xref: /illumos-gate/usr/src/uts/i86pc/os/lgrpplat.c (revision 22a84b8d79248a611e4ba663a268d3c4bed054ac)
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 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * LOCALITY GROUP (LGROUP) PLATFORM SUPPORT FOR X86/AMD64 PLATFORMS
29  * ================================================================
30  * Multiprocessor AMD and Intel systems may have Non Uniform Memory Access
31  * (NUMA).  A NUMA machine consists of one or more "nodes" that each consist of
32  * one or more CPUs and some local memory.  The CPUs in each node can access
33  * the memory in the other nodes but at a higher latency than accessing their
34  * local memory.  Typically, a system with only one node has Uniform Memory
35  * Access (UMA), but it may be possible to have a one node system that has
36  * some global memory outside of the node which is higher latency.
37  *
38  * Module Description
39  * ------------------
40  * This module provides a platform interface for determining which CPUs and
41  * which memory (and how much) are in a NUMA node and how far each node is from
42  * each other.  The interface is used by the Virtual Memory (VM) system and the
43  * common lgroup framework.  The VM system uses the plat_*() routines to fill
44  * in its memory node (memnode) array with the physical address range spanned
45  * by each NUMA node to know which memory belongs to which node, so it can
46  * build and manage a physical page free list for each NUMA node and allocate
47  * local memory from each node as needed.  The common lgroup framework uses the
48  * exported lgrp_plat_*() routines to figure out which CPUs and memory belong
49  * to each node (leaf lgroup) and how far each node is from each other, so it
50  * can build the latency (lgroup) topology for the machine in order to optimize
51  * for locality.  Also, an lgroup platform handle instead of lgroups are used
52  * in the interface with this module, so this module shouldn't need to know
53  * anything about lgroups.  Instead, it just needs to know which CPUs, memory,
54  * etc. are in each NUMA node, how far each node is from each other, and to use
55  * a unique lgroup platform handle to refer to each node through the interface.
56  *
57  * Determining NUMA Configuration
58  * ------------------------------
59  * By default, this module will try to determine the NUMA configuration of the
60  * machine by reading the ACPI System Resource Affinity Table (SRAT) and System
61  * Locality Information Table (SLIT).  The SRAT contains info to tell which
62  * CPUs and memory are local to a given proximity domain (NUMA node).  The SLIT
63  * is a matrix that gives the distance between each system locality (which is
64  * a NUMA node and should correspond to proximity domains in the SRAT).  For
65  * more details on the SRAT and SLIT, please refer to an ACPI 3.0 or newer
66  * specification.
67  *
68  * If the SRAT doesn't exist on a system with AMD Opteron processors, we
69  * examine registers in PCI configuration space to determine how many nodes are
70  * in the system and which CPUs and memory are in each node.
71  * do while booting the kernel.
72  *
73  * NOTE: Using these PCI configuration space registers to determine this
74  *       locality info is not guaranteed to work or be compatible across all
75  *	 Opteron processor families.
76  *
77  * If the SLIT does not exist or look right, the kernel will probe to determine
78  * the distance between nodes as long as the NUMA CPU and memory configuration
79  * has been determined (see lgrp_plat_probe() for details).
80  *
81  * Data Structures
82  * ---------------
83  * The main data structures used by this code are the following:
84  *
85  * - lgrp_plat_cpu_node[]		CPU to node ID mapping table indexed by
86  *					CPU ID (only used for SRAT)
87  *
88  * - lgrp_plat_lat_stats.latencies[][]	Table of latencies between same and
89  *					different nodes indexed by node ID
90  *
91  * - lgrp_plat_node_cnt			Number of NUMA nodes in system
92  *
93  * - lgrp_plat_node_domain[]		Node ID to proximity domain ID mapping
94  *					table indexed by node ID (only used
95  *					for SRAT)
96  *
97  * - lgrp_plat_node_memory[]		Table with physical address range for
98  *					each node indexed by node ID
99  *
100  * The code is implemented to make the following always be true:
101  *
102  *	lgroup platform handle == node ID == memnode ID
103  *
104  * Moreover, it allows for the proximity domain ID to be equal to all of the
105  * above as long as the proximity domains IDs are numbered from 0 to <number of
106  * nodes - 1>.  This is done by hashing each proximity domain ID into the range
107  * from 0 to <number of nodes - 1>.  Then proximity ID N will hash into node ID
108  * N and proximity domain ID N will be entered into lgrp_plat_node_domain[N]
109  * and be assigned node ID N.  If the proximity domain IDs aren't numbered
110  * from 0 to <number of nodes - 1>, then hashing the proximity domain IDs into
111  * lgrp_plat_node_domain[] will still work for assigning proximity domain IDs
112  * to node IDs.  However, the proximity domain IDs may not map to the
113  * equivalent node ID since we want to keep the node IDs numbered from 0 to
114  * <number of nodes - 1> to minimize cost of searching and potentially space.
115  *
116  * The code below really tries to do the above.  However, the virtual memory
117  * system expects the memnodes which describe the physical address range for
118  * each NUMA node to be arranged in ascending order by physical address.  (:-(
119  * Otherwise, the kernel will panic in different semi-random places in the VM
120  * system.
121  *
122  * Consequently, this module has to try to sort the nodes in ascending order by
123  * each node's starting physical address to try to meet this "constraint" in
124  * the VM system (see lgrp_plat_node_sort()).  Also, the lowest numbered
125  * proximity domain ID in the system is deteremined and used to make the lowest
126  * numbered proximity domain map to node 0 in hopes that the proximity domains
127  * are sorted in ascending order by physical address already even if their IDs
128  * don't start at 0 (see NODE_DOMAIN_HASH() and lgrp_plat_srat_domains()).
129  * Finally, it is important to note that these workarounds may not be
130  * sufficient if/when memory hotplugging is supported and the VM system may
131  * ultimately need to be fixed to handle this....
132  */
133 
134 
135 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
136 #include <sys/bootconf.h>
137 #include <sys/cmn_err.h>
138 #include <sys/controlregs.h>
139 #include <sys/cpupart.h>
140 #include <sys/cpuvar.h>
141 #include <sys/lgrp.h>
142 #include <sys/machsystm.h>
143 #include <sys/memlist.h>
144 #include <sys/memnode.h>
145 #include <sys/mman.h>
146 #include <sys/pci_cfgspace.h>
147 #include <sys/pci_impl.h>
148 #include <sys/param.h>
149 #include <sys/pghw.h>
150 #include <sys/promif.h>		/* for prom_printf() */
151 #include <sys/sysmacros.h>
152 #include <sys/systm.h>
153 #include <sys/thread.h>
154 #include <sys/types.h>
155 #include <sys/var.h>
156 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
157 #include <vm/hat_i86.h>
158 #include <vm/seg_kmem.h>
159 #include <vm/vm_dep.h>
160 
161 #include "acpi_fw.h"		/* for SRAT and SLIT */
162 
163 
164 #define	MAX_NODES		8
165 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
166 
167 /*
168  * Constants for configuring probing
169  */
170 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
171 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
172 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
173 
174 /*
175  * Flags for probing
176  */
177 #define	LGRP_PLAT_PROBE_ENABLE		0x1	/* enable probing */
178 #define	LGRP_PLAT_PROBE_PGCPY		0x2	/* probe using page copy */
179 #define	LGRP_PLAT_PROBE_VENDOR		0x4	/* probe vendor ID register */
180 
181 /*
182  * Hash proximity domain ID into node to domain mapping table "mod" number of
183  * nodes to minimize span of entries used and try to have lowest numbered
184  * proximity domain be node 0
185  */
186 #define	NODE_DOMAIN_HASH(domain, node_cnt) \
187 	((lgrp_plat_prox_domain_min == UINT32_MAX) ? (domain) % node_cnt : \
188 	    ((domain) - lgrp_plat_prox_domain_min) % node_cnt)
189 
190 
191 /*
192  * CPU to node ID mapping structure (only used with SRAT)
193  */
194 typedef	struct cpu_node_map {
195 	int		exists;
196 	uint_t		node;
197 	uint32_t	apicid;
198 	uint32_t	prox_domain;
199 } cpu_node_map_t;
200 
201 /*
202  * Latency statistics
203  */
204 typedef struct lgrp_plat_latency_stats {
205 	hrtime_t	latencies[MAX_NODES][MAX_NODES];
206 	hrtime_t	latency_max;
207 	hrtime_t	latency_min;
208 } lgrp_plat_latency_stats_t;
209 
210 /*
211  * Memory configuration for probing
212  */
213 typedef struct lgrp_plat_probe_mem_config {
214 	size_t	probe_memsize;		/* how much memory to probe per node */
215 	caddr_t	probe_va[MAX_NODES];	/* where memory mapped for probing */
216 	pfn_t	probe_pfn[MAX_NODES];	/* physical pages to map for probing */
217 } lgrp_plat_probe_mem_config_t;
218 
219 /*
220  * Statistics kept for probing
221  */
222 typedef struct lgrp_plat_probe_stats {
223 	hrtime_t	flush_cost;
224 	hrtime_t	probe_cost;
225 	hrtime_t	probe_cost_total;
226 	hrtime_t	probe_error_code;
227 	hrtime_t	probe_errors[MAX_NODES][MAX_NODES];
228 	int		probe_suspect[MAX_NODES][MAX_NODES];
229 	hrtime_t	probe_max[MAX_NODES][MAX_NODES];
230 	hrtime_t	probe_min[MAX_NODES][MAX_NODES];
231 } lgrp_plat_probe_stats_t;
232 
233 /*
234  * Node to proximity domain ID mapping structure (only used with SRAT)
235  */
236 typedef	struct node_domain_map {
237 	int		exists;
238 	uint32_t	prox_domain;
239 } node_domain_map_t;
240 
241 /*
242  * Node ID and starting and ending page for physical memory in node
243  */
244 typedef	struct node_phys_addr_map {
245 	pfn_t		start;
246 	pfn_t		end;
247 	int		exists;
248 	uint32_t	prox_domain;
249 } node_phys_addr_map_t;
250 
251 /*
252  * Number of CPUs for which we got APIC IDs
253  */
254 static int				lgrp_plat_apic_ncpus = 0;
255 
256 /*
257  * CPU to node ID mapping table (only used for SRAT) and its max number of
258  * entries
259  */
260 static cpu_node_map_t			*lgrp_plat_cpu_node = NULL;
261 static uint_t				lgrp_plat_cpu_node_nentries = 0;
262 
263 /*
264  * Latency statistics
265  */
266 lgrp_plat_latency_stats_t		lgrp_plat_lat_stats;
267 
268 /*
269  * Whether memory is interleaved across nodes causing MPO to be disabled
270  */
271 static int				lgrp_plat_mem_intrlv = 0;
272 
273 /*
274  * Node ID to proximity domain ID mapping table (only used for SRAT)
275  */
276 static node_domain_map_t		lgrp_plat_node_domain[MAX_NODES];
277 
278 /*
279  * Physical address range for memory in each node
280  */
281 static node_phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
282 
283 /*
284  * Statistics gotten from probing
285  */
286 static lgrp_plat_probe_stats_t		lgrp_plat_probe_stats;
287 
288 /*
289  * Memory configuration for probing
290  */
291 static lgrp_plat_probe_mem_config_t	lgrp_plat_probe_mem_config;
292 
293 /*
294  * Lowest proximity domain ID seen in ACPI SRAT
295  */
296 static uint32_t				lgrp_plat_prox_domain_min = UINT32_MAX;
297 
298 /*
299  * Error code from processing ACPI SRAT
300  */
301 static int				lgrp_plat_srat_error = 0;
302 
303 /*
304  * Error code from processing ACPI SLIT
305  */
306 static int				lgrp_plat_slit_error = 0;
307 
308 /*
309  * Allocate lgroup array statically
310  */
311 static lgrp_t				lgrp_space[NLGRP];
312 static int				nlgrps_alloc;
313 
314 
315 /*
316  * Enable finding and using minimum proximity domain ID when hashing
317  */
318 int			lgrp_plat_domain_min_enable = 1;
319 
320 /*
321  * Number of nodes in system
322  */
323 uint_t			lgrp_plat_node_cnt = 1;
324 
325 /*
326  * Enable sorting nodes in ascending order by starting physical address
327  */
328 int			lgrp_plat_node_sort_enable = 1;
329 
330 /*
331  * Configuration Parameters for Probing
332  * - lgrp_plat_probe_flags	Flags to specify enabling probing, probe
333  *				operation, etc.
