xref: /linux/arch/x86/hyperv/hv_apic.c (revision ca853314e78b0a65c20b6a889a23c31f918d4aa2)
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 		return reg_val;
64 	case APIC_TASKPRI:
65 		rdmsr(HV_X64_MSR_TPR, reg_val, hi);
66 		return reg_val;
67 
68 	default:
69 		return native_apic_mem_read(reg);
70 	}
71 }
72 
73 static void hv_apic_write(u32 reg, u32 val)
74 {
75 	switch (reg) {
76 	case APIC_EOI:
77 		wrmsr(HV_X64_MSR_EOI, val, 0);
78 		break;
79 	case APIC_TASKPRI:
80 		wrmsr(HV_X64_MSR_TPR, val, 0);
81 		break;
82 	default:
83 		native_apic_mem_write(reg, val);
84 	}
85 }
86 
87 static void hv_apic_eoi_write(u32 reg, u32 val)
88 {
89 	struct hv_vp_assist_page *hvp = hv_vp_assist_page[smp_processor_id()];
90 
91 	if (hvp && (xchg(&hvp->apic_assist, 0) & 0x1))
92 		return;
93 
94 	wrmsr(HV_X64_MSR_EOI, val, 0);
95 }
96 
97 /*
98  * IPI implementation on Hyper-V.
99  */
100 static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector)
101 {
102 	struct hv_send_ipi_ex **arg;
103 	struct hv_send_ipi_ex *ipi_arg;
104 	unsigned long flags;
105 	int nr_bank = 0;
106 	int ret = 1;
107 
108 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
109 		return false;
110 
111 	local_irq_save(flags);
112 	arg = (struct hv_send_ipi_ex **)this_cpu_ptr(hyperv_pcpu_input_arg);
113 
114 	ipi_arg = *arg;
115 	if (unlikely(!ipi_arg))
116 		goto ipi_mask_ex_done;
117 
118 	ipi_arg->vector = vector;
119 	ipi_arg->reserved = 0;
120 	ipi_arg->vp_set.valid_bank_mask = 0;
121 
122 	if (!cpumask_equal(mask, cpu_present_mask)) {
123 		ipi_arg->vp_set.format = HV_GENERIC_SET_SPARSE_4K;
124 		nr_bank = cpumask_to_vpset(&(ipi_arg->vp_set), mask);
125 	}
126 	if (nr_bank < 0)
127 		goto ipi_mask_ex_done;
128 	if (!nr_bank)
129 		ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
130 
131 	ret = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
132 			      ipi_arg, NULL);
133 
134 ipi_mask_ex_done:
135 	local_irq_restore(flags);
136 	return ((ret == 0) ? true : false);
137 }
138 
139 static bool __send_ipi_mask(const struct cpumask *mask, int vector)
140 {
141 	int cur_cpu, vcpu;
142 	struct hv_send_ipi ipi_arg;
143 	int ret = 1;
144 
145 	trace_hyperv_send_ipi_mask(mask, vector);
146 
147 	if (cpumask_empty(mask))
148 		return true;
149 
150 	if (!hv_hypercall_pg)
151 		return false;
152 
153 	if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
154 		return false;
155 
156 	/*
157 	 * From the supplied CPU set we need to figure out if we can get away
158 	 * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the
159 	 * highest VP number in the set is < 64. As VP numbers are usually in
160 	 * ascending order and match Linux CPU ids, here is an optimization:
161 	 * we check the VP number for the highest bit in the supplied set first
162 	 * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is
163 	 * a must. We will also check all VP numbers when walking the supplied
164 	 * CPU set to remain correct in all cases.
165 	 */
166 	if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64)
167 		goto do_ex_hypercall;
168 
169 	ipi_arg.vector = vector;
170 	ipi_arg.cpu_mask = 0;
171 
172 	for_each_cpu(cur_cpu, mask) {
173 		vcpu = hv_cpu_number_to_vp_number(cur_cpu);
174 		if (vcpu == VP_INVAL)
175 			return false;
176 
177 		/*
178 		 * This particular version of the IPI hypercall can
179 		 * only target upto 64 CPUs.
