xref: /illumos-gate/usr/src/uts/i86pc/cpu/amd_opteron/ao_mca.c (revision d6bb6a8465e557cb946ef49d56ed3202f6218652)
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 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/regset.h>
31 #include <sys/privregs.h>
32 #include <sys/pci_impl.h>
33 #include <sys/cpuvar.h>
34 #include <sys/x86_archext.h>
35 #include <sys/cmn_err.h>
36 #include <sys/systm.h>
37 #include <sys/sysmacros.h>
38 #include <sys/chip.h>
39 #include <sys/cyclic.h>
40 #include <sys/cpu_module_impl.h>
41 #include <sys/pci_cfgspace_impl.h>
42 #include <sys/sysevent.h>
43 #include <sys/smbios.h>
44 #include <sys/mca_x86.h>
45 #include <sys/mca_amd.h>
46 #include <sys/mc.h>
47 #include <sys/psw.h>
48 #include <sys/ddi.h>
49 #include <sys/sunddi.h>
50 #include <sys/sdt.h>
51 #include <sys/fm/util.h>
52 #include <sys/fm/protocol.h>
53 #include <sys/fm/cpu/AMD.h>
54 
55 #include "ao.h"
56 #include "ao_mca_disp.h"
57 
58 errorq_t *ao_mca_queue;			/* machine-check ereport queue */
59 int ao_mca_stack_flag = 0;		/* record stack trace in ereports */
60 int ao_mca_smi_disable = 1;		/* attempt to disable SMI polling */
61 
62 ao_bank_regs_t ao_bank_regs[AMD_MCA_BANK_COUNT] = {
63 	{ AMD_MSR_DC_STATUS, AMD_MSR_DC_ADDR },
64 	{ AMD_MSR_IC_STATUS, AMD_MSR_IC_ADDR },
65 	{ AMD_MSR_BU_STATUS, AMD_MSR_BU_ADDR },
66 	{ AMD_MSR_LS_STATUS, AMD_MSR_LS_ADDR },
67 	{ AMD_MSR_NB_STATUS, AMD_MSR_NB_ADDR }
68 };
69 
70 typedef struct ao_bank_cfg {
71 	uint_t bank_ctl;
72 	uint_t bank_ctl_mask;
73 	uint64_t bank_ctl_init;
74 	uint_t bank_status;
75 	uint_t bank_addr;
76 } ao_bank_cfg_t;
77 
78 static const ao_bank_cfg_t ao_bank_cfgs[] = {
79 	{ AMD_MSR_DC_CTL, AMD_MSR_DC_MASK, AMD_DC_CTL_INIT, AMD_MSR_DC_STATUS,
80 	    AMD_MSR_DC_ADDR },
81 	{ AMD_MSR_IC_CTL, AMD_MSR_IC_MASK, AMD_IC_CTL_INIT, AMD_MSR_IC_STATUS,
82 	    AMD_MSR_IC_ADDR },
83 	{ AMD_MSR_BU_CTL, AMD_MSR_BU_MASK, AMD_BU_CTL_INIT, AMD_MSR_BU_STATUS,
84 	    AMD_MSR_BU_ADDR },
85 	{ AMD_MSR_LS_CTL, AMD_MSR_LS_MASK, AMD_LS_CTL_INIT, AMD_MSR_LS_STATUS,
86 	    AMD_MSR_LS_ADDR },
87 	{ AMD_MSR_NB_CTL, AMD_MSR_NB_MASK, AMD_NB_CTL_INIT, AMD_MSR_NB_STATUS,
88 	    AMD_MSR_NB_ADDR }
89 };
90 
91 static const ao_error_disp_t ao_disp_unknown = {
92 	FM_EREPORT_CPU_AMD_UNKNOWN,
93 	FM_EREPORT_PAYLOAD_FLAGS_CPU_AMD_UNKNOWN
94 };
95 
96 /*
97  * This is quite awful but necessary to work around x86 system vendor's view of
98  * the world.  Other operating systems (you know who you are) don't understand
99  * Opteron-specific error handling, so BIOS and system vendors often hide these
100  * conditions from them by using SMI polling to copy out any errors from the
101  * machine-check registers.  When Solaris runs on a system with this feature,
102  * we want to disable the SMI polling so we can use FMA instead.  Sadly, there
103  * isn't even a standard self-describing way to express the whole situation,
104  * so we have to resort to hard-coded values.  This should all be changed to
105  * be a self-describing vendor-specific SMBIOS structure in the future.
