xref: /titanic_41/usr/src/uts/i86pc/os/lgrpplat.c (revision 49218d4f8e4d84d1c08aeb267bcf6e451f2056dc)
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 
30 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
31 #include <sys/cmn_err.h>
32 #include <sys/cpupart.h>
33 #include <sys/cpuvar.h>
34 #include <sys/lgrp.h>
35 #include <sys/machsystm.h>
36 #include <sys/memlist.h>
37 #include <sys/memnode.h>
38 #include <sys/mman.h>
39 #include <sys/pci_impl.h>	/* for PCI configuration space macros */
40 #include <sys/param.h>
41 #include <sys/promif.h>		/* for prom_printf() */
42 #include <sys/systm.h>
43 #include <sys/thread.h>
44 #include <sys/types.h>
45 #include <sys/var.h>
46 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
47 #include <vm/hat_i86.h>
48 #include <vm/seg_kmem.h>
49 #include <vm/vm_dep.h>
50 
51 
52 
53 /*
54  * lgroup platform support for x86 platforms.
55  */
56 
57 #define	MAX_NODES		8
58 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
59 
60 #define	LGRP_PLAT_CPU_TO_NODE(cpu)	(chip_plat_get_chipid(cpu))
61 
62 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
63 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
64 
65 
66 /*
67  * Multiprocessor Opteron machines have Non Uniform Memory Access (NUMA).
68  *
69  * Until System Affinity Resource Table (SRAT) becomes part of ACPI standard,
70  * we need to examine registers in PCI configuration space to determine how
71  * many nodes are in the system and which CPUs and memory are in each node.
72  * This could be determined by probing all memory from each CPU, but that is
73  * too expensive to do while booting the kernel.
74  *
75  * NOTE: Using these PCI configuration space registers to determine this
76  *       locality info is Opteron K8 specific and not guaranteed to work on
77  *       the next generation Opteron processor.  Furthermore, we assume that
78  *	 there is one CPU per node and CPU 0 is in node 0, CPU 1 is in node 1,
79  *	 etc. which should be true for Opteron K8....
80  */
81 
82 /*
83  * Opteron DRAM Address Map in PCI configuration space gives base and limit
84  * of physical memory in each node for Opteron K8.  The following constants
85  * and macros define their contents, structure, and access.
86  */
87 
88 /*
89  * How many bits to shift Opteron DRAM Address Map base and limit registers
90  * to get actual value
91  */
92 #define	OPT_DRAMADDR_LSHIFT_ADDR	8	/* shift left for address */
93 
94 #define	OPT_DRAMADDR_MASK_OFF	0xFFFFFF	/* offset for address */
95 
96 /*
97  * Bit masks defining what's in Opteron DRAM Address Map base register
98  */
99 #define	OPT_DRAMBASE_MASK_RE		0x1		/* read enable */
100 #define	OPT_DRAMBASE_MASK_WE		0x2		/* write enable */
101 #define	OPT_DRAMBASE_MASK_INTRLVEN	0x700		/* interleave */
102 
103 #define	OPT_DRAMBASE_MASK_ADDR	0xFFFF0000	/* address bits 39-24 */
104 
105 /*
106  * Macros to get values from Opteron DRAM Address Map base register
107  */
108 #define	OPT_DRAMBASE(reg) \
109 	(((u_longlong_t)reg & OPT_DRAMBASE_MASK_ADDR) << \
110 	    OPT_DRAMADDR_LSHIFT_ADDR)
111 
112 
113 /*
114  * Bit masks defining what's in Opteron DRAM Address Map limit register
115  */
116 #define	OPT_DRAMLIMIT_MASK_DSTNODE	0x7		/* destination node */
117 #define	OPT_DRAMLIMIT_MASK_INTRLVSEL	0x70		/* interleave select */
118 #define	OPT_DRAMLIMIT_MASK_ADDR		0xFFFF0000	/* addr bits 39-24 */
119 
120 /*
121  * Macros to get values from Opteron DRAM Address Map limit register
122  */
123 #define	OPT_DRAMLIMIT(reg) \
124 	(((u_longlong_t)reg & OPT_DRAMLIMIT_MASK_ADDR) << \
125 	    OPT_DRAMADDR_LSHIFT_ADDR)
126 
127 
128 /*
129  * Opteron Node ID register in PCI configuration space contains
130  * number of nodes in system, etc. for Opteron K8.  The following
131  * constants and macros define its contents, structure, and access.
