1 // SPDX-License-Identifier: GPL-2.0 2 3 /* 4 * Hyper-V specific APIC code. 5 * 6 * Copyright (C) 2018, Microsoft, Inc. 7 * 8 * Author : K. Y. Srinivasan <kys@microsoft.com> 9 * 10 * This program is free software; you can redistribute it and/or modify it 11 * under the terms of the GNU General Public License version 2 as published 12 * by the Free Software Foundation. 13 * 14 * This program is distributed in the hope that it will be useful, but 15 * WITHOUT ANY WARRANTY; without even the implied warranty of 16 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or 17 * NON INFRINGEMENT. See the GNU General Public License for more 18 * details. 19 * 20 */ 21 22 #include <linux/types.h> 23 #include <linux/vmalloc.h> 24 #include <linux/mm.h> 25 #include <linux/clockchips.h> 26 #include <linux/hyperv.h> 27 #include <linux/slab.h> 28 #include <linux/cpuhotplug.h> 29 #include <asm/hypervisor.h> 30 #include <asm/mshyperv.h> 31 #include <asm/apic.h> 32 33 #include <asm/trace/hyperv.h> 34 35 static struct apic orig_apic; 36 37 static u64 hv_apic_icr_read(void) 38 { 39 u64 reg_val; 40 41 rdmsrl(HV_X64_MSR_ICR, reg_val); 42 return reg_val; 43 } 44 45 static void hv_apic_icr_write(u32 low, u32 id) 46 { 47 u64 reg_val; 48 49 reg_val = SET_XAPIC_DEST_FIELD(id); 50 reg_val = reg_val << 32; 51 reg_val |= low; 52 53 wrmsrl(HV_X64_MSR_ICR, reg_val); 54 } 55 56 static u32 hv_apic_read(u32 reg) 57 { 58 u32 reg_val, hi; 59 60 switch (reg) { 61 case APIC_EOI: 62 rdmsr(HV_X64_MSR_EOI, reg_val, hi); 63 (void)hi; 64 return reg_val; 65 case APIC_TASKPRI: 66 rdmsr(HV_X64_MSR_TPR, reg_val, hi); 67 (void)hi; 68 return reg_val; 69 70 default: 71 return native_apic_mem_read(reg); 72 } 73 } 74 75 static void hv_apic_write(u32 reg, u32 val) 76 { 77 switch (reg) { 78 case APIC_EOI: 79 wrmsr(HV_X64_MSR_EOI, val, 0); 80 break; 81 case APIC_TASKPRI: 82 wrmsr(HV_X64_MSR_TPR, val, 0); 83 break; 84 default: 85 native_apic_mem_write(reg, val); 86 } 87 } 88 89 static void hv_apic_eoi_write(u32 reg, u32 val) 90 { 91 struct hv_vp_assist_page *hvp = hv_vp_assist_page[smp_processor_id()]; 92 93 if (hvp && (xchg(&hvp->apic_assist, 0) & 0x1)) 94 return; 95 96 wrmsr(HV_X64_MSR_EOI, val, 0); 97 } 98 99 /* 100 * IPI implementation on Hyper-V. 101 */ 102 static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector, 103 bool exclude_self) 104 { 105 struct hv_send_ipi_ex **arg; 106 struct hv_send_ipi_ex *ipi_arg; 107 unsigned long flags; 108 int nr_bank = 0; 109 u64 status = HV_STATUS_INVALID_PARAMETER; 110 111 if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED)) 112 return false; 113 114 local_irq_save(flags); 115 arg = (struct hv_send_ipi_ex **)this_cpu_ptr(hyperv_pcpu_input_arg); 116 117 ipi_arg = *arg; 118 if (unlikely(!ipi_arg)) 119 goto ipi_mask_ex_done; 120 121 ipi_arg->vector = vector; 122 ipi_arg->reserved = 0; 123 ipi_arg->vp_set.valid_bank_mask = 0; 124 125 /* 126 * Use HV_GENERIC_SET_ALL and avoid converting cpumask to VP_SET 127 * when the IPI is sent to all currently present CPUs. 128 */ 129 if (!cpumask_equal(mask, cpu_present_mask) || exclude_self) { 130 ipi_arg->vp_set.format = HV_GENERIC_SET_SPARSE_4K; 131 if (exclude_self) 132 nr_bank = cpumask_to_vpset_noself(&(ipi_arg->vp_set), mask); 133 else 134 nr_bank = cpumask_to_vpset(&(ipi_arg->vp_set), mask); 135 136 /* 137 * 'nr_bank <= 0' means some CPUs in cpumask can't be 138 * represented in VP_SET. Return an error and fall back to 139 * native (architectural) method of sending IPIs. 140 */ 141 if (nr_bank <= 0) 142 goto ipi_mask_ex_done; 143 } else { 144 ipi_arg->vp_set.format = HV_GENERIC_SET_ALL; 145 } 146 147 status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank, 148 ipi_arg, NULL); 149 150 ipi_mask_ex_done: 151 local_irq_restore(flags); 152 return hv_result_success(status); 153 } 154 155 static bool __send_ipi_mask(const struct cpumask *mask, int vector, 156 bool exclude_self) 157 { 158 int cur_cpu, vcpu, this_cpu = smp_processor_id(); 159 struct hv_send_ipi ipi_arg; 160 u64 status; 161 unsigned int weight; 162 163 trace_hyperv_send_ipi_mask(mask, vector); 164 165 weight = cpumask_weight(mask); 166 167 /* 168 * Do nothing if 169 * 1. the mask is empty 170 * 2. the mask only contains self when exclude_self is true 171 */ 172 if (weight == 0 || 173 (exclude_self && weight == 1 && cpumask_test_cpu(this_cpu, mask))) 174 return true; 175 176 if (!hv_hypercall_pg) 177 return false; 178 179 if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR)) 180 return false; 181 182 /* 183 * From the supplied CPU set we need to figure out if we can get away 184 * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the 185 * highest VP number in the set is < 64. As VP numbers are usually in 186 * ascending order and match Linux CPU ids, here is an optimization: 187 * we check the VP number for the highest bit in the supplied set first 188 * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is 189 * a must. We will also check all VP numbers when walking the supplied 190 * CPU set to remain correct in all cases. 