Skip to contentSkip to conversation

small adjustment to asm constraints for c/s 19400

PATCH In review Jan Beulich 4 participants 19 messages Mar 27, 2009 – Mar 31, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 small adjustment to asm constraints for c/s 19400

Properly indicate the variable that the asm writes to, and allow compiler more freedom in selecting a suitable input operand for the ltr instruction.

xen/arch/x86/traps.c +2 -2
@@ -3048,8 +3048,8 @@ void load_TR(void)
 
     /* Switch to non-compat GDT (which has B bit clear) to execute LTR. */
     asm volatile (
-        "sgdt %1; lgdt %2; ltr %%ax; lgdt %1"
-        : : "a" (TSS_ENTRY << 3), "m" (old_gdt), "m" (tss_gdt) : "memory" );
+        "sgdt %0; lgdt %2; ltr %w1; lgdt %0"
+        : "=m" (old_gdt) : "rm" (TSS_ENTRY << 3), "m" (tss_gdt) : "memory" );
 }
 
 void __devinit percpu_traps_init(void)
#1 lore raw

Hi Jan,

Quoted 19 lines · Jan Beulich

Properly indicate the variable that the asm writes to, and allow compiler more freedom in selecting a suitable input operand for the ltr instruction.

--- 2009-03-27.orig/xen/arch/x86/traps.c 2009-03-24
09:04:02.000000000 +0100 +++
2009-03-27/xen/arch/x86/traps.c 2009-03-27 10:10:57.000000000 +0100

(quoted diff) +0 -2
@@ -3048,8 +3048,8 @@ void load_TR(void) 
 
     /* Switch to non-compat GDT (which has B bit clear) to execute
     LTR. */ asm volatile (
-        "sgdt %1; lgdt %2; ltr %%ax; lgdt %1"
-        : : "a" (TSS_ENTRY << 3), "m" (old_gdt), "m" (tss_gdt) :

"memory" ); + "sgdt %0; lgdt %2; ltr %w1; lgdt %0"
+ : "=m" (old_gdt) : "rm" (TSS_ENTRY << 3), "m" (tss_gdt) :
"memory" ); }

Quoted 2 lines

void __devinit percpu_traps_init(void)

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

Thanks.

Quoted 7 lines

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Signature · 1 line
Guanqun
#2 lore raw

RE: [PATCH] small adjustment to asm constraints for c/s19400

Quoted 3 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 30.03.09 06:45 >>>

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

And it's still failing at the ltr? Then no, I don't have any clue without additional data. I'd say try dumping out relevant variables and descriptors.

Jan

#3 lore raw

RE: [PATCH] small adjustment to asm constraints for c/s19400

Quoted 9 lines · Jan Beulich
Quoted 3 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 30.03.09 06:45 >>>

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

And it's still failing at the ltr? Then no, I don't have any clue without additional data. I'd say try dumping out relevant variables and descriptors.

Jan

it doesn't fail at ltr. please see my previous mail, thanks.

Signature · 1 line
Guanqun
#4 lore raw
In reply to Lu, Guanqun
Quoted 3 lines · Guanqun

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0.
Do you have any idea what will cause this problem?

Does it still fail at the LTR? Does the alternative fix of settign the type to 9 rather than 11 work?

-- Keir

#5 lore raw
Quoted 9 lines · Keir Fraser
Quoted 2 lines · Guanqun

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

Does it still fail at the LTR? Does the alternative fix of settign the type to 9 rather than 11 work?

-- Keir

It doesn't fail at the LTR.
Xen resumes back, but dom0 hangs there(technically speaking, it doen't hang, eip changes, but can't come back to shell console), I used xen serial line to dump the dom0 registers,
the eip it runs to from time to time is 'general_protection' in dom0.

Any idea?

Fixing type to 9 doesn't really solves the problem, it has the same phenomenon.
I'll have a double check, and try to dump more registers and variable according to Jan's advice and see if something un-expected happens.

Thanks.

Signature · 1 line
Guanqun
#6 lore raw

RE: [PATCH] small adjustment to asm constraints for c/s19400

Quoted 11 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 30.03.09 10:56 >>>

Quoted 8 lines · Keir Fraser
Quoted 2 lines · Guanqun

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

Does it still fail at the LTR? Does the alternative fix of settign the type to 9 rather than 11 work?