132  */
133 
134 /*
135  * Bit masks defining what's in Opteron Node ID register
136  */
137 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
138 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
139 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
140 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
141 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
142 
143 /*
144  * How many bits in Opteron Node ID register to shift right to get actual value
145  */
146 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
147 
148 /*
149  * Macros to get values from Opteron Node ID register
150  */
151 #define	OPT_NODE_CNT(reg) \
152 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
153 
154 
155 /*
156  * PCI configuration space registers accessed by specifying
157  * a bus, device, function, and offset.  The following constants
158  * define the values needed to access Opteron K8 configuration
159  * info to determine its node topology
160  */
161 
162 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
163 
164 /*
165  * Opteron PCI configuration space register function values
166  */
167 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
168 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
169 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
170 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
171 
172 /*
173  * PCI Configuration Space register offsets
174  */
175 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
176 #define	OPT_PCS_OFF_DRAMBASE	0x40	/* DRAM Base register (node 0) */
177 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
178 
179 /*
180  * Opteron PCI Configuration Space device IDs for nodes
181  */
182 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
183 
184 
185 /*
186  * Bookkeeping for latencies seen during probing (used for verification)
187  */
188 typedef	struct lgrp_plat_latency_acct {
189 	hrtime_t	la_value;	/* latency value */
190 	int		la_count;	/* occurrences */
191 } lgrp_plat_latency_acct_t;
192 
193 
194 /*
195  * Choices for probing to determine lgroup topology
196  */
197 typedef	enum lgrp_plat_probe_op {
198 	LGRP_PLAT_PROBE_PGCPY,		/* Use page copy */
199 	LGRP_PLAT_PROBE_VENDOR		/* Read vendor ID on Northbridge */
200 } lgrp_plat_probe_op_t;
201 
202 
203 /*
204  * Opteron DRAM address map gives base and limit for physical memory in a node
205  */
206 typedef	struct opt_dram_addr_map {
207 	uint32_t	base;
208 	uint32_t	limit;
209 } opt_dram_addr_map_t;
210 
211 
212 /*
213  * Starting and ending page for physical memory in node
214  */
215 typedef	struct phys_addr_map {
216 	pfn_t	start;
217 	pfn_t	end;
218 } phys_addr_map_t;
219 
220 
221 /*
222  * Opteron DRAM address map for each node
223  */
224 struct opt_dram_addr_map	opt_dram_map[MAX_NODES];
225 
226 /*
227  * Node ID register contents for each node
228  */
229 uint_t				opt_node_info[MAX_NODES];
230 
231 /*
232  * Whether memory is interleaved across nodes causing MPO to be disabled
233  */
234 int			lgrp_plat_mem_intrlv = 0;
235 
236 /*
237  * Number of nodes in system
238  */
239 uint_t			lgrp_plat_node_cnt = 1;
240 
241 /*
242  * Physical address range for memory in each node
243  */
244 phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
245 
246 /*
247  * Probe costs (individual and total) and flush cost
248  */
249 hrtime_t		lgrp_plat_flush_cost = 0;
250 hrtime_t		lgrp_plat_probe_cost = 0;
251 hrtime_t		lgrp_plat_probe_cost_total = 0;
252 
253 /*
254  * Error code for latency adjustment and verification
255  */
256 int			lgrp_plat_probe_error_code = 0;
257 
258 /*
259  * How much latencies were off from minimum values gotten
260  */
261 hrtime_t		lgrp_plat_probe_errors[MAX_NODES][MAX_NODES];
262 
263 /*
264  * Unique probe latencies and number of occurrences of each
265  */
266 lgrp_plat_latency_acct_t	lgrp_plat_probe_lat_acct[MAX_NODES];
267 
268 /*
269  * Size of memory buffer in each node for probing
270  */
271 size_t			lgrp_plat_probe_memsize = 0;
272 
273 /*
274  * Virtual address of page in each node for probing
275  */
276 caddr_t			lgrp_plat_probe_memory[MAX_NODES];
277 
278 /*
279  * Number of unique latencies in probe times
280  */
281 int			lgrp_plat_probe_nlatencies = 0;
282 
283 /*
284  * How many rounds of probing to do
285  */
286 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
287 
288 /*
289  * Number of samples to take when probing each node
290  */
291 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
292 
293 /*
294  * How to probe to determine lgroup topology
295  */
296 lgrp_plat_probe_op_t	lgrp_plat_probe_op = LGRP_PLAT_PROBE_VENDOR;
297 
298 /*
299  * PFN of page in each node for probing
300  */
301 pfn_t			lgrp_plat_probe_pfn[MAX_NODES];
302 
303 /*
304  * Whether probe time was suspect (ie. not within tolerance of value that it
305  * should match)
306  */
307 int			lgrp_plat_probe_suspect[MAX_NODES][MAX_NODES];
308 
309 /*
310  * How long it takes to access memory from each node
311  */
312 hrtime_t		lgrp_plat_probe_times[MAX_NODES][MAX_NODES];
313 
314 /*
315  * Min and max node memory probe times seen
316  */
317 hrtime_t		lgrp_plat_probe_time_max = 0;
318 hrtime_t		lgrp_plat_probe_time_min = -1;
319 hrtime_t		lgrp_plat_probe_max[MAX_NODES][MAX_NODES];
320 hrtime_t		lgrp_plat_probe_min[MAX_NODES][MAX_NODES];
321 
322 
323 /*
324  * Allocate lgrp and lgrp stat arrays statically.
325  */
326 static lgrp_t	lgrp_space[NLGRP];
327 static int	nlgrps_alloc;
328 
329 struct lgrp_stats lgrp_stats[NLGRP];
330 
331 #define	CPUID_FAMILY_OPTERON	15
332 
333 uint_t	opt_family = 0;
334 uint_t	opt_model = 0;
335 uint_t	opt_probe_func = OPT_PCS_FUNC_DRAM;
336 
337 
338 /*
339  * Determine whether we're running on an AMD Opteron K8 machine
340  */
341 int
342 is_opteron(void)
343 {
344 	if (x86_vendor != X86_VENDOR_AMD)
345 		return (0);
346 
347 	if (cpuid_getfamily(CPU) == CPUID_FAMILY_OPTERON)
348 		return (1);
349 	else
350 		return (0);
351 }
352 
353 int
354 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
355 {
356 	if (max_mem_nodes == 1)
357 		return (0);
358 
359 	return ((int)hand);
360 }
361 
362 lgrp_handle_t
363 plat_mem_node_to_lgrphand(int mnode)
364 {
365 	if (max_mem_nodes == 1)
366 		return (LGRP_DEFAULT_HANDLE);
367 
368 	return ((lgrp_handle_t)mnode);
369 }
370 
371 int
372 plat_pfn_to_mem_node(pfn_t pfn)
373 {
374 	int	node;
375 
376 	if (max_mem_nodes == 1)
377 		return (0);
378 
379 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
380 		if (pfn >= lgrp_plat_node_memory[node].start &&
381 		    pfn <= lgrp_plat_node_memory[node].end)
382 			return (node);
383 	}
384 
385 	ASSERT(node < lgrp_plat_node_cnt);
386 	return (-1);
387 }
388 
389 /*
390  * Configure memory nodes for machines with more than one node (ie NUMA)
391  */
392 void
393 plat_build_mem_nodes(struct memlist *list)
394 {
395 	pfn_t	cur_start, cur_end;	/* start & end addr of subrange */
396 	pfn_t	start, end;		/* start & end addr of whole range */
397 
398 	/*
399 	 * Boot install lists are arranged <addr, len>, ...