106  */
107 static const struct ao_smi_disable {
108 	const char *asd_sys_vendor;	/* SMB_TYPE_SYSTEM vendor prefix */
109 	const char *asd_bios_vendor;	/* SMB_TYPE_BIOS vendor prefix */
110 	uint32_t asd_port;		/* output port for SMI disable */
111 	uint32_t asd_code;		/* output code for SMI disable */
112 } ao_smi_disable[] = {
113 	{ "Sun Microsystems", "American Megatrends", 0x502F, 0x59 },
114 	{ NULL, NULL, 0, 0 }
115 };
116 
117 static int
118 ao_disp_match_r4(uint16_t ref, uint8_t r4)
119 {
120 	static const uint16_t ao_r4_map[] = {
121 		AO_MCA_R4_BIT_GEN,	/* AMD_ERRCODE_R4_GEN */
122 		AO_MCA_R4_BIT_RD,	/* AMD_ERRCODE_R4_RD */
123 		AO_MCA_R4_BIT_WR,	/* AMD_ERRCODE_R4_WR */
124 		AO_MCA_R4_BIT_DRD,	/* AMD_ERRCODE_R4_DRD */
125 		AO_MCA_R4_BIT_DWR,	/* AMD_ERRCODE_R4_DWR */
126 		AO_MCA_R4_BIT_IRD,	/* AMD_ERRCODE_R4_IRD */
127 		AO_MCA_R4_BIT_PREFETCH,	/* AMD_ERRCODE_R4_PREFETCH */
128 		AO_MCA_R4_BIT_EVICT,	/* AMD_ERRCODE_R4_EVICT */
129 		AO_MCA_R4_BIT_SNOOP	/* AMD_ERRCODE_R4_SNOOP */
130 	};
131 
132 	ASSERT(r4 < sizeof (ao_r4_map) / sizeof (uint16_t));
133 
134 	return ((ref & ao_r4_map[r4]) != 0);
135 }
136 
137 static int
138 ao_disp_match_pp(uint8_t ref, uint8_t pp)
139 {
140 	static const uint8_t ao_pp_map[] = {
141 		AO_MCA_PP_BIT_SRC,	/* AMD_ERRCODE_PP_SRC */
142 		AO_MCA_PP_BIT_RSP,	/* AMD_ERRCODE_PP_RSP */
143 		AO_MCA_PP_BIT_OBS,	/* AMD_ERRCODE_PP_OBS */
144 		AO_MCA_PP_BIT_GEN	/* AMD_ERRCODE_PP_GEN */
145 	};
146 
147 	ASSERT(pp < sizeof (ao_pp_map) / sizeof (uint8_t));
148 
149 	return ((ref & ao_pp_map[pp]) != 0);
150 }
151 
152 static int
153 ao_disp_match_ii(uint8_t ref, uint8_t ii)
154 {
155 	static const uint8_t ao_ii_map[] = {
156 		AO_MCA_II_BIT_MEM,	/* AMD_ERRCODE_II_MEM */
157 		0,
158 		AO_MCA_II_BIT_IO,	/* AMD_ERRCODE_II_IO */
159 		AO_MCA_II_BIT_GEN	/* AMD_ERRCODE_II_GEN */
160 	};
161 
162 	ASSERT(ii < sizeof (ao_ii_map) / sizeof (uint8_t));
163 
164 	return ((ref & ao_ii_map[ii]) != 0);
165 }
166 
167 static uint8_t
168 bit_strip(uint16_t *codep, uint16_t mask, uint16_t shift)
169 {
170 	uint8_t val = (*codep & mask) >> shift;
171 	*codep &= ~mask;
172 	return (val);
173 }
174 
175 #define	BIT_STRIP(codep, name) \
176 	bit_strip(codep, AMD_ERRCODE_##name##_MASK, AMD_ERRCODE_##name##_SHIFT)
177 
178 static int
179 ao_disp_match_one(const ao_error_disp_t *aed, uint64_t status)
180 {
181 	uint16_t code = status & AMD_ERRCODE_MASK;
182 	uint8_t extcode = (status & AMD_ERREXT_MASK) >> AMD_ERREXT_SHIFT;
183 	uint64_t stat_mask = aed->aed_stat_mask;
184 	uint64_t stat_mask_res = aed->aed_stat_mask_res;
185 
186 	/*
187 	 * If the bank's status register indicates overflow, then we can no
188 	 * longer rely on the value of CECC: our experience with actual fault
189 	 * injection has shown that multiple CE's overwriting each other shows
190 	 * AMD_BANK_STAT_CECC and AMD_BANK_STAT_UECC both set to zero.  This
191 	 * should be clarified in a future BKDG or by the Revision Guide.
