xref: /titanic_51/usr/src/uts/i86pc/os/lgrpplat.c (revision 853389e344087efe23a5cc688693e5301a1cb90d)
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 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 
30 /*
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
95  *
96  * - lgrp_plat_node_domain[]		Node ID to proximity domain ID mapping
97  *					table indexed by node ID (only used
98  *					for SRAT)
99  *
100  * - lgrp_plat_node_memory[]		Table with physical address range for
101  *					each node indexed by node ID
102  *
103  * The code is implemented to make the following always be true:
104  *
105  *	lgroup platform handle == node ID == memnode ID
106  *
107  * Moreover, it allows for the proximity domain ID to be equal to all of the
108  * above as long as the proximity domains IDs are numbered from 0 to <number of
109  * nodes - 1>.  This is done by hashing each proximity domain ID into the range
110  * from 0 to <number of nodes - 1>.  Then proximity ID N will hash into node ID
111  * N and proximity domain ID N will be entered into lgrp_plat_node_domain[N]
112  * and be assigned node ID N.  If the proximity domain IDs aren't numbered
113  * from 0 to <number of nodes - 1>, then hashing the proximity domain IDs into
114  * lgrp_plat_node_domain[] will still work for assigning proximity domain IDs
115  * to node IDs.  However, the proximity domain IDs may not map to the
116  * equivalent node ID since we want to keep the node IDs numbered from 0 to
117  * <number of nodes - 1> to minimize cost of searching and potentially space.
118  */
119 
120 
121 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
122 #include <sys/bootconf.h>
123 #include <sys/cmn_err.h>
124 #include <sys/controlregs.h>
125 #include <sys/cpupart.h>
126 #include <sys/cpuvar.h>
127 #include <sys/lgrp.h>
128 #include <sys/machsystm.h>
129 #include <sys/memlist.h>
130 #include <sys/memnode.h>
131 #include <sys/mman.h>
132 #include <sys/pci_cfgspace.h>
133 #include <sys/pci_impl.h>
134 #include <sys/param.h>
135 #include <sys/pghw.h>
136 #include <sys/promif.h>		/* for prom_printf() */
137 #include <sys/sysmacros.h>
138 #include <sys/systm.h>
139 #include <sys/thread.h>
140 #include <sys/types.h>
141 #include <sys/var.h>
142 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
143 #include <vm/hat_i86.h>
144 #include <vm/seg_kmem.h>
145 #include <vm/vm_dep.h>
146 
147 #include "acpi_fw.h"		/* for SRAT and SLIT */
148 
149 
150 #define	MAX_NODES		8
151 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
152 
153 /*
154  * Constants for configuring probing
155  */
156 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
157 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
158 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
159 
160 /*
161  * Flags for probing
162  */
163 #define	LGRP_PLAT_PROBE_ENABLE		0x1	/* enable probing */
164 #define	LGRP_PLAT_PROBE_PGCPY		0x2	/* probe using page copy */
165 #define	LGRP_PLAT_PROBE_VENDOR		0x4	/* probe vendor ID register */
166 
167 /*
168  * Hash proximity domain ID into node to domain mapping table using to minimize
169  * span of entries used
170  */
171 #define	NODE_DOMAIN_HASH(domain)	((domain) % lgrp_plat_node_cnt)
172 
173 
174 /*
175  * CPU to node ID mapping structure (only used with SRAT)
176  */
177 typedef	struct cpu_node_map {
178 	int		exists;
179 	uint_t		node;
180 	uint32_t	apicid;
181 	uint32_t	prox_domain;
182 } cpu_node_map_t;
183 
184 /*
185  * Latency statistics
186  */
187 typedef struct lgrp_plat_latency_stats {
188 	hrtime_t	latencies[MAX_NODES][MAX_NODES];
189 	hrtime_t	latency_max;
190 	hrtime_t	latency_min;
191 } lgrp_plat_latency_stats_t;
192 
193 /*
194  * Memory configuration for probing
195  */
196 typedef struct lgrp_plat_probe_mem_config {
197 	size_t	probe_memsize;		/* how much memory to probe per node */
198 	caddr_t	probe_va[MAX_NODES];	/* where memory mapped for probing */
199 	pfn_t	probe_pfn[MAX_NODES];	/* physical pages to map for probing */
200 } lgrp_plat_probe_mem_config_t;
201 
202 /*
203  * Statistics kept for probing
204  */
205 typedef struct lgrp_plat_probe_stats {
206 	hrtime_t	flush_cost;
207 	hrtime_t	probe_cost;
208 	hrtime_t	probe_cost_total;
209 	hrtime_t	probe_error_code;
210 	hrtime_t	probe_errors[MAX_NODES][MAX_NODES];
211 	int		probe_suspect[MAX_NODES][MAX_NODES];
212 	hrtime_t	probe_max[MAX_NODES][MAX_NODES];
213 	hrtime_t	probe_min[MAX_NODES][MAX_NODES];
214 } lgrp_plat_probe_stats_t;
215 
216 /*
217  * Node to proximity domain ID mapping structure (only used with SRAT)
218  */
219 typedef	struct node_domain_map {
220 	int		exists;
221 	uint32_t	prox_domain;
222 } node_domain_map_t;
223 
224 /*
225  * Node ID and starting and ending page for physical memory in node
226  */
227 typedef	struct node_phys_addr_map {
228 	pfn_t		start;
229 	pfn_t		end;
230 	int		exists;
231 	uint32_t	prox_domain;
232 } node_phys_addr_map_t;
233 
234 /*
235  * Error code from processing CPU to APIC ID array boot property
236  */
237 static int				lgrp_plat_cpu_apicid_error = 0;
238 
239 /*
240  * CPU to node ID mapping table (only used for SRAT)
241  */
242 static cpu_node_map_t			lgrp_plat_cpu_node[NCPU];
243 
244 /*
245  * Latency statistics
246  */
247 lgrp_plat_latency_stats_t		lgrp_plat_lat_stats;
248 
249 /*
250  * Whether memory is interleaved across nodes causing MPO to be disabled
251  */
252 static int				lgrp_plat_mem_intrlv = 0;
253 
254 /*
255  * Node ID to proximity domain ID mapping table (only used for SRAT)
256  */
257 static node_domain_map_t		lgrp_plat_node_domain[MAX_NODES];
258 
259 /*
260  * Physical address range for memory in each node
261  */
262 static node_phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
263 
264 /*
265  * Statistics gotten from probing
266  */
267 static lgrp_plat_probe_stats_t		lgrp_plat_probe_stats;
268 
269 /*
270  * Memory configuration for probing
271  */
272 static lgrp_plat_probe_mem_config_t	lgrp_plat_probe_mem_config;
273 
274 /*
275  * Error code from processing ACPI SRAT
276  */
277 static int				lgrp_plat_srat_error = 0;
278 
279 /*
280  * Error code from processing ACPI SLIT
281  */
282 static int				lgrp_plat_slit_error = 0;
283 
284 /*
285  * Allocate lgroup array statically
286  */
287 static lgrp_t				lgrp_space[NLGRP];
288 static int				nlgrps_alloc;
289 
290 
291 /*
292  * Number of nodes in system
293  */
294 uint_t			lgrp_plat_node_cnt = 1;
295 
296 /*
297  * Configuration Parameters for Probing
298  * - lgrp_plat_probe_flags	Flags to specify enabling probing, probe
299  *				operation, etc.
