xref: /openbmc/linux/arch/x86/events/amd/brs.c (revision c059ee9d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Implement support for AMD Fam19h Branch Sampling feature
4  * Based on specifications published in AMD PPR Fam19 Model 01
5  *
6  * Copyright 2021 Google LLC
7  * Contributed by Stephane Eranian <eranian@google.com>
8  */
9 #include <linux/kernel.h>
10 #include <linux/jump_label.h>
11 #include <asm/msr.h>
12 #include <asm/cpufeature.h>
13 
14 #include "../perf_event.h"
15 
16 #define BRS_POISON	0xFFFFFFFFFFFFFFFEULL /* mark limit of valid entries */
17 
18 /* Debug Extension Configuration register layout */
19 union amd_debug_extn_cfg {
20 	__u64 val;
21 	struct {
22 		__u64	rsvd0:2,  /* reserved */
23 			brsmen:1, /* branch sample enable */
24 			rsvd4_3:2,/* reserved - must be 0x3 */
25 			vb:1,     /* valid branches recorded */
26 			rsvd2:10, /* reserved */
27 			msroff:4, /* index of next entry to write */
28 			rsvd3:4,  /* reserved */
29 			pmc:3,    /* #PMC holding the sampling event */
30 			rsvd4:37; /* reserved */
31 	};
32 };
33 
34 static inline unsigned int brs_from(int idx)
35 {
36 	return MSR_AMD_SAMP_BR_FROM + 2 * idx;
37 }
38 
39 static inline unsigned int brs_to(int idx)
40 {
41 	return MSR_AMD_SAMP_BR_FROM + 2 * idx + 1;
42 }
43 
44 static inline void set_debug_extn_cfg(u64 val)
45 {
46 	/* bits[4:3] must always be set to 11b */
47 	wrmsrl(MSR_AMD_DBG_EXTN_CFG, val | 3ULL << 3);
48 }
49 
50 static inline u64 get_debug_extn_cfg(void)
51 {
52 	u64 val;
53 
54 	rdmsrl(MSR_AMD_DBG_EXTN_CFG, val);
55 	return val;
56 }
57 
58 static bool __init amd_brs_detect(void)
59 {
60 	if (!cpu_feature_enabled(X86_FEATURE_BRS))
61 		return false;
62 
63 	switch (boot_cpu_data.x86) {
64 	case 0x19: /* AMD Fam19h (Zen3) */
65 		x86_pmu.lbr_nr = 16;
66 
67 		/* No hardware filtering supported */
68 		x86_pmu.lbr_sel_map = NULL;
69 		x86_pmu.lbr_sel_mask = 0;
70 		break;
71 	default:
72 		return false;
73 	}
74 
75 	return true;
76 }
77 
78 /*
79  * Current BRS implementation does not support branch type or privilege level
80  * filtering. Therefore, this function simply enforces these limitations. No need for
81  * a br_sel_map. Software filtering is not supported because it would not correlate well
82  * with a sampling period.
83  */
84 int amd_brs_setup_filter(struct perf_event *event)
85 {
86 	u64 type = event->attr.branch_sample_type;
87 
88 	/* No BRS support */
89 	if (!x86_pmu.lbr_nr)
90 		return -EOPNOTSUPP;
91 
92 	/* Can only capture all branches, i.e., no filtering */
93 	if ((type & ~PERF_SAMPLE_BRANCH_PLM_ALL) != PERF_SAMPLE_BRANCH_ANY)
94 		return -EINVAL;
95 
96 	return 0;
97 }
98 
99 /* tos = top of stack, i.e., last valid entry written */
100 static inline int amd_brs_get_tos(union amd_debug_extn_cfg *cfg)
101 {
102 	/*
103 	 * msroff: index of next entry to write so top-of-stack is one off
104 	 * if BRS is full then msroff is set back to 0.
