[patch] x86: fix ESP corruption CPU bug

[patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

[-- Attachment #1: Type: text/plain, Size: 1817 bytes --]

Hi Alan.

Attached patch works around the corruption
of the high word of the ESP register, which
is the official bug of x86 CPUs. The bug
triggers only when the one is using the
16bit stack segment, and is described here:
http://www.intel.com/design/intarch/specupdt/27287402.PDF

Patch helps running many apps under dosemu,
and, according to the comments found in
Wine sources, also helps Wine.

Also the patch closes the "information leak",
which is that the half of the kernel's %esp
value gets leaked to user-space.

The initial discussion about the problem
can be found here:
http://lkml.org/lkml/2004/9/16/254
On a later stages the development of that
patch was driven mainly by Linus, but it
was in private.

This patch adds only 6 instructions for
the fast kernel-->user return path, 6
instructions on an exception path and
5 instructions on an NMI path. All the
rest of the patch gets executed only
for the two apps in that world, namely,
dosemu and wine (and VMWare?).
That's why it should be relatevely easy
to make sure the patch doesn't do any
harm for the normal apps, and so it is
safe, and probably fits into an -ac tree.

How it works:
- Allocates the per-cpu data for 16bit
stacks and sets the per-cpu GDT entries
for them.
- On a return to userspace, checks whether
the SS is from LDT and is 16 bit.
- If it is, the iret frame gets copied
to the 16bit per-cpu stack and stack gets
switched to the 16bit one, while the
higher word of %esp gets preloaded with
the proper value.
- On an exceptions the check is performed
to see if we are on a 16bit stack, and
if we are - switch to the 32bit one.


Alan, would it be possible to apply that
patch to an -ac tree?


Acked-by: Linus Torvalds <torvalds@osdl.org>
Acked-by: Petr Vandrovec <vandrove@vc.cvut.cz>
Signed-off-by: Stas Sergeev <stsp@aknet.ru>


[-- Attachment #2: linux-2.6.11-ac2-stk7.diff --]
[-- Type: text/x-patch, Size: 11163 bytes --]

diff -ur linux-2.6.11-ac2/arch/i386/kernel/cpu/common.c linux-2.6.11-ac2-stk/arch/i386/kernel/cpu/common.c
--- linux-2.6.11-ac2/arch/i386/kernel/cpu/common.c	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/cpu/common.c	2005-03-13 18:17:47.000000000 +0300
@@ -21,6 +21,10 @@
 DEFINE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 EXPORT_PER_CPU_SYMBOL(cpu_gdt_table);
 
+unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DEFINE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+EXPORT_PER_CPU_SYMBOL(cpu_16bit_stack);
+
 static int cachesize_override __initdata = -1;
 static int disable_x86_fxsr __initdata = 0;
 static int disable_x86_serial_nr __initdata = 1;
@@ -557,6 +561,7 @@
 	int cpu = smp_processor_id();
 	struct tss_struct * t = &per_cpu(init_tss, cpu);
 	struct thread_struct *thread = &current->thread;
+	__u32 stk16_off = (__u32)&per_cpu(cpu_16bit_stack, cpu);
 
 	if (cpu_test_and_set(cpu, cpu_initialized)) {
 		printk(KERN_WARNING "CPU#%d already initialized!\n", cpu);
@@ -579,6 +584,13 @@
 	 */
 	memcpy(&per_cpu(cpu_gdt_table, cpu), cpu_gdt_table,
 	       GDT_SIZE);
+
+	/* Set up GDT entry for 16bit stack */
+	*(__u64 *)&(per_cpu(cpu_gdt_table, cpu)[GDT_ENTRY_ESPFIX_SS]) |=
+		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
+		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
+		(CPU_16BIT_STACK_SIZE - 1);
+
 	cpu_gdt_descr[cpu].size = GDT_SIZE - 1;
 	cpu_gdt_descr[cpu].address =
 	    (unsigned long)&per_cpu(cpu_gdt_table, cpu);
diff -ur linux-2.6.11-ac2/arch/i386/kernel/entry.S linux-2.6.11-ac2-stk/arch/i386/kernel/entry.S
--- linux-2.6.11-ac2/arch/i386/kernel/entry.S	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/entry.S	2005-03-13 18:17:47.000000000 +0300
@@ -47,6 +47,7 @@
 #include <asm/segment.h>
 #include <asm/smp.h>
 #include <asm/page.h>
+#include <asm/desc.h>
 #include "irq_vectors.h"
 
 #define nr_syscalls ((syscall_table_size)/4)
@@ -78,7 +79,7 @@
 #define preempt_stop		cli
 #else
 #define preempt_stop
-#define resume_kernel		restore_all
+#define resume_kernel		restore_nocheck
 #endif
 
 #define SAVE_ALL \
@@ -122,24 +123,6 @@
 .previous
 
 
-#define RESTORE_ALL	\
-	RESTORE_REGS	\
-	addl $4, %esp;	\
-1:	iret;		\
-.section .fixup,"ax";   \
-2:	sti;		\
-	movl $(__USER_DS), %edx; \
-	movl %edx, %ds; \
-	movl %edx, %es; \
-	movl $11,%eax;	\
-	call do_exit;	\
-.previous;		\
-.section __ex_table,"a";\
-	.align 4;	\
-	.long 1b,2b;	\
-.previous
-
-
 ENTRY(ret_from_fork)
 	pushl %eax
 	call schedule_tail
@@ -163,7 +146,7 @@
 	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
 	movb CS(%esp), %al
 	testl $(VM_MASK | 3), %eax
-	jz resume_kernel		# returning to kernel or vm86-space
+	jz resume_kernel
 ENTRY(resume_userspace)
  	cli				# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
@@ -178,7 +161,7 @@
 ENTRY(resume_kernel)
 	cli
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
-	jnz restore_all
+	jnz restore_nocheck
 need_resched:
 	movl TI_flags(%ebp), %ecx	# need_resched set ?
 	testb $_TIF_NEED_RESCHED, %cl
@@ -189,6 +172,31 @@
 	jmp need_resched
 #endif
 
+ldt_ss:
+	larl OLDSS(%esp), %eax
+	jnz restore_nocheck
+	testl $0x00400000, %eax		# returning to 32bit stack?
+	jnz restore_nocheck		# allright, normal return
+	/* If returning to userspace with 16bit stack,
+	 * try to fix the higher word of ESP, as the CPU
+	 * won't restore it.
+	 * This is an "official" bug of all the x86-compatible
+	 * CPUs, which we can try to work around to make
+	 * dosemu and wine happy. */
+	subl $8, %esp		# reserve space for switch16 pointer
+	cli
+	movl %esp, %eax
+	/* Set up the 16bit stack frame with switch32 pointer on top,
+	 * and a switch16 pointer on top of the current frame. */
+	call setup_x86_bogus_stack
+	RESTORE_REGS
+	lss 20+4(%esp), %esp	# switch to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 /* SYSENTER_RETURN points to after the "sysenter" instruction in
    the vsyscall page.  See vsyscall-sysentry.S, which defines the symbol.  */
 
@@ -257,8 +265,31 @@
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx	# current->work
 	jne syscall_exit_work
+
 restore_all:
-	RESTORE_ALL
+	movl EFLAGS(%esp), %eax		# mix EFLAGS, SS and CS
+	movb OLDSS(%esp), %ah
+	movb CS(%esp), %al
+	andl $(VM_MASK | (4 << 8) | 3), %eax
+	cmpl $((4 << 8) | 3), %eax
+	je ldt_ss			# returning to user-space with LDT SS
+restore_nocheck:
+	RESTORE_REGS
+	addl $4, %esp
+1:	iret
+.section .fixup,"ax"
+iret_exc:
+	sti
+	movl $__USER_DS, %edx
+	movl %edx, %ds
+	movl %edx, %es
+	movl $11,%eax
+	call do_exit
+.previous
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
 
 	# perform work that needs to be done immediately before resumption
 	ALIGN
@@ -334,6 +365,27 @@
 	movl $-ENOSYS,EAX(%esp)
 	jmp resume_userspace
 
+#define FIXUP_ESPFIX_STACK \
+	movl %esp, %eax; \
+	/* switch to 32bit stack using the pointer on top of 16bit stack */ \
+	lss %ss:CPU_16BIT_STACK_SIZE-8, %esp; \
+	/* copy data from 16bit stack to 32bit stack */ \
+	call fixup_x86_bogus_stack; \
+	/* put ESP to the proper location */ \
+	movl %eax, %esp;
+#define UNWIND_ESPFIX_STACK \
+	pushl %eax; \
+	movl %ss, %eax; \
+	/* see if on 16bit stack */ \
+	cmpw $__ESPFIX_SS, %ax; \
+	jne 28f; \
+	movl $__KERNEL_DS, %edx; \
+	movl %edx, %ds; \
+	movl %edx, %es; \
+	/* switch to 32bit stack */ \
+	FIXUP_ESPFIX_STACK \
+28:	popl %eax;
+
 /*
  * Build the entry stubs and pointer table with
  * some assembler magic.
@@ -388,7 +440,9 @@
 	pushl %ecx
 	pushl %ebx
 	cld
-	movl %es, %ecx
+	pushl %es
+	UNWIND_ESPFIX_STACK
+	popl %ecx
 	movl ES(%esp), %edi		# get the function address
 	movl ORIG_EAX(%esp), %edx	# get the error code
 	movl %eax, ORIG_EAX(%esp)
@@ -470,6 +524,11 @@
  * fault happened on the sysenter path.
  */
 ENTRY(nmi)
+	pushl %eax
+	movl %ss, %eax
+	cmpw $__ESPFIX_SS, %ax
+	popl %eax
+	je nmi_16bit_stack
 	cmpl $sysenter_entry,(%esp)
 	je nmi_stack_fixup
 	pushl %eax
@@ -489,7 +548,7 @@
 	xorl %edx,%edx		# zero error code
 	movl %esp,%eax		# pt_regs pointer
 	call do_nmi
-	RESTORE_ALL
+	jmp restore_all
 
 nmi_stack_fixup:
 	FIX_STACK(12,nmi_stack_correct, 1)
@@ -505,6 +564,29 @@
 	FIX_STACK(24,nmi_stack_correct, 1)
 	jmp nmi_stack_correct
 
+nmi_16bit_stack:
+	/* create the pointer to lss back */
+	pushl %ss
+	pushl %esp
+	movzwl %sp, %esp
+	addw $4, (%esp)
+	/* copy the iret frame of 12 bytes */
+	.rept 3
+	pushl 16(%esp)
+	.endr
+	pushl %eax
+	SAVE_ALL
+	FIXUP_ESPFIX_STACK		# %eax == %esp
+	xorl %edx,%edx			# zero error code
+	call do_nmi
+	RESTORE_REGS
+	lss 12+4(%esp), %esp		# back to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 ENTRY(int3)
 	pushl $-1			# mark this as an int
 	SAVE_ALL
diff -ur linux-2.6.11-ac2/arch/i386/kernel/head.S linux-2.6.11-ac2-stk/arch/i386/kernel/head.S
--- linux-2.6.11-ac2/arch/i386/kernel/head.S	2005-03-13 17:56:19.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/head.S	2005-03-13 18:17:47.000000000 +0300
@@ -512,7 +512,7 @@
 	.quad 0x00009a0000000000	/* 0xc0 APM CS 16 code (16 bit) */
 	.quad 0x0040920000000000	/* 0xc8 APM DS    data */
 