Quoted 6 lines
Quoted 1 line

-- Keir

It doesn't fail at the LTR.
Xen resumes back, but dom0 hangs there(technically speaking, it doen't hang, eip changes, but can't come back to shell console), I used xen serial line to dump the dom0 registers,
the eip it runs to from time to time is 'general_protection' in dom0.

Since under normal circumstances GP faults are rare, why don't you just print out the faulting vCPU's state from Xen at the point where the fault gets forwarded to the guest? That'll tell you exactly what the guest is dying on.

Jan

#7 lore raw

RE: [PATCH] small adjustment to asm constraints for c/s19400

Quoted 19 lines · Jan Beulich
Quoted 11 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 30.03.09 10:56 >>>

Quoted 8 lines · Keir Fraser
Quoted 2 lines · Guanqun

With this adjustment or previous patch 19400, S3 still fails on 64 xen / 32 dom0. Do you have any idea what will cause this problem?

Does it still fail at the LTR? Does the alternative fix of settign the type to 9 rather than 11 work?

Quoted 7 lines
Quoted 1 line

-- Keir

It doesn't fail at the LTR.
Xen resumes back, but dom0 hangs there(technically speaking, it doen't hang, eip changes, but can't come back to shell console), I used xen serial line to dump the dom0 registers,
the eip it runs to from time to time is 'general_protection' in dom0.

Quoted 6 lines

Since under normal circumstances GP faults are rare, why don't you just print out the faulting vCPU's state from Xen at the point where the fault gets forwarded to the guest? That'll tell you exactly what the guest is dying on.

Jan

Hi Jan,

The point when S3 resumes back, it's running in compat mode.
But the code doesn't differentiate it, as in xen/arch/x86/acpi/wakeup_prot.S. Maybe we lose some context during the switches which may causes the problem.

When it resumes back, hitting '0' on xen serial console gives us:

(XEN) '0' pressed -> dumping Dom0's registers
(XEN) *** Dumping Dom0 vcpu#0 state: ***
(XEN) ----[ Xen-3.4-unstable x86_64 debug=y Not tainted ]----
(XEN) CPU: 0
(XEN) RIP: 0061:[<00000000c02a9208>]
(XEN) RFLAGS: 0000000000200246 EM: 0 CONTEXT: pv guest

(XEN) rax: 00000000000000af   rbx: 0000000000000000   rcx: 00000000bfe7d8e4
(XEN) rdx: 00000000bfe7d864   rsi: 0000000000000008   rdi: 0000000000d3cff4
(XEN) rbp: 00000000bfe7d790   rsp: 00000000c6aa3fe0   r8:  0000000000000000
(XEN) r9:  0000000000000000   r10: 0000000000000000   r11: 0000000000000000
(XEN) r12: 0000000000000000   r13: 0000000000000000   r14: 0000000000000000
(XEN) r15: 0000000000000000   cr0: 000000008005003b   cr4: 00000000000026f0
(XEN) cr3: 00000000bd6e4000   cr2: 00000000b7fbe000
(XEN) ds: 007b   es: 007b   fs: 0000   gs: 0033   ss: 0069   cs: 0061
(XEN) Guest stack trace from esp=c6aa3fe0:
(XEN)   00000000 b7f9d405 00000073 00210246 bfe7d790 0000007b 00000000 00000000
(XEN)   c6aa1080 c02fd6a0 00000000 00000000 00000000 00000000 ffffffff 00000000
(XEN)   c0126a47 00000000 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   ffffffff ffffffbf ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffffffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   ffffffff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffbfffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   fffffeff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffffffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   ffffffff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffffffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   ffffffff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   bfffffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   fffbffff ffffffff ffefffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ff7fffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   ffffffff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffffffff ffffffff 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   fff7dfff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff
(XEN)   ffffffff fffffff7 00000000 00000000 00000000 00000000 00000000 00000000
(XEN)   bfffffff ffffffff ffffffff ffffffff 00000000 00000000 ffffffff ffffffff

Comparing the eip with the dom0 symbol table,
0xc02a9208 -> general_protection
0xc02fd6a0 -> default_exec_domain
0xc0126a47 -> do_no_restart_syscall

Quoted 1 line

From the stack, such invocation chain is observed.

