xref: /titanic_51/usr/src/uts/sun4u/opl/os/opl.c (revision a60fc142342386d0b786e65fba901234400d7020)
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 
341 	if (kernel_cage_enable) {
342 		pgcnt_t preferred_cage_size;
343 
344 		preferred_cage_size =
345 			MAX(opl_startup_cage_size, total_pages / 256);
346 
347 		ml = opl_memlist_per_board(phys_avail);
348 
349 		/*
350 		 * Note: we are assuming that post has load the
351 		 * whole show in to the high end of memory. Having
352 		 * taken this leap, we copy the whole of phys_avail
353 		 * the glist and arrange for the cage to grow
354 		 * downward (descending pfns).
355 		 */
356 		kcage_range_init(ml, KCAGE_DOWN, preferred_cage_size);
357 
358 		/* free the memlist */
359 		do {
360 			tml = ml->next;
361 			kmem_free(ml, sizeof (struct memlist));
362 			ml = tml;
363 		} while (ml != NULL);
364 	}
365 
366 	if (kcage_on)
367 		cmn_err(CE_NOTE, "!DR Kernel Cage is ENABLED");
368 	else
369 		cmn_err(CE_NOTE, "!DR Kernel Cage is DISABLED");
370 }
371 
372 /*ARGSUSED*/
373 int
374 plat_cpu_poweron(struct cpu *cp)
375 {
376 	int (*opl_cpu_poweron)(struct cpu *) = NULL;
377 
378 	opl_cpu_poweron =
379 	    (int (*)(struct cpu *))kobj_getsymvalue("drmach_cpu_poweron", 0);
380 
381 	if (opl_cpu_poweron == NULL)
382 		return (ENOTSUP);
383 	else
384 		return ((opl_cpu_poweron)(cp));
385 
386 }
387 
388 /*ARGSUSED*/
389 int
390 plat_cpu_poweroff(struct cpu *cp)
391 {
392 	int (*opl_cpu_poweroff)(struct cpu *) = NULL;
393 
394 	opl_cpu_poweroff =
395 	    (int (*)(struct cpu *))kobj_getsymvalue("drmach_cpu_poweroff", 0);
396 
397 	if (opl_cpu_poweroff == NULL)
398 		return (ENOTSUP);
399 	else
400 		return ((opl_cpu_poweroff)(cp));
401 
402 }
403 
404 int
405 plat_max_boards(void)
406 {
407 	return (OPL_MAX_BOARDS);
408 }
409 
410 int
411 plat_max_cpu_units_per_board(void)
412 {
413 	return (OPL_MAX_CPU_PER_BOARD);
414 }
415 
416 int
417 plat_max_mem_units_per_board(void)
418 {
419 	return (OPL_MAX_MEM_UNITS_PER_BOARD);
420 }
421 
422 int
423 plat_max_io_units_per_board(void)
424 {
425 	return (OPL_MAX_IO_UNITS_PER_BOARD);
426 }
427 
428 int
429 plat_max_cmp_units_per_board(void)
430 {
431 	return (OPL_MAX_CMP_UNITS_PER_BOARD);
432 }
433 
434 int
435 plat_max_core_units_per_board(void)
436 {
437 	return (OPL_MAX_CORE_UNITS_PER_BOARD);
438 }
439 
440 int
441 plat_pfn_to_mem_node(pfn_t pfn)
442 {
443 	return (pfn >> mem_node_pfn_shift);
444 }
445 
446 /* ARGSUSED */
447 void
448 plat_build_mem_nodes(u_longlong_t *list, size_t nelems)
449 {
450 	size_t	elem;
451 	pfn_t	basepfn;
452 	pgcnt_t	npgs;
453 	uint64_t	boundary, ssize;
454 	uint64_t	low, high;
455 
456 	/*
457 	 * OPL mem slices are always aligned on a 256GB boundary.
458 	 */
459 	mem_node_pfn_shift = OPL_MC_MEMBOARD_SHIFT - MMU_PAGESHIFT;
460 	mem_node_physalign = 0;
461 
462 	/*
463 	 * Boot install lists are arranged <addr, len>, <addr, len>, ...
