xref: /titanic_41/usr/src/uts/sun4u/cpu/us3_common.c (revision 0d282d1376eb7ba06504448622a6d65726e4bd3e)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/types.h>
29 #include <sys/systm.h>
30 #include <sys/ddi.h>
31 #include <sys/sysmacros.h>
32 #include <sys/archsystm.h>
33 #include <sys/vmsystm.h>
34 #include <sys/machparam.h>
35 #include <sys/machsystm.h>
36 #include <sys/machthread.h>
37 #include <sys/cpu.h>
38 #include <sys/cmp.h>
39 #include <sys/elf_SPARC.h>
40 #include <vm/vm_dep.h>
41 #include <vm/hat_sfmmu.h>
42 #include <vm/seg_kpm.h>
43 #include <sys/cpuvar.h>
44 #include <sys/cheetahregs.h>
45 #include <sys/us3_module.h>
46 #include <sys/async.h>
47 #include <sys/cmn_err.h>
48 #include <sys/debug.h>
49 #include <sys/dditypes.h>
50 #include <sys/prom_debug.h>
51 #include <sys/prom_plat.h>
52 #include <sys/cpu_module.h>
53 #include <sys/sysmacros.h>
54 #include <sys/intreg.h>
55 #include <sys/clock.h>
56 #include <sys/platform_module.h>
57 #include <sys/machtrap.h>
58 #include <sys/ontrap.h>
59 #include <sys/panic.h>
60 #include <sys/memlist.h>
61 #include <sys/bootconf.h>
62 #include <sys/ivintr.h>
63 #include <sys/atomic.h>
64 #include <sys/taskq.h>
65 #include <sys/note.h>
66 #include <sys/ndifm.h>
67 #include <sys/ddifm.h>
68 #include <sys/fm/protocol.h>
69 #include <sys/fm/util.h>
70 #include <sys/fm/cpu/UltraSPARC-III.h>
71 #include <sys/fpras_impl.h>
72 #include <sys/dtrace.h>
73 #include <sys/watchpoint.h>
74 #include <sys/plat_ecc_unum.h>
75 #include <sys/cyclic.h>
76 #include <sys/errorq.h>
77 #include <sys/errclassify.h>
78 
79 #ifdef	CHEETAHPLUS_ERRATUM_25
80 #include <sys/xc_impl.h>
81 #endif	/* CHEETAHPLUS_ERRATUM_25 */
82 
83 /*
84  * Note that 'Cheetah PRM' refers to:
85  *   SPARC V9 JPS1 Implementation Supplement: Sun UltraSPARC-III
86  */
87 
88 /*
89  * Per CPU pointers to physical address of TL>0 logout data areas.
90  * These pointers have to be in the kernel nucleus to avoid MMU
91  * misses.
92  */
93 uint64_t ch_err_tl1_paddrs[NCPU];
94 
95 /*
96  * One statically allocated structure to use during startup/DR
97  * to prevent unnecessary panics.
98  */
99 ch_err_tl1_data_t ch_err_tl1_data;
100 
101 /*
102  * Per CPU pending error at TL>0, used by level15 softint handler
103  */
104 uchar_t ch_err_tl1_pending[NCPU];
105 
106 /*
107  * For deferred CE re-enable after trap.
108  */
109 taskq_t		*ch_check_ce_tq;
110 
111 /*
112  * Internal functions.
113  */
114 static int cpu_async_log_err(void *flt, errorq_elem_t *eqep);
115 static void cpu_log_diag_info(ch_async_flt_t *ch_flt);
116 static void cpu_queue_one_event(ch_async_flt_t *ch_flt, char *reason,
117     ecc_type_to_info_t *eccp, ch_diag_data_t *cdp);
118 static int cpu_flt_in_memory_one_event(ch_async_flt_t *ch_flt,
119     uint64_t t_afsr_bit);
120 static int clear_ecc(struct async_flt *ecc);
121 #if defined(CPU_IMP_ECACHE_ASSOC)
122 static int cpu_ecache_line_valid(ch_async_flt_t *ch_flt);
123 #endif
124 static int cpu_ecache_set_size(struct cpu *cp);
125 static int cpu_ectag_line_invalid(int cachesize, uint64_t tag);
126 static int cpu_ectag_pa_to_subblk(int cachesize, uint64_t subaddr);
127 static uint64_t cpu_ectag_to_pa(int setsize, uint64_t tag);
128 static int cpu_ectag_pa_to_subblk_state(int cachesize,
129 				uint64_t subaddr, uint64_t tag);
130 static void cpu_flush_ecache_line(ch_async_flt_t *ch_flt);
131 static int afsr_to_afar_status(uint64_t afsr, uint64_t afsr_bit);
132 static int afsr_to_esynd_status(uint64_t afsr, uint64_t afsr_bit);
133 static int afsr_to_msynd_status(uint64_t afsr, uint64_t afsr_bit);
134 static int afsr_to_synd_status(uint_t cpuid, uint64_t afsr, uint64_t afsr_bit);
135 static int synd_to_synd_code(int synd_status, ushort_t synd, uint64_t afsr_bit);
136 static int cpu_get_mem_unum_synd(int synd_code, struct async_flt *, char *buf);
137 static void cpu_uninit_ecache_scrub_dr(struct cpu *cp);
138 static void cpu_scrubphys(struct async_flt *aflt);
139 static void cpu_payload_add_aflt(struct async_flt *, nvlist_t *, nvlist_t *,
140     int *, int *);
141 static void cpu_payload_add_ecache(struct async_flt *, nvlist_t *);
142 static void cpu_ereport_init(struct async_flt *aflt);
143 static int cpu_check_secondary_errors(ch_async_flt_t *, uint64_t, uint64_t);
144 static uint8_t cpu_flt_bit_to_plat_error(struct async_flt *aflt);
145 static void cpu_log_fast_ecc_error(caddr_t tpc, int priv, int tl, uint64_t ceen,
146     uint64_t nceen, ch_cpu_logout_t *clop);
147 static int cpu_ce_delayed_ec_logout(uint64_t);
148 static int cpu_matching_ecache_line(uint64_t, void *, int, int *);
149 static int cpu_error_is_ecache_data(int, uint64_t);
150 static void cpu_fmri_cpu_set(nvlist_t *, int);
151 static int cpu_error_to_resource_type(struct async_flt *aflt);
152 
153 #ifdef	CHEETAHPLUS_ERRATUM_25
154 static int mondo_recover_proc(uint16_t, int);
155 static void cheetah_nudge_init(void);
156 static void cheetah_nudge_onln(void *arg, cpu_t *cpu, cyc_handler_t *hdlr,
157     cyc_time_t *when);
158 static void cheetah_nudge_buddy(void);
159 #endif	/* CHEETAHPLUS_ERRATUM_25 */
160 
161 #if defined(CPU_IMP_L1_CACHE_PARITY)
162 static void cpu_dcache_parity_info(ch_async_flt_t *ch_flt);
163 static void cpu_dcache_parity_check(ch_async_flt_t *ch_flt, int index);
164 static void cpu_record_dc_data_parity(ch_async_flt_t *ch_flt,
165     ch_dc_data_t *dest_dcp, ch_dc_data_t *src_dcp, int way, int word);
166 static void cpu_icache_parity_info(ch_async_flt_t *ch_flt);
167 static void cpu_icache_parity_check(ch_async_flt_t *ch_flt, int index);
168 static void cpu_pcache_parity_info(ch_async_flt_t *ch_flt);
169 static void cpu_pcache_parity_check(ch_async_flt_t *ch_flt, int index);
170 static void cpu_payload_add_dcache(struct async_flt *, nvlist_t *);
171 static void cpu_payload_add_icache(struct async_flt *, nvlist_t *);
172 #endif	/* CPU_IMP_L1_CACHE_PARITY */
173 
174 int (*p2get_mem_info)(int synd_code, uint64_t paddr,
175     uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
176     int *segsp, int *banksp, int *mcidp);
177 
178 /*
179  * This table is used to determine which bit(s) is(are) bad when an ECC
180  * error occurs.  The array is indexed by an 9-bit syndrome.  The entries
181  * of this array have the following semantics:
182  *
183  *      00-127  The number of the bad bit, when only one bit is bad.
184  *      128     ECC bit C0 is bad.
185  *      129     ECC bit C1 is bad.
186  *      130     ECC bit C2 is bad.
187  *      131     ECC bit C3 is bad.
188  *      132     ECC bit C4 is bad.
189  *      133     ECC bit C5 is bad.
190  *      134     ECC bit C6 is bad.
191  *      135     ECC bit C7 is bad.
192  *      136     ECC bit C8 is bad.
193  *	137-143 reserved for Mtag Data and ECC.
194  *      144(M2) Two bits are bad within a nibble.
195  *      145(M3) Three bits are bad within a nibble.
196  *      146(M3) Four bits are bad within a nibble.
197  *      147(M)  Multiple bits (5 or more) are bad.
198  *      148     NO bits are bad.
199  * Based on "Cheetah Programmer's Reference Manual" rev 1.1, Tables 11-4,11-5.
200  */
201 
202 #define	C0	128
203 #define	C1	129
204 #define	C2	130
205 #define	C3	131
206 #define	C4	132
207 #define	C5	133
208 #define	C6	134
209 #define	C7	135
210 #define	C8	136
211 #define	MT0	137	/* Mtag Data bit 0 */
212 #define	MT1	138
213 #define	MT2	139
214 #define	MTC0	140	/* Mtag Check bit 0 */
215 #define	MTC1	141
216 #define	MTC2	142
217 #define	MTC3	143
218 #define	M2	144
219 #define	M3	145
220 #define	M4	146
221 #define	M	147
222 #define	NA	148
223 #if defined(JALAPENO) || defined(SERRANO)
224 #define	S003	149	/* Syndrome 0x003 => likely from CPU/EDU:ST/FRU/BP */
225 #define	S003MEM	150	/* Syndrome 0x003 => likely from WDU/WBP */
226 #define	SLAST	S003MEM	/* last special syndrome */
227 #else /* JALAPENO || SERRANO */
228 #define	S003	149	/* Syndrome 0x003 => likely from EDU:ST */
229 #define	S071	150	/* Syndrome 0x071 => likely from WDU/CPU */
230 #define	S11C	151	/* Syndrome 0x11c => likely from BERR/DBERR */
231 #define	SLAST	S11C	/* last special syndrome */
232 #endif /* JALAPENO || SERRANO */
233 #if defined(JALAPENO) || defined(SERRANO)
234 #define	BPAR0	152	/* syndrom 152 through 167 for bus parity */
235 #define	BPAR15	167
236 #endif	/* JALAPENO || SERRANO */
237 
238 static uint8_t ecc_syndrome_tab[] =
239 {
240 NA,  C0,  C1, S003, C2,  M2,  M3,  47,  C3,  M2,  M2,  53,  M2,  41,  29,   M,
241 C4,   M,   M,  50,  M2,  38,  25,  M2,  M2,  33,  24,  M2,  11,   M,  M2,  16,
242 C5,   M,   M,  46,  M2,  37,  19,  M2,   M,  31,  32,   M,   7,  M2,  M2,  10,
243 M2,  40,  13,  M2,  59,   M,  M2,  66,   M,  M2,  M2,   0,  M2,  67,  71,   M,
244 C6,   M,   M,  43,   M,  36,  18,   M,  M2,  49,  15,   M,  63,  M2,  M2,   6,
245 M2,  44,  28,  M2,   M,  M2,  M2,  52,  68,  M2,  M2,  62,  M2,  M3,  M3,  M4,
246 M2,  26, 106,  M2,  64,   M,  M2,   2, 120,   M,  M2,  M3,   M,  M3,  M3,  M4,
247 #if defined(JALAPENO) || defined(SERRANO)
248 116, M2,  M2,  M3,  M2,  M3,   M,  M4,  M2,  58,  54,  M2,   M,  M4,  M4,  M3,
249 #else	/* JALAPENO || SERRANO */
250 116, S071, M2,  M3,  M2,  M3,   M,  M4,  M2,  58,  54,  M2,   M,  M4,  M4,  M3,
251 #endif	/* JALAPENO || SERRANO */
252 C7,  M2,   M,  42,   M,  35,  17,  M2,   M,  45,  14,  M2,  21,  M2,  M2,   5,
253 M,   27,   M,   M,  99,   M,   M,   3, 114,  M2,  M2,  20,  M2,  M3,  M3,   M,
254 M2,  23, 113,  M2, 112,  M2,   M,  51,  95,   M,  M2,  M3,  M2,  M3,  M3,  M2,
255 103,  M,  M2,  M3,  M2,  M3,  M3,  M4,  M2,  48,   M,   M,  73,  M2,   M,  M3,
256 M2,  22, 110,  M2, 109,  M2,   M,   9, 108,  M2,   M,  M3,  M2,  M3,  M3,   M,
257 102, M2,   M,   M,  M2,  M3,  M3,   M,  M2,  M3,  M3,  M2,   M,  M4,   M,  M3,
258 98,   M,  M2,  M3,  M2,   M,  M3,  M4,  M2,  M3,  M3,  M4,  M3,   M,   M,   M,
259 M2,  M3,  M3,   M,  M3,   M,   M,   M,  56,  M4,   M,  M3,  M4,   M,   M,   M,
260 C8,   M,  M2,  39,   M,  34, 105,  M2,   M,  30, 104,   M, 101,   M,   M,   4,
261 #if defined(JALAPENO) || defined(SERRANO)
262 M,    M, 100,   M,  83,   M,  M2,  12,  87,   M,   M,  57,  M2,   M,  M3,   M,
263 #else	/* JALAPENO || SERRANO */
264 M,    M, 100,   M,  83,   M,  M2,  12,  87,   M,   M,  57, S11C,  M,  M3,   M,
265 #endif	/* JALAPENO || SERRANO */
266 M2,  97,  82,  M2,  78,  M2,  M2,   1,  96,   M,   M,   M,   M,   M,  M3,  M2,
267 94,   M,  M2,  M3,  M2,   M,  M3,   M,  M2,   M,  79,   M,  69,   M,  M4,   M,
268 M2,  93,  92,   M,  91,   M,  M2,   8,  90,  M2,  M2,   M,   M,   M,   M,  M4,
269 89,   M,   M,  M3,  M2,  M3,  M3,   M,   M,   M,  M3,  M2,  M3,  M2,   M,  M3,
270 86,   M,  M2,  M3,  M2,   M,  M3,   M,  M2,   M,  M3,   M,  M3,   M,   M,  M3,
271 M,    M,  M3,  M2,  M3,  M2,  M4,   M,  60,   M,  M2,  M3,  M4,   M,   M,  M2,
272 M2,  88,  85,  M2,  84,   M,  M2,  55,  81,  M2,  M2,  M3,  M2,  M3,  M3,  M4,
273 77,   M,   M,   M,  M2,  M3,   M,   M,  M2,  M3,  M3,  M4,  M3,  M2,   M,   M,
274 74,   M,  M2,  M3,   M,   M,  M3,   M,   M,   M,  M3,   M,  M3,   M,  M4,  M3,
275 M2,  70, 107,  M4,  65,  M2,  M2,   M, 127,   M,   M,   M,  M2,  M3,  M3,   M,
276 80,  M2,  M2,  72,   M, 119, 118,   M,  M2, 126,  76,   M, 125,   M,  M4,  M3,
277 M2, 115, 124,   M,  75,   M,   M,  M3,  61,   M,  M4,   M,  M4,   M,   M,   M,
278 M,  123, 122,  M4, 121,  M4,   M,  M3, 117,  M2,  M2,  M3,  M4,  M3,   M,   M,
279 111,  M,   M,   M,  M4,  M3,  M3,   M,   M,   M,  M3,   M,  M3,  M2,   M,   M
280 };
281 
282 #define	ESYND_TBL_SIZE	(sizeof (ecc_syndrome_tab) / sizeof (uint8_t))
283 
284 #if !(defined(JALAPENO) || defined(SERRANO))
285 /*
286  * This table is used to determine which bit(s) is(are) bad when a Mtag
287  * error occurs.  The array is indexed by an 4-bit ECC syndrome. The entries
288  * of this array have the following semantics:
289  *
290  *      -1	Invalid mtag syndrome.
291  *      137     Mtag Data 0 is bad.
292  *      138     Mtag Data 1 is bad.
293  *      139     Mtag Data 2 is bad.
294  *      140     Mtag ECC 0 is bad.
295  *      141     Mtag ECC 1 is bad.
296  *      142     Mtag ECC 2 is bad.
297  *      143     Mtag ECC 3 is bad.
298  * Based on "Cheetah Programmer's Reference Manual" rev 1.1, Tables 11-6.
299  */
300 short mtag_syndrome_tab[] =
301 {
302 NA, MTC0, MTC1, M2, MTC2, M2, M2, MT0, MTC3, M2, M2,  MT1, M2, MT2, M2, M2
303 };
304 
305 #define	MSYND_TBL_SIZE	(sizeof (mtag_syndrome_tab) / sizeof (short))
306 
307 #else /* !(JALAPENO || SERRANO) */
308 
309 #define	BSYND_TBL_SIZE	16
310 
311 #endif /* !(JALAPENO || SERRANO) */
312 
313 /*
314  * Types returned from cpu_error_to_resource_type()
315  */
316 #define	ERRTYPE_UNKNOWN		0
317 #define	ERRTYPE_CPU		1
318 #define	ERRTYPE_MEMORY		2
319 #define	ERRTYPE_ECACHE_DATA	3
320 
321 /*
322  * CE initial classification and subsequent action lookup table
323  */
324 static ce_dispact_t ce_disp_table[CE_INITDISPTBL_SIZE];
325 static int ce_disp_inited;
326 
327 /*
328  * Set to disable leaky and partner check for memory correctables
329  */
330 int ce_xdiag_off;
331 
332 /*
333  * The following are not incremented atomically so are indicative only
334  */
335 static int ce_xdiag_drops;
336 static int ce_xdiag_lkydrops;
337 static int ce_xdiag_ptnrdrops;
338 static int ce_xdiag_bad;
339 
340 /*
341  * CE leaky check callback structure
342  */
343 typedef struct {
344 	struct async_flt *lkycb_aflt;
345 	errorq_t *lkycb_eqp;
346 	errorq_elem_t *lkycb_eqep;
347 } ce_lkychk_cb_t;
348 
349 /*
350  * defines for various ecache_flush_flag's
351  */
352 #define	ECACHE_FLUSH_LINE	1
353 #define	ECACHE_FLUSH_ALL	2
354 
355 /*
356  * STICK sync
357  */
358 #define	STICK_ITERATION 10
359 #define	MAX_TSKEW	1
360 #define	EV_A_START	0
361 #define	EV_A_END	1
362 #define	EV_B_START	2
363 #define	EV_B_END	3
364 #define	EVENTS		4
365 
366 static int64_t stick_iter = STICK_ITERATION;
367 static int64_t stick_tsk = MAX_TSKEW;
368 
369 typedef enum {
370 	EVENT_NULL = 0,
371 	SLAVE_START,
372 	SLAVE_CONT,
373 	MASTER_START
374 } event_cmd_t;
375 
376 static volatile event_cmd_t stick_sync_cmd = EVENT_NULL;
377 static int64_t timestamp[EVENTS];
378 static volatile int slave_done;
379 
380 #ifdef DEBUG
381 #define	DSYNC_ATTEMPTS 64
382 typedef struct {
383 	int64_t	skew_val[DSYNC_ATTEMPTS];
384 } ss_t;
385 
386 ss_t stick_sync_stats[NCPU];
387 #endif /* DEBUG */
388 
389 uint_t cpu_impl_dual_pgsz = 0;
390 #if defined(CPU_IMP_DUAL_PAGESIZE)
391 uint_t disable_dual_pgsz = 0;
392 #endif	/* CPU_IMP_DUAL_PAGESIZE */
393 
394 /*
395  * Save the cache bootup state for use when internal
396  * caches are to be re-enabled after an error occurs.
397  */
398 uint64_t cache_boot_state;
399 
400 /*
401  * PA[22:0] represent Displacement in Safari configuration space.
402  */
403 uint_t	root_phys_addr_lo_mask = 0x7fffffu;
404 
405 bus_config_eclk_t bus_config_eclk[] = {
406 #if defined(JALAPENO) || defined(SERRANO)
407 	{JBUS_CONFIG_ECLK_1_DIV, JBUS_CONFIG_ECLK_1},
408 	{JBUS_CONFIG_ECLK_2_DIV, JBUS_CONFIG_ECLK_2},
409 	{JBUS_CONFIG_ECLK_32_DIV, JBUS_CONFIG_ECLK_32},
410 #else /* JALAPENO || SERRANO */
411 	{SAFARI_CONFIG_ECLK_1_DIV, SAFARI_CONFIG_ECLK_1},
412 	{SAFARI_CONFIG_ECLK_2_DIV, SAFARI_CONFIG_ECLK_2},
413 	{SAFARI_CONFIG_ECLK_32_DIV, SAFARI_CONFIG_ECLK_32},
414 #endif /* JALAPENO || SERRANO */
415 	{0, 0}
416 };
417 
418 /*
419  * Interval for deferred CEEN reenable
420  */
421 int cpu_ceen_delay_secs = CPU_CEEN_DELAY_SECS;
422 
423 /*
424  * set in /etc/system to control logging of user BERR/TO's
425  */
426 int cpu_berr_to_verbose = 0;
427 
428 /*
429  * set to 0 in /etc/system to defer CEEN reenable for all CEs
430  */
431 uint64_t cpu_ce_not_deferred = CPU_CE_NOT_DEFERRED;
432 uint64_t cpu_ce_not_deferred_ext = CPU_CE_NOT_DEFERRED_EXT;
433 
434 /*
435  * Set of all offline cpus
436  */
437 cpuset_t cpu_offline_set;
438 
439 static void cpu_delayed_check_ce_errors(void *);
440 static void cpu_check_ce_errors(void *);
441 void cpu_error_ecache_flush(ch_async_flt_t *);
442 static int cpu_error_ecache_flush_required(ch_async_flt_t *);
443 static void cpu_log_and_clear_ce(ch_async_flt_t *);
444 void cpu_ce_detected(ch_cpu_errors_t *, int);
445 
446 /*
447  * CE Leaky check timeout in microseconds.  This is chosen to be twice the
448  * memory refresh interval of current DIMMs (64ms).  After initial fix that
449  * gives at least one full refresh cycle in which the cell can leak
450  * (whereafter further refreshes simply reinforce any incorrect bit value).
451  */
452 clock_t cpu_ce_lkychk_timeout_usec = 128000;
453 
454 /*
455  * CE partner check partner caching period in seconds
456  */
457 int cpu_ce_ptnr_cachetime_sec = 60;
458 
459 /*
460  * Sets trap table entry ttentry by overwriting eight instructions from ttlabel
461  */
462 #define	CH_SET_TRAP(ttentry, ttlabel)			\
463 		bcopy((const void *)&ttlabel, &ttentry, 32);		\
464 		flush_instr_mem((caddr_t)&ttentry, 32);
465 
466 static int min_ecache_size;
467 static uint_t priv_hcl_1;
468 static uint_t priv_hcl_2;
469 static uint_t priv_hcl_4;
470 static uint_t priv_hcl_8;
471 
472 void
473 cpu_setup(void)
474 {
475 	extern int at_flags;
476 	extern int disable_delay_tlb_flush, delay_tlb_flush;
477 	extern int cpc_has_overflow_intr;
478 	extern int disable_text_largepages;
479 	extern int use_text_pgsz4m;
480 
481 	/*
482 	 * Setup chip-specific trap handlers.
483 	 */
484 	cpu_init_trap();
485 
486 	cache |= (CACHE_VAC | CACHE_PTAG | CACHE_IOCOHERENT);
487 
488 	at_flags = EF_SPARC_32PLUS | EF_SPARC_SUN_US1 | EF_SPARC_SUN_US3;
489 
490 	/*
491 	 * save the cache bootup state.
492 	 */
493 	cache_boot_state = get_dcu() & DCU_CACHE;
494 
495 	/*
496 	 * Due to the number of entries in the fully-associative tlb
497 	 * this may have to be tuned lower than in spitfire.
498 	 */
499 	pp_slots = MIN(8, MAXPP_SLOTS);
500 
501 	/*
502 	 * Block stores do not invalidate all pages of the d$, pagecopy
503 	 * et. al. need virtual translations with virtual coloring taken
504 	 * into consideration.  prefetch/ldd will pollute the d$ on the
505 	 * load side.
506 	 */
507 	pp_consistent_coloring = PPAGE_STORE_VCOLORING | PPAGE_LOADS_POLLUTE;
508 
509 	if (use_page_coloring) {
510 		do_pg_coloring = 1;
511 		if (use_virtual_coloring)
512 			do_virtual_coloring = 1;
513 	}
514 
515 	isa_list =
516 	    "sparcv9+vis2 sparcv9+vis sparcv9 "
517 	    "sparcv8plus+vis2 sparcv8plus+vis sparcv8plus "
518 	    "sparcv8 sparcv8-fsmuld sparcv7 sparc";
519 
520 	/*
521 	 * On Panther-based machines, this should
522 	 * also include AV_SPARC_POPC too
523 	 */
524 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2;
525 
526 	/*
527 	 * On cheetah, there's no hole in the virtual address space
528 	 */
529 	hole_start = hole_end = 0;
530 
531 	/*
532 	 * The kpm mapping window.
533 	 * kpm_size:
534 	 *	The size of a single kpm range.
535 	 *	The overall size will be: kpm_size * vac_colors.
536 	 * kpm_vbase:
537 	 *	The virtual start address of the kpm range within the kernel
538 	 *	virtual address space. kpm_vbase has to be kpm_size aligned.
539 	 */
540 	kpm_size = (size_t)(8ull * 1024 * 1024 * 1024 * 1024); /* 8TB */
541 	kpm_size_shift = 43;
542 	kpm_vbase = (caddr_t)0x8000000000000000ull; /* 8EB */
543 	kpm_smallpages = 1;
544 
545 	/*
546 	 * The traptrace code uses either %tick or %stick for
547 	 * timestamping.  We have %stick so we can use it.
548 	 */
549 	traptrace_use_stick = 1;
550 
551 	/*
552 	 * Cheetah has a performance counter overflow interrupt
553 	 */
554 	cpc_has_overflow_intr = 1;
555 
556 	/*
557 	 * Use cheetah flush-all support
558 	 */
559 	if (!disable_delay_tlb_flush)
560 		delay_tlb_flush = 1;
561 
562 #if defined(CPU_IMP_DUAL_PAGESIZE)
563 	/*
564 	 * Use Cheetah+ and later dual page size support.
565 	 */
566 	if (!disable_dual_pgsz) {
567 		cpu_impl_dual_pgsz = 1;
568 	}
569 #endif	/* CPU_IMP_DUAL_PAGESIZE */
570 
571 	/*
572 	 * Declare that this architecture/cpu combination does fpRAS.
573 	 */
574 	fpras_implemented = 1;
575 
576 	/*
577 	 * Enable 4M pages to be used for mapping user text by default.  Don't
578 	 * use large pages for initialized data segments since we may not know
579 	 * at exec() time what should be the preferred large page size for DTLB
580 	 * programming.
581 	 */
582 	use_text_pgsz4m = 1;
583 	disable_text_largepages = (1 << TTE64K) | (1 << TTE512K) |
584 	    (1 << TTE32M) | (1 << TTE256M);
585 
586 	/*
587 	 * Setup CE lookup table
588 	 */
589 	CE_INITDISPTBL_POPULATE(ce_disp_table);
590 	ce_disp_inited = 1;
591 }
592 
593 /*
594  * Called by setcpudelay
595  */
596 void
597 cpu_init_tick_freq(void)
598 {
599 	/*
600 	 * For UltraSPARC III and beyond we want to use the
601 	 * system clock rate as the basis for low level timing,
602 	 * due to support of mixed speed CPUs and power managment.
603 	 */
604 	if (system_clock_freq == 0)
605 		cmn_err(CE_PANIC, "setcpudelay: invalid system_clock_freq");
606 
607 	sys_tick_freq = system_clock_freq;
608 }
609 
610 #ifdef CHEETAHPLUS_ERRATUM_25
611 /*
612  * Tunables
613  */
614 int cheetah_bpe_off = 0;
615 int cheetah_sendmondo_recover = 1;
616 int cheetah_sendmondo_fullscan = 0;
617 int cheetah_sendmondo_recover_delay = 5;
618 
619 #define	CHEETAH_LIVELOCK_MIN_DELAY	1
620 
621 /*
622  * Recovery Statistics
623  */
624 typedef struct cheetah_livelock_entry	{
625 	int cpuid;		/* fallen cpu */
626 	int buddy;		/* cpu that ran recovery */
627 	clock_t lbolt;		/* when recovery started */
628 	hrtime_t recovery_time;	/* time spent in recovery */
629 } cheetah_livelock_entry_t;
630 
631 #define	CHEETAH_LIVELOCK_NENTRY	32
632 
633 cheetah_livelock_entry_t cheetah_livelock_hist[CHEETAH_LIVELOCK_NENTRY];
634 int cheetah_livelock_entry_nxt;
635 
636 #define	CHEETAH_LIVELOCK_ENTRY_NEXT(statp)	{			\
637 	statp = cheetah_livelock_hist + cheetah_livelock_entry_nxt;	\
638 	if (++cheetah_livelock_entry_nxt >= CHEETAH_LIVELOCK_NENTRY) {	\
639 		cheetah_livelock_entry_nxt = 0;				\
640 	}								\
641 }
642 
643 #define	CHEETAH_LIVELOCK_ENTRY_SET(statp, item, val)	statp->item = val
644 
645 struct {
646 	hrtime_t hrt;		/* maximum recovery time */
647 	int recovery;		/* recovered */
648 	int full_claimed;	/* maximum pages claimed in full recovery */
649 	int proc_entry;		/* attempted to claim TSB */
650 	int proc_tsb_scan;	/* tsb scanned */
651 	int proc_tsb_partscan;	/* tsb partially scanned */
652 	int proc_tsb_fullscan;	/* whole tsb scanned */
653 	int proc_claimed;	/* maximum pages claimed in tsb scan */
654 	int proc_user;		/* user thread */
655 	int proc_kernel;	/* kernel thread */
656 	int proc_onflt;		/* bad stack */
657 	int proc_cpu;		/* null cpu */
658 	int proc_thread;	/* null thread */
659 	int proc_proc;		/* null proc */
660 	int proc_as;		/* null as */
661 	int proc_hat;		/* null hat */
662 	int proc_hat_inval;	/* hat contents don't make sense */
663 	int proc_hat_busy;	/* hat is changing TSBs */
664 	int proc_tsb_reloc;	/* TSB skipped because being relocated */
665 	int proc_cnum_bad;	/* cnum out of range */
666 	int proc_cnum;		/* last cnum processed */
667 	tte_t proc_tte;		/* last tte processed */
668 } cheetah_livelock_stat;
669 
670 #define	CHEETAH_LIVELOCK_STAT(item)	cheetah_livelock_stat.item++
671 
672 #define	CHEETAH_LIVELOCK_STATSET(item, value)		\
673 	cheetah_livelock_stat.item = value
674 
675 #define	CHEETAH_LIVELOCK_MAXSTAT(item, value)	{	\
676 	if (value > cheetah_livelock_stat.item)		\
677 		cheetah_livelock_stat.item = value;	\
678 }
679 
680 /*
681  * Attempt to recover a cpu by claiming every cache line as saved
682  * in the TSB that the non-responsive cpu is using. Since we can't
683  * grab any adaptive lock, this is at best an attempt to do so. Because
684  * we don't grab any locks, we must operate under the protection of
685  * on_fault().
686  *
687  * Return 1 if cpuid could be recovered, 0 if failed.
688  */
689 int
690 mondo_recover_proc(uint16_t cpuid, int bn)
691 {
692 	label_t ljb;
693 	cpu_t *cp;
694 	kthread_t *t;
695 	proc_t *p;
696 	struct as *as;
697 	struct hat *hat;
698 	uint_t  cnum;
699 	struct tsb_info *tsbinfop;
700 	struct tsbe *tsbep;
701 	caddr_t tsbp;
702 	caddr_t end_tsbp;
703 	uint64_t paddr;
704 	uint64_t idsr;
705 	u_longlong_t pahi, palo;
706 	int pages_claimed = 0;
707 	tte_t tsbe_tte;
708 	int tried_kernel_tsb = 0;
709 	mmu_ctx_t *mmu_ctxp;
710 
711 	CHEETAH_LIVELOCK_STAT(proc_entry);
712 
713 	if (on_fault(&ljb)) {
714 		CHEETAH_LIVELOCK_STAT(proc_onflt);
715 		goto badstruct;
716 	}
717 
718 	if ((cp = cpu[cpuid]) == NULL) {
719 		CHEETAH_LIVELOCK_STAT(proc_cpu);
720 		goto badstruct;
721 	}
722 
723 	if ((t = cp->cpu_thread) == NULL) {
724 		CHEETAH_LIVELOCK_STAT(proc_thread);
725 		goto badstruct;
726 	}
727 
728 	if ((p = ttoproc(t)) == NULL) {
729 		CHEETAH_LIVELOCK_STAT(proc_proc);
730 		goto badstruct;
731 	}
732 
733 	if ((as = p->p_as) == NULL) {
734 		CHEETAH_LIVELOCK_STAT(proc_as);
735 		goto badstruct;
736 	}
737 
738 	if ((hat = as->a_hat) == NULL) {
739 		CHEETAH_LIVELOCK_STAT(proc_hat);
740 		goto badstruct;
741 	}
742 
743 	if (hat != ksfmmup) {
744 		CHEETAH_LIVELOCK_STAT(proc_user);
745 		if (hat->sfmmu_flags & (HAT_BUSY | HAT_SWAPPED | HAT_SWAPIN)) {
746 			CHEETAH_LIVELOCK_STAT(proc_hat_busy);
747 			goto badstruct;
748 		}
749 		tsbinfop = hat->sfmmu_tsb;
750 		if (tsbinfop == NULL) {
751 			CHEETAH_LIVELOCK_STAT(proc_hat_inval);
752 			goto badstruct;
753 		}
754 		tsbp = tsbinfop->tsb_va;
755 		end_tsbp = tsbp + TSB_BYTES(tsbinfop->tsb_szc);
756 	} else {
757 		CHEETAH_LIVELOCK_STAT(proc_kernel);
758 		tsbinfop = NULL;
759 		tsbp = ktsb_base;
760 		end_tsbp = tsbp + TSB_BYTES(ktsb_sz);
761 	}
762 
763 	/* Verify as */
764 	if (hat->sfmmu_as != as) {
765 		CHEETAH_LIVELOCK_STAT(proc_hat_inval);
766 		goto badstruct;
767 	}
768 
769 	mmu_ctxp = CPU_MMU_CTXP(cp);
770 	ASSERT(mmu_ctxp);
771 	cnum = hat->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
772 	CHEETAH_LIVELOCK_STATSET(proc_cnum, cnum);
773 
774 	if ((cnum < 0) || (cnum == INVALID_CONTEXT) ||
775 	    (cnum >= mmu_ctxp->mmu_nctxs)) {
776 		CHEETAH_LIVELOCK_STAT(proc_cnum_bad);
777 		goto badstruct;
778 	}
779 
780 	do {
781 		CHEETAH_LIVELOCK_STAT(proc_tsb_scan);
782 
783 		/*
784 		 * Skip TSBs being relocated.  This is important because
785 		 * we want to avoid the following deadlock scenario:
786 		 *
787 		 * 1) when we came in we set ourselves to "in recover" state.
788 		 * 2) when we try to touch TSB being relocated the mapping
789 		 *    will be in the suspended state so we'll spin waiting
790 		 *    for it to be unlocked.
791 		 * 3) when the CPU that holds the TSB mapping locked tries to
792 		 *    unlock it it will send a xtrap which will fail to xcall
793 		 *    us or the CPU we're trying to recover, and will in turn
794 		 *    enter the mondo code.
795 		 * 4) since we are still spinning on the locked mapping
796 		 *    no further progress will be made and the system will
797 		 *    inevitably hard hang.
798 		 *
799 		 * A TSB not being relocated can't begin being relocated
800 		 * while we're accessing it because we check
801 		 * sendmondo_in_recover before relocating TSBs.
802 		 */
803 		if (hat != ksfmmup &&
804 		    (tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
805 			CHEETAH_LIVELOCK_STAT(proc_tsb_reloc);
806 			goto next_tsbinfo;
807 		}
808 
809 		for (tsbep = (struct tsbe *)tsbp;
810 		    tsbep < (struct tsbe *)end_tsbp; tsbep++) {
811 			tsbe_tte = tsbep->tte_data;
812 
813 			if (tsbe_tte.tte_val == 0) {
814 				/*
815 				 * Invalid tte
816 				 */
817 				continue;
818 			}
819 			if (tsbe_tte.tte_se) {
820 				/*
821 				 * Don't want device registers
822 				 */
823 				continue;
824 			}
825 			if (tsbe_tte.tte_cp == 0) {
826 				/*
827 				 * Must be cached in E$
828 				 */
829 				continue;
830 			}
831 			CHEETAH_LIVELOCK_STATSET(proc_tte, tsbe_tte);
832 			idsr = getidsr();
833 			if ((idsr & (IDSR_NACK_BIT(bn) |
834 			    IDSR_BUSY_BIT(bn))) == 0) {
835 				CHEETAH_LIVELOCK_STAT(proc_tsb_partscan);
836 				goto done;
837 			}
838 			pahi = tsbe_tte.tte_pahi;
839 			palo = tsbe_tte.tte_palo;
840 			paddr = (uint64_t)((pahi << 32) |
841 			    (palo << MMU_PAGESHIFT));
842 			claimlines(paddr, TTEBYTES(TTE_CSZ(&tsbe_tte)),
843 			    CH_ECACHE_SUBBLK_SIZE);
844 			if ((idsr & IDSR_BUSY_BIT(bn)) == 0) {
845 				shipit(cpuid, bn);
846 			}
847 			pages_claimed++;
848 		}
849 next_tsbinfo:
850 		if (tsbinfop != NULL)
851 			tsbinfop = tsbinfop->tsb_next;
852 		if (tsbinfop != NULL) {
853 			tsbp = tsbinfop->tsb_va;
854 			end_tsbp = tsbp + TSB_BYTES(tsbinfop->tsb_szc);
855 		} else if (tsbp == ktsb_base) {
856 			tried_kernel_tsb = 1;
857 		} else if (!tried_kernel_tsb) {
858 			tsbp = ktsb_base;
859 			end_tsbp = tsbp + TSB_BYTES(ktsb_sz);
860 			hat = ksfmmup;
861 			tsbinfop = NULL;
862 		}
863 	} while (tsbinfop != NULL ||
864 			((tsbp == ktsb_base) && !tried_kernel_tsb));
865 
866 	CHEETAH_LIVELOCK_STAT(proc_tsb_fullscan);
867 	CHEETAH_LIVELOCK_MAXSTAT(proc_claimed, pages_claimed);
868 	no_fault();
869 	idsr = getidsr();
870 	if ((idsr & (IDSR_NACK_BIT(bn) |
871 	    IDSR_BUSY_BIT(bn))) == 0) {
872 		return (1);
873 	} else {
874 		return (0);
875 	}
876 
877 done:
878 	no_fault();
879 	CHEETAH_LIVELOCK_MAXSTAT(proc_claimed, pages_claimed);
880 	return (1);
881 
882 badstruct:
883 	no_fault();
884 	return (0);
885 }
886 
887 /*
888  * Attempt to claim ownership, temporarily, of every cache line that a
889  * non-responsive cpu might be using.  This might kick that cpu out of
890  * this state.
