I'm trying to figure out a reason why your change below should help, and
offhand, modulo tool bugs, I don't see it. I'm assuming that your diff
below is a diff relative to the pre-patch stackframe.h. I wouldn't
bless it as an alternative because it moves code and comments
unnecessarily - all you should really have to do is to move the
190 mfc0 v1, CP0_STATUS
191 LONG_S $2, PT_R2(sp)
to be just after the #endif /* CONFIG_MIPS_MT_SMTC */ at around line 201.
If moving the save of zero to PT_R0(sp) actually makes a difference,
it's evidence that you've got problems in your toolchain (or, heaven
forbid, your pipeline)!
But I'd really like to see what your assembler is doing to the original
patch for it to be broken. Assembler instruction reordering is armed,
but it ought not to move register moves and stores around in ways where
your sequence
197 .set mips32
198 mfc0 v1, CP0_TCSTATUS
199 .set mips0
200 LONG_S v1, PT_TCSTATUS(sp)
189 LONG_S $0, PT_R0(sp)
190 mfc0 v1, CP0_STATUS
191 LONG_S $2, PT_R2(sp)
202 LONG_S $4, PT_R4(sp)
203 LONG_S $5, PT_R5(sp)
204 LONG_S v1, PT_STATUS(sp)
to work while
189 LONG_S $0, PT_R0(sp)
190 mfc0 v1, CP0_STATUS
191 LONG_S $2, PT_R2(sp)
197 .set mips32
198 mfc0 v0, CP0_TCSTATUS
199 .set mips0
200 LONG_S v0, PT_TCSTATUS(sp)
202 LONG_S $4, PT_R4(sp)
203 LONG_S $5, PT_R5(sp)
204 LONG_S v1, PT_STATUS(sp)
does not, provided that the identity of v0=$2, v1=$3 is respected.
One thing that does stick out as being different - though, again, I'd
need to see the disassembly of an instance of the macro to know what it
could have done - is that the SMTC conditiona code brackets the mfc0 of
TCStatus with .set mips32/.set mips0. Given that the code no longer has
a .set mips0 early in the macro, it would be more correct to make it:
.set push
.set mips32
mfc0 v0, CP0_TCSTATUS (or v1, if we move the mfc0
v1,CP0_STATUS)
.set pop
and presumably make a similar chage for the block from line 334 to 429.
But I don't see any causal path from that funniness to failure.
Regards,
Kevin K.
On 01/04/11 06:37, Anoop P A wrote:
Hi Kevin,
the stackframe patch that you have suggested had some side effects I was
unable execute init. When I changed some thing like below it started
working .Could you kindly review it ?.
diff --git a/arch/mips/include/asm/stackframe.h
b/arch/mips/include/asm/stackframe.h
index 58730c5..da786ed 100644
--- a/arch/mips/include/asm/stackframe.h
+++ b/arch/mips/include/asm/stackframe.h
@@ -181,14 +181,6 @@
#endif
LONG_S k0, PT_R29(sp)
LONG_S $3, PT_R3(sp)
- /*
- * You might think that you don't need to save $0,
- * but the FPU emulator and gdb remote debug stub
- * need it to operate correctly
- */
- LONG_S $0, PT_R0(sp)
- mfc0 v1, CP0_STATUS
- LONG_S $2, PT_R2(sp)
#ifdef CONFIG_MIPS_MT_SMTC
/*
* Ideally, these instructions would be shuffled in
@@ -199,6 +191,14 @@
.set mips0
LONG_S v1, PT_TCSTATUS(sp)
#endif /* CONFIG_MIPS_MT_SMTC */
+ /*
+ * You might think that you don't need to save $0,
+ * but the FPU emulator and gdb remote debug stub
+ * need it to operate correctly
+ */
+ LONG_S $0, PT_R0(sp)
+ mfc0 v1, CP0_STATUS
+ LONG_S $2, PT_R2(sp)
LONG_S $4, PT_R4(sp)
LONG_S $5, PT_R5(sp)
LONG_S v1, PT_STATUS(sp)
Linux-2.6.37-rc7 boots all the way if I specify maxvpes=1 in command
line.
