Re: linux-2.5.7

Re: linux-2.5.7

[Neil Brown]

On Saturday March 30, viro@math.psu.edu wrote:
> 
> 
> On Sat, 30 Mar 2002, Timothy Murphy wrote:
> 
> > I'm sure this has been recognised,
> > but I would point out that sys_nfsservctl is not "undefined"
> > if NFSD is not chosen.
> > 
> > The following patch to .../arch/i386/kernel/entry.S corrects this,
> > though this is obviously not the right place to put it:
> 
> Wrong fix.  Using weak aliases would do it in cleaner way, but there's
> a sparc64 toolchain bugs that don't allow that.

I cannot see the weak aliases being a real fix either.
If you compile with NFSD as a module, and with CONFIG_KMOD, then the
nfssvc_ctl systemcall is suppose to auto-load nfsd.o.  How can this be
achieved with weak aliases?

NeilBrown

Re: linux-2.5.7

[Keith Owens]

On Sun, 31 Mar 2002 09:48:38 +1000 (EST), 
Neil Brown <neilb@cse.unsw.edu.au> wrote:
>I cannot see the weak aliases being a real fix either.
>If you compile with NFSD as a module, and with CONFIG_KMOD, then the
>nfssvc_ctl systemcall is suppose to auto-load nfsd.o.  How can this be
>achieved with weak aliases?

System calls cannot be in modules.  Linus forbids it (that way lies
"extend and embrace") and at least two architectures (ia64, ppc64)
break when a syscall is in a module.

Re: linux-2.5.7

[Alexander Viro]

On Sun, 31 Mar 2002, Keith Owens wrote:

> On Sun, 31 Mar 2002 09:48:38 +1000 (EST), 
> Neil Brown <neilb@cse.unsw.edu.au> wrote:
> >I cannot see the weak aliases being a real fix either.
> >If you compile with NFSD as a module, and with CONFIG_KMOD, then the
> >nfssvc_ctl systemcall is suppose to auto-load nfsd.o.  How can this be
> >achieved with weak aliases?
> 
> System calls cannot be in modules.  Linus forbids it (that way lies
> "extend and embrace") and at least two architectures (ia64, ppc64)
> break when a syscall is in a module.

Yup.  The logics being:

if we have neither CONFIG_NFSD nor CONFIG_NFSD_MODULE
	sys_nfsservctl() is alias for sys_ni_syscall()
else
	sys_nfsservctl() is defined in fs/nfsct.c and does do_kern_mount()
	with type "nfsd", which triggers autoload if nfsd is modular.  Whether
	nfsd is modular or comipled-in, syscall itself is in kernel (and
	is nothing but a wrapper for write()/read()).

Re: linux-2.5.7

[Alvaro Figueroa]

> Rather, "the ancient version of gcc commonly used on sparc64".

I'm sort of forwading a message I wrote at March 13th that was never
answered and that I think has a huge priority for the sparc64 port.

--
>    GCC 3.0.3 works.
>    
> (Dave S. Miller wrote)
> I won't accept kernel bug reports for people using it though.
> The old egcs64 compiler is the only thing I trust right now, still.

In that case, I would repectfully comment that a special case comment
should be added to the lkml FAQ (chapter 8, section 2), because there,
it is stated:

"
    2. What are the recommended compiler/binutils for building kernels?

* (REG) This depends on the kernel version. Until 26-OCT-2000, gcc
2.7.2.3 was the recommended compiler for all kernels. On this date,
Linus announced that gcc 2.91.66 (aka egcs 1.1.2) is the recommended
compiler for 2.4.x kernels up to version 2.4.9. Gcc 2.95.3 is the
recommended compiler for kernel 2.4.10 and later.

(...)
            Always see the Documentation/Changes file for details.
"

At Documentation/Changes you can read some negative comments about this
version of the compiler, and it even says that is likely that 2.5 will
drop egcs workarrounds.

How will this be managed? What is the recommended compiler for 2.5?

-- 
Alvaro Figueroa

Re: linux-2.5.7

[Richard Henderson]

On Sun, Mar 31, 2002 at 02:54:14AM -0500, Alexander Viro wrote:
> + * What we want is __attribute__((weak,alias("sys_ni_syscall"))),
> + * but it doesn't work on sparc64, so we just do it by hand

Rather, "the ancient version of gcc commonly used on sparc64".


r~

Re: linux-2.5.7

[Alexander Viro]

On Sun, 31 Mar 2002, Andrew Morton wrote:

> Could you remind us what problem this is solving?  The
> #ifdef approach seemed reasonable and there's no indication
> here why weak linkage is needed.