300  * - lgrp_plat_probe_nrounds	How many rounds of probing to do
301  * - lgrp_plat_probe_nsamples	Number of samples to take when probing each
302  *				node
303  * - lgrp_plat_probe_nreads	Number of times to read vendor ID from
304  *				Northbridge for each probe
305  */
306 uint_t			lgrp_plat_probe_flags = 0;
307 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
308 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
309 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
310 
311 /*
312  * Enable use of ACPI System Resource Affinity Table (SRAT) and System
313  * Locality Information Table (SLIT)
314  */
315 int			lgrp_plat_srat_enable = 1;
316 int			lgrp_plat_slit_enable = 1;
317 
318 /*
319  * Static array to hold lgroup statistics
320  */
321 struct lgrp_stats	lgrp_stats[NLGRP];
322 
323 
324 /*
325  * Forward declarations of platform interface routines
326  */
327 void		plat_build_mem_nodes(struct memlist *list);
328 
329 int		plat_lgrphand_to_mem_node(lgrp_handle_t hand);
330 
331 lgrp_handle_t	plat_mem_node_to_lgrphand(int mnode);
332 
333 int		plat_mnode_xcheck(pfn_t pfncnt);
334 
335 int		plat_pfn_to_mem_node(pfn_t pfn);
336 
337 /*
338  * Forward declarations of lgroup platform interface routines
339  */
340 lgrp_t		*lgrp_plat_alloc(lgrp_id_t lgrpid);
341 
342 void		lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg);
343 
344 lgrp_handle_t	lgrp_plat_cpu_to_hand(processorid_t id);
345 
346 void		lgrp_plat_init(void);
347 
348 int		lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to);
349 
350 void		lgrp_plat_main_init(void);
351 
352 int		lgrp_plat_max_lgrps(void);
353 
354 pgcnt_t		lgrp_plat_mem_size(lgrp_handle_t plathand,
355     lgrp_mem_query_t query);
356 
357 lgrp_handle_t	lgrp_plat_pfn_to_hand(pfn_t pfn);
358 
359 void		lgrp_plat_probe(void);
360 
361 lgrp_handle_t	lgrp_plat_root_hand(void);
362 
363 
364 /*
365  * Forward declarations of local routines
366  */
367 static int	is_opteron(void);
368 
369 static int	lgrp_plat_cpu_node_update(node_domain_map_t *node_domain,
370     cpu_node_map_t *cpu_node, int nentries, uint32_t apicid, uint32_t domain);
371 
372 static int	lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node);
373 
374 static int	lgrp_plat_domain_to_node(node_domain_map_t *node_domain,
375     uint32_t domain);
376 
377 static void	lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
378     lgrp_plat_latency_stats_t *lat_stats,
379     lgrp_plat_probe_stats_t *probe_stats);
380 
381 static int	lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
382     lgrp_plat_latency_stats_t *lat_stats);
383 
384 static pgcnt_t	lgrp_plat_mem_size_default(lgrp_handle_t, lgrp_mem_query_t);
385 
386 static int	lgrp_plat_node_domain_update(node_domain_map_t *node_domain,
387     uint32_t domain);
388 
389 static int	lgrp_plat_node_memory_update(node_domain_map_t *node_domain,
390     node_phys_addr_map_t *node_memory, uint64_t start, uint64_t end,
391     uint32_t domain);
392 
393 static hrtime_t	lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
394     lgrp_plat_probe_mem_config_t *probe_mem_config,
395     lgrp_plat_latency_stats_t *lat_stats,
396     lgrp_plat_probe_stats_t *probe_stats);
397 
398 static int	lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node,
399     int boot_ncpus);
400 
401 static int	lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
402     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats);
403 
404 static int	lgrp_plat_process_srat(struct srat *tp, int cpu_count,
405     uint_t *node_cnt, node_domain_map_t *node_domain, cpu_node_map_t *cpu_node,
406     node_phys_addr_map_t *node_memory);
407 
408 static int	lgrp_plat_srat_domains(struct srat *tp);
409 
410 static void	lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
411     lgrp_plat_latency_stats_t *lat_stats);
412 
413 static void	opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
414     node_phys_addr_map_t *node_memory);
415 
416 static hrtime_t	opt_probe_vendor(int dest_node, int nreads);
417 
418 
419 /*
420  * PLATFORM INTERFACE ROUTINES
421  */
422 
423 /*
424  * Configure memory nodes for machines with more than one node (ie NUMA)
425  */
426 void
427 plat_build_mem_nodes(struct memlist *list)
428 {
429 	pfn_t		cur_start;	/* start addr of subrange */
430 	pfn_t		cur_end;	/* end addr of subrange */
431 	pfn_t		start;		/* start addr of whole range */
432 	pfn_t		end;		/* end addr of whole range */
433 
434 	/*
435 	 * Boot install lists are arranged <addr, len>, ...
436 	 */
437 	while (list) {
438 		int	node;
439 
440 		start = list->address >> PAGESHIFT;
441 		end = (list->address + list->size - 1) >> PAGESHIFT;
442 
443 		if (start > physmax) {
444 			list = list->next;
445 			continue;
446 		}
447 		if (end > physmax)
448 			end = physmax;
449 
450 		/*
451 		 * When there is only one memnode, just add memory to memnode
452 		 */
453 		if (max_mem_nodes == 1) {
454 			mem_node_add_slice(start, end);
455 			list = list->next;
456 			continue;
457 		}
458 
459 		/*
460 		 * mem_node_add_slice() expects to get a memory range that
461 		 * is within one memnode, so need to split any memory range
462 		 * that spans multiple memnodes into subranges that are each
463 		 * contained within one memnode when feeding them to
464 		 * mem_node_add_slice()
465 		 */
466 		cur_start = start;
467 		do {
468 			node = plat_pfn_to_mem_node(cur_start);
469 
470 			/*
471 			 * Panic if DRAM address map registers or SRAT say
472 			 * memory in node doesn't exist or address from
473 			 * boot installed memory list entry isn't in this node.
474 			 * This shouldn't happen and rest of code can't deal
475 			 * with this if it does.
476 			 */
477 			if (node < 0 || node >= lgrp_plat_node_cnt ||
478 			    !lgrp_plat_node_memory[node].exists ||
479 			    cur_start < lgrp_plat_node_memory[node].start ||
480 			    cur_start > lgrp_plat_node_memory[node].end) {
481 				cmn_err(CE_PANIC, "Don't know which memnode "
482 				    "to add installed memory address 0x%lx\n",
483 				    cur_start);
484 			}
485 
486 			/*
487 			 * End of current subrange should not span memnodes
488 			 */
489 			cur_end = end;
490 			if (lgrp_plat_node_memory[node].exists &&
491 			    cur_end > lgrp_plat_node_memory[node].end)
492 				cur_end = lgrp_plat_node_memory[node].end;
493 
494 			mem_node_add_slice(cur_start, cur_end);
495 
496 			/*
497 			 * Next subrange starts after end of current one
498 			 */
499 			cur_start = cur_end + 1;
500 		} while (cur_end < end);
501 
502 		list = list->next;
503 	}
504 	mem_node_physalign = 0;
505 	mem_node_pfn_shift = 0;
506 }
507 
508 
509 int
510 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
511 {
512 	if (max_mem_nodes == 1)
513 		return (0);
514 
515 	return ((int)hand);
516 }
517 
518 
519 /*
520  * plat_mnode_xcheck: checks the node memory ranges to see if there is a pfncnt
521  * range of pages aligned on pfncnt that crosses an node boundary. Returns 1 if
522  * a crossing is found and returns 0 otherwise.
523  */
524 int
525 plat_mnode_xcheck(pfn_t pfncnt)
526 {
527 	int	node, prevnode = -1, basenode;
528 	pfn_t	ea, sa;
529 
530 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
531 
532 		if (lgrp_plat_node_memory[node].exists == 0)
533 			continue;
534 
535 		if (prevnode == -1) {
536 			prevnode = node;
537 			basenode = node;
538 			continue;
539 		}
540 
541 		/* assume x86 node pfn ranges are in increasing order */
542 		ASSERT(lgrp_plat_node_memory[node].start >
543 		    lgrp_plat_node_memory[prevnode].end);
544 
545 		/*
546 		 * continue if the starting address of node is not contiguous
547 		 * with the previous node.
