xref: /illumos-gate/usr/src/uts/i86pc/os/lgrpplat.c (revision ddb365bfc9e868ad24ccdcb0dc91af18b10df082)
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 (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 /*
26  * Copyright (c) 2010, Intel Corporation.
27  * All rights reserved.
28  */
29 
30 /*
31  * LOCALITY GROUP (LGROUP) PLATFORM SUPPORT FOR X86/AMD64 PLATFORMS
32  * ================================================================
33  * Multiprocessor AMD and Intel systems may have Non Uniform Memory Access
34  * (NUMA).  A NUMA machine consists of one or more "nodes" that each consist of
35  * one or more CPUs and some local memory.  The CPUs in each node can access
36  * the memory in the other nodes but at a higher latency than accessing their
37  * local memory.  Typically, a system with only one node has Uniform Memory
38  * Access (UMA), but it may be possible to have a one node system that has
39  * some global memory outside of the node which is higher latency.
40  *
41  * Module Description
42  * ------------------
43  * This module provides a platform interface for determining which CPUs and
44  * which memory (and how much) are in a NUMA node and how far each node is from
45  * each other.  The interface is used by the Virtual Memory (VM) system and the
46  * common lgroup framework.  The VM system uses the plat_*() routines to fill
47  * in its memory node (memnode) array with the physical address range spanned
48  * by each NUMA node to know which memory belongs to which node, so it can
49  * build and manage a physical page free list for each NUMA node and allocate
50  * local memory from each node as needed.  The common lgroup framework uses the
51  * exported lgrp_plat_*() routines to figure out which CPUs and memory belong
52  * to each node (leaf lgroup) and how far each node is from each other, so it
53  * can build the latency (lgroup) topology for the machine in order to optimize
54  * for locality.  Also, an lgroup platform handle instead of lgroups are used
55  * in the interface with this module, so this module shouldn't need to know
56  * anything about lgroups.  Instead, it just needs to know which CPUs, memory,
57  * etc. are in each NUMA node, how far each node is from each other, and to use
58  * a unique lgroup platform handle to refer to each node through the interface.
59  *
60  * Determining NUMA Configuration
61  * ------------------------------
62  * By default, this module will try to determine the NUMA configuration of the
63  * machine by reading the ACPI System Resource Affinity Table (SRAT) and System
64  * Locality Information Table (SLIT).  The SRAT contains info to tell which
65  * CPUs and memory are local to a given proximity domain (NUMA node).  The SLIT
66  * is a matrix that gives the distance between each system locality (which is
67  * a NUMA node and should correspond to proximity domains in the SRAT).  For
68  * more details on the SRAT and SLIT, please refer to an ACPI 3.0 or newer
69  * specification.
70  *
71  * If the SRAT doesn't exist on a system with AMD Opteron processors, we
72  * examine registers in PCI configuration space to determine how many nodes are
73  * in the system and which CPUs and memory are in each node.
74  * do while booting the kernel.
75  *
76  * NOTE: Using these PCI configuration space registers to determine this
77  *       locality info is not guaranteed to work or be compatible across all
78  *	 Opteron processor families.
79  *
80  * If the SLIT does not exist or look right, the kernel will probe to determine
81  * the distance between nodes as long as the NUMA CPU and memory configuration
82  * has been determined (see lgrp_plat_probe() for details).
83  *
84  * Data Structures
85  * ---------------
86  * The main data structures used by this code are the following:
87  *
88  * - lgrp_plat_cpu_node[]		CPU to node ID mapping table indexed by
89  *					CPU ID (only used for SRAT)
90  *
91  * - lgrp_plat_lat_stats.latencies[][]	Table of latencies between same and
92  *					different nodes indexed by node ID
93  *
94  * - lgrp_plat_node_cnt			Number of NUMA nodes in system for
95  *					non-DR-capable systems,
96  *					maximum possible number of NUMA nodes
97  *					in system for DR capable systems.
98  *
99  * - lgrp_plat_node_domain[]		Node ID to proximity domain ID mapping
100  *					table indexed by node ID (only used
101  *					for SRAT)
102  *
103  * - lgrp_plat_memnode_info[]		Table with physical address range for
104  *					each memory node indexed by memory node
105  *					ID
106  *
107  * The code is implemented to make the following always be true:
108  *
109  *	lgroup platform handle == node ID == memnode ID
110  *
111  * Moreover, it allows for the proximity domain ID to be equal to all of the
112  * above as long as the proximity domains IDs are numbered from 0 to <number of
113  * nodes - 1>.  This is done by hashing each proximity domain ID into the range
114  * from 0 to <number of nodes - 1>.  Then proximity ID N will hash into node ID
115  * N and proximity domain ID N will be entered into lgrp_plat_node_domain[N]
116  * and be assigned node ID N.  If the proximity domain IDs aren't numbered
117  * from 0 to <number of nodes - 1>, then hashing the proximity domain IDs into
118  * lgrp_plat_node_domain[] will still work for assigning proximity domain IDs
119  * to node IDs.  However, the proximity domain IDs may not map to the
120  * equivalent node ID since we want to keep the node IDs numbered from 0 to
121  * <number of nodes - 1> to minimize cost of searching and potentially space.
122  *
123  * With the introduction of support of memory DR operations on x86 platforms,
124  * things get a little complicated. The addresses of hot-added memory may not
125  * be continuous with other memory connected to the same lgrp node. In other
126  * words, memory addresses may get interleaved among lgrp nodes after memory
127  * DR operations. To work around this limitation, we have extended the
128  * relationship between lgrp node and memory node from 1:1 map to 1:N map,
129  * that means there may be multiple memory nodes associated with a lgrp node
130  * after memory DR operations.
131  *
132  * To minimize the code changes to support memory DR operations, the
133  * following policies have been adopted.
134  * 1) On non-DR-capable systems, the relationship among lgroup platform handle,
135  *    node ID and memnode ID is still kept as:
136  *	lgroup platform handle == node ID == memnode ID
137  * 2) For memory present at boot time on DR capable platforms, the relationship
138  *    is still kept as is.
139  *	lgroup platform handle == node ID == memnode ID
140  * 3) For hot-added memory, the relationship between lgrp ID and memnode ID have
141  *    been changed from 1:1 map to 1:N map. Memnode IDs [0 - lgrp_plat_node_cnt)
142  *    are reserved for memory present at boot time, and memnode IDs
143  *    [lgrp_plat_node_cnt, max_mem_nodes) are used to dynamically allocate
144  *    memnode ID for hot-added memory.
145  * 4) All boot code having the assumption "node ID == memnode ID" can live as
146  *    is, that's because node ID is always equal to memnode ID at boot time.
147  * 5) The lgrp_plat_memnode_info_update(), plat_pfn_to_mem_node() and
148  *    lgrp_plat_mem_size() related logics have been enhanced to deal with
149  *    the 1:N map relationship.
150  * 6) The latency probing related logics, which have the assumption
151  *    "node ID == memnode ID" and may be called at run time, is disabled if
152  *    memory DR operation is enabled.
153  */
154 
155 
156 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
157 #include <sys/atomic.h>
158 #include <sys/bootconf.h>
159 #include <sys/cmn_err.h>
160 #include <sys/controlregs.h>
161 #include <sys/cpupart.h>
162 #include <sys/cpuvar.h>
163 #include <sys/lgrp.h>
164 #include <sys/machsystm.h>
165 #include <sys/memlist.h>
166 #include <sys/memnode.h>
167 #include <sys/mman.h>
168 #include <sys/note.h>
169 #include <sys/pci_cfgspace.h>
170 #include <sys/pci_impl.h>
171 #include <sys/param.h>
172 #include <sys/pghw.h>
173 #include <sys/promif.h>		/* for prom_printf() */
174 #include <sys/sysmacros.h>
175 #include <sys/systm.h>
176 #include <sys/thread.h>
177 #include <sys/types.h>
178 #include <sys/var.h>
179 #include <sys/x86_archext.h>
180 #include <vm/hat_i86.h>
181 #include <vm/seg_kmem.h>
182 #include <vm/vm_dep.h>
183 
184 #include <sys/acpidev.h>
185 #include <sys/acpi/acpi.h>		/* for SRAT, SLIT and MSCT */
186 
187 /* from fakebop.c */
188 extern ACPI_TABLE_SRAT *srat_ptr;
189 extern ACPI_TABLE_SLIT *slit_ptr;
190 extern ACPI_TABLE_MSCT *msct_ptr;
191 
192 #define	MAX_NODES		8
193 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
194 
195 /*
196  * Constants for configuring probing
197  */
198 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
199 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
200 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
201 
202 /*
203  * Flags for probing
204  */
205 #define	LGRP_PLAT_PROBE_ENABLE		0x1	/* enable probing */
206 #define	LGRP_PLAT_PROBE_PGCPY		0x2	/* probe using page copy */
207 #define	LGRP_PLAT_PROBE_VENDOR		0x4	/* probe vendor ID register */
208 
209 /*
210  * Hash proximity domain ID into node to domain mapping table "mod" number of
211  * nodes to minimize span of entries used and try to have lowest numbered
212  * proximity domain be node 0
213  */
214 #define	NODE_DOMAIN_HASH(domain, node_cnt) \
215 	((lgrp_plat_prox_domain_min == UINT32_MAX) ? (domain) % node_cnt : \
216 	    ((domain) - lgrp_plat_prox_domain_min) % node_cnt)
217 
218 /*
219  * CPU to node ID mapping structure (only used with SRAT)
220  */
221 typedef	struct cpu_node_map {
222 	int		exists;
223 	uint_t		node;
224 	uint32_t	apicid;
225 	uint32_t	prox_domain;
226 } cpu_node_map_t;
227 
228 /*
229  * Latency statistics
230  */
231 typedef struct lgrp_plat_latency_stats {
232 	hrtime_t	latencies[MAX_NODES][MAX_NODES];
233 	hrtime_t	latency_max;
234 	hrtime_t	latency_min;
235 } lgrp_plat_latency_stats_t;
236 
237 /*
238  * Memory configuration for probing
239  */
240 typedef struct lgrp_plat_probe_mem_config {
241 	size_t	probe_memsize;		/* how much memory to probe per node */
242 	caddr_t	probe_va[MAX_NODES];	/* where memory mapped for probing */
243 	pfn_t	probe_pfn[MAX_NODES];	/* physical pages to map for probing */
244 } lgrp_plat_probe_mem_config_t;
245 
246 /*
247  * Statistics kept for probing
248  */
249 typedef struct lgrp_plat_probe_stats {
250 	hrtime_t	flush_cost;
251 	hrtime_t	probe_cost;
252 	hrtime_t	probe_cost_total;
253 	hrtime_t	probe_error_code;
254 	hrtime_t	probe_errors[MAX_NODES][MAX_NODES];
255 	int		probe_suspect[MAX_NODES][MAX_NODES];
256 	hrtime_t	probe_max[MAX_NODES][MAX_NODES];
257 	hrtime_t	probe_min[MAX_NODES][MAX_NODES];
258 } lgrp_plat_probe_stats_t;
259 
260 /*
261  * Node to proximity domain ID mapping structure (only used with SRAT)
262  */
263 typedef	struct node_domain_map {
264 	int		exists;
265 	uint32_t	prox_domain;
266 } node_domain_map_t;
267 
268 /*
269  * Node ID and starting and ending page for physical memory in memory node
270  */
271 typedef	struct memnode_phys_addr_map {
272 	pfn_t		start;
273 	pfn_t		end;
274 	int		exists;
275 	uint32_t	prox_domain;
276 	uint32_t	device_id;
277 	uint_t		lgrphand;
278 } memnode_phys_addr_map_t;
279 
280 /*
281  * Number of CPUs for which we got APIC IDs
282  */
283 static int				lgrp_plat_apic_ncpus = 0;
284 
285 /*
286  * CPU to node ID mapping table (only used for SRAT) and its max number of
287  * entries
288  */
289 static cpu_node_map_t			*lgrp_plat_cpu_node = NULL;
290 static uint_t				lgrp_plat_cpu_node_nentries = 0;
291 
292 /*
293  * Latency statistics
294  */
295 lgrp_plat_latency_stats_t		lgrp_plat_lat_stats;
296 
297 /*
298  * Whether memory is interleaved across nodes causing MPO to be disabled
299  */
300 static int				lgrp_plat_mem_intrlv = 0;
301 
302 /*
303  * Node ID to proximity domain ID mapping table (only used for SRAT)
304  */
305 static node_domain_map_t		lgrp_plat_node_domain[MAX_NODES];
306 
307 /*
308  * Physical address range for memory in each node
309  */
310 static memnode_phys_addr_map_t		lgrp_plat_memnode_info[MAX_MEM_NODES];
311 
312 /*
313  * Statistics gotten from probing
314  */
315 static lgrp_plat_probe_stats_t		lgrp_plat_probe_stats;
316 
317 /*
318  * Memory configuration for probing
319  */
320 static lgrp_plat_probe_mem_config_t	lgrp_plat_probe_mem_config;
321 
322 /*
323  * Lowest proximity domain ID seen in ACPI SRAT
324  */
325 static uint32_t				lgrp_plat_prox_domain_min = UINT32_MAX;
326 
327 /*
328  * Error code from processing ACPI SRAT
329  */
330 static int				lgrp_plat_srat_error = 0;
331 
332 /*
333  * Error code from processing ACPI SLIT
334  */
335 static int				lgrp_plat_slit_error = 0;
336 
337 /*
338  * Whether lgrp topology has been flattened to 2 levels.
339  */
340 static int				lgrp_plat_topo_flatten = 0;
341 
342 
343 /*
344  * Maximum memory node ID in use.
345  */
346 static uint_t				lgrp_plat_max_mem_node;
347 
348 /*
349  * Allocate lgroup array statically
350  */
351 static lgrp_t				lgrp_space[NLGRP];
352 static int				nlgrps_alloc;
353 
354 
355 /*
356  * Enable finding and using minimum proximity domain ID when hashing
357  */
358 int			lgrp_plat_domain_min_enable = 1;
359 
360 /*
361  * Maximum possible number of nodes in system
362  */
363 uint_t			lgrp_plat_node_cnt = 1;
364 
365 /*
366  * Enable sorting nodes in ascending order by starting physical address
367  */
368 int			lgrp_plat_node_sort_enable = 1;
369 
370 /*
371  * Configuration Parameters for Probing
372  * - lgrp_plat_probe_flags	Flags to specify enabling probing, probe
373  *				operation, etc.
374  * - lgrp_plat_probe_nrounds	How many rounds of probing to do
375  * - lgrp_plat_probe_nsamples	Number of samples to take when probing each
376  *				node
377  * - lgrp_plat_probe_nreads	Number of times to read vendor ID from
378  *				Northbridge for each probe
379  */
380 uint_t			lgrp_plat_probe_flags = 0;
381 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
382 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
383 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
384 
385 /*
386  * Enable use of ACPI System Resource Affinity Table (SRAT), System
387  * Locality Information Table (SLIT) and Maximum System Capability Table (MSCT)
388  */
389 int			lgrp_plat_srat_enable = 1;
390 int			lgrp_plat_slit_enable = 1;
391 int			lgrp_plat_msct_enable = 1;
392 
393 /*
394  * mnode_xwa: set to non-zero value to initiate workaround if large pages are
395  * found to be crossing memory node boundaries. The workaround will eliminate
396  * a base size page at the end of each memory node boundary to ensure that
397  * a large page with constituent pages that span more than 1 memory node
398  * can never be formed.
