xref: /titanic_41/usr/src/uts/i86pc/os/lgrpplat.c (revision f841f6ad96ea6675d6c6b35c749eaac601799fdf)
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 /*
23  * Copyright 2006 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_cfgspace.h>
40 #include <sys/pci_impl.h>
41 #include <sys/param.h>
42 #include <sys/promif.h>		/* for prom_printf() */
43 #include <sys/systm.h>
44 #include <sys/thread.h>
45 #include <sys/types.h>
46 #include <sys/var.h>
47 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
48 #include <vm/hat_i86.h>
49 #include <vm/seg_kmem.h>
50 #include <vm/vm_dep.h>
51 
52 
53 
54 /*
55  * lgroup platform support for x86 platforms.
56  */
57 
58 #define	MAX_NODES		8
59 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
60 
61 #define	LGRP_PLAT_CPU_TO_NODE(cpu)	(chip_plat_get_chipid(cpu))
62 
63 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
64 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
65 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
66 
67 /*
68  * Multiprocessor Opteron machines have Non Uniform Memory Access (NUMA).
69  *
70  * Until System Affinity Resource Table (SRAT) becomes part of ACPI standard,
71  * we need to examine registers in PCI configuration space to determine how
72  * many nodes are in the system and which CPUs and memory are in each node.
73  * This could be determined by probing all memory from each CPU, but that is
74  * too expensive to do while booting the kernel.
75  *
76  * NOTE: Using these PCI configuration space registers to determine this
77  *       locality info is Opteron K8 specific and not guaranteed to work on
78  *       the next generation Opteron processor.  Furthermore, we assume that
79  *	 there is one CPU per node and CPU 0 is in node 0, CPU 1 is in node 1,
80  *	 etc. which should be true for Opteron K8....
81  */
82 
83 /*
84  * Opteron DRAM Address Map in PCI configuration space gives base and limit
85  * of physical memory in each node for Opteron K8.  The following constants
86  * and macros define their contents, structure, and access.
87  */
88 
89 /*
90  * How many bits to shift Opteron DRAM Address Map base and limit registers
91  * to get actual value
92  */
93 #define	OPT_DRAMADDR_LSHIFT_ADDR	8	/* shift left for address */
94 
95 #define	OPT_DRAMADDR_MASK_OFF	0xFFFFFF	/* offset for address */
96 
97 /*
98  * Bit masks defining what's in Opteron DRAM Address Map base register
99  */
100 #define	OPT_DRAMBASE_MASK_RE		0x1	/* read enable */
101 #define	OPT_DRAMBASE_MASK_WE		0x2	/* write enable */
102 #define	OPT_DRAMBASE_MASK_INTRLVEN	0x700	/* interleave */
103 
104 #define	OPT_DRAMBASE_MASK_ADDR	0xFFFF0000	/* address bits 39-24 */
105 
106 /*
107  * Macros to get values from Opteron DRAM Address Map base register
108  */
109 #define	OPT_DRAMBASE(reg) \
110 	(((u_longlong_t)reg & OPT_DRAMBASE_MASK_ADDR) << \
111 	    OPT_DRAMADDR_LSHIFT_ADDR)
112 
113 
114 /*
115  * Bit masks defining what's in Opteron DRAM Address Map limit register
116  */
117 #define	OPT_DRAMLIMIT_MASK_DSTNODE	0x7		/* destination node */
118 #define	OPT_DRAMLIMIT_MASK_INTRLVSEL	0x70		/* interleave select */
119 #define	OPT_DRAMLIMIT_MASK_ADDR		0xFFFF0000	/* addr bits 39-24 */
120 
121 /*
122  * Macros to get values from Opteron DRAM Address Map limit register
123  */
124 #define	OPT_DRAMLIMIT(reg) \
125 	(((u_longlong_t)reg & OPT_DRAMLIMIT_MASK_ADDR) << \
126 	    OPT_DRAMADDR_LSHIFT_ADDR)
127 
128 
129 /*
130  * Opteron Node ID register in PCI configuration space contains
131  * number of nodes in system, etc. for Opteron K8.  The following
132  * constants and macros define its contents, structure, and access.
133  */
134 
135 /*
136  * Bit masks defining what's in Opteron Node ID register
137  */
138 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
139 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
140 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
141 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
142 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
143 
144 /*
145  * How many bits in Opteron Node ID register to shift right to get actual value
146  */
147 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
148 
149 /*
150  * Macros to get values from Opteron Node ID register
151  */
152 #define	OPT_NODE_CNT(reg) \
153 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
154 
155 
156 /*
157  * PCI configuration space registers accessed by specifying
158  * a bus, device, function, and offset.  The following constants
159  * define the values needed to access Opteron K8 configuration
160  * info to determine its node topology
161  */
162 
163 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
164 
165 /*
166  * Opteron PCI configuration space register function values
167  */
168 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
169 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
170 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
171 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
172 
173 /*
174  * PCI Configuration Space register offsets
175  */
176 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
177 #define	OPT_PCS_OFF_DRAMBASE	0x40	/* DRAM Base register (node 0) */
178 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
179 
180 /*
181  * Opteron PCI Configuration Space device IDs for nodes
182  */
183 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
184 
185 
186 /*
187  * Bookkeeping for latencies seen during probing (used for verification)
188  */