/ # cat /proc/interrupts
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6
1: 249 218024 218286 218263 218235 218208
218179 MIPS SMTC_IPI
6: 0 0 0 0 0 0
0 MIPS MSP CIC cascade
8: 0 0 0 0 0 0
0 MSP_CIC Softreset button
9: 0 0 0 0 0 0
0 MSP_CIC Standby switch
21: 0 0 0 0 0 0
0 MSP_CIC MSP PER cascade
25: 218128 711 11 0 0 0
0 MSP_CIC timer
27: 341 22 0 0 2 0
6 MSP_CIC serial
ERR: 0
/ # uname -a
Linux (none) 2.6.37-rc7-pmc-00001-g9cff2d6-dirty #289 SMP PREEMPT Tue
Jan 4 19:48:31 IST 2011 mips GNU/Linux
So clock setup / distribution on VPE1 is some thing need fix.
Thanks
Anoop
On Tue, 2011-01-04 at 18:32 +0530, Anoop P A wrote:
On Tue, 2011-01-04 at 00:17 -0800, Kevin D. Kissell wrote:
Those interrupt counters show that IPIs are being taken everywhere,
though very few by CPUs 5 and 6. If I understand the configuration
correctly, CPU 4 is a TC in VPE 1, and it's getting a reasonable IPI
Yes CPU4 is in second VPE
rate, *if* we're looking at a tickless kernel under low load. But there
No it was not the tickless kernel.I had selected 250 MHz timer. can't we
expect IPI / timer interrupt for all the threads in this case ?.
may be a clue there to part of your problem. I have no idea why the
behavior would have changed from 2.6.36 to 2.6.37, but it looks as if
you're getting your clock interrupts through the MSP CIC interrupt
controller on VPE 0. There's nothing symmetric for VPE1. The Malta
example code is perhaps deceptively simple, in that both VPEs have their
count/compare indication wired directly to the 2 clock interrupt inputs,
so that having both of them running with only a single set of irq state
just works. I don't know whether the MSP CIC timer interrupt is a
In my case it is separate irq. MSP_INT_VPE1_TIMER (34) and
MSP_INT_VPE0_TIMER (25) are wired to CIC . CIC interrupt has been
connected to cpu irq 6.
I can reproduce cpu stall in VSMP mode If I don't setup VPE1 timer
interrupt . Don't we have support for separate irq in SMTC
implementation ?..
gating of the VPE0 count/compare output, or whether it's it's own
interval timer, but I suspect that you may need to do some further
low-level initialization in the platform-specific code to set up an
interrupt on the VPE1 side. I don't think the snippet you've got below
would work as written.
The routine which I copied works fine for VSMP mode .
/ # cat /proc/interrupts
CPU0 CPU1
0: 187 254 MIPS IPI_resched
1: 77 174 MIPS IPI_call
6: 0 0 MIPS MSP CIC cascade
8: 0 0 MSP_CIC Softreset button
9: 0 0 MSP_CIC Standby switch
21: 0 0 MSP_CIC MSP PER cascade
25: 37077 0 MSP_CIC timer
27: 188 0 MSP_CIC serial
34: 0 36986 MSP_CIC timer
Do I want to change anything specific for SMTC ? .
If it's purely an issue with clock distribution on VPE1, then a boot
with maxvpes=1 maxtcs=4 should be stable.
Yes the kernel seems to be stable if I boot with maxvpes=1 maxtcs=4 .
/K.
On 1/3/2011 11:20 AM, Anoop P A wrote:
Hi Kevin,
On Mon, 2011-01-03 at 08:14 -0800, Kevin D. Kissell wrote:
The very first SMTC implementations didn't support full kernel-mode
preemption, which anyway wasn't a priority, given the hardware event
response support in MIPS MT. I believe it was later made compatible,
but it was never extensively exercised. Since SMTC has fingers in some
pretty low-level atomicity mechanisms, if a new, parallel set was
implemented for RCU, I can easily imagine that nobody has yet
implemented SMTC-ified variants of that set.
Your last statement isn't very clear, though. Are you saying that if
you configure for no forced preemption and with TREE_CPU, the 2.6.37
kernel boots all the way up, or that it simply hangs later? What's the
last rev kernel that actually boots all the way up?
I have debugged this a bit more. It seems that kernel getting stalled
while executing on TC's of second VPE .
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=2504 jiffies)
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=10036 jiffies)
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=17568 jiffies)
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=25100 jiffies)
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=32632 jiffies)
INFO: rcu_sched_state detected stalls on CPUs/tasks: { 4 5 6} (detected
by 1, t=40164 jiffies)
With CONFIG_TREE_CPU we were not hitting this scenario very often.
However with CONFIG_PREEMPT_TREE_CPU stall happens most of the time.
