Created attachment 90581 [details]
In kernels 3.7.0 and 3.7.1 gpu sometimes switches in permanent "always-powered on" state after resume from suspend.
This results in increased power consumpution and a noise from fans.
An output from powertop:
| GPU |
| Powered On100.0% |
| RC6 0.0% |
| RC6p 0.0% |
| RC6pp 0.0% |
intel_gpu_top says that there are no programs who use gpu:
render busy: 0%: render space: 0/131072
bitstream busy: 0%: bitstream space: 0/131072
blitter busy: 0%: blitter space: 0/131072
task percent busy
vert fetch: 0 (0/sec)
prim fetch: 0 (0/sec)
VS invocations: 0 (0/sec)
GS invocations: 0 (0/sec)
GS prims: 0 (0/sec)
CL invocations: 0 (0/sec)
CL prims: 0 (0/sec)
PS invocations: 0 (0/sec)
PS depth pass: 0 (0/sec)
I haven't seen this on older kernels(3.6.8, 3.6.6, 3.6.1, 3.6.0, 3.5.2, 3.5.0, 3.4.5, 3.4.0, 3.3.0, 3.2.5, 3.2.0, 3.1.0, 3.0.3)
Some data about the system:
# uname -a
Linux BaysGentooNotebook 3.7.1-gentoo #1 SMP Thu Jan 3 22:16:59 YEKT 2013 x86_64 Intel(R) Core(TM) i5-2410M CPU @ 2.30GHz GenuineIntel GNU/Linux
Vendor: 0x8086, Device: 0x0116, Revision: 0x09 (??)
# dmesg | grep RC6
[ 3388.580819] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
[ 5458.208680] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
[27481.114489] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
[29492.097527] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
[29909.872220] [drm] Enabling RC6 states: RC6 on, RC6p off, RC6pp off
Kernel command line: root=/dev/sda7 rootfstype=ext4 resume=/dev/sda6 rw pcie_aspm=force i915.i915_enable_rc6=1 quiet
Created attachment 90591 [details]
intel_reg_checker output when all is good
Created attachment 90601 [details]
intel_reg_checker output when in permanent "powered on state"
Since this seems rather clear-cut, can you attempt to bisect where this regression has been introduced?
612a9aab56a93533e76e3ad91642db7033e03b69 is the first bad commit
(In reply to comment #4)
> 612a9aab56a93533e76e3ad91642db7033e03b69 is the first bad commit
Hmm, that's a merge commit, not very helpful. :(
Created attachment 90941 [details]
I've attached a patch. It fixes bug.
(In reply to comment #6)
> Created an attachment (id=90941) [details]
> I've attached a patch. It fixes bug.
Interesting. That's effectively a revert of:
Author: Ben Widawsky <email@example.com>
Date: Sat Sep 1 22:59:50 2012 -0700
drm/i915: Never read FORCEWAKE
CC Ben. Any thoughts?
May be, how follows from commit:
- POSTING_READ(ECOBUS); /* something from same cacheline, but !FORCEWAKE */
driver should read something from same cacheline after writing to FORCEWAKE register, but in another place:
- /* gen6_gt_check_fifodbg doubles as the POSTING_READ */
it is said that gen6_gt_check_fifodbg doing POSING_READ, but gen6_gt_check_fifodbg is reading from GTFIFODBG register and it is not on the same cacheline
Created attachment 90971 [details]
Second version of fix patch
Here is a second version of fix patch
Created attachment 90981 [details]
drm/i915: fix FORCEWAKE posting reads
Ahah, seems like you beat me to it, but here's my version anyway.
Hope to see this patch in next kernel versions.
I think the bug can be closed as fixed.
This patch has been submitted to our developer's mailing list. While the patch itself is fairly benign, I am worried that there is another problem here which we are missing.
Personally, I fear the extra read is what's actually fixing the problem. Alexander, could you please try replacing the read of ECOBUS with some other register. GEN6_GT_MODE for instance. If that works, please try a second read of GTFIFODBG. If it doesn't work, can you try some other register in the same CL that isn't ECOBUS or FORCEWAKE?
I'd just like to get as much info into the commit message of the fix before we upstream the patch.
It works good with one read operation of ECOBUS or GEN6_GT_MODE register. Also it works with any two read operations, for example with 2xGTFIFODBG. It doesn't work in the right way with one read operation of GTFIFODBG or FORCEWAKE_VLV.
The part about any two operations was wrong. It works 2xGTFIFODBG, but not works with some other registers.
(In reply to comment #14)
> The part about any two operations was wrong. It works 2xGTFIFODBG, but not
> works with some other registers.
Which registers do not work (that you've tried)?
Please only try registers < 0x40000 as to not require forcewake for those.
I tried to insert just after I915_WRITE_NOTRACE(FORCEWAKE, 0) posting readings of:
ECOBUS + GEN6_GT_MODE (GOOD)
GTFIFODBG + GTFIFODBG (GOOD)
ECOBUS + GTFIFODBG (GOOD)
GEN6_GT_MODE + GTFIFODBG (GOOD)
<SOME REGISTER > 0x40000> <SOME REGISTER > 0x40000> (BAD)
I not remember what was the value of SOME REGISTER exactly but is definately > 0x40000.
I can check any other readings after I915_WRITE_NOTRACE if it helps.
Checked few more:
udelay(1000) and no read (BAD)
FORCEWAKE_VLV makes sense, since this platform is not Valleyview, and won't have that register.
So what have we learned then, GTFIFODBG and registers requiring forcewake are bad?
I wonder if it's not always valid to read GTFIFODBG.
By the way, GTFIFODBG is 0x120000, this is more than 0x40000.
May be my platform doesn't have GTFIFODBG regsiter as well(or the value of GTFIFODBG is wrong)?
Alexander, if it's not too much trouble, please try these registers, one at a time:
BAD means that I managed to catch permanent 100% powered on state for 10-15 minutes when there was no applications who uses gpu.
GOOD - means that I tried for quite long time and bug wasn't occured.
Thanks for the resuts. The registers were all > 0x40000.
(In reply to comment #21)
The above two are GTTMMADR.
This is a CPU reg.
These are both supposed to be PCH regs.
Ben, Daniel, any ideas for explaining this? Or shall I just "think of something" for the commit message...?
(In reply to comment #22)
> Ben, Daniel, any ideas for explaining this? Or shall I just "think of
> something" for the commit message...?
I think just summarize the findings here and if you want to be creative, give this new elephant a cute name.
Author: Jani Nikula <firstname.lastname@example.org>
Date: Thu Jan 17 10:24:09 2013 +0200
drm/i915: fix FORCEWAKE posting reads
Isn't this a dup of bug #48791 ?