192 	 */
193 	if (status & AMD_BANK_STAT_OVER) {
194 		stat_mask &= ~AMD_BANK_STAT_CECC;
195 		stat_mask_res &= ~AMD_BANK_STAT_CECC;
196 	}
197 
198 	if ((status & stat_mask) != stat_mask_res)
199 		return (0);
200 
201 	/*
202 	 * r4 and pp bits are stored separately, so we mask off and compare them
203 	 * for the code types that use them.  Once we've taken the r4 and pp
204 	 * bits out of the equation, we can directly compare the resulting code
205 	 * with the one stored in the ao_error_disp_t.
206 	 */
207 	if (AMD_ERRCODE_ISMEM(code)) {
208 		uint8_t r4 = BIT_STRIP(&code, R4);
209 
210 		if (!ao_disp_match_r4(aed->aed_stat_r4_bits, r4))
211 			return (0);
212 
213 	} else if (AMD_ERRCODE_ISBUS(code)) {
214 		uint8_t r4 = BIT_STRIP(&code, R4);
215 		uint8_t pp = BIT_STRIP(&code, PP);
216 		uint8_t ii = BIT_STRIP(&code, II);
217 
218 		if (!ao_disp_match_r4(aed->aed_stat_r4_bits, r4) ||
219 		    !ao_disp_match_pp(aed->aed_stat_pp_bits, pp) ||
220 		    !ao_disp_match_ii(aed->aed_stat_ii_bits, ii))
221 			return (0);
222 	}
223 
224 	return (code == aed->aed_stat_code && extcode == aed->aed_stat_extcode);
225 }
226 
227 static const ao_error_disp_t *
228 ao_disp_match(uint_t bankno, uint64_t status)
229 {
230 	const ao_error_disp_t *aed;
231 
232 	for (aed = ao_error_disp[bankno]; aed->aed_stat_mask != 0; aed++) {
233 		if (ao_disp_match_one(aed, status))
234 			return (aed);
235 	}
236 
237 	return (&ao_disp_unknown);
238 }
239 
240 void
241 ao_pcicfg_write(uint_t chipid, uint_t func, uint_t reg, uint32_t val)
242 {
243 	ASSERT(chipid + 24 <= 31);
244 	ASSERT((func & 7) == func);
245 	ASSERT((reg & 3) == 0 && reg < 256);
246 
247 	pci_mech1_putl(0, chipid + 24, func, reg, val);
248 }
249 
250 uint32_t
251 ao_pcicfg_read(uint_t chipid, uint_t func, uint_t reg)
252 {
253 	ASSERT(chipid + 24 <= 31);
254 	ASSERT((func & 7) == func);
255 	ASSERT((reg & 3) == 0 && reg < 256);
256 
257 	return (pci_mech1_getl(0, chipid + 24, func, reg));
258 }
259 
260 /*
261  * Setup individual bank detectors after stashing their bios settings.
262  */
263 static void
264 ao_bank_cfg(ao_mca_t *mca)
265 {
266 	ao_bios_cfg_t *bioscfg = &mca->ao_mca_bios_cfg;
267 	const ao_bank_cfg_t *bankcfg = ao_bank_cfgs;
268 	int i;
269 
270 	for (i = 0; i < AMD_MCA_BANK_COUNT; i++, bankcfg++) {
271 		bioscfg->bcfg_bank_ctl[i] = rdmsr(bankcfg->bank_ctl);
272 		bioscfg->bcfg_bank_mask[i] = rdmsr(bankcfg->bank_ctl_mask);
273 		wrmsr(bankcfg->bank_ctl, bankcfg->bank_ctl_init);
274 	}
275 }
276 
277 /*
278  * Bits to be added to the NorthBridge (NB) configuration register.
279  * See BKDG 3.29 Section 3.6.4.2 for more information.
280  */
281 uint32_t ao_nb_cfg_add =
282     AMD_NB_CFG_NBMCATOMSTCPUEN |
283     AMD_NB_CFG_DISPCICFGCPUERRRSP |
284     AMD_NB_CFG_SYNCONUCECCEN |
285     AMD_NB_CFG_CPUECCERREN;
286 
287 /*
288  * Bits to be cleared from the NorthBridge (NB) configuration register.
289  * See BKDG 3.29 Section 3.6.4.2 for more information.