-	.quad 0x0000000000000000	/* 0xd0 - unused */
+	.quad 0x0000920000000000	/* 0xd0 - ESPFIX 16-bit SS */
 	.quad 0x0000000000000000	/* 0xd8 - unused */
 	.quad 0x0000000000000000	/* 0xe0 - unused */
 	.quad 0x0000000000000000	/* 0xe8 - unused */
diff -ur linux-2.6.11-ac2/arch/i386/kernel/traps.c linux-2.6.11-ac2-stk/arch/i386/kernel/traps.c
--- linux-2.6.11-ac2/arch/i386/kernel/traps.c	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/traps.c	2005-03-13 18:17:47.000000000 +0300
@@ -895,6 +895,51 @@
 #endif
 }
 
+fastcall void setup_x86_bogus_stack(unsigned char * stk)
+{
+	unsigned long *switch16_ptr, *switch32_ptr;
+	struct pt_regs *regs;
+	unsigned long stack_top, stack_bot;
+	unsigned short iret_frame16_off;
+	int cpu = smp_processor_id();
+	/* reserve the space on 32bit stack for the magic switch16 pointer */
+	memmove(stk, stk + 8, sizeof(struct pt_regs));
+	switch16_ptr = (unsigned long *)(stk + sizeof(struct pt_regs));
+	regs = (struct pt_regs *)stk;
+	/* now the switch32 on 16bit stack */
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	iret_frame16_off = CPU_16BIT_STACK_SIZE - 8 - 20;
+	/* copy iret frame on 16bit stack */
+	memcpy((void *)(stack_bot + iret_frame16_off), &regs->eip, 20);
+	/* fill in the switch pointers */
+	switch16_ptr[0] = (regs->esp & 0xffff0000) | iret_frame16_off;
+	switch16_ptr[1] = __ESPFIX_SS;
+	switch32_ptr[0] = (unsigned long)stk + sizeof(struct pt_regs) +
+		8 - CPU_16BIT_STACK_SIZE;
+	switch32_ptr[1] = __KERNEL_DS;
+}
+
+fastcall unsigned char * fixup_x86_bogus_stack(unsigned short sp)
+{
+	unsigned long *switch32_ptr;
+	unsigned char *stack16, *stack32;
+	unsigned long stack_top, stack_bot;
+	int len;
+	int cpu = smp_processor_id();
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	/* copy the data from 16bit stack to 32bit stack */
+	len = CPU_16BIT_STACK_SIZE - 8 - sp;
+	stack16 = (unsigned char *)(stack_bot + sp);
+	stack32 = (unsigned char *)
+		(switch32_ptr[0] + CPU_16BIT_STACK_SIZE - 8 - len);
+	memcpy(stack32, stack16, len);
+	return stack32;
+}
+
 /*
  *  'math_state_restore()' saves the current math information in the
  * old math state array, and gets the new ones from the current task
diff -ur linux-2.6.11-ac2/include/asm-i386/desc.h linux-2.6.11-ac2-stk/include/asm-i386/desc.h
--- linux-2.6.11-ac2/include/asm-i386/desc.h	2005-03-13 17:57:12.000000000 +0300
+++ linux-2.6.11-ac2-stk/include/asm-i386/desc.h	2005-03-13 18:17:47.000000000 +0300
@@ -4,6 +4,8 @@
 #include <asm/ldt.h>
 #include <asm/segment.h>
 
+#define CPU_16BIT_STACK_SIZE 1024
+
 #ifndef __ASSEMBLY__
 
 #include <linux/preempt.h>
@@ -15,6 +17,9 @@
 extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
 DECLARE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 
+extern unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DECLARE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+
 struct Xgt_desc_struct {
 	unsigned short size;
 	unsigned long address __attribute__((packed));
diff -ur linux-2.6.11-ac2/include/asm-i386/segment.h linux-2.6.11-ac2-stk/include/asm-i386/segment.h
--- linux-2.6.11-ac2/include/asm-i386/segment.h	2005-01-04 03:59:37.000000000 +0300
+++ linux-2.6.11-ac2-stk/include/asm-i386/segment.h	2005-03-13 18:17:47.000000000 +0300
@@ -38,7 +38,7 @@
  *  24 - APM BIOS support
  *  25 - APM BIOS support 
  *
- *  26 - unused
+ *  26 - ESPFIX small SS
  *  27 - unused
  *  28 - unused
  *  29 - unused
@@ -71,6 +71,9 @@
 #define GDT_ENTRY_PNPBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 6)
 #define GDT_ENTRY_APMBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 11)
 
+#define GDT_ENTRY_ESPFIX_SS		(GDT_ENTRY_KERNEL_BASE + 14)
+#define __ESPFIX_SS (GDT_ENTRY_ESPFIX_SS * 8)
+
 #define GDT_ENTRY_DOUBLEFAULT_TSS	31
 
 /*

Re: [patch] x86: fix ESP corruption CPU bug

[Grzegorz Kulewski]

On Sun, 13 Mar 2005, Stas Sergeev wrote:

> Hi Alan.
>
> Attached patch works around the corruption
> of the high word of the ESP register, which
> is the official bug of x86 CPUs. The bug
> triggers only when the one is using the
> 16bit stack segment, and is described here:
> http://www.intel.com/design/intarch/specupdt/27287402.PDF

Does the bug also egsist on AMD CPU's? Does the patch add anything to 
kernels compiled for AMD CPU's?


Thanks,

Grzegorz Kulewski

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hello.

Grzegorz Kulewski wrote:
> Does the bug also egsist on AMD CPU's?
Yes. As well as the ones of a Transmeta etc.
I just haven't tested the old Cyrixes, that
AFAIK were trying to ignore some Intel bugs.
The test-case for the bug is here:
http://www.ussg.iu.edu/hypermail/linux/kernel/0409.2/0690.html

> Does the patch add anything to 
> kernels compiled for AMD CPU's?
Same as for the Intel ones - unless you are
a dosemu or, in a lesser extent, Wine user -
nothing except for fixing the small "information
leak".
If you are the dosemu user however, then this
patch adds a lot. Whether or not it adds
something for the VMWare users I can't say, since
I am not the one of them, but my guess is that
it can help with the DOS programs under it. 

Re: [patch] x86: fix ESP corruption CPU bug

[Ondrej Zary]

Stas Sergeev wrote:
> Hello.
> 
> Grzegorz Kulewski wrote:
> 
>> Does the bug also egsist on AMD CPU's?
> 
> Yes. As well as the ones of a Transmeta etc.
> I just haven't tested the old Cyrixes, that
> AFAIK were trying to ignore some Intel bugs.
> The test-case for the bug is here:
> http://www.ussg.iu.edu/hypermail/linux/kernel/0409.2/0690.html

I've just ran that on my Cyrix MII PR300 and the bug is present:
old_ss=0x7b new_ss=0x7f
In sighandler: esp=bffff780
old_esp=0xbffff780 new_esp=0xc1a6f780
BUG!

I have also an older Cyrix CPU - 6x86 PR166 - but can't test it now as 
it's sitting in a plastic box on the shelf :-)

UMC U5SX/33 in my router - also present:
old_ss=0x7b new_ss=0x7f
In sighandler: esp=bffff820
old_esp=0xbffff820 new_esp=0xc003f820
BUG!


-- 
Ondrej Zary

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hello.

Ondrej Zary wrote:
> I've just ran that on my Cyrix MII PR300 and the bug is present:<>
> UMC U5SX/33 in my router - also present:
Thanks, now I know for sure that it exist
everywhere.
Now you can apply the patch and make sure
the bug goes away:)

Re: [patch] x86: fix ESP corruption CPU bug

[Pavel Machek]

Hi!

> >Attached patch works around the corruption
> >of the high word of the ESP register, which
> >is the official bug of x86 CPUs. The bug
> >triggers only when the one is using the
> >16bit stack segment, and is described here:
> >http://www.intel.com/design/intarch/specupdt/27287402.PDF
> 
> Does the bug also egsist on AMD CPU's? Does the patch add anything to 
> kernels compiled for AMD CPU's?

Yes, same workaround is needed on AMDs, Cyrixes, ...

								Pavel

-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Re: [patch] x86: fix ESP corruption CPU bug

[Pavel Machek]

Hi!

> @@ -257,8 +265,31 @@
>  	movl TI_flags(%ebp), %ecx
>  	testw $_TIF_ALLWORK_MASK, %cx	# current->work
>  	jne syscall_exit_work
> +
>  restore_all:
> -	RESTORE_ALL
> +	movl EFLAGS(%esp), %eax		# mix EFLAGS, SS and CS
> +	movb OLDSS(%esp), %ah
> +	movb CS(%esp), %al
> +	andl $(VM_MASK | (4 << 8) | 3), %eax
> +	cmpl $((4 << 8) | 3), %eax
> +	je ldt_ss			# returning to user-space with LDT SS

All common linux apps use same %ss, no? Perhaps it would be more
efficient to just check if %ss == 0x7b, and proceed directly to
restore_nocheck if not?

Or perhaps we could only enable this code after application loads
custom ldt?

								Pavel

-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hi.

Pavel Machek wrote:
>> +	andl $(VM_MASK | (4 << 8) | 3), %eax
>> +	cmpl $((4 << 8) | 3), %eax
>> +	je ldt_ss			# returning to user-space with LDT SS
> All common linux apps use same %ss, no? Perhaps it would be more
> efficient to just check if %ss == 0x7b, and proceed directly to
> restore_nocheck if not?
Such an optimization will cost three more
instructions, one of which is a "taken"
jump. It seems like the "taken" jump on
a fast path is not good, while now it is
only 5 instructions with a not-taken jump.
I am not sure here, but I think the current
solution is better (depends on how bad the
"taken" jump is, and how bad it is to have
the three extra insns for that optimization
purpose).

> Or perhaps we could only enable this code
> after application loads custom ldt?
The good thing here is that the code
actually does what you say, i.e. it jumps
to ldt_ss only when the app has loaded
the custom ldt and loaded that selector
to %ss. The way it is implemented, is
probably different from what you mean,
I assume you mean the new per-thread flag?
But I don't see how it can be more optimal,
i.e. you propose to check whether or not
the app altered the ldt (which can just be
the old glibc I think), while the current
solution is to also check whether it was
loaded to %ss (so the glibc case is avoided,
IIRC glibc used to load %gs with LDT selector).
I.e. since right now we jump to ldt_ss only
when the %ss is loaded with an LDT selector,
I think the extra checks are not needed.

Re: [patch] x86: fix ESP corruption CPU bug

[Linus Torvalds]

On Sun, 13 Mar 2005, Stas Sergeev wrote:
>
> Such an optimization will cost three more
> instructions, one of which is a "taken"
> jump.

I think Pavel missed the fact that you have to check the VM86 bit in
eflags before you check SS, since otherwise SS doesn't mean anything
special at all (ie checking for just the normal SS isn't correct: you
could have a 16-bit SS that looks normal, but is actually a vm86 segment).

Pavel: for the same reason you have to check the low two bits of CS too, 
since if they are zero, then SS hasn't been saved on the stack at all, so 
comparing it against some normal value is meaningless.

That said, the "ldt_ss" case should be moved _after_ the conditional
tests, since most CPU's out there will do static prediction based on
forward/backwards direction if the branch predictor isn't primed. And so
since it's an unusual case, the branch should be a forward branch, which
is usually the not-taken one (this depends on the core, of course, and you
could also add the prefix byte to mark it explicitly predict-not-taken for 
the newer cores that support it).

			Linus

[patch] x86: fix ESP corruption CPU bug (take 2)

[Stas Sergeev]

[-- Attachment #1: Type: text/plain, Size: 964 bytes --]

Hello.

Linus Torvalds wrote:
> I think Pavel missed the fact that you have to check the VM86 bit in
Ah, thanks, I must have forgotten the
essentials of the own patch :(

> That said, the "ldt_ss" case should be moved _after_ the conditional
> tests, since most CPU's out there will do static prediction based on<>
> forward/backwards direction if the branch predictor isn't primed.
OK, I moved the "ldt_ss" below the check.
The new patch is attached.


