arm64: perf: Don't register user access sysctl handler multiple times
[platform/kernel/linux-starfive.git] / arch / arm64 / kernel / perf_event.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * ARMv8 PMUv3 Performance Events handling code.
4  *
5  * Copyright (C) 2012 ARM Limited
6  * Author: Will Deacon <will.deacon@arm.com>
7  *
8  * This code is based heavily on the ARMv7 perf event code.
9  */
10
11 #include <asm/irq_regs.h>
12 #include <asm/perf_event.h>
13 #include <asm/sysreg.h>
14 #include <asm/virt.h>
15
16 #include <clocksource/arm_arch_timer.h>
17
18 #include <linux/acpi.h>
19 #include <linux/clocksource.h>
20 #include <linux/kvm_host.h>
21 #include <linux/of.h>
22 #include <linux/perf/arm_pmu.h>
23 #include <linux/platform_device.h>
24 #include <linux/sched_clock.h>
25 #include <linux/smp.h>
26
27 /* ARMv8 Cortex-A53 specific event types. */
28 #define ARMV8_A53_PERFCTR_PREF_LINEFILL                         0xC2
29
30 /* ARMv8 Cavium ThunderX specific event types. */
31 #define ARMV8_THUNDER_PERFCTR_L1D_CACHE_MISS_ST                 0xE9
32 #define ARMV8_THUNDER_PERFCTR_L1D_CACHE_PREF_ACCESS             0xEA
33 #define ARMV8_THUNDER_PERFCTR_L1D_CACHE_PREF_MISS               0xEB
34 #define ARMV8_THUNDER_PERFCTR_L1I_CACHE_PREF_ACCESS             0xEC
35 #define ARMV8_THUNDER_PERFCTR_L1I_CACHE_PREF_MISS               0xED
36
37 /*
38  * ARMv8 Architectural defined events, not all of these may
39  * be supported on any given implementation. Unsupported events will
40  * be disabled at run-time based on the PMCEID registers.
41  */
42 static const unsigned armv8_pmuv3_perf_map[PERF_COUNT_HW_MAX] = {
43         PERF_MAP_ALL_UNSUPPORTED,
44         [PERF_COUNT_HW_CPU_CYCLES]              = ARMV8_PMUV3_PERFCTR_CPU_CYCLES,
45         [PERF_COUNT_HW_INSTRUCTIONS]            = ARMV8_PMUV3_PERFCTR_INST_RETIRED,
46         [PERF_COUNT_HW_CACHE_REFERENCES]        = ARMV8_PMUV3_PERFCTR_L1D_CACHE,
47         [PERF_COUNT_HW_CACHE_MISSES]            = ARMV8_PMUV3_PERFCTR_L1D_CACHE_REFILL,
48         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]     = ARMV8_PMUV3_PERFCTR_PC_WRITE_RETIRED,
49         [PERF_COUNT_HW_BRANCH_MISSES]           = ARMV8_PMUV3_PERFCTR_BR_MIS_PRED,
50         [PERF_COUNT_HW_BUS_CYCLES]              = ARMV8_PMUV3_PERFCTR_BUS_CYCLES,
51         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV8_PMUV3_PERFCTR_STALL_FRONTEND,
52         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND]  = ARMV8_PMUV3_PERFCTR_STALL_BACKEND,
53 };
54
55 static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
56                                                 [PERF_COUNT_HW_CACHE_OP_MAX]
57                                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
58         PERF_CACHE_MAP_ALL_UNSUPPORTED,
59
60         [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_PMUV3_PERFCTR_L1D_CACHE,
61         [C(L1D)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_PMUV3_PERFCTR_L1D_CACHE_REFILL,
62
63         [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_PMUV3_PERFCTR_L1I_CACHE,
64         [C(L1I)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_PMUV3_PERFCTR_L1I_CACHE_REFILL,
65
66         [C(DTLB)][C(OP_READ)][C(RESULT_MISS)]   = ARMV8_PMUV3_PERFCTR_L1D_TLB_REFILL,
67         [C(DTLB)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1D_TLB,
68
69         [C(ITLB)][C(OP_READ)][C(RESULT_MISS)]   = ARMV8_PMUV3_PERFCTR_L1I_TLB_REFILL,
70         [C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1I_TLB,
71
72         [C(LL)][C(OP_READ)][C(RESULT_MISS)]     = ARMV8_PMUV3_PERFCTR_LL_CACHE_MISS_RD,
73         [C(LL)][C(OP_READ)][C(RESULT_ACCESS)]   = ARMV8_PMUV3_PERFCTR_LL_CACHE_RD,
74
75         [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_PMUV3_PERFCTR_BR_PRED,
76         [C(BPU)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_PMUV3_PERFCTR_BR_MIS_PRED,
77 };
78
79 static const unsigned armv8_a53_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
80                                               [PERF_COUNT_HW_CACHE_OP_MAX]
81                                               [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
82         PERF_CACHE_MAP_ALL_UNSUPPORTED,
83
84         [C(L1D)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV8_A53_PERFCTR_PREF_LINEFILL,
85
86         [C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_RD,
87         [C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_WR,
88 };
89
90 static const unsigned armv8_a57_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
91                                               [PERF_COUNT_HW_CACHE_OP_MAX]
92                                               [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
93         PERF_CACHE_MAP_ALL_UNSUPPORTED,
94
95         [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_RD,
96         [C(L1D)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_REFILL_RD,
97         [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_WR,
98         [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)]   = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_REFILL_WR,
99
100         [C(DTLB)][C(OP_READ)][C(RESULT_MISS)]   = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_RD,
101         [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)]  = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_WR,
102
103         [C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_RD,
104         [C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_WR,
105 };
106
107 static const unsigned armv8_a73_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
108                                               [PERF_COUNT_HW_CACHE_OP_MAX]
109                                               [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
110         PERF_CACHE_MAP_ALL_UNSUPPORTED,
111
112         [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_RD,
113         [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_WR,
114 };
115
116 static const unsigned armv8_thunder_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
117                                                    [PERF_COUNT_HW_CACHE_OP_MAX]
118                                                    [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
119         PERF_CACHE_MAP_ALL_UNSUPPORTED,
120
121         [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_RD,
122         [C(L1D)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_REFILL_RD,
123         [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_WR,
124         [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)]   = ARMV8_THUNDER_PERFCTR_L1D_CACHE_MISS_ST,
125         [C(L1D)][C(OP_PREFETCH)][C(RESULT_ACCESS)] = ARMV8_THUNDER_PERFCTR_L1D_CACHE_PREF_ACCESS,
