xref: /illumos-gate/usr/src/uts/i86pc/cpu/generic_cpu/gcpu_main.c (revision b445c7c6f2f09c2296534b5ccda2c05321c474b1)
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 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  * Copyright (c) 2018, Joyent, Inc.
26  */
27 /*
28  * Copyright (c) 2010, Intel Corporation.
29  * All rights reserved.
30  */
31 
32 /*
33  * Copyright (c) 2018, Joyent, Inc.
34  * Copyright 2020 RackTop Systems, Inc.
35  */
36 
37 /*
38  * Generic x86 CPU Module
39  *
40  * This CPU module is used for generic x86 CPUs when Solaris has no other
41  * CPU-specific support module available.  Code in this module should be the
42  * absolute bare-bones support and must be cognizant of both Intel and AMD etc.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/cpu_module_impl.h>
47 #include <sys/cpuvar.h>
48 #include <sys/kmem.h>
49 #include <sys/modctl.h>
50 #include <sys/pghw.h>
51 #include <sys/x86_archext.h>
52 
53 #include "gcpu.h"
54 
55 /*
56  * Prevent generic cpu support from loading.
57  */
58 int gcpu_disable = 0;
59 
60 #define	GCPU_MAX_CHIPID		32
61 static struct gcpu_chipshared *gcpu_shared[GCPU_MAX_CHIPID];
62 #ifdef	DEBUG
63 int gcpu_id_disable = 0;
64 static const char *gcpu_id_override[GCPU_MAX_CHIPID] = { NULL };
65 #endif
66 
67 #ifndef	__xpv
68 
69 /*
70  * The purpose of this is to construct a unique identifier for a given processor
71  * that can be used by things like FMA to determine when a FRU has been
72  * replaced. It is supported on Intel Xeon Platforms since Ivy Bridge and AMD
73  * 17h processors since Rome. See cpuid_pass1_ppin() for how we determine if a
74  * CPU is supported.
75  *
76  * The protected processor inventory number (PPIN) can be used to create a
77  * unique identifier when combined with the processor's cpuid signature. We
78  * create a versioned, synthetic ID using the following scheme for the
79  * identifier: iv0-<vendor>-<signature>-<PPIN>. The iv0 is the illumos version
80  * zero of the ID. If we have a new scheme for a new generation of processors,
81  * then that should rev the version field, otherwise for a given processor, this
82  * synthetic ID should not change.
83  *
84  * We use the string "INTC" for Intel and "AMD" for AMD. None of these or the
85  * formatting of the values can change without changing the version string.
86  */
87 static char *
88 gcpu_init_ident_ppin(cmi_hdl_t hdl)
89 {
90 	uint_t ppin_ctl_msr, ppin_msr;
91 	uint64_t value;
92 	const char *vendor;
93 
94 	/*
95 	 * This list should be extended as new Intel Xeon family processors come
96 	 * out.
97 	 */
98 	switch (cmi_hdl_vendor(hdl)) {
99 	case X86_VENDOR_Intel:
100 		ppin_ctl_msr = MSR_PPIN_CTL_INTC;
101 		ppin_msr = MSR_PPIN_INTC;
102 		vendor = "INTC";
103 		break;
104 	case X86_VENDOR_AMD:
105 		ppin_ctl_msr = MSR_PPIN_CTL_AMD;
106 		ppin_msr = MSR_PPIN_AMD;
107 		vendor = "AMD";
108 		break;
109 	default:
110 		return (NULL);
111 	}
112 
113 	if (cmi_hdl_rdmsr(hdl, ppin_ctl_msr, &value) != CMI_SUCCESS) {
114 		return (NULL);
115 	}
116 
117 	/*
118 	 * If the PPIN is not enabled and not locked, attempt to enable it.
119 	 * Note: in some environments such as Amazon EC2 the PPIN appears
120 	 * to be disabled and unlocked but our attempts to enable it don't
121 	 * stick, and when we attempt to read the PPIN we get an uncaught
122 	 * #GP. To avoid that happening we read the MSR back and verify it
123 	 * has taken the new value.
124 	 */
125 	if ((value & MSR_PPIN_CTL_ENABLED) == 0) {
126 		if ((value & MSR_PPIN_CTL_LOCKED) != 0) {
127 			return (NULL);
128 		}
129 
130 		if (cmi_hdl_wrmsr(hdl, ppin_ctl_msr, MSR_PPIN_CTL_ENABLED) !=
131 		    CMI_SUCCESS) {
132 			return (NULL);
133 		}
134 
135 		if (cmi_hdl_rdmsr(hdl, ppin_ctl_msr, &value) != CMI_SUCCESS) {
136 			return (NULL);
137 		}
138 
139 		if ((value & MSR_PPIN_CTL_ENABLED) == 0) {
140 			return (NULL);
141 		}
142 	}
143 
144 	if (cmi_hdl_rdmsr(hdl, ppin_msr, &value) != CMI_SUCCESS) {
145 		return (NULL);
146 	}
147 
148 	/*
149 	 * Now that we've read data, lock the PPIN. Don't worry about success or
150 	 * failure of this part, as we will have gotten everything that we need.