399  *
400  */
401 int	mnode_xwa = 1;
402 
403 /*
404  * Static array to hold lgroup statistics
405  */
406 struct lgrp_stats	lgrp_stats[NLGRP];
407 
408 
409 /*
410  * Forward declarations of platform interface routines
411  */
412 void		plat_build_mem_nodes(struct memlist *list);
413 
414 int		plat_mnode_xcheck(pfn_t pfncnt);
415 
416 lgrp_handle_t	plat_mem_node_to_lgrphand(int mnode);
417 
418 int		plat_pfn_to_mem_node(pfn_t pfn);
419 
420 /*
421  * Forward declarations of lgroup platform interface routines
422  */
423 lgrp_t		*lgrp_plat_alloc(lgrp_id_t lgrpid);
424 
425 void		lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg);
426 
427 lgrp_handle_t	lgrp_plat_cpu_to_hand(processorid_t id);
428 
429 void		lgrp_plat_init(lgrp_init_stages_t stage);
430 
431 int		lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to);
432 
433 int		lgrp_plat_max_lgrps(void);
434 
435 pgcnt_t		lgrp_plat_mem_size(lgrp_handle_t plathand,
436     lgrp_mem_query_t query);
437 
438 lgrp_handle_t	lgrp_plat_pfn_to_hand(pfn_t pfn);
439 
440 void		lgrp_plat_probe(void);
441 
442 lgrp_handle_t	lgrp_plat_root_hand(void);
443 
444 
445 /*
446  * Forward declarations of local routines
447  */
448 static int	is_opteron(void);
449 
450 static int	lgrp_plat_cpu_node_update(node_domain_map_t *node_domain,
451     int node_cnt, cpu_node_map_t *cpu_node, int nentries, uint32_t apicid,
452     uint32_t domain);
453 
454 static int	lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node,
455     int cpu_node_nentries);
456 
457 static int	lgrp_plat_domain_to_node(node_domain_map_t *node_domain,
458     int node_cnt, uint32_t domain);
459 
460 static void	lgrp_plat_get_numa_config(void);
461 
462 static void	lgrp_plat_latency_adjust(memnode_phys_addr_map_t *memnode_info,
463     lgrp_plat_latency_stats_t *lat_stats,
464     lgrp_plat_probe_stats_t *probe_stats);
465 
466 static int	lgrp_plat_latency_verify(memnode_phys_addr_map_t *memnode_info,
467     lgrp_plat_latency_stats_t *lat_stats);
468 
469 static void	lgrp_plat_main_init(void);
470 
471 static pgcnt_t	lgrp_plat_mem_size_default(lgrp_handle_t, lgrp_mem_query_t);
472 
473 static int	lgrp_plat_node_domain_update(node_domain_map_t *node_domain,
474     int node_cnt, uint32_t domain);
475 
476 static int	lgrp_plat_memnode_info_update(node_domain_map_t *node_domain,
477     int node_cnt, memnode_phys_addr_map_t *memnode_info, int memnode_cnt,
478     uint64_t start, uint64_t end, uint32_t domain, uint32_t device_id);
479 
480 static void	lgrp_plat_node_sort(node_domain_map_t *node_domain,
481     int node_cnt, cpu_node_map_t *cpu_node, int cpu_count,
482     memnode_phys_addr_map_t *memnode_info);
483 
484 static hrtime_t	lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
485     int cpu_node_nentries, lgrp_plat_probe_mem_config_t *probe_mem_config,
486     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats);
487 
488 static int	lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node);
489 
490 static int	lgrp_plat_process_slit(ACPI_TABLE_SLIT *tp,
491     node_domain_map_t *node_domain, uint_t node_cnt,
492     memnode_phys_addr_map_t *memnode_info,
493     lgrp_plat_latency_stats_t *lat_stats);
494 
495 static int	lgrp_plat_process_sli(uint32_t domain, uchar_t *sli_info,
496     uint32_t sli_cnt, node_domain_map_t *node_domain, uint_t node_cnt,
497     lgrp_plat_latency_stats_t *lat_stats);
498 
499 static int	lgrp_plat_process_srat(ACPI_TABLE_SRAT *tp, ACPI_TABLE_MSCT *mp,
500     uint32_t *prox_domain_min, node_domain_map_t *node_domain,
501     cpu_node_map_t *cpu_node, int cpu_count,
502     memnode_phys_addr_map_t *memnode_info);
503 
504 static void	lgrp_plat_release_bootstrap(void);
505 
506 static int	lgrp_plat_srat_domains(ACPI_TABLE_SRAT *tp,
507     uint32_t *prox_domain_min);
508 
509 static int	lgrp_plat_msct_domains(ACPI_TABLE_MSCT *tp,
510     uint32_t *prox_domain_min);
511 
512 static void	lgrp_plat_2level_setup(lgrp_plat_latency_stats_t *lat_stats);
513 
514 static void	opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
515     memnode_phys_addr_map_t *memnode_info);
516 
517 static hrtime_t	opt_probe_vendor(int dest_node, int nreads);
518 
519 
520 /*
521  * PLATFORM INTERFACE ROUTINES
522  */
523 
524 /*
525  * Configure memory nodes for machines with more than one node (ie NUMA)
526  */
527 void
528 plat_build_mem_nodes(struct memlist *list)
529 {
530 	pfn_t		cur_start;	/* start addr of subrange */
531 	pfn_t		cur_end;	/* end addr of subrange */
532 	pfn_t		start;		/* start addr of whole range */
533 	pfn_t		end;		/* end addr of whole range */
534 	pgcnt_t		endcnt;		/* pages to sacrifice */
535 
536 	/*
537 	 * Boot install lists are arranged <addr, len>, ...
538 	 */
539 	while (list) {
540 		int	node;
541 
542 		start = list->ml_address >> PAGESHIFT;
543 		end = (list->ml_address + list->ml_size - 1) >> PAGESHIFT;
544 
545 		if (start > physmax) {
546 			list = list->ml_next;
547 			continue;
548 		}
549 		if (end > physmax)
550 			end = physmax;
551 
552 		/*
553 		 * When there is only one memnode, just add memory to memnode
554 		 */
555 		if (max_mem_nodes == 1) {
556 			mem_node_add_slice(start, end);
557 			list = list->ml_next;
558 			continue;
559 		}
560 
561 		/*
562 		 * mem_node_add_slice() expects to get a memory range that
563 		 * is within one memnode, so need to split any memory range
564 		 * that spans multiple memnodes into subranges that are each
565 		 * contained within one memnode when feeding them to
566 		 * mem_node_add_slice()
567 		 */
568 		cur_start = start;
569 		do {
570 			node = plat_pfn_to_mem_node(cur_start);
571 
572 			/*
573 			 * Panic if DRAM address map registers or SRAT say
574 			 * memory in node doesn't exist or address from
575 			 * boot installed memory list entry isn't in this node.
576 			 * This shouldn't happen and rest of code can't deal
577 			 * with this if it does.
578 			 */
579 			if (node < 0 || node >= lgrp_plat_max_mem_node ||
580 			    !lgrp_plat_memnode_info[node].exists ||
581 			    cur_start < lgrp_plat_memnode_info[node].start ||
582 			    cur_start > lgrp_plat_memnode_info[node].end) {
583 				cmn_err(CE_PANIC, "Don't know which memnode "
584 				    "to add installed memory address 0x%lx\n",
585 				    cur_start);
586 			}
587 
588 			/*
589 			 * End of current subrange should not span memnodes
590 			 */
591 			cur_end = end;
592 			endcnt = 0;
593 			if (lgrp_plat_memnode_info[node].exists &&
594 			    cur_end > lgrp_plat_memnode_info[node].end) {
595 				cur_end = lgrp_plat_memnode_info[node].end;
596 				if (mnode_xwa > 1) {
597 					/*
598 					 * sacrifice the last page in each
599 					 * node to eliminate large pages
600 					 * that span more than 1 memory node.
601 					 */
602 					endcnt = 1;
603 					physinstalled--;
604 				}
605 			}
606 
607 			mem_node_add_slice(cur_start, cur_end - endcnt);
608 
609 			/*
610 			 * Next subrange starts after end of current one
611 			 */
612 			cur_start = cur_end + 1;
613 		} while (cur_end < end);
614 
615 		list = list->ml_next;
616 	}
617 	mem_node_physalign = 0;
618 	mem_node_pfn_shift = 0;
619 }
620 
621 
622 /*
623  * plat_mnode_xcheck: checks the node memory ranges to see if there is a pfncnt
624  * range of pages aligned on pfncnt that crosses an node boundary. Returns 1 if
625  * a crossing is found and returns 0 otherwise.
626  */
627 int
628 plat_mnode_xcheck(pfn_t pfncnt)
629 {
630 	int	node, prevnode = -1, basenode;
631 	pfn_t	ea, sa;
632 
633 	for (node = 0; node < lgrp_plat_max_mem_node; node++) {
634 
635 		if (lgrp_plat_memnode_info[node].exists == 0)
636 			continue;
637 
638 		if (prevnode == -1) {
639 			prevnode = node;
640 			basenode = node;
641 			continue;
642 		}
643 
644 		/* assume x86 node pfn ranges are in increasing order */
645 		ASSERT(lgrp_plat_memnode_info[node].start >
646 		    lgrp_plat_memnode_info[prevnode].end);
647 
648 		/*
649 		 * continue if the starting address of node is not contiguous
650 		 * with the previous node.
651 		 */
652 
653 		if (lgrp_plat_memnode_info[node].start !=
654 		    (lgrp_plat_memnode_info[prevnode].end + 1)) {
655 			basenode = node;
656 			prevnode = node;
657 			continue;
658 		}
659 
660 		/* check if the starting address of node is pfncnt aligned */
661 		if ((lgrp_plat_memnode_info[node].start & (pfncnt - 1)) != 0) {
662 
663 			/*
664 			 * at this point, node starts at an unaligned boundary
665 			 * and is contiguous with the previous node(s) to
666 			 * basenode. Check if there is an aligned contiguous
667 			 * range of length pfncnt that crosses this boundary.
668 			 */
669 
670 			sa = P2ALIGN(lgrp_plat_memnode_info[prevnode].end,
671 			    pfncnt);
672 			ea = P2ROUNDUP((lgrp_plat_memnode_info[node].start),
673 			    pfncnt);
674 
675 			ASSERT((ea - sa) == pfncnt);
676 			if (sa >= lgrp_plat_memnode_info[basenode].start &&
677 			    ea <= (lgrp_plat_memnode_info[node].end + 1)) {
678 				/*
679 				 * large page found to cross mnode boundary.
680 				 * Return Failure if workaround not enabled.
681 				 */
682 				if (mnode_xwa == 0)
683 					return (1);
684 				mnode_xwa++;
685 			}
686 		}
687 		prevnode = node;
688 	}
689 	return (0);
690 }
691 
692 
693 lgrp_handle_t
694 plat_mem_node_to_lgrphand(int mnode)
695 {
696 	if (max_mem_nodes == 1)
697 		return (LGRP_DEFAULT_HANDLE);
698 
699 	ASSERT(0 <= mnode && mnode < lgrp_plat_max_mem_node);
700 
701 	return ((lgrp_handle_t)(lgrp_plat_memnode_info[mnode].lgrphand));
702 }
703 
704 int
705 plat_pfn_to_mem_node(pfn_t pfn)
706 {
707 	int	node;
708 
709 	if (max_mem_nodes == 1)
710 		return (0);
711 
712 	for (node = 0; node < lgrp_plat_max_mem_node; node++) {
713 		/*
714 		 * Skip nodes with no memory
715 		 */
716 		if (!lgrp_plat_memnode_info[node].exists)
717 			continue;
718 
719 		membar_consumer();
720 		if (pfn >= lgrp_plat_memnode_info[node].start &&
721 		    pfn <= lgrp_plat_memnode_info[node].end)
722 			return (node);
723 	}
724 
725 	/*
726 	 * Didn't find memnode where this PFN lives which should never happen
727 	 */
728 	ASSERT(node < lgrp_plat_max_mem_node);
729 	return (-1);
730 }
731 
732 
733 /*
734  * LGROUP PLATFORM INTERFACE ROUTINES
735  */
736 
737 /*
738  * Allocate additional space for an lgroup.
739  */
740 lgrp_t *
741 lgrp_plat_alloc(lgrp_id_t lgrpid)
742 {
743 	lgrp_t *lgrp;
744 
745 	lgrp = &lgrp_space[nlgrps_alloc++];
746 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
747 		return (NULL);
748 	return (lgrp);
749 }
750 
751 
752 /*
753  * Platform handling for (re)configuration changes
754  *
755  * Mechanism to protect lgrp_plat_cpu_node[] at CPU hotplug:
756  * 1) Use cpu_lock to synchronize between lgrp_plat_config() and
757  *    lgrp_plat_cpu_to_hand().
758  * 2) Disable latency probing logic by making sure that the flag
759  *    LGRP_PLAT_PROBE_ENABLE is cleared.
760  *
761  * Mechanism to protect lgrp_plat_memnode_info[] at memory hotplug:
762  * 1) Only inserts into lgrp_plat_memnode_info at memory hotplug, no removal.
763  * 2) Only expansion to existing entries, no shrinking.
764  * 3) On writing side, DR framework ensures that lgrp_plat_config() is called
765  *    in single-threaded context. And membar_producer() is used to ensure that
766  *    all changes are visible to other CPUs before setting the "exists" flag.
767  * 4) On reading side, membar_consumer() after checking the "exists" flag
768  *    ensures that right values are retrieved.
769  *
770  * Mechanism to protect lgrp_plat_node_domain[] at hotplug:
771  * 1) Only insertion into lgrp_plat_node_domain at hotplug, no removal.
772  * 2) On writing side, it's single-threaded and membar_producer() is used to
773  *    ensure all changes are visible to other CPUs before setting the "exists"
774  *    flag.
775  * 3) On reading side, membar_consumer() after checking the "exists" flag
776  *    ensures that right values are retrieved.
777  */
778 void
779 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
780 {
781 #ifdef	__xpv
782 	_NOTE(ARGUNUSED(flag, arg));
783 #else
784 	int	rc, node;
785 	cpu_t	*cp;
786 	void	*hdl = NULL;
787 	uchar_t	*sliptr = NULL;
788 	uint32_t domain, apicid, slicnt = 0;
789 	update_membounds_t *mp;
790 
791 	extern int acpidev_dr_get_cpu_numa_info(cpu_t *, void **, uint32_t *,
792 	    uint32_t *, uint32_t *, uchar_t **);
793 	extern void acpidev_dr_free_cpu_numa_info(void *);
794 
795 	/*
796 	 * This interface is used to support CPU/memory DR operations.
797 	 * Don't bother here if it's still during boot or only one lgrp node
798 	 * is supported.
