xref: /illumos-gate/usr/src/uts/sun4u/serengeti/os/serengeti.c (revision 10a4fa49f51ed9ae1c857a626de6ce9ebf41661a)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/time.h>
30 #include <sys/cpuvar.h>
31 #include <sys/dditypes.h>
32 #include <sys/ddipropdefs.h>
33 #include <sys/ddi_impldefs.h>
34 #include <sys/sunddi.h>
35 #include <sys/esunddi.h>
36 #include <sys/sunndi.h>
37 #include <sys/platform_module.h>
38 #include <sys/errno.h>
39 #include <sys/conf.h>
40 #include <sys/modctl.h>
41 #include <sys/promif.h>
42 #include <sys/promimpl.h>
43 #include <sys/prom_plat.h>
44 #include <sys/cmn_err.h>
45 #include <sys/sysmacros.h>
46 #include <sys/mem_cage.h>
47 #include <sys/kobj.h>
48 #include <sys/utsname.h>
49 #include <sys/cpu_sgnblk_defs.h>
50 #include <sys/atomic.h>
51 #include <sys/kdi_impl.h>
52 
53 #include <sys/sgsbbc.h>
54 #include <sys/sgsbbc_iosram.h>
55 #include <sys/sgsbbc_iosram_priv.h>
56 #include <sys/sgsbbc_mailbox.h>
57 #include <sys/sgsgn.h>
58 #include <sys/sgcn.h>
59 #include <sys/serengeti.h>
60 #include <sys/sgfrutypes.h>
61 #include <sys/machsystm.h>
62 #include <sys/sbd_ioctl.h>
63 #include <sys/sbd.h>
64 #include <sys/sbdp_mem.h>
65 
66 #include <sys/memnode.h>
67 #include <vm/vm_dep.h>
68 #include <vm/page.h>
69 
70 #include <sys/cheetahregs.h>
71 #include <sys/plat_ecc_unum.h>
72 #include <sys/plat_ecc_dimm.h>
73 
74 #include <sys/lgrp.h>
75 
76 static int sg_debug = 0;
77 
78 #ifdef DEBUG
79 #define	DCMNERR if (sg_debug) cmn_err
80 #else
81 #define	DCMNERR
82 #endif
83 
84 int (*p2get_mem_unum)(int, uint64_t, char *, int, int *);
85 
86 /* local functions */
87 static void cpu_sgn_update(ushort_t sgn, uchar_t state,
88     uchar_t sub_state, int cpuid);
89 
90 
91 /*
92  * Local data.
93  *
94  * iosram_write_ptr is a pointer to iosram_write().  Because of
95  * kernel dynamic linking, we can't get to the function by name,
96  * but we can look up its address, and store it in this variable
97  * instead.
98  *
99  * We include the extern for iosram_write() here not because we call
100  * it, but to force compilation errors if its prototype doesn't
101  * match the prototype of iosram_write_ptr.
102  *
103  * The same issues apply to iosram_read() and iosram_read_ptr.
104  */
105 /*CSTYLED*/
106 extern int   iosram_write     (int, uint32_t, caddr_t, uint32_t);
107 static int (*iosram_write_ptr)(int, uint32_t, caddr_t, uint32_t) = NULL;
108 /*CSTYLED*/
109 extern int   iosram_read     (int, uint32_t, caddr_t, uint32_t);
110 static int (*iosram_read_ptr)(int, uint32_t, caddr_t, uint32_t) = NULL;
111 
112 
113 /*
114  * Variable to indicate if the date should be obtained from the SC or not.
115  */
116 int todsg_use_sc = FALSE;	/* set the false at the beginning */
117 
118 /*
119  * Preallocation of spare tsb's for DR
120  *
121  * We don't allocate spares for Wildcat since TSBs should come
122  * out of memory local to the node.
123  */
124 #define	IOMMU_PER_SCHIZO	2
125 int serengeti_tsb_spares = (SG_MAX_IO_BDS * SG_SCHIZO_PER_IO_BD *
126 	IOMMU_PER_SCHIZO);
127 
128 /*
129  * sg_max_ncpus is the maximum number of CPUs supported on Serengeti
130  * and Wildcat at GA.  We assume that the maximum number of SSM nodes
131  * supported at GA is 4.  sg_max_ncpus is set to be smaller than NCPU
132  * to reduce the amount of memory the logs take up until we have a
133  * dynamic log memory allocation solution.
134  */
135 int sg_max_ncpus = (24 * 4);	/* (CPUs per node * max number of nodes) */
136 
137 /*
138  * variables to control mailbox message timeouts.
139  * These can be patched via /etc/system or mdb.
140  */
141 int	sbbc_mbox_default_timeout = MBOX_DEFAULT_TIMEOUT;
142 int	sbbc_mbox_min_timeout = MBOX_MIN_TIMEOUT;
143 
144 /* cached 'chosen' node_id */
145 pnode_t chosen_nodeid = (pnode_t)0;
146 
147 static void (*sg_ecc_taskq_func)(sbbc_ecc_mbox_t *) = NULL;
148 static int (*sg_ecc_mbox_func)(sbbc_ecc_mbox_t *) = NULL;
149 
150 /*
151  * Table that maps memory slices to a specific memnode.
152  */
153 int slice_to_memnode[SG_MAX_SLICE];
154 
155 plat_dimm_sid_board_t	domain_dimm_sids[SG_MAX_CPU_BDS];
156 
157 
158 int
159 set_platform_tsb_spares()
160 {
161 	return (MIN(serengeti_tsb_spares, MAX_UPA));
162 }
163 
164 #pragma weak mmu_init_large_pages
165 
166 void
167 set_platform_defaults(void)
168 {
169 	extern int watchdog_enable;
170 	extern uint64_t xc_tick_limit_scale;
171 	extern void mmu_init_large_pages(size_t);
172 
173 #ifdef DEBUG
174 	char *todsg_name = "todsg";
175 	ce_verbose_memory = 2;
176 	ce_verbose_other = 2;
177 #endif /* DEBUG */
178 
179 	watchdog_enable = TRUE;
180 	watchdog_available = TRUE;
181 
182 	cpu_sgn_func = cpu_sgn_update;
183 
184 #ifdef DEBUG
185 	/* tod_module_name should be set to "todsg" from OBP property */
186 	if (tod_module_name && (strcmp(tod_module_name, todsg_name) == 0))
187 		prom_printf("Using todsg driver\n");
188 	else {
189 		prom_printf("Force using todsg driver\n");
190 		tod_module_name = todsg_name;
191 	}
192 #endif /* DEBUG */
193 
194 	/* Serengeti does not support forthdebug */
195 	forthdebug_supported = 0;
196 
197 
198 	/*
199 	 * Some DR operations require the system to be sync paused.
200 	 * Sync pause on Serengeti could potentially take up to 4
201 	 * seconds to complete depending on the load on the SC.  To
202 	 * avoid send_mond panics during such operations, we need to
203 	 * increase xc_tick_limit to a larger value on Serengeti by
204 	 * setting xc_tick_limit_scale to 5.