The thing we want here _is_ weak linkage - "return -ENOSYS unless
you have the real thing".  You can emulate that with ifdefs,
but that's what it is - emulation.  IOW, what we want actually
belongs to linker, not compiler.

When file looks like

#ifdef FOO
<lots of stuff>
<function calling that stuff>
#else
<make that function an equivalent of sys_ni_syscall()>
#endif

we are really talking about "make it an alias of sys_ni_syscall() and let
<all this stuff> override that".

Re: linux-2.5.7

[Andrew Morton]

Alexander Viro wrote:
> 
> ...
> +/*
> + * "Conditional" syscalls
> + *
> + * What we want is __attribute__((weak,alias("sys_ni_syscall"))),
> + * but it doesn't work on sparc64, so we just do it by hand
> + */
> +#define cond_syscall(x) asm(".weak\t" #x "\n\t.set\t" #x ",sys_ni_syscall");
> +
> +cond_syscall(sys_nfsservctl)
> +cond_syscall(sys_quotactl)
> +cond_syscall(sys_acct)
> +

Could you remind us what problem this is solving?  The
#ifdef approach seemed reasonable and there's no indication
here why weak linkage is needed.

Weak linkage could perhaps be useful elsewhere.  Maybe this
should be implemented as

	weak_symbol(sym, default_sym)

in some generic header somewhere...

-

Re: linux-2.5.7

[Alexander Viro]

On Sat, 30 Mar 2002, David S. Miller wrote:

>    From: Alexander Viro <viro@math.psu.edu>
>    Date: Sat, 30 Mar 2002 18:41:46 -0500 (EST)
>    
>    	Dave, you've mentioned doing the equivalent of
>    __attribute__((weak,alias("foo")) by hand.  Could you give an example?
>    
> 
> #define make_weak(foo,bar)	__asm__(".weak foo, bar")
> 
> Or however the assembler syntax works.  GLIBC has some header file it
> at least used to use which had macros doing something similar.

OK...  How about the following?

diff -urN C7-0/fs/Makefile C7-1/fs/Makefile
--- C7-0/fs/Makefile	Tue Mar 19 16:05:56 2002
+++ C7-1/fs/Makefile	Sun Mar 31 02:52:35 2002
@@ -16,12 +16,6 @@
 		dcache.o inode.o attr.o bad_inode.o file.o iobuf.o dnotify.o \
 		filesystems.o namespace.o seq_file.o xattr.o libfs.o
 
-ifeq ($(CONFIG_QUOTA),y)
-obj-y += dquot.o
-else
-obj-y += noquot.o
-endif
-
 ifneq ($(CONFIG_NFSD),n)
 ifneq ($(CONFIG_NFSD),)
 obj-y += nfsctl.o
@@ -84,6 +78,8 @@
 obj-y				+= binfmt_script.o
 
 obj-$(CONFIG_BINFMT_ELF)	+= binfmt_elf.o
+
+obj-$(CONFIG_QUOTA) += dquot.o
 
 # persistent filesystems
 obj-y += $(join $(subdir-y),$(subdir-y:%=/%.o))
diff -urN C7-0/fs/dquot.c C7-1/fs/dquot.c
--- C7-0/fs/dquot.c	Tue Feb 19 22:33:03 2002
+++ C7-1/fs/dquot.c	Sun Mar 31 02:52:35 2002
@@ -59,6 +59,7 @@
 #include <linux/tty.h>
 #include <linux/file.h>
 #include <linux/slab.h>
+#include <linux/sysctl.h>
 #include <linux/smp_lock.h>
 #include <linux/init.h>
 
@@ -1240,9 +1241,22 @@
 	return ret;
 }
 
+static ctl_table fs_table[] = {
+	{FS_NRDQUOT, "dquot-nr", &nr_dquots, 2*sizeof(int),
+	 0444, NULL, &proc_dointvec},
+	{},
+};
+
+static ctl_table dquot_table[] = {
+	{CTL_FS, "fs", NULL, 0, 0555, fs_table},
+	{},
+};
+
 static int __init dquot_init(void)
 {
 	int i;
+
+	register_sysctl_table(dquot_table, 0);
 