891  *
892  * The return value indicates to the caller if we have exhausted all recovery
893  * techniques. If 1 is returned, it is useless to call this function again
894  * even for a different target CPU.
895  */
896 int
897 mondo_recover(uint16_t cpuid, int bn)
898 {
899 	struct memseg *seg;
900 	uint64_t begin_pa, end_pa, cur_pa;
901 	hrtime_t begin_hrt, end_hrt;
902 	int retval = 0;
903 	int pages_claimed = 0;
904 	cheetah_livelock_entry_t *histp;
905 	uint64_t idsr;
906 
907 	if (cas32(&sendmondo_in_recover, 0, 1) != 0) {
908 		/*
909 		 * Wait while recovery takes place
910 		 */
911 		while (sendmondo_in_recover) {
912 			drv_usecwait(1);
913 		}
914 		/*
915 		 * Assume we didn't claim the whole memory. If
916 		 * the target of this caller is not recovered,
917 		 * it will come back.
918 		 */
919 		return (retval);
920 	}
921 
922 	CHEETAH_LIVELOCK_ENTRY_NEXT(histp)
923 	CHEETAH_LIVELOCK_ENTRY_SET(histp, lbolt, lbolt);
924 	CHEETAH_LIVELOCK_ENTRY_SET(histp, cpuid, cpuid);
925 	CHEETAH_LIVELOCK_ENTRY_SET(histp, buddy, CPU->cpu_id);
926 
927 	begin_hrt = gethrtime_waitfree();
928 	/*
929 	 * First try to claim the lines in the TSB the target
930 	 * may have been using.
931 	 */
932 	if (mondo_recover_proc(cpuid, bn) == 1) {
933 		/*
934 		 * Didn't claim the whole memory
935 		 */
936 		goto done;
937 	}
938 
939 	/*
940 	 * We tried using the TSB. The target is still
941 	 * not recovered. Check if complete memory scan is
942 	 * enabled.
943 	 */
944 	if (cheetah_sendmondo_fullscan == 0) {
945 		/*
946 		 * Full memory scan is disabled.
947 		 */
948 		retval = 1;
949 		goto done;
950 	}
951 
952 	/*
953 	 * Try claiming the whole memory.
954 	 */
955 	for (seg = memsegs; seg; seg = seg->next) {
956 		begin_pa = (uint64_t)(seg->pages_base) << MMU_PAGESHIFT;
957 		end_pa = (uint64_t)(seg->pages_end) << MMU_PAGESHIFT;
958 		for (cur_pa = begin_pa; cur_pa < end_pa;
959 		    cur_pa += MMU_PAGESIZE) {
960 			idsr = getidsr();
961 			if ((idsr & (IDSR_NACK_BIT(bn) |
962 			    IDSR_BUSY_BIT(bn))) == 0) {
963 				/*
964 				 * Didn't claim all memory
965 				 */
966 				goto done;
967 			}
968 			claimlines(cur_pa, MMU_PAGESIZE,
969 			    CH_ECACHE_SUBBLK_SIZE);
970 			if ((idsr & IDSR_BUSY_BIT(bn)) == 0) {
971 				shipit(cpuid, bn);
972 			}
973 			pages_claimed++;
974 		}
975 	}
976 
977 	/*
978 	 * We did all we could.
979 	 */
980 	retval = 1;
981 
982 done:
983 	/*
984 	 * Update statistics
985 	 */
986 	end_hrt = gethrtime_waitfree();
987 	CHEETAH_LIVELOCK_STAT(recovery);
988 	CHEETAH_LIVELOCK_MAXSTAT(hrt, (end_hrt - begin_hrt));
989 	CHEETAH_LIVELOCK_MAXSTAT(full_claimed, pages_claimed);
990 	CHEETAH_LIVELOCK_ENTRY_SET(histp, recovery_time, \
991 	    (end_hrt -  begin_hrt));
992 
993 	while (cas32(&sendmondo_in_recover, 1, 0) != 1);
994 
995 	return (retval);
996 }
997 
998 /*
999  * This is called by the cyclic framework when this CPU becomes online
1000  */
1001 /*ARGSUSED*/
1002 static void
1003 cheetah_nudge_onln(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
1004 {
1005 
1006 	hdlr->cyh_func = (cyc_func_t)cheetah_nudge_buddy;
1007 	hdlr->cyh_level = CY_LOW_LEVEL;
1008 	hdlr->cyh_arg = NULL;
1009 
1010 	/*
1011 	 * Stagger the start time
1012 	 */
1013 	when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
1014 	if (cheetah_sendmondo_recover_delay < CHEETAH_LIVELOCK_MIN_DELAY) {
1015 		cheetah_sendmondo_recover_delay = CHEETAH_LIVELOCK_MIN_DELAY;
1016 	}
1017 	when->cyt_interval = cheetah_sendmondo_recover_delay * NANOSEC;
1018 }
1019 
1020 /*
1021  * Create a low level cyclic to send a xtrap to the next cpu online.
1022  * However, there's no need to have this running on a uniprocessor system.
1023  */
1024 static void
1025 cheetah_nudge_init(void)
1026 {
1027 	cyc_omni_handler_t hdlr;
1028 
1029 	if (max_ncpus == 1) {
1030 		return;
1031 	}
1032 
1033 	hdlr.cyo_online = cheetah_nudge_onln;
1034 	hdlr.cyo_offline = NULL;
1035 	hdlr.cyo_arg = NULL;
1036 
1037 	mutex_enter(&cpu_lock);
1038 	(void) cyclic_add_omni(&hdlr);
1039 	mutex_exit(&cpu_lock);
1040 }
1041 
1042 /*
1043  * Cyclic handler to wake up buddy
1044  */
1045 void
1046 cheetah_nudge_buddy(void)
1047 {
1048 	/*
1049 	 * Disable kernel preemption to protect the cpu list
1050 	 */
1051 	kpreempt_disable();
1052 	if ((CPU->cpu_next_onln != CPU) && (sendmondo_in_recover == 0)) {
1053 		xt_one(CPU->cpu_next_onln->cpu_id, (xcfunc_t *)xt_sync_tl1,
1054 		    0, 0);
1055 	}
1056 	kpreempt_enable();
1057 }
1058 
1059 #endif	/* CHEETAHPLUS_ERRATUM_25 */
1060 
1061 #ifdef SEND_MONDO_STATS
1062 uint32_t x_one_stimes[64];
1063 uint32_t x_one_ltimes[16];
1064 uint32_t x_set_stimes[64];
1065 uint32_t x_set_ltimes[16];
1066 uint32_t x_set_cpus[NCPU];
1067 uint32_t x_nack_stimes[64];
1068 #endif
1069 
1070 /*
1071  * Note: A version of this function is used by the debugger via the KDI,
1072  * and must be kept in sync with this version.  Any changes made to this
1073  * function to support new chips or to accomodate errata must also be included
1074  * in the KDI-specific version.  See us3_kdi.c.
1075  */
1076 void
1077 send_one_mondo(int cpuid)
1078 {
1079 	int busy, nack;
1080 	uint64_t idsr, starttick, endtick, tick, lasttick;
1081 	uint64_t busymask;
1082 #ifdef	CHEETAHPLUS_ERRATUM_25
1083 	int recovered = 0;
1084 #endif
1085 
1086 	CPU_STATS_ADDQ(CPU, sys, xcalls, 1);
1087 	starttick = lasttick = gettick();
1088 	shipit(cpuid, 0);
1089 	endtick = starttick + xc_tick_limit;
1090 	busy = nack = 0;
1091 #if defined(JALAPENO) || defined(SERRANO)
1092 	/*
1093 	 * Lower 2 bits of the agent ID determine which BUSY/NACK pair
1094 	 * will be used for dispatching interrupt. For now, assume
1095 	 * there are no more than IDSR_BN_SETS CPUs, hence no aliasing
1096 	 * issues with respect to BUSY/NACK pair usage.
1097 	 */
1098 	busymask  = IDSR_BUSY_BIT(cpuid);
1099 #else /* JALAPENO || SERRANO */
1100 	busymask = IDSR_BUSY;
1101 #endif /* JALAPENO || SERRANO */
1102 	for (;;) {
1103 		idsr = getidsr();
1104 		if (idsr == 0)
1105 			break;
1106 
1107 		tick = gettick();
1108 		/*
1109 		 * If there is a big jump between the current tick
1110 		 * count and lasttick, we have probably hit a break
1111 		 * point.  Adjust endtick accordingly to avoid panic.
1112 		 */
1113 		if (tick > (lasttick + xc_tick_jump_limit))
1114 			endtick += (tick - lasttick);
1115 		lasttick = tick;
1116 		if (tick > endtick) {
1117 			if (panic_quiesce)
1118 				return;
1119 #ifdef	CHEETAHPLUS_ERRATUM_25
1120 			if (cheetah_sendmondo_recover && recovered == 0) {
1121 				if (mondo_recover(cpuid, 0)) {
1122 					/*
1123 					 * We claimed the whole memory or
1124 					 * full scan is disabled.
1125 					 */
1126 					recovered++;
1127 				}
1128 				tick = gettick();
1129 				endtick = tick + xc_tick_limit;
1130 				lasttick = tick;
1131 				/*
1132 				 * Recheck idsr
1133 				 */
1134 				continue;
1135 			} else
1136 #endif	/* CHEETAHPLUS_ERRATUM_25 */
1137 			{
1138 				cmn_err(CE_PANIC, "send mondo timeout "
1139 				    "(target 0x%x) [%d NACK %d BUSY]",
1140 				    cpuid, nack, busy);
1141 			}
1142 		}
1143 
1144 		if (idsr & busymask) {
1145 			busy++;
1146 			continue;
1147 		}
1148 		drv_usecwait(1);
1149 		shipit(cpuid, 0);
1150 		nack++;
1151 		busy = 0;
1152 	}
1153 #ifdef SEND_MONDO_STATS
1154 	{
1155 		int n = gettick() - starttick;
1156 		if (n < 8192)
1157 			x_one_stimes[n >> 7]++;
1158 		else
1159 			x_one_ltimes[(n >> 13) & 0xf]++;
1160 	}
1161 #endif
1162 }
1163 
1164 void
1165 syncfpu(void)
1166 {
1167 }
1168 
1169 /*
1170  * Return processor specific async error structure
1171  * size used.
1172  */
1173 int
1174 cpu_aflt_size(void)
1175 {
1176 	return (sizeof (ch_async_flt_t));
1177 }
1178 
1179 /*
1180  * Tunable to disable the checking of other cpu logout areas during panic for
1181  * potential syndrome 71 generating errors.
1182  */
1183 int enable_check_other_cpus_logout = 1;
1184 
1185 /*
1186  * Check other cpus logout area for potential synd 71 generating
1187  * errors.
1188  */
1189 static void
1190 cpu_check_cpu_logout(int cpuid, caddr_t tpc, int tl, int ecc_type,
1191     ch_cpu_logout_t *clop)
1192 {
1193 	struct async_flt *aflt;
1194 	ch_async_flt_t ch_flt;
1195 	uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs;
1196 
1197 	if (clop == NULL || clop->clo_data.chd_afar == LOGOUT_INVALID) {
1198 		return;
1199 	}
1200 
1201 	bzero(&ch_flt, sizeof (ch_async_flt_t));
1202 
1203 	t_afar = clop->clo_data.chd_afar;
1204 	t_afsr = clop->clo_data.chd_afsr;
1205 	t_afsr_ext = clop->clo_data.chd_afsr_ext;
1206 #if defined(SERRANO)
1207 	ch_flt.afar2 = clop->clo_data.chd_afar2;
1208 #endif	/* SERRANO */
1209 
1210 	/*
1211 	 * In order to simplify code, we maintain this afsr_errs
1212 	 * variable which holds the aggregate of AFSR and AFSR_EXT
1213 	 * sticky bits.
1214 	 */
1215 	t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) |
1216 	    (t_afsr & C_AFSR_ALL_ERRS);
1217 
1218 	/* Setup the async fault structure */
1219 	aflt = (struct async_flt *)&ch_flt;
1220 	aflt->flt_id = gethrtime_waitfree();
1221 	ch_flt.afsr_ext = t_afsr_ext;
1222 	ch_flt.afsr_errs = t_afsr_errs;
1223 	aflt->flt_stat = t_afsr;
1224 	aflt->flt_addr = t_afar;
1225 	aflt->flt_bus_id = cpuid;
1226 	aflt->flt_inst = cpuid;
1227 	aflt->flt_pc = tpc;
1228 	aflt->flt_prot = AFLT_PROT_NONE;
1229 	aflt->flt_class = CPU_FAULT;
1230 	aflt->flt_priv = ((t_afsr & C_AFSR_PRIV) != 0);
1231 	aflt->flt_tl = tl;
1232 	aflt->flt_status = ecc_type;
1233 	aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs);
1234 
1235 	/*
1236 	 * Queue events on the async event queue, one event per error bit.
1237 	 * If no events are queued, queue an event to complain.
1238 	 */
1239 	if (cpu_queue_events(&ch_flt, NULL, t_afsr_errs, clop) == 0) {
1240 		ch_flt.flt_type = CPU_INV_AFSR;
1241 		cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR,
1242 		    (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue,
1243 		    aflt->flt_panic);
1244 	}
1245 
1246 	/*
1247 	 * Zero out + invalidate CPU logout.
1248 	 */
1249 	bzero(clop, sizeof (ch_cpu_logout_t));
1250 	clop->clo_data.chd_afar = LOGOUT_INVALID;
1251 }
1252 
1253 /*
1254  * Check the logout areas of all other cpus for unlogged errors.
1255  */
1256 static void
1257 cpu_check_other_cpus_logout(void)
1258 {
1259 	int i, j;
1260 	processorid_t myid;
1261 	struct cpu *cp;
1262 	ch_err_tl1_data_t *cl1p;
1263 
1264 	myid = CPU->cpu_id;
1265 	for (i = 0; i < NCPU; i++) {
1266 		cp = cpu[i];
1267 
1268 		if ((cp == NULL) || !(cp->cpu_flags & CPU_EXISTS) ||
1269 		    (cp->cpu_id == myid) || (CPU_PRIVATE(cp) == NULL)) {
1270 			continue;
1271 		}
1272 
1273 		/*
1274 		 * Check each of the tl>0 logout areas
1275 		 */
1276 		cl1p = CPU_PRIVATE_PTR(cp, chpr_tl1_err_data[0]);
1277 		for (j = 0; j < CH_ERR_TL1_TLMAX; j++, cl1p++) {
1278 			if (cl1p->ch_err_tl1_flags == 0)
1279 				continue;
1280 
1281 			cpu_check_cpu_logout(i, (caddr_t)cl1p->ch_err_tl1_tpc,
1282 			    1, ECC_F_TRAP, &cl1p->ch_err_tl1_logout);
1283 		}
1284 
1285 		/*
1286 		 * Check each of the remaining logout areas
1287 		 */
1288 		cpu_check_cpu_logout(i, NULL, 0, ECC_F_TRAP,
1289 		    CPU_PRIVATE_PTR(cp, chpr_fecctl0_logout));
1290 		cpu_check_cpu_logout(i, NULL, 0, ECC_C_TRAP,
1291 		    CPU_PRIVATE_PTR(cp, chpr_cecc_logout));
1292 		cpu_check_cpu_logout(i, NULL, 0, ECC_D_TRAP,
1293 		    CPU_PRIVATE_PTR(cp, chpr_async_logout));
1294 	}
1295 }
1296 
1297 /*
1298  * The fast_ecc_err handler transfers control here for UCU, UCC events.
1299  * Note that we flush Ecache twice, once in the fast_ecc_err handler to
1300  * flush the error that caused the UCU/UCC, then again here at the end to
1301  * flush the TL=1 trap handler code out of the Ecache, so we can minimize
1302  * the probability of getting a TL>1 Fast ECC trap when we're fielding
1303  * another Fast ECC trap.
1304  *
1305  * Cheetah+ also handles: TSCE: No additional processing required.
1306  * Panther adds L3_UCU and L3_UCC which are reported in AFSR_EXT.
1307  *
1308  * Note that the p_clo_flags input is only valid in cases where the
1309  * cpu_private struct is not yet initialized (since that is the only
1310  * time that information cannot be obtained from the logout struct.)
1311  */
1312 /*ARGSUSED*/
1313 void
1314 cpu_fast_ecc_error(struct regs *rp, ulong_t p_clo_flags)
1315 {
1316 	ch_cpu_logout_t *clop;
1317 	uint64_t ceen, nceen;
1318 
1319 	/*
1320 	 * Get the CPU log out info. If we can't find our CPU private
1321 	 * pointer, then we will have to make due without any detailed
1322 	 * logout information.
1323 	 */
1324 	if (CPU_PRIVATE(CPU) == NULL) {
1325 		clop = NULL;
1326 		ceen = p_clo_flags & EN_REG_CEEN;
1327 		nceen = p_clo_flags & EN_REG_NCEEN;
1328 	} else {
1329 		clop = CPU_PRIVATE_PTR(CPU, chpr_fecctl0_logout);
1330 		ceen = clop->clo_flags & EN_REG_CEEN;
1331 		nceen = clop->clo_flags & EN_REG_NCEEN;
1332 	}
1333 
1334 	cpu_log_fast_ecc_error((caddr_t)rp->r_pc,
1335 	    (rp->r_tstate & TSTATE_PRIV) ? 1 : 0, 0, ceen, nceen, clop);
1336 }
1337 
1338 /*
1339  * Log fast ecc error, called from either Fast ECC at TL=0 or Fast
1340  * ECC at TL>0.  Need to supply either a error register pointer or a
1341  * cpu logout structure pointer.
1342  */
1343 static void
1344 cpu_log_fast_ecc_error(caddr_t tpc, int priv, int tl, uint64_t ceen,
1345     uint64_t nceen, ch_cpu_logout_t *clop)
1346 {
1347 	struct async_flt *aflt;
1348 	ch_async_flt_t ch_flt;
1349 	uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs;
1350 	char pr_reason[MAX_REASON_STRING];
1351 	ch_cpu_errors_t cpu_error_regs;
1352 
1353 	bzero(&ch_flt, sizeof (ch_async_flt_t));
1354 	/*
1355 	 * If no cpu logout data, then we will have to make due without
1356 	 * any detailed logout information.
1357 	 */
1358 	if (clop == NULL) {
1359 		ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID;
1360 		get_cpu_error_state(&cpu_error_regs);
1361 		set_cpu_error_state(&cpu_error_regs);
1362 		t_afar = cpu_error_regs.afar;
1363 		t_afsr = cpu_error_regs.afsr;
1364 		t_afsr_ext = cpu_error_regs.afsr_ext;
1365 #if defined(SERRANO)
1366 		ch_flt.afar2 = cpu_error_regs.afar2;
1367 #endif	/* SERRANO */
1368 	} else {
1369 		t_afar = clop->clo_data.chd_afar;
1370 		t_afsr = clop->clo_data.chd_afsr;
1371 		t_afsr_ext = clop->clo_data.chd_afsr_ext;
1372 #if defined(SERRANO)
1373 		ch_flt.afar2 = clop->clo_data.chd_afar2;
1374 #endif	/* SERRANO */
1375 	}
1376 
1377 	/*
1378 	 * In order to simplify code, we maintain this afsr_errs
1379 	 * variable which holds the aggregate of AFSR and AFSR_EXT
1380 	 * sticky bits.
1381 	 */
1382 	t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) |
1383 	    (t_afsr & C_AFSR_ALL_ERRS);
1384 	pr_reason[0] = '\0';
1385 
1386 	/* Setup the async fault structure */
1387 	aflt = (struct async_flt *)&ch_flt;
1388 	aflt->flt_id = gethrtime_waitfree();
1389 	ch_flt.afsr_ext = t_afsr_ext;
1390 	ch_flt.afsr_errs = t_afsr_errs;
1391 	aflt->flt_stat = t_afsr;
1392 	aflt->flt_addr = t_afar;
1393 	aflt->flt_bus_id = getprocessorid();
1394 	aflt->flt_inst = CPU->cpu_id;
1395 	aflt->flt_pc = tpc;
1396 	aflt->flt_prot = AFLT_PROT_NONE;
1397 	aflt->flt_class = CPU_FAULT;
1398 	aflt->flt_priv = priv;
1399 	aflt->flt_tl = tl;
1400 	aflt->flt_status = ECC_F_TRAP;
1401 	aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs);
1402 
1403 	/*
1404 	 * XXXX - Phenomenal hack to get around Solaris not getting all the
1405 	 * cmn_err messages out to the console.  The situation is a UCU (in
1406 	 * priv mode) which causes a WDU which causes a UE (on the retry).
1407 	 * The messages for the UCU and WDU are enqueued and then pulled off
1408 	 * the async queue via softint and syslogd starts to process them
1409 	 * but doesn't get them to the console.  The UE causes a panic, but
1410 	 * since the UCU/WDU messages are already in transit, those aren't
1411 	 * on the async queue.  The hack is to check if we have a matching
1412 	 * WDU event for the UCU, and if it matches, we're more than likely
1413 	 * going to panic with a UE, unless we're under protection.  So, we
1414 	 * check to see if we got a matching WDU event and if we're under
1415 	 * protection.
1416 	 *
1417 	 * For Cheetah/Cheetah+/Jaguar/Jalapeno, the sequence we care about
1418 	 * looks like this:
1419 	 *    UCU->WDU->UE
1420 	 * For Panther, it could look like either of these:
1421 	 *    UCU---->WDU->L3_WDU->UE
1422 	 *    L3_UCU->WDU->L3_WDU->UE
1423 	 */
1424 	if ((t_afsr_errs & (C_AFSR_UCU | C_AFSR_L3_UCU)) &&
1425 	    aflt->flt_panic == 0 && aflt->flt_priv != 0 &&
1426 	    curthread->t_ontrap == NULL && curthread->t_lofault == NULL) {
1427 		get_cpu_error_state(&cpu_error_regs);
1428 		aflt->flt_panic |= ((cpu_error_regs.afsr & C_AFSR_WDU) &&
1429 		    (cpu_error_regs.afar == t_afar));
1430 		aflt->flt_panic |= ((clop == NULL) &&
1431 		    (t_afsr_errs & C_AFSR_WDU));
1432 	}
1433 
1434 	/*
1435 	 * Queue events on the async event queue, one event per error bit.
1436 	 * If no events are queued or no Fast ECC events are on in the AFSR,
1437 	 * queue an event to complain.
1438 	 */
1439 	if (cpu_queue_events(&ch_flt, pr_reason, t_afsr_errs, clop) == 0 ||
1440 	    ((t_afsr_errs & (C_AFSR_FECC_ERRS | C_AFSR_EXT_FECC_ERRS)) == 0)) {
1441 		ch_flt.flt_type = CPU_INV_AFSR;
1442 		cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR,
1443 		    (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue,
1444 		    aflt->flt_panic);
1445 	}
1446 
1447 	/*
1448 	 * Zero out + invalidate CPU logout.
1449 	 */
1450 	if (clop) {
1451 		bzero(clop, sizeof (ch_cpu_logout_t));
1452 		clop->clo_data.chd_afar = LOGOUT_INVALID;
1453 	}
1454 
1455 	/*
1456 	 * We carefully re-enable NCEEN and CEEN and then check if any deferred
1457 	 * or disrupting errors have happened.  We do this because if a
1458 	 * deferred or disrupting error had occurred with NCEEN/CEEN off, the
1459 	 * trap will not be taken when NCEEN/CEEN is re-enabled.  Note that
1460 	 * CEEN works differently on Cheetah than on Spitfire.  Also, we enable
1461 	 * NCEEN/CEEN *before* checking the AFSR to avoid the small window of a
1462 	 * deferred or disrupting error happening between checking the AFSR and
1463 	 * enabling NCEEN/CEEN.
1464 	 *
1465 	 * Note: CEEN and NCEEN are only reenabled if they were on when trap
1466 	 * taken.
1467 	 */
1468 	set_error_enable(get_error_enable() | (nceen | ceen));
1469 	if (clear_errors(&ch_flt)) {
1470 		aflt->flt_panic |= ((ch_flt.afsr_errs &
1471 		    (C_AFSR_EXT_ASYNC_ERRS | C_AFSR_ASYNC_ERRS)) != 0);
1472 		(void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs,
1473 		    NULL);
1474 	}
1475 
1476 	/*
1477 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
1478 	 * be logged as part of the panic flow.
1479 	 */
1480 	if (aflt->flt_panic)
1481 		fm_panic("%sError(s)", pr_reason);
1482 
1483 	/*
1484 	 * Flushing the Ecache here gets the part of the trap handler that
1485 	 * is run at TL=1 out of the Ecache.
1486 	 */
1487 	cpu_flush_ecache();
1488 }
1489 
1490 /*
1491  * This is called via sys_trap from pil15_interrupt code if the
1492  * corresponding entry in ch_err_tl1_pending is set.  Checks the
1493  * various ch_err_tl1_data structures for valid entries based on the bit
1494  * settings in the ch_err_tl1_flags entry of the structure.
1495  */
1496 /*ARGSUSED*/
1497 void
1498 cpu_tl1_error(struct regs *rp, int panic)
1499 {
1500 	ch_err_tl1_data_t *cl1p, cl1;
1501 	int i, ncl1ps;
1502 	uint64_t me_flags;
1503 	uint64_t ceen, nceen;
1504 
1505 	if (ch_err_tl1_paddrs[CPU->cpu_id] == 0) {
1506 		cl1p = &ch_err_tl1_data;
1507 		ncl1ps = 1;
1508 	} else if (CPU_PRIVATE(CPU) != NULL) {
1509 		cl1p = CPU_PRIVATE_PTR(CPU, chpr_tl1_err_data[0]);
1510 		ncl1ps = CH_ERR_TL1_TLMAX;
1511 	} else {
1512 		ncl1ps = 0;
1513 	}
1514 
1515 	for (i = 0; i < ncl1ps; i++, cl1p++) {
1516 		if (cl1p->ch_err_tl1_flags == 0)
1517 			continue;
1518 
1519 		/*
1520 		 * Grab a copy of the logout data and invalidate
1521 		 * the logout area.
1522 		 */
1523 		cl1 = *cl1p;
1524 		bzero(cl1p, sizeof (ch_err_tl1_data_t));
1525 		cl1p->ch_err_tl1_logout.clo_data.chd_afar = LOGOUT_INVALID;
1526 		me_flags = CH_ERR_ME_FLAGS(cl1.ch_err_tl1_flags);
1527 
1528 		/*
1529 		 * Log "first error" in ch_err_tl1_data.
1530 		 */
1531 		if (cl1.ch_err_tl1_flags & CH_ERR_FECC) {
1532 			ceen = get_error_enable() & EN_REG_CEEN;
1533 			nceen = get_error_enable() & EN_REG_NCEEN;
1534 			cpu_log_fast_ecc_error((caddr_t)cl1.ch_err_tl1_tpc, 1,
1535 			    1, ceen, nceen, &cl1.ch_err_tl1_logout);
1536 		}
1537 #if defined(CPU_IMP_L1_CACHE_PARITY)
1538 		if (cl1.ch_err_tl1_flags & (CH_ERR_IPE | CH_ERR_DPE)) {
1539 			cpu_parity_error(rp, cl1.ch_err_tl1_flags,
1540 			    (caddr_t)cl1.ch_err_tl1_tpc);
1541 		}
1542 #endif	/* CPU_IMP_L1_CACHE_PARITY */
1543 
1544 		/*
1545 		 * Log "multiple events" in ch_err_tl1_data.  Note that
1546 		 * we don't read and clear the AFSR/AFAR in the TL>0 code
1547 		 * if the structure is busy, we just do the cache flushing
1548 		 * we have to do and then do the retry.  So the AFSR/AFAR
1549 		 * at this point *should* have some relevant info.  If there
1550 		 * are no valid errors in the AFSR, we'll assume they've
1551 		 * already been picked up and logged.  For I$/D$ parity,
1552 		 * we just log an event with an "Unknown" (NULL) TPC.
1553 		 */
1554 		if (me_flags & CH_ERR_FECC) {
1555 			ch_cpu_errors_t cpu_error_regs;
1556 			uint64_t t_afsr_errs;
1557 
1558 			/*
1559 			 * Get the error registers and see if there's
1560 			 * a pending error.  If not, don't bother
1561 			 * generating an "Invalid AFSR" error event.
1562 			 */
1563 			get_cpu_error_state(&cpu_error_regs);
1564 			t_afsr_errs = (cpu_error_regs.afsr_ext &
1565 			    C_AFSR_EXT_ALL_ERRS) |
1566 			    (cpu_error_regs.afsr & C_AFSR_ALL_ERRS);
1567 			if (t_afsr_errs != 0) {
1568 				ceen = get_error_enable() & EN_REG_CEEN;
1569 				nceen = get_error_enable() & EN_REG_NCEEN;
1570 				cpu_log_fast_ecc_error((caddr_t)NULL, 1,
1571 				    1, ceen, nceen, NULL);
1572 			}
1573 		}
1574 #if defined(CPU_IMP_L1_CACHE_PARITY)
1575 		if (me_flags & (CH_ERR_IPE | CH_ERR_DPE)) {
1576 			cpu_parity_error(rp, me_flags, (caddr_t)NULL);
1577 		}
1578 #endif	/* CPU_IMP_L1_CACHE_PARITY */
1579 	}
1580 }
1581 
1582 /*
1583  * Called from Fast ECC TL>0 handler in case of fatal error.
1584  * cpu_tl1_error should always find an associated ch_err_tl1_data structure,
1585  * but if we don't, we'll panic with something reasonable.
1586  */
1587 /*ARGSUSED*/
1588 void
1589 cpu_tl1_err_panic(struct regs *rp, ulong_t flags)
1590 {
1591 	cpu_tl1_error(rp, 1);
1592 	/*
1593 	 * Should never return, but just in case.
1594 	 */
1595 	fm_panic("Unsurvivable ECC Error at TL>0");
1596 }
1597 
1598 /*
1599  * The ce_err/ce_err_tl1 handlers transfer control here for CE, EMC, EDU:ST,
1600  * EDC, WDU, WDC, CPU, CPC, IVU, IVC events.
1601  * Disrupting errors controlled by NCEEN: EDU:ST, WDU, CPU, IVU
1602  * Disrupting errors controlled by CEEN: CE, EMC, EDC, WDC, CPC, IVC
1603  *
1604  * Cheetah+ also handles (No additional processing required):
1605  *    DUE, DTO, DBERR	(NCEEN controlled)
1606  *    THCE		(CEEN and ET_ECC_en controlled)
1607  *    TUE		(ET_ECC_en controlled)
1608  *
1609  * Panther further adds:
1610  *    IMU, L3_EDU, L3_WDU, L3_CPU		(NCEEN controlled)
1611  *    IMC, L3_EDC, L3_WDC, L3_CPC, L3_THCE	(CEEN controlled)
1612  *    TUE_SH, TUE		(NCEEN and L2_tag_ECC_en controlled)
1613  *    L3_TUE, L3_TUE_SH		(NCEEN and ET_ECC_en controlled)
1614  *    THCE			(CEEN and L2_tag_ECC_en controlled)
1615  *    L3_THCE			(CEEN and ET_ECC_en controlled)
1616  *
1617  * Note that the p_clo_flags input is only valid in cases where the
1618  * cpu_private struct is not yet initialized (since that is the only
1619  * time that information cannot be obtained from the logout struct.)
1620  */
1621 /*ARGSUSED*/
1622 void
1623 cpu_disrupting_error(struct regs *rp, ulong_t p_clo_flags)
1624 {
1625 	struct async_flt *aflt;
1626 	ch_async_flt_t ch_flt;
1627 	char pr_reason[MAX_REASON_STRING];
1628 	ch_cpu_logout_t *clop;
1629 	uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs;
1630 	ch_cpu_errors_t cpu_error_regs;
1631 
1632 	bzero(&ch_flt, sizeof (ch_async_flt_t));
1633 	/*
1634 	 * Get the CPU log out info. If we can't find our CPU private
1635 	 * pointer, then we will have to make due without any detailed
1636 	 * logout information.
1637 	 */
1638 	if (CPU_PRIVATE(CPU) == NULL) {
1639 		clop = NULL;
1640 		ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID;
1641 		get_cpu_error_state(&cpu_error_regs);
1642 		set_cpu_error_state(&cpu_error_regs);
1643 		t_afar = cpu_error_regs.afar;
1644 		t_afsr = cpu_error_regs.afsr;
1645 		t_afsr_ext = cpu_error_regs.afsr_ext;
1646 #if defined(SERRANO)
1647 		ch_flt.afar2 = cpu_error_regs.afar2;
1648 #endif	/* SERRANO */
1649 	} else {
1650 		clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout);
1651 		t_afar = clop->clo_data.chd_afar;
1652 		t_afsr = clop->clo_data.chd_afsr;
1653 		t_afsr_ext = clop->clo_data.chd_afsr_ext;
1654 #if defined(SERRANO)
1655 		ch_flt.afar2 = clop->clo_data.chd_afar2;
1656 #endif	/* SERRANO */
1657 	}
1658 
1659 	/*
1660 	 * In order to simplify code, we maintain this afsr_errs
1661 	 * variable which holds the aggregate of AFSR and AFSR_EXT
1662 	 * sticky bits.
1663 	 */
1664 	t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) |
1665 	    (t_afsr & C_AFSR_ALL_ERRS);
1666 
1667 	pr_reason[0] = '\0';
1668 	/* Setup the async fault structure */
1669 	aflt = (struct async_flt *)&ch_flt;
1670 	ch_flt.afsr_ext = t_afsr_ext;
1671 	ch_flt.afsr_errs = t_afsr_errs;
1672 	aflt->flt_stat = t_afsr;
1673 	aflt->flt_addr = t_afar;
1674 	aflt->flt_pc = (caddr_t)rp->r_pc;
1675 	aflt->flt_priv = (rp->r_tstate & TSTATE_PRIV) ?  1 : 0;
1676 	aflt->flt_tl = 0;
1677 	aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs);
1678 
1679 	/*
1680 	 * If this trap is a result of one of the errors not masked
1681 	 * by cpu_ce_not_deferred, we don't reenable CEEN. Instead
1682 	 * indicate that a timeout is to be set later.
1683 	 */
1684 	if (!(t_afsr_errs & (cpu_ce_not_deferred | cpu_ce_not_deferred_ext)) &&
1685 	    !aflt->flt_panic)
1686 		ch_flt.flt_trapped_ce = CE_CEEN_DEFER | CE_CEEN_TRAPPED;
1687 	else
1688 		ch_flt.flt_trapped_ce = CE_CEEN_NODEFER | CE_CEEN_TRAPPED;
1689 
1690 	/*
1691 	 * log the CE and clean up
1692 	 */
1693 	cpu_log_and_clear_ce(&ch_flt);
1694 
1695 	/*
1696 	 * We re-enable CEEN (if required) and check if any disrupting errors
1697 	 * have happened.  We do this because if a disrupting error had occurred
1698 	 * with CEEN off, the trap will not be taken when CEEN is re-enabled.
1699 	 * Note that CEEN works differently on Cheetah than on Spitfire.  Also,
1700 	 * we enable CEEN *before* checking the AFSR to avoid the small window
1701 	 * of a error happening between checking the AFSR and enabling CEEN.
1702 	 */
1703 	if (ch_flt.flt_trapped_ce & CE_CEEN_NODEFER)
1704 	    set_error_enable(get_error_enable() | EN_REG_CEEN);
1705 	if (clear_errors(&ch_flt)) {
1706 		(void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs,
1707 		    NULL);
1708 	}
1709 
1710 	/*
1711 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
1712 	 * be logged as part of the panic flow.
1713 	 */
1714 	if (aflt->flt_panic)
1715 		fm_panic("%sError(s)", pr_reason);
1716 }
1717 
1718 /*
1719  * The async_err handler transfers control here for UE, EMU, EDU:BLD,
1720  * L3_EDU:BLD, TO, and BERR events.
1721  * Deferred errors controlled by NCEEN: UE, EMU, EDU:BLD, L3_EDU:BLD, TO, BERR
1722  *
1723  * Cheetah+: No additional errors handled.
1724  *
1725  * Note that the p_clo_flags input is only valid in cases where the
1726  * cpu_private struct is not yet initialized (since that is the only
1727  * time that information cannot be obtained from the logout struct.)
1728  */
1729 /*ARGSUSED*/
1730 void
1731 cpu_deferred_error(struct regs *rp, ulong_t p_clo_flags)
1732 {
1733 	ushort_t ttype, tl;
1734 	ch_async_flt_t ch_flt;
1735 	struct async_flt *aflt;
1736 	int trampolined = 0;
1737 	char pr_reason[MAX_REASON_STRING];
1738 	ch_cpu_logout_t *clop;
1739 	uint64_t ceen, clo_flags;
1740 	uint64_t log_afsr;
1741 	uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs;
1742 	ch_cpu_errors_t cpu_error_regs;
1743 	int expected = DDI_FM_ERR_UNEXPECTED;
1744 	ddi_acc_hdl_t *hp;
1745 
1746 	/*
1747 	 * We need to look at p_flag to determine if the thread detected an
1748 	 * error while dumping core.  We can't grab p_lock here, but it's ok
1749 	 * because we just need a consistent snapshot and we know that everyone
1750 	 * else will store a consistent set of bits while holding p_lock.  We
1751 	 * don't have to worry about a race because SDOCORE is set once prior
1752 	 * to doing i/o from the process's address space and is never cleared.
1753 	 */
1754 	uint_t pflag = ttoproc(curthread)->p_flag;
1755 
1756 	bzero(&ch_flt, sizeof (ch_async_flt_t));
1757 	/*
1758 	 * Get the CPU log out info. If we can't find our CPU private
1759 	 * pointer then we will have to make due without any detailed
1760 	 * logout information.
