xref: /illumos-gate/usr/src/uts/intel/os/cpc_subr.c (revision ddb365bfc9e868ad24ccdcb0dc91af18b10df082)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2021 Joyent, Inc.
24  */
25 
26 /*
27  * x86-specific routines used by the CPU Performance counter driver.
28  */
29 
30 #include <sys/types.h>
31 #include <sys/time.h>
32 #include <sys/atomic.h>
33 #include <sys/regset.h>
34 #include <sys/privregs.h>
35 #include <sys/x86_archext.h>
36 #include <sys/cpuvar.h>
37 #include <sys/machcpuvar.h>
38 #include <sys/archsystm.h>
39 #include <sys/cpc_pcbe.h>
40 #include <sys/cpc_impl.h>
41 #include <sys/x_call.h>
42 #include <sys/cmn_err.h>
43 #include <sys/cmt.h>
44 #include <sys/spl.h>
45 #include <sys/apic.h>
46 
47 static kcpc_ctx_t *(*overflow_intr_handler)(caddr_t);
48 
49 /* Do threads share performance monitoring hardware? */
50 static int strands_perfmon_shared = 0;
51 
52 int kcpc_hw_overflow_intr_installed;		/* set by APIC code */
53 extern kcpc_ctx_t *kcpc_overflow_intr(caddr_t arg, uint64_t bitmap);
54 
55 extern int kcpc_counts_include_idle; /* Project Private /etc/system variable */
56 
57 void (*kcpc_hw_enable_cpc_intr)(void);		/* set by APIC code */
58 
59 int
60 kcpc_hw_add_ovf_intr(kcpc_ctx_t *(*handler)(caddr_t))
61 {
62 	if (x86_type != X86_TYPE_P6)
63 		return (0);
64 	overflow_intr_handler = handler;
65 	return (ipltospl(APIC_PCINT_IPL));
66 }
67 
68 void
69 kcpc_hw_rem_ovf_intr(void)
70 {
71 	overflow_intr_handler = NULL;
72 }
73 
74 /*
75  * Hook used on P4 systems to catch online/offline events.
76  */
77 /*ARGSUSED*/
78 static int
79 kcpc_cpu_setup(cpu_setup_t what, int cpuid, void *arg)
80 {
81 	pg_cmt_t	*chip_pg;
82 	int		active_cpus_cnt;
83 
84 	if (what != CPU_ON)
85 		return (0);
86 
87 	/*
88 	 * If any CPU-bound contexts exist, we don't need to invalidate
89 	 * anything, as no per-LWP contexts can coexist.
90 	 */
91 	if (kcpc_cpuctx || dtrace_cpc_in_use)
92 		return (0);
93 
94 	/*
95 	 * If this chip now has more than 1 active cpu, we must invalidate all
96 	 * contexts in the system.
97 	 */
98 	chip_pg = (pg_cmt_t *)pghw_find_pg(cpu[cpuid], PGHW_CHIP);
99 	if (chip_pg != NULL) {
100 		active_cpus_cnt = GROUP_SIZE(&chip_pg->cmt_cpus_actv);
101 		if (active_cpus_cnt > 1)
102 			kcpc_invalidate_all();
103 	}
104 
105 	return (0);
106 }
107 
108 static kmutex_t cpu_setup_lock;	/* protects setup_registered */
109 static int setup_registered;
110 
111 
112 void
113 kcpc_hw_init(cpu_t *cp)
114 {
115 	kthread_t *t = cp->cpu_idle_thread;
116 	uint32_t versionid;
117 	struct cpuid_regs cpuid;
118 
119 	strands_perfmon_shared = 0;
120 	if (is_x86_feature(x86_featureset, X86FSET_HTT)) {
121 		if (cpuid_getvendor(cpu[0]) == X86_VENDOR_Intel) {
122 			/*
123 			 * Intel processors that support Architectural
124 			 * Performance Monitoring Version 3 have per strand
125 			 * performance monitoring hardware.
126 			 * Hence we can allow use of performance counters on
127 			 * multiple strands on the same core simultaneously.
128 			 */
129 			cpuid.cp_eax = 0x0;
130 			(void) __cpuid_insn(&cpuid);
131 			if (cpuid.cp_eax < 0xa) {
132 				strands_perfmon_shared = 1;
133 			} else {
134 				cpuid.cp_eax = 0xa;
135 				(void) __cpuid_insn(&cpuid);
136 
137 				versionid = cpuid.cp_eax & 0xFF;
138 				if (versionid < 3) {
139 					strands_perfmon_shared = 1;
140 				}
141 			}
142 		} else if (cpuid_getvendor(cpu[0]) == X86_VENDOR_AMD ||
143 		    cpuid_getvendor(cpu[0]) == X86_VENDOR_HYGON) {
144 			/*
145 			 * On AMD systems with HT, all of the performance
146 			 * monitors exist on a per-logical CPU basis.
