Re: Machine Check in P2010(e500v2)

From: Joakim Tjernlund <hidden>
Date: 2017-09-07 08:41:29

On Thu, 2017-09-07 at 00:50 +0200, Joakim Tjernlund wrote:

On Wed, 2017-09-06 at 21:13 +0000, Leo Li wrote:

quoted

-----Original Message-----
From: Joakim Tjernlund [mailto:Joakim.Tjernlund@infinera.com]
Sent: Wednesday, September 06, 2017 3:54 PM
To: linuxppc-dev@lists.ozlabs.org; Leo Li <redacted>; York =

Sun

quoted

[off-list ref]
Subject: Re: Machine Check in P2010(e500v2)
=20
On Wed, 2017-09-06 at 20:28 +0000, Leo Li wrote:

quoted

-----Original Message-----
From: Joakim Tjernlund [mailto:Joakim.Tjernlund@infinera.com]
Sent: Wednesday, September 06, 2017 3:17 PM
To: linuxppc-dev@lists.ozlabs.org; Leo Li <redacted>; Y=

ork

quoted

Sun [off-list ref]
Subject: Re: Machine Check in P2010(e500v2)
=20
On Wed, 2017-09-06 at 19:31 +0000, Leo Li wrote:

quoted

-----Original Message-----
From: York Sun
Sent: Wednesday, September 06, 2017 10:38 AM
To: Joakim Tjernlund <redacted>; linuxpp=

c-

quoted

dev@lists.ozlabs.org; Leo Li [off-list ref]
Subject: Re: Machine Check in P2010(e500v2)
=20
Scott is no longer with Freescale/NXP. Adding Leo.
=20
On 09/05/2017 01:40 AM, Joakim Tjernlund wrote:

quoted

So after some debugging I found this bug:

@@ -996,7 +998,7 @@ int fsl_pci_mcheck_exception(struct pt_=

regs

quoted

=20
*regs)

quoted

         if (is_in_pci_mem_space(addr)) {
                 if (user_mode(regs)) {
                         pagefault_disable();
-                       ret =3D get_user(regs->nip, &inst);
+                       ret =3D get_user(inst, (__u32 __use=

quoted

+ *)regs->nip);
                         pagefault_enable();
                 } else {
                         ret =3D probe_kernel_address(regs-=

nip,

quoted

inst);
=20
However, the kernel still locked up after fixing that.
Now I wonder why this fixup is there in the first place? Th=

quoted
quoted
quoted
quoted
quoted
quoted
quoted
routine will not really fixup the insn, just return 0xfffff=

fff

quoted

for the failing read and then advance the process NIP.

=20
You are right.  The code here only gives 0xffffffff to the load
instructions and

=20
continue with the next instruction when the load instruction is
causing the machine check.  This will prevent a system lockup whe=

quoted

reading from PCI/RapidIO device which is link down.

quoted

=20
I don't know what is actual problem in your case.  Maybe it is =

quoted

write

=20
instruction instead of read?   Or the code is in a infinite loop =

waiting for a

quoted

=20
valid

quoted

read result?  Are you able to do some further debugging with the =

NIP

quoted

correctly printed?

quoted

=20

=20
According to the MC it is a Read and the NIP also leads to a read=

 in the

quoted

=20
program.

quoted

ATM, I have disabled the fixup but I will enable that again.
Question, is it safe add a small printk when this MC happens(afte=

quoted

fixing up)? I need to see that it has happened as the error is so=

mewhat

quoted

=20
random.

quoted

=20
I think it is safe to add printk as the current machine check handl=

ers are also

quoted

=20
using printk.
=20
I hope so, but if the fixup fires there is no printk at all so I was =

a bit unsure.

quoted

Don't like this fixup though, is there not a better way than faking a=

 read to user

quoted

space(or kernel for that matter) ?

=20
I don't have a better idea.  Without the fixup, the offending load inst=

ruction will never finish if there is anything wrong with the backing devic=
e and freeze the whole system.  Do you have any suggestion in mind?

quoted

=20

=20
But it never finishes the load, it just fakes a load of 0xfffffffff, for =

user space I rather have it signal

a SIGBUS but that does not seem to work either, at least not for us but t=

hat could be a bug in general MC code

 maybe.
This fixup might be valid for kernel only as it has never worked for user=

 space due to the bug I found.

=20
Where can I read about this errata ?

I have look high and low an cannot find an errata which maps to this fixup.
The closest I get is A-005125 which seems to have another workaround, I can=
not find
any evidence that this workaround has been applied in Linux, can you?

 Jocke=

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help