Alan, can you please apply that to an -ac
tree?

----
Summary:
Attached patch works around the corruption
of the high word of the ESP register, which
is the official bug of x86 CPUs. The bug
triggers only when the one is using the
16bit stack segment, and is described here:
http://www.intel.com/design/intarch/specupdt/27287402.PDF


Acked-by: Linus Torvalds <torvalds@osdl.org>
Acked-by: Petr Vandrovec <vandrove@vc.cvut.cz>
Acked-by: Pavel Machek <pavel@ucw.cz>
Signed-off-by: Stas Sergeev <stsp@aknet.ru>


[-- Attachment #2: linux-2.6.11-ac2-stk7a.diff --]
[-- Type: text/x-patch, Size: 10969 bytes --]

diff -ur linux-2.6.11-ac2/arch/i386/kernel/cpu/common.c linux-2.6.11-ac2-stk/arch/i386/kernel/cpu/common.c
--- linux-2.6.11-ac2/arch/i386/kernel/cpu/common.c	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/cpu/common.c	2005-03-13 20:36:12.000000000 +0300
@@ -21,6 +21,10 @@
 DEFINE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 EXPORT_PER_CPU_SYMBOL(cpu_gdt_table);
 
+unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DEFINE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+EXPORT_PER_CPU_SYMBOL(cpu_16bit_stack);
+
 static int cachesize_override __initdata = -1;
 static int disable_x86_fxsr __initdata = 0;
 static int disable_x86_serial_nr __initdata = 1;
@@ -557,6 +561,7 @@
 	int cpu = smp_processor_id();
 	struct tss_struct * t = &per_cpu(init_tss, cpu);
 	struct thread_struct *thread = &current->thread;
+	__u32 stk16_off = (__u32)&per_cpu(cpu_16bit_stack, cpu);
 
 	if (cpu_test_and_set(cpu, cpu_initialized)) {
 		printk(KERN_WARNING "CPU#%d already initialized!\n", cpu);
@@ -579,6 +584,13 @@
 	 */
 	memcpy(&per_cpu(cpu_gdt_table, cpu), cpu_gdt_table,
 	       GDT_SIZE);
+
+	/* Set up GDT entry for 16bit stack */
+	*(__u64 *)&(per_cpu(cpu_gdt_table, cpu)[GDT_ENTRY_ESPFIX_SS]) |=
+		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
+		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
+		(CPU_16BIT_STACK_SIZE - 1);
+
 	cpu_gdt_descr[cpu].size = GDT_SIZE - 1;
 	cpu_gdt_descr[cpu].address =
 	    (unsigned long)&per_cpu(cpu_gdt_table, cpu);
diff -ur linux-2.6.11-ac2/arch/i386/kernel/entry.S linux-2.6.11-ac2-stk/arch/i386/kernel/entry.S
--- linux-2.6.11-ac2/arch/i386/kernel/entry.S	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/entry.S	2005-03-14 00:45:16.000000000 +0300
@@ -47,6 +47,7 @@
 #include <asm/segment.h>
 #include <asm/smp.h>
 #include <asm/page.h>
+#include <asm/desc.h>
 #include "irq_vectors.h"
 
 #define nr_syscalls ((syscall_table_size)/4)
@@ -78,7 +79,7 @@
 #define preempt_stop		cli
 #else
 #define preempt_stop
-#define resume_kernel		restore_all
+#define resume_kernel		restore_nocheck
 #endif
 
 #define SAVE_ALL \
@@ -122,24 +123,6 @@
 .previous
 
 
-#define RESTORE_ALL	\
-	RESTORE_REGS	\
-	addl $4, %esp;	\
-1:	iret;		\
-.section .fixup,"ax";   \
-2:	sti;		\
-	movl $(__USER_DS), %edx; \
-	movl %edx, %ds; \
-	movl %edx, %es; \
-	movl $11,%eax;	\
-	call do_exit;	\
-.previous;		\
-.section __ex_table,"a";\
-	.align 4;	\
-	.long 1b,2b;	\
-.previous
-
-
 ENTRY(ret_from_fork)
 	pushl %eax
 	call schedule_tail
@@ -163,7 +146,7 @@
 	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
 	movb CS(%esp), %al
 	testl $(VM_MASK | 3), %eax
-	jz resume_kernel		# returning to kernel or vm86-space
+	jz resume_kernel
 ENTRY(resume_userspace)
  	cli				# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
@@ -178,7 +161,7 @@
 ENTRY(resume_kernel)
 	cli
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
-	jnz restore_all
+	jnz restore_nocheck
 need_resched:
 	movl TI_flags(%ebp), %ecx	# need_resched set ?
 	testb $_TIF_NEED_RESCHED, %cl
@@ -257,8 +240,56 @@
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx	# current->work
 	jne syscall_exit_work
+
 restore_all:
-	RESTORE_ALL
+	movl EFLAGS(%esp), %eax		# mix EFLAGS, SS and CS
+	movb OLDSS(%esp), %ah
+	movb CS(%esp), %al
+	andl $(VM_MASK | (4 << 8) | 3), %eax
+	cmpl $((4 << 8) | 3), %eax
+	je ldt_ss			# returning to user-space with LDT SS
+restore_nocheck:
+	RESTORE_REGS
+	addl $4, %esp
+1:	iret
+.section .fixup,"ax"
+iret_exc:
+	sti
+	movl $__USER_DS, %edx
+	movl %edx, %ds
+	movl %edx, %es
+	movl $11,%eax
+	call do_exit
+.previous
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
+ldt_ss:
+	larl OLDSS(%esp), %eax
+	jnz restore_nocheck
+	testl $0x00400000, %eax		# returning to 32bit stack?
+	jnz restore_nocheck		# allright, normal return
+	/* If returning to userspace with 16bit stack,
+	 * try to fix the higher word of ESP, as the CPU
+	 * won't restore it.
+	 * This is an "official" bug of all the x86-compatible
+	 * CPUs, which we can try to work around to make
+	 * dosemu and wine happy. */
+	subl $8, %esp		# reserve space for switch16 pointer
+	cli
+	movl %esp, %eax
+	/* Set up the 16bit stack frame with switch32 pointer on top,
+	 * and a switch16 pointer on top of the current frame. */
+	call setup_x86_bogus_stack
+	RESTORE_REGS
+	lss 20+4(%esp), %esp	# switch to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
 
 	# perform work that needs to be done immediately before resumption
 	ALIGN
@@ -334,6 +365,27 @@
 	movl $-ENOSYS,EAX(%esp)
 	jmp resume_userspace
 
+#define FIXUP_ESPFIX_STACK \
+	movl %esp, %eax; \
+	/* switch to 32bit stack using the pointer on top of 16bit stack */ \
+	lss %ss:CPU_16BIT_STACK_SIZE-8, %esp; \
+	/* copy data from 16bit stack to 32bit stack */ \
+	call fixup_x86_bogus_stack; \
+	/* put ESP to the proper location */ \
+	movl %eax, %esp;
+#define UNWIND_ESPFIX_STACK \
+	pushl %eax; \
+	movl %ss, %eax; \
+	/* see if on 16bit stack */ \
+	cmpw $__ESPFIX_SS, %ax; \
+	jne 28f; \
+	movl $__KERNEL_DS, %edx; \
+	movl %edx, %ds; \
+	movl %edx, %es; \
+	/* switch to 32bit stack */ \
+	FIXUP_ESPFIX_STACK \
+28:	popl %eax;
+
 /*
  * Build the entry stubs and pointer table with
  * some assembler magic.
@@ -388,7 +440,9 @@
 	pushl %ecx
 	pushl %ebx
 	cld
-	movl %es, %ecx
+	pushl %es
+	UNWIND_ESPFIX_STACK
+	popl %ecx
 	movl ES(%esp), %edi		# get the function address
 	movl ORIG_EAX(%esp), %edx	# get the error code
 	movl %eax, ORIG_EAX(%esp)
@@ -470,6 +524,11 @@
  * fault happened on the sysenter path.
  */
 ENTRY(nmi)
+	pushl %eax
+	movl %ss, %eax
+	cmpw $__ESPFIX_SS, %ax
+	popl %eax
+	je nmi_16bit_stack
 	cmpl $sysenter_entry,(%esp)
 	je nmi_stack_fixup
 	pushl %eax
@@ -489,7 +548,7 @@
 	xorl %edx,%edx		# zero error code
 	movl %esp,%eax		# pt_regs pointer
 	call do_nmi
-	RESTORE_ALL
+	jmp restore_all
 
 nmi_stack_fixup:
 	FIX_STACK(12,nmi_stack_correct, 1)
@@ -505,6 +564,29 @@
 	FIX_STACK(24,nmi_stack_correct, 1)
 	jmp nmi_stack_correct
 
+nmi_16bit_stack:
+	/* create the pointer to lss back */
+	pushl %ss
+	pushl %esp
+	movzwl %sp, %esp
+	addw $4, (%esp)
+	/* copy the iret frame of 12 bytes */
+	.rept 3
+	pushl 16(%esp)
+	.endr
+	pushl %eax
+	SAVE_ALL
+	FIXUP_ESPFIX_STACK		# %eax == %esp
+	xorl %edx,%edx			# zero error code
+	call do_nmi
+	RESTORE_REGS
+	lss 12+4(%esp), %esp		# back to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 ENTRY(int3)
 	pushl $-1			# mark this as an int
 	SAVE_ALL
diff -ur linux-2.6.11-ac2/arch/i386/kernel/head.S linux-2.6.11-ac2-stk/arch/i386/kernel/head.S
--- linux-2.6.11-ac2/arch/i386/kernel/head.S	2005-03-13 17:56:19.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/head.S	2005-03-13 20:36:12.000000000 +0300
@@ -512,7 +512,7 @@
 	.quad 0x00009a0000000000	/* 0xc0 APM CS 16 code (16 bit) */
 	.quad 0x0040920000000000	/* 0xc8 APM DS    data */
 
-	.quad 0x0000000000000000	/* 0xd0 - unused */
+	.quad 0x0000920000000000	/* 0xd0 - ESPFIX 16-bit SS */
 	.quad 0x0000000000000000	/* 0xd8 - unused */
 	.quad 0x0000000000000000	/* 0xe0 - unused */
 	.quad 0x0000000000000000	/* 0xe8 - unused */
diff -ur linux-2.6.11-ac2/arch/i386/kernel/traps.c linux-2.6.11-ac2-stk/arch/i386/kernel/traps.c
--- linux-2.6.11-ac2/arch/i386/kernel/traps.c	2005-03-13 17:59:45.000000000 +0300
+++ linux-2.6.11-ac2-stk/arch/i386/kernel/traps.c	2005-03-13 20:36:12.000000000 +0300
@@ -895,6 +895,51 @@
 #endif
 }
 