189 typedef	struct lgrp_plat_latency_acct {
190 	hrtime_t	la_value;	/* latency value */
191 	int		la_count;	/* occurrences */
192 } lgrp_plat_latency_acct_t;
193 
194 
195 /*
196  * Choices for probing to determine lgroup topology
197  */
198 typedef	enum lgrp_plat_probe_op {
199 	LGRP_PLAT_PROBE_PGCPY,		/* Use page copy */
200 	LGRP_PLAT_PROBE_VENDOR		/* Read vendor ID on Northbridge */
201 } lgrp_plat_probe_op_t;
202 
203 
204 /*
205  * Opteron DRAM address map gives base and limit for physical memory in a node
206  */
207 typedef	struct opt_dram_addr_map {
208 	uint32_t	base;
209 	uint32_t	limit;
210 } opt_dram_addr_map_t;
211 
212 
213 /*
214  * Starting and ending page for physical memory in node
215  */
216 typedef	struct phys_addr_map {
217 	pfn_t	start;
218 	pfn_t	end;
219 	int	exists;
220 } phys_addr_map_t;
221 
222 
223 /*
224  * Opteron DRAM address map for each node
225  */
226 struct opt_dram_addr_map	opt_dram_map[MAX_NODES];
227 
228 /*
229  * Node ID register contents for each node
230  */
231 uint_t				opt_node_info[MAX_NODES];
232 
233 /*
234  * Whether memory is interleaved across nodes causing MPO to be disabled
235  */
236 int			lgrp_plat_mem_intrlv = 0;
237 
238 /*
239  * Number of nodes in system
240  */
241 uint_t			lgrp_plat_node_cnt = 1;
242 
243 /*
244  * Physical address range for memory in each node
245  */
246 phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
247 
248 /*
249  * Probe costs (individual and total) and flush cost
250  */
251 hrtime_t		lgrp_plat_flush_cost = 0;
252 hrtime_t		lgrp_plat_probe_cost = 0;
253 hrtime_t		lgrp_plat_probe_cost_total = 0;
254 
255 /*
256  * Error code for latency adjustment and verification
257  */
258 int			lgrp_plat_probe_error_code = 0;
259 
260 /*
261  * How much latencies were off from minimum values gotten
262  */
263 hrtime_t		lgrp_plat_probe_errors[MAX_NODES][MAX_NODES];
264 
265 /*
266  * Unique probe latencies and number of occurrences of each
267  */
268 lgrp_plat_latency_acct_t	lgrp_plat_probe_lat_acct[MAX_NODES];
269 
270 /*
271  * Size of memory buffer in each node for probing
272  */
273 size_t			lgrp_plat_probe_memsize = 0;
274 
275 /*
276  * Virtual address of page in each node for probing
277  */
278 caddr_t			lgrp_plat_probe_memory[MAX_NODES];
279 
280 /*
281  * Number of unique latencies in probe times
282  */
283 int			lgrp_plat_probe_nlatencies = 0;
284 
285 /*
286  * How many rounds of probing to do
287  */
288 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
289 
290 /*
291  * Number of samples to take when probing each node
292  */
293 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
294 
295 /*
296  * Number of times to read vendor ID from Northbridge for each probe.
297  */
298 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
299 
300 /*
301  * How to probe to determine lgroup topology
302  */
303 lgrp_plat_probe_op_t	lgrp_plat_probe_op = LGRP_PLAT_PROBE_VENDOR;
304 
305 /*
306  * PFN of page in each node for probing
307  */
308 pfn_t			lgrp_plat_probe_pfn[MAX_NODES];
309 
310 /*
311  * Whether probe time was suspect (ie. not within tolerance of value that it
312  * should match)
313  */
314 int			lgrp_plat_probe_suspect[MAX_NODES][MAX_NODES];
315 
316 /*
317  * How long it takes to access memory from each node
318  */
319 hrtime_t		lgrp_plat_probe_times[MAX_NODES][MAX_NODES];
320 
321 /*
322  * Min and max node memory probe times seen
323  */
324 hrtime_t		lgrp_plat_probe_time_max = 0;
325 hrtime_t		lgrp_plat_probe_time_min = -1;
326 hrtime_t		lgrp_plat_probe_max[MAX_NODES][MAX_NODES];
327 hrtime_t		lgrp_plat_probe_min[MAX_NODES][MAX_NODES];
328 
329 
330 /*
331  * Allocate lgrp and lgrp stat arrays statically.
332  */
333 static lgrp_t	lgrp_space[NLGRP];
334 static int	nlgrps_alloc;
335 
336 struct lgrp_stats lgrp_stats[NLGRP];
337 
338 #define	CPUID_FAMILY_OPTERON	15
339 
340 uint_t	opt_family = 0;
341 uint_t	opt_model = 0;
342 uint_t	opt_probe_func = OPT_PCS_FUNC_DRAM;
343 
344 
345 /*
346  * Determine whether we're running on an AMD Opteron K8 machine
347  */
348 int
349 is_opteron(void)
350 {
351 	if (x86_vendor != X86_VENDOR_AMD)
352 		return (0);
353 
354 	if (cpuid_getfamily(CPU) == CPUID_FAMILY_OPTERON)
355 		return (1);
356 	else
357 		return (0);
358 }
359 
360 int
361 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
362 {
363 	if (max_mem_nodes == 1)
364 		return (0);
365 
366 	return ((int)hand);
367 }
368 
369 lgrp_handle_t
370 plat_mem_node_to_lgrphand(int mnode)
371 {
372 	if (max_mem_nodes == 1)
373 		return (LGRP_DEFAULT_HANDLE);
374 
375 	return ((lgrp_handle_t)mnode);
376 }
377 
378 int
379 plat_pfn_to_mem_node(pfn_t pfn)
380 {
381 	int	node;
382 
383 	if (max_mem_nodes == 1)
384 		return (0);
385 
386 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
387 		/*
388 		 * Skip nodes with no memory
389 		 */
390 		if (!lgrp_plat_node_memory[node].exists)
391 			continue;
392 
393 		if (pfn >= lgrp_plat_node_memory[node].start &&
394 		    pfn <= lgrp_plat_node_memory[node].end)
395 			return (node);
396 	}
397 
398 	ASSERT(node < lgrp_plat_node_cnt);
399 	return (-1);
400 }
401 
402 /*
403  * Configure memory nodes for machines with more than one node (ie NUMA)
404  */
405 void
406 plat_build_mem_nodes(struct memlist *list)
407 {
408 	pfn_t		cur_start;	/* start addr of subrange */
409 	pfn_t		cur_end;	/* end addr of subrange */
410 	pfn_t		start;		/* start addr of whole range */
411 	pfn_t		end;		/* end addr of whole range */
412 
413 	/*
414 	 * Boot install lists are arranged <addr, len>, ...