799 	 */
800 	if (!lgrp_topo_initialized || lgrp_plat_node_cnt == 1)
801 		return;
802 
803 	switch (flag) {
804 	case LGRP_CONFIG_CPU_ADD:
805 		cp = (cpu_t *)arg;
806 		ASSERT(cp != NULL);
807 		ASSERT(MUTEX_HELD(&cpu_lock));
808 
809 		/* Check whether CPU already exists. */
810 		ASSERT(!lgrp_plat_cpu_node[cp->cpu_id].exists);
811 		if (lgrp_plat_cpu_node[cp->cpu_id].exists) {
812 			cmn_err(CE_WARN,
813 			    "!lgrp: CPU(%d) already exists in cpu_node map.",
814 			    cp->cpu_id);
815 			break;
816 		}
817 
818 		/* Query CPU lgrp information. */
819 		rc = acpidev_dr_get_cpu_numa_info(cp, &hdl, &apicid, &domain,
820 		    &slicnt, &sliptr);
821 		ASSERT(rc == 0);
822 		if (rc != 0) {
823 			cmn_err(CE_WARN,
824 			    "!lgrp: failed to query lgrp info for CPU(%d).",
825 			    cp->cpu_id);
826 			break;
827 		}
828 
829 		/* Update node to proximity domain mapping */
830 		node = lgrp_plat_domain_to_node(lgrp_plat_node_domain,
831 		    lgrp_plat_node_cnt, domain);
832 		if (node == -1) {
833 			node = lgrp_plat_node_domain_update(
834 			    lgrp_plat_node_domain, lgrp_plat_node_cnt, domain);
835 			ASSERT(node != -1);
836 			if (node == -1) {
837 				acpidev_dr_free_cpu_numa_info(hdl);
838 				cmn_err(CE_WARN, "!lgrp: failed to update "
839 				    "node_domain map for domain(%u).", domain);
840 				break;
841 			}
842 		}
843 
844 		/* Update latency information among lgrps. */
845 		if (slicnt != 0 && sliptr != NULL) {
846 			if (lgrp_plat_process_sli(domain, sliptr, slicnt,
847 			    lgrp_plat_node_domain, lgrp_plat_node_cnt,
848 			    &lgrp_plat_lat_stats) != 0) {
849 				cmn_err(CE_WARN, "!lgrp: failed to update "
850 				    "latency information for domain (%u).",
851 				    domain);
852 			}
853 		}
854 
855 		/* Update CPU to node mapping. */
856 		lgrp_plat_cpu_node[cp->cpu_id].prox_domain = domain;
857 		lgrp_plat_cpu_node[cp->cpu_id].node = node;
858 		lgrp_plat_cpu_node[cp->cpu_id].apicid = apicid;
859 		lgrp_plat_cpu_node[cp->cpu_id].exists = 1;
860 		lgrp_plat_apic_ncpus++;
861 
862 		acpidev_dr_free_cpu_numa_info(hdl);
863 		break;
864 
865 	case LGRP_CONFIG_CPU_DEL:
866 		cp = (cpu_t *)arg;
867 		ASSERT(cp != NULL);
868 		ASSERT(MUTEX_HELD(&cpu_lock));
869 
870 		/* Check whether CPU exists. */
871 		ASSERT(lgrp_plat_cpu_node[cp->cpu_id].exists);
872 		if (!lgrp_plat_cpu_node[cp->cpu_id].exists) {
873 			cmn_err(CE_WARN,
874 			    "!lgrp: CPU(%d) doesn't exist in cpu_node map.",
875 			    cp->cpu_id);
876 			break;
877 		}
878 
879 		/* Query CPU lgrp information. */
880 		rc = acpidev_dr_get_cpu_numa_info(cp, &hdl, &apicid, &domain,
881 		    NULL, NULL);
882 		ASSERT(rc == 0);
883 		if (rc != 0) {
884 			cmn_err(CE_WARN,
885 			    "!lgrp: failed to query lgrp info for CPU(%d).",
886 			    cp->cpu_id);
887 			break;
888 		}
889 
890 		/* Update map. */
891 		ASSERT(lgrp_plat_cpu_node[cp->cpu_id].apicid == apicid);
892 		ASSERT(lgrp_plat_cpu_node[cp->cpu_id].prox_domain == domain);
893 		lgrp_plat_cpu_node[cp->cpu_id].exists = 0;
894 		lgrp_plat_cpu_node[cp->cpu_id].apicid = UINT32_MAX;
895 		lgrp_plat_cpu_node[cp->cpu_id].prox_domain = UINT32_MAX;
896 		lgrp_plat_cpu_node[cp->cpu_id].node = UINT_MAX;
897 		lgrp_plat_apic_ncpus--;
898 
899 		acpidev_dr_free_cpu_numa_info(hdl);
900 		break;
901 
902 	case LGRP_CONFIG_MEM_ADD:
903 		mp = (update_membounds_t *)arg;
904 		ASSERT(mp != NULL);
905 
906 		/* Update latency information among lgrps. */
907 		if (mp->u_sli_cnt != 0 && mp->u_sli_ptr != NULL) {
908 			if (lgrp_plat_process_sli(mp->u_domain,
909 			    mp->u_sli_ptr, mp->u_sli_cnt,
910 			    lgrp_plat_node_domain, lgrp_plat_node_cnt,
911 			    &lgrp_plat_lat_stats) != 0) {
912 				cmn_err(CE_WARN, "!lgrp: failed to update "
913 				    "latency information for domain (%u).",
914 				    domain);
915 			}
916 		}
917 
918 		if (lgrp_plat_memnode_info_update(lgrp_plat_node_domain,
919 		    lgrp_plat_node_cnt, lgrp_plat_memnode_info, max_mem_nodes,
920 		    mp->u_base, mp->u_base + mp->u_length,
921 		    mp->u_domain, mp->u_device_id) < 0) {
922 			cmn_err(CE_WARN,
923 			    "!lgrp: failed to update latency  information for "
924 			    "memory (0x%" PRIx64 " - 0x%" PRIx64 ").",
925 			    mp->u_base, mp->u_base + mp->u_length);
926 		}
927 		break;
928 
929 	default:
930 		break;
931 	}
932 #endif	/* __xpv */
933 }
934 
935 
936 /*
937  * Return the platform handle for the lgroup containing the given CPU
938  */
939 lgrp_handle_t
940 lgrp_plat_cpu_to_hand(processorid_t id)
941 {
942 	lgrp_handle_t	hand;
943 
944 	ASSERT(!lgrp_initialized || MUTEX_HELD(&cpu_lock));
945 
946 	if (lgrp_plat_node_cnt == 1)
947 		return (LGRP_DEFAULT_HANDLE);
948 
949 	hand = (lgrp_handle_t)lgrp_plat_cpu_to_node(cpu[id],
950 	    lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries);
951 
952 	ASSERT(hand != (lgrp_handle_t)-1);
953 	if (hand == (lgrp_handle_t)-1)
954 		return (LGRP_NULL_HANDLE);
955 
956 	return (hand);
957 }
958 
959 
960 /*
961  * Platform-specific initialization of lgroups
962  */
963 void
964 lgrp_plat_init(lgrp_init_stages_t stage)
965 {
966 #if defined(__xpv)
967 #else	/* __xpv */
968 	u_longlong_t	value;
969 #endif	/* __xpv */
970 
971 	switch (stage) {
972 	case LGRP_INIT_STAGE1:
973 #if defined(__xpv)
974 		/*
975 		 * XXPV	For now, the hypervisor treats all memory equally.
976 		 */
977 		lgrp_plat_node_cnt = max_mem_nodes = 1;
978 #else	/* __xpv */
979 
980 		/*
981 		 * Get boot property for lgroup topology height limit
982 		 */
983 		if (bootprop_getval(BP_LGRP_TOPO_LEVELS, &value) == 0)
984 			(void) lgrp_topo_ht_limit_set((int)value);
985 
986 		/*
987 		 * Get boot property for enabling/disabling SRAT
988 		 */
989 		if (bootprop_getval(BP_LGRP_SRAT_ENABLE, &value) == 0)
990 			lgrp_plat_srat_enable = (int)value;
991 
992 		/*
993 		 * Get boot property for enabling/disabling SLIT
994 		 */
995 		if (bootprop_getval(BP_LGRP_SLIT_ENABLE, &value) == 0)
996 			lgrp_plat_slit_enable = (int)value;
997 
998 		/*
999 		 * Get boot property for enabling/disabling MSCT
1000 		 */
1001 		if (bootprop_getval(BP_LGRP_MSCT_ENABLE, &value) == 0)
1002 			lgrp_plat_msct_enable = (int)value;
1003 
1004 		/*
1005 		 * Initialize as a UMA machine
1006 		 */
1007 		if (lgrp_topo_ht_limit() == 1) {
1008 			lgrp_plat_node_cnt = max_mem_nodes = 1;
1009 			lgrp_plat_max_mem_node = 1;
1010 			return;
1011 		}
1012 
1013 		lgrp_plat_get_numa_config();
1014 
1015 		/*
1016 		 * Each lgrp node needs MAX_MEM_NODES_PER_LGROUP memnodes
1017 		 * to support memory DR operations if memory DR is enabled.
1018 		 */
1019 		lgrp_plat_max_mem_node = lgrp_plat_node_cnt;
1020 		if (plat_dr_support_memory() && lgrp_plat_node_cnt != 1) {
1021 			max_mem_nodes = MAX_MEM_NODES_PER_LGROUP *
1022 			    lgrp_plat_node_cnt;
1023 			ASSERT(max_mem_nodes <= MAX_MEM_NODES);
1024 		}
1025 #endif	/* __xpv */
1026 		break;
1027 
1028 	case LGRP_INIT_STAGE3:
1029 		lgrp_plat_probe();
1030 		lgrp_plat_release_bootstrap();
1031 		break;
1032 
1033 	case LGRP_INIT_STAGE4:
1034 		lgrp_plat_main_init();
1035 		break;
1036 
1037 	default:
1038 		break;
1039 	}
1040 }
1041 
1042 
1043 /*
1044  * Return latency between "from" and "to" lgroups
1045  *
1046  * This latency number can only be used for relative comparison
1047  * between lgroups on the running system, cannot be used across platforms,
1048  * and may not reflect the actual latency.  It is platform and implementation
1049  * specific, so platform gets to decide its value.  It would be nice if the
1050  * number was at least proportional to make comparisons more meaningful though.
1051  */
1052 int
1053 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1054 {
1055 	lgrp_handle_t	src, dest;
1056 	int		node;
1057 
1058 	if (max_mem_nodes == 1)
1059 		return (0);
1060 
1061 	/*
1062 	 * Return max latency for root lgroup
1063 	 */
1064 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1065 		return (lgrp_plat_lat_stats.latency_max);
1066 
1067 	src = from;
1068 	dest = to;
1069 
1070 	/*
1071 	 * Return 0 for nodes (lgroup platform handles) out of range
1072 	 */
1073 	if (src >= MAX_NODES || dest >= MAX_NODES)
1074 		return (0);
1075 
1076 	/*
1077 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1078 	 * and we are trying to get latency from current CPU to some node.
1079 	 * Avoid probing if CPU/memory DR is enabled.
1080 	 */
1081 	if (lgrp_plat_lat_stats.latencies[src][src] == 0) {
1082 		/*
1083 		 * Latency information should be updated by lgrp_plat_config()
1084 		 * for DR operations. Something is wrong if reaches here.
1085 		 * For safety, flatten lgrp topology to two levels.
1086 		 */
1087 		if (plat_dr_support_cpu() || plat_dr_support_memory()) {
1088 			ASSERT(lgrp_plat_lat_stats.latencies[src][src]);
1089 			cmn_err(CE_WARN,
1090 			    "lgrp: failed to get latency information, "
1091 			    "fall back to two-level topology.");
1092 			lgrp_plat_2level_setup(&lgrp_plat_lat_stats);
1093 		} else {
1094 			node = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
1095 			    lgrp_plat_cpu_node_nentries);
1096 			ASSERT(node >= 0 && node < lgrp_plat_node_cnt);
1097 			if (node == src)
1098 				lgrp_plat_probe();
1099 		}
1100 	}
1101 
1102 	return (lgrp_plat_lat_stats.latencies[src][dest]);
1103 }
1104 
1105 
1106 /*
1107  * Return the maximum number of lgrps supported by the platform.
1108  * Before lgrp topology is known it returns an estimate based on the number of
1109  * nodes. Once topology is known it returns:
1110  * 1) the actual maximim number of lgrps created if CPU/memory DR operations
1111  *    are not suppported.
1112  * 2) the maximum possible number of lgrps if CPU/memory DR operations are
1113  *    supported.
1114  */
1115 int
1116 lgrp_plat_max_lgrps(void)
1117 {
1118 	if (!lgrp_topo_initialized || plat_dr_support_cpu() ||
1119 	    plat_dr_support_memory()) {
1120 		return (lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1121 	} else {
1122 		return (lgrp_alloc_max + 1);
1123 	}
1124 }
1125 
1126 
1127 /*
1128  * Count number of memory pages (_t) based on mnode id (_n) and query type (_t).
1129  */
1130 #define	_LGRP_PLAT_MEM_SIZE(_n, _q, _t)					\
1131 	if (mem_node_config[_n].exists) {				\
1132 		switch (_q) {						\
1133 		case LGRP_MEM_SIZE_FREE:				\
1134 			_t += MNODE_PGCNT(_n);				\
1135 			break;						\
1136 		case LGRP_MEM_SIZE_AVAIL:				\
1137 			_t += mem_node_memlist_pages(_n, phys_avail);	\
1138 				break;					\
1139 		case LGRP_MEM_SIZE_INSTALL:				\
1140 			_t += mem_node_memlist_pages(_n, phys_install);	\
1141 			break;						\
1142 		default:						\
1143 			break;						\
1144 		}							\
1145 	}
1146 
1147 /*
1148  * Return the number of free pages in an lgroup.
1149  *
1150  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1151  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1152  * number of allocatable base pagesize pages corresponding to the
1153  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1154  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1155  * memory installed, regardless of whether or not it's usable.
1156  */
1157 pgcnt_t
1158 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1159 {
1160 	int	mnode;
1161 	pgcnt_t npgs = (pgcnt_t)0;
1162 	extern struct memlist *phys_avail;
1163 	extern struct memlist *phys_install;
1164 
1165 
1166 	if (plathand == LGRP_DEFAULT_HANDLE)
1167 		return (lgrp_plat_mem_size_default(plathand, query));
1168 
1169 	if (plathand != LGRP_NULL_HANDLE) {
1170 		/* Count memory node present at boot. */
1171 		mnode = (int)plathand;
1172 		ASSERT(mnode < lgrp_plat_node_cnt);
1173 		_LGRP_PLAT_MEM_SIZE(mnode, query, npgs);
1174 
1175 		/* Count possible hot-added memory nodes. */
1176 		for (mnode = lgrp_plat_node_cnt;
1177 		    mnode < lgrp_plat_max_mem_node; mnode++) {
1178 			if (lgrp_plat_memnode_info[mnode].lgrphand == plathand)
1179 				_LGRP_PLAT_MEM_SIZE(mnode, query, npgs);
1180 		}
1181 	}
1182 
1183 	return (npgs);
1184 }
1185 
1186 
1187 /*
1188  * Return the platform handle of the lgroup that contains the physical memory
1189  * corresponding to the given page frame number
1190  */
1191 lgrp_handle_t
1192 lgrp_plat_pfn_to_hand(pfn_t pfn)
1193 {
1194 	int	mnode;
1195 
1196 	if (max_mem_nodes == 1)
1197 		return (LGRP_DEFAULT_HANDLE);
1198 
1199 	if (pfn > physmax)
1200 		return (LGRP_NULL_HANDLE);
1201 
1202 	mnode = plat_pfn_to_mem_node(pfn);
1203 	if (mnode < 0)
1204 		return (LGRP_NULL_HANDLE);
1205 
1206 	return (MEM_NODE_2_LGRPHAND(mnode));
1207 }
1208 
1209 
1210 /*
1211  * Probe memory in each node from current CPU to determine latency topology
1212  *
1213  * The probing code will probe the vendor ID register on the Northbridge of
1214  * Opteron processors and probe memory for other processors by default.
1215  *
1216  * Since probing is inherently error prone, the code takes laps across all the
1217  * nodes probing from each node to each of the other nodes some number of
1218  * times.  Furthermore, each node is probed some number of times before moving
1219  * onto the next one during each lap.  The minimum latency gotten between nodes
1220  * is kept as the latency between the nodes.
1221  *
1222  * After all that,  the probe times are adjusted by normalizing values that are
1223  * close to each other and local latencies are made the same.  Lastly, the
1224  * latencies are verified to make sure that certain conditions are met (eg.
1225  * local < remote, latency(a, b) == latency(b, a), etc.).
1226  *
1227  * If any of the conditions aren't met, the code will export a NUMA
1228  * configuration with the local CPUs and memory given by the SRAT or PCI config
1229  * space registers and one remote memory latency since it can't tell exactly
1230  * how far each node is from each other.
1231  */
1232 void
1233 lgrp_plat_probe(void)
1234 {
1235 	int				from;
1236 	int				i;
1237 	lgrp_plat_latency_stats_t	*lat_stats;
1238 	boolean_t			probed;
1239 	hrtime_t			probe_time;
1240 	int				to;
1241 
1242 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
1243 	    max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1244 		return;
1245 
1246 	/* SRAT and SLIT should be enabled if DR operations are enabled. */
1247 	if (plat_dr_support_cpu() || plat_dr_support_memory())
1248 		return;
1249 
1250 	/*
1251 	 * Determine ID of node containing current CPU
1252 	 */
1253 	from = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
1254 	    lgrp_plat_cpu_node_nentries);
1255 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
1256 	if (srat_ptr && lgrp_plat_srat_enable && !lgrp_plat_srat_error)
1257 		ASSERT(lgrp_plat_node_domain[from].exists);
1258 
1259 	/*
1260 	 * Don't need to probe if got times already
1261 	 */
1262 	lat_stats = &lgrp_plat_lat_stats;
1263 	if (lat_stats->latencies[from][from] != 0)
1264 		return;
1265 
1266 	/*
1267 	 * Read vendor ID in Northbridge or read and write page(s)
1268 	 * in each node from current CPU and remember how long it takes,
1269 	 * so we can build latency topology of machine later.
1270 	 * This should approximate the memory latency between each node.
