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