415 	 */
416 	while (list) {
417 		int	node;
418 
419 		start = list->address >> PAGESHIFT;
420 		end = (list->address + list->size - 1) >> PAGESHIFT;
421 
422 		if (start > physmax) {
423 			list = list->next;
424 			continue;
425 		}
426 		if (end > physmax)
427 			end = physmax;
428 
429 		/*
430 		 * When there is only one memnode, just add memory to memnode
431 		 */
432 		if (max_mem_nodes == 1) {
433 			mem_node_add_slice(start, end);
434 			list = list->next;
435 			continue;
436 		}
437 
438 		/*
439 		 * mem_node_add_slice() expects to get a memory range that
440 		 * is within one memnode, so need to split any memory range
441 		 * that spans multiple memnodes into subranges that are each
442 		 * contained within one memnode when feeding them to
443 		 * mem_node_add_slice()
444 		 */
445 		cur_start = start;
446 		do {
447 			node = plat_pfn_to_mem_node(cur_start);
448 
449 			/*
450 			 * Panic if DRAM address map registers or SRAT say
451 			 * memory in node doesn't exist or address from
452 			 * boot installed memory list entry isn't in this node.
453 			 * This shouldn't happen and rest of code can't deal
454 			 * with this if it does.
455 			 */
456 			if (node < 0 || node >= lgrp_plat_node_cnt ||
457 			    !lgrp_plat_node_memory[node].exists ||
458 			    cur_start < lgrp_plat_node_memory[node].start ||
459 			    cur_start > lgrp_plat_node_memory[node].end) {
460 				cmn_err(CE_PANIC, "Don't know which memnode "
461 				    "to add installed memory address 0x%lx\n",
462 				    cur_start);
463 			}
464 
465 			/*
466 			 * End of current subrange should not span memnodes
467 			 */
468 			cur_end = end;
469 			if (lgrp_plat_node_memory[node].exists &&
470 			    cur_end > lgrp_plat_node_memory[node].end)
471 				cur_end = lgrp_plat_node_memory[node].end;
472 
473 			mem_node_add_slice(cur_start, cur_end);
474 
475 			/*
476 			 * Next subrange starts after end of current one
477 			 */
478 			cur_start = cur_end + 1;
479 		} while (cur_end < end);
480 
481 		list = list->next;
482 	}
483 	mem_node_physalign = 0;
484 	mem_node_pfn_shift = 0;
485 }
486 
487 
488 /*
489  * Platform-specific initialization of lgroups
490  */
491 void
492 lgrp_plat_init(void)
493 {
494 	uint_t		bus;
495 	uint_t		dev;
496 	uint_t		node;
497 	uint_t		off;
498 
499 	extern lgrp_load_t	lgrp_expand_proc_thresh;
500 	extern lgrp_load_t	lgrp_expand_proc_diff;
501 
502 	/*
503 	 * Initialize as a UMA machine if this isn't an Opteron
504 	 */
505 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
506 		lgrp_plat_node_cnt = max_mem_nodes = 1;
507 		return;
508 	}
509 
510 	/*
511 	 * Read configuration registers from PCI configuration space to
512 	 * determine node information, which memory is in each node, etc.
513 	 *
514 	 * Write to PCI configuration space address register to specify
515 	 * which configuration register to read and read/write PCI
516 	 * configuration space data register to get/set contents
517 	 */
518 	bus = OPT_PCS_BUS_CONFIG;
519 	dev = OPT_PCS_DEV_NODE0;
520 	off = OPT_PCS_OFF_DRAMBASE;
521 
522 	/*
523 	 * Read node ID register for node 0 to get node count
524 	 */
525 	opt_node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
526 	    OPT_PCS_OFF_NODEID);
527 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
528 
529 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
530 		/*
531 		 * Read node ID register (except for node 0 which we just read)
532 		 */
533 		if (node > 0) {
534 			opt_node_info[node] = pci_getl_func(bus, dev,
535 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
536 		}
537 
538 		/*
539 		 * Read DRAM base and limit registers which specify
540 		 * physical memory range of each node
541 		 */
542 		opt_dram_map[node].base = pci_getl_func(bus, dev,
543 		    OPT_PCS_FUNC_ADDRMAP, off);
544 		if (opt_dram_map[node].base & OPT_DRAMBASE_MASK_INTRLVEN)
545 			lgrp_plat_mem_intrlv++;
546 
547 		off += 4;	/* limit register offset */
548 		opt_dram_map[node].limit = pci_getl_func(bus, dev,
549 		    OPT_PCS_FUNC_ADDRMAP, off);
550 
551 		/*
552 		 * Increment device number to next node and register offset for
553 		 * DRAM base register of next node
554 		 */
555 		off += 4;
556 		dev++;
557 
558 		/*
559 		 * Both read and write enable bits must be enabled in DRAM
560 		 * address map base register for physical memory to exist in
561 		 * node
562 		 */
563 		if ((opt_dram_map[node].base & OPT_DRAMBASE_MASK_RE) == 0 ||
564 		    (opt_dram_map[node].base & OPT_DRAMBASE_MASK_WE) == 0) {
565 			/*
566 			 * Mark node memory as non-existent and set start and
567 			 * end addresses to be same in lgrp_plat_node_memory[]
568 			 */
569 			lgrp_plat_node_memory[node].exists = 0;
570 			lgrp_plat_node_memory[node].start =
571 			    lgrp_plat_node_memory[node].end = (pfn_t)-1;
572 			continue;
573 		}
574 
575 		/*
576 		 * Get PFN for first page in each node,
577 		 * so we can probe memory to determine latency topology
578 		 */
579 		lgrp_plat_probe_pfn[node] =
580 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
581 
582 		/*
583 		 * Mark node memory as existing and remember physical address
584 		 * range of each node for use later
585 		 */
586 		lgrp_plat_node_memory[node].exists = 1;
587 		lgrp_plat_node_memory[node].start =
588 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
589 		lgrp_plat_node_memory[node].end =
590 		    btop(OPT_DRAMLIMIT(opt_dram_map[node].limit) |
591 		    OPT_DRAMADDR_MASK_OFF);
592 	}
593 
594 	/*
595 	 * Only use one memory node if memory is interleaved between any nodes
596 	 */
597 	if (lgrp_plat_mem_intrlv) {
598 		lgrp_plat_node_cnt = max_mem_nodes = 1;
599 		(void) lgrp_topo_ht_limit_set(1);
600 	} else {
601 		max_mem_nodes = lgrp_plat_node_cnt;
602 
603 		/*
604 		 * Probing errors can mess up the lgroup topology and force us
605 		 * fall back to a 2 level lgroup topology.  Here we bound how
606 		 * tall the lgroup topology can grow in hopes of avoiding any
607 		 * anamolies in probing from messing up the lgroup topology
608 		 * by limiting the accuracy of the latency topology.