1761 	 */
1762 	if (CPU_PRIVATE(CPU) == NULL) {
1763 		clop = NULL;
1764 		ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID;
1765 		get_cpu_error_state(&cpu_error_regs);
1766 		set_cpu_error_state(&cpu_error_regs);
1767 		t_afar = cpu_error_regs.afar;
1768 		t_afsr = cpu_error_regs.afsr;
1769 		t_afsr_ext = cpu_error_regs.afsr_ext;
1770 #if defined(SERRANO)
1771 		ch_flt.afar2 = cpu_error_regs.afar2;
1772 #endif	/* SERRANO */
1773 		clo_flags = p_clo_flags;
1774 	} else {
1775 		clop = CPU_PRIVATE_PTR(CPU, chpr_async_logout);
1776 		t_afar = clop->clo_data.chd_afar;
1777 		t_afsr = clop->clo_data.chd_afsr;
1778 		t_afsr_ext = clop->clo_data.chd_afsr_ext;
1779 #if defined(SERRANO)
1780 		ch_flt.afar2 = clop->clo_data.chd_afar2;
1781 #endif	/* SERRANO */
1782 		clo_flags = clop->clo_flags;
1783 	}
1784 
1785 	/*
1786 	 * In order to simplify code, we maintain this afsr_errs
1787 	 * variable which holds the aggregate of AFSR and AFSR_EXT
1788 	 * sticky bits.
1789 	 */
1790 	t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) |
1791 	    (t_afsr & C_AFSR_ALL_ERRS);
1792 	pr_reason[0] = '\0';
1793 
1794 	/*
1795 	 * Grab information encoded into our clo_flags field.
1796 	 */
1797 	ceen = clo_flags & EN_REG_CEEN;
1798 	tl = (clo_flags & CLO_FLAGS_TL_MASK) >> CLO_FLAGS_TL_SHIFT;
1799 	ttype = (clo_flags & CLO_FLAGS_TT_MASK) >> CLO_FLAGS_TT_SHIFT;
1800 
1801 	/*
1802 	 * handle the specific error
1803 	 */
1804 	aflt = (struct async_flt *)&ch_flt;
1805 	aflt->flt_id = gethrtime_waitfree();
1806 	aflt->flt_bus_id = getprocessorid();
1807 	aflt->flt_inst = CPU->cpu_id;
1808 	ch_flt.afsr_ext = t_afsr_ext;
1809 	ch_flt.afsr_errs = t_afsr_errs;
1810 	aflt->flt_stat = t_afsr;
1811 	aflt->flt_addr = t_afar;
1812 	aflt->flt_pc = (caddr_t)rp->r_pc;
1813 	aflt->flt_prot = AFLT_PROT_NONE;
1814 	aflt->flt_class = CPU_FAULT;
1815 	aflt->flt_priv = (rp->r_tstate & TSTATE_PRIV) ?  1 : 0;
1816 	aflt->flt_tl = (uchar_t)tl;
1817 	aflt->flt_panic = ((tl != 0) || (aft_testfatal != 0) ||
1818 	    C_AFSR_PANIC(t_afsr_errs));
1819 	aflt->flt_core = (pflag & SDOCORE) ? 1 : 0;
1820 	aflt->flt_status = ((ttype == T_DATA_ERROR) ? ECC_D_TRAP : ECC_I_TRAP);
1821 
1822 	/*
1823 	 * If the trap occurred in privileged mode at TL=0, we need to check to
1824 	 * see if we were executing in the kernel under on_trap() or t_lofault
1825 	 * protection.  If so, modify the saved registers so that we return
1826 	 * from the trap to the appropriate trampoline routine.
1827 	 */
1828 	if (aflt->flt_priv && tl == 0) {
1829 		if (curthread->t_ontrap != NULL) {
1830 			on_trap_data_t *otp = curthread->t_ontrap;
1831 
1832 			if (otp->ot_prot & OT_DATA_EC) {
1833 				aflt->flt_prot = AFLT_PROT_EC;
1834 				otp->ot_trap |= OT_DATA_EC;
1835 				rp->r_pc = otp->ot_trampoline;
1836 				rp->r_npc = rp->r_pc + 4;
1837 				trampolined = 1;
1838 			}
1839 
1840 			if ((t_afsr & (C_AFSR_TO | C_AFSR_BERR)) &&
1841 			    (otp->ot_prot & OT_DATA_ACCESS)) {
1842 				aflt->flt_prot = AFLT_PROT_ACCESS;
1843 				otp->ot_trap |= OT_DATA_ACCESS;
1844 				rp->r_pc = otp->ot_trampoline;
1845 				rp->r_npc = rp->r_pc + 4;
1846 				trampolined = 1;
1847 				/*
1848 				 * for peeks and caut_gets errors are expected
1849 				 */
1850 				hp = (ddi_acc_hdl_t *)otp->ot_handle;
1851 				if (!hp)
1852 					expected = DDI_FM_ERR_PEEK;
1853 				else if (hp->ah_acc.devacc_attr_access ==
1854 				    DDI_CAUTIOUS_ACC)
1855 					expected = DDI_FM_ERR_EXPECTED;
1856 			}
1857 
1858 		} else if (curthread->t_lofault) {
1859 			aflt->flt_prot = AFLT_PROT_COPY;
1860 			rp->r_g1 = EFAULT;
1861 			rp->r_pc = curthread->t_lofault;
1862 			rp->r_npc = rp->r_pc + 4;
1863 			trampolined = 1;
1864 		}
1865 	}
1866 
1867 	/*
1868 	 * If we're in user mode or we're doing a protected copy, we either
1869 	 * want the ASTON code below to send a signal to the user process
1870 	 * or we want to panic if aft_panic is set.
1871 	 *
1872 	 * If we're in privileged mode and we're not doing a copy, then we
1873 	 * need to check if we've trampolined.  If we haven't trampolined,
1874 	 * we should panic.
1875 	 */
1876 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
1877 		if (t_afsr_errs &
1878 		    ((C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS) &
1879 		    ~(C_AFSR_BERR | C_AFSR_TO)))
1880 			aflt->flt_panic |= aft_panic;
1881 	} else if (!trampolined) {
1882 			aflt->flt_panic = 1;
1883 	}
1884 
1885 	/*
1886 	 * If we've trampolined due to a privileged TO or BERR, or if an
1887 	 * unprivileged TO or BERR occurred, we don't want to enqueue an
1888 	 * event for that TO or BERR.  Queue all other events (if any) besides
1889 	 * the TO/BERR.  Since we may not be enqueing any events, we need to
1890 	 * ignore the number of events queued.  If we haven't trampolined due
1891 	 * to a TO or BERR, just enqueue events normally.
1892 	 */
1893 	log_afsr = t_afsr_errs;
1894 	if (trampolined) {
1895 		log_afsr &= ~(C_AFSR_TO | C_AFSR_BERR);
1896 	} else if (!aflt->flt_priv) {
1897 		/*
1898 		 * User mode, suppress messages if
1899 		 * cpu_berr_to_verbose is not set.
1900 		 */
1901 		if (!cpu_berr_to_verbose)
1902 			log_afsr &= ~(C_AFSR_TO | C_AFSR_BERR);
1903 	}
1904 
1905 	/*
1906 	 * Log any errors that occurred
1907 	 */
1908 	if (((log_afsr &
1909 		((C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS) & ~C_AFSR_ME)) &&
1910 		cpu_queue_events(&ch_flt, pr_reason, log_afsr, clop) == 0) ||
1911 		(t_afsr_errs &
1912 		(C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS)) == 0) {
1913 		ch_flt.flt_type = CPU_INV_AFSR;
1914 		cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR,
1915 		    (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue,
1916 		    aflt->flt_panic);
1917 	}
1918 
1919 	/*
1920 	 * Zero out + invalidate CPU logout.
1921 	 */
1922 	if (clop) {
1923 		bzero(clop, sizeof (ch_cpu_logout_t));
1924 		clop->clo_data.chd_afar = LOGOUT_INVALID;
1925 	}
1926 
1927 #if defined(JALAPENO) || defined(SERRANO)
1928 	/*
1929 	 * UE/RUE/BERR/TO: Call our bus nexus friends to check for
1930 	 * IO errors that may have resulted in this trap.
1931 	 */
1932 	if (t_afsr & (C_AFSR_UE|C_AFSR_RUE|C_AFSR_TO|C_AFSR_BERR)) {
1933 		cpu_run_bus_error_handlers(aflt, expected);
1934 	}
1935 
1936 	/*
1937 	 * UE/RUE: If UE or RUE is in memory, we need to flush the bad
1938 	 * line from the Ecache.  We also need to query the bus nexus for
1939 	 * fatal errors.  Attempts to do diagnostic read on caches may
1940 	 * introduce more errors (especially when the module is bad).
1941 	 */
1942 	if (t_afsr & (C_AFSR_UE|C_AFSR_RUE)) {
1943 		/*
1944 		 * Ask our bus nexus friends if they have any fatal errors.  If
1945 		 * so, they will log appropriate error messages.
1946 		 */
1947 		if (bus_func_invoke(BF_TYPE_UE) == BF_FATAL)
1948 			aflt->flt_panic = 1;
1949 
1950 		/*
1951 		 * We got a UE or RUE and are panicking, save the fault PA in
1952 		 * a known location so that the platform specific panic code
1953 		 * can check for copyback errors.
1954 		 */
1955 		if (aflt->flt_panic && cpu_flt_in_memory(&ch_flt, C_AFSR_UE)) {
1956 			panic_aflt = *aflt;
1957 		}
1958 	}
1959 
1960 	/*
1961 	 * Flush Ecache line or entire Ecache
1962 	 */
1963 	if (t_afsr & (C_AFSR_UE | C_AFSR_RUE | C_AFSR_EDU | C_AFSR_BERR))
1964 		cpu_error_ecache_flush(&ch_flt);
1965 #else /* JALAPENO || SERRANO */
1966 	/*
1967 	 * UE/BERR/TO: Call our bus nexus friends to check for
1968 	 * IO errors that may have resulted in this trap.
1969 	 */
1970 	if (t_afsr & (C_AFSR_UE|C_AFSR_TO|C_AFSR_BERR)) {
1971 		cpu_run_bus_error_handlers(aflt, expected);
1972 	}
1973 
1974 	/*
1975 	 * UE: If the UE is in memory, we need to flush the bad
1976 	 * line from the Ecache.  We also need to query the bus nexus for
1977 	 * fatal errors.  Attempts to do diagnostic read on caches may
1978 	 * introduce more errors (especially when the module is bad).
1979 	 */
1980 	if (t_afsr & C_AFSR_UE) {
1981 		/*
1982 		 * Ask our legacy bus nexus friends if they have any fatal
1983 		 * errors.  If so, they will log appropriate error messages.
1984 		 */
1985 		if (bus_func_invoke(BF_TYPE_UE) == BF_FATAL)
1986 			aflt->flt_panic = 1;
1987 
1988 		/*
1989 		 * We got a UE and are panicking, save the fault PA in a known
1990 		 * location so that the platform specific panic code can check
1991 		 * for copyback errors.
1992 		 */
1993 		if (aflt->flt_panic && cpu_flt_in_memory(&ch_flt, C_AFSR_UE)) {
1994 			panic_aflt = *aflt;
1995 		}
1996 	}
1997 
1998 	/*
1999 	 * Flush Ecache line or entire Ecache
2000 	 */
2001 	if (t_afsr_errs &
2002 	    (C_AFSR_UE | C_AFSR_EDU | C_AFSR_BERR | C_AFSR_L3_EDU))
2003 		cpu_error_ecache_flush(&ch_flt);
2004 #endif /* JALAPENO || SERRANO */
2005 
2006 	/*
2007 	 * We carefully re-enable NCEEN and CEEN and then check if any deferred
2008 	 * or disrupting errors have happened.  We do this because if a
2009 	 * deferred or disrupting error had occurred with NCEEN/CEEN off, the
2010 	 * trap will not be taken when NCEEN/CEEN is re-enabled.  Note that
2011 	 * CEEN works differently on Cheetah than on Spitfire.  Also, we enable
2012 	 * NCEEN/CEEN *before* checking the AFSR to avoid the small window of a
2013 	 * deferred or disrupting error happening between checking the AFSR and
2014 	 * enabling NCEEN/CEEN.
2015 	 *
2016 	 * Note: CEEN reenabled only if it was on when trap taken.
2017 	 */
2018 	set_error_enable(get_error_enable() | (EN_REG_NCEEN | ceen));
2019 	if (clear_errors(&ch_flt)) {
2020 		/*
2021 		 * Check for secondary errors, and avoid panicking if we
2022 		 * have them
2023 		 */
2024 		if (cpu_check_secondary_errors(&ch_flt, t_afsr_errs,
2025 		    t_afar) == 0) {
2026 			aflt->flt_panic |= ((ch_flt.afsr_errs &
2027 			    (C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS)) != 0);
2028 		}
2029 		(void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs,
2030 		    NULL);
2031 	}
2032 
2033 	/*
2034 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
2035 	 * be logged as part of the panic flow.
2036 	 */
2037 	if (aflt->flt_panic)
2038 		fm_panic("%sError(s)", pr_reason);
2039 
2040 	/*
2041 	 * If we queued an error and we are going to return from the trap and
2042 	 * the error was in user mode or inside of a copy routine, set AST flag
2043 	 * so the queue will be drained before returning to user mode.  The
2044 	 * AST processing will also act on our failure policy.
2045 	 */
2046 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
2047 		int pcb_flag = 0;
2048 
2049 		if (t_afsr_errs &
2050 		    (C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS &
2051 		    ~(C_AFSR_BERR | C_AFSR_TO)))
2052 			pcb_flag |= ASYNC_HWERR;
2053 
2054 		if (t_afsr & C_AFSR_BERR)
2055 			pcb_flag |= ASYNC_BERR;
2056 
2057 		if (t_afsr & C_AFSR_TO)
2058 			pcb_flag |= ASYNC_BTO;
2059 
2060 		ttolwp(curthread)->lwp_pcb.pcb_flags |= pcb_flag;
2061 		aston(curthread);
2062 	}
2063 }
2064 
2065 #if defined(CPU_IMP_L1_CACHE_PARITY)
2066 /*
2067  * Handling of data and instruction parity errors (traps 0x71, 0x72).
2068  *
2069  * For Panther, P$ data parity errors during floating point load hits
2070  * are also detected (reported as TT 0x71) and handled by this trap
2071  * handler.
2072  *
2073  * AFSR/AFAR are not set for parity errors, only TPC (a virtual address)
2074  * is available.
2075  */
2076 /*ARGSUSED*/
2077 void
2078 cpu_parity_error(struct regs *rp, uint_t flags, caddr_t tpc)
2079 {
2080 	ch_async_flt_t ch_flt;
2081 	struct async_flt *aflt;
2082 	uchar_t tl = ((flags & CH_ERR_TL) != 0);
2083 	uchar_t iparity = ((flags & CH_ERR_IPE) != 0);
2084 	uchar_t panic = ((flags & CH_ERR_PANIC) != 0);
2085 	char *error_class;
2086 
2087 	/*
2088 	 * Log the error.
2089 	 * For icache parity errors the fault address is the trap PC.
2090 	 * For dcache/pcache parity errors the instruction would have to
2091 	 * be decoded to determine the address and that isn't possible
2092 	 * at high PIL.
2093 	 */
2094 	bzero(&ch_flt, sizeof (ch_async_flt_t));
2095 	aflt = (struct async_flt *)&ch_flt;
2096 	aflt->flt_id = gethrtime_waitfree();
2097 	aflt->flt_bus_id = getprocessorid();
2098 	aflt->flt_inst = CPU->cpu_id;
2099 	aflt->flt_pc = tpc;
2100 	aflt->flt_addr = iparity ? (uint64_t)tpc : AFLT_INV_ADDR;
2101 	aflt->flt_prot = AFLT_PROT_NONE;
2102 	aflt->flt_class = CPU_FAULT;
2103 	aflt->flt_priv = (tl || (rp->r_tstate & TSTATE_PRIV)) ?  1 : 0;
2104 	aflt->flt_tl = tl;
2105 	aflt->flt_panic = panic;
2106 	aflt->flt_status = iparity ? ECC_IP_TRAP : ECC_DP_TRAP;
2107 	ch_flt.flt_type = iparity ? CPU_IC_PARITY : CPU_DC_PARITY;
2108 
2109 	if (iparity) {
2110 		cpu_icache_parity_info(&ch_flt);
2111 		if (ch_flt.parity_data.ipe.cpl_off != -1)
2112 			error_class = FM_EREPORT_CPU_USIII_IDSPE;
2113 		else if (ch_flt.parity_data.ipe.cpl_way != -1)
2114 			error_class = FM_EREPORT_CPU_USIII_ITSPE;
2115 		else
2116 			error_class = FM_EREPORT_CPU_USIII_IPE;
2117 		aflt->flt_payload = FM_EREPORT_PAYLOAD_ICACHE_PE;
2118 	} else {
2119 		cpu_dcache_parity_info(&ch_flt);
2120 		if (ch_flt.parity_data.dpe.cpl_off != -1)
2121 			error_class = FM_EREPORT_CPU_USIII_DDSPE;
2122 		else if (ch_flt.parity_data.dpe.cpl_way != -1)
2123 			error_class = FM_EREPORT_CPU_USIII_DTSPE;
2124 		else
2125 			error_class = FM_EREPORT_CPU_USIII_DPE;
2126 		aflt->flt_payload = FM_EREPORT_PAYLOAD_DCACHE_PE;
2127 		/*
2128 		 * For panther we also need to check the P$ for parity errors.
2129 		 */
2130 		if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
2131 			cpu_pcache_parity_info(&ch_flt);
2132 			if (ch_flt.parity_data.dpe.cpl_cache == CPU_PC_PARITY) {
2133 				error_class = FM_EREPORT_CPU_USIII_PDSPE;
2134 				aflt->flt_payload =
2135 				    FM_EREPORT_PAYLOAD_PCACHE_PE;
2136 			}
2137 		}
2138 	}
2139 
2140 	cpu_errorq_dispatch(error_class, (void *)&ch_flt,
2141 	    sizeof (ch_async_flt_t), ue_queue, aflt->flt_panic);
2142 
2143 	if (iparity) {
2144 		/*
2145 		 * Invalidate entire I$.
2146 		 * This is required due to the use of diagnostic ASI
2147 		 * accesses that may result in a loss of I$ coherency.
2148 		 */
2149 		if (cache_boot_state & DCU_IC) {
2150 			flush_icache();
2151 		}
2152 		/*
2153 		 * According to section P.3.1 of the Panther PRM, we
2154 		 * need to do a little more for recovery on those
2155 		 * CPUs after encountering an I$ parity error.
2156 		 */
2157 		if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
2158 			flush_ipb();
2159 			correct_dcache_parity(dcache_size,
2160 			    dcache_linesize);
2161 			flush_pcache();
2162 		}
2163 	} else {
2164 		/*
2165 		 * Since the valid bit is ignored when checking parity the
2166 		 * D$ data and tag must also be corrected.  Set D$ data bits
2167 		 * to zero and set utag to 0, 1, 2, 3.
2168 		 */
2169 		correct_dcache_parity(dcache_size, dcache_linesize);
2170 
2171 		/*
2172 		 * According to section P.3.3 of the Panther PRM, we
2173 		 * need to do a little more for recovery on those
2174 		 * CPUs after encountering a D$ or P$ parity error.
2175 		 *
2176 		 * As far as clearing P$ parity errors, it is enough to
2177 		 * simply invalidate all entries in the P$ since P$ parity
2178 		 * error traps are only generated for floating point load
2179 		 * hits.
2180 		 */
2181 		if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
2182 			flush_icache();
2183 			flush_ipb();
2184 			flush_pcache();
2185 		}
2186 	}
2187 
2188 	/*
2189 	 * Invalidate entire D$ if it was enabled.
2190 	 * This is done to avoid stale data in the D$ which might
2191 	 * occur with the D$ disabled and the trap handler doing
2192 	 * stores affecting lines already in the D$.
2193 	 */
2194 	if (cache_boot_state & DCU_DC) {
2195 		flush_dcache();
2196 	}
2197 
2198 	/*
2199 	 * Restore caches to their bootup state.
2200 	 */
2201 	set_dcu(get_dcu() | cache_boot_state);
2202 
2203 	/*
2204 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
2205 	 * be logged as part of the panic flow.
2206 	 */
2207 	if (aflt->flt_panic)
2208 		fm_panic("%sError(s)", iparity ? "IPE " : "DPE ");
2209 
2210 	/*
2211 	 * If this error occurred at TL>0 then flush the E$ here to reduce
2212 	 * the chance of getting an unrecoverable Fast ECC error.  This
2213 	 * flush will evict the part of the parity trap handler that is run
2214 	 * at TL>1.
2215 	 */
2216 	if (tl) {
2217 		cpu_flush_ecache();
2218 	}
2219 }
2220 
2221 /*
2222  * On an I$ parity error, mark the appropriate entries in the ch_async_flt_t
2223  * to indicate which portions of the captured data should be in the ereport.
2224  */
2225 void
2226 cpu_async_log_ic_parity_err(ch_async_flt_t *ch_flt)
2227 {
2228 	int way = ch_flt->parity_data.ipe.cpl_way;
2229 	int offset = ch_flt->parity_data.ipe.cpl_off;
2230 	int tag_index;
2231 	struct async_flt *aflt = (struct async_flt *)ch_flt;
2232 
2233 
2234 	if ((offset != -1) || (way != -1)) {
2235 		/*
2236 		 * Parity error in I$ tag or data
2237 		 */
2238 		tag_index = ch_flt->parity_data.ipe.cpl_ic[way].ic_idx;
2239 		if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation))
2240 			ch_flt->parity_data.ipe.cpl_ic[way].ic_way =
2241 			    PN_ICIDX_TO_WAY(tag_index);
2242 		else
2243 			ch_flt->parity_data.ipe.cpl_ic[way].ic_way =
2244 			    CH_ICIDX_TO_WAY(tag_index);
2245 		ch_flt->parity_data.ipe.cpl_ic[way].ic_logflag =
2246 		    IC_LOGFLAG_MAGIC;
2247 	} else {
2248 		/*
2249 		 * Parity error was not identified.
2250 		 * Log tags and data for all ways.
2251 		 */
2252 		for (way = 0; way < CH_ICACHE_NWAY; way++) {
2253 			tag_index = ch_flt->parity_data.ipe.cpl_ic[way].ic_idx;
2254 			if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation))
2255 				ch_flt->parity_data.ipe.cpl_ic[way].ic_way =
2256 				    PN_ICIDX_TO_WAY(tag_index);
2257 			else
2258 				ch_flt->parity_data.ipe.cpl_ic[way].ic_way =
2259 				    CH_ICIDX_TO_WAY(tag_index);
2260 			ch_flt->parity_data.ipe.cpl_ic[way].ic_logflag =
2261 			    IC_LOGFLAG_MAGIC;
2262 		}
2263 	}
2264 }
2265 
2266 /*
2267  * On an D$ parity error, mark the appropriate entries in the ch_async_flt_t
2268  * to indicate which portions of the captured data should be in the ereport.
2269  */
2270 void
2271 cpu_async_log_dc_parity_err(ch_async_flt_t *ch_flt)
2272 {
2273 	int way = ch_flt->parity_data.dpe.cpl_way;
2274 	int offset = ch_flt->parity_data.dpe.cpl_off;
2275 	int tag_index;
2276 
2277 	if (offset != -1) {
2278 		/*
2279 		 * Parity error in D$ or P$ data array.
2280 		 *
2281 		 * First check to see whether the parity error is in D$ or P$
2282 		 * since P$ data parity errors are reported in Panther using
2283 		 * the same trap.
2284 		 */
2285 		if (ch_flt->parity_data.dpe.cpl_cache == CPU_PC_PARITY) {
2286 			tag_index = ch_flt->parity_data.dpe.cpl_pc[way].pc_idx;
2287 			ch_flt->parity_data.dpe.cpl_pc[way].pc_way =
2288 			    CH_PCIDX_TO_WAY(tag_index);
2289 			ch_flt->parity_data.dpe.cpl_pc[way].pc_logflag =
2290 			    PC_LOGFLAG_MAGIC;
2291 		} else {
2292 			tag_index = ch_flt->parity_data.dpe.cpl_dc[way].dc_idx;
2293 			ch_flt->parity_data.dpe.cpl_dc[way].dc_way =
2294 			    CH_DCIDX_TO_WAY(tag_index);
2295 			ch_flt->parity_data.dpe.cpl_dc[way].dc_logflag =
2296 			    DC_LOGFLAG_MAGIC;
2297 		}
2298 	} else if (way != -1) {
2299 		/*
2300 		 * Parity error in D$ tag.
2301 		 */
2302 		tag_index = ch_flt->parity_data.dpe.cpl_dc[way].dc_idx;
2303 		ch_flt->parity_data.dpe.cpl_dc[way].dc_way =
2304 		    CH_DCIDX_TO_WAY(tag_index);
2305 		ch_flt->parity_data.dpe.cpl_dc[way].dc_logflag =
2306 		    DC_LOGFLAG_MAGIC;
2307 	}
2308 }
2309 #endif	/* CPU_IMP_L1_CACHE_PARITY */
2310 
2311 /*
2312  * The cpu_async_log_err() function is called via the [uc]e_drain() function to
2313  * post-process CPU events that are dequeued.  As such, it can be invoked
2314  * from softint context, from AST processing in the trap() flow, or from the
2315  * panic flow.  We decode the CPU-specific data, and take appropriate actions.
2316  * Historically this entry point was used to log the actual cmn_err(9F) text;
2317  * now with FMA it is used to prepare 'flt' to be converted into an ereport.
2318  * With FMA this function now also returns a flag which indicates to the
2319  * caller whether the ereport should be posted (1) or suppressed (0).
2320  */
2321 static int
2322 cpu_async_log_err(void *flt, errorq_elem_t *eqep)
2323 {
2324 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)flt;
2325 	struct async_flt *aflt = (struct async_flt *)flt;
2326 	uint64_t errors;
2327 
2328 	switch (ch_flt->flt_type) {
2329 	case CPU_INV_AFSR:
2330 		/*
2331 		 * If it is a disrupting trap and the AFSR is zero, then
2332 		 * the event has probably already been noted. Do not post
2333 		 * an ereport.
2334 		 */
2335 		if ((aflt->flt_status & ECC_C_TRAP) &&
2336 		    (!(aflt->flt_stat & C_AFSR_MASK)))
2337 			return (0);
2338 		else
2339 			return (1);
2340 	case CPU_TO:
2341 	case CPU_BERR:
2342 	case CPU_FATAL:
2343 	case CPU_FPUERR:
2344 		return (1);
2345 
2346 	case CPU_UE_ECACHE_RETIRE:
2347 		cpu_log_err(aflt);
2348 		cpu_page_retire(ch_flt);
2349 		return (1);
2350 
2351 	/*
2352 	 * Cases where we may want to suppress logging or perform
2353 	 * extended diagnostics.
2354 	 */
2355 	case CPU_CE:
2356 	case CPU_EMC:
2357 		/*
2358 		 * We want to skip logging and further classification
2359 		 * only if ALL the following conditions are true:
2360 		 *
2361 		 *	1. There is only one error
2362 		 *	2. That error is a correctable memory error
2363 		 *	3. The error is caused by the memory scrubber (in
2364 		 *	   which case the error will have occurred under
2365 		 *	   on_trap protection)
2366 		 *	4. The error is on a retired page
2367 		 *
2368 		 * Note: AFLT_PROT_EC is used places other than the memory
2369 		 * scrubber.  However, none of those errors should occur
2370 		 * on a retired page.
2371 		 */
2372 		if ((ch_flt->afsr_errs &
2373 		    (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) == C_AFSR_CE &&
2374 		    aflt->flt_prot == AFLT_PROT_EC) {
2375 
2376 			if (page_retire_check(aflt->flt_addr, NULL) == 0) {
2377 			    if (ch_flt->flt_trapped_ce & CE_CEEN_DEFER) {
2378 
2379 				/*
2380 				 * Since we're skipping logging, we'll need
2381 				 * to schedule the re-enabling of CEEN
2382 				 */
2383 				(void) timeout(cpu_delayed_check_ce_errors,
2384 				    (void *)(uintptr_t)aflt->flt_inst,
2385 				    drv_usectohz((clock_t)cpu_ceen_delay_secs
2386 						 * MICROSEC));
2387 			    }
2388 			    return (0);
2389 			}
2390 		}
2391 
2392 		/*
2393 		 * Perform/schedule further classification actions, but
2394 		 * only if the page is healthy (we don't want bad
2395 		 * pages inducing too much diagnostic activity).  If we could
2396 		 * not find a page pointer then we also skip this.  If
2397 		 * ce_scrub_xdiag_recirc returns nonzero then it has chosen
2398 		 * to copy and recirculate the event (for further diagnostics)
2399 		 * and we should not proceed to log it here.
2400 		 *
2401 		 * This must be the last step here before the cpu_log_err()
2402 		 * below - if an event recirculates cpu_ce_log_err() will
2403 		 * not call the current function but just proceed directly
2404 		 * to cpu_ereport_post after the cpu_log_err() avoided below.
2405 		 *
2406 		 * Note: Check cpu_impl_async_log_err if changing this
2407 		 */
2408 		if (page_retire_check(aflt->flt_addr, &errors) == EINVAL) {
2409 			CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
2410 			    CE_XDIAG_SKIP_NOPP);
2411 		} else {
2412 			if (errors != PR_OK) {
2413 				CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
2414 				    CE_XDIAG_SKIP_PAGEDET);
2415 			} else if (ce_scrub_xdiag_recirc(aflt, ce_queue, eqep,
2416 			    offsetof(ch_async_flt_t, cmn_asyncflt))) {
2417 				return (0);
2418 			}
2419 		}
2420 		/*FALLTHRU*/
2421 
2422 	/*
2423 	 * Cases where we just want to report the error and continue.
2424 	 */
2425 	case CPU_CE_ECACHE:
2426 	case CPU_UE_ECACHE:
2427 	case CPU_IV:
2428 	case CPU_ORPH:
2429 		cpu_log_err(aflt);
2430 		return (1);
2431 
2432 	/*
2433 	 * Cases where we want to fall through to handle panicking.
2434 	 */
2435 	case CPU_UE:
2436 		/*
2437 		 * We want to skip logging in the same conditions as the
2438 		 * CE case.  In addition, we want to make sure we're not
2439 		 * panicking.
2440 		 */
2441 		if (!panicstr && (ch_flt->afsr_errs &
2442 		    (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) == C_AFSR_UE &&
2443 		    aflt->flt_prot == AFLT_PROT_EC) {
2444 			if (page_retire_check(aflt->flt_addr, NULL) == 0) {
2445 				/* Zero the address to clear the error */
2446 				softcall(ecc_page_zero, (void *)aflt->flt_addr);
2447 				return (0);
2448 			}
2449 		}
2450 		cpu_log_err(aflt);
2451 		break;
2452 
2453 	default:
2454 		/*
2455 		 * If the us3_common.c code doesn't know the flt_type, it may
2456 		 * be an implementation-specific code.  Call into the impldep
2457 		 * backend to find out what to do: if it tells us to continue,
2458 		 * break and handle as if falling through from a UE; if not,
2459 		 * the impldep backend has handled the error and we're done.
2460 		 */
2461 		switch (cpu_impl_async_log_err(flt, eqep)) {
2462 		case CH_ASYNC_LOG_DONE:
2463 			return (1);
2464 		case CH_ASYNC_LOG_RECIRC:
2465 			return (0);
2466 		case CH_ASYNC_LOG_CONTINUE:
2467 			break; /* continue on to handle UE-like error */
2468 		default:
2469 			cmn_err(CE_WARN, "discarding error 0x%p with "
2470 			    "invalid fault type (0x%x)",
2471 			    (void *)aflt, ch_flt->flt_type);
2472 			return (0);
2473 		}
2474 	}
2475 
2476 	/* ... fall through from the UE case */
2477 
2478 	if (aflt->flt_addr != AFLT_INV_ADDR && aflt->flt_in_memory) {
2479 		if (!panicstr) {
2480 			cpu_page_retire(ch_flt);
2481 		} else {
2482 			/*
2483 			 * Clear UEs on panic so that we don't
2484 			 * get haunted by them during panic or
2485 			 * after reboot
2486 			 */
2487 			cpu_clearphys(aflt);
2488 			(void) clear_errors(NULL);
2489 		}
2490 	}
2491 
2492 	return (1);
2493 }
2494 
2495 /*
2496  * Retire the bad page that may contain the flushed error.
2497  */
2498 void
2499 cpu_page_retire(ch_async_flt_t *ch_flt)
2500 {
2501 	struct async_flt *aflt = (struct async_flt *)ch_flt;
2502 	(void) page_retire(aflt->flt_addr, PR_UE);
2503 }
2504 
2505 /*
2506  * Return true if the error specified in the AFSR indicates
2507  * an E$ data error (L2$ for Cheetah/Cheetah+/Jaguar, L3$
2508  * for Panther, none for Jalapeno/Serrano).
2509  */
2510 /* ARGSUSED */
2511 static int
2512 cpu_error_is_ecache_data(int cpuid, uint64_t t_afsr)
2513 {
2514 #if defined(JALAPENO) || defined(SERRANO)
2515 	return (0);
2516 #elif defined(CHEETAH_PLUS)
2517 	if (IS_PANTHER(cpunodes[cpuid].implementation))
2518 		return ((t_afsr & C_AFSR_EXT_L3_DATA_ERRS) != 0);
2519 	return ((t_afsr & C_AFSR_EC_DATA_ERRS) != 0);
2520 #else	/* CHEETAH_PLUS */
2521 	return ((t_afsr & C_AFSR_EC_DATA_ERRS) != 0);
2522 #endif
2523 }
2524 
2525 /*
2526  * The cpu_log_err() function is called by cpu_async_log_err() to perform the
2527  * generic event post-processing for correctable and uncorrectable memory,
2528  * E$, and MTag errors.  Historically this entry point was used to log bits of
2529  * common cmn_err(9F) text; now with FMA it is used to prepare 'flt' to be
2530  * converted into an ereport.  In addition, it transmits the error to any
2531  * platform-specific service-processor FRU logging routines, if available.
2532  */
2533 void
2534 cpu_log_err(struct async_flt *aflt)
2535 {
2536 	char unum[UNUM_NAMLEN];
2537 	int synd_status, synd_code, afar_status;
2538 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
2539 
2540 	if (cpu_error_is_ecache_data(aflt->flt_inst, ch_flt->flt_bit))
2541 		aflt->flt_status |= ECC_ECACHE;
2542 	else
2543 		aflt->flt_status &= ~ECC_ECACHE;
2544 	/*
2545 	 * Determine syndrome status.
2546 	 */
2547 	synd_status = afsr_to_synd_status(aflt->flt_inst,
2548 	    ch_flt->afsr_errs, ch_flt->flt_bit);
2549 
2550 	/*
2551 	 * Determine afar status.
2552 	 */
2553 	if (pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT))
2554 		afar_status = afsr_to_afar_status(ch_flt->afsr_errs,
2555 				ch_flt->flt_bit);
2556 	else
2557 		afar_status = AFLT_STAT_INVALID;
2558 
2559 	synd_code = synd_to_synd_code(synd_status,
2560 	    aflt->flt_synd, ch_flt->flt_bit);
2561 
2562 	/*
2563 	 * If afar status is not invalid do a unum lookup.
2564 	 */
2565 	if (afar_status != AFLT_STAT_INVALID) {
2566 		(void) cpu_get_mem_unum_synd(synd_code, aflt, unum);
2567 	} else {
2568 		unum[0] = '\0';
2569 	}
2570 
2571 	/*
2572 	 * Do not send the fruid message (plat_ecc_error_data_t)
2573 	 * to the SC if it can handle the enhanced error information
2574 	 * (plat_ecc_error2_data_t) or when the tunable
2575 	 * ecc_log_fruid_enable is set to 0.
2576 	 */
2577 
2578 	if (&plat_ecc_capability_sc_get &&
2579 	    plat_ecc_capability_sc_get(PLAT_ECC_ERROR_MESSAGE)) {
2580 		if (&plat_log_fruid_error)
2581 			plat_log_fruid_error(synd_code, aflt, unum,
2582 			    ch_flt->flt_bit);
2583 	}
2584 
2585 	if (aflt->flt_func != NULL)
2586 		aflt->flt_func(aflt, unum);
2587 
2588 	if (afar_status != AFLT_STAT_INVALID)
2589 		cpu_log_diag_info(ch_flt);
2590 
2591 	/*
2592 	 * If we have a CEEN error , we do not reenable CEEN until after
2593 	 * we exit the trap handler. Otherwise, another error may
2594 	 * occur causing the handler to be entered recursively.
2595 	 * We set a timeout to trigger in cpu_ceen_delay_secs seconds,
2596 	 * to try and ensure that the CPU makes progress in the face
2597 	 * of a CE storm.
2598 	 */
2599 	if (ch_flt->flt_trapped_ce & CE_CEEN_DEFER) {
2600 		(void) timeout(cpu_delayed_check_ce_errors,
2601 		    (void *)(uintptr_t)aflt->flt_inst,
2602 		    drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC));
2603 	}
2604 }
2605 
2606 /*
2607  * Invoked by error_init() early in startup and therefore before
2608  * startup_errorq() is called to drain any error Q -
2609  *
2610  * startup()
2611  *   startup_end()
2612  *     error_init()
2613  *       cpu_error_init()
2614  * errorq_init()
2615  *   errorq_drain()
2616  * start_other_cpus()
2617  *
2618  * The purpose of this routine is to create error-related taskqs.  Taskqs
2619  * are used for this purpose because cpu_lock can't be grabbed from interrupt
2620  * context.
