When calling smp_call_ipl_cpu() from the IPL CPU, we will try to read
from pcpu_devices->lowcore. However, due to prefixing, that will result
in reading from absolute address 0 on that CPU. We have to go via the
actual lowcore instead.
This means that right now, we will read lc->nodat_stack == 0 and
therfore work on a very wrong stack.
This BUG essentially broke rebooting under QEMU TCG (which will report
a low address protection exception). And checking under KVM, it is
also broken under KVM. With 1 VCPU it can be easily triggered.
:/# echo 1 > /proc/sys/kernel/sysrq
:/# echo b > /proc/sysrq-trigger
[ 28.476745] sysrq: SysRq : Resetting
[ 28.476793] Kernel stack overflow.
[ 28.476817] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
[ 28.476820] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
[ 28.476826] Krnl PSW : 0400c00180000000 0000000000115c0c (pcpu_delegate+0x12c/0x140)
[ 28.476861] R:0 T:1 IO:0 EX:0 Key:0 M:0 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
[ 28.476863] Krnl GPRS: ffffffffffffffff 0000000000000000 000000000010dff8 0000000000000000
[ 28.476864] 0000000000000000 0000000000000000 0000000000ab7090 000003e0006efbf0
[ 28.476864] 000000000010dff8 0000000000000000 0000000000000000 0000000000000000
[ 28.476865] 000000007fffc000 0000000000730408 000003e0006efc58 0000000000000000
[ 28.476887] Krnl Code: 0000000000115bfe: 4170f000 la %r7,0(%r15)
[ 28.476887] 0000000000115c02: 41f0a000 la %r15,0(%r10)
[ 28.476887] #0000000000115c06: e370f0980024 stg %r7,152(%r15)
[ 28.476887] >0000000000115c0c: c0e5fffff86e brasl %r14,114ce8
[ 28.476887] 0000000000115c12: 41f07000 la %r15,0(%r7)
[ 28.476887] 0000000000115c16: a7f4ffa8 brc 15,115b66
[ 28.476887] 0000000000115c1a: 0707 bcr 0,%r7
[ 28.476887] 0000000000115c1c: 0707 bcr 0,%r7
[ 28.476901] Call Trace:
[ 28.476902] Last Breaking-Event-Address:
[ 28.476920] [<0000000000a01c4a>] arch_call_rest_init+0x22/0x80
[ 28.476927] Kernel panic - not syncing: Corrupt kernel stack, can't continue.
[ 28.476930] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
[ 28.476932] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
[ 28.476932] Call Trace:
Reported-by: Cornelia Huck <[email protected]>
Signed-off-by: David Hildenbrand <[email protected]>
---
arch/s390/kernel/smp.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/arch/s390/kernel/smp.c b/arch/s390/kernel/smp.c
index f82b3d3c36e2..be32dd0b4191 100644
--- a/arch/s390/kernel/smp.c
+++ b/arch/s390/kernel/smp.c
@@ -381,8 +381,13 @@ void smp_call_online_cpu(void (*func)(void *), void *data)
*/
void smp_call_ipl_cpu(void (*func)(void *), void *data)
{
+ struct lowcore *lc = pcpu_devices->lowcore;
+
+ if (pcpu_devices[0].address == stap())
+ lc = &S390_lowcore;
+
pcpu_delegate(&pcpu_devices[0], func, data,
- pcpu_devices->lowcore->nodat_stack);
+ lc->nodat_stack);
}
int smp_find_processor_id(u16 address)
--
2.17.2
On 11.01.19 15:18, David Hildenbrand wrote:
> When calling smp_call_ipl_cpu() from the IPL CPU, we will try to read
> from pcpu_devices->lowcore. However, due to prefixing, that will result
> in reading from absolute address 0 on that CPU. We have to go via the
> actual lowcore instead.
>
> This means that right now, we will read lc->nodat_stack == 0 and
> therfore work on a very wrong stack.
>
> This BUG essentially broke rebooting under QEMU TCG (which will report
> a low address protection exception). And checking under KVM, it is
> also broken under KVM. With 1 VCPU it can be easily triggered.
>
> :/# echo 1 > /proc/sys/kernel/sysrq
> :/# echo b > /proc/sysrq-trigger
> [ 28.476745] sysrq: SysRq : Resetting
> [ 28.476793] Kernel stack overflow.