609 		 *
610 		 * Assume that nodes will at least be configured in a ring,
611 		 * so limit height of lgroup topology to be less than number
612 		 * of nodes on a system with 4 or more nodes
613 		 */
614 		if (lgrp_plat_node_cnt >= 4 &&
615 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
616 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
617 	}
618 
619 	/*
620 	 * Lgroups on Opteron architectures have but a single physical
621 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
622 	 * so that lgrp_choose() will spread things out aggressively.
623 	 */
624 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
625 	lgrp_expand_proc_diff = 0;
626 }
627 
628 
629 /*
630  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
631  * be considered same
632  */
633 #define	LGRP_LAT_TOLERANCE_SHIFT	4
634 
635 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
636 
637 
638 /*
639  * Adjust latencies between nodes to be symmetric, normalize latencies between
640  * any nodes that are within some tolerance to be same, and make local
641  * latencies be same
642  */
643 static void
644 lgrp_plat_latency_adjust(void)
645 {
646 	int				i;
647 	int				j;
648 	int				k;
649 	int				l;
650 	u_longlong_t			max;
651 	u_longlong_t			min;
652 	u_longlong_t			t;
653 	u_longlong_t			t1;
654 	u_longlong_t			t2;
655 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LATENCY_CHANGE;
656 	int				lat_corrected[MAX_NODES][MAX_NODES];
657 
658 	/*
659 	 * Nothing to do when this is an UMA machine
660 	 */
661 	if (max_mem_nodes == 1)
662 		return;
663 
664 	/*
665 	 * Make sure that latencies are symmetric between any two nodes
666 	 * (ie. latency(node0, node1) == latency(node1, node0))
667 	 */
668 	for (i = 0; i < lgrp_plat_node_cnt; i++)
669 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
670 			t1 = lgrp_plat_probe_times[i][j];
671 			t2 = lgrp_plat_probe_times[j][i];
672 
673 			if (t1 == 0 || t2 == 0 || t1 == t2)
674 				continue;
675 
676 			/*
677 			 * Latencies should be same
678 			 * - Use minimum of two latencies which should be same
679 			 * - Track suspect probe times not within tolerance of
680 			 *   min value
681 			 * - Remember how much values are corrected by
682 			 */
683 			if (t1 > t2) {
684 				t = t2;
685 				lgrp_plat_probe_errors[i][j] += t1 - t2;
686 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
687 					lgrp_plat_probe_suspect[i][j]++;
688 					lgrp_plat_probe_suspect[j][i]++;
689 				}
690 			} else if (t2 > t1) {
691 				t = t1;
692 				lgrp_plat_probe_errors[j][i] += t2 - t1;
693 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
694 					lgrp_plat_probe_suspect[i][j]++;
695 					lgrp_plat_probe_suspect[j][i]++;
696 				}
697 			}
698 
699 			lgrp_plat_probe_times[i][j] =
700 			    lgrp_plat_probe_times[j][i] = t;
701 			lgrp_config(cflag, t1, t);
702 			lgrp_config(cflag, t2, t);
703 		}
704 
705 	/*
706 	 * Keep track of which latencies get corrected
707 	 */
708 	for (i = 0; i < MAX_NODES; i++)
709 		for (j = 0; j < MAX_NODES; j++)
710 			lat_corrected[i][j] = 0;
711 
712 	/*
713 	 * For every two nodes, see whether there is another pair of nodes which
714 	 * are about the same distance apart and make the latencies be the same
715 	 * if they are close enough together
716 	 */
717 	for (i = 0; i < lgrp_plat_node_cnt; i++)
718 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
719 			/*
720 			 * Pick one pair of nodes (i, j)
721 			 * and get latency between them
722 			 */
723 			t1 = lgrp_plat_probe_times[i][j];
724 
725 			/*
726 			 * Skip this pair of nodes if there isn't a latency
727 			 * for it yet
728 			 */
729 			if (t1 == 0)
730 				continue;
731 
732 			for (k = 0; k < lgrp_plat_node_cnt; k++)
733 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
734 					/*
735 					 * Pick another pair of nodes (k, l)
736 					 * not same as (i, j) and get latency
737 					 * between them
738 					 */
739 					if (k == i && l == j)
740 						continue;
741 
742 					t2 = lgrp_plat_probe_times[k][l];
743 
744 					/*
745 					 * Skip this pair of nodes if there
746 					 * isn't a latency for it yet
747 					 */
748 
749 					if (t2 == 0)
750 						continue;
751 
752 					/*
753 					 * Skip nodes (k, l) if they already
754 					 * have same latency as (i, j) or
755 					 * their latency isn't close enough to
756 					 * be considered/made the same
757 					 */
758 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
759 					    t1 >> lgrp_plat_probe_lt_shift) ||
760 					    (t2 > t1 && t2 - t1 >
761 					    t2 >> lgrp_plat_probe_lt_shift))
762 						continue;
763 
764 					/*
765 					 * Make latency(i, j) same as
766 					 * latency(k, l), try to use latency
767 					 * that has been adjusted already to get
768 					 * more consistency (if possible), and
769 					 * remember which latencies were
770 					 * adjusted for next time
771 					 */
772 					if (lat_corrected[i][j]) {
773 						t = t1;
774 						lgrp_config(cflag, t2, t);
775 						t2 = t;
776 					} else if (lat_corrected[k][l]) {
777 						t = t2;
778 						lgrp_config(cflag, t1, t);
779 						t1 = t;
780 					} else {
781 						if (t1 > t2)
782 							t = t2;
783 						else
784 							t = t1;
785 						lgrp_config(cflag, t1, t);
786 						lgrp_config(cflag, t2, t);
787 						t1 = t2 = t;
788 					}
789 
790 					lgrp_plat_probe_times[i][j] =
791 					    lgrp_plat_probe_times[k][l] = t;
792 
793 					lat_corrected[i][j] =
794 					    lat_corrected[k][l] = 1;
795 				}
796 		}
797 
798 	/*
799 	 * Local latencies should be same