+fastcall void setup_x86_bogus_stack(unsigned char * stk)
+{
+	unsigned long *switch16_ptr, *switch32_ptr;
+	struct pt_regs *regs;
+	unsigned long stack_top, stack_bot;
+	unsigned short iret_frame16_off;
+	int cpu = smp_processor_id();
+	/* reserve the space on 32bit stack for the magic switch16 pointer */
+	memmove(stk, stk + 8, sizeof(struct pt_regs));
+	switch16_ptr = (unsigned long *)(stk + sizeof(struct pt_regs));
+	regs = (struct pt_regs *)stk;
+	/* now the switch32 on 16bit stack */
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	iret_frame16_off = CPU_16BIT_STACK_SIZE - 8 - 20;
+	/* copy iret frame on 16bit stack */
+	memcpy((void *)(stack_bot + iret_frame16_off), &regs->eip, 20);
+	/* fill in the switch pointers */
+	switch16_ptr[0] = (regs->esp & 0xffff0000) | iret_frame16_off;
+	switch16_ptr[1] = __ESPFIX_SS;
+	switch32_ptr[0] = (unsigned long)stk + sizeof(struct pt_regs) +
+		8 - CPU_16BIT_STACK_SIZE;
+	switch32_ptr[1] = __KERNEL_DS;
+}
+
+fastcall unsigned char * fixup_x86_bogus_stack(unsigned short sp)
+{
+	unsigned long *switch32_ptr;
+	unsigned char *stack16, *stack32;
+	unsigned long stack_top, stack_bot;
+	int len;
+	int cpu = smp_processor_id();
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	/* copy the data from 16bit stack to 32bit stack */
+	len = CPU_16BIT_STACK_SIZE - 8 - sp;
+	stack16 = (unsigned char *)(stack_bot + sp);
+	stack32 = (unsigned char *)
+		(switch32_ptr[0] + CPU_16BIT_STACK_SIZE - 8 - len);
+	memcpy(stack32, stack16, len);
+	return stack32;
+}
+
 /*
  *  'math_state_restore()' saves the current math information in the
  * old math state array, and gets the new ones from the current task
diff -ur linux-2.6.11-ac2/include/asm-i386/desc.h linux-2.6.11-ac2-stk/include/asm-i386/desc.h
--- linux-2.6.11-ac2/include/asm-i386/desc.h	2005-03-13 17:57:12.000000000 +0300
+++ linux-2.6.11-ac2-stk/include/asm-i386/desc.h	2005-03-13 20:36:12.000000000 +0300
@@ -4,6 +4,8 @@
 #include <asm/ldt.h>
 #include <asm/segment.h>
 
+#define CPU_16BIT_STACK_SIZE 1024
+
 #ifndef __ASSEMBLY__
 
 #include <linux/preempt.h>
@@ -15,6 +17,9 @@
 extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
 DECLARE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 
+extern unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DECLARE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+
 struct Xgt_desc_struct {
 	unsigned short size;
 	unsigned long address __attribute__((packed));
diff -ur linux-2.6.11-ac2/include/asm-i386/segment.h linux-2.6.11-ac2-stk/include/asm-i386/segment.h
--- linux-2.6.11-ac2/include/asm-i386/segment.h	2005-01-04 03:59:37.000000000 +0300
+++ linux-2.6.11-ac2-stk/include/asm-i386/segment.h	2005-03-13 20:36:12.000000000 +0300
@@ -38,7 +38,7 @@
  *  24 - APM BIOS support
  *  25 - APM BIOS support 
  *
- *  26 - unused
+ *  26 - ESPFIX small SS
  *  27 - unused
  *  28 - unused
  *  29 - unused
@@ -71,6 +71,9 @@
 #define GDT_ENTRY_PNPBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 6)
 #define GDT_ENTRY_APMBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 11)
 
+#define GDT_ENTRY_ESPFIX_SS		(GDT_ENTRY_KERNEL_BASE + 14)
+#define __ESPFIX_SS (GDT_ENTRY_ESPFIX_SS * 8)
+
 #define GDT_ENTRY_DOUBLEFAULT_TSS	31
 
 /*

Re: [patch] x86: fix ESP corruption CPU bug (take 2)

[Alan Cox]

On Mon, Mar 14, 2005 at 01:06:36AM +0300, Stas Sergeev wrote:
> Alan, can you please apply that to an -ac
> tree?

Ask Andrew Morton as it belongs in the -mm tree

Re: [patch] x86: fix ESP corruption CPU bug (take 2)

[Stas Sergeev]

Hi!

Alan Cox wrote:
>> Alan, can you please apply that to an -ac
>> tree?
> Ask Andrew Morton as it belongs in the -mm tree
Actually I tried that already. Andrew
had nothing against that patch personally,
as well as Linus, but after all that didn't
work:
http://lkml.org/lkml/2005/1/3/260

So it can't be applied to -mm, and not
depending on the kgdb-ga patch allowed for
some extra optimization.
So I have to try the -ac as the last resort.
I realized you won't be pleased with that
too much. If you are confident it doesn't
fit the -ac tree by whatever means, I can
understand that. (In that case I'll much
appreciate the suggestion what other tree
should I try.)

Re: [patch] x86: fix ESP corruption CPU bug (take 2)

[Andrew Morton]

Stas Sergeev <stsp@aknet.ru> wrote:
>
> Alan Cox wrote:
>  >> Alan, can you please apply that to an -ac
>  >> tree?
>  > Ask Andrew Morton as it belongs in the -mm tree
>  Actually I tried that already.

I added this patch to -mm.

> Andrew
>  had nothing against that patch personally,
>  as well as Linus, but after all that didn't
>  work:
>  http://lkml.org/lkml/2005/1/3/260
> 
>  So it can't be applied to -mm, and not
>  depending on the kgdb-ga patch allowed for
>  some extra optimization.

The rule is:

- If the patch patches something which is in Linus's kernel, prepare a
  diff against Linus's latest kernel.

- If the patch patches something which is only in -mm, prepare a patch
  against -mm.

In this case, I merged the patch prior to the kgdb patch and then fixed
up the fallout.

(If that causes kgdb to break in non-obvious-to-me ways then I might come
calling "help".  We'll see)

x86: spin_unlock(), spin_unlock_irq() & others are out of line ?

[Eric Dumazet]

Hi all

I noticed that in current linux kernel versions (2.6.11), some basic 
functions are out of line (not inlined)

Example of a call to spin_unlock(&somelock)
c01069fa:   b8 e8 7b 35 c0          mov    $0xc0357be8,%eax
c01069ff:   e8 3c e4 1f 00          call   c0304e40 <_spin_unlock>


c0304e40 <_spin_unlock>:
c0304e40:   c6 00 01                movb   $0x1,(%eax)
c0304e43:   c3                      ret


Same problem for _write_unlock(), _read_unlock(), _spin_unlock_irq(), ...

That seems odd, and I fail to see the reason for that. (It's OK for 
complex functions, but not for very short ones...)

Is it a regression, or is it needed ?

configuration :
	- SMP
	- Processor family (Pentium-4/Celeron(P4-based)/Pentium-4 M/Xeon)
	- No "Generic x86 support"

Thank you
Eric Dumazet

Re: x86: spin_unlock(), spin_unlock_irq() & others are out of line ?

[Lee Revell]

On Tue, 2005-03-15 at 11:48 +0100, Eric Dumazet wrote:
> Is it a regression, or is it needed ?
> 

Please see the "Completely out of line spinlocks" thread from about a
month ago.

Lee

Re: [patch] x86: fix ESP corruption CPU bug (take 2)

[Stas Sergeev]

Hello.

Andrew Morton wrote:
> I added this patch to -mm.
Many thanks!

Alan, sorry for bothering you with that.


> - If the patch patches something which is in Linus's kernel, prepare a
>   diff against Linus's latest kernel.
> - If the patch patches something which is only in -mm, prepare a patch
>   against -mm.
> In this case, I merged the patch prior to the kgdb patch and then fixed
> up the fallout.
> (If that causes kgdb to break in non-obvious-to-me ways then I might come
> calling "help".
Thanks for explanations. That's what I
tried to find in your "The perfect patch"
guidelines many times, but it is not there
in such details.

Re: [patch] x86: fix ESP corruption CPU bug

[Pavel Machek]

On Ne 13-03-05 13:13:16, Linus Torvalds wrote:
> 
> 
> On Sun, 13 Mar 2005, Stas Sergeev wrote:
> >
> > Such an optimization will cost three more
> > instructions, one of which is a "taken"
> > jump.
> 
> I think Pavel missed the fact that you have to check the VM86 bit in
> eflags before you check SS, since otherwise SS doesn't mean anything
> special at all (ie checking for just the normal SS isn't correct: you
> could have a 16-bit SS that looks normal, but is actually a vm86 segment).
> 
> Pavel: for the same reason you have to check the low two bits of CS too, 
> since if they are zero, then SS hasn't been saved on the stack at all, so 
> comparing it against some normal value is meaningless.

Yes, I missed that one, thanks.

What about flag similar to _TIF_SYSCALL_TRACE (call it
_TIF_THIS_BEAST_USES_V86 or something), and only do the tests in the
slowpath if it is set? As normal applications do not use v86, we could
make this 0 instructions in syscall fast path...
								Pavel
-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Re: [patch] x86: fix ESP corruption CPU bug

[Linus Torvalds]

On Mon, 14 Mar 2005, Pavel Machek wrote:
> 
> What about flag similar to _TIF_SYSCALL_TRACE (call it
> _TIF_THIS_BEAST_USES_V86 or something), and only do the tests in the
> slowpath if it is set? As normal applications do not use v86, we could
> make this 0 instructions in syscall fast path...

It wouldn't help you. You'd need to mix in two of the values anyway, so at 
most you'd save one instruction. And the cost would be that anything that 
has ever used vm86 mode (can you say "X server"?) would be slower. Not a 
good trade-off.

Oh, I guess you could clear the flag when you know there's no vm86 state
anywhere (easy enough, those things never nest), but then it still comes
back to "extra complexity that you can get wrong, just to save a single
"mov" instruction - that "mov" may have partial-register-stall issues, 
but still..).

		Linus

Re: [patch] x86: fix ESP corruption CPU bug

[Linus Torvalds]

On Sun, 13 Mar 2005, Linus Torvalds wrote:
> 
> That said, the "ldt_ss" case should be moved _after_ the conditional
> tests, since most CPU's out there will do static prediction based on
> forward/backwards direction 

Btw, Stas, one thing I'd really like to see is even a partial list of 
anything that actually cares about this. Ie, if there is some known 
Windows app where Wine works better or something like that, just adding 
that information to the comments would be hugely appreciated. 

Another way of saying the same thing: I absolutely hate seeing patches 
that fix some theoretical issue that no Linux apps will ever care about. 
So I'd like to have a bit more of a case for this patch, since I know what 
the case against it is ;)

		Linus

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hi,

Linus Torvalds wrote:
> Btw, Stas, one thing I'd really like to see is even a partial list of 
> anything that actually cares about this. Ie, if there is some known 
> Windows app where Wine works better or something like that, just adding
I am not using Wine too much, but I've
found this:
http://cvs.winehq.org/cvsweb/~checkout~/wine/dlls/winedos/int31.c?rev=1.41&content-type=text/plain
---
/* due to a flaw in some CPUs (at least mine), it is best to mark stack segments as 32-bit if they
   can be used in 32-bit code. Otherwise, these CPUs may not set the high word of esp during a
   ring transition (from kernel code) to the 16-bit stack, and this causes trouble if executing
   32-bit code using this stack. */
---
I added win-devel to CC, maybe people there
can tell if that patch has any value for them
or not.
The reference to the original patch:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0503.1/1794.html

Dosemu looks a little better on that, the
whole chapter of the docs is dedicated to
that problem:
http://dosemu.sourceforge.net/docs/EMUfailure/t1.html#AEN55
There you can find a (relatively small)
list of the programs that are affected,
but I personally have the old Microsoft
linker that crashes, and a few more DOS
games.

> Another way of saying the same thing: I absolutely hate seeing patches 
> that fix some theoretical issue that no Linux apps will ever care about. 
No, it is not theoretical, but it is mainly
about a DOS games and an MS linker, as for
me. The things I'd like to get working, but
the ones you may not care too much about:)
The particular game I want to get working,
is "Master of Orion 2" for DOS.