2621  */
2622 void
2623 cpu_error_init(int items)
2624 {
2625 	/*
2626 	 * Create taskq(s) to reenable CE
2627 	 */
2628 	ch_check_ce_tq = taskq_create("cheetah_check_ce", 1, minclsyspri,
2629 	    items, items, TASKQ_PREPOPULATE);
2630 }
2631 
2632 void
2633 cpu_ce_log_err(struct async_flt *aflt, errorq_elem_t *eqep)
2634 {
2635 	char unum[UNUM_NAMLEN];
2636 	int len;
2637 
2638 	switch (aflt->flt_class) {
2639 	case CPU_FAULT:
2640 		cpu_ereport_init(aflt);
2641 		if (cpu_async_log_err(aflt, eqep))
2642 			cpu_ereport_post(aflt);
2643 		break;
2644 
2645 	case BUS_FAULT:
2646 		if (aflt->flt_func != NULL) {
2647 			(void) cpu_get_mem_unum_aflt(AFLT_STAT_VALID, aflt,
2648 			    unum, UNUM_NAMLEN, &len);
2649 			aflt->flt_func(aflt, unum);
2650 		}
2651 		break;
2652 
2653 	case RECIRC_CPU_FAULT:
2654 		aflt->flt_class = CPU_FAULT;
2655 		cpu_log_err(aflt);
2656 		cpu_ereport_post(aflt);
2657 		break;
2658 
2659 	case RECIRC_BUS_FAULT:
2660 		ASSERT(aflt->flt_class != RECIRC_BUS_FAULT);
2661 		/*FALLTHRU*/
2662 	default:
2663 		cmn_err(CE_WARN, "discarding CE error 0x%p with invalid "
2664 		    "fault class (0x%x)", (void *)aflt, aflt->flt_class);
2665 		return;
2666 	}
2667 }
2668 
2669 /*
2670  * Scrub and classify a CE.  This function must not modify the
2671  * fault structure passed to it but instead should return the classification
2672  * information.
2673  */
2674 
2675 static uchar_t
2676 cpu_ce_scrub_mem_err_common(struct async_flt *ecc, boolean_t logout_tried)
2677 {
2678 	uchar_t disp = CE_XDIAG_EXTALG;
2679 	on_trap_data_t otd;
2680 	uint64_t orig_err;
2681 	ch_cpu_logout_t *clop;
2682 
2683 	/*
2684 	 * Clear CEEN.  CPU CE TL > 0 trap handling will already have done
2685 	 * this, but our other callers have not.  Disable preemption to
2686 	 * avoid CPU migration so that we restore CEEN on the correct
2687 	 * cpu later.
2688 	 *
2689 	 * CEEN is cleared so that further CEs that our instruction and
2690 	 * data footprint induce do not cause use to either creep down
2691 	 * kernel stack to the point of overflow, or do so much CE
2692 	 * notification as to make little real forward progress.
2693 	 *
2694 	 * NCEEN must not be cleared.  However it is possible that
2695 	 * our accesses to the flt_addr may provoke a bus error or timeout
2696 	 * if the offending address has just been unconfigured as part of
2697 	 * a DR action.  So we must operate under on_trap protection.
2698 	 */
2699 	kpreempt_disable();
2700 	orig_err = get_error_enable();
2701 	if (orig_err & EN_REG_CEEN)
2702 	    set_error_enable(orig_err & ~EN_REG_CEEN);
2703 
2704 	/*
2705 	 * Our classification algorithm includes the line state before
2706 	 * the scrub; we'd like this captured after the detection and
2707 	 * before the algorithm below - the earlier the better.
2708 	 *
2709 	 * If we've come from a cpu CE trap then this info already exists
2710 	 * in the cpu logout area.
2711 	 *
2712 	 * For a CE detected by memscrub for which there was no trap
2713 	 * (running with CEEN off) cpu_log_and_clear_ce has called
2714 	 * cpu_ce_delayed_ec_logout to capture some cache data, and
2715 	 * marked the fault structure as incomplete as a flag to later
2716 	 * logging code.
2717 	 *
2718 	 * If called directly from an IO detected CE there has been
2719 	 * no line data capture.  In this case we logout to the cpu logout
2720 	 * area - that's appropriate since it's the cpu cache data we need
2721 	 * for classification.  We thus borrow the cpu logout area for a
2722 	 * short time, and cpu_ce_delayed_ec_logout will mark it as busy in
2723 	 * this time (we will invalidate it again below).
2724 	 *
2725 	 * If called from the partner check xcall handler then this cpu
2726 	 * (the partner) has not necessarily experienced a CE at this
2727 	 * address.  But we want to capture line state before its scrub
2728 	 * attempt since we use that in our classification.
2729 	 */
2730 	if (logout_tried == B_FALSE) {
2731 		if (!cpu_ce_delayed_ec_logout(ecc->flt_addr))
2732 			disp |= CE_XDIAG_NOLOGOUT;
2733 	}
2734 
2735 	/*
2736 	 * Scrub memory, then check AFSR for errors.  The AFAR we scrub may
2737 	 * no longer be valid (if DR'd since the initial event) so we
2738 	 * perform this scrub under on_trap protection.  If this access is
2739 	 * ok then further accesses below will also be ok - DR cannot
2740 	 * proceed while this thread is active (preemption is disabled);
2741 	 * to be safe we'll nonetheless use on_trap again below.
2742 	 */
2743 	if (!on_trap(&otd, OT_DATA_ACCESS)) {
2744 		cpu_scrubphys(ecc);
2745 	} else {
2746 		no_trap();
2747 		if (orig_err & EN_REG_CEEN)
2748 		    set_error_enable(orig_err);
2749 		kpreempt_enable();
2750 		return (disp);
2751 	}
2752 	no_trap();
2753 
2754 	/*
2755 	 * Did the casx read of the scrub log a CE that matches the AFAR?
2756 	 * Note that it's quite possible that the read sourced the data from
2757 	 * another cpu.
2758 	 */
2759 	if (clear_ecc(ecc))
2760 		disp |= CE_XDIAG_CE1;
2761 
2762 	/*
2763 	 * Read the data again.  This time the read is very likely to
2764 	 * come from memory since the scrub induced a writeback to memory.
2765 	 */
2766 	if (!on_trap(&otd, OT_DATA_ACCESS)) {
2767 		(void) lddphys(P2ALIGN(ecc->flt_addr, 8));
2768 	} else {
2769 		no_trap();
2770 		if (orig_err & EN_REG_CEEN)
2771 		    set_error_enable(orig_err);
2772 		kpreempt_enable();
2773 		return (disp);
2774 	}
2775 	no_trap();
2776 
2777 	/* Did that read induce a CE that matches the AFAR? */
2778 	if (clear_ecc(ecc))
2779 		disp |= CE_XDIAG_CE2;
2780 
2781 	/*
2782 	 * Look at the logout information and record whether we found the
2783 	 * line in l2/l3 cache.  For Panther we are interested in whether
2784 	 * we found it in either cache (it won't reside in both but
2785 	 * it is possible to read it that way given the moving target).
2786 	 */
2787 	clop = CPU_PRIVATE(CPU) ? CPU_PRIVATE_PTR(CPU, chpr_cecc_logout) : NULL;
2788 	if (!(disp & CE_XDIAG_NOLOGOUT) && clop &&
2789 	    clop->clo_data.chd_afar != LOGOUT_INVALID) {
2790 		int hit, level;
2791 		int state;
2792 		int totalsize;
2793 		ch_ec_data_t *ecp;
2794 
2795 		/*
2796 		 * If hit is nonzero then a match was found and hit will
2797 		 * be one greater than the index which hit.  For Panther we
2798 		 * also need to pay attention to level to see which of l2$ or
2799 		 * l3$ it hit in.
2800 		 */
2801 		hit = cpu_matching_ecache_line(ecc->flt_addr, &clop->clo_data,
2802 		    0, &level);
2803 
2804 		if (hit) {
2805 			--hit;
2806 			disp |= CE_XDIAG_AFARMATCH;
2807 
2808 			if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
2809 				if (level == 2)
2810 					ecp = &clop->clo_data.chd_l2_data[hit];
2811 				else
2812 					ecp = &clop->clo_data.chd_ec_data[hit];
2813 			} else {
2814 				ASSERT(level == 2);
2815 				ecp = &clop->clo_data.chd_ec_data[hit];
2816 			}
2817 			totalsize = cpunodes[CPU->cpu_id].ecache_size;
2818 			state = cpu_ectag_pa_to_subblk_state(totalsize,
2819 			    ecc->flt_addr, ecp->ec_tag);
2820 
2821 			/*
2822 			 * Cheetah variants use different state encodings -
2823 			 * the CH_ECSTATE_* defines vary depending on the
2824 			 * module we're compiled for.  Translate into our
2825 			 * one true version.  Conflate Owner-Shared state
2826 			 * of SSM mode with Owner as victimisation of such
2827 			 * lines may cause a writeback.
2828 			 */
2829 			switch (state) {
2830 			case CH_ECSTATE_MOD:
2831 				disp |= EC_STATE_M;
2832 				break;
2833 
2834 			case CH_ECSTATE_OWN:
2835 			case CH_ECSTATE_OWS:
2836 				disp |= EC_STATE_O;
2837 				break;
2838 
2839 			case CH_ECSTATE_EXL:
2840 				disp |= EC_STATE_E;
2841 				break;
2842 
2843 			case CH_ECSTATE_SHR:
2844 				disp |= EC_STATE_S;
2845 				break;
2846 
2847 			default:
2848 				disp |= EC_STATE_I;
2849 				break;
2850 			}
2851 		}
2852 
2853 		/*
2854 		 * If we initiated the delayed logout then we are responsible
2855 		 * for invalidating the logout area.
2856 		 */
2857 		if (logout_tried == B_FALSE) {
2858 			bzero(clop, sizeof (ch_cpu_logout_t));
2859 			clop->clo_data.chd_afar = LOGOUT_INVALID;
2860 		}
2861 	}
2862 
2863 	/*
2864 	 * Re-enable CEEN if we turned it off.
2865 	 */
2866 	if (orig_err & EN_REG_CEEN)
2867 	    set_error_enable(orig_err);
2868 	kpreempt_enable();
2869 
2870 	return (disp);
2871 }
2872 
2873 /*
2874  * Scrub a correctable memory error and collect data for classification
2875  * of CE type.  This function is called in the detection path, ie tl0 handling
2876  * of a correctable error trap (cpus) or interrupt (IO) at high PIL.
2877  */
2878 void
2879 cpu_ce_scrub_mem_err(struct async_flt *ecc, boolean_t logout_tried)
2880 {
2881 	/*
2882 	 * Cheetah CE classification does not set any bits in flt_status.
2883 	 * Instead we will record classification datapoints in flt_disp.
2884 	 */
2885 	ecc->flt_status &= ~(ECC_INTERMITTENT | ECC_PERSISTENT | ECC_STICKY);
2886 
2887 	/*
2888 	 * To check if the error detected by IO is persistent, sticky or
2889 	 * intermittent.  This is noticed by clear_ecc().
2890 	 */
2891 	if (ecc->flt_status & ECC_IOBUS)
2892 		ecc->flt_stat = C_AFSR_MEMORY;
2893 
2894 	/*
2895 	 * Record information from this first part of the algorithm in
2896 	 * flt_disp.
2897 	 */
2898 	ecc->flt_disp = cpu_ce_scrub_mem_err_common(ecc, logout_tried);
2899 }
2900 
2901 /*
2902  * Select a partner to perform a further CE classification check from.
2903  * Must be called with kernel preemption disabled (to stop the cpu list
2904  * from changing).  The detecting cpu we are partnering has cpuid
2905  * aflt->flt_inst; we might not be running on the detecting cpu.
2906  *
2907  * Restrict choice to active cpus in the same cpu partition as ourselves in
2908  * an effort to stop bad cpus in one partition causing other partitions to
2909  * perform excessive diagnostic activity.  Actually since the errorq drain
2910  * is run from a softint most of the time and that is a global mechanism
2911  * this isolation is only partial.  Return NULL if we fail to find a
2912  * suitable partner.
2913  *
2914  * We prefer a partner that is in a different latency group to ourselves as
2915  * we will share fewer datapaths.  If such a partner is unavailable then
2916  * choose one in the same lgroup but prefer a different chip and only allow
2917  * a sibling core if flags includes PTNR_SIBLINGOK.  If all else fails and
2918  * flags includes PTNR_SELFOK then permit selection of the original detector.
2919  *
2920  * We keep a cache of the last partner selected for a cpu, and we'll try to
2921  * use that previous partner if no more than cpu_ce_ptnr_cachetime_sec seconds
2922  * have passed since that selection was made.  This provides the benefit
2923  * of the point-of-view of different partners over time but without
2924  * requiring frequent cpu list traversals.
2925  */
2926 
2927 #define	PTNR_SIBLINGOK	0x1	/* Allow selection of sibling core */
2928 #define	PTNR_SELFOK	0x2	/* Allow selection of cpu to "partner" itself */
2929 
2930 static cpu_t *
2931 ce_ptnr_select(struct async_flt *aflt, int flags, int *typep)
2932 {
2933 	cpu_t *sp, *dtcr, *ptnr, *locptnr, *sibptnr;
2934 	hrtime_t lasttime, thistime;
2935 
2936 	ASSERT(curthread->t_preempt > 0 || getpil() >= DISP_LEVEL);
2937 
2938 	dtcr = cpu[aflt->flt_inst];
2939 
2940 	/*
2941 	 * Short-circuit for the following cases:
2942 	 *	. the dtcr is not flagged active
2943 	 *	. there is just one cpu present
2944 	 *	. the detector has disappeared
2945 	 *	. we were given a bad flt_inst cpuid; this should not happen
2946 	 *	  (eg PCI code now fills flt_inst) but if it does it is no
2947 	 *	  reason to panic.
2948 	 *	. there is just one cpu left online in the cpu partition
2949 	 *
2950 	 * If we return NULL after this point then we do not update the
2951 	 * chpr_ceptnr_seltime which will cause us to perform a full lookup
2952 	 * again next time; this is the case where the only other cpu online
2953 	 * in the detector's partition is on the same chip as the detector
2954 	 * and since CEEN re-enable is throttled even that case should not
2955 	 * hurt performance.
2956 	 */
2957 	if (dtcr == NULL || !cpu_flagged_active(dtcr->cpu_flags)) {
2958 		return (NULL);
2959 	}
2960 	if (ncpus == 1 || dtcr->cpu_part->cp_ncpus == 1) {
2961 		if (flags & PTNR_SELFOK) {
2962 			*typep = CE_XDIAG_PTNR_SELF;
2963 			return (dtcr);
2964 		} else {
2965 			return (NULL);
2966 		}
2967 	}
2968 
2969 	thistime = gethrtime();
2970 	lasttime = CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime);
2971 
2972 	/*
2973 	 * Select a starting point.
2974 	 */
2975 	if (!lasttime) {
2976 		/*
2977 		 * We've never selected a partner for this detector before.
2978 		 * Start the scan at the next online cpu in the same cpu
2979 		 * partition.
2980 		 */
2981 		sp = dtcr->cpu_next_part;
2982 	} else if (thistime - lasttime < cpu_ce_ptnr_cachetime_sec * NANOSEC) {
2983 		/*
2984 		 * Our last selection has not aged yet.  If this partner:
2985 		 *	. is still a valid cpu,
2986 		 *	. is still in the same partition as the detector
2987 		 *	. is still marked active
2988 		 *	. satisfies the 'flags' argument criteria
2989 		 * then select it again without updating the timestamp.
2990 		 */
2991 		sp = cpu[CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id)];
2992 		if (sp == NULL || sp->cpu_part != dtcr->cpu_part ||
2993 		    !cpu_flagged_active(sp->cpu_flags) ||
2994 		    (sp == dtcr && !(flags & PTNR_SELFOK)) ||
2995 		    (sp->cpu_chip->chip_id == dtcr->cpu_chip->chip_id &&
2996 		    !(flags & PTNR_SIBLINGOK))) {
2997 			sp = dtcr->cpu_next_part;
2998 		} else {
2999 			if (sp->cpu_lpl->lpl_lgrp != dtcr->cpu_lpl->lpl_lgrp) {
3000 				*typep = CE_XDIAG_PTNR_REMOTE;
3001 			} else if (sp == dtcr) {
3002 				*typep = CE_XDIAG_PTNR_SELF;
3003 			} else if (sp->cpu_chip->chip_id ==
3004 			    dtcr->cpu_chip->chip_id) {
3005 				*typep = CE_XDIAG_PTNR_SIBLING;
3006 			} else {
3007 				*typep = CE_XDIAG_PTNR_LOCAL;
3008 			}
3009 			return (sp);
3010 		}
3011 	} else {
3012 		/*
3013 		 * Our last selection has aged.  If it is nonetheless still a
3014 		 * valid cpu then start the scan at the next cpu in the
3015 		 * partition after our last partner.  If the last selection
3016 		 * is no longer a valid cpu then go with our default.  In
3017 		 * this way we slowly cycle through possible partners to
3018 		 * obtain multiple viewpoints over time.
3019 		 */
3020 		sp = cpu[CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id)];
3021 		if (sp == NULL) {
3022 			sp = dtcr->cpu_next_part;
3023 		} else {
3024 			sp = sp->cpu_next_part;		/* may be dtcr */
3025 			if (sp->cpu_part != dtcr->cpu_part)
3026 				sp = dtcr;
3027 		}
3028 	}
3029 
3030 	/*
3031 	 * We have a proposed starting point for our search, but if this
3032 	 * cpu is offline then its cpu_next_part will point to itself
3033 	 * so we can't use that to iterate over cpus in this partition in
3034 	 * the loop below.  We still want to avoid iterating over cpus not
3035 	 * in our partition, so in the case that our starting point is offline
3036 	 * we will repoint it to be the detector itself;  and if the detector
3037 	 * happens to be offline we'll return NULL from the following loop.
3038 	 */
3039 	if (!cpu_flagged_active(sp->cpu_flags)) {
3040 		sp = dtcr;
3041 	}
3042 
3043 	ptnr = sp;
3044 	locptnr = NULL;
3045 	sibptnr = NULL;
3046 	do {
3047 		if (ptnr == dtcr || !cpu_flagged_active(ptnr->cpu_flags))
3048 			continue;
3049 		if (ptnr->cpu_lpl->lpl_lgrp != dtcr->cpu_lpl->lpl_lgrp) {
3050 			CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = ptnr->cpu_id;
3051 			CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime;
3052 			*typep = CE_XDIAG_PTNR_REMOTE;
3053 			return (ptnr);
3054 		}
3055 		if (ptnr->cpu_chip->chip_id == dtcr->cpu_chip->chip_id) {
3056 			if (sibptnr == NULL)
3057 				sibptnr = ptnr;
3058 			continue;
3059 		}
3060 		if (locptnr == NULL)
3061 			locptnr = ptnr;
3062 	} while ((ptnr = ptnr->cpu_next_part) != sp);
3063 
3064 	/*
3065 	 * A foreign partner has already been returned if one was available.
3066 	 *
3067 	 * If locptnr is not NULL it is a cpu in the same lgroup as the
3068 	 * detector, is active, and is not a sibling of the detector.
3069 	 *
3070 	 * If sibptnr is not NULL it is a sibling of the detector, and is
3071 	 * active.
3072 	 *
3073 	 * If we have to resort to using the detector itself we have already
3074 	 * checked that it is active.
3075 	 */
3076 	if (locptnr) {
3077 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = locptnr->cpu_id;
3078 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime;
3079 		*typep = CE_XDIAG_PTNR_LOCAL;
3080 		return (locptnr);
3081 	} else if (sibptnr && flags & PTNR_SIBLINGOK) {
3082 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = sibptnr->cpu_id;
3083 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime;
3084 		*typep = CE_XDIAG_PTNR_SIBLING;
3085 		return (sibptnr);
3086 	} else if (flags & PTNR_SELFOK) {
3087 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = dtcr->cpu_id;
3088 		CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime;
3089 		*typep = CE_XDIAG_PTNR_SELF;
3090 		return (dtcr);
3091 	}
3092 
3093 	return (NULL);
3094 }
3095 
3096 /*
3097  * Cross call handler that is requested to run on the designated partner of
3098  * a cpu that experienced a possibly sticky or possibly persistnet CE.
3099  */
3100 static void
3101 ce_ptnrchk_xc(struct async_flt *aflt, uchar_t *dispp)
3102 {
3103 	*dispp = cpu_ce_scrub_mem_err_common(aflt, B_FALSE);
3104 }
3105 
3106 /*
3107  * The associated errorqs are never destroyed so we do not need to deal with
3108  * them disappearing before this timeout fires.  If the affected memory
3109  * has been DR'd out since the original event the scrub algrithm will catch
3110  * any errors and return null disposition info.  If the original detecting
3111  * cpu has been DR'd out then ereport detector info will not be able to
3112  * lookup CPU type;  with a small timeout this is unlikely.
3113  */
3114 static void
3115 ce_lkychk_cb(ce_lkychk_cb_t *cbarg)
3116 {
3117 	struct async_flt *aflt = cbarg->lkycb_aflt;
3118 	uchar_t disp;
3119 	cpu_t *cp;
3120 	int ptnrtype;
3121 
3122 	kpreempt_disable();
3123 	if (cp = ce_ptnr_select(aflt, PTNR_SIBLINGOK | PTNR_SELFOK,
3124 	    &ptnrtype)) {
3125 		xc_one(cp->cpu_id, (xcfunc_t *)ce_ptnrchk_xc, (uint64_t)aflt,
3126 		    (uint64_t)&disp);
3127 		CE_XDIAG_SETLKYINFO(aflt->flt_disp, disp);
3128 		CE_XDIAG_SETPTNRID(aflt->flt_disp, cp->cpu_id);
3129 		CE_XDIAG_SETPTNRTYPE(aflt->flt_disp, ptnrtype);
3130 	} else {
3131 		ce_xdiag_lkydrops++;
3132 		if (ncpus > 1)
3133 			CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
3134 			    CE_XDIAG_SKIP_NOPTNR);
3135 	}
3136 	kpreempt_enable();
3137 
3138 	errorq_commit(cbarg->lkycb_eqp, cbarg->lkycb_eqep, ERRORQ_ASYNC);
3139 	kmem_free(cbarg, sizeof (ce_lkychk_cb_t));
3140 }
3141 
3142 /*
3143  * Called from errorq drain code when processing a CE error, both from
3144  * CPU and PCI drain functions.  Decide what further classification actions,
3145  * if any, we will perform.  Perform immediate actions now, and schedule
3146  * delayed actions as required.  Note that we are no longer necessarily running
3147  * on the detecting cpu, and that the async_flt structure will not persist on
3148  * return from this function.
3149  *
3150  * Calls to this function should aim to be self-throtlling in some way.  With
3151  * the delayed re-enable of CEEN the absolute rate of calls should not
3152  * be excessive.  Callers should also avoid performing in-depth classification
3153  * for events in pages that are already known to be suspect.
3154  *
3155  * We return nonzero to indicate that the event has been copied and
3156  * recirculated for further testing.  The caller should not log the event
3157  * in this case - it will be logged when further test results are available.
3158  *
3159  * Our possible contexts are that of errorq_drain: below lock level or from
3160  * panic context.  We can assume that the cpu we are running on is online.
3161  */
3162 
3163 
3164 #ifdef DEBUG
3165 static int ce_xdiag_forceaction;
3166 #endif
3167 
3168 int
3169 ce_scrub_xdiag_recirc(struct async_flt *aflt, errorq_t *eqp,
3170     errorq_elem_t *eqep, size_t afltoffset)
3171 {
3172 	ce_dispact_t dispact, action;
3173 	cpu_t *cp;
3174 	uchar_t dtcrinfo, disp;
3175 	int ptnrtype;
3176 
3177 	if (!ce_disp_inited || panicstr || ce_xdiag_off) {
3178 		ce_xdiag_drops++;
3179 		return (0);
3180 	} else if (!aflt->flt_in_memory) {
3181 		ce_xdiag_drops++;
3182 		CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_NOTMEM);
3183 		return (0);
3184 	}
3185 
3186 	dtcrinfo = CE_XDIAG_DTCRINFO(aflt->flt_disp);
3187 
3188 	/*
3189 	 * Some correctable events are not scrubbed/classified, such as those
3190 	 * noticed at the tail of cpu_deferred_error.  So if there is no
3191 	 * initial detector classification go no further.
3192 	 */
3193 	if (!CE_XDIAG_EXT_ALG_APPLIED(dtcrinfo)) {
3194 		ce_xdiag_drops++;
3195 		CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_NOSCRUB);
3196 		return (0);
3197 	}
3198 
3199 	dispact = CE_DISPACT(ce_disp_table,
3200 	    CE_XDIAG_AFARMATCHED(dtcrinfo),
3201 	    CE_XDIAG_STATE(dtcrinfo),
3202 	    CE_XDIAG_CE1SEEN(dtcrinfo),
3203 	    CE_XDIAG_CE2SEEN(dtcrinfo));
3204 
3205 
3206 	action = CE_ACT(dispact);	/* bad lookup caught below */
3207 #ifdef DEBUG
3208 	if (ce_xdiag_forceaction != 0)
3209 		action = ce_xdiag_forceaction;
3210 #endif
3211 
3212 	switch (action) {
3213 	case CE_ACT_LKYCHK: {
3214 		caddr_t ndata;
3215 		errorq_elem_t *neqep;
3216 		struct async_flt *ecc;
3217 		ce_lkychk_cb_t *cbargp;
3218 
3219 		if ((ndata = errorq_elem_dup(eqp, eqep, &neqep)) == NULL) {
3220 			ce_xdiag_lkydrops++;
3221 			CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
3222 			    CE_XDIAG_SKIP_DUPFAIL);
3223 			break;
3224 		}
3225 		ecc = (struct async_flt *)(ndata + afltoffset);
3226 
3227 		ASSERT(ecc->flt_class == CPU_FAULT ||
3228 		    ecc->flt_class == BUS_FAULT);
3229 		ecc->flt_class = (ecc->flt_class == CPU_FAULT) ?
3230 		    RECIRC_CPU_FAULT : RECIRC_BUS_FAULT;
3231 
3232 		cbargp = kmem_alloc(sizeof (ce_lkychk_cb_t), KM_SLEEP);
3233 		cbargp->lkycb_aflt = ecc;
3234 		cbargp->lkycb_eqp = eqp;
3235 		cbargp->lkycb_eqep = neqep;
3236 
3237 		(void) timeout((void (*)(void *))ce_lkychk_cb,
3238 		    (void *)cbargp, drv_usectohz(cpu_ce_lkychk_timeout_usec));
3239 		return (1);
3240 	}
3241 
3242 	case CE_ACT_PTNRCHK:
3243 		kpreempt_disable();	/* stop cpu list changing */
3244 		if ((cp = ce_ptnr_select(aflt, 0, &ptnrtype)) != NULL) {
3245 			xc_one(cp->cpu_id, (xcfunc_t *)ce_ptnrchk_xc,
3246 			    (uint64_t)aflt, (uint64_t)&disp);
3247 			CE_XDIAG_SETPTNRINFO(aflt->flt_disp, disp);
3248 			CE_XDIAG_SETPTNRID(aflt->flt_disp, cp->cpu_id);
3249 			CE_XDIAG_SETPTNRTYPE(aflt->flt_disp, ptnrtype);
3250 		} else if (ncpus > 1) {
3251 			ce_xdiag_ptnrdrops++;
3252 			CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
3253 			    CE_XDIAG_SKIP_NOPTNR);
3254 		} else {
3255 			ce_xdiag_ptnrdrops++;
3256 			CE_XDIAG_SETSKIPCODE(aflt->flt_disp,
3257 			    CE_XDIAG_SKIP_UNIPROC);
3258 		}
3259 		kpreempt_enable();
3260 		break;
3261 
3262 	case CE_ACT_DONE:
3263 		break;
3264 
3265 	case CE_ACT(CE_DISP_BAD):
3266 	default:
3267 #ifdef DEBUG
3268 		cmn_err(CE_PANIC, "ce_scrub_post: Bad action '%d'", action);
3269 #endif
3270 		ce_xdiag_bad++;
3271 		CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_ACTBAD);
3272 		break;
3273 	}
3274 
3275 	return (0);
3276 }
3277 
3278 /*
3279  * We route all errors through a single switch statement.
3280  */
3281 void
3282 cpu_ue_log_err(struct async_flt *aflt)
3283 {
3284 	switch (aflt->flt_class) {
3285 	case CPU_FAULT:
3286 		cpu_ereport_init(aflt);
3287 		if (cpu_async_log_err(aflt, NULL))
3288 			cpu_ereport_post(aflt);
3289 		break;
3290 
3291 	case BUS_FAULT:
3292 		bus_async_log_err(aflt);
3293 		break;
3294 
3295 	default:
3296 		cmn_err(CE_WARN, "discarding async error %p with invalid "
3297 		    "fault class (0x%x)", (void *)aflt, aflt->flt_class);
3298 		return;
3299 	}
3300 }
3301 
3302 /*
3303  * Routine for panic hook callback from panic_idle().
3304  */
3305 void
3306 cpu_async_panic_callb(void)
3307 {
3308 	ch_async_flt_t ch_flt;
3309 	struct async_flt *aflt;
3310 	ch_cpu_errors_t cpu_error_regs;
3311 	uint64_t afsr_errs;
3312 
3313 	get_cpu_error_state(&cpu_error_regs);
3314 
3315 	afsr_errs = (cpu_error_regs.afsr & C_AFSR_ALL_ERRS) |
3316 	    (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS);
3317 
3318 	if (afsr_errs) {
3319 
3320 		bzero(&ch_flt, sizeof (ch_async_flt_t));
3321 		aflt = (struct async_flt *)&ch_flt;
3322 		aflt->flt_id = gethrtime_waitfree();
3323 		aflt->flt_bus_id = getprocessorid();
3324 		aflt->flt_inst = CPU->cpu_id;
3325 		aflt->flt_stat = cpu_error_regs.afsr;
3326 		aflt->flt_addr = cpu_error_regs.afar;
3327 		aflt->flt_prot = AFLT_PROT_NONE;
3328 		aflt->flt_class = CPU_FAULT;
3329 		aflt->flt_priv = ((cpu_error_regs.afsr & C_AFSR_PRIV) != 0);
3330 		aflt->flt_panic = 1;
3331 		ch_flt.afsr_ext = cpu_error_regs.afsr_ext;
3332 		ch_flt.afsr_errs = afsr_errs;
3333 #if defined(SERRANO)
3334 		ch_flt.afar2 = cpu_error_regs.afar2;
3335 #endif	/* SERRANO */
3336 		(void) cpu_queue_events(&ch_flt, NULL, afsr_errs, NULL);
3337 	}
3338 }
3339 
3340 /*
3341  * Routine to convert a syndrome into a syndrome code.
3342  */
3343 static int
3344 synd_to_synd_code(int synd_status, ushort_t synd, uint64_t afsr_bit)
3345 {
3346 	if (synd_status == AFLT_STAT_INVALID)
3347 		return (-1);
3348 
3349 	/*
3350 	 * Use the syndrome to index the appropriate syndrome table,
3351 	 * to get the code indicating which bit(s) is(are) bad.
3352 	 */
3353 	if (afsr_bit &
3354 	    (C_AFSR_MSYND_ERRS | C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) {
3355 		if (afsr_bit & C_AFSR_MSYND_ERRS) {
3356 #if defined(JALAPENO) || defined(SERRANO)
3357 			if ((synd == 0) || (synd >= BSYND_TBL_SIZE))
3358 				return (-1);
3359 			else
3360 				return (BPAR0 + synd);
3361 #else /* JALAPENO || SERRANO */
3362 			if ((synd == 0) || (synd >= MSYND_TBL_SIZE))
3363 				return (-1);
3364 			else
3365 				return (mtag_syndrome_tab[synd]);
3366 #endif /* JALAPENO || SERRANO */
3367 		} else {
3368 			if ((synd == 0) || (synd >= ESYND_TBL_SIZE))
3369 				return (-1);
3370 			else
3371 				return (ecc_syndrome_tab[synd]);
3372 		}
3373 	} else {
3374 		return (-1);
3375 	}
3376 }
3377 
3378 int
3379 cpu_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
3380 {
3381 	if (&plat_get_mem_sid)
3382 		return (plat_get_mem_sid(unum, buf, buflen, lenp));
3383 	else
3384 		return (ENOTSUP);
3385 }
3386 
3387 int
3388 cpu_get_mem_offset(uint64_t flt_addr, uint64_t *offp)
3389 {
3390 	if (&plat_get_mem_offset)
3391 		return (plat_get_mem_offset(flt_addr, offp));
3392 	else
3393 		return (ENOTSUP);
3394 }
3395 
3396 int
3397 cpu_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
3398 {
3399 	if (&plat_get_mem_addr)
3400 		return (plat_get_mem_addr(unum, sid, offset, addrp));
3401 	else
3402 		return (ENOTSUP);
3403 }
3404 
3405 /*
3406  * Routine to return a string identifying the physical name
3407  * associated with a memory/cache error.
3408  */
3409 int
3410 cpu_get_mem_unum(int synd_status, ushort_t flt_synd, uint64_t flt_stat,
3411     uint64_t flt_addr, int flt_bus_id, int flt_in_memory,
3412     ushort_t flt_status, char *buf, int buflen, int *lenp)
3413 {
3414 	int synd_code;
3415 	int ret;
3416 
3417 	/*
3418 	 * An AFSR of -1 defaults to a memory syndrome.
3419 	 */
3420 	if (flt_stat == (uint64_t)-1)
3421 		flt_stat = C_AFSR_CE;
3422 
3423 	synd_code = synd_to_synd_code(synd_status, flt_synd, flt_stat);
3424 
3425 	/*
3426 	 * Syndrome code must be either a single-bit error code
3427 	 * (0...143) or -1 for unum lookup.
3428 	 */
3429 	if (synd_code < 0 || synd_code >= M2)
3430 		synd_code = -1;
3431 	if (&plat_get_mem_unum) {
3432 		if ((ret = plat_get_mem_unum(synd_code, flt_addr, flt_bus_id,
3433 		    flt_in_memory, flt_status, buf, buflen, lenp)) != 0) {
3434 			buf[0] = '\0';
3435 			*lenp = 0;
3436 		}
3437 
3438 		return (ret);
3439 	}
3440 
3441 	return (ENOTSUP);
3442 }
3443 
3444 /*
3445  * Wrapper for cpu_get_mem_unum() routine that takes an
3446  * async_flt struct rather than explicit arguments.
3447  */
3448 int
3449 cpu_get_mem_unum_aflt(int synd_status, struct async_flt *aflt,
3450     char *buf, int buflen, int *lenp)
3451 {
3452 	/*
3453 	 * If we come thru here for an IO bus error aflt->flt_stat will
3454 	 * not be the CPU AFSR, and we pass in a -1 to cpu_get_mem_unum()
3455 	 * so it will interpret this as a memory error.
3456 	 */
3457 	return (cpu_get_mem_unum(synd_status, aflt->flt_synd,
3458 	    (aflt->flt_class == BUS_FAULT) ?
3459 	    (uint64_t)-1 : ((ch_async_flt_t *)aflt)->flt_bit,
3460 	    aflt->flt_addr, aflt->flt_bus_id, aflt->flt_in_memory,
3461 	    aflt->flt_status, buf, buflen, lenp));
3462 }
3463 
3464 /*
3465  * Return unum string given synd_code and async_flt into
3466  * the buf with size UNUM_NAMLEN
3467  */
3468 static int
3469 cpu_get_mem_unum_synd(int synd_code, struct async_flt *aflt, char *buf)
3470 {
3471 	int ret, len;
3472 
3473 	/*
3474 	 * Syndrome code must be either a single-bit error code
3475 	 * (0...143) or -1 for unum lookup.
3476 	 */
3477 	if (synd_code < 0 || synd_code >= M2)
3478 		synd_code = -1;
3479 	if (&plat_get_mem_unum) {
3480 		if ((ret = plat_get_mem_unum(synd_code, aflt->flt_addr,
3481 		    aflt->flt_bus_id, aflt->flt_in_memory,
3482 		    aflt->flt_status, buf, UNUM_NAMLEN, &len)) != 0) {
3483 			buf[0] = '\0';
3484 		}
3485 		return (ret);
3486 	}
3487 
3488 	buf[0] = '\0';
3489 	return (ENOTSUP);
3490 }
3491 
3492 /*
3493  * This routine is a more generic interface to cpu_get_mem_unum()
3494  * that may be used by other modules (e.g. the 'mm' driver, through
3495  * the 'MEM_NAME' ioctl, which is used by fmd to resolve unum's
3496  * for Jalapeno/Serrano FRC/RCE or FRU/RUE paired events).
3497  */
3498 int
3499 cpu_get_mem_name(uint64_t synd, uint64_t *afsr, uint64_t afar,
3500     char *buf, int buflen, int *lenp)
3501 {
3502 	int synd_status, flt_in_memory, ret;
3503 	ushort_t flt_status = 0;
3504 	char unum[UNUM_NAMLEN];
3505 	uint64_t t_afsr_errs;
3506 
3507 	/*
3508 	 * Check for an invalid address.
3509 	 */
3510 	if (afar == (uint64_t)-1)
3511 		return (ENXIO);
3512 
3513 	if (synd == (uint64_t)-1)
3514 		synd_status = AFLT_STAT_INVALID;
3515 	else
3516 		synd_status = AFLT_STAT_VALID;
3517 
3518 	flt_in_memory = (*afsr & C_AFSR_MEMORY) &&
3519 	    pf_is_memory(afar >> MMU_PAGESHIFT);
3520 
3521 	/*
3522 	 * Get aggregate AFSR for call to cpu_error_is_ecache_data.
3523 	 */
3524 	if (*afsr == (uint64_t)-1)
3525 		t_afsr_errs = C_AFSR_CE;
3526 	else {
3527 		t_afsr_errs = (*afsr & C_AFSR_ALL_ERRS);
3528 #if defined(CHEETAH_PLUS)
3529 		if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation))
3530 			t_afsr_errs |= (*(afsr + 1) & C_AFSR_EXT_ALL_ERRS);
3531 #endif	/* CHEETAH_PLUS */
3532 	}
3533 
3534 	/*
3535 	 * Turn on ECC_ECACHE if error type is E$ Data.
3536 	 */
3537 	if (cpu_error_is_ecache_data(CPU->cpu_id, t_afsr_errs))
3538 		flt_status |= ECC_ECACHE;
3539 
3540 	ret = cpu_get_mem_unum(synd_status, (ushort_t)synd, t_afsr_errs, afar,
3541 	    CPU->cpu_id, flt_in_memory, flt_status, unum, UNUM_NAMLEN, lenp);
3542 	if (ret != 0)
3543 		return (ret);
3544 
3545 	if (*lenp >= buflen)
3546 		return (ENAMETOOLONG);
3547 
3548 	(void) strncpy(buf, unum, buflen);
3549 
3550 	return (0);
3551 }
3552 
3553 /*
3554  * Routine to return memory information associated
3555  * with a physical address and syndrome.
