Le 24/09/2017 à 18:05, Guenter Roeck a écrit :
On 09/21/2017 11:44 AM, Christophe LEROY wrote:
Le 20/09/2017 à 05:45, Guenter Roeck a écrit :
On 09/19/2017 08:05 PM, Michael Ellerman wrote:
Guenter Roeck <li...@roeck-us.net> writes:
Hi,
I see a the following traceback when running an SMP image based on
85xx/mpc85xx_cds_defconfig in qemu.
------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at kernel/smp.c:416
smp_call_function_many+0xcc/0x2fc
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.14.0-rc1-00009-g0666f56 #1
task: cf830000 task.stack: cf82e000
NIP: c00a93c8 LR: c00a9634 CTR: 00000001
REGS: cf82fde0 TRAP: 0700 Not tainted (4.14.0-rc1-00009-g0666f56)
MSR: 00021000 <CE,ME> CR: 24000082 XER: 00000000
GPR00: c00a9634 cf82fe90 cf830000 c050ad3c c0015a54 00000000
00000001 00000001
GPR08: 00000001 00000000 00000000 cf82e000 24000084 00000000
c0003150 00000000
GPR16: 00000000 00000000 00000000 00000000 00000000 00000001
00000000 c0510000
GPR24: 00000000 c0015a54 00000000 c050ad3c c051823c c050ad3c
00000025 00000000
NIP [c00a93c8] smp_call_function_many+0xcc/0x2fc
LR [c00a9634] smp_call_function+0x3c/0x50
Call Trace:
[cf82fe90] [00000010] 0x10 (unreliable)
[cf82fed0] [c00a9634] smp_call_function+0x3c/0x50
[cf82fee0] [c0015d2c] flush_tlb_kernel_range+0x20/0x38
[cf82fef0] [c001524c] mark_initmem_nx+0x154/0x16c
[cf82ff20] [c001484c] free_initmem+0x20/0x4c
[cf82ff30] [c000316c] kernel_init+0x1c/0x108
[cf82ff40] [c000f3a8] ret_from_kernel_thread+0x5c/0x64
Instruction dump:
7c0803a6 7d808120 38210040 4e800020 3d20c052 812981a0 2f890000
40beffac
3d20c051 8929ac64 2f890000 40beff9c <0fe00000> 4bffff94 7fc3f378
7f64db78
---[ end trace 7da7bdcf8b15ddb3 ]---
Thanks.
I guess the system still runs OK otherwise, you're just seeing the
warning?
Yes, though I am not sure if that is because there is only one active
CPU (there is
still only one if I say "-smp 4" on the qemu command line).
A complete log is available at:
http://kerneltests.org/builders/qemu-ppc-master/builds/814/steps/qemubuildcommand/logs/stdio
Bisect points to commit 3184cc4b6f6a1dc0 ("powerpc/mm: Fix kernel
RAM protection
after freeing unused memory on PPC32"). Bisect log is attached. A
quick look
suggests that mark_initmem_nx() is called with interrupts disabled,
which
triggers the traceback.
Hmm. Yes the MSR says you have interrupts disabled (EE missing).
But I don't see why. start_kernel() did local_irq_enable(), so I don't
understand why we got to mark_initmem_nx() with them disabled. I'll
hope
that Christophe has some idea.
Good question. I only see this with one of 9 ppc emulations, with
85xx/mpc85xx_cds_defconfig
+CONFIG_DEVTMPFS=y +CONFIG_SMP=y. Maybe there is a platform specific
init function
which leaves interrupts disabled. Question is which one that might be.
Unfortunatly no, I have no idea. My three platforms (860, 885 and
8321) are not SMPs so that warning would not appear, but I added a
WARN_ON(1) just become calling mark_initmem_nx(), and I can confirm
that MSR has EE set on all three at that time.
You should still be able to compile and run a SMP kernel.
mpc85xx_cds_defconfig
SMP doesn't support the 8xx, and the 83xx has hash MMU.
without CONFIG_SMP=y does not show the warning either.
Yes that's normal, as the smp_call_function() is not called in that
case, hence my test with a WARN_ON(1) just before calling mark_initram_nx()
Turns out interrupts are disabled in change_page_attr(), called by
mark_initmem_nx().
Oops, you're right, I missed it.
change_page_attr() calls flush_tlb_kernel_range() with interrupts disabled.
This only happens if CONFIG_PPC_MMU_NOHASH=y.
Given that, I would assume that this will be seen with every 32 bit ppc
build which has
CONFIG_SMP=y and CONFIG_PPC_MMU_NOHASH=y.
Maybe the problem was really introduced with commit e611939fc8ec1
("powerpc/mm: Ensure
change_page_attr() doesn't invalidate pinned TLBs"). From the context it
appears that
flush_tlb_kernel_range() should not be called with interrupts disabled.
Right, it looks like that warning was introduced by this commit.
However, by looking at flush_tlb_page() which was the function that was
called instead before that commit, there was most likely also an issue
with SMP because flush_tlb_page() called with a NULL vma results in a
warning in the SMP NOHASH version of flush_tlb_page().
Indeed, moving flush_tlb_kernel_range() outside the irq disabled code fixes
the problem for me.
Yes that's likely the solution it seems.
Thanks
Christophe
Thanks,
Guenter
So as you suggest, there must be a platform specific stuff leaving the
interrupts disabled.
Christophe
Guenter
