Re: [PATCH v6 1/2] x86/kexec: Build identity mapping for EFI systab and ACPI tables

From: Dave Young
Date: Tue May 14 2019 - 09:00:24 EST


On 05/14/19 at 01:38pm, Peter Zijlstra wrote:
> On Tue, May 14, 2019 at 04:48:41PM +0800, Dave Young wrote:
>
> > > I did some tests on the laptop, thing is:
> > > 1. apply the 3 patches (two you posted + Boris's revert commit 52b922c3d49c)
> > > on latest Linus master branch, everything works fine.
> > >
> > > 2. build and test the tip/next-merge-window branch, kernel hangs early
> > > without output, (both 1st boot and kexec boot)
> >
> > Update about 2. It should be not early rsdp related, I got the boot log
> > Since can not reproduce with Linus master branch it may have been fixed.
>
> Nothing was changed here since PTI.
>
> > [ 0.685374][ T1] rcu: Hierarchical SRCU implementation.
> > [ 0.686414][ T1] general protection fault: 0000 [#1] SMP PTI
> > [ 0.687328][ T1] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.1.0-rc6+ #877
> > [ 0.687328][ T1] Hardware name: LENOVO 4236NUC/4236NUC, BIOS 83ET82WW (1.52 ) 06/04/2018
> > [ 0.687328][ T1] RIP: 0010:reserve_ds_buffers+0x34e/0x450
>
> > [ 0.687328][ T1] Call Trace:
> > [ 0.687328][ T1] ? hardlockup_detector_event_create+0x50/0x50
> > [ 0.687328][ T1] x86_reserve_hardware+0x173/0x180
> > [ 0.687328][ T1] x86_pmu_event_init+0x39/0x220
>
> The DS buffers are special in that they're part of cpu_entrt_area. If
> this comes apart it might mean your pagetables are dodgy.

Hmm, it seems caused by some WIP branch patches, I suspect below:
commit 124d6af5a5f559e516ed2c6ea857e889ed293b43
x86/paravirt: Standardize 'insn_buff' variable names

The suspicious line is "per_cpu(insn_buff, cpu) = insn_buff;"

I can help to test if need to try anything, eg. debug patch.

I do not know anything of the pti and ds buffer logic, but below chunk
make the next-merge-window branch booting fine on the laptop.
---
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index ad47f6415b17..fa254c576032 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -337,7 +337,7 @@ static int alloc_pebs_buffer(int cpu)
struct debug_store *ds = hwev->ds;
size_t bsiz = x86_pmu.pebs_buffer_size;
int max, node = cpu_to_node(cpu);
- void *buffer, *insn_buff, *cea;
+ void *buffer, *ibuff, *cea;

if (!x86_pmu.pebs)
return 0;
@@ -351,12 +351,12 @@ static int alloc_pebs_buffer(int cpu)
* buffer then.
*/
if (x86_pmu.intel_cap.pebs_format < 2) {
- insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
- if (!insn_buff) {
+ ibuff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
+ if (!ibuff) {
dsfree_pages(buffer, bsiz);
return -ENOMEM;
}
- per_cpu(insn_buff, cpu) = insn_buff;
+ per_cpu(insn_buff, cpu) = ibuff;
}
hwev->ds_pebs_vaddr = buffer;
/* Update the cpu entry area mapping */