290  */
291 uint32_t ao_nb_cfg_remove =
292     AMD_NB_CFG_IORDDATERREN |
293     AMD_NB_CFG_SYNCONANYERREN |
294     AMD_NB_CFG_SYNCONWDOGEN |
295     AMD_NB_CFG_IOERRDIS |
296     AMD_NB_CFG_IOMSTABORTDIS |
297     AMD_NB_CFG_SYNCPKTPROPDIS |
298     AMD_NB_CFG_SYNCPKTGENDIS;
299 
300 /*
301  * Bits to be used if we configure the NorthBridge (NB) Watchdog.  The watchdog
302  * triggers a machine check exception when no response to an NB system access
303  * occurs within a specified time interval.  If the BIOS (i.e. platform design)
304  * has enabled the watchdog, we leave its rate alone.  If the BIOS has not
305  * enabled the watchdog, we enable it and set the rate to one specified below.
306  * To disable the watchdog, add the AMD_NB_CFG_WDOGTMRDIS bit to ao_nb_cfg_add.
307  */
308 uint32_t ao_nb_cfg_wdog =
309     AMD_NB_CFG_WDOGTMRCNTSEL_4095 |
310     AMD_NB_CFG_WDOGTMRBASESEL_1MS;
311 
312 static void
313 ao_nb_cfg(ao_mca_t *mca)
314 {
315 	uint_t chipid = chip_plat_get_chipid(CPU);
316 	uint32_t val;
317 
318 	if (chip_plat_get_clogid(CPU) != 0)
319 		return; /* only configure NB once per CPU */
320 
321 	/*
322 	 * Read the NorthBridge (NB) configuration register in PCI space,
323 	 * modify the settings accordingly, and store the new value back.
324 	 */
325 	mca->ao_mca_bios_cfg.bcfg_nb_cfg = val =
326 	    ao_pcicfg_read(chipid, AMD_NB_FUNC, AMD_NB_REG_CFG);
327 
328 	/*
329 	 * If the watchdog was disabled, enable it according to the policy
330 	 * described above.  Then apply the ao_nb_cfg_[add|remove] masks.
331 	 */
332 	if (val & AMD_NB_CFG_WDOGTMRDIS) {
333 		val &= ~AMD_NB_CFG_WDOGTMRBASESEL_MASK;
334 		val &= ~AMD_NB_CFG_WDOGTMRCNTSEL_MASK;
335 		val &= ~AMD_NB_CFG_WDOGTMRDIS;
336 		val |= ao_nb_cfg_wdog;
337 	}
338 
339 	val &= ~ao_nb_cfg_remove;
340 	val |= ao_nb_cfg_add;
341 
342 	ao_pcicfg_write(chipid, AMD_NB_FUNC, AMD_NB_REG_CFG, val);
343 }
344 
345 /*
346  * Capture the machine-check exception state into our per-CPU logout area, and
347  * dispatch a copy of the logout area to our error queue for ereport creation.
348  * If 'rp' is non-NULL, we're being called from trap context; otherwise we're
349  * being polled or poked by the injector.  We return the number of errors
350  * found through 'np', and a boolean indicating whether the error is fatal.
351  * The caller is expected to call fm_panic() if we return fatal (non-zero).
352  */
353 int
354 ao_mca_logout(ao_cpu_logout_t *acl, struct regs *rp, int *np)
355 {
356 	int i, fatal = 0, n = 0;
357 
358 	acl->acl_timestamp = gethrtime_waitfree();
359 	acl->acl_mcg_status = rdmsr(IA32_MSR_MCG_STATUS);
360 	acl->acl_ip = rp ? rp->r_pc : 0;
361 	acl->acl_flags = 0;
362 
363 	/*
364 	 * Iterate over the banks of machine-check registers, read the address
365 	 * and status registers into the logout area, and clear them as we go.
366 	 */
367 	for (i = 0; i < AMD_MCA_BANK_COUNT; i++) {
368 		ao_bank_logout_t *abl = &acl->acl_banks[i];
369 
370 		abl->abl_addr = rdmsr(ao_bank_regs[i].abr_addr);
371 		abl->abl_status = rdmsr(ao_bank_regs[i].abr_status);
372 
373 		if (abl->abl_status & AMD_BANK_STAT_VALID)
374 			wrmsr(ao_bank_regs[i].abr_status, 0);
375 	}
376 
377 	if (rp == NULL || !USERMODE(rp->r_cs))
378 		acl->acl_flags |= AO_ACL_F_PRIV;
379 
380 	if (ao_mca_stack_flag)
381 		acl->acl_stackdepth = getpcstack(acl->acl_stack, FM_STK_DEPTH);
382 	else
383 		acl->acl_stackdepth = 0;
384 
385 	/*
386 	 * Clear MCG_STATUS, indicating that machine-check trap processing is
387 	 * complete.  Once we do this, another machine-check trap can occur.
388 	 */
389 	wrmsr(IA32_MSR_MCG_STATUS, 0);
390 
391 	/*
392 	 * If we took a machine-check trap, then the error is fatal if the
393 	 * return instruction pointer is not valid in the global register.