3556  */
3557 int
3558 cpu_get_mem_info(uint64_t synd, uint64_t afar,
3559     uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
3560     int *segsp, int *banksp, int *mcidp)
3561 {
3562 	int synd_status, synd_code;
3563 
3564 	if (afar == (uint64_t)-1)
3565 		return (ENXIO);
3566 
3567 	if (synd == (uint64_t)-1)
3568 		synd_status = AFLT_STAT_INVALID;
3569 	else
3570 		synd_status = AFLT_STAT_VALID;
3571 
3572 	synd_code = synd_to_synd_code(synd_status, synd, C_AFSR_CE);
3573 
3574 	if (p2get_mem_info != NULL)
3575 		return ((p2get_mem_info)(synd_code, afar,
3576 			mem_sizep, seg_sizep, bank_sizep,
3577 			segsp, banksp, mcidp));
3578 	else
3579 		return (ENOTSUP);
3580 }
3581 
3582 /*
3583  * Routine to return a string identifying the physical
3584  * name associated with a cpuid.
3585  */
3586 int
3587 cpu_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
3588 {
3589 	int ret;
3590 	char unum[UNUM_NAMLEN];
3591 
3592 	if (&plat_get_cpu_unum) {
3593 		if ((ret = plat_get_cpu_unum(cpuid, unum, UNUM_NAMLEN, lenp))
3594 		    != 0)
3595 			return (ret);
3596 	} else {
3597 		return (ENOTSUP);
3598 	}
3599 
3600 	if (*lenp >= buflen)
3601 		return (ENAMETOOLONG);
3602 
3603 	(void) strncpy(buf, unum, buflen);
3604 
3605 	return (0);
3606 }
3607 
3608 /*
3609  * This routine exports the name buffer size.
3610  */
3611 size_t
3612 cpu_get_name_bufsize()
3613 {
3614 	return (UNUM_NAMLEN);
3615 }
3616 
3617 /*
3618  * Historical function, apparantly not used.
3619  */
3620 /* ARGSUSED */
3621 void
3622 cpu_read_paddr(struct async_flt *ecc, short verbose, short ce_err)
3623 {}
3624 
3625 /*
3626  * Historical function only called for SBus errors in debugging.
3627  */
3628 /*ARGSUSED*/
3629 void
3630 read_ecc_data(struct async_flt *aflt, short verbose, short ce_err)
3631 {}
3632 
3633 /*
3634  * Clear the AFSR sticky bits.  The routine returns a non-zero value if
3635  * any of the AFSR's sticky errors are detected.  If a non-null pointer to
3636  * an async fault structure argument is passed in, the captured error state
3637  * (AFSR, AFAR) info will be returned in the structure.
3638  */
3639 int
3640 clear_errors(ch_async_flt_t *ch_flt)
3641 {
3642 	struct async_flt *aflt = (struct async_flt *)ch_flt;
3643 	ch_cpu_errors_t	cpu_error_regs;
3644 
3645 	get_cpu_error_state(&cpu_error_regs);
3646 
3647 	if (ch_flt != NULL) {
3648 		aflt->flt_stat = cpu_error_regs.afsr & C_AFSR_MASK;
3649 		aflt->flt_addr = cpu_error_regs.afar;
3650 		ch_flt->afsr_ext = cpu_error_regs.afsr_ext;
3651 		ch_flt->afsr_errs = (cpu_error_regs.afsr & C_AFSR_ALL_ERRS) |
3652 		    (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS);
3653 #if defined(SERRANO)
3654 		ch_flt->afar2 = cpu_error_regs.afar2;
3655 #endif	/* SERRANO */
3656 	}
3657 
3658 	set_cpu_error_state(&cpu_error_regs);
3659 
3660 	return (((cpu_error_regs.afsr & C_AFSR_ALL_ERRS) |
3661 	    (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS)) != 0);
3662 }
3663 
3664 /*
3665  * Clear any AFSR error bits, and check for persistence.
3666  *
3667  * It would be desirable to also insist that syndrome match.  PCI handling
3668  * has already filled flt_synd.  For errors trapped by CPU we only fill
3669  * flt_synd when we queue the event, so we do not have a valid flt_synd
3670  * during initial classification (it is valid if we're called as part of
3671  * subsequent low-pil additional classification attempts).  We could try
3672  * to determine which syndrome to use: we know we're only called for
3673  * CE/RCE (Jalapeno & Serrano) and CE/EMC (others) so the syndrome to use
3674  * would be esynd/none and esynd/msynd, respectively.  If that is
3675  * implemented then what do we do in the case that we do experience an
3676  * error on the same afar but with different syndrome?  At the very least
3677  * we should count such occurences.  Anyway, for now, we'll leave it as
3678  * it has been for ages.
3679  */
3680 static int
3681 clear_ecc(struct async_flt *aflt)
3682 {
3683 	ch_cpu_errors_t	cpu_error_regs;
3684 
3685 	/*
3686 	 * Snapshot the AFSR and AFAR and clear any errors
3687 	 */
3688 	get_cpu_error_state(&cpu_error_regs);
3689 	set_cpu_error_state(&cpu_error_regs);
3690 
3691 	/*
3692 	 * If any of the same memory access error bits are still on and
3693 	 * the AFAR matches, return that the error is persistent.
3694 	 */
3695 	return ((cpu_error_regs.afsr & (C_AFSR_MEMORY & aflt->flt_stat)) != 0 &&
3696 	    cpu_error_regs.afar == aflt->flt_addr);
3697 }
3698 
3699 /*
3700  * Turn off all cpu error detection, normally only used for panics.
3701  */
3702 void
3703 cpu_disable_errors(void)
3704 {
3705 	xt_all(set_error_enable_tl1, EN_REG_DISABLE, EER_SET_ABSOLUTE);
3706 
3707 	/*
3708 	 * With error detection now turned off, check the other cpus
3709 	 * logout areas for any unlogged errors.
3710 	 */
3711 	if (enable_check_other_cpus_logout) {
3712 		cpu_check_other_cpus_logout();
3713 		/*
3714 		 * Make a second pass over the logout areas, in case
3715 		 * there is a failing CPU in an error-trap loop which
3716 		 * will write to the logout area once it is emptied.
3717 		 */
3718 		cpu_check_other_cpus_logout();
3719 	}
3720 }
3721 
3722 /*
3723  * Enable errors.
3724  */
3725 void
3726 cpu_enable_errors(void)
3727 {
3728 	xt_all(set_error_enable_tl1, EN_REG_ENABLE, EER_SET_ABSOLUTE);
3729 }
3730 
3731 /*
3732  * Flush the entire ecache using displacement flush by reading through a
3733  * physical address range twice as large as the Ecache.
3734  */
3735 void
3736 cpu_flush_ecache(void)
3737 {
3738 	flush_ecache(ecache_flushaddr, cpunodes[CPU->cpu_id].ecache_size,
3739 	    cpunodes[CPU->cpu_id].ecache_linesize);
3740 }
3741 
3742 /*
3743  * Return CPU E$ set size - E$ size divided by the associativity.
3744  * We use this function in places where the CPU_PRIVATE ptr may not be
3745  * initialized yet.  Note that for send_mondo and in the Ecache scrubber,
3746  * we're guaranteed that CPU_PRIVATE is initialized.  Also, cpunodes is set
3747  * up before the kernel switches from OBP's to the kernel's trap table, so
3748  * we don't have to worry about cpunodes being unitialized.
3749  */
3750 int
3751 cpu_ecache_set_size(struct cpu *cp)
3752 {
3753 	if (CPU_PRIVATE(cp))
3754 		return (CPU_PRIVATE_VAL(cp, chpr_ec_set_size));
3755 
3756 	return (cpunodes[cp->cpu_id].ecache_size / cpu_ecache_nway());
3757 }
3758 
3759 /*
3760  * Flush Ecache line.
3761  * Uses ASI_EC_DIAG for Cheetah+ and Jalapeno.
3762  * Uses normal displacement flush for Cheetah.
3763  */
3764 static void
3765 cpu_flush_ecache_line(ch_async_flt_t *ch_flt)
3766 {
3767 	struct async_flt *aflt = (struct async_flt *)ch_flt;
3768 	int ec_set_size = cpu_ecache_set_size(CPU);
3769 
3770 	ecache_flush_line(aflt->flt_addr, ec_set_size);
3771 }
3772 
3773 /*
3774  * Scrub physical address.
3775  * Scrub code is different depending upon whether this a Cheetah+ with 2-way
3776  * Ecache or direct-mapped Ecache.
3777  */
3778 static void
3779 cpu_scrubphys(struct async_flt *aflt)
3780 {
3781 	int ec_set_size = cpu_ecache_set_size(CPU);
3782 
3783 	scrubphys(aflt->flt_addr, ec_set_size);
3784 }
3785 
3786 /*
3787  * Clear physical address.
3788  * Scrub code is different depending upon whether this a Cheetah+ with 2-way
3789  * Ecache or direct-mapped Ecache.
3790  */
3791 void
3792 cpu_clearphys(struct async_flt *aflt)
3793 {
3794 	int lsize = cpunodes[CPU->cpu_id].ecache_linesize;
3795 	int ec_set_size = cpu_ecache_set_size(CPU);
3796 
3797 
3798 	clearphys(P2ALIGN(aflt->flt_addr, lsize), ec_set_size, lsize);
3799 }
3800 
3801 #if defined(CPU_IMP_ECACHE_ASSOC)
3802 /*
3803  * Check for a matching valid line in all the sets.
3804  * If found, return set# + 1. Otherwise return 0.
3805  */
3806 static int
3807 cpu_ecache_line_valid(ch_async_flt_t *ch_flt)
3808 {
3809 	struct async_flt *aflt = (struct async_flt *)ch_flt;
3810 	int totalsize = cpunodes[CPU->cpu_id].ecache_size;
3811 	int ec_set_size = cpu_ecache_set_size(CPU);
3812 	ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0];
3813 	int nway = cpu_ecache_nway();
3814 	int i;
3815 
3816 	for (i = 0; i < nway; i++, ecp++) {
3817 		if (!cpu_ectag_line_invalid(totalsize, ecp->ec_tag) &&
3818 		    (aflt->flt_addr & P2ALIGN(C_AFAR_PA, ec_set_size)) ==
3819 		    cpu_ectag_to_pa(ec_set_size, ecp->ec_tag))
3820 			return (i+1);
3821 	}
3822 	return (0);
3823 }
3824 #endif /* CPU_IMP_ECACHE_ASSOC */
3825 
3826 /*
3827  * Check whether a line in the given logout info matches the specified
3828  * fault address.  If reqval is set then the line must not be Invalid.
3829  * Returns 0 on failure;  on success (way + 1) is returned an *level is
3830  * set to 2 for l2$ or 3 for l3$.
3831  */
3832 static int
3833 cpu_matching_ecache_line(uint64_t faddr, void *data, int reqval, int *level)
3834 {
3835 	ch_diag_data_t *cdp = data;
3836 	ch_ec_data_t *ecp;
3837 	int totalsize, ec_set_size;
3838 	int i, ways;
3839 	int match = 0;
3840 	int tagvalid;
3841 	uint64_t addr, tagpa;
3842 	int ispanther = IS_PANTHER(cpunodes[CPU->cpu_id].implementation);
3843 
3844 	/*
3845 	 * Check the l2$ logout data
3846 	 */
3847 	if (ispanther) {
3848 		ecp = &cdp->chd_l2_data[0];
3849 		ec_set_size = PN_L2_SET_SIZE;
3850 		ways = PN_L2_NWAYS;
3851 	} else {
3852 		ecp = &cdp->chd_ec_data[0];
3853 		ec_set_size = cpu_ecache_set_size(CPU);
3854 		ways = cpu_ecache_nway();
3855 		totalsize = cpunodes[CPU->cpu_id].ecache_size;
3856 	}
3857 	/* remove low order PA bits from fault address not used in PA tag */
3858 	addr = faddr & P2ALIGN(C_AFAR_PA, ec_set_size);
3859 	for (i = 0; i < ways; i++, ecp++) {
3860 		if (ispanther) {
3861 			tagpa = PN_L2TAG_TO_PA(ecp->ec_tag);
3862 			tagvalid = !PN_L2_LINE_INVALID(ecp->ec_tag);
3863 		} else {
3864 			tagpa = cpu_ectag_to_pa(ec_set_size, ecp->ec_tag);
3865 			tagvalid = !cpu_ectag_line_invalid(totalsize,
3866 			    ecp->ec_tag);
3867 		}
3868 		if (tagpa == addr && (!reqval || tagvalid)) {
3869 			match = i + 1;
3870 			*level = 2;
3871 			break;
3872 		}
3873 	}
3874 
3875 	if (match || !ispanther)
3876 		return (match);
3877 
3878 	/* For Panther we also check the l3$ */
3879 	ecp = &cdp->chd_ec_data[0];
3880 	ec_set_size = PN_L3_SET_SIZE;
3881 	ways = PN_L3_NWAYS;
3882 	addr = faddr & P2ALIGN(C_AFAR_PA, ec_set_size);
3883 
3884 	for (i = 0; i < ways; i++, ecp++) {
3885 		if (PN_L3TAG_TO_PA(ecp->ec_tag) == addr && (!reqval ||
3886 		    !PN_L3_LINE_INVALID(ecp->ec_tag))) {
3887 			match = i + 1;
3888 			*level = 3;
3889 			break;
3890 		}
3891 	}
3892 
3893 	return (match);
3894 }
3895 
3896 #if defined(CPU_IMP_L1_CACHE_PARITY)
3897 /*
3898  * Record information related to the source of an Dcache Parity Error.
3899  */
3900 static void
3901 cpu_dcache_parity_info(ch_async_flt_t *ch_flt)
3902 {
3903 	int dc_set_size = dcache_size / CH_DCACHE_NWAY;
3904 	int index;
3905 
3906 	/*
3907 	 * Since instruction decode cannot be done at high PIL
3908 	 * just examine the entire Dcache to locate the error.
3909 	 */
3910 	if (ch_flt->parity_data.dpe.cpl_lcnt == 0) {
3911 		ch_flt->parity_data.dpe.cpl_way = -1;
3912 		ch_flt->parity_data.dpe.cpl_off = -1;
3913 	}
3914 	for (index = 0; index < dc_set_size; index += dcache_linesize)
3915 		cpu_dcache_parity_check(ch_flt, index);
3916 }
3917 
3918 /*
3919  * Check all ways of the Dcache at a specified index for good parity.
3920  */
3921 static void
3922 cpu_dcache_parity_check(ch_async_flt_t *ch_flt, int index)
3923 {
3924 	int dc_set_size = dcache_size / CH_DCACHE_NWAY;
3925 	uint64_t parity_bits, pbits, data_word;
3926 	static int parity_bits_popc[] = { 0, 1, 1, 0 };
3927 	int way, word, data_byte;
3928 	ch_dc_data_t *dcp = &ch_flt->parity_data.dpe.cpl_dc[0];
3929 	ch_dc_data_t tmp_dcp;
3930 
3931 	for (way = 0; way < CH_DCACHE_NWAY; way++, dcp++) {
3932 		/*
3933 		 * Perform diagnostic read.
3934 		 */
3935 		get_dcache_dtag(index + way * dc_set_size,
3936 				(uint64_t *)&tmp_dcp);
3937 
3938 		/*
3939 		 * Check tag for even parity.
3940 		 * Sum of 1 bits (including parity bit) should be even.
3941 		 */
3942 		if (popc64(tmp_dcp.dc_tag & CHP_DCTAG_PARMASK) & 1) {
3943 			/*
3944 			 * If this is the first error log detailed information
3945 			 * about it and check the snoop tag. Otherwise just
3946 			 * record the fact that we found another error.
3947 			 */
3948 			if (ch_flt->parity_data.dpe.cpl_lcnt == 0) {
3949 				ch_flt->parity_data.dpe.cpl_way = way;
3950 				ch_flt->parity_data.dpe.cpl_cache =
3951 				    CPU_DC_PARITY;
3952 				ch_flt->parity_data.dpe.cpl_tag |= CHP_DC_TAG;
3953 
3954 				if (popc64(tmp_dcp.dc_sntag &
3955 						CHP_DCSNTAG_PARMASK) & 1) {
3956 					ch_flt->parity_data.dpe.cpl_tag |=
3957 								CHP_DC_SNTAG;
3958 					ch_flt->parity_data.dpe.cpl_lcnt++;
3959 				}
3960 
3961 				bcopy(&tmp_dcp, dcp, sizeof (ch_dc_data_t));
3962 			}
3963 
3964 			ch_flt->parity_data.dpe.cpl_lcnt++;
3965 		}
3966 
3967 		if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
3968 			/*
3969 			 * Panther has more parity bits than the other
3970 			 * processors for covering dcache data and so each
3971 			 * byte of data in each word has its own parity bit.
3972 			 */
3973 			parity_bits = tmp_dcp.dc_pn_data_parity;
3974 			for (word = 0; word < 4; word++) {
3975 				data_word = tmp_dcp.dc_data[word];
3976 				pbits = parity_bits & PN_DC_DATA_PARITY_MASK;
3977 				for (data_byte = 0; data_byte < 8;
3978 				    data_byte++) {
3979 					if (((popc64(data_word &
3980 					    PN_DC_DATA_PARITY_MASK)) & 1) ^
3981 					    (pbits & 1)) {
3982 						cpu_record_dc_data_parity(
3983 						ch_flt, dcp, &tmp_dcp, way,
3984 						word);
3985 					}
3986 					pbits >>= 1;
3987 					data_word >>= 8;
3988 				}
3989 				parity_bits >>= 8;
3990 			}
3991 		} else {
3992 			/*
3993 			 * Check data array for even parity.
3994 			 * The 8 parity bits are grouped into 4 pairs each
3995 			 * of which covers a 64-bit word.  The endianness is
3996 			 * reversed -- the low-order parity bits cover the
3997 			 * high-order data words.
3998 			 */
3999 			parity_bits = tmp_dcp.dc_utag >> 8;
4000 			for (word = 0; word < 4; word++) {
4001 				pbits = (parity_bits >> (6 - word * 2)) & 3;
4002 				if ((popc64(tmp_dcp.dc_data[word]) +
4003 				    parity_bits_popc[pbits]) & 1) {
4004 					cpu_record_dc_data_parity(ch_flt, dcp,
4005 					    &tmp_dcp, way, word);
4006 				}
4007 			}
4008 		}
4009 	}
4010 }
4011 
4012 static void
4013 cpu_record_dc_data_parity(ch_async_flt_t *ch_flt,
4014     ch_dc_data_t *dest_dcp, ch_dc_data_t *src_dcp, int way, int word)
4015 {
4016 	/*
4017 	 * If this is the first error log detailed information about it.
4018 	 * Otherwise just record the fact that we found another error.
4019 	 */
4020 	if (ch_flt->parity_data.dpe.cpl_lcnt == 0) {
4021 		ch_flt->parity_data.dpe.cpl_way = way;
4022 		ch_flt->parity_data.dpe.cpl_cache = CPU_DC_PARITY;
4023 		ch_flt->parity_data.dpe.cpl_off = word * 8;
4024 		bcopy(src_dcp, dest_dcp, sizeof (ch_dc_data_t));
4025 	}
4026 	ch_flt->parity_data.dpe.cpl_lcnt++;
4027 }
4028 
4029 /*
4030  * Record information related to the source of an Icache Parity Error.
4031  *
4032  * Called with the Icache disabled so any diagnostic accesses are safe.
4033  */
4034 static void
4035 cpu_icache_parity_info(ch_async_flt_t *ch_flt)
4036 {
4037 	int	ic_set_size;
4038 	int	ic_linesize;
4039 	int	index;
4040 
4041 	if (CPU_PRIVATE(CPU)) {
4042 		ic_set_size = CPU_PRIVATE_VAL(CPU, chpr_icache_size) /
4043 		    CH_ICACHE_NWAY;
4044 		ic_linesize = CPU_PRIVATE_VAL(CPU, chpr_icache_linesize);
4045 	} else {
4046 		ic_set_size = icache_size / CH_ICACHE_NWAY;
4047 		ic_linesize = icache_linesize;
4048 	}
4049 
4050 	ch_flt->parity_data.ipe.cpl_way = -1;
4051 	ch_flt->parity_data.ipe.cpl_off = -1;
4052 
4053 	for (index = 0; index < ic_set_size; index += ic_linesize)
4054 		cpu_icache_parity_check(ch_flt, index);
4055 }
4056 
4057 /*
4058  * Check all ways of the Icache at a specified index for good parity.
4059  */
4060 static void
4061 cpu_icache_parity_check(ch_async_flt_t *ch_flt, int index)
4062 {
4063 	uint64_t parmask, pn_inst_parity;
4064 	int ic_set_size;
4065 	int ic_linesize;
4066 	int flt_index, way, instr, num_instr;
4067 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4068 	ch_ic_data_t *icp = &ch_flt->parity_data.ipe.cpl_ic[0];
4069 	ch_ic_data_t tmp_icp;
4070 
4071 	if (CPU_PRIVATE(CPU)) {
4072 		ic_set_size = CPU_PRIVATE_VAL(CPU, chpr_icache_size) /
4073 		    CH_ICACHE_NWAY;
4074 		ic_linesize = CPU_PRIVATE_VAL(CPU, chpr_icache_linesize);
4075 	} else {
4076 		ic_set_size = icache_size / CH_ICACHE_NWAY;
4077 		ic_linesize = icache_linesize;
4078 	}
4079 
4080 	/*
4081 	 * Panther has twice as many instructions per icache line and the
4082 	 * instruction parity bit is in a different location.
4083 	 */
4084 	if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
4085 		num_instr = PN_IC_DATA_REG_SIZE / sizeof (uint64_t);
4086 		pn_inst_parity = PN_ICDATA_PARITY_BIT_MASK;
4087 	} else {
4088 		num_instr = CH_IC_DATA_REG_SIZE / sizeof (uint64_t);
4089 		pn_inst_parity = 0;
4090 	}
4091 
4092 	/*
4093 	 * Index at which we expect to find the parity error.
4094 	 */
4095 	flt_index = P2ALIGN(aflt->flt_addr % ic_set_size, ic_linesize);
4096 
4097 	for (way = 0; way < CH_ICACHE_NWAY; way++, icp++) {
4098 		/*
4099 		 * Diagnostic reads expect address argument in ASI format.
4100 		 */
4101 		get_icache_dtag(2 * (index + way * ic_set_size),
4102 				(uint64_t *)&tmp_icp);
4103 
4104 		/*
4105 		 * If this is the index in which we expect to find the
4106 		 * error log detailed information about each of the ways.
4107 		 * This information will be displayed later if we can't
4108 		 * determine the exact way in which the error is located.
4109 		 */
4110 		if (flt_index == index)
4111 			bcopy(&tmp_icp, icp, sizeof (ch_ic_data_t));
4112 
4113 		/*
4114 		 * Check tag for even parity.
4115 		 * Sum of 1 bits (including parity bit) should be even.
4116 		 */
4117 		if (popc64(tmp_icp.ic_patag & CHP_ICPATAG_PARMASK) & 1) {
4118 			/*
4119 			 * If this way is the one in which we expected
4120 			 * to find the error record the way and check the
4121 			 * snoop tag. Otherwise just record the fact we
4122 			 * found another error.
4123 			 */
4124 			if (flt_index == index) {
4125 				ch_flt->parity_data.ipe.cpl_way = way;
4126 				ch_flt->parity_data.ipe.cpl_tag |= CHP_IC_TAG;
4127 
4128 				if (popc64(tmp_icp.ic_sntag &
4129 						CHP_ICSNTAG_PARMASK) & 1) {
4130 					ch_flt->parity_data.ipe.cpl_tag |=
4131 								CHP_IC_SNTAG;
4132 					ch_flt->parity_data.ipe.cpl_lcnt++;
4133 				}
4134 
4135 			}
4136 			ch_flt->parity_data.ipe.cpl_lcnt++;
4137 			continue;
4138 		}
4139 
4140 		/*
4141 		 * Check instruction data for even parity.
4142 		 * Bits participating in parity differ for PC-relative
4143 		 * versus non-PC-relative instructions.
4144 		 */
4145 		for (instr = 0; instr < num_instr; instr++) {
4146 			parmask = (tmp_icp.ic_data[instr] &
4147 					CH_ICDATA_PRED_ISPCREL) ?
4148 				(CHP_ICDATA_PCREL_PARMASK | pn_inst_parity) :
4149 				(CHP_ICDATA_NPCREL_PARMASK | pn_inst_parity);
4150 			if (popc64(tmp_icp.ic_data[instr] & parmask) & 1) {
4151 				/*
4152 				 * If this way is the one in which we expected
4153 				 * to find the error record the way and offset.
4154 				 * Otherwise just log the fact we found another
4155 				 * error.
4156 				 */
4157 				if (flt_index == index) {
4158 					ch_flt->parity_data.ipe.cpl_way = way;
4159 					ch_flt->parity_data.ipe.cpl_off =
4160 								instr * 4;
4161 				}
4162 				ch_flt->parity_data.ipe.cpl_lcnt++;
4163 				continue;
4164 			}
4165 		}
4166 	}
4167 }
4168 
4169 /*
4170  * Record information related to the source of an Pcache Parity Error.
4171  */
4172 static void
4173 cpu_pcache_parity_info(ch_async_flt_t *ch_flt)
4174 {
4175 	int pc_set_size = CH_PCACHE_SIZE / CH_PCACHE_NWAY;
4176 	int index;
4177 
4178 	/*
4179 	 * Since instruction decode cannot be done at high PIL just
4180 	 * examine the entire Pcache to check for any parity errors.
4181 	 */
4182 	if (ch_flt->parity_data.dpe.cpl_lcnt == 0) {
4183 		ch_flt->parity_data.dpe.cpl_way = -1;
4184 		ch_flt->parity_data.dpe.cpl_off = -1;
4185 	}
4186 	for (index = 0; index < pc_set_size; index += CH_PCACHE_LSIZE)
4187 		cpu_pcache_parity_check(ch_flt, index);
4188 }
4189 
4190 /*
4191  * Check all ways of the Pcache at a specified index for good parity.
4192  */
4193 static void
4194 cpu_pcache_parity_check(ch_async_flt_t *ch_flt, int index)
4195 {
4196 	int pc_set_size = CH_PCACHE_SIZE / CH_PCACHE_NWAY;
4197 	int pc_data_words = CH_PC_DATA_REG_SIZE / sizeof (uint64_t);
4198 	int way, word, pbit, parity_bits;
4199 	ch_pc_data_t *pcp = &ch_flt->parity_data.dpe.cpl_pc[0];
4200 	ch_pc_data_t tmp_pcp;
4201 
4202 	for (way = 0; way < CH_PCACHE_NWAY; way++, pcp++) {
4203 		/*
4204 		 * Perform diagnostic read.
4205 		 */
4206 		get_pcache_dtag(index + way * pc_set_size,
4207 				(uint64_t *)&tmp_pcp);
4208 		/*
4209 		 * Check data array for odd parity. There are 8 parity
4210 		 * bits (bits 57:50 of ASI_PCACHE_STATUS_DATA) and each
4211 		 * of those bits covers exactly 8 bytes of the data
4212 		 * array:
4213 		 *
4214 		 *	parity bit	P$ data bytes covered
4215 		 *	----------	---------------------
4216 		 *	50		63:56
4217 		 *	51		55:48
4218 		 *	52		47:40
4219 		 *	53		39:32
4220 		 *	54		31:24
4221 		 *	55		23:16
4222 		 *	56		15:8
4223 		 *	57		7:0
4224 		 */
4225 		parity_bits = PN_PC_PARITY_BITS(tmp_pcp.pc_status);
4226 		for (word = 0; word < pc_data_words; word++) {
4227 			pbit = (parity_bits >> (pc_data_words - word - 1)) & 1;
4228 			if ((popc64(tmp_pcp.pc_data[word]) & 1) ^ pbit) {
4229 				/*
4230 				 * If this is the first error log detailed
4231 				 * information about it. Otherwise just record
4232 				 * the fact that we found another error.
4233 				 */
4234 				if (ch_flt->parity_data.dpe.cpl_lcnt == 0) {
4235 					ch_flt->parity_data.dpe.cpl_way = way;
4236 					ch_flt->parity_data.dpe.cpl_cache =
4237 					    CPU_PC_PARITY;
4238 					ch_flt->parity_data.dpe.cpl_off =
4239 					    word * sizeof (uint64_t);
4240 					bcopy(&tmp_pcp, pcp,
4241 							sizeof (ch_pc_data_t));
4242 				}
4243 				ch_flt->parity_data.dpe.cpl_lcnt++;
4244 			}
4245 		}
4246 	}
4247 }
4248 
4249 
4250 /*
4251  * Add L1 Data cache data to the ereport payload.
4252  */
4253 static void
4254 cpu_payload_add_dcache(struct async_flt *aflt, nvlist_t *nvl)
4255 {
4256 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4257 	ch_dc_data_t *dcp;
4258 	ch_dc_data_t dcdata[CH_DCACHE_NWAY];
4259 	uint_t nelem;
4260 	int i, ways_to_check, ways_logged = 0;
4261 
4262 	/*
4263 	 * If this is an D$ fault then there may be multiple
4264 	 * ways captured in the ch_parity_log_t structure.
4265 	 * Otherwise, there will be at most one way captured
4266 	 * in the ch_diag_data_t struct.
4267 	 * Check each way to see if it should be encoded.
4268 	 */
4269 	if (ch_flt->flt_type == CPU_DC_PARITY)
4270 		ways_to_check = CH_DCACHE_NWAY;
4271 	else
4272 		ways_to_check = 1;
4273 	for (i = 0; i < ways_to_check; i++) {
4274 		if (ch_flt->flt_type == CPU_DC_PARITY)
4275 			dcp = &ch_flt->parity_data.dpe.cpl_dc[i];
4276 		else
4277 			dcp = &ch_flt->flt_diag_data.chd_dc_data;
4278 		if (dcp->dc_logflag == DC_LOGFLAG_MAGIC) {
4279 			bcopy(dcp, &dcdata[ways_logged],
4280 				sizeof (ch_dc_data_t));
4281 			ways_logged++;
4282 		}
4283 	}
4284 
4285 	/*
4286 	 * Add the dcache data to the payload.
4287 	 */
4288 	fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1D_WAYS,
4289 	    DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL);
4290 	if (ways_logged != 0) {
4291 		nelem = sizeof (ch_dc_data_t) / sizeof (uint64_t) * ways_logged;
4292 		fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1D_DATA,
4293 		    DATA_TYPE_UINT64_ARRAY, nelem, (uint64_t *)dcdata, NULL);
4294 	}
4295 }
4296 
4297 /*
4298  * Add L1 Instruction cache data to the ereport payload.
4299  */
4300 static void
4301 cpu_payload_add_icache(struct async_flt *aflt, nvlist_t *nvl)
4302 {
4303 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4304 	ch_ic_data_t *icp;
4305 	ch_ic_data_t icdata[CH_ICACHE_NWAY];
4306 	uint_t nelem;
4307 	int i, ways_to_check, ways_logged = 0;
4308 
4309 	/*
4310 	 * If this is an I$ fault then there may be multiple
4311 	 * ways captured in the ch_parity_log_t structure.
4312 	 * Otherwise, there will be at most one way captured
4313 	 * in the ch_diag_data_t struct.
4314 	 * Check each way to see if it should be encoded.
4315 	 */
4316 	if (ch_flt->flt_type == CPU_IC_PARITY)
4317 		ways_to_check = CH_ICACHE_NWAY;
4318 	else
4319 		ways_to_check = 1;
4320 	for (i = 0; i < ways_to_check; i++) {
4321 		if (ch_flt->flt_type == CPU_IC_PARITY)
4322 			icp = &ch_flt->parity_data.ipe.cpl_ic[i];
4323 		else
4324 			icp = &ch_flt->flt_diag_data.chd_ic_data;
4325 		if (icp->ic_logflag == IC_LOGFLAG_MAGIC) {
4326 			bcopy(icp, &icdata[ways_logged],
4327 				sizeof (ch_ic_data_t));
4328 			ways_logged++;
4329 		}
4330 	}
4331 
4332 	/*
4333 	 * Add the icache data to the payload.
4334 	 */
4335 	fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1I_WAYS,
4336 	    DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL);
4337 	if (ways_logged != 0) {
4338 		nelem = sizeof (ch_ic_data_t) / sizeof (uint64_t) * ways_logged;
4339 		fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1I_DATA,
4340 		    DATA_TYPE_UINT64_ARRAY, nelem, (uint64_t *)icdata, NULL);
4341 	}
4342 }
4343 
4344 #endif	/* CPU_IMP_L1_CACHE_PARITY */
4345 
4346 /*
4347  * Add ecache data to payload.
4348  */
4349 static void
4350 cpu_payload_add_ecache(struct async_flt *aflt, nvlist_t *nvl)
4351 {
4352 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4353 	ch_ec_data_t *ecp;
4354 	ch_ec_data_t ecdata[CHD_EC_DATA_SETS];
4355 	uint_t nelem;
4356 	int i, ways_logged = 0;
4357 
4358 	/*
4359 	 * Check each way to see if it should be encoded
4360 	 * and concatinate it into a temporary buffer.
4361 	 */
4362 	for (i = 0; i < CHD_EC_DATA_SETS; i++) {
4363 		ecp = &ch_flt->flt_diag_data.chd_ec_data[i];
4364 		if (ecp->ec_logflag == EC_LOGFLAG_MAGIC) {
4365 			bcopy(ecp, &ecdata[ways_logged],
4366 				sizeof (ch_ec_data_t));
4367 			ways_logged++;
4368 		}
4369 	}
4370 
4371 	/*
4372 	 * Panther CPUs have an additional level of cache and so
4373 	 * what we just collected was the L3 (ecache) and not the
4374 	 * L2 cache.
4375 	 */
4376 	if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) {
4377 		/*
4378 		 * Add the L3 (ecache) data to the payload.
4379 		 */
4380 		fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L3_WAYS,
4381 		    DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL);
4382 		if (ways_logged != 0) {
4383 			nelem = sizeof (ch_ec_data_t) /
4384 			    sizeof (uint64_t) * ways_logged;
4385 			fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L3_DATA,
4386 			    DATA_TYPE_UINT64_ARRAY, nelem,
4387 			    (uint64_t *)ecdata, NULL);
4388 		}
4389 
4390 		/*
4391 		 * Now collect the L2 cache.
4392 		 */
4393 		ways_logged = 0;
4394 		for (i = 0; i < PN_L2_NWAYS; i++) {
4395 			ecp = &ch_flt->flt_diag_data.chd_l2_data[i];
4396 			if (ecp->ec_logflag == EC_LOGFLAG_MAGIC) {
4397 				bcopy(ecp, &ecdata[ways_logged],
4398 				    sizeof (ch_ec_data_t));
4399 				ways_logged++;
4400 			}
4401 		}
4402 	}
4403 
4404 	/*
4405 	 * Add the L2 cache data to the payload.
4406 	 */
4407 	fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L2_WAYS,
4408 	    DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL);
4409 	if (ways_logged != 0) {
4410 		nelem = sizeof (ch_ec_data_t) /
4411 			sizeof (uint64_t) * ways_logged;
4412 		fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L2_DATA,
4413 		    DATA_TYPE_UINT64_ARRAY, nelem,  (uint64_t *)ecdata, NULL);
4414 	}
4415 }
4416 
4417 /*
4418  * Initialize cpu scheme for specified cpu.
4419  */
4420 static void
4421 cpu_fmri_cpu_set(nvlist_t *cpu_fmri, int cpuid)
4422 {
4423 	char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */
4424 	uint8_t mask;
4425 
4426 	mask = cpunodes[cpuid].version;
4427 	(void) snprintf(sbuf, sizeof (sbuf), "%llX",
4428 	    (u_longlong_t)cpunodes[cpuid].device_id);
4429 	(void) fm_fmri_cpu_set(cpu_fmri, FM_CPU_SCHEME_VERSION, NULL,
4430 	    cpuid, &mask, (const char *)sbuf);
4431 }
4432 
4433 /*
4434  * Returns ereport resource type.
4435  */
4436 static int
4437 cpu_error_to_resource_type(struct async_flt *aflt)
4438 {
4439 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4440 
4441 	switch (ch_flt->flt_type) {
4442 
4443 	case CPU_CE_ECACHE:
4444 	case CPU_UE_ECACHE:
4445 	case CPU_UE_ECACHE_RETIRE:
4446 	case CPU_ORPH:
4447 		/*
4448 		 * If AFSR error bit indicates L2$ Data for Cheetah,
4449 		 * Cheetah+ or Jaguar, or L3$ Data for Panther, return
4450 		 * E$ Data type, otherwise, return CPU type.
4451 		 */
4452 		if (cpu_error_is_ecache_data(aflt->flt_inst,
4453 		    ch_flt->flt_bit))
4454 			return (ERRTYPE_ECACHE_DATA);
4455 		return (ERRTYPE_CPU);
4456 
4457 	case CPU_CE:
4458 	case CPU_UE:
4459 	case CPU_EMC:
4460 	case CPU_DUE:
4461 	case CPU_RCE:
4462 	case CPU_RUE:
4463 	case CPU_FRC:
4464 	case CPU_FRU:
4465 		return (ERRTYPE_MEMORY);
4466 
4467 	case CPU_IC_PARITY:
4468 	case CPU_DC_PARITY:
4469 	case CPU_FPUERR:
4470 	case CPU_PC_PARITY:
4471 	case CPU_ITLB_PARITY:
4472 	case CPU_DTLB_PARITY:
4473 		return (ERRTYPE_CPU);
4474 	}
4475 	return (ERRTYPE_UNKNOWN);
4476 }
4477 
4478 /*
4479  * Encode the data saved in the ch_async_flt_t struct into
4480  * the FM ereport payload.