334  * - lgrp_plat_probe_nrounds	How many rounds of probing to do
335  * - lgrp_plat_probe_nsamples	Number of samples to take when probing each
336  *				node
337  * - lgrp_plat_probe_nreads	Number of times to read vendor ID from
338  *				Northbridge for each probe
339  */
340 uint_t			lgrp_plat_probe_flags = 0;
341 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
342 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
343 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
344 
345 /*
346  * Enable use of ACPI System Resource Affinity Table (SRAT) and System
347  * Locality Information Table (SLIT)
348  */
349 int			lgrp_plat_srat_enable = 1;
350 int			lgrp_plat_slit_enable = 1;
351 
352 /*
353  * mnode_xwa: set to non-zero value to initiate workaround if large pages are
354  * found to be crossing memory node boundaries. The workaround will eliminate
355  * a base size page at the end of each memory node boundary to ensure that
356  * a large page with constituent pages that span more than 1 memory node
357  * can never be formed.
358  *
359  */
360 int	mnode_xwa = 1;
361 
362 /*
363  * Static array to hold lgroup statistics
364  */
365 struct lgrp_stats	lgrp_stats[NLGRP];
366 
367 
368 /*
369  * Forward declarations of platform interface routines
370  */
371 void		plat_build_mem_nodes(struct memlist *list);
372 
373 int		plat_lgrphand_to_mem_node(lgrp_handle_t hand);
374 
375 lgrp_handle_t	plat_mem_node_to_lgrphand(int mnode);
376 
377 int		plat_mnode_xcheck(pfn_t pfncnt);
378 
379 int		plat_pfn_to_mem_node(pfn_t pfn);
380 
381 /*
382  * Forward declarations of lgroup platform interface routines
383  */
384 lgrp_t		*lgrp_plat_alloc(lgrp_id_t lgrpid);
385 
386 void		lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg);
387 
388 lgrp_handle_t	lgrp_plat_cpu_to_hand(processorid_t id);
389 
390 void		lgrp_plat_init(lgrp_init_stages_t stage);
391 
392 int		lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to);
393 
394 int		lgrp_plat_max_lgrps(void);
395 
396 pgcnt_t		lgrp_plat_mem_size(lgrp_handle_t plathand,
397     lgrp_mem_query_t query);
398 
399 lgrp_handle_t	lgrp_plat_pfn_to_hand(pfn_t pfn);
400 
401 void		lgrp_plat_probe(void);
402 
403 lgrp_handle_t	lgrp_plat_root_hand(void);
404 
405 
406 /*
407  * Forward declarations of local routines
408  */
409 static int	is_opteron(void);
410 
411 static int	lgrp_plat_cpu_node_update(node_domain_map_t *node_domain,
412     int node_cnt, cpu_node_map_t *cpu_node, int nentries, uint32_t apicid,
413     uint32_t domain);
414 
415 static int	lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node,
416     int cpu_node_nentries);
417 
418 static int	lgrp_plat_domain_to_node(node_domain_map_t *node_domain,
419     int node_cnt, uint32_t domain);
420 
421 static void	lgrp_plat_get_numa_config(void);
422 
423 static void	lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
424     lgrp_plat_latency_stats_t *lat_stats,
425     lgrp_plat_probe_stats_t *probe_stats);
426 
427 static int	lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
428     lgrp_plat_latency_stats_t *lat_stats);
429 
430 static void	lgrp_plat_main_init(void);
431 
432 static pgcnt_t	lgrp_plat_mem_size_default(lgrp_handle_t, lgrp_mem_query_t);
433 
434 static int	lgrp_plat_node_domain_update(node_domain_map_t *node_domain,
435     int node_cnt, uint32_t domain);
436 
437 static int	lgrp_plat_node_memory_update(node_domain_map_t *node_domain,
438     int node_cnt, node_phys_addr_map_t *node_memory, uint64_t start,
439     uint64_t end, uint32_t domain);
440 
441 static void	lgrp_plat_node_sort(node_domain_map_t *node_domain,
442     int node_cnt, cpu_node_map_t *cpu_node, int cpu_count,
443     node_phys_addr_map_t *node_memory);
444 
445 static hrtime_t	lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
446     int cpu_node_nentries, lgrp_plat_probe_mem_config_t *probe_mem_config,
447     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats);
448 
449 static int	lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node);
450 
451 static int	lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
452     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats);
453 
454 static int	lgrp_plat_process_srat(struct srat *tp,
455     uint32_t *prox_domain_min, node_domain_map_t *node_domain,
456     cpu_node_map_t *cpu_node, int cpu_count,
457     node_phys_addr_map_t *node_memory);
458 
459 static void	lgrp_plat_release_bootstrap(void);
460 
461 static int	lgrp_plat_srat_domains(struct srat *tp,
462     uint32_t *prox_domain_min);
463 
464 static void	lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
465     lgrp_plat_latency_stats_t *lat_stats);
466 
467 static void	opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
468     node_phys_addr_map_t *node_memory);
469 
470 static hrtime_t	opt_probe_vendor(int dest_node, int nreads);
471 
472 
473 /*
474  * PLATFORM INTERFACE ROUTINES
475  */
476 
477 /*
478  * Configure memory nodes for machines with more than one node (ie NUMA)
479  */
480 void
481 plat_build_mem_nodes(struct memlist *list)
482 {
483 	pfn_t		cur_start;	/* start addr of subrange */
484 	pfn_t		cur_end;	/* end addr of subrange */
485 	pfn_t		start;		/* start addr of whole range */
486 	pfn_t		end;		/* end addr of whole range */
487 	pgcnt_t		endcnt;		/* pages to sacrifice */
488 
489 	/*
490 	 * Boot install lists are arranged <addr, len>, ...
491 	 */
492 	while (list) {
493 		int	node;
494 
495 		start = list->address >> PAGESHIFT;
496 		end = (list->address + list->size - 1) >> PAGESHIFT;
497 
498 		if (start > physmax) {
499 			list = list->next;
500 			continue;
501 		}
502 		if (end > physmax)
503 			end = physmax;
504 
505 		/*
506 		 * When there is only one memnode, just add memory to memnode
507 		 */
508 		if (max_mem_nodes == 1) {
509 			mem_node_add_slice(start, end);
510 			list = list->next;
511 			continue;
512 		}
513 
514 		/*
515 		 * mem_node_add_slice() expects to get a memory range that
516 		 * is within one memnode, so need to split any memory range
517 		 * that spans multiple memnodes into subranges that are each
518 		 * contained within one memnode when feeding them to
519 		 * mem_node_add_slice()
520 		 */
521 		cur_start = start;
522 		do {
523 			node = plat_pfn_to_mem_node(cur_start);
524 
525 			/*
526 			 * Panic if DRAM address map registers or SRAT say
527 			 * memory in node doesn't exist or address from
528 			 * boot installed memory list entry isn't in this node.
529 			 * This shouldn't happen and rest of code can't deal
530 			 * with this if it does.
531 			 */
532 			if (node < 0 || node >= lgrp_plat_node_cnt ||
533 			    !lgrp_plat_node_memory[node].exists ||
534 			    cur_start < lgrp_plat_node_memory[node].start ||
535 			    cur_start > lgrp_plat_node_memory[node].end) {
536 				cmn_err(CE_PANIC, "Don't know which memnode "
537 				    "to add installed memory address 0x%lx\n",
538 				    cur_start);
539 			}
540 
541 			/*
542 			 * End of current subrange should not span memnodes
543 			 */
544 			cur_end = end;
545 			endcnt = 0;
546 			if (lgrp_plat_node_memory[node].exists &&
547 			    cur_end > lgrp_plat_node_memory[node].end) {
548 				cur_end = lgrp_plat_node_memory[node].end;
549 				if (mnode_xwa > 1) {
550 					/*
551 					 * sacrifice the last page in each
552 					 * node to eliminate large pages
553 					 * that span more than 1 memory node.
554 					 */
555 					endcnt = 1;
556 					physinstalled--;
557 				}
558 			}
559 
560 			mem_node_add_slice(cur_start, cur_end - endcnt);
561 
562 			/*
563 			 * Next subrange starts after end of current one
564 			 */
565 			cur_start = cur_end + 1;
566 		} while (cur_end < end);
567 
568 		list = list->next;
569 	}
570 	mem_node_physalign = 0;
571 	mem_node_pfn_shift = 0;
572 }
573 
574 
575 int
576 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
577 {
578 	if (max_mem_nodes == 1)
579 		return (0);
580 
581 	return ((int)hand);
582 }
583 
584 
585 /*
586  * plat_mnode_xcheck: checks the node memory ranges to see if there is a pfncnt
587  * range of pages aligned on pfncnt that crosses an node boundary. Returns 1 if
588  * a crossing is found and returns 0 otherwise.
589  */
590 int
591 plat_mnode_xcheck(pfn_t pfncnt)
592 {
593 	int	node, prevnode = -1, basenode;
594 	pfn_t	ea, sa;
595 
596 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
597 
598 		if (lgrp_plat_node_memory[node].exists == 0)
599 			continue;
600 
601 		if (prevnode == -1) {
602 			prevnode = node;
603 			basenode = node;
604 			continue;
605 		}
606 
607 		/* assume x86 node pfn ranges are in increasing order */
608 		ASSERT(lgrp_plat_node_memory[node].start >
609 		    lgrp_plat_node_memory[prevnode].end);
610 
611 		/*
612 		 * continue if the starting address of node is not contiguous
613 		 * with the previous node.
614 		 */
615 
616 		if (lgrp_plat_node_memory[node].start !=
617 		    (lgrp_plat_node_memory[prevnode].end + 1)) {
618 			basenode = node;
619 			prevnode = node;
620 			continue;
621 		}
622 
623 		/* check if the starting address of node is pfncnt aligned */
624 		if ((lgrp_plat_node_memory[node].start & (pfncnt - 1)) != 0) {
625 
626 			/*
627 			 * at this point, node starts at an unaligned boundary
628 			 * and is contiguous with the previous node(s) to
629 			 * basenode. Check if there is an aligned contiguous
630 			 * range of length pfncnt that crosses this boundary.
631 			 */
632 
633 			sa = P2ALIGN(lgrp_plat_node_memory[prevnode].end,
634 			    pfncnt);
635 			ea = P2ROUNDUP((lgrp_plat_node_memory[node].start),
636 			    pfncnt);
637 
638 			ASSERT((ea - sa) == pfncnt);
639 			if (sa >= lgrp_plat_node_memory[basenode].start &&
640 			    ea <= (lgrp_plat_node_memory[node].end + 1)) {
641 				/*
642 				 * large page found to cross mnode boundary.
643 				 * Return Failure if workaround not enabled.
644 				 */
645 				if (mnode_xwa == 0)
646 					return (1);
647 				mnode_xwa++;
648 			}
649 		}
650 		prevnode = node;
651 	}
652 	return (0);
653 }
654 
655 
656 lgrp_handle_t
657 plat_mem_node_to_lgrphand(int mnode)
658 {
659 	if (max_mem_nodes == 1)
660 		return (LGRP_DEFAULT_HANDLE);
661 
662 	return ((lgrp_handle_t)mnode);
663 }
664 
665 
666 int
667 plat_pfn_to_mem_node(pfn_t pfn)
668 {
669 	int	node;
670 
671 	if (max_mem_nodes == 1)
672 		return (0);
673 
674 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
675 		/*
676 		 * Skip nodes with no memory
677 		 */
678 		if (!lgrp_plat_node_memory[node].exists)
679 			continue;
680 
681 		if (pfn >= lgrp_plat_node_memory[node].start &&
682 		    pfn <= lgrp_plat_node_memory[node].end)
683 			return (node);
684 	}
685 
686 	/*
687 	 * Didn't find memnode where this PFN lives which should never happen
688 	 */
689 	ASSERT(node < lgrp_plat_node_cnt);
690 	return (-1);
691 }
692 
693 
694 /*
695  * LGROUP PLATFORM INTERFACE ROUTINES
696  */
697 
698 /*
699  * Allocate additional space for an lgroup.
700  */
701 /* ARGSUSED */
702 lgrp_t *
703 lgrp_plat_alloc(lgrp_id_t lgrpid)
704 {
705 	lgrp_t *lgrp;
706 
707 	lgrp = &lgrp_space[nlgrps_alloc++];
708 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
709 		return (NULL);
710 	return (lgrp);
711 }
712 
713 
714 /*
715  * Platform handling for (re)configuration changes
716  */
717 /* ARGSUSED */
718 void
719 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
720 {
721 }
722 
723 
724 /*
725  * Return the platform handle for the lgroup containing the given CPU
726  */
727 /* ARGSUSED */
728 lgrp_handle_t
729 lgrp_plat_cpu_to_hand(processorid_t id)
730 {
731 	lgrp_handle_t	hand;
732 
733 	if (lgrp_plat_node_cnt == 1)
734 		return (LGRP_DEFAULT_HANDLE);
735 
736 	hand = (lgrp_handle_t)lgrp_plat_cpu_to_node(cpu[id],
737 	    lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries);
738 
739 	ASSERT(hand != (lgrp_handle_t)-1);
740 	if (hand == (lgrp_handle_t)-1)
741 		return (LGRP_NULL_HANDLE);
742 
743 	return (hand);
744 }
745 
746 
747 /*
748  * Platform-specific initialization of lgroups
749  */
750 void
751 lgrp_plat_init(lgrp_init_stages_t stage)
752 {
753 #if defined(__xpv)
754 #else	/* __xpv */
755 	u_longlong_t	value;
756 #endif	/* __xpv */
757 
758 	switch (stage) {
759 	case LGRP_INIT_STAGE1:
760 #if defined(__xpv)
761 		/*
762 		 * XXPV	For now, the hypervisor treats all memory equally.
