| 1 | // SPDX-License-Identifier: GPL-2.0 | 
|---|
| 2 | /* | 
|---|
| 3 | * TSC frequency enumeration via MSR | 
|---|
| 4 | * | 
|---|
| 5 | * Copyright (C) 2013, 2018 Intel Corporation | 
|---|
| 6 | * Author: Bin Gao <bin.gao@intel.com> | 
|---|
| 7 | */ | 
|---|
| 8 |  | 
|---|
| 9 | #include <linux/kernel.h> | 
|---|
| 10 | #include <linux/thread_info.h> | 
|---|
| 11 |  | 
|---|
| 12 | #include <asm/apic.h> | 
|---|
| 13 | #include <asm/cpu_device_id.h> | 
|---|
| 14 | #include <asm/intel-family.h> | 
|---|
| 15 | #include <asm/msr.h> | 
|---|
| 16 | #include <asm/param.h> | 
|---|
| 17 | #include <asm/tsc.h> | 
|---|
| 18 |  | 
|---|
| 19 | #define MAX_NUM_FREQS	16 /* 4 bits to select the frequency */ | 
|---|
| 20 |  | 
|---|
| 21 | /* | 
|---|
| 22 | * The frequency numbers in the SDM are e.g. 83.3 MHz, which does not contain a | 
|---|
| 23 | * lot of accuracy which leads to clock drift. As far as we know Bay Trail SoCs | 
|---|
| 24 | * use a 25 MHz crystal and Cherry Trail uses a 19.2 MHz crystal, the crystal | 
|---|
| 25 | * is the source clk for a root PLL which outputs 1600 and 100 MHz. It is | 
|---|
| 26 | * unclear if the root PLL outputs are used directly by the CPU clock PLL or | 
|---|
| 27 | * if there is another PLL in between. | 
|---|
| 28 | * This does not matter though, we can model the chain of PLLs as a single PLL | 
|---|
| 29 | * with a quotient equal to the quotients of all PLLs in the chain multiplied. | 
|---|
| 30 | * So we can create a simplified model of the CPU clock setup using a reference | 
|---|
| 31 | * clock of 100 MHz plus a quotient which gets us as close to the frequency | 
|---|
| 32 | * from the SDM as possible. | 
|---|
| 33 | * For the 83.3 MHz example from above this would give us 100 MHz * 5 / 6 = | 
|---|
| 34 | * 83 and 1/3 MHz, which matches exactly what has been measured on actual hw. | 
|---|
| 35 | */ | 
|---|
| 36 | #define TSC_REFERENCE_KHZ 100000 | 
|---|
| 37 |  | 
|---|
| 38 | struct muldiv { | 
|---|
| 39 | u32 multiplier; | 
|---|
| 40 | u32 divider; | 
|---|
| 41 | }; | 
|---|
| 42 |  | 
|---|
| 43 | /* | 
|---|
| 44 | * If MSR_PERF_STAT[31] is set, the maximum resolved bus ratio can be | 
|---|
| 45 | * read in MSR_PLATFORM_ID[12:8], otherwise in MSR_PERF_STAT[44:40]. | 
|---|
| 46 | * Unfortunately some Intel Atom SoCs aren't quite compliant to this, | 
|---|
| 47 | * so we need manually differentiate SoC families. This is what the | 
|---|
| 48 | * field use_msr_plat does. | 
|---|
| 49 | */ | 
|---|
| 50 | struct freq_desc { | 
|---|
| 51 | bool use_msr_plat; | 
|---|
| 52 | struct muldiv muldiv[MAX_NUM_FREQS]; | 
|---|
| 53 | /* | 
|---|
| 54 | * Some CPU frequencies in the SDM do not map to known PLL freqs, in | 
|---|