548 		 */
549 
550 		if (lgrp_plat_node_memory[node].start !=
551 		    (lgrp_plat_node_memory[prevnode].end + 1)) {
552 			basenode = node;
553 			prevnode = node;
554 			continue;
555 		}
556 
557 		/* check if the starting address of node is pfncnt aligned */
558 		if ((lgrp_plat_node_memory[node].start & (pfncnt - 1)) != 0) {
559 
560 			/*
561 			 * at this point, node starts at an unaligned boundary
562 			 * and is contiguous with the previous node(s) to
563 			 * basenode. Check if there is an aligned contiguous
564 			 * range of length pfncnt that crosses this boundary.
565 			 */
566 
567 			sa = P2ALIGN(lgrp_plat_node_memory[prevnode].end,
568 			    pfncnt);
569 			ea = P2ROUNDUP((lgrp_plat_node_memory[node].start),
570 			    pfncnt);
571 
572 			ASSERT((ea - sa) == pfncnt);
573 			if (sa >= lgrp_plat_node_memory[basenode].start &&
574 			    ea <= (lgrp_plat_node_memory[node].end + 1))
575 				return (1);
576 		}
577 		prevnode = node;
578 	}
579 	return (0);
580 }
581 
582 
583 lgrp_handle_t
584 plat_mem_node_to_lgrphand(int mnode)
585 {
586 	if (max_mem_nodes == 1)
587 		return (LGRP_DEFAULT_HANDLE);
588 
589 	return ((lgrp_handle_t)mnode);
590 }
591 
592 
593 int
594 plat_pfn_to_mem_node(pfn_t pfn)
595 {
596 	int	node;
597 
598 	if (max_mem_nodes == 1)
599 		return (0);
600 
601 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
602 		/*
603 		 * Skip nodes with no memory
604 		 */
605 		if (!lgrp_plat_node_memory[node].exists)
606 			continue;
607 
608 		if (pfn >= lgrp_plat_node_memory[node].start &&
609 		    pfn <= lgrp_plat_node_memory[node].end)
610 			return (node);
611 	}
612 
613 	/*
614 	 * Didn't find memnode where this PFN lives which should never happen
615 	 */
616 	ASSERT(node < lgrp_plat_node_cnt);
617 	return (-1);
618 }
619 
620 
621 /*
622  * LGROUP PLATFORM INTERFACE ROUTINES
623  */
624 
625 /*
626  * Allocate additional space for an lgroup.
627  */
628 /* ARGSUSED */
629 lgrp_t *
630 lgrp_plat_alloc(lgrp_id_t lgrpid)
631 {
632 	lgrp_t *lgrp;
633 
634 	lgrp = &lgrp_space[nlgrps_alloc++];
635 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
636 		return (NULL);
637 	return (lgrp);
638 }
639 
640 
641 /*
642  * Platform handling for (re)configuration changes
643  */
644 /* ARGSUSED */
645 void
646 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
647 {
648 }
649 
650 
651 /*
652  * Return the platform handle for the lgroup containing the given CPU
653  */
654 /* ARGSUSED */
655 lgrp_handle_t
656 lgrp_plat_cpu_to_hand(processorid_t id)
657 {
658 	lgrp_handle_t	hand;
659 
660 	if (lgrp_plat_node_cnt == 1)
661 		return (LGRP_DEFAULT_HANDLE);
662 
663 	hand = (lgrp_handle_t)lgrp_plat_cpu_to_node(cpu[id],
664 	    lgrp_plat_cpu_node);
665 
666 	ASSERT(hand != (lgrp_handle_t)-1);
667 	if (hand == (lgrp_handle_t)-1)
668 		return (LGRP_NULL_HANDLE);
669 
670 	return (hand);
671 }
672 
673 
674 /*
675  * Platform-specific initialization of lgroups
676  */
677 void
678 lgrp_plat_init(void)
679 {
680 #if defined(__xpv)
681 	/*
682 	 * XXPV	For now, the hypervisor treats all memory equally.
683 	 */
684 	lgrp_plat_node_cnt = max_mem_nodes = 1;
685 #else	/* __xpv */
686 	uint_t	probe_op;
687 
688 	/*
689 	 * Initialize as a UMA machine
690 	 */
691 	if (lgrp_topo_ht_limit() == 1) {
692 		lgrp_plat_node_cnt = max_mem_nodes = 1;
693 		return;
694 	}
695 
696 	/*
697 	 * Read boot property with CPU to APIC ID mapping table/array and fill
698 	 * in CPU to node ID mapping table with APIC ID for each CPU
699 	 */
700 	lgrp_plat_cpu_apicid_error =
701 	    lgrp_plat_process_cpu_apicids(lgrp_plat_cpu_node, boot_max_ncpus);
702 
703 	/*
704 	 * Determine which CPUs and memory are local to each other and number
705 	 * of NUMA nodes by reading ACPI System Resource Affinity Table (SRAT)
706 	 */
707 	if (!lgrp_plat_cpu_apicid_error) {
708 		lgrp_plat_srat_error = lgrp_plat_process_srat(srat_ptr,
709 		    boot_max_ncpus, &lgrp_plat_node_cnt, lgrp_plat_node_domain,
710 		    lgrp_plat_cpu_node, lgrp_plat_node_memory);
711 	}
712 
713 	/*
714 	 * Try to use PCI config space registers on Opteron if there's an error
715 	 * processing CPU to APIC ID mapping or SRAT
716 	 */
717 	if ((lgrp_plat_cpu_apicid_error != 0 || lgrp_plat_srat_error != 0) &&
718 	    is_opteron())
719 		opt_get_numa_config(&lgrp_plat_node_cnt, &lgrp_plat_mem_intrlv,
720 		    lgrp_plat_node_memory);
721 
722 	/*
723 	 * Don't bother to setup system for multiple lgroups and only use one
724 	 * memory node when memory is interleaved between any nodes or there is
725 	 * only one NUMA node
726 	 *
727 	 * NOTE: May need to change this for Dynamic Reconfiguration (DR)
728 	 *	 when and if it happens for x86/x64
729 	 */
730 	if (lgrp_plat_mem_intrlv || lgrp_plat_node_cnt == 1) {
731 		lgrp_plat_node_cnt = max_mem_nodes = 1;
732 		(void) lgrp_topo_ht_limit_set(1);
733 		return;
734 	}
735 
736 	/*
737 	 * Leaf lgroups on x86/x64 architectures contain one physical
738 	 * processor chip. Tune lgrp_expand_proc_thresh and
739 	 * lgrp_expand_proc_diff so that lgrp_choose() will spread
740 	 * things out aggressively.
741 	 */
742 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
743 	lgrp_expand_proc_diff = 0;
744 
745 	/*
746 	 * There should be one memnode (physical page free list(s)) for
747 	 * each node
748 	 */
749 	max_mem_nodes = lgrp_plat_node_cnt;
750 
751 	/*
752 	 * Initialize min and max latency before reading SLIT or probing
753 	 */
754 	lgrp_plat_lat_stats.latency_min = -1;
755 	lgrp_plat_lat_stats.latency_max = 0;
756 
757 	/*
758 	 * Determine how far each NUMA node is from each other by
759 	 * reading ACPI System Locality Information Table (SLIT) if it
760 	 * exists
761 	 */
762 	lgrp_plat_slit_error = lgrp_plat_process_slit(slit_ptr,
763 	    lgrp_plat_node_cnt, lgrp_plat_node_memory,
764 	    &lgrp_plat_lat_stats);
765 	if (lgrp_plat_slit_error == 0)
766 		return;
767 
768 	/*
769 	 * Probe to determine latency between NUMA nodes when SLIT
770 	 * doesn't exist or make sense
771 	 */
772 	lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_ENABLE;
773 
774 	/*
775 	 * Specify whether to probe using vendor ID register or page copy
776 	 * if hasn't been specified already or is overspecified
777 	 */
778 	probe_op = lgrp_plat_probe_flags &
779 	    (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
780 
781 	if (probe_op == 0 ||
782 	    probe_op == (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR)) {
783 		lgrp_plat_probe_flags &=
784 		    ~(LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
785 		if (is_opteron())
786 			lgrp_plat_probe_flags |=
787 			    LGRP_PLAT_PROBE_VENDOR;
788 		else
789 			lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_PGCPY;
790 	}
791 
792 	/*
793 	 * Probing errors can mess up the lgroup topology and
794 	 * force us fall back to a 2 level lgroup topology.
