xref: /linux/arch/x86/hyperv/hv_apic.c (revision a713222906e4f77b5fb1b5346d4f5de1adc639b4)
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/slab.h>
27 #include <linux/cpuhotplug.h>
28 #include <asm/hypervisor.h>
29 #include <asm/mshyperv.h>
30 #include <asm/apic.h>
31 #include <asm/msr.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 	rdmsrq(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 	wrmsrq(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 		wrmsrq(HV_X64_MSR_EOI, val);
80 		break;
81 	case APIC_TASKPRI:
82 		wrmsrq(HV_X64_MSR_TPR, val);
83 		break;
84 	default:
85 		native_apic_mem_write(reg, val);
86 	}
87 }
88 
89 static void hv_apic_eoi_write(void)
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 	wrmsrq(HV_X64_MSR_EOI, APIC_EOI_ACK);
97 }
98 
99 static bool cpu_is_self(int cpu)
100 {
101 	return cpu == smp_processor_id();
102 }
103 
104 /*
105  * IPI implementation on Hyper-V.
106  */
107 static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector,
108 			       bool exclude_self)
109 {
110 	struct hv_send_ipi_ex *ipi_arg;
111 	unsigned long flags;
112 	int nr_bank = 0;
113 	u64 status = HV_STATUS_INVALID_PARAMETER;
114 
115 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
116 		return false;
117 
118 	local_irq_save(flags);
119 	ipi_arg = *this_cpu_ptr(hyperv_pcpu_input_arg);
120 
121 	if (unlikely(!ipi_arg))
122 		goto ipi_mask_ex_done;
123 
124 	ipi_arg->vector = vector;
125 	ipi_arg->reserved = 0;
126 	ipi_arg->vp_set.valid_bank_mask = 0;
127 
128 	/*
129 	 * Use HV_GENERIC_SET_ALL and avoid converting cpumask to VP_SET
130 	 * when the IPI is sent to all currently present CPUs.
131 	 */
132 	if (!cpumask_equal(mask, cpu_present_mask) || exclude_self) {
133 		ipi_arg->vp_set.format = HV_GENERIC_SET_SPARSE_4K;
134 
135 		nr_bank = cpumask_to_vpset_skip(&ipi_arg->vp_set, mask,
136 						exclude_self ? cpu_is_self : NULL);
137 
138 		/*
139 		 * 'nr_bank <= 0' means some CPUs in cpumask can't be
140 		 * represented in VP_SET. Return an error and fall back to
141 		 * native (architectural) method of sending IPIs.
142 		 */
143 		if (nr_bank <= 0)
144 			goto ipi_mask_ex_done;
145 	} else {
146 		ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
147 	}
148 
149 	/*
150 	 * For this hypercall, Hyper-V treats the valid_bank_mask field
151 	 * of ipi_arg->vp_set as part of the fixed size input header.
152 	 * So the variable input header size is equal to nr_bank.
153 	 */
154 	status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
155 				     ipi_arg, NULL);
156 
157 ipi_mask_ex_done:
158 	local_irq_restore(flags);
159 	return hv_result_success(status);
160 }
161 
162 static bool __send_ipi_mask(const struct cpumask *mask, int vector,
163 			    bool exclude_self)
164 {
165 	int cur_cpu, vcpu, this_cpu = smp_processor_id();
166 	struct hv_send_ipi ipi_arg;
167 	u64 status;
168 	unsigned int weight;
169 
170 	trace_hyperv_send_ipi_mask(mask, vector);
171 
172 	weight = cpumask_weight(mask);
173 
174 	/*
175 	 * Do nothing if
176 	 *   1. the mask is empty
177 	 *   2. the mask only contains self when exclude_self is true
178 	 */
179 	if (weight == 0 ||
180 	    (exclude_self && weight == 1 && cpumask_test_cpu(this_cpu, mask)))
181 		return true;
182 
183 	/* A fully enlightened TDX VM uses GHCI rather than hv_hypercall_pg. */
184 	if (!hv_hypercall_pg) {
185 		if (ms_hyperv.paravisor_present || !hv_isolation_type_tdx())
186 			return false;
187 	}
188 
189 	if (vector < HV_IPI_LOW_VECTOR || vector > HV_IPI_HIGH_VECTOR)
190 		return false;
191 
192 	/*
193 	 * From the supplied CPU set we need to figure out if we can get away
194 	 * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the
195 	 * highest VP number in the set is < 64. As VP numbers are usually in
196 	 * ascending order and match Linux CPU ids, here is an optimization:
197 	 * we check the VP number for the highest bit in the supplied set first
198 	 * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is
199 	 * a must. We will also check all VP numbers when walking the supplied
200 	 * CPU set to remain correct in all cases.