1271 	 */
1272 	probed = B_FALSE;
1273 	for (i = 0; i < lgrp_plat_probe_nrounds; i++) {
1274 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1275 			/*
1276 			 * Get probe time and skip over any nodes that can't be
1277 			 * probed yet or don't have memory
1278 			 */
1279 			probe_time = lgrp_plat_probe_time(to,
1280 			    lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries,
1281 			    &lgrp_plat_probe_mem_config, &lgrp_plat_lat_stats,
1282 			    &lgrp_plat_probe_stats);
1283 			if (probe_time == 0)
1284 				continue;
1285 
1286 			probed = B_TRUE;
1287 
1288 			/*
1289 			 * Keep lowest probe time as latency between nodes
1290 			 */
1291 			if (lat_stats->latencies[from][to] == 0 ||
1292 			    probe_time < lat_stats->latencies[from][to])
1293 				lat_stats->latencies[from][to] = probe_time;
1294 
1295 			/*
1296 			 * Update overall minimum and maximum probe times
1297 			 * across all nodes
1298 			 */
1299 			if (probe_time < lat_stats->latency_min ||
1300 			    lat_stats->latency_min == -1)
1301 				lat_stats->latency_min = probe_time;
1302 			if (probe_time > lat_stats->latency_max)
1303 				lat_stats->latency_max = probe_time;
1304 		}
1305 	}
1306 
1307 	/*
1308 	 * Bail out if weren't able to probe any nodes from current CPU
1309 	 */
1310 	if (probed == B_FALSE)
1311 		return;
1312 
1313 	/*
1314 	 * - Fix up latencies such that local latencies are same,
1315 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1316 	 *
1317 	 * - Verify that latencies look ok
1318 	 *
1319 	 * - Fallback to just optimizing for local and remote if
1320 	 *   latencies didn't look right
1321 	 */
1322 	lgrp_plat_latency_adjust(lgrp_plat_memnode_info, &lgrp_plat_lat_stats,
1323 	    &lgrp_plat_probe_stats);
1324 	lgrp_plat_probe_stats.probe_error_code =
1325 	    lgrp_plat_latency_verify(lgrp_plat_memnode_info,
1326 	    &lgrp_plat_lat_stats);
1327 	if (lgrp_plat_probe_stats.probe_error_code)
1328 		lgrp_plat_2level_setup(&lgrp_plat_lat_stats);
1329 }
1330 
1331 
1332 /*
1333  * Return platform handle for root lgroup
1334  */
1335 lgrp_handle_t
1336 lgrp_plat_root_hand(void)
1337 {
1338 	return (LGRP_DEFAULT_HANDLE);
1339 }
1340 
1341 
1342 /*
1343  * INTERNAL ROUTINES
1344  */
1345 
1346 
1347 /*
1348  * Update CPU to node mapping for given CPU and proximity domain.
1349  * Return values:
1350  *	- zero for success
1351  *	- positive numbers for warnings
1352  *	- negative numbers for errors
1353  */
1354 static int
1355 lgrp_plat_cpu_node_update(node_domain_map_t *node_domain, int node_cnt,
1356     cpu_node_map_t *cpu_node, int nentries, uint32_t apicid, uint32_t domain)
1357 {
1358 	uint_t	i;
1359 	int	node;
1360 
1361 	/*
1362 	 * Get node number for proximity domain
1363 	 */
1364 	node = lgrp_plat_domain_to_node(node_domain, node_cnt, domain);
1365 	if (node == -1) {
1366 		node = lgrp_plat_node_domain_update(node_domain, node_cnt,
1367 		    domain);
1368 		if (node == -1)
1369 			return (-1);
1370 	}
1371 
1372 	/*
1373 	 * Search for entry with given APIC ID and fill in its node and
1374 	 * proximity domain IDs (if they haven't been set already)
1375 	 */
1376 	for (i = 0; i < nentries; i++) {
1377 		/*
1378 		 * Skip nonexistent entries and ones without matching APIC ID
1379 		 */
1380 		if (!cpu_node[i].exists || cpu_node[i].apicid != apicid)
1381 			continue;
1382 
1383 		/*
1384 		 * Just return if entry completely and correctly filled in
1385 		 * already
1386 		 */
1387 		if (cpu_node[i].prox_domain == domain &&
1388 		    cpu_node[i].node == node)
1389 			return (1);
1390 
1391 		/*
1392 		 * It's invalid to have more than one entry with the same
1393 		 * local APIC ID in SRAT table.
1394 		 */
1395 		if (cpu_node[i].node != UINT_MAX)
1396 			return (-2);
1397 
1398 		/*
1399 		 * Fill in node and proximity domain IDs
1400 		 */
1401 		cpu_node[i].prox_domain = domain;
1402 		cpu_node[i].node = node;
1403 
1404 		return (0);
1405 	}
1406 
1407 	/*
1408 	 * It's possible that an apicid doesn't exist in the cpu_node map due
1409 	 * to user limits number of CPUs powered on at boot by specifying the
1410 	 * boot_ncpus kernel option.
1411 	 */
1412 	return (2);
1413 }
1414 
1415 
1416 /*
1417  * Get node ID for given CPU
1418  */
1419 static int
1420 lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node,
1421     int cpu_node_nentries)
1422 {
1423 	processorid_t	cpuid;
1424 
1425 	if (cp == NULL)
1426 		return (-1);
1427 
1428 	cpuid = cp->cpu_id;
1429 	if (cpuid < 0 || cpuid >= max_ncpus)
1430 		return (-1);
1431 
1432 	/*
1433 	 * SRAT doesn't exist, isn't enabled, or there was an error processing
1434 	 * it, so return node ID for Opteron and -1 otherwise.
1435 	 */
1436 	if (srat_ptr == NULL || !lgrp_plat_srat_enable ||
1437 	    lgrp_plat_srat_error) {
1438 		if (is_opteron())
1439 			return (pg_plat_hw_instance_id(cp, PGHW_PROCNODE));
1440 		return (-1);
1441 	}
1442 
1443 	/*
1444 	 * Return -1 when CPU to node ID mapping entry doesn't exist for given
1445 	 * CPU
1446 	 */
1447 	if (cpuid >= cpu_node_nentries || !cpu_node[cpuid].exists)
1448 		return (-1);
1449 
1450 	return (cpu_node[cpuid].node);
1451 }
1452 
1453 
1454 /*
1455  * Return node number for given proximity domain/system locality
1456  */
1457 static int
1458 lgrp_plat_domain_to_node(node_domain_map_t *node_domain, int node_cnt,
1459     uint32_t domain)
1460 {
1461 	uint_t	node;
1462 	uint_t	start;
1463 
1464 	/*
1465 	 * Hash proximity domain ID into node to domain mapping table (array),
1466 	 * search for entry with matching proximity domain ID, and return index
1467 	 * of matching entry as node ID.
1468 	 */
1469 	node = start = NODE_DOMAIN_HASH(domain, node_cnt);
1470 	do {
1471 		if (node_domain[node].exists) {
1472 			membar_consumer();
1473 			if (node_domain[node].prox_domain == domain)
1474 				return (node);
1475 		}
1476 		node = (node + 1) % node_cnt;
1477 	} while (node != start);
1478 	return (-1);
1479 }
1480 
1481 
1482 /*
1483  * Get NUMA configuration of machine
1484  */
1485 static void
1486 lgrp_plat_get_numa_config(void)
1487 {
1488 	uint_t		probe_op;
1489 
1490 	/*
1491 	 * Read boot property with CPU to APIC ID mapping table/array to
1492 	 * determine number of CPUs
1493 	 */
1494 	lgrp_plat_apic_ncpus = lgrp_plat_process_cpu_apicids(NULL);
1495 
1496 	/*
1497 	 * Determine which CPUs and memory are local to each other and number
1498 	 * of NUMA nodes by reading ACPI System Resource Affinity Table (SRAT)
1499 	 */
1500 	if (lgrp_plat_apic_ncpus > 0) {
1501 		int	retval;
1502 
1503 		/* Reserve enough resources if CPU DR is enabled. */
1504 		if (plat_dr_support_cpu() && max_ncpus > lgrp_plat_apic_ncpus)
1505 			lgrp_plat_cpu_node_nentries = max_ncpus;
1506 		else
1507 			lgrp_plat_cpu_node_nentries = lgrp_plat_apic_ncpus;
1508 
1509 		/*
1510 		 * Temporarily allocate boot memory to use for CPU to node
1511 		 * mapping since kernel memory allocator isn't alive yet
1512 		 */
1513 		lgrp_plat_cpu_node = (cpu_node_map_t *)BOP_ALLOC(bootops,
1514 		    NULL, lgrp_plat_cpu_node_nentries * sizeof (cpu_node_map_t),
1515 		    sizeof (int));
1516 
1517 		ASSERT(lgrp_plat_cpu_node != NULL);
1518 		if (lgrp_plat_cpu_node) {
1519 			bzero(lgrp_plat_cpu_node, lgrp_plat_cpu_node_nentries *
1520 			    sizeof (cpu_node_map_t));
1521 		} else {
1522 			lgrp_plat_cpu_node_nentries = 0;
1523 		}
1524 
1525 		/*
1526 		 * Fill in CPU to node ID mapping table with APIC ID for each
1527 		 * CPU
1528 		 */
1529 		(void) lgrp_plat_process_cpu_apicids(lgrp_plat_cpu_node);
1530 
1531 		retval = lgrp_plat_process_srat(srat_ptr, msct_ptr,
1532 		    &lgrp_plat_prox_domain_min,
1533 		    lgrp_plat_node_domain, lgrp_plat_cpu_node,
1534 		    lgrp_plat_apic_ncpus, lgrp_plat_memnode_info);
1535 		if (retval <= 0) {
1536 			lgrp_plat_srat_error = retval;
1537 			lgrp_plat_node_cnt = 1;
1538 		} else {
1539 			lgrp_plat_srat_error = 0;
1540 			lgrp_plat_node_cnt = retval;
1541 		}
1542 	}
1543 
1544 	/*
1545 	 * Try to use PCI config space registers on Opteron if there's an error
1546 	 * processing CPU to APIC ID mapping or SRAT
1547 	 */
1548 	if ((lgrp_plat_apic_ncpus <= 0 || lgrp_plat_srat_error != 0) &&
1549 	    is_opteron())
1550 		opt_get_numa_config(&lgrp_plat_node_cnt, &lgrp_plat_mem_intrlv,
1551 		    lgrp_plat_memnode_info);
1552 
1553 	/*
1554 	 * Don't bother to setup system for multiple lgroups and only use one
1555 	 * memory node when memory is interleaved between any nodes or there is
1556 	 * only one NUMA node
1557 	 */
1558 	if (lgrp_plat_mem_intrlv || lgrp_plat_node_cnt == 1) {
1559 		lgrp_plat_node_cnt = max_mem_nodes = 1;
1560 		(void) lgrp_topo_ht_limit_set(1);
1561 		return;
1562 	}
1563 
1564 	/*
1565 	 * Leaf lgroups on x86/x64 architectures contain one physical
1566 	 * processor chip. Tune lgrp_expand_proc_thresh and
1567 	 * lgrp_expand_proc_diff so that lgrp_choose() will spread
1568 	 * things out aggressively.
1569 	 */
1570 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
1571 	lgrp_expand_proc_diff = 0;
1572 
1573 	/*
1574 	 * There should be one memnode (physical page free list(s)) for
1575 	 * each node if memory DR is disabled.
1576 	 */
1577 	max_mem_nodes = lgrp_plat_node_cnt;
1578 
1579 	/*
1580 	 * Initialize min and max latency before reading SLIT or probing
1581 	 */
1582 	lgrp_plat_lat_stats.latency_min = -1;
1583 	lgrp_plat_lat_stats.latency_max = 0;
1584 
1585 	/*
1586 	 * Determine how far each NUMA node is from each other by
1587 	 * reading ACPI System Locality Information Table (SLIT) if it
1588 	 * exists
1589 	 */
1590 	lgrp_plat_slit_error = lgrp_plat_process_slit(slit_ptr,
1591 	    lgrp_plat_node_domain, lgrp_plat_node_cnt, lgrp_plat_memnode_info,
1592 	    &lgrp_plat_lat_stats);
1593 
1594 	/*
1595 	 * Disable support of CPU/memory DR operations if multiple locality
1596 	 * domains exist in system and either of following is true.
1597 	 * 1) Failed to process SLIT table.
1598 	 * 2) Latency probing is enabled by user.
1599 	 */
1600 	if (lgrp_plat_node_cnt > 1 &&
1601 	    (plat_dr_support_cpu() || plat_dr_support_memory())) {
1602 		if (!lgrp_plat_slit_enable || lgrp_plat_slit_error != 0 ||
1603 		    !lgrp_plat_srat_enable || lgrp_plat_srat_error != 0 ||
1604 		    lgrp_plat_apic_ncpus <= 0) {
1605 			cmn_err(CE_CONT,
1606 			    "?lgrp: failed to process ACPI SRAT/SLIT table, "
1607 			    "disable support of CPU/memory DR operations.");
1608 			plat_dr_disable_cpu();
1609 			plat_dr_disable_memory();
1610 		} else if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) {
1611 			cmn_err(CE_CONT,
1612 			    "?lgrp: latency probing enabled by user, "
1613 			    "disable support of CPU/memory DR operations.");
1614 			plat_dr_disable_cpu();
1615 			plat_dr_disable_memory();
1616 		}
1617 	}
1618 
1619 	/* Done if succeeded to process SLIT table. */
1620 	if (lgrp_plat_slit_error == 0)
1621 		return;
1622 
1623 	/*
1624 	 * Probe to determine latency between NUMA nodes when SLIT
1625 	 * doesn't exist or make sense
1626 	 */
1627 	lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_ENABLE;
1628 
1629 	/*
1630 	 * Specify whether to probe using vendor ID register or page copy
1631 	 * if hasn't been specified already or is overspecified
1632 	 */
1633 	probe_op = lgrp_plat_probe_flags &
1634 	    (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
1635 
1636 	if (probe_op == 0 ||
1637 	    probe_op == (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR)) {
1638 		lgrp_plat_probe_flags &=
1639 		    ~(LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
1640 		if (is_opteron())
1641 			lgrp_plat_probe_flags |=
1642 			    LGRP_PLAT_PROBE_VENDOR;
1643 		else
1644 			lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_PGCPY;
1645 	}
1646 
1647 	/*
1648 	 * Probing errors can mess up the lgroup topology and
1649 	 * force us fall back to a 2 level lgroup topology.
1650 	 * Here we bound how tall the lgroup topology can grow
1651 	 * in hopes of avoiding any anamolies in probing from
1652 	 * messing up the lgroup topology by limiting the
1653 	 * accuracy of the latency topology.