800 	 * - Find min and max local latencies
801 	 * - Make all local latencies be minimum
802 	 */
803 	min = -1;
804 	max = 0;
805 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
806 		t = lgrp_plat_probe_times[i][i];
807 		if (t == 0)
808 			continue;
809 		if (min == -1 || t < min)
810 			min = t;
811 		if (t > max)
812 			max = t;
813 	}
814 	if (min != max) {
815 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
816 			int	local;
817 
818 			local = lgrp_plat_probe_times[i][i];
819 			if (local == 0)
820 				continue;
821 
822 			/*
823 			 * Track suspect probe times that aren't within
824 			 * tolerance of minimum local latency and how much
825 			 * probe times are corrected by
826 			 */
827 			if (local - min > min >> lgrp_plat_probe_lt_shift)
828 				lgrp_plat_probe_suspect[i][i]++;
829 
830 			lgrp_plat_probe_errors[i][i] += local - min;
831 
832 			/*
833 			 * Make local latencies be minimum
834 			 */
835 			lgrp_config(cflag, local, min);
836 			lgrp_plat_probe_times[i][i] = min;
837 		}
838 	}
839 
840 	/*
841 	 * Determine max probe time again since just adjusted latencies
842 	 */
843 	lgrp_plat_probe_time_max = 0;
844 	for (i = 0; i < lgrp_plat_node_cnt; i++)
845 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
846 			t = lgrp_plat_probe_times[i][j];
847 			if (t > lgrp_plat_probe_time_max)
848 				lgrp_plat_probe_time_max = t;
849 		}
850 }
851 
852 
853 /*
854  * Verify following about latencies between nodes:
855  *
856  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
857  * - Local latencies same
858  * - Local < remote
859  * - Number of latencies seen is reasonable
860  * - Number of occurrences of a given latency should be more than 1
861  *
862  * Returns:
863  *	0	Success
864  *	-1	Not symmetric
865  *	-2	Local latencies not same
866  *	-3	Local >= remote
867  *	-4	Wrong number of latencies
868  *	-5	Not enough occurrences of given latency
869  */
870 static int
871 lgrp_plat_latency_verify(void)
872 {
873 	int				i;
874 	int				j;
875 	lgrp_plat_latency_acct_t	*l;
876 	int				probed;
877 	u_longlong_t			t1;
878 	u_longlong_t			t2;
879 
880 	/*
881 	 * Nothing to do when this is an UMA machine, lgroup topology is
882 	 * limited to 2 levels, or there aren't any probe times yet
883 	 */
884 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
885 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
886 		return (0);
887 
888 	/*
889 	 * Make sure that latencies are symmetric between any two nodes
890 	 * (ie. latency(node0, node1) == latency(node1, node0))
891 	 */
892 	for (i = 0; i < lgrp_plat_node_cnt; i++)
893 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
894 			t1 = lgrp_plat_probe_times[i][j];
895 			t2 = lgrp_plat_probe_times[j][i];
896 
897 			if (t1 == 0 || t2 == 0 || t1 == t2)
898 				continue;
899 
900 			return (-1);
901 		}
902 
903 	/*
904 	 * Local latencies should be same
905 	 */
906 	t1 = lgrp_plat_probe_times[0][0];
907 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
908 		t2 = lgrp_plat_probe_times[i][i];
909 		if (t2 == 0)
910 			continue;
911 
912 		if (t1 == 0) {
913 			t1 = t2;
914 			continue;
915 		}
916 
917 		if (t1 != t2)
918 			return (-2);
919 	}
920 
921 	/*
922 	 * Local latencies should be less than remote
923 	 */
924 	if (t1) {
925 		for (i = 0; i < lgrp_plat_node_cnt; i++)
926 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
927 				t2 = lgrp_plat_probe_times[i][j];
928 				if (i == j || t2 == 0)
929 					continue;
930 
931 				if (t1 >= t2)
932 					return (-3);
933 			}
934 	}
935 
936 	/*
937 	 * Rest of checks are not very useful for machines with less than
938 	 * 4 nodes (which means less than 3 latencies on Opteron)
939 	 */
940 	if (lgrp_plat_node_cnt < 4)
941 		return (0);
942 
943 	/*
944 	 * Need to see whether done probing in order to verify number of
945 	 * latencies are correct
946 	 */
947 	probed = 0;
948 	for (i = 0; i < lgrp_plat_node_cnt; i++)
949 		if (lgrp_plat_probe_times[i][i])
950 			probed++;
951 
952 	if (probed != lgrp_plat_node_cnt)
953 		return (0);
954 
955 	/*
956 	 * Determine number of unique latencies seen in probe times,
957 	 * their values, and number of occurrences of each
958 	 */
959 	lgrp_plat_probe_nlatencies = 0;
960 	bzero(lgrp_plat_probe_lat_acct,
961 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
962 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
963 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
964 			int	k;
965 
966 			/*
967 			 * Look at each probe time
968 			 */
969 			t1 = lgrp_plat_probe_times[i][j];
970 			if (t1 == 0)
971 				continue;
972 
973 			/*
974 			 * Account for unique latencies
975 			 */
976 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
977 				l = &lgrp_plat_probe_lat_acct[k];
978 				if (t1 == l->la_value) {
979 					/*
980 					 * Increment number of occurrences
981 					 * if seen before
982 					 */
983 					l->la_count++;
984 					break;
985 				} else if (l->la_value == 0) {
986 					/*
987 					 * Record latency if haven't seen before
988 					 */
989 					l->la_value = t1;
990 					l->la_count++;
991 					lgrp_plat_probe_nlatencies++;
992 					break;
993 				}
994 			}
995 		}
996 	}
997 
998 	/*
999 	 * Number of latencies should be relative to number of
1000 	 * nodes in system:
1001 	 * - Same as nodes when nodes <= 2
1002 	 * - Less than nodes when nodes > 2
1003 	 * - Greater than 2 when nodes >= 4
1004 	 */
1005 	if ((lgrp_plat_node_cnt <= 2 &&
1006 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
1007 	    (lgrp_plat_node_cnt > 2 &&