I presume some issue in my timer setup . I am not seeing timer interrupt
(or IPI interrupt) getting incremented for VPE1 tcs on a completely
booted 2.6.32-stable kernel.
/ # cat /proc/interrupts
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6
1: 148 15023 15140 15093 3779 8
2 MIPS SMTC_IPI
6: 0 0 0 0 0 0
0 MIPS MSP CIC cascade
8: 0 0 0 0 0 0
0 MSP_CIC Softreset button
9: 0 0 0 0 0 0
0 MSP_CIC Standby switch
21: 0 0 0 0 0 0
0 MSP_CIC MSP PER cascade
25: 15113 341 4 7 0 0
0 MSP_CIC timer
27: 260 9 0 1 0 0
0 MSP_CIC serial
34: 0 0 0 0 0 0
0 MSP_CIC timer
Can't we use separate timer interrupts for VPE1 and VPE0 in SMTC ?.
I have tried setting up VPE1 timer from get_co_compare_int as follows
unsigned int __cpuinit get_c0_compare_int(void)
{
if ((1==get_current_vpe())&& !vpe1_timr_installed){
memcpy(&timer_vpe1,&c0_compare_irqaction,sizeof(timer_vpe1));
setup_irq(MSP_INT_VPE1_TIMER,&timer_vpe1);
vpe1_timr_installed++;
}
return (get_current_vpe() ? MSP_INT_VPE1_TIMER :
MSP_INT_VPE0_TIMER);
}
Thanks
Anoop
Regards,
Kevin K.
On 1/3/2011 7:12 AM, Anoop P A wrote:
Hi ,
Following patch restricts TREE_CPU RCU implementation only for !PREEMPT
SMP kernel.
http://git.linux-mips.org/?p=linux.git;a=commit;h=687d7a960aea46e016182c7ce346d62c4dbd0366
CONFIG_TREE_PREEMPT_RCU option seems to be not working for SMTC kernel
( which will be only available RCU implementation for SMTC kernel from
2.6.37 onwards) .
With no forced preemption and selecting TREE_CPU I am able to boot
further to the hang that I have reported.
Thanks
Anoop
On Sat, 2011-01-01 at 00:42 -0800, Kevin D. Kissell wrote:
At this point the logical thing to do would seem to look at your kernel
image and disassemble smtc_ipi_replay(), which is where the EPC of VPE 0
shows the last exception to have been taken. That's a critical SMTC
routine that gets called whenever an xxx_irq_restore() enables
interrupts, so that virtual per-TC IPI interrupts that were posted while
the TC had interrupts disabled can be handled deterministically. As I
mentioned in an earlier message, there was some cleanup work from David
Howell that changed a number of irq management-related function names
and prototypes across all architectures, which went into linux-mips.org
at very roughly the time of the breakage. The SMTC overlay over the irq
implementation has been pretty robust, but it's written in a perhaps
doomed attempt to be both efficient and using a maximum amount of common
code with the general case. A mechanical or semi-mechanical change
could conceivably have broken things.
Regards,
Kevin K.
On 12/31/2010 4:27 AM, Anoop P A wrote:
Hi ,
Kernel hangs on stop_machine call. Please find mt reg dump below.
Another important observation is even though 2.6.33 kernel + stackframe
patch well passes calibration hang , I am still unable boot in to a
initramfs root ( verified ramfs working with VSMP). So it looks like
still some issue to fix between 2.6.32 and 2.6.33 .