4481  */
4482 static void
4483 cpu_payload_add_aflt(struct async_flt *aflt, nvlist_t *payload,
4484 	nvlist_t *resource, int *afar_status, int *synd_status)
4485 {
4486 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4487 	*synd_status = AFLT_STAT_INVALID;
4488 	*afar_status = AFLT_STAT_INVALID;
4489 
4490 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFSR) {
4491 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFSR,
4492 		    DATA_TYPE_UINT64, aflt->flt_stat, NULL);
4493 	}
4494 
4495 	if ((aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFSR_EXT) &&
4496 	    IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) {
4497 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFSR_EXT,
4498 		    DATA_TYPE_UINT64, ch_flt->afsr_ext, NULL);
4499 	}
4500 
4501 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFAR_STATUS) {
4502 		*afar_status = afsr_to_afar_status(ch_flt->afsr_errs,
4503 		    ch_flt->flt_bit);
4504 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFAR_STATUS,
4505 		    DATA_TYPE_UINT8, (uint8_t)*afar_status, NULL);
4506 	}
4507 
4508 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFAR) {
4509 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFAR,
4510 		    DATA_TYPE_UINT64, aflt->flt_addr, NULL);
4511 	}
4512 
4513 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PC) {
4514 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PC,
4515 		    DATA_TYPE_UINT64, (uint64_t)aflt->flt_pc, NULL);
4516 	}
4517 
4518 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TL) {
4519 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TL,
4520 		    DATA_TYPE_UINT8, (uint8_t)aflt->flt_tl, NULL);
4521 	}
4522 
4523 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TT) {
4524 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TT,
4525 		    DATA_TYPE_UINT8, flt_to_trap_type(aflt), NULL);
4526 	}
4527 
4528 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PRIV) {
4529 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PRIV,
4530 		    DATA_TYPE_BOOLEAN_VALUE,
4531 		    (aflt->flt_priv ? B_TRUE : B_FALSE), NULL);
4532 	}
4533 
4534 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ME) {
4535 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ME,
4536 		    DATA_TYPE_BOOLEAN_VALUE,
4537 		    (aflt->flt_stat & C_AFSR_ME) ? B_TRUE : B_FALSE, NULL);
4538 	}
4539 
4540 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SYND_STATUS) {
4541 		*synd_status = afsr_to_synd_status(aflt->flt_inst,
4542 		    ch_flt->afsr_errs, ch_flt->flt_bit);
4543 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND_STATUS,
4544 		    DATA_TYPE_UINT8, (uint8_t)*synd_status, NULL);
4545 	}
4546 
4547 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SYND) {
4548 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND,
4549 		    DATA_TYPE_UINT16, (uint16_t)aflt->flt_synd, NULL);
4550 	}
4551 
4552 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ERR_TYPE) {
4553 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ERR_TYPE,
4554 		    DATA_TYPE_STRING, flt_to_error_type(aflt), NULL);
4555 	}
4556 
4557 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ERR_DISP) {
4558 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ERR_DISP,
4559 		    DATA_TYPE_UINT64, aflt->flt_disp, NULL);
4560 	}
4561 
4562 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L2)
4563 		cpu_payload_add_ecache(aflt, payload);
4564 
4565 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_COPYFUNCTION) {
4566 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_COPYFUNCTION,
4567 		    DATA_TYPE_UINT8, (uint8_t)aflt->flt_status & 0xff, NULL);
4568 	}
4569 
4570 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_HOWDETECTED) {
4571 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_HOWDETECTED,
4572 		    DATA_TYPE_UINT8, (uint8_t)(aflt->flt_status >> 8), NULL);
4573 	}
4574 
4575 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_INSTRBLOCK) {
4576 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_INSTRBLOCK,
4577 		    DATA_TYPE_UINT32_ARRAY, 16,
4578 		    (uint32_t *)&ch_flt->flt_fpdata, NULL);
4579 	}
4580 
4581 #if defined(CPU_IMP_L1_CACHE_PARITY)
4582 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1D)
4583 		cpu_payload_add_dcache(aflt, payload);
4584 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1I)
4585 		cpu_payload_add_icache(aflt, payload);
4586 #endif	/* CPU_IMP_L1_CACHE_PARITY */
4587 
4588 #if defined(CHEETAH_PLUS)
4589 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1P)
4590 		cpu_payload_add_pcache(aflt, payload);
4591 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_TLB)
4592 		cpu_payload_add_tlb(aflt, payload);
4593 #endif	/* CHEETAH_PLUS */
4594 	/*
4595 	 * Create the FMRI that goes into the payload
4596 	 * and contains the unum info if necessary.
4597 	 */
4598 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_RESOURCE) {
4599 		char unum[UNUM_NAMLEN] = "";
4600 		char sid[DIMM_SERIAL_ID_LEN] = "";
4601 		int len, ret, rtype, synd_code;
4602 		uint64_t offset = (uint64_t)-1;
4603 
4604 		rtype = cpu_error_to_resource_type(aflt);
4605 		switch (rtype) {
4606 
4607 		case ERRTYPE_MEMORY:
4608 		case ERRTYPE_ECACHE_DATA:
4609 
4610 			/*
4611 			 * Memory errors, do unum lookup
4612 			 */
4613 			if (*afar_status == AFLT_STAT_INVALID)
4614 				break;
4615 
4616 			if (rtype == ERRTYPE_ECACHE_DATA)
4617 				aflt->flt_status |= ECC_ECACHE;
4618 			else
4619 				aflt->flt_status &= ~ECC_ECACHE;
4620 
4621 			synd_code = synd_to_synd_code(*synd_status,
4622 			    aflt->flt_synd, ch_flt->flt_bit);
4623 
4624 			if (cpu_get_mem_unum_synd(synd_code, aflt, unum) != 0)
4625 				break;
4626 
4627 			ret = cpu_get_mem_sid(unum, sid, DIMM_SERIAL_ID_LEN,
4628 			    &len);
4629 
4630 			if (ret == 0) {
4631 				(void) cpu_get_mem_offset(aflt->flt_addr,
4632 				    &offset);
4633 			}
4634 
4635 			fm_fmri_mem_set(resource, FM_MEM_SCHEME_VERSION,
4636 			    NULL, unum, (ret == 0) ? sid : NULL, offset);
4637 			fm_payload_set(payload,
4638 			    FM_EREPORT_PAYLOAD_NAME_RESOURCE,
4639 			    DATA_TYPE_NVLIST, resource, NULL);
4640 			break;
4641 
4642 		case ERRTYPE_CPU:
4643 			/*
4644 			 * On-board processor array error, add cpu resource.
4645 			 */
4646 			cpu_fmri_cpu_set(resource, aflt->flt_inst);
4647 			fm_payload_set(payload,
4648 			    FM_EREPORT_PAYLOAD_NAME_RESOURCE,
4649 			    DATA_TYPE_NVLIST, resource, NULL);
4650 			break;
4651 		}
4652 	}
4653 }
4654 
4655 /*
4656  * Initialize the way info if necessary.
4657  */
4658 void
4659 cpu_ereport_init(struct async_flt *aflt)
4660 {
4661 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
4662 	ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0];
4663 	ch_ec_data_t *l2p = &ch_flt->flt_diag_data.chd_l2_data[0];
4664 	int i;
4665 
4666 	/*
4667 	 * Initialize the info in the CPU logout structure.
4668 	 * The I$/D$ way information is not initialized here
4669 	 * since it is captured in the logout assembly code.
4670 	 */
4671 	for (i = 0; i < CHD_EC_DATA_SETS; i++)
4672 		(ecp + i)->ec_way = i;
4673 
4674 	for (i = 0; i < PN_L2_NWAYS; i++)
4675 		(l2p + i)->ec_way = i;
4676 }
4677 
4678 /*
4679  * Returns whether fault address is valid for this error bit and
4680  * whether the address is "in memory" (i.e. pf_is_memory returns 1).
4681  */
4682 int
4683 cpu_flt_in_memory(ch_async_flt_t *ch_flt, uint64_t t_afsr_bit)
4684 {
4685 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4686 
4687 	return ((t_afsr_bit & C_AFSR_MEMORY) &&
4688 	    afsr_to_afar_status(ch_flt->afsr_errs, t_afsr_bit) ==
4689 	    AFLT_STAT_VALID &&
4690 	    pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT));
4691 }
4692 
4693 /*
4694  * Returns whether fault address is valid based on the error bit for the
4695  * one event being queued and whether the address is "in memory".
4696  */
4697 static int
4698 cpu_flt_in_memory_one_event(ch_async_flt_t *ch_flt, uint64_t t_afsr_bit)
4699 {
4700 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4701 	int afar_status;
4702 	uint64_t afsr_errs, afsr_ow, *ow_bits;
4703 
4704 	if (!(t_afsr_bit & C_AFSR_MEMORY) ||
4705 	    !pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT))
4706 		return (0);
4707 
4708 	afsr_errs = ch_flt->afsr_errs;
4709 	afar_status = afsr_to_afar_status(afsr_errs, t_afsr_bit);
4710 
4711 	switch (afar_status) {
4712 	case AFLT_STAT_VALID:
4713 		return (1);
4714 
4715 	case AFLT_STAT_AMBIGUOUS:
4716 		/*
4717 		 * Status is ambiguous since another error bit (or bits)
4718 		 * of equal priority to the specified bit on in the afsr,
4719 		 * so check those bits. Return 1 only if the bits on in the
4720 		 * same class as the t_afsr_bit are also C_AFSR_MEMORY bits.
4721 		 * Otherwise not all the equal priority bits are for memory
4722 		 * errors, so return 0.
4723 		 */
4724 		ow_bits = afar_overwrite;
4725 		while ((afsr_ow = *ow_bits++) != 0) {
4726 			/*
4727 			 * Get other bits that are on in t_afsr_bit's priority
4728 			 * class to check for Memory Error bits only.
4729 			 */
4730 			if (afsr_ow & t_afsr_bit) {
4731 				if ((afsr_errs & afsr_ow) & ~C_AFSR_MEMORY)
4732 					return (0);
4733 				else
4734 					return (1);
4735 			}
4736 		}
4737 		/*FALLTHRU*/
4738 
4739 	default:
4740 		return (0);
4741 	}
4742 }
4743 
4744 static void
4745 cpu_log_diag_info(ch_async_flt_t *ch_flt)
4746 {
4747 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4748 	ch_dc_data_t *dcp = &ch_flt->flt_diag_data.chd_dc_data;
4749 	ch_ic_data_t *icp = &ch_flt->flt_diag_data.chd_ic_data;
4750 	ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0];
4751 #if defined(CPU_IMP_ECACHE_ASSOC)
4752 	int i, nway;
4753 #endif /* CPU_IMP_ECACHE_ASSOC */
4754 
4755 	/*
4756 	 * Check if the CPU log out captured was valid.
4757 	 */
4758 	if (ch_flt->flt_diag_data.chd_afar == LOGOUT_INVALID ||
4759 	    ch_flt->flt_data_incomplete)
4760 		return;
4761 
4762 #if defined(CPU_IMP_ECACHE_ASSOC)
4763 	nway = cpu_ecache_nway();
4764 	i =  cpu_ecache_line_valid(ch_flt);
4765 	if (i == 0 || i > nway) {
4766 		for (i = 0; i < nway; i++)
4767 			ecp[i].ec_logflag = EC_LOGFLAG_MAGIC;
4768 	} else
4769 		ecp[i - 1].ec_logflag = EC_LOGFLAG_MAGIC;
4770 #else /* CPU_IMP_ECACHE_ASSOC */
4771 	ecp->ec_logflag = EC_LOGFLAG_MAGIC;
4772 #endif /* CPU_IMP_ECACHE_ASSOC */
4773 
4774 #if defined(CHEETAH_PLUS)
4775 	pn_cpu_log_diag_l2_info(ch_flt);
4776 #endif /* CHEETAH_PLUS */
4777 
4778 	if (CH_DCTAG_MATCH(dcp->dc_tag, aflt->flt_addr)) {
4779 		dcp->dc_way = CH_DCIDX_TO_WAY(dcp->dc_idx);
4780 		dcp->dc_logflag = DC_LOGFLAG_MAGIC;
4781 	}
4782 
4783 	if (CH_ICTAG_MATCH(icp, aflt->flt_addr)) {
4784 		if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation))
4785 			icp->ic_way = PN_ICIDX_TO_WAY(icp->ic_idx);
4786 		else
4787 			icp->ic_way = CH_ICIDX_TO_WAY(icp->ic_idx);
4788 		icp->ic_logflag = IC_LOGFLAG_MAGIC;
4789 	}
4790 }
4791 
4792 /*
4793  * Cheetah ECC calculation.
4794  *
4795  * We only need to do the calculation on the data bits and can ignore check
4796  * bit and Mtag bit terms in the calculation.
4797  */
4798 static uint64_t ch_ecc_table[9][2] = {
4799 	/*
4800 	 * low order 64-bits   high-order 64-bits
4801 	 */
4802 	{ 0x46bffffeccd1177f, 0x488800022100014c },
4803 	{ 0x42fccc81331ff77f, 0x14424f1010249184 },
4804 	{ 0x8898827c222f1ffe, 0x22c1222808184aaf },
4805 	{ 0xf7632203e131ccf1, 0xe1241121848292b8 },
4806 	{ 0x7f5511421b113809, 0x901c88d84288aafe },
4807 	{ 0x1d49412184882487, 0x8f338c87c044c6ef },
4808 	{ 0xf552181014448344, 0x7ff8f4443e411911 },
4809 	{ 0x2189240808f24228, 0xfeeff8cc81333f42 },
4810 	{ 0x3280008440001112, 0xfee88b337ffffd62 },
4811 };
4812 
4813 /*
4814  * 64-bit population count, use well-known popcnt trick.
4815  * We could use the UltraSPARC V9 POPC instruction, but some
4816  * CPUs including Cheetahplus and Jaguar do not support that
4817  * instruction.
4818  */
4819 int
4820 popc64(uint64_t val)
4821 {
4822 	int cnt;
4823 
4824 	for (cnt = 0; val != 0; val &= val - 1)
4825 		cnt++;
4826 	return (cnt);
4827 }
4828 
4829 /*
4830  * Generate the 9 ECC bits for the 128-bit chunk based on the table above.
4831  * Note that xor'ing an odd number of 1 bits == 1 and xor'ing an even number
4832  * of 1 bits == 0, so we can just use the least significant bit of the popcnt
4833  * instead of doing all the xor's.
4834  */
4835 uint32_t
4836 us3_gen_ecc(uint64_t data_low, uint64_t data_high)
4837 {
4838 	int bitno, s;
4839 	int synd = 0;
4840 
4841 	for (bitno = 0; bitno < 9; bitno++) {
4842 		s = (popc64(data_low & ch_ecc_table[bitno][0]) +
4843 		    popc64(data_high & ch_ecc_table[bitno][1])) & 1;
4844 		synd |= (s << bitno);
4845 	}
4846 	return (synd);
4847 
4848 }
4849 
4850 /*
4851  * Queue one event based on ecc_type_to_info entry.  If the event has an AFT1
4852  * tag associated with it or is a fatal event (aflt_panic set), it is sent to
4853  * the UE event queue.  Otherwise it is dispatched to the CE event queue.
4854  */
4855 static void
4856 cpu_queue_one_event(ch_async_flt_t *ch_flt, char *reason,
4857     ecc_type_to_info_t *eccp, ch_diag_data_t *cdp)
4858 {
4859 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4860 
4861 	if (reason &&
4862 	    strlen(reason) + strlen(eccp->ec_reason) < MAX_REASON_STRING) {
4863 		(void) strcat(reason, eccp->ec_reason);
4864 	}
4865 
4866 	ch_flt->flt_bit = eccp->ec_afsr_bit;
4867 	ch_flt->flt_type = eccp->ec_flt_type;
4868 	if (cdp != NULL && cdp->chd_afar != LOGOUT_INVALID)
4869 		ch_flt->flt_diag_data = *cdp;
4870 	else
4871 		ch_flt->flt_diag_data.chd_afar = LOGOUT_INVALID;
4872 	aflt->flt_in_memory =
4873 	    cpu_flt_in_memory_one_event(ch_flt, ch_flt->flt_bit);
4874 
4875 	if (ch_flt->flt_bit & C_AFSR_MSYND_ERRS)
4876 		aflt->flt_synd = GET_M_SYND(aflt->flt_stat);
4877 	else if (ch_flt->flt_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS))
4878 		aflt->flt_synd = GET_E_SYND(aflt->flt_stat);
4879 	else
4880 		aflt->flt_synd = 0;
4881 
4882 	aflt->flt_payload = eccp->ec_err_payload;
4883 
4884 	if (aflt->flt_panic || (eccp->ec_afsr_bit &
4885 	    (C_AFSR_LEVEL1 | C_AFSR_EXT_LEVEL1)))
4886 		cpu_errorq_dispatch(eccp->ec_err_class,
4887 		    (void *)ch_flt, sizeof (ch_async_flt_t), ue_queue,
4888 		    aflt->flt_panic);
4889 	else
4890 		cpu_errorq_dispatch(eccp->ec_err_class,
4891 		    (void *)ch_flt, sizeof (ch_async_flt_t), ce_queue,
4892 		    aflt->flt_panic);
4893 }
4894 
4895 /*
4896  * Queue events on async event queue one event per error bit.  First we
4897  * queue the events that we "expect" for the given trap, then we queue events
4898  * that we may not expect.  Return number of events queued.
4899  */
4900 int
4901 cpu_queue_events(ch_async_flt_t *ch_flt, char *reason, uint64_t t_afsr_errs,
4902     ch_cpu_logout_t *clop)
4903 {
4904 	struct async_flt *aflt = (struct async_flt *)ch_flt;
4905 	ecc_type_to_info_t *eccp;
4906 	int nevents = 0;
4907 	uint64_t primary_afar = aflt->flt_addr, primary_afsr = aflt->flt_stat;
4908 #if defined(CHEETAH_PLUS)
4909 	uint64_t orig_t_afsr_errs;
4910 #endif
4911 	uint64_t primary_afsr_ext = ch_flt->afsr_ext;
4912 	uint64_t primary_afsr_errs = ch_flt->afsr_errs;
4913 	ch_diag_data_t *cdp = NULL;
4914 
4915 	t_afsr_errs &= ((C_AFSR_ALL_ERRS & ~C_AFSR_ME) | C_AFSR_EXT_ALL_ERRS);
4916 
4917 #if defined(CHEETAH_PLUS)
4918 	orig_t_afsr_errs = t_afsr_errs;
4919 
4920 	/*
4921 	 * For Cheetah+, log the shadow AFSR/AFAR bits first.
4922 	 */
4923 	if (clop != NULL) {
4924 		/*
4925 		 * Set the AFSR and AFAR fields to the shadow registers.  The
4926 		 * flt_addr and flt_stat fields will be reset to the primaries
4927 		 * below, but the sdw_addr and sdw_stat will stay as the
4928 		 * secondaries.
4929 		 */
4930 		cdp = &clop->clo_sdw_data;
4931 		aflt->flt_addr = ch_flt->flt_sdw_afar = cdp->chd_afar;
4932 		aflt->flt_stat = ch_flt->flt_sdw_afsr = cdp->chd_afsr;
4933 		ch_flt->afsr_ext = ch_flt->flt_sdw_afsr_ext = cdp->chd_afsr_ext;
4934 		ch_flt->afsr_errs = (cdp->chd_afsr_ext & C_AFSR_EXT_ALL_ERRS) |
4935 		    (cdp->chd_afsr & C_AFSR_ALL_ERRS);
4936 
4937 		/*
4938 		 * If the primary and shadow AFSR differ, tag the shadow as
4939 		 * the first fault.
4940 		 */
4941 		if ((primary_afar != cdp->chd_afar) ||
4942 		    (primary_afsr_errs != ch_flt->afsr_errs)) {
4943 			aflt->flt_stat |= (1ull << C_AFSR_FIRSTFLT_SHIFT);
4944 		}
4945 
4946 		/*
4947 		 * Check AFSR bits as well as AFSR_EXT bits in order of
4948 		 * the AFAR overwrite priority. Our stored AFSR_EXT value
4949 		 * is expected to be zero for those CPUs which do not have
4950 		 * an AFSR_EXT register.
4951 		 */
4952 		for (eccp = ecc_type_to_info; eccp->ec_desc != NULL; eccp++) {
4953 			if ((eccp->ec_afsr_bit &
4954 			    (ch_flt->afsr_errs & t_afsr_errs)) &&
4955 			    ((eccp->ec_flags & aflt->flt_status) != 0)) {
4956 				cpu_queue_one_event(ch_flt, reason, eccp, cdp);
4957 				cdp = NULL;
4958 				t_afsr_errs &= ~eccp->ec_afsr_bit;
4959 				nevents++;
4960 			}
4961 		}
4962 
4963 		/*
4964 		 * If the ME bit is on in the primary AFSR turn all the
4965 		 * error bits on again that may set the ME bit to make
4966 		 * sure we see the ME AFSR error logs.
4967 		 */
4968 		if ((primary_afsr & C_AFSR_ME) != 0)
4969 			t_afsr_errs = (orig_t_afsr_errs & C_AFSR_ALL_ME_ERRS);
4970 	}
4971 #endif	/* CHEETAH_PLUS */
4972 
4973 	if (clop != NULL)
4974 		cdp = &clop->clo_data;
4975 
4976 	/*
4977 	 * Queue expected errors, error bit and fault type must match
4978 	 * in the ecc_type_to_info table.
4979 	 */
4980 	for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL;
4981 	    eccp++) {
4982 		if ((eccp->ec_afsr_bit & t_afsr_errs) != 0 &&
4983 		    (eccp->ec_flags & aflt->flt_status) != 0) {
4984 #if defined(SERRANO)
4985 			/*
4986 			 * For FRC/FRU errors on Serrano the afar2 captures
4987 			 * the address and the associated data is
4988 			 * in the shadow logout area.
4989 			 */
4990 			if (eccp->ec_afsr_bit  & (C_AFSR_FRC | C_AFSR_FRU)) {
4991 				if (clop != NULL)
4992 					cdp = &clop->clo_sdw_data;
4993 				aflt->flt_addr = ch_flt->afar2;
4994 			} else {
4995 				if (clop != NULL)
4996 					cdp = &clop->clo_data;
4997 				aflt->flt_addr = primary_afar;
4998 			}
4999 #else	/* SERRANO */
5000 			aflt->flt_addr = primary_afar;
5001 #endif	/* SERRANO */
5002 			aflt->flt_stat = primary_afsr;
5003 			ch_flt->afsr_ext = primary_afsr_ext;
5004 			ch_flt->afsr_errs = primary_afsr_errs;
5005 			cpu_queue_one_event(ch_flt, reason, eccp, cdp);
5006 			cdp = NULL;
5007 			t_afsr_errs &= ~eccp->ec_afsr_bit;
5008 			nevents++;
5009 		}
5010 	}
5011 
5012 	/*
5013 	 * Queue unexpected errors, error bit only match.
5014 	 */
5015 	for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL;
5016 	    eccp++) {
5017 		if (eccp->ec_afsr_bit & t_afsr_errs) {
5018 #if defined(SERRANO)
5019 			/*
5020 			 * For FRC/FRU errors on Serrano the afar2 captures
5021 			 * the address and the associated data is
5022 			 * in the shadow logout area.
5023 			 */
5024 			if (eccp->ec_afsr_bit  & (C_AFSR_FRC | C_AFSR_FRU)) {
5025 				if (clop != NULL)
5026 					cdp = &clop->clo_sdw_data;
5027 				aflt->flt_addr = ch_flt->afar2;
5028 			} else {
5029 				if (clop != NULL)
5030 					cdp = &clop->clo_data;
5031 				aflt->flt_addr = primary_afar;
5032 			}
5033 #else	/* SERRANO */
5034 			aflt->flt_addr = primary_afar;
5035 #endif	/* SERRANO */
5036 			aflt->flt_stat = primary_afsr;
5037 			ch_flt->afsr_ext = primary_afsr_ext;
5038 			ch_flt->afsr_errs = primary_afsr_errs;
5039 			cpu_queue_one_event(ch_flt, reason, eccp, cdp);
5040 			cdp = NULL;
5041 			t_afsr_errs &= ~eccp->ec_afsr_bit;
5042 			nevents++;
5043 		}
5044 	}
5045 	return (nevents);
5046 }
5047 
5048 /*
5049  * Return trap type number.
5050  */
5051 uint8_t
5052 flt_to_trap_type(struct async_flt *aflt)
5053 {
5054 	if (aflt->flt_status & ECC_I_TRAP)
5055 		return (TRAP_TYPE_ECC_I);
5056 	if (aflt->flt_status & ECC_D_TRAP)
5057 		return (TRAP_TYPE_ECC_D);
5058 	if (aflt->flt_status & ECC_F_TRAP)
5059 		return (TRAP_TYPE_ECC_F);
5060 	if (aflt->flt_status & ECC_C_TRAP)
5061 		return (TRAP_TYPE_ECC_C);
5062 	if (aflt->flt_status & ECC_DP_TRAP)
5063 		return (TRAP_TYPE_ECC_DP);
5064 	if (aflt->flt_status & ECC_IP_TRAP)
5065 		return (TRAP_TYPE_ECC_IP);
5066 	if (aflt->flt_status & ECC_ITLB_TRAP)
5067 		return (TRAP_TYPE_ECC_ITLB);
5068 	if (aflt->flt_status & ECC_DTLB_TRAP)
5069 		return (TRAP_TYPE_ECC_DTLB);
5070 	return (TRAP_TYPE_UNKNOWN);
5071 }
5072 
5073 /*
5074  * Decide an error type based on detector and leaky/partner tests.
5075  * The following array is used for quick translation - it must
5076  * stay in sync with ce_dispact_t.
5077  */
5078 
5079 static char *cetypes[] = {
5080 	CE_DISP_DESC_U,
5081 	CE_DISP_DESC_I,
5082 	CE_DISP_DESC_PP,
5083 	CE_DISP_DESC_P,
5084 	CE_DISP_DESC_L,
5085 	CE_DISP_DESC_PS,
5086 	CE_DISP_DESC_S
5087 };
5088 
5089 char *
5090 flt_to_error_type(struct async_flt *aflt)
5091 {
5092 	ce_dispact_t dispact, disp;
5093 	uchar_t dtcrinfo, ptnrinfo, lkyinfo;
5094 
5095 	/*
5096 	 * The memory payload bundle is shared by some events that do
5097 	 * not perform any classification.  For those flt_disp will be
5098 	 * 0 and we will return "unknown".
5099 	 */
5100 	if (!ce_disp_inited || !aflt->flt_in_memory || aflt->flt_disp == 0)
5101 		return (cetypes[CE_DISP_UNKNOWN]);
5102 
5103 	dtcrinfo = CE_XDIAG_DTCRINFO(aflt->flt_disp);
5104 
5105 	/*
5106 	 * It is also possible that no scrub/classification was performed
5107 	 * by the detector, for instance where a disrupting error logged
5108 	 * in the AFSR while CEEN was off in cpu_deferred_error.
5109 	 */
5110 	if (!CE_XDIAG_EXT_ALG_APPLIED(dtcrinfo))
5111 		return (cetypes[CE_DISP_UNKNOWN]);
5112 
5113 	/*
5114 	 * Lookup type in initial classification/action table
5115 	 */
5116 	dispact = CE_DISPACT(ce_disp_table,
5117 	    CE_XDIAG_AFARMATCHED(dtcrinfo),
5118 	    CE_XDIAG_STATE(dtcrinfo),
5119 	    CE_XDIAG_CE1SEEN(dtcrinfo),
5120 	    CE_XDIAG_CE2SEEN(dtcrinfo));
5121 
5122 	/*
5123 	 * A bad lookup is not something to panic production systems for.
5124 	 */
5125 	ASSERT(dispact != CE_DISP_BAD);
5126 	if (dispact == CE_DISP_BAD)
5127 		return (cetypes[CE_DISP_UNKNOWN]);
5128 
5129 	disp = CE_DISP(dispact);
5130 
5131 	switch (disp) {
5132 	case CE_DISP_UNKNOWN:
5133 	case CE_DISP_INTERMITTENT:
5134 		break;
5135 
5136 	case CE_DISP_POSS_PERS:
5137 		/*
5138 		 * "Possible persistent" errors to which we have applied a valid
5139 		 * leaky test can be separated into "persistent" or "leaky".
5140 		 */
5141 		lkyinfo = CE_XDIAG_LKYINFO(aflt->flt_disp);
5142 		if (CE_XDIAG_TESTVALID(lkyinfo)) {
5143 			if (CE_XDIAG_CE1SEEN(lkyinfo) ||
5144 			    CE_XDIAG_CE2SEEN(lkyinfo))
5145 				disp = CE_DISP_LEAKY;
5146 			else
5147 				disp = CE_DISP_PERS;
5148 		}
5149 		break;
5150 
5151 	case CE_DISP_POSS_STICKY:
5152 		/*
5153 		 * Promote "possible sticky" results that have been
5154 		 * confirmed by a partner test to "sticky".  Unconfirmed
5155 		 * "possible sticky" events are left at that status - we do not
5156 		 * guess at any bad reader/writer etc status here.
5157 		 */
5158 		ptnrinfo = CE_XDIAG_PTNRINFO(aflt->flt_disp);
5159 		if (CE_XDIAG_TESTVALID(ptnrinfo) &&
5160 		    CE_XDIAG_CE1SEEN(ptnrinfo) && CE_XDIAG_CE2SEEN(ptnrinfo))
5161 			disp = CE_DISP_STICKY;
5162 
5163 		/*
5164 		 * Promote "possible sticky" results on a uniprocessor
5165 		 * to "sticky"
5166 		 */
5167 		if (disp == CE_DISP_POSS_STICKY &&
5168 		    CE_XDIAG_SKIPCODE(disp) == CE_XDIAG_SKIP_UNIPROC)
5169 			disp = CE_DISP_STICKY;
5170 		break;
5171 
5172 	default:
5173 		disp = CE_DISP_UNKNOWN;
5174 		break;
5175 	}
5176 
5177 	return (cetypes[disp]);
5178 }
5179 
5180 /*
5181  * Given the entire afsr, the specific bit to check and a prioritized list of
5182  * error bits, determine the validity of the various overwrite priority
5183  * features of the AFSR/AFAR: AFAR, ESYND and MSYND, each of which have
5184  * different overwrite priorities.
5185  *
5186  * Given a specific afsr error bit and the entire afsr, there are three cases:
5187  *   INVALID:	The specified bit is lower overwrite priority than some other
5188  *		error bit which is on in the afsr (or IVU/IVC).
5189  *   VALID:	The specified bit is higher priority than all other error bits
5190  *		which are on in the afsr.
5191  *   AMBIGUOUS: Another error bit (or bits) of equal priority to the specified
5192  *		bit is on in the afsr.
5193  */
5194 int
5195 afsr_to_overw_status(uint64_t afsr, uint64_t afsr_bit, uint64_t *ow_bits)
5196 {
5197 	uint64_t afsr_ow;
5198 
5199 	while ((afsr_ow = *ow_bits++) != 0) {
5200 		/*
5201 		 * If bit is in the priority class, check to see if another
5202 		 * bit in the same class is on => ambiguous.  Otherwise,
5203 		 * the value is valid.  If the bit is not on at this priority
5204 		 * class, but a higher priority bit is on, then the value is
5205 		 * invalid.
5206 		 */
5207 		if (afsr_ow & afsr_bit) {
5208 			/*
5209 			 * If equal pri bit is on, ambiguous.
5210 			 */
5211 			if (afsr & (afsr_ow & ~afsr_bit))
5212 				return (AFLT_STAT_AMBIGUOUS);
5213 			return (AFLT_STAT_VALID);
5214 		} else if (afsr & afsr_ow)
5215 			break;
5216 	}
5217 
5218 	/*
5219 	 * We didn't find a match or a higher priority bit was on.  Not
5220 	 * finding a match handles the case of invalid AFAR for IVC, IVU.
5221 	 */
5222 	return (AFLT_STAT_INVALID);
5223 }
5224 
5225 static int
5226 afsr_to_afar_status(uint64_t afsr, uint64_t afsr_bit)
5227 {
5228 #if defined(SERRANO)
5229 	if (afsr_bit & (C_AFSR_FRC | C_AFSR_FRU))
5230 		return (afsr_to_overw_status(afsr, afsr_bit, afar2_overwrite));
5231 	else
5232 #endif	/* SERRANO */
5233 		return (afsr_to_overw_status(afsr, afsr_bit, afar_overwrite));
5234 }
5235 
5236 static int
5237 afsr_to_esynd_status(uint64_t afsr, uint64_t afsr_bit)
5238 {
5239 	return (afsr_to_overw_status(afsr, afsr_bit, esynd_overwrite));
5240 }
5241 
5242 static int
5243 afsr_to_msynd_status(uint64_t afsr, uint64_t afsr_bit)
5244 {
5245 	return (afsr_to_overw_status(afsr, afsr_bit, msynd_overwrite));
5246 }
5247 
5248 static int
5249 afsr_to_synd_status(uint_t cpuid, uint64_t afsr, uint64_t afsr_bit)
5250 {
5251 #ifdef lint
5252 	cpuid = cpuid;
5253 #endif
5254 #if defined(CHEETAH_PLUS)
5255 	/*
5256 	 * The M_SYND overwrite policy is combined with the E_SYND overwrite
5257 	 * policy for Cheetah+ and separate for Panther CPUs.
5258 	 */
5259 	if (afsr_bit & C_AFSR_MSYND_ERRS) {
5260 		if (IS_PANTHER(cpunodes[cpuid].implementation))
5261 			return (afsr_to_msynd_status(afsr, afsr_bit));
5262 		else
5263 			return (afsr_to_esynd_status(afsr, afsr_bit));
5264 	} else if (afsr_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) {
5265 		if (IS_PANTHER(cpunodes[cpuid].implementation))
5266 			return (afsr_to_pn_esynd_status(afsr, afsr_bit));
5267 		else
5268 			return (afsr_to_esynd_status(afsr, afsr_bit));
5269 #else /* CHEETAH_PLUS */
5270 	if (afsr_bit & C_AFSR_MSYND_ERRS) {
5271 		return (afsr_to_msynd_status(afsr, afsr_bit));
5272 	} else if (afsr_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) {
5273 		return (afsr_to_esynd_status(afsr, afsr_bit));
5274 #endif /* CHEETAH_PLUS */
5275 	} else {
5276 		return (AFLT_STAT_INVALID);
5277 	}
5278 }
5279 
5280 /*
5281  * Slave CPU stick synchronization.
5282  */
5283 void
5284 sticksync_slave(void)
5285 {
5286 	int 		i;
5287 	int		tries = 0;
5288 	int64_t		tskew;
5289 	int64_t		av_tskew;
5290 
5291 	kpreempt_disable();
5292 	/* wait for the master side */
5293 	while (stick_sync_cmd != SLAVE_START)
5294 		;
5295 	/*
5296 	 * Synchronization should only take a few tries at most. But in the
5297 	 * odd case where the cpu isn't cooperating we'll keep trying. A cpu
5298 	 * without it's stick synchronized wouldn't be a good citizen.
5299 	 */
5300 	while (slave_done == 0) {
5301 		/*
5302 		 * Time skew calculation.
5303 		 */
5304 		av_tskew = tskew = 0;
5305 
5306 		for (i = 0; i < stick_iter; i++) {
5307 			/* make location hot */
5308 			timestamp[EV_A_START] = 0;
5309 			stick_timestamp(&timestamp[EV_A_START]);
5310 
5311 			/* tell the master we're ready */
5312 			stick_sync_cmd = MASTER_START;
5313 
5314 			/* and wait */
5315 			while (stick_sync_cmd != SLAVE_CONT)
5316 				;
5317 			/* Event B end */
5318 			stick_timestamp(&timestamp[EV_B_END]);
5319 
5320 			/* calculate time skew */
5321 			tskew = ((timestamp[EV_B_END] - timestamp[EV_B_START])
5322 				- (timestamp[EV_A_END] -
5323 				timestamp[EV_A_START])) / 2;
5324 
5325 			/* keep running count */
5326 			av_tskew += tskew;
5327 		} /* for */
5328 
5329 		/*
5330 		 * Adjust stick for time skew if not within the max allowed;
5331 		 * otherwise we're all done.
5332 		 */
5333 		if (stick_iter != 0)
5334 			av_tskew = av_tskew/stick_iter;
5335 		if (ABS(av_tskew) > stick_tsk) {
5336 			/*
5337 			 * If the skew is 1 (the slave's STICK register
5338 			 * is 1 STICK ahead of the master's), stick_adj
5339 			 * could fail to adjust the slave's STICK register
5340 			 * if the STICK read on the slave happens to
5341 			 * align with the increment of the STICK.
5342 			 * Therefore, we increment the skew to 2.
5343 			 */
5344 			if (av_tskew == 1)
5345 				av_tskew++;
5346 			stick_adj(-av_tskew);
5347 		} else
5348 			slave_done = 1;
5349 #ifdef DEBUG
5350 		if (tries < DSYNC_ATTEMPTS)
5351 			stick_sync_stats[CPU->cpu_id].skew_val[tries] =
5352 				av_tskew;
5353 		++tries;
5354 #endif /* DEBUG */
5355 #ifdef lint
5356 		tries = tries;
5357 #endif
5358 
5359 	} /* while */
5360 
5361 	/* allow the master to finish */
5362 	stick_sync_cmd = EVENT_NULL;
5363 	kpreempt_enable();
5364 }
5365 
5366 /*
5367  * Master CPU side of stick synchronization.
5368  *  - timestamp end of Event A
5369  *  - timestamp beginning of Event B
5370  */
5371 void
5372 sticksync_master(void)
5373 {
5374 	int		i;
5375 
5376 	kpreempt_disable();
5377 	/* tell the slave we've started */
5378 	slave_done = 0;
5379 	stick_sync_cmd = SLAVE_START;
5380 
5381 	while (slave_done == 0) {
5382 		for (i = 0; i < stick_iter; i++) {
5383 			/* wait for the slave */
5384 			while (stick_sync_cmd != MASTER_START)
5385 				;
5386 			/* Event A end */
5387 			stick_timestamp(&timestamp[EV_A_END]);
5388 
5389 			/* make location hot */
5390 			timestamp[EV_B_START] = 0;
5391 			stick_timestamp(&timestamp[EV_B_START]);
5392 
5393 			/* tell the slave to continue */
5394 			stick_sync_cmd = SLAVE_CONT;
5395 		} /* for */
5396 
5397 		/* wait while slave calculates time skew */
5398 		while (stick_sync_cmd == SLAVE_CONT)
5399 			;
5400 	} /* while */
5401 	kpreempt_enable();
5402 }
5403 
5404 /*
5405  * Cheetah/Cheetah+ have disrupting error for copyback's, so we don't need to
5406  * do Spitfire hack of xcall'ing all the cpus to ask to check for them.  Also,
5407  * in cpu_async_panic_callb, each cpu checks for CPU events on its way to
5408  * panic idle.
5409  */
5410 /*ARGSUSED*/
5411 void
5412 cpu_check_allcpus(struct async_flt *aflt)
5413 {}
5414 
5415 struct kmem_cache *ch_private_cache;
5416 
5417 /*
5418  * Cpu private unitialization.  Uninitialize the Ecache scrubber and
5419  * deallocate the scrubber data structures and cpu_private data structure.