795 	 * Here we bound how tall the lgroup topology can grow
796 	 * in hopes of avoiding any anamolies in probing from
797 	 * messing up the lgroup topology by limiting the
798 	 * accuracy of the latency topology.
799 	 *
800 	 * Assume that nodes will at least be configured in a
801 	 * ring, so limit height of lgroup topology to be less
802 	 * than number of nodes on a system with 4 or more
803 	 * nodes
804 	 */
805 	if (lgrp_plat_node_cnt >= 4 && lgrp_topo_ht_limit() ==
806 	    lgrp_topo_ht_limit_default())
807 		(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
808 #endif	/* __xpv */
809 }
810 
811 
812 /*
813  * Return latency between "from" and "to" lgroups
814  *
815  * This latency number can only be used for relative comparison
816  * between lgroups on the running system, cannot be used across platforms,
817  * and may not reflect the actual latency.  It is platform and implementation
818  * specific, so platform gets to decide its value.  It would be nice if the
819  * number was at least proportional to make comparisons more meaningful though.
820  */
821 /* ARGSUSED */
822 int
823 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
824 {
825 	lgrp_handle_t	src, dest;
826 	int		node;
827 
828 	if (max_mem_nodes == 1)
829 		return (0);
830 
831 	/*
832 	 * Return max latency for root lgroup
833 	 */
834 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
835 		return (lgrp_plat_lat_stats.latency_max);
836 
837 	src = from;
838 	dest = to;
839 
840 	/*
841 	 * Return 0 for nodes (lgroup platform handles) out of range
842 	 */
843 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
844 		return (0);
845 
846 	/*
847 	 * Probe from current CPU if its lgroup latencies haven't been set yet
848 	 * and we are trying to get latency from current CPU to some node
849 	 */
850 	node = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
851 	ASSERT(node >= 0 && node < lgrp_plat_node_cnt);
852 	if (lgrp_plat_lat_stats.latencies[src][src] == 0 && node == src)
853 		lgrp_plat_probe();
854 
855 	return (lgrp_plat_lat_stats.latencies[src][dest]);
856 }
857 
858 
859 /*
860  * Platform-specific initialization
861  */
862 void
863 lgrp_plat_main_init(void)
864 {
865 	int	curnode;
866 	int	ht_limit;
867 	int	i;
868 
869 	/*
870 	 * Print a notice that MPO is disabled when memory is interleaved
871 	 * across nodes....Would do this when it is discovered, but can't
872 	 * because it happens way too early during boot....
873 	 */
874 	if (lgrp_plat_mem_intrlv)
875 		cmn_err(CE_NOTE,
876 		    "MPO disabled because memory is interleaved\n");
877 
878 	/*
879 	 * Don't bother to do any probing if it is disabled, there is only one
880 	 * node, or the height of the lgroup topology less than or equal to 2
881 	 */
882 	ht_limit = lgrp_topo_ht_limit();
883 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
884 	    max_mem_nodes == 1 || ht_limit <= 2) {
885 		/*
886 		 * Setup lgroup latencies for 2 level lgroup topology
887 		 * (ie. local and remote only) if they haven't been set yet
888 		 */
889 		if (ht_limit == 2 && lgrp_plat_lat_stats.latency_min == -1 &&
890 		    lgrp_plat_lat_stats.latency_max == 0)
891 			lgrp_plat_2level_setup(lgrp_plat_node_memory,
892 			    &lgrp_plat_lat_stats);
893 		return;
894 	}
895 
896 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
897 		/*
898 		 * Should have been able to probe from CPU 0 when it was added
899 		 * to lgroup hierarchy, but may not have been able to then
900 		 * because it happens so early in boot that gethrtime() hasn't
901 		 * been initialized.  (:-(
902 		 */
903 		curnode = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
904 		ASSERT(curnode >= 0 && curnode < lgrp_plat_node_cnt);
905 		if (lgrp_plat_lat_stats.latencies[curnode][curnode] == 0)
906 			lgrp_plat_probe();
907 
908 		return;
909 	}
910 
911 	/*
912 	 * When probing memory, use one page for every sample to determine
913 	 * lgroup topology and taking multiple samples
914 	 */
915 	if (lgrp_plat_probe_mem_config.probe_memsize == 0)
916 		lgrp_plat_probe_mem_config.probe_memsize = PAGESIZE *
917 		    lgrp_plat_probe_nsamples;
918 
919 	/*
920 	 * Map memory in each node needed for probing to determine latency
921 	 * topology
922 	 */
923 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
924 		int	mnode;
925 
926 		/*
927 		 * Skip this node and leave its probe page NULL
928 		 * if it doesn't have any memory
929 		 */
930 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
931 		if (!mem_node_config[mnode].exists) {
932 			lgrp_plat_probe_mem_config.probe_va[i] = NULL;
933 			continue;
934 		}
935 
936 		/*
937 		 * Allocate one kernel virtual page
938 		 */
939 		lgrp_plat_probe_mem_config.probe_va[i] = vmem_alloc(heap_arena,
940 		    lgrp_plat_probe_mem_config.probe_memsize, VM_NOSLEEP);
941 		if (lgrp_plat_probe_mem_config.probe_va[i] == NULL) {
942 			cmn_err(CE_WARN,
943 			    "lgrp_plat_main_init: couldn't allocate memory");
944 			return;
945 		}
946 
947 		/*
948 		 * Get PFN for first page in each node
949 		 */
950 		lgrp_plat_probe_mem_config.probe_pfn[i] =
951 		    mem_node_config[mnode].physbase;
952 
953 		/*
954 		 * Map virtual page to first page in node
955 		 */
956 		hat_devload(kas.a_hat, lgrp_plat_probe_mem_config.probe_va[i],
957 		    lgrp_plat_probe_mem_config.probe_memsize,
958 		    lgrp_plat_probe_mem_config.probe_pfn[i],
959 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
960 		    HAT_LOAD_NOCONSIST);
961 	}
962 
963 	/*
964 	 * Probe from current CPU
965 	 */
966 	lgrp_plat_probe();
967 }
968 
969 
970 /*
971  * Return the maximum number of lgrps supported by the platform.
972  * Before lgrp topology is known it returns an estimate based on the number of
973  * nodes. Once topology is known it returns the actual maximim number of lgrps
974  * created. Since x86/x64 doesn't support Dynamic Reconfiguration (DR) and
975  * dynamic addition of new nodes, this number may not grow during system
976  * lifetime (yet).
977  */
978 int
979 lgrp_plat_max_lgrps(void)
980 {
981 	return (lgrp_topo_initialized ?
982 	    lgrp_alloc_max + 1 :
983 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
984 }
985 
986 
987 /*
988  * Return the number of free pages in an lgroup.
989  *
990  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
991  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
992  * number of allocatable base pagesize pages corresponding to the
993  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
994  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
995  * memory installed, regardless of whether or not it's usable.