1008 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
1009 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
1010 	    lgrp_plat_probe_nlatencies <= 2))
1011 		return (-4);
1012 
1013 	/*
1014 	 * There should be more than one occurrence of every latency
1015 	 * as long as probing is complete
1016 	 */
1017 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
1018 		l = &lgrp_plat_probe_lat_acct[i];
1019 		if (l->la_count <= 1)
1020 			return (-5);
1021 	}
1022 	return (0);
1023 }
1024 
1025 
1026 /*
1027  * Set lgroup latencies for 2 level lgroup topology
1028  */
1029 static void
1030 lgrp_plat_2level_setup(void)
1031 {
1032 	int	i;
1033 
1034 	if (lgrp_plat_node_cnt >= 4)
1035 		cmn_err(CE_NOTE,
1036 		    "MPO only optimizing for local and remote\n");
1037 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1038 		int	j;
1039 
1040 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1041 			if (i == j)
1042 				lgrp_plat_probe_times[i][j] = 2;
1043 			else
1044 				lgrp_plat_probe_times[i][j] = 3;
1045 		}
1046 	}
1047 	lgrp_plat_probe_time_min = 2;
1048 	lgrp_plat_probe_time_max = 3;
1049 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1050 }
1051 
1052 
1053 /*
1054  * Return time needed to probe from current CPU to memory in given node
1055  */
1056 static hrtime_t
1057 lgrp_plat_probe_time(int to)
1058 {
1059 	caddr_t		buf;
1060 	uint_t		dev;
1061 	/* LINTED: set but not used in function */
1062 	volatile uint_t	dev_vendor;
1063 	hrtime_t	elapsed;
1064 	hrtime_t	end;
1065 	int		from;
1066 	int		i;
1067 	int		ipl;
1068 	hrtime_t	max;
1069 	hrtime_t	min;
1070 	hrtime_t	start;
1071 	int		cnt;
1072 	extern int	use_sse_pagecopy;
1073 
1074 	/*
1075 	 * Determine ID of node containing current CPU
1076 	 */
1077 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1078 
1079 	/*
1080 	 * Do common work for probing main memory
1081 	 */
1082 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1083 		/*
1084 		 * Skip probing any nodes without memory and
1085 		 * set probe time to 0
1086 		 */
1087 		if (lgrp_plat_probe_memory[to] == NULL) {
1088 			lgrp_plat_probe_times[from][to] = 0;
1089 			return (0);
1090 		}
1091 
1092 		/*
1093 		 * Invalidate caches once instead of once every sample
1094 		 * which should cut cost of probing by a lot
1095 		 */
1096 		lgrp_plat_flush_cost = gethrtime();
1097 		invalidate_cache();
1098 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1099 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1100 	}
1101 
1102 	/*
1103 	 * Probe from current CPU to given memory using specified operation
1104 	 * and take specified number of samples
1105 	 */
1106 	max = 0;
1107 	min = -1;
1108 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1109 		lgrp_plat_probe_cost = gethrtime();
1110 
1111 		/*
1112 		 * Can't measure probe time if gethrtime() isn't working yet
1113 		 */
1114 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1115 			return (0);
1116 
1117 		switch (lgrp_plat_probe_op) {
1118 
1119 		case LGRP_PLAT_PROBE_PGCPY:
1120 		default:
1121 			/*
1122 			 * Measure how long it takes to copy page
1123 			 * on top of itself
1124 			 */
1125 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1126 
1127 			kpreempt_disable();
1128 			ipl = splhigh();
1129 			start = gethrtime();
1130 			if (use_sse_pagecopy)
1131 				hwblkpagecopy(buf, buf);
1132 			else
1133 				bcopy(buf, buf, PAGESIZE);
1134 			end = gethrtime();
1135 			elapsed = end - start;
1136 			splx(ipl);
1137 			kpreempt_enable();
1138 			break;
1139 
1140 		case LGRP_PLAT_PROBE_VENDOR:
1141 			/*
1142 			 * Measure how long it takes to read vendor ID from
1143 			 * Northbridge
1144 			 */
1145 			dev = OPT_PCS_DEV_NODE0 + to;
1146 			kpreempt_disable();
1147 			ipl = spl8();
1148 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1149 			    OPT_PCS_OFF_VENDOR));
1150 			start = gethrtime();
1151 			for (cnt = 0; cnt < lgrp_plat_probe_nreads; cnt++)
1152 				dev_vendor = inl(PCI_CONFDATA);
1153 			end = gethrtime();
1154 			elapsed = (end - start) / lgrp_plat_probe_nreads;
1155 			splx(ipl);
1156 			kpreempt_enable();
1157 			break;
1158 		}
1159 
1160 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1161 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1162 
1163 		if (min == -1 || elapsed < min)
1164 			min = elapsed;
1165 		if (elapsed > max)
1166 			max = elapsed;
1167 	}
1168 
1169 	/*
1170 	 * Update minimum and maximum probe times between
1171 	 * these two nodes
1172 	 */
1173 	if (min < lgrp_plat_probe_min[from][to] ||
1174 	    lgrp_plat_probe_min[from][to] == 0)
1175 		lgrp_plat_probe_min[from][to] = min;
1176 
1177 	if (max > lgrp_plat_probe_max[from][to])
1178 		lgrp_plat_probe_max[from][to] = max;
1179 
1180 	return (min);
1181 }
1182 
1183 
1184 /*
1185  * Probe memory in each node from current CPU to determine latency topology
1186  */
1187 void
1188 lgrp_plat_probe(void)
1189 {
1190 	int		from;
1191 	int		i;
1192 	hrtime_t	probe_time;
1193 	int		to;
1194 
1195 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1196 		return;
1197 
1198 	/*
1199 	 * Determine ID of node containing current CPU
1200 	 */
1201 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1202 
1203 	/*
1204 	 * Don't need to probe if got times already
1205 	 */
1206 	if (lgrp_plat_probe_times[from][from] != 0)
1207 		return;
1208 
1209 	/*
1210 	 * Read vendor ID in Northbridge or read and write page(s)
1211 	 * in each node from current CPU and remember how long it takes,
1212 	 * so we can build latency topology of machine later.