763 		 */
764 		lgrp_plat_node_cnt = max_mem_nodes = 1;
765 #else	/* __xpv */
766 		/*
767 		 * Get boot property for lgroup topology height limit
768 		 */
769 		if (bootprop_getval(BP_LGRP_TOPO_LEVELS, &value) == 0)
770 			(void) lgrp_topo_ht_limit_set((int)value);
771 
772 		/*
773 		 * Get boot property for enabling/disabling SRAT
774 		 */
775 		if (bootprop_getval(BP_LGRP_SRAT_ENABLE, &value) == 0)
776 			lgrp_plat_srat_enable = (int)value;
777 
778 		/*
779 		 * Get boot property for enabling/disabling SLIT
780 		 */
781 		if (bootprop_getval(BP_LGRP_SLIT_ENABLE, &value) == 0)
782 			lgrp_plat_slit_enable = (int)value;
783 
784 		/*
785 		 * Initialize as a UMA machine
786 		 */
787 		if (lgrp_topo_ht_limit() == 1) {
788 			lgrp_plat_node_cnt = max_mem_nodes = 1;
789 			return;
790 		}
791 
792 		lgrp_plat_get_numa_config();
793 #endif	/* __xpv */
794 		break;
795 
796 	case LGRP_INIT_STAGE3:
797 		lgrp_plat_probe();
798 		lgrp_plat_release_bootstrap();
799 		break;
800 
801 	case LGRP_INIT_STAGE4:
802 		lgrp_plat_main_init();
803 		break;
804 
805 	default:
806 		break;
807 	}
808 }
809 
810 
811 /*
812  * Return latency between "from" and "to" lgroups
813  *
814  * This latency number can only be used for relative comparison
815  * between lgroups on the running system, cannot be used across platforms,
816  * and may not reflect the actual latency.  It is platform and implementation
817  * specific, so platform gets to decide its value.  It would be nice if the
818  * number was at least proportional to make comparisons more meaningful though.
819  */
820 /* ARGSUSED */
821 int
822 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
823 {
824 	lgrp_handle_t	src, dest;
825 	int		node;
826 
827 	if (max_mem_nodes == 1)
828 		return (0);
829 
830 	/*
831 	 * Return max latency for root lgroup
832 	 */
833 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
834 		return (lgrp_plat_lat_stats.latency_max);
835 
836 	src = from;
837 	dest = to;
838 
839 	/*
840 	 * Return 0 for nodes (lgroup platform handles) out of range
841 	 */
842 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
843 		return (0);
844 
845 	/*
846 	 * Probe from current CPU if its lgroup latencies haven't been set yet
847 	 * and we are trying to get latency from current CPU to some node
848 	 */
849 	node = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
850 	    lgrp_plat_cpu_node_nentries);
851 	ASSERT(node >= 0 && node < lgrp_plat_node_cnt);
852 	if (lgrp_plat_lat_stats.latencies[src][src] == 0 && node == src)
853 		lgrp_plat_probe();
854 
855 	return (lgrp_plat_lat_stats.latencies[src][dest]);
856 }
857 
858 
859 /*
860  * Return the maximum number of lgrps supported by the platform.
861  * Before lgrp topology is known it returns an estimate based on the number of
862  * nodes. Once topology is known it returns the actual maximim number of lgrps
863  * created. Since x86/x64 doesn't support Dynamic Reconfiguration (DR) and
864  * dynamic addition of new nodes, this number may not grow during system
865  * lifetime (yet).
866  */
867 int
868 lgrp_plat_max_lgrps(void)
869 {
870 	return (lgrp_topo_initialized ?
871 	    lgrp_alloc_max + 1 :
872 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
873 }
874 
875 
876 /*
877  * Return the number of free pages in an lgroup.
878  *
879  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
880  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
881  * number of allocatable base pagesize pages corresponding to the
882  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
883  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
884  * memory installed, regardless of whether or not it's usable.
885  */
886 pgcnt_t
887 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
888 {
889 	int	mnode;
890 	pgcnt_t npgs = (pgcnt_t)0;
891 	extern struct memlist *phys_avail;
892 	extern struct memlist *phys_install;
893 
894 
895 	if (plathand == LGRP_DEFAULT_HANDLE)
896 		return (lgrp_plat_mem_size_default(plathand, query));
897 
898 	if (plathand != LGRP_NULL_HANDLE) {
899 		mnode = plat_lgrphand_to_mem_node(plathand);
900 		if (mnode >= 0 && mem_node_config[mnode].exists) {
901 			switch (query) {
902 			case LGRP_MEM_SIZE_FREE:
903 				npgs = MNODE_PGCNT(mnode);
904 				break;
905 			case LGRP_MEM_SIZE_AVAIL:
906 				npgs = mem_node_memlist_pages(mnode,
907 				    phys_avail);
908 				break;
909 			case LGRP_MEM_SIZE_INSTALL:
910 				npgs = mem_node_memlist_pages(mnode,
911 				    phys_install);
912 				break;
913 			default:
914 				break;
915 			}
916 		}
917 	}
918 	return (npgs);
919 }
920 
921 
922 /*
923  * Return the platform handle of the lgroup that contains the physical memory
924  * corresponding to the given page frame number
925  */
926 /* ARGSUSED */
927 lgrp_handle_t
928 lgrp_plat_pfn_to_hand(pfn_t pfn)
929 {
930 	int	mnode;
931 
932 	if (max_mem_nodes == 1)
933 		return (LGRP_DEFAULT_HANDLE);
934 
935 	if (pfn > physmax)
936 		return (LGRP_NULL_HANDLE);
937 
938 	mnode = plat_pfn_to_mem_node(pfn);
939 	if (mnode < 0)
940 		return (LGRP_NULL_HANDLE);
941 
942 	return (MEM_NODE_2_LGRPHAND(mnode));
943 }
944 
945 
946 /*
947  * Probe memory in each node from current CPU to determine latency topology
948  *
949  * The probing code will probe the vendor ID register on the Northbridge of
950  * Opteron processors and probe memory for other processors by default.
951  *
952  * Since probing is inherently error prone, the code takes laps across all the
953  * nodes probing from each node to each of the other nodes some number of
954  * times.  Furthermore, each node is probed some number of times before moving
955  * onto the next one during each lap.  The minimum latency gotten between nodes
956  * is kept as the latency between the nodes.
957  *
958  * After all that,  the probe times are adjusted by normalizing values that are
959  * close to each other and local latencies are made the same.  Lastly, the
960  * latencies are verified to make sure that certain conditions are met (eg.
961  * local < remote, latency(a, b) == latency(b, a), etc.).
962  *
963  * If any of the conditions aren't met, the code will export a NUMA
964  * configuration with the local CPUs and memory given by the SRAT or PCI config
965  * space registers and one remote memory latency since it can't tell exactly
966  * how far each node is from each other.
967  */
968 void
969 lgrp_plat_probe(void)
970 {
971 	int				from;
972 	int				i;
973 	lgrp_plat_latency_stats_t	*lat_stats;
974 	boolean_t			probed;
975 	hrtime_t			probe_time;
976 	int				to;
977 
978 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
979 	    max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
980 		return;
981 
982 	/*
983 	 * Determine ID of node containing current CPU
984 	 */
985 	from = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
986 	    lgrp_plat_cpu_node_nentries);
987 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
988 	if (srat_ptr && lgrp_plat_srat_enable && !lgrp_plat_srat_error)
989 		ASSERT(lgrp_plat_node_domain[from].exists);
990 
991 	/*
992 	 * Don't need to probe if got times already
993 	 */
994 	lat_stats = &lgrp_plat_lat_stats;
995 	if (lat_stats->latencies[from][from] != 0)
996 		return;
997 
998 	/*
999 	 * Read vendor ID in Northbridge or read and write page(s)
1000 	 * in each node from current CPU and remember how long it takes,
1001 	 * so we can build latency topology of machine later.
1002 	 * This should approximate the memory latency between each node.
1003 	 */
1004 	probed = B_FALSE;
1005 	for (i = 0; i < lgrp_plat_probe_nrounds; i++) {
1006 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1007 			/*
1008 			 * Get probe time and skip over any nodes that can't be
1009 			 * probed yet or don't have memory
1010 			 */
1011 			probe_time = lgrp_plat_probe_time(to,
1012 			    lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries,
1013 			    &lgrp_plat_probe_mem_config, &lgrp_plat_lat_stats,
1014 			    &lgrp_plat_probe_stats);
1015 			if (probe_time == 0)
1016 				continue;
1017 
1018 			probed = B_TRUE;
1019 
1020 			/*
1021 			 * Keep lowest probe time as latency between nodes
1022 			 */
1023 			if (lat_stats->latencies[from][to] == 0 ||
1024 			    probe_time < lat_stats->latencies[from][to])
1025 				lat_stats->latencies[from][to] = probe_time;
1026 
1027 			/*
1028 			 * Update overall minimum and maximum probe times
1029 			 * across all nodes
1030 			 */
1031 			if (probe_time < lat_stats->latency_min ||
1032 			    lat_stats->latency_min == -1)
1033 				lat_stats->latency_min = probe_time;
1034 			if (probe_time > lat_stats->latency_max)
1035 				lat_stats->latency_max = probe_time;
1036 		}
1037 	}
1038 
1039 	/*
1040 	 * Bail out if weren't able to probe any nodes from current CPU
1041 	 */
1042 	if (probed == B_FALSE)
1043 		return;
1044 
1045 	/*
1046 	 * - Fix up latencies such that local latencies are same,
1047 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1048 	 *
1049 	 * - Verify that latencies look ok
1050 	 *
1051 	 * - Fallback to just optimizing for local and remote if
1052 	 *   latencies didn't look right
1053 	 */
1054 	lgrp_plat_latency_adjust(lgrp_plat_node_memory, &lgrp_plat_lat_stats,
1055 	    &lgrp_plat_probe_stats);
1056 	lgrp_plat_probe_stats.probe_error_code =
1057 	    lgrp_plat_latency_verify(lgrp_plat_node_memory,
1058 	    &lgrp_plat_lat_stats);
1059 	if (lgrp_plat_probe_stats.probe_error_code)
1060 		lgrp_plat_2level_setup(lgrp_plat_node_memory,
1061 		    &lgrp_plat_lat_stats);
1062 }
1063 
1064 
1065 /*
1066  * Return platform handle for root lgroup
1067  */
1068 lgrp_handle_t
1069 lgrp_plat_root_hand(void)
1070 {
1071 	return (LGRP_DEFAULT_HANDLE);
1072 }
1073 
1074 
1075 /*
1076  * INTERNAL ROUTINES
1077  */
1078 
1079 
1080 /*
1081  * Update CPU to node mapping for given CPU and proximity domain (and returns
1082  * negative numbers for errors and positive ones for success)
1083  */
1084 static int
1085 lgrp_plat_cpu_node_update(node_domain_map_t *node_domain, int node_cnt,
1086     cpu_node_map_t *cpu_node, int nentries, uint32_t apicid, uint32_t domain)
1087 {
1088 	uint_t	i;
1089 	int	node;
1090 
1091 	/*
1092 	 * Get node number for proximity domain
1093 	 */
1094 	node = lgrp_plat_domain_to_node(node_domain, node_cnt, domain);
1095 	if (node == -1) {
1096 		node = lgrp_plat_node_domain_update(node_domain, node_cnt,
1097 		    domain);
1098 		if (node == -1)
1099 			return (-1);
1100 	}
1101 
1102 	/*
1103 	 * Search for entry with given APIC ID and fill in its node and
1104 	 * proximity domain IDs (if they haven't been set already)
1105 	 */
1106 	for (i = 0; i < nentries; i++) {
1107 		/*
1108 		 * Skip nonexistent entries and ones without matching APIC ID
1109 		 */
1110 		if (!cpu_node[i].exists || cpu_node[i].apicid != apicid)
1111 			continue;
1112 
1113 		/*
1114 		 * Just return if entry completely and correctly filled in
1115 		 * already
1116 		 */
1117 		if (cpu_node[i].prox_domain == domain &&
1118 		    cpu_node[i].node == node)
1119 			return (1);
1120 
1121 		/*
1122 		 * Fill in node and proximity domain IDs
1123 		 */
1124 		cpu_node[i].prox_domain = domain;
1125 		cpu_node[i].node = node;
1126 
1127 		return (0);
1128 	}
1129 
1130 	/*
1131 	 * Return error when entry for APIC ID wasn't found in table
1132 	 */
1133 	return (-2);
1134 }
1135 
1136 
1137 /*
1138  * Get node ID for given CPU
1139  */
1140 static int
1141 lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node,
1142     int cpu_node_nentries)
1143 {
1144 	processorid_t	cpuid;
1145 
1146 	if (cp == NULL)
1147 		return (-1);
1148 
1149 	cpuid = cp->cpu_id;
1150 	if (cpuid < 0 || cpuid >= max_ncpus)
1151 		return (-1);
1152 
1153 	/*
1154 	 * SRAT doesn't exist, isn't enabled, or there was an error processing
1155 	 * it, so return chip ID for Opteron and -1 otherwise.