126         [C(L1D)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV8_THUNDER_PERFCTR_L1D_CACHE_PREF_MISS,
127
128         [C(L1I)][C(OP_PREFETCH)][C(RESULT_ACCESS)] = ARMV8_THUNDER_PERFCTR_L1I_CACHE_PREF_ACCESS,
129         [C(L1I)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV8_THUNDER_PERFCTR_L1I_CACHE_PREF_MISS,
130
131         [C(DTLB)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_TLB_RD,
132         [C(DTLB)][C(OP_READ)][C(RESULT_MISS)]   = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_RD,
133         [C(DTLB)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_TLB_WR,
134         [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)]  = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_WR,
135 };
136
137 static const unsigned armv8_vulcan_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
138                                               [PERF_COUNT_HW_CACHE_OP_MAX]
139                                               [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
140         PERF_CACHE_MAP_ALL_UNSUPPORTED,
141
142         [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)]  = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_RD,
143         [C(L1D)][C(OP_READ)][C(RESULT_MISS)]    = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_REFILL_RD,
144         [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_WR,
145         [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)]   = ARMV8_IMPDEF_PERFCTR_L1D_CACHE_REFILL_WR,
146
147         [C(DTLB)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_TLB_RD,
148         [C(DTLB)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_L1D_TLB_WR,
149         [C(DTLB)][C(OP_READ)][C(RESULT_MISS)]   = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_RD,
150         [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)]  = ARMV8_IMPDEF_PERFCTR_L1D_TLB_REFILL_WR,
151
152         [C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_RD,
153         [C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV8_IMPDEF_PERFCTR_BUS_ACCESS_WR,
154 };
155
156 static ssize_t
157 armv8pmu_events_sysfs_show(struct device *dev,
158                            struct device_attribute *attr, char *page)
159 {
160         struct perf_pmu_events_attr *pmu_attr;
161
162         pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
163
164         return sprintf(page, "event=0x%04llx\n", pmu_attr->id);
165 }
166
167 #define ARMV8_EVENT_ATTR(name, config)                                          \
168         PMU_EVENT_ATTR_ID(name, armv8pmu_events_sysfs_show, config)
169
170 static struct attribute *armv8_pmuv3_event_attrs[] = {
171         ARMV8_EVENT_ATTR(sw_incr, ARMV8_PMUV3_PERFCTR_SW_INCR),
172         ARMV8_EVENT_ATTR(l1i_cache_refill, ARMV8_PMUV3_PERFCTR_L1I_CACHE_REFILL),
173         ARMV8_EVENT_ATTR(l1i_tlb_refill, ARMV8_PMUV3_PERFCTR_L1I_TLB_REFILL),
174         ARMV8_EVENT_ATTR(l1d_cache_refill, ARMV8_PMUV3_PERFCTR_L1D_CACHE_REFILL),
175         ARMV8_EVENT_ATTR(l1d_cache, ARMV8_PMUV3_PERFCTR_L1D_CACHE),
176         ARMV8_EVENT_ATTR(l1d_tlb_refill, ARMV8_PMUV3_PERFCTR_L1D_TLB_REFILL),
177         ARMV8_EVENT_ATTR(ld_retired, ARMV8_PMUV3_PERFCTR_LD_RETIRED),
178         ARMV8_EVENT_ATTR(st_retired, ARMV8_PMUV3_PERFCTR_ST_RETIRED),
179         ARMV8_EVENT_ATTR(inst_retired, ARMV8_PMUV3_PERFCTR_INST_RETIRED),
180         ARMV8_EVENT_ATTR(exc_taken, ARMV8_PMUV3_PERFCTR_EXC_TAKEN),
181         ARMV8_EVENT_ATTR(exc_return, ARMV8_PMUV3_PERFCTR_EXC_RETURN),
182         ARMV8_EVENT_ATTR(cid_write_retired, ARMV8_PMUV3_PERFCTR_CID_WRITE_RETIRED),
183         ARMV8_EVENT_ATTR(pc_write_retired, ARMV8_PMUV3_PERFCTR_PC_WRITE_RETIRED),
184         ARMV8_EVENT_ATTR(br_immed_retired, ARMV8_PMUV3_PERFCTR_BR_IMMED_RETIRED),
185         ARMV8_EVENT_ATTR(br_return_retired, ARMV8_PMUV3_PERFCTR_BR_RETURN_RETIRED),
186         ARMV8_EVENT_ATTR(unaligned_ldst_retired, ARMV8_PMUV3_PERFCTR_UNALIGNED_LDST_RETIRED),
187         ARMV8_EVENT_ATTR(br_mis_pred, ARMV8_PMUV3_PERFCTR_BR_MIS_PRED),
188         ARMV8_EVENT_ATTR(cpu_cycles, ARMV8_PMUV3_PERFCTR_CPU_CYCLES),
189         ARMV8_EVENT_ATTR(br_pred, ARMV8_PMUV3_PERFCTR_BR_PRED),
190         ARMV8_EVENT_ATTR(mem_access, ARMV8_PMUV3_PERFCTR_MEM_ACCESS),
191         ARMV8_EVENT_ATTR(l1i_cache, ARMV8_PMUV3_PERFCTR_L1I_CACHE),
192         ARMV8_EVENT_ATTR(l1d_cache_wb, ARMV8_PMUV3_PERFCTR_L1D_CACHE_WB),
193         ARMV8_EVENT_ATTR(l2d_cache, ARMV8_PMUV3_PERFCTR_L2D_CACHE),
194         ARMV8_EVENT_ATTR(l2d_cache_refill, ARMV8_PMUV3_PERFCTR_L2D_CACHE_REFILL),
195         ARMV8_EVENT_ATTR(l2d_cache_wb, ARMV8_PMUV3_PERFCTR_L2D_CACHE_WB),
196         ARMV8_EVENT_ATTR(bus_access, ARMV8_PMUV3_PERFCTR_BUS_ACCESS),
197         ARMV8_EVENT_ATTR(memory_error, ARMV8_PMUV3_PERFCTR_MEMORY_ERROR),
198         ARMV8_EVENT_ATTR(inst_spec, ARMV8_PMUV3_PERFCTR_INST_SPEC),
199         ARMV8_EVENT_ATTR(ttbr_write_retired, ARMV8_PMUV3_PERFCTR_TTBR_WRITE_RETIRED),
200         ARMV8_EVENT_ATTR(bus_cycles, ARMV8_PMUV3_PERFCTR_BUS_CYCLES),
201         /* Don't expose the chain event in /sys, since it's useless in isolation */
202         ARMV8_EVENT_ATTR(l1d_cache_allocate, ARMV8_PMUV3_PERFCTR_L1D_CACHE_ALLOCATE),
203         ARMV8_EVENT_ATTR(l2d_cache_allocate, ARMV8_PMUV3_PERFCTR_L2D_CACHE_ALLOCATE),
204         ARMV8_EVENT_ATTR(br_retired, ARMV8_PMUV3_PERFCTR_BR_RETIRED),
205         ARMV8_EVENT_ATTR(br_mis_pred_retired, ARMV8_PMUV3_PERFCTR_BR_MIS_PRED_RETIRED),
206         ARMV8_EVENT_ATTR(stall_frontend, ARMV8_PMUV3_PERFCTR_STALL_FRONTEND),
207         ARMV8_EVENT_ATTR(stall_backend, ARMV8_PMUV3_PERFCTR_STALL_BACKEND),
208         ARMV8_EVENT_ATTR(l1d_tlb, ARMV8_PMUV3_PERFCTR_L1D_TLB),
209         ARMV8_EVENT_ATTR(l1i_tlb, ARMV8_PMUV3_PERFCTR_L1I_TLB),
210         ARMV8_EVENT_ATTR(l2i_cache, ARMV8_PMUV3_PERFCTR_L2I_CACHE),
211         ARMV8_EVENT_ATTR(l2i_cache_refill, ARMV8_PMUV3_PERFCTR_L2I_CACHE_REFILL),
212         ARMV8_EVENT_ATTR(l3d_cache_allocate, ARMV8_PMUV3_PERFCTR_L3D_CACHE_ALLOCATE),
213         ARMV8_EVENT_ATTR(l3d_cache_refill, ARMV8_PMUV3_PERFCTR_L3D_CACHE_REFILL),
214         ARMV8_EVENT_ATTR(l3d_cache, ARMV8_PMUV3_PERFCTR_L3D_CACHE),
215         ARMV8_EVENT_ATTR(l3d_cache_wb, ARMV8_PMUV3_PERFCTR_L3D_CACHE_WB),
216         ARMV8_EVENT_ATTR(l2d_tlb_refill, ARMV8_PMUV3_PERFCTR_L2D_TLB_REFILL),
217         ARMV8_EVENT_ATTR(l2i_tlb_refill, ARMV8_PMUV3_PERFCTR_L2I_TLB_REFILL),
218         ARMV8_EVENT_ATTR(l2d_tlb, ARMV8_PMUV3_PERFCTR_L2D_TLB),
219         ARMV8_EVENT_ATTR(l2i_tlb, ARMV8_PMUV3_PERFCTR_L2I_TLB),
220         ARMV8_EVENT_ATTR(remote_access, ARMV8_PMUV3_PERFCTR_REMOTE_ACCESS),
221         ARMV8_EVENT_ATTR(ll_cache, ARMV8_PMUV3_PERFCTR_LL_CACHE),