464 	 */
465 	ssize = (1ull << OPL_MC_MEMBOARD_SHIFT);
466 	for (elem = 0; elem < nelems; elem += 2) {
467 		low  = (uint64_t)list[elem];
468 		high = low+(uint64_t)(list[elem+1]);
469 		while (low < high) {
470 			boundary = roundup(low+1, ssize);
471 			boundary = MIN(high, boundary);
472 			basepfn = btop(low);
473 			npgs = btop(boundary - low);
474 			mem_node_add_slice(basepfn, basepfn + npgs - 1);
475 			low = boundary;
476 		}
477 	}
478 }
479 
480 /*
481  * Find the CPU associated with a slice at boot-time.
482  */
483 void
484 plat_fill_mc(pnode_t nodeid)
485 {
486 	int board;
487 	int memnode;
488 	struct {
489 		uint64_t	addr;
490 		uint64_t	size;
491 	} mem_range;
492 
493 	if (prom_getprop(nodeid, "board#", (caddr_t)&board) < 0) {
494 		panic("Can not find board# property in mc node %x", nodeid);
495 	}
496 	if (prom_getprop(nodeid, "sb-mem-ranges", (caddr_t)&mem_range) < 0) {
497 		panic("Can not find sb-mem-ranges property in mc node %x",
498 			nodeid);
499 	}
500 	memnode = mem_range.addr >> OPL_MC_MEMBOARD_SHIFT;
501 	plat_assign_lgrphand_to_mem_node(board, memnode);
502 }
503 
504 /*
505  * Return the platform handle for the lgroup containing the given CPU
506  *
507  * For OPL, lgroup platform handle == board #.
508  */
509 
510 extern int mpo_disabled;
511 extern lgrp_handle_t lgrp_default_handle;
512 
513 lgrp_handle_t
514 plat_lgrp_cpu_to_hand(processorid_t id)
515 {
516 	lgrp_handle_t plathand;
517 
518 	/*
519 	 * Return the real platform handle for the CPU until
520 	 * such time as we know that MPO should be disabled.
521 	 * At that point, we set the "mpo_disabled" flag to true,
522 	 * and from that point on, return the default handle.
523 	 *
524 	 * By the time we know that MPO should be disabled, the
525 	 * first CPU will have already been added to a leaf
526 	 * lgroup, but that's ok. The common lgroup code will
527 	 * double check that the boot CPU is in the correct place,
528 	 * and in the case where mpo should be disabled, will move
529 	 * it to the root if necessary.
530 	 */
531 	if (mpo_disabled) {
532 		/* If MPO is disabled, return the default (UMA) handle */
533 		plathand = lgrp_default_handle;
534 	} else
535 		plathand = (lgrp_handle_t)LSB_ID(id);
536 	return (plathand);
537 }
538 
539 /*
540  * Platform specific lgroup initialization
541  */
542 void
543 plat_lgrp_init(void)
544 {
545 	extern uint32_t lgrp_expand_proc_thresh;
546 	extern uint32_t lgrp_expand_proc_diff;
547 
548 	/*
549 	 * Set tuneables for the OPL architecture
550 	 *
551 	 * lgrp_expand_proc_thresh is the minimum load on the lgroups
552 	 * this process is currently running on before considering
553 	 * expanding threads to another lgroup.
554 	 *
555 	 * lgrp_expand_proc_diff determines how much less the remote lgroup
556 	 * must be loaded before expanding to it.
557 	 *
558 	 * Since remote latencies can be costly, attempt to keep 3 threads
559 	 * within the same lgroup before expanding to the next lgroup.
560 	 */
561 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX * 3;
562 	lgrp_expand_proc_diff = LGRP_LOADAVG_THREAD_MAX;
563 }
564 
565 /*
566  * Platform notification of lgroup (re)configuration changes
567  */
568 /*ARGSUSED*/
569 void
570 plat_lgrp_config(lgrp_config_flag_t evt, uintptr_t arg)
571 {
572 	update_membounds_t *umb;
573 	lgrp_config_mem_rename_t lmr;
574 	int sbd, tbd;
575 	lgrp_handle_t hand, shand, thand;
576 	int mnode, snode, tnode;
577 	pfn_t start, end;
578 
579 	if (mpo_disabled)
580 		return;
581 
582 	switch (evt) {
583 
584 	case LGRP_CONFIG_MEM_ADD:
585 		/*
586 		 * Establish the lgroup handle to memnode translation.
587 		 */
588 		umb = (update_membounds_t *)arg;
589 
590 		hand = umb->u_board;
591 		mnode = plat_pfn_to_mem_node(umb->u_base >> MMU_PAGESHIFT);
592 		plat_assign_lgrphand_to_mem_node(hand, mnode);
593 
594 		break;
595 
596 	case LGRP_CONFIG_MEM_DEL:
597 		/*
598 		 * Special handling for possible memory holes.