1654 	 *
1655 	 * Assume that nodes will at least be configured in a
1656 	 * ring, so limit height of lgroup topology to be less
1657 	 * than number of nodes on a system with 4 or more
1658 	 * nodes
1659 	 */
1660 	if (lgrp_plat_node_cnt >= 4 && lgrp_topo_ht_limit() ==
1661 	    lgrp_topo_ht_limit_default())
1662 		(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
1663 }
1664 
1665 
1666 /*
1667  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
1668  * be considered same
1669  */
1670 #define	LGRP_LAT_TOLERANCE_SHIFT	4
1671 
1672 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
1673 
1674 
1675 /*
1676  * Adjust latencies between nodes to be symmetric, normalize latencies between
1677  * any nodes that are within some tolerance to be same, and make local
1678  * latencies be same
1679  */
1680 static void
1681 lgrp_plat_latency_adjust(memnode_phys_addr_map_t *memnode_info,
1682     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1683 {
1684 	int				i;
1685 	int				j;
1686 	int				k;
1687 	int				l;
1688 	u_longlong_t			max;
1689 	u_longlong_t			min;
1690 	u_longlong_t			t;
1691 	u_longlong_t			t1;
1692 	u_longlong_t			t2;
1693 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
1694 	int				lat_corrected[MAX_NODES][MAX_NODES];
1695 
1696 	t = 0;
1697 	/*
1698 	 * Nothing to do when this is an UMA machine or don't have args needed
1699 	 */
1700 	if (max_mem_nodes == 1)
1701 		return;
1702 
1703 	ASSERT(memnode_info != NULL && lat_stats != NULL &&
1704 	    probe_stats != NULL);
1705 
1706 	/*
1707 	 * Make sure that latencies are symmetric between any two nodes
1708 	 * (ie. latency(node0, node1) == latency(node1, node0))
1709 	 */
1710 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1711 		if (!memnode_info[i].exists)
1712 			continue;
1713 
1714 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1715 			if (!memnode_info[j].exists)
1716 				continue;
1717 
1718 			t1 = lat_stats->latencies[i][j];
1719 			t2 = lat_stats->latencies[j][i];
1720 
1721 			if (t1 == 0 || t2 == 0 || t1 == t2)
1722 				continue;
1723 
1724 			/*
1725 			 * Latencies should be same
1726 			 * - Use minimum of two latencies which should be same
1727 			 * - Track suspect probe times not within tolerance of
1728 			 *   min value
1729 			 * - Remember how much values are corrected by
1730 			 */
1731 			if (t1 > t2) {
1732 				t = t2;
1733 				probe_stats->probe_errors[i][j] += t1 - t2;
1734 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
1735 					probe_stats->probe_suspect[i][j]++;
1736 					probe_stats->probe_suspect[j][i]++;
1737 				}
1738 			} else if (t2 > t1) {
1739 				t = t1;
1740 				probe_stats->probe_errors[j][i] += t2 - t1;
1741 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
1742 					probe_stats->probe_suspect[i][j]++;
1743 					probe_stats->probe_suspect[j][i]++;
1744 				}
1745 			}
1746 
1747 			lat_stats->latencies[i][j] =
1748 			    lat_stats->latencies[j][i] = t;
1749 			lgrp_config(cflag, t1, t);
1750 			lgrp_config(cflag, t2, t);
1751 		}
1752 	}
1753 
1754 	/*
1755 	 * Keep track of which latencies get corrected
1756 	 */
1757 	for (i = 0; i < MAX_NODES; i++)
1758 		for (j = 0; j < MAX_NODES; j++)
1759 			lat_corrected[i][j] = 0;
1760 
1761 	/*
1762 	 * For every two nodes, see whether there is another pair of nodes which
1763 	 * are about the same distance apart and make the latencies be the same
1764 	 * if they are close enough together
1765 	 */
1766 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1767 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1768 			if (!memnode_info[j].exists)
1769 				continue;
1770 			/*
1771 			 * Pick one pair of nodes (i, j)
1772 			 * and get latency between them
1773 			 */
1774 			t1 = lat_stats->latencies[i][j];
1775 
1776 			/*
1777 			 * Skip this pair of nodes if there isn't a latency
1778 			 * for it yet
1779 			 */
1780 			if (t1 == 0)
1781 				continue;
1782 
1783 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1784 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
1785 					if (!memnode_info[l].exists)
1786 						continue;
1787 					/*
1788 					 * Pick another pair of nodes (k, l)
1789 					 * not same as (i, j) and get latency
1790 					 * between them
1791 					 */
1792 					if (k == i && l == j)
1793 						continue;
1794 
1795 					t2 = lat_stats->latencies[k][l];
1796 
1797 					/*
1798 					 * Skip this pair of nodes if there
1799 					 * isn't a latency for it yet
1800 					 */
1801 
1802 					if (t2 == 0)
1803 						continue;
1804 
1805 					/*
1806 					 * Skip nodes (k, l) if they already
1807 					 * have same latency as (i, j) or
1808 					 * their latency isn't close enough to
1809 					 * be considered/made the same
1810 					 */
1811 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
1812 					    t1 >> lgrp_plat_probe_lt_shift) ||
1813 					    (t2 > t1 && t2 - t1 >
1814 					    t2 >> lgrp_plat_probe_lt_shift))
1815 						continue;
1816 
1817 					/*
1818 					 * Make latency(i, j) same as
1819 					 * latency(k, l), try to use latency
1820 					 * that has been adjusted already to get
1821 					 * more consistency (if possible), and
1822 					 * remember which latencies were
1823 					 * adjusted for next time
1824 					 */
1825 					if (lat_corrected[i][j]) {
1826 						t = t1;
1827 						lgrp_config(cflag, t2, t);
1828 						t2 = t;
1829 					} else if (lat_corrected[k][l]) {
1830 						t = t2;
1831 						lgrp_config(cflag, t1, t);
1832 						t1 = t;
1833 					} else {
1834 						if (t1 > t2)
1835 							t = t2;
1836 						else
1837 							t = t1;
1838 						lgrp_config(cflag, t1, t);
1839 						lgrp_config(cflag, t2, t);
1840 						t1 = t2 = t;
1841 					}
1842 
1843 					lat_stats->latencies[i][j] =
1844 					    lat_stats->latencies[k][l] = t;
1845 
1846 					lat_corrected[i][j] =
1847 					    lat_corrected[k][l] = 1;
1848 				}
1849 			}
1850 		}
1851 	}
1852 
1853 	/*
1854 	 * Local latencies should be same
1855 	 * - Find min and max local latencies
1856 	 * - Make all local latencies be minimum
1857 	 */
1858 	min = -1;
1859 	max = 0;
1860 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1861 		if (!memnode_info[i].exists)
1862 			continue;
1863 		t = lat_stats->latencies[i][i];
1864 		if (t == 0)
1865 			continue;
1866 		if (min == -1 || t < min)
1867 			min = t;
1868 		if (t > max)
1869 			max = t;
1870 	}
1871 	if (min != max) {
1872 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1873 			int	local;
1874 
1875 			if (!memnode_info[i].exists)
1876 				continue;
1877 
1878 			local = lat_stats->latencies[i][i];
1879 			if (local == 0)
1880 				continue;
1881 
1882 			/*
1883 			 * Track suspect probe times that aren't within
1884 			 * tolerance of minimum local latency and how much
1885 			 * probe times are corrected by
1886 			 */
1887 			if (local - min > min >> lgrp_plat_probe_lt_shift)
1888 				probe_stats->probe_suspect[i][i]++;
1889 
1890 			probe_stats->probe_errors[i][i] += local - min;
1891 
1892 			/*
1893 			 * Make local latencies be minimum
1894 			 */
1895 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
1896 			lat_stats->latencies[i][i] = min;
1897 		}
1898 	}
1899 
1900 	/*
1901 	 * Determine max probe time again since just adjusted latencies
1902 	 */
1903 	lat_stats->latency_max = 0;
1904 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1905 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1906 			if (!memnode_info[j].exists)
1907 				continue;
1908 			t = lat_stats->latencies[i][j];
1909 			if (t > lat_stats->latency_max)
1910 				lat_stats->latency_max = t;
1911 		}
1912 	}
1913 }
1914 
1915 
1916 /*
1917  * Verify following about latencies between nodes:
1918  *
1919  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
1920  * - Local latencies same
1921  * - Local < remote
1922  * - Number of latencies seen is reasonable
1923  * - Number of occurrences of a given latency should be more than 1
1924  *
1925  * Returns:
1926  *	0	Success
1927  *	-1	Not symmetric
1928  *	-2	Local latencies not same
1929  *	-3	Local >= remote
1930  */
1931 static int
1932 lgrp_plat_latency_verify(memnode_phys_addr_map_t *memnode_info,
1933     lgrp_plat_latency_stats_t *lat_stats)
1934 {
1935 	int				i;
1936 	int				j;
1937 	u_longlong_t			t1;
1938 	u_longlong_t			t2;
1939 
1940 	ASSERT(memnode_info != NULL && lat_stats != NULL);
1941 
1942 	/*
1943 	 * Nothing to do when this is an UMA machine, lgroup topology is
1944 	 * limited to 2 levels, or there aren't any probe times yet
1945 	 */
1946 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
1947 	    lat_stats->latencies[0][0] == 0)
1948 		return (0);
1949 
1950 	/*
1951 	 * Make sure that latencies are symmetric between any two nodes
1952 	 * (ie. latency(node0, node1) == latency(node1, node0))
1953 	 */
1954 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1955 		if (!memnode_info[i].exists)
1956 			continue;
1957 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1958 			if (!memnode_info[j].exists)
1959 				continue;
1960 			t1 = lat_stats->latencies[i][j];
1961 			t2 = lat_stats->latencies[j][i];
1962 
1963 			if (t1 == 0 || t2 == 0 || t1 == t2)
1964 				continue;
1965 
1966 			return (-1);
1967 		}
1968 	}
1969 
1970 	/*
1971 	 * Local latencies should be same
1972 	 */
1973 	t1 = lat_stats->latencies[0][0];
1974 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
1975 		if (!memnode_info[i].exists)
1976 			continue;
1977 
1978 		t2 = lat_stats->latencies[i][i];
1979 		if (t2 == 0)
1980 			continue;
1981 
1982 		if (t1 == 0) {
1983 			t1 = t2;
1984 			continue;
1985 		}
1986 
1987 		if (t1 != t2)
1988 			return (-2);
1989 	}
1990 
1991 	/*
1992 	 * Local latencies should be less than remote
1993 	 */
1994 	if (t1) {
1995 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1996 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1997 				if (!memnode_info[j].exists)
1998 					continue;
1999 				t2 = lat_stats->latencies[i][j];
2000 				if (i == j || t2 == 0)
2001 					continue;
2002 
2003 				if (t1 >= t2)
2004 					return (-3);
2005 			}
2006 		}
2007 	}
2008 
2009 	return (0);
2010 }
2011 
2012 
2013 /*
2014  * Platform-specific initialization
2015  */
2016 static void
2017 lgrp_plat_main_init(void)
2018 {
2019 	int	curnode;
2020 	int	ht_limit;
2021 	int	i;
2022 
2023 	/*
2024 	 * Print a notice that MPO is disabled when memory is interleaved
2025 	 * across nodes....Would do this when it is discovered, but can't
2026 	 * because it happens way too early during boot....
2027 	 */
2028 	if (lgrp_plat_mem_intrlv)
2029 		cmn_err(CE_NOTE,
2030 		    "MPO disabled because memory is interleaved\n");
2031 
2032 	/*
2033 	 * Don't bother to do any probing if it is disabled, there is only one
2034 	 * node, or the height of the lgroup topology less than or equal to 2
2035 	 */
2036 	ht_limit = lgrp_topo_ht_limit();
2037 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
2038 	    max_mem_nodes == 1 || ht_limit <= 2) {
2039 		/*
2040 		 * Setup lgroup latencies for 2 level lgroup topology
2041 		 * (ie. local and remote only) if they haven't been set yet
2042 		 */
2043 		if (ht_limit == 2 && lgrp_plat_lat_stats.latency_min == -1 &&
2044 		    lgrp_plat_lat_stats.latency_max == 0)
2045 			lgrp_plat_2level_setup(&lgrp_plat_lat_stats);
2046 		return;
2047 	}
2048 
2049 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
2050 		/*
2051 		 * Should have been able to probe from CPU 0 when it was added
2052 		 * to lgroup hierarchy, but may not have been able to then
2053 		 * because it happens so early in boot that gethrtime() hasn't
2054 		 * been initialized.  (:-(
2055 		 */
2056 		curnode = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node,
2057 		    lgrp_plat_cpu_node_nentries);
2058 		ASSERT(curnode >= 0 && curnode < lgrp_plat_node_cnt);
2059 		if (lgrp_plat_lat_stats.latencies[curnode][curnode] == 0)
2060 			lgrp_plat_probe();
2061 
2062 		return;
2063 	}
2064 
2065 	/*
2066 	 * When probing memory, use one page for every sample to determine
2067 	 * lgroup topology and taking multiple samples
2068 	 */
2069 	if (lgrp_plat_probe_mem_config.probe_memsize == 0)
2070 		lgrp_plat_probe_mem_config.probe_memsize = PAGESIZE *
2071 		    lgrp_plat_probe_nsamples;
2072 
2073 	/*
2074 	 * Map memory in each node needed for probing to determine latency
2075 	 * topology
2076 	 */
2077 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
2078 		int	mnode;
2079 
2080 		/*
2081 		 * Skip this node and leave its probe page NULL
2082 		 * if it doesn't have any memory
2083 		 */
2084 		mnode = i;
2085 		if (!mem_node_config[mnode].exists) {
2086 			lgrp_plat_probe_mem_config.probe_va[i] = NULL;
2087 			continue;
2088 		}
2089 
2090 		/*
2091 		 * Allocate one kernel virtual page
2092 		 */
2093 		lgrp_plat_probe_mem_config.probe_va[i] = vmem_alloc(heap_arena,
2094 		    lgrp_plat_probe_mem_config.probe_memsize, VM_NOSLEEP);
2095 		if (lgrp_plat_probe_mem_config.probe_va[i] == NULL) {
2096 			cmn_err(CE_WARN,
2097 			    "lgrp_plat_main_init: couldn't allocate memory");
2098 			return;
2099 		}
2100 
2101 		/*
2102 		 * Get PFN for first page in each node
2103 		 */
2104 		lgrp_plat_probe_mem_config.probe_pfn[i] =
2105 		    mem_node_config[mnode].physbase;
2106 
2107 		/*
2108 		 * Map virtual page to first page in node
2109 		 */
2110 		hat_devload(kas.a_hat, lgrp_plat_probe_mem_config.probe_va[i],
2111 		    lgrp_plat_probe_mem_config.probe_memsize,
2112 		    lgrp_plat_probe_mem_config.probe_pfn[i],
2113 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
2114 		    HAT_LOAD_NOCONSIST);
2115 	}
2116 
2117 	/*
2118 	 * Probe from current CPU
2119 	 */
2120 	lgrp_plat_probe();
2121 }
2122 
2123 
2124 /*
2125  * Return the number of free, allocatable, or installed
2126  * pages in an lgroup
2127  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
2128  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
2129  */
2130 static pgcnt_t
2131 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
2132 {
2133 	_NOTE(ARGUNUSED(lgrphand));
2134 
2135 	struct memlist *mlist;
2136 	pgcnt_t npgs = 0;
2137 	extern struct memlist *phys_avail;
2138 	extern struct memlist *phys_install;
2139 
2140 	switch (query) {
2141 	case LGRP_MEM_SIZE_FREE:
2142 		return ((pgcnt_t)freemem);
2143 	case LGRP_MEM_SIZE_AVAIL:
2144 		memlist_read_lock();
2145 		for (mlist = phys_avail; mlist; mlist = mlist->ml_next)
2146 			npgs += btop(mlist->ml_size);
2147 		memlist_read_unlock();
2148 		return (npgs);
2149 	case LGRP_MEM_SIZE_INSTALL:
2150 		memlist_read_lock();
2151 		for (mlist = phys_install; mlist; mlist = mlist->ml_next)
2152 			npgs += btop(mlist->ml_size);
2153 		memlist_read_unlock();
2154 		return (npgs);
2155 	default:
2156 		return ((pgcnt_t)0);
2157 	}
2158 }
2159 
2160 
2161 /*
2162  * Update node to proximity domain mappings for given domain and return node ID
2163  */
2164 static int
2165 lgrp_plat_node_domain_update(node_domain_map_t *node_domain, int node_cnt,
2166     uint32_t domain)
2167 {
2168 	uint_t	node;
2169 	uint_t	start;
2170 
2171 	/*
2172 	 * Hash proximity domain ID into node to domain mapping table (array)
2173 	 * and add entry for it into first non-existent or matching entry found
2174 	 */
2175 	node = start = NODE_DOMAIN_HASH(domain, node_cnt);
2176 	do {
2177 		/*
2178 		 * Entry doesn't exist yet, so create one for this proximity
2179 		 * domain and return node ID which is index into mapping table.
2180 		 */
2181 		if (!node_domain[node].exists) {
2182 			node_domain[node].prox_domain = domain;
2183 			membar_producer();
2184 			node_domain[node].exists = 1;
2185 			return (node);
2186 		}
2187 
2188 		/*
2189 		 * Entry exists for this proximity domain already, so just
2190 		 * return node ID (index into table).
2191 		 */
2192 		if (node_domain[node].prox_domain == domain)
2193 			return (node);
2194 		node = NODE_DOMAIN_HASH(node + 1, node_cnt);
2195 	} while (node != start);
2196 
2197 	/*
2198 	 * Ran out of supported number of entries which shouldn't happen....
2199 	 */
2200 	ASSERT(node != start);
2201 	return (-1);
2202 }
2203 
2204 /*
2205  * Update node memory information for given proximity domain with specified
2206  * starting and ending physical address range (and return positive numbers for
2207  * success and negative ones for errors)
2208  */
2209 static int
2210 lgrp_plat_memnode_info_update(node_domain_map_t *node_domain, int node_cnt,
2211     memnode_phys_addr_map_t *memnode_info, int memnode_cnt, uint64_t start,
2212     uint64_t end, uint32_t domain, uint32_t device_id)
2213 {
2214 	int	node, mnode;
2215 
2216 	/*
2217 	 * Get node number for proximity domain
2218 	 */
2219 	node = lgrp_plat_domain_to_node(node_domain, node_cnt, domain);
2220 	if (node == -1) {
2221 		node = lgrp_plat_node_domain_update(node_domain, node_cnt,
2222 		    domain);
2223 		if (node == -1)
2224 			return (-1);
2225 	}
2226 
2227 	/*
2228 	 * This function is called during boot if device_id is
2229 	 * ACPI_MEMNODE_DEVID_BOOT, otherwise it's called at runtime for
2230 	 * memory DR operations.
