Skip to contentSkip to conversation

x86: unify sys_iopl

GIT PULL Jeremy Fitzhardinge 4 participants 11 messages Sep 25, 2009 – Oct 13, 2009 x86

lore
Share & tools
lore
#0 lore raw

Hi Ingo,

The x86-64 implementation of iopl was buggy because it never ended up calling set_iopl_mask(). This had no effect on native sys_iopl (because set_iopl_mask is normally no-op on 64-bit), but it ended up never calling the pvop, which caused iopl to have no effect on 64-bit Xen guests.

The two functions are needlessly different anyway. This patch just unifies them into a single function which is mostly derived from the 32-bit version.

Thanks,
J

The following changes since commit c44c9ec0f38b939b3200436e3aa95c1aa83c41c7:
Jeremy Fitzhardinge (1):
x86: split NX setup into separate file to limit unstack-protected code

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git bugfix

Jeremy Fitzhardinge (1):
x86: unify sys_iopl

 arch/x86/include/asm/syscalls.h |    8 +++-----
 arch/x86/kernel/ioport.c        |   11 ++---------
 2 files changed, 5 insertions(+), 14 deletions(-)
Quoted 1 line

From 8dbb1d168eb8be6bf43d123fdfb3ba3b03762e62 Mon Sep 17 00:00:00 2001

From: Jeremy Fitzhardinge <[email protected]>
Date: Wed, 23 Sep 2009 16:35:24 -0700
Subject: [PATCH] x86: unify sys_iopl

The 32 and 64-bit versions of sys_iopl were needlessly different:
- The 32-bit version ignored its function argument and directly fetched the parameter from struct regs, presumably code dating from the neolithic era.
- The 64-bit version never called set_iopl_mask(). Usually this is a no-op for 64-bit, but it is also a pvop, which meant that that iopl never worked for 64-bit guests under Xen.
- Both versions use the archaic style of getting pt_regs passed to them. We can get the current pt_regs with task_pt_regs, which avoids worrying about the fine details of stack layout.

We can use the body of the 32-bit function with a few small adjustments to the function definition.

arch/x86/include/asm/syscalls.h +3 -5
index 372b76e..5336ce2 100644
@@ -33,11 +33,11 @@ long sys_rt_sigreturn(struct pt_regs *);
 asmlinkage int sys_set_thread_area(struct user_desc __user *);
 asmlinkage int sys_get_thread_area(struct user_desc __user *);
 
-/* X86_32 only */
-#ifdef CONFIG_X86_32
 /* kernel/ioport.c */
-long sys_iopl(struct pt_regs *);
+asmlinkage long sys_iopl(unsigned int);
 
+/* X86_32 only */
+#ifdef CONFIG_X86_32
 /* kernel/process_32.c */
 int sys_clone(struct pt_regs *);
 int sys_execve(struct pt_regs *);
@@ -70,8 +70,6 @@ int sys_vm86(struct pt_regs *);
 #else /* CONFIG_X86_32 */
 
 /* X86_64 only */
-/* kernel/ioport.c */
-asmlinkage long sys_iopl(unsigned int, struct pt_regs *);
 
 /* kernel/process_64.c */
 asmlinkage long sys_clone(unsigned long, unsigned long,
arch/x86/kernel/ioport.c +2 -9
index 99c4d30..bad4f22 100644
@@ -119,11 +119,10 @@ static int do_iopl(unsigned int level, struct pt_regs *regs)
 	return 0;
 }
 
-#ifdef CONFIG_X86_32
-long sys_iopl(struct pt_regs *regs)
+asmlinkage long sys_iopl(unsigned int level)
 {
-	unsigned int level = regs->bx;
 	struct thread_struct *t = &current->thread;
+	struct pt_regs *regs = task_pt_regs(current);
 	int rc;
 
 	rc = do_iopl(level, regs);
@@ -135,9 +134,3 @@ long sys_iopl(struct pt_regs *regs)
 out:
 	return rc;
 }
-#else
-asmlinkage long sys_iopl(unsigned int level, struct pt_regs *regs)
-{
-	return do_iopl(level, regs);
-}
-#endif
#1 lore raw

The correct fix is to move set_iopl_mask into do_iopl. sys_iopl should only be a wrapper around do_iopl to manage the different calling convention used by 32-bit and 64-bit.

