| 1 | // SPDX-License-Identifier: GPL-2.0+ | 
|---|
| 2 | /* | 
|---|
| 3 | * Hygon Processor Support for Linux | 
|---|
| 4 | * | 
|---|
| 5 | * Copyright (C) 2018 Chengdu Haiguang IC Design Co., Ltd. | 
|---|
| 6 | * | 
|---|
| 7 | * Author: Pu Wen <puwen@hygon.cn> | 
|---|
| 8 | */ | 
|---|
| 9 | #include <linux/io.h> | 
|---|
| 10 |  | 
|---|
| 11 | #include <asm/apic.h> | 
|---|
| 12 | #include <asm/cpu.h> | 
|---|
| 13 | #include <asm/smp.h> | 
|---|
| 14 | #include <asm/numa.h> | 
|---|
| 15 | #include <asm/cacheinfo.h> | 
|---|
| 16 | #include <asm/spec-ctrl.h> | 
|---|
| 17 | #include <asm/delay.h> | 
|---|
| 18 | #include <asm/msr.h> | 
|---|
| 19 | #include <asm/resctrl.h> | 
|---|
| 20 |  | 
|---|
| 21 | #include "cpu.h" | 
|---|
| 22 |  | 
|---|
| 23 | #ifdef CONFIG_NUMA | 
|---|
| 24 | /* | 
|---|
| 25 | * To workaround broken NUMA config.  Read the comment in | 
|---|
| 26 | * srat_detect_node(). | 
|---|
| 27 | */ | 
|---|
| 28 | static int nearby_node(int apicid) | 
|---|
| 29 | { | 
|---|
| 30 | int i, node; | 
|---|
| 31 |  | 
|---|
| 32 | for (i = apicid - 1; i >= 0; i--) { | 
|---|
| 33 | node = __apicid_to_node[i]; | 
|---|
| 34 | if (node != NUMA_NO_NODE && node_online(node)) | 
|---|
| 35 | return node; | 
|---|
| 36 | } | 
|---|
| 37 | for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { | 
|---|
| 38 | node = __apicid_to_node[i]; | 
|---|
| 39 | if (node != NUMA_NO_NODE && node_online(node)) | 
|---|
| 40 | return node; | 
|---|
| 41 | } | 
|---|
| 42 | return first_node(node_online_map); /* Shouldn't happen */ | 
|---|
| 43 | } | 
|---|
| 44 | #endif | 
|---|
| 45 |  | 
|---|
| 46 | static void srat_detect_node(struct cpuinfo_x86 *c) | 
|---|
| 47 | { | 
|---|
| 48 | #ifdef CONFIG_NUMA | 
|---|
| 49 | int cpu = smp_processor_id(); | 
|---|
| 50 | int node; | 
|---|
| 51 | unsigned int apicid = c->topo.apicid; | 
|---|
| 52 |  | 
|---|
| 53 | node = numa_cpu_node(cpu); | 
|---|
| 54 | if (node == NUMA_NO_NODE) | 
|---|
| 55 | node = c->topo.llc_id; | 
|---|
| 56 |  | 
|---|
| 57 | /* | 
|---|
| 58 | * On multi-fabric platform (e.g. Numascale NumaChip) a | 
|---|
| 59 | * platform-specific handler needs to be called to fixup some | 
|---|
| 60 | * IDs of the CPU. | 
|---|
| 61 | */ | 
|---|
| 62 | if (x86_cpuinit.fixup_cpu_id) | 
|---|
| 63 | x86_cpuinit.fixup_cpu_id(c, node); | 
|---|
| 64 |  | 
|---|
| 65 | if (!node_online(node)) { | 
|---|
| 66 | /* | 
|---|
| 67 | * Two possibilities here: | 
|---|
| 68 | * | 
|---|
| 69 | * - The CPU is missing memory and no node was created.  In | 
|---|
| 70 | *   that case try picking one from a nearby CPU. | 
|---|
| 71 | * | 
|---|
| 72 | * - The APIC IDs differ from the HyperTransport node IDs. | 
|---|
| 73 | *   Assume they are all increased by a constant offset, but | 
|---|
| 74 | *   in the same order as the HT nodeids.  If that doesn't | 
|---|