2231 	 */
2232 	if (device_id != ACPI_MEMNODE_DEVID_BOOT) {
2233 		ASSERT(lgrp_plat_max_mem_node <= memnode_cnt);
2234 
2235 		for (mnode = lgrp_plat_node_cnt;
2236 		    mnode < lgrp_plat_max_mem_node; mnode++) {
2237 			if (memnode_info[mnode].exists &&
2238 			    memnode_info[mnode].prox_domain == domain &&
2239 			    memnode_info[mnode].device_id == device_id) {
2240 				if (btop(start) < memnode_info[mnode].start)
2241 					memnode_info[mnode].start = btop(start);
2242 				if (btop(end) > memnode_info[mnode].end)
2243 					memnode_info[mnode].end = btop(end);
2244 				return (1);
2245 			}
2246 		}
2247 
2248 		if (lgrp_plat_max_mem_node >= memnode_cnt) {
2249 			return (-3);
2250 		} else {
2251 			lgrp_plat_max_mem_node++;
2252 			memnode_info[mnode].start = btop(start);
2253 			memnode_info[mnode].end = btop(end);
2254 			memnode_info[mnode].prox_domain = domain;
2255 			memnode_info[mnode].device_id = device_id;
2256 			memnode_info[mnode].lgrphand = node;
2257 			membar_producer();
2258 			memnode_info[mnode].exists = 1;
2259 			return (0);
2260 		}
2261 	}
2262 
2263 	/*
2264 	 * Create entry in table for node if it doesn't exist
2265 	 */
2266 	ASSERT(node < memnode_cnt);
2267 	if (!memnode_info[node].exists) {
2268 		memnode_info[node].start = btop(start);
2269 		memnode_info[node].end = btop(end);
2270 		memnode_info[node].prox_domain = domain;
2271 		memnode_info[node].device_id = device_id;
2272 		memnode_info[node].lgrphand = node;
2273 		membar_producer();
2274 		memnode_info[node].exists = 1;
2275 		return (0);
2276 	}
2277 
2278 	/*
2279 	 * Entry already exists for this proximity domain
2280 	 *
2281 	 * There may be more than one SRAT memory entry for a domain, so we may
2282 	 * need to update existing start or end address for the node.
2283 	 */
2284 	if (memnode_info[node].prox_domain == domain) {
2285 		if (btop(start) < memnode_info[node].start)
2286 			memnode_info[node].start = btop(start);
2287 		if (btop(end) > memnode_info[node].end)
2288 			memnode_info[node].end = btop(end);
2289 		return (1);
2290 	}
2291 	return (-2);
2292 }
2293 
2294 
2295 /*
2296  * Have to sort nodes by starting physical address because plat_mnode_xcheck()
2297  * assumes and expects memnodes to be sorted in ascending order by physical
2298  * address.
2299  */
2300 static void
2301 lgrp_plat_node_sort(node_domain_map_t *node_domain, int node_cnt,
2302     cpu_node_map_t *cpu_node, int cpu_count,
2303     memnode_phys_addr_map_t *memnode_info)
2304 {
2305 	boolean_t	found;
2306 	int		i;
2307 	int		j;
2308 	int		n;
2309 	boolean_t	sorted;
2310 	boolean_t	swapped;
2311 
2312 	if (!lgrp_plat_node_sort_enable || node_cnt <= 1 ||
2313 	    node_domain == NULL || memnode_info == NULL)
2314 		return;
2315 
2316 	/*
2317 	 * Sorted already?
2318 	 */
2319 	sorted = B_TRUE;
2320 	for (i = 0; i < node_cnt - 1; i++) {
2321 		/*
2322 		 * Skip entries that don't exist
2323 		 */
2324 		if (!memnode_info[i].exists)
2325 			continue;
2326 
2327 		/*
2328 		 * Try to find next existing entry to compare against
2329 		 */
2330 		found = B_FALSE;
2331 		for (j = i + 1; j < node_cnt; j++) {
2332 			if (memnode_info[j].exists) {
2333 				found = B_TRUE;
2334 				break;
2335 			}
2336 		}
2337 
2338 		/*
2339 		 * Done if no more existing entries to compare against
2340 		 */
2341 		if (found == B_FALSE)
2342 			break;
2343 
2344 		/*
2345 		 * Not sorted if starting address of current entry is bigger
2346 		 * than starting address of next existing entry
2347 		 */
2348 		if (memnode_info[i].start > memnode_info[j].start) {
2349 			sorted = B_FALSE;
2350 			break;
2351 		}
2352 	}
2353 
2354 	/*
2355 	 * Don't need to sort if sorted already
2356 	 */
2357 	if (sorted == B_TRUE)
2358 		return;
2359 
2360 	/*
2361 	 * Just use bubble sort since number of nodes is small
2362 	 */
2363 	n = node_cnt;
2364 	do {
2365 		swapped = B_FALSE;
2366 		n--;
2367 		for (i = 0; i < n; i++) {
2368 			/*
2369 			 * Skip entries that don't exist
2370 			 */
2371 			if (!memnode_info[i].exists)
2372 				continue;
2373 
2374 			/*
2375 			 * Try to find next existing entry to compare against
2376 			 */
2377 			found = B_FALSE;
2378 			for (j = i + 1; j <= n; j++) {
2379 				if (memnode_info[j].exists) {
2380 					found = B_TRUE;
2381 					break;
2382 				}
2383 			}
2384 
2385 			/*
2386 			 * Done if no more existing entries to compare against
2387 			 */
2388 			if (found == B_FALSE)
2389 				break;
2390 
2391 			if (memnode_info[i].start > memnode_info[j].start) {
2392 				memnode_phys_addr_map_t	save_addr;
2393 				node_domain_map_t	save_node;
2394 
2395 				/*
2396 				 * Swap node to proxmity domain ID assignments
2397 				 */
2398 				bcopy(&node_domain[i], &save_node,
2399 				    sizeof (node_domain_map_t));
2400 				bcopy(&node_domain[j], &node_domain[i],
2401 				    sizeof (node_domain_map_t));
2402 				bcopy(&save_node, &node_domain[j],
2403 				    sizeof (node_domain_map_t));
2404 
2405 				/*
2406 				 * Swap node to physical memory assignments
2407 				 */
2408 				bcopy(&memnode_info[i], &save_addr,
2409 				    sizeof (memnode_phys_addr_map_t));
2410 				bcopy(&memnode_info[j], &memnode_info[i],
2411 				    sizeof (memnode_phys_addr_map_t));
2412 				bcopy(&save_addr, &memnode_info[j],
2413 				    sizeof (memnode_phys_addr_map_t));
2414 				swapped = B_TRUE;
2415 			}
2416 		}
2417 	} while (swapped == B_TRUE);
2418 
2419 	/*
2420 	 * Check to make sure that CPUs assigned to correct node IDs now since
2421 	 * node to proximity domain ID assignments may have been changed above
2422 	 */
2423 	if (n == node_cnt - 1 || cpu_node == NULL || cpu_count < 1)
2424 		return;
2425 	for (i = 0; i < cpu_count; i++) {
2426 		int		node;
2427 
2428 		node = lgrp_plat_domain_to_node(node_domain, node_cnt,
2429 		    cpu_node[i].prox_domain);
2430 		if (cpu_node[i].node != node)
2431 			cpu_node[i].node = node;
2432 	}
2433 
2434 }
2435 
2436 
2437 /*
2438  * Return time needed to probe from current CPU to memory in given node
2439  */
2440 static hrtime_t
2441 lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node, int cpu_node_nentries,
2442     lgrp_plat_probe_mem_config_t *probe_mem_config,
2443     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
2444 {
2445 	caddr_t			buf;
2446 	hrtime_t		elapsed;
2447 	hrtime_t		end;
2448 	int			from;
2449 	int			i;
2450 	int			ipl;
2451 	hrtime_t		max;
2452 	hrtime_t		min;
2453 	hrtime_t		start;
2454 	extern int		use_sse_pagecopy;
2455 
2456 	/*
2457 	 * Determine ID of node containing current CPU
2458 	 */
2459 	from = lgrp_plat_cpu_to_node(CPU, cpu_node, cpu_node_nentries);
2460 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
2461 
2462 	/*
2463 	 * Do common work for probing main memory
2464 	 */
2465 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_PGCPY) {
2466 		/*
2467 		 * Skip probing any nodes without memory and
2468 		 * set probe time to 0
2469 		 */
2470 		if (probe_mem_config->probe_va[to] == NULL) {
2471 			lat_stats->latencies[from][to] = 0;
2472 			return (0);
2473 		}
2474 
2475 		/*
2476 		 * Invalidate caches once instead of once every sample
2477 		 * which should cut cost of probing by a lot
2478 		 */
2479 		probe_stats->flush_cost = gethrtime();
2480 		invalidate_cache();
2481 		probe_stats->flush_cost = gethrtime() -
2482 		    probe_stats->flush_cost;
2483 		probe_stats->probe_cost_total += probe_stats->flush_cost;
2484 	}
2485 
2486 	/*
2487 	 * Probe from current CPU to given memory using specified operation
2488 	 * and take specified number of samples
2489 	 */
2490 	max = 0;
2491 	min = -1;
2492 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
2493 		probe_stats->probe_cost = gethrtime();
2494 
2495 		/*
2496 		 * Can't measure probe time if gethrtime() isn't working yet
2497 		 */
2498 		if (probe_stats->probe_cost == 0 && gethrtime() == 0)
2499 			return (0);
2500 
2501 		if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
2502 			/*
2503 			 * Measure how long it takes to read vendor ID from
2504 			 * Northbridge
2505 			 */
2506 			elapsed = opt_probe_vendor(to, lgrp_plat_probe_nreads);
2507 		} else {
2508 			/*
2509 			 * Measure how long it takes to copy page
2510 			 * on top of itself
2511 			 */
2512 			buf = probe_mem_config->probe_va[to] + (i * PAGESIZE);
2513 
2514 			kpreempt_disable();
2515 			ipl = splhigh();
2516 			start = gethrtime();
2517 			if (use_sse_pagecopy)
2518 				hwblkpagecopy(buf, buf);
2519 			else
2520 				bcopy(buf, buf, PAGESIZE);
2521 			end = gethrtime();
2522 			elapsed = end - start;
2523 			splx(ipl);
2524 			kpreempt_enable();
2525 		}
2526 
2527 		probe_stats->probe_cost = gethrtime() -
2528 		    probe_stats->probe_cost;
2529 		probe_stats->probe_cost_total += probe_stats->probe_cost;
2530 
2531 		if (min == -1 || elapsed < min)
2532 			min = elapsed;
2533 		if (elapsed > max)
2534 			max = elapsed;
2535 	}
2536 
2537 	/*
2538 	 * Update minimum and maximum probe times between
2539 	 * these two nodes
2540 	 */
2541 	if (min < probe_stats->probe_min[from][to] ||
2542 	    probe_stats->probe_min[from][to] == 0)
2543 		probe_stats->probe_min[from][to] = min;
2544 
2545 	if (max > probe_stats->probe_max[from][to])
2546 		probe_stats->probe_max[from][to] = max;
2547 
2548 	return (min);
2549 }
2550 
2551 
2552 /*
2553  * Read boot property with CPU to APIC ID array, fill in CPU to node ID
2554  * mapping table with APIC ID for each CPU (if pointer to table isn't NULL),
2555  * and return number of CPU APIC IDs.
2556  *
2557  * NOTE: This code assumes that CPU IDs are assigned in order that they appear
2558  *       in in cpu_apicid_array boot property which is based on and follows
2559  *	 same ordering as processor list in ACPI MADT.  If the code in
2560  *	 usr/src/uts/i86pc/io/pcplusmp/apic.c that reads MADT and assigns
2561  *	 CPU IDs ever changes, then this code will need to change too....
2562  */
2563 static int
2564 lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node)
2565 {
2566 	int	boot_prop_len;
2567 	char	*boot_prop_name = BP_CPU_APICID_ARRAY;
2568 	uint32_t *cpu_apicid_array;
2569 	int	i;
2570 	int	n;
2571 
2572 	/*
2573 	 * Check length of property value
2574 	 */
2575 	boot_prop_len = BOP_GETPROPLEN(bootops, boot_prop_name);
2576 	if (boot_prop_len <= 0)
2577 		return (-1);
2578 
2579 	/*
2580 	 * Calculate number of entries in array and return when the system is
2581 	 * not very interesting for NUMA. It's not interesting for NUMA if
2582 	 * system has only one CPU and doesn't support CPU hotplug.
2583 	 */
2584 	n = boot_prop_len / sizeof (*cpu_apicid_array);
2585 	if (n == 1 && !plat_dr_support_cpu())
2586 		return (-2);
2587 
2588 	cpu_apicid_array = (uint32_t *)BOP_ALLOC(bootops, NULL, boot_prop_len,
2589 	    sizeof (*cpu_apicid_array));
2590 	/*
2591 	 * Get CPU to APIC ID property value
2592 	 */
2593 	if (cpu_apicid_array == NULL ||
2594 	    BOP_GETPROP(bootops, boot_prop_name, cpu_apicid_array) < 0)
2595 		return (-3);
2596 
2597 	/*
2598 	 * Just return number of CPU APIC IDs if CPU to node mapping table is
2599 	 * NULL
2600 	 */
2601 	if (cpu_node == NULL) {
2602 		if (plat_dr_support_cpu() && n >= boot_ncpus) {
2603 			return (boot_ncpus);
2604 		} else {
2605 			return (n);
2606 		}
2607 	}
2608 
2609 	/*
2610 	 * Fill in CPU to node ID mapping table with APIC ID for each CPU
2611 	 */
2612 	for (i = 0; i < n; i++) {
2613 		/* Only add boot CPUs into the map if CPU DR is enabled. */
2614 		if (plat_dr_support_cpu() && i >= boot_ncpus)
2615 			break;
2616 		cpu_node[i].exists = 1;
2617 		cpu_node[i].apicid = cpu_apicid_array[i];
2618 		cpu_node[i].prox_domain = UINT32_MAX;
2619 		cpu_node[i].node = UINT_MAX;
2620 	}
2621 
2622 	/*
2623 	 * Return number of CPUs based on number of APIC IDs
2624 	 */
2625 	return (i);
2626 }
2627 
2628 
2629 /*
2630  * Read ACPI System Locality Information Table (SLIT) to determine how far each
2631  * NUMA node is from each other
2632  */
2633 static int
2634 lgrp_plat_process_slit(ACPI_TABLE_SLIT *tp,
2635     node_domain_map_t *node_domain, uint_t node_cnt,
2636     memnode_phys_addr_map_t *memnode_info, lgrp_plat_latency_stats_t *lat_stats)
2637 {
2638 	int		i;
2639 	int		j;
2640 	int		src;
2641 	int		dst;
2642 	int		localities;
2643 	hrtime_t	max;
2644 	hrtime_t	min;
2645 	int		retval;
2646 	uint8_t		*slit_entries;
2647 
2648 	if (tp == NULL || !lgrp_plat_slit_enable)
2649 		return (1);
2650 
2651 	if (lat_stats == NULL)
2652 		return (2);
2653 
2654 	localities = tp->LocalityCount;
2655 
2656 	min = lat_stats->latency_min;
2657 	max = lat_stats->latency_max;
2658 
2659 	/*
2660 	 * Fill in latency matrix based on SLIT entries
2661 	 */
2662 	slit_entries = tp->Entry;
2663 	for (i = 0; i < localities; i++) {
2664 		src = lgrp_plat_domain_to_node(node_domain,
2665 		    node_cnt, i);
2666 		if (src == -1)
2667 			continue;
2668 
2669 		for (j = 0; j < localities; j++) {
2670 			uint8_t	latency;
2671 
2672 			dst = lgrp_plat_domain_to_node(node_domain,
2673 			    node_cnt, j);
2674 			if (dst == -1)
2675 				continue;
2676 
2677 			latency = slit_entries[(i * localities) + j];
2678 			lat_stats->latencies[src][dst] = latency;
2679 			if (latency < min || min == -1)
2680 				min = latency;
2681 			if (latency > max)
2682 				max = latency;
2683 		}
2684 	}
2685 
2686 	/*
2687 	 * Verify that latencies/distances given in SLIT look reasonable
2688 	 */
2689 	retval = lgrp_plat_latency_verify(memnode_info, lat_stats);
2690 
2691 	if (retval) {
2692 		/*
2693 		 * Reinitialize (zero) latency table since SLIT doesn't look
2694 		 * right
2695 		 */
2696 		for (i = 0; i < localities; i++) {
2697 			for (j = 0; j < localities; j++)
2698 				lat_stats->latencies[i][j] = 0;
2699 		}
2700 	} else {
2701 		/*
2702 		 * Update min and max latencies seen since SLIT looks valid
2703 		 */
2704 		lat_stats->latency_min = min;
2705 		lat_stats->latency_max = max;
2706 	}
2707 
2708 	return (retval);
2709 }
2710 
2711 
2712 /*
2713  * Update lgrp latencies according to information returned by ACPI _SLI method.