191 */ 192 if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64) 193 goto do_ex_hypercall; 194 195 ipi_arg.vector = vector; 196 ipi_arg.cpu_mask = 0; 197 198 for_each_cpu(cur_cpu, mask) { 199 if (exclude_self && cur_cpu == this_cpu) 200 continue; 201 vcpu = hv_cpu_number_to_vp_number(cur_cpu); 202 if (vcpu == VP_INVAL) 203 return false; 204 205 /* 206 * This particular version of the IPI hypercall can 207 * only target upto 64 CPUs. 208 */ 209 if (vcpu >= 64) 210 goto do_ex_hypercall; 211 212 __set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask); 213 } 214 215 status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector, 216 ipi_arg.cpu_mask); 217 return hv_result_success(status); 218 219 do_ex_hypercall: 220 return __send_ipi_mask_ex(mask, vector, exclude_self); 221 } 222 223 static bool __send_ipi_one(int cpu, int vector) 224 { 225 int vp = hv_cpu_number_to_vp_number(cpu); 226 u64 status; 227 228 trace_hyperv_send_ipi_one(cpu, vector); 229 230 if (!hv_hypercall_pg || (vp == VP_INVAL)) 231 return false; 232 233 if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR)) 234 return false; 235 236 if (vp >= 64) 237 return __send_ipi_mask_ex(cpumask_of(cpu), vector, false); 238 239 status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp)); 240 return hv_result_success(status); 241 } 242 243 static void hv_send_ipi(int cpu, int vector) 244 { 245 if (!__send_ipi_one(cpu, vector)) 246 orig_apic.send_IPI(cpu, vector); 247 } 248 249 static void hv_send_ipi_mask(const struct cpumask *mask, int vector) 250 { 251 if (!__send_ipi_mask(mask, vector, false)) 252 orig_apic.send_IPI_mask(mask, vector); 253 } 254 255 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector) 256 { 257 if (!__send_ipi_mask(mask, vector, true)) 258 orig_apic.send_IPI_mask_allbutself(mask, vector); 259 } 260 261 static void hv_send_ipi_allbutself(int vector) 262 { 263 hv_send_ipi_mask_allbutself(cpu_online_mask, vector); 264 } 265 266 static void hv_send_ipi_all(int vector) 267 { 268 if (!__send_ipi_mask(cpu_online_mask, vector, false)) 269 orig_apic.send_IPI_all(vector); 270 } 271 272 static void hv_send_ipi_self(int vector) 273 { 274 if (!__send_ipi_one(smp_processor_id(), vector)) 275 orig_apic.send_IPI_self(vector); 276 } 277 278 void __init hv_apic_init(void) 279 { 280 if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) { 281 pr_info("Hyper-V: Using IPI hypercalls\n"); 282 /* 283 * Set the IPI entry points. 284 */ 285 orig_apic = *apic; 286 287 apic->send_IPI = hv_send_ipi; 288 apic->send_IPI_mask = hv_send_ipi_mask; 289 apic->send_IPI_mask_allbutself = hv_send_ipi_mask_allbutself; 290 apic->send_IPI_allbutself = hv_send_ipi_allbutself; 291 apic->send_IPI_all = hv_send_ipi_all; 292 apic->send_IPI_self = hv_send_ipi_self; 293 } 294 295 if (ms_hyperv.hints & HV_X64_APIC_ACCESS_RECOMMENDED) { 296 pr_info("Hyper-V: Using enlightened APIC (%s mode)", 297 x2apic_enabled() ? "x2apic" : "xapic"); 298 /* 299 * When in x2apic mode, don't use the Hyper-V specific APIC 300 * accessors since the field layout in the ICR register is 301 * different in x2apic mode. Furthermore, the architectural 302 * x2apic MSRs function just as well as the Hyper-V 303 * synthetic APIC MSRs, so there's no benefit in having 304 * separate Hyper-V accessors for x2apic mode. The only 305 * exception is hv_apic_eoi_write, because it benefits from 306 * lazy EOI when available, but the same accessor works for 307 * both xapic and x2apic because the field layout is the same. 308 */ 309 apic_set_eoi_write(hv_apic_eoi_write); 310 if (!x2apic_enabled()) { 311 apic->read = hv_apic_read; 312 apic->write = hv_apic_write; 313 apic->icr_write = hv_apic_icr_write; 314 apic->icr_read = hv_apic_icr_read; 315 } 316 } 317 } 318