| 75 | *   result in a usable node fall back to the path for the | 
|---|
| 76 | *   previous case. | 
|---|
| 77 | * | 
|---|
| 78 | * This workaround operates directly on the mapping between | 
|---|
| 79 | * APIC ID and NUMA node, assuming certain relationship | 
|---|
| 80 | * between APIC ID, HT node ID and NUMA topology.  As going | 
|---|
| 81 | * through CPU mapping may alter the outcome, directly | 
|---|
| 82 | * access __apicid_to_node[]. | 
|---|
| 83 | */ | 
|---|
| 84 | int ht_nodeid = c->topo.initial_apicid; | 
|---|
| 85 |  | 
|---|
| 86 | if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE) | 
|---|
| 87 | node = __apicid_to_node[ht_nodeid]; | 
|---|
| 88 | /* Pick a nearby node */ | 
|---|
| 89 | if (!node_online(node)) | 
|---|
| 90 | node = nearby_node(apicid); | 
|---|
| 91 | } | 
|---|
| 92 | numa_set_node(cpu, node); | 
|---|
| 93 | #endif | 
|---|
| 94 | } | 
|---|
| 95 |  | 
|---|
| 96 | static void bsp_init_hygon(struct cpuinfo_x86 *c) | 
|---|
| 97 | { | 
|---|
| 98 | if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) { | 
|---|
| 99 | u64 val; | 
|---|
| 100 |  | 
|---|
| 101 | rdmsrq(MSR_K7_HWCR, val); | 
|---|
| 102 | if (!(val & BIT(24))) | 
|---|
| 103 | pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n"); | 
|---|
| 104 | } | 
|---|
| 105 |  | 
|---|
| 106 | if (cpu_has(c, X86_FEATURE_MWAITX)) | 
|---|
| 107 | use_mwaitx_delay(); | 
|---|
| 108 |  | 
|---|
| 109 | if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) && | 
|---|
| 110 | !boot_cpu_has(X86_FEATURE_VIRT_SSBD)) { | 
|---|
| 111 | /* | 
|---|
| 112 | * Try to cache the base value so further operations can | 
|---|
| 113 | * avoid RMW. If that faults, do not enable SSBD. | 
|---|
| 114 | */ | 
|---|
| 115 | if (!rdmsrq_safe(MSR_AMD64_LS_CFG, p: &x86_amd_ls_cfg_base)) { | 
|---|
| 116 | setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD); | 
|---|
| 117 | setup_force_cpu_cap(X86_FEATURE_SSBD); | 
|---|
| 118 | x86_amd_ls_cfg_ssbd_mask = 1ULL << 10; | 
|---|
| 119 | } | 
|---|
| 120 | } | 
|---|
| 121 |  | 
|---|
| 122 | resctrl_cpu_detect(c); | 
|---|
| 123 | } | 
|---|
| 124 |  | 
|---|
| 125 | static void early_init_hygon(struct cpuinfo_x86 *c) | 
|---|
| 126 | { | 
|---|
| 127 | u32 dummy; | 
|---|
| 128 |  | 
|---|
| 129 | set_cpu_cap(c, X86_FEATURE_K8); | 
|---|
| 130 |  | 
|---|
| 131 | rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy); | 
|---|
| 132 |  | 
|---|
| 133 | /* | 
|---|
| 134 | * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate | 
|---|
| 135 | * with P/T states and does not stop in deep C-states | 
|---|
| 136 | */ | 
|---|
| 137 | if (c->x86_power & (1 << 8)) { | 
|---|
| 138 | set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); | 
|---|
| 139 | set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); | 
|---|
| 140 | } | 
|---|
| 141 |  | 
|---|
| 142 | /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */ | 