2714  */
2715 static int
2716 lgrp_plat_process_sli(uint32_t domain_id, uchar_t *sli_info,
2717     uint32_t sli_cnt, node_domain_map_t *node_domain, uint_t node_cnt,
2718     lgrp_plat_latency_stats_t *lat_stats)
2719 {
2720 	int		i;
2721 	int		src, dst;
2722 	uint8_t		latency;
2723 	hrtime_t	max, min;
2724 
2725 	if (lat_stats == NULL || sli_info == NULL ||
2726 	    sli_cnt == 0 || domain_id >= sli_cnt)
2727 		return (-1);
2728 
2729 	src = lgrp_plat_domain_to_node(node_domain, node_cnt, domain_id);
2730 	if (src == -1) {
2731 		src = lgrp_plat_node_domain_update(node_domain, node_cnt,
2732 		    domain_id);
2733 		if (src == -1)
2734 			return (-1);
2735 	}
2736 
2737 	/*
2738 	 * Don't update latency info if topology has been flattened to 2 levels.
2739 	 */
2740 	if (lgrp_plat_topo_flatten != 0) {
2741 		return (0);
2742 	}
2743 
2744 	/*
2745 	 * Latency information for proximity domain is ready.
2746 	 * TODO: support adjusting latency information at runtime.
2747 	 */
2748 	if (lat_stats->latencies[src][src] != 0) {
2749 		return (0);
2750 	}
2751 
2752 	/* Validate latency information. */
2753 	for (i = 0; i < sli_cnt; i++) {
2754 		if (i == domain_id) {
2755 			if (sli_info[i] != ACPI_SLIT_SELF_LATENCY ||
2756 			    sli_info[sli_cnt + i] != ACPI_SLIT_SELF_LATENCY) {
2757 				return (-1);
2758 			}
2759 		} else {
2760 			if (sli_info[i] <= ACPI_SLIT_SELF_LATENCY ||
2761 			    sli_info[sli_cnt + i] <= ACPI_SLIT_SELF_LATENCY ||
2762 			    sli_info[i] != sli_info[sli_cnt + i]) {
2763 				return (-1);
2764 			}
2765 		}
2766 	}
2767 
2768 	min = lat_stats->latency_min;
2769 	max = lat_stats->latency_max;
2770 	for (i = 0; i < sli_cnt; i++) {
2771 		dst = lgrp_plat_domain_to_node(node_domain, node_cnt, i);
2772 		if (dst == -1)
2773 			continue;
2774 
2775 		ASSERT(sli_info[i] == sli_info[sli_cnt + i]);
2776 
2777 		/* Update row in latencies matrix. */
2778 		latency = sli_info[i];
2779 		lat_stats->latencies[src][dst] = latency;
2780 		if (latency < min || min == -1)
2781 			min = latency;
2782 		if (latency > max)
2783 			max = latency;
2784 
2785 		/* Update column in latencies matrix. */
2786 		latency = sli_info[sli_cnt + i];
2787 		lat_stats->latencies[dst][src] = latency;
2788 		if (latency < min || min == -1)
2789 			min = latency;
2790 		if (latency > max)
2791 			max = latency;
2792 	}
2793 	lat_stats->latency_min = min;
2794 	lat_stats->latency_max = max;
2795 
2796 	return (0);
2797 }
2798 
2799 
2800 /*
2801  * Read ACPI System Resource Affinity Table (SRAT) to determine which CPUs
2802  * and memory are local to each other in the same NUMA node and return number
2803  * of nodes
2804  */
2805 static int
2806 lgrp_plat_process_srat(ACPI_TABLE_SRAT *tp, ACPI_TABLE_MSCT *mp,
2807     uint32_t *prox_domain_min, node_domain_map_t *node_domain,
2808     cpu_node_map_t *cpu_node, int cpu_count,
2809     memnode_phys_addr_map_t *memnode_info)
2810 {
2811 	ACPI_SUBTABLE_HEADER	*item, *srat_end;
2812 	int			i;
2813 	int			node_cnt;
2814 	int			proc_entry_count;
2815 	int			rc;
2816 
2817 	/*
2818 	 * Nothing to do when no SRAT or disabled
2819 	 */
2820 	if (tp == NULL || !lgrp_plat_srat_enable)
2821 		return (-1);
2822 
2823 	/*
2824 	 * Try to get domain information from MSCT table.
2825 	 * ACPI4.0: OSPM will use information provided by the MSCT only
2826 	 * when the System Resource Affinity Table (SRAT) exists.
2827 	 */
2828 	node_cnt = lgrp_plat_msct_domains(mp, prox_domain_min);
2829 	if (node_cnt <= 0) {
2830 		/*
2831 		 * Determine number of nodes by counting number of proximity
2832 		 * domains in SRAT.
2833 		 */
2834 		node_cnt = lgrp_plat_srat_domains(tp, prox_domain_min);
2835 	}
2836 	/*
2837 	 * Return if number of nodes is 1 or less since don't need to read SRAT.
2838 	 */
2839 	if (node_cnt == 1)
2840 		return (1);
2841 	else if (node_cnt <= 0)
2842 		return (-2);
2843 
2844 	/*
2845 	 * Walk through SRAT, examining each CPU and memory entry to determine
2846 	 * which CPUs and memory belong to which node.
2847 	 */
2848 	item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)tp + sizeof (*tp));
2849 	srat_end = (ACPI_SUBTABLE_HEADER *)(tp->Header.Length + (uintptr_t)tp);
2850 	proc_entry_count = 0;
2851 	while (item < srat_end) {
2852 		uint32_t	apic_id;
2853 		uint32_t	domain;
2854 		uint64_t	end;
2855 		uint64_t	length;
2856 		uint64_t	start;
2857 
2858 		switch (item->Type) {
2859 		case ACPI_SRAT_TYPE_CPU_AFFINITY: {	/* CPU entry */
2860 			ACPI_SRAT_CPU_AFFINITY *cpu =
2861 			    (ACPI_SRAT_CPU_AFFINITY *) item;
2862 
2863 			if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED) ||
2864 			    cpu_node == NULL)
2865 				break;
2866 
2867 			/*
2868 			 * Calculate domain (node) ID and fill in APIC ID to
2869 			 * domain/node mapping table
2870 			 */
2871 			domain = cpu->ProximityDomainLo;
2872 			for (i = 0; i < 3; i++) {
2873 				domain += cpu->ProximityDomainHi[i] <<
2874 				    ((i + 1) * 8);
2875 			}
2876 			apic_id = cpu->ApicId;
2877 
2878 			rc = lgrp_plat_cpu_node_update(node_domain, node_cnt,
2879 			    cpu_node, cpu_count, apic_id, domain);
2880 			if (rc < 0)
2881 				return (-3);
2882 			else if (rc == 0)
2883 				proc_entry_count++;
2884 			break;
2885 		}
2886 		case ACPI_SRAT_TYPE_MEMORY_AFFINITY: {	/* memory entry */
2887 			ACPI_SRAT_MEM_AFFINITY *mem =
2888 			    (ACPI_SRAT_MEM_AFFINITY *)item;
2889 
2890 			if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED) ||
2891 			    memnode_info == NULL)
2892 				break;
2893 
2894 			/*
2895 			 * Get domain (node) ID and fill in domain/node
2896 			 * to memory mapping table
2897 			 */
2898 			domain = mem->ProximityDomain;
2899 			start = mem->BaseAddress;
2900 			length = mem->Length;
2901 			end = start + length - 1;
2902 
2903 			/*
2904 			 * According to ACPI 4.0, both ENABLE and HOTPLUG flags
2905 			 * may be set for memory address range entries in SRAT
2906 			 * table which are reserved for memory hot plug.
2907 			 * We intersect memory address ranges in SRAT table
2908 			 * with memory ranges in physinstalled to filter out
2909 			 * memory address ranges reserved for hot plug.
2910 			 */
2911 			if (mem->Flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) {
2912 				uint64_t	rstart = UINT64_MAX;
2913 				uint64_t	rend = 0;
2914 				struct memlist	*ml;
2915 				extern struct bootops	*bootops;
2916 
2917 				memlist_read_lock();
2918 				for (ml = bootops->boot_mem->physinstalled;
2919 				    ml; ml = ml->ml_next) {
2920 					uint64_t tstart = ml->ml_address;
2921 					uint64_t tend;
2922 
2923 					tend = ml->ml_address + ml->ml_size;
2924 					if (tstart > end || tend < start)
2925 						continue;
2926 					if (start > tstart)
2927 						tstart = start;
2928 					if (rstart > tstart)
2929 						rstart = tstart;
2930 					if (end < tend)
2931 						tend = end;
2932 					if (rend < tend)
2933 						rend = tend;
2934 				}
2935 				memlist_read_unlock();
2936 				start = rstart;
2937 				end = rend;
2938 				/* Skip this entry if no memory installed. */
2939 				if (start > end)
2940 					break;
2941 			}
2942 
2943 			if (lgrp_plat_memnode_info_update(node_domain,
2944 			    node_cnt, memnode_info, node_cnt,
2945 			    start, end, domain, ACPI_MEMNODE_DEVID_BOOT) < 0)
2946 				return (-4);
2947 			break;
2948 		}
2949 		case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY: {	/* x2apic CPU */
2950 			ACPI_SRAT_X2APIC_CPU_AFFINITY *x2cpu =
2951 			    (ACPI_SRAT_X2APIC_CPU_AFFINITY *) item;
2952 
2953 			if (!(x2cpu->Flags & ACPI_SRAT_CPU_ENABLED) ||
2954 			    cpu_node == NULL)
2955 				break;
2956 
2957 			/*
2958 			 * Calculate domain (node) ID and fill in APIC ID to
2959 			 * domain/node mapping table
2960 			 */
2961 			domain = x2cpu->ProximityDomain;
2962 			apic_id = x2cpu->ApicId;
2963 
2964 			rc = lgrp_plat_cpu_node_update(node_domain, node_cnt,
2965 			    cpu_node, cpu_count, apic_id, domain);
2966 			if (rc < 0)
2967 				return (-3);
2968 			else if (rc == 0)
2969 				proc_entry_count++;
2970 			break;
2971 		}
2972 		default:
2973 			break;
2974 		}
2975 
2976 		item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)item + item->Length);
2977 	}
2978 
2979 	/*
2980 	 * Should have seen at least as many SRAT processor entries as CPUs
2981 	 */
2982 	if (proc_entry_count < cpu_count)
2983 		return (-5);
2984 
2985 	/*
2986 	 * Need to sort nodes by starting physical address since VM system
2987 	 * assumes and expects memnodes to be sorted in ascending order by
2988 	 * physical address
2989 	 */
2990 	lgrp_plat_node_sort(node_domain, node_cnt, cpu_node, cpu_count,
2991 	    memnode_info);
2992 
2993 	return (node_cnt);
2994 }
2995 
2996 
2997 /*
2998  * Allocate permanent memory for any temporary memory that we needed to
2999  * allocate using BOP_ALLOC() before kmem_alloc() and VM system were
3000  * initialized and copy everything from temporary to permanent memory since
3001  * temporary boot memory will eventually be released during boot
3002  */
3003 static void
3004 lgrp_plat_release_bootstrap(void)
3005 {
3006 	void	*buf;
3007 	size_t	size;
3008 
3009 	if (lgrp_plat_cpu_node_nentries > 0) {
3010 		size = lgrp_plat_cpu_node_nentries * sizeof (cpu_node_map_t);
3011 		buf = kmem_alloc(size, KM_SLEEP);
3012 		bcopy(lgrp_plat_cpu_node, buf, size);
3013 		lgrp_plat_cpu_node = buf;
3014 	}
3015 }
3016 
3017 
3018 /*
3019  * Return number of proximity domains given in ACPI SRAT
3020  */
3021 static int
3022 lgrp_plat_srat_domains(ACPI_TABLE_SRAT *tp, uint32_t *prox_domain_min)
3023 {
3024 	int			domain_cnt;
3025 	uint32_t		domain_min;
3026 	ACPI_SUBTABLE_HEADER	*item, *end;
3027 	int			i;
3028 	node_domain_map_t	node_domain[MAX_NODES];
3029 
3030 
3031 	if (tp == NULL || !lgrp_plat_srat_enable)
3032 		return (1);
3033 
3034 	/*
3035 	 * Walk through SRAT to find minimum proximity domain ID
3036 	 */
3037 	domain_min = UINT32_MAX;
3038 	item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)tp + sizeof (*tp));
3039 	end = (ACPI_SUBTABLE_HEADER *)(tp->Header.Length + (uintptr_t)tp);
3040 	while (item < end) {
3041 		uint32_t	domain;
3042 
3043 		switch (item->Type) {
3044 		case ACPI_SRAT_TYPE_CPU_AFFINITY: {	/* CPU entry */
3045 			ACPI_SRAT_CPU_AFFINITY *cpu =
3046 			    (ACPI_SRAT_CPU_AFFINITY *) item;
3047 
3048 			if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED)) {
3049 				item = (ACPI_SUBTABLE_HEADER *)
3050 				    ((uintptr_t)item + item->Length);
3051 				continue;
3052 			}
3053 			domain = cpu->ProximityDomainLo;
3054 			for (i = 0; i < 3; i++) {
3055 				domain += cpu->ProximityDomainHi[i] <<
3056 				    ((i + 1) * 8);
3057 			}
3058 			break;
3059 		}
3060 		case ACPI_SRAT_TYPE_MEMORY_AFFINITY: {	/* memory entry */
3061 			ACPI_SRAT_MEM_AFFINITY *mem =
3062 			    (ACPI_SRAT_MEM_AFFINITY *)item;
3063 
3064 			if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED)) {
3065 				item = (ACPI_SUBTABLE_HEADER *)
3066 				    ((uintptr_t)item + item->Length);
3067 				continue;
3068 			}
3069 			domain = mem->ProximityDomain;
3070 			break;
3071 		}
3072 		case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY: {	/* x2apic CPU */
3073 			ACPI_SRAT_X2APIC_CPU_AFFINITY *x2cpu =
3074 			    (ACPI_SRAT_X2APIC_CPU_AFFINITY *) item;
3075 
3076 			if (!(x2cpu->Flags & ACPI_SRAT_CPU_ENABLED)) {
3077 				item = (ACPI_SUBTABLE_HEADER *)
3078 				    ((uintptr_t)item + item->Length);
3079 				continue;
3080 			}
3081 			domain = x2cpu->ProximityDomain;
3082 			break;
3083 		}
3084 		default:
3085 			item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)item +
3086 			    item->Length);
3087 			continue;
3088 		}
3089 
3090 		/*
3091 		 * Keep track of minimum proximity domain ID
3092 		 */
3093 		if (domain < domain_min)
3094 			domain_min = domain;
3095 
3096 		item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)item + item->Length);
3097 	}
3098 	if (lgrp_plat_domain_min_enable && prox_domain_min != NULL)
3099 		*prox_domain_min = domain_min;
3100 
3101 	/*
3102 	 * Walk through SRAT, examining each CPU and memory entry to determine
3103 	 * proximity domain ID for each.