151 	 * It is possible that it locked open, for example.
152 	 */
153 	(void) cmi_hdl_wrmsr(hdl, ppin_ctl_msr, MSR_PPIN_CTL_LOCKED);
154 
155 	return (kmem_asprintf("iv0-%s-%x-%llx", vendor, cmi_hdl_chipsig(hdl),
156 	    value));
157 }
158 #endif	/* __xpv */
159 
160 static void
161 gcpu_init_ident(cmi_hdl_t hdl, struct gcpu_chipshared *sp)
162 {
163 #ifdef	DEBUG
164 	uint_t chipid;
165 
166 	/*
167 	 * On debug, allow a developer to override the string to more
168 	 * easily test CPU autoreplace without needing to physically
169 	 * replace a CPU.
170 	 */
171 	if (gcpu_id_disable != 0) {
172 		return;
173 	}
174 
175 	chipid = cmi_hdl_chipid(hdl);
176 	if (gcpu_id_override[chipid] != NULL) {
177 		sp->gcpus_ident = strdup(gcpu_id_override[chipid]);
178 		return;
179 	}
180 #endif
181 
182 #ifndef __xpv
183 	if (is_x86_feature(x86_featureset, X86FSET_PPIN)) {
184 		sp->gcpus_ident = gcpu_init_ident_ppin(hdl);
185 	}
186 #endif	/* __xpv */
187 }
188 
189 /*
190  * Our cmi_init entry point, called during startup of each cpu instance.
191  */
192 int
193 gcpu_init(cmi_hdl_t hdl, void **datap)
194 {
195 	uint_t chipid = cmi_hdl_chipid(hdl);
196 	struct gcpu_chipshared *sp, *osp;
197 	gcpu_data_t *gcpu;
198 
199 	if (gcpu_disable || chipid >= GCPU_MAX_CHIPID)
200 		return (ENOTSUP);
201 
202 	/*
203 	 * Allocate the state structure for this cpu.  We will only
204 	 * allocate the bank logout areas in gcpu_mca_init once we
205 	 * know how many banks there are.
206 	 */
207 	gcpu = *datap = kmem_zalloc(sizeof (gcpu_data_t), KM_SLEEP);
208 	cmi_hdl_hold(hdl);	/* release in gcpu_fini */
209 	gcpu->gcpu_hdl = hdl;
210 
211 	/*
212 	 * Allocate a chipshared structure if no sibling cpu has already
213 	 * allocated it, but allow for the fact that a sibling core may
214 	 * be starting up in parallel.
215 	 */
216 	if ((sp = gcpu_shared[chipid]) == NULL) {
217 		sp = kmem_zalloc(sizeof (struct gcpu_chipshared), KM_SLEEP);
218 		mutex_init(&sp->gcpus_poll_lock, NULL, MUTEX_DRIVER, NULL);
219 		mutex_init(&sp->gcpus_cfglock, NULL, MUTEX_DRIVER, NULL);
220 		osp = atomic_cas_ptr(&gcpu_shared[chipid], NULL, sp);
221 		if (osp != NULL) {
222 			mutex_destroy(&sp->gcpus_cfglock);
223 			mutex_destroy(&sp->gcpus_poll_lock);
224 			kmem_free(sp, sizeof (struct gcpu_chipshared));
225 			sp = osp;
226 		} else {
227 			gcpu_init_ident(hdl, sp);
228 		}
229 	}
230 
231 	atomic_inc_32(&sp->gcpus_actv_cnt);
232 	gcpu->gcpu_shared = sp;
233 
234 	return (0);
235 }
236 
237 /*
238  * deconfigure gcpu_init()
239  */
240 void
241 gcpu_fini(cmi_hdl_t hdl)
242 {
243 	uint_t chipid = cmi_hdl_chipid(hdl);
244 	gcpu_data_t *gcpu = cmi_hdl_getcmidata(hdl);
245 	struct gcpu_chipshared *sp;
246 
247 	if (gcpu_disable || chipid >= GCPU_MAX_CHIPID)
248 		return;
249 
250 	gcpu_mca_fini(hdl);
251 
252 	/*
253 	 * Keep shared data in cache for reuse.
