Amd10Cpu freeze

[Rem0o]

Just to report that the AMD10Cpu LHM refactor seems to have freezing issues, all FX/bulldozers processor reporting the same thing:
Rem0o/FanControl.Releases#3521 (comment)

[PhyxionNL]

I've temporarily disabled this until this is sorted :)
LibreHardwareMonitor/LibreHardwareMonitor@df2c5eb

[Serial95]

If that can helps to identify the problem, I have the same issue on an A8-7600 (Steamroller/Kaveri, a Bulldozer refresh according to a wiki).

The computer completely freezes. When I restarts it, it's like a sudden shutdown, no errors or BSOD appearing in the event log (except the sudden shutdown itself).

[LightingBasket4959]

I have the same issue as described above on an FX-6300 (Piledriver)

[eberhardweber]

Reporting immediate freeze on my FX CPUs, too, both with FX-8350 and FX-9590, on two separate motherboards, Gigabyte GA-990XA-UD3 (rev. 3.0) and Asus M5A99FX PRO R2.0.

[irusanov]

I can reproduce it and pinpoint it to this ioctl ioctl_measure_tsc_multiplier

PawnIO.Modules/AMDFamily10.p

Line 166 in f216654

DEFINE_IOCTL_SIZED(ioctl_measure_tsc_multiplier, 0, 2) {

I guess some of the MSR writes locks the system (potentially restoring hwcr msr and p-states related which could make the cpu boost without proper vid adjustment). K10 (except Thuban) has no CPB and maybe that's why it works on Athlon II.

Everything else works.

It will take time to debug it fully as the system is really slow. My board is in costant PROCHOT state and I have to disable it with some registers manipulation every boot. Also using nvme through usb adapter.

[irusanov]

@namazso, @Rem0o The initial single tick get seems to work.

new tick_old = get_tick_count();

it seems to be locking on the two while loops:

PawnIO.Modules/AMDFamily10.p

Line 195 in f216654

while (get_tick_count() < tick_start) {}

PawnIO.Modules/AMDFamily10.p

Line 202 in f216654

while (get_tick_count() < tick_end) {}

Additionally I've added a second write here to enable the counter as per 15h BKDG, but not sure it makes a difference, however the last vesion of the module I tested includes this change. I won't be able to test more in the following week or so.

    new event_select = (1 << 16) | 0x76;
    msr_write(MSR_K7_EVNTSEL0, (0 << 22) | event_select);
    msr_write(MSR_K7_PERFCTR0, 0);
    msr_write(MSR_K7_EVNTSEL0, (1 << 22) | event_select);

[namazso]

I think this is actually a really stupid mistake: I'm waiting on the tick count to be increased by the timer interrupt... The one I disabled about 10 lines before.

[irusanov]

hmm, now when you wrote about it, that's probably the reason. Adding couple more interrupts disable and enable around the msr operations might do it. Unfortunately don't know if I will have time to test tomorrow and then I won't have access to this system for a while.

[namazso]

I added the interrupt disabling to ensure we don't get descheduled which would mess up the measurements, which is a bug the original LHM version had. I should probably figure out what exactly the code is meant to do, and maybe try to find a better way to do it.

[irusanov]

I've used TSC directly with RDTSC instruction in the past, but I think it might be skewed in certain conditions if the base clock is changed in OS. Maybe this implementation addresses it, but I'm not sure.

[namazso]

Yeah, it seems like it's meant to measure something different than the TSC frequency
https://github.com/LibreHardwareMonitor/LibreHardwareMonitor/blob/master/LibreHardwareMonitorLib/Hardware/Cpu/Amd10Cpu.cs#L115 in reality it is multiplied by tick frequency for use. That means we could probably do a shorter interval and just use microsleep2 and QueryPerformanceCounter to derive the same number. Those work on the highest IRQL as well.

[irusanov]

@namazso I can confirm the sample program runs fine on a FX-6300 cpu and Asus Crosshair V Formula motherboard. Seems to produce a correct frequency.

CPU: AMD64 Family 21 Model 2 Stepping 0, AuthenticAMD
Measuring TSC frequency
TSC Frequency: 4515.05 MHz (rounded: 4500 MHz)

Unfortunately I am not able to compile librehardwaremonitor at the moment, can't figure out why. It used to compile before. It seems to be something with CsWin32 as the imports VS is complaining about appear to be from this package, but even downgrading it didn't fix the problem.

[namazso]

That doesn't seem complete, the full program is supposed to be this:

https://gist.github.com/namazso/42d7daa931114832b1f3d72d2c77e37e

Does it get stuck, or crashes?

[irusanov]

Sorry, didn't notice test signing was off for some reason

CPU: AMD64 Family 21 Model 2 Stepping 0, AuthenticAMD
Measuring TSC frequency
TSC Frequency: 4515.05 MHz (rounded: 4500 MHz)
Opened handle: 00000000000000A8
Loaded AMDFamily10.bin
Measured info:
  COFVID: 0x20000063E02361D
  Control TSC Delta: 7181
  Control Event Counter: 2960
  Measured TSC Delta: 9929
  Measured Event Counter: 5707
Estimated clock frequency: 4498.36 MHz (rounded: 4500 MHz)
Everything done!

PS: Also managed to compile LHM by checking out older commit before CsWin32 version bump. Replaced AmdFamily10h.bin with a compiled binary from source and enabled it to load in the library - seems to work, except LHM needs to adapt the code to use the new module and show correct frequency.

[irusanov]

Changing the multiplier (reducing to 21.5x from 22.5x) in Windows does not produce correct results (TSC Frequency is matching the current clock):

CPU: AMD64 Family 21 Model 2 Stepping 0, AuthenticAMD
Measuring TSC frequency
TSC Frequency: 4314.38 MHz (rounded: 4300 MHz)
Opened handle: 00000000000000AC
Loaded AMDFamily10.bin
Measured info:
  COFVID: 0x28000063E02361B
  Control TSC Delta: 6586
  Control Event Counter: 2897
  Measured TSC Delta: 9346
  Measured Event Counter: 6114
Estimated clock frequency: 5011.99 MHz (rounded: 5000 MHz)
Everything done!

increased to 23.5x

CPU: AMD64 Family 21 Model 2 Stepping 0, AuthenticAMD
Measuring TSC frequency
TSC Frequency: 4715.73 MHz (rounded: 4700 MHz)
Opened handle: 00000000000000AC
Loaded AMDFamily10.bin
Measured info:
  COFVID: 0x28000063E02361F
  Control TSC Delta: 5262
  Control Event Counter: 2318
  Measured TSC Delta: 7611
  Measured Event Counter: 4756
Estimated clock frequency: 4878.08 MHz (rounded: 4900 MHz)
Everything done!

22.5x and bclk increased to 204:

CPU: AMD64 Family 21 Model 2 Stepping 0, AuthenticAMD
Measuring TSC frequency
TSC Frequency: 4605.35 MHz (rounded: 4600 MHz)
Opened handle: 00000000000000AC
Loaded AMDFamily10.bin
Measured info:
  COFVID: 0x28000063E02361D
  Control TSC Delta: 5179
  Control Event Counter: 2265
  Measured TSC Delta: 7544
  Measured Event Counter: 4775
Estimated clock frequency: 4882.03 MHz (rounded: 4900 MHz)
Everything done!

My cpu is currently running in forced P0 mode as otherwise the motherboard is stuck in PROCHOT condition, so I have to forcibly clear the status and disable HTC.