394 	 */
395 	if (rp != NULL && !(acl->acl_mcg_status & MCG_STATUS_RIPV))
396 		fatal++;
397 
398 	/*
399 	 * Now iterate over the saved logout area, determining whether the
400 	 * error that we saw is fatal or not based upon our dispositions
401 	 * and the hardware's indicators of whether or not we can resume.
402 	 */
403 	for (i = 0; i < AMD_MCA_BANK_COUNT; i++) {
404 		ao_bank_logout_t *abl = &acl->acl_banks[i];
405 		const ao_error_disp_t *aed;
406 		uint8_t when;
407 
408 		if (!(abl->abl_status & AMD_BANK_STAT_VALID))
409 			continue;
410 
411 		aed = ao_disp_match(i, abl->abl_status);
412 		if ((when = aed->aed_panic_when) != AO_AED_PANIC_NEVER) {
413 			if ((when & AO_AED_PANIC_ALWAYS) ||
414 			    ((when & AO_AED_PANIC_IFMCE) && rp != NULL)) {
415 					fatal++;
416 			}
417 		}
418 
419 		/*
420 		 * If we are taking a machine-check exception and the overflow
421 		 * bit is set or our context is corrupt, then we must die.
422 		 * NOTE: This code assumes that if the overflow bit is set and
423 		 * we didn't take a #mc exception (i.e. the poller found it),
424 		 * then multiple correctable errors overwrote each other.
425 		 * This will need to change if we eventually use the Opteron
426 		 * Rev E exception mechanism for detecting correctable errors.
427 		 */
428 		if (rp != NULL && (abl->abl_status &
429 		    (AMD_BANK_STAT_OVER | AMD_BANK_STAT_PCC)))
430 			fatal++;
431 
432 		/*
433 		 * If we are taking a machine-check exception and we don't
434 		 * recognize the error case at all, then assume it's fatal.
435 		 * This will need to change if we eventually use the Opteron
436 		 * Rev E exception mechanism for detecting correctable errors.
437 		 */
438 		if (rp != NULL && aed == &ao_disp_unknown)
439 			fatal++;
440 
441 		n++;
442 	}
443 
444 	if (n > 0) {
445 		errorq_dispatch(ao_mca_queue, acl, sizeof (ao_cpu_logout_t),
446 		    fatal && cmi_panic_on_uncorrectable_error ?
447 		    ERRORQ_SYNC : ERRORQ_ASYNC);
448 	}
449 
450 	if (np != NULL)
451 		*np = n; /* return number of errors found to caller */
452 
453 	return (fatal != 0);
454 }
455 
456 static uint_t
457 ao_ereport_synd(ao_mca_t *mca,
458     const ao_bank_logout_t *abl, uint_t *typep, int is_nb)
459 {
460 	if (is_nb) {
461 		if ((mca->ao_mca_bios_cfg.bcfg_nb_cfg &
462 		    AMD_NB_CFG_CHIPKILLECCEN) != 0) {
463 			*typep = AMD_SYNDTYPE_CHIPKILL;
464 			return (AMD_NB_STAT_CKSYND(abl->abl_status));
465 		} else {
466 			*typep = AMD_SYNDTYPE_ECC;
467 			return (AMD_BANK_SYND(abl->abl_status));
468 		}
469 	} else {
470 		*typep = AMD_SYNDTYPE_ECC;
471 		return (AMD_BANK_SYND(abl->abl_status));
472 	}
473 }
474 
475 static void
476 ao_ereport_create_resource_elem(nvlist_t **nvlp, nv_alloc_t *nva,
477     mc_unum_t *unump, int dimmnum)
478 {
479 	nvlist_t *snvl;
480 	*nvlp = fm_nvlist_create(nva);		/* freed by caller */
481 
482 	snvl = fm_nvlist_create(nva);
483 
484 	(void) nvlist_add_uint64(snvl, FM_FMRI_HC_SPECIFIC_OFFSET,
485 	    unump->unum_offset);
486 
487 	fm_fmri_hc_set(*nvlp, FM_HC_SCHEME_VERSION, NULL, snvl, 4,
488 	    "motherboard", unump->unum_board,
489 	    "chip", unump->unum_chip,
490 	    "memory-controller", unump->unum_mc,
491 	    "dimm", unump->unum_dimms[dimmnum]);
492 
493 	fm_nvlist_destroy(snvl, FM_NVA_FREE);
494 }
495 
496 static void
497 ao_ereport_add_resource(nvlist_t *payload, nv_alloc_t *nva, mc_unum_t *unump)
498 {
499 
500 	nvlist_t *elems[MC_UNUM_NDIMM];
501 	int nelems = 0;
502 	int i;
503 
504 	for (i = 0; i < MC_UNUM_NDIMM; i++) {
505 		if (unump->unum_dimms[i] == -1)
506 			break;
507 		ao_ereport_create_resource_elem(&elems[nelems++], nva,
508 		    unump, i);