180 		 */
181 		if (vcpu >= 64)
182 			goto do_ex_hypercall;
183 
184 		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
185 	}
186 
187 	ret = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
188 				     ipi_arg.cpu_mask);
189 	return ((ret == 0) ? true : false);
190 
191 do_ex_hypercall:
192 	return __send_ipi_mask_ex(mask, vector);
193 }
194 
195 static bool __send_ipi_one(int cpu, int vector)
196 {
197 	int vp = hv_cpu_number_to_vp_number(cpu);
198 
199 	trace_hyperv_send_ipi_one(cpu, vector);
200 
201 	if (!hv_hypercall_pg || (vp == VP_INVAL))
202 		return false;
203 
204 	if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
205 		return false;
206 
207 	if (vp >= 64)
208 		return __send_ipi_mask_ex(cpumask_of(cpu), vector);
209 
210 	return !hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
211 }
212 
213 static void hv_send_ipi(int cpu, int vector)
214 {
215 	if (!__send_ipi_one(cpu, vector))
216 		orig_apic.send_IPI(cpu, vector);
217 }
218 
219 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
220 {
221 	if (!__send_ipi_mask(mask, vector))
222 		orig_apic.send_IPI_mask(mask, vector);
223 }
224 
225 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
226 {
227 	unsigned int this_cpu = smp_processor_id();
228 	struct cpumask new_mask;
229 	const struct cpumask *local_mask;
230 
231 	cpumask_copy(&new_mask, mask);
232 	cpumask_clear_cpu(this_cpu, &new_mask);
233 	local_mask = &new_mask;
234 	if (!__send_ipi_mask(local_mask, vector))
235 		orig_apic.send_IPI_mask_allbutself(mask, vector);
236 }
237 
238 static void hv_send_ipi_allbutself(int vector)
239 {
240 	hv_send_ipi_mask_allbutself(cpu_online_mask, vector);
241 }
242 
243 static void hv_send_ipi_all(int vector)
244 {
245 	if (!__send_ipi_mask(cpu_online_mask, vector))
246 		orig_apic.send_IPI_all(vector);
247 }
248 
249 static void hv_send_ipi_self(int vector)
250 {
251 	if (!__send_ipi_one(smp_processor_id(), vector))
252 		orig_apic.send_IPI_self(vector);
253 }
254 
255 void __init hv_apic_init(void)
256 {
257 	if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) {
258 		pr_info("Hyper-V: Using IPI hypercalls\n");
259 		/*
260 		 * Set the IPI entry points.
261 		 */
262 		orig_apic = *apic;
263 
264 		apic->send_IPI = hv_send_ipi;
265 		apic->send_IPI_mask = hv_send_ipi_mask;
266 		apic->send_IPI_mask_allbutself = hv_send_ipi_mask_allbutself;
267 		apic->send_IPI_allbutself = hv_send_ipi_allbutself;
268 		apic->send_IPI_all = hv_send_ipi_all;
269 		apic->send_IPI_self = hv_send_ipi_self;
270 	}
271 
272 	if (ms_hyperv.hints & HV_X64_APIC_ACCESS_RECOMMENDED) {
273 		pr_info("Hyper-V: Using enlightened APIC (%s mode)",
274 			x2apic_enabled() ? "x2apic" : "xapic");
275 		/*
276 		 * When in x2apic mode, don't use the Hyper-V specific APIC
277 		 * accessors since the field layout in the ICR register is
278 		 * different in x2apic mode. Furthermore, the architectural
279 		 * x2apic MSRs function just as well as the Hyper-V
280 		 * synthetic APIC MSRs, so there's no benefit in having
281 		 * separate Hyper-V accessors for x2apic mode. The only
282 		 * exception is hv_apic_eoi_write, because it benefits from
283 		 * lazy EOI when available, but the same accessor works for
284 		 * both xapic and x2apic because the field layout is the same.
285 		 */
286 		apic_set_eoi_write(hv_apic_eoi_write);
287 		if (!x2apic_enabled()) {
288 			apic->read      = hv_apic_read;
289 			apic->write     = hv_apic_write;
290 			apic->icr_write = hv_apic_icr_write;
291 			apic->icr_read  = hv_apic_icr_read;
292 		}
293 	}
294 }
295