######################## Log ###########################
=== MIPS MT State Dump ===
-- Global State --
MVPControl Passed: 00000005
MVPControl Read: 00000004
MVPConf0 : a8008406
-- per-VPE State --
VPE 0
VPEControl : 00008000
VPEConf0 : 800f0003
VPE0.Status : 11004201
VPE0.EPC : 8010dc54 smtc_ipi_replay+0xcc/0x108
VPE0.Cause : 50804000
VPE0.Config7 : 00010000
VPE 1
VPEControl : 00068006
VPEConf0 : 80cf0003
VPE1.Status : 11008301
VPE1.EPC : 801022a0 r4k_wait+0x20/0x40
VPE1.Cause : 50800000
VPE1.Config7 : 00010000
-- per-TC State --
TC 0 (current TC with VPE EPC above)
TCStatus : 18102000
TCBind : 00000000
TCRestart : 803fa19c printk+0xc/0x30
TCHalt : 00000000
TCContext : 00000000
TC 1
TCStatus : 18902000
TCBind : 00200000
TCRestart : 801022a0 r4k_wait+0x20/0x40
TCHalt : 00000000
TCContext : 00140000
TC 2
TCStatus : 18902000
TCBind : 00400000
TCRestart : 801022a0 r4k_wait+0x20/0x40
TCHalt : 00000000
TCContext : 00280000
TC 3
TCStatus : 18902000
TCBind : 00600000
TCRestart : 801022a0 r4k_wait+0x20/0x40
TCHalt : 00000000
TCContext : 003c0000
TC 4
TCStatus : 18902000
TCBind : 00800001
TCRestart : 8010229c r4k_wait+0x1c/0x40
TCHalt : 00000000
TCContext : 00500000
TC 5
TCStatus : 18902000
TCBind : 00a00001
TCRestart : 8010229c r4k_wait+0x1c/0x40
TCHalt : 00000000
TCContext : 00640000
TC 6
TCStatus : 18902000
TCBind : 00c00001
TCRestart : 8010229c r4k_wait+0x1c/0x40
TCHalt : 00000000
TCContext : 00780000
Counter Interrupts taken per CPU (TC)
0: 0
1: 0
2: 0
3: 0
4: 0
5: 0
6: 0
7: 0
Self-IPI invocations:
0: 12
1: 0
2: 0
3: 0
4: 0
5: 5
6: 4
7: 0
IPIQ[0]: head = 0x0, tail = 0x0, depth = 0
IPIQ[1]: head = 0x0, tail = 0x0, depth = 0
IPIQ[2]: head = 0x0, tail = 0x0, depth = 0
IPIQ[3]: head = 0x0, tail = 0x0, depth = 0
IPIQ[4]: head = 0x0, tail = 0x0, depth = 0
IPIQ[5]: head = 0x0, tail = 0x0, depth = 0
IPIQ[6]: head = 0x0, tail = 0x0, depth = 0
IPIQ[7]: head = 0x0, tail = 0x0, depth = 0
0 Recoveries of "stolen" FPU
===========================
################################################################
Thanks
Anoop
On Tue, 2010-12-28 at 00:43 -0800, Kevin D. Kissell wrote:
I took a quick look last night, and the only thing that looked vaguely
dangerous in changes since the timer changes I alluded to earlier was
the global naming cleanup of irq-related function names that David
Howell submitted. The diff didn't look dangerous in itself, but some of
the definitions are nested subtly for SMTC to maximize the amount of
common code, and I could imagine something getting lost in translation
there. If that were really the problem, it would of course affect much
more than just the timer subsystem, but early in the boot process,
timers are pretty much the only interrupts that have to be handled
correctly.
I'm travelling today, but will take a look at timekeeping_notify()
tomorrow or the next day...
/K.
On 12/28/10 12:19 AM, Anoop P A wrote:
Hi,
I had a glance into the code diff without notice of any suspect-able
code .
Tracing the hang showed that it is getting hanged in timekeeping_notify
function.
Thanks,
Anoop
PS: I may not be available until Thursday
On Mon, 2010-12-27 at 22:49 +0530, Anoop P A wrote:
Hi Kevin,
It is very unlikely that the patch you pointed has any impact on the the
hang I am seeing. The patch you have mentioned got into kernel around
2.6.32 timeframe. I am able to boot both 2.6.32 and 2.6.33 kernel ( +
stackframe patch) .
Hi Stuart,
I haven't got much time to spend on this today.
I had got 2.6.36-stable(+ stack frame patch) booting last day and I have
observed hang issue with 2.6.37-rc1 ( Same as rc6 and current git head)
So probably some patches in 2.6.37 branch introduced this hang.
Hopefully I will get some free slot tomorrow so that I can look into
code diff .
Thanks
Anoop
On Mon, 2010-12-27 at 09:49 -0600, STUART VENTERS wrote:
Kevin,
Outstanding, sometimes it's better to be lucky than good.
Anoop,
Maybe we can get lucky again.
If you can isolate the .33 works/.37 works_not bug to a specific pair of versions,
I'll be happy to do another diff.
Hope you'll have had a good Christmas as well.
We've had snow in Alabama since Christmas eve!
Regards,
Stuart
-----Original Message-----
From: Kevin D. Kissell [mailto:kevink@xxxxxxxxxxxxx]
Sent: Friday, December 24, 2010 5:34 PM
To: Anoop P A
Cc: STUART VENTERS; Anoop P.A.; linux-mips@xxxxxxxxxxxxxx
Subject: Re: SMTC support status in latest git head.
