| Naveen Krishna Chatradhi | b968279 | 2020-05-19 21:20:10 +0530 | [diff] [blame] | 1 | .. SPDX-License-Identifier: GPL-2.0 |
| 2 | |
| 3 | Kernel driver amd_energy |
| 4 | ========================== |
| 5 | |
| 6 | Supported chips: |
| 7 | |
| 8 | * AMD Family 17h Processors |
| 9 | |
| 10 | Prefix: 'amd_energy' |
| 11 | |
| 12 | Addresses used: RAPL MSRs |
| 13 | |
| 14 | Datasheets: |
| 15 | |
| 16 | - Processor Programming Reference (PPR) for AMD Family 17h Model 01h, Revision B1 Processors |
| 17 | |
| 18 | https://developer.amd.com/wp-content/resources/55570-B1_PUB.zip |
| 19 | |
| 20 | - Preliminary Processor Programming Reference (PPR) for AMD Family 17h Model 31h, Revision B0 Processors |
| 21 | |
| 22 | https://developer.amd.com/wp-content/resources/56176_ppr_Family_17h_Model_71h_B0_pub_Rev_3.06.zip |
| 23 | |
| 24 | Author: Naveen Krishna Chatradhi <nchatrad@amd.com> |
| 25 | |
| 26 | Description |
| 27 | ----------- |
| 28 | |
| 29 | The Energy driver exposes the energy counters that are |
| 30 | reported via the Running Average Power Limit (RAPL) |
| 31 | Model-specific Registers (MSRs) via the hardware monitor |
| 32 | (HWMON) sysfs interface. |
| 33 | |
| 34 | 1. Power, Energy and Time Units |
| 35 | MSR_RAPL_POWER_UNIT/ C001_0299: |
| 36 | shared with all cores in the socket |
| 37 | |
| 38 | 2. Energy consumed by each Core |
| 39 | MSR_CORE_ENERGY_STATUS/ C001_029A: |
| 40 | 32-bitRO, Accumulator, core-level power reporting |
| 41 | |
| 42 | 3. Energy consumed by Socket |
| 43 | MSR_PACKAGE_ENERGY_STATUS/ C001_029B: |
| 44 | 32-bitRO, Accumulator, socket-level power reporting, |
| 45 | shared with all cores in socket |
| 46 | |
| 47 | These registers are updated every 1ms and cleared on |
| 48 | reset of the system. |
| 49 | |
| 50 | Note: If SMT is enabled, Linux enumerates all threads as cpus. |
| 51 | Since, the energy status registers are accessed at core level, |
| 52 | reading those registers from the sibling threads would result |
| 53 | in duplicate values. Hence, energy counter entries are not |
| 54 | populated for the siblings. |
| 55 | |
| 56 | Energy Caluclation |
| 57 | ------------------ |
| 58 | |
| 59 | Energy information (in Joules) is based on the multiplier, |
| 60 | 1/2^ESU; where ESU is an unsigned integer read from |
| 61 | MSR_RAPL_POWER_UNIT register. Default value is 10000b, |
| 62 | indicating energy status unit is 15.3 micro-Joules increment. |
| 63 | |
| 64 | Reported values are scaled as per the formula |
| 65 | |
| 66 | scaled value = ((1/2^ESU) * (Raw value) * 1000000UL) in uJoules |
| 67 | |
| 68 | Users calculate power for a given domain by calculating |
| 69 | dEnergy/dTime for that domain. |
| 70 | |
| 71 | Energy accumulation |
| 72 | -------------------------- |
| 73 | |
| 74 | Current, Socket energy status register is 32bit, assuming a 240W |
| 75 | 2P system, the register would wrap around in |
| 76 | |
| 77 | 2^32*15.3 e-6/240 * 2 = 547.60833024 secs to wrap(~9 mins) |
| 78 | |
| 79 | The Core energy register may wrap around after several days. |
| 80 | |
| 81 | To improve the wrap around time, a kernel thread is implemented |
| 82 | to accumulate the socket energy counters and one core energy counter |
| 83 | per run to a respective 64-bit counter. The kernel thread starts |
| 84 | running during probe, wakes up every 100secs and stops running |
| 85 | when driver is removed. |
| 86 | |
| 87 | A socket and core energy read would return the current register |
| 88 | value added to the respective energy accumulator. |
| 89 | |
| 90 | Sysfs attributes |
| 91 | ---------------- |
| 92 | |
| 93 | =============== ======== ===================================== |
| 94 | Attribute Label Description |
| 95 | =============== ======== ===================================== |
| 96 | |
| 97 | * For index N between [1] and [nr_cpus] |
| 98 | |
| 99 | =============== ======== ====================================== |
| 100 | energy[N]_input EcoreX Core Energy X = [0] to [nr_cpus - 1] |
| 101 | Measured input core energy |
| 102 | =============== ======== ====================================== |
| 103 | |
| 104 | * For N between [nr_cpus] and [nr_cpus + nr_socks] |
| 105 | |
| 106 | =============== ======== ====================================== |
| 107 | energy[N]_input EsocketX Socket Energy X = [0] to [nr_socks -1] |
| 108 | Measured input socket energy |
| 109 | =============== ======== ====================================== |