xref: /linux/arch/x86/hyperv/hv_apic.c (revision 17b121ad0c43342bc894632f6710b894849ca372)
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_APIC_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 {
104 	struct hv_send_ipi_ex **arg;
105 	struct hv_send_ipi_ex *ipi_arg;
106 	unsigned long flags;
107 	int nr_bank = 0;
108 	u64 status = HV_STATUS_INVALID_PARAMETER;
109 
110 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
111 		return false;
112 
113 	local_irq_save(flags);
114 	arg = (struct hv_send_ipi_ex **)this_cpu_ptr(hyperv_pcpu_input_arg);
115 
116 	ipi_arg = *arg;
117 	if (unlikely(!ipi_arg))
118 		goto ipi_mask_ex_done;
119 
120 	ipi_arg->vector = vector;
121 	ipi_arg->reserved = 0;
122 	ipi_arg->vp_set.valid_bank_mask = 0;
123 
124 	if (!cpumask_equal(mask, cpu_present_mask)) {
125 		ipi_arg->vp_set.format = HV_GENERIC_SET_SPARSE_4K;
126 		nr_bank = cpumask_to_vpset(&(ipi_arg->vp_set), mask);
127 	}
128 	if (nr_bank < 0)
129 		goto ipi_mask_ex_done;
130 	if (!nr_bank)
131 		ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
132 
133 	status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
134 			      ipi_arg, NULL);
135 
136 ipi_mask_ex_done:
137 	local_irq_restore(flags);
138 	return hv_result_success(status);
139 }
140 
141 static bool __send_ipi_mask(const struct cpumask *mask, int vector)
142 {
143 	int cur_cpu, vcpu;
144 	struct hv_send_ipi ipi_arg;
145 	u64 status;
146 
147 	trace_hyperv_send_ipi_mask(mask, vector);
148 
149 	if (cpumask_empty(mask))
150 		return true;
151 
152 	if (!hv_hypercall_pg)
153 		return false;
154 
155 	if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
156 		return false;
157 
158 	/*
159 	 * From the supplied CPU set we need to figure out if we can get away
160 	 * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the
161 	 * highest VP number in the set is < 64. As VP numbers are usually in
162 	 * ascending order and match Linux CPU ids, here is an optimization:
163 	 * we check the VP number for the highest bit in the supplied set first
164 	 * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is
165 	 * a must. We will also check all VP numbers when walking the supplied
166 	 * CPU set to remain correct in all cases.
167 	 */
168 	if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64)
169 		goto do_ex_hypercall;
170 
171 	ipi_arg.vector = vector;
172 	ipi_arg.cpu_mask = 0;
173 
174 	for_each_cpu(cur_cpu, mask) {
175 		vcpu = hv_cpu_number_to_vp_number(cur_cpu);
176 		if (vcpu == VP_INVAL)
177 			return false;
178 
179 		/*
180 		 * This particular version of the IPI hypercall can
181 		 * only target upto 64 CPUs.
182 		 */
183 		if (vcpu >= 64)
184 			goto do_ex_hypercall;
185 
186 		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
187 	}
188 
189 	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
190 				     ipi_arg.cpu_mask);
191 	return hv_result_success(status);
192 
193 do_ex_hypercall:
194 	return __send_ipi_mask_ex(mask, vector);
195 }
196 
197 static bool __send_ipi_one(int cpu, int vector)
198 {
199 	int vp = hv_cpu_number_to_vp_number(cpu);
200 	u64 status;
201 
202 	trace_hyperv_send_ipi_one(cpu, vector);
203 
204 	if (!hv_hypercall_pg || (vp == VP_INVAL))
205 		return false;
206 
207 	if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
208 		return false;
209 
210 	if (vp >= 64)
211 		return __send_ipi_mask_ex(cpumask_of(cpu), vector);
212 
213 	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
214 	return hv_result_success(status);
215 }
216 
217 static void hv_send_ipi(int cpu, int vector)
218 {
219 	if (!__send_ipi_one(cpu, vector))
220 		orig_apic.send_IPI(cpu, vector);
221 }
222 
223 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
224 {
225 	if (!__send_ipi_mask(mask, vector))
226 		orig_apic.send_IPI_mask(mask, vector);
227 }
228 
229 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
230 {
231 	unsigned int this_cpu = smp_processor_id();
232 	struct cpumask new_mask;
233 	const struct cpumask *local_mask;
234 
235 	cpumask_copy(&new_mask, mask);
236 	cpumask_clear_cpu(this_cpu, &new_mask);
237 	local_mask = &new_mask;
238 	if (!__send_ipi_mask(local_mask, vector))
239 		orig_apic.send_IPI_mask_allbutself(mask, vector);
240 }
241 
242 static void hv_send_ipi_allbutself(int vector)
243 {
244 	hv_send_ipi_mask_allbutself(cpu_online_mask, vector);
245 }
246 
247 static void hv_send_ipi_all(int vector)
248 {
249 	if (!__send_ipi_mask(cpu_online_mask, vector))
250 		orig_apic.send_IPI_all(vector);
251 }
252 
253 static void hv_send_ipi_self(int vector)
254 {
255 	if (!__send_ipi_one(smp_processor_id(), vector))
256 		orig_apic.send_IPI_self(vector);
257 }
258 
259 void __init hv_apic_init(void)
260 {
261 	if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) {
262 		pr_info("Hyper-V: Using IPI hypercalls\n");
263 		/*
264 		 * Set the IPI entry points.
265 		 */
266 		orig_apic = *apic;
267 
268 		apic->send_IPI = hv_send_ipi;
269 		apic->send_IPI_mask = hv_send_ipi_mask;
270 		apic->send_IPI_mask_allbutself = hv_send_ipi_mask_allbutself;
271 		apic->send_IPI_allbutself = hv_send_ipi_allbutself;
272 		apic->send_IPI_all = hv_send_ipi_all;
273 		apic->send_IPI_self = hv_send_ipi_self;
274 	}
275 
276 	if (ms_hyperv.hints & HV_X64_APIC_ACCESS_RECOMMENDED) {
277 		pr_info("Hyper-V: Using enlightened APIC (%s mode)",
278 			x2apic_enabled() ? "x2apic" : "xapic");
279 		/*
280 		 * When in x2apic mode, don't use the Hyper-V specific APIC
281 		 * accessors since the field layout in the ICR register is
282 		 * different in x2apic mode. Furthermore, the architectural
283 		 * x2apic MSRs function just as well as the Hyper-V
284 		 * synthetic APIC MSRs, so there's no benefit in having
285 		 * separate Hyper-V accessors for x2apic mode. The only
286 		 * exception is hv_apic_eoi_write, because it benefits from
287 		 * lazy EOI when available, but the same accessor works for
288 		 * both xapic and x2apic because the field layout is the same.
289 		 */
290 		apic_set_eoi_write(hv_apic_eoi_write);
291 		if (!x2apic_enabled()) {
292 			apic->read      = hv_apic_read;
293 			apic->write     = hv_apic_write;
294 			apic->icr_write = hv_apic_icr_write;
295 			apic->icr_read  = hv_apic_icr_read;
296 		}
297 	}
298 }
299