105 	 */
106 	return (cfg->msroff ? cfg->msroff : x86_pmu.lbr_nr) - 1;
107 }
108 
109 /*
110  * make sure we have a sane BRS offset to begin with
111  * especially with kexec
112  */
113 void amd_brs_reset(void)
114 {
115 	if (!cpu_feature_enabled(X86_FEATURE_BRS))
116 		return;
117 
118 	/*
119 	 * Reset config
120 	 */
121 	set_debug_extn_cfg(0);
122 
123 	/*
124 	 * Mark first entry as poisoned
125 	 */
126 	wrmsrl(brs_to(0), BRS_POISON);
127 }
128 
129 int __init amd_brs_init(void)
130 {
131 	if (!amd_brs_detect())
132 		return -EOPNOTSUPP;
133 
134 	pr_cont("%d-deep BRS, ", x86_pmu.lbr_nr);
135 
136 	return 0;
137 }
138 
139 void amd_brs_enable(void)
140 {
141 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
142 	union amd_debug_extn_cfg cfg;
143 
144 	/* Activate only on first user */
145 	if (++cpuc->brs_active > 1)
146 		return;
147 
148 	cfg.val    = 0; /* reset all fields */
149 	cfg.brsmen = 1; /* enable branch sampling */
150 
151 	/* Set enable bit */
152 	set_debug_extn_cfg(cfg.val);
153 }
154 
155 void amd_brs_enable_all(void)
156 {
157 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
158 	if (cpuc->lbr_users)
159 		amd_brs_enable();
160 }
161 
162 void amd_brs_disable(void)
163 {
164 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
165 	union amd_debug_extn_cfg cfg;
166 
167 	/* Check if active (could be disabled via x86_pmu_disable_all()) */
168 	if (!cpuc->brs_active)
169 		return;
170 
171 	/* Only disable for last user */
172 	if (--cpuc->brs_active)
173 		return;
174 
175 	/*
176 	 * Clear the brsmen bit but preserve the others as they contain
177 	 * useful state such as vb and msroff
178 	 */
179 	cfg.val = get_debug_extn_cfg();
180 
181 	/*
182 	 * When coming in on interrupt and BRS is full, then hw will have
183 	 * already stopped BRS, no need to issue wrmsr again
184 	 */
185 	if (cfg.brsmen) {
186 		cfg.brsmen = 0;
187 		set_debug_extn_cfg(cfg.val);
188 	}
189 }
190 
191 void amd_brs_disable_all(void)
192 {
193 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
194 	if (cpuc->lbr_users)
195 		amd_brs_disable();
196 }
197 
198 static bool amd_brs_match_plm(struct perf_event *event, u64 to)
199 {
200 	int type = event->attr.branch_sample_type;
201 	int plm_k = PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_HV;
202 	int plm_u = PERF_SAMPLE_BRANCH_USER;
203 
204 	if (!(type & plm_k) && kernel_ip(to))
205 		return 0;
206 
207 	if (!(type & plm_u) && !kernel_ip(to))
208 		return 0;
209 
210 	return 1;
211 }
212 
213 /*
214  * Caller must ensure amd_brs_inuse() is true before calling
215  * return:
216  */
217 void amd_brs_drain(void)
218 {
219 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
220 	struct perf_event *event = cpuc->events[0];
221 	struct perf_branch_entry *br = cpuc->lbr_entries;
222 	union amd_debug_extn_cfg cfg;
223 	u32 i, nr = 0, num, tos, start;
224 	u32 shift = 64 - boot_cpu_data.x86_virt_bits;
225 
226 	/*
227 	 * BRS event forced on PMC0,
228 	 * so check if there is an event.