400 	 */
401 	while (list) {
402 		int	node;
403 
404 		start = list->address >> PAGESHIFT;
405 		end = (list->address + list->size - 1) >> PAGESHIFT;
406 
407 		if (start > physmax) {
408 			list = list->next;
409 			continue;
410 		}
411 		if (end > physmax)
412 			end = physmax;
413 
414 		/*
415 		 * When there is only one memnode, just add memory to memnode
416 		 */
417 		if (max_mem_nodes == 1) {
418 			mem_node_add_slice(start, end);
419 			list = list->next;
420 			continue;
421 		}
422 
423 		/*
424 		 * mem_node_add_slice() expects to get a memory range that
425 		 * is within one memnode, so need to split any memory range
426 		 * that spans multiple memnodes into subranges that are each
427 		 * contained within one memnode when feeding them to
428 		 * mem_node_add_slice()
429 		 */
430 		cur_start = start;
431 		do {
432 			node = plat_pfn_to_mem_node(cur_start);
433 			ASSERT(cur_start >=
434 			    lgrp_plat_node_memory[node].start &&
435 			    cur_start <= lgrp_plat_node_memory[node].end);
436 
437 			cur_end = end;
438 
439 			/*
440 			 * End of current subrange should not span memnodes
441 			 */
442 			if (cur_end > lgrp_plat_node_memory[node].end)
443 				cur_end = lgrp_plat_node_memory[node].end;
444 
445 			mem_node_add_slice(cur_start, cur_end);
446 
447 			/*
448 			 * Next subrange starts after end of current one
449 			 */
450 			cur_start = cur_end + 1;
451 		} while (cur_end < end);
452 
453 		list = list->next;
454 	}
455 	mem_node_physalign = 0;
456 	mem_node_pfn_shift = 0;
457 }
458 
459 
460 /*
461  * Platform-specific initialization of lgroups
462  */
463 void
464 lgrp_plat_init(void)
465 {
466 	uint_t		bus;
467 	uint_t		dev;
468 	uint_t		node;
469 	uint_t		off;
470 
471 	extern lgrp_load_t	lgrp_expand_proc_thresh;
472 	extern lgrp_load_t	lgrp_expand_proc_diff;
473 
474 	/*
475 	 * Initialize as a UMA machine if this isn't an Opteron
476 	 */
477 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
478 		lgrp_plat_node_cnt = max_mem_nodes = 1;
479 		return;
480 	}
481 
482 	/*
483 	 * Read configuration registers from PCI configuration space to
484 	 * determine node information, which memory is in each node, etc.
485 	 *
486 	 * Write to PCI configuration space address register to specify
487 	 * which configuration register to read and read/write PCI
488 	 * configuration space data register to get/set contents
489 	 */
490 	bus = OPT_PCS_BUS_CONFIG;
491 	dev = OPT_PCS_DEV_NODE0;
492 	off = OPT_PCS_OFF_DRAMBASE;
493 
494 	/*
495 	 * Read node ID register for node 0 to get node count
496 	 */
497 	outl(PCI_CONFADD, PCI_CADDR1(bus, dev, OPT_PCS_FUNC_HT,
498 	    OPT_PCS_OFF_NODEID));
499 	opt_node_info[0] = inl(PCI_CONFDATA);
500 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
501 
502 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
503 		/*
504 		 * Read node ID register (except for node 0 which we just read)
505 		 */
506 		if (node > 0) {
507 			outl(PCI_CONFADD, PCI_CADDR1(bus, dev,
508 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID));
509 			opt_node_info[node] = inl(PCI_CONFDATA);
510 		}
511 
512 		/*
513 		 * Read DRAM base and limit registers which specify
514 		 * physical memory range of each node
515 		 */
516 		outl(PCI_CONFADD, PCI_CADDR1(bus, dev, OPT_PCS_FUNC_ADDRMAP,
517 		    off));
518 		opt_dram_map[node].base = inl(PCI_CONFDATA);
519 		if (opt_dram_map[node].base & OPT_DRAMBASE_MASK_INTRLVEN)
520 			lgrp_plat_mem_intrlv++;
521 
522 		off += 4;	/* limit register offset */
523 		outl(PCI_CONFADD, PCI_CADDR1(bus, dev, OPT_PCS_FUNC_ADDRMAP,
524 		    off));
525 		opt_dram_map[node].limit = inl(PCI_CONFDATA);
526 
527 		/*
528 		 * Increment device number to next node and register offset for
529 		 * DRAM base register of next node
530 		 */
531 		off += 4;
532 		dev++;
533 
534 		/*
535 		 * Get PFN for first page in each node,
536 		 * so we can probe memory to determine latency topology
537 		 */
538 		lgrp_plat_probe_pfn[node] =
539 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
540 
541 		/*
542 		 * Remember physical address range of each node for use later
543 		 */
544 		lgrp_plat_node_memory[node].start =
545 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
546 		lgrp_plat_node_memory[node].end =
547 		    btop(OPT_DRAMLIMIT(opt_dram_map[node].limit) |
548 		    OPT_DRAMADDR_MASK_OFF);
549 	}
550 
551 	/*
552 	 * Only use one memory node if memory is interleaved between any nodes
553 	 */
554 	if (lgrp_plat_mem_intrlv) {
555 		lgrp_plat_node_cnt = max_mem_nodes = 1;
556 		(void) lgrp_topo_ht_limit_set(1);
557 	} else {
558 		max_mem_nodes = lgrp_plat_node_cnt;
559 
560 		/*
561 		 * Probing errors can mess up the lgroup topology and force us
562 		 * fall back to a 2 level lgroup topology.  Here we bound how
563 		 * tall the lgroup topology can grow in hopes of avoiding any
564 		 * anamolies in probing from messing up the lgroup topology
565 		 * by limiting the accuracy of the latency topology.