599 		 */
600 		umb = (update_membounds_t *)arg;
601 		hand = umb->u_board;
602 		if ((mnode = plat_lgrphand_to_mem_node(hand)) != -1) {
603 			if (mem_node_config[mnode].exists) {
604 				start = mem_node_config[mnode].physbase;
605 				end = mem_node_config[mnode].physmax;
606 				mem_node_pre_del_slice(start, end);
607 				mem_node_post_del_slice(start, end, 0);
608 			}
609 		}
610 
611 		break;
612 
613 	case LGRP_CONFIG_MEM_RENAME:
614 		/*
615 		 * During a DR copy-rename operation, all of the memory
616 		 * on one board is moved to another board -- but the
617 		 * addresses/pfns and memnodes don't change. This means
618 		 * the memory has changed locations without changing identity.
619 		 *
620 		 * Source is where we are copying from and target is where we
621 		 * are copying to.  After source memnode is copied to target
622 		 * memnode, the physical addresses of the target memnode are
623 		 * renamed to match what the source memnode had.  Then target
624 		 * memnode can be removed and source memnode can take its
625 		 * place.
626 		 *
627 		 * To do this, swap the lgroup handle to memnode mappings for
628 		 * the boards, so target lgroup will have source memnode and
629 		 * source lgroup will have empty target memnode which is where
630 		 * its memory will go (if any is added to it later).
631 		 *
632 		 * Then source memnode needs to be removed from its lgroup
633 		 * and added to the target lgroup where the memory was living
634 		 * but under a different name/memnode.  The memory was in the
635 		 * target memnode and now lives in the source memnode with
636 		 * different physical addresses even though it is the same
637 		 * memory.
638 		 */
639 		sbd = arg & 0xffff;
640 		tbd = (arg & 0xffff0000) >> 16;
641 		shand = sbd;
642 		thand = tbd;
643 		snode = plat_lgrphand_to_mem_node(shand);
644 		tnode = plat_lgrphand_to_mem_node(thand);
645 
646 		/*
647 		 * Special handling for possible memory holes.
648 		 */
649 		if (tnode != -1 && mem_node_config[tnode].exists) {
650 			start = mem_node_config[tnode].physbase;
651 			end = mem_node_config[tnode].physmax;
652 			mem_node_pre_del_slice(start, end);
653 			mem_node_post_del_slice(start, end, 0);
654 		}
655 
656 		plat_assign_lgrphand_to_mem_node(thand, snode);
657 		plat_assign_lgrphand_to_mem_node(shand, tnode);
658 
659 		lmr.lmem_rename_from = shand;
660 		lmr.lmem_rename_to = thand;
661 
662 		/*
663 		 * Remove source memnode of copy rename from its lgroup
664 		 * and add it to its new target lgroup
665 		 */
666 		lgrp_config(LGRP_CONFIG_MEM_RENAME, (uintptr_t)snode,
667 		    (uintptr_t)&lmr);
668 
669 		break;
670 
671 	default:
672 		break;
673 	}
674 }
675 
676 /*
677  * Return latency between "from" and "to" lgroups
678  *
679  * This latency number can only be used for relative comparison
680  * between lgroups on the running system, cannot be used across platforms,
681  * and may not reflect the actual latency.  It is platform and implementation
682  * specific, so platform gets to decide its value.  It would be nice if the
683  * number was at least proportional to make comparisons more meaningful though.
684  * NOTE: The numbers below are supposed to be load latencies for uncached
685  * memory divided by 10.