5420  */
5421 void
5422 cpu_uninit_private(struct cpu *cp)
5423 {
5424 	cheetah_private_t *chprp = CPU_PRIVATE(cp);
5425 
5426 	ASSERT(chprp);
5427 	cpu_uninit_ecache_scrub_dr(cp);
5428 	CPU_PRIVATE(cp) = NULL;
5429 	ch_err_tl1_paddrs[cp->cpu_id] = NULL;
5430 	kmem_cache_free(ch_private_cache, chprp);
5431 	cmp_delete_cpu(cp->cpu_id);
5432 
5433 }
5434 
5435 /*
5436  * Cheetah Cache Scrubbing
5437  *
5438  * The primary purpose of Cheetah cache scrubbing is to reduce the exposure
5439  * of E$ tags, D$ data, and I$ data to cosmic ray events since they are not
5440  * protected by either parity or ECC.
5441  *
5442  * We currently default the E$ and D$ scan rate to 100 (scan 10% of the
5443  * cache per second). Due to the the specifics of how the I$ control
5444  * logic works with respect to the ASI used to scrub I$ lines, the entire
5445  * I$ is scanned at once.
5446  */
5447 
5448 /*
5449  * Tuneables to enable and disable the scrubbing of the caches, and to tune
5450  * scrubbing behavior.  These may be changed via /etc/system or using mdb
5451  * on a running system.
5452  */
5453 int dcache_scrub_enable = 1;		/* D$ scrubbing is on by default */
5454 
5455 /*
5456  * The following are the PIL levels that the softints/cross traps will fire at.
5457  */
5458 uint_t ecache_scrub_pil = PIL_9;	/* E$ scrub PIL for cross traps */
5459 uint_t dcache_scrub_pil = PIL_9;	/* D$ scrub PIL for cross traps */
5460 uint_t icache_scrub_pil = PIL_9;	/* I$ scrub PIL for cross traps */
5461 
5462 #if defined(JALAPENO)
5463 
5464 /*
5465  * Due to several errata (82, 85, 86), we don't enable the L2$ scrubber
5466  * on Jalapeno.
5467  */
5468 int ecache_scrub_enable = 0;
5469 
5470 #else	/* JALAPENO */
5471 
5472 /*
5473  * With all other cpu types, E$ scrubbing is on by default
5474  */
5475 int ecache_scrub_enable = 1;
5476 
5477 #endif	/* JALAPENO */
5478 
5479 
5480 #if defined(CHEETAH_PLUS) || defined(JALAPENO) || defined(SERRANO)
5481 
5482 /*
5483  * The I$ scrubber tends to cause latency problems for real-time SW, so it
5484  * is disabled by default on non-Cheetah systems
5485  */
5486 int icache_scrub_enable = 0;
5487 
5488 /*
5489  * Tuneables specifying the scrub calls per second and the scan rate
5490  * for each cache
5491  *
5492  * The cyclic times are set during boot based on the following values.
5493  * Changing these values in mdb after this time will have no effect.  If
5494  * a different value is desired, it must be set in /etc/system before a
5495  * reboot.
5496  */
5497 int ecache_calls_a_sec = 1;
5498 int dcache_calls_a_sec = 2;
5499 int icache_calls_a_sec = 2;
5500 
5501 int ecache_scan_rate_idle = 1;
5502 int ecache_scan_rate_busy = 1;
5503 int dcache_scan_rate_idle = 1;
5504 int dcache_scan_rate_busy = 1;
5505 int icache_scan_rate_idle = 1;
5506 int icache_scan_rate_busy = 1;
5507 
5508 #else	/* CHEETAH_PLUS || JALAPENO || SERRANO */
5509 
5510 int icache_scrub_enable = 1;		/* I$ scrubbing is on by default */
5511 
5512 int ecache_calls_a_sec = 100;		/* E$ scrub calls per seconds */
5513 int dcache_calls_a_sec = 100;		/* D$ scrub calls per seconds */
5514 int icache_calls_a_sec = 100;		/* I$ scrub calls per seconds */
5515 
5516 int ecache_scan_rate_idle = 100;	/* E$ scan rate (in tenths of a %) */
5517 int ecache_scan_rate_busy = 100;	/* E$ scan rate (in tenths of a %) */
5518 int dcache_scan_rate_idle = 100;	/* D$ scan rate (in tenths of a %) */
5519 int dcache_scan_rate_busy = 100;	/* D$ scan rate (in tenths of a %) */
5520 int icache_scan_rate_idle = 100;	/* I$ scan rate (in tenths of a %) */
5521 int icache_scan_rate_busy = 100;	/* I$ scan rate (in tenths of a %) */
5522 
5523 #endif	/* CHEETAH_PLUS || JALAPENO || SERRANO */
5524 
5525 /*
5526  * In order to scrub on offline cpus, a cross trap is sent.  The handler will
5527  * increment the outstanding request counter and schedule a softint to run
5528  * the scrubber.
5529  */
5530 extern xcfunc_t cache_scrubreq_tl1;
5531 
5532 /*
5533  * These are the softint functions for each cache scrubber
5534  */
5535 static uint_t scrub_ecache_line_intr(caddr_t arg1, caddr_t arg2);
5536 static uint_t scrub_dcache_line_intr(caddr_t arg1, caddr_t arg2);
5537 static uint_t scrub_icache_line_intr(caddr_t arg1, caddr_t arg2);
5538 
5539 /*
5540  * The cache scrub info table contains cache specific information
5541  * and allows for some of the scrub code to be table driven, reducing
5542  * duplication of cache similar code.
5543  *
5544  * This table keeps a copy of the value in the calls per second variable
5545  * (?cache_calls_a_sec).  This makes it much more difficult for someone
5546  * to cause us problems (for example, by setting ecache_calls_a_sec to 0 in
5547  * mdb in a misguided attempt to disable the scrubber).
5548  */
5549 struct scrub_info {
5550 	int		*csi_enable;	/* scrubber enable flag */
5551 	int		csi_freq;	/* scrubber calls per second */
5552 	int		csi_index;	/* index to chsm_outstanding[] */
5553 	uint_t		csi_inum;	/* scrubber interrupt number */
5554 	cyclic_id_t	csi_omni_cyc_id;	/* omni cyclic ID */
5555 	cyclic_id_t	csi_offline_cyc_id;	/* offline cyclic ID */
5556 	char		csi_name[3];	/* cache name for this scrub entry */
5557 } cache_scrub_info[] = {
5558 { &ecache_scrub_enable, 0, CACHE_SCRUBBER_INFO_E, 0, 0, 0, "E$"},
5559 { &dcache_scrub_enable, 0, CACHE_SCRUBBER_INFO_D, 0, 0, 0, "D$"},
5560 { &icache_scrub_enable, 0, CACHE_SCRUBBER_INFO_I, 0, 0, 0, "I$"}
5561 };
5562 
5563 /*
5564  * If scrubbing is enabled, increment the outstanding request counter.  If it
5565  * is 1 (meaning there were no previous requests outstanding), call
5566  * setsoftint_tl1 through xt_one_unchecked, which eventually ends up doing
5567  * a self trap.
5568  */
5569 static void
5570 do_scrub(struct scrub_info *csi)
5571 {
5572 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5573 	int index = csi->csi_index;
5574 	uint32_t *outstanding = &csmp->chsm_outstanding[index];
5575 
5576 	if (*(csi->csi_enable) && (csmp->chsm_enable[index])) {
5577 		if (atomic_add_32_nv(outstanding, 1) == 1) {
5578 			xt_one_unchecked(CPU->cpu_id, setsoftint_tl1,
5579 			    csi->csi_inum, 0);
5580 		}
5581 	}
5582 }
5583 
5584 /*
5585  * Omni cyclics don't fire on offline cpus, so we use another cyclic to
5586  * cross-trap the offline cpus.
5587  */
5588 static void
5589 do_scrub_offline(struct scrub_info *csi)
5590 {
5591 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5592 
5593 	if (CPUSET_ISNULL(cpu_offline_set)) {
5594 		/*
5595 		 * No offline cpus - nothing to do
5596 		 */
5597 		return;
5598 	}
5599 
5600 	if (*(csi->csi_enable) && (csmp->chsm_enable[csi->csi_index])) {
5601 		xt_some(cpu_offline_set, cache_scrubreq_tl1, csi->csi_inum,
5602 		    csi->csi_index);
5603 	}
5604 }
5605 
5606 /*
5607  * This is the initial setup for the scrubber cyclics - it sets the
5608  * interrupt level, frequency, and function to call.
5609  */
5610 /*ARGSUSED*/
5611 static void
5612 cpu_scrub_cyclic_setup(void *arg, cpu_t *cpu, cyc_handler_t *hdlr,
5613     cyc_time_t *when)
5614 {
5615 	struct scrub_info *csi = (struct scrub_info *)arg;
5616 
5617 	ASSERT(csi != NULL);
5618 	hdlr->cyh_func = (cyc_func_t)do_scrub;
5619 	hdlr->cyh_level = CY_LOW_LEVEL;
5620 	hdlr->cyh_arg = arg;
5621 
5622 	when->cyt_when = 0;	/* Start immediately */
5623 	when->cyt_interval = NANOSEC / csi->csi_freq;
5624 }
5625 
5626 /*
5627  * Initialization for cache scrubbing.
5628  * This routine is called AFTER all cpus have had cpu_init_private called
5629  * to initialize their private data areas.
5630  */
5631 void
5632 cpu_init_cache_scrub(void)
5633 {
5634 	int i;
5635 	struct scrub_info *csi;
5636 	cyc_omni_handler_t omni_hdlr;
5637 	cyc_handler_t offline_hdlr;
5638 	cyc_time_t when;
5639 
5640 	/*
5641 	 * save away the maximum number of lines for the D$
5642 	 */
5643 	dcache_nlines = dcache_size / dcache_linesize;
5644 
5645 	/*
5646 	 * register the softints for the cache scrubbing
5647 	 */
5648 	cache_scrub_info[CACHE_SCRUBBER_INFO_E].csi_inum =
5649 	    add_softintr(ecache_scrub_pil, scrub_ecache_line_intr,
5650 	    (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_E]);
5651 	cache_scrub_info[CACHE_SCRUBBER_INFO_E].csi_freq = ecache_calls_a_sec;
5652 
5653 	cache_scrub_info[CACHE_SCRUBBER_INFO_D].csi_inum =
5654 	    add_softintr(dcache_scrub_pil, scrub_dcache_line_intr,
5655 	    (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_D]);
5656 	cache_scrub_info[CACHE_SCRUBBER_INFO_D].csi_freq = dcache_calls_a_sec;
5657 
5658 	cache_scrub_info[CACHE_SCRUBBER_INFO_I].csi_inum =
5659 	    add_softintr(icache_scrub_pil, scrub_icache_line_intr,
5660 	    (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_I]);
5661 	cache_scrub_info[CACHE_SCRUBBER_INFO_I].csi_freq = icache_calls_a_sec;
5662 
5663 	/*
5664 	 * start the scrubbing for all the caches
5665 	 */
5666 	mutex_enter(&cpu_lock);
5667 	for (i = 0; i < CACHE_SCRUBBER_COUNT; i++) {
5668 
5669 		csi = &cache_scrub_info[i];
5670 
5671 		if (!(*csi->csi_enable))
5672 			continue;
5673 
5674 		/*
5675 		 * force the following to be true:
5676 		 *	1 <= calls_a_sec <= hz
5677 		 */
5678 		if (csi->csi_freq > hz) {
5679 			cmn_err(CE_NOTE, "%s scrub calls_a_sec set too high "
5680 				"(%d); resetting to hz (%d)", csi->csi_name,
5681 				csi->csi_freq, hz);
5682 			csi->csi_freq = hz;
5683 		} else if (csi->csi_freq < 1) {
5684 			cmn_err(CE_NOTE, "%s scrub calls_a_sec set too low "
5685 				"(%d); resetting to 1", csi->csi_name,
5686 				csi->csi_freq);
5687 			csi->csi_freq = 1;
5688 		}
5689 
5690 		omni_hdlr.cyo_online = cpu_scrub_cyclic_setup;
5691 		omni_hdlr.cyo_offline = NULL;
5692 		omni_hdlr.cyo_arg = (void *)csi;
5693 
5694 		offline_hdlr.cyh_func = (cyc_func_t)do_scrub_offline;
5695 		offline_hdlr.cyh_arg = (void *)csi;
5696 		offline_hdlr.cyh_level = CY_LOW_LEVEL;
5697 
5698 		when.cyt_when = 0;	/* Start immediately */
5699 		when.cyt_interval = NANOSEC / csi->csi_freq;
5700 
5701 		csi->csi_omni_cyc_id = cyclic_add_omni(&omni_hdlr);
5702 		csi->csi_offline_cyc_id = cyclic_add(&offline_hdlr, &when);
5703 	}
5704 	register_cpu_setup_func(cpu_scrub_cpu_setup, NULL);
5705 	mutex_exit(&cpu_lock);
5706 }
5707 
5708 /*
5709  * Indicate that the specified cpu is idle.
5710  */
5711 void
5712 cpu_idle_ecache_scrub(struct cpu *cp)
5713 {
5714 	if (CPU_PRIVATE(cp) != NULL) {
5715 		ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc);
5716 		csmp->chsm_ecache_busy = ECACHE_CPU_IDLE;
5717 	}
5718 }
5719 
5720 /*
5721  * Indicate that the specified cpu is busy.
5722  */
5723 void
5724 cpu_busy_ecache_scrub(struct cpu *cp)
5725 {
5726 	if (CPU_PRIVATE(cp) != NULL) {
5727 		ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc);
5728 		csmp->chsm_ecache_busy = ECACHE_CPU_BUSY;
5729 	}
5730 }
5731 
5732 /*
5733  * Initialization for cache scrubbing for the specified cpu.
5734  */
5735 void
5736 cpu_init_ecache_scrub_dr(struct cpu *cp)
5737 {
5738 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc);
5739 	int cpuid = cp->cpu_id;
5740 
5741 	/* initialize the number of lines in the caches */
5742 	csmp->chsm_ecache_nlines = cpunodes[cpuid].ecache_size /
5743 	    cpunodes[cpuid].ecache_linesize;
5744 	csmp->chsm_icache_nlines = CPU_PRIVATE_VAL(cp, chpr_icache_size) /
5745 	    CPU_PRIVATE_VAL(cp, chpr_icache_linesize);
5746 
5747 	/*
5748 	 * do_scrub() and do_scrub_offline() check both the global
5749 	 * ?cache_scrub_enable and this per-cpu enable variable.  All scrubbers
5750 	 * check this value before scrubbing.  Currently, we use it to
5751 	 * disable the E$ scrubber on multi-core cpus or while running at
5752 	 * slowed speed.  For now, just turn everything on and allow
5753 	 * cpu_init_private() to change it if necessary.
5754 	 */
5755 	csmp->chsm_enable[CACHE_SCRUBBER_INFO_E] = 1;
5756 	csmp->chsm_enable[CACHE_SCRUBBER_INFO_D] = 1;
5757 	csmp->chsm_enable[CACHE_SCRUBBER_INFO_I] = 1;
5758 
5759 	cpu_busy_ecache_scrub(cp);
5760 }
5761 
5762 /*
5763  * Un-initialization for cache scrubbing for the specified cpu.
5764  */
5765 static void
5766 cpu_uninit_ecache_scrub_dr(struct cpu *cp)
5767 {
5768 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc);
5769 
5770 	/*
5771 	 * un-initialize bookkeeping for cache scrubbing
5772 	 */
5773 	bzero(csmp, sizeof (ch_scrub_misc_t));
5774 
5775 	cpu_idle_ecache_scrub(cp);
5776 }
5777 
5778 /*
5779  * Called periodically on each CPU to scrub the D$.
5780  */
5781 static void
5782 scrub_dcache(int how_many)
5783 {
5784 	int i;
5785 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5786 	int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_D];
5787 
5788 	/*
5789 	 * scrub the desired number of lines
5790 	 */
5791 	for (i = 0; i < how_many; i++) {
5792 		/*
5793 		 * scrub a D$ line
5794 		 */
5795 		dcache_inval_line(index);
5796 
5797 		/*
5798 		 * calculate the next D$ line to scrub, assumes
5799 		 * that dcache_nlines is a power of 2
5800 		 */
5801 		index = (index + 1) & (dcache_nlines - 1);
5802 	}
5803 
5804 	/*
5805 	 * set the scrub index for the next visit
5806 	 */
5807 	csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_D] = index;
5808 }
5809 
5810 /*
5811  * Handler for D$ scrub inum softint. Call scrub_dcache until
5812  * we decrement the outstanding request count to zero.
5813  */
5814 /*ARGSUSED*/
5815 static uint_t
5816 scrub_dcache_line_intr(caddr_t arg1, caddr_t arg2)
5817 {
5818 	int i;
5819 	int how_many;
5820 	int outstanding;
5821 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5822 	uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_D];
5823 	struct scrub_info *csi = (struct scrub_info *)arg1;
5824 	int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ?
5825 		dcache_scan_rate_idle : dcache_scan_rate_busy;
5826 
5827 	/*
5828 	 * The scan rates are expressed in units of tenths of a
5829 	 * percent.  A scan rate of 1000 (100%) means the whole
5830 	 * cache is scanned every second.
5831 	 */
5832 	how_many = (dcache_nlines * scan_rate) / (1000 * csi->csi_freq);
5833 
5834 	do {
5835 		outstanding = *countp;
5836 		for (i = 0; i < outstanding; i++) {
5837 			scrub_dcache(how_many);
5838 		}
5839 	} while (atomic_add_32_nv(countp, -outstanding));
5840 
5841 	return (DDI_INTR_CLAIMED);
5842 }
5843 
5844 /*
5845  * Called periodically on each CPU to scrub the I$. The I$ is scrubbed
5846  * by invalidating lines. Due to the characteristics of the ASI which
5847  * is used to invalidate an I$ line, the entire I$ must be invalidated
5848  * vs. an individual I$ line.
5849  */
5850 static void
5851 scrub_icache(int how_many)
5852 {
5853 	int i;
5854 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5855 	int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_I];
5856 	int icache_nlines = csmp->chsm_icache_nlines;
5857 
5858 	/*
5859 	 * scrub the desired number of lines
5860 	 */
5861 	for (i = 0; i < how_many; i++) {
5862 		/*
5863 		 * since the entire I$ must be scrubbed at once,
5864 		 * wait until the index wraps to zero to invalidate
5865 		 * the entire I$
5866 		 */
5867 		if (index == 0) {
5868 			icache_inval_all();
5869 		}
5870 
5871 		/*
5872 		 * calculate the next I$ line to scrub, assumes
5873 		 * that chsm_icache_nlines is a power of 2
5874 		 */
5875 		index = (index + 1) & (icache_nlines - 1);
5876 	}
5877 
5878 	/*
5879 	 * set the scrub index for the next visit
5880 	 */
5881 	csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_I] = index;
5882 }
5883 
5884 /*
5885  * Handler for I$ scrub inum softint. Call scrub_icache until
5886  * we decrement the outstanding request count to zero.
5887  */
5888 /*ARGSUSED*/
5889 static uint_t
5890 scrub_icache_line_intr(caddr_t arg1, caddr_t arg2)
5891 {
5892 	int i;
5893 	int how_many;
5894 	int outstanding;
5895 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5896 	uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_I];
5897 	struct scrub_info *csi = (struct scrub_info *)arg1;
5898 	int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ?
5899 	    icache_scan_rate_idle : icache_scan_rate_busy;
5900 	int icache_nlines = csmp->chsm_icache_nlines;
5901 
5902 	/*
5903 	 * The scan rates are expressed in units of tenths of a
5904 	 * percent.  A scan rate of 1000 (100%) means the whole
5905 	 * cache is scanned every second.
5906 	 */
5907 	how_many = (icache_nlines * scan_rate) / (1000 * csi->csi_freq);
5908 
5909 	do {
5910 		outstanding = *countp;
5911 		for (i = 0; i < outstanding; i++) {
5912 			scrub_icache(how_many);
5913 		}
5914 	} while (atomic_add_32_nv(countp, -outstanding));
5915 
5916 	return (DDI_INTR_CLAIMED);
5917 }
5918 
5919 /*
5920  * Called periodically on each CPU to scrub the E$.
5921  */
5922 static void
5923 scrub_ecache(int how_many)
5924 {
5925 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5926 	int i;
5927 	int cpuid = CPU->cpu_id;
5928 	int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_E];
5929 	int nlines = csmp->chsm_ecache_nlines;
5930 	int linesize = cpunodes[cpuid].ecache_linesize;
5931 	int ec_set_size = cpu_ecache_set_size(CPU);
5932 
5933 	/*
5934 	 * scrub the desired number of lines
5935 	 */
5936 	for (i = 0; i < how_many; i++) {
5937 		/*
5938 		 * scrub the E$ line
5939 		 */
5940 		ecache_flush_line(ecache_flushaddr + (index * linesize),
5941 		    ec_set_size);
5942 
5943 		/*
5944 		 * calculate the next E$ line to scrub based on twice
5945 		 * the number of E$ lines (to displace lines containing
5946 		 * flush area data), assumes that the number of lines
5947 		 * is a power of 2
5948 		 */
5949 		index = (index + 1) & ((nlines << 1) - 1);
5950 	}
5951 
5952 	/*
5953 	 * set the ecache scrub index for the next visit
5954 	 */
5955 	csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_E] = index;
5956 }
5957 
5958 /*
5959  * Handler for E$ scrub inum softint. Call the E$ scrubber until
5960  * we decrement the outstanding request count to zero.
5961  *
5962  * Due to interactions with cpu_scrub_cpu_setup(), the outstanding count may
5963  * become negative after the atomic_add_32_nv().  This is not a problem, as
5964  * the next trip around the loop won't scrub anything, and the next add will
5965  * reset the count back to zero.
5966  */
5967 /*ARGSUSED*/
5968 static uint_t
5969 scrub_ecache_line_intr(caddr_t arg1, caddr_t arg2)
5970 {
5971 	int i;
5972 	int how_many;
5973 	int outstanding;
5974 	ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc);
5975 	uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_E];
5976 	struct scrub_info *csi = (struct scrub_info *)arg1;
5977 	int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ?
5978 		ecache_scan_rate_idle : ecache_scan_rate_busy;
5979 	int ecache_nlines = csmp->chsm_ecache_nlines;
5980 
5981 	/*
5982 	 * The scan rates are expressed in units of tenths of a
5983 	 * percent.  A scan rate of 1000 (100%) means the whole
5984 	 * cache is scanned every second.
5985 	 */
5986 	how_many = (ecache_nlines * scan_rate) / (1000 * csi->csi_freq);
5987 
5988 	do {
5989 		outstanding = *countp;
5990 		for (i = 0; i < outstanding; i++) {
5991 			scrub_ecache(how_many);
5992 		}
5993 	} while (atomic_add_32_nv(countp, -outstanding));
5994 
5995 	return (DDI_INTR_CLAIMED);
5996 }
5997 
5998 /*
5999  * Timeout function to reenable CE
6000  */
6001 static void
6002 cpu_delayed_check_ce_errors(void *arg)
6003 {
6004 	if (!taskq_dispatch(ch_check_ce_tq, cpu_check_ce_errors, arg,
6005 	    TQ_NOSLEEP)) {
6006 		(void) timeout(cpu_delayed_check_ce_errors, arg,
6007 		    drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC));
6008 	}
6009 }
6010 
6011 /*
6012  * CE Deferred Re-enable after trap.
6013  *
6014  * When the CPU gets a disrupting trap for any of the errors
6015  * controlled by the CEEN bit, CEEN is disabled in the trap handler
6016  * immediately. To eliminate the possibility of multiple CEs causing
6017  * recursive stack overflow in the trap handler, we cannot
6018  * reenable CEEN while still running in the trap handler. Instead,
6019  * after a CE is logged on a CPU, we schedule a timeout function,
6020  * cpu_check_ce_errors(), to trigger after cpu_ceen_delay_secs
6021  * seconds. This function will check whether any further CEs
6022  * have occurred on that CPU, and if none have, will reenable CEEN.
6023  *
6024  * If further CEs have occurred while CEEN is disabled, another
6025  * timeout will be scheduled. This is to ensure that the CPU can
6026  * make progress in the face of CE 'storms', and that it does not
6027  * spend all its time logging CE errors.
6028  */
6029 static void
6030 cpu_check_ce_errors(void *arg)
6031 {
6032 	int	cpuid = (int)(uintptr_t)arg;
6033 	cpu_t	*cp;
6034 
6035 	/*
6036 	 * We acquire cpu_lock.
6037 	 */
6038 	ASSERT(curthread->t_pil == 0);
6039 
6040 	/*
6041 	 * verify that the cpu is still around, DR
6042 	 * could have got there first ...
6043 	 */
6044 	mutex_enter(&cpu_lock);
6045 	cp = cpu_get(cpuid);
6046 	if (cp == NULL) {
6047 		mutex_exit(&cpu_lock);
6048 		return;
6049 	}
6050 	/*
6051 	 * make sure we don't migrate across CPUs
6052 	 * while checking our CE status.
6053 	 */
6054 	kpreempt_disable();
6055 
6056 	/*
6057 	 * If we are running on the CPU that got the
6058 	 * CE, we can do the checks directly.
6059 	 */
6060 	if (cp->cpu_id == CPU->cpu_id) {
6061 		mutex_exit(&cpu_lock);
6062 		cpu_check_ce(TIMEOUT_CEEN_CHECK, 0, 0, 0);
6063 		kpreempt_enable();
6064 		return;
6065 	}
6066 	kpreempt_enable();
6067 
6068 	/*
6069 	 * send an x-call to get the CPU that originally
6070 	 * got the CE to do the necessary checks. If we can't
6071 	 * send the x-call, reschedule the timeout, otherwise we
6072 	 * lose CEEN forever on that CPU.
6073 	 */
6074 	if (CPU_XCALL_READY(cp->cpu_id) && (!(cp->cpu_flags & CPU_QUIESCED))) {
6075 		xc_one(cp->cpu_id, (xcfunc_t *)cpu_check_ce,
6076 		    TIMEOUT_CEEN_CHECK, 0);
6077 		mutex_exit(&cpu_lock);
6078 	} else {
6079 		/*
6080 		 * When the CPU is not accepting xcalls, or
6081 		 * the processor is offlined, we don't want to
6082 		 * incur the extra overhead of trying to schedule the
6083 		 * CE timeout indefinitely. However, we don't want to lose
6084 		 * CE checking forever.
6085 		 *
6086 		 * Keep rescheduling the timeout, accepting the additional
6087 		 * overhead as the cost of correctness in the case where we get
6088 		 * a CE, disable CEEN, offline the CPU during the
6089 		 * the timeout interval, and then online it at some
6090 		 * point in the future. This is unlikely given the short
6091 		 * cpu_ceen_delay_secs.
6092 		 */
6093 		mutex_exit(&cpu_lock);
6094 		(void) timeout(cpu_delayed_check_ce_errors,
6095 		    (void *)(uintptr_t)cp->cpu_id,
6096 		    drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC));
6097 	}
6098 }
6099 
6100 /*
6101  * This routine will check whether CEs have occurred while
6102  * CEEN is disabled. Any CEs detected will be logged and, if
6103  * possible, scrubbed.
6104  *
6105  * The memscrubber will also use this routine to clear any errors
6106  * caused by its scrubbing with CEEN disabled.
6107  *
6108  * flag == SCRUBBER_CEEN_CHECK
6109  *		called from memscrubber, just check/scrub, no reset
6110  *		paddr 	physical addr. for start of scrub pages
6111  *		vaddr 	virtual addr. for scrub area
6112  *		psz	page size of area to be scrubbed
6113  *
6114  * flag == TIMEOUT_CEEN_CHECK
6115  *		timeout function has triggered, reset timeout or CEEN
6116  *
6117  * Note: We must not migrate cpus during this function.  This can be
6118  * achieved by one of:
6119  *    - invoking as target of an x-call in which case we're at XCALL_PIL
6120  *	The flag value must be first xcall argument.
6121  *    - disabling kernel preemption.  This should be done for very short
6122  *	periods so is not suitable for SCRUBBER_CEEN_CHECK where we might
6123  *	scrub an extended area with cpu_check_block.  The call for
6124  *	TIMEOUT_CEEN_CHECK uses this so cpu_check_ce must be kept
6125  *	brief for this case.
6126  *    - binding to a cpu, eg with thread_affinity_set().  This is used
6127  *	in the SCRUBBER_CEEN_CHECK case, but is not practical for
6128  *	the TIMEOUT_CEEN_CHECK because both need cpu_lock.
6129  */
6130 void
6131 cpu_check_ce(int flag, uint64_t pa, caddr_t va, uint_t psz)
6132 {
6133 	ch_cpu_errors_t	cpu_error_regs;
6134 	uint64_t	ec_err_enable;
6135 	uint64_t	page_offset;
6136 
6137 	/* Read AFSR */
6138 	get_cpu_error_state(&cpu_error_regs);
6139 
6140 	/*
6141 	 * If no CEEN errors have occurred during the timeout
6142 	 * interval, it is safe to re-enable CEEN and exit.
6143 	 */
6144 	if ((cpu_error_regs.afsr & C_AFSR_CECC_ERRS) == 0) {
6145 		if (flag == TIMEOUT_CEEN_CHECK &&
6146 		    !((ec_err_enable = get_error_enable()) & EN_REG_CEEN))
6147 			set_error_enable(ec_err_enable | EN_REG_CEEN);
6148 		return;
6149 	}
6150 
6151 	/*
6152 	 * Ensure that CEEN was not reenabled (maybe by DR) before
6153 	 * we log/clear the error.
6154 	 */
6155 	if ((ec_err_enable = get_error_enable()) & EN_REG_CEEN)
6156 	    set_error_enable(ec_err_enable & ~EN_REG_CEEN);
6157 
6158 	/*
6159 	 * log/clear the CE. If CE_CEEN_DEFER is passed, the
6160 	 * timeout will be rescheduled when the error is logged.
6161 	 */
6162 	if (!(cpu_error_regs.afsr & cpu_ce_not_deferred))
6163 	    cpu_ce_detected(&cpu_error_regs,
6164 		CE_CEEN_DEFER | CE_CEEN_TIMEOUT);
6165 	else
6166 	    cpu_ce_detected(&cpu_error_regs, CE_CEEN_TIMEOUT);
6167 
6168 	/*
6169 	 * If the memory scrubber runs while CEEN is
6170 	 * disabled, (or if CEEN is disabled during the
6171 	 * scrub as a result of a CE being triggered by
6172 	 * it), the range being scrubbed will not be
6173 	 * completely cleaned. If there are multiple CEs
6174 	 * in the range at most two of these will be dealt
6175 	 * with, (one by the trap handler and one by the
6176 	 * timeout). It is also possible that none are dealt
6177 	 * with, (CEEN disabled and another CE occurs before
6178 	 * the timeout triggers). So to ensure that the
6179 	 * memory is actually scrubbed, we have to access each
6180 	 * memory location in the range and then check whether
6181 	 * that access causes a CE.
6182 	 */
6183 	if (flag == SCRUBBER_CEEN_CHECK && va) {
6184 		if ((cpu_error_regs.afar >= pa) &&
6185 		    (cpu_error_regs.afar < (pa + psz))) {
6186 			/*
6187 			 * Force a load from physical memory for each
6188 			 * 64-byte block, then check AFSR to determine
6189 			 * whether this access caused an error.
6190 			 *
6191 			 * This is a slow way to do a scrub, but as it will
6192 			 * only be invoked when the memory scrubber actually
6193 			 * triggered a CE, it should not happen too
6194 			 * frequently.
6195 			 *
6196 			 * cut down what we need to check as the scrubber
6197 			 * has verified up to AFAR, so get it's offset
6198 			 * into the page and start there.
6199 			 */
6200 			page_offset = (uint64_t)(cpu_error_regs.afar &
6201 			    (psz - 1));
6202 			va = (caddr_t)(va + (P2ALIGN(page_offset, 64)));
6203 			psz -= (uint_t)(P2ALIGN(page_offset, 64));
6204 			cpu_check_block((caddr_t)(P2ALIGN((uint64_t)va, 64)),
6205 			    psz);
6206 		}
6207 	}
6208 
6209 	/*
6210 	 * Reset error enable if this CE is not masked.
6211 	 */
6212 	if ((flag == TIMEOUT_CEEN_CHECK) &&
6213 	    (cpu_error_regs.afsr & cpu_ce_not_deferred))
6214 	    set_error_enable(ec_err_enable | EN_REG_CEEN);
6215 
6216 }
6217 
6218 /*
6219  * Attempt a cpu logout for an error that we did not trap for, such
6220  * as a CE noticed with CEEN off.  It is assumed that we are still running
6221  * on the cpu that took the error and that we cannot migrate.  Returns
6222  * 0 on success, otherwise nonzero.
6223  */
6224 static int
6225 cpu_ce_delayed_ec_logout(uint64_t afar)
6226 {
6227 	ch_cpu_logout_t *clop;
6228 
6229 	if (CPU_PRIVATE(CPU) == NULL)
6230 		return (0);
6231 
6232 	clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout);
6233 	if (cas64(&clop->clo_data.chd_afar, LOGOUT_INVALID, afar) !=
6234 	    LOGOUT_INVALID)
6235 		return (0);
6236 
6237 	cpu_delayed_logout(afar, clop);
6238 	return (1);
6239 }
6240 
6241 /*
6242  * We got an error while CEEN was disabled. We
6243  * need to clean up after it and log whatever
6244  * information we have on the CE.
6245  */
6246 void
6247 cpu_ce_detected(ch_cpu_errors_t *cpu_error_regs, int flag)
6248 {
6249 	ch_async_flt_t 	ch_flt;
6250 	struct async_flt *aflt;
6251 	char 		pr_reason[MAX_REASON_STRING];
6252 
6253 	bzero(&ch_flt, sizeof (ch_async_flt_t));
6254 	ch_flt.flt_trapped_ce = flag;
6255 	aflt = (struct async_flt *)&ch_flt;
6256 	aflt->flt_stat = cpu_error_regs->afsr & C_AFSR_MASK;
6257 	ch_flt.afsr_ext = cpu_error_regs->afsr_ext;
6258 	ch_flt.afsr_errs = (cpu_error_regs->afsr_ext & C_AFSR_EXT_ALL_ERRS) |
6259 	    (cpu_error_regs->afsr & C_AFSR_ALL_ERRS);
6260 	aflt->flt_addr = cpu_error_regs->afar;
6261 #if defined(SERRANO)
6262 	ch_flt.afar2 = cpu_error_regs->afar2;
6263 #endif	/* SERRANO */
6264 	aflt->flt_pc = NULL;
6265 	aflt->flt_priv = ((cpu_error_regs->afsr & C_AFSR_PRIV) != 0);
6266 	aflt->flt_tl = 0;
6267 	aflt->flt_panic = 0;
6268 	cpu_log_and_clear_ce(&ch_flt);
6269 
6270 	/*
6271 	 * check if we caused any errors during cleanup
6272 	 */
6273 	if (clear_errors(&ch_flt)) {
6274 		pr_reason[0] = '\0';
6275 		(void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs,
6276 		    NULL);
6277 	}
6278 }
6279 
6280 /*
6281  * Log/clear CEEN-controlled disrupting errors
6282  */
6283 static void
6284 cpu_log_and_clear_ce(ch_async_flt_t *ch_flt)
6285 {
6286 	struct async_flt *aflt;
6287 	uint64_t afsr, afsr_errs;
6288 	ch_cpu_logout_t *clop;
6289 	char 		pr_reason[MAX_REASON_STRING];
6290 	on_trap_data_t	*otp = curthread->t_ontrap;
6291 
6292 	aflt = (struct async_flt *)ch_flt;
6293 	afsr = aflt->flt_stat;
6294 	afsr_errs = ch_flt->afsr_errs;
6295 	aflt->flt_id = gethrtime_waitfree();
6296 	aflt->flt_bus_id = getprocessorid();
6297 	aflt->flt_inst = CPU->cpu_id;
6298 	aflt->flt_prot = AFLT_PROT_NONE;
6299 	aflt->flt_class = CPU_FAULT;
6300 	aflt->flt_status = ECC_C_TRAP;
6301 
6302 	pr_reason[0] = '\0';
6303 	/*
6304 	 * Get the CPU log out info for Disrupting Trap.
6305 	 */
6306 	if (CPU_PRIVATE(CPU) == NULL) {
6307 		clop = NULL;
6308 		ch_flt->flt_diag_data.chd_afar = LOGOUT_INVALID;
6309 	} else {
6310 		clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout);
6311 	}
6312 
6313 	if (clop && ch_flt->flt_trapped_ce & CE_CEEN_TIMEOUT) {
6314 		ch_cpu_errors_t cpu_error_regs;
6315 
6316 		get_cpu_error_state(&cpu_error_regs);
6317 		(void) cpu_ce_delayed_ec_logout(cpu_error_regs.afar);
6318 		clop->clo_data.chd_afsr = cpu_error_regs.afsr;
6319 		clop->clo_data.chd_afar = cpu_error_regs.afar;
6320 		clop->clo_data.chd_afsr_ext = cpu_error_regs.afsr_ext;
6321 		clop->clo_sdw_data.chd_afsr = cpu_error_regs.shadow_afsr;
6322 		clop->clo_sdw_data.chd_afar = cpu_error_regs.shadow_afar;
6323 		clop->clo_sdw_data.chd_afsr_ext =
6324 		    cpu_error_regs.shadow_afsr_ext;
6325 #if defined(SERRANO)
6326 		clop->clo_data.chd_afar2 = cpu_error_regs.afar2;
6327 #endif	/* SERRANO */
6328 		ch_flt->flt_data_incomplete = 1;
6329 
6330 		/*
6331 		 * The logging/clear code expects AFSR/AFAR to be cleared.
6332 		 * The trap handler does it for CEEN enabled errors
6333 		 * so we need to do it here.
6334 		 */
6335 		set_cpu_error_state(&cpu_error_regs);
6336 	}
6337 
6338 #if defined(JALAPENO) || defined(SERRANO)
6339 	/*
6340 	 * FRC: Can't scrub memory as we don't have AFAR for Jalapeno.
6341 	 * For Serrano, even thou we do have the AFAR, we still do the
6342 	 * scrub on the RCE side since that's where the error type can
6343 	 * be properly classified as intermittent, persistent, etc.
6344 	 *
6345 	 * CE/RCE:  If error is in memory and AFAR is valid, scrub the memory.
6346 	 * Must scrub memory before cpu_queue_events, as scrubbing memory sets
6347 	 * the flt_status bits.
6348 	 */
6349 	if ((afsr & (C_AFSR_CE|C_AFSR_RCE)) &&
6350 	    (cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_CE)) ||
6351 	    cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_RCE)))) {
6352 		cpu_ce_scrub_mem_err(aflt, B_TRUE);
6353 	}
6354 #else /* JALAPENO || SERRANO */
6355 	/*
6356 	 * CE/EMC:  If error is in memory and AFAR is valid, scrub the memory.