| 55 | * that case the muldiv array is empty and the freqs array is used. | 
|---|
| 56 | */ | 
|---|
| 57 | u32 freqs[MAX_NUM_FREQS]; | 
|---|
| 58 | u32 mask; | 
|---|
| 59 | }; | 
|---|
| 60 |  | 
|---|
| 61 | /* | 
|---|
| 62 | * Penwell and Clovertrail use spread spectrum clock, | 
|---|
| 63 | * so the freq number is not exactly the same as reported | 
|---|
| 64 | * by MSR based on SDM. | 
|---|
| 65 | */ | 
|---|
| 66 | static const struct freq_desc freq_desc_pnw = { | 
|---|
| 67 | .use_msr_plat = false, | 
|---|
| 68 | .freqs = { 0, 0, 0, 0, 0, 99840, 0, 83200 }, | 
|---|
| 69 | .mask = 0x07, | 
|---|
| 70 | }; | 
|---|
| 71 |  | 
|---|
| 72 | static const struct freq_desc freq_desc_clv = { | 
|---|
| 73 | .use_msr_plat = false, | 
|---|
| 74 | .freqs = { 0, 133200, 0, 0, 0, 99840, 0, 83200 }, | 
|---|
| 75 | .mask = 0x07, | 
|---|
| 76 | }; | 
|---|
| 77 |  | 
|---|
| 78 | /* | 
|---|
| 79 | * Bay Trail SDM MSR_FSB_FREQ frequencies simplified PLL model: | 
|---|
| 80 | *  000:   100 *  5 /  6  =  83.3333 MHz | 
|---|
| 81 | *  001:   100 *  1 /  1  = 100.0000 MHz | 
|---|
| 82 | *  010:   100 *  4 /  3  = 133.3333 MHz | 
|---|
| 83 | *  011:   100 *  7 /  6  = 116.6667 MHz | 
|---|
| 84 | *  100:   100 *  4 /  5  =  80.0000 MHz | 
|---|
| 85 | */ | 
|---|
| 86 | static const struct freq_desc freq_desc_byt = { | 
|---|
| 87 | .use_msr_plat = true, | 
|---|
| 88 | .muldiv = { { 5, 6 }, { 1, 1 }, { 4, 3 }, { 7, 6 }, | 
|---|
| 89 | { 4, 5 } }, | 
|---|
| 90 | .mask = 0x07, | 
|---|
| 91 | }; | 
|---|
| 92 |  | 
|---|
| 93 | /* | 
|---|
| 94 | * Cherry Trail SDM MSR_FSB_FREQ frequencies simplified PLL model: | 
|---|
| 95 | * 0000:   100 *  5 /  6  =  83.3333 MHz | 
|---|
| 96 | * 0001:   100 *  1 /  1  = 100.0000 MHz | 
|---|
| 97 | * 0010:   100 *  4 /  3  = 133.3333 MHz | 
|---|
| 98 | * 0011:   100 *  7 /  6  = 116.6667 MHz | 
|---|
| 99 | * 0100:   100 *  4 /  5  =  80.0000 MHz | 
|---|
| 100 | * 0101:   100 * 14 / 15  =  93.3333 MHz | 
|---|
| 101 | * 0110:   100 *  9 / 10  =  90.0000 MHz | 
|---|
| 102 | * 0111:   100 *  8 /  9  =  88.8889 MHz | 
|---|
| 103 | * 1000:   100 *  7 /  8  =  87.5000 MHz | 
|---|
| 104 | */ | 
|---|
| 105 | static const struct freq_desc freq_desc_cht = { | 
|---|
| 106 | .use_msr_plat = true, | 
|---|
| 107 | .muldiv = { { 5, 6 }, {  1,  1 }, { 4,  3 }, { 7, 6 }, | 
|---|
| 108 | { 4, 5 }, { 14, 15 }, { 9, 10 }, { 8, 9 }, | 
|---|
| 109 | { 7, 8 } }, | 
|---|
| 110 | .mask = 0x0f, | 
|---|
| 111 | }; | 
|---|
| 112 |  | 
|---|
| 113 | /* | 
|---|
| 114 | * Merriefield SDM MSR_FSB_FREQ frequencies simplified PLL model: | 
|---|
| 115 | * 0001:   100 *  1 /  1  = 100.0000 MHz | 
|---|
| 116 | * 0010:   100 *  4 /  3  = 133.3333 MHz | 
|---|
| 117 | */ | 