1156 	 */
1157 	if (srat_ptr == NULL || !lgrp_plat_srat_enable ||
1158 	    lgrp_plat_srat_error) {
1159 		if (is_opteron())
1160 			return (pg_plat_hw_instance_id(cp, PGHW_CHIP));
1161 		return (-1);
1162 	}
1163 
1164 	/*
1165 	 * Return -1 when CPU to node ID mapping entry doesn't exist for given
1166 	 * CPU
1167 	 */
1168 	if (cpuid >= cpu_node_nentries || !cpu_node[cpuid].exists)
1169 		return (-1);
1170 
1171 	return (cpu_node[cpuid].node);
1172 }
1173 
1174 
1175 /*
1176  * Return node number for given proximity domain/system locality
1177  */
1178 static int
1179 lgrp_plat_domain_to_node(node_domain_map_t *node_domain, int node_cnt,
1180     uint32_t domain)
1181 {
1182 	uint_t	node;
1183 	uint_t	start;
1184 
1185 	/*
1186 	 * Hash proximity domain ID into node to domain mapping table (array),
1187 	 * search for entry with matching proximity domain ID, and return index
1188 	 * of matching entry as node ID.
1189 	 */
1190 	node = start = NODE_DOMAIN_HASH(domain, node_cnt);
1191 	do {
1192 		if (node_domain[node].prox_domain == domain &&
1193 		    node_domain[node].exists)
1194 			return (node);
1195 		node = (node + 1) % node_cnt;
1196 	} while (node != start);
1197 	return (-1);
1198 }
1199 
1200 
1201 /*
1202  * Get NUMA configuration of machine
1203  */
1204 static void
1205 lgrp_plat_get_numa_config(void)
1206 {
1207 	uint_t		probe_op;
1208 
1209 	/*
1210 	 * Read boot property with CPU to APIC ID mapping table/array to
1211 	 * determine number of CPUs
1212 	 */
1213 	lgrp_plat_apic_ncpus = lgrp_plat_process_cpu_apicids(NULL);
1214 
1215 	/*
1216 	 * Determine which CPUs and memory are local to each other and number
1217 	 * of NUMA nodes by reading ACPI System Resource Affinity Table (SRAT)
1218 	 */
1219 	if (lgrp_plat_apic_ncpus > 0) {
1220 		int	retval;
1221 
1222 		/*
1223 		 * Temporarily allocate boot memory to use for CPU to node
1224 		 * mapping since kernel memory allocator isn't alive yet
1225 		 */
1226 		lgrp_plat_cpu_node = (cpu_node_map_t *)BOP_ALLOC(bootops,
1227 		    NULL, lgrp_plat_apic_ncpus * sizeof (cpu_node_map_t),
1228 		    sizeof (int));
1229 
1230 		ASSERT(lgrp_plat_cpu_node != NULL);
1231 		if (lgrp_plat_cpu_node) {
1232 			lgrp_plat_cpu_node_nentries = lgrp_plat_apic_ncpus;
1233 			bzero(lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries *
1234 			    sizeof (cpu_node_map_t));
1235 		}
1236 
1237 		/*
1238 		 * Fill in CPU to node ID mapping table with APIC ID for each
1239 		 * CPU
1240 		 */
1241 		(void) lgrp_plat_process_cpu_apicids(lgrp_plat_cpu_node);
1242 
1243 		retval = lgrp_plat_process_srat(srat_ptr,
1244 		    &lgrp_plat_prox_domain_min,
1245 		    lgrp_plat_node_domain, lgrp_plat_cpu_node,
1246 		    lgrp_plat_apic_ncpus, lgrp_plat_node_memory);
1247 		if (retval <= 0) {
1248 			lgrp_plat_srat_error = retval;
1249 			lgrp_plat_node_cnt = 1;
1250 		} else {
1251 			lgrp_plat_srat_error = 0;
1252 			lgrp_plat_node_cnt = retval;
1253 		}
1254 	}
1255 
1256 	/*
1257 	 * Try to use PCI config space registers on Opteron if there's an error
1258 	 * processing CPU to APIC ID mapping or SRAT
1259 	 */
1260 	if ((lgrp_plat_apic_ncpus <= 0 || lgrp_plat_srat_error != 0) &&
1261 	    is_opteron())
1262 		opt_get_numa_config(&lgrp_plat_node_cnt, &lgrp_plat_mem_intrlv,
1263 		    lgrp_plat_node_memory);
1264 
1265 	/*
1266 	 * Don't bother to setup system for multiple lgroups and only use one
1267 	 * memory node when memory is interleaved between any nodes or there is
1268 	 * only one NUMA node
1269 	 *
1270 	 * NOTE: May need to change this for Dynamic Reconfiguration (DR)
1271 	 *	 when and if it happens for x86/x64
1272 	 */
1273 	if (lgrp_plat_mem_intrlv || lgrp_plat_node_cnt == 1) {
1274 		lgrp_plat_node_cnt = max_mem_nodes = 1;
1275 		(void) lgrp_topo_ht_limit_set(1);
1276 		return;
1277 	}
1278 
1279 	/*
1280 	 * Leaf lgroups on x86/x64 architectures contain one physical
1281 	 * processor chip. Tune lgrp_expand_proc_thresh and
1282 	 * lgrp_expand_proc_diff so that lgrp_choose() will spread
1283 	 * things out aggressively.
1284 	 */
1285 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
1286 	lgrp_expand_proc_diff = 0;
1287 
1288 	/*
1289 	 * There should be one memnode (physical page free list(s)) for
1290 	 * each node
1291 	 */
1292 	max_mem_nodes = lgrp_plat_node_cnt;
1293 
1294 	/*
1295 	 * Initialize min and max latency before reading SLIT or probing
1296 	 */
1297 	lgrp_plat_lat_stats.latency_min = -1;
1298 	lgrp_plat_lat_stats.latency_max = 0;
1299 
1300 	/*
1301 	 * Determine how far each NUMA node is from each other by
1302 	 * reading ACPI System Locality Information Table (SLIT) if it
1303 	 * exists
1304 	 */
1305 	lgrp_plat_slit_error = lgrp_plat_process_slit(slit_ptr,
1306 	    lgrp_plat_node_cnt, lgrp_plat_node_memory,
1307 	    &lgrp_plat_lat_stats);
1308 	if (lgrp_plat_slit_error == 0)
1309 		return;
1310 
1311 	/*
1312 	 * Probe to determine latency between NUMA nodes when SLIT
1313 	 * doesn't exist or make sense
1314 	 */
1315 	lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_ENABLE;
1316 
1317 	/*
1318 	 * Specify whether to probe using vendor ID register or page copy
1319 	 * if hasn't been specified already or is overspecified
1320 	 */
1321 	probe_op = lgrp_plat_probe_flags &
1322 	    (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
1323 
1324 	if (probe_op == 0 ||
1325 	    probe_op == (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR)) {
1326 		lgrp_plat_probe_flags &=
1327 		    ~(LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
1328 		if (is_opteron())
1329 			lgrp_plat_probe_flags |=
1330 			    LGRP_PLAT_PROBE_VENDOR;
1331 		else
1332 			lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_PGCPY;
1333 	}
1334 
1335 	/*
1336 	 * Probing errors can mess up the lgroup topology and
1337 	 * force us fall back to a 2 level lgroup topology.
1338 	 * Here we bound how tall the lgroup topology can grow
1339 	 * in hopes of avoiding any anamolies in probing from
1340 	 * messing up the lgroup topology by limiting the
1341 	 * accuracy of the latency topology.
1342 	 *
1343 	 * Assume that nodes will at least be configured in a
1344 	 * ring, so limit height of lgroup topology to be less
1345 	 * than number of nodes on a system with 4 or more
1346 	 * nodes
1347 	 */
1348 	if (lgrp_plat_node_cnt >= 4 && lgrp_topo_ht_limit() ==
1349 	    lgrp_topo_ht_limit_default())
1350 		(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
1351 }
1352 
1353 
1354 /*
1355  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
1356  * be considered same
1357  */
1358 #define	LGRP_LAT_TOLERANCE_SHIFT	4
1359 
1360 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
1361 
1362 
1363 /*
1364  * Adjust latencies between nodes to be symmetric, normalize latencies between
1365  * any nodes that are within some tolerance to be same, and make local
1366  * latencies be same
1367  */
1368 static void
1369 lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
1370     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1371 {
1372 	int				i;
1373 	int				j;
1374 	int				k;
1375 	int				l;
1376 	u_longlong_t			max;
1377 	u_longlong_t			min;
1378 	u_longlong_t			t;
1379 	u_longlong_t			t1;
1380 	u_longlong_t			t2;
1381 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
1382 	int				lat_corrected[MAX_NODES][MAX_NODES];
1383 
1384 	/*
1385 	 * Nothing to do when this is an UMA machine or don't have args needed
1386 	 */
1387 	if (max_mem_nodes == 1)
1388 		return;
1389 
1390 	ASSERT(node_memory != NULL && lat_stats != NULL &&
1391 	    probe_stats != NULL);
1392 
1393 	/*
1394 	 * Make sure that latencies are symmetric between any two nodes
1395 	 * (ie. latency(node0, node1) == latency(node1, node0))
1396 	 */
1397 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1398 		if (!node_memory[i].exists)
1399 			continue;
1400 
1401 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1402 			if (!node_memory[j].exists)
1403 				continue;
1404 
1405 			t1 = lat_stats->latencies[i][j];
1406 			t2 = lat_stats->latencies[j][i];
1407 
1408 			if (t1 == 0 || t2 == 0 || t1 == t2)
1409 				continue;
1410 
1411 			/*
1412 			 * Latencies should be same
1413 			 * - Use minimum of two latencies which should be same
1414 			 * - Track suspect probe times not within tolerance of
1415 			 *   min value
1416 			 * - Remember how much values are corrected by
1417 			 */
1418 			if (t1 > t2) {
1419 				t = t2;
1420 				probe_stats->probe_errors[i][j] += t1 - t2;
1421 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
1422 					probe_stats->probe_suspect[i][j]++;
1423 					probe_stats->probe_suspect[j][i]++;
1424 				}
1425 			} else if (t2 > t1) {
1426 				t = t1;
1427 				probe_stats->probe_errors[j][i] += t2 - t1;
1428 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
1429 					probe_stats->probe_suspect[i][j]++;
1430 					probe_stats->probe_suspect[j][i]++;
1431 				}
1432 			}
1433 
1434 			lat_stats->latencies[i][j] =
1435 			    lat_stats->latencies[j][i] = t;
1436 			lgrp_config(cflag, t1, t);
1437 			lgrp_config(cflag, t2, t);
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * Keep track of which latencies get corrected
1443 	 */
1444 	for (i = 0; i < MAX_NODES; i++)
1445 		for (j = 0; j < MAX_NODES; j++)
1446 			lat_corrected[i][j] = 0;
1447 
1448 	/*
1449 	 * For every two nodes, see whether there is another pair of nodes which
1450 	 * are about the same distance apart and make the latencies be the same
1451 	 * if they are close enough together
1452 	 */
1453 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1454 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1455 			if (!node_memory[j].exists)
1456 				continue;
1457 			/*
1458 			 * Pick one pair of nodes (i, j)
1459 			 * and get latency between them
1460 			 */
1461 			t1 = lat_stats->latencies[i][j];
1462 
1463 			/*
1464 			 * Skip this pair of nodes if there isn't a latency
1465 			 * for it yet
1466 			 */
1467 			if (t1 == 0)
1468 				continue;
1469 
1470 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1471 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
1472 					if (!node_memory[l].exists)
1473 						continue;
1474 					/*
1475 					 * Pick another pair of nodes (k, l)
1476 					 * not same as (i, j) and get latency
1477 					 * between them
1478 					 */
1479 					if (k == i && l == j)
1480 						continue;
1481 
1482 					t2 = lat_stats->latencies[k][l];
1483 
1484 					/*
1485 					 * Skip this pair of nodes if there
1486 					 * isn't a latency for it yet
1487 					 */
1488 
1489 					if (t2 == 0)
1490 						continue;
1491 
1492 					/*
1493 					 * Skip nodes (k, l) if they already
1494 					 * have same latency as (i, j) or
1495 					 * their latency isn't close enough to
1496 					 * be considered/made the same
1497 					 */
1498 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
1499 					    t1 >> lgrp_plat_probe_lt_shift) ||
1500 					    (t2 > t1 && t2 - t1 >
1501 					    t2 >> lgrp_plat_probe_lt_shift))
1502 						continue;
1503 
1504 					/*