|---|
| 143 | if (c->x86_power & BIT(12)) | 
|---|
| 144 | set_cpu_cap(c, X86_FEATURE_ACC_POWER); | 
|---|
| 145 |  | 
|---|
| 146 | /* Bit 14 indicates the Runtime Average Power Limit interface. */ | 
|---|
| 147 | if (c->x86_power & BIT(14)) | 
|---|
| 148 | set_cpu_cap(c, X86_FEATURE_RAPL); | 
|---|
| 149 |  | 
|---|
| 150 | #ifdef CONFIG_X86_64 | 
|---|
| 151 | set_cpu_cap(c, X86_FEATURE_SYSCALL32); | 
|---|
| 152 | #endif | 
|---|
| 153 |  | 
|---|
| 154 | #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI) | 
|---|
| 155 | /* | 
|---|
| 156 | * ApicID can always be treated as an 8-bit value for Hygon APIC So, we | 
|---|
| 157 | * can safely set X86_FEATURE_EXTD_APICID unconditionally. | 
|---|
| 158 | */ | 
|---|
| 159 | if (boot_cpu_has(X86_FEATURE_APIC)) | 
|---|
| 160 | set_cpu_cap(c, X86_FEATURE_EXTD_APICID); | 
|---|
| 161 | #endif | 
|---|
| 162 |  | 
|---|
| 163 | /* | 
|---|
| 164 | * This is only needed to tell the kernel whether to use VMCALL | 
|---|
| 165 | * and VMMCALL.  VMMCALL is never executed except under virt, so | 
|---|
| 166 | * we can set it unconditionally. | 
|---|
| 167 | */ | 
|---|
| 168 | set_cpu_cap(c, X86_FEATURE_VMMCALL); | 
|---|
| 169 | } | 
|---|
| 170 |  | 
|---|
| 171 | static void init_hygon(struct cpuinfo_x86 *c) | 
|---|
| 172 | { | 
|---|
| 173 | u64 vm_cr; | 
|---|
| 174 |  | 
|---|
| 175 | early_init_hygon(c); | 
|---|
| 176 |  | 
|---|
| 177 | /* | 
|---|
| 178 | * Bit 31 in normal CPUID used for nonstandard 3DNow ID; | 
|---|
| 179 | * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway | 
|---|
| 180 | */ | 
|---|
| 181 | clear_cpu_cap(c, bit: 0*32+31); | 
|---|
| 182 |  | 
|---|
| 183 | set_cpu_cap(c, X86_FEATURE_REP_GOOD); | 
|---|
| 184 |  | 
|---|
| 185 | /* | 
|---|
| 186 | * XXX someone from Hygon needs to confirm this DTRT | 
|---|
| 187 | * | 
|---|
| 188 | init_spectral_chicken(c); | 
|---|
| 189 | */ | 
|---|
| 190 |  | 
|---|
| 191 | set_cpu_cap(c, X86_FEATURE_ZEN); | 
|---|
| 192 | set_cpu_cap(c, X86_FEATURE_CPB); | 
|---|
| 193 |  | 
|---|
| 194 | cpu_detect_cache_sizes(c); | 
|---|
| 195 |  | 
|---|
| 196 | srat_detect_node(c); | 
|---|
| 197 |  | 
|---|
| 198 | init_hygon_cacheinfo(c); | 
|---|
| 199 |  | 
|---|
| 200 | if (cpu_has(c, X86_FEATURE_SVM)) { | 
|---|
| 201 | rdmsrq(MSR_VM_CR, vm_cr); | 
|---|
| 202 | if (vm_cr & SVM_VM_CR_SVM_DIS_MASK) { | 
|---|
| 203 | pr_notice_once( "SVM disabled (by BIOS) in MSR_VM_CR\n"); | 
|---|
| 204 | clear_cpu_cap(c, X86_FEATURE_SVM); | 
|---|
| 205 | } | 
|---|
| 206 | } | 
|---|
| 207 |  | 
|---|
| 208 | if (cpu_has(c, X86_FEATURE_XMM2)) { | 
|---|
| 209 | /* | 
|---|
| 210 | * Use LFENCE for execution serialization.  On families which | 
|---|
| 211 | * don't have that MSR, LFENCE is already serializing. | 
|---|
| 212 | * msr_set_bit() uses the safe accessors, too, even if the MSR | 
|---|