Ah, well, at least we have a stackframe.h fix that preserves David's
performance tweak for the deeper pipelined processors. In looking for
this, I did notice that someone did some modification to the SMTC clock
tick logic that I was skeptical had ever been tested. If you've still
got that kernel binary handy, you might check to see if it boots with
maxtcs=1 maxvpes=1, maxtcs=2 maxvpes=1, and/or maxtcs=2 maxvpes=2.
Oh, yes, and Merry Christmas one and all!
Regards,
Kevin K.
On 12/24/10 8:02 AM, Anoop P A wrote:
On Fri, 2010-12-24 at 06:53 -0800, Kevin D. Kissell wrote:
Excellent! Now, does the attached patch (relative to 2.6.37.11) also
fix things, while preserving the other fixes and performance enhancements?
I have tested that patch with 2.6.37 branch it well passes calibration
loop but hangs after switching to mips closource
TC 6 going on-line as CPU 6
Brought up 7 CPUs
bio: create slab<bio-0> at 0
SCSI subsystem initialized
Switching to clocksource MIPS
I Presume this is a different issue as restoring older file didn't help
much to get rid of this hang.
diff --git a/arch/mips/include/asm/stackframe.h
b/arch/mips/include/asm/stackframe.h
index 58730c5..7fc9f10 100644
--- a/arch/mips/include/asm/stackframe.h
+++ b/arch/mips/include/asm/stackframe.h
@@ -195,9 +195,9 @@
* to cover the pipeline delay.
*/
.set mips32
- mfc0 v1, CP0_TCSTATUS
+ mfc0 v0, CP0_TCSTATUS
.set mips0
- LONG_S v1, PT_TCSTATUS(sp)
+ LONG_S v0, PT_TCSTATUS(sp)
#endif /* CONFIG_MIPS_MT_SMTC */
LONG_S $4, PT_R4(sp)
LONG_S $5, PT_R5(sp)
/K.
On 12/24/10 6:39 AM, Anoop P A wrote:
Hi Kevin, Stuart ,
Woohooo You guys spotted !.
http://git.linux-mips.org/?p=linux.git;a=commit;h=d5ec6e3c seems to be
the culprit
Once I restored previous version of stackframe.h 2.6.33-stable started
booting !.
Thanks,
Anoop
On Fri, 2010-12-24 at 04:32 -0800, Kevin D. Kissell wrote:
Thank you, Stuart! I've spotted some definite breakage to SMTC between
those versions. In arch/mips/include/asm/stackframe.h, someone moved
the store of the Status register value in SAVE_SOME (line 169 or 204,
depending on the version) from two instructions after the mfc0 to a
point after the #ifdef for SMTC, presumably to get better pipelining of
the register access. Unfortunately, the v1 register is also used in the
SMTC-specific fragment to save TCStatus, so the Status value gets
clobbered before it gets stored. This will eventually result in the
Status register getting a TCStatus value, which has some bits on common,
but isn't identical and sooner or later Bad Things will happen.
I'm a little surprised this wasn't caught by visual inspection of the patch.
Possible solutions would include reverting the store of the CP0_STATUS
value to the block above the #ifdef, or, to retain whatever performance
advantage was obtained by moving the store downward, to use v0/$2
instead of v1/$3, as the staging register for the TCStatus value. I'd
lean toward the second option, but I'm not in a position to test and
submit a patch just now.
Regards,
Kevin K.
On 12/23/10 1:09 PM, STUART VENTERS wrote:
Kevin,
I'm not sure if it's useful,
but finally I got the time to look at the two kernel versions Anoop pointed out.
works 2.6.32-stable with patch 804
works_not 2.6.33-stable
greping for files with CONFIG_MIPS_MT_SMTC
and looking for timer interrupt related stuff found the following differences:
arch/mips/include/asm/irq.h
arch/mips/kernel/irq.c
do_IRQ
arch/mips/include/asm/stackframe.h
SAVE_SOME SAVE_TEMP get/set_saved_sp
arch/mips/include/asm/time.h
clocksource_set_clock
arch/mips/kernel/process.c
cpu_idle
arch/mips/kernel/smtc.c
__irq_entry
ipi_decode
SMTC_CLOCK_TICK
Enclosed are the two subsets of files for a more expert look.
I'll try to look in more detail after Christmas.
Cheers,
Stuart