Quoted 113 lines · Jeremy Fitzhardinge

Hi Ingo,

The x86-64 implementation of iopl was buggy because it never ended up calling set_iopl_mask().  This had no effect on native sys_iopl (because set_iopl_mask is normally no-op on 64-bit), but it ended up never calling the pvop, which caused iopl to have no effect on 64-bit Xen guests.

The two functions are needlessly different anyway.  This patch just unifies them into a single function which is mostly derived from the 32-bit version.

Thanks,
       J

The following changes since commit c44c9ec0f38b939b3200436e3aa95c1aa83c41c7:
 Jeremy Fitzhardinge (1):
       x86: split NX setup into separate file to limit unstack-protected code

are available in the git repository at:

 git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git bugfix

Jeremy Fitzhardinge (1):
     x86: unify sys_iopl

 arch/x86/include/asm/syscalls.h |    8 +++----- arch/x86/kernel/ioport.c        |   11 ++---------

 2 files changed, 5 insertions(+), 14 deletions(-)

From 8dbb1d168eb8be6bf43d123fdfb3ba3b03762e62 Mon Sep 17 00:00:00 2001
From: Jeremy Fitzhardinge <[email protected]>
Date: Wed, 23 Sep 2009 16:35:24 -0700
Subject: [PATCH] x86: unify sys_iopl

The 32 and 64-bit versions of sys_iopl were needlessly different:
 - The 32-bit version ignored its function argument and directly fetched the parameter from struct regs, presumably code dating from the neolithic era.
 - The 64-bit version never called set_iopl_mask().  Usually this is a no-op for 64-bit, but it is also a pvop, which meant that that iopl never worked for 64-bit guests under Xen.
 - Both versions use the archaic style of getting pt_regs passed to them.  We can get the current pt_regs with task_pt_regs, which avoids worrying about the fine details of stack layout.

We can use the body of the 32-bit function with a few small adjustments to the function definition.

arch/x86/include/asm/syscalls.h +3 -5
index 372b76e..5336ce2 100644
@@ -33,11 +33,11 @@ long sys_rt_sigreturn(struct pt_regs *);
  asmlinkage int sys_set_thread_area(struct user_desc __user *);
  asmlinkage int sys_get_thread_area(struct user_desc __user *);
 
-/* X86_32 only */
-#ifdef CONFIG_X86_32
  /* kernel/ioport.c */
-long sys_iopl(struct pt_regs *);
+asmlinkage long sys_iopl(unsigned int);
 
+/* X86_32 only */
+#ifdef CONFIG_X86_32
  /* kernel/process_32.c */
  int sys_clone(struct pt_regs *);
  int sys_execve(struct pt_regs *);
@@ -70,8 +70,6 @@ int sys_vm86(struct pt_regs *);
  #else /* CONFIG_X86_32 */
 
  /* X86_64 only */
-/* kernel/ioport.c */
-asmlinkage long sys_iopl(unsigned int, struct pt_regs *);
 
  /* kernel/process_64.c */
  asmlinkage long sys_clone(unsigned long, unsigned long,
arch/x86/kernel/ioport.c +2 -3
index 99c4d30..bad4f22 100644
@@ -119,11 +119,10 @@ static int do_iopl(unsigned int level, struct pt_regs *regs)
        return 0;
  }
 