222         ARMV8_EVENT_ATTR(ll_cache_miss, ARMV8_PMUV3_PERFCTR_LL_CACHE_MISS),
223         ARMV8_EVENT_ATTR(dtlb_walk, ARMV8_PMUV3_PERFCTR_DTLB_WALK),
224         ARMV8_EVENT_ATTR(itlb_walk, ARMV8_PMUV3_PERFCTR_ITLB_WALK),
225         ARMV8_EVENT_ATTR(ll_cache_rd, ARMV8_PMUV3_PERFCTR_LL_CACHE_RD),
226         ARMV8_EVENT_ATTR(ll_cache_miss_rd, ARMV8_PMUV3_PERFCTR_LL_CACHE_MISS_RD),
227         ARMV8_EVENT_ATTR(remote_access_rd, ARMV8_PMUV3_PERFCTR_REMOTE_ACCESS_RD),
228         ARMV8_EVENT_ATTR(l1d_cache_lmiss_rd, ARMV8_PMUV3_PERFCTR_L1D_CACHE_LMISS_RD),
229         ARMV8_EVENT_ATTR(op_retired, ARMV8_PMUV3_PERFCTR_OP_RETIRED),
230         ARMV8_EVENT_ATTR(op_spec, ARMV8_PMUV3_PERFCTR_OP_SPEC),
231         ARMV8_EVENT_ATTR(stall, ARMV8_PMUV3_PERFCTR_STALL),
232         ARMV8_EVENT_ATTR(stall_slot_backend, ARMV8_PMUV3_PERFCTR_STALL_SLOT_BACKEND),
233         ARMV8_EVENT_ATTR(stall_slot_frontend, ARMV8_PMUV3_PERFCTR_STALL_SLOT_FRONTEND),
234         ARMV8_EVENT_ATTR(stall_slot, ARMV8_PMUV3_PERFCTR_STALL_SLOT),
235         ARMV8_EVENT_ATTR(sample_pop, ARMV8_SPE_PERFCTR_SAMPLE_POP),
236         ARMV8_EVENT_ATTR(sample_feed, ARMV8_SPE_PERFCTR_SAMPLE_FEED),
237         ARMV8_EVENT_ATTR(sample_filtrate, ARMV8_SPE_PERFCTR_SAMPLE_FILTRATE),
238         ARMV8_EVENT_ATTR(sample_collision, ARMV8_SPE_PERFCTR_SAMPLE_COLLISION),
239         ARMV8_EVENT_ATTR(cnt_cycles, ARMV8_AMU_PERFCTR_CNT_CYCLES),
240         ARMV8_EVENT_ATTR(stall_backend_mem, ARMV8_AMU_PERFCTR_STALL_BACKEND_MEM),
241         ARMV8_EVENT_ATTR(l1i_cache_lmiss, ARMV8_PMUV3_PERFCTR_L1I_CACHE_LMISS),
242         ARMV8_EVENT_ATTR(l2d_cache_lmiss_rd, ARMV8_PMUV3_PERFCTR_L2D_CACHE_LMISS_RD),
243         ARMV8_EVENT_ATTR(l2i_cache_lmiss, ARMV8_PMUV3_PERFCTR_L2I_CACHE_LMISS),
244         ARMV8_EVENT_ATTR(l3d_cache_lmiss_rd, ARMV8_PMUV3_PERFCTR_L3D_CACHE_LMISS_RD),
245         ARMV8_EVENT_ATTR(ldst_align_lat, ARMV8_PMUV3_PERFCTR_LDST_ALIGN_LAT),
246         ARMV8_EVENT_ATTR(ld_align_lat, ARMV8_PMUV3_PERFCTR_LD_ALIGN_LAT),
247         ARMV8_EVENT_ATTR(st_align_lat, ARMV8_PMUV3_PERFCTR_ST_ALIGN_LAT),
248         ARMV8_EVENT_ATTR(mem_access_checked, ARMV8_MTE_PERFCTR_MEM_ACCESS_CHECKED),
249         ARMV8_EVENT_ATTR(mem_access_checked_rd, ARMV8_MTE_PERFCTR_MEM_ACCESS_CHECKED_RD),
250         ARMV8_EVENT_ATTR(mem_access_checked_wr, ARMV8_MTE_PERFCTR_MEM_ACCESS_CHECKED_WR),
251         NULL,
252 };
253
254 static umode_t
255 armv8pmu_event_attr_is_visible(struct kobject *kobj,
256                                struct attribute *attr, int unused)
257 {
258         struct device *dev = kobj_to_dev(kobj);
259         struct pmu *pmu = dev_get_drvdata(dev);
260         struct arm_pmu *cpu_pmu = container_of(pmu, struct arm_pmu, pmu);
261         struct perf_pmu_events_attr *pmu_attr;
262
263         pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
264
265         if (pmu_attr->id < ARMV8_PMUV3_MAX_COMMON_EVENTS &&
266             test_bit(pmu_attr->id, cpu_pmu->pmceid_bitmap))
267                 return attr->mode;
268
269         if (pmu_attr->id >= ARMV8_PMUV3_EXT_COMMON_EVENT_BASE) {
270                 u64 id = pmu_attr->id - ARMV8_PMUV3_EXT_COMMON_EVENT_BASE;
271
272                 if (id < ARMV8_PMUV3_MAX_COMMON_EVENTS &&
273                     test_bit(id, cpu_pmu->pmceid_ext_bitmap))
274                         return attr->mode;
275         }
276
277         return 0;
278 }
279
280 static const struct attribute_group armv8_pmuv3_events_attr_group = {
281         .name = "events",
282         .attrs = armv8_pmuv3_event_attrs,
283         .is_visible = armv8pmu_event_attr_is_visible,
284 };
285
286 PMU_FORMAT_ATTR(event, "config:0-15");
287 PMU_FORMAT_ATTR(long, "config1:0");
288 PMU_FORMAT_ATTR(rdpmc, "config1:1");
289
290 static int sysctl_perf_user_access __read_mostly;
291
292 static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
293 {
294         return event->attr.config1 & 0x1;
295 }
296
297 static inline bool armv8pmu_event_want_user_access(struct perf_event *event)
298 {
299         return event->attr.config1 & 0x2;
300 }
301
302 static struct attribute *armv8_pmuv3_format_attrs[] = {
303         &format_attr_event.attr,
304         &format_attr_long.attr,
305         &format_attr_rdpmc.attr,
306         NULL,
307 };
308
309 static const struct attribute_group armv8_pmuv3_format_attr_group = {
310         .name = "format",
311         .attrs = armv8_pmuv3_format_attrs,
312 };
313
314 static ssize_t slots_show(struct device *dev, struct device_attribute *attr,
315                           char *page)
316 {
317         struct pmu *pmu = dev_get_drvdata(dev);
318         struct arm_pmu *cpu_pmu = container_of(pmu, struct arm_pmu, pmu);
319         u32 slots = cpu_pmu->reg_pmmir & ARMV8_PMU_SLOTS_MASK;
320
321         return sysfs_emit(page, "0x%08x\n", slots);
322 }
323
324 static DEVICE_ATTR_RO(slots);
325
326 static ssize_t bus_slots_show(struct device *dev, struct device_attribute *attr,
327                               char *page)
328 {
329         struct pmu *pmu = dev_get_drvdata(dev);
330         struct arm_pmu *cpu_pmu = container_of(pmu, struct arm_pmu, pmu);
331         u32 bus_slots = (cpu_pmu->reg_pmmir >> ARMV8_PMU_BUS_SLOTS_SHIFT)
332                         & ARMV8_PMU_BUS_SLOTS_MASK;
333
334         return sysfs_emit(page, "0x%08x\n", bus_slots);
335 }
336
337 static DEVICE_ATTR_RO(bus_slots);
338
339 static ssize_t bus_width_show(struct device *dev, struct device_attribute *attr,
340                               char *page)
341 {
342         struct pmu *pmu = dev_get_drvdata(dev);
343         struct arm_pmu *cpu_pmu = container_of(pmu, struct arm_pmu, pmu);
344         u32 bus_width = (cpu_pmu->reg_pmmir >> ARMV8_PMU_BUS_WIDTH_SHIFT)
345                         & ARMV8_PMU_BUS_WIDTH_MASK;
346         u32 val = 0;
347
348         /* Encoded as Log2(number of bytes), plus one */
349         if (bus_width > 2 && bus_width < 13)
350                 val = 1 << (bus_width - 1);
351
352         return sysfs_emit(page, "0x%08x\n", val);
353 }
354
355 static DEVICE_ATTR_RO(bus_width);
356
357 static struct attribute *armv8_pmuv3_caps_attrs[] = {
358         &dev_attr_slots.attr,
359         &dev_attr_bus_slots.attr,
360         &dev_attr_bus_width.attr,
361         NULL,
362 };
363
364 static const struct attribute_group armv8_pmuv3_caps_attr_group = {
365         .name = "caps",
366         .attrs = armv8_pmuv3_caps_attrs,
367 };
368
369 /*
370  * Perf Events' indices
371  */
372 #define ARMV8_IDX_CYCLE_COUNTER 0
373 #define ARMV8_IDX_COUNTER0      1
374 #define ARMV8_IDX_CYCLE_COUNTER_USER    32
375
376 /*
377  * We unconditionally enable ARMv8.5-PMU long event counter support
378  * (64-bit events) where supported. Indicate if this arm_pmu has long
379  * event counter support.