1213 	 * This should approximate the memory latency between each node.
1214 	 */
1215 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1216 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1217 			/*
1218 			 * Get probe time and bail out if can't get it yet
1219 			 */
1220 			probe_time = lgrp_plat_probe_time(to);
1221 			if (probe_time == 0)
1222 				return;
1223 
1224 			/*
1225 			 * Keep lowest probe time as latency between nodes
1226 			 */
1227 			if (lgrp_plat_probe_times[from][to] == 0 ||
1228 			    probe_time < lgrp_plat_probe_times[from][to])
1229 				lgrp_plat_probe_times[from][to] = probe_time;
1230 
1231 			/*
1232 			 * Update overall minimum and maximum probe times
1233 			 * across all nodes
1234 			 */
1235 			if (probe_time < lgrp_plat_probe_time_min ||
1236 			    lgrp_plat_probe_time_min == -1)
1237 				lgrp_plat_probe_time_min = probe_time;
1238 			if (probe_time > lgrp_plat_probe_time_max)
1239 				lgrp_plat_probe_time_max = probe_time;
1240 		}
1241 
1242 	/*
1243 	 * - Fix up latencies such that local latencies are same,
1244 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1245 	 *
1246 	 * - Verify that latencies look ok
1247 	 *
1248 	 * - Fallback to just optimizing for local and remote if
1249 	 *   latencies didn't look right
1250 	 */
1251 	lgrp_plat_latency_adjust();
1252 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1253 	if (lgrp_plat_probe_error_code)
1254 		lgrp_plat_2level_setup();
1255 }
1256 
1257 
1258 /*
1259  * Platform-specific initialization
1260  */
1261 void
1262 lgrp_plat_main_init(void)
1263 {
1264 	int	curnode;
1265 	int	ht_limit;
1266 	int	i;
1267 
1268 	/*
1269 	 * Print a notice that MPO is disabled when memory is interleaved
1270 	 * across nodes....Would do this when it is discovered, but can't
1271 	 * because it happens way too early during boot....
1272 	 */
1273 	if (lgrp_plat_mem_intrlv)
1274 		cmn_err(CE_NOTE,
1275 		    "MPO disabled because memory is interleaved\n");
1276 
1277 	/*
1278 	 * Don't bother to do any probing if there is only one node or the
1279 	 * height of the lgroup topology less than or equal to 2
1280 	 */
1281 	ht_limit = lgrp_topo_ht_limit();
1282 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1283 		/*
1284 		 * Setup lgroup latencies for 2 level lgroup topology
1285 		 * (ie. local and remote only) if they haven't been set yet
1286 		 */
1287 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1288 		    lgrp_plat_probe_time_max == 0)
1289 			lgrp_plat_2level_setup();
1290 		return;
1291 	}
1292 
1293 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1294 		/*
1295 		 * Should have been able to probe from CPU 0 when it was added
1296 		 * to lgroup hierarchy, but may not have been able to then
1297 		 * because it happens so early in boot that gethrtime() hasn't
1298 		 * been initialized.  (:-(
1299 		 */
1300 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1301 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1302 			lgrp_plat_probe();
1303 
1304 		return;
1305 	}
1306 
1307 	/*
1308 	 * When probing memory, use one page for every sample to determine
1309 	 * lgroup topology and taking multiple samples
1310 	 */
1311 	if (lgrp_plat_probe_memsize == 0)
1312 		lgrp_plat_probe_memsize = PAGESIZE *
1313 		    lgrp_plat_probe_nsamples;
1314 
1315 	/*
1316 	 * Map memory in each node needed for probing to determine latency
1317 	 * topology
1318 	 */
1319 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1320 		int	mnode;
1321 
1322 		/*
1323 		 * Skip this node and leave its probe page NULL
1324 		 * if it doesn't have any memory
1325 		 */
1326 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1327 		if (!mem_node_config[mnode].exists) {
1328 			lgrp_plat_probe_memory[i] = NULL;
1329 			continue;
1330 		}
1331 
1332 		/*
1333 		 * Allocate one kernel virtual page
1334 		 */
1335 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1336 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1337 		if (lgrp_plat_probe_memory[i] == NULL) {
1338 			cmn_err(CE_WARN,
1339 			    "lgrp_plat_main_init: couldn't allocate memory");
1340 			return;
1341 		}
1342 
1343 		/*
1344 		 * Map virtual page to first page in node
1345 		 */
1346 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1347 		    lgrp_plat_probe_memsize,
1348 		    lgrp_plat_probe_pfn[i],
1349 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1350 		    HAT_LOAD_NOCONSIST);
1351 	}
1352 
1353 	/*
1354 	 * Probe from current CPU
1355 	 */
1356 	lgrp_plat_probe();
1357 }
1358 
1359 /*
1360  * Allocate additional space for an lgroup.