Any idea what's happening there?

And I also dump the MSR just before and after S3, there're no difference. patch:

xen/arch/x86/acpi/power.c Tue Mar 31 12:26:12 2009 +0800 renamed from xen/arch/x86/acpi/power.c Mon Mar 30 16:48:26 2009 +0100 +24 -0
diff -r d5ddc782bc49 xen/arch/x86/acpi/power.c
@@ -42,6 +42,28 @@ struct acpi_sleep_info acpi_sinfo;
 
 void do_suspend_lowlevel(void);
 
+static void dump_msr_registers(void)
+{
+    unsigned long msr;
+
+    rdmsrl(MSR_EFER, msr);
+    printk("<0> EFER: %lx", msr);
+    rdmsrl(MSR_STAR, msr);
+    printk("<0> STAR: %lx", msr);
+    rdmsrl(MSR_LSTAR, msr);
+    printk("<0> LSTAR: %lx", msr);
+    rdmsrl(MSR_CSTAR, msr);
+    printk("<0> CSTAR: %lx", msr);
+    rdmsrl(MSR_SYSCALL_MASK, msr);
+    printk("<0> SYSCALL_MASK: %lx", msr);
+    rdmsrl(MSR_FS_BASE, msr);
+    printk("<0> FS_BASE: %lx", msr);
+    rdmsrl(MSR_GS_BASE, msr);
+    printk("<0> GS_BASE: %lx", msr);
+    rdmsrl(MSR_SHADOW_GS_BASE, msr);
+    printk("<0> SH_GS_BASE: %lx", msr);
+}
+
 static int device_power_down(void)
 {
     console_suspend();
@@ -173,6 +195,7 @@ static int enter_state(u32 state)
 
     console_start_sync();
     printk("Entering ACPI S%d state.\n", state);
+    dump_msr_registers();
 
     local_irq_save(flags);
     spin_debug_disable();
@@ -208,6 +231,7 @@ static int enter_state(u32 state)
     device_power_up();
 
     printk(XENLOG_INFO "Finishing wakeup from ACPI S%d state.\n", state);
+    dump_msr_registers();
 
     if ( (state == ACPI_STATE_S3) && error )
         panic("Memory integrity was lost on resume (%d)\n", error);

the result is:
[before] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR: ffff828c8029b000<0> CSTAR: ffff828c8029b080<0> SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0 [after] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR: ffff828c8029b000<0> CSTAR: ffff828c8029b080<0> SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0

They're the same.

And aslo I dump GDT, IDT, LDT and TR register and tss struct, there are no difference seen.

So can you give some advice some more registers should be dumped?
Or maybe our S3 code doesn't take the 32dom0 into consideration from the beginning, is there anyway to fix this?

Thanks!

Signature · 1 line
Guanqun
#8 lore raw

RE: [PATCH] small adjustment to asm constraints for c/s19400

Quoted 9 lines

From: Lu, Guanqun
Sent: 2009年3月31日 16:55
Hi Jan,

The point when S3 resumes back, it's running in compat mode. But the code doesn't differentiate it, as in xen/arch/x86/acpi/wakeup_prot.S.
Maybe we lose some context during the switches which may causes the problem.

One correction here. pCPU is of course in long mode when resuming from S3 since it happens in Xen. It's current vcpu as a 32bit vcpu with compatibility context (32bit dom0). There may be some registers only used in compatibility mode, but kept even when Xen is running in long mode. However current low level context save/restore across S3 suspend/resume, as you can see, have nothing specific for whether current context is 32bit vcpu. So we're not sure whether some registers content specific for compatibility mode get lost (cpu is reset in S3) and not be recovered by current resume code.