380  */
381 static bool armv8pmu_has_long_event(struct arm_pmu *cpu_pmu)
382 {
383         return (cpu_pmu->pmuver >= ID_AA64DFR0_PMUVER_8_5);
384 }
385
386 static inline bool armv8pmu_event_has_user_read(struct perf_event *event)
387 {
388         return event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT;
389 }
390
391 /*
392  * We must chain two programmable counters for 64 bit events,
393  * except when we have allocated the 64bit cycle counter (for CPU
394  * cycles event) or when user space counter access is enabled.
395  */
396 static inline bool armv8pmu_event_is_chained(struct perf_event *event)
397 {
398         int idx = event->hw.idx;
399         struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
400
401         return !armv8pmu_event_has_user_read(event) &&
402                armv8pmu_event_is_64bit(event) &&
403                !armv8pmu_has_long_event(cpu_pmu) &&
404                (idx != ARMV8_IDX_CYCLE_COUNTER);
405 }
406
407 /*
408  * ARMv8 low level PMU access
409  */
410
411 /*
412  * Perf Event to low level counters mapping
413  */
414 #define ARMV8_IDX_TO_COUNTER(x) \
415         (((x) - ARMV8_IDX_COUNTER0) & ARMV8_PMU_COUNTER_MASK)
416
417 /*
418  * This code is really good
419  */
420
421 #define PMEVN_CASE(n, case_macro) \
422         case n: case_macro(n); break
423
424 #define PMEVN_SWITCH(x, case_macro)                             \
425         do {                                                    \
426                 switch (x) {                                    \
427                 PMEVN_CASE(0,  case_macro);                     \
428                 PMEVN_CASE(1,  case_macro);                     \
429                 PMEVN_CASE(2,  case_macro);                     \
430                 PMEVN_CASE(3,  case_macro);                     \
431                 PMEVN_CASE(4,  case_macro);                     \
432                 PMEVN_CASE(5,  case_macro);                     \
433                 PMEVN_CASE(6,  case_macro);                     \
434                 PMEVN_CASE(7,  case_macro);                     \
435                 PMEVN_CASE(8,  case_macro);                     \
436                 PMEVN_CASE(9,  case_macro);                     \
437                 PMEVN_CASE(10, case_macro);                     \
438                 PMEVN_CASE(11, case_macro);                     \
439                 PMEVN_CASE(12, case_macro);                     \
440                 PMEVN_CASE(13, case_macro);                     \
441                 PMEVN_CASE(14, case_macro);                     \
442                 PMEVN_CASE(15, case_macro);                     \
443                 PMEVN_CASE(16, case_macro);                     \
444                 PMEVN_CASE(17, case_macro);                     \
445                 PMEVN_CASE(18, case_macro);                     \
446                 PMEVN_CASE(19, case_macro);                     \
447                 PMEVN_CASE(20, case_macro);                     \
448                 PMEVN_CASE(21, case_macro);                     \
449                 PMEVN_CASE(22, case_macro);                     \
450                 PMEVN_CASE(23, case_macro);                     \
451                 PMEVN_CASE(24, case_macro);                     \
452                 PMEVN_CASE(25, case_macro);                     \
453                 PMEVN_CASE(26, case_macro);                     \
454                 PMEVN_CASE(27, case_macro);                     \
455                 PMEVN_CASE(28, case_macro);                     \
456                 PMEVN_CASE(29, case_macro);                     \
457                 PMEVN_CASE(30, case_macro);                     \
458                 default: WARN(1, "Invalid PMEV* index\n");      \
459                 }                                               \
460         } while (0)
461
462 #define RETURN_READ_PMEVCNTRN(n) \
463         return read_sysreg(pmevcntr##n##_el0)
464 static unsigned long read_pmevcntrn(int n)
465 {
466         PMEVN_SWITCH(n, RETURN_READ_PMEVCNTRN);
467         return 0;
468 }
469
470 #define WRITE_PMEVCNTRN(n) \
471         write_sysreg(val, pmevcntr##n##_el0)
472 static void write_pmevcntrn(int n, unsigned long val)
473 {
474         PMEVN_SWITCH(n, WRITE_PMEVCNTRN);
475 }
476
477 #define WRITE_PMEVTYPERN(n) \
478         write_sysreg(val, pmevtyper##n##_el0)
479 static void write_pmevtypern(int n, unsigned long val)
480 {
481         PMEVN_SWITCH(n, WRITE_PMEVTYPERN);
482 }
483
484 static inline u32 armv8pmu_pmcr_read(void)
485 {
486         return read_sysreg(pmcr_el0);
487 }
488
489 static inline void armv8pmu_pmcr_write(u32 val)
490 {
491         val &= ARMV8_PMU_PMCR_MASK;
492         isb();
493         write_sysreg(val, pmcr_el0);
494 }
495
496 static inline int armv8pmu_has_overflowed(u32 pmovsr)
497 {
498         return pmovsr & ARMV8_PMU_OVERFLOWED_MASK;
499 }
500
501 static inline int armv8pmu_counter_has_overflowed(u32 pmnc, int idx)
502 {
503         return pmnc & BIT(ARMV8_IDX_TO_COUNTER(idx));
504 }
505
506 static inline u64 armv8pmu_read_evcntr(int idx)
507 {
508         u32 counter = ARMV8_IDX_TO_COUNTER(idx);
509
510         return read_pmevcntrn(counter);
511 }
512
513 static inline u64 armv8pmu_read_hw_counter(struct perf_event *event)
514 {
515         int idx = event->hw.idx;
516         u64 val = armv8pmu_read_evcntr(idx);
517
518         if (armv8pmu_event_is_chained(event))
519                 val = (val << 32) | armv8pmu_read_evcntr(idx - 1);
520         return val;
521 }
522
523 /*
524  * The cycle counter is always a 64-bit counter. When ARMV8_PMU_PMCR_LP
525  * is set the event counters also become 64-bit counters. Unless the
526  * user has requested a long counter (attr.config1) then we want to
527  * interrupt upon 32-bit overflow - we achieve this by applying a bias.
528  */
529 static bool armv8pmu_event_needs_bias(struct perf_event *event)
530 {
531         struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
532         struct hw_perf_event *hwc = &event->hw;
533         int idx = hwc->idx;
534
535         if (armv8pmu_event_is_64bit(event))
536                 return false;
537
538         if (armv8pmu_has_long_event(cpu_pmu) ||
539             idx == ARMV8_IDX_CYCLE_COUNTER)
540                 return true;
541
542         return false;
543 }
544
545 static u64 armv8pmu_bias_long_counter(struct perf_event *event, u64 value)
546 {
547         if (armv8pmu_event_needs_bias(event))
548                 value |= GENMASK(63, 32);
549
550         return value;
551 }
552
553 static u64 armv8pmu_unbias_long_counter(struct perf_event *event, u64 value)
554 {
555         if (armv8pmu_event_needs_bias(event))
556                 value &= ~GENMASK(63, 32);
557
558         return value;
559 }
560
561 static u64 armv8pmu_read_counter(struct perf_event *event)
562 {
563         struct hw_perf_event *hwc = &event->hw;
564         int idx = hwc->idx;
565         u64 value;
566
567         if (idx == ARMV8_IDX_CYCLE_COUNTER)
568                 value = read_sysreg(pmccntr_el0);
569         else
570                 value = armv8pmu_read_hw_counter(event);
571
572         return  armv8pmu_unbias_long_counter(event, value);
573 }
574
575 static inline void armv8pmu_write_evcntr(int idx, u64 value)
576 {
577         u32 counter = ARMV8_IDX_TO_COUNTER(idx);
578
579         write_pmevcntrn(counter, value);
580 }
581
582 static inline void armv8pmu_write_hw_counter(struct perf_event *event,
583                                              u64 value)
584 {
585         int idx = event->hw.idx;
586
587         if (armv8pmu_event_is_chained(event)) {
588                 armv8pmu_write_evcntr(idx, upper_32_bits(value));
589                 armv8pmu_write_evcntr(idx - 1, lower_32_bits(value));
590         } else {
591                 armv8pmu_write_evcntr(idx, value);
592         }
593 }
594
595 static void armv8pmu_write_counter(struct perf_event *event, u64 value)
596 {
597         struct hw_perf_event *hwc = &event->hw;
598         int idx = hwc->idx;
599
600         value = armv8pmu_bias_long_counter(event, value);
601
602         if (idx == ARMV8_IDX_CYCLE_COUNTER)
603                 write_sysreg(value, pmccntr_el0);
604         else
605                 armv8pmu_write_hw_counter(event, value);
606 }
607
608 static inline void armv8pmu_write_evtype(int idx, u32 val)
609 {
610         u32 counter = ARMV8_IDX_TO_COUNTER(idx);
611
612         val &= ARMV8_PMU_EVTYPE_MASK;
613         write_pmevtypern(counter, val);
614 }
615
616 static inline void armv8pmu_write_event_type(struct perf_event *event)
617 {
618         struct hw_perf_event *hwc = &event->hw;
619         int idx = hwc->idx;
620
621         /*
622          * For chained events, the low counter is programmed to count
623          * the event of interest and the high counter is programmed
624          * with CHAIN event code with filters set to count at all ELs.