566 		 *
567 		 * Assume that nodes will at least be configured in a ring,
568 		 * so limit height of lgroup topology to be less than number
569 		 * of nodes on a system with 4 or more nodes
570 		 */
571 		if (lgrp_plat_node_cnt >= 4 &&
572 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
573 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
574 	}
575 
576 	/*
577 	 * Lgroups on Opteron architectures have but a single physical
578 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
579 	 * so that lgrp_choose() will spread things out aggressively.
580 	 */
581 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
582 	lgrp_expand_proc_diff = 0;
583 }
584 
585 
586 /*
587  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
588  * be considered same
589  */
590 #define	LGRP_LAT_TOLERANCE_SHIFT	4
591 
592 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
593 
594 
595 /*
596  * Adjust latencies between nodes to be symmetric, normalize latencies between
597  * any nodes that are within some tolerance to be same, and make local
598  * latencies be same
599  */
600 static void
601 lgrp_plat_latency_adjust(void)
602 {
603 	int				i;
604 	int				j;
605 	int				k;
606 	int				l;
607 	u_longlong_t			max;
608 	u_longlong_t			min;
609 	u_longlong_t			t;
610 	u_longlong_t			t1;
611 	u_longlong_t			t2;
612 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LATENCY_CHANGE;
613 	int				lat_corrected[MAX_NODES][MAX_NODES];
614 
615 	/*
616 	 * Nothing to do when this is an UMA machine
617 	 */
618 	if (max_mem_nodes == 1)
619 		return;
620 
621 	/*
622 	 * Make sure that latencies are symmetric between any two nodes
623 	 * (ie. latency(node0, node1) == latency(node1, node0))
624 	 */
625 	for (i = 0; i < lgrp_plat_node_cnt; i++)
626 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
627 			t1 = lgrp_plat_probe_times[i][j];
628 			t2 = lgrp_plat_probe_times[j][i];
629 
630 			if (t1 == 0 || t2 == 0 || t1 == t2)
631 				continue;
632 
633 			/*
634 			 * Latencies should be same
635 			 * - Use minimum of two latencies which should be same
636 			 * - Track suspect probe times not within tolerance of
637 			 *   min value
638 			 * - Remember how much values are corrected by
639 			 */
640 			if (t1 > t2) {
641 				t = t2;
642 				lgrp_plat_probe_errors[i][j] += t1 - t2;
643 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
644 					lgrp_plat_probe_suspect[i][j]++;
645 					lgrp_plat_probe_suspect[j][i]++;
646 				}
647 			} else if (t2 > t1) {
648 				t = t1;
649 				lgrp_plat_probe_errors[j][i] += t2 - t1;
650 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
651 					lgrp_plat_probe_suspect[i][j]++;
652 					lgrp_plat_probe_suspect[j][i]++;
653 				}
654 			}
655 
656 			lgrp_plat_probe_times[i][j] =
657 			    lgrp_plat_probe_times[j][i] = t;
658 			lgrp_config(cflag, t1, t);
659 			lgrp_config(cflag, t2, t);
660 		}
661 
662 	/*
663 	 * Keep track of which latencies get corrected
664 	 */
665 	for (i = 0; i < MAX_NODES; i++)
666 		for (j = 0; j < MAX_NODES; j++)
667 			lat_corrected[i][j] = 0;
668 
669 	/*
670 	 * For every two nodes, see whether there is another pair of nodes which
671 	 * are about the same distance apart and make the latencies be the same
672 	 * if they are close enough together
673 	 */
674 	for (i = 0; i < lgrp_plat_node_cnt; i++)
675 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
676 			/*
677 			 * Pick one pair of nodes (i, j)
678 			 * and get latency between them
679 			 */
680 			t1 = lgrp_plat_probe_times[i][j];
681 
682 			/*
683 			 * Skip this pair of nodes if there isn't a latency
684 			 * for it yet
685 			 */
686 			if (t1 == 0)
687 				continue;
688 
689 			for (k = 0; k < lgrp_plat_node_cnt; k++)
690 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
691 					/*
692 					 * Pick another pair of nodes (k, l)
693 					 * not same as (i, j) and get latency
694 					 * between them
695 					 */
696 					if (k == i && l == j)
697 						continue;
698 
699 					t2 = lgrp_plat_probe_times[k][l];
700 
701 					/*
702 					 * Skip this pair of nodes if there
703 					 * isn't a latency for it yet
704 					 */
705 
706 					if (t2 == 0)
707 						continue;
708 
709 					/*
710 					 * Skip nodes (k, l) if they already
711 					 * have same latency as (i, j) or
712 					 * their latency isn't close enough to
713 					 * be considered/made the same
714 					 */
715 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
716 					    t1 >> lgrp_plat_probe_lt_shift) ||
717 					    (t2 > t1 && t2 - t1 >
718 					    t2 >> lgrp_plat_probe_lt_shift))
719 						continue;
720 
721 					/*
722 					 * Make latency(i, j) same as
723 					 * latency(k, l), try to use latency
724 					 * that has been adjusted already to get
725 					 * more consistency (if possible), and
726 					 * remember which latencies were
727 					 * adjusted for next time
728 					 */
729 					if (lat_corrected[i][j]) {
730 						t = t1;
731 						lgrp_config(cflag, t2, t);
732 						t2 = t;
733 					} else if (lat_corrected[k][l]) {
734 						t = t2;
735 						lgrp_config(cflag, t1, t);
736 						t1 = t;
737 					} else {
738 						if (t1 > t2)
739 							t = t2;
740 						else
741 							t = t1;
742 						lgrp_config(cflag, t1, t);
743 						lgrp_config(cflag, t2, t);
744 						t1 = t2 = t;
745 					}
746 
747 					lgrp_plat_probe_times[i][j] =
748 					    lgrp_plat_probe_times[k][l] = t;
749 
750 					lat_corrected[i][j] =
751 					    lat_corrected[k][l] = 1;
752 				}
753 		}
754 