Jan, since you're the No.1 expert in this area, we'd appreciate your possible suggestions here. :-)

Thanks
Kevin

Quoted 141 lines

When it resumes back, hitting '0' on xen serial console gives us:

(XEN) '0' pressed -> dumping Dom0's registers
(XEN) *** Dumping Dom0 vcpu#0 state: ***
(XEN) ----[ Xen-3.4-unstable x86_64 debug=y Not tainted ]----
(XEN) CPU: 0
(XEN) RIP: 0061:[<00000000c02a9208>]
(XEN) RFLAGS: 0000000000200246 EM: 0 CONTEXT: pv guest
(XEN) rax: 00000000000000af rbx: 0000000000000000 rcx:
00000000bfe7d8e4
(XEN) rdx: 00000000bfe7d864 rsi: 0000000000000008 rdi:
0000000000d3cff4
(XEN) rbp: 00000000bfe7d790 rsp: 00000000c6aa3fe0 r8:
0000000000000000
(XEN) r9: 0000000000000000 r10: 0000000000000000 r11:
0000000000000000
(XEN) r12: 0000000000000000 r13: 0000000000000000 r14:
0000000000000000
(XEN) r15: 0000000000000000 cr0: 000000008005003b cr4:
00000000000026f0
(XEN) cr3: 00000000bd6e4000 cr2: 00000000b7fbe000
(XEN) ds: 007b es: 007b fs: 0000 gs: 0033 ss: 0069 cs: 0061
(XEN) Guest stack trace from esp=c6aa3fe0:
(XEN) 00000000 b7f9d405 00000073 00210246 bfe7d790 0000007b
00000000 00000000
(XEN) c6aa1080 c02fd6a0 00000000 00000000 00000000 00000000
ffffffff 00000000
(XEN) c0126a47 00000000 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) ffffffff ffffffbf ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffffffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) ffffffff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffbfffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) fffffeff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffffffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) ffffffff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffffffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) ffffffff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) bfffffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) fffbffff ffffffff ffefffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ff7fffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) ffffffff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffffffff ffffffff 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) fff7dfff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff
(XEN) ffffffff fffffff7 00000000 00000000 00000000 00000000
00000000 00000000
(XEN) bfffffff ffffffff ffffffff ffffffff 00000000 00000000
ffffffff ffffffff

Comparing the eip with the dom0 symbol table,
0xc02a9208 -> general_protection
0xc02fd6a0 -> default_exec_domain
0xc0126a47 -> do_no_restart_syscall

Quoted 1 line

From the stack, such invocation chain is observed.

Any idea what's happening there?

And I also dump the MSR just before and after S3, there're no difference.
patch:

diff -r d5ddc782bc49 xen/arch/x86/acpi/power.c

xen/arch/x86/acpi/power.c Tue Mar 31 12:26:12 2009 +0800 renamed from xen/arch/x86/acpi/power.c Mon Mar 30 16:48:26 2009 +0100 +24 -0
@@ -42,6 +42,28 @@ struct acpi_sleep_info acpi_sinfo;
 
 void do_suspend_lowlevel(void);
 
+static void dump_msr_registers(void)
+{
+    unsigned long msr;
+
+    rdmsrl(MSR_EFER, msr);
+    printk("<0> EFER: %lx", msr);
+    rdmsrl(MSR_STAR, msr);
+    printk("<0> STAR: %lx", msr);
+    rdmsrl(MSR_LSTAR, msr);
+    printk("<0> LSTAR: %lx", msr);
+    rdmsrl(MSR_CSTAR, msr);
+    printk("<0> CSTAR: %lx", msr);
+    rdmsrl(MSR_SYSCALL_MASK, msr);
+    printk("<0> SYSCALL_MASK: %lx", msr);
+    rdmsrl(MSR_FS_BASE, msr);
+    printk("<0> FS_BASE: %lx", msr);
+    rdmsrl(MSR_GS_BASE, msr);
+    printk("<0> GS_BASE: %lx", msr);
+    rdmsrl(MSR_SHADOW_GS_BASE, msr);
+    printk("<0> SH_GS_BASE: %lx", msr);
+}
+
 static int device_power_down(void)
 {
    console_suspend();
@@ -173,6 +195,7 @@ static int enter_state(u32 state)
 
    console_start_sync();
    printk("Entering ACPI S%d state.\n", state);
+    dump_msr_registers();
 
    local_irq_save(flags);
    spin_debug_disable();
@@ -208,6 +231,7 @@ static int enter_state(u32 state)
    device_power_up();
 
    printk(XENLOG_INFO "Finishing wakeup from ACPI S%d 
 state.\n", state);
+    dump_msr_registers();
 
    if ( (state == ACPI_STATE_S3) && error )
        panic("Memory integrity was lost on resume (%d)\n", error);

the result is:
[before] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:
ffff828c8029b000<0> CSTAR: ffff828c8029b080<0> SYSCALL_MASK:
34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0 [after] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:
ffff828c8029b000<0> CSTAR: ffff828c8029b080<0> SYSCALL_MASK:
34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0

They're the same.