625          */
626         if (armv8pmu_event_is_chained(event)) {
627                 u32 chain_evt = ARMV8_PMUV3_PERFCTR_CHAIN |
628                                 ARMV8_PMU_INCLUDE_EL2;
629
630                 armv8pmu_write_evtype(idx - 1, hwc->config_base);
631                 armv8pmu_write_evtype(idx, chain_evt);
632         } else {
633                 if (idx == ARMV8_IDX_CYCLE_COUNTER)
634                         write_sysreg(hwc->config_base, pmccfiltr_el0);
635                 else
636                         armv8pmu_write_evtype(idx, hwc->config_base);
637         }
638 }
639
640 static u32 armv8pmu_event_cnten_mask(struct perf_event *event)
641 {
642         int counter = ARMV8_IDX_TO_COUNTER(event->hw.idx);
643         u32 mask = BIT(counter);
644
645         if (armv8pmu_event_is_chained(event))
646                 mask |= BIT(counter - 1);
647         return mask;
648 }
649
650 static inline void armv8pmu_enable_counter(u32 mask)
651 {
652         /*
653          * Make sure event configuration register writes are visible before we
654          * enable the counter.
655          * */
656         isb();
657         write_sysreg(mask, pmcntenset_el0);
658 }
659
660 static inline void armv8pmu_enable_event_counter(struct perf_event *event)
661 {
662         struct perf_event_attr *attr = &event->attr;
663         u32 mask = armv8pmu_event_cnten_mask(event);
664
665         kvm_set_pmu_events(mask, attr);
666
667         /* We rely on the hypervisor switch code to enable guest counters */
668         if (!kvm_pmu_counter_deferred(attr))
669                 armv8pmu_enable_counter(mask);
670 }
671
672 static inline void armv8pmu_disable_counter(u32 mask)
673 {
674         write_sysreg(mask, pmcntenclr_el0);
675         /*
676          * Make sure the effects of disabling the counter are visible before we
677          * start configuring the event.
678          */
679         isb();
680 }
681
682 static inline void armv8pmu_disable_event_counter(struct perf_event *event)
683 {
684         struct perf_event_attr *attr = &event->attr;
685         u32 mask = armv8pmu_event_cnten_mask(event);
686
687         kvm_clr_pmu_events(mask);
688
689         /* We rely on the hypervisor switch code to disable guest counters */
690         if (!kvm_pmu_counter_deferred(attr))
691                 armv8pmu_disable_counter(mask);
692 }
693
694 static inline void armv8pmu_enable_intens(u32 mask)
695 {
696         write_sysreg(mask, pmintenset_el1);
697 }
698
699 static inline void armv8pmu_enable_event_irq(struct perf_event *event)
700 {
701         u32 counter = ARMV8_IDX_TO_COUNTER(event->hw.idx);
702         armv8pmu_enable_intens(BIT(counter));
703 }
704
705 static inline void armv8pmu_disable_intens(u32 mask)
706 {
707         write_sysreg(mask, pmintenclr_el1);
708         isb();
709         /* Clear the overflow flag in case an interrupt is pending. */
710         write_sysreg(mask, pmovsclr_el0);
711         isb();
712 }
713
714 static inline void armv8pmu_disable_event_irq(struct perf_event *event)
715 {
716         u32 counter = ARMV8_IDX_TO_COUNTER(event->hw.idx);
717         armv8pmu_disable_intens(BIT(counter));
718 }
719
720 static inline u32 armv8pmu_getreset_flags(void)
721 {
722         u32 value;
723
724         /* Read */
725         value = read_sysreg(pmovsclr_el0);
726
727         /* Write to clear flags */
728         value &= ARMV8_PMU_OVSR_MASK;
729         write_sysreg(value, pmovsclr_el0);
730
731         return value;
732 }
733
734 static void armv8pmu_disable_user_access(void)
735 {
736         write_sysreg(0, pmuserenr_el0);
737 }
738
739 static void armv8pmu_enable_user_access(struct arm_pmu *cpu_pmu)
740 {
741         int i;
742         struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
743
744         /* Clear any unused counters to avoid leaking their contents */
745         for_each_clear_bit(i, cpuc->used_mask, cpu_pmu->num_events) {
746                 if (i == ARMV8_IDX_CYCLE_COUNTER)
747                         write_sysreg(0, pmccntr_el0);
748                 else
749                         armv8pmu_write_evcntr(i, 0);
750         }
751
752         write_sysreg(0, pmuserenr_el0);
753         write_sysreg(ARMV8_PMU_USERENR_ER | ARMV8_PMU_USERENR_CR, pmuserenr_el0);
754 }
755
756 static void armv8pmu_enable_event(struct perf_event *event)
757 {
758         /*
759          * Enable counter and interrupt, and set the counter to count
760          * the event that we're interested in.
761          */
762
763         /*
764          * Disable counter
765          */
766         armv8pmu_disable_event_counter(event);
767
768         /*
769          * Set event.
770          */
771         armv8pmu_write_event_type(event);
772
773         /*
774          * Enable interrupt for this counter
775          */
776         armv8pmu_enable_event_irq(event);
777
778         /*
779          * Enable counter
780          */
781         armv8pmu_enable_event_counter(event);
782 }
783
784 static void armv8pmu_disable_event(struct perf_event *event)
785 {
786         /*
787          * Disable counter
788          */
789         armv8pmu_disable_event_counter(event);
790
791         /*
792          * Disable interrupt for this counter
793          */
794         armv8pmu_disable_event_irq(event);
795 }
796
797 static void armv8pmu_start(struct arm_pmu *cpu_pmu)
798 {
799         struct perf_event_context *task_ctx =
800                 this_cpu_ptr(cpu_pmu->pmu.pmu_cpu_context)->task_ctx;
801
802         if (sysctl_perf_user_access && task_ctx && task_ctx->nr_user)
803                 armv8pmu_enable_user_access(cpu_pmu);
804         else
805                 armv8pmu_disable_user_access();
806
807         /* Enable all counters */
808         armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMU_PMCR_E);
809 }
810
811 static void armv8pmu_stop(struct arm_pmu *cpu_pmu)
812 {
813         /* Disable all counters */
814         armv8pmu_pmcr_write(armv8pmu_pmcr_read() & ~ARMV8_PMU_PMCR_E);
815 }
816
817 static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
818 {
819         u32 pmovsr;
820         struct perf_sample_data data;
821         struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
822         struct pt_regs *regs;
823         int idx;
824
825         /*
826          * Get and reset the IRQ flags
827          */
828         pmovsr = armv8pmu_getreset_flags();
829
830         /*
831          * Did an overflow occur?
832          */
833         if (!armv8pmu_has_overflowed(pmovsr))
834                 return IRQ_NONE;
835
836         /*
837          * Handle the counter(s) overflow(s)
838          */
839         regs = get_irq_regs();
840
841         /*
842          * Stop the PMU while processing the counter overflows
843          * to prevent skews in group events.
844          */
845         armv8pmu_stop(cpu_pmu);
846         for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
847                 struct perf_event *event = cpuc->events[idx];
848                 struct hw_perf_event *hwc;
849
850                 /* Ignore if we don't have an event. */
851                 if (!event)
852                         continue;
853
854                 /*
855                  * We have a single interrupt for all counters. Check that
856                  * each counter has overflowed before we process it.
857                  */
858                 if (!armv8pmu_counter_has_overflowed(pmovsr, idx))
859                         continue;
860
861                 hwc = &event->hw;
862                 armpmu_event_update(event);
863                 perf_sample_data_init(&data, 0, hwc->last_period);
864                 if (!armpmu_event_set_period(event))
865                         continue;
866
867                 /*
868                  * Perf event overflow will queue the processing of the event as
869                  * an irq_work which will be taken care of in the handling of
870                  * IPI_IRQ_WORK.