> [ 28.476817] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476820] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476826] Krnl PSW : 0400c00180000000 0000000000115c0c (pcpu_delegate+0x12c/0x140)
> [ 28.476861] R:0 T:1 IO:0 EX:0 Key:0 M:0 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
> [ 28.476863] Krnl GPRS: ffffffffffffffff 0000000000000000 000000000010dff8 0000000000000000
> [ 28.476864] 0000000000000000 0000000000000000 0000000000ab7090 000003e0006efbf0
> [ 28.476864] 000000000010dff8 0000000000000000 0000000000000000 0000000000000000
> [ 28.476865] 000000007fffc000 0000000000730408 000003e0006efc58 0000000000000000
> [ 28.476887] Krnl Code: 0000000000115bfe: 4170f000 la %r7,0(%r15)
> [ 28.476887] 0000000000115c02: 41f0a000 la %r15,0(%r10)
> [ 28.476887] #0000000000115c06: e370f0980024 stg %r7,152(%r15)
> [ 28.476887] >0000000000115c0c: c0e5fffff86e brasl %r14,114ce8
> [ 28.476887] 0000000000115c12: 41f07000 la %r15,0(%r7)
> [ 28.476887] 0000000000115c16: a7f4ffa8 brc 15,115b66
> [ 28.476887] 0000000000115c1a: 0707 bcr 0,%r7
> [ 28.476887] 0000000000115c1c: 0707 bcr 0,%r7
> [ 28.476901] Call Trace:
> [ 28.476902] Last Breaking-Event-Address:
> [ 28.476920] [<0000000000a01c4a>] arch_call_rest_init+0x22/0x80
> [ 28.476927] Kernel panic - not syncing: Corrupt kernel stack, can't continue.
> [ 28.476930] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476932] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476932] Call Trace:
>
I guess this has been broken for a long time. Most probably
ce3dc447493f ("s390: add support for virtually mapped kernel stacks")
finally made it trigger the low address protection.
@@ -372,8 +382,7 @@ void smp_call_online_cpu(void (*func)(void *), void
*data)
void smp_call_ipl_cpu(void (*func)(void *), void *data)
{
pcpu_delegate(&pcpu_devices[0], func, data,
- pcpu_devices->lowcore->panic_stack -
- PANIC_FRAME_OFFSET + PAGE_SIZE);
+ pcpu_devices->lowcore->nodat_stack);
}
The old "- PANIC_FRAME_OFFSET + PAGE_SIZE" Made sure that we were not
trying to write at absolute address 0, but instead something >= 512,
whereby low address protection would not trigger.
> Reported-by: Cornelia Huck <[email protected]>
> Signed-off-by: David Hildenbrand <[email protected]>
> ---
> arch/s390/kernel/smp.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/arch/s390/kernel/smp.c b/arch/s390/kernel/smp.c
> index f82b3d3c36e2..be32dd0b4191 100644
> --- a/arch/s390/kernel/smp.c
> +++ b/arch/s390/kernel/smp.c
> @@ -381,8 +381,13 @@ void smp_call_online_cpu(void (*func)(void *), void *data)
> */
> void smp_call_ipl_cpu(void (*func)(void *), void *data)
> {
> + struct lowcore *lc = pcpu_devices->lowcore;
> +
> + if (pcpu_devices[0].address == stap())
> + lc = &S390_lowcore;
> +
> pcpu_delegate(&pcpu_devices[0], func, data,
> - pcpu_devices->lowcore->nodat_stack);
> + lc->nodat_stack);
> }
>
> int smp_find_processor_id(u16 address)
>
--
Thanks,
David / dhildenb
On Fri, 11 Jan 2019 15:18:22 +0100
David Hildenbrand <[email protected]> wrote:
> When calling smp_call_ipl_cpu() from the IPL CPU, we will try to read
> from pcpu_devices->lowcore. However, due to prefixing, that will result
> in reading from absolute address 0 on that CPU. We have to go via the
> actual lowcore instead.
>
> This means that right now, we will read lc->nodat_stack == 0 and
> therfore work on a very wrong stack.
>
> This BUG essentially broke rebooting under QEMU TCG (which will report
> a low address protection exception). And checking under KVM, it is
> also broken under KVM. With 1 VCPU it can be easily triggered.
>
> :/# echo 1 > /proc/sys/kernel/sysrq
> :/# echo b > /proc/sysrq-trigger
> [ 28.476745] sysrq: SysRq : Resetting
> [ 28.476793] Kernel stack overflow.
> [ 28.476817] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476820] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476826] Krnl PSW : 0400c00180000000 0000000000115c0c (pcpu_delegate+0x12c/0x140)
> [ 28.476861] R:0 T:1 IO:0 EX:0 Key:0 M:0 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
> [ 28.476863] Krnl GPRS: ffffffffffffffff 0000000000000000 000000000010dff8 0000000000000000
> [ 28.476864] 0000000000000000 0000000000000000 0000000000ab7090 000003e0006efbf0
> [ 28.476864] 000000000010dff8 0000000000000000 0000000000000000 0000000000000000
> [ 28.476865] 000000007fffc000 0000000000730408 000003e0006efc58 0000000000000000
> [ 28.476887] Krnl Code: 0000000000115bfe: 4170f000 la %r7,0(%r15)
> [ 28.476887] 0000000000115c02: 41f0a000 la %r15,0(%r10)
> [ 28.476887] #0000000000115c06: e370f0980024 stg %r7,152(%r15)
> [ 28.476887] >0000000000115c0c: c0e5fffff86e brasl %r14,114ce8
> [ 28.476887] 0000000000115c12: 41f07000 la %r15,0(%r7)
> [ 28.476887] 0000000000115c16: a7f4ffa8 brc 15,115b66
> [ 28.476887] 0000000000115c1a: 0707 bcr 0,%r7
> [ 28.476887] 0000000000115c1c: 0707 bcr 0,%r7
> [ 28.476901] Call Trace:
> [ 28.476902] Last Breaking-Event-Address:
> [ 28.476920] [<0000000000a01c4a>] arch_call_rest_init+0x22/0x80
> [ 28.476927] Kernel panic - not syncing: Corrupt kernel stack, can't continue.