 	for (i = 0; i < NR_DQHASH; i++)
 		INIT_LIST_HEAD(dquot_hash + i);
diff -urN C7-0/fs/noquot.c C7-1/fs/noquot.c
--- C7-0/fs/noquot.c	Fri May 12 14:21:20 2000
+++ C7-1/fs/noquot.c	Wed Dec 31 19:00:00 1969
@@ -1,15 +0,0 @@
-/* noquot.c: Quota stubs necessary for when quotas are not
- *           compiled into the kernel.
- */
-
-#include <linux/kernel.h>
-#include <linux/types.h>
-#include <linux/errno.h>
-
-int nr_dquots, nr_free_dquots;
-int max_dquots;
-
-asmlinkage long sys_quotactl(int cmd, const char *special, int id, caddr_t addr)
-{
-	return(-ENOSYS);
-}
diff -urN C7-0/include/linux/quota.h C7-1/include/linux/quota.h
--- C7-0/include/linux/quota.h	Mon Jan 14 23:13:52 2002
+++ C7-1/include/linux/quota.h	Sun Mar 31 02:52:35 2002
@@ -144,7 +144,6 @@
 
 #ifdef __KERNEL__
 
-extern int nr_dquots, nr_free_dquots;
 extern int dquot_root_squash;
 
 #define NR_DQHASH 43            /* Just an arbitrary number */
diff -urN C7-0/kernel/Makefile C7-1/kernel/Makefile
--- C7-0/kernel/Makefile	Tue Mar 19 16:06:01 2002
+++ C7-1/kernel/Makefile	Sun Mar 31 02:52:35 2002
@@ -14,12 +14,13 @@
 
 obj-y     = sched.o dma.o fork.o exec_domain.o panic.o printk.o \
 	    module.o exit.o itimer.o info.o time.o softirq.o resource.o \
-	    sysctl.o acct.o capability.o ptrace.o timer.o user.o \
+	    sysctl.o capability.o ptrace.o timer.o user.o \
 	    signal.o sys.o kmod.o context.o futex.o
 
 obj-$(CONFIG_UID16) += uid16.o
 obj-$(CONFIG_MODULES) += ksyms.o
 obj-$(CONFIG_PM) += pm.o
+obj-$(CONFIG_BSD_PROCESS_ACCT) += acct.o
 
 ifneq ($(CONFIG_IA64),y)
 # According to Alan Modra <alan@linuxcare.com.au>, the -fno-omit-frame-pointer is
diff -urN C7-0/kernel/acct.c C7-1/kernel/acct.c
--- C7-0/kernel/acct.c	Tue Mar 19 16:06:01 2002
+++ C7-1/kernel/acct.c	Sun Mar 31 02:52:35 2002
@@ -44,17 +44,11 @@
  */
 
 #include <linux/config.h>
-#include <linux/errno.h>
-#include <linux/kernel.h>
-
-#ifdef CONFIG_BSD_PROCESS_ACCT
 #include <linux/mm.h>
 #include <linux/slab.h>
 #include <linux/acct.h>
-#include <linux/smp_lock.h>
 #include <linux/file.h>
 #include <linux/tty.h>
-
 #include <asm/uaccess.h>
 
 /*
@@ -397,15 +391,3 @@
 		spin_unlock(&acct_globals.lock);
 	return 0;
 }
-
-#else
-/*
- * Dummy system call when BSD process accounting is not configured
- * into the kernel.
- */
-
-asmlinkage long sys_acct(const char * filename)
-{
-	return -ENOSYS;
-}
-#endif
diff -urN C7-0/kernel/exit.c C7-1/kernel/exit.c
--- C7-0/kernel/exit.c	Tue Mar 19 16:06:01 2002
+++ C7-1/kernel/exit.c	Sun Mar 31 02:52:35 2002
@@ -14,9 +14,7 @@
 #include <linux/personality.h>
 #include <linux/tty.h>
 #include <linux/namespace.h>
-#ifdef CONFIG_BSD_PROCESS_ACCT
 #include <linux/acct.h>
-#endif
 #include <linux/file.h>
 #include <linux/binfmts.h>
 
@@ -493,9 +491,7 @@
 	del_timer_sync(&tsk->real_timer);
 
 fake_volatile:
-#ifdef CONFIG_BSD_PROCESS_ACCT
 	acct_process(code);
-#endif
 	__exit_mm(tsk);
 
 	lock_kernel();
diff -urN C7-0/kernel/sys.c C7-1/kernel/sys.c
--- C7-0/kernel/sys.c	Tue Mar 19 16:06:01 2002
+++ C7-1/kernel/sys.c	Sun Mar 31 02:52:49 2002
@@ -173,6 +173,18 @@
 	return -ENOSYS;
 }
 