996  */
997 pgcnt_t
998 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
999 {
1000 	int	mnode;
1001 	pgcnt_t npgs = (pgcnt_t)0;
1002 	extern struct memlist *phys_avail;
1003 	extern struct memlist *phys_install;
1004 
1005 
1006 	if (plathand == LGRP_DEFAULT_HANDLE)
1007 		return (lgrp_plat_mem_size_default(plathand, query));
1008 
1009 	if (plathand != LGRP_NULL_HANDLE) {
1010 		mnode = plat_lgrphand_to_mem_node(plathand);
1011 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1012 			switch (query) {
1013 			case LGRP_MEM_SIZE_FREE:
1014 				npgs = MNODE_PGCNT(mnode);
1015 				break;
1016 			case LGRP_MEM_SIZE_AVAIL:
1017 				npgs = mem_node_memlist_pages(mnode,
1018 				    phys_avail);
1019 				break;
1020 			case LGRP_MEM_SIZE_INSTALL:
1021 				npgs = mem_node_memlist_pages(mnode,
1022 				    phys_install);
1023 				break;
1024 			default:
1025 				break;
1026 			}
1027 		}
1028 	}
1029 	return (npgs);
1030 }
1031 
1032 
1033 /*
1034  * Return the platform handle of the lgroup that contains the physical memory
1035  * corresponding to the given page frame number
1036  */
1037 /* ARGSUSED */
1038 lgrp_handle_t
1039 lgrp_plat_pfn_to_hand(pfn_t pfn)
1040 {
1041 	int	mnode;
1042 
1043 	if (max_mem_nodes == 1)
1044 		return (LGRP_DEFAULT_HANDLE);
1045 
1046 	if (pfn > physmax)
1047 		return (LGRP_NULL_HANDLE);
1048 
1049 	mnode = plat_pfn_to_mem_node(pfn);
1050 	if (mnode < 0)
1051 		return (LGRP_NULL_HANDLE);
1052 
1053 	return (MEM_NODE_2_LGRPHAND(mnode));
1054 }
1055 
1056 
1057 /*
1058  * Probe memory in each node from current CPU to determine latency topology
1059  *
1060  * The probing code will probe the vendor ID register on the Northbridge of
1061  * Opteron processors and probe memory for other processors by default.
1062  *
1063  * Since probing is inherently error prone, the code takes laps across all the
1064  * nodes probing from each node to each of the other nodes some number of
1065  * times.  Furthermore, each node is probed some number of times before moving
1066  * onto the next one during each lap.  The minimum latency gotten between nodes
1067  * is kept as the latency between the nodes.
1068  *
1069  * After all that,  the probe times are adjusted by normalizing values that are
1070  * close to each other and local latencies are made the same.  Lastly, the
1071  * latencies are verified to make sure that certain conditions are met (eg.
1072  * local < remote, latency(a, b) == latency(b, a), etc.).
1073  *
1074  * If any of the conditions aren't met, the code will export a NUMA
1075  * configuration with the local CPUs and memory given by the SRAT or PCI config
1076  * space registers and one remote memory latency since it can't tell exactly
1077  * how far each node is from each other.
1078  */
1079 void
1080 lgrp_plat_probe(void)
1081 {
1082 	int				from;
1083 	int				i;
1084 	lgrp_plat_latency_stats_t	*lat_stats;
1085 	hrtime_t			probe_time;
1086 	int				to;
1087 
1088 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
1089 	    max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1090 		return;
1091 
1092 	/*
1093 	 * Determine ID of node containing current CPU
1094 	 */
1095 	from = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
1096 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
1097 	if (srat_ptr && lgrp_plat_srat_enable && !lgrp_plat_srat_error)
1098 		ASSERT(lgrp_plat_node_domain[from].exists);
1099 
1100 	/*
1101 	 * Don't need to probe if got times already
1102 	 */
1103 	lat_stats = &lgrp_plat_lat_stats;
1104 	if (lat_stats->latencies[from][from] != 0)
1105 		return;
1106 
1107 	/*
1108 	 * Read vendor ID in Northbridge or read and write page(s)
1109 	 * in each node from current CPU and remember how long it takes,
1110 	 * so we can build latency topology of machine later.
1111 	 * This should approximate the memory latency between each node.
1112 	 */
1113 	for (i = 0; i < lgrp_plat_probe_nrounds; i++) {
1114 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1115 			/*
1116 			 * Get probe time and bail out if can't get it yet
1117 			 */
1118 			probe_time = lgrp_plat_probe_time(to,
1119 			    lgrp_plat_cpu_node, &lgrp_plat_probe_mem_config,
1120 			    &lgrp_plat_lat_stats, &lgrp_plat_probe_stats);
1121 			if (probe_time == 0)
1122 				return;
1123 
1124 			/*
1125 			 * Keep lowest probe time as latency between nodes
1126 			 */
1127 			if (lat_stats->latencies[from][to] == 0 ||
1128 			    probe_time < lat_stats->latencies[from][to])
1129 				lat_stats->latencies[from][to] = probe_time;
1130 
1131 			/*
1132 			 * Update overall minimum and maximum probe times
1133 			 * across all nodes
1134 			 */
1135 			if (probe_time < lat_stats->latency_min ||
1136 			    lat_stats->latency_min == -1)
1137 				lat_stats->latency_min = probe_time;
1138 			if (probe_time > lat_stats->latency_max)
1139 				lat_stats->latency_max = probe_time;
1140 		}
1141 	}
1142 
1143 	/*
1144 	 * - Fix up latencies such that local latencies are same,
1145 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1146 	 *
1147 	 * - Verify that latencies look ok
1148 	 *
1149 	 * - Fallback to just optimizing for local and remote if
1150 	 *   latencies didn't look right
1151 	 */
1152 	lgrp_plat_latency_adjust(lgrp_plat_node_memory, &lgrp_plat_lat_stats,
1153 	    &lgrp_plat_probe_stats);
1154 	lgrp_plat_probe_stats.probe_error_code =
1155 	    lgrp_plat_latency_verify(lgrp_plat_node_memory,
1156 	    &lgrp_plat_lat_stats);
1157 	if (lgrp_plat_probe_stats.probe_error_code)
1158 		lgrp_plat_2level_setup(lgrp_plat_node_memory,
1159 		    &lgrp_plat_lat_stats);
1160 }
1161 
1162 
1163 /*
1164  * Return platform handle for root lgroup
1165  */
1166 lgrp_handle_t
1167 lgrp_plat_root_hand(void)
1168 {
1169 	return (LGRP_DEFAULT_HANDLE);
1170 }
1171 
1172 
1173 /*
1174  * INTERNAL ROUTINES
1175  */
1176 
1177 
1178 /*
1179  * Update CPU to node mapping for given CPU and proximity domain (and returns
1180  * negative numbers for errors and positive ones for success)
1181  */
1182 static int
1183 lgrp_plat_cpu_node_update(node_domain_map_t *node_domain,
1184     cpu_node_map_t *cpu_node, int nentries, uint32_t apicid, uint32_t domain)
1185 {
1186 	uint_t	i;
1187 	int	node;
1188 
1189 	/*
1190 	 * Get node number for proximity domain
1191 	 */
1192 	node = lgrp_plat_domain_to_node(node_domain, domain);
1193 	if (node == -1) {
1194 		node = lgrp_plat_node_domain_update(node_domain, domain);
1195 		if (node == -1)
1196 			return (-1);
1197 	}
1198 
1199 	/*
1200 	 * Search for entry with given APIC ID and fill in its node and
1201 	 * proximity domain IDs (if they haven't been set already)
1202 	 */
1203 	for (i = 0; i < nentries; i++) {
1204 		/*
1205 		 * Skip nonexistent entries and ones without matching APIC ID
1206 		 */
1207 		if (!cpu_node[i].exists || cpu_node[i].apicid != apicid)
1208 			continue;
1209 
1210 		/*
1211 		 * Just return if entry completely and correctly filled in
1212 		 * already
1213 		 */
1214 		if (cpu_node[i].prox_domain == domain &&
1215 		    cpu_node[i].node == node)
1216 			return (1);
1217 
1218 		/*
1219 		 * Fill in node and proximity domain IDs
1220 		 */
1221 		cpu_node[i].prox_domain = domain;
1222 		cpu_node[i].node = node;
1223 
1224 		return (0);
1225 	}
1226 
1227 	/*
1228 	 * Return error when entry for APIC ID wasn't found in table
1229 	 */
1230 	return (-2);
1231 }
1232 
1233 
1234 /*
1235  * Get node ID for given CPU
1236  */
1237 static int
1238 lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node)
1239 {
1240 	processorid_t	cpuid;
1241 
1242 	if (cp == NULL)
1243 		return (-1);
1244 
1245 	cpuid = cp->cpu_id;
1246 	if (cpuid < 0 || cpuid >= max_ncpus)
1247 		return (-1);
1248 
1249 	/*
1250 	 * SRAT doesn't exist, isn't enabled, or there was an error processing
1251 	 * it, so return chip ID for Opteron and -1 otherwise.