> So I'd like to have a bit more of a case for this patch, since I know what 
> the case against it is ;)
Yep, and the informational leak it closes,
looks also rather minor.
So it is only a matter of how do you care
about the dosemu and the DOS games under
linux. Considering the amount of the
dosemu-related code in vm86.c, I guess you
care:)
And uhm, adding the list of the DOS games
to the comments of the Linux kernel code,
doesn't sound like a good idea to me:)

Re: [patch] x86: fix ESP corruption CPU bug

[Andi Kleen]

Stas Sergeev <stsp@aknet.ru> writes:
>
>> Another way of saying the same thing: I absolutely hate seeing
>> patches that fix some theoretical issue that no Linux apps will ever
>> care about.
> No, it is not theoretical, but it is mainly
> about a DOS games and an MS linker, as for
> me. The things I'd like to get working, but
> the ones you may not care too much about:)
> The particular game I want to get working,
> is "Master of Orion 2" for DOS.

How about you just run it in dosbox instead of dosemu ?

-Andi

[BUG?] x86_64 : Can not read /dev/kmem ?

[Eric Dumazet]

Hi Andi

I tried to read /dev/kmem on x86_64 (linux-2.6.11) and got no success.

read() or pread() returns EINVAL

I tried mmap() too : mmap() calls succeed, but as soon the user process 
dereference memory, we get :

tinfo: Corrupted page table at address 2aaaaaabf800
PGD 8a983067 PUD c7e5a067 PMD 91588067 PTE ffffffff8048a025
Bad pagetable: 000d [1] SMP
CPU 0
Modules linked in: ipt_REJECT
Pid: 10892, comm: tinfo Not tainted 2.6.11
RIP: 0033:[<0000000000100562>] [<0000000000100562>]
RSP: 002b:00007ffffffff790  EFLAGS: 00010217
RAX: 00002aaaaaabf000 RBX: 00002aaaaabbe000 RCX: 00002aaaaac8fc0c
RDX: 0000000000000001 RSI: 0000000000001000 RDI: 0000000000000000
RBP: 00007ffffffff7f8 R08: 0000000000000003 R09: ffffffff8048a000
R10: 0000000000000001 R11: 0000000000000206 R12: 00000000001005b0
R13: 0000000000000001 R14: 00002aaaaadfdfe8 R15: 0000000000100530
FS:  00002aaaaabcb970(0000) GS:ffffffff804866c0(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 00002aaaaaabf800 CR3: 0000000090368000 CR4: 00000000000006e0
Process tinfo (pid: 10892, threadinfo ffff8100901b0000, task 
ffff8100c7d976c0)

RIP [<0000000000100562>] RSP <00007ffffffff790>



Thank you

Eric Dumazet
----------------------------------------------------------------
# cat tinfo.c

#define _XOPEN_SOURCE 500
#include <unistd.h>
#include <stdio.h>
#include <fcntl.h>

struct tcp_hashinfo {
         struct tcp_ehash_bucket *__tcp_ehash;
         struct tcp_bind_hashbucket *__tcp_bhash;

         int __tcp_bhash_size;
         int __tcp_ehash_size;

} tcp_hashinfo;

#define TCPINFO_ADDR 0xffffffff8048a000 /* tcp_hashinfo */

int main()
{
int fd = open("/dev/kmem", O_RDONLY) ;


if (pread(fd, &tcp_hashinfo, sizeof(tcp_hashinfo), TCPINFO_ADDR) == -1) {
         lseek(fd, TCPINFO_ADDR, 0) ;
         if (read(fd, &tcp_hashinfo, sizeof(tcp_hashinfo)) == -1) {
                 perror("Can not read /dev/kmem ?") ;
                 return 1 ;
                 }
         }
printf("ehash=%p esize=%d bhash=%p bsize=%d\n",
         tcp_hashinfo.__tcp_ehash,
         tcp_hashinfo.__tcp_ehash_size,
         tcp_hashinfo.__tcp_bhash,
         tcp_hashinfo.__tcp_bhash_size) ;
return 0 ;
}

Re: [BUG?] x86_64 : Can not read /dev/kmem ?

[Andi Kleen]

Sorry for the late answer.

On Mon, Mar 14, 2005 at 11:37:19AM +0100, Eric Dumazet wrote:
> 
> Hi Andi
> 
> I tried to read /dev/kmem on x86_64 (linux-2.6.11) and got no success.
> 
> read() or pread() returns EINVAL

Yes. I fixed this once for 2.4, but somehow the changes never made 
it into 2.6. The VFS code doesn't like negative offsets, and the kernel
addresses are negative.

> 
> I tried mmap() too : mmap() calls succeed, but as soon the user process 
> dereference memory, we get :

Hmm, looks like a bug yes. I can reproduce it. Thanks for the report.
Will investigate later.

-Andi

Re: [BUG?] x86_64 : Can not read /dev/kmem ?

[Andi Kleen]

On Mon, Mar 14, 2005 at 11:37:19AM +0100, Eric Dumazet wrote:
> 
> Hi Andi
> 
> I tried to mmap /dev/kmem on x86_64 (linux-2.6.11) and got no success.
> 

Here's a patch that fixes the problem to me.


Fix mmap of /dev/kmem. It cannot ever have worked before.

vmalloc is still not supported because that would be more
complicated.

Signed-off-by: Andi Kleen <ak@suse.de>


diff -u linux-2.6.11/drivers/char/mem.c-o linux-2.6.11/drivers/char/mem.c
--- linux-2.6.11/drivers/char/mem.c-o	2004-12-24 22:34:47.000000000 +0100
+++ linux-2.6.11/drivers/char/mem.c	2005-03-22 12:36:33.852319000 +0100
@@ -211,6 +211,23 @@
 	return 0;
 }
 
+static int mmap_kmem(struct file * file, struct vm_area_struct * vma)
+{
+        unsigned long long val;
+	/* 
+	 * RED-PEN: on some architectures there is more mapped memory
+	 * than available in mem_map which pfn_valid checks
+	 * for. Perhaps should add a new macro here.
+	 * 
+	 * RED-PEN: vmalloc is not supported right now. 
+	 */
+	if (!pfn_valid(vma->vm_pgoff))
+		return -EIO;
+	val = (u64)vma->vm_pgoff << PAGE_SHIFT;
+	vma->vm_pgoff = __pa(val) >> PAGE_SHIFT; 
+	return mmap_mem(file, vma);
+}
+
 extern long vread(char *buf, char *addr, unsigned long count);
 extern long vwrite(char *buf, char *addr, unsigned long count);
 
@@ -567,7 +584,6 @@
 	return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
 }
 
-#define mmap_kmem	mmap_mem
 #define zero_lseek	null_lseek
 #define full_lseek      null_lseek
 #define write_zero	write_null

Re: [patch] x86: fix ESP corruption CPU bug

[Jakob Eriksson]

Andi Kleen wrote:

>Stas Sergeev <stsp@aknet.ru> writes:
>  
>
>>>Another way of saying the same thing: I absolutely hate seeing
>>>patches that fix some theoretical issue that no Linux apps will ever
>>>care about.
>>>      
>>>
>>No, it is not theoretical, but it is mainly
>>about a DOS games and an MS linker, as for
>>me. The things I'd like to get working, but
>>the ones you may not care too much about:)
>>The particular game I want to get working,
>>is "Master of Orion 2" for DOS.
>>    
>>
>
>How about you just run it in dosbox instead of dosemu ?
>  
>

Yes, that's a solution of course, but it is a bit like saying why
not use Open Office instead of MS Word.

A long term goal of wine is to support DOS apps to. Of course
it's not a priority, but it's there.

regards,
Jakob

Re: [patch] x86: fix ESP corruption CPU bug

[linux-os]

On Mon, 14 Mar 2005, Jakob Eriksson wrote:

> Andi Kleen wrote:
>
>> Stas Sergeev <stsp@aknet.ru> writes:
>> 
>>>> Another way of saying the same thing: I absolutely hate seeing
>>>> patches that fix some theoretical issue that no Linux apps will ever
>>>> care about.
>>>> 
>>> No, it is not theoretical, but it is mainly
>>> about a DOS games and an MS linker, as for
>>> me. The things I'd like to get working, but
>>> the ones you may not care too much about:)
>>> The particular game I want to get working,
>>> is "Master of Orion 2" for DOS.
>>> 
>> 
>> How about you just run it in dosbox instead of dosemu ?
>> 
>
> Yes, that's a solution of course, but it is a bit like saying why
> not use Open Office instead of MS Word.
>
> A long term goal of wine is to support DOS apps to. Of course
> it's not a priority, but it's there.
>
> regards,
> Jakob
>

Can you tell me how the invisible high-word (invisible in VM-86, and
in real mode) could possibly harm something running in VM-86 or
read-mode ???  I don't even think it's a BUG. If the transition
into and out of VM-86 doesn't handle the fact that the high-word
of the stack hasn't been used in VM-86, then that piece of code
is bad (the SP isn't even the same stack, BTW).


Cheers,
Dick Johnson
Penguin : Linux version 2.6.11 on an i686 machine (5537.79 BogoMips).
  Notice : All mail here is now cached for review by Dictator Bush.
                  98.36% of all statistics are fiction.

Re: [patch] x86: fix ESP corruption CPU bug

[Pavel Machek]

Hi!

> Can you tell me how the invisible high-word (invisible in VM-86, and
> in real mode) could possibly harm something running in VM-86 or
> read-mode ???  I don't even think it's a BUG. If the transition

You can have protected-mode application running in dosemu with 16-bit
stack segment.
								Pavel
-- 
People were complaining that M$ turns users into beta-testers...
...jr ghea gurz vagb qrirybcref, naq gurl frrz gb yvxr vg gung jnl!

Re: [patch] x86: fix ESP corruption CPU bug

[Brian Gerst]

linux-os wrote:
> On Mon, 14 Mar 2005, Jakob Eriksson wrote:
> 
>> Andi Kleen wrote:
>>
>>> Stas Sergeev <stsp@aknet.ru> writes:
>>>
>>>>> Another way of saying the same thing: I absolutely hate seeing
>>>>> patches that fix some theoretical issue that no Linux apps will ever
>>>>> care about.
>>>>>
>>>> No, it is not theoretical, but it is mainly
>>>> about a DOS games and an MS linker, as for
>>>> me. The things I'd like to get working, but
>>>> the ones you may not care too much about:)
>>>> The particular game I want to get working,
>>>> is "Master of Orion 2" for DOS.
>>>>
>>>
>>> How about you just run it in dosbox instead of dosemu ?
>>>
>>
>> Yes, that's a solution of course, but it is a bit like saying why
>> not use Open Office instead of MS Word.
>>
>> A long term goal of wine is to support DOS apps to. Of course
>> it's not a priority, but it's there.
>>
>> regards,
>> Jakob
>>
> 
> Can you tell me how the invisible high-word (invisible in VM-86, and
> in real mode) could possibly harm something running in VM-86 or
> read-mode ???  I don't even think it's a BUG. If the transition
> into and out of VM-86 doesn't handle the fact that the high-word
> of the stack hasn't been used in VM-86, then that piece of code
> is bad (the SP isn't even the same stack, BTW).

Because even in 16-bit mode (real, vm86 or 16-bit protected) you can use 
32-bit instructions, with an operand and/or address size override 
prefix.  Of course this only works on 386 or later.

--
				Brian Gerst

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hello.

Brian Gerst wrote:
>> Can you tell me how the invisible high-word (invisible in VM-86, and
>> in real mode) could possibly harm something running in VM-86 or
>> read-mode ???  I don't even think it's a BUG. If the transition
>> into and out of VM-86 doesn't handle the fact that the high-word
>> of the stack hasn't been used in VM-86, then that piece of code
>> is bad (the SP isn't even the same stack, BTW).
> Because even in 16-bit mode (real, vm86 or 16-bit protected) you can use 
> 32-bit instructions, with an operand and/or address size override 
> prefix.
And the real problem is when the pure
32bit code is starting to use the 16bit
stack for some strange reasons. Looks like
the common technique for the early dos4gw
-based apps...