509 	}
510 
511 	fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_RESOURCE,
512 	    DATA_TYPE_NVLIST_ARRAY, nelems, elems, NULL);
513 
514 	for (i = 0; i < nelems; i++)
515 		fm_nvlist_destroy(elems[i], FM_NVA_FREE);
516 }
517 
518 static void
519 ao_ereport_add_logout(ao_data_t *ao, nvlist_t *payload, nv_alloc_t *nva,
520     const ao_cpu_logout_t *acl, uint_t bankno, const ao_error_disp_t *aed)
521 {
522 	uint64_t members = aed->aed_ereport_members;
523 	ao_mca_t *mca = &ao->ao_mca;
524 	const ao_bank_logout_t *abl = &acl->acl_banks[bankno];
525 	uint_t synd, syndtype;
526 
527 	synd = ao_ereport_synd(mca, abl, &syndtype, bankno == AMD_MCA_BANK_NB);
528 
529 	if (members & FM_EREPORT_PAYLOAD_FLAG_BANK_STAT) {
530 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_BANK_STAT,
531 		    DATA_TYPE_UINT64, abl->abl_status, NULL);
532 	}
533 
534 	if (members & FM_EREPORT_PAYLOAD_FLAG_BANK_NUM) {
535 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_BANK_NUM,
536 		    DATA_TYPE_UINT8, bankno, NULL);
537 	}
538 
539 	if (members & FM_EREPORT_PAYLOAD_FLAG_ADDR) {
540 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ADDR,
541 		    DATA_TYPE_UINT64, abl->abl_addr, NULL);
542 	}
543 
544 	if (members & FM_EREPORT_PAYLOAD_FLAG_ADDR_VALID) {
545 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ADDR_VALID,
546 		    DATA_TYPE_BOOLEAN_VALUE, (abl->abl_status &
547 		    AMD_BANK_STAT_ADDRV) ? B_TRUE : B_FALSE, NULL);
548 	}
549 
550 	if (members & FM_EREPORT_PAYLOAD_FLAG_SYND) {
551 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND,
552 		    DATA_TYPE_UINT16, synd, NULL);
553 	}
554 
555 	if (members & FM_EREPORT_PAYLOAD_FLAG_SYND_TYPE) {
556 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND_TYPE,
557 		    DATA_TYPE_STRING, (syndtype == AMD_SYNDTYPE_CHIPKILL ?
558 		    "C" : "E"), NULL);
559 	}
560 
561 	if (members & FM_EREPORT_PAYLOAD_FLAG_IP) {
562 		uint64_t ip = (acl->acl_mcg_status & MCG_STATUS_EIPV) ?
563 		    acl->acl_ip : 0;
564 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_IP,
565 		    DATA_TYPE_UINT64, ip, NULL);
566 	}
567 
568 	if (members & FM_EREPORT_PAYLOAD_FLAG_PRIV) {
569 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PRIV,
570 		    DATA_TYPE_BOOLEAN_VALUE, (acl->acl_flags & AO_ACL_F_PRIV) ?
571 		    B_TRUE : B_FALSE, NULL);
572 	}
573 
574 	if (members & FM_EREPORT_PAYLOAD_FLAG_RESOURCE) {
575 		mc_unum_t unum;
576 		int addrvalid;
577 
578 		addrvalid = (members & FM_EREPORT_PAYLOAD_FLAG_ADDR) &&
579 		    (members & FM_EREPORT_PAYLOAD_FLAG_ADDR_VALID) &&
580 		    (abl->abl_status & AMD_BANK_STAT_ADDRV);
581 
582 		if (addrvalid && ao_mc_patounum(ao, abl->abl_addr, synd,
583 		    syndtype, &unum))
584 			ao_ereport_add_resource(payload, nva, &unum);
585 	}
586 
587 	if (ao_mca_stack_flag && members & FM_EREPORT_PAYLOAD_FLAG_STACK) {
588 		fm_payload_stack_add(payload, acl->acl_stack,
589 		    acl->acl_stackdepth);
590 	}
591 }
592 
593 static void
594 ao_ereport_post(const ao_cpu_logout_t *acl,
595     int bankno, const ao_error_disp_t *aed)
596 {
597 	ao_data_t *ao = acl->acl_ao;
598 	errorq_elem_t *eqep;
599 	nvlist_t *ereport, *detector;
600 	nv_alloc_t *nva = NULL;
601 	char buf[FM_MAX_CLASS];
602 
603 	if (panicstr) {
604 		if ((eqep = errorq_reserve(ereport_errorq)) == NULL)
605 			return;
606 		ereport = errorq_elem_nvl(ereport_errorq, eqep);
607 		nva = errorq_elem_nva(ereport_errorq, eqep);
608 	} else {
609 		ereport = fm_nvlist_create(nva);
610 	}
611 
612 	/*
613 	 * Create the scheme "cpu" FMRI
614 	 */
615 	detector = ao_fmri_create(ao, nva);
616 
617 	/*
618 	 * Encode all the common data into the ereport.