755 	/*
756 	 * Local latencies should be same
757 	 * - Find min and max local latencies
758 	 * - Make all local latencies be minimum
759 	 */
760 	min = -1;
761 	max = 0;
762 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
763 		t = lgrp_plat_probe_times[i][i];
764 		if (t == 0)
765 			continue;
766 		if (min == -1 || t < min)
767 			min = t;
768 		if (t > max)
769 			max = t;
770 	}
771 	if (min != max) {
772 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
773 			int	local;
774 
775 			local = lgrp_plat_probe_times[i][i];
776 			if (local == 0)
777 				continue;
778 
779 			/*
780 			 * Track suspect probe times that aren't within
781 			 * tolerance of minimum local latency and how much
782 			 * probe times are corrected by
783 			 */
784 			if (local - min > min >> lgrp_plat_probe_lt_shift)
785 				lgrp_plat_probe_suspect[i][i]++;
786 
787 			lgrp_plat_probe_errors[i][i] += local - min;
788 
789 			/*
790 			 * Make local latencies be minimum
791 			 */
792 			lgrp_config(cflag, local, min);
793 			lgrp_plat_probe_times[i][i] = min;
794 		}
795 	}
796 
797 	/*
798 	 * Determine max probe time again since just adjusted latencies
799 	 */
800 	lgrp_plat_probe_time_max = 0;
801 	for (i = 0; i < lgrp_plat_node_cnt; i++)
802 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
803 			t = lgrp_plat_probe_times[i][j];
804 			if (t > lgrp_plat_probe_time_max)
805 				lgrp_plat_probe_time_max = t;
806 		}
807 }
808 
809 
810 /*
811  * Verify following about latencies between nodes:
812  *
813  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
814  * - Local latencies same
815  * - Local < remote
816  * - Number of latencies seen is reasonable
817  * - Number of occurrences of a given latency should be more than 1
818  *
819  * Returns:
820  *	0	Success
821  *	-1	Not symmetric
822  *	-2	Local latencies not same
823  *	-3	Local >= remote
824  *	-4	Wrong number of latencies
825  *	-5	Not enough occurrences of given latency
826  */
827 static int
828 lgrp_plat_latency_verify(void)
829 {
830 	int				i;
831 	int				j;
832 	lgrp_plat_latency_acct_t	*l;
833 	int				probed;
834 	u_longlong_t			t1;
835 	u_longlong_t			t2;
836 
837 	/*
838 	 * Nothing to do when this is an UMA machine, lgroup topology is
839 	 * limited to 2 levels, or there aren't any probe times yet
840 	 */
841 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
842 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
843 		return (0);
844 
845 	/*
846 	 * Make sure that latencies are symmetric between any two nodes
847 	 * (ie. latency(node0, node1) == latency(node1, node0))
848 	 */
849 	for (i = 0; i < lgrp_plat_node_cnt; i++)
850 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
851 			t1 = lgrp_plat_probe_times[i][j];
852 			t2 = lgrp_plat_probe_times[j][i];
853 
854 			if (t1 == 0 || t2 == 0 || t1 == t2)
855 				continue;
856 
857 			return (-1);
858 		}
859 
860 	/*
861 	 * Local latencies should be same
862 	 */
863 	t1 = lgrp_plat_probe_times[0][0];
864 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
865 		t2 = lgrp_plat_probe_times[i][i];
866 		if (t2 == 0)
867 			continue;
868 
869 		if (t1 == 0) {
870 			t1 = t2;
871 			continue;
872 		}
873 
874 		if (t1 != t2)
875 			return (-2);
876 	}
877 
878 	/*
879 	 * Local latencies should be less than remote
880 	 */
881 	if (t1) {
882 		for (i = 0; i < lgrp_plat_node_cnt; i++)
883 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
884 				t2 = lgrp_plat_probe_times[i][j];
885 				if (i == j || t2 == 0)
886 					continue;
887 
888 				if (t1 >= t2)
889 					return (-3);
890 			}
891 	}
892 
893 	/*
894 	 * Rest of checks are not very useful for machines with less than
895 	 * 4 nodes (which means less than 3 latencies on Opteron)
896 	 */
897 	if (lgrp_plat_node_cnt < 4)
898 		return (0);
899 
900 	/*
901 	 * Need to see whether done probing in order to verify number of
902 	 * latencies are correct
903 	 */
904 	probed = 0;
905 	for (i = 0; i < lgrp_plat_node_cnt; i++)
906 		if (lgrp_plat_probe_times[i][i])
907 			probed++;
908 
909 	if (probed != lgrp_plat_node_cnt)
910 		return (0);
911 
912 	/*
913 	 * Determine number of unique latencies seen in probe times,
914 	 * their values, and number of occurrences of each
915 	 */
916 	lgrp_plat_probe_nlatencies = 0;
917 	bzero(lgrp_plat_probe_lat_acct,
918 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
919 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
920 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
921 			int	k;
922 
923 			/*
924 			 * Look at each probe time
925 			 */
926 			t1 = lgrp_plat_probe_times[i][j];
927 			if (t1 == 0)
928 				continue;
929 
930 			/*
931 			 * Account for unique latencies
932 			 */
933 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
934 				l = &lgrp_plat_probe_lat_acct[k];
935 				if (t1 == l->la_value) {
936 					/*
937 					 * Increment number of occurrences
938 					 * if seen before
939 					 */
940 					l->la_count++;
941 					break;
942 				} else if (l->la_value == 0) {
943 					/*
944 					 * Record latency if haven't seen before
945 					 */
946 					l->la_value = t1;
947 					l->la_count++;
948 					lgrp_plat_probe_nlatencies++;
949 					break;
950 				}
951 			}
952 		}
953 	}
954 
955 	/*
956 	 * Number of latencies should be relative to number of
957 	 * nodes in system:
958 	 * - Same as nodes when nodes <= 2
959 	 * - Less than nodes when nodes > 2
960 	 * - Greater than 2 when nodes >= 4
961 	 */
962 	if ((lgrp_plat_node_cnt <= 2 &&