6357 	 * Must scrub memory before cpu_queue_events, as scrubbing memory sets
6358 	 * the flt_status bits.
6359 	 */
6360 	if (afsr & (C_AFSR_CE|C_AFSR_EMC)) {
6361 		if (cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_CE)) ||
6362 		    cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_EMC))) {
6363 			cpu_ce_scrub_mem_err(aflt, B_TRUE);
6364 		}
6365 	}
6366 
6367 #endif /* JALAPENO || SERRANO */
6368 
6369 	/*
6370 	 * Update flt_prot if this error occurred under on_trap protection.
6371 	 */
6372 	if (otp != NULL && (otp->ot_prot & OT_DATA_EC))
6373 		aflt->flt_prot = AFLT_PROT_EC;
6374 
6375 	/*
6376 	 * Queue events on the async event queue, one event per error bit.
6377 	 */
6378 	if (cpu_queue_events(ch_flt, pr_reason, afsr_errs, clop) == 0 ||
6379 	    (afsr_errs & (C_AFSR_CECC_ERRS | C_AFSR_EXT_CECC_ERRS)) == 0) {
6380 		ch_flt->flt_type = CPU_INV_AFSR;
6381 		cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR,
6382 		    (void *)ch_flt, sizeof (ch_async_flt_t), ue_queue,
6383 		    aflt->flt_panic);
6384 	}
6385 
6386 	/*
6387 	 * Zero out + invalidate CPU logout.
6388 	 */
6389 	if (clop) {
6390 		bzero(clop, sizeof (ch_cpu_logout_t));
6391 		clop->clo_data.chd_afar = LOGOUT_INVALID;
6392 	}
6393 
6394 	/*
6395 	 * If either a CPC, WDC or EDC error has occurred while CEEN
6396 	 * was disabled, we need to flush either the entire
6397 	 * E$ or an E$ line.
6398 	 */
6399 #if defined(JALAPENO) || defined(SERRANO)
6400 	if (afsr & (C_AFSR_EDC | C_AFSR_CPC | C_AFSR_CPU | C_AFSR_WDC))
6401 #else	/* JALAPENO || SERRANO */
6402 	if (afsr_errs & (C_AFSR_EDC | C_AFSR_CPC | C_AFSR_WDC | C_AFSR_L3_EDC |
6403 	    C_AFSR_L3_CPC | C_AFSR_L3_WDC))
6404 #endif	/* JALAPENO || SERRANO */
6405 		cpu_error_ecache_flush(ch_flt);
6406 
6407 }
6408 
6409 /*
6410  * depending on the error type, we determine whether we
6411  * need to flush the entire ecache or just a line.
6412  */
6413 static int
6414 cpu_error_ecache_flush_required(ch_async_flt_t *ch_flt)
6415 {
6416 	struct async_flt *aflt;
6417 	uint64_t	afsr;
6418 	uint64_t	afsr_errs = ch_flt->afsr_errs;
6419 
6420 	aflt = (struct async_flt *)ch_flt;
6421 	afsr = aflt->flt_stat;
6422 
6423 	/*
6424 	 * If we got multiple errors, no point in trying
6425 	 * the individual cases, just flush the whole cache
6426 	 */
6427 	if (afsr & C_AFSR_ME) {
6428 		return (ECACHE_FLUSH_ALL);
6429 	}
6430 
6431 	/*
6432 	 * If either a CPC, WDC or EDC error has occurred while CEEN
6433 	 * was disabled, we need to flush entire E$. We can't just
6434 	 * flush the cache line affected as the ME bit
6435 	 * is not set when multiple correctable errors of the same
6436 	 * type occur, so we might have multiple CPC or EDC errors,
6437 	 * with only the first recorded.
6438 	 */
6439 #if defined(JALAPENO) || defined(SERRANO)
6440 	if (afsr & (C_AFSR_CPC | C_AFSR_CPU | C_AFSR_EDC | C_AFSR_WDC)) {
6441 #else	/* JALAPENO || SERRANO */
6442 	if (afsr_errs & (C_AFSR_CPC | C_AFSR_EDC | C_AFSR_WDC | C_AFSR_L3_CPC |
6443 	    C_AFSR_L3_EDC | C_AFSR_L3_WDC)) {
6444 #endif	/* JALAPENO || SERRANO */
6445 		return (ECACHE_FLUSH_ALL);
6446 	}
6447 
6448 #if defined(JALAPENO) || defined(SERRANO)
6449 	/*
6450 	 * If only UE or RUE is set, flush the Ecache line, otherwise
6451 	 * flush the entire Ecache.
6452 	 */
6453 	if (afsr & (C_AFSR_UE|C_AFSR_RUE)) {
6454 		if ((afsr & C_AFSR_ALL_ERRS) == C_AFSR_UE ||
6455 		    (afsr & C_AFSR_ALL_ERRS) == C_AFSR_RUE) {
6456 			return (ECACHE_FLUSH_LINE);
6457 		} else {
6458 			return (ECACHE_FLUSH_ALL);
6459 		}
6460 	}
6461 #else /* JALAPENO || SERRANO */
6462 	/*
6463 	 * If UE only is set, flush the Ecache line, otherwise
6464 	 * flush the entire Ecache.
6465 	 */
6466 	if (afsr_errs & C_AFSR_UE) {
6467 		if ((afsr_errs & (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) ==
6468 		    C_AFSR_UE) {
6469 			return (ECACHE_FLUSH_LINE);
6470 		} else {
6471 			return (ECACHE_FLUSH_ALL);
6472 		}
6473 	}
6474 #endif /* JALAPENO || SERRANO */
6475 
6476 	/*
6477 	 * EDU: If EDU only is set, flush the ecache line, otherwise
6478 	 * flush the entire Ecache.
6479 	 */
6480 	if (afsr_errs & (C_AFSR_EDU | C_AFSR_L3_EDU)) {
6481 		if (((afsr_errs & ~C_AFSR_EDU) == 0) ||
6482 		    ((afsr_errs & ~C_AFSR_L3_EDU) == 0)) {
6483 			return (ECACHE_FLUSH_LINE);
6484 		} else {
6485 			return (ECACHE_FLUSH_ALL);
6486 		}
6487 	}
6488 
6489 	/*
6490 	 * BERR: If BERR only is set, flush the Ecache line, otherwise
6491 	 * flush the entire Ecache.
6492 	 */
6493 	if (afsr_errs & C_AFSR_BERR) {
6494 		if ((afsr_errs & ~C_AFSR_BERR) == 0) {
6495 			return (ECACHE_FLUSH_LINE);
6496 		} else {
6497 			return (ECACHE_FLUSH_ALL);
6498 		}
6499 	}
6500 
6501 	return (0);
6502 }
6503 
6504 void
6505 cpu_error_ecache_flush(ch_async_flt_t *ch_flt)
6506 {
6507 	int	ecache_flush_flag =
6508 	    cpu_error_ecache_flush_required(ch_flt);
6509 
6510 	/*
6511 	 * Flush Ecache line or entire Ecache based on above checks.
6512 	 */
6513 	if (ecache_flush_flag == ECACHE_FLUSH_ALL)
6514 		cpu_flush_ecache();
6515 	else if (ecache_flush_flag == ECACHE_FLUSH_LINE) {
6516 		cpu_flush_ecache_line(ch_flt);
6517 	}
6518 
6519 }
6520 
6521 /*
6522  * Extract the PA portion from the E$ tag.
6523  */
6524 uint64_t
6525 cpu_ectag_to_pa(int setsize, uint64_t tag)
6526 {
6527 	if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation))
6528 		return (JG_ECTAG_TO_PA(setsize, tag));
6529 	else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation))
6530 		return (PN_L3TAG_TO_PA(tag));
6531 	else
6532 		return (CH_ECTAG_TO_PA(setsize, tag));
6533 }
6534 
6535 /*
6536  * Convert the E$ tag PA into an E$ subblock index.
6537  */
6538 static int
6539 cpu_ectag_pa_to_subblk(int cachesize, uint64_t subaddr)
6540 {
6541 	if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation))
6542 		return (JG_ECTAG_PA_TO_SUBBLK(cachesize, subaddr));
6543 	else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation))
6544 		/* Panther has only one subblock per line */
6545 		return (0);
6546 	else
6547 		return (CH_ECTAG_PA_TO_SUBBLK(cachesize, subaddr));
6548 }
6549 
6550 /*
6551  * All subblocks in an E$ line must be invalid for
6552  * the line to be invalid.
6553  */
6554 int
6555 cpu_ectag_line_invalid(int cachesize, uint64_t tag)
6556 {
6557 	if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation))
6558 		return (JG_ECTAG_LINE_INVALID(cachesize, tag));
6559 	else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation))
6560 		return (PN_L3_LINE_INVALID(tag));
6561 	else
6562 		return (CH_ECTAG_LINE_INVALID(cachesize, tag));
6563 }
6564 
6565 /*
6566  * Extract state bits for a subblock given the tag.  Note that for Panther
6567  * this works on both l2 and l3 tags.
6568  */
6569 static int
6570 cpu_ectag_pa_to_subblk_state(int cachesize, uint64_t subaddr, uint64_t tag)
6571 {
6572 	if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation))
6573 		return (JG_ECTAG_PA_TO_SUBBLK_STATE(cachesize, subaddr, tag));
6574 	else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation))
6575 		return (tag & CH_ECSTATE_MASK);
6576 	else
6577 		return (CH_ECTAG_PA_TO_SUBBLK_STATE(cachesize, subaddr, tag));
6578 }
6579 
6580 /*
6581  * Cpu specific initialization.
6582  */
6583 void
6584 cpu_mp_init(void)
6585 {
6586 #ifdef	CHEETAHPLUS_ERRATUM_25
6587 	if (cheetah_sendmondo_recover) {
6588 		cheetah_nudge_init();
6589 	}
6590 #endif
6591 }
6592 
6593 void
6594 cpu_ereport_post(struct async_flt *aflt)
6595 {
6596 	char *cpu_type, buf[FM_MAX_CLASS];
6597 	nv_alloc_t *nva = NULL;
6598 	nvlist_t *ereport, *detector, *resource;
6599 	errorq_elem_t *eqep;
6600 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
6601 	char unum[UNUM_NAMLEN];
6602 	int synd_code;
6603 	uint8_t msg_type;
6604 	plat_ecc_ch_async_flt_t	plat_ecc_ch_flt;
6605 
6606 	if (aflt->flt_panic || panicstr) {
6607 		eqep = errorq_reserve(ereport_errorq);
6608 		if (eqep == NULL)
6609 			return;
6610 		ereport = errorq_elem_nvl(ereport_errorq, eqep);
6611 		nva = errorq_elem_nva(ereport_errorq, eqep);
6612 	} else {
6613 		ereport = fm_nvlist_create(nva);
6614 	}
6615 
6616 	/*
6617 	 * Create the scheme "cpu" FMRI.
6618 	 */
6619 	detector = fm_nvlist_create(nva);
6620 	resource = fm_nvlist_create(nva);
6621 	switch (cpunodes[aflt->flt_inst].implementation) {
6622 	case CHEETAH_IMPL:
6623 		cpu_type = FM_EREPORT_CPU_USIII;
6624 		break;
6625 	case CHEETAH_PLUS_IMPL:
6626 		cpu_type = FM_EREPORT_CPU_USIIIplus;
6627 		break;
6628 	case JALAPENO_IMPL:
6629 		cpu_type = FM_EREPORT_CPU_USIIIi;
6630 		break;
6631 	case SERRANO_IMPL:
6632 		cpu_type = FM_EREPORT_CPU_USIIIiplus;
6633 		break;
6634 	case JAGUAR_IMPL:
6635 		cpu_type = FM_EREPORT_CPU_USIV;
6636 		break;
6637 	case PANTHER_IMPL:
6638 		cpu_type = FM_EREPORT_CPU_USIVplus;
6639 		break;
6640 	default:
6641 		cpu_type = FM_EREPORT_CPU_UNSUPPORTED;
6642 		break;
6643 	}
6644 
6645 	cpu_fmri_cpu_set(detector, aflt->flt_inst);
6646 
6647 	/*
6648 	 * Encode all the common data into the ereport.
6649 	 */
6650 	(void) snprintf(buf, FM_MAX_CLASS, "%s.%s.%s",
6651 		FM_ERROR_CPU, cpu_type, aflt->flt_erpt_class);
6652 
6653 	fm_ereport_set(ereport, FM_EREPORT_VERSION, buf,
6654 	    fm_ena_generate_cpu(aflt->flt_id, aflt->flt_inst, FM_ENA_FMT1),
6655 	    detector, NULL);
6656 
6657 	/*
6658 	 * Encode the error specific data that was saved in
6659 	 * the async_flt structure into the ereport.
6660 	 */
6661 	cpu_payload_add_aflt(aflt, ereport, resource,
6662 	    &plat_ecc_ch_flt.ecaf_afar_status,
6663 	    &plat_ecc_ch_flt.ecaf_synd_status);
6664 
6665 	if (aflt->flt_panic || panicstr) {
6666 		errorq_commit(ereport_errorq, eqep, ERRORQ_SYNC);
6667 	} else {
6668 		(void) fm_ereport_post(ereport, EVCH_TRYHARD);
6669 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
6670 		fm_nvlist_destroy(detector, FM_NVA_FREE);
6671 		fm_nvlist_destroy(resource, FM_NVA_FREE);
6672 	}
6673 	/*
6674 	 * Send the enhanced error information (plat_ecc_error2_data_t)
6675 	 * to the SC olny if it can process it.
6676 	 */
6677 
6678 	if (&plat_ecc_capability_sc_get &&
6679 	    plat_ecc_capability_sc_get(PLAT_ECC_ERROR2_MESSAGE)) {
6680 		msg_type = cpu_flt_bit_to_plat_error(aflt);
6681 		if (msg_type != PLAT_ECC_ERROR2_NONE) {
6682 			/*
6683 			 * If afar status is not invalid do a unum lookup.
6684 			 */
6685 			if (plat_ecc_ch_flt.ecaf_afar_status !=
6686 			    AFLT_STAT_INVALID) {
6687 				synd_code = synd_to_synd_code(
6688 				    plat_ecc_ch_flt.ecaf_synd_status,
6689 				    aflt->flt_synd, ch_flt->flt_bit);
6690 				(void) cpu_get_mem_unum_synd(synd_code,
6691 				    aflt, unum);
6692 			} else {
6693 				unum[0] = '\0';
6694 			}
6695 			plat_ecc_ch_flt.ecaf_sdw_afar = ch_flt->flt_sdw_afar;
6696 			plat_ecc_ch_flt.ecaf_sdw_afsr = ch_flt->flt_sdw_afsr;
6697 			plat_ecc_ch_flt.ecaf_afsr_ext = ch_flt->afsr_ext;
6698 			plat_ecc_ch_flt.ecaf_sdw_afsr_ext =
6699 			    ch_flt->flt_sdw_afsr_ext;
6700 
6701 			if (&plat_log_fruid_error2)
6702 				plat_log_fruid_error2(msg_type, unum, aflt,
6703 				    &plat_ecc_ch_flt);
6704 		}
6705 	}
6706 }
6707 
6708 void
6709 cpu_run_bus_error_handlers(struct async_flt *aflt, int expected)
6710 {
6711 	int status;
6712 	ddi_fm_error_t de;
6713 
6714 	bzero(&de, sizeof (ddi_fm_error_t));
6715 
6716 	de.fme_version = DDI_FME_VERSION;
6717 	de.fme_ena = fm_ena_generate_cpu(aflt->flt_id, aflt->flt_inst,
6718 	    FM_ENA_FMT1);
6719 	de.fme_flag = expected;
6720 	de.fme_bus_specific = (void *)aflt->flt_addr;
6721 	status = ndi_fm_handler_dispatch(ddi_root_node(), NULL, &de);
6722 	if ((aflt->flt_prot == AFLT_PROT_NONE) && (status == DDI_FM_FATAL))
6723 		aflt->flt_panic = 1;
6724 }
6725 
6726 void
6727 cpu_errorq_dispatch(char *error_class, void *payload, size_t payload_sz,
6728     errorq_t *eqp, uint_t flag)
6729 {
6730 	struct async_flt *aflt = (struct async_flt *)payload;
6731 
6732 	aflt->flt_erpt_class = error_class;
6733 	errorq_dispatch(eqp, payload, payload_sz, flag);
6734 }
6735 
6736 /*
6737  * This routine may be called by the IO module, but does not do
6738  * anything in this cpu module. The SERD algorithm is handled by
6739  * cpumem-diagnosis engine instead.
6740  */
6741 /*ARGSUSED*/
6742 void
6743 cpu_ce_count_unum(struct async_flt *ecc, int len, char *unum)
6744 {}
6745 
6746 void
6747 adjust_hw_copy_limits(int ecache_size)
6748 {
6749 	/*
6750 	 * Set hw copy limits.
6751 	 *
6752 	 * /etc/system will be parsed later and can override one or more
6753 	 * of these settings.
6754 	 *
6755 	 * At this time, ecache size seems only mildly relevant.
6756 	 * We seem to run into issues with the d-cache and stalls
6757 	 * we see on misses.
6758 	 *
6759 	 * Cycle measurement indicates that 2 byte aligned copies fare
6760 	 * little better than doing things with VIS at around 512 bytes.
6761 	 * 4 byte aligned shows promise until around 1024 bytes. 8 Byte
6762 	 * aligned is faster whenever the source and destination data
6763 	 * in cache and the total size is less than 2 Kbytes.  The 2K
6764 	 * limit seems to be driven by the 2K write cache.
6765 	 * When more than 2K of copies are done in non-VIS mode, stores
6766 	 * backup in the write cache.  In VIS mode, the write cache is
6767 	 * bypassed, allowing faster cache-line writes aligned on cache
6768 	 * boundaries.
6769 	 *
6770 	 * In addition, in non-VIS mode, there is no prefetching, so
6771 	 * for larger copies, the advantage of prefetching to avoid even
6772 	 * occasional cache misses is enough to justify using the VIS code.
6773 	 *
6774 	 * During testing, it was discovered that netbench ran 3% slower
6775 	 * when hw_copy_limit_8 was 2K or larger.  Apparently for server
6776 	 * applications, data is only used once (copied to the output
6777 	 * buffer, then copied by the network device off the system).  Using
6778 	 * the VIS copy saves more L2 cache state.  Network copies are
6779 	 * around 1.3K to 1.5K in size for historical reasons.
6780 	 *
6781 	 * Therefore, a limit of 1K bytes will be used for the 8 byte
6782 	 * aligned copy even for large caches and 8 MB ecache.  The
6783 	 * infrastructure to allow different limits for different sized
6784 	 * caches is kept to allow further tuning in later releases.
6785 	 */
6786 
6787 	if (min_ecache_size == 0 && use_hw_bcopy) {
6788 		/*
6789 		 * First time through - should be before /etc/system
6790 		 * is read.
6791 		 * Could skip the checks for zero but this lets us
6792 		 * preserve any debugger rewrites.
6793 		 */
6794 		if (hw_copy_limit_1 == 0) {
6795 			hw_copy_limit_1 = VIS_COPY_THRESHOLD;
6796 			priv_hcl_1 = hw_copy_limit_1;
6797 		}
6798 		if (hw_copy_limit_2 == 0) {
6799 			hw_copy_limit_2 = 2 * VIS_COPY_THRESHOLD;
6800 			priv_hcl_2 = hw_copy_limit_2;
6801 		}
6802 		if (hw_copy_limit_4 == 0) {
6803 			hw_copy_limit_4 = 4 * VIS_COPY_THRESHOLD;
6804 			priv_hcl_4 = hw_copy_limit_4;
6805 		}
6806 		if (hw_copy_limit_8 == 0) {
6807 			hw_copy_limit_8 = 4 * VIS_COPY_THRESHOLD;
6808 			priv_hcl_8 = hw_copy_limit_8;
6809 		}
6810 		min_ecache_size = ecache_size;
6811 	} else {
6812 		/*
6813 		 * MP initialization. Called *after* /etc/system has
6814 		 * been parsed. One CPU has already been initialized.
6815 		 * Need to cater for /etc/system having scragged one
6816 		 * of our values.
6817 		 */
6818 		if (ecache_size == min_ecache_size) {
6819 			/*
6820 			 * Same size ecache. We do nothing unless we
6821 			 * have a pessimistic ecache setting. In that
6822 			 * case we become more optimistic (if the cache is
6823 			 * large enough).
6824 			 */
6825 			if (hw_copy_limit_8 == 4 * VIS_COPY_THRESHOLD) {
6826 				/*
6827 				 * Need to adjust hw_copy_limit* from our
6828 				 * pessimistic uniprocessor value to a more
6829 				 * optimistic UP value *iff* it hasn't been
6830 				 * reset.
6831 				 */
6832 				if ((ecache_size > 1048576) &&
6833 				    (priv_hcl_8 == hw_copy_limit_8)) {
6834 					if (ecache_size <= 2097152)
6835 						hw_copy_limit_8 = 4 *
6836 						    VIS_COPY_THRESHOLD;
6837 					else if (ecache_size <= 4194304)
6838 						hw_copy_limit_8 = 4 *
6839 						    VIS_COPY_THRESHOLD;
6840 					else
6841 						hw_copy_limit_8 = 4 *
6842 						    VIS_COPY_THRESHOLD;
6843 					priv_hcl_8 = hw_copy_limit_8;
6844 				}
6845 			}
6846 		} else if (ecache_size < min_ecache_size) {
6847 			/*
6848 			 * A different ecache size. Can this even happen?
6849 			 */
6850 			if (priv_hcl_8 == hw_copy_limit_8) {
6851 				/*
6852 				 * The previous value that we set
6853 				 * is unchanged (i.e., it hasn't been
6854 				 * scragged by /etc/system). Rewrite it.
6855 				 */
6856 				if (ecache_size <= 1048576)
6857 					hw_copy_limit_8 = 8 *
6858 					    VIS_COPY_THRESHOLD;
6859 				else if (ecache_size <= 2097152)
6860 					hw_copy_limit_8 = 8 *
6861 					    VIS_COPY_THRESHOLD;
6862 				else if (ecache_size <= 4194304)
6863 					hw_copy_limit_8 = 8 *
6864 					    VIS_COPY_THRESHOLD;
6865 				else
6866 					hw_copy_limit_8 = 10 *
6867 					    VIS_COPY_THRESHOLD;
6868 				priv_hcl_8 = hw_copy_limit_8;
6869 				min_ecache_size = ecache_size;
6870 			}
6871 		}
6872 	}
6873 }
6874 
6875 /*
6876  * Called from illegal instruction trap handler to see if we can attribute
6877  * the trap to a fpras check.
6878  */
6879 int
6880 fpras_chktrap(struct regs *rp)
6881 {
6882 	int op;
6883 	struct fpras_chkfngrp *cgp;
6884 	uintptr_t tpc = (uintptr_t)rp->r_pc;
6885 
6886 	if (fpras_chkfngrps == NULL)
6887 		return (0);
6888 
6889 	cgp = &fpras_chkfngrps[CPU->cpu_id];
6890 	for (op = 0; op < FPRAS_NCOPYOPS; ++op) {
6891 		if (tpc >= (uintptr_t)&cgp->fpras_fn[op].fpras_blk0 &&
6892 		    tpc < (uintptr_t)&cgp->fpras_fn[op].fpras_chkresult)
6893 			break;
6894 	}
6895 	if (op == FPRAS_NCOPYOPS)
6896 		return (0);
6897 
6898 	/*
6899 	 * This is an fpRAS failure caught through an illegal
6900 	 * instruction - trampoline.
6901 	 */
6902 	rp->r_pc = (uintptr_t)&cgp->fpras_fn[op].fpras_trampoline;
6903 	rp->r_npc = rp->r_pc + 4;
6904 	return (1);
6905 }
6906 
6907 /*
6908  * fpras_failure is called when a fpras check detects a bad calculation
6909  * result or an illegal instruction trap is attributed to an fpras
6910  * check.  In all cases we are still bound to CPU.
6911  */
6912 int
6913 fpras_failure(int op, int how)
6914 {
6915 	int use_hw_bcopy_orig, use_hw_bzero_orig;
6916 	uint_t hcl1_orig, hcl2_orig, hcl4_orig, hcl8_orig;
6917 	ch_async_flt_t ch_flt;
6918 	struct async_flt *aflt = (struct async_flt *)&ch_flt;
6919 	struct fpras_chkfn *sfp, *cfp;
6920 	uint32_t *sip, *cip;
6921 	int i;
6922 
6923 	/*
6924 	 * We're running on a sick CPU.  Avoid further FPU use at least for
6925 	 * the time in which we dispatch an ereport and (if applicable) panic.
6926 	 */
6927 	use_hw_bcopy_orig = use_hw_bcopy;
6928 	use_hw_bzero_orig = use_hw_bzero;
6929 	hcl1_orig = hw_copy_limit_1;
6930 	hcl2_orig = hw_copy_limit_2;
6931 	hcl4_orig = hw_copy_limit_4;
6932 	hcl8_orig = hw_copy_limit_8;
6933 	use_hw_bcopy = use_hw_bzero = 0;
6934 	hw_copy_limit_1 = hw_copy_limit_2 = hw_copy_limit_4 =
6935 	    hw_copy_limit_8 = 0;
6936 
6937 	bzero(&ch_flt, sizeof (ch_async_flt_t));
6938 	aflt->flt_id = gethrtime_waitfree();
6939 	aflt->flt_class = CPU_FAULT;
6940 	aflt->flt_inst = CPU->cpu_id;
6941 	aflt->flt_status = (how << 8) | op;
6942 	aflt->flt_payload = FM_EREPORT_PAYLOAD_FPU_HWCOPY;
6943 	ch_flt.flt_type = CPU_FPUERR;
6944 
6945 	/*
6946 	 * We must panic if the copy operation had no lofault protection -
6947 	 * ie, don't panic for copyin, copyout, kcopy and bcopy called
6948 	 * under on_fault and do panic for unprotected bcopy and hwblkpagecopy.
6949 	 */
6950 	aflt->flt_panic = (curthread->t_lofault == NULL);
6951 
6952 	/*
6953 	 * XOR the source instruction block with the copied instruction
6954 	 * block - this will show us which bit(s) are corrupted.
6955 	 */
6956 	sfp = (struct fpras_chkfn *)fpras_chkfn_type1;
6957 	cfp = &fpras_chkfngrps[CPU->cpu_id].fpras_fn[op];
6958 	if (op == FPRAS_BCOPY || op == FPRAS_COPYOUT) {
6959 		sip = &sfp->fpras_blk0[0];
6960 		cip = &cfp->fpras_blk0[0];
6961 	} else {
6962 		sip = &sfp->fpras_blk1[0];
6963 		cip = &cfp->fpras_blk1[0];
6964 	}
6965 	for (i = 0; i < 16; ++i, ++sip, ++cip)
6966 		ch_flt.flt_fpdata[i] = *sip ^ *cip;
6967 
6968 	cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_FPU_HWCOPY, (void *)&ch_flt,
6969 	    sizeof (ch_async_flt_t), ue_queue, aflt->flt_panic);
6970 
6971 	if (aflt->flt_panic)
6972 		fm_panic("FPU failure on CPU %d", CPU->cpu_id);
6973 
6974 	/*
6975 	 * We get here for copyin/copyout and kcopy or bcopy where the
6976 	 * caller has used on_fault.  We will flag the error so that
6977 	 * the process may be killed  The trap_async_hwerr mechanism will
6978 	 * take appropriate further action (such as a reboot, contract
6979 	 * notification etc).  Since we may be continuing we will
6980 	 * restore the global hardware copy acceleration switches.
6981 	 *
6982 	 * When we return from this function to the copy function we want to
6983 	 * avoid potentially bad data being used, ie we want the affected
6984 	 * copy function to return an error.  The caller should therefore
6985 	 * invoke its lofault handler (which always exists for these functions)
6986 	 * which will return the appropriate error.
6987 	 */
6988 	ttolwp(curthread)->lwp_pcb.pcb_flags |= ASYNC_HWERR;
6989 	aston(curthread);
6990 
6991 	use_hw_bcopy = use_hw_bcopy_orig;
6992 	use_hw_bzero = use_hw_bzero_orig;
6993 	hw_copy_limit_1 = hcl1_orig;
6994 	hw_copy_limit_2 = hcl2_orig;
6995 	hw_copy_limit_4 = hcl4_orig;
6996 	hw_copy_limit_8 = hcl8_orig;
6997 
6998 	return (1);
6999 }
7000 
7001 #define	VIS_BLOCKSIZE		64
7002 
7003 int
7004 dtrace_blksuword32_err(uintptr_t addr, uint32_t *data)
7005 {
7006 	int ret, watched;
7007 
7008 	watched = watch_disable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
7009 	ret = dtrace_blksuword32(addr, data, 0);
7010 	if (watched)
7011 		watch_enable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
7012 
7013 	return (ret);
7014 }
7015 
7016 /*
7017  * Called when a cpu enters the CPU_FAULTED state (by the cpu placing the
7018  * faulted cpu into that state).  Cross-trap to the faulted cpu to clear
7019  * CEEN from the EER to disable traps for further disrupting error types
7020  * on that cpu.  We could cross-call instead, but that has a larger
7021  * instruction and data footprint than cross-trapping, and the cpu is known
7022  * to be faulted.
7023  */
7024 
7025 void
7026 cpu_faulted_enter(struct cpu *cp)
7027 {
7028 	xt_one(cp->cpu_id, set_error_enable_tl1, EN_REG_CEEN, EER_SET_CLRBITS);
7029 }
7030 
7031 /*
7032  * Called when a cpu leaves the CPU_FAULTED state to return to one of
7033  * offline, spare, or online (by the cpu requesting this state change).
7034  * First we cross-call to clear the AFSR (and AFSR_EXT on Panther) of
7035  * disrupting error bits that have accumulated without trapping, then
7036  * we cross-trap to re-enable CEEN controlled traps.
7037  */
7038 void
7039 cpu_faulted_exit(struct cpu *cp)
7040 {
7041 	ch_cpu_errors_t cpu_error_regs;
7042 
7043 	cpu_error_regs.afsr = C_AFSR_CECC_ERRS;
7044 	if (IS_PANTHER(cpunodes[cp->cpu_id].implementation))
7045 		cpu_error_regs.afsr_ext &= C_AFSR_EXT_CECC_ERRS;
7046 	xc_one(cp->cpu_id, (xcfunc_t *)set_cpu_error_state,
7047 	    (uint64_t)&cpu_error_regs, 0);
7048 
7049 	xt_one(cp->cpu_id, set_error_enable_tl1, EN_REG_CEEN, EER_SET_SETBITS);
7050 }
7051 
7052 /*
7053  * Return 1 if the errors in ch_flt's AFSR are secondary errors caused by
7054  * the errors in the original AFSR, 0 otherwise.
7055  *
7056  * For all procs if the initial error was a BERR or TO, then it is possible
7057  * that we may have caused a secondary BERR or TO in the process of logging the
7058  * inital error via cpu_run_bus_error_handlers().  If this is the case then
7059  * if the request was protected then a panic is still not necessary, if not
7060  * protected then aft_panic is already set - so either way there's no need
7061  * to set aft_panic for the secondary error.
7062  *
7063  * For Cheetah and Jalapeno, if the original error was a UE which occurred on
7064  * a store merge, then the error handling code will call cpu_deferred_error().
7065  * When clear_errors() is called, it will determine that secondary errors have
7066  * occurred - in particular, the store merge also caused a EDU and WDU that
7067  * weren't discovered until this point.
7068  *
7069  * We do three checks to verify that we are in this case.  If we pass all three
7070  * checks, we return 1 to indicate that we should not panic.  If any unexpected
7071  * errors occur, we return 0.
7072  *
7073  * For Cheetah+ and derivative procs, the store merge causes a DUE, which is
7074  * handled in cpu_disrupting_errors().  Since this function is not even called
7075  * in the case we are interested in, we just return 0 for these processors.
7076  */
7077 /*ARGSUSED*/
7078 static int
7079 cpu_check_secondary_errors(ch_async_flt_t *ch_flt, uint64_t t_afsr_errs,
7080     uint64_t t_afar)
7081 {
7082 #if defined(CHEETAH_PLUS)
7083 #else	/* CHEETAH_PLUS */
7084 	struct async_flt *aflt = (struct async_flt *)ch_flt;
7085 #endif	/* CHEETAH_PLUS */
7086 
7087 	/*
7088 	 * Was the original error a BERR or TO and only a BERR or TO
7089 	 * (multiple errors are also OK)
7090 	 */
7091 	if ((t_afsr_errs & ~(C_AFSR_BERR | C_AFSR_TO | C_AFSR_ME)) == 0) {
7092 		/*
7093 		 * Is the new error a BERR or TO and only a BERR or TO
7094 		 * (multiple errors are also OK)
7095 		 */
7096 		if ((ch_flt->afsr_errs &
7097 		    ~(C_AFSR_BERR | C_AFSR_TO | C_AFSR_ME)) == 0)
7098 			return (1);
7099 	}
7100 
7101 #if defined(CHEETAH_PLUS)
7102 	return (0);
7103 #else	/* CHEETAH_PLUS */
7104 	/*
7105 	 * Now look for secondary effects of a UE on cheetah/jalapeno
7106 	 *
7107 	 * Check the original error was a UE, and only a UE.  Note that
7108 	 * the ME bit will cause us to fail this check.
7109 	 */
7110 	if (t_afsr_errs != C_AFSR_UE)
7111 		return (0);
7112 
7113 	/*
7114 	 * Check the secondary errors were exclusively an EDU and/or WDU.
7115 	 */
7116 	if ((ch_flt->afsr_errs & ~(C_AFSR_EDU|C_AFSR_WDU)) != 0)
7117 		return (0);
7118 
7119 	/*
7120 	 * Check the AFAR of the original error and secondary errors
7121 	 * match to the 64-byte boundary
7122 	 */
7123 	if (P2ALIGN(aflt->flt_addr, 64) != P2ALIGN(t_afar, 64))
7124 		return (0);
7125 
7126 	/*
7127 	 * We've passed all the checks, so it's a secondary error!
7128 	 */
7129 	return (1);
7130 #endif	/* CHEETAH_PLUS */
7131 }
7132 
7133 /*
7134  * Translate the flt_bit or flt_type into an error type.  First, flt_bit
7135  * is checked for any valid errors.  If found, the error type is
7136  * returned. If not found, the flt_type is checked for L1$ parity errors.
7137  */
7138 /*ARGSUSED*/
7139 static uint8_t
7140 cpu_flt_bit_to_plat_error(struct async_flt *aflt)
7141 {
7142 #if defined(JALAPENO)
7143 	/*
7144 	 * Currently, logging errors to the SC is not supported on Jalapeno
7145 	 */
7146 	return (PLAT_ECC_ERROR2_NONE);
7147 #else
7148 	ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt;
7149 
7150 	switch (ch_flt->flt_bit) {
7151 	case C_AFSR_CE:
7152 		return (PLAT_ECC_ERROR2_CE);
7153 	case C_AFSR_UCC:
7154 	case C_AFSR_EDC:
7155 	case C_AFSR_WDC:
7156 	case C_AFSR_CPC:
7157 		return (PLAT_ECC_ERROR2_L2_CE);
7158 	case C_AFSR_EMC:
7159 		return (PLAT_ECC_ERROR2_EMC);
7160 	case C_AFSR_IVC:
7161 		return (PLAT_ECC_ERROR2_IVC);
7162 	case C_AFSR_UE:
7163 		return (PLAT_ECC_ERROR2_UE);
7164 	case C_AFSR_UCU:
7165 	case C_AFSR_EDU:
7166 	case C_AFSR_WDU:
7167 	case C_AFSR_CPU:
7168 		return (PLAT_ECC_ERROR2_L2_UE);
7169 	case C_AFSR_IVU:
7170 		return (PLAT_ECC_ERROR2_IVU);
7171 	case C_AFSR_TO:
7172 		return (PLAT_ECC_ERROR2_TO);
7173 	case C_AFSR_BERR:
7174 		return (PLAT_ECC_ERROR2_BERR);
7175 #if defined(CHEETAH_PLUS)
7176 	case C_AFSR_L3_EDC:
7177 	case C_AFSR_L3_UCC:
7178 	case C_AFSR_L3_CPC:
7179 	case C_AFSR_L3_WDC:
7180 		return (PLAT_ECC_ERROR2_L3_CE);
7181 	case C_AFSR_IMC:
7182 		return (PLAT_ECC_ERROR2_IMC);
7183 	case C_AFSR_TSCE:
7184 		return (PLAT_ECC_ERROR2_L2_TSCE);
7185 	case C_AFSR_THCE:
7186 		return (PLAT_ECC_ERROR2_L2_THCE);
7187 	case C_AFSR_L3_MECC:
7188 		return (PLAT_ECC_ERROR2_L3_MECC);
7189 	case C_AFSR_L3_THCE:
7190 		return (PLAT_ECC_ERROR2_L3_THCE);
7191 	case C_AFSR_L3_CPU:
7192 	case C_AFSR_L3_EDU:
7193 	case C_AFSR_L3_UCU:
7194 	case C_AFSR_L3_WDU:
7195 		return (PLAT_ECC_ERROR2_L3_UE);
7196 	case C_AFSR_DUE:
7197 		return (PLAT_ECC_ERROR2_DUE);
7198 	case C_AFSR_DTO:
7199 		return (PLAT_ECC_ERROR2_DTO);
7200 	case C_AFSR_DBERR:
7201 		return (PLAT_ECC_ERROR2_DBERR);
7202 #endif	/* CHEETAH_PLUS */
7203 	default:
7204 		switch (ch_flt->flt_type) {
7205 #if defined(CPU_IMP_L1_CACHE_PARITY)
7206 		case CPU_IC_PARITY:
7207 			return (PLAT_ECC_ERROR2_IPE);
7208 		case CPU_DC_PARITY:
7209 			if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) {
7210 				if (ch_flt->parity_data.dpe.cpl_cache ==
7211 				    CPU_PC_PARITY) {
7212 					return (PLAT_ECC_ERROR2_PCACHE);
7213 				}
7214 			}
7215 			return (PLAT_ECC_ERROR2_DPE);
7216 #endif /* CPU_IMP_L1_CACHE_PARITY */
7217 		case CPU_ITLB_PARITY:
7218 			return (PLAT_ECC_ERROR2_ITLB);
7219 		case CPU_DTLB_PARITY:
7220 			return (PLAT_ECC_ERROR2_DTLB);
7221 		default:
7222 			return (PLAT_ECC_ERROR2_NONE);
7223 		}
7224 	}
7225 #endif	/* JALAPENO */
7226 }
7227