Merge tag 'linux-kselftest-fixes-5.15-rc5' of git://git.kernel.org/pub/scm/linux...
[platform/kernel/linux-rpi.git] / arch / x86 / hyperv / hv_apic.c
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                 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         if (!cpumask_equal(mask, cpu_present_mask)) {
126                 ipi_arg->vp_set.format = HV_GENERIC_SET_SPARSE_4K;
127                 if (exclude_self)
128                         nr_bank = cpumask_to_vpset_noself(&(ipi_arg->vp_set), mask);
129                 else
130                         nr_bank = cpumask_to_vpset(&(ipi_arg->vp_set), mask);
131         }
132         if (nr_bank < 0)
133                 goto ipi_mask_ex_done;
134         if (!nr_bank)
135                 ipi_arg->vp_set.format = HV_GENERIC_SET_ALL;
136
137         status = hv_do_rep_hypercall(HVCALL_SEND_IPI_EX, 0, nr_bank,
138                               ipi_arg, NULL);
139
140 ipi_mask_ex_done:
141         local_irq_restore(flags);
142         return hv_result_success(status);
143 }
144
145 static bool __send_ipi_mask(const struct cpumask *mask, int vector,
146                 bool exclude_self)
147 {
148         int cur_cpu, vcpu, this_cpu = smp_processor_id();
149         struct hv_send_ipi ipi_arg;
150         u64 status;
151         unsigned int weight;
152
153         trace_hyperv_send_ipi_mask(mask, vector);
154
155         weight = cpumask_weight(mask);
156
157         /*
158          * Do nothing if
159          *   1. the mask is empty
160          *   2. the mask only contains self when exclude_self is true
161          */
162         if (weight == 0 ||
163             (exclude_self && weight == 1 && cpumask_test_cpu(this_cpu, mask)))
164                 return true;
165
166         if (!hv_hypercall_pg)
167                 return false;
168
169         if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
170                 return false;
171
172         /*
173          * From the supplied CPU set we need to figure out if we can get away
174          * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the
175          * highest VP number in the set is < 64. As VP numbers are usually in
176          * ascending order and match Linux CPU ids, here is an optimization:
177          * we check the VP number for the highest bit in the supplied set first
178          * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is
179          * a must. We will also check all VP numbers when walking the supplied
180          * CPU set to remain correct in all cases.
181          */
182         if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64)
183                 goto do_ex_hypercall;
184
185         ipi_arg.vector = vector;
186         ipi_arg.cpu_mask = 0;
187
188         for_each_cpu(cur_cpu, mask) {
189                 if (exclude_self && cur_cpu == this_cpu)
190                         continue;
191                 vcpu = hv_cpu_number_to_vp_number(cur_cpu);
192                 if (vcpu == VP_INVAL)
193                         return false;
194
195                 /*
196                  * This particular version of the IPI hypercall can
197                  * only target upto 64 CPUs.
198                  */
199                 if (vcpu >= 64)
200                         goto do_ex_hypercall;
201
202                 __set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
203         }
204
205         status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
206                                      ipi_arg.cpu_mask);
207         return hv_result_success(status);
208
209 do_ex_hypercall:
210         return __send_ipi_mask_ex(mask, vector, exclude_self);
211 }
212
213 static bool __send_ipi_one(int cpu, int vector)
214 {
215         int vp = hv_cpu_number_to_vp_number(cpu);
216         u64 status;
217
218         trace_hyperv_send_ipi_one(cpu, vector);
219
220         if (!hv_hypercall_pg || (vp == VP_INVAL))
221                 return false;
222
223         if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
224                 return false;
225
226         if (vp >= 64)
227                 return __send_ipi_mask_ex(cpumask_of(cpu), vector, false);
228
229         status = hv_do_fast_hypercall16(HVCALL_SEND_IPI, vector, BIT_ULL(vp));
230         return hv_result_success(status);
231 }
232
233 static void hv_send_ipi(int cpu, int vector)
234 {
235         if (!__send_ipi_one(cpu, vector))
236                 orig_apic.send_IPI(cpu, vector);
237 }
238
239 static void hv_send_ipi_mask(const struct cpumask *mask, int vector)
240 {
241         if (!__send_ipi_mask(mask, vector, false))
242                 orig_apic.send_IPI_mask(mask, vector);
243 }
244
245 static void hv_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
246 {
247         if (!__send_ipi_mask(mask, vector, true))
248                 orig_apic.send_IPI_mask_allbutself(mask, vector);
249 }
250
251 static void hv_send_ipi_allbutself(int vector)
252 {
253         hv_send_ipi_mask_allbutself(cpu_online_mask, vector);
254 }
255
256 static void hv_send_ipi_all(int vector)
257 {
258         if (!__send_ipi_mask(cpu_online_mask, vector, false))
259                 orig_apic.send_IPI_all(vector);
260 }
261
262 static void hv_send_ipi_self(int vector)
263 {
264         if (!__send_ipi_one(smp_processor_id(), vector))
265                 orig_apic.send_IPI_self(vector);
266 }
267
268 void __init hv_apic_init(void)
269 {
270         if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) {
271                 pr_info("Hyper-V: Using IPI hypercalls\n");
272                 /*
273                  * Set the IPI entry points.
274                  */
275                 orig_apic = *apic;
276
277                 apic->send_IPI = hv_send_ipi;
278                 apic->send_IPI_mask = hv_send_ipi_mask;
279                 apic->send_IPI_mask_allbutself = hv_send_ipi_mask_allbutself;
280                 apic->send_IPI_allbutself = hv_send_ipi_allbutself;
281                 apic->send_IPI_all = hv_send_ipi_all;
282                 apic->send_IPI_self = hv_send_ipi_self;
283         }
284
285         if (ms_hyperv.hints & HV_X64_APIC_ACCESS_RECOMMENDED) {
286                 pr_info("Hyper-V: Using enlightened APIC (%s mode)",
287                         x2apic_enabled() ? "x2apic" : "xapic");
288                 /*
289                  * When in x2apic mode, don't use the Hyper-V specific APIC
290                  * accessors since the field layout in the ICR register is
291                  * different in x2apic mode. Furthermore, the architectural
292                  * x2apic MSRs function just as well as the Hyper-V
293                  * synthetic APIC MSRs, so there's no benefit in having
294                  * separate Hyper-V accessors for x2apic mode. The only
295                  * exception is hv_apic_eoi_write, because it benefits from
296                  * lazy EOI when available, but the same accessor works for
297                  * both xapic and x2apic because the field layout is the same.
298                  */
299                 apic_set_eoi_write(hv_apic_eoi_write);
300                 if (!x2apic_enabled()) {
301                         apic->read      = hv_apic_read;
302                         apic->write     = hv_apic_write;
303                         apic->icr_write = hv_apic_icr_write;
304                         apic->icr_read  = hv_apic_icr_read;
305                 }
306         }
307 }