1505 					 * Make latency(i, j) same as
1506 					 * latency(k, l), try to use latency
1507 					 * that has been adjusted already to get
1508 					 * more consistency (if possible), and
1509 					 * remember which latencies were
1510 					 * adjusted for next time
1511 					 */
1512 					if (lat_corrected[i][j]) {
1513 						t = t1;
1514 						lgrp_config(cflag, t2, t);
1515 						t2 = t;
1516 					} else if (lat_corrected[k][l]) {
1517 						t = t2;
1518 						lgrp_config(cflag, t1, t);
1519 						t1 = t;
1520 					} else {
1521 						if (t1 > t2)
1522 							t = t2;
1523 						else
1524 							t = t1;
1525 						lgrp_config(cflag, t1, t);
1526 						lgrp_config(cflag, t2, t);
1527 						t1 = t2 = t;
1528 					}
1529 
1530 					lat_stats->latencies[i][j] =
1531 					    lat_stats->latencies[k][l] = t;
1532 
1533 					lat_corrected[i][j] =
1534 					    lat_corrected[k][l] = 1;
1535 				}
1536 			}
1537 		}
1538 	}
1539 
1540 	/*
1541 	 * Local latencies should be same
1542 	 * - Find min and max local latencies
1543 	 * - Make all local latencies be minimum
1544 	 */
1545 	min = -1;
1546 	max = 0;
1547 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1548 		if (!node_memory[i].exists)
1549 			continue;
1550 		t = lat_stats->latencies[i][i];
1551 		if (t == 0)
1552 			continue;
1553 		if (min == -1 || t < min)
1554 			min = t;
1555 		if (t > max)
1556 			max = t;
1557 	}
1558 	if (min != max) {
1559 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1560 			int	local;
1561 
1562 			if (!node_memory[i].exists)
1563 				continue;
1564 
1565 			local = lat_stats->latencies[i][i];
1566 			if (local == 0)
1567 				continue;
1568 
1569 			/*
1570 			 * Track suspect probe times that aren't within
1571 			 * tolerance of minimum local latency and how much
1572 			 * probe times are corrected by
1573 			 */
1574 			if (local - min > min >> lgrp_plat_probe_lt_shift)
1575 				probe_stats->probe_suspect[i][i]++;
1576 
1577 			probe_stats->probe_errors[i][i] += local - min;
1578 
1579 			/*
1580 			 * Make local latencies be minimum
1581 			 */
1582 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
1583 			lat_stats->latencies[i][i] = min;
1584 		}
1585 	}
1586 
1587 	/*
1588 	 * Determine max probe time again since just adjusted latencies
1589 	 */
1590 	lat_stats->latency_max = 0;
1591 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1592 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1593 			if (!node_memory[j].exists)
1594 				continue;
1595 			t = lat_stats->latencies[i][j];
1596 			if (t > lat_stats->latency_max)
1597 				lat_stats->latency_max = t;
1598 		}
1599 	}
1600 }
1601 
1602 
1603 /*
1604  * Verify following about latencies between nodes:
1605  *
1606  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
1607  * - Local latencies same
1608  * - Local < remote
1609  * - Number of latencies seen is reasonable
1610  * - Number of occurrences of a given latency should be more than 1
1611  *
1612  * Returns:
1613  *	0	Success
1614  *	-1	Not symmetric
1615  *	-2	Local latencies not same
1616  *	-3	Local >= remote
1617  */
1618 static int
1619 lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
1620     lgrp_plat_latency_stats_t *lat_stats)
1621 {
1622 	int				i;
1623 	int				j;
1624 	u_longlong_t			t1;
1625 	u_longlong_t			t2;
1626 
1627 	ASSERT(node_memory != NULL && lat_stats != NULL);
1628 
1629 	/*
1630 	 * Nothing to do when this is an UMA machine, lgroup topology is
1631 	 * limited to 2 levels, or there aren't any probe times yet
1632 	 */
1633 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
1634 	    lat_stats->latencies[0][0] == 0)
1635 		return (0);
1636 
1637 	/*
1638 	 * Make sure that latencies are symmetric between any two nodes
1639 	 * (ie. latency(node0, node1) == latency(node1, node0))
1640 	 */
1641 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1642 		if (!node_memory[i].exists)
1643 			continue;
1644 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1645 			if (!node_memory[j].exists)
1646 				continue;
1647 			t1 = lat_stats->latencies[i][j];
1648 			t2 = lat_stats->latencies[j][i];
1649 
1650 			if (t1 == 0 || t2 == 0 || t1 == t2)
1651 				continue;
1652 
1653 			return (-1);
1654 		}
1655 	}
1656 
1657 	/*
1658 	 * Local latencies should be same
1659 	 */
1660 	t1 = lat_stats->latencies[0][0];
1661 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
1662 		if (!node_memory[i].exists)
1663 			continue;
1664 
1665 		t2 = lat_stats->latencies[i][i];
1666 		if (t2 == 0)
1667 			continue;
1668 
1669 		if (t1 == 0) {
1670 			t1 = t2;
1671 			continue;
1672 		}
1673 
1674 		if (t1 != t2)
1675 			return (-2);
1676 	}
1677 
1678 	/*
1679 	 * Local latencies should be less than remote
1680 	 */
1681 	if (t1) {
1682 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1683 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1684 				if (!node_memory[j].exists)
1685 					continue;
1686 				t2 = lat_stats->latencies[i][j];
1687 				if (i == j || t2 == 0)
1688 					continue;
1689 
1690 				if (t1 >= t2)
1691 					return (-3);
1692 			}
1693 		}
1694 	}
1695 
1696 	return (0);
1697 }
1698 
1699 
1700 /*
1701  * Platform-specific initialization
1702  */
1703 static void
1704 lgrp_plat_main_init(void)
1705 {
1706 	int	curnode;
1707 	int	ht_limit;
1708 	int	i;
1709 
1710 	/*
1711 	 * Print a notice that MPO is disabled when memory is interleaved
1712 	 * across nodes....Would do this when it is discovered, but can't
1713 	 * because it happens way too early during boot....
1714 	 */
1715 	if (lgrp_plat_mem_intrlv)
1716 		cmn_err(CE_NOTE,
1717 		    "MPO disabled because memory is interleaved\n");
1718 
1719 	/*
1720 	 * Don't bother to do any probing if it is disabled, there is only one
1721 	 * node, or the height of the lgroup topology less than or equal to 2
1722 	 */
1723 	ht_limit = lgrp_topo_ht_limit();
1724 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
1725 	    max_mem_nodes == 1 || ht_limit <= 2) {
1726 		/*
1727 		 * Setup lgroup latencies for 2 level lgroup topology
1728 		 * (ie. local and remote only) if they haven't been set yet
1729 		 */
1730 		if (ht_limit == 2 && lgrp_plat_lat_stats.latency_min == -1 &&
1731 		    lgrp_plat_lat_stats.latency_max == 0)
1732 			lgrp_plat_2level_setup(lgrp_plat_node_memory,
1733 			    &lgrp_plat_lat_stats);
1734 		return;
1735 	}
1736 
1737 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
1738 		/*
1739 		 * Should have been able to probe from CPU 0 when it was added
1740 		 * to lgroup hierarchy, but may not have been able to then
1741 		 * because it happens so early in boot that gethrtime() hasn't
1742 		 * been initialized.  (:-(
1743 		 */
1744 		curnode = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
1745 		    lgrp_plat_cpu_node_nentries);
1746 		ASSERT(curnode >= 0 && curnode < lgrp_plat_node_cnt);
1747 		if (lgrp_plat_lat_stats.latencies[curnode][curnode] == 0)
1748 			lgrp_plat_probe();
1749 
1750 		return;
1751 	}
1752 
1753 	/*
1754 	 * When probing memory, use one page for every sample to determine
1755 	 * lgroup topology and taking multiple samples
1756 	 */
1757 	if (lgrp_plat_probe_mem_config.probe_memsize == 0)
1758 		lgrp_plat_probe_mem_config.probe_memsize = PAGESIZE *
1759 		    lgrp_plat_probe_nsamples;
1760 
1761 	/*
1762 	 * Map memory in each node needed for probing to determine latency
1763 	 * topology
1764 	 */
1765 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1766 		int	mnode;
1767 
1768 		/*
1769 		 * Skip this node and leave its probe page NULL
1770 		 * if it doesn't have any memory
1771 		 */
1772 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1773 		if (!mem_node_config[mnode].exists) {
1774 			lgrp_plat_probe_mem_config.probe_va[i] = NULL;
1775 			continue;
1776 		}
1777 
1778 		/*
1779 		 * Allocate one kernel virtual page
1780 		 */
1781 		lgrp_plat_probe_mem_config.probe_va[i] = vmem_alloc(heap_arena,
1782 		    lgrp_plat_probe_mem_config.probe_memsize, VM_NOSLEEP);
1783 		if (lgrp_plat_probe_mem_config.probe_va[i] == NULL) {
1784 			cmn_err(CE_WARN,
1785 			    "lgrp_plat_main_init: couldn't allocate memory");
1786 			return;
1787 		}
1788 
1789 		/*
1790 		 * Get PFN for first page in each node
1791 		 */
1792 		lgrp_plat_probe_mem_config.probe_pfn[i] =
1793 		    mem_node_config[mnode].physbase;
1794 
1795 		/*
1796 		 * Map virtual page to first page in node
1797 		 */
1798 		hat_devload(kas.a_hat, lgrp_plat_probe_mem_config.probe_va[i],
1799 		    lgrp_plat_probe_mem_config.probe_memsize,
1800 		    lgrp_plat_probe_mem_config.probe_pfn[i],
1801 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1802 		    HAT_LOAD_NOCONSIST);
1803 	}
1804 
1805 	/*
1806 	 * Probe from current CPU
1807 	 */
1808 	lgrp_plat_probe();
1809 }
1810 
1811 
1812 /*
1813  * Return the number of free, allocatable, or installed
1814  * pages in an lgroup
1815  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1816  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1817  */
1818 /* ARGSUSED */
1819 static pgcnt_t
1820 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1821 {
1822 	struct memlist *mlist;
1823 	pgcnt_t npgs = 0;
1824 	extern struct memlist *phys_avail;
1825 	extern struct memlist *phys_install;
1826 
1827 	switch (query) {
1828 	case LGRP_MEM_SIZE_FREE:
1829 		return ((pgcnt_t)freemem);
1830 	case LGRP_MEM_SIZE_AVAIL:
1831 		memlist_read_lock();
1832 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1833 			npgs += btop(mlist->size);
1834 		memlist_read_unlock();
1835 		return (npgs);
1836 	case LGRP_MEM_SIZE_INSTALL:
1837 		memlist_read_lock();
1838 		for (mlist = phys_install; mlist; mlist = mlist->next)
1839 			npgs += btop(mlist->size);
1840 		memlist_read_unlock();
1841 		return (npgs);
1842 	default:
1843 		return ((pgcnt_t)0);
1844 	}
1845 }
1846 
1847 
1848 /*
1849  * Update node to proximity domain mappings for given domain and return node ID
1850  */
1851 static int
1852 lgrp_plat_node_domain_update(node_domain_map_t *node_domain, int node_cnt,
1853     uint32_t domain)
1854 {
1855 	uint_t	node;
1856 	uint_t	start;
1857 
1858 	/*
1859 	 * Hash proximity domain ID into node to domain mapping table (array)
1860 	 * and add entry for it into first non-existent or matching entry found
1861 	 */
1862 	node = start = NODE_DOMAIN_HASH(domain, node_cnt);
1863 	do {
1864 		/*
1865 		 * Entry doesn't exist yet, so create one for this proximity
1866 		 * domain and return node ID which is index into mapping table.
1867 		 */
1868 		if (!node_domain[node].exists) {
1869 			node_domain[node].exists = 1;
1870 			node_domain[node].prox_domain = domain;
1871 			return (node);
1872 		}
1873 
1874 		/*
1875 		 * Entry exists for this proximity domain already, so just
1876 		 * return node ID (index into table).
1877 		 */
1878 		if (node_domain[node].prox_domain == domain)
1879 			return (node);
1880 		node = NODE_DOMAIN_HASH(node + 1, node_cnt);
1881 	} while (node != start);
1882 
1883 	/*
1884 	 * Ran out of supported number of entries which shouldn't happen....
