Occasional userland crashes on sparc in 4.1.0+

Occasional userland crashes on sparc in 4.1.0+

[Meelis Roos]

While compiling 4.2-rc3 on my test machines, I got segfaults during 
compilation on V240 and V440. Running repeated compilations encountered 
some more of these, on both 4.1.0 and 4.2-rc3.

At the first glance, no UltraSparc II machines had this but I have not 
done any more stresstesting on them.

Symptions from dmesg:

[111825.922395] as[22148]: segfault at 2424f128 ip 0000000000021f30 (rpc 0000000000050c85) sp 00000000ffcea018 error 30001 in as[10000+46000]
[76756.534860] git[20101]: segfault at 0 ip 00000000f7752448 (rpc 00000000f76e7d68) sp 00000000f6fbeef0 error 30001 in libc-2.19.so[f7668000+16a000]

These happen once per several days for continuous compilation. So they 
might be slightly older than 4.1.0.

Additionally, v240 still sees the possible irq lock inversion dependency 
as reported in https://lkml.org/lkml/2015/5/11/348 but this is probably 
not related.

-- 
Meelis Roos (mroos@linux.ee)

Re: Occasional userland crashes on sparc in 4.1.0+

[Meelis Roos]

> While compiling 4.2-rc3 on my test machines, I got segfaults during 
> compilation on V240 and V440. Running repeated compilations encountered 
> some more of these, on both 4.1.0 and 4.2-rc3.
> 
> At the first glance, no UltraSparc II machines had this but I have not 
> done any more stresstesting on them.

Reconnected and tested my Ultra 1, with problems:

[63949.058730] resolvconf[11593]: segfault at 3860 ip 00000000f7c4e790 (rpc 00000000f7c4e714) sp 00000000fffeaf60 error 30001 in libc-2.19.so[f7b98000+16a000]
[63949.389191] BUG: Bad rss-counter state mm:fffff800335b4440 idx:1 val:1
[63949.465567] BUG: non-zero nr_ptes on freeing mm: 1

This was with 4.1.0 while compiling newer git release.

-- 
Meelis Roos (mroos@linux.ee)

Re: Occasional userland crashes on sparc in 4.1.0+

[Meelis Roos]

> > While compiling 4.2-rc3 on my test machines, I got segfaults during 
> > compilation on V240 and V440. Running repeated compilations encountered 
> > some more of these, on both 4.1.0 and 4.2-rc3.
> > 
> > At the first glance, no UltraSparc II machines had this but I have not 
> > done any more stresstesting on them.
> 
> Reconnected and tested my Ultra 1, with problems:
> 
> [63949.058730] resolvconf[11593]: segfault at 3860 ip 00000000f7c4e790 (rpc 00000000f7c4e714) sp 00000000fffeaf60 error 30001 in libc-2.19.so[f7b98000+16a000]
> [63949.389191] BUG: Bad rss-counter state mm:fffff800335b4440 idx:1 val:1
> [63949.465567] BUG: non-zero nr_ptes on freeing mm: 1
> 
> This was with 4.1.0 while compiling newer git release.

Tested todays git with the FPU fix on most of my sparcs and got the same 
error on Ultra 5 running 4.2.0-rc5-00078-g49d7c65-dirty (dirty because 
of test_rhashtable.c cond_resxhed() additions) while doing repeated 
make; make clean:

BUG: Bad rss-counter state mm:fffff8001c982ea0 idx:0 val:8
BUG: non-zero nr_ptes on freeing mm: 1

-- 
Meelis Roos (mroos@linux.ee)

Re: Occasional userland crashes on sparc in 4.1.0+

[Meelis Roos]

> Additionally, v240 still sees the possible irq lock inversion dependency 
> as reported in https://lkml.org/lkml/2015/5/11/348 but this is probably 
> not related.

Today I also got a spontaneous reboot and a hang from the same V240, one 
from 4.2.0-rc3+ and one from todays 4.2.0-rc5+ kernel.

-- 
Meelis Roos (mroos@linux.ee)

Re: Occasional userland crashes on sparc in 4.1.0+

[Meelis Roos]

> > > While compiling 4.2-rc3 on my test machines, I got segfaults during 
> > > compilation on V240 and V440. Running repeated compilations encountered 
> > > some more of these, on both 4.1.0 and 4.2-rc3.
> > > 
> > > At the first glance, no UltraSparc II machines had this but I have not 
> > > done any more stresstesting on them.
> > 
> > Reconnected and tested my Ultra 1, with problems:
> > 
> > [63949.058730] resolvconf[11593]: segfault at 3860 ip 00000000f7c4e790 (rpc 00000000f7c4e714) sp 00000000fffeaf60 error 30001 in libc-2.19.so[f7b98000+16a000]
> > [63949.389191] BUG: Bad rss-counter state mm:fffff800335b4440 idx:1 val:1
> > [63949.465567] BUG: non-zero nr_ptes on freeing mm: 1
> > 
> > This was with 4.1.0 while compiling newer git release.
> 
> Tested todays git with the FPU fix on most of my sparcs and got the same 
> error on Ultra 5 running 4.2.0-rc5-00078-g49d7c65-dirty (dirty because 
> of test_rhashtable.c cond_resxhed() additions) while doing repeated 
> make; make clean:
> 
> BUG: Bad rss-counter state mm:fffff8001c982ea0 idx:0 val:8
> BUG: non-zero nr_ptes on freeing mm: 1

And similar but slightly different on Fire V100 with UltraSparc IIe with 
4.2.0-rc5-00078-g49d7c65:

[ 9271.941651] BUG: Bad rss-counter state mm:fffff8006c8b20c0 idx:1 val:1
[ 9272.027556] BUG: non-zero nr_ptes on freeing mm: 1
[ 9272.090591] BUG: non-zero nr_pmds on freeing mm: 1

Not bisecting yet because I have another sparc bisection in progress - 
will report when done.

-- 
Meelis Roos (mroos@linux.ee)