201 	 */
202 	if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64)
203 		goto do_ex_hypercall;
204 
205 	ipi_arg.vector = vector;
206 	ipi_arg.cpu_mask = 0;
207 
208 	for_each_cpu(cur_cpu, mask) {
209 		if (exclude_self && cur_cpu == this_cpu)
210 			continue;
211 		vcpu = hv_cpu_number_to_vp_number(cur_cpu);
212 		if (vcpu == VP_INVAL)
213 			return false;
214 
215 		/*
216 		 * This particular version of the IPI hypercall can
217 		 * only target up to 64 CPUs.
218 		 */
219 		if (vcpu >= 64)
220 			goto do_ex_hypercall;
221 
222 		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
223 	}
224 
225 	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
226 					ipi_arg.cpu_mask);
227 	return hv_result_success(status);
228 
229 do_ex_hypercall:
230 	return __send_ipi_mask_ex(mask, vector, exclude_self);
231 }
232 
233 static bool __send_ipi_one(int cpu, int vector)
234 {
235 	int vp = hv_cpu_number_to_vp_number(cpu);
236 	u64 status;
237 
238 	trace_hyperv_send_ipi_one(cpu, vector);
239 
240 	if (vp == VP_INVAL)
241 		return false;
242 
243 	/* A fully enlightened TDX VM uses GHCI rather than hv_hypercall_pg. */
244 	if (!hv_hypercall_pg) {
245 		if (ms_hyperv.paravisor_present || !hv_isolation_type_tdx())
246 			return false;
247 	}
248 
249 	if (vector < HV_IPI_LOW_VECTOR || vector > HV_IPI_HIGH_VECTOR)
250 		return false;
251 
252 	if (vp >= 64)
253 		return __send_ipi_mask_ex(cpumask_of(cpu), vector, false);
254 
255 	status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
256 	return hv_result_success(status);
257 }
258 
259 static void hv_send_ipi(int cpu, int vector)
260 {
261 	if (!__send_ipi_one(cpu, vector))
262 		orig_apic.send_IPI(cpu, vector);
263 }
264 
265 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
266 {
267 	if (!__send_ipi_mask(mask, vector, false))
268 		orig_apic.send_IPI_mask(mask, vector);
269 }
270 
271 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
272 {
273 	if (!__send_ipi_mask(mask, vector, true))
274 		orig_apic.send_IPI_mask_allbutself(mask, vector);
275 }
276 
277 static void hv_send_ipi_allbutself(int vector)
278 {
279 	hv_send_ipi_mask_allbutself(cpu_online_mask, vector);
280 }
281 
282 static void hv_send_ipi_all(int vector)
283 {
284 	if (!__send_ipi_mask(cpu_online_mask, vector, false))
285 		orig_apic.send_IPI_all(vector);
286 }
287 
288 static void hv_send_ipi_self(int vector)
289 {
290 	if (!__send_ipi_one(smp_processor_id(), vector))
291 		orig_apic.send_IPI_self(vector);
292 }
293 
294 void __init hv_apic_init(void)
295 {
296 	if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) {
297 		pr_info("Hyper-V: Using IPI hypercalls\n");
298 		/*
299 		 * Set the IPI entry points.
300 		 */
301 		orig_apic = *apic;
302 
303 		apic_update_callback(send_IPI, hv_send_ipi);
304 		apic_update_callback(send_IPI_mask, hv_send_ipi_mask);
305 		apic_update_callback(send_IPI_mask_allbutself, hv_send_ipi_mask_allbutself);
306 		apic_update_callback(send_IPI_allbutself, hv_send_ipi_allbutself);
307 		apic_update_callback(send_IPI_all, hv_send_ipi_all);
308 		apic_update_callback(send_IPI_self, hv_send_ipi_self);
309 	}
310 
311 	if (ms_hyperv.hints & HV_X64_APIC_ACCESS_RECOMMENDED) {
312 		pr_info("Hyper-V: Using enlightened APIC (%s mode)",
313 			x2apic_enabled() ? "x2apic" : "xapic");
314 		/*
315 		 * When in x2apic mode, don't use the Hyper-V specific APIC
316 		 * accessors since the field layout in the ICR register is
317 		 * different in x2apic mode. Furthermore, the architectural
318 		 * x2apic MSRs function just as well as the Hyper-V
319 		 * synthetic APIC MSRs, so there's no benefit in having
320 		 * separate Hyper-V accessors for x2apic mode. The only
321 		 * exception is hv_apic_eoi_write, because it benefits from
322 		 * lazy EOI when available, but the same accessor works for
323 		 * both xapic and x2apic because the field layout is the same.
324 		 */
325 		apic_update_callback(eoi, hv_apic_eoi_write);
326 		if (!x2apic_enabled()) {
327 			apic_update_callback(read, hv_apic_read);
328 			apic_update_callback(write, hv_apic_write);
329 			apic_update_callback(icr_write, hv_apic_icr_write);
330 			apic_update_callback(icr_read, hv_apic_icr_read);
331 		}
332 	}
333 }
334