Skip to contentSkip to conversation

x86/smpboot: Don't unconditionally call memguard_guard_stack() in cpu_smpboot_alloc()

PATCH In review Andrew Cooper 2 participants 6 messages Oct 14, 2020 – Oct 16, 2020

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 x86/smpboot: Don't unconditionally call memguard_guard_stack() in cpu_smpboot_alloc()

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes. For CPU Parking, it would invalidate the shadow stack of the parked CPUs, causing a #DF on the next NMI/#MC to hit the thread.

---
 xen/arch/x86/smpboot.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
xen/arch/x86/smpboot.c +6 -4
index 5708573c41..67e727cebd 100644
@@ -997,16 +997,18 @@ static int cpu_smpboot_alloc(unsigned int cpu)
         memflags = MEMF_node(node);
 
     if ( stack_base[cpu] == NULL )
+    {
         stack_base[cpu] = alloc_xenheap_pages(STACK_ORDER, memflags);
-    if ( stack_base[cpu] == NULL )
-        goto out;
+        if ( !stack_base[cpu] )
+            goto out;
+
+        memguard_guard_stack(stack_base[cpu]);
+    }
 
     info = get_cpu_info_from_stack((unsigned long)stack_base[cpu]);
     info->processor_id = cpu;
     info->per_cpu_offset = __per_cpu_offset[cpu];
 
-    memguard_guard_stack(stack_base[cpu]);
-
     gdt = per_cpu(gdt, cpu) ?: alloc_xenheap_pages(0, memflags);
     if ( gdt == NULL )
         goto out;
Signature · 1 line
2.11.0
#1 lore raw
Quoted 28 lines · Andrew Cooper

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

Quoted 7 lines

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes.

The code change looks correct to me, but since I don't understand this part I'm afraid I may be overlooking something. I understand the "turn the BSP shadow stack into regular mappings" relates to cpu_smpboot_free()'s call to memguard_unguard_stack(), but I didn't think we come through cpu_smpboot_free() for the BSP upon entering or leaving S3.

Jan

#2 lore raw
Quoted 36 lines · Jan Beulich
Quoted 28 lines · Andrew Cooper

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

Quoted 7 lines

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes.

Quoted 6 lines

The code change looks correct to me, but since I don't understand this part I'm afraid I may be overlooking something. I understand the "turn the BSP shadow stack into regular mappings" relates to cpu_smpboot_free()'s call to memguard_unguard_stack(), but I didn't think we come through cpu_smpboot_free() for the BSP upon entering or leaving S3.

The v1 really did fix Marek's repro of the problem.

The only possible way this can occur is if, somewhere, there is a call to cpu_smpboot_free() for CPU0 with remove=0 on the S3 path

I have to admit that I can't actually spot where it is.

Either way - it doesn't impact the fix, which attempts to make "the stack" into a single object.  I experimented with introducing smpboot_{alloc,free}_stack(), but the result wasn't clean and I abandoned that approach.

~Andrew

#3 lore raw
Quoted 44 lines · Andrew Cooper
Quoted 36 lines · Jan Beulich
Quoted 28 lines · Andrew Cooper

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

Quoted 7 lines

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes.

Quoted 6 lines

The code change looks correct to me, but since I don't understand this part I'm afraid I may be overlooking something. I understand the "turn the BSP shadow stack into regular mappings" relates to cpu_smpboot_free()'s call to memguard_unguard_stack(), but I didn't think we come through cpu_smpboot_free() for the BSP upon entering or leaving S3.

Quoted 4 lines

The v1 really did fix Marek's repro of the problem.

The only possible way this can occur is if, somewhere, there is a call to cpu_smpboot_free() for CPU0 with remove=0 on the S3 path

I didn't think it was the BSP's stack that got written to, but the first AP's before letting it run.

Jan

#4 lore raw
Quoted 44 lines · Jan Beulich
Quoted 43 lines · Andrew Cooper
Quoted 42 lines · Jan Beulich
Quoted 28 lines · Andrew Cooper

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

Quoted 7 lines

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes.

The code change looks correct to me, but since I don't understand this part I'm afraid I may be overlooking something. I understand the "turn the BSP shadow stack into regular mappings" relates to cpu_smpboot_free()'s call to memguard_unguard_stack(), but I didn't think we come through cpu_smpboot_free() for the BSP upon entering or leaving S3.

Quoted 6 lines
Quoted 4 lines

The v1 really did fix Marek's repro of the problem.

The only possible way this can occur is if, somewhere, there is a call to cpu_smpboot_free() for CPU0 with remove=0 on the S3 path

I didn't think it was the BSP's stack that got written to, but the first AP's before letting it run.

Oh yes - my analysis was wrong.  The CPU notifier for CPU 1 to come up runs on CPU 0.

So only the --- text was wrong.  Are you happy with the fix now?

~Andrew

Reviewed-by
#5 lore raw
Quoted 49 lines · Andrew Cooper
Quoted 48 lines · Jan Beulich
Quoted 47 lines · Andrew Cooper
Quoted 36 lines · Jan Beulich
Quoted 28 lines · Andrew Cooper

cpu_smpboot_alloc() is designed to be idempotent with respect to partially initialised state. This occurs for S3 and CPU parking, where enough state to handle NMIs/#MCs needs to remain valid for the entire lifetime of Xen, even when we otherwise want to offline the CPU.

For simplicity between various configuration, Xen always uses shadow stack mappings (Read-only + Dirty) for the guard page, irrespective of whether CET-SS is enabled.

Unfortunately, the CET-SS changes in memguard_guard_stack() broke idempotency by first writing out the supervisor shadow stack tokens with plain writes, then changing the mapping to being read-only.

This ordering is strictly necessary to configure the BSP, which cannot have the supervisor tokens be written with WRSS.

Instead of calling memguard_guard_stack() unconditionally, call it only when actually allocating a new stack. Xenheap allocates are guaranteed to be writeable, and the net result is idempotency WRT configuring stack_base[].

Fixes 91d26ed304f ("x86/shstk: Create shadow stacks")

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Wei Liu <[email protected]>

Quoted 7 lines

This can more easily be demonstrated with CPU hotplug than S3, and the absence of bug reports goes to show how rarely hotplug is used.

v2:
* Don't break S3/CPU parking in combination with CET-SS. v1 would, for S3, turn the BSP shadow stack into regular mappings, and #DF as soon as the TLB shootdown completes.

Quoted 6 lines

The code change looks correct to me, but since I don't understand this part I'm afraid I may be overlooking something. I understand the "turn the BSP shadow stack into regular mappings" relates to cpu_smpboot_free()'s call to memguard_unguard_stack(), but I didn't think we come through cpu_smpboot_free() for the BSP upon entering or leaving S3.

The v1 really did fix Marek's repro of the problem.

The only possible way this can occur is if, somewhere, there is a call to cpu_smpboot_free() for CPU0 with remove=0 on the S3 path

Quoted 7 lines
Quoted 2 lines

I didn't think it was the BSP's stack that got written to, but the first AP's before letting it run.

Oh yes - my analysis was wrong.  The CPU notifier for CPU 1 to come up runs on CPU 0.

So only the --- text was wrong.  Are you happy with the fix now?

Indeed I am:

Jan