229 	 * It is possible to have lbr_users > 0 but the event
230 	 * not yet scheduled due to long latency PMU irq
231 	 */
232 	if (!event)
233 		goto empty;
234 
235 	cfg.val = get_debug_extn_cfg();
236 
237 	/* Sanity check [0-x86_pmu.lbr_nr] */
238 	if (WARN_ON_ONCE(cfg.msroff >= x86_pmu.lbr_nr))
239 		goto empty;
240 
241 	/* No valid branch */
242 	if (cfg.vb == 0)
243 		goto empty;
244 
245 	/*
246 	 * msr.off points to next entry to be written
247 	 * tos = most recent entry index = msr.off - 1
248 	 * BRS register buffer saturates, so we know we have
249 	 * start < tos and that we have to read from start to tos
250 	 */
251 	start = 0;
252 	tos = amd_brs_get_tos(&cfg);
253 
254 	num = tos - start + 1;
255 
256 	/*
257 	 * BRS is only one pass (saturation) from MSROFF to depth-1
258 	 * MSROFF wraps to zero when buffer is full
259 	 */
260 	for (i = 0; i < num; i++) {
261 		u32 brs_idx = tos - i;
262 		u64 from, to;
263 
264 		rdmsrl(brs_to(brs_idx), to);
265 
266 		/* Entry does not belong to us (as marked by kernel) */
267 		if (to == BRS_POISON)
268 			break;
269 
270 		/*
271 		 * Sign-extend SAMP_BR_TO to 64 bits, bits 61-63 are reserved.
272 		 * Necessary to generate proper virtual addresses suitable for
273 		 * symbolization
274 		 */
275 		to = (u64)(((s64)to << shift) >> shift);
276 
277 		if (!amd_brs_match_plm(event, to))
278 			continue;
279 
280 		rdmsrl(brs_from(brs_idx), from);
281 
282 		perf_clear_branch_entry_bitfields(br+nr);
283 
284 		br[nr].from = from;
285 		br[nr].to   = to;
286 
287 		nr++;
288 	}
289 empty:
290 	/* Record number of sampled branches */
291 	cpuc->lbr_stack.nr = nr;
292 }
293 
294 /*
295  * Poison most recent entry to prevent reuse by next task
296  * required because BRS entry are not tagged by PID
297  */
298 static void amd_brs_poison_buffer(void)
299 {
300 	union amd_debug_extn_cfg cfg;
301 	unsigned int idx;
302 
303 	/* Get current state */
304 	cfg.val = get_debug_extn_cfg();
305 
306 	/* idx is most recently written entry */
307 	idx = amd_brs_get_tos(&cfg);
308 
309 	/* Poison target of entry */
310 	wrmsrl(brs_to(idx), BRS_POISON);
311 }
312 
313 /*
314  * On context switch in, we need to make sure no samples from previous user
315  * are left in the BRS.
316  *
317  * On ctxswin, sched_in = true, called after the PMU has started
318  * On ctxswout, sched_in = false, called before the PMU is stopped
319  */
320 void amd_pmu_brs_sched_task(struct perf_event_context *ctx, bool sched_in)
321 {
322 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
323 
324 	/* no active users */
325 	if (!cpuc->lbr_users)
326 		return;
327 
328 	/*
329 	 * On context switch in, we need to ensure we do not use entries
330 	 * from previous BRS user on that CPU, so we poison the buffer as
331 	 * a faster way compared to resetting all entries.
332 	 */
333 	if (sched_in)
334 		amd_brs_poison_buffer();
335 }
336 
337 /*
338  * called from ACPI processor_idle.c or acpi_pad.c
339  * with interrupts disabled
340  */
341 void perf_amd_brs_lopwr_cb(bool lopwr_in)
342 {
343 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
344 	union amd_debug_extn_cfg cfg;
345 
346 	/*
347 	 * on mwait in, we may end up in non C0 state.
348 	 * we must disable branch sampling to avoid holding the NMI
349 	 * for too long. We disable it in hardware but we
350 	 * keep the state in cpuc, so we can re-enable.
351 	 *
352 	 * The hardware will deliver the NMI if needed when brsmen cleared
353 	 */
354 	if (cpuc->brs_active) {
355 		cfg.val = get_debug_extn_cfg();
356 		cfg.brsmen = !lopwr_in;
357 		set_debug_extn_cfg(cfg.val);
358 	}
359 }
360 
361 DEFINE_STATIC_CALL_NULL(perf_lopwr_cb, perf_amd_brs_lopwr_cb);
362 EXPORT_STATIC_CALL_TRAMP_GPL(perf_lopwr_cb);
363 
364 void __init amd_brs_lopwr_init(void)
365 {
366 	static_call_update(perf_lopwr_cb, perf_amd_brs_lopwr_cb);
367 }
368