1885 	 */
1886 	ASSERT(node != start);
1887 	return (-1);
1888 }
1889 
1890 
1891 /*
1892  * Update node memory information for given proximity domain with specified
1893  * starting and ending physical address range (and return positive numbers for
1894  * success and negative ones for errors)
1895  */
1896 static int
1897 lgrp_plat_node_memory_update(node_domain_map_t *node_domain, int node_cnt,
1898     node_phys_addr_map_t *node_memory, uint64_t start, uint64_t end,
1899     uint32_t domain)
1900 {
1901 	int	node;
1902 
1903 	/*
1904 	 * Get node number for proximity domain
1905 	 */
1906 	node = lgrp_plat_domain_to_node(node_domain, node_cnt, domain);
1907 	if (node == -1) {
1908 		node = lgrp_plat_node_domain_update(node_domain, node_cnt,
1909 		    domain);
1910 		if (node == -1)
1911 			return (-1);
1912 	}
1913 
1914 	/*
1915 	 * Create entry in table for node if it doesn't exist
1916 	 */
1917 	if (!node_memory[node].exists) {
1918 		node_memory[node].exists = 1;
1919 		node_memory[node].start = btop(start);
1920 		node_memory[node].end = btop(end);
1921 		node_memory[node].prox_domain = domain;
1922 		return (0);
1923 	}
1924 
1925 	/*
1926 	 * Entry already exists for this proximity domain
1927 	 *
1928 	 * There may be more than one SRAT memory entry for a domain, so we may
1929 	 * need to update existing start or end address for the node.
1930 	 */
1931 	if (node_memory[node].prox_domain == domain) {
1932 		if (btop(start) < node_memory[node].start)
1933 			node_memory[node].start = btop(start);
1934 		if (btop(end) > node_memory[node].end)
1935 			node_memory[node].end = btop(end);
1936 		return (1);
1937 	}
1938 	return (-2);
1939 }
1940 
1941 
1942 /*
1943  * Have to sort node by starting physical address because VM system (physical
1944  * page free list management) assumes and expects memnodes to be sorted in
1945  * ascending order by physical address.  If not, the kernel will panic in
1946  * potentially a number of different places.  (:-(
1947  * NOTE: This workaround will not be sufficient if/when hotplugging memory is
1948  *	 supported on x86/x64.
1949  */
1950 static void
1951 lgrp_plat_node_sort(node_domain_map_t *node_domain, int node_cnt,
1952     cpu_node_map_t *cpu_node, int cpu_count, node_phys_addr_map_t *node_memory)
1953 {
1954 	boolean_t	found;
1955 	int		i;
1956 	int		j;
1957 	int		n;
1958 	boolean_t	sorted;
1959 	boolean_t	swapped;
1960 
1961 	if (!lgrp_plat_node_sort_enable || node_cnt <= 1 ||
1962 	    node_domain == NULL || node_memory == NULL)
1963 		return;
1964 
1965 	/*
1966 	 * Sorted already?
1967 	 */
1968 	sorted = B_TRUE;
1969 	for (i = 0; i < node_cnt - 1; i++) {
1970 		/*
1971 		 * Skip entries that don't exist
1972 		 */
1973 		if (!node_memory[i].exists)
1974 			continue;
1975 
1976 		/*
1977 		 * Try to find next existing entry to compare against
1978 		 */
1979 		found = B_FALSE;
1980 		for (j = i + 1; j < node_cnt; j++) {
1981 			if (node_memory[j].exists) {
1982 				found = B_TRUE;
1983 				break;
1984 			}
1985 		}
1986 
1987 		/*
1988 		 * Done if no more existing entries to compare against
1989 		 */
1990 		if (found == B_FALSE)
1991 			break;
1992 
1993 		/*
1994 		 * Not sorted if starting address of current entry is bigger
1995 		 * than starting address of next existing entry
1996 		 */
1997 		if (node_memory[i].start > node_memory[j].start) {
1998 			sorted = B_FALSE;
1999 			break;
2000 		}
2001 	}
2002 
2003 	/*
2004 	 * Don't need to sort if sorted already
2005 	 */
2006 	if (sorted == B_TRUE)
2007 		return;
2008 
2009 	/*
2010 	 * Just use bubble sort since number of nodes is small
2011 	 */
2012 	n = node_cnt;
2013 	do {
2014 		swapped = B_FALSE;
2015 		n--;
2016 		for (i = 0; i < n; i++) {
2017 			/*
2018 			 * Skip entries that don't exist
2019 			 */
2020 			if (!node_memory[i].exists)
2021 				continue;
2022 
2023 			/*
2024 			 * Try to find next existing entry to compare against
2025 			 */
2026 			found = B_FALSE;
2027 			for (j = i + 1; j <= n; j++) {
2028 				if (node_memory[j].exists) {
2029 					found = B_TRUE;
2030 					break;
2031 				}
2032 			}
2033 
2034 			/*
2035 			 * Done if no more existing entries to compare against
2036 			 */
2037 			if (found == B_FALSE)
2038 				break;
2039 
2040 			if (node_memory[i].start > node_memory[j].start) {
2041 				node_phys_addr_map_t	save_addr;
2042 				node_domain_map_t	save_node;
2043 
2044 				/*
2045 				 * Swap node to proxmity domain ID assignments
2046 				 */
2047 				bcopy(&node_domain[i], &save_node,
2048 				    sizeof (node_domain_map_t));
2049 				bcopy(&node_domain[j], &node_domain[i],
2050 				    sizeof (node_domain_map_t));
2051 				bcopy(&save_node, &node_domain[j],
2052 				    sizeof (node_domain_map_t));
2053 
2054 				/*
2055 				 * Swap node to physical memory assignments
2056 				 */
2057 				bcopy(&node_memory[i], &save_addr,
2058 				    sizeof (node_phys_addr_map_t));
2059 				bcopy(&node_memory[j], &node_memory[i],
2060 				    sizeof (node_phys_addr_map_t));
2061 				bcopy(&save_addr, &node_memory[j],
2062 				    sizeof (node_phys_addr_map_t));
2063 				swapped = B_TRUE;
2064 			}
2065 		}
2066 	} while (swapped == B_TRUE);
2067 
2068 	/*
2069 	 * Check to make sure that CPUs assigned to correct node IDs now since
2070 	 * node to proximity domain ID assignments may have been changed above
2071 	 */
2072 	if (n == node_cnt - 1 || cpu_node == NULL || cpu_count < 1)
2073 		return;
2074 	for (i = 0; i < cpu_count; i++) {
2075 		int		node;
2076 
2077 		node = lgrp_plat_domain_to_node(node_domain, node_cnt,
2078 		    cpu_node[i].prox_domain);
2079 		if (cpu_node[i].node != node)
2080 			cpu_node[i].node = node;
2081 	}
2082 
2083 }
2084 
2085 
2086 /*
2087  * Return time needed to probe from current CPU to memory in given node
2088  */
2089 static hrtime_t
2090 lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node, int cpu_node_nentries,
2091     lgrp_plat_probe_mem_config_t *probe_mem_config,
2092     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
2093 {
2094 	caddr_t			buf;
2095 	hrtime_t		elapsed;
2096 	hrtime_t		end;
2097 	int			from;
2098 	int			i;
2099 	int			ipl;
2100 	hrtime_t		max;
2101 	hrtime_t		min;
2102 	hrtime_t		start;
2103 	extern int		use_sse_pagecopy;
2104 
2105 	/*
2106 	 * Determine ID of node containing current CPU
2107 	 */
2108 	from = lgrp_plat_cpu_to_node(CPU, cpu_node, cpu_node_nentries);
2109 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
2110 
2111 	/*
2112 	 * Do common work for probing main memory
2113 	 */
2114 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_PGCPY) {
2115 		/*
2116 		 * Skip probing any nodes without memory and
2117 		 * set probe time to 0
2118 		 */
2119 		if (probe_mem_config->probe_va[to] == NULL) {
2120 			lat_stats->latencies[from][to] = 0;
2121 			return (0);
2122 		}
2123 
2124 		/*
2125 		 * Invalidate caches once instead of once every sample
2126 		 * which should cut cost of probing by a lot
2127 		 */
2128 		probe_stats->flush_cost = gethrtime();
2129 		invalidate_cache();
2130 		probe_stats->flush_cost = gethrtime() -
2131 		    probe_stats->flush_cost;
2132 		probe_stats->probe_cost_total += probe_stats->flush_cost;
2133 	}
2134 
2135 	/*
2136 	 * Probe from current CPU to given memory using specified operation
2137 	 * and take specified number of samples
2138 	 */
2139 	max = 0;
2140 	min = -1;
2141 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
2142 		probe_stats->probe_cost = gethrtime();
2143 
2144 		/*
2145 		 * Can't measure probe time if gethrtime() isn't working yet
2146 		 */
2147 		if (probe_stats->probe_cost == 0 && gethrtime() == 0)
2148 			return (0);
2149 
2150 		if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
2151 			/*
2152 			 * Measure how long it takes to read vendor ID from
2153 			 * Northbridge
2154 			 */
2155 			elapsed = opt_probe_vendor(to, lgrp_plat_probe_nreads);
2156 		} else {
2157 			/*
2158 			 * Measure how long it takes to copy page
2159 			 * on top of itself
2160 			 */
2161 			buf = probe_mem_config->probe_va[to] + (i * PAGESIZE);
2162 
2163 			kpreempt_disable();
2164 			ipl = splhigh();
2165 			start = gethrtime();
2166 			if (use_sse_pagecopy)
2167 				hwblkpagecopy(buf, buf);
2168 			else
2169 				bcopy(buf, buf, PAGESIZE);
2170 			end = gethrtime();
2171 			elapsed = end - start;
2172 			splx(ipl);
2173 			kpreempt_enable();
2174 		}
2175 
2176 		probe_stats->probe_cost = gethrtime() -
2177 		    probe_stats->probe_cost;
2178 		probe_stats->probe_cost_total += probe_stats->probe_cost;
2179 
2180 		if (min == -1 || elapsed < min)
2181 			min = elapsed;
2182 		if (elapsed > max)
2183 			max = elapsed;
2184 	}
2185 
2186 	/*
2187 	 * Update minimum and maximum probe times between
2188 	 * these two nodes
2189 	 */
2190 	if (min < probe_stats->probe_min[from][to] ||
2191 	    probe_stats->probe_min[from][to] == 0)
2192 		probe_stats->probe_min[from][to] = min;
2193 
2194 	if (max > probe_stats->probe_max[from][to])
2195 		probe_stats->probe_max[from][to] = max;
2196 
2197 	return (min);
2198 }
2199 
2200 
2201 /*
2202  * Read boot property with CPU to APIC ID array, fill in CPU to node ID
2203  * mapping table with APIC ID for each CPU (if pointer to table isn't NULL),
2204  * and return number of CPU APIC IDs.
2205  *
2206  * NOTE: This code assumes that CPU IDs are assigned in order that they appear
2207  *       in in cpu_apicid_array boot property which is based on and follows
2208  *	 same ordering as processor list in ACPI MADT.  If the code in
2209  *	 usr/src/uts/i86pc/io/pcplusmp/apic.c that reads MADT and assigns
2210  *	 CPU IDs ever changes, then this code will need to change too....