686  *
687  */
688 int
689 plat_lgrp_latency(lgrp_handle_t from, lgrp_handle_t to)
690 {
691 	/*
692 	 * Return min remote latency when there are more than two lgroups
693 	 * (root and child) and getting latency between two different lgroups
694 	 * or root is involved
695 	 */
696 	if (lgrp_optimizations() && (from != to ||
697 	    from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE))
698 		return (42);
699 	else
700 		return (35);
701 }
702 
703 /*
704  * Return platform handle for root lgroup
705  */
706 lgrp_handle_t
707 plat_lgrp_root_hand(void)
708 {
709 	if (mpo_disabled)
710 		return (lgrp_default_handle);
711 
712 	return (LGRP_DEFAULT_HANDLE);
713 }
714 
715 /*ARGSUSED*/
716 void
717 plat_freelist_process(int mnode)
718 {
719 }
720 
721 void
722 load_platform_drivers(void)
723 {
724 	(void) i_ddi_attach_pseudo_node("dr");
725 }
726 
727 /*
728  * No platform drivers on this platform
729  */
730 char *platform_module_list[] = {
731 	(char *)0
732 };
733 
734 /*ARGSUSED*/
735 void
736 plat_tod_fault(enum tod_fault_type tod_bad)
737 {
738 }
739 
740 /*ARGSUSED*/
741 void
742 cpu_sgn_update(ushort_t sgn, uchar_t state, uchar_t sub_state, int cpuid)
743 {
744 	static void (*scf_panic_callback)(int);
745 	static void (*scf_shutdown_callback)(int);
746 
747 	/*
748 	 * This is for notifing system panic/shutdown to SCF.
749 	 * In case of shutdown and panic, SCF call back
750 	 * function should be called.
751 	 *  <SCF call back functions>
752 	 *   scf_panic_callb()   : panicsys()->panic_quiesce_hw()
753 	 *   scf_shutdown_callb(): halt() or power_down() or reboot_machine()
754 	 * cpuid should be -1 and state should be SIGST_EXIT.
755 	 */
756 	if (state == SIGST_EXIT && cpuid == -1) {
757 
758 		/*
759 		 * find the symbol for the SCF panic callback routine in driver
760 		 */
761 		if (scf_panic_callback == NULL)
762 			scf_panic_callback = (void (*)(int))
763 				modgetsymvalue("scf_panic_callb", 0);
764 		if (scf_shutdown_callback == NULL)
765 			scf_shutdown_callback = (void (*)(int))
766 				modgetsymvalue("scf_shutdown_callb", 0);
767 
768 		switch (sub_state) {
769 		case SIGSUBST_PANIC:
770 			if (scf_panic_callback == NULL) {
771 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
772 				    "scf_panic_callb not found\n");
773 				return;
774 			}
775 			scf_panic_callback(SIGSUBST_PANIC);
776 			break;
777 
778 		case SIGSUBST_HALT:
779 			if (scf_shutdown_callback == NULL) {
780 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
781 				    "scf_shutdown_callb not found\n");
782 				return;
783 			}
784 			scf_shutdown_callback(SIGSUBST_HALT);
785 			break;
786 
787 		case SIGSUBST_ENVIRON:
788 			if (scf_shutdown_callback == NULL) {
789 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
790 				    "scf_shutdown_callb not found\n");
791 				return;
792 			}
793 			scf_shutdown_callback(SIGSUBST_ENVIRON);
794 			break;
795 
796 		case SIGSUBST_REBOOT:
797 			if (scf_shutdown_callback == NULL) {
798 				cmn_err(CE_NOTE, "!cpu_sgn_update: "
799 				    "scf_shutdown_callb not found\n");
800 				return;
801 			}
802 			scf_shutdown_callback(SIGSUBST_REBOOT);
803 			break;
804 		}
805 	}
806 }
807 
808 /*ARGSUSED*/
809 int
810 plat_get_mem_unum(int synd_code, uint64_t flt_addr, int flt_bus_id,
811 	int flt_in_memory, ushort_t flt_status,
812 	char *buf, int buflen, int *lenp)
813 {
814 	/*
815 	 * check if it's a Memory error.