And aslo I dump GDT, IDT, LDT and TR register and tss struct, there are no difference seen.

Quoted 15 lines

So can you give some advice some more registers should be dumped? Or maybe our S3 code doesn't take the 32dom0 into consideration from the beginning, is there anyway to fix this?

Thanks!

Signature · 5 lines
Guanqun
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#9 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

In reply to Lu, Guanqun
Quoted 6 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 31.03.09 10:54 >>>

[before] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR: ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>
SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0
[after] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR: ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>
SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0> SH_GS_BASE: 0

Quoted 6 lines

They're the same.

And aslo I dump GDT, IDT, LDT and TR register and tss struct, there are no difference seen.

So can you give some advice some more registers should be dumped?

Assuming this is an Intel box, the guest is certainly using sysenter (which the user mode EIP saved on the stack would also suggest), so you're really after the three SYSENTER MSRs. And indeed - restore_rest_processor_state() seems to only care for the STAR ones.

Jan

#10 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

Quoted 19 lines

From: Jan Beulich
Sent: 2009年3月31日 17:24

Quoted 2 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 31.03.09 10:54 >>>

[before] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:

ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>

Quoted 1 line

SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0>

SH_GS_BASE: 0

Quoted 1 line

[after] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:

ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>

Quoted 1 line

SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0>

SH_GS_BASE: 0

Quoted 4 lines

They're the same.

And aslo I dump GDT, IDT, LDT and TR register and tss struct,

there are no difference seen.

Quoted 2 lines
Quoted 8 lines
Quoted 1 line

So can you give some advice some more registers should be dumped?

Assuming this is an Intel box, the guest is certainly using sysenter (which the user mode EIP saved on the stack would also suggest), so you're really after the three SYSENTER MSRs. And indeed - restore_rest_processor_state() seems to only care for the STAR ones.

There are also some set of registers save/restored in assembly code in wakeup_prot.S, such as segment selectors, CRs, etc.

Thanks
Kevin

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#11 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

Quoted 11 lines
Quoted 8 lines
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 11:26 >>>

From: Jan Beulich
Assuming this is an Intel box, the guest is certainly using sysenter (which the user mode EIP saved on the stack would also suggest), so you're really after the three SYSENTER MSRs. And indeed - restore_rest_processor_state() seems to only care for the STAR ones.

There are also some set of registers save/restored in assembly code in wakeup_prot.S, such as segment selectors, CRs, etc.

Sure, but just grep-ing for MSR_IA32_SYSENTER across all of xen/ shows that the suspend/resume paths don't seem to consider them at all.

Jan

#12 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

In reply to Jan Beulich
Quoted 19 lines

From: Jan Beulich
Sent: 2009年3月31日 17:24

Quoted 2 lines
Quoted 1 line
Quoted 1 line

"Lu, Guanqun" <[email protected]> 31.03.09 10:54 >>>

[before] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:

ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>

Quoted 1 line

SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0>

SH_GS_BASE: 0

Quoted 1 line

[after] (XEN) EFER: d01<0> STAR: e023e00800000000<0> LSTAR:

ffff828c8029b000<0> CSTAR: ffff828c8029b080<0>

Quoted 1 line

SYSCALL_MASK: 34700<0> FS_BASE: 0<0> GS_BASE: b7f4a6c0<0>

SH_GS_BASE: 0

Quoted 4 lines

They're the same.

And aslo I dump GDT, IDT, LDT and TR register and tss struct,

there are no difference seen.

Quoted 2 lines
Quoted 8 lines
Quoted 1 line

So can you give some advice some more registers should be dumped?

Assuming this is an Intel box, the guest is certainly using sysenter (which the user mode EIP saved on the stack would also suggest), so you're really after the three SYSENTER MSRs. And indeed - restore_rest_processor_state() seems to only care for the STAR ones.

Well, got your point. Those three SYSENTER MSRs are lost after resume. We'll verify whether this is the only missing parts. BTW, in our box 32-on -32 S3 works fine, does that mean SYSENTER/SYSEXIT not used in pure 32bit pv environment, since in Xen 32bit path also misses that important recovery?