-#ifdef CONFIG_X86_32
-long sys_iopl(struct pt_regs *regs)
+asmlinkage long sys_iopl(unsigned int level)
  {
-       unsigned int level = regs->bx;
        struct thread_struct *t = &current->thread;
+       struct pt_regs *regs = task_pt_regs(current);
        int rc;
 
        rc = do_iopl(level, regs);
@@ -135,9 +134,3 @@ long sys_iopl(struct pt_regs *regs)

 out:
       return rc;
 }

arch/x86/kernel/ioport.c +0 -7
-#else
-asmlinkage long sys_iopl(unsigned int level, struct pt_regs *regs)
-{
-       return do_iopl(level, regs);
-}
-#endif
 
 
--

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

#2 lore raw
Quoted 4 lines · Brian Gerst

The correct fix is to move set_iopl_mask into do_iopl. sys_iopl should only be a wrapper around do_iopl to manage the different calling convention used by 32-bit and 64-bit.

The patch eliminates that distinction, so I guess do_iopl should really be folded into sys_iopl (though the compiler effectively does that anyway by inlining it).

J

#3 lore raw
Quoted 9 lines · Jeremy Fitzhardinge
Quoted 4 lines · Brian Gerst

The correct fix is to move set_iopl_mask into do_iopl.  sys_iopl should only be a wrapper around do_iopl to manage the different calling convention used by 32-bit and 64-bit.

The patch eliminates that distinction, so I guess do_iopl should really be folded into sys_iopl (though the compiler effectively does that anyway by inlining it).

Using task_pt_regs is less effecient than getting the pointer directly from the syscall entry code. On 32-bit it probably doesn't matter too much, but on 64-bit, you still need the special stub code in entry_64.S to put the full pt_regs struct on the stack. Changing the calling conventions should be a seperate patch, and should be done for all pt_regs-using syscalls.

#4 lore raw
Quoted 2 lines · Brian Gerst

Using task_pt_regs is less effecient than getting the pointer directly from the syscall entry code.

(really not an issue in this case)

Quoted 6 lines

On 32-bit it probably doesn't matter too
much, but on 64-bit, you still need the special stub code in entry_64.S to put the full pt_regs struct on the stack. Changing the calling conventions should be a seperate patch, and should be done for all pt_regs-using syscalls.

iopl only really cares about [er]flags, so it probably doesn't need a full ptregs structure anyway. But I'm not sure what changes you'd consider for the rest of the ptregs calls; sigaltstack probably doesn't need full ptregs either (it just updates [er]sp I think), but fork/clone definitely do.

J

#5 lore raw
In reply to Jeremy Fitzhardinge
Quoted 21 lines · Jeremy Fitzhardinge

Hi Ingo,

The x86-64 implementation of iopl was buggy because it never ended up calling set_iopl_mask(). This had no effect on native sys_iopl (because set_iopl_mask is normally no-op on 64-bit), but it ended up never calling the pvop, which caused iopl to have no effect on 64-bit Xen guests.

The two functions are needlessly different anyway. This patch just unifies them into a single function which is mostly derived from the 32-bit version.

Thanks,
J

The following changes since commit c44c9ec0f38b939b3200436e3aa95c1aa83c41c7:
Jeremy Fitzhardinge (1):
x86: split NX setup into separate file to limit unstack-protected code

are available in the git repository at:

Quoted 8 lines

git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git bugfix

Jeremy Fitzhardinge (1):
x86: unify sys_iopl

 arch/x86/include/asm/syscalls.h |    8 +++-----
 arch/x86/kernel/ioport.c        |   11 ++---------
 2 files changed, 5 insertions(+), 14 deletions(-)

Pulled, thanks a lot Jeremy!

It's in tip:x86/paravirt right now. I'm uneasy about pushing this into .32 - we had this status quo forever. Peter, Thomas, what do you think?

Ingo

#6 lore raw
Quoted 6 lines · Ingo Molnar

Pulled, thanks a lot Jeremy!