1252 	 */
1253 	if (srat_ptr == NULL || !lgrp_plat_srat_enable ||
1254 	    lgrp_plat_srat_error) {
1255 		if (is_opteron())
1256 			return (pg_plat_hw_instance_id(cp, PGHW_CHIP));
1257 		return (-1);
1258 	}
1259 
1260 	/*
1261 	 * Return -1 when CPU to node ID mapping entry doesn't exist for given
1262 	 * CPU
1263 	 */
1264 	if (!cpu_node[cpuid].exists)
1265 		return (-1);
1266 
1267 	return (cpu_node[cpuid].node);
1268 }
1269 
1270 
1271 /*
1272  * Return node number for given proximity domain/system locality
1273  */
1274 static int
1275 lgrp_plat_domain_to_node(node_domain_map_t *node_domain, uint32_t domain)
1276 {
1277 	uint_t	node;
1278 	uint_t	start;
1279 
1280 	/*
1281 	 * Hash proximity domain ID into node to domain mapping table (array),
1282 	 * search for entry with matching proximity domain ID, and return index
1283 	 * of matching entry as node ID.
1284 	 */
1285 	node = start = NODE_DOMAIN_HASH(domain);
1286 	do {
1287 		if (node_domain[node].prox_domain == domain &&
1288 		    node_domain[node].exists)
1289 			return (node);
1290 		node = NODE_DOMAIN_HASH(node + 1);
1291 	} while (node != start);
1292 	return (-1);
1293 }
1294 
1295 
1296 /*
1297  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
1298  * be considered same
1299  */
1300 #define	LGRP_LAT_TOLERANCE_SHIFT	4
1301 
1302 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
1303 
1304 
1305 /*
1306  * Adjust latencies between nodes to be symmetric, normalize latencies between
1307  * any nodes that are within some tolerance to be same, and make local
1308  * latencies be same
1309  */
1310 static void
1311 lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
1312     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1313 {
1314 	int				i;
1315 	int				j;
1316 	int				k;
1317 	int				l;
1318 	u_longlong_t			max;
1319 	u_longlong_t			min;
1320 	u_longlong_t			t;
1321 	u_longlong_t			t1;
1322 	u_longlong_t			t2;
1323 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
1324 	int				lat_corrected[MAX_NODES][MAX_NODES];
1325 
1326 	/*
1327 	 * Nothing to do when this is an UMA machine or don't have args needed
1328 	 */
1329 	if (max_mem_nodes == 1)
1330 		return;
1331 
1332 	ASSERT(node_memory != NULL && lat_stats != NULL &&
1333 	    probe_stats != NULL);
1334 
1335 	/*
1336 	 * Make sure that latencies are symmetric between any two nodes
1337 	 * (ie. latency(node0, node1) == latency(node1, node0))
1338 	 */
1339 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1340 		if (!node_memory[i].exists)
1341 			continue;
1342 
1343 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1344 			if (!node_memory[j].exists)
1345 				continue;
1346 
1347 			t1 = lat_stats->latencies[i][j];
1348 			t2 = lat_stats->latencies[j][i];
1349 
1350 			if (t1 == 0 || t2 == 0 || t1 == t2)
1351 				continue;
1352 
1353 			/*
1354 			 * Latencies should be same
1355 			 * - Use minimum of two latencies which should be same
1356 			 * - Track suspect probe times not within tolerance of
1357 			 *   min value
1358 			 * - Remember how much values are corrected by
1359 			 */
1360 			if (t1 > t2) {
1361 				t = t2;
1362 				probe_stats->probe_errors[i][j] += t1 - t2;
1363 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
1364 					probe_stats->probe_suspect[i][j]++;
1365 					probe_stats->probe_suspect[j][i]++;
1366 				}
1367 			} else if (t2 > t1) {
1368 				t = t1;
1369 				probe_stats->probe_errors[j][i] += t2 - t1;
1370 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
1371 					probe_stats->probe_suspect[i][j]++;
1372 					probe_stats->probe_suspect[j][i]++;
1373 				}
1374 			}
1375 
1376 			lat_stats->latencies[i][j] =
1377 			    lat_stats->latencies[j][i] = t;
1378 			lgrp_config(cflag, t1, t);
1379 			lgrp_config(cflag, t2, t);
1380 		}
1381 	}
1382 
1383 	/*
1384 	 * Keep track of which latencies get corrected
1385 	 */
1386 	for (i = 0; i < MAX_NODES; i++)
1387 		for (j = 0; j < MAX_NODES; j++)
1388 			lat_corrected[i][j] = 0;
1389 
1390 	/*
1391 	 * For every two nodes, see whether there is another pair of nodes which
1392 	 * are about the same distance apart and make the latencies be the same
1393 	 * if they are close enough together
1394 	 */
1395 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1396 		if (!node_memory[i].exists)
1397 			continue;
1398 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1399 			if (!node_memory[j].exists)
1400 				continue;
1401 			/*
1402 			 * Pick one pair of nodes (i, j)
1403 			 * and get latency between them
1404 			 */
1405 			t1 = lat_stats->latencies[i][j];
1406 
1407 			/*
1408 			 * Skip this pair of nodes if there isn't a latency
1409 			 * for it yet
1410 			 */
1411 			if (t1 == 0)
1412 				continue;
1413 
1414 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1415 				if (!node_memory[k].exists)
1416 					continue;
1417 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
1418 					if (!node_memory[l].exists)
1419 						continue;
1420 					/*
1421 					 * Pick another pair of nodes (k, l)
1422 					 * not same as (i, j) and get latency
1423 					 * between them
1424 					 */
1425 					if (k == i && l == j)
1426 						continue;
1427 
1428 					t2 = lat_stats->latencies[k][l];
1429 
1430 					/*
1431 					 * Skip this pair of nodes if there
1432 					 * isn't a latency for it yet
1433 					 */
1434 
1435 					if (t2 == 0)
1436 						continue;
1437 
1438 					/*
1439 					 * Skip nodes (k, l) if they already
1440 					 * have same latency as (i, j) or
1441 					 * their latency isn't close enough to
1442 					 * be considered/made the same
1443 					 */
1444 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
1445 					    t1 >> lgrp_plat_probe_lt_shift) ||
1446 					    (t2 > t1 && t2 - t1 >
1447 					    t2 >> lgrp_plat_probe_lt_shift))
1448 						continue;
1449 
1450 					/*
1451 					 * Make latency(i, j) same as
1452 					 * latency(k, l), try to use latency
1453 					 * that has been adjusted already to get
1454 					 * more consistency (if possible), and
1455 					 * remember which latencies were
1456 					 * adjusted for next time
1457 					 */
1458 					if (lat_corrected[i][j]) {
1459 						t = t1;
1460 						lgrp_config(cflag, t2, t);
1461 						t2 = t;
1462 					} else if (lat_corrected[k][l]) {
1463 						t = t2;
1464 						lgrp_config(cflag, t1, t);
1465 						t1 = t;
1466 					} else {
1467 						if (t1 > t2)
1468 							t = t2;
1469 						else
1470 							t = t1;
1471 						lgrp_config(cflag, t1, t);
1472 						lgrp_config(cflag, t2, t);
1473 						t1 = t2 = t;
1474 					}
1475 
1476 					lat_stats->latencies[i][j] =
1477 					    lat_stats->latencies[k][l] = t;
1478 