3104 	 */
3105 	domain_cnt = 0;
3106 	item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)tp + sizeof (*tp));
3107 	end = (ACPI_SUBTABLE_HEADER *)(tp->Header.Length + (uintptr_t)tp);
3108 	bzero(node_domain, MAX_NODES * sizeof (node_domain_map_t));
3109 	while (item < end) {
3110 		uint32_t	domain;
3111 		boolean_t	overflow;
3112 		uint_t		start;
3113 
3114 		switch (item->Type) {
3115 		case ACPI_SRAT_TYPE_CPU_AFFINITY: {	/* CPU entry */
3116 			ACPI_SRAT_CPU_AFFINITY *cpu =
3117 			    (ACPI_SRAT_CPU_AFFINITY *) item;
3118 
3119 			if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED)) {
3120 				item = (ACPI_SUBTABLE_HEADER *)
3121 				    ((uintptr_t)item + item->Length);
3122 				continue;
3123 			}
3124 			domain = cpu->ProximityDomainLo;
3125 			for (i = 0; i < 3; i++) {
3126 				domain += cpu->ProximityDomainHi[i] <<
3127 				    ((i + 1) * 8);
3128 			}
3129 			break;
3130 		}
3131 		case ACPI_SRAT_TYPE_MEMORY_AFFINITY: {	/* memory entry */
3132 			ACPI_SRAT_MEM_AFFINITY *mem =
3133 			    (ACPI_SRAT_MEM_AFFINITY *)item;
3134 
3135 			if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED)) {
3136 				item = (ACPI_SUBTABLE_HEADER *)
3137 				    ((uintptr_t)item + item->Length);
3138 				continue;
3139 			}
3140 			domain = mem->ProximityDomain;
3141 			break;
3142 		}
3143 		case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY: {	/* x2apic CPU */
3144 			ACPI_SRAT_X2APIC_CPU_AFFINITY *x2cpu =
3145 			    (ACPI_SRAT_X2APIC_CPU_AFFINITY *) item;
3146 
3147 			if (!(x2cpu->Flags & ACPI_SRAT_CPU_ENABLED)) {
3148 				item = (ACPI_SUBTABLE_HEADER *)
3149 				    ((uintptr_t)item + item->Length);
3150 				continue;
3151 			}
3152 			domain = x2cpu->ProximityDomain;
3153 			break;
3154 		}
3155 		default:
3156 			item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)item +
3157 			    item->Length);
3158 			continue;
3159 		}
3160 
3161 		/*
3162 		 * Count and keep track of which proximity domain IDs seen
3163 		 */
3164 		start = i = domain % MAX_NODES;
3165 		overflow = B_TRUE;
3166 		do {
3167 			/*
3168 			 * Create entry for proximity domain and increment
3169 			 * count when no entry exists where proximity domain
3170 			 * hashed
3171 			 */
3172 			if (!node_domain[i].exists) {
3173 				node_domain[i].exists = 1;
3174 				node_domain[i].prox_domain = domain;
3175 				domain_cnt++;
3176 				overflow = B_FALSE;
3177 				break;
3178 			}
3179 
3180 			/*
3181 			 * Nothing to do when proximity domain seen already
3182 			 * and its entry exists
3183 			 */
3184 			if (node_domain[i].prox_domain == domain) {
3185 				overflow = B_FALSE;
3186 				break;
3187 			}
3188 
3189 			/*
3190 			 * Entry exists where proximity domain hashed, but for
3191 			 * different proximity domain so keep search for empty
3192 			 * slot to put it or matching entry whichever comes
3193 			 * first.
3194 			 */
3195 			i = (i + 1) % MAX_NODES;
3196 		} while (i != start);
3197 
3198 		/*
3199 		 * Didn't find empty or matching entry which means have more
3200 		 * proximity domains than supported nodes (:-(
3201 		 */
3202 		ASSERT(overflow != B_TRUE);
3203 		if (overflow == B_TRUE)
3204 			return (-1);
3205 
3206 		item = (ACPI_SUBTABLE_HEADER *)((uintptr_t)item + item->Length);
3207 	}
3208 	return (domain_cnt);
3209 }
3210 
3211 
3212 /*
3213  * Parse domain information in ACPI Maximum System Capability Table (MSCT).
3214  * MSCT table has been verified in function process_msct() in fakebop.c.
3215  */
3216 static int
3217 lgrp_plat_msct_domains(ACPI_TABLE_MSCT *tp, uint32_t *prox_domain_min)
3218 {
3219 	int last_seen = 0;
3220 	uint32_t proxmin = UINT32_MAX;
3221 	ACPI_MSCT_PROXIMITY *item, *end;
3222 
3223 	if (tp == NULL || lgrp_plat_msct_enable == 0)
3224 		return (-1);
3225 
3226 	if (tp->MaxProximityDomains >= MAX_NODES) {
3227 		cmn_err(CE_CONT,
3228 		    "?lgrp: too many proximity domains (%d), max %d supported, "
3229 		    "disable support of CPU/memory DR operations.",
3230 		    tp->MaxProximityDomains + 1, MAX_NODES);
3231 		plat_dr_disable_cpu();
3232 		plat_dr_disable_memory();
3233 		return (-1);
3234 	}
3235 
3236 	if (prox_domain_min != NULL) {
3237 		end = (void *)(tp->Header.Length + (uintptr_t)tp);
3238 		for (item = (void *)((uintptr_t)tp +
3239 		    tp->ProximityOffset); item < end;
3240 		    item = (void *)(item->Length + (uintptr_t)item)) {
3241 			if (item->RangeStart < proxmin) {
3242 				proxmin = item->RangeStart;
3243 			}
3244 
3245 			last_seen = item->RangeEnd - item->RangeStart + 1;
3246 			/*
3247 			 * Break out if all proximity domains have been
3248 			 * processed. Some BIOSes may have unused items
3249 			 * at the end of MSCT table.
3250 			 */
3251 			if (last_seen > tp->MaxProximityDomains) {
3252 				break;
3253 			}
3254 		}
3255 		*prox_domain_min = proxmin;
3256 	}
3257 
3258 	return (tp->MaxProximityDomains + 1);
3259 }
3260 
3261 
3262 /*
3263  * Set lgroup latencies for 2 level lgroup topology
3264  */
3265 static void
3266 lgrp_plat_2level_setup(lgrp_plat_latency_stats_t *lat_stats)
3267 {
3268 	int	i, j;
3269 
3270 	ASSERT(lat_stats != NULL);
3271 
3272 	if (lgrp_plat_node_cnt >= 4)
3273 		cmn_err(CE_NOTE,
3274 		    "MPO only optimizing for local and remote\n");
3275 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
3276 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
3277 			if (i == j)
3278 				lat_stats->latencies[i][j] = 2;
3279 			else
3280 				lat_stats->latencies[i][j] = 3;
3281 		}
3282 	}
3283 	lat_stats->latency_min = 2;
3284 	lat_stats->latency_max = 3;
3285 	/* TODO: check it. */
3286 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
3287 	lgrp_plat_topo_flatten = 1;
3288 }
3289 
3290 
3291 /*
3292  * The following Opteron specific constants, macros, types, and routines define
3293  * PCI configuration space registers and how to read them to determine the NUMA
3294  * configuration of *supported* Opteron processors.  They provide the same
3295  * information that may be gotten from the ACPI System Resource Affinity Table
3296  * (SRAT) if it exists on the machine of interest.
3297  *
3298  * The AMD BIOS and Kernel Developer's Guide (BKDG) for the processor family
3299  * of interest describes all of these registers and their contents.  The main
3300  * registers used by this code to determine the NUMA configuration of the
3301  * machine are the node ID register for the number of NUMA nodes and the DRAM
3302  * address map registers for the physical address range of each node.
3303  *
3304  * NOTE: The format and how to determine the NUMA configuration using PCI
3305  *	 config space registers may change or may not be supported in future
3306  *	 Opteron processor families.
3307  */
3308 
3309 /*
3310  * How many bits to shift Opteron DRAM Address Map base and limit registers
3311  * to get actual value
3312  */
3313 #define	OPT_DRAMADDR_HI_LSHIFT_ADDR	40	/* shift left for address */
3314 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
3315 
3316 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
3317 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
3318 
3319 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
3320 
3321 /*
3322  * Macros to derive addresses from Opteron DRAM Address Map registers
3323  */
3324 #define	OPT_DRAMADDR_HI(reg) \
3325 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
3326 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
3327 
3328 #define	OPT_DRAMADDR_LO(reg) \
3329 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
3330 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
3331 
3332 #define	OPT_DRAMADDR(high, low) \
3333 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
3334 
3335 /*
3336  * Bit masks defining what's in Opteron DRAM Address Map base register
3337  */
3338 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
3339 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
3340 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
3341 
3342 /*
3343  * Bit masks defining what's in Opteron DRAM Address Map limit register
3344  */
3345 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
3346 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
3347 
3348 
3349 /*
3350  * Opteron Node ID register in PCI configuration space contains
3351  * number of nodes in system, etc. for Opteron K8.  The following
3352  * constants and macros define its contents, structure, and access.
3353  */
3354 
3355 /*
3356  * Bit masks defining what's in Opteron Node ID register
3357  */
3358 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
3359 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
3360 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
3361 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
3362 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
3363 
3364 /*
3365  * How many bits in Opteron Node ID register to shift right to get actual value
3366  */
3367 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
3368 
3369 /*
3370  * Macros to get values from Opteron Node ID register
3371  */
3372 #define	OPT_NODE_CNT(reg) \
3373 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
3374 
3375 /*
3376  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
3377  * "in/out" instructions
3378  *
3379  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
3380  *	 other uses should just do MMIO to access PCI ECS.
3381  *	 Must enable special bit in Northbridge Configuration Register on
3382  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
3383  *	 using "in/out" instructions and restore special bit after done
3384  *	 accessing PCI ECS.
3385  */
3386 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
3387 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
3388 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
3389 	    ((((reg) >> 8) & 0xf) << 24))
3390 
3391 /*
3392  * PCI configuration space registers accessed by specifying
3393  * a bus, device, function, and offset.  The following constants
3394  * define the values needed to access Opteron K8 configuration
3395  * info to determine its node topology
3396  */
3397 
3398 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
3399 
3400 /*
3401  * Opteron PCI configuration space register function values
3402  */
3403 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
3404 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
3405 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
3406 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
3407 
3408 /*
3409  * PCI Configuration Space register offsets
3410  */
3411 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
3412 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
3413 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
3414 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
3415 
3416 /*
3417  * Opteron PCI Configuration Space device IDs for nodes
3418  */
3419 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
3420 
3421 
3422 /*
3423  * Opteron DRAM address map gives base and limit for physical memory in a node
3424  */
3425 typedef	struct opt_dram_addr_map {
3426 	uint32_t	base_hi;
3427 	uint32_t	base_lo;
3428 	uint32_t	limit_hi;
3429 	uint32_t	limit_lo;
3430 } opt_dram_addr_map_t;
3431 
3432 
3433 /*
3434  * Supported AMD processor families
3435  */
3436 #define	AMD_FAMILY_HAMMER	15
3437 #define	AMD_FAMILY_GREYHOUND	16
3438 
3439 /*
3440  * Whether to have is_opteron() return 1 even when processor isn't supported
3441  */
3442 uint_t	is_opteron_override = 0;
3443 
3444 /*
3445  * AMD processor family for current CPU
3446  */
3447 uint_t	opt_family = 0;
3448 
3449 
3450 /*
3451  * Determine whether we're running on a supported AMD Opteron since reading
3452  * node count and DRAM address map registers may have different format or
3453  * may not be supported across processor families
3454  */
3455 static int
3456 is_opteron(void)
3457 {
3458 
3459 	if (x86_vendor != X86_VENDOR_AMD)
3460 		return (0);
3461 
3462 	opt_family = cpuid_getfamily(CPU);
3463 	if (opt_family == AMD_FAMILY_HAMMER ||
3464 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
3465 		return (1);
3466 	else
3467 		return (0);
3468 }
3469 
3470 
3471 /*
3472  * Determine NUMA configuration for Opteron from registers that live in PCI
3473  * configuration space
3474  */
3475 static void
3476 opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
3477     memnode_phys_addr_map_t *memnode_info)
3478 {
3479 	uint_t				bus;
3480 	uint_t				dev;
3481 	struct opt_dram_addr_map	dram_map[MAX_NODES];
3482 	uint_t				node;
3483 	uint_t				node_info[MAX_NODES];
3484 	uint_t				off_hi;
3485 	uint_t				off_lo;
3486 	uint64_t nb_cfg_reg;
3487 
3488 	/*
3489 	 * Read configuration registers from PCI configuration space to
3490 	 * determine node information, which memory is in each node, etc.
3491 	 *
3492 	 * Write to PCI configuration space address register to specify
3493 	 * which configuration register to read and read/write PCI
3494 	 * configuration space data register to get/set contents
3495 	 */
3496 	bus = OPT_PCS_BUS_CONFIG;
3497 	dev = OPT_PCS_DEV_NODE0;
3498 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
3499 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
3500 
3501 	/*
3502 	 * Read node ID register for node 0 to get node count
3503 	 */
3504 	node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
3505 	    OPT_PCS_OFF_NODEID);
3506 	*node_cnt = OPT_NODE_CNT(node_info[0]) + 1;
3507 
3508 	/*
3509 	 * If number of nodes is more than maximum supported, then set node
3510 	 * count to 1 and treat system as UMA instead of NUMA.
3511 	 */
3512 	if (*node_cnt > MAX_NODES) {
3513 		*node_cnt = 1;
3514 		return;
3515 	}
3516 
3517 	/*
3518 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
3519 	 * read high DRAM address map base and limit registers
3520 	 */
3521 	nb_cfg_reg = 0;
3522 	if (opt_family == AMD_FAMILY_GREYHOUND) {
3523 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
3524 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
3525 			wrmsr(MSR_AMD_NB_CFG,
3526 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
3527 	}
3528 
3529 	for (node = 0; node < *node_cnt; node++) {
3530 		uint32_t	base_hi;
3531 		uint32_t	base_lo;
3532 		uint32_t	limit_hi;
3533 		uint32_t	limit_lo;
3534 
3535 		/*
3536 		 * Read node ID register (except for node 0 which we just read)
3537 		 */
3538 		if (node > 0) {
3539 			node_info[node] = pci_getl_func(bus, dev,
3540 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
3541 		}
3542 
3543 		/*
3544 		 * Read DRAM base and limit registers which specify
3545 		 * physical memory range of each node
3546 		 */
3547 		if (opt_family != AMD_FAMILY_GREYHOUND)
3548 			base_hi = 0;
3549 		else {
3550 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
3551 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
3552 			base_hi = dram_map[node].base_hi =
3553 			    inl(PCI_CONFDATA);
3554 		}
3555 		base_lo = dram_map[node].base_lo = pci_getl_func(bus, dev,
3556 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
3557 
3558 		if ((dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN) &&
3559 		    mem_intrlv)
3560 			*mem_intrlv = *mem_intrlv + 1;
3561 
3562 		off_hi += 4;	/* high limit register offset */
3563 		if (opt_family != AMD_FAMILY_GREYHOUND)
3564 			limit_hi = 0;
3565 		else {
3566 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
3567 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
3568 			limit_hi = dram_map[node].limit_hi =
3569 			    inl(PCI_CONFDATA);
3570 		}
3571 
3572 		off_lo += 4;	/* low limit register offset */
3573 		limit_lo = dram_map[node].limit_lo = pci_getl_func(bus,
3574 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
3575 
3576 		/*
3577 		 * Increment device number to next node and register offsets
3578 		 * for DRAM base register of next node
3579 		 */
3580 		off_hi += 4;
3581 		off_lo += 4;
3582 		dev++;
3583 
3584 		/*
3585 		 * Both read and write enable bits must be enabled in DRAM
3586 		 * address map base register for physical memory to exist in
3587 		 * node
3588 		 */
3589 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
3590 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
3591 			/*
3592 			 * Mark node memory as non-existent and set start and
3593 			 * end addresses to be same in memnode_info[]
3594 			 */
3595 			memnode_info[node].exists = 0;
3596 			memnode_info[node].start = memnode_info[node].end =
3597 			    (pfn_t)-1;
3598 			continue;
3599 		}
3600 
3601 		/*
3602 		 * Mark node memory as existing and remember physical address
3603 		 * range of each node for use later
3604 		 */
3605 		memnode_info[node].exists = 1;
3606 
3607 		memnode_info[node].start = btop(OPT_DRAMADDR(base_hi, base_lo));
3608 
3609 		memnode_info[node].end = btop(OPT_DRAMADDR(limit_hi, limit_lo) |
3610 		    OPT_DRAMADDR_LO_MASK_OFF);
3611 	}
3612 
3613 	/*
3614 	 * Restore PCI Extended Configuration Space enable bit
3615 	 */
3616 	if (opt_family == AMD_FAMILY_GREYHOUND) {
3617 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
3618 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
3619 	}
3620 }
3621 
3622 
3623 /*
3624  * Return average amount of time to read vendor ID register on Northbridge
3625  * N times on specified destination node from current CPU
3626  */
3627 static hrtime_t
3628 opt_probe_vendor(int dest_node, int nreads)
3629 {
3630 	int		cnt;
3631 	uint_t		dev;
3632 	/* LINTED: set but not used in function */
3633 	volatile uint_t	dev_vendor __unused;
3634 	hrtime_t	elapsed;
3635 	hrtime_t	end;
3636 	int		ipl;
3637 	hrtime_t	start;
3638 
3639 	dev = OPT_PCS_DEV_NODE0 + dest_node;
3640 	kpreempt_disable();
3641 	ipl = spl8();
3642 	outl(PCI_CONFADD, PCI_CADDR1(0, dev, OPT_PCS_FUNC_DRAM,
3643 	    OPT_PCS_OFF_VENDOR));
3644 	start = gethrtime();
3645 	for (cnt = 0; cnt < nreads; cnt++)
3646 		dev_vendor = inl(PCI_CONFDATA);
3647 	end = gethrtime();
3648 	elapsed = (end - start) / nreads;
3649 	splx(ipl);
3650 	kpreempt_enable();
3651 	return (elapsed);
3652 }
3653