It's in tip:x86/paravirt right now. I'm uneasy about pushing this into .32 - we had this status quo forever. Peter, Thomas, what do you think?

First of all, the unification looks good.

As far as .32 is concerned... this *is* a bug even if this is only for paravirt, and given the small amount of code I am personally OK with taking the whole patch for .32.

However, the patch is not complete! The patch incidentally eliminates the need to have assembly stubs for sys_iopl, and those assembly stubs should be removed. I have a patch for that currently test building.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#7 lore raw
Quoted 10 lines · H. Peter Anvin

First of all, the unification looks good.

As far as .32 is concerned... this *is* a bug even if this is only for paravirt, and given the small amount of code I am personally OK with taking the whole patch for .32.

However, the patch is not complete! The patch incidentally eliminates the need to have assembly stubs for sys_iopl, and those assembly stubs should be removed. I have a patch for that currently test building.

I wasn't sure whether task_pt_regs() needed the full register set to be saved to correctly return rflags (that is, does PTREGSCALL change the shape of the stack, or just the contents?).

J

#8 lore raw
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 6 lines · H. Peter Anvin

First of all, the unification looks good.

As far as .32 is concerned... this *is* a bug even if this is only for paravirt, and given the small amount of code I am personally OK with taking the whole patch for .32.

Quoted 9 lines
Quoted 4 lines

However, the patch is not complete! The patch incidentally eliminates the need to have assembly stubs for sys_iopl, and those assembly stubs should be removed. I have a patch for that currently test building.

I wasn't sure whether task_pt_regs() needed the full register set to be saved to correctly return rflags (that is, does PTREGSCALL change the shape of the stack, or just the contents?).

PTREGSCALL does change the shape of the stack on 64 bits, but not on 32 bits. However, I believe it only matters for the additional fields, since the structure is anchored at the top. I will double-check.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#9 lore raw
In reply to Jeremy Fitzhardinge
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 6 lines · H. Peter Anvin

First of all, the unification looks good.

As far as .32 is concerned... this *is* a bug even if this is only for paravirt, and given the small amount of code I am personally OK with taking the whole patch for .32.

Quoted 9 lines
Quoted 4 lines

However, the patch is not complete! The patch incidentally eliminates the need to have assembly stubs for sys_iopl, and those assembly stubs should be removed. I have a patch for that currently test building.

I wasn't sure whether task_pt_regs() needed the full register set to be saved to correctly return rflags (that is, does PTREGSCALL change the shape of the stack, or just the contents?).

A double-check confirmed my previous statement: on 64 bits, PTREGSCALL does populate the fields at the beginning of struct pt_regs, but task_pt_regs(current) is valid even without those fields populated, since it is anchored at the top "hanging down" from the kernel entry stack pointer address.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.
#10 lore raw
In reply to Jeremy Fitzhardinge
Quoted 7 lines · Jeremy Fitzhardinge
Quoted 6 lines · H. Peter Anvin

First of all, the unification looks good.

As far as .32 is concerned... this *is* a bug even if this is only for paravirt, and given the small amount of code I am personally OK with taking the whole patch for .32.

Quoted 9 lines
Quoted 4 lines

However, the patch is not complete! The patch incidentally eliminates the need to have assembly stubs for sys_iopl, and those assembly stubs should be removed. I have a patch for that currently test building.

I wasn't sure whether task_pt_regs() needed the full register set to be saved to correctly return rflags (that is, does PTREGSCALL change the shape of the stack, or just the contents?).

Erk, I was right but still wrong... it isn't safe to modify the partial pt_regs because of the magic sync stuff that goes on in entry_64.S.

I have to say the entry_64.S stuff gives me a headache in the extreme. It really can't be the sanest way to do this stuff.

/me makes a mental note to try to work through this code at some point.

-hpa

Signature · 2 lines
H. Peter Anvin, Intel Open Source Technology Center
I work for Intel.  I don't speak on their behalf.