871                  */
872                 if (perf_event_overflow(event, &data, regs))
873                         cpu_pmu->disable(event);
874         }
875         armv8pmu_start(cpu_pmu);
876
877         return IRQ_HANDLED;
878 }
879
880 static int armv8pmu_get_single_idx(struct pmu_hw_events *cpuc,
881                                     struct arm_pmu *cpu_pmu)
882 {
883         int idx;
884
885         for (idx = ARMV8_IDX_COUNTER0; idx < cpu_pmu->num_events; idx++) {
886                 if (!test_and_set_bit(idx, cpuc->used_mask))
887                         return idx;
888         }
889         return -EAGAIN;
890 }
891
892 static int armv8pmu_get_chain_idx(struct pmu_hw_events *cpuc,
893                                    struct arm_pmu *cpu_pmu)
894 {
895         int idx;
896
897         /*
898          * Chaining requires two consecutive event counters, where
899          * the lower idx must be even.
900          */
901         for (idx = ARMV8_IDX_COUNTER0 + 1; idx < cpu_pmu->num_events; idx += 2) {
902                 if (!test_and_set_bit(idx, cpuc->used_mask)) {
903                         /* Check if the preceding even counter is available */
904                         if (!test_and_set_bit(idx - 1, cpuc->used_mask))
905                                 return idx;
906                         /* Release the Odd counter */
907                         clear_bit(idx, cpuc->used_mask);
908                 }
909         }
910         return -EAGAIN;
911 }
912
913 static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
914                                   struct perf_event *event)
915 {
916         struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
917         struct hw_perf_event *hwc = &event->hw;
918         unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
919
920         /* Always prefer to place a cycle counter into the cycle counter. */
921         if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
922                 if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
923                         return ARMV8_IDX_CYCLE_COUNTER;
924                 else if (armv8pmu_event_is_64bit(event) &&
925                            armv8pmu_event_want_user_access(event) &&
926                            !armv8pmu_has_long_event(cpu_pmu))
927                                 return -EAGAIN;
928         }
929
930         /*
931          * Otherwise use events counters
932          */
933         if (armv8pmu_event_is_chained(event))
934                 return  armv8pmu_get_chain_idx(cpuc, cpu_pmu);
935         else
936                 return armv8pmu_get_single_idx(cpuc, cpu_pmu);
937 }
938
939 static void armv8pmu_clear_event_idx(struct pmu_hw_events *cpuc,
940                                      struct perf_event *event)
941 {
942         int idx = event->hw.idx;
943
944         clear_bit(idx, cpuc->used_mask);
945         if (armv8pmu_event_is_chained(event))
946                 clear_bit(idx - 1, cpuc->used_mask);
947 }
948
949 static int armv8pmu_user_event_idx(struct perf_event *event)
950 {
951         if (!sysctl_perf_user_access || !armv8pmu_event_has_user_read(event))
952                 return 0;
953
954         /*
955          * We remap the cycle counter index to 32 to
956          * match the offset applied to the rest of
957          * the counter indices.
958          */
959         if (event->hw.idx == ARMV8_IDX_CYCLE_COUNTER)
960                 return ARMV8_IDX_CYCLE_COUNTER_USER;
961
962         return event->hw.idx;
963 }
964
965 /*
966  * Add an event filter to a given event.
967  */
968 static int armv8pmu_set_event_filter(struct hw_perf_event *event,
969                                      struct perf_event_attr *attr)
970 {
971         unsigned long config_base = 0;
972
973         if (attr->exclude_idle)
974                 return -EPERM;
975
976         /*
977          * If we're running in hyp mode, then we *are* the hypervisor.
978          * Therefore we ignore exclude_hv in this configuration, since
979          * there's no hypervisor to sample anyway. This is consistent
980          * with other architectures (x86 and Power).
981          */
982         if (is_kernel_in_hyp_mode()) {
983                 if (!attr->exclude_kernel && !attr->exclude_host)
984                         config_base |= ARMV8_PMU_INCLUDE_EL2;
985                 if (attr->exclude_guest)
986                         config_base |= ARMV8_PMU_EXCLUDE_EL1;
987                 if (attr->exclude_host)
988                         config_base |= ARMV8_PMU_EXCLUDE_EL0;
989         } else {
990                 if (!attr->exclude_hv && !attr->exclude_host)
991                         config_base |= ARMV8_PMU_INCLUDE_EL2;
992         }
993
994         /*
995          * Filter out !VHE kernels and guest kernels
996          */
997         if (attr->exclude_kernel)
998                 config_base |= ARMV8_PMU_EXCLUDE_EL1;
999
1000         if (attr->exclude_user)
1001                 config_base |= ARMV8_PMU_EXCLUDE_EL0;
1002
1003         /*
1004          * Install the filter into config_base as this is used to
1005          * construct the event type.
1006          */
1007         event->config_base = config_base;
1008
1009         return 0;
1010 }
1011
1012 static int armv8pmu_filter_match(struct perf_event *event)
1013 {
1014         unsigned long evtype = event->hw.config_base & ARMV8_PMU_EVTYPE_EVENT;
1015         return evtype != ARMV8_PMUV3_PERFCTR_CHAIN;
1016 }
1017
1018 static void armv8pmu_reset(void *info)
1019 {
1020         struct arm_pmu *cpu_pmu = (struct arm_pmu *)info;
1021         u32 pmcr;
1022
1023         /* The counter and interrupt enable registers are unknown at reset. */
1024         armv8pmu_disable_counter(U32_MAX);
1025         armv8pmu_disable_intens(U32_MAX);
1026
1027         /* Clear the counters we flip at guest entry/exit */
1028         kvm_clr_pmu_events(U32_MAX);
1029
1030         /*
1031          * Initialize & Reset PMNC. Request overflow interrupt for
1032          * 64 bit cycle counter but cheat in armv8pmu_write_counter().
1033          */
1034         pmcr = ARMV8_PMU_PMCR_P | ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_LC;
1035
1036         /* Enable long event counter support where available */
1037         if (armv8pmu_has_long_event(cpu_pmu))
1038                 pmcr |= ARMV8_PMU_PMCR_LP;
1039
1040         armv8pmu_pmcr_write(pmcr);
1041 }
1042
1043 static int __armv8_pmuv3_map_event(struct perf_event *event,
1044                                    const unsigned (*extra_event_map)
1045                                                   [PERF_COUNT_HW_MAX],
1046                                    const unsigned (*extra_cache_map)
1047                                                   [PERF_COUNT_HW_CACHE_MAX]
1048                                                   [PERF_COUNT_HW_CACHE_OP_MAX]
1049                                                   [PERF_COUNT_HW_CACHE_RESULT_MAX])
1050 {
1051         int hw_event_id;
1052         struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
1053
1054         hw_event_id = armpmu_map_event(event, &armv8_pmuv3_perf_map,
1055                                        &armv8_pmuv3_perf_cache_map,
1056                                        ARMV8_PMU_EVTYPE_EVENT);
1057
1058         if (armv8pmu_event_is_64bit(event))
1059                 event->hw.flags |= ARMPMU_EVT_64BIT;
1060
1061         /*
1062          * User events must be allocated into a single counter, and so
1063          * must not be chained.
1064          *
1065          * Most 64-bit events require long counter support, but 64-bit
1066          * CPU_CYCLES events can be placed into the dedicated cycle
1067          * counter when this is free.