619 	 */
620 	(void) snprintf(buf, FM_MAX_CLASS, "%s.%s.%s",
621 	    FM_ERROR_CPU, "amd", aed->aed_class);
622 
623 	fm_ereport_set(ereport, FM_EREPORT_VERSION, buf,
624 	    fm_ena_generate_cpu(acl->acl_timestamp, ao->ao_cpu->cpu_id,
625 	    FM_ENA_FMT1), detector, NULL);
626 
627 	/*
628 	 * Encode the error-specific data that was saved in the logout area.
629 	 */
630 	ao_ereport_add_logout(ao, ereport, nva, acl, bankno, aed);
631 
632 	if (panicstr) {
633 		errorq_commit(ereport_errorq, eqep, ERRORQ_SYNC);
634 	} else {
635 		(void) fm_ereport_post(ereport, EVCH_TRYHARD);
636 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
637 		fm_nvlist_destroy(detector, FM_NVA_FREE);
638 	}
639 }
640 
641 /*ARGSUSED*/
642 void
643 ao_mca_drain(void *ignored, const void *data, const errorq_elem_t *eqe)
644 {
645 	const ao_cpu_logout_t *acl = data;
646 	int i;
647 
648 	for (i = 0; i < AMD_MCA_BANK_COUNT; i++) {
649 		const ao_bank_logout_t *abl = &acl->acl_banks[i];
650 		const ao_error_disp_t *aed;
651 
652 		if (abl->abl_status & AMD_BANK_STAT_VALID) {
653 			aed = ao_disp_match(i, abl->abl_status);
654 			ao_ereport_post(acl, i, aed);
655 		}
656 	}
657 }
658 
659 int
660 ao_mca_trap(void *data, struct regs *rp)
661 {
662 	ao_data_t *ao = data;
663 	ao_mca_t *mca = &ao->ao_mca;
664 	ao_cpu_logout_t *acl = &mca->ao_mca_logout[AO_MCA_LOGOUT_EXCEPTION];
665 	return (ao_mca_logout(acl, rp, NULL));
666 }
667 
668 /*ARGSUSED*/
669 int
670 ao_mca_inject(void *data, cmi_mca_regs_t *regs, uint_t nregs)
671 {
672 	uint64_t hwcr, oldhwcr;
673 	int i;
674 
675 	oldhwcr = rdmsr(MSR_AMD_HWCR);
676 	hwcr = oldhwcr | AMD_HWCR_MCI_STATUS_WREN;
677 	wrmsr(MSR_AMD_HWCR, hwcr);
678 
679 	for (i = 0; i < nregs; i++)
680 		wrmsr(regs[i].cmr_msrnum, regs[i].cmr_msrval);
681 
682 	wrmsr(MSR_AMD_HWCR, oldhwcr);
683 	return (0);
684 }
685 
686 void
687 ao_mca_init(void *data)
688 {
689 	ao_data_t *ao = data;
690 	ao_mca_t *mca = &ao->ao_mca;
691 	uint64_t cap;
692 	int i;
693 
694 	ao_mca_poll_init(mca);
695 
696 	ASSERT(x86_feature & X86_MCA);
697 	cap = rdmsr(IA32_MSR_MCG_CAP);
698 	ASSERT(cap & MCG_CAP_CTL_P);
699 
700 	/*
701 	 * If the hardware's bank count is different than what we expect, then
702 	 * we're running on some Opteron variant that we don't understand yet.
703 	 */
704 	if ((cap & MCG_CAP_COUNT_MASK) != AMD_MCA_BANK_COUNT) {
705 		cmn_err(CE_WARN, "CPU %d has %llu MCA banks; expected %u: "
706 		    "disabling MCA on this CPU", ao->ao_cpu->cpu_id,
707 		    (u_longlong_t)cap & MCG_CAP_COUNT_MASK, AMD_MCA_BANK_COUNT);
708 		return;
709 	}
710 
711 	/*
712 	 * Configure the logout areas.  We preset every logout area's acl_ao
713 	 * pointer to refer back to our per-CPU state for errorq drain usage.