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hi,

Jakob Eriksson wrote:
> A long term goal of wine is to support DOS apps to. Of course
> it's not a priority, but it's there.
Yes, that's exactly what I was hoping
for, thanks!
Even if no Windows apps do such a thing
(which wasn't confirmed yet), Wine may
still need that fix for the DOS support
in the future, and dosemu seems to be in
need of that fix already and for long
(that's where I am getting the use of it).
And we haven't heard a word for a VMWare
yet, and I think they may appreciate that
too.
Also, since the first version of the path,
I've been contacted by a few people asking
me to provide an updated version for 2.6.9
and 2.6.10. I don't know the reason, but
I know it was used (I think they were more
concerned about an aforementioned "information
leak", rather than about the %esp corruption).
So if the last problem with that patch was
that it is not really needed, I think it
no longer stays:)

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

Hello.

Andi Kleen wrote:
>> The particular game I want to get working,
>> is "Master of Orion 2" for DOS.
> How about you just run it in dosbox instead of dosemu ?
Way too slow, and there are other reasons
too, like the better networking support on
dosemu side (not for Orion, but for other
games is important), but that might be an
off-topic here:)

Re: [patch] x86: fix ESP corruption CPU bug

[Zoltan Boszormenyi]

> Stas Sergeev <stsp@aknet.ru> writes:
>>
>>> Another way of saying the same thing: I absolutely hate seeing
>>> patches that fix some theoretical issue that no Linux apps will ever
>>> care about.
>> No, it is not theoretical, but it is mainly
>> about a DOS games and an MS linker, as for
>> me. The things I'd like to get working, but
>> the ones you may not care too much about:)
>> The particular game I want to get working,
>> is "Master of Orion 2" for DOS.
> 
> How about you just run it in dosbox instead of dosemu ?
> 
> -Andi

Nah, don't insult a DOSemu developer. ;-) Stas is one of them...

Best regards,
Zoltán Böszörményi

[patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

[-- Attachment #1: Type: text/plain, Size: 744 bytes --]

Hi Andrew.

Attached patch works around the corruption
of the high word of the ESP register, which
is the official bug of x86 CPUs. The bug
triggers only when the one is using the
16bit stack segment.
Patch helps running many apps under dosemu,
and, according to the comments found in
Wine sources, also helps Wine.

The patch defines the per-CPU 16bit stacks,
and every time the process that uses 16bit
stack, returns to the userspace, we switch
to our 16bit stack and preload the high word
of ESP.
This also closes the "informational leak",
which is that the user process is not
supposed to know the kernel's ESP value.

Can this please be applied?

Acked-by: Linus Torvalds <torvalds@osdl.org>
|Signed-off-by: Stas Sergeev <stsp@aknet.ru>|


[-- Attachment #2: linux-2.6.10-mm1-stk4.diff --]
[-- Type: text/x-patch, Size: 11529 bytes --]

diff -urN linux-2.6.10-mm1/arch/i386/kernel/cpu/common.c linux-2.6.10-mm1-stk/arch/i386/kernel/cpu/common.c
--- linux-2.6.10-mm1/arch/i386/kernel/cpu/common.c	2005-01-03 22:56:03.000000000 +0300
+++ linux-2.6.10-mm1-stk/arch/i386/kernel/cpu/common.c	2005-01-03 23:13:21.000000000 +0300
@@ -16,6 +16,10 @@
 DEFINE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 EXPORT_PER_CPU_SYMBOL(cpu_gdt_table);
 
+unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DEFINE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+EXPORT_PER_CPU_SYMBOL(cpu_16bit_stack);
+
 static int cachesize_override __initdata = -1;
 static int disable_x86_fxsr __initdata = 0;
 static int disable_x86_serial_nr __initdata = 1;
@@ -505,6 +509,7 @@
 	int cpu = smp_processor_id();
 	struct tss_struct * t = &per_cpu(init_tss, cpu);
 	struct thread_struct *thread = &current->thread;
+	__u32 stk16_off = (__u32)&per_cpu(cpu_16bit_stack, cpu);
 
 	if (test_and_set_bit(cpu, &cpu_initialized)) {
 		printk(KERN_WARNING "CPU#%d already initialized!\n", cpu);
@@ -527,6 +532,13 @@
 	 */
 	memcpy(&per_cpu(cpu_gdt_table, cpu), cpu_gdt_table,
 	       GDT_SIZE);
+
+	/* Set up GDT entry for 16bit stack */
+	*(__u64 *)&(per_cpu(cpu_gdt_table, cpu)[GDT_ENTRY_ESPFIX_SS]) |=
+		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
+		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
+		(CPU_16BIT_STACK_SIZE - 1);
+
 	cpu_gdt_descr[cpu].size = GDT_SIZE - 1;
 	cpu_gdt_descr[cpu].address =
 	    (unsigned long)&per_cpu(cpu_gdt_table, cpu);
diff -urN linux-2.6.10-mm1/arch/i386/kernel/entry.S linux-2.6.10-mm1-stk/arch/i386/kernel/entry.S
--- linux-2.6.10-mm1/arch/i386/kernel/entry.S	2005-01-03 22:56:03.000000000 +0300
+++ linux-2.6.10-mm1-stk/arch/i386/kernel/entry.S	2005-01-04 00:47:03.000000000 +0300
@@ -47,6 +47,7 @@
 #include <asm/segment.h>
 #include <asm/smp.h>
 #include <asm/page.h>
+#include <asm/desc.h>
 #include "irq_vectors.h"
         /* We do not recover from a stack overflow, but at least
          * we know it happened and should be able to track it down.
@@ -90,7 +91,7 @@
 #define preempt_stop		cli
 #else
 #define preempt_stop
-#define resume_kernel		restore_all
+#define resume_kernel		resume_kernelX
 #endif
 
 #define SAVE_ALL \
@@ -135,24 +136,6 @@
 .previous
 
 
-#define RESTORE_ALL	\
-	RESTORE_REGS	\
-	addl $4, %esp;	\
-1:	iret;		\
-.section .fixup,"ax";   \
-2:	sti;		\
-	movl $(__USER_DS), %edx; \
-	movl %edx, %ds; \
-	movl %edx, %es; \
-	movl $11,%eax;	\
-	call do_exit;	\
-.previous;		\
-.section __ex_table,"a";\
-	.align 4;	\
-	.long 1b,2b;	\
-.previous
-
-
 ENTRY(ret_from_fork)
 	pushl %eax
 	call schedule_tail
@@ -176,7 +159,7 @@
 	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
 	movb CS(%esp), %al
 	testl $(VM_MASK | 3), %eax
-	jz resume_kernel		# returning to kernel or vm86-space
+	jz resume_kernel
 ENTRY(resume_userspace)
  	cli				# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
@@ -190,7 +173,7 @@
 #ifdef CONFIG_PREEMPT
 ENTRY(resume_kernel)
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
-	jnz restore_all
+	jnz resume_kernelX
 need_resched:
 	movl TI_flags(%ebp), %ecx	# need_resched set ?
 	testb $_TIF_NEED_RESCHED, %cl
@@ -204,6 +187,31 @@
 	jmp need_resched
 #endif
 
+ldt_ss:
+	larl OLDSS(%esp), %eax
+	jnz resume_kernelX
+	testl $0x00400000, %eax		# returning to 32bit stack?
+	jnz resume_kernelX		# allright, normal return
+	/* If returning to userspace with 16bit stack,
+	 * try to fix the higher word of ESP, as the CPU
+	 * won't restore it.
+	 * This is an "official" bug of all the x86-compatible
+	 * CPUs, which we can try to work around to make
+	 * dosemu and wine happy. */
+	subl $8, %esp		# reserve space for switch16 pointer
+	cli
+	movl %esp, %eax
+	/* Set up the 16bit stack frame with switch32 pointer on top,
+	 * and a switch16 pointer on top of the current frame. */
+	call setup_x86_bogus_stack
+	RESTORE_REGS
+	lss 20+4(%esp), %esp	# The magic pointer above iret frame
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 /* SYSENTER_RETURN points to after the "sysenter" instruction in
    the vsyscall page.  See vsyscall-sysentry.S, which defines the symbol.  */
 
@@ -273,21 +281,40 @@
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx	# current->work
 	jne syscall_exit_work
+
 restore_all:
-#ifdef CONFIG_TRAP_BAD_SYSCALL_EXITS
 	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
 	movb CS(%esp), %al
-	testl $(VM_MASK | 3), %eax
-	jz resume_kernelX		# returning to kernel or vm86-space
+	andl $(VM_MASK | 3), %eax
+	cmpl $3, %eax
+	jne resume_kernelX		# returning to kernel or vm86-space
 
+#ifdef CONFIG_TRAP_BAD_SYSCALL_EXITS
 	cmpl $0,TI_preempt_count(%ebp)  # non-zero preempt_count ?
-	jz resume_kernelX
-
+	jz 47f
         int $3
+47:
+#endif
 
+	testl $4, OLDSS(%esp)
+	jnz ldt_ss
 resume_kernelX:
-#endif
-	RESTORE_ALL
+	RESTORE_REGS
+	addl $4, %esp
+1:	iret
+.section .fixup,"ax"
+iret_exc:
+	sti
+	movl $(__USER_DS), %edx
+	movl %edx, %ds
+	movl %edx, %es
+	movl $11,%eax
+	call do_exit
+.previous
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
 
 	# perform work that needs to be done immediately before resumption
 	ALIGN
@@ -363,6 +390,25 @@
 	movl $-ENOSYS,EAX(%esp)
 	jmp resume_userspace
 
+/* Check if we are on 16bit stack. Can happen either on iret of ESPFIX,
+ * or in an exception handler after that iret... */
+#define FIXUP_ESPFIX_STACK \
+	movl %esp, %eax; \
+	/* Magic pointer is at the top of the 16bit stack */ \
+	lss %ss:CPU_16BIT_STACK_SIZE-8, %esp; \
+	call fixup_x86_bogus_stack; \
+	movl %eax, %esp;
+#define UNWIND_ESPFIX_STACK \
+	pushl %eax; \
+	movl %ss, %eax; \
+	cmpw $__ESPFIX_SS, %ax; \
+	jne 28f; \
+	movl $(__KERNEL_DS), %edx; \
+	movl %edx, %ds; \
+	movl %edx, %es; \
+	FIXUP_ESPFIX_STACK \
+28:	popl %eax;
+
 /*
  * Build the entry stubs and pointer table with
  * some assembler magic.
@@ -427,7 +473,9 @@
 	pushl %ecx
 	pushl %ebx
 	cld
-	movl %es, %ecx
+	pushl %es
+	UNWIND_ESPFIX_STACK
+	popl %ecx
 	movl ES(%esp), %edi		# get the function address
 	movl ORIG_EAX(%esp), %edx	# get the error code
 	movl %eax, ORIG_EAX(%esp)
@@ -509,6 +557,11 @@
  * fault happened on the sysenter path.
  */
 ENTRY(nmi)
+	pushl %eax
+	movl %ss, %eax
+	cmpw $__ESPFIX_SS, %ax
+	popl %eax
+	je nmi_16bit_stack
 	cmpl $sysenter_entry,(%esp)
 	je nmi_stack_fixup
 	pushl %eax
@@ -528,7 +581,7 @@
 	xorl %edx,%edx		# zero error code
 	movl %esp,%eax		# pt_regs pointer
 	call do_nmi
-	RESTORE_ALL
+	jmp restore_all
 
 nmi_stack_fixup:
 	FIX_STACK(12,nmi_stack_correct, 1)
@@ -544,6 +597,29 @@
 	FIX_STACK(24,nmi_stack_correct, 1)
 	jmp nmi_stack_correct
 
