again
> this can't change.

changing those would break our binary compatibility...
 
> Can we change _NSIG to 69 so there are 32 RT signals as on other
> arches? [Dave followup]: It looks like it needs to be a power of 2.
> MIPS uses 128.

127 seems to be the better choice then... (signal #128 seems to conflict with
the core dump bit)

So, how should we proceed here?
If I haven't overlooked something, it seems that changing the parisc kernel
 & glibc is possible.
Downside is, that people of course need newer kernel/glibc and at least 
systemd recompiled.
Another downside is that we add more overhead in the signal paths, but I'm not
sure if this really hurts performance (is signal handling performance-relevant
at all?).

I can come up with a kernel/glibc patch (which seem to be trivial) if we agree 
to really go that way...

Helge

>> [1] http://www.freedesktop.org/software/systemd/man/systemd.html 
>> [2]
>> http://h21007.www2.hp.com/portal/download/files/unprot/STK/Linux_STK/impacts/i60.html

Re: systemd real-time signals choices clash with Linux/PARISC available SIGRT range, WAS: fanotify_mark()

[John David Anglin]

Regarding _NSIG, I realized after I wrote that the number has to be 
divisible by 64 because
of the way various words are allocated (one bit per signal). There is no 
roundup in the allocation.
We need to tweak the RT max value so we at least 32 RT signals. This 
should keep userspace
happy.  This will leave a whole bunch of numbers free.  So, probably 
there wouldn't be a conflict
with the core dump bit.

I had the sense the glibc doesn't need to change because of the way 
POSIX specifies
the number of available realtime signals.

Dave

On 9/18/2013 2:40 PM, Helge Deller wrote:
> On 08/27/2013 05:26 PM, John David Anglin wrote:
>> On 8/27/2013 10:46 AM, Jeroen Roovers wrote:
>>> On Mon, 26 Aug 2013 18:37:54 -0400 John David Anglin
>>> <dave.anglin@bell.net> wrote:
>>>
>>>> On 26-Aug-13, at 5:27 PM, Helge Deller wrote:
>>>>>> [1] Sneak preview:
>>>>>> https://bugs.gentoo.org/show_bug.cgi?id=482214
>>>>> Did you already filed this signal-problem upstream as suggested
>>>>> in comment #3
>>>>> (https://bugs.gentoo.org/show_bug.cgi?id=482214#c3)?
>>> Well, there are two ways to resolve this problem, and seeing who
>>> is developing systemd and seeing the generous way the "available"
>>> signal range is used, I'm pretty doubtful about changes there.
> I can see why they wouldn't like to change this again.
> It might break systemd binaries on other existing arches which already
> support systemd.
>
> Anyway, has there been any further upstream discussion?
>
>>> As I said in comment #2, that range could be compacted (a lot) and
>>> then fit easily on any future platform. Since it was a design
>>> choice even reflected in man pages[1] ("[...], SIGRTMIN+29   Sets
>>> the log level to [...]"), I'm very afraid they will not change it
>>> easily.
>>>
>>>> I believe two of the signal numbers come from HP-UX.
>>> #define SIGXCPU         33 #define SIGXFSZ         34
>> The signal numbers for these two signals come from HP-UX but the
>> signals are used by Linux, so I can't see how they can change.
>>> #define SIGSTKFLT       36
>>>
>>> According to [2], SIGSTKFLT isn't used.
>> This signal isn't used by HP-UX but it is used by Linux, so again
>> this can't change.
> changing those would break our binary compatibility...
>   
>> Can we change _NSIG to 69 so there are 32 RT signals as on other
>> arches? [Dave followup]: It looks like it needs to be a power of 2.
>> MIPS uses 128.
> 127 seems to be the better choice then... (signal #128 seems to conflict with
> the core dump bit)
>
> So, how should we proceed here?
> If I haven't overlooked something, it seems that changing the parisc kernel
>   & glibc is possible.
> Downside is, that people of course need newer kernel/glibc and at least
> systemd recompiled.
> Another downside is that we add more overhead in the signal paths, but I'm not
> sure if this really hurts performance (is signal handling performance-relevant
> at all?).
>
> I can come up with a kernel/glibc patch (which seem to be trivial) if we agree
> to really go that way...
>
> Helge
>
>>> [1] http://www.freedesktop.org/software/systemd/man/systemd.html
>>> [2]
>>> http://h21007.www2.hp.com/portal/download/files/unprot/STK/Linux_STK/impacts/i60.html
> --
> To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
>


-- 
John David Anglin    dave.anglin@bell.net