1361  */
1362 /* ARGSUSED */
1363 lgrp_t *
1364 lgrp_plat_alloc(lgrp_id_t lgrpid)
1365 {
1366 	lgrp_t *lgrp;
1367 
1368 	lgrp = &lgrp_space[nlgrps_alloc++];
1369 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1370 		return (NULL);
1371 	return (lgrp);
1372 }
1373 
1374 /*
1375  * Platform handling for (re)configuration changes
1376  */
1377 /* ARGSUSED */
1378 void
1379 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1380 {
1381 }
1382 
1383 /*
1384  * Return the platform handle for the lgroup containing the given CPU
1385  */
1386 /* ARGSUSED */
1387 lgrp_handle_t
1388 lgrp_plat_cpu_to_hand(processorid_t id)
1389 {
1390 	if (lgrp_plat_node_cnt == 1)
1391 		return (LGRP_DEFAULT_HANDLE);
1392 
1393 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1394 }
1395 
1396 /*
1397  * Return the platform handle of the lgroup that contains the physical memory
1398  * corresponding to the given page frame number
1399  */
1400 /* ARGSUSED */
1401 lgrp_handle_t
1402 lgrp_plat_pfn_to_hand(pfn_t pfn)
1403 {
1404 	int	mnode;
1405 
1406 	if (max_mem_nodes == 1)
1407 		return (LGRP_DEFAULT_HANDLE);
1408 
1409 	if (pfn > physmax)
1410 		return (LGRP_NULL_HANDLE);
1411 
1412 	mnode = plat_pfn_to_mem_node(pfn);
1413 	if (mnode < 0)
1414 		return (LGRP_NULL_HANDLE);
1415 
1416 	return (MEM_NODE_2_LGRPHAND(mnode));
1417 }
1418 
1419 /*
1420  * Return the maximum number of lgrps supported by the platform.
1421  * Before lgrp topology is known it returns an estimate based on the number of
1422  * nodes. Once topology is known it returns the actual maximim number of lgrps
1423  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1424  * may not grow during system lifetime.
1425  */
1426 int
1427 lgrp_plat_max_lgrps()
1428 {
1429 	return (lgrp_topo_initialized ?
1430 	    lgrp_alloc_max + 1 :
1431 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1432 }
1433 
1434 /*
1435  * Return the number of free, allocatable, or installed
1436  * pages in an lgroup
1437  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1438  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1439  */
1440 /* ARGSUSED */
1441 static pgcnt_t
1442 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1443 {
1444 	struct memlist *mlist;
1445 	pgcnt_t npgs = 0;
1446 	extern struct memlist *phys_avail;
1447 	extern struct memlist *phys_install;
1448 
1449 	switch (query) {
1450 	case LGRP_MEM_SIZE_FREE:
1451 		return ((pgcnt_t)freemem);
1452 	case LGRP_MEM_SIZE_AVAIL:
1453 		memlist_read_lock();
1454 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1455 			npgs += btop(mlist->size);
1456 		memlist_read_unlock();
1457 		return (npgs);
1458 	case LGRP_MEM_SIZE_INSTALL:
1459 		memlist_read_lock();
1460 		for (mlist = phys_install; mlist; mlist = mlist->next)
1461 			npgs += btop(mlist->size);
1462 		memlist_read_unlock();
1463 		return (npgs);
1464 	default:
1465 		return ((pgcnt_t)0);
1466 	}
1467 }
1468 
1469 /*
1470  * Return the number of free pages in an lgroup.
1471  *
1472  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1473  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1474  * number of allocatable base pagesize pages corresponding to the
1475  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1476  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1477  * memory installed, regardless of whether or not it's usable.
1478  */
1479 pgcnt_t
1480 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1481 {
1482 	int	mnode;
1483 	pgcnt_t npgs = (pgcnt_t)0;
1484 	extern struct memlist *phys_avail;
1485 	extern struct memlist *phys_install;
1486 
1487 
1488 	if (plathand == LGRP_DEFAULT_HANDLE)
1489 		return (lgrp_plat_mem_size_default(plathand, query));
1490 
1491 	if (plathand != LGRP_NULL_HANDLE) {
1492 		mnode = plat_lgrphand_to_mem_node(plathand);
1493 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1494 			switch (query) {
1495 			case LGRP_MEM_SIZE_FREE:
1496 				npgs = MNODE_PGCNT(mnode);
1497 				break;
1498 			case LGRP_MEM_SIZE_AVAIL:
1499 				npgs = mem_node_memlist_pages(mnode,
1500 				    phys_avail);
1501 				break;
1502 			case LGRP_MEM_SIZE_INSTALL:
1503 				npgs = mem_node_memlist_pages(mnode,
1504 				    phys_install);
1505 				break;
1506 			default:
1507 				break;
1508 			}
1509 		}
1510 	}
1511 	return (npgs);
1512 }
1513 
1514 /*
1515  * Return latency between "from" and "to" lgroups
1516  *
1517  * This latency number can only be used for relative comparison
1518  * between lgroups on the running system, cannot be used across platforms,
1519  * and may not reflect the actual latency.  It is platform and implementation
1520  * specific, so platform gets to decide its value.  It would be nice if the
1521  * number was at least proportional to make comparisons more meaningful though.
1522  */
1523 /* ARGSUSED */
1524 int
1525 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1526 {
1527 	lgrp_handle_t	src, dest;
1528 
1529 	if (max_mem_nodes == 1)
1530 		return (0);
1531 
1532 	/*
1533 	 * Return max latency for root lgroup
1534 	 */
1535 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1536 		return (lgrp_plat_probe_time_max);
1537 
1538 	src = from;
1539 	dest = to;
1540 
1541 	/*
1542 	 * Return 0 for nodes (lgroup platform handles) out of range
1543 	 */
1544 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1545 		return (0);
1546 
1547 	/*
1548 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1549 	 * and we are trying to get latency from current CPU to some node
1550 	 */
1551 	if (lgrp_plat_probe_times[src][src] == 0 &&
1552 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1553 		lgrp_plat_probe();
1554 
1555 	return (lgrp_plat_probe_times[src][dest]);
1556 }
1557 
1558 /*
1559  * Return platform handle for root lgroup
1560  */
1561 lgrp_handle_t
1562 lgrp_plat_root_hand(void)
1563 {
1564 	return (LGRP_DEFAULT_HANDLE);
1565 }
1566