816 	 */
817 	if (flt_in_memory) {
818 		if (opl_get_mem_unum != NULL) {
819 			return (opl_get_mem_unum(synd_code, flt_addr,
820 				buf, buflen, lenp));
821 		} else {
822 			return (ENOTSUP);
823 		}
824 	} else {
825 		return (ENOTSUP);
826 	}
827 }
828 
829 /*ARGSUSED*/
830 int
831 plat_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
832 {
833 	int	ret = 0;
834 	uint_t	sb;
835 	int	plen;
836 
837 	sb = opl_get_physical_board(LSB_ID(cpuid));
838 	if (sb == -1) {
839 		return (ENXIO);
840 	}
841 
842 	/*
843 	 * opl_cur_model is assigned here
844 	 */
845 	if (opl_cur_model == NULL) {
846 		set_model_info();
847 	}
848 
849 	ASSERT((opl_cur_model - opl_models) == (opl_cur_model->model_type));
850 
851 	switch (opl_cur_model->model_type) {
852 	case FF1:
853 		plen = snprintf(buf, buflen, "/%s/CPUM%d", "MBU_A",
854 		    CHIP_ID(cpuid) / 2);
855 		break;
856 
857 	case FF2:
858 		plen = snprintf(buf, buflen, "/%s/CPUM%d", "MBU_B",
859 		    (CHIP_ID(cpuid) / 2) + (sb * 2));
860 		break;
861 
862 	case DC1:
863 	case DC2:
864 	case DC3:
865 		plen = snprintf(buf, buflen, "/%s%02d/CPUM%d", "CMU", sb,
866 		    CHIP_ID(cpuid));
867 		break;
868 
869 	default:
870 		/* This should never happen */
871 		return (ENODEV);
872 	}
873 
874 	if (plen >= buflen) {
875 		ret = ENOSPC;
876 	} else {
877 		if (lenp)
878 			*lenp = strlen(buf);
879 	}
880 	return (ret);
881 }
882 
883 #define	SCF_PUTINFO(f, s, p)	\
884 	f(KEY_ESCF, 0x01, 0, s, p)
885 void
886 plat_nodename_set(void)
887 {
888 	void *datap;
889 	static int (*scf_service_function)(uint32_t, uint8_t,
890 	    uint32_t, uint32_t, void *);
891 	int counter = 5;
892 
893 	/*
894 	 * find the symbol for the SCF put routine in driver
895 	 */
896 	if (scf_service_function == NULL)
897 		scf_service_function =
898 			(int (*)(uint32_t, uint8_t, uint32_t, uint32_t, void *))
899 			modgetsymvalue("scf_service_putinfo", 0);
900 
901 	/*
902 	 * If the symbol was found, call it.  Otherwise, log a note (but not to
903 	 * the console).
904 	 */
905 
906 	if (scf_service_function == NULL) {
907 		cmn_err(CE_NOTE,
908 		    "!plat_nodename_set: scf_service_putinfo not found\n");
909 		return;
910 	}
911 
912 	datap =
913 	    (struct utsname *)kmem_zalloc(sizeof (struct utsname), KM_SLEEP);
914 
915 	if (datap == NULL) {
916 		return;
917 	}
918 
919 	bcopy((struct utsname *)&utsname,
920 	    (struct utsname *)datap, sizeof (struct utsname));
921 
922 	while ((SCF_PUTINFO(scf_service_function,
923 		sizeof (struct utsname), datap) == EBUSY) && (counter-- > 0)) {
924 		delay(10 * drv_usectohz(1000000));
925 	}
926 	if (counter == 0)
927 		cmn_err(CE_NOTE,
928 			"!plat_nodename_set: "
929 			"scf_service_putinfo not responding\n");
930 
931 	kmem_free(datap, sizeof (struct utsname));
932 }
933 
934 caddr_t	efcode_vaddr = NULL;
935 
936 /*
937  * Preallocate enough memory for fcode claims.
938  */
939 
940 caddr_t
941 efcode_alloc(caddr_t alloc_base)
942 {
943 	caddr_t efcode_alloc_base = (caddr_t)roundup((uintptr_t)alloc_base,
944 	    MMU_PAGESIZE);
945 	caddr_t vaddr;
946 
947 	/*
948 	 * allocate the physical memory for the Oberon fcode.
949 	 */
950 	if ((vaddr = (caddr_t)BOP_ALLOC(bootops, efcode_alloc_base,
951 	    efcode_size, MMU_PAGESIZE)) == NULL)
952 		cmn_err(CE_PANIC, "Cannot allocate Efcode Memory");
953 
954 	efcode_vaddr = vaddr;
955 
956 	return (efcode_alloc_base + efcode_size);
957 }
958 
959 caddr_t
960 plat_startup_memlist(caddr_t alloc_base)
961 {
962 	caddr_t tmp_alloc_base;
963 
964 	tmp_alloc_base = efcode_alloc(alloc_base);
965 	tmp_alloc_base =
966 	    (caddr_t)roundup((uintptr_t)tmp_alloc_base, ecache_alignsize);
967 	return (tmp_alloc_base);
968 }
969 
970 void
971 startup_platform(void)
972 {
973 }
974 
975 void
976 plat_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *info)
977 {
978 	int	impl;
979 
980 	impl = cpunodes[cpuid].implementation;
981 	if (IS_OLYMPUS_C(impl)) {
982 		info->mmu_idx = MMU_ID(cpuid);
983 		info->mmu_nctxs = 8192;
984 	} else {
985 		cmn_err(CE_PANIC, "Unknown processor %d", impl);
986 	}
987 }
988 
989 int
990 plat_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
991 {
992 	if (opl_get_mem_sid == NULL) {
993 		return (ENOTSUP);
994 	}
995 	return (opl_get_mem_sid(unum, buf, buflen, lenp));
996 }
997 
998 int
999 plat_get_mem_offset(uint64_t paddr, uint64_t *offp)
1000 {
1001 	if (opl_get_mem_offset == NULL) {
1002 		return (ENOTSUP);
1003 	}
1004 	return (opl_get_mem_offset(paddr, offp));
1005 }
1006 
1007 int
1008 plat_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
1009 {
1010 	if (opl_get_mem_addr == NULL) {
1011 		return (ENOTSUP);
1012 	}
1013 	return (opl_get_mem_addr(unum, sid, offset, addrp));
1014 }
1015 
1016 void
1017 plat_lock_delay(int *backoff)
1018 {
1019 	int i;
1020 	int cnt;
1021 	int flag;
1022 	int ctr;
1023 	hrtime_t delay_start;
1024 	/*
1025 	 * Platform specific lock delay code for OPL
1026 	 *
1027 	 * Using staged linear increases in the delay.