963 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
964 	    (lgrp_plat_node_cnt > 2 &&
965 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
966 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
967 	    lgrp_plat_probe_nlatencies <= 2))
968 		return (-4);
969 
970 	/*
971 	 * There should be more than one occurrence of every latency
972 	 * as long as probing is complete
973 	 */
974 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
975 		l = &lgrp_plat_probe_lat_acct[i];
976 		if (l->la_count <= 1)
977 			return (-5);
978 	}
979 	return (0);
980 }
981 
982 
983 /*
984  * Set lgroup latencies for 2 level lgroup topology
985  */
986 static void
987 lgrp_plat_2level_setup(void)
988 {
989 	int	i;
990 
991 	if (lgrp_plat_node_cnt >= 4)
992 		cmn_err(CE_NOTE,
993 		    "MPO only optimizing for local and remote\n");
994 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
995 		int	j;
996 
997 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
998 			if (i == j)
999 				lgrp_plat_probe_times[i][j] = 2;
1000 			else
1001 				lgrp_plat_probe_times[i][j] = 3;
1002 		}
1003 	}
1004 	lgrp_plat_probe_time_min = 2;
1005 	lgrp_plat_probe_time_max = 3;
1006 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1007 }
1008 
1009 
1010 /*
1011  * Return time needed to probe from current CPU to memory in given node
1012  */
1013 static hrtime_t
1014 lgrp_plat_probe_time(int to)
1015 {
1016 	caddr_t		buf;
1017 	uint_t		dev;
1018 	/* LINTED: set but not used in function */
1019 	volatile uint_t	dev_vendor;
1020 	hrtime_t	elapsed;
1021 	hrtime_t	end;
1022 	int		from;
1023 	int		i;
1024 	int		ipl;
1025 	hrtime_t	max;
1026 	hrtime_t	min;
1027 	hrtime_t	start;
1028 	extern int	use_sse_pagecopy;
1029 
1030 	/*
1031 	 * Determine ID of node containing current CPU
1032 	 */
1033 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1034 
1035 	/*
1036 	 * Do common work for probing main memory
1037 	 */
1038 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1039 		/*
1040 		 * Skip probing any nodes without memory and
1041 		 * set probe time to 0
1042 		 */
1043 		if (lgrp_plat_probe_memory[to] == NULL) {
1044 			lgrp_plat_probe_times[from][to] = 0;
1045 			return (0);
1046 		}
1047 
1048 		/*
1049 		 * Invalidate caches once instead of once every sample
1050 		 * which should cut cost of probing by a lot
1051 		 */
1052 		lgrp_plat_flush_cost = gethrtime();
1053 		invalidate_cache();
1054 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1055 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1056 	}
1057 
1058 	/*
1059 	 * Probe from current CPU to given memory using specified operation
1060 	 * and take specified number of samples
1061 	 */
1062 	max = 0;
1063 	min = -1;
1064 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1065 		lgrp_plat_probe_cost = gethrtime();
1066 
1067 		/*
1068 		 * Can't measure probe time if gethrtime() isn't working yet
1069 		 */
1070 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1071 			return (0);
1072 
1073 		switch (lgrp_plat_probe_op) {
1074 
1075 		case LGRP_PLAT_PROBE_PGCPY:
1076 		default:
1077 			/*
1078 			 * Measure how long it takes to copy page
1079 			 * on top of itself
1080 			 */
1081 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1082 
1083 			kpreempt_disable();
1084 			ipl = splhigh();
1085 			start = gethrtime();
1086 			if (use_sse_pagecopy)
1087 				hwblkpagecopy(buf, buf);
1088 			else
1089 				bcopy(buf, buf, PAGESIZE);
1090 			end = gethrtime();
1091 			elapsed = end - start;
1092 			splx(ipl);
1093 			kpreempt_enable();
1094 			break;
1095 
1096 		case LGRP_PLAT_PROBE_VENDOR:
1097 			/*
1098 			 * Measure how long it takes to read vendor ID from
1099 			 * Northbridge
1100 			 */
1101 			dev = OPT_PCS_DEV_NODE0 + to;
1102 			kpreempt_disable();
1103 			ipl = spl8();
1104 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1105 			    OPT_PCS_OFF_VENDOR));
1106 			start = gethrtime();
1107 			dev_vendor = inl(PCI_CONFDATA);
1108 			end = gethrtime();
1109 			elapsed = end - start;
1110 			splx(ipl);
1111 			kpreempt_enable();
1112 			break;
1113 		}
1114 
1115 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1116 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1117 
1118 		if (min == -1 || elapsed < min)
1119 			min = elapsed;
1120 		if (elapsed > max)
1121 			max = elapsed;
1122 	}
1123 
1124 	/*
1125 	 * Update minimum and maximum probe times between
1126 	 * these two nodes
1127 	 */
1128 	if (min < lgrp_plat_probe_min[from][to] ||
1129 	    lgrp_plat_probe_min[from][to] == 0)
1130 		lgrp_plat_probe_min[from][to] = min;
1131 
1132 	if (max > lgrp_plat_probe_max[from][to])
1133 		lgrp_plat_probe_max[from][to] = max;
1134 
1135 	return (min);
1136 }
1137 
1138 
1139 /*
1140  * Probe memory in each node from current CPU to determine latency topology
1141  */
1142 void
1143 lgrp_plat_probe(void)
1144 {
1145 	int		from;
1146 	int		i;
1147 	hrtime_t	probe_time;
1148 	int		to;
1149 
1150 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1151 		return;
1152 
1153 	/*
1154 	 * Determine ID of node containing current CPU
1155 	 */
1156 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1157 
1158 	/*
1159 	 * Don't need to probe if got times already
1160 	 */
1161 	if (lgrp_plat_probe_times[from][from] != 0)
1162 		return;
1163 
1164 	/*
1165 	 * Read vendor ID in Northbridge or read and write page(s)
1166 	 * in each node from current CPU and remember how long it takes,
1167 	 * so we can build latency topology of machine later.