1068          */
1069         if (armv8pmu_event_want_user_access(event)) {
1070                 if (!(event->attach_state & PERF_ATTACH_TASK))
1071                         return -EINVAL;
1072                 if (armv8pmu_event_is_64bit(event) &&
1073                     (hw_event_id != ARMV8_PMUV3_PERFCTR_CPU_CYCLES) &&
1074                     !armv8pmu_has_long_event(armpmu))
1075                         return -EOPNOTSUPP;
1076
1077                 event->hw.flags |= PERF_EVENT_FLAG_USER_READ_CNT;
1078         }
1079
1080         /* Only expose micro/arch events supported by this PMU */
1081         if ((hw_event_id > 0) && (hw_event_id < ARMV8_PMUV3_MAX_COMMON_EVENTS)
1082             && test_bit(hw_event_id, armpmu->pmceid_bitmap)) {
1083                 return hw_event_id;
1084         }
1085
1086         return armpmu_map_event(event, extra_event_map, extra_cache_map,
1087                                 ARMV8_PMU_EVTYPE_EVENT);
1088 }
1089
1090 static int armv8_pmuv3_map_event(struct perf_event *event)
1091 {
1092         return __armv8_pmuv3_map_event(event, NULL, NULL);
1093 }
1094
1095 static int armv8_a53_map_event(struct perf_event *event)
1096 {
1097         return __armv8_pmuv3_map_event(event, NULL, &armv8_a53_perf_cache_map);
1098 }
1099
1100 static int armv8_a57_map_event(struct perf_event *event)
1101 {
1102         return __armv8_pmuv3_map_event(event, NULL, &armv8_a57_perf_cache_map);
1103 }
1104
1105 static int armv8_a73_map_event(struct perf_event *event)
1106 {
1107         return __armv8_pmuv3_map_event(event, NULL, &armv8_a73_perf_cache_map);
1108 }
1109
1110 static int armv8_thunder_map_event(struct perf_event *event)
1111 {
1112         return __armv8_pmuv3_map_event(event, NULL,
1113                                        &armv8_thunder_perf_cache_map);
1114 }
1115
1116 static int armv8_vulcan_map_event(struct perf_event *event)
1117 {
1118         return __armv8_pmuv3_map_event(event, NULL,
1119                                        &armv8_vulcan_perf_cache_map);
1120 }
1121
1122 struct armv8pmu_probe_info {
1123         struct arm_pmu *pmu;
1124         bool present;
1125 };
1126
1127 static void __armv8pmu_probe_pmu(void *info)
1128 {
1129         struct armv8pmu_probe_info *probe = info;
1130         struct arm_pmu *cpu_pmu = probe->pmu;
1131         u64 dfr0;
1132         u64 pmceid_raw[2];
1133         u32 pmceid[2];
1134         int pmuver;
1135
1136         dfr0 = read_sysreg(id_aa64dfr0_el1);
1137         pmuver = cpuid_feature_extract_unsigned_field(dfr0,
1138                         ID_AA64DFR0_PMUVER_SHIFT);
1139         if (pmuver == ID_AA64DFR0_PMUVER_IMP_DEF || pmuver == 0)
1140                 return;
1141
1142         cpu_pmu->pmuver = pmuver;
1143         probe->present = true;
1144
1145         /* Read the nb of CNTx counters supported from PMNC */
1146         cpu_pmu->num_events = (armv8pmu_pmcr_read() >> ARMV8_PMU_PMCR_N_SHIFT)
1147                 & ARMV8_PMU_PMCR_N_MASK;
1148
1149         /* Add the CPU cycles counter */
1150         cpu_pmu->num_events += 1;
1151
1152         pmceid[0] = pmceid_raw[0] = read_sysreg(pmceid0_el0);
1153         pmceid[1] = pmceid_raw[1] = read_sysreg(pmceid1_el0);
1154
1155         bitmap_from_arr32(cpu_pmu->pmceid_bitmap,
1156                              pmceid, ARMV8_PMUV3_MAX_COMMON_EVENTS);
1157
1158         pmceid[0] = pmceid_raw[0] >> 32;
1159         pmceid[1] = pmceid_raw[1] >> 32;
1160
1161         bitmap_from_arr32(cpu_pmu->pmceid_ext_bitmap,
1162                              pmceid, ARMV8_PMUV3_MAX_COMMON_EVENTS);
1163
1164         /* store PMMIR_EL1 register for sysfs */
1165         if (pmuver >= ID_AA64DFR0_PMUVER_8_4 && (pmceid_raw[1] & BIT(31)))
1166                 cpu_pmu->reg_pmmir = read_cpuid(PMMIR_EL1);
1167         else
1168                 cpu_pmu->reg_pmmir = 0;
1169 }
1170
1171 static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
1172 {
1173         struct armv8pmu_probe_info probe = {
1174                 .pmu = cpu_pmu,
1175                 .present = false,
1176         };
1177         int ret;
1178
1179         ret = smp_call_function_any(&cpu_pmu->supported_cpus,
1180                                     __armv8pmu_probe_pmu,
1181                                     &probe, 1);
1182         if (ret)
1183                 return ret;
1184
1185         return probe.present ? 0 : -ENODEV;
1186 }
1187
1188 static void armv8pmu_disable_user_access_ipi(void *unused)
1189 {
1190         armv8pmu_disable_user_access();
1191 }
1192
1193 static int armv8pmu_proc_user_access_handler(struct ctl_table *table, int write,
1194                 void *buffer, size_t *lenp, loff_t *ppos)
1195 {
1196         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1197         if (ret || !write || sysctl_perf_user_access)
1198                 return ret;
1199
1200         on_each_cpu(armv8pmu_disable_user_access_ipi, NULL, 1);
1201         return 0;
1202 }
1203
1204 static struct ctl_table armv8_pmu_sysctl_table[] = {
1205         {
1206                 .procname       = "perf_user_access",
1207                 .data           = &sysctl_perf_user_access,
1208                 .maxlen         = sizeof(unsigned int),
1209                 .mode           = 0644,
1210                 .proc_handler   = armv8pmu_proc_user_access_handler,
1211                 .extra1         = SYSCTL_ZERO,
1212                 .extra2         = SYSCTL_ONE,
1213         },
1214         { }
1215 };
1216
1217 static void armv8_pmu_register_sysctl_table(void)
1218 {
1219         static u32 tbl_registered = 0;
1220
1221         if (!cmpxchg_relaxed(&tbl_registered, 0, 1))
1222                 register_sysctl("kernel", armv8_pmu_sysctl_table);
1223 }
1224
1225 static int armv8_pmu_init(struct arm_pmu *cpu_pmu, char *name,
1226                           int (*map_event)(struct perf_event *event),
1227                           const struct attribute_group *events,
1228                           const struct attribute_group *format,
1229                           const struct attribute_group *caps)
1230 {
1231         int ret = armv8pmu_probe_pmu(cpu_pmu);
1232         if (ret)
1233                 return ret;
1234
1235         cpu_pmu->handle_irq             = armv8pmu_handle_irq;
1236         cpu_pmu->enable                 = armv8pmu_enable_event;
1237         cpu_pmu->disable                = armv8pmu_disable_event;
1238         cpu_pmu->read_counter           = armv8pmu_read_counter;
1239         cpu_pmu->write_counter          = armv8pmu_write_counter;
1240         cpu_pmu->get_event_idx          = armv8pmu_get_event_idx;
1241         cpu_pmu->clear_event_idx        = armv8pmu_clear_event_idx;
1242         cpu_pmu->start                  = armv8pmu_start;
1243         cpu_pmu->stop                   = armv8pmu_stop;
1244         cpu_pmu->reset                  = armv8pmu_reset;
1245         cpu_pmu->set_event_filter       = armv8pmu_set_event_filter;
1246         cpu_pmu->filter_match           = armv8pmu_filter_match;
1247
1248         cpu_pmu->pmu.event_idx          = armv8pmu_user_event_idx;
1249
1250         cpu_pmu->name                   = name;
1251         cpu_pmu->map_event              = map_event;
1252         cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = events ?
1253                         events : &armv8_pmuv3_events_attr_group;
1254         cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_FORMATS] = format ?
1255                         format : &armv8_pmuv3_format_attr_group;
1256         cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_CAPS] = caps ?