Thanks
Kevin

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#13 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

Quoted 6 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 11:47 >>>

Well, got your point. Those three SYSENTER MSRs are lost after resume. We'll verify whether this is the only missing parts. BTW, in our box 32-on -32 S3 works fine, does that mean SYSENTER/SYSEXIT not used in pure 32bit pv environment, since in Xen 32bit path also misses that important recovery?

A 32-bit hypervisor does not expose sysenter to the guest (the feature bit gets cleared) - iirc this is to prevent the guest from thinking that using sysenter would provide a performance benefit. Hence restoring these MSRs in that case would only be needed if Xen itself made use of sysenter, which however it doesn't (except for the now only theoretical case of the supervisor-mode-kernel feature).

Jan

#14 lore raw

RE: [PATCH] small adjustment to asm constraints forc/s19400

Quoted 12 lines

From: Jan Beulich [mailto:[email protected]]
Sent: 2009年3月31日 18:15

Quoted 3 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 11:47 >>>

Well, got your point. Those three SYSENTER MSRs are lost after resume. We'll verify whether this is the only missing parts. BTW, in

our box 32-on

Quoted 2 lines

-32 S3 works fine, does that mean SYSENTER/SYSEXIT not used in pure 32bit pv environment, since in Xen 32bit path also misses

that important

Quoted 1 line

recovery?

Quoted 11 lines

A 32-bit hypervisor does not expose sysenter to the guest (the feature bit
gets cleared) - iirc this is to prevent the guest from thinking that using
sysenter would provide a performance benefit. Hence restoring these MSRs
in that case would only be needed if Xen itself made use of sysenter, which
however it doesn't (except for the now only theoretical case of the supervisor-mode-kernel feature).

Clear enough. BTW, another question comes when I'm clearing the code. in load_segments, fs/gs base MSRs are only updated for non 32-on-64 domain. I guess that it's just for performance reason as 32bit guest doesn't require to switch gs base. I want to get your confirmation since in current low level context S3 resume code, we always restore those base MSRs even for 32-on-64 case.

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#15 lore raw

RE: [PATCH] small adjustment to asm constraintsforc/s19400

Quoted 7 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 12:21 >>>

BTW, another question comes when I'm clearing the code. in load_segments, fs/gs base MSRs are only updated for non 32-on-64 domain. I guess that it's just for performance reason as 32bit guest doesn't require to switch gs base. I want to get your confirmation since in current low level context S3 resume code, we always restore those base MSRs even for 32-on-64 case.

Correct - your code isn't performance sensitive, so keeping it simple is probably preferable. load_segments() sits in the context switch code, and thus it tries to avoid things not needed for a particular guest.

Jan

#16 lore raw

RE: [PATCH] small adjustment to asm constraintsforc/s19400

Quoted 11 lines

From: Jan Beulich [mailto:[email protected]]
Sent: 2009年3月31日 18:55

Quoted 7 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 12:21 >>>

BTW, another question comes when I'm clearing the code. in load_segments, fs/gs base MSRs are only updated for non 32-on-64 domain. I guess that it's just for performance reason as 32bit guest doesn't require to switch gs base. I want to get your confirmation since in current low level context S3 resume code, we always restore those base MSRs even for 32-on-64 case.

Quoted 6 lines

Correct - your code isn't performance sensitive, so keeping it simple is
probably preferable. load_segments() sits in the context switch code, and
thus it tries to avoid things not needed for a particular guest.

Thanks. After recovering those 3 SYSENTER MSRs, now 32-on-64 S3 could succeeds several times which is a big step. However there seems to be some other instability issues to cause machine occasionally reboot after more S3 actions. Both screen and serial port is dark and thus no more info is available. We'll continue to dig out the exact reason. But this shouldn't prevent sending out this reasonable fix anyway.

Thanks,
Kevin

-----

Enable S3 for 32bit dom0 on 64bit Xen

Three SYSENTER MSRs should be taken care of at save/restore BSP context, or else 32bit dom0 rejects working after S3 resume. Thanks for Jan's help to find this missing part.