+nmi_16bit_stack:
+	/* create the pointer to lss back */
+	pushl %ss
+	pushl %esp
+	andl $0xffff, %esp
+	addw $4, (%esp)
+	/* copy the iret frame of 12 bytes */
+	.rept 3
+	pushl 16(%esp)
+	.endr
+	pushl %eax
+	SAVE_ALL
+	FIXUP_ESPFIX_STACK		# %eax == %esp
+	xorl %edx,%edx			# zero error code
+	call do_nmi
+	RESTORE_REGS
+	lss 12+4(%esp), %esp		# back to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 ENTRY(int3)
 	pushl $-1			# mark this as an int
 	SAVE_ALL
diff -urN linux-2.6.10-mm1/arch/i386/kernel/head.S linux-2.6.10-mm1-stk/arch/i386/kernel/head.S
--- linux-2.6.10-mm1/arch/i386/kernel/head.S	2005-01-03 22:56:03.000000000 +0300
+++ linux-2.6.10-mm1-stk/arch/i386/kernel/head.S	2005-01-03 23:13:21.000000000 +0300
@@ -512,7 +512,7 @@
 	.quad 0x00009a0000000000	/* 0xc0 APM CS 16 code (16 bit) */
 	.quad 0x0040920000000000	/* 0xc8 APM DS    data */
 
-	.quad 0x0000000000000000	/* 0xd0 - unused */
+	.quad 0x0000920000000000	/* 0xd0 - ESPFIX 16-bit SS */
 	.quad 0x0000000000000000	/* 0xd8 - unused */
 	.quad 0x0000000000000000	/* 0xe0 - unused */
 	.quad 0x0000000000000000	/* 0xe8 - unused */
diff -urN linux-2.6.10-mm1/arch/i386/kernel/traps.c linux-2.6.10-mm1-stk/arch/i386/kernel/traps.c
--- linux-2.6.10-mm1/arch/i386/kernel/traps.c	2005-01-03 22:56:03.000000000 +0300
+++ linux-2.6.10-mm1-stk/arch/i386/kernel/traps.c	2005-01-03 23:13:21.000000000 +0300
@@ -972,6 +972,51 @@
 #endif
 }
 
+fastcall void setup_x86_bogus_stack(unsigned char * stk)
+{
+	unsigned long *switch16_ptr, *switch32_ptr;
+	struct pt_regs *regs;
+	unsigned long stack_top, stack_bot;
+	unsigned short iret_frame16_off;
+	int cpu = smp_processor_id();
+	/* reserve the space on 32bit stack for the magic switch16 pointer */
+	memmove(stk, stk + 8, sizeof(struct pt_regs));
+	switch16_ptr = (unsigned long *)(stk + sizeof(struct pt_regs));
+	regs = (struct pt_regs *)stk;
+	/* now the switch32 on 16bit stack */
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	iret_frame16_off = CPU_16BIT_STACK_SIZE - 8 - 20;
+	/* copy iret frame on 16bit stack */
+	memcpy((void *)(stack_bot + iret_frame16_off), &regs->eip, 20);
+	/* fill in the switch pointers */
+	switch16_ptr[0] = (regs->esp & 0xffff0000) | iret_frame16_off;
+	switch16_ptr[1] = __ESPFIX_SS;
+	switch32_ptr[0] = (unsigned long)stk + sizeof(struct pt_regs) +
+		8 - CPU_16BIT_STACK_SIZE;
+	switch32_ptr[1] = __KERNEL_DS;
+}
+
+fastcall unsigned char * fixup_x86_bogus_stack(unsigned short sp)
+{
+	unsigned long *switch32_ptr;
+	unsigned char *stack16, *stack32;
+	unsigned long stack_top, stack_bot;
+	int len;
+	int cpu = smp_processor_id();
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	/* copy the data from 16bit stack to 32bit stack */
+	len = CPU_16BIT_STACK_SIZE - 8 - sp;
+	stack16 = (unsigned char *)(stack_bot + sp);
+	stack32 = (unsigned char *)
+		(switch32_ptr[0] + CPU_16BIT_STACK_SIZE - 8 - len);
+	memcpy(stack32, stack16, len);
+	return stack32;
+}
+
 /*
  *  'math_state_restore()' saves the current math information in the
  * old math state array, and gets the new ones from the current task
diff -urN linux-2.6.10-mm1/include/asm-i386/desc.h linux-2.6.10-mm1-stk/include/asm-i386/desc.h
--- linux-2.6.10-mm1/include/asm-i386/desc.h	2005-01-03 22:55:08.000000000 +0300
+++ linux-2.6.10-mm1-stk/include/asm-i386/desc.h	2005-01-03 23:13:21.000000000 +0300
@@ -4,6 +4,8 @@
 #include <asm/ldt.h>
 #include <asm/segment.h>
 
+#define CPU_16BIT_STACK_SIZE 1024
+
 #ifndef __ASSEMBLY__
 
 #include <linux/preempt.h>
@@ -15,6 +17,9 @@
 extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
 DECLARE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 
+extern unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DECLARE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+
 struct Xgt_desc_struct {
 	unsigned short size;
 	unsigned long address __attribute__((packed));
diff -urN linux-2.6.10-mm1/include/asm-i386/segment.h linux-2.6.10-mm1-stk/include/asm-i386/segment.h
--- linux-2.6.10-mm1/include/asm-i386/segment.h	2005-01-03 22:56:32.000000000 +0300
+++ linux-2.6.10-mm1-stk/include/asm-i386/segment.h	2005-01-03 23:13:21.000000000 +0300
@@ -38,7 +38,7 @@
  *  24 - APM BIOS support
  *  25 - APM BIOS support 
  *
- *  26 - unused
+ *  26 - ESPFIX small SS
  *  27 - unused
  *  28 - unused
  *  29 - unused
@@ -71,6 +71,9 @@
 #define GDT_ENTRY_PNPBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 6)
 #define GDT_ENTRY_APMBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 11)
 
+#define GDT_ENTRY_ESPFIX_SS		(GDT_ENTRY_KERNEL_BASE + 14)
+#define __ESPFIX_SS (GDT_ENTRY_ESPFIX_SS * 8)
+
 #define GDT_ENTRY_DOUBLEFAULT_TSS	31
 
 /*

Re: [patch] x86: fix ESP corruption CPU bug

[Linus Torvalds]

On Tue, 4 Jan 2005, Stas Sergeev wrote:
> 
> Can this please be applied?

Please don't do it like this - you made the patch now depend on the 
ugliest code in the universe, namely that horribly crappy kgdb-ga sh*t 
("Don't hold back, Linus, tell us how you really feel").

The 16-bit stack code may not be the prettiest either, but it doesn't hold 
a candle to the asm-crap that is entry.S after kgdb-ga.

"resume_kernelX"? What crud.

		Linus

Re: [patch] x86: fix ESP corruption CPU bug

[Stas Sergeev]

[-- Attachment #1: Type: text/plain, Size: 1075 bytes --]

Hi Linus.

Linus Torvalds wrote:
> Please don't do it like this - you made the patch now depend on the 
> ugliest code in the universe, namely that horribly crappy kgdb-ga sh*t 
I didn't do that. I just re-targeted the
patch to -mm tree and it clashed with the
kgdb-ga patch. And somehow it happened
that a few lines I would have to insert
myself otherwise anyway, appeared to be
already there for me to re-use. I wouldn't
call that a dependancy. Only 3 lines are
re-used in fact.

> The 16-bit stack code may not be the prettiest either, but it doesn't hold 
> a candle to the asm-crap that is entry.S after kgdb-ga.
>From what I can see, kgdb-ga have only
3 small hunks in entry.S, so the crap is
probably very dense there.
Clashing into one of these hunks looks
unavoidable for my needs.

> "resume_kernelX"? What crud.
Does "restore_nocheck" sound better?
Yes, maybe, but then I don't see the way
to provide my patch for -mm. So the attached
one is for plain 2.6.10. I don't know how
Andrew can apply it, so maybe you will?


Signed-off-by: Stas Sergeev <stsp@aknet.ru>


[-- Attachment #2: linux-2.6.10-stk5.diff --]
[-- Type: text/x-patch, Size: 11056 bytes --]

diff -urN linux-2.6.10/arch/i386/kernel/cpu/common.c linux-2.6.10-stk/arch/i386/kernel/cpu/common.c
--- linux-2.6.10/arch/i386/kernel/cpu/common.c	2004-10-21 21:35:21.000000000 +0400
+++ linux-2.6.10-stk/arch/i386/kernel/cpu/common.c	2004-12-31 11:16:28.000000000 +0300
@@ -16,6 +16,10 @@
 DEFINE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 EXPORT_PER_CPU_SYMBOL(cpu_gdt_table);
 
+unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DEFINE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+EXPORT_PER_CPU_SYMBOL(cpu_16bit_stack);
+
 static int cachesize_override __initdata = -1;
 static int disable_x86_fxsr __initdata = 0;
 static int disable_x86_serial_nr __initdata = 1;
@@ -508,6 +512,7 @@
 	int cpu = smp_processor_id();
 	struct tss_struct * t = &per_cpu(init_tss, cpu);
 	struct thread_struct *thread = &current->thread;
+	__u32 stk16_off = (__u32)&per_cpu(cpu_16bit_stack, cpu);
 
 	if (test_and_set_bit(cpu, &cpu_initialized)) {
 		printk(KERN_WARNING "CPU#%d already initialized!\n", cpu);
@@ -530,6 +535,13 @@
 	 */
 	memcpy(&per_cpu(cpu_gdt_table, cpu), cpu_gdt_table,
 	       GDT_SIZE);
+
+	/* Set up GDT entry for 16bit stack */
+	*(__u64 *)&(per_cpu(cpu_gdt_table, cpu)[GDT_ENTRY_ESPFIX_SS]) |=
+		((((__u64)stk16_off) << 16) & 0x000000ffffff0000ULL) |
+		((((__u64)stk16_off) << 32) & 0xff00000000000000ULL) |
+		(CPU_16BIT_STACK_SIZE - 1);
+
 	cpu_gdt_descr[cpu].size = GDT_SIZE - 1;
 	cpu_gdt_descr[cpu].address =
 	    (unsigned long)&per_cpu(cpu_gdt_table, cpu);
diff -urN linux-2.6.10/arch/i386/kernel/entry.S linux-2.6.10-stk/arch/i386/kernel/entry.S
--- linux-2.6.10/arch/i386/kernel/entry.S	2004-12-26 00:37:55.000000000 +0300
+++ linux-2.6.10-stk/arch/i386/kernel/entry.S	2004-12-31 11:30:18.000000000 +0300
@@ -47,6 +47,7 @@
 #include <asm/segment.h>
 #include <asm/smp.h>
 #include <asm/page.h>
+#include <asm/desc.h>
 #include "irq_vectors.h"
 
 #define nr_syscalls ((syscall_table_size)/4)
@@ -78,7 +79,7 @@
 #define preempt_stop		cli
 #else
 #define preempt_stop
-#define resume_kernel		restore_all
+#define resume_kernel		restore_nocheck
 #endif
 
 #define SAVE_ALL \
@@ -122,24 +123,6 @@
 .previous
 
 
-#define RESTORE_ALL	\
-	RESTORE_REGS	\
-	addl $4, %esp;	\
-1:	iret;		\
-.section .fixup,"ax";   \
-2:	sti;		\
-	movl $(__USER_DS), %edx; \
-	movl %edx, %ds; \
-	movl %edx, %es; \
-	movl $11,%eax;	\
-	call do_exit;	\
-.previous;		\
-.section __ex_table,"a";\
-	.align 4;	\
-	.long 1b,2b;	\
-.previous
-
-
 ENTRY(ret_from_fork)
 	pushl %eax
 	call schedule_tail
@@ -163,7 +146,7 @@
 	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
 	movb CS(%esp), %al
 	testl $(VM_MASK | 3), %eax
-	jz resume_kernel		# returning to kernel or vm86-space
+	jz resume_kernel
 ENTRY(resume_userspace)
  	cli				# make sure we don't miss an interrupt
 					# setting need_resched or sigpending
@@ -177,7 +160,7 @@
 #ifdef CONFIG_PREEMPT
 ENTRY(resume_kernel)
 	cmpl $0,TI_preempt_count(%ebp)	# non-zero preempt_count ?
-	jnz restore_all
+	jnz restore_nocheck
 need_resched:
 	movl TI_flags(%ebp), %ecx	# need_resched set ?
 	testb $_TIF_NEED_RESCHED, %cl
@@ -192,6 +175,31 @@
 	jmp need_resched
 #endif
 
