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!
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 126.96.36.199) 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
index 58730c5..7fc9f10 100644
@@ -195,9 +195,9 @@
* to cover the pipeline delay.
- mfc0 v1, CP0_TCSTATUS
+ mfc0 v0, CP0_TCSTATUS
- 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)
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
Once I restored previous version of stackframe.h 2.6.33-stable started
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.
On 12/23/10 1:09 PM, STUART VENTERS wrote:
I'm not sure if it's useful,
but finally I got the time to look at the two kernel versions Anoop
works 2.6.32-stable with patch 804
greping for files with CONFIG_MIPS_MT_SMTC
and looking for timer interrupt related stuff found the following
SAVE_SOME SAVE_TEMP get/set_saved_sp
Enclosed are the two subsets of files for a more expert look.
I'll try to look in more detail after Christmas.