Signed-off-by Guanqun Lu <[email protected]>
Signed-off-by Kevin Tian <[email protected]>

xen/arch/x86/acpi/suspend.c +8 -0
diff -r e152892a0b1e xen/arch/x86/acpi/suspend.c
@@ -16,6 +16,7 @@
 
 #if defined(CONFIG_X86_64)
 static unsigned long saved_lstar, saved_cstar;
+static unsigned long saved_sysenter_esp, saved_sysenter_eip;
 #endif
 
 void save_rest_processor_state(void)
@@ -26,6 +27,8 @@
 #if defined(CONFIG_X86_64)
     rdmsrl(MSR_CSTAR, saved_cstar);
     rdmsrl(MSR_LSTAR, saved_lstar);
+    rdmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+    rdmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
 #endif
 }
 
@@ -41,6 +44,11 @@
     wrmsrl(MSR_CSTAR, saved_cstar);
     wrmsr(MSR_STAR, 0, (FLAT_RING3_CS32<<16) | __HYPERVISOR_CS);
     wrmsr(MSR_SYSCALL_MASK, EF_VM|EF_RF|EF_NT|EF_DF|EF_IE|EF_TF, 0U);    
+
+    /* Recover sysenter MSRs */
+    wrmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+    wrmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+    wrmsr(MSR_IA32_SYSENTER_CS, __HYPERVISOR_CS, 0);
 #else /* !defined(CONFIG_X86_64) */
     if ( supervisor_mode_kernel && cpu_has_sep )
         wrmsr(MSR_IA32_SYSENTER_ESP, &init_tss[smp_processor_id()].esp1, 0);
xen/arch/x86/acpi/suspend.c +8 -0
diff -r e152892a0b1e xen/arch/x86/acpi/suspend.c
@@ -16,6 +16,7 @@
 
 #if defined(CONFIG_X86_64)
 static unsigned long saved_lstar, saved_cstar;
+static unsigned long saved_sysenter_esp, saved_sysenter_eip;
 #endif
 
 void save_rest_processor_state(void)
@@ -26,6 +27,8 @@
 #if defined(CONFIG_X86_64)
     rdmsrl(MSR_CSTAR, saved_cstar);
     rdmsrl(MSR_LSTAR, saved_lstar);
+    rdmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+    rdmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
 #endif
 }
 
@@ -41,6 +44,11 @@
     wrmsrl(MSR_CSTAR, saved_cstar);
     wrmsr(MSR_STAR, 0, (FLAT_RING3_CS32<<16) | __HYPERVISOR_CS);
     wrmsr(MSR_SYSCALL_MASK, EF_VM|EF_RF|EF_NT|EF_DF|EF_IE|EF_TF, 0U);    
+
+    /* Recover sysenter MSRs */
+    wrmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+    wrmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+    wrmsr(MSR_IA32_SYSENTER_CS, __HYPERVISOR_CS, 0);
 #else /* !defined(CONFIG_X86_64) */
     if ( supervisor_mode_kernel && cpu_has_sep )
         wrmsr(MSR_IA32_SYSENTER_ESP, &init_tss[smp_processor_id()].esp1, 0);

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#17 lore raw

RE: [PATCH] small adjustment to asm constraintsforc/s19400

Quoted 8 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 13:09 >>>

Thanks. After recovering those 3 SYSENTER MSRs, now 32-on-64 S3 could succeeds several times which is a big step. However there seems to be some other instability issues to cause machine occasionally reboot after more S3 actions. Both screen and serial port is dark and thus no more info is available. We'll continue to dig out the exact reason. But this shouldn't prevent sending out this reasonable fix anyway.

I think you need to be a little more careful here - subarch_percpu_traps_init() writes them only on Intel CPUs, and I think you ought to follow this for both saving and restoring them.

Jan

#18 lore raw

RE: [PATCH] small adjustment to asm constraintsforc/s19400

Quoted 12 lines

From: Jan Beulich [mailto:[email protected]]
Sent: 2009年3月31日 19:30

Quoted 6 lines
Quoted 1 line
Quoted 1 line

"Tian, Kevin" <[email protected]> 31.03.09 13:09 >>>

Thanks. After recovering those 3 SYSENTER MSRs, now 32-on-64 S3 could succeeds several times which is a big step. However there seems to be some other instability issues to cause machine occasionally reboot after more S3 actions. Both screen and serial port is dark and thus no more info is available. We'll

continue to dig out

Quoted 9 lines
Quoted 2 lines

the exact reason. But this shouldn't prevent sending out this reasonable fix anyway.