+/*
+ * "Conditional" syscalls
+ *
+ * What we want is __attribute__((weak,alias("sys_ni_syscall"))),
+ * but it doesn't work on sparc64, so we just do it by hand
+ */
+#define cond_syscall(x) asm(".weak\t" #x "\n\t.set\t" #x ",sys_ni_syscall");
+
+cond_syscall(sys_nfsservctl)
+cond_syscall(sys_quotactl)
+cond_syscall(sys_acct)
+
 static int proc_sel(struct task_struct *p, int which, int who)
 {
 	if(p->pid)
diff -urN C7-0/kernel/sysctl.c C7-1/kernel/sysctl.c
--- C7-0/kernel/sysctl.c	Tue Feb 19 22:33:08 2002
+++ C7-1/kernel/sysctl.c	Sun Mar 31 02:52:35 2002
@@ -284,8 +284,6 @@
 	 0444, NULL, &proc_dointvec},
 	{FS_MAXFILE, "file-max", &files_stat.max_files, sizeof(int),
 	 0644, NULL, &proc_dointvec},
-	{FS_NRDQUOT, "dquot-nr", &nr_dquots, 2*sizeof(int),
-	 0444, NULL, &proc_dointvec},
 	{FS_DENTRY, "dentry-state", &dentry_stat, 6*sizeof(int),
 	 0444, NULL, &proc_dointvec},
 	{FS_OVERFLOWUID, "overflowuid", &fs_overflowuid, sizeof(int), 0644, NULL,

Re: linux-2.5.7

[David S. Miller]

   From: Pierre Rousselet <pierre.rousselet@wanadoo.fr>
   Date: Sun, 31 Mar 2002 05:25:19 +0200

   I've noticed 2.5.7 fails to build without tcp/ip enabled :
   sock.c:559: `TCP_LISTEN' undeclared
   sock.c:1192: `TCP_CLOSE' undeclared

Just add an include of linux/tcp.h to net/core/sock.c, that
should clear it up.

I'll fix this in my sources.

Re: linux-2.5.7

[Pierre Rousselet]

Timothy Murphy wrote:
> I'm sure this has been recognised,
> but I would point out that sys_nfsservctl is not "undefined"
> if NFSD is not chosen.

I've noticed 2.5.7 fails to build without tcp/ip enabled :
sock.c:559: `TCP_LISTEN' undeclared
sock.c:1192: `TCP_CLOSE' undeclared
*and* without nfs choosen for the reason you give.

Pierre
-- 
------------------------------------------------
  Pierre Rousselet <pierre.rousselet@wanadoo.fr>
------------------------------------------------

Re: linux-2.5.7

[David S. Miller]

   From: Alexander Viro <viro@math.psu.edu>
   Date: Sat, 30 Mar 2002 18:41:46 -0500 (EST)
   
   	Dave, you've mentioned doing the equivalent of
   __attribute__((weak,alias("foo")) by hand.  Could you give an example?
   

#define make_weak(foo,bar)	__asm__(".weak foo, bar")

Or however the assembler syntax works.  GLIBC has some header file it
at least used to use which had macros doing something similar.

Re: linux-2.5.7

[Alexander Viro]

On Sat, 30 Mar 2002, Timothy Murphy wrote:

> I'm sure this has been recognised,
> but I would point out that sys_nfsservctl is not "undefined"
> if NFSD is not chosen.
> 
> The following patch to .../arch/i386/kernel/entry.S corrects this,
> though this is obviously not the right place to put it:

Wrong fix.  Using weak aliases would do it in cleaner way, but there's
a sparc64 toolchain bugs that don't allow that.

	Dave, you've mentioned doing the equivalent of
__attribute__((weak,alias("foo")) by hand.  Could you give an example?

linux-2.5.7

[Timothy Murphy]

I'm sure this has been recognised,
but I would point out that sys_nfsservctl is not "undefined"
if NFSD is not chosen.

The following patch to .../arch/i386/kernel/entry.S corrects this,
though this is obviously not the right place to put it:

===============================================================
--- entry.S.bak	Mon Mar 18 20:37:09 2002
+++ entry.S	Thu Mar 28 15:59:20 2002
@@ -40,6 +40,11 @@
  * "current" is in register %ebx during any slow entries.
  */
 
+#if defined(CONFIG_NFSD) || defined(CONFIG_NFSD_MODULE)
+#else
+#define sys_nfsservctl		sys_ni_syscall
+#endif
+
 #include <linux/config.h>
 #include <linux/sys.h>
 #include <linux/linkage.h>
===============================================================

-- 
Timothy Murphy  
e-mail: tim@birdsnest.maths.tcd.ie
tel: 086-233 6090
s-mail: School of Mathematics, Trinity College, Dublin 2, Ireland