254 	 */
255 	sp = gcpu_shared[chipid];
256 	ASSERT(sp != NULL);
257 	atomic_dec_32(&sp->gcpus_actv_cnt);
258 
259 	if (gcpu != NULL)
260 		kmem_free(gcpu, sizeof (gcpu_data_t));
261 
262 	/* Release reference count held in gcpu_init(). */
263 	cmi_hdl_rele(hdl);
264 }
265 
266 void
267 gcpu_post_startup(cmi_hdl_t hdl)
268 {
269 	gcpu_data_t *gcpu = cmi_hdl_getcmidata(hdl);
270 
271 	if (gcpu_disable)
272 		return;
273 
274 	if (gcpu != NULL)
275 		cms_post_startup(hdl);
276 #ifdef __xpv
277 	/*
278 	 * All cpu handles are initialized so we can begin polling now.
279 	 * Furthermore, our virq mechanism requires that everything
280 	 * be run on cpu 0 so we can assure that by starting from here.
281 	 */
282 	gcpu_mca_poll_start(hdl);
283 #else
284 	/*
285 	 * The boot CPU has a bit of a chicken and egg problem for CMCI. Its MCA
286 	 * initialization is run before we have initialized the PSM module that
287 	 * we would use for enabling CMCI. Therefore, we use this as a chance to
288 	 * enable CMCI for the boot CPU. For all other CPUs, this chicken and
289 	 * egg problem will have already been solved.
290 	 */
291 	gcpu_mca_cmci_enable(hdl);
292 #endif
293 }
294 
295 void
296 gcpu_post_mpstartup(cmi_hdl_t hdl)
297 {
298 	if (gcpu_disable)
299 		return;
300 
301 	cms_post_mpstartup(hdl);
302 
303 #ifndef __xpv
304 	/*
305 	 * All cpu handles are initialized only once all cpus are started, so we
306 	 * can begin polling post mp startup.
307 	 */
308 	gcpu_mca_poll_start(hdl);
309 #endif
310 }
311 
312 const char *
313 gcpu_ident(cmi_hdl_t hdl)
314 {
315 	uint_t chipid;
316 	struct gcpu_chipshared *sp;
317 
318 	if (gcpu_disable)
319 		return (NULL);
320 
321 	chipid = cmi_hdl_chipid(hdl);
322 	if (chipid >= GCPU_MAX_CHIPID)
323 		return (NULL);
324 
325 	if (cmi_hdl_getcmidata(hdl) == NULL)
326 		return (NULL);
327 
328 	sp = gcpu_shared[cmi_hdl_chipid(hdl)];
329 	return (sp->gcpus_ident);
330 }
331 
332 #ifdef __xpv
333 #define	GCPU_OP(ntvop, xpvop)	xpvop
334 #else
335 #define	GCPU_OP(ntvop, xpvop)	ntvop
336 #endif
337 
338 cmi_api_ver_t _cmi_api_version = CMI_API_VERSION_3;
339 
340 const cmi_ops_t _cmi_ops = {
341 	gcpu_init,				/* cmi_init */
342 	gcpu_post_startup,			/* cmi_post_startup */
343 	gcpu_post_mpstartup,			/* cmi_post_mpstartup */
344 	gcpu_faulted_enter,			/* cmi_faulted_enter */
345 	gcpu_faulted_exit,			/* cmi_faulted_exit */
346 	gcpu_mca_init,				/* cmi_mca_init */
347 	GCPU_OP(gcpu_mca_trap, NULL),		/* cmi_mca_trap */
348 	GCPU_OP(gcpu_cmci_trap, NULL),		/* cmi_cmci_trap */
349 	gcpu_msrinject,				/* cmi_msrinject */
350 	GCPU_OP(gcpu_hdl_poke, NULL),		/* cmi_hdl_poke */
351 	gcpu_fini,				/* cmi_fini */
352 	GCPU_OP(NULL, gcpu_xpv_panic_callback),	/* cmi_panic_callback */
353 	gcpu_ident				/* cmi_ident */
354 };
355 
356 static struct modlcpu modlcpu = {
357 	&mod_cpuops,
358 	"Generic x86 CPU Module"
359 };
360 
361 static struct modlinkage modlinkage = {
362 	MODREV_1,
363 	(void *)&modlcpu,
364 	NULL
365 };
366 
367 int
368 _init(void)
369 {
370 	return (mod_install(&modlinkage));
371 }
372 
373 int
374 _info(struct modinfo *modinfop)
375 {
376 	return (mod_info(&modlinkage, modinfop));
377 }
378 
379 int
380 _fini(void)
381 {
382 	return (mod_remove(&modlinkage));
383 }
384