2211  */
2212 static int
2213 lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node)
2214 {
2215 	int	boot_prop_len;
2216 	char	*boot_prop_name = BP_CPU_APICID_ARRAY;
2217 	uint8_t	cpu_apicid_array[UINT8_MAX + 1];
2218 	int	i;
2219 	int	n;
2220 
2221 	/*
2222 	 * Check length of property value
2223 	 */
2224 	boot_prop_len = BOP_GETPROPLEN(bootops, boot_prop_name);
2225 	if (boot_prop_len <= 0 || boot_prop_len > sizeof (cpu_apicid_array))
2226 		return (-1);
2227 
2228 	/*
2229 	 * Calculate number of entries in array and return when there's just
2230 	 * one CPU since that's not very interesting for NUMA
2231 	 */
2232 	n = boot_prop_len / sizeof (uint8_t);
2233 	if (n == 1)
2234 		return (-2);
2235 
2236 	/*
2237 	 * Get CPU to APIC ID property value
2238 	 */
2239 	if (BOP_GETPROP(bootops, boot_prop_name, cpu_apicid_array) < 0)
2240 		return (-3);
2241 
2242 	/*
2243 	 * Just return number of CPU APIC IDs if CPU to node mapping table is
2244 	 * NULL
2245 	 */
2246 	if (cpu_node == NULL)
2247 		return (n);
2248 
2249 	/*
2250 	 * Fill in CPU to node ID mapping table with APIC ID for each CPU
2251 	 */
2252 	for (i = 0; i < n; i++) {
2253 		cpu_node[i].exists = 1;
2254 		cpu_node[i].apicid = cpu_apicid_array[i];
2255 	}
2256 
2257 	/*
2258 	 * Return number of CPUs based on number of APIC IDs
2259 	 */
2260 	return (n);
2261 }
2262 
2263 
2264 /*
2265  * Read ACPI System Locality Information Table (SLIT) to determine how far each
2266  * NUMA node is from each other
2267  */
2268 static int
2269 lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
2270     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats)
2271 {
2272 	int		i;
2273 	int		j;
2274 	int		localities;
2275 	hrtime_t	max;
2276 	hrtime_t	min;
2277 	int		retval;
2278 	uint8_t		*slit_entries;
2279 
2280 	if (tp == NULL || !lgrp_plat_slit_enable)
2281 		return (1);
2282 
2283 	if (lat_stats == NULL)
2284 		return (2);
2285 
2286 	localities = tp->number;
2287 	if (localities != node_cnt)
2288 		return (3);
2289 
2290 	min = lat_stats->latency_min;
2291 	max = lat_stats->latency_max;
2292 
2293 	/*
2294 	 * Fill in latency matrix based on SLIT entries
2295 	 */
2296 	slit_entries = tp->entry;
2297 	for (i = 0; i < localities; i++) {
2298 		for (j = 0; j < localities; j++) {
2299 			uint8_t	latency;
2300 
2301 			latency = slit_entries[(i * localities) + j];
2302 			lat_stats->latencies[i][j] = latency;
2303 			if (latency < min || min == -1)
2304 				min = latency;
2305 			if (latency > max)
2306 				max = latency;
2307 		}
2308 	}
2309 
2310 	/*
2311 	 * Verify that latencies/distances given in SLIT look reasonable
2312 	 */
2313 	retval = lgrp_plat_latency_verify(node_memory, lat_stats);
2314 
2315 	if (retval) {
2316 		/*
2317 		 * Reinitialize (zero) latency table since SLIT doesn't look
2318 		 * right
2319 		 */
2320 		for (i = 0; i < localities; i++) {
2321 			for (j = 0; j < localities; j++)
2322 				lat_stats->latencies[i][j] = 0;
2323 		}
2324 	} else {
2325 		/*
2326 		 * Update min and max latencies seen since SLIT looks valid
2327 		 */
2328 		lat_stats->latency_min = min;
2329 		lat_stats->latency_max = max;
2330 	}
2331 
2332 	return (retval);
2333 }
2334 
2335 
2336 /*
2337  * Read ACPI System Resource Affinity Table (SRAT) to determine which CPUs
2338  * and memory are local to each other in the same NUMA node and return number
2339  * of nodes
2340  */
2341 static int
2342 lgrp_plat_process_srat(struct srat *tp, uint32_t *prox_domain_min,
2343     node_domain_map_t *node_domain, cpu_node_map_t *cpu_node, int cpu_count,
2344     node_phys_addr_map_t *node_memory)
2345 {
2346 	struct srat_item	*srat_end;
2347 	int			i;
2348 	struct srat_item	*item;
2349 	int			node_cnt;
2350 	int			proc_entry_count;
2351 
2352 	/*
2353 	 * Nothing to do when no SRAT or disabled
2354 	 */
2355 	if (tp == NULL || !lgrp_plat_srat_enable)
2356 		return (-1);
2357 
2358 	/*
2359 	 * Determine number of nodes by counting number of proximity domains in
2360 	 * SRAT and return if number of nodes is 1 or less since don't need to
2361 	 * read SRAT then
2362 	 */
2363 	node_cnt = lgrp_plat_srat_domains(tp, prox_domain_min);
2364 	if (node_cnt == 1)
2365 		return (1);
2366 	else if (node_cnt <= 0)
2367 		return (-2);
2368 
2369 	/*
2370 	 * Walk through SRAT, examining each CPU and memory entry to determine
2371 	 * which CPUs and memory belong to which node.
2372 	 */
2373 	item = tp->list;
2374 	srat_end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2375 	proc_entry_count = 0;
2376 	while (item < srat_end) {
2377 		uint32_t	apic_id;
2378 		uint32_t	domain;
2379 		uint64_t	end;
2380 		uint64_t	length;
2381 		uint64_t	start;
2382 
2383 		switch (item->type) {
2384 		case SRAT_PROCESSOR:	/* CPU entry */
2385 			if (!(item->i.p.flags & SRAT_ENABLED) ||
2386 			    cpu_node == NULL)
2387 				break;
2388 
2389 			/*
2390 			 * Calculate domain (node) ID and fill in APIC ID to
2391 			 * domain/node mapping table
2392 			 */
2393 			domain = item->i.p.domain1;
2394 			for (i = 0; i < 3; i++) {
2395 				domain += item->i.p.domain2[i] <<
2396 				    ((i + 1) * 8);
2397 			}
2398 			apic_id = item->i.p.apic_id;
2399 
2400 			if (lgrp_plat_cpu_node_update(node_domain, node_cnt,
2401 			    cpu_node, cpu_count, apic_id, domain) < 0)
2402 				return (-3);
2403 
2404 			proc_entry_count++;
2405 			break;
2406 
2407 		case SRAT_MEMORY:	/* memory entry */
2408 			if (!(item->i.m.flags & SRAT_ENABLED) ||
2409 			    node_memory == NULL)
2410 				break;
2411 
2412 			/*
2413 			 * Get domain (node) ID and fill in domain/node
2414 			 * to memory mapping table
2415 			 */
2416 			domain = item->i.m.domain;
2417 			start = item->i.m.base_addr;
2418 			length = item->i.m.len;
2419 			end = start + length - 1;
2420 
2421 			if (lgrp_plat_node_memory_update(node_domain, node_cnt,
2422 			    node_memory, start, end, domain) < 0)
2423 				return (-4);
2424 			break;
2425 		case SRAT_X2APIC:	/* x2apic CPU entry */
2426 			if (!(item->i.xp.flags & SRAT_ENABLED) ||
2427 			    cpu_node == NULL)
2428 				break;
2429 
2430 			/*
2431 			 * Calculate domain (node) ID and fill in APIC ID to
2432 			 * domain/node mapping table
2433 			 */
2434 			domain = item->i.xp.domain;
2435 			apic_id = item->i.xp.x2apic_id;
2436 
2437 			if (lgrp_plat_cpu_node_update(node_domain, node_cnt,
2438 			    cpu_node, cpu_count, apic_id, domain) < 0)
2439 				return (-3);
2440 
2441 			proc_entry_count++;
2442 			break;
2443 
2444 		default:
2445 			break;
2446 		}
2447 
2448 		item = (struct srat_item *)((uintptr_t)item + item->len);
2449 	}
2450 
2451 	/*
2452 	 * Should have seen at least as many SRAT processor entries as CPUs
2453 	 */
2454 	if (proc_entry_count < cpu_count)
2455 		return (-5);
2456 
2457 	/*
2458 	 * Need to sort nodes by starting physical address since VM system
2459 	 * assumes and expects memnodes to be sorted in ascending order by
2460 	 * physical address
2461 	 */
2462 	lgrp_plat_node_sort(node_domain, node_cnt, cpu_node, cpu_count,
2463 	    node_memory);
2464 
2465 	return (node_cnt);
2466 }
2467 
2468 
2469 /*
2470  * Allocate permanent memory for any temporary memory that we needed to
2471  * allocate using BOP_ALLOC() before kmem_alloc() and VM system were
2472  * initialized and copy everything from temporary to permanent memory since
2473  * temporary boot memory will eventually be released during boot
2474  */
2475 static void
2476 lgrp_plat_release_bootstrap(void)
2477 {
2478 	void	*buf;
2479 	size_t	size;
2480 
2481 	if (lgrp_plat_cpu_node_nentries > 0) {
2482 		size = lgrp_plat_cpu_node_nentries * sizeof (cpu_node_map_t);
2483 		buf = kmem_alloc(size, KM_SLEEP);
2484 		bcopy(lgrp_plat_cpu_node, buf, size);
2485 		lgrp_plat_cpu_node = buf;
2486 	}
2487 }
2488 
2489 
2490 /*
2491  * Return number of proximity domains given in ACPI SRAT
2492  */
2493 static int
2494 lgrp_plat_srat_domains(struct srat *tp, uint32_t *prox_domain_min)
2495 {
2496 	int			domain_cnt;
2497 	uint32_t		domain_min;
2498 	struct srat_item	*end;
2499 	int			i;
2500 	struct srat_item	*item;
2501 	node_domain_map_t	node_domain[MAX_NODES];
2502 
2503 
2504 	if (tp == NULL || !lgrp_plat_srat_enable)
2505 		return (1);
2506 
2507 	/*
2508 	 * Walk through SRAT to find minimum proximity domain ID
2509 	 */
2510 	domain_min = UINT32_MAX;
2511 	item = tp->list;
2512 	end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2513 	while (item < end) {
2514 		uint32_t	domain;
2515 
2516 		switch (item->type) {
2517 		case SRAT_PROCESSOR:	/* CPU entry */
2518 			if (!(item->i.p.flags & SRAT_ENABLED)) {
2519 				item = (struct srat_item *)((uintptr_t)item +
2520 				    item->len);
2521 				continue;
2522 			}
2523 			domain = item->i.p.domain1;
2524 			for (i = 0; i < 3; i++) {
2525 				domain += item->i.p.domain2[i] <<
2526 				    ((i + 1) * 8);
2527 			}
2528 			break;
2529 
2530 		case SRAT_MEMORY:	/* memory entry */
2531 			if (!(item->i.m.flags & SRAT_ENABLED)) {
2532 				item = (struct srat_item *)((uintptr_t)item +
2533 				    item->len);
2534 				continue;
2535 			}
2536 			domain = item->i.m.domain;
2537 			break;
2538 
2539 		case SRAT_X2APIC:	/* x2apic CPU entry */
2540 			if (!(item->i.xp.flags & SRAT_ENABLED)) {
2541 				item = (struct srat_item *)((uintptr_t)item +
2542 				    item->len);
2543 				continue;
2544 			}
2545 			domain = item->i.xp.domain;
2546 			break;
2547 
2548 		default:
2549 			item = (struct srat_item *)((uintptr_t)item +
2550 			    item->len);
2551 			continue;
2552 		}
2553 
2554 		/*
2555 		 * Keep track of minimum proximity domain ID
2556 		 */
2557 		if (domain < domain_min)
2558 			domain_min = domain;
2559 
2560 		item = (struct srat_item *)((uintptr_t)item + item->len);
2561 	}
2562 	if (lgrp_plat_domain_min_enable && prox_domain_min != NULL)
2563 		*prox_domain_min = domain_min;
2564 
2565 	/*
2566 	 * Walk through SRAT, examining each CPU and memory entry to determine
2567 	 * proximity domain ID for each.
2568 	 */
2569 	domain_cnt = 0;
2570 	item = tp->list;
2571 	end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2572 	bzero(node_domain, MAX_NODES * sizeof (node_domain_map_t));
2573 	while (item < end) {
2574 		uint32_t	domain;
2575 		boolean_t	overflow;
2576 		uint_t		start;
2577 
2578 		switch (item->type) {
2579 		case SRAT_PROCESSOR:	/* CPU entry */
2580 			if (!(item->i.p.flags & SRAT_ENABLED)) {
2581 				item = (struct srat_item *)((uintptr_t)item +
2582 				    item->len);
2583 				continue;
2584 			}
2585 			domain = item->i.p.domain1;
2586 			for (i = 0; i < 3; i++) {
2587 				domain += item->i.p.domain2[i] <<
2588 				    ((i + 1) * 8);
2589 			}
2590 			break;
2591 
2592 		case SRAT_MEMORY:	/* memory entry */
2593 			if (!(item->i.m.flags & SRAT_ENABLED)) {
2594 				item = (struct srat_item *)((uintptr_t)item +
2595 				    item->len);
2596 				continue;
2597 			}
2598 			domain = item->i.m.domain;
2599 			break;
2600 
2601 		case SRAT_X2APIC:	/* x2apic CPU entry */
2602 			if (!(item->i.xp.flags & SRAT_ENABLED)) {
2603 				item = (struct srat_item *)((uintptr_t)item +
2604 				    item->len);
2605 				continue;
2606 			}
2607 			domain = item->i.xp.domain;
2608 			break;
2609 
2610 		default:
2611 			item = (struct srat_item *)((uintptr_t)item +
2612 			    item->len);
2613 			continue;
2614 		}
2615 
2616 		/*
2617 		 * Count and keep track of which proximity domain IDs seen
2618 		 */
2619 		start = i = domain % MAX_NODES;
2620 		overflow = B_TRUE;
2621 		do {
2622 			/*
2623 			 * Create entry for proximity domain and increment
2624 			 * count when no entry exists where proximity domain
2625 			 * hashed
2626 			 */
2627 			if (!node_domain[i].exists) {
2628 				node_domain[i].exists = 1;
2629 				node_domain[i].prox_domain = domain;
2630 				domain_cnt++;
2631 				overflow = B_FALSE;
2632 				break;
2633 			}
2634 
2635 			/*
2636 			 * Nothing to do when proximity domain seen already
2637 			 * and its entry exists
2638 			 */
2639 			if (node_domain[i].prox_domain == domain) {
2640 				overflow = B_FALSE;
2641 				break;
2642 			}
2643 
2644 			/*
2645 			 * Entry exists where proximity domain hashed, but for
2646 			 * different proximity domain so keep search for empty
2647 			 * slot to put it or matching entry whichever comes
2648 			 * first.