1479 					lat_corrected[i][j] =
1480 					    lat_corrected[k][l] = 1;
1481 				}
1482 			}
1483 		}
1484 	}
1485 
1486 	/*
1487 	 * Local latencies should be same
1488 	 * - Find min and max local latencies
1489 	 * - Make all local latencies be minimum
1490 	 */
1491 	min = -1;
1492 	max = 0;
1493 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1494 		if (!node_memory[i].exists)
1495 			continue;
1496 		t = lat_stats->latencies[i][i];
1497 		if (t == 0)
1498 			continue;
1499 		if (min == -1 || t < min)
1500 			min = t;
1501 		if (t > max)
1502 			max = t;
1503 	}
1504 	if (min != max) {
1505 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1506 			int	local;
1507 
1508 			if (!node_memory[i].exists)
1509 				continue;
1510 
1511 			local = lat_stats->latencies[i][i];
1512 			if (local == 0)
1513 				continue;
1514 
1515 			/*
1516 			 * Track suspect probe times that aren't within
1517 			 * tolerance of minimum local latency and how much
1518 			 * probe times are corrected by
1519 			 */
1520 			if (local - min > min >> lgrp_plat_probe_lt_shift)
1521 				probe_stats->probe_suspect[i][i]++;
1522 
1523 			probe_stats->probe_errors[i][i] += local - min;
1524 
1525 			/*
1526 			 * Make local latencies be minimum
1527 			 */
1528 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
1529 			lat_stats->latencies[i][i] = min;
1530 		}
1531 	}
1532 
1533 	/*
1534 	 * Determine max probe time again since just adjusted latencies
1535 	 */
1536 	lat_stats->latency_max = 0;
1537 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1538 		if (!node_memory[i].exists)
1539 			continue;
1540 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1541 			if (!node_memory[j].exists)
1542 				continue;
1543 			t = lat_stats->latencies[i][j];
1544 			if (t > lat_stats->latency_max)
1545 				lat_stats->latency_max = t;
1546 		}
1547 	}
1548 }
1549 
1550 
1551 /*
1552  * Verify following about latencies between nodes:
1553  *
1554  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
1555  * - Local latencies same
1556  * - Local < remote
1557  * - Number of latencies seen is reasonable
1558  * - Number of occurrences of a given latency should be more than 1
1559  *
1560  * Returns:
1561  *	0	Success
1562  *	-1	Not symmetric
1563  *	-2	Local latencies not same
1564  *	-3	Local >= remote
1565  */
1566 static int
1567 lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
1568     lgrp_plat_latency_stats_t *lat_stats)
1569 {
1570 	int				i;
1571 	int				j;
1572 	u_longlong_t			t1;
1573 	u_longlong_t			t2;
1574 
1575 	ASSERT(node_memory != NULL && lat_stats != NULL);
1576 
1577 	/*
1578 	 * Nothing to do when this is an UMA machine, lgroup topology is
1579 	 * limited to 2 levels, or there aren't any probe times yet
1580 	 */
1581 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
1582 	    lat_stats->latencies[0][0] == 0)
1583 		return (0);
1584 
1585 	/*
1586 	 * Make sure that latencies are symmetric between any two nodes
1587 	 * (ie. latency(node0, node1) == latency(node1, node0))
1588 	 */
1589 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1590 		if (!node_memory[i].exists)
1591 			continue;
1592 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1593 			if (!node_memory[j].exists)
1594 				continue;
1595 			t1 = lat_stats->latencies[i][j];
1596 			t2 = lat_stats->latencies[j][i];
1597 
1598 			if (t1 == 0 || t2 == 0 || t1 == t2)
1599 				continue;
1600 
1601 			return (-1);
1602 		}
1603 	}
1604 
1605 	/*
1606 	 * Local latencies should be same
1607 	 */
1608 	t1 = lat_stats->latencies[0][0];
1609 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
1610 		if (!node_memory[i].exists)
1611 			continue;
1612 
1613 		t2 = lat_stats->latencies[i][i];
1614 		if (t2 == 0)
1615 			continue;
1616 
1617 		if (t1 == 0) {
1618 			t1 = t2;
1619 			continue;
1620 		}
1621 
1622 		if (t1 != t2)
1623 			return (-2);
1624 	}
1625 
1626 	/*
1627 	 * Local latencies should be less than remote
1628 	 */
1629 	if (t1) {
1630 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1631 			if (!node_memory[i].exists)
1632 				continue;
1633 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1634 				if (!node_memory[j].exists)
1635 					continue;
1636 				t2 = lat_stats->latencies[i][j];
1637 				if (i == j || t2 == 0)
1638 					continue;
1639 
1640 				if (t1 >= t2)
1641 					return (-3);
1642 			}
1643 		}
1644 	}
1645 
1646 	return (0);
1647 }
1648 
1649 
1650 /*
1651  * Return the number of free, allocatable, or installed
1652  * pages in an lgroup
1653  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1654  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1655  */
1656 /* ARGSUSED */
1657 static pgcnt_t
1658 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1659 {
1660 	struct memlist *mlist;
1661 	pgcnt_t npgs = 0;
1662 	extern struct memlist *phys_avail;
1663 	extern struct memlist *phys_install;
1664 
1665 	switch (query) {
1666 	case LGRP_MEM_SIZE_FREE:
1667 		return ((pgcnt_t)freemem);
1668 	case LGRP_MEM_SIZE_AVAIL:
1669 		memlist_read_lock();
1670 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1671 			npgs += btop(mlist->size);
1672 		memlist_read_unlock();
1673 		return (npgs);
1674 	case LGRP_MEM_SIZE_INSTALL:
1675 		memlist_read_lock();
1676 		for (mlist = phys_install; mlist; mlist = mlist->next)
1677 			npgs += btop(mlist->size);
1678 		memlist_read_unlock();
1679 		return (npgs);
1680 	default:
1681 		return ((pgcnt_t)0);
1682 	}
1683 }
1684 
1685 
1686 /*
1687  * Update node to proximity domain mappings for given domain and return node ID
1688  */
1689 static int
1690 lgrp_plat_node_domain_update(node_domain_map_t *node_domain, uint32_t domain)
1691 {
1692 	uint_t	node;
1693 	uint_t	start;
1694 
1695 	/*
1696 	 * Hash proximity domain ID into node to domain mapping table (array)
1697 	 * and add entry for it into first non-existent or matching entry found
1698 	 */
1699 	node = start = NODE_DOMAIN_HASH(domain);
1700 	do {
1701 		/*
1702 		 * Entry doesn't exist yet, so create one for this proximity
1703 		 * domain and return node ID which is index into mapping table.
1704 		 */
1705 		if (!node_domain[node].exists) {
1706 			node_domain[node].exists = 1;
1707 			node_domain[node].prox_domain = domain;
1708 			return (node);
1709 		}
1710 
1711 		/*
1712 		 * Entry exists for this proximity domain already, so just
1713 		 * return node ID (index into table).
1714 		 */
1715 		if (node_domain[node].prox_domain == domain)
1716 			return (node);
1717 		node = NODE_DOMAIN_HASH(node + 1);
1718 	} while (node != start);
1719 
1720 	/*
1721 	 * Ran out of supported number of entries which shouldn't happen....