1028 	 * The sleep instruction is the preferred method of delay,
1029 	 * but is too large of granularity for the initial backoff.
1030 	 */
1031 
1032 	if (*backoff == 0) *backoff = OPL_BOFF_BASE;
1033 
1034 	flag = !*backoff;
1035 
1036 	if (*backoff < OPL_BOFF_CAP1) {
1037 		/*
1038 		 * If desired backoff is long enough,
1039 		 * use sleep for most of it
1040 		 */
1041 		for (cnt = *backoff;
1042 			cnt >= OPL_BOFF_SLEEP;
1043 			cnt -= OPL_BOFF_SLEEP) {
1044 			cpu_smt_pause();
1045 		}
1046 		/*
1047 		 * spin for small remainder of backoff
1048 		 *
1049 		 * fake call to nulldev included to prevent
1050 		 * compiler from optimizing out the spin loop
1051 		 */
1052 		for (ctr = cnt * OPL_BOFF_SPIN; ctr; ctr--) {
1053 			if (flag) (void) nulldev();
1054 		}
1055 	} else {
1056 		/* backoff is very large.  Fill it by sleeping */
1057 		delay_start = gethrtime();
1058 		cnt = *backoff/OPL_BOFF_SLEEP;
1059 		/*
1060 		 * use sleep instructions for delay
1061 		 */
1062 		for (i = 0; i < cnt; i++) {
1063 			cpu_smt_pause();
1064 		}
1065 
1066 		/*
1067 		 * Note: if the other strand executes a sleep instruction,
1068 		 * then the sleep ends immediately with a minimum time of
1069 		 * 42 clocks.  We check gethrtime to insure we have
1070 		 * waited long enough.  And we include both a short
1071 		 * spin loop and a sleep for any final delay time.
1072 		 */
1073 
1074 		while ((gethrtime() - delay_start) < cnt * OPL_BOFF_TM) {
1075 			cpu_smt_pause();
1076 			for (ctr = OPL_BOFF_SPIN; ctr; ctr--) {
1077 				if (flag) (void) nulldev();
1078 			}
1079 		}
1080 	}
1081 
1082 	/*
1083 	 * We adjust the backoff in three linear stages
1084 	 * The initial stage has small increases as this phase is
1085 	 * usually handle locks with light contention.  We don't want
1086 	 * to have a long backoff on a lock that is available.
1087 	 *
1088 	 * In the second stage, we are in transition, unsure whether
1089 	 * the lock is under heavy contention.  As the failures to
1090 	 * obtain the lock increase, we back off further.
1091 	 *
1092 	 * For the final stage, we are in a heavily contended or
1093 	 * long held long so we want to reduce the number of tries.
1094 	 */
1095 	if (*backoff < OPL_BOFF_CAP1) {
1096 		*backoff += 1;
1097 	} else {
1098 		if (*backoff < OPL_BOFF_CAP2) {
1099 			*backoff += OPL_BOFF_SLEEP;
1100 		} else {
1101 			*backoff += 2 * OPL_BOFF_SLEEP;
1102 		}
1103 		if (*backoff > OPL_BOFF_MAX) {
1104 			*backoff = OPL_BOFF_MAX;
1105 		}
1106 	}
1107 }
1108