2649 			 */
2650 			i = (i + 1) % MAX_NODES;
2651 		} while (i != start);
2652 
2653 		/*
2654 		 * Didn't find empty or matching entry which means have more
2655 		 * proximity domains than supported nodes (:-(
2656 		 */
2657 		ASSERT(overflow != B_TRUE);
2658 		if (overflow == B_TRUE)
2659 			return (-1);
2660 
2661 		item = (struct srat_item *)((uintptr_t)item + item->len);
2662 	}
2663 	return (domain_cnt);
2664 }
2665 
2666 
2667 /*
2668  * Set lgroup latencies for 2 level lgroup topology
2669  */
2670 static void
2671 lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
2672     lgrp_plat_latency_stats_t *lat_stats)
2673 {
2674 	int	i;
2675 
2676 	ASSERT(node_memory != NULL && lat_stats != NULL);
2677 
2678 	if (lgrp_plat_node_cnt >= 4)
2679 		cmn_err(CE_NOTE,
2680 		    "MPO only optimizing for local and remote\n");
2681 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
2682 		int	j;
2683 
2684 		if (!node_memory[i].exists)
2685 			continue;
2686 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
2687 			if (!node_memory[j].exists)
2688 				continue;
2689 			if (i == j)
2690 				lat_stats->latencies[i][j] = 2;
2691 			else
2692 				lat_stats->latencies[i][j] = 3;
2693 		}
2694 	}
2695 	lat_stats->latency_min = 2;
2696 	lat_stats->latency_max = 3;
2697 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
2698 }
2699 
2700 
2701 /*
2702  * The following Opteron specific constants, macros, types, and routines define
2703  * PCI configuration space registers and how to read them to determine the NUMA
2704  * configuration of *supported* Opteron processors.  They provide the same
2705  * information that may be gotten from the ACPI System Resource Affinity Table
2706  * (SRAT) if it exists on the machine of interest.
2707  *
2708  * The AMD BIOS and Kernel Developer's Guide (BKDG) for the processor family
2709  * of interest describes all of these registers and their contents.  The main
2710  * registers used by this code to determine the NUMA configuration of the
2711  * machine are the node ID register for the number of NUMA nodes and the DRAM
2712  * address map registers for the physical address range of each node.
2713  *
2714  * NOTE: The format and how to determine the NUMA configuration using PCI
2715  *	 config space registers may change or may not be supported in future
2716  *	 Opteron processor families.
2717  */
2718 
2719 /*
2720  * How many bits to shift Opteron DRAM Address Map base and limit registers
2721  * to get actual value
2722  */
2723 #define	OPT_DRAMADDR_HI_LSHIFT_ADDR	40	/* shift left for address */
2724 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
2725 
2726 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
2727 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
2728 
2729 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
2730 
2731 /*
2732  * Macros to derive addresses from Opteron DRAM Address Map registers
2733  */
2734 #define	OPT_DRAMADDR_HI(reg) \
2735 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
2736 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
2737 
2738 #define	OPT_DRAMADDR_LO(reg) \
2739 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
2740 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
2741 
2742 #define	OPT_DRAMADDR(high, low) \
2743 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
2744 
2745 /*
2746  * Bit masks defining what's in Opteron DRAM Address Map base register
2747  */
2748 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
2749 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
2750 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
2751 
2752 /*
2753  * Bit masks defining what's in Opteron DRAM Address Map limit register
2754  */
2755 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
2756 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
2757 
2758 
2759 /*
2760  * Opteron Node ID register in PCI configuration space contains
2761  * number of nodes in system, etc. for Opteron K8.  The following
2762  * constants and macros define its contents, structure, and access.
2763  */
2764 
2765 /*
2766  * Bit masks defining what's in Opteron Node ID register
2767  */
2768 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
2769 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
2770 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
2771 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
2772 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
2773 
2774 /*
2775  * How many bits in Opteron Node ID register to shift right to get actual value
2776  */
2777 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
2778 
2779 /*
2780  * Macros to get values from Opteron Node ID register
2781  */
2782 #define	OPT_NODE_CNT(reg) \
2783 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
2784 
2785 /*
2786  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
2787  * "in/out" instructions
2788  *
2789  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
2790  *	 other uses should just do MMIO to access PCI ECS.
2791  *	 Must enable special bit in Northbridge Configuration Register on
2792  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
2793  *	 using "in/out" instructions and restore special bit after done
2794  *	 accessing PCI ECS.
2795  */
2796 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
2797 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
2798 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
2799 	    ((((reg) >> 8) & 0xf) << 24))
2800 
2801 /*
2802  * PCI configuration space registers accessed by specifying
2803  * a bus, device, function, and offset.  The following constants
2804  * define the values needed to access Opteron K8 configuration
2805  * info to determine its node topology
2806  */
2807 
2808 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
2809 
2810 /*
2811  * Opteron PCI configuration space register function values
2812  */
2813 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
2814 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
2815 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
2816 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
2817 
2818 /*
2819  * PCI Configuration Space register offsets
2820  */
2821 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
2822 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
2823 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
2824 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
2825 
2826 /*
2827  * Opteron PCI Configuration Space device IDs for nodes
2828  */
2829 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
2830 
2831 
2832 /*
2833  * Opteron DRAM address map gives base and limit for physical memory in a node
2834  */
2835 typedef	struct opt_dram_addr_map {
2836 	uint32_t	base_hi;
2837 	uint32_t	base_lo;
2838 	uint32_t	limit_hi;
2839 	uint32_t	limit_lo;
2840 } opt_dram_addr_map_t;
2841 
2842 
2843 /*
2844  * Supported AMD processor families
2845  */
2846 #define	AMD_FAMILY_HAMMER	15
2847 #define	AMD_FAMILY_GREYHOUND	16
2848 
2849 /*
2850  * Whether to have is_opteron() return 1 even when processor isn't supported
2851  */
2852 uint_t	is_opteron_override = 0;
2853 
2854 /*
2855  * AMD processor family for current CPU
2856  */
2857 uint_t	opt_family = 0;
2858 
2859 
2860 /*
2861  * Determine whether we're running on a supported AMD Opteron since reading
2862  * node count and DRAM address map registers may have different format or
2863  * may not be supported across processor families
2864  */
2865 static int
2866 is_opteron(void)
2867 {
2868 
2869 	if (x86_vendor != X86_VENDOR_AMD)
2870 		return (0);
2871 
2872 	opt_family = cpuid_getfamily(CPU);
2873 	if (opt_family == AMD_FAMILY_HAMMER ||
2874 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
2875 		return (1);
2876 	else
2877 		return (0);
2878 }
2879 
2880 
2881 /*
2882  * Determine NUMA configuration for Opteron from registers that live in PCI
2883  * configuration space
2884  */
2885 static void
2886 opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
2887     node_phys_addr_map_t *node_memory)
2888 {
2889 	uint_t				bus;
2890 	uint_t				dev;
2891 	struct opt_dram_addr_map	dram_map[MAX_NODES];
2892 	uint_t				node;
2893 	uint_t				node_info[MAX_NODES];
2894 	uint_t				off_hi;
2895 	uint_t				off_lo;
2896 	uint64_t			nb_cfg_reg;
2897 
2898 	/*
2899 	 * Read configuration registers from PCI configuration space to
2900 	 * determine node information, which memory is in each node, etc.
2901 	 *
2902 	 * Write to PCI configuration space address register to specify
2903 	 * which configuration register to read and read/write PCI
2904 	 * configuration space data register to get/set contents
2905 	 */
2906 	bus = OPT_PCS_BUS_CONFIG;
2907 	dev = OPT_PCS_DEV_NODE0;
2908 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
2909 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
2910 
2911 	/*
2912 	 * Read node ID register for node 0 to get node count
2913 	 */
2914 	node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
2915 	    OPT_PCS_OFF_NODEID);
2916 	*node_cnt = OPT_NODE_CNT(node_info[0]) + 1;
2917 
2918 	/*
2919 	 * If number of nodes is more than maximum supported, then set node
2920 	 * count to 1 and treat system as UMA instead of NUMA.
2921 	 */
2922 	if (*node_cnt > MAX_NODES) {
2923 		*node_cnt = 1;
2924 		return;
2925 	}
2926 
2927 	/*
2928 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
2929 	 * read high DRAM address map base and limit registers
2930 	 */
2931 	if (opt_family == AMD_FAMILY_GREYHOUND) {
2932 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
2933 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
2934 			wrmsr(MSR_AMD_NB_CFG,
2935 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
2936 	}
2937 
2938 	for (node = 0; node < *node_cnt; node++) {
2939 		uint32_t	base_hi;
2940 		uint32_t	base_lo;
2941 		uint32_t	limit_hi;
2942 		uint32_t	limit_lo;
2943 
2944 		/*
2945 		 * Read node ID register (except for node 0 which we just read)
2946 		 */
2947 		if (node > 0) {
2948 			node_info[node] = pci_getl_func(bus, dev,
2949 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
2950 		}
2951 
2952 		/*
2953 		 * Read DRAM base and limit registers which specify
2954 		 * physical memory range of each node
2955 		 */
2956 		if (opt_family != AMD_FAMILY_GREYHOUND)
2957 			base_hi = 0;
2958 		else {
2959 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2960 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2961 			base_hi = dram_map[node].base_hi =
2962 			    inl(PCI_CONFDATA);
2963 		}
2964 		base_lo = dram_map[node].base_lo = pci_getl_func(bus, dev,
2965 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
2966 
2967 		if ((dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN) &&
2968 		    mem_intrlv)
2969 			*mem_intrlv = *mem_intrlv + 1;
2970 
2971 		off_hi += 4;	/* high limit register offset */
2972 		if (opt_family != AMD_FAMILY_GREYHOUND)
2973 			limit_hi = 0;
2974 		else {
2975 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2976 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2977 			limit_hi = dram_map[node].limit_hi =
2978 			    inl(PCI_CONFDATA);
2979 		}
2980 
2981 		off_lo += 4;	/* low limit register offset */
2982 		limit_lo = dram_map[node].limit_lo = pci_getl_func(bus,
2983 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
2984 
2985 		/*
2986 		 * Increment device number to next node and register offsets
2987 		 * for DRAM base register of next node
2988 		 */
2989 		off_hi += 4;
2990 		off_lo += 4;
2991 		dev++;
2992 
2993 		/*
2994 		 * Both read and write enable bits must be enabled in DRAM
2995 		 * address map base register for physical memory to exist in
2996 		 * node
2997 		 */
2998 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
2999 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
3000 			/*
3001 			 * Mark node memory as non-existent and set start and
3002 			 * end addresses to be same in node_memory[]
3003 			 */
3004 			node_memory[node].exists = 0;
3005 			node_memory[node].start = node_memory[node].end =
3006 			    (pfn_t)-1;
3007 			continue;
3008 		}
3009 
3010 		/*
3011 		 * Mark node memory as existing and remember physical address
3012 		 * range of each node for use later
3013 		 */
3014 		node_memory[node].exists = 1;
3015 
3016 		node_memory[node].start = btop(OPT_DRAMADDR(base_hi, base_lo));
3017 
3018 		node_memory[node].end = btop(OPT_DRAMADDR(limit_hi, limit_lo) |
3019 		    OPT_DRAMADDR_LO_MASK_OFF);
3020 	}
3021 
3022 	/*
3023 	 * Restore PCI Extended Configuration Space enable bit
3024 	 */
3025 	if (opt_family == AMD_FAMILY_GREYHOUND) {
3026 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
3027 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
3028 	}
3029 }
3030 
3031 
3032 /*
3033  * Return average amount of time to read vendor ID register on Northbridge
3034  * N times on specified destination node from current CPU
3035  */
3036 static hrtime_t
3037 opt_probe_vendor(int dest_node, int nreads)
3038 {
3039 	int		cnt;
3040 	uint_t		dev;
3041 	/* LINTED: set but not used in function */
3042 	volatile uint_t	dev_vendor;
3043 	hrtime_t	elapsed;
3044 	hrtime_t	end;
3045 	int		ipl;
3046 	hrtime_t	start;
3047 
3048 	dev = OPT_PCS_DEV_NODE0 + dest_node;
3049 	kpreempt_disable();
3050 	ipl = spl8();
3051 	outl(PCI_CONFADD, PCI_CADDR1(0, dev, OPT_PCS_FUNC_DRAM,
3052 	    OPT_PCS_OFF_VENDOR));
3053 	start = gethrtime();
3054 	for (cnt = 0; cnt < nreads; cnt++)
3055 		dev_vendor = inl(PCI_CONFDATA);
3056 	end = gethrtime();
3057 	elapsed = (end - start) / nreads;
3058 	splx(ipl);
3059 	kpreempt_enable();
3060 	return (elapsed);
3061 }
3062