147 			 */
148 			strands_perfmon_shared = 0;
149 		} else {
150 			strands_perfmon_shared = 1;
151 		}
152 	}
153 
154 	if (strands_perfmon_shared) {
155 		mutex_enter(&cpu_setup_lock);
156 		if (setup_registered == 0) {
157 			mutex_enter(&cpu_lock);
158 			register_cpu_setup_func(kcpc_cpu_setup, NULL);
159 			mutex_exit(&cpu_lock);
160 			setup_registered = 1;
161 		}
162 		mutex_exit(&cpu_setup_lock);
163 	}
164 
165 	mutex_init(&cp->cpu_cpc_ctxlock, "cpu_cpc_ctxlock", MUTEX_DEFAULT, 0);
166 
167 	if (kcpc_counts_include_idle)
168 		return;
169 
170 	kcpc_idle_ctxop_install(t, cp);
171 }
172 
173 void
174 kcpc_hw_fini(cpu_t *cp)
175 {
176 	ASSERT(cp->cpu_idle_thread == NULL);
177 
178 	mutex_destroy(&cp->cpu_cpc_ctxlock);
179 }
180 
181 #define	BITS(v, u, l)	\
182 	(((v) >> (l)) & ((1 << (1 + (u) - (l))) - 1))
183 
184 #define	PCBE_NAMELEN 30	/* Enough Room for pcbe.manuf.model.family.stepping */
185 
186 /*
187  * Examine the processor and load an appropriate PCBE.
188  */
189 int
190 kcpc_hw_load_pcbe(void)
191 {
192 	return (kcpc_pcbe_tryload(cpuid_getvendorstr(CPU), cpuid_getfamily(CPU),
193 	    cpuid_getmodel(CPU), cpuid_getstep(CPU)));
194 }
195 
196 /*
197  * Called by the generic framework to check if it's OK to bind a set to a CPU.
198  */
199 int
200 kcpc_hw_cpu_hook(processorid_t cpuid, ulong_t *kcpc_cpumap)
201 {
202 	cpu_t		*cpu, *p;
203 	pg_t		*chip_pg;
204 	pg_cpu_itr_t	itr;
205 
206 	if (!strands_perfmon_shared)
207 		return (0);
208 
209 	/*
210 	 * Only one logical CPU on each Pentium 4 HT CPU may be bound to at
211 	 * once.
212 	 *
213 	 * This loop is protected by holding cpu_lock, in order to properly
214 	 * access the cpu_t of the desired cpu.
215 	 */
216 	mutex_enter(&cpu_lock);
217 	if ((cpu = cpu_get(cpuid)) == NULL) {
218 		mutex_exit(&cpu_lock);
219 		return (-1);
220 	}
221 
222 	chip_pg = (pg_t *)pghw_find_pg(cpu, PGHW_CHIP);
223 
224 	PG_CPU_ITR_INIT(chip_pg, itr);
225 	while ((p = pg_cpu_next(&itr)) != NULL) {
226 		if (p == cpu)
227 			continue;
228 		if (BT_TEST(kcpc_cpumap, p->cpu_id)) {
229 			mutex_exit(&cpu_lock);
230 			return (-1);
231 		}
232 	}
233 
234 	mutex_exit(&cpu_lock);
235 	return (0);
236 }
237 
238 /*
239  * Called by the generic framework to check if it's OK to bind a set to an LWP.
240  */
241 int
242 kcpc_hw_lwp_hook(void)
243 {
244 	pg_cmt_t	*chip;
245 	group_t		*chips;
246 	group_iter_t	i;
247 
248 	if (!strands_perfmon_shared)
249 		return (0);
250 
251 	/*
252 	 * Only one CPU per chip may be online.
253 	 */
254 	mutex_enter(&cpu_lock);
255 
256 	chips = pghw_set_lookup(PGHW_CHIP);
257 	if (chips == NULL) {
258 		mutex_exit(&cpu_lock);
259 		return (0);
260 	}
261 
262 	group_iter_init(&i);
263 	while ((chip = group_iterate(chips, &i)) != NULL) {
264 		if (GROUP_SIZE(&chip->cmt_cpus_actv) > 1) {
265 			mutex_exit(&cpu_lock);
266 			return (-1);
267 		}
268 	}
269 
270 	mutex_exit(&cpu_lock);
271 	return (0);
272 }
273