205 	 */
206 	xc_tick_limit_scale = 5;
207 
208 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
209 	    (mmu_ism_pagesize != MMU_PAGESIZE32M)) {
210 		if (&mmu_init_large_pages)
211 			mmu_init_large_pages(mmu_ism_pagesize);
212 	}
213 }
214 
215 void
216 load_platform_modules(void)
217 {
218 	if (modload("misc", "pcihp") < 0) {
219 		cmn_err(CE_NOTE, "pcihp driver failed to load");
220 	}
221 }
222 
223 /*ARGSUSED*/
224 int
225 plat_cpu_poweron(struct cpu *cp)
226 {
227 	int (*serengeti_cpu_poweron)(struct cpu *) = NULL;
228 
229 	serengeti_cpu_poweron =
230 	    (int (*)(struct cpu *))modgetsymvalue("sbdp_cpu_poweron", 0);
231 
232 	if (serengeti_cpu_poweron == NULL)
233 		return (ENOTSUP);
234 	else
235 		return ((serengeti_cpu_poweron)(cp));
236 }
237 
238 /*ARGSUSED*/
239 int
240 plat_cpu_poweroff(struct cpu *cp)
241 {
242 	int (*serengeti_cpu_poweroff)(struct cpu *) = NULL;
243 
244 	serengeti_cpu_poweroff =
245 	    (int (*)(struct cpu *))modgetsymvalue("sbdp_cpu_poweroff", 0);
246 
247 	if (serengeti_cpu_poweroff == NULL)
248 		return (ENOTSUP);
249 	else
250 		return ((serengeti_cpu_poweroff)(cp));
251 }
252 
253 #ifdef DEBUG
254 pgcnt_t serengeti_cage_size_limit;
255 #endif
256 
257 /* Preferred minimum cage size (expressed in pages)... for DR */
258 pgcnt_t serengeti_minimum_cage_size = 0;
259 
260 void
261 set_platform_cage_params(void)
262 {
263 	extern pgcnt_t total_pages;
264 	extern struct memlist *phys_avail;
265 	int ret;
266 
267 	if (kernel_cage_enable) {
268 		pgcnt_t preferred_cage_size;
269 
270 		preferred_cage_size =
271 		    MAX(serengeti_minimum_cage_size, total_pages / 256);
272 #ifdef DEBUG
273 		if (serengeti_cage_size_limit)
274 			preferred_cage_size = serengeti_cage_size_limit;
275 #endif
276 		kcage_range_lock();
277 		/*
278 		 * Post copies obp into the lowest slice.  This requires the
279 		 * cage to grow upwards
280 		 */
281 		ret = kcage_range_init(phys_avail, 0);
282 		if (ret == 0)
283 			kcage_init(preferred_cage_size);
284 		kcage_range_unlock();
285 	}
286 
287 	/* Only note when the cage is off since it should always be on. */
288 	if (!kcage_on)
289 		cmn_err(CE_NOTE, "!DR Kernel Cage is DISABLED");
290 }
291 
292 #define	ALIGN(x, a)	((a) == 0 ? (uint64_t)(x) : \
293 	(((uint64_t)(x) + (uint64_t)(a) - 1l) & ~((uint64_t)(a) - 1l)))
294 
295 void
296 update_mem_bounds(int brd, uint64_t base, uint64_t sz)
297 {
298 	uint64_t	end;
299 	int		mnode;
300 
301 	end = base + sz - 1;
302 
303 	/*
304 	 * First see if this board already has a memnode associated
305 	 * with it.  If not, see if this slice has a memnode.  This
306 	 * covers the cases where a single slice covers multiple
307 	 * boards (cross-board interleaving) and where a single
308 	 * board has multiple slices (1+GB DIMMs).
309 	 */
310 	if ((mnode = plat_lgrphand_to_mem_node(brd)) == -1) {
311 		if ((mnode = slice_to_memnode[PA_2_SLICE(base)]) == -1)
312 			mnode = mem_node_alloc();
313 		plat_assign_lgrphand_to_mem_node(brd, mnode);
314 	}
315 
316 	/*
317 	 * Align base at 16GB boundary
318 	 */
319 	base = ALIGN(base, (1ul << PA_SLICE_SHIFT));
320 
321 	while (base < end) {
322 		slice_to_memnode[PA_2_SLICE(base)] = mnode;
323 		base += (1ul << PA_SLICE_SHIFT);
324 	}
325 }
326 
327 /*
328  * Dynamically detect memory slices in the system by decoding
329  * the cpu memory decoder registers at boot time.
330  */
331 void
332 plat_fill_mc(pnode_t nodeid)
333 {
334 	uint64_t	mc_addr, mask;
335 	uint64_t	mc_decode[SG_MAX_BANKS_PER_MC];
336 	uint64_t	base, size;
337 	uint32_t	regs[4];
338 	int		len;
339 	int		local_mc;
340 	int		portid;
341 	int		boardid;
342 	int		i;
343 
344 	if ((prom_getprop(nodeid, "portid", (caddr_t)&portid) < 0) ||
345 	    (portid == -1))
346 		return;
347 
348 	/*
349 	 * Decode the board number from the MC portid
350 	 */
351 	boardid = SG_PORTID_TO_BOARD_NUM(portid);
352 
353 	/*
354 	 * The "reg" property returns 4 32-bit values. The first two are
355 	 * combined to form a 64-bit address.  The second two are for a
356 	 * 64-bit size, but we don't actually need to look at that value.
357 	 */
358 	len = prom_getproplen(nodeid, "reg");
359 	if (len != (sizeof (uint32_t) * 4)) {
360 		prom_printf("Warning: malformed 'reg' property\n");
361 		return;
362 	}
363 	if (prom_getprop(nodeid, "reg", (caddr_t)regs) < 0)
364 		return;
365 	mc_addr = ((uint64_t)regs[0]) << 32;
366 	mc_addr |= (uint64_t)regs[1];
367 
368 	/*
369 	 * Figure out whether the memory controller we are examining
370 	 * belongs to this CPU or a different one.
371 	 */
372 	if (portid == cpunodes[CPU->cpu_id].portid)
373 		local_mc = 1;
374 	else
375 		local_mc = 0;
376 
377 	for (i = 0; i < SG_MAX_BANKS_PER_MC; i++) {
378 		mask = SG_REG_2_OFFSET(i);
379 
380 		/*
381 		 * If the memory controller is local to this CPU, we use
382 		 * the special ASI to read the decode registers.
383 		 * Otherwise, we load the values from a magic address in
384 		 * I/O space.
385 		 */
386 		if (local_mc)
387 			mc_decode[i] = lddmcdecode(mask & MC_OFFSET_MASK);
388 		else
389 			mc_decode[i] = lddphysio((mc_addr | mask));
390 
391 		if (mc_decode[i] >> MC_VALID_SHIFT) {
392 			/*
393 			 * The memory decode register is a bitmask field,
394 			 * so we can decode that into both a base and
395 			 * a span.
396 			 */
397 			base = MC_BASE(mc_decode[i]) << PHYS2UM_SHIFT;
398 			size = MC_UK2SPAN(mc_decode[i]);
399 			update_mem_bounds(boardid, base, size);
400 		}
401 	}
402 }
403 
404 /*
405  * This routine is run midway through the boot process.  By the time we get
406  * here, we know about all the active CPU boards in the system, and we have
407  * extracted information about each board's memory from the memory
408  * controllers.  We have also figured out which ranges of memory will be
409  * assigned to which memnodes, so we walk the slice table to build the table
410  * of memnodes.
411  */
412 /* ARGSUSED */
413 void
414 plat_build_mem_nodes(u_longlong_t *list, size_t  nelems)
415 {
416 	int	slice;
417 	pfn_t	basepfn;
418 	pgcnt_t	npgs;
419 
420 	mem_node_pfn_shift = PFN_SLICE_SHIFT;
421 	mem_node_physalign = (1ull << PA_SLICE_SHIFT);
422 
423 	for (slice = 0; slice < SG_MAX_SLICE; slice++) {
424 		if (slice_to_memnode[slice] == -1)
425 			continue;
426 		basepfn = (uint64_t)slice << PFN_SLICE_SHIFT;
427 		npgs = 1ull << PFN_SLICE_SHIFT;
428 		mem_node_add_slice(basepfn, basepfn + npgs - 1);
429 	}
430 }
431 
432 int
433 plat_pfn_to_mem_node(pfn_t pfn)
434 {
435 	int node;
436 
437 	node = slice_to_memnode[PFN_2_SLICE(pfn)];
438 
439 	ASSERT(node >= 0);
440 	return (node);
441 }
442 
443 /*
444  * Serengeti support for lgroups.
445  *
446  * On Serengeti, an lgroup platform handle == board number.
447  *
448  * Mappings between lgroup handles and memnodes are managed
449  * in addition to mappings between memory slices and memnodes
450  * to support cross-board interleaving as well as multiple
451  * slices per board (e.g. >1GB DIMMs). The initial mapping
452  * of memnodes to lgroup handles is determined at boot time.
453  * A DR addition of memory adds a new mapping. A DR copy-rename
454  * swaps mappings.
455  */
456 
457 /*
458  * Macro for extracting the board number from the CPU id
459  */
460 #define	CPUID_TO_BOARD(id)	(((id) >> 2) & 0x7)
461 
462 /*
463  * Return the platform handle for the lgroup containing the given CPU
464  *
465  * For Serengeti, lgroup platform handle == board number
466  */
467 lgrp_handle_t
468 plat_lgrp_cpu_to_hand(processorid_t id)
469 {
470 	return (CPUID_TO_BOARD(id));
471 }
472 
473 /*
474  * Platform specific lgroup initialization
475  */
476 void
477 plat_lgrp_init(void)
478 {
479 	int i;
480 	extern uint32_t lgrp_expand_proc_thresh;
481 	extern uint32_t lgrp_expand_proc_diff;
482 
483 	/*
484 	 * Initialize lookup tables to invalid values so we catch
485 	 * any illegal use of them.
486 	 */
487 	for (i = 0; i < SG_MAX_SLICE; i++) {
488 		slice_to_memnode[i] = -1;
489 	}
490 
491 	/*
492 	 * Set tuneables for Serengeti architecture
493 	 *
494 	 * lgrp_expand_proc_thresh is the minimum load on the lgroups
495 	 * this process is currently running on before considering
496 	 * expanding threads to another lgroup.
497 	 *
498 	 * lgrp_expand_proc_diff determines how much less the remote lgroup
499 	 * must be loaded before expanding to it.
500 	 *
501 	 * Bandwidth is maximized on Serengeti by spreading load across
502 	 * the machine. The impact to inter-thread communication isn't
503 	 * too costly since remote latencies are relatively low.  These
504 	 * values equate to one CPU's load and so attempt to spread the
505 	 * load out across as many lgroups as possible one CPU at a time.
506 	 */
507 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX;
508 	lgrp_expand_proc_diff = LGRP_LOADAVG_THREAD_MAX;
509 }
510 
511 /*
512  * Platform notification of lgroup (re)configuration changes
513  */
514 /*ARGSUSED*/
515 void
516 plat_lgrp_config(lgrp_config_flag_t evt, uintptr_t arg)
517 {
518 	update_membounds_t	*umb;
519 	lgrp_config_mem_rename_t lmr;
520 	lgrp_handle_t		shand, thand;
521 	int			snode, tnode;
522 
523 	switch (evt) {
524 
525 	case LGRP_CONFIG_MEM_ADD:
526 		umb = (update_membounds_t *)arg;
527 		update_mem_bounds(umb->u_board, umb->u_base, umb->u_len);
528 
529 		break;
530 
531 	case LGRP_CONFIG_MEM_DEL:
532 		/* We don't have to do anything */
533 		break;
534 
535 	case LGRP_CONFIG_MEM_RENAME:
536 		/*
537 		 * During a DR copy-rename operation, all of the memory
538 		 * on one board is moved to another board -- but the
539 		 * addresses/pfns and memnodes don't change. This means
540 		 * the memory has changed locations without changing identity.
541 		 *
542 		 * Source is where we are copying from and target is where we
543 		 * are copying to.  After source memnode is copied to target
544 		 * memnode, the physical addresses of the target memnode are
545 		 * renamed to match what the source memnode had.  Then target
546 		 * memnode can be removed and source memnode can take its
547 		 * place.
548 		 *
549 		 * To do this, swap the lgroup handle to memnode mappings for
550 		 * the boards, so target lgroup will have source memnode and
551 		 * source lgroup will have empty target memnode which is where
552 		 * its memory will go (if any is added to it later).
553 		 *
554 		 * Then source memnode needs to be removed from its lgroup
555 		 * and added to the target lgroup where the memory was living
556 		 * but under a different name/memnode.  The memory was in the
557 		 * target memnode and now lives in the source memnode with
558 		 * different physical addresses even though it is the same
559 		 * memory.
560 		 */
561 		shand = arg & 0xffff;
562 		thand = (arg & 0xffff0000) >> 16;
563 		snode = plat_lgrphand_to_mem_node(shand);
564 		tnode = plat_lgrphand_to_mem_node(thand);
565 
566 		plat_assign_lgrphand_to_mem_node(thand, snode);
567 		plat_assign_lgrphand_to_mem_node(shand, tnode);
568 
569 		/*
570 		 * Remove source memnode of copy rename from its lgroup
571 		 * and add it to its new target lgroup
572 		 */
573 		lmr.lmem_rename_from = shand;
574 		lmr.lmem_rename_to = thand;
575 
576 		lgrp_config(LGRP_CONFIG_MEM_RENAME, (uintptr_t)snode,
577 		    (uintptr_t)&lmr);
578 
579 		break;
580 
581 	default:
582 		break;
583 	}
584 }
585 
586 /*
587  * Return latency between "from" and "to" lgroups
588  *
589  * This latency number can only be used for relative comparison
590  * between lgroups on the running system, cannot be used across platforms,
591  * and may not reflect the actual latency.  It is platform and implementation
592  * specific, so platform gets to decide its value.  It would be nice if the
593  * number was at least proportional to make comparisons more meaningful though.
594  * NOTE: The numbers below are supposed to be load latencies for uncached
595  * memory divided by 10.
596  */
597 int
598 plat_lgrp_latency(lgrp_handle_t from, lgrp_handle_t to)
599 {
600 	/*
601 	 * Return min remote latency when there are more than two lgroups
602 	 * (root and child) and getting latency between two different lgroups
603 	 * or root is involved
604 	 */
605 	if (lgrp_optimizations() && (from != to ||
606 	    from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE))
607 		return (28);
608 	else
609 		return (23);
610 }
611 
612 /* ARGSUSED */
613 void
614 plat_freelist_process(int mnode)
615 {
616 }
617 
618 /*
619  * Find dip for chosen IOSRAM
620  */
621 dev_info_t *
622 find_chosen_dip(void)
623 {
624 	dev_info_t	*dip;
625 	char		master_sbbc[MAXNAMELEN];
626 	pnode_t		nodeid;
627 	uint_t		tunnel;
628 
629 	/*
630 	 * find the /chosen SBBC node, prom interface will handle errors
631 	 */
632 	nodeid = prom_chosennode();
633 
634 	/*
635 	 * get the 'iosram' property from the /chosen node
636 	 */
637 	if (prom_getprop(nodeid, IOSRAM_CHOSEN_PROP, (caddr_t)&tunnel) <= 0) {
638 		SBBC_ERR(CE_PANIC, "No iosram property found! \n");
639 	}
640 
641 	if (prom_phandle_to_path((phandle_t)tunnel, master_sbbc,
642 	    sizeof (master_sbbc)) < 0) {
643 		SBBC_ERR1(CE_PANIC, "prom_phandle_to_path(%d) failed\n",
644 		    tunnel);
645 	}
646 
647 	chosen_nodeid = nodeid;
648 
649 	/*
650 	 * load and attach the sgsbbc driver.
651 	 * This will also attach all the sgsbbc driver instances
652 	 */
653 	if (i_ddi_attach_hw_nodes("sgsbbc") != DDI_SUCCESS) {
654 		cmn_err(CE_WARN, "sgsbbc failed to load\n");
655 	}
656 
657 	/* translate a path name to a dev_info_t */
658 	dip = e_ddi_hold_devi_by_path(master_sbbc, 0);
659 	if ((dip == NULL) || (ddi_get_nodeid(dip) != tunnel)) {
660 		cmn_err(CE_PANIC, "i_ddi_path_to_devi(%x) failed for SBBC\n",
661 		    tunnel);
662 	}
663 
664 	/* make sure devi_ref is ZERO */
665 	ndi_rele_devi(dip);
666 
667 	DCMNERR(CE_CONT, "Chosen IOSRAM is at %s \n", master_sbbc);
668 
669 	return (dip);
670 }
671 
672 void
673 load_platform_drivers(void)
674 {
675 	int ret;
676 
677 	/*
678 	 * Load and attach the mc-us3 memory driver.
679 	 */
680 	if (i_ddi_attach_hw_nodes("mc-us3") != DDI_SUCCESS)
681 		cmn_err(CE_WARN, "mc-us3 failed to load");
682 	else
683 		(void) ddi_hold_driver(ddi_name_to_major("mc-us3"));
684 
685 	/*
686 	 * Initialize the chosen IOSRAM before its clients
687 	 * are loaded.
688 	 */
689 	(void) find_chosen_dip();
690 
691 	/*
692 	 * Ideally, we'd do this in set_platform_defaults(), but
693 	 * at that point it's too early to look up symbols.
694 	 */
695 	iosram_write_ptr = (int (*)(int, uint32_t, caddr_t, uint32_t))
696 	    modgetsymvalue("iosram_write", 0);
697 
698 	if (iosram_write_ptr == NULL) {
699 		DCMNERR(CE_WARN, "load_platform_defaults: iosram_write()"
700 		    " not found; signatures will not be updated\n");
701 	} else {
702 		/*
703 		 * The iosram read ptr is only needed if we can actually
704 		 * write CPU signatures, so only bother setting it if we
705 		 * set a valid write pointer, above.
706 		 */
707 		iosram_read_ptr = (int (*)(int, uint32_t, caddr_t, uint32_t))
708 		    modgetsymvalue("iosram_read", 0);
709 
710 		if (iosram_read_ptr == NULL)
711 			DCMNERR(CE_WARN, "load_platform_defaults: iosram_read()"
712 			    " not found\n");
713 	}
714 
715 	/*
716 	 * Set todsg_use_sc to TRUE so that we will be getting date
717 	 * from the SC.
718 	 */
719 	todsg_use_sc = TRUE;
720 
721 	/*
722 	 * Now is a good time to activate hardware watchdog (if one exists).
723 	 */
724 	mutex_enter(&tod_lock);
725 	if (watchdog_enable)
726 		ret = tod_ops.tod_set_watchdog_timer(watchdog_timeout_seconds);
727 	mutex_exit(&tod_lock);
728 	if (ret != 0)
729 		printf("Hardware watchdog enabled\n");
730 
731 	/*
732 	 * Load and attach the schizo pci bus nexus driver.
733 	 */
734 	if (i_ddi_attach_hw_nodes("pcisch") != DDI_SUCCESS)
735 		cmn_err(CE_WARN, "pcisch failed to load");
736 
737 	plat_ecc_init();
738 }
739 
740 /*
741  * No platform drivers on this platform
742  */
743 char *platform_module_list[] = {
744 	(char *)0
745 };
746 
747 /*ARGSUSED*/
748 void
749 plat_tod_fault(enum tod_fault_type tod_bad)
750 {
751 }
752 int
753 plat_max_boards()
754 {
755 	return (SG_MAX_BDS);
756 }
757 int
758 plat_max_io_units_per_board()
759 {
760 	return (SG_MAX_IO_PER_BD);
761 }
762 int
763 plat_max_cmp_units_per_board()
764 {
765 	return (SG_MAX_CMPS_PER_BD);
766 }
767 int
768 plat_max_cpu_units_per_board()
769 {
770 	return (SG_MAX_CPUS_PER_BD);
771 }
772 
773 int
774 plat_max_mc_units_per_board()
775 {
776 	return (SG_MAX_CMPS_PER_BD); /* each CPU die has a memory controller */
777 }
778 
779 int
780 plat_max_mem_units_per_board()
781 {
782 	return (SG_MAX_MEM_PER_BD);
783 }
784 
785 int
786 plat_max_cpumem_boards(void)
787 {
788 	return (SG_MAX_CPU_BDS);
789 }
790 
791 int
792 set_platform_max_ncpus(void)
793 {
794 	return (sg_max_ncpus);
795 }
796 
797 void
798 plat_dmv_params(uint_t *hwint, uint_t *swint)
799 {
800 	*hwint = MAX_UPA;
801 	*swint = 0;
802 }
803 
804 /*
805  * Our nodename has been set, pass it along to the SC.
806  */
807 void
808 plat_nodename_set(void)
809 {
810 	sbbc_msg_t	req;	/* request */
811 	sbbc_msg_t	resp;	/* response */
812 	int		rv;	/* return value from call to mbox */
813 	struct nodename_info {
814 		int32_t	namelen;
815 		char	nodename[_SYS_NMLN];
816 	} nni;
817 	int (*sg_mbox)(sbbc_msg_t *, sbbc_msg_t *, time_t) = NULL;
818 
819 	/*
820 	 * find the symbol for the mailbox routine
821 	 */
822 	sg_mbox = (int (*)(sbbc_msg_t *, sbbc_msg_t *, time_t))
823 		modgetsymvalue("sbbc_mbox_request_response", 0);
824 
825 	if (sg_mbox == NULL) {
826 		cmn_err(CE_NOTE, "!plat_nodename_set: sg_mbox not found\n");
827 		return;
828 	}
829 
830 	/*
831 	 * construct the message telling the SC our nodename
832 	 */
833 	(void) strcpy(nni.nodename, utsname.nodename);
834 	nni.namelen = (int32_t)strlen(nni.nodename);
835 
836 	req.msg_type.type = INFO_MBOX;
837 	req.msg_type.sub_type = INFO_MBOX_NODENAME;
838 	req.msg_status = 0;
839 	req.msg_len = (int)(nni.namelen + sizeof (nni.namelen));
840 	req.msg_bytes = 0;
841 	req.msg_buf = (caddr_t)&nni;
842 	req.msg_data[0] = 0;
843 	req.msg_data[1] = 0;
844 
845 	/*
846 	 * initialize the response back from the SC
847 	 */
848 	resp.msg_type.type = INFO_MBOX;
849 	resp.msg_type.sub_type = INFO_MBOX_NODENAME;
850 	resp.msg_status = 0;
851 	resp.msg_len = 0;
852 	resp.msg_bytes = 0;
853 	resp.msg_buf = (caddr_t)0;
854 	resp.msg_data[0] = 0;
855 	resp.msg_data[1] = 0;
856 
857 	/*
858 	 * ship it and check for success
859 	 */
860 	rv = (sg_mbox)(&req, &resp, sbbc_mbox_default_timeout);
861 
862 	if (rv != 0) {
863 		cmn_err(CE_NOTE, "!plat_nodename_set: sg_mbox retval %d\n", rv);
864 	} else if (resp.msg_status != 0) {
865 		cmn_err(CE_NOTE, "!plat_nodename_set: msg_status %d\n",
866 			resp.msg_status);
867 	} else {
868 		DCMNERR(CE_NOTE, "!plat_nodename_set was successful\n");
869 
870 		/*
871 		 * It is necessary to exchange the capability bitmap
872 		 * with SC before sending any ecc error information and
873 		 * indictment. We are calling the plat_ecc_capability_send()
874 		 * here just after sending the nodename successfully.
875 		 */
876 		rv = plat_ecc_capability_send();
877 		if (rv == 0) {
878 			DCMNERR(CE_NOTE, "!plat_ecc_capability_send was"
879 			    " successful\n");
880 		}
881 	}
882 }
883 
884 /*
885  * flag to allow users switch between using OBP's
886  * prom_get_unum() and mc-us3 driver's p2get_mem_unum()
887  * (for main memory errors only).
888  */
889 int sg_use_prom_get_unum = 0;
890 
891 /*
892  * Debugging flag: set to 1 to call into obp for get_unum, or set it to 0
893  * to call into the unum cache system.  This is the E$ equivalent of
894  * sg_use_prom_get_unum.
895  */
896 int sg_use_prom_ecache_unum = 0;
897 
898 /* used for logging ECC errors to the SC */
899 #define	SG_MEMORY_ECC	1
900 #define	SG_ECACHE_ECC	2
901 #define	SG_UNKNOWN_ECC	(-1)
902 
903 /*
904  * plat_get_mem_unum() generates a string identifying either the
905  * memory or E$ DIMM(s) during error logging. Depending on whether
906  * the error is E$ or memory related, the appropriate support
907  * routine is called to assist in the string generation.
908  *
909  * - For main memory errors we can use the mc-us3 drivers p2getunum()
910  *   (or prom_get_unum() for debugging purposes).
911  *
912  * - For E$ errors we call sg_get_ecacheunum() to generate the unum (or
913  *   prom_serengeti_get_ecacheunum() for debugging purposes).
914  */
915 
916 static int
917 sg_prom_get_unum(int synd_code, uint64_t paddr, char *buf, int buflen,
918     int *lenp)
919 {
920 	if ((prom_get_unum(synd_code, (unsigned long long)paddr,
921 	    buf, buflen, lenp)) != 0)
922 		return (EIO);
923 	else if (*lenp <= 1)
924 		return (EINVAL);
925 	else
926 		return (0);
927 }
928 
929 /*ARGSUSED*/
930 int
931 plat_get_mem_unum(int synd_code, uint64_t flt_addr, int flt_bus_id,
932     int flt_in_memory, ushort_t flt_status, char *buf, int buflen, int *lenp)
933 {
934 	/*
935 	 * unum_func will either point to the memory drivers p2get_mem_unum()
936 	 * or to prom_get_unum() for memory errors.
937 	 */
938 	int (*unum_func)(int synd_code, uint64_t paddr, char *buf,
939 	    int buflen, int *lenp) = p2get_mem_unum;
940 
941 	/*
942 	 * check if it's a Memory or an Ecache error.
943 	 */
944 	if (flt_in_memory) {
945 		/*
946 		 * It's a main memory error.
947 		 *
948 		 * For debugging we allow the user to switch between
949 		 * using OBP's get_unum and the memory driver's get_unum
950 		 * so we create a pointer to the functions and switch
951 		 * depending on the sg_use_prom_get_unum flag.
952 		 */
953 		if (sg_use_prom_get_unum) {
954 			DCMNERR(CE_NOTE, "Using prom_get_unum from OBP");
955 			return (sg_prom_get_unum(synd_code,
956 			    P2ALIGN(flt_addr, 8), buf, buflen, lenp));
957 		} else if (unum_func != NULL) {
958 			return (unum_func(synd_code, P2ALIGN(flt_addr, 8),
959 			    buf, buflen, lenp));
960 		} else {
961 			return (ENOTSUP);
962 		}
963 	} else if (flt_status & ECC_ECACHE) {
964 		/*
965 		 * It's an E$ error.
966 		 */
967 		if (sg_use_prom_ecache_unum) {
968 			/*
969 			 * We call to OBP to handle this.
970 			 */
971 			DCMNERR(CE_NOTE,
972 			    "Using prom_serengeti_get_ecacheunum from OBP");
973 			if (prom_serengeti_get_ecacheunum(flt_bus_id,
974 			    P2ALIGN(flt_addr, 8), buf, buflen, lenp) != 0) {
975 				return (EIO);
976 			}
977 		} else {
978 			return (sg_get_ecacheunum(flt_bus_id, flt_addr,
979 			    buf, buflen, lenp));
980 		}
981 	} else {
982 		return (ENOTSUP);
983 	}
984 
985 	return (0);
986 }
987 
988 /*
989  * This platform hook gets called from mc_add_mem_unum_label() in the mc-us3
990  * driver giving each platform the opportunity to add platform
991  * specific label information to the unum for ECC error logging purposes.
992  */
993 void
994 plat_add_mem_unum_label(char *unum, int mcid, int bank, int dimm)
995 {
996 	char	new_unum[UNUM_NAMLEN] = "";
997 	int	node = SG_PORTID_TO_NODEID(mcid);
998 	int	board = SG_CPU_BD_PORTID_TO_BD_NUM(mcid);
999 	int	position = SG_PORTID_TO_CPU_POSN(mcid);
1000 
1001 	/*
1002 	 * The mc-us3 driver deals with logical banks but for unum
1003 	 * purposes we need to use physical banks so that the correct
1004 	 * dimm can be physically located. Logical banks 0 and 2
1005 	 * make up physical bank 0. Logical banks 1 and 3 make up
1006 	 * physical bank 1. Here we do the necessary conversion.
1007 	 */
1008 	bank = (bank % 2);
1009 
1010 	if (dimm == -1) {
1011 		SG_SET_FRU_NAME_NODE(new_unum, node);
1012 		SG_SET_FRU_NAME_CPU_BOARD(new_unum, board);
1013 		SG_SET_FRU_NAME_MODULE(new_unum, position);
1014 		SG_SET_FRU_NAME_BANK(new_unum, bank);
1015 
1016 	} else {
1017 		SG_SET_FRU_NAME_NODE(new_unum, node);
1018 		SG_SET_FRU_NAME_CPU_BOARD(new_unum, board);
1019 		SG_SET_FRU_NAME_MODULE(new_unum, position);
1020 		SG_SET_FRU_NAME_BANK(new_unum, bank);
1021 		SG_SET_FRU_NAME_DIMM(new_unum, dimm);
1022 
1023 		strcat(new_unum, " ");
1024 		strcat(new_unum, unum);
1025 	}
1026 
1027 	strcpy(unum, new_unum);
1028 }
1029 
1030 int
1031 plat_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
1032 {
1033 	int	node = SG_PORTID_TO_NODEID(cpuid);
1034 	int	board = SG_CPU_BD_PORTID_TO_BD_NUM(cpuid);
1035 
1036 	if (snprintf(buf, buflen, "/N%d/%s%d", node,
1037 	    SG_HPU_TYPE_CPU_BOARD_ID, board) >= buflen) {
1038 		return (ENOSPC);
1039 	} else {
1040 		*lenp = strlen(buf);
1041 		return (0);
1042 	}
1043 }
1044 
1045 /*
1046  * We log all ECC events to the SC so we send a mailbox
1047  * message to the SC passing it the relevant data.
1048  * ECC mailbox messages are sent via a taskq mechanism to
1049  * prevent impaired system performance during ECC floods.
1050  * Indictments have already passed through a taskq, so they
1051  * are not queued here.
1052  */
1053 int
1054 plat_send_ecc_mailbox_msg(plat_ecc_message_type_t msg_type, void *datap)
1055 {
1056 	sbbc_ecc_mbox_t	*msgp;
1057 	size_t		msg_size;
1058 	uint16_t	msg_subtype;
1059 	int		sleep_flag, log_error;
1060 
1061 	if (sg_ecc_taskq_func == NULL) {
1062 		sg_ecc_taskq_func = (void (*)(sbbc_ecc_mbox_t *))
1063 		    modgetsymvalue("sbbc_mbox_queue_ecc_event", 0);
1064 		if (sg_ecc_taskq_func == NULL) {
1065 			cmn_err(CE_NOTE, "!plat_send_ecc_mailbox_msg: "
1066 			    "sbbc_mbox_queue_ecc_event not found");
1067 			return (ENODEV);
1068 		}
1069 	}
1070 	if (sg_ecc_mbox_func == NULL) {
1071 		sg_ecc_mbox_func = (int (*)(sbbc_ecc_mbox_t *))
1072 		    modgetsymvalue("sbbc_mbox_ecc_output", 0);
1073 		if (sg_ecc_mbox_func == NULL) {
1074 			cmn_err(CE_NOTE, "!plat_send_ecc_mailbox_msg: "
1075 			    "sbbc_mbox_ecc_output not found");
1076 			return (ENODEV);
1077 		}
1078 	}
1079 
1080 	/*
1081 	 * Initialize the request and response structures
1082 	 */
1083 	switch (msg_type) {
1084 	case PLAT_ECC_ERROR_MESSAGE:
1085 		msg_subtype = INFO_MBOX_ERROR_ECC;
1086 		msg_size = sizeof (plat_ecc_error_data_t);
1087 		sleep_flag = KM_NOSLEEP;
1088 		log_error = 1;
1089 		break;
1090 	case PLAT_ECC_ERROR2_MESSAGE:
1091 		msg_subtype = INFO_MBOX_ECC;
1092 		msg_size = sizeof (plat_ecc_error2_data_t);
1093 		sleep_flag = KM_NOSLEEP;
1094 		log_error = 1;
1095 		break;
1096 	case PLAT_ECC_INDICTMENT_MESSAGE:
1097 		msg_subtype = INFO_MBOX_ERROR_INDICT;
1098 		msg_size = sizeof (plat_ecc_indictment_data_t);
1099 		sleep_flag = KM_SLEEP;
1100 		log_error = 0;
1101 		break;
1102 	case PLAT_ECC_INDICTMENT2_MESSAGE:
1103 		msg_subtype = INFO_MBOX_ECC;
1104 		msg_size = sizeof (plat_ecc_indictment2_data_t);
1105 		sleep_flag = KM_SLEEP;
1106 		log_error = 0;
1107 		break;
1108 	case PLAT_ECC_CAPABILITY_MESSAGE:
1109 		msg_subtype = INFO_MBOX_ECC_CAP;
1110 		msg_size = sizeof (plat_capability_data_t) +
1111 		    strlen(utsname.release) + strlen(utsname.version) + 2;
1112 		sleep_flag = KM_SLEEP;
1113 		log_error = 0;
1114 		break;
1115 	case PLAT_ECC_DIMM_SID_MESSAGE:
1116 		msg_subtype = INFO_MBOX_ECC;
1117 		msg_size = sizeof (plat_dimm_sid_request_data_t);
1118 		sleep_flag = KM_SLEEP;
1119 		log_error = 0;
1120 		break;
1121 	default:
1122 		return (EINVAL);
1123 	}
1124 
1125 	msgp = (sbbc_ecc_mbox_t	*)kmem_zalloc(sizeof (sbbc_ecc_mbox_t),
1126 		sleep_flag);
1127 	if (msgp == NULL) {
1128 		cmn_err(CE_NOTE, "!plat_send_ecc_mailbox_msg: "
1129 				"unable to allocate sbbc_ecc_mbox");
1130 		return (ENOMEM);
1131 	}
1132 
1133 	msgp->ecc_log_error = log_error;
1134 
1135 	msgp->ecc_req.msg_type.type = INFO_MBOX;
1136 	msgp->ecc_req.msg_type.sub_type = msg_subtype;
1137 	msgp->ecc_req.msg_status = 0;
1138 	msgp->ecc_req.msg_len = (int)msg_size;
1139 	msgp->ecc_req.msg_bytes = 0;
1140 	msgp->ecc_req.msg_buf = (caddr_t)kmem_zalloc(msg_size, sleep_flag);
1141 	msgp->ecc_req.msg_data[0] = 0;
1142 	msgp->ecc_req.msg_data[1] = 0;
1143 
1144 	if (msgp->ecc_req.msg_buf == NULL) {
1145 		cmn_err(CE_NOTE, "!plat_send_ecc_mailbox_msg: "
1146 				"unable to allocate request msg_buf");
1147 		kmem_free((void *)msgp, sizeof (sbbc_ecc_mbox_t));
1148 		return (ENOMEM);
1149 	}
1150 	bcopy(datap, (void *)msgp->ecc_req.msg_buf, msg_size);
1151 
1152 	/*
1153 	 * initialize the response back from the SC
1154 	 */
1155 	msgp->ecc_resp.msg_type.type = INFO_MBOX;
1156 	msgp->ecc_resp.msg_type.sub_type = msg_subtype;
1157 	msgp->ecc_resp.msg_status = 0;
1158 	msgp->ecc_resp.msg_len = 0;
1159 	msgp->ecc_resp.msg_bytes = 0;
1160 	msgp->ecc_resp.msg_buf = NULL;
1161 	msgp->ecc_resp.msg_data[0] = 0;
1162 	msgp->ecc_resp.msg_data[1] = 0;
1163 
1164 	switch (msg_type) {
1165 	case PLAT_ECC_ERROR_MESSAGE:
1166 	case PLAT_ECC_ERROR2_MESSAGE:
1167 		/*
1168 		 * For Error Messages, we go through a taskq.
1169 		 * Queue up the message for processing
1170 		 */
1171 		(*sg_ecc_taskq_func)(msgp);
1172 		return (0);
1173 
1174 	case PLAT_ECC_CAPABILITY_MESSAGE:
1175 		/*
1176 		 * For indictment and capability messages, we've already gone
1177 		 * through the taskq, so we can call the mailbox routine
1178 		 * directly.  Find the symbol for the routine that sends
1179 		 * the mailbox msg
1180 		 */
1181 		msgp->ecc_resp.msg_len = (int)msg_size;
1182 		msgp->ecc_resp.msg_buf = (caddr_t)kmem_zalloc(msg_size,
1183 		    sleep_flag);
1184 		/* FALLTHRU */
1185 
1186 	case PLAT_ECC_INDICTMENT_MESSAGE:
1187 	case PLAT_ECC_INDICTMENT2_MESSAGE:
1188 		return ((*sg_ecc_mbox_func)(msgp));
1189 
1190 	case PLAT_ECC_DIMM_SID_MESSAGE:
1191 		msgp->ecc_resp.msg_len = sizeof (plat_dimm_sid_board_data_t);
1192 		msgp->ecc_resp.msg_buf = (caddr_t)kmem_zalloc(
1193 		    sizeof (plat_dimm_sid_board_data_t), sleep_flag);
1194 		return ((*sg_ecc_mbox_func)(msgp));
1195 
1196 	default:
1197 		ASSERT(0);
1198 		return (EINVAL);
1199 	}
1200 }
1201 
1202 /*
1203  * m is redundant on serengeti as the multiplier is always 4
1204  */
1205 /*ARGSUSED*/
1206 int
1207 plat_make_fru_cpuid(int sb, int m, int proc)
1208 {
1209 	return (MAKE_CPUID(sb, proc));
1210 }
1211 
1212 /*
1213  * board number for a given proc
1214  */
1215 int
1216 plat_make_fru_boardnum(int proc)
1217 {
1218 	return (SG_CPU_BD_PORTID_TO_BD_NUM(proc));
1219 }
1220 
1221 static
1222 void
1223 cpu_sgn_update(ushort_t sig, uchar_t state, uchar_t sub_state, int cpuid)
1224 {
1225 	uint32_t signature = CPU_SIG_BLD(sig, state, sub_state);
1226 	sig_state_t current_sgn;
1227 	int i;
1228 
1229 	if (iosram_write_ptr == NULL) {
1230 		/*
1231 		 * If the IOSRAM write pointer isn't set, we won't be able
1232 		 * to write signatures to ANYTHING, so we may as well just
1233 		 * write out an error message (if desired) and exit this
1234 		 * routine now...
1235 		 */
1236 		DCMNERR(CE_WARN,
1237 		    "cpu_sgn_update: iosram_write() not found;"
1238 		    " cannot write signature 0x%x for CPU(s) or domain\n",
1239 		    signature);
1240 		return;
1241 	}
1242 
1243 
1244 	/*
1245 	 * Differentiate a panic reboot from a non-panic reboot in the
1246 	 * setting of the substate of the signature.
1247 	 *
1248 	 * If the new substate is REBOOT and we're rebooting due to a panic,
1249 	 * then set the new substate to a special value indicating a panic
1250 	 * reboot, SIGSUBST_PANIC_REBOOT.
1251 	 *
1252 	 * A panic reboot is detected by a current (previous) domain signature
1253 	 * state of SIGST_EXIT, and a new signature substate of SIGSUBST_REBOOT.
1254 	 * The domain signature state SIGST_EXIT is used as the panic flow
1255 	 * progresses.
1256 	 *
1257 	 * At the end of the panic flow, the reboot occurs but we should now
1258 	 * one that was involuntary, something that may be quite useful to know
1259 	 * at OBP level.
1260 	 */
1261 	if (sub_state == SIGSUBST_REBOOT) {
1262 		if (iosram_read_ptr == NULL) {
1263 			DCMNERR(CE_WARN,
1264 			    "cpu_sgn_update: iosram_read() not found;"
1265 			    " could not check current domain signature\n");
1266 		} else {
1267 			(void) (*iosram_read_ptr)(SBBC_SIGBLCK_KEY,
1268 				SG_SGNBLK_DOMAINSIG_OFFSET,
1269 				(char *)&current_sgn, sizeof (current_sgn));
1270 			if (current_sgn.state_t.state == SIGST_EXIT)
1271 				signature = CPU_SIG_BLD(sig, state,
1272 					SIGSUBST_PANIC_REBOOT);
1273 		}
1274 	}
1275 
1276 	/*
1277 	 * cpuid == -1 indicates that the operation applies to all cpus.
1278 	 */
1279 	if (cpuid >= 0) {
1280 		(void) (*iosram_write_ptr)(SBBC_SIGBLCK_KEY,
1281 			SG_SGNBLK_CPUSIG_OFFSET(cpuid), (char *)&signature,
1282 			sizeof (signature));
1283 	} else {
1284 		for (i = 0; i < NCPU; i++) {
1285 			if (cpu[i] == NULL || !(cpu[i]->cpu_flags &
1286 				(CPU_EXISTS|CPU_QUIESCED))) {
1287 				continue;
1288 			}
1289 			(void) (*iosram_write_ptr)(SBBC_SIGBLCK_KEY,
1290 				SG_SGNBLK_CPUSIG_OFFSET(i), (char *)&signature,
1291 				sizeof (signature));
1292 		}
1293 	}
1294 
1295 	if (state == SIGST_OFFLINE || state == SIGST_DETACHED) {
1296 		return;
1297 	}
1298 
1299 	(void) (*iosram_write_ptr)(SBBC_SIGBLCK_KEY,
1300 		SG_SGNBLK_DOMAINSIG_OFFSET, (char *)&signature,
1301 		sizeof (signature));
1302 }
1303 
1304 void
1305 startup_platform(void)
1306 {
1307 }
1308 
1309 /*
1310  * A routine to convert a number (represented as a string) to
1311  * the integer value it represents.
1312  */
1313 
1314 static int
1315 isdigit(int ch)
1316 {
1317 	return (ch >= '0' && ch <= '9');
1318 }
1319 
1320 #define	isspace(c)	((c) == ' ' || (c) == '\t' || (c) == '\n')
1321 
1322 static int
1323 strtoi(char *p, char **pos)
1324 {
1325 	int n;
1326 	int c, neg = 0;
1327 
1328 	if (!isdigit(c = *p)) {
1329 		while (isspace(c))
1330 			c = *++p;
1331 		switch (c) {
1332 			case '-':
1333 				neg++;
1334 				/* FALLTHROUGH */
1335 			case '+':
1336 			c = *++p;
1337 		}
1338 		if (!isdigit(c)) {
1339 			if (pos != NULL)
1340 				*pos = p;
1341 			return (0);
1342 		}
1343 	}
1344 	for (n = '0' - c; isdigit(c = *++p); ) {
1345 		n *= 10; /* two steps to avoid unnecessary overflow */
1346 		n += '0' - c; /* accum neg to avoid surprises at MAX */
1347 	}
1348 	if (pos != NULL)
1349 		*pos = p;
1350 	return (neg ? n : -n);
1351 }
1352 
1353 /*
1354  * Get the three parts of the Serengeti PROM version.
1355  * Used for feature readiness tests.
1356  *
1357  * Return 0 if version extracted successfully, -1 otherwise.
1358  */
1359 
1360 int
1361 sg_get_prom_version(int *sysp, int *intfp, int *bldp)
1362 {
1363 	int plen;
1364 	char vers[512];
1365 	static pnode_t node;
1366 	static char version[] = "version";
1367 	char *verp, *ep;
1368 
1369 	node = prom_finddevice("/openprom");
1370 	if (node == OBP_BADNODE)
1371 		return (-1);
1372 
1373 	plen = prom_getproplen(node, version);
1374 	if (plen <= 0 || plen >= sizeof (vers))
1375 		return (-1);
1376 	(void) prom_getprop(node, version, vers);
1377 	vers[plen] = '\0';
1378 
1379 	/* Make sure it's an OBP flashprom */
1380 	if (vers[0] != 'O' && vers[1] != 'B' && vers[2] != 'P') {
1381 		cmn_err(CE_WARN, "sg_get_prom_version: "
1382 		    "unknown <version> string in </openprom>\n");
1383 		return (-1);
1384 	}
1385 	verp = &vers[4];
1386 
1387 	*sysp = strtoi(verp, &ep);
1388 	if (ep == verp || *ep != '.')
1389 		return (-1);
1390 	verp = ep + 1;
1391 
1392 	*intfp = strtoi(verp, &ep);
1393 	if (ep == verp || *ep != '.')
1394 		return (-1);
1395 	verp = ep + 1;
1396 
1397 	*bldp = strtoi(verp, &ep);
1398 	if (ep == verp || (*ep != '\0' && !isspace(*ep)))
1399 		return (-1);
1400 	return (0);
1401 }
1402 
1403 /*
1404  * Return 0 if system board Dynamic Reconfiguration
1405  * is supported by the firmware, -1 otherwise.
1406  */
1407 int
1408 sg_prom_sb_dr_check(void)
1409 {
1410 	static int prom_res = 1;
1411 
1412 	if (prom_res == 1) {
1413 		int sys, intf, bld;
1414 		int rv;
1415 
1416 		rv = sg_get_prom_version(&sys, &intf, &bld);
1417 		if (rv == 0 && sys == 5 &&
1418 		    (intf >= 12 || (intf == 11 && bld >= 200))) {
1419 			prom_res = 0;
1420 		} else {
1421 			prom_res = -1;
1422 		}
1423 	}
1424 	return (prom_res);
1425 }
1426 
1427 /*
1428  * Return 0 if cPCI Dynamic Reconfiguration
1429  * is supported by the firmware, -1 otherwise.
1430  */
1431 int
1432 sg_prom_cpci_dr_check(void)
1433 {
1434 	/*
1435 	 * The version check is currently the same as for
1436 	 * system boards. Since the two DR sub-systems are
1437 	 * independent, this could change.
1438 	 */
1439 	return (sg_prom_sb_dr_check());
1440 }
1441 
1442 /*
1443  * KDI functions - used by the in-situ kernel debugger (kmdb) to perform
1444  * platform-specific operations.  These functions execute when the world is
1445  * stopped, and as such cannot make any blocking calls, hold locks, etc.
1446  * promif functions are a special case, and may be used.
1447  */
1448 
1449 /*
1450  * Our implementation of this KDI op updates the CPU signature in the system
1451  * controller.  Note that we set the signature to OBP_SIG, rather than DBG_SIG.
1452  * The Forth words we execute will, among other things, transform our OBP_SIG
1453  * into DBG_SIG.  They won't function properly if we try to use DBG_SIG.
1454  */
1455 static void
1456 sg_system_claim(void)
1457 {
1458 	prom_interpret("sigb-sig! my-sigb-sig!", OBP_SIG, OBP_SIG, 0, 0, 0);
1459 }
1460 
1461 static void
1462 sg_system_release(void)
1463 {
1464 	prom_interpret("sigb-sig! my-sigb-sig!", OS_SIG, OS_SIG, 0, 0, 0);
1465 }
1466 
1467 static void
1468 sg_console_claim(void)
1469 {
1470 	prom_serengeti_set_console_input(SGCN_OBP_STR);
1471 }
1472 
1473 static void
1474 sg_console_release(void)
1475 {
1476 	prom_serengeti_set_console_input(SGCN_CLNT_STR);
1477 }
1478 
1479 void
1480 plat_kdi_init(kdi_t *kdi)
1481 {
1482 	kdi->pkdi_system_claim = sg_system_claim;
1483 	kdi->pkdi_system_release = sg_system_release;
1484 	kdi->pkdi_console_claim = sg_console_claim;
1485 	kdi->pkdi_console_release = sg_console_release;
1486 }
1487