1168 	 * This should approximate the memory latency between each node.
1169 	 */
1170 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1171 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1172 			/*
1173 			 * Get probe time and bail out if can't get it yet
1174 			 */
1175 			probe_time = lgrp_plat_probe_time(to);
1176 			if (probe_time == 0)
1177 				return;
1178 
1179 			/*
1180 			 * Keep lowest probe time as latency between nodes
1181 			 */
1182 			if (lgrp_plat_probe_times[from][to] == 0 ||
1183 			    probe_time < lgrp_plat_probe_times[from][to])
1184 				lgrp_plat_probe_times[from][to] = probe_time;
1185 
1186 			/*
1187 			 * Update overall minimum and maximum probe times
1188 			 * across all nodes
1189 			 */
1190 			if (probe_time < lgrp_plat_probe_time_min ||
1191 			    lgrp_plat_probe_time_min == -1)
1192 				lgrp_plat_probe_time_min = probe_time;
1193 			if (probe_time > lgrp_plat_probe_time_max)
1194 				lgrp_plat_probe_time_max = probe_time;
1195 		}
1196 
1197 	/*
1198 	 * - Fix up latencies such that local latencies are same,
1199 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1200 	 *
1201 	 * - Verify that latencies look ok
1202 	 *
1203 	 * - Fallback to just optimizing for local and remote if
1204 	 *   latencies didn't look right
1205 	 */
1206 	lgrp_plat_latency_adjust();
1207 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1208 	if (lgrp_plat_probe_error_code)
1209 		lgrp_plat_2level_setup();
1210 }
1211 
1212 
1213 /*
1214  * Platform-specific initialization
1215  */
1216 void
1217 lgrp_plat_main_init(void)
1218 {
1219 	int	curnode;
1220 	int	ht_limit;
1221 	int	i;
1222 
1223 	/*
1224 	 * Print a notice that MPO is disabled when memory is interleaved
1225 	 * across nodes....Would do this when it is discovered, but can't
1226 	 * because it happens way too early during boot....
1227 	 */
1228 	if (lgrp_plat_mem_intrlv)
1229 		cmn_err(CE_NOTE,
1230 		    "MPO disabled because memory is interleaved\n");
1231 
1232 	/*
1233 	 * Don't bother to do any probing if there is only one node or the
1234 	 * height of the lgroup topology less than or equal to 2
1235 	 */
1236 	ht_limit = lgrp_topo_ht_limit();
1237 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1238 		/*
1239 		 * Setup lgroup latencies for 2 level lgroup topology
1240 		 * (ie. local and remote only) if they haven't been set yet
1241 		 */
1242 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1243 		    lgrp_plat_probe_time_max == 0)
1244 			lgrp_plat_2level_setup();
1245 		return;
1246 	}
1247 
1248 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1249 		/*
1250 		 * Should have been able to probe from CPU 0 when it was added
1251 		 * to lgroup hierarchy, but may not have been able to then
1252 		 * because it happens so early in boot that gethrtime() hasn't
1253 		 * been initialized.  (:-(
1254 		 */
1255 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1256 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1257 			lgrp_plat_probe();
1258 
1259 		return;
1260 	}
1261 
1262 	/*
1263 	 * When probing memory, use one page for every sample to determine
1264 	 * lgroup topology and taking multiple samples
1265 	 */
1266 	if (lgrp_plat_probe_memsize == 0)
1267 		lgrp_plat_probe_memsize = PAGESIZE *
1268 		    lgrp_plat_probe_nsamples;
1269 
1270 	/*
1271 	 * Map memory in each node needed for probing to determine latency
1272 	 * topology
1273 	 */
1274 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1275 		int	mnode;
1276 
1277 		/*
1278 		 * Skip this node and leave its probe page NULL
1279 		 * if it doesn't have any memory
1280 		 */
1281 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1282 		if (!mem_node_config[mnode].exists) {
1283 			lgrp_plat_probe_memory[i] = NULL;
1284 			continue;
1285 		}
1286 
1287 		/*
1288 		 * Allocate one kernel virtual page
1289 		 */
1290 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1291 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1292 		if (lgrp_plat_probe_memory[i] == NULL) {
1293 			cmn_err(CE_WARN,
1294 			    "lgrp_plat_main_init: couldn't allocate memory");
1295 			return;
1296 		}
1297 
1298 		/*
1299 		 * Map virtual page to first page in node
1300 		 */
1301 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1302 		    lgrp_plat_probe_memsize,
1303 		    lgrp_plat_probe_pfn[i],
1304 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1305 		    HAT_LOAD_NOCONSIST);
1306 	}
1307 
1308 	/*
1309 	 * Probe from current CPU
1310 	 */
1311 	lgrp_plat_probe();
1312 }
1313 
1314 /*
1315  * Allocate additional space for an lgroup.