|---|
| 118 | static const struct freq_desc freq_desc_tng = { | 
|---|
| 119 | .use_msr_plat = true, | 
|---|
| 120 | .muldiv = { { 0, 0 }, { 1, 1 }, { 4, 3 } }, | 
|---|
| 121 | .mask = 0x07, | 
|---|
| 122 | }; | 
|---|
| 123 |  | 
|---|
| 124 | /* | 
|---|
| 125 | * Moorefield SDM MSR_FSB_FREQ frequencies simplified PLL model: | 
|---|
| 126 | * 0000:   100 *  5 /  6  =  83.3333 MHz | 
|---|
| 127 | * 0001:   100 *  1 /  1  = 100.0000 MHz | 
|---|
| 128 | * 0010:   100 *  4 /  3  = 133.3333 MHz | 
|---|
| 129 | * 0011:   100 *  1 /  1  = 100.0000 MHz | 
|---|
| 130 | */ | 
|---|
| 131 | static const struct freq_desc freq_desc_ann = { | 
|---|
| 132 | .use_msr_plat = true, | 
|---|
| 133 | .muldiv = { { 5, 6 }, { 1, 1 }, { 4, 3 }, { 1, 1 } }, | 
|---|
| 134 | .mask = 0x0f, | 
|---|
| 135 | }; | 
|---|
| 136 |  | 
|---|
| 137 | /* | 
|---|
| 138 | * 24 MHz crystal? : 24 * 13 / 4 = 78 MHz | 
|---|
| 139 | * Frequency step for Lightning Mountain SoC is fixed to 78 MHz, | 
|---|
| 140 | * so all the frequency entries are 78000. | 
|---|
| 141 | */ | 
|---|
| 142 | static const struct freq_desc freq_desc_lgm = { | 
|---|
| 143 | .use_msr_plat = true, | 
|---|
| 144 | .freqs = { 78000, 78000, 78000, 78000, 78000, 78000, 78000, 78000, | 
|---|
| 145 | 78000, 78000, 78000, 78000, 78000, 78000, 78000, 78000 }, | 
|---|
| 146 | .mask = 0x0f, | 
|---|
| 147 | }; | 
|---|
| 148 |  | 
|---|
| 149 | static const struct x86_cpu_id tsc_msr_cpu_ids[] = { | 
|---|
| 150 | X86_MATCH_VFM(INTEL_ATOM_SALTWELL_MID,	&freq_desc_pnw), | 
|---|
| 151 | X86_MATCH_VFM(INTEL_ATOM_SALTWELL_TABLET, &freq_desc_clv), | 
|---|
| 152 | X86_MATCH_VFM(INTEL_ATOM_SILVERMONT,	&freq_desc_byt), | 
|---|
| 153 | X86_MATCH_VFM(INTEL_ATOM_SILVERMONT_MID,	&freq_desc_tng), | 
|---|
| 154 | X86_MATCH_VFM(INTEL_ATOM_AIRMONT,	&freq_desc_cht), | 
|---|
| 155 | X86_MATCH_VFM(INTEL_ATOM_SILVERMONT_MID2,	&freq_desc_ann), | 
|---|
| 156 | X86_MATCH_VFM(INTEL_ATOM_AIRMONT_NP,	&freq_desc_lgm), | 
|---|
| 157 | {} | 
|---|
| 158 | }; | 
|---|
| 159 |  | 
|---|
| 160 | /* | 
|---|
| 161 | * MSR-based CPU/TSC frequency discovery for certain CPUs. | 
|---|
| 162 | * | 
|---|
| 163 | * Set global "lapic_timer_period" to bus_clock_cycles/jiffy | 
|---|
| 164 | * Return processor base frequency in KHz, or 0 on failure. | 
|---|
| 165 | */ | 
|---|
| 166 | unsigned long cpu_khz_from_msr(void) | 
|---|
| 167 | { | 
|---|
| 168 | u32 lo, hi, ratio, freq, tscref; | 
|---|
| 169 | const struct freq_desc *freq_desc; | 
|---|
| 170 | const struct x86_cpu_id *id; | 
|---|
| 171 | const struct muldiv *md; | 
|---|
| 172 | unsigned long res; | 
|---|
| 173 | int index; | 
|---|
| 174 |  | 
|---|
| 175 | id = x86_match_cpu(match: tsc_msr_cpu_ids); | 
|---|
| 176 | if (!id) | 
