During power failure, Asynchronous DRAM Refresh (ADR) flushes
the write buffer in memory controllers into NVDIMM, but does not
flush processor caches. While the kernel and application code
need to take care of processor cache flush, they may not be able
to do so during panic or reboot.
Add processor cache flush (wbinvd) to the stop-CPUs interfaces,
native_stop_other_cpus() and nmi_shootdown_cpus(), which are
called during panic and reboot as follows. These wbinvd()s are
called on each CPU after its irq/APIC is disabled.
- panic()
+ smp_send_stop()
o native_stop_other_cpus()
o stop_this_cpu()
- native_machine_restart()
- native_machine_halt()
- native_machine_power_off()
+ native_machine_shutdown()
+ stop_other_cpus()
o native_stop_other_cpus()
o stop_this_cpu()
- native_machine_crash_shutdown()
+ kdump_nmi_shootdown_cpus()
o nmi_shootdown_cpus()
o crash_nmi_callback()
Note, the cpu offline path, mwait_play_dead(), already calls
wbinvd().
Signed-off-by: Toshi Kani <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: H. Peter Anvin <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Borislav Petkov <[email protected]>
Cc: linux-nvdimm <[email protected]>
---
arch/x86/kernel/process.c | 2 ++
arch/x86/kernel/reboot.c | 5 +++++
arch/x86/kernel/smp.c | 2 ++
3 files changed, 9 insertions(+)
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 397688b..3a1f381 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -320,6 +320,8 @@ void stop_this_cpu(void *dummy)
set_cpu_online(smp_processor_id(), false);
disable_local_APIC();
+ wbinvd();
+
for (;;)
halt();
}
diff --git a/arch/x86/kernel/reboot.c b/arch/x86/kernel/reboot.c
index 86db4bc..5ef4d4b 100644
--- a/arch/x86/kernel/reboot.c
+++ b/arch/x86/kernel/reboot.c
@@ -737,6 +737,9 @@ static int crash_nmi_callback(unsigned int val, struct pt_regs *regs)
shootdown_callback(cpu, regs);
atomic_dec(&waiting_for_crash_ipi);
+
+ wbinvd();
+
/* Assume hlt works */
halt();
for (;;)
@@ -780,6 +783,8 @@ void nmi_shootdown_cpus(nmi_shootdown_cb callback)
smp_send_nmi_allbutself();
+ wbinvd();
+
msecs = 1000; /* Wait at most a second for the other cpus to stop */
while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
mdelay(1);
diff --git a/arch/x86/kernel/smp.c b/arch/x86/kernel/smp.c
index 15aaa69..41e7ca8 100644
--- a/arch/x86/kernel/smp.c
+++ b/arch/x86/kernel/smp.c
@@ -244,6 +244,8 @@ finish:
local_irq_save(flags);
disable_local_APIC();
local_irq_restore(flags);
+
+ wbinvd();
}
/*
On Thu, Jul 23, 2015 at 10:33 AM, Toshi Kani <[email protected]> wrote:
> During power failure, Asynchronous DRAM Refresh (ADR) flushes
> the write buffer in memory controllers into NVDIMM, but does not
> flush processor caches. While the kernel and application code
> need to take care of processor cache flush, they may not be able
> to do so during panic or reboot.
>
> Add processor cache flush (wbinvd) to the stop-CPUs interfaces,
> native_stop_other_cpus() and nmi_shootdown_cpus(), which are
> called during panic and reboot as follows. These wbinvd()s are
> called on each CPU after its irq/APIC is disabled.
>
> - panic()
> + smp_send_stop()
> o native_stop_other_cpus()
> o stop_this_cpu()
>
> - native_machine_restart()
> - native_machine_halt()
> - native_machine_power_off()
> + native_machine_shutdown()
> + stop_other_cpus()
> o native_stop_other_cpus()
> o stop_this_cpu()
>
> - native_machine_crash_shutdown()
> + kdump_nmi_shootdown_cpus()
> o nmi_shootdown_cpus()
> o crash_nmi_callback()
>
> Note, the cpu offline path, mwait_play_dead(), already calls
> wbinvd().
>
If the application is already prepared for surprise power loss what
additional benefit is there to flushing caches on panic? In other
words, if the application needs this for correctness then it is broken
with respect to surprise power loss, otherwise these flushes are not
necessary.
On Thu, 2015-07-23 at 10:40 -0700, Dan Williams wrote:
> On Thu, Jul 23, 2015 at 10:33 AM, Toshi Kani <[email protected]> wrote:
> > During power failure, Asynchronous DRAM Refresh (ADR) flushes
> > the write buffer in memory controllers into NVDIMM, but does not
> > flush processor caches. While the kernel and application code
> > need to take care of processor cache flush, they may not be able
> > to do so during panic or reboot.
> >
> > Add processor cache flush (wbinvd) to the stop-CPUs interfaces,
> > native_stop_other_cpus() and nmi_shootdown_cpus(), which are
> > called during panic and reboot as follows. These wbinvd()s are
> > called on each CPU after its irq/APIC is disabled.
> >
> > - panic()
> > + smp_send_stop()
> > o native_stop_other_cpus()
> > o stop_this_cpu()
> >
> > - native_machine_restart()
> > - native_machine_halt()
> > - native_machine_power_off()
> > + native_machine_shutdown()
> > + stop_other_cpus()
> > o native_stop_other_cpus()
> > o stop_this_cpu()
> >
> > - native_machine_crash_shutdown()
> > + kdump_nmi_shootdown_cpus()
> > o nmi_shootdown_cpus()
> > o crash_nmi_callback()
> >
> > Note, the cpu offline path, mwait_play_dead(), already calls
> > wbinvd().
> >
>
> If the application is already prepared for surprise power loss what
> additional benefit is there to flushing caches on panic? In other
> words, if the application needs this for correctness then it is broken
> with respect to surprise power loss, otherwise these flushes are not
> necessary.
I agree that well-written applications should withstand with their own
journaling mechanisms whey they access NVDIMM directly. But not all
applications are well-written or perfect on this regard. msync does not
flush processor caches at this point, either. So, we want to save all
updates to NVDIMM as much as possible to minimize inconsistency.
Thanks,
-Toshi