+ldt_ss:
+	larl OLDSS(%esp), %eax
+	jnz restore_nocheck
+	testl $0x00400000, %eax		# returning to 32bit stack?
+	jnz restore_nocheck		# allright, normal return
+	/* If returning to userspace with 16bit stack,
+	 * try to fix the higher word of ESP, as the CPU
+	 * won't restore it.
+	 * This is an "official" bug of all the x86-compatible
+	 * CPUs, which we can try to work around to make
+	 * dosemu and wine happy. */
+	subl $8, %esp		# reserve space for switch16 pointer
+	cli
+	movl %esp, %eax
+	/* Set up the 16bit stack frame with switch32 pointer on top,
+	 * and a switch16 pointer on top of the current frame. */
+	call setup_x86_bogus_stack
+	RESTORE_REGS
+	lss 20+4(%esp), %esp	# The magic pointer above iret frame
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 /* SYSENTER_RETURN points to after the "sysenter" instruction in
    the vsyscall page.  See vsyscall-sysentry.S, which defines the symbol.  */
 
@@ -260,8 +268,32 @@
 	movl TI_flags(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx	# current->work
 	jne syscall_exit_work
+
 restore_all:
-	RESTORE_ALL
+	movl EFLAGS(%esp), %eax		# mix EFLAGS and CS
+	movb CS(%esp), %al
+	andl $(VM_MASK | 3), %eax
+	cmpl $3, %eax
+	jne restore_nocheck		# returning to kernel or vm86-space
+	testl $4, OLDSS(%esp)
+	jnz ldt_ss
+restore_nocheck:
+	RESTORE_REGS
+	addl $4, %esp
+1:	iret
+.section .fixup,"ax"
+iret_exc:
+	sti
+	movl $(__USER_DS), %edx
+	movl %edx, %ds
+	movl %edx, %es
+	movl $11,%eax
+	call do_exit
+.previous
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
 
 	# perform work that needs to be done immediately before resumption
 	ALIGN
@@ -337,6 +369,25 @@
 	movl $-ENOSYS,EAX(%esp)
 	jmp resume_userspace
 
+/* Check if we are on 16bit stack. Can happen either on iret of ESPFIX,
+ * or in an exception handler after that iret... */
+#define FIXUP_ESPFIX_STACK \
+	movl %esp, %eax; \
+	/* Magic pointer is at the top of the 16bit stack */ \
+	lss %ss:CPU_16BIT_STACK_SIZE-8, %esp; \
+	call fixup_x86_bogus_stack; \
+	movl %eax, %esp;
+#define UNWIND_ESPFIX_STACK \
+	pushl %eax; \
+	movl %ss, %eax; \
+	cmpw $__ESPFIX_SS, %ax; \
+	jne 28f; \
+	movl $(__KERNEL_DS), %edx; \
+	movl %edx, %ds; \
+	movl %edx, %es; \
+	FIXUP_ESPFIX_STACK \
+28:	popl %eax;
+
 /*
  * Build the entry stubs and pointer table with
  * some assembler magic.
@@ -391,7 +442,9 @@
 	pushl %ecx
 	pushl %ebx
 	cld
-	movl %es, %ecx
+	pushl %es
+	UNWIND_ESPFIX_STACK
+	popl %ecx
 	movl ES(%esp), %edi		# get the function address
 	movl ORIG_EAX(%esp), %edx	# get the error code
 	movl %eax, ORIG_EAX(%esp)
@@ -473,6 +526,11 @@
  * fault happened on the sysenter path.
  */
 ENTRY(nmi)
+	pushl %eax
+	movl %ss, %eax
+	cmpw $__ESPFIX_SS, %ax
+	popl %eax
+	je nmi_16bit_stack
 	cmpl $sysenter_entry,(%esp)
 	je nmi_stack_fixup
 	pushl %eax
@@ -492,7 +550,7 @@
 	xorl %edx,%edx		# zero error code
 	movl %esp,%eax		# pt_regs pointer
 	call do_nmi
-	RESTORE_ALL
+	jmp restore_all
 
 nmi_stack_fixup:
 	FIX_STACK(12,nmi_stack_correct, 1)
@@ -508,6 +566,29 @@
 	FIX_STACK(24,nmi_stack_correct, 1)
 	jmp nmi_stack_correct
 
+nmi_16bit_stack:
+	/* create the pointer to lss back */
+	pushl %ss
+	pushl %esp
+	andl $0xffff, %esp
+	addw $4, (%esp)
+	/* copy the iret frame of 12 bytes */
+	.rept 3
+	pushl 16(%esp)
+	.endr
+	pushl %eax
+	SAVE_ALL
+	FIXUP_ESPFIX_STACK		# %eax == %esp
+	xorl %edx,%edx			# zero error code
+	call do_nmi
+	RESTORE_REGS
+	lss 12+4(%esp), %esp		# back to 16bit stack
+1:	iret
+.section __ex_table,"a"
+	.align 4
+	.long 1b,iret_exc
+.previous
+
 ENTRY(int3)
 	pushl $-1			# mark this as an int
 	SAVE_ALL
diff -urN linux-2.6.10/arch/i386/kernel/head.S linux-2.6.10-stk/arch/i386/kernel/head.S
--- linux-2.6.10/arch/i386/kernel/head.S	2004-10-21 21:35:21.000000000 +0400
+++ linux-2.6.10-stk/arch/i386/kernel/head.S	2005-01-01 13:02:39.000000000 +0300
@@ -514,7 +514,7 @@
 	.quad 0x00009a0000000000	/* 0xc0 APM CS 16 code (16 bit) */
 	.quad 0x0040920000000000	/* 0xc8 APM DS    data */
 
-	.quad 0x0000000000000000	/* 0xd0 - unused */
+	.quad 0x0000920000000000	/* 0xd0 - ESPFIX 16-bit SS */
 	.quad 0x0000000000000000	/* 0xd8 - unused */
 	.quad 0x0000000000000000	/* 0xe0 - unused */
 	.quad 0x0000000000000000	/* 0xe8 - unused */
diff -urN linux-2.6.10/arch/i386/kernel/traps.c linux-2.6.10-stk/arch/i386/kernel/traps.c
--- linux-2.6.10/arch/i386/kernel/traps.c	2004-12-26 00:37:31.000000000 +0300
+++ linux-2.6.10-stk/arch/i386/kernel/traps.c	2004-12-31 11:27:02.000000000 +0300
@@ -904,6 +904,51 @@
 #endif
 }
 
+fastcall void setup_x86_bogus_stack(unsigned char * stk)
+{
+	unsigned long *switch16_ptr, *switch32_ptr;
+	struct pt_regs *regs;
+	unsigned long stack_top, stack_bot;
+	unsigned short iret_frame16_off;
+	int cpu = smp_processor_id();
+	/* reserve the space on 32bit stack for the magic switch16 pointer */
+	memmove(stk, stk + 8, sizeof(struct pt_regs));
+	switch16_ptr = (unsigned long *)(stk + sizeof(struct pt_regs));
+	regs = (struct pt_regs *)stk;
+	/* now the switch32 on 16bit stack */
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	iret_frame16_off = CPU_16BIT_STACK_SIZE - 8 - 20;
+	/* copy iret frame on 16bit stack */
+	memcpy((void *)(stack_bot + iret_frame16_off), &regs->eip, 20);
+	/* fill in the switch pointers */
+	switch16_ptr[0] = (regs->esp & 0xffff0000) | iret_frame16_off;
+	switch16_ptr[1] = __ESPFIX_SS;
+	switch32_ptr[0] = (unsigned long)stk + sizeof(struct pt_regs) +
+		8 - CPU_16BIT_STACK_SIZE;
+	switch32_ptr[1] = __KERNEL_DS;
+}
+
+fastcall unsigned char * fixup_x86_bogus_stack(unsigned short sp)
+{
+	unsigned long *switch32_ptr;
+	unsigned char *stack16, *stack32;
+	unsigned long stack_top, stack_bot;
+	int len;
+	int cpu = smp_processor_id();
+	stack_bot = (unsigned long)&per_cpu(cpu_16bit_stack, cpu);
+	stack_top = stack_bot +	CPU_16BIT_STACK_SIZE;
+	switch32_ptr = (unsigned long *)(stack_top - 8);
+	/* copy the data from 16bit stack to 32bit stack */
+	len = CPU_16BIT_STACK_SIZE - 8 - sp;
+	stack16 = (unsigned char *)(stack_bot + sp);
+	stack32 = (unsigned char *)
+		(switch32_ptr[0] + CPU_16BIT_STACK_SIZE - 8 - len);
+	memcpy(stack32, stack16, len);
+	return stack32;
+}
+
 /*
  *  'math_state_restore()' saves the current math information in the
  * old math state array, and gets the new ones from the current task
diff -urN linux-2.6.10/include/asm-i386/desc.h linux-2.6.10-stk/include/asm-i386/desc.h
--- linux-2.6.10/include/asm-i386/desc.h	2004-10-21 21:35:55.000000000 +0400
+++ linux-2.6.10-stk/include/asm-i386/desc.h	2005-01-01 13:02:39.000000000 +0300
@@ -4,6 +4,8 @@
 #include <asm/ldt.h>
 #include <asm/segment.h>
 
+#define CPU_16BIT_STACK_SIZE 1024
+
 #ifndef __ASSEMBLY__
 
 #include <linux/preempt.h>
@@ -15,6 +17,9 @@
 extern struct desc_struct cpu_gdt_table[GDT_ENTRIES];
 DECLARE_PER_CPU(struct desc_struct, cpu_gdt_table[GDT_ENTRIES]);
 
+extern unsigned char cpu_16bit_stack[CPU_16BIT_STACK_SIZE];
+DECLARE_PER_CPU(unsigned char, cpu_16bit_stack[CPU_16BIT_STACK_SIZE]);
+
 struct Xgt_desc_struct {
 	unsigned short size;
 	unsigned long address __attribute__((packed));
diff -urN linux-2.6.10/include/asm-i386/segment.h linux-2.6.10-stk/include/asm-i386/segment.h
--- linux-2.6.10/include/asm-i386/segment.h	2004-01-09 09:59:19.000000000 +0300
+++ linux-2.6.10-stk/include/asm-i386/segment.h	2005-01-01 13:02:39.000000000 +0300
@@ -38,7 +38,7 @@
  *  24 - APM BIOS support
  *  25 - APM BIOS support 
  *
- *  26 - unused
+ *  26 - ESPFIX small SS
  *  27 - unused
  *  28 - unused
  *  29 - unused
@@ -71,6 +71,9 @@
 #define GDT_ENTRY_PNPBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 6)
 #define GDT_ENTRY_APMBIOS_BASE		(GDT_ENTRY_KERNEL_BASE + 11)
 
+#define GDT_ENTRY_ESPFIX_SS		(GDT_ENTRY_KERNEL_BASE + 14)
+#define __ESPFIX_SS (GDT_ENTRY_ESPFIX_SS * 8)
+
 #define GDT_ENTRY_DOUBLEFAULT_TSS	31
 
 /*