1316  */
1317 /* ARGSUSED */
1318 lgrp_t *
1319 lgrp_plat_alloc(lgrp_id_t lgrpid)
1320 {
1321 	lgrp_t *lgrp;
1322 
1323 	lgrp = &lgrp_space[nlgrps_alloc++];
1324 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1325 		return (NULL);
1326 	return (lgrp);
1327 }
1328 
1329 /*
1330  * Platform handling for (re)configuration changes
1331  */
1332 /* ARGSUSED */
1333 void
1334 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1335 {
1336 }
1337 
1338 /*
1339  * Return the platform handle for the lgroup containing the given CPU
1340  */
1341 /* ARGSUSED */
1342 lgrp_handle_t
1343 lgrp_plat_cpu_to_hand(processorid_t id)
1344 {
1345 	if (lgrp_plat_node_cnt == 1)
1346 		return (LGRP_DEFAULT_HANDLE);
1347 
1348 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1349 }
1350 
1351 /*
1352  * Return the platform handle of the lgroup that contains the physical memory
1353  * corresponding to the given page frame number
1354  */
1355 /* ARGSUSED */
1356 lgrp_handle_t
1357 lgrp_plat_pfn_to_hand(pfn_t pfn)
1358 {
1359 	int	mnode;
1360 
1361 	if (max_mem_nodes == 1)
1362 		return (LGRP_DEFAULT_HANDLE);
1363 
1364 	mnode = plat_pfn_to_mem_node(pfn);
1365 	return (MEM_NODE_2_LGRPHAND(mnode));
1366 }
1367 
1368 /*
1369  * Return the maximum number of lgrps supported by the platform.
1370  * Before lgrp topology is known it returns an estimate based on the number of
1371  * nodes. Once topology is known it returns the actual maximim number of lgrps
1372  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1373  * may not grow during system lifetime.
1374  */
1375 int
1376 lgrp_plat_max_lgrps()
1377 {
1378 	return (lgrp_topo_initialized ?
1379 	    lgrp_alloc_max + 1 :
1380 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1381 }
1382 
1383 /*
1384  * Return the number of free, allocatable, or installed
1385  * pages in an lgroup
1386  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1387  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1388  */
1389 /* ARGSUSED */
1390 static pgcnt_t
1391 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1392 {
1393 	struct memlist *mlist;
1394 	pgcnt_t npgs = 0;
1395 	extern struct memlist *phys_avail;
1396 	extern struct memlist *phys_install;
1397 
1398 	switch (query) {
1399 	case LGRP_MEM_SIZE_FREE:
1400 		return ((pgcnt_t)freemem);
1401 	case LGRP_MEM_SIZE_AVAIL:
1402 		memlist_read_lock();
1403 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1404 			npgs += btop(mlist->size);
1405 		memlist_read_unlock();
1406 		return (npgs);
1407 	case LGRP_MEM_SIZE_INSTALL:
1408 		memlist_read_lock();
1409 		for (mlist = phys_install; mlist; mlist = mlist->next)
1410 			npgs += btop(mlist->size);
1411 		memlist_read_unlock();
1412 		return (npgs);
1413 	default:
1414 		return ((pgcnt_t)0);
1415 	}
1416 }
1417 
1418 /*
1419  * Return the number of free pages in an lgroup.
1420  *
1421  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1422  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1423  * number of allocatable base pagesize pages corresponding to the
1424  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1425  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1426  * memory installed, regardless of whether or not it's usable.
1427  */
1428 pgcnt_t
1429 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1430 {
1431 	int	mnode;
1432 	pgcnt_t npgs = (pgcnt_t)0;
1433 	extern struct memlist *phys_avail;
1434 	extern struct memlist *phys_install;
1435 
1436 
1437 	if (plathand == LGRP_DEFAULT_HANDLE)
1438 		return (lgrp_plat_mem_size_default(plathand, query));
1439 
1440 	if (plathand != LGRP_NULL_HANDLE) {
1441 		mnode = plat_lgrphand_to_mem_node(plathand);
1442 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1443 			switch (query) {
1444 			case LGRP_MEM_SIZE_FREE:
1445 				npgs = MNODE_PGCNT(mnode);
1446 				break;
1447 			case LGRP_MEM_SIZE_AVAIL:
1448 				npgs = mem_node_memlist_pages(mnode,
1449 				    phys_avail);
1450 				break;
1451 			case LGRP_MEM_SIZE_INSTALL:
1452 				npgs = mem_node_memlist_pages(mnode,
1453 				    phys_install);
1454 				break;
1455 			default:
1456 				break;
1457 			}
1458 		}
1459 	}
1460 	return (npgs);
1461 }
1462 
1463 /*
1464  * Return latency between "from" and "to" lgroups
1465  *
1466  * This latency number can only be used for relative comparison
1467  * between lgroups on the running system, cannot be used across platforms,
1468  * and may not reflect the actual latency.  It is platform and implementation
1469  * specific, so platform gets to decide its value.  It would be nice if the
1470  * number was at least proportional to make comparisons more meaningful though.
1471  */
1472 /* ARGSUSED */
1473 int
1474 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1475 {
1476 	lgrp_handle_t	src, dest;
1477 
1478 	if (max_mem_nodes == 1)
1479 		return (0);
1480 
1481 	/*
1482 	 * Return max latency for root lgroup
1483 	 */
1484 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1485 		return (lgrp_plat_probe_time_max);
1486 
1487 	src = from;
1488 	dest = to;
1489 
1490 	/*
1491 	 * Return 0 for nodes (lgroup platform handles) out of range
1492 	 */
1493 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1494 		return (0);
1495 
1496 	/*
1497 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1498 	 * and we are trying to get latency from current CPU to some node
1499 	 */
1500 	if (lgrp_plat_probe_times[src][src] == 0 &&
1501 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1502 		lgrp_plat_probe();
1503 
1504 	return (lgrp_plat_probe_times[src][dest]);
1505 }
1506 
1507 /*
1508  * Return platform handle for root lgroup
1509  */
1510 lgrp_handle_t
1511 lgrp_plat_root_hand(void)
1512 {
1513 	return (LGRP_DEFAULT_HANDLE);
1514 }
1515