1722 	 */
1723 	ASSERT(node != start);
1724 	return (-1);
1725 }
1726 
1727 
1728 /*
1729  * Update node memory information for given proximity domain with specified
1730  * starting and ending physical address range (and return positive numbers for
1731  * success and negative ones for errors)
1732  */
1733 static int
1734 lgrp_plat_node_memory_update(node_domain_map_t *node_domain,
1735     node_phys_addr_map_t *node_memory, uint64_t start, uint64_t end,
1736     uint32_t domain)
1737 {
1738 	int	node;
1739 
1740 	/*
1741 	 * Get node number for proximity domain
1742 	 */
1743 	node = lgrp_plat_domain_to_node(node_domain, domain);
1744 	if (node == -1) {
1745 		node = lgrp_plat_node_domain_update(node_domain, domain);
1746 		if (node == -1)
1747 			return (-1);
1748 	}
1749 
1750 	/*
1751 	 * Create entry in table for node if it doesn't exist
1752 	 */
1753 	if (!node_memory[node].exists) {
1754 		node_memory[node].exists = 1;
1755 		node_memory[node].start = btop(start);
1756 		node_memory[node].end = btop(end);
1757 		node_memory[node].prox_domain = domain;
1758 		return (0);
1759 	}
1760 
1761 	/*
1762 	 * Entry already exists for this proximity domain
1763 	 *
1764 	 * There may be more than one SRAT memory entry for a domain, so we may
1765 	 * need to update existing start or end address for the node.
1766 	 */
1767 	if (node_memory[node].prox_domain == domain) {
1768 		if (btop(start) < node_memory[node].start)
1769 			node_memory[node].start = btop(start);
1770 		if (btop(end) > node_memory[node].end)
1771 			node_memory[node].end = btop(end);
1772 		return (1);
1773 	}
1774 	return (-2);
1775 }
1776 
1777 
1778 /*
1779  * Return time needed to probe from current CPU to memory in given node
1780  */
1781 static hrtime_t
1782 lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
1783     lgrp_plat_probe_mem_config_t *probe_mem_config,
1784     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1785 {
1786 	caddr_t			buf;
1787 	hrtime_t		elapsed;
1788 	hrtime_t		end;
1789 	int			from;
1790 	int			i;
1791 	int			ipl;
1792 	hrtime_t		max;
1793 	hrtime_t		min;
1794 	hrtime_t		start;
1795 	extern int		use_sse_pagecopy;
1796 
1797 	/*
1798 	 * Determine ID of node containing current CPU
1799 	 */
1800 	from = lgrp_plat_cpu_to_node(CPU, cpu_node);
1801 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
1802 
1803 	/*
1804 	 * Do common work for probing main memory
1805 	 */
1806 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_PGCPY) {
1807 		/*
1808 		 * Skip probing any nodes without memory and
1809 		 * set probe time to 0
1810 		 */
1811 		if (probe_mem_config->probe_va[to] == NULL) {
1812 			lat_stats->latencies[from][to] = 0;
1813 			return (0);
1814 		}
1815 
1816 		/*
1817 		 * Invalidate caches once instead of once every sample
1818 		 * which should cut cost of probing by a lot
1819 		 */
1820 		probe_stats->flush_cost = gethrtime();
1821 		invalidate_cache();
1822 		probe_stats->flush_cost = gethrtime() -
1823 		    probe_stats->flush_cost;
1824 		probe_stats->probe_cost_total += probe_stats->flush_cost;
1825 	}
1826 
1827 	/*
1828 	 * Probe from current CPU to given memory using specified operation
1829 	 * and take specified number of samples
1830 	 */
1831 	max = 0;
1832 	min = -1;
1833 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1834 		probe_stats->probe_cost = gethrtime();
1835 
1836 		/*
1837 		 * Can't measure probe time if gethrtime() isn't working yet
1838 		 */
1839 		if (probe_stats->probe_cost == 0 && gethrtime() == 0)
1840 			return (0);
1841 
1842 		if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
1843 			/*
1844 			 * Measure how long it takes to read vendor ID from
1845 			 * Northbridge
1846 			 */
1847 			elapsed = opt_probe_vendor(to, lgrp_plat_probe_nreads);
1848 		} else {
1849 			/*
1850 			 * Measure how long it takes to copy page
1851 			 * on top of itself
1852 			 */
1853 			buf = probe_mem_config->probe_va[to] + (i * PAGESIZE);
1854 
1855 			kpreempt_disable();
1856 			ipl = splhigh();
1857 			start = gethrtime();
1858 			if (use_sse_pagecopy)
1859 				hwblkpagecopy(buf, buf);
1860 			else
1861 				bcopy(buf, buf, PAGESIZE);
1862 			end = gethrtime();
1863 			elapsed = end - start;
1864 			splx(ipl);
1865 			kpreempt_enable();
1866 		}
1867 
1868 		probe_stats->probe_cost = gethrtime() -
1869 		    probe_stats->probe_cost;
1870 		probe_stats->probe_cost_total += probe_stats->probe_cost;
1871 
1872 		if (min == -1 || elapsed < min)
1873 			min = elapsed;
1874 		if (elapsed > max)
1875 			max = elapsed;
1876 	}
1877 
1878 	/*
1879 	 * Update minimum and maximum probe times between
1880 	 * these two nodes
1881 	 */
1882 	if (min < probe_stats->probe_min[from][to] ||
1883 	    probe_stats->probe_min[from][to] == 0)
1884 		probe_stats->probe_min[from][to] = min;
1885 
1886 	if (max > probe_stats->probe_max[from][to])
1887 		probe_stats->probe_max[from][to] = max;
1888 
1889 	return (min);
1890 }
1891 
1892 
1893 /*
1894  * Read boot property with CPU to APIC ID array and fill in CPU to node ID
1895  * mapping table with APIC ID for each CPU
1896  *
1897  * NOTE: This code assumes that CPU IDs are assigned in order that they appear
1898  *       in in cpu_apicid_array boot property which is based on and follows
1899  *	 same ordering as processor list in ACPI MADT.  If the code in
1900  *	 usr/src/uts/i86pc/io/pcplusmp/apic.c that reads MADT and assigns
1901  *	 CPU IDs ever changes, then this code will need to change too....
1902  */
1903 static int
1904 lgrp_plat_process_cpu_apicids(cpu_node_map_t *cpu_node, int boot_ncpus)
1905 {
1906 	char	*boot_prop_name = BP_CPU_APICID_ARRAY;
1907 	uint8_t	cpu_apicid_array[UINT8_MAX + 1];
1908 	int	i;
1909 	int	boot_prop_len;
1910 
1911 	/*
1912 	 * Nothing to do when no array to fill in or not enough CPUs
1913 	 */
1914 	if (cpu_node == NULL || boot_ncpus <= 1)
1915 		return (1);
1916 
1917 	/*
1918 	 * Check length of property value
1919 	 */
1920 	boot_prop_len = BOP_GETPROPLEN(bootops, boot_prop_name);
1921 	if (boot_prop_len <= 0 || boot_prop_len > UINT8_MAX)
1922 		return (2);
1923 
1924 	/*
1925 	 * Get CPU to APIC ID property value
1926 	 */
1927 	if (BOP_GETPROP(bootops, boot_prop_name, cpu_apicid_array) < 0)
1928 		return (3);
1929 
1930 	/*
1931 	 * Fill in CPU to node ID mapping table with APIC ID for each CPU
1932 	 */
1933 	for (i = 0; i < boot_ncpus; i++) {
1934 		cpu_node[i].exists = 1;
1935 		cpu_node[i].apicid = cpu_apicid_array[i];
1936 	}
1937 
1938 	return (0);
1