714 	 */
715 	for (i = 0; i < AO_MCA_LOGOUT_NUM; i++)
716 		mca->ao_mca_logout[i].acl_ao = ao;
717 
718 	ao_bank_cfg(mca);
719 	ao_nb_cfg(mca);
720 
721 	wrmsr(IA32_MSR_MCG_CTL, AMD_MCG_EN_ALL);
722 
723 	/*
724 	 * Throw away all existing bank state.  We do this because some BIOSes,
725 	 * perhaps during POST, do things to the machine that cause MCA state
726 	 * to be updated.  If we interpret this state as an actual error, we
727 	 * may end up indicting something that's not actually broken.
728 	 */
729 	for (i = 0; i < sizeof (ao_bank_cfgs) / sizeof (ao_bank_cfg_t); i++)
730 		wrmsr(ao_bank_cfgs[i].bank_status, 0ULL);
731 
732 	wrmsr(IA32_MSR_MCG_STATUS, 0ULL);
733 	membar_producer();
734 
735 	setcr4(getcr4() | CR4_MCE); /* enable #mc exceptions */
736 }
737 
738 /*ARGSUSED*/
739 void
740 ao_mca_post_init(void *data)
741 {
742 	const struct ao_smi_disable *asd;
743 	id_t id;
744 
745 	smbios_system_t sy;
746 	smbios_bios_t sb;
747 	smbios_info_t si;
748 
749 	/*
750 	 * Fetch the System and BIOS vendor strings from SMBIOS and see if they
751 	 * match a value in our table.  If so, disable SMI error polling.  This
752 	 * is grotesque and should be replaced by self-describing vendor-
753 	 * specific SMBIOS data or a specification enhancement instead.
754 	 */
755 	if (ao_mca_smi_disable && ksmbios != NULL &&
756 	    smbios_info_bios(ksmbios, &sb) != SMB_ERR &&
757 	    (id = smbios_info_system(ksmbios, &sy)) != SMB_ERR &&
758 	    smbios_info_common(ksmbios, id, &si) != SMB_ERR) {
759 
760 		for (asd = ao_smi_disable; asd->asd_sys_vendor != NULL; asd++) {
761 			if (strncmp(asd->asd_sys_vendor, si.smbi_manufacturer,
762 			    strlen(asd->asd_sys_vendor)) != 0 ||
763 			    strncmp(asd->asd_bios_vendor, sb.smbb_vendor,
764 			    strlen(asd->asd_bios_vendor)) != 0)
765 				continue;
766 
767 			cmn_err(CE_CONT, "?SMI polling disabled in favor of "
768 			    "Solaris Fault Management for AMD Processors");
769 
770 			outl(asd->asd_port, asd->asd_code);
771 			break;
772 		}
773 	}
774 
775 	ao_mca_poll_start();
776 }
777 
778 /*
779  * Called after a CPU has been marked with CPU_FAULTED.  Not called on the
780  * faulted CPU.  cpu_lock is held.
781  */
782 /*ARGSUSED*/
783 void
784 ao_faulted_enter(void *data)
785 {
786 	/*
787 	 * Nothing to do here.  We'd like to turn off the faulted CPU's
788 	 * correctable error detectors, but that can only be done by the
789 	 * faulted CPU itself.  cpu_get_state() will now return P_FAULTED,
790 	 * allowing the poller to skip this CPU until it is re-enabled.
791 	 */
792 }
793 
794 /*
795  * Called after the CPU_FAULTED bit has been cleared from a previously-faulted
796  * CPU.  Not called on the faulted CPU.  cpu_lock is held.
797  */
798 void
799 ao_faulted_exit(void *data)
800 {
801 	ao_data_t *ao = data;
802 
803 	/*
804 	 * We'd like to clear the faulted CPU's MCi_STATUS registers so as to
805 	 * avoid generating ereports for errors which occurred while the CPU was
806 	 * officially faulted.  Unfortunately, those registers can only be
807 	 * cleared by the CPU itself, so we can't do it here.
808 	 *
809 	 * We're going to set the UNFAULTING bit on the formerly-faulted CPU's
810 	 * MCA state.  This will tell the poller that the MCi_STATUS registers
811 	 * can't yet be trusted.  The poller, which is the first thing we
812 	 * control that'll execute on that CPU, will clear the registers, and
813 	 * will then clear the bit.
814 	 */
815 
816 	ao->ao_mca.ao_mca_flags |= AO_MCA_F_UNFAULTING;
817 }
818