I think you need to be a little more careful here - subarch_percpu_traps_init()
writes them only on Intel CPUs, and I think you ought to follow this for
both saving and restoring them.

Ah, yes, I need add same condition check here. Below is updated one:

-----

Enable S3 for 32bit dom0 on 64bit Xen

Three SYSENTER MSRs should be taken care of at save/restore BSP context, or else 32bit dom0 rejects working after S3 resume. Thanks for Jan's help to find this missing part.

Signed-off-by Guanqun Lu <[email protected]>
Signed-off-by Kevin Tian <[email protected]>

xen/arch/x86/acpi/suspend.c +14 -0
diff -r e152892a0b1e xen/arch/x86/acpi/suspend.c
@@ -16,6 +16,7 @@
 
 #if defined(CONFIG_X86_64)
 static unsigned long saved_lstar, saved_cstar;
+static unsigned long saved_sysenter_esp, saved_sysenter_eip;
 #endif
 
 void save_rest_processor_state(void)
@@ -26,6 +27,11 @@
 #if defined(CONFIG_X86_64)
     rdmsrl(MSR_CSTAR, saved_cstar);
     rdmsrl(MSR_LSTAR, saved_lstar);
+    if ( boot_cpu_data.x86_vendor == X86_VENDOR_INTEL )
+    {
+        rdmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+        rdmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+    }
 #endif
 }
 
@@ -41,6 +47,14 @@
     wrmsrl(MSR_CSTAR, saved_cstar);
     wrmsr(MSR_STAR, 0, (FLAT_RING3_CS32<<16) | __HYPERVISOR_CS);
     wrmsr(MSR_SYSCALL_MASK, EF_VM|EF_RF|EF_NT|EF_DF|EF_IE|EF_TF, 0U);    
+
+    if ( boot_cpu_data.x86_vendor == X86_VENDOR_INTEL )
+    {
+        /* Recover sysenter MSRs */
+        wrmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+        wrmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+        wrmsr(MSR_IA32_SYSENTER_CS, __HYPERVISOR_CS, 0);
+    }
 #else /* !defined(CONFIG_X86_64) */
     if ( supervisor_mode_kernel && cpu_has_sep )
         wrmsr(MSR_IA32_SYSENTER_ESP, &init_tss[smp_processor_id()].esp1, 0);
xen/arch/x86/acpi/suspend.c +14 -0
diff -r e152892a0b1e xen/arch/x86/acpi/suspend.c
@@ -16,6 +16,7 @@
 
 #if defined(CONFIG_X86_64)
 static unsigned long saved_lstar, saved_cstar;
+static unsigned long saved_sysenter_esp, saved_sysenter_eip;
 #endif
 
 void save_rest_processor_state(void)
@@ -26,6 +27,11 @@
 #if defined(CONFIG_X86_64)
     rdmsrl(MSR_CSTAR, saved_cstar);
     rdmsrl(MSR_LSTAR, saved_lstar);
+    if ( boot_cpu_data.x86_vendor == X86_VENDOR_INTEL )
+    {
+        rdmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+        rdmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+    }
 #endif
 }
 
@@ -41,6 +47,14 @@
     wrmsrl(MSR_CSTAR, saved_cstar);
     wrmsr(MSR_STAR, 0, (FLAT_RING3_CS32<<16) | __HYPERVISOR_CS);
     wrmsr(MSR_SYSCALL_MASK, EF_VM|EF_RF|EF_NT|EF_DF|EF_IE|EF_TF, 0U);    
+
+    if ( boot_cpu_data.x86_vendor == X86_VENDOR_INTEL )
+    {
+        /* Recover sysenter MSRs */
+        wrmsrl(MSR_IA32_SYSENTER_ESP, saved_sysenter_esp);
+        wrmsrl(MSR_IA32_SYSENTER_EIP, saved_sysenter_eip);
+        wrmsr(MSR_IA32_SYSENTER_CS, __HYPERVISOR_CS, 0);
+    }
 #else /* !defined(CONFIG_X86_64) */
     if ( supervisor_mode_kernel && cpu_has_sep )
         wrmsr(MSR_IA32_SYSENTER_ESP, &init_tss[smp_processor_id()].esp1, 0);

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel