xref: /titanic_44/usr/src/uts/sun4u/opl/os/opl.c (revision b0fe7b8fa79924061f3bdf7f240ea116c2c0b704)
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 2007 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/cpuvar.h>
29 #include <sys/systm.h>
30 #include <sys/sysmacros.h>
31 #include <sys/promif.h>
32 #include <sys/platform_module.h>
33 #include <sys/cmn_err.h>
34 #include <sys/errno.h>
35 #include <sys/machsystm.h>
36 #include <sys/bootconf.h>
37 #include <sys/nvpair.h>
38 #include <sys/kobj.h>
39 #include <sys/mem_cage.h>
40 #include <sys/opl.h>
41 #include <sys/scfd/scfostoescf.h>
42 #include <sys/cpu_sgnblk_defs.h>
43 #include <sys/utsname.h>
44 #include <sys/ddi.h>
45 #include <sys/sunndi.h>
46 #include <sys/lgrp.h>
47 #include <sys/memnode.h>
48 #include <sys/sysmacros.h>
49 #include <sys/time.h>
50 #include <sys/cpu.h>
51 #include <vm/vm_dep.h>
52 
53 int (*opl_get_mem_unum)(int, uint64_t, char *, int, int *);
54 int (*opl_get_mem_sid)(char *unum, char *buf, int buflen, int *lenp);
55 int (*opl_get_mem_offset)(uint64_t paddr, uint64_t *offp);
56 int (*opl_get_mem_addr)(char *unum, char *sid,
57     uint64_t offset, uint64_t *paddr);
58 
59 /* Memory for fcode claims.  16k times # maximum possible IO units */
60 #define	EFCODE_SIZE	(OPL_MAX_BOARDS * OPL_MAX_IO_UNITS_PER_BOARD * 0x4000)
61 int efcode_size = EFCODE_SIZE;
62 
63 #define	OPL_MC_MEMBOARD_SHIFT 38	/* Boards on 256BG boundary */
64 
65 /* Set the maximum number of boards for DR */
66 int opl_boards = OPL_MAX_BOARDS;
67 
68 void sgn_update_all_cpus(ushort_t, uchar_t, uchar_t);
69 
70 extern int tsb_lgrp_affinity;
71 
72 int opl_tsb_spares = (OPL_MAX_BOARDS) * (OPL_MAX_PCICH_UNITS_PER_BOARD) *
73 	(OPL_MAX_TSBS_PER_PCICH);
74 
75 pgcnt_t opl_startup_cage_size = 0;
76 
77 static opl_model_info_t opl_models[] = {
78 	{ "FF1", OPL_MAX_BOARDS_FF1, FF1, STD_DISPATCH_TABLE },
79 	{ "FF2", OPL_MAX_BOARDS_FF2, FF2, STD_DISPATCH_TABLE },
80 	{ "DC1", OPL_MAX_BOARDS_DC1, DC1, STD_DISPATCH_TABLE },
81 	{ "DC2", OPL_MAX_BOARDS_DC2, DC2, EXT_DISPATCH_TABLE },
82 	{ "DC3", OPL_MAX_BOARDS_DC3, DC3, EXT_DISPATCH_TABLE },
83 };
84 static	int	opl_num_models = sizeof (opl_models)/sizeof (opl_model_info_t);
85 
86 /*
87  * opl_cur_model
88  */
89 static	opl_model_info_t *opl_cur_model = NULL;
90 
91 static struct memlist *opl_memlist_per_board(struct memlist *ml);
92 
93 /*
94  * Note FF/DC out-of-order instruction engine takes only a
95  * single cycle to execute each spin loop
96  * for comparison, Panther takes 6 cycles for same loop
97  * 1500 approx nsec for OPL sleep instruction
98  * if spin count = OPL_BOFF_SLEEP*OPL_BOFF_SPIN then
99  * spin time should be equal to OPL_BOFF_TM nsecs
100  * Listed values tuned for 2.15GHz to 2.4GHz systems
101  * Value may change for future systems
102  */
103 #define	OPL_BOFF_SPIN 720
104 #define	OPL_BOFF_BASE 1
105 #define	OPL_BOFF_SLEEP 5
106 #define	OPL_BOFF_CAP1 20
107 #define	OPL_BOFF_CAP2 60
108 #define	OPL_BOFF_MAX (40 * OPL_BOFF_SLEEP)
109 #define	OPL_BOFF_TM 1500
110 
111 int
112 set_platform_max_ncpus(void)
113 {
114 	return (OPL_MAX_CPU_PER_BOARD * OPL_MAX_BOARDS);
115 }
116 
117 int
118 set_platform_tsb_spares(void)
119 {
120 	return (MIN(opl_tsb_spares, MAX_UPA));
121 }
122 
123 static void
124 set_model_info()
125 {
126 	extern int ts_dispatch_extended;
127 	char	name[MAXSYSNAME];
128 	int	i;
129 
130 	/*
131 	 * Get model name from the root node.
132 	 *
133 	 * We are using the prom device tree since, at this point,
134 	 * the Solaris device tree is not yet setup.
135 	 */
136 	(void) prom_getprop(prom_rootnode(), "model", (caddr_t)name);
137 
138 	for (i = 0; i < opl_num_models; i++) {
139 		if (strncmp(name, opl_models[i].model_name, MAXSYSNAME) == 0) {
140 			opl_cur_model = &opl_models[i];
141 			break;
142 		}
143 	}
144 
145 	if (i == opl_num_models)
146 		halt("No valid OPL model is found!");
147 
148 	if ((opl_cur_model->model_cmds & EXT_DISPATCH_TABLE) &&
149 				(ts_dispatch_extended == -1)) {
150 		/*
151 		 * Based on a platform model, select a dispatch table.
152 		 * Only DC2 and DC3 systems uses the alternate/extended
153 		 * TS dispatch table.
154 		 * FF1, FF2 and DC1 systems used standard dispatch tables.
155 		 */
156 		ts_dispatch_extended = 1;
157 	}
158 
159 }
160 
161 static void
162 set_max_mmu_ctxdoms()
163 {
164 	extern uint_t	max_mmu_ctxdoms;
165 	int		max_boards;
166 
167 	/*
168 	 * From the model, get the maximum number of boards
169 	 * supported and set the value accordingly. If the model
170 	 * could not be determined or recognized, we assume the max value.
171 	 */
172 	if (opl_cur_model == NULL)
173 		max_boards = OPL_MAX_BOARDS;
174 	else
175 		max_boards = opl_cur_model->model_max_boards;
176 
177 	/*
178 	 * On OPL, cores and MMUs are one-to-one.
179 	 */
180 	max_mmu_ctxdoms = OPL_MAX_CORE_UNITS_PER_BOARD * max_boards;
181 }
182 
183 #pragma weak mmu_init_large_pages
184 
185 void
186 set_platform_defaults(void)
187 {
188 	extern char *tod_module_name;
189 	extern void cpu_sgn_update(ushort_t, uchar_t, uchar_t, int);
190 	extern void mmu_init_large_pages(size_t);
191 
192 	/* Set the CPU signature function pointer */
193 	cpu_sgn_func = cpu_sgn_update;
194 
195 	/* Set appropriate tod module for OPL platform */
196 	ASSERT(tod_module_name == NULL);
197 	tod_module_name = "todopl";
198 
199 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
200 	    (mmu_ism_pagesize != DEFAULT_ISM_PAGESIZE)) {
201 		if (&mmu_init_large_pages)
202 			mmu_init_large_pages(mmu_ism_pagesize);
203 	}
204 
205 	tsb_lgrp_affinity = 1;
206 
207 	set_max_mmu_ctxdoms();
208 }
209 
210 /*
211  * Convert logical a board number to a physical one.
212  */
213 
214 #define	LSBPROP		"board#"
215 #define	PSBPROP		"physical-board#"
216 
217 int
218 opl_get_physical_board(int id)
219 {
220 	dev_info_t	*root_dip, *dip = NULL;
221 	char		*dname = NULL;
222 	int		circ;
223 
224 	pnode_t		pnode;
225 	char		pname[MAXSYSNAME] = {0};
226 
227 	int		lsb_id;	/* Logical System Board ID */
228 	int		psb_id;	/* Physical System Board ID */
229 
230 
231 	/*
232 	 * This function is called on early stage of bootup when the
233 	 * kernel device tree is not initialized yet, and also
234 	 * later on when the device tree is up. We want to try
235 	 * the fast track first.
236 	 */
237 	root_dip = ddi_root_node();
238 	if (root_dip) {
239 		/* Get from devinfo node */
240 		ndi_devi_enter(root_dip, &circ);
241 		for (dip = ddi_get_child(root_dip); dip;
242 		    dip = ddi_get_next_sibling(dip)) {
243 
244 			dname = ddi_node_name(dip);
245 			if (strncmp(dname, "pseudo-mc", 9) != 0)
246 				continue;
247 
248 			if ((lsb_id = (int)ddi_getprop(DDI_DEV_T_ANY, dip,
249 			    DDI_PROP_DONTPASS, LSBPROP, -1)) == -1)
250 				continue;
251 
252 			if (id == lsb_id) {
253 				if ((psb_id = (int)ddi_getprop(DDI_DEV_T_ANY,
254 				    dip, DDI_PROP_DONTPASS, PSBPROP, -1))
255 				    == -1) {
256 					ndi_devi_exit(root_dip, circ);
257 					return (-1);
258 				} else {
259 					ndi_devi_exit(root_dip, circ);
260 					return (psb_id);
261 				}
262 			}
263 		}
264 		ndi_devi_exit(root_dip, circ);
265 	}
266 
267 	/*
268 	 * We do not have the kernel device tree, or we did not
269 	 * find the node for some reason (let's say the kernel
270 	 * device tree was modified), let's try the OBP tree.
271 	 */
272 	pnode = prom_rootnode();
273 	for (pnode = prom_childnode(pnode); pnode;
274 	    pnode = prom_nextnode(pnode)) {
275 
276 		if ((prom_getprop(pnode, "name", (caddr_t)pname) == -1) ||
277 		    (strncmp(pname, "pseudo-mc", 9) != 0))
278 			continue;
279 
280 		if (prom_getprop(pnode, LSBPROP, (caddr_t)&lsb_id) == -1)
281 			continue;
282 
283 		if (id == lsb_id) {
284 			if (prom_getprop(pnode, PSBPROP,
285 			    (caddr_t)&psb_id) == -1) {
286 				return (-1);
287 			} else {
288 				return (psb_id);
289 			}
290 		}
291 	}
292 
293 	return (-1);
294 }
295 
296 /*
297  * For OPL it's possible that memory from two or more successive boards
298  * will be contiguous across the boards, and therefore represented as a
299  * single chunk.
300  * This function splits such chunks down the board boundaries.
301  */
302 static struct memlist *
303 opl_memlist_per_board(struct memlist *ml)
304 {
305 	uint64_t ssize, low, high, boundary;
306 	struct memlist *head, *tail, *new;
307 
308 	ssize = (1ull << OPL_MC_MEMBOARD_SHIFT);
309 
310 	head = tail = NULL;
311 
312 	for (; ml; ml = ml->next) {
313 		low  = (uint64_t)ml->address;
314 		high = low+(uint64_t)(ml->size);
315 		while (low < high) {
316 			boundary = roundup(low+1, ssize);
317 			boundary = MIN(high, boundary);
318 			new = kmem_zalloc(sizeof (struct memlist), KM_SLEEP);
319 			new->address = low;
320 			new->size = boundary - low;
321 			if (head == NULL)
322 				head = new;
323 			if (tail) {
324 				tail->next = new;
325 				new->prev = tail;
326 			}
327 			tail = new;
328 			low = boundary;
329 		}
330 	}
331 	return (head);
332 }
333 
334 void
335 set_platform_cage_params(void)
336 {
337 	extern pgcnt_t total_pages;
338 	extern struct memlist *phys_avail;
339 	struct memlist *ml, *tml;
340 	int ret;
341 
342 	if (kernel_cage_enable) {
343 		pgcnt_t preferred_cage_size;
344 
345 		preferred_cage_size =
346 			MAX(opl_startup_cage_size, total_pages / 256);
347 
348 		ml = opl_memlist_per_board(phys_avail);
349 
350 		kcage_range_lock();
351 		/*
352 		 * Note: we are assuming that post has load the
353 		 * whole show in to the high end of memory. Having
354 		 * taken this leap, we copy the whole of phys_avail
355 		 * the glist and arrange for the cage to grow
356 		 * downward (descending pfns).
357 		 */
358 		ret = kcage_range_init(ml, 1);
359 
360 		/* free the memlist */
361 		do {
362 			tml = ml->next;
363 			kmem_free(ml, sizeof (struct memlist));
364 			ml = tml;
365 		} while (ml != NULL);
366 
367 		if (ret == 0)
368 			kcage_init(preferred_cage_size);
369 		kcage_range_unlock();
370 	}
371 
372 	if (kcage_on)
373 		cmn_err(CE_NOTE, "!DR Kernel Cage is ENABLED");
374 	else
375 		cmn_err(CE_NOTE, "!DR Kernel Cage is DISABLED");
376 }
377 
378 /*ARGSUSED*/
379 int
380 plat_cpu_poweron(struct cpu *cp)
381 {
382 	int (*opl_cpu_poweron)(struct cpu *) = NULL;
383 
384 	opl_cpu_poweron =
385 	    (int (*)(struct cpu *))kobj_getsymvalue("drmach_cpu_poweron", 0);
386 
387 	if (opl_cpu_poweron == NULL)
388 		return (ENOTSUP);
389 	else
390 		return ((opl_cpu_poweron)(cp));
391 
392 }
393 
394 /*ARGSUSED*/
395 int
396 plat_cpu_poweroff(struct cpu *cp)
397 {
398 	int (*opl_cpu_poweroff)(struct cpu *) = NULL;
399 
400 	opl_cpu_poweroff =
401 	    (int (*)(struct cpu *))kobj_getsymvalue("drmach_cpu_poweroff", 0);
402 
403 	if (opl_cpu_poweroff == NULL)
404 		return (ENOTSUP);
405 	else
406 		return ((opl_cpu_poweroff)(cp));
407 
408 }
409 
410 int
411 plat_max_boards(void)
412 {
413 	return (OPL_MAX_BOARDS);
414 }
415 
416 int
417 plat_max_cpu_units_per_board(void)
418 {
419 	return (OPL_MAX_CPU_PER_BOARD);
420 }
421 
422 int
423 plat_max_mem_units_per_board(void)
424 {
425 	return (OPL_MAX_MEM_UNITS_PER_BOARD);
426 }
427 
428 int
429 plat_max_io_units_per_board(void)
430 {
431 	return (OPL_MAX_IO_UNITS_PER_BOARD);
432 }
433 
434 int
435 plat_max_cmp_units_per_board(void)
436 {
437 	return (OPL_MAX_CMP_UNITS_PER_BOARD);
438 }
439 
440 int
441 plat_max_core_units_per_board(void)
442 {
443 	return (OPL_MAX_CORE_UNITS_PER_BOARD);
444 }
445 
446 int
447 plat_pfn_to_mem_node(pfn_t pfn)
448 {
449 	return (pfn >> mem_node_pfn_shift);
450 }
451 
452 /* ARGSUSED */
453 void
454 plat_build_mem_nodes(u_longlong_t *list, size_t nelems)
455 {
456 	size_t	elem;
457 	pfn_t	basepfn;
458 	pgcnt_t	npgs;
459 	uint64_t	boundary, ssize;
460 	uint64_t	low, high;
461 
462 	/*
463 	 * OPL mem slices are always aligned on a 256GB boundary.
464 	 */
465 	mem_node_pfn_shift = OPL_MC_MEMBOARD_SHIFT - MMU_PAGESHIFT;
466 	mem_node_physalign = 0;
467 
468 	/*
469 	 * Boot install lists are arranged <addr, len>, <addr, len>, ...
470 	 */
471 	ssize = (1ull << OPL_MC_MEMBOARD_SHIFT);
472 	for (elem = 0; elem < nelems; elem += 2) {
473 		low  = (uint64_t)list[elem];
474 		high = low+(uint64_t)(list[elem+1]);
475 		while (low < high) {
476 			boundary = roundup(low+1, ssize);
477 			boundary = MIN(high, boundary);
478 			basepfn = btop(low);
479 			npgs = btop(boundary - low);
480 			mem_node_add_slice(basepfn, basepfn + npgs - 1);
481 			low = boundary;
482 		}
483 	}
484 }
485 
486 /*
487  * Find the CPU associated with a slice at boot-time.
488  */
489 void
490 plat_fill_mc(pnode_t nodeid)
491 {
492 	int board;
493 	int memnode;
494 	struct {
495 		uint64_t	addr;
496 		uint64_t	size;
497 	} mem_range;
498 
499 	if (prom_getprop(nodeid, "board#", (caddr_t)&board) < 0) {
500 		panic("Can not find board# property in mc node %x", nodeid);
501 	}
502 	if (prom_getprop(nodeid, "sb-mem-ranges", (caddr_t)&mem_range) < 0) {
503 		panic("Can not find sb-mem-ranges property in mc node %x",
504 			nodeid);
505 	}
506 	memnode = mem_range.addr >> OPL_MC_MEMBOARD_SHIFT;
507 	plat_assign_lgrphand_to_mem_node(board, memnode);
508 }
509 
510 /*
511  * Return the platform handle for the lgroup containing the given CPU
512  *
513  * For OPL, lgroup platform handle == board #.
514  */
515 
516 extern int mpo_disabled;
517 extern lgrp_handle_t lgrp_default_handle;
518 
519 lgrp_handle_t
520 plat_lgrp_cpu_to_hand(processorid_t id)
521 {
522 	lgrp_handle_t plathand;
523 
524 	/*
525 	 * Return the real platform handle for the CPU until
526 	 * such time as we know that MPO should be disabled.
527 	 * At that point, we set the "mpo_disabled" flag to true,
528 	 * and from that point on, return the default handle.
529 	 *
530 	 * By the time we know that MPO should be disabled, the
531 	 * first CPU will have already been added to a leaf
532 	 * lgroup, but that's ok. The common lgroup code will
533 	 * double check that the boot CPU is in the correct place,
534 	 * and in the case where mpo should be disabled, will move
535 	 * it to the root if necessary.
536 	 */
537 	if (mpo_disabled) {
538 		/* If MPO is disabled, return the default (UMA) handle */
539 		plathand = lgrp_default_handle;
540 	} else
541 		plathand = (lgrp_handle_t)LSB_ID(id);
542 	return (plathand);
543 }
544 
545 /*
546  * Platform specific lgroup initialization
547  */
548 void
549 plat_lgrp_init(void)
550 {
551 	extern uint32_t lgrp_expand_proc_thresh;
552 	extern uint32_t lgrp_expand_proc_diff;
553 
554 	/*
555 	 * Set tuneables for the OPL architecture
556 	 *
557 	 * lgrp_expand_proc_thresh is the minimum load on the lgroups
558 	 * this process is currently running on before considering
559 	 * expanding threads to another lgroup.
560 	 *
561 	 * lgrp_expand_proc_diff determines how much less the remote lgroup
562 	 * must be loaded before expanding to it.
563 	 *
564 	 * Since remote latencies can be costly, attempt to keep 3 threads
565 	 * within the same lgroup before expanding to the next lgroup.
566 	 */
567 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX * 3;
568 	lgrp_expand_proc_diff = LGRP_LOADAVG_THREAD_MAX;
569 }
570 
571 /*
572  * Platform notification of lgroup (re)configuration changes
573  */
574 /*ARGSUSED*/
575 void
576 plat_lgrp_config(lgrp_config_flag_t evt, uintptr_t arg)
577 {
578 	update_membounds_t *umb;
579 	lgrp_config_mem_rename_t lmr;
580 	int sbd, tbd;
581 	lgrp_handle_t hand, shand, thand;
582 	int mnode, snode, tnode;
583 	pfn_t start, end;
584 
585 	if (mpo_disabled)
586 		return;
587 
588 	switch (evt) {
589 
590 	case LGRP_CONFIG_MEM_ADD:
591 		/*
592 		 * Establish the lgroup handle to memnode translation.
593 		 */
594 		umb = (update_membounds_t *)arg;
595 
596 		hand = umb->u_board;
597 		mnode = plat_pfn_to_mem_node(umb->u_base >> MMU_PAGESHIFT);
598 		plat_assign_lgrphand_to_mem_node(hand, mnode);
599 
600 		break;
601 
602 	case LGRP_CONFIG_MEM_DEL:
603 		/*
604 		 * Special handling for possible memory holes.
605 		 */
606 		umb = (update_membounds_t *)arg;
607 		hand = umb->u_board;
608 		if ((mnode = plat_lgrphand_to_mem_node(hand)) != -1) {
609 			if (mem_node_config[mnode].exists) {
610 				start = mem_node_config[mnode].physbase;
611 				end = mem_node_config[mnode].physmax;
612 				mem_node_pre_del_slice(start, end);
613 				mem_node_post_del_slice(start, end, 0);
614 			}
615 		}
616 
617 		break;
618 
619 	case LGRP_CONFIG_MEM_RENAME:
620 		/*
621 		 * During a DR copy-rename operation, all of the memory
622 		 * on one board is moved to another board -- but the
623 		 * addresses/pfns and memnodes don't change. This means
624 		 * the memory has changed locations without changing identity.
625 		 *
626 		 * Source is where we are copying from and target is where we
627 		 * are copying to.  After source memnode is copied to target
628 		 * memnode, the physical addresses of the target memnode are
629 		 * renamed to match what the source memnode had.  Then target
630 		 * memnode can be removed and source memnode can take its
631 		 * place.
632 		 *
633 		 * To do this, swap the lgroup handle to memnode mappings for
634 		 * the boards, so target lgroup will have source memnode and
635 		 * source lgroup will have empty target memnode which is where
636 		 * its memory will go (if any is added to it later).
637 		 *
638 		 * Then source memnode needs to be removed from its lgroup
639 		 * and added to the target lgroup where the memory was living
640 		 * but under a different name/memnode.  The memory was in the
641 		 * target memnode and now lives in the source memnode with
642 		 * different physical addresses even though it is the same
643 		 * memory.
644 		 */
645 		sbd = arg & 0xffff;
646 		tbd = (arg & 0xffff0000) >> 16;
647 		shand = sbd;
648 		thand = tbd;
649 		snode = plat_lgrphand_to_mem_node(shand);
650 		tnode = plat_lgrphand_to_mem_node(thand);
651 
652 		/*
653 		 * Special handling for possible memory holes.
654 		 */
655 		if (tnode != -1 && mem_node_config[tnode].exists) {
656 			start = mem_node_config[tnode].physbase;
657 			end = mem_node_config[tnode].physmax;
658 			mem_node_pre_del_slice(start, end);
659 			mem_node_post_del_slice(start, end, 0);
660 		}
661 
662 		plat_assign_lgrphand_to_mem_node(thand, snode);
663 		plat_assign_lgrphand_to_mem_node(shand, tnode);
664 
665 		lmr.lmem_rename_from = shand;
666 		lmr.lmem_rename_to = thand;
667 
668 		/*
669 		 * Remove source memnode of copy rename from its lgroup
670 		 * and add it to its new target lgroup
671 		 */
672 		lgrp_config(LGRP_CONFIG_MEM_RENAME, (uintptr_t)snode,
673 		    (uintptr_t)&lmr);
674 
675 		break;
676 
677 	default:
678 		break;
679 	}
680 }
681 
682 /*
683  * Return latency between "from" and "to" lgroups
684  *
685  * This latency number can only be used for relative comparison
686  * between lgroups on the running system, cannot be used across platforms,
687  * and may not reflect the actual latency.  It is platform and implementation
688  * specific, so platform gets to decide its value.  It would be nice if the
689  * number was at least proportional to make comparisons more meaningful though.
690  * NOTE: The numbers below are supposed to be load latencies for uncached
691  * memory divided by 10.
692  *
693  */
694 int
695 plat_lgrp_latency(lgrp_handle_t from, lgrp_handle_t to)
696 {
697 	/*
698 	 * Return min remote latency when there are more than two lgroups
699 	 * (root and child) and getting latency between two different lgroups
700 	 * or root is involved
701 	 */
702 	if (lgrp_optimizations() && (from != to ||
703 	    from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE))
704 		return (42);
705 	else
706 		return (35);
707 }
708 
709 /*
710  * Return platform handle for root lgroup
711  */
712 lgrp_handle_t
713 plat_lgrp_root_hand(void)
714 {
715 	if (mpo_disabled)
716 		return (lgrp_default_handle);
717 
718 	return (LGRP_DEFAULT_HANDLE);
719 }
720 
721 /*ARGSUSED*/
722 void
723 plat_freelist_process(int mnode)
724 {
725 }
726 
727 void
728 load_platform_drivers(void)
729 {
730 	(void) i_ddi_attach_pseudo_node("dr");
731 }
732 
733 /*
734  * No platform drivers on this platform
735  */
736 char *platform_module_list[] = {
737 	(char *)0
738 };
739 
740 /*ARGSUSED*/
741 void
742 plat_tod_fault(enum tod_fault_type tod_bad)
743 {
744 }
745 
746 /*ARGSUSED*/
747 void
748 cpu_sgn_update(ushort_t sgn, uchar_t state, uchar_t sub_state, int cpuid)
749 {
750 	static void (*scf_panic_callback)(int);
751 	static void (*scf_shutdown_callback)(int);
752 
753 	/*
754 	 * This is for notifing system panic/shutdown to SCF.
755 	 * In case of shutdown and panic, SCF call back
756 	 * function should be called.
757 	 *  <SCF call back functions>
758 	 *   scf_panic_callb()   : panicsys()->panic_quiesce_hw()
759 	 *   scf_shutdown_callb(): halt() or power_down() or reboot_machine()
760 	 * cpuid should be -1 and state should be SIGST_EXIT.
761 	 */
762 	if (state == SIGST_EXIT && cpuid == -1) {
763 
764 		/*
765 		 * find the symbol for the SCF panic callback routine in driver
766 		 */
767 		if (scf_panic_callback == NULL)
768 			scf_panic_callback = (void (*)(int))
769 				modgetsymvalue("scf_panic_callb", 0);
770 		if (scf_shutdown_callback == NULL)
771 			scf_shutdown_callback = (void (*)(int))
772 				modgetsymvalue("scf_shutdown_callb", 0);
773 
774 		switch (sub_state) {
775 		case SIGSUBST_PANIC:
776 			if (scf_panic_callback == NULL) {
777 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
778 				    "scf_panic_callb not found\n");
779 				return;
780 			}
781 			scf_panic_callback(SIGSUBST_PANIC);
782 			break;
783 
784 		case SIGSUBST_HALT:
785 			if (scf_shutdown_callback == NULL) {
786 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
787 				    "scf_shutdown_callb not found\n");
788 				return;
789 			}
790 			scf_shutdown_callback(SIGSUBST_HALT);
791 			break;
792 
793 		case SIGSUBST_ENVIRON:
794 			if (scf_shutdown_callback == NULL) {
795 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
796 				    "scf_shutdown_callb not found\n");
797 				return;
798 			}
799 			scf_shutdown_callback(SIGSUBST_ENVIRON);
800 			break;
801 
802 		case SIGSUBST_REBOOT:
803 			if (scf_shutdown_callback == NULL) {
804 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
805 				    "scf_shutdown_callb not found\n");
806 				return;
807 			}
808 			scf_shutdown_callback(SIGSUBST_REBOOT);
809 			break;
810 		}
811 	}
812 }
813 
814 /*ARGSUSED*/
815 int
816 plat_get_mem_unum(int synd_code, uint64_t flt_addr, int flt_bus_id,
817 	int flt_in_memory, ushort_t flt_status,
818 	char *buf, int buflen, int *lenp)
819 {
820 	/*
821 	 * check if it's a Memory error.
822 	 */
823 	if (flt_in_memory) {
824 		if (opl_get_mem_unum != NULL) {
825 			return (opl_get_mem_unum(synd_code, flt_addr,
826 				buf, buflen, lenp));
827 		} else {
828 			return (ENOTSUP);
829 		}
830 	} else {
831 		return (ENOTSUP);
832 	}
833 }
834 
835 /*ARGSUSED*/
836 int
837 plat_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
838 {
839 	int	ret = 0;
840 	uint_t	sb;
841 	int	plen;
842 
843 	sb = opl_get_physical_board(LSB_ID(cpuid));
844 	if (sb == -1) {
845 		return (ENXIO);
846 	}
847 
848 	/*
849 	 * opl_cur_model is assigned here
850 	 */
851 	if (opl_cur_model == NULL) {
852 		set_model_info();
853 	}
854 
855 	ASSERT((opl_cur_model - opl_models) == (opl_cur_model->model_type));
856 
857 	switch (opl_cur_model->model_type) {
858 	case FF1:
859 		plen = snprintf(buf, buflen, "/%s/CPUM%d", "MBU_A",
860 		    CHIP_ID(cpuid) / 2);
861 		break;
862 
863 	case FF2:
864 		plen = snprintf(buf, buflen, "/%s/CPUM%d", "MBU_B",
865 		    (CHIP_ID(cpuid) / 2) + (sb * 2));
866 		break;
867 
868 	case DC1:
869 	case DC2:
870 	case DC3:
871 		plen = snprintf(buf, buflen, "/%s%02d/CPUM%d", "CMU", sb,
872 		    CHIP_ID(cpuid));
873 		break;
874 
875 	default:
876 		/* This should never happen */
877 		return (ENODEV);
878 	}
879 
880 	if (plen >= buflen) {
881 		ret = ENOSPC;
882 	} else {
883 		if (lenp)
884 			*lenp = strlen(buf);
885 	}
886 	return (ret);
887 }
888 
889 #define	SCF_PUTINFO(f, s, p)	\
890 	f(KEY_ESCF, 0x01, 0, s, p)
891 void
892 plat_nodename_set(void)
893 {
894 	void *datap;
895 	static int (*scf_service_function)(uint32_t, uint8_t,
896 	    uint32_t, uint32_t, void *);
897 	int counter = 5;
898 
899 	/*
900 	 * find the symbol for the SCF put routine in driver
901 	 */
902 	if (scf_service_function == NULL)
903 		scf_service_function =
904 			(int (*)(uint32_t, uint8_t, uint32_t, uint32_t, void *))
905 			modgetsymvalue("scf_service_putinfo", 0);
906 
907 	/*
908 	 * If the symbol was found, call it.  Otherwise, log a note (but not to
909 	 * the console).
910 	 */
911 
912 	if (scf_service_function == NULL) {
913 		cmn_err(CE_NOTE,
914 		    "!plat_nodename_set: scf_service_putinfo not found\n");
915 		return;
916 	}
917 
918 	datap =
919 	    (struct utsname *)kmem_zalloc(sizeof (struct utsname), KM_SLEEP);
920 
921 	if (datap == NULL) {
922 		return;
923 	}
924 
925 	bcopy((struct utsname *)&utsname,
926 	    (struct utsname *)datap, sizeof (struct utsname));
927 
928 	while ((SCF_PUTINFO(scf_service_function,
929 		sizeof (struct utsname), datap) == EBUSY) && (counter-- > 0)) {
930 		delay(10 * drv_usectohz(1000000));
931 	}
932 	if (counter == 0)
933 		cmn_err(CE_NOTE,
934 			"!plat_nodename_set: "
935 			"scf_service_putinfo not responding\n");
936 
937 	kmem_free(datap, sizeof (struct utsname));
938 }
939 
940 caddr_t	efcode_vaddr = NULL;
941 
942 /*
943  * Preallocate enough memory for fcode claims.
944  */
945 
946 caddr_t
947 efcode_alloc(caddr_t alloc_base)
948 {
949 	caddr_t efcode_alloc_base = (caddr_t)roundup((uintptr_t)alloc_base,
950 	    MMU_PAGESIZE);
951 	caddr_t vaddr;
952 
953 	/*
954 	 * allocate the physical memory for the Oberon fcode.
955 	 */
956 	if ((vaddr = (caddr_t)BOP_ALLOC(bootops, efcode_alloc_base,
957 	    efcode_size, MMU_PAGESIZE)) == NULL)
958 		cmn_err(CE_PANIC, "Cannot allocate Efcode Memory");
959 
960 	efcode_vaddr = vaddr;
961 
962 	return (efcode_alloc_base + efcode_size);
963 }
964 
965 caddr_t
966 plat_startup_memlist(caddr_t alloc_base)
967 {
968 	caddr_t tmp_alloc_base;
969 
970 	tmp_alloc_base = efcode_alloc(alloc_base);
971 	tmp_alloc_base =
972 	    (caddr_t)roundup((uintptr_t)tmp_alloc_base, ecache_alignsize);
973 	return (tmp_alloc_base);
974 }
975 
976 void
977 startup_platform(void)
978 {
979 }
980 
981 void
982 plat_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *info)
983 {
984 	int	impl;
985 
986 	impl = cpunodes[cpuid].implementation;
987 	if (IS_OLYMPUS_C(impl)) {
988 		info->mmu_idx = MMU_ID(cpuid);
989 		info->mmu_nctxs = 8192;
990 	} else {
991 		cmn_err(CE_PANIC, "Unknown processor %d", impl);
992 	}
993 }
994 
995 int
996 plat_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
997 {
998 	if (opl_get_mem_sid == NULL) {
999 		return (ENOTSUP);
1000 	}
1001 	return (opl_get_mem_sid(unum, buf, buflen, lenp));
1002 }
1003 
1004 int
1005 plat_get_mem_offset(uint64_t paddr, uint64_t *offp)
1006 {
1007 	if (opl_get_mem_offset == NULL) {
1008 		return (ENOTSUP);
1009 	}
1010 	return (opl_get_mem_offset(paddr, offp));
1011 }
1012 
1013 int
1014 plat_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
1015 {
1016 	if (opl_get_mem_addr == NULL) {
1017 		return (ENOTSUP);
1018 	}
1019 	return (opl_get_mem_addr(unum, sid, offset, addrp));
1020 }
1021 
1022 void
1023 plat_lock_delay(int *backoff)
1024 {
1025 	int i;
1026 	int cnt;
1027 	int flag;
1028 	int ctr;
1029 	hrtime_t delay_start;
1030 	/*
1031 	 * Platform specific lock delay code for OPL
1032 	 *
1033 	 * Using staged linear increases in the delay.
1034 	 * The sleep instruction is the preferred method of delay,
1035 	 * but is too large of granularity for the initial backoff.
1036 	 */
1037 
1038 	if (*backoff == 0) *backoff = OPL_BOFF_BASE;
1039 
1040 	flag = !*backoff;
1041 
1042 	if (*backoff < OPL_BOFF_CAP1) {
1043 		/*
1044 		 * If desired backoff is long enough,
1045 		 * use sleep for most of it
1046 		 */
1047 		for (cnt = *backoff;
1048 			cnt >= OPL_BOFF_SLEEP;
1049 			cnt -= OPL_BOFF_SLEEP) {
1050 			cpu_smt_pause();
1051 		}
1052 		/*
1053 		 * spin for small remainder of backoff
1054 		 *
1055 		 * fake call to nulldev included to prevent
1056 		 * compiler from optimizing out the spin loop
1057 		 */
1058 		for (ctr = cnt * OPL_BOFF_SPIN; ctr; ctr--) {
1059 			if (flag) (void) nulldev();
1060 		}
1061 	} else {
1062 		/* backoff is very large.  Fill it by sleeping */
1063 		delay_start = gethrtime();
1064 		cnt = *backoff/OPL_BOFF_SLEEP;
1065 		/*
1066 		 * use sleep instructions for delay
1067 		 */
1068 		for (i = 0; i < cnt; i++) {
1069 			cpu_smt_pause();
1070 		}
1071 
1072 		/*
1073 		 * Note: if the other strand executes a sleep instruction,
1074 		 * then the sleep ends immediately with a minimum time of
1075 		 * 42 clocks.  We check gethrtime to insure we have
1076 		 * waited long enough.  And we include both a short
1077 		 * spin loop and a sleep for any final delay time.
1078 		 */
1079 
1080 		while ((gethrtime() - delay_start) < cnt * OPL_BOFF_TM) {
1081 			cpu_smt_pause();
1082 			for (ctr = OPL_BOFF_SPIN; ctr; ctr--) {
1083 				if (flag) (void) nulldev();
1084 			}
1085 		}
1086 	}
1087 
1088 	/*
1089 	 * We adjust the backoff in three linear stages
1090 	 * The initial stage has small increases as this phase is
1091 	 * usually handle locks with light contention.  We don't want
1092 	 * to have a long backoff on a lock that is available.
1093 	 *
1094 	 * In the second stage, we are in transition, unsure whether
1095 	 * the lock is under heavy contention.  As the failures to
1096 	 * obtain the lock increase, we back off further.
1097 	 *
1098 	 * For the final stage, we are in a heavily contended or
1099 	 * long held long so we want to reduce the number of tries.
1100 	 */
1101 	if (*backoff < OPL_BOFF_CAP1) {
1102 		*backoff += 1;
1103 	} else {
1104 		if (*backoff < OPL_BOFF_CAP2) {
1105 			*backoff += OPL_BOFF_SLEEP;
1106 		} else {
1107 			*backoff += 2 * OPL_BOFF_SLEEP;
1108 		}
1109 		if (*backoff > OPL_BOFF_MAX) {
1110 			*backoff = OPL_BOFF_MAX;
1111 		}
1112 	}
1113 }
1114