|---|
| 177 | return 0; | 
|---|
| 178 |  | 
|---|
| 179 | freq_desc = (struct freq_desc *)id->driver_data; | 
|---|
| 180 | if (freq_desc->use_msr_plat) { | 
|---|
| 181 | rdmsr(MSR_PLATFORM_INFO, lo, hi); | 
|---|
| 182 | ratio = (lo >> 8) & 0xff; | 
|---|
| 183 | } else { | 
|---|
| 184 | rdmsr(MSR_IA32_PERF_STATUS, lo, hi); | 
|---|
| 185 | ratio = (hi >> 8) & 0x1f; | 
|---|
| 186 | } | 
|---|
| 187 |  | 
|---|
| 188 | /* Get FSB FREQ ID */ | 
|---|
| 189 | rdmsr(MSR_FSB_FREQ, lo, hi); | 
|---|
| 190 | index = lo & freq_desc->mask; | 
|---|
| 191 | md = &freq_desc->muldiv[index]; | 
|---|
| 192 |  | 
|---|
| 193 | /* | 
|---|
| 194 | * Note this also catches cases where the index points to an unpopulated | 
|---|
| 195 | * part of muldiv, in that case the else will set freq and res to 0. | 
|---|
| 196 | */ | 
|---|
| 197 | if (md->divider) { | 
|---|
| 198 | tscref = TSC_REFERENCE_KHZ * md->multiplier; | 
|---|
| 199 | freq = DIV_ROUND_CLOSEST(tscref, md->divider); | 
|---|
| 200 | /* | 
|---|
| 201 | * Multiplying by ratio before the division has better | 
|---|
| 202 | * accuracy than just calculating freq * ratio. | 
|---|
| 203 | */ | 
|---|
| 204 | res = DIV_ROUND_CLOSEST(tscref * ratio, md->divider); | 
|---|
| 205 | } else { | 
|---|
| 206 | freq = freq_desc->freqs[index]; | 
|---|
| 207 | res = freq * ratio; | 
|---|
| 208 | } | 
|---|
| 209 |  | 
|---|
| 210 | if (freq == 0) | 
|---|
| 211 | pr_err( "Error MSR_FSB_FREQ index %d is unknown\n", index); | 
|---|
| 212 |  | 
|---|
| 213 | #ifdef CONFIG_X86_LOCAL_APIC | 
|---|
| 214 | lapic_timer_period = (freq * 1000) / HZ; | 
|---|
| 215 | #endif | 
|---|
| 216 |  | 
|---|
| 217 | /* | 
|---|
| 218 | * TSC frequency determined by MSR is always considered "known" | 
|---|
| 219 | * because it is reported by HW. | 
|---|
| 220 | * Another fact is that on MSR capable platforms, PIT/HPET is | 
|---|
| 221 | * generally not available so calibration won't work at all. | 
|---|
| 222 | */ | 
|---|
| 223 | setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ); | 
|---|
| 224 |  | 
|---|
| 225 | /* | 
|---|
| 226 | * Unfortunately there is no way for hardware to tell whether the | 
|---|
| 227 | * TSC is reliable.  We were told by silicon design team that TSC | 
|---|
| 228 | * on Atom SoCs are always "reliable". TSC is also the only | 
|---|
| 229 | * reliable clocksource on these SoCs (HPET is either not present | 
|---|
| 230 | * or not functional) so mark TSC reliable which removes the | 
|---|
| 231 | * requirement for a watchdog clocksource. | 
|---|
| 232 | */ | 
|---|
| 233 | setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE); | 
|---|
| 234 |  | 
|---|
| 235 | return res; | 
|---|
| 236 | } | 
|---|
| 237 |  | 
|---|