> [ 28.476930] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476932] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476932] Call Trace:
>
> Reported-by: Cornelia Huck <[email protected]>
> Signed-off-by: David Hildenbrand <[email protected]>
> ---
> arch/s390/kernel/smp.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
I can confirm that this patch fixes rebooting for me, both under
QEMU/TCG and under QEMU/KVM (1 vcpu).
Tested-by: Cornelia Huck <[email protected]>
On Fri, 11 Jan 2019 15:18:22 +0100
David Hildenbrand <[email protected]> wrote:
> When calling smp_call_ipl_cpu() from the IPL CPU, we will try to read
> from pcpu_devices->lowcore. However, due to prefixing, that will result
> in reading from absolute address 0 on that CPU. We have to go via the
> actual lowcore instead.
>
> This means that right now, we will read lc->nodat_stack == 0 and
> therfore work on a very wrong stack.
>
> This BUG essentially broke rebooting under QEMU TCG (which will report
> a low address protection exception). And checking under KVM, it is
> also broken under KVM. With 1 VCPU it can be easily triggered.
>
> :/# echo 1 > /proc/sys/kernel/sysrq
> :/# echo b > /proc/sysrq-trigger
> [ 28.476745] sysrq: SysRq : Resetting
> [ 28.476793] Kernel stack overflow.
> [ 28.476817] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476820] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476826] Krnl PSW : 0400c00180000000 0000000000115c0c (pcpu_delegate+0x12c/0x140)
> [ 28.476861] R:0 T:1 IO:0 EX:0 Key:0 M:0 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
> [ 28.476863] Krnl GPRS: ffffffffffffffff 0000000000000000 000000000010dff8 0000000000000000
> [ 28.476864] 0000000000000000 0000000000000000 0000000000ab7090 000003e0006efbf0
> [ 28.476864] 000000000010dff8 0000000000000000 0000000000000000 0000000000000000
> [ 28.476865] 000000007fffc000 0000000000730408 000003e0006efc58 0000000000000000
> [ 28.476887] Krnl Code: 0000000000115bfe: 4170f000 la %r7,0(%r15)
> [ 28.476887] 0000000000115c02: 41f0a000 la %r15,0(%r10)
> [ 28.476887] #0000000000115c06: e370f0980024 stg %r7,152(%r15)
> [ 28.476887] >0000000000115c0c: c0e5fffff86e brasl %r14,114ce8
> [ 28.476887] 0000000000115c12: 41f07000 la %r15,0(%r7)
> [ 28.476887] 0000000000115c16: a7f4ffa8 brc 15,115b66
> [ 28.476887] 0000000000115c1a: 0707 bcr 0,%r7
> [ 28.476887] 0000000000115c1c: 0707 bcr 0,%r7
> [ 28.476901] Call Trace:
> [ 28.476902] Last Breaking-Event-Address:
> [ 28.476920] [<0000000000a01c4a>] arch_call_rest_init+0x22/0x80
> [ 28.476927] Kernel panic - not syncing: Corrupt kernel stack, can't continue.
> [ 28.476930] CPU: 0 PID: 424 Comm: sh Not tainted 5.0.0-rc1+ #13
> [ 28.476932] Hardware name: IBM 2964 NE1 716 (KVM/Linux)
> [ 28.476932] Call Trace:
>
> Reported-by: Cornelia Huck <[email protected]>
> Signed-off-by: David Hildenbrand <[email protected]>
> ---
> arch/s390/kernel/smp.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/arch/s390/kernel/smp.c b/arch/s390/kernel/smp.c
> index f82b3d3c36e2..be32dd0b4191 100644
> --- a/arch/s390/kernel/smp.c
> +++ b/arch/s390/kernel/smp.c
> @@ -381,8 +381,13 @@ void smp_call_online_cpu(void (*func)(void *), void *data)
> */
> void smp_call_ipl_cpu(void (*func)(void *), void *data)
> {
> + struct lowcore *lc = pcpu_devices->lowcore;
> +
> + if (pcpu_devices[0].address == stap())
> + lc = &S390_lowcore;
> +
> pcpu_delegate(&pcpu_devices[0], func, data,
> - pcpu_devices->lowcore->nodat_stack);
> + lc->nodat_stack);
> }
>
> int smp_find_processor_id(u16 address)
Uhh, subtle. With EDAT-1 this sort of works because the lowcore of CPU #0
is allocated in an area of the kernel address space that is backed with 1M
pages. It is nevertheless broken.
I have added
Fixes: 2f859d0dad81 ("s390/smp: reduce size of struct pcpu")
Cc: [email protected] # 4.0+
and will queue the patch for the next pull. Thanks!
--
blue skies,
Martin.
"Reality continues to ruin my life." - Calvin.