| 213 | * is not present. | 
|---|
| 214 | */ | 
|---|
| 215 | msr_set_bit(MSR_AMD64_DE_CFG, | 
|---|
| 216 | MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT); | 
|---|
| 217 |  | 
|---|
| 218 | /* A serializing LFENCE stops RDTSC speculation */ | 
|---|
| 219 | set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC); | 
|---|
| 220 | } | 
|---|
| 221 |  | 
|---|
| 222 | /* | 
|---|
| 223 | * Hygon processors have APIC timer running in deep C states. | 
|---|
| 224 | */ | 
|---|
| 225 | set_cpu_cap(c, X86_FEATURE_ARAT); | 
|---|
| 226 |  | 
|---|
| 227 | /* Hygon CPUs don't reset SS attributes on SYSRET, Xen does. */ | 
|---|
| 228 | if (!cpu_feature_enabled(X86_FEATURE_XENPV)) | 
|---|
| 229 | set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS); | 
|---|
| 230 |  | 
|---|
| 231 | check_null_seg_clears_base(c); | 
|---|
| 232 |  | 
|---|
| 233 | /* Hygon CPUs don't need fencing after x2APIC/TSC_DEADLINE MSR writes. */ | 
|---|
| 234 | clear_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE); | 
|---|
| 235 | } | 
|---|
| 236 |  | 
|---|
| 237 | static void cpu_detect_tlb_hygon(struct cpuinfo_x86 *c) | 
|---|
| 238 | { | 
|---|
| 239 | u32 ebx, eax, ecx, edx; | 
|---|
| 240 | u16 mask = 0xfff; | 
|---|
| 241 |  | 
|---|
| 242 | if (c->extended_cpuid_level < 0x80000006) | 
|---|
| 243 | return; | 
|---|
| 244 |  | 
|---|
| 245 | cpuid(op: 0x80000006, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); | 
|---|
| 246 |  | 
|---|
| 247 | tlb_lld_4k = (ebx >> 16) & mask; | 
|---|
| 248 | tlb_lli_4k = ebx & mask; | 
|---|
| 249 |  | 
|---|
| 250 | /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ | 
|---|
| 251 | if (!((eax >> 16) & mask)) | 
|---|
| 252 | tlb_lld_2m = (cpuid_eax(op: 0x80000005) >> 16) & 0xff; | 
|---|
| 253 | else | 
|---|
| 254 | tlb_lld_2m = (eax >> 16) & mask; | 
|---|
| 255 |  | 
|---|
| 256 | /* a 4M entry uses two 2M entries */ | 
|---|
| 257 | tlb_lld_4m = tlb_lld_2m >> 1; | 
|---|
| 258 |  | 
|---|
| 259 | /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ | 
|---|
| 260 | if (!(eax & mask)) { | 
|---|
| 261 | cpuid(op: 0x80000005, eax: &eax, ebx: &ebx, ecx: &ecx, edx: &edx); | 
|---|
| 262 | tlb_lli_2m = eax & 0xff; | 
|---|
| 263 | } else | 
|---|
| 264 | tlb_lli_2m = eax & mask; | 
|---|
| 265 |  | 
|---|
| 266 | tlb_lli_4m = tlb_lli_2m >> 1; | 
|---|
| 267 | } | 
|---|
| 268 |  | 
|---|
| 269 | static const struct cpu_dev hygon_cpu_dev = { | 
|---|
| 270 | .c_vendor	= "Hygon", | 
|---|
| 271 | .c_ident	= { "HygonGenuine"}, | 
|---|
| 272 | .c_early_init   = early_init_hygon, | 
|---|
| 273 | .c_detect_tlb	= cpu_detect_tlb_hygon, | 
|---|
| 274 | .c_bsp_init	= bsp_init_hygon, | 
|---|
| 275 | .c_init		= init_hygon, | 
|---|
| 276 | .c_x86_vendor	= X86_VENDOR_HYGON, | 
|---|
| 277 | }; | 
|---|
| 278 |  | 
|---|
| 279 | cpu_dev_register(hygon_cpu_dev); | 
|---|
| 280 |  | 
|---|