1257                         caps : &armv8_pmuv3_caps_attr_group;
1258
1259         armv8_pmu_register_sysctl_table();
1260         return 0;
1261 }
1262
1263 static int armv8_pmu_init_nogroups(struct arm_pmu *cpu_pmu, char *name,
1264                                    int (*map_event)(struct perf_event *event))
1265 {
1266         return armv8_pmu_init(cpu_pmu, name, map_event, NULL, NULL, NULL);
1267 }
1268
1269 #define PMUV3_INIT_SIMPLE(name)                                         \
1270 static int name##_pmu_init(struct arm_pmu *cpu_pmu)                     \
1271 {                                                                       \
1272         return armv8_pmu_init_nogroups(cpu_pmu, #name, armv8_pmuv3_map_event);\
1273 }
1274
1275 PMUV3_INIT_SIMPLE(armv8_pmuv3)
1276
1277 PMUV3_INIT_SIMPLE(armv8_cortex_a34)
1278 PMUV3_INIT_SIMPLE(armv8_cortex_a55)
1279 PMUV3_INIT_SIMPLE(armv8_cortex_a65)
1280 PMUV3_INIT_SIMPLE(armv8_cortex_a75)
1281 PMUV3_INIT_SIMPLE(armv8_cortex_a76)
1282 PMUV3_INIT_SIMPLE(armv8_cortex_a77)
1283 PMUV3_INIT_SIMPLE(armv8_cortex_a78)
1284 PMUV3_INIT_SIMPLE(armv9_cortex_a510)
1285 PMUV3_INIT_SIMPLE(armv9_cortex_a710)
1286 PMUV3_INIT_SIMPLE(armv8_cortex_x1)
1287 PMUV3_INIT_SIMPLE(armv9_cortex_x2)
1288 PMUV3_INIT_SIMPLE(armv8_neoverse_e1)
1289 PMUV3_INIT_SIMPLE(armv8_neoverse_n1)
1290 PMUV3_INIT_SIMPLE(armv9_neoverse_n2)
1291 PMUV3_INIT_SIMPLE(armv8_neoverse_v1)
1292
1293 PMUV3_INIT_SIMPLE(armv8_nvidia_carmel)
1294 PMUV3_INIT_SIMPLE(armv8_nvidia_denver)
1295
1296 static int armv8_a35_pmu_init(struct arm_pmu *cpu_pmu)
1297 {
1298         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cortex_a35",
1299                                        armv8_a53_map_event);
1300 }
1301
1302 static int armv8_a53_pmu_init(struct arm_pmu *cpu_pmu)
1303 {
1304         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cortex_a53",
1305                                        armv8_a53_map_event);
1306 }
1307
1308 static int armv8_a57_pmu_init(struct arm_pmu *cpu_pmu)
1309 {
1310         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cortex_a57",
1311                                        armv8_a57_map_event);
1312 }
1313
1314 static int armv8_a72_pmu_init(struct arm_pmu *cpu_pmu)
1315 {
1316         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cortex_a72",
1317                                        armv8_a57_map_event);
1318 }
1319
1320 static int armv8_a73_pmu_init(struct arm_pmu *cpu_pmu)
1321 {
1322         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cortex_a73",
1323                                        armv8_a73_map_event);
1324 }
1325
1326 static int armv8_thunder_pmu_init(struct arm_pmu *cpu_pmu)
1327 {
1328         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_cavium_thunder",
1329                                        armv8_thunder_map_event);
1330 }
1331
1332 static int armv8_vulcan_pmu_init(struct arm_pmu *cpu_pmu)
1333 {
1334         return armv8_pmu_init_nogroups(cpu_pmu, "armv8_brcm_vulcan",
1335                                        armv8_vulcan_map_event);
1336 }
1337
1338 static const struct of_device_id armv8_pmu_of_device_ids[] = {
1339         {.compatible = "arm,armv8-pmuv3",       .data = armv8_pmuv3_pmu_init},
1340         {.compatible = "arm,cortex-a34-pmu",    .data = armv8_cortex_a34_pmu_init},
1341         {.compatible = "arm,cortex-a35-pmu",    .data = armv8_a35_pmu_init},
1342         {.compatible = "arm,cortex-a53-pmu",    .data = armv8_a53_pmu_init},
1343         {.compatible = "arm,cortex-a55-pmu",    .data = armv8_cortex_a55_pmu_init},
1344         {.compatible = "arm,cortex-a57-pmu",    .data = armv8_a57_pmu_init},
1345         {.compatible = "arm,cortex-a65-pmu",    .data = armv8_cortex_a65_pmu_init},
1346         {.compatible = "arm,cortex-a72-pmu",    .data = armv8_a72_pmu_init},
1347         {.compatible = "arm,cortex-a73-pmu",    .data = armv8_a73_pmu_init},
1348         {.compatible = "arm,cortex-a75-pmu",    .data = armv8_cortex_a75_pmu_init},
1349         {.compatible = "arm,cortex-a76-pmu",    .data = armv8_cortex_a76_pmu_init},
1350         {.compatible = "arm,cortex-a77-pmu",    .data = armv8_cortex_a77_pmu_init},
1351         {.compatible = "arm,cortex-a78-pmu",    .data = armv8_cortex_a78_pmu_init},
1352         {.compatible = "arm,cortex-a510-pmu",   .data = armv9_cortex_a510_pmu_init},
1353         {.compatible = "arm,cortex-a710-pmu",   .data = armv9_cortex_a710_pmu_init},
1354         {.compatible = "arm,cortex-x1-pmu",     .data = armv8_cortex_x1_pmu_init},
1355         {.compatible = "arm,cortex-x2-pmu",     .data = armv9_cortex_x2_pmu_init},
1356         {.compatible = "arm,neoverse-e1-pmu",   .data = armv8_neoverse_e1_pmu_init},
1357         {.compatible = "arm,neoverse-n1-pmu",   .data = armv8_neoverse_n1_pmu_init},
1358         {.compatible = "arm,neoverse-n2-pmu",   .data = armv9_neoverse_n2_pmu_init},
1359         {.compatible = "arm,neoverse-v1-pmu",   .data = armv8_neoverse_v1_pmu_init},
1360         {.compatible = "cavium,thunder-pmu",    .data = armv8_thunder_pmu_init},
1361         {.compatible = "brcm,vulcan-pmu",       .data = armv8_vulcan_pmu_init},
1362         {.compatible = "nvidia,carmel-pmu",     .data = armv8_nvidia_carmel_pmu_init},
1363         {.compatible = "nvidia,denver-pmu",     .data = armv8_nvidia_denver_pmu_init},
1364         {},
1365 };
1366
1367 static int armv8_pmu_device_probe(struct platform_device *pdev)
1368 {
1369         return arm_pmu_device_probe(pdev, armv8_pmu_of_device_ids, NULL);
1370 }
1371
1372 static struct platform_driver armv8_pmu_driver = {
1373         .driver         = {
1374                 .name   = ARMV8_PMU_PDEV_NAME,
1375                 .of_match_table = armv8_pmu_of_device_ids,
1376                 .suppress_bind_attrs = true,
1377         },
1378         .probe          = armv8_pmu_device_probe,
1379 };
1380
1381 static int __init armv8_pmu_driver_init(void)
1382 {
1383         if (acpi_disabled)
1384                 return platform_driver_register(&armv8_pmu_driver);
1385         else
1386                 return arm_pmu_acpi_probe(armv8_pmuv3_pmu_init);
1387 }
1388 device_initcall(armv8_pmu_driver_init)
1389
1390 void arch_perf_update_userpage(struct perf_event *event,
1391                                struct perf_event_mmap_page *userpg, u64 now)
1392 {
1393         struct clock_read_data *rd;
1394         unsigned int seq;
1395         u64 ns;
1396
1397         userpg->cap_user_time = 0;
1398         userpg->cap_user_time_zero = 0;
1399         userpg->cap_user_time_short = 0;
1400         userpg->cap_user_rdpmc = armv8pmu_event_has_user_read(event);
1401
1402         if (userpg->cap_user_rdpmc) {
1403                 if (event->hw.flags & ARMPMU_EVT_64BIT)
1404                         userpg->pmc_width = 64;
1405                 else
1406                         userpg->pmc_width = 32;
1407         }
1408
1409         do {
1410                 rd = sched_clock_read_begin(&seq);
1411
1412                 if (rd->read_sched_clock != arch_timer_read_counter)
1413                         return;
1414
1415                 userpg->time_mult = rd->mult;
1416                 userpg->time_shift = rd->shift;
1417                 userpg->time_zero = rd->epoch_ns;
1418                 userpg->time_cycles = rd->epoch_cyc;
1419                 userpg->time_mask = rd->sched_clock_mask;
1420
1421                 /*
1422                  * Subtract the cycle base, such that software that
1423                  * doesn't know about cap_user_time_short still 'works'
1424                  * assuming no wraps.
1425                  */
1426                 ns = mul_u64_u32_shr(rd->epoch_cyc, rd->mult, rd->shift);
1427                 userpg->time_zero -= ns;
1428
1429         } while (sched_clock_read_retry(seq));
1430
1431         userpg->time_offset = userpg->time_zero - now;
1432
1433         /*
1434          * time_shift is not expected to be greater than 31 due to
1435          * the original published conversion algorithm shifting a
1436          * 32-bit value (now specifies a 64-bit value) - refer
1437          * perf_event_mmap_page documentation in perf_event.h.
1438          */
1439         if (userpg->time_shift == 32) {
1440                 userpg->time_shift = 31;
1441                 userpg->time_mult >>= 1;
1442         }
1443
1444         /*
1445          * Internal timekeeping for enabled/running/stopped times
1446          * is always computed with the sched_clock.
1447          */
1448         userpg->cap_user_time = 1;
1449         userpg->cap_user_time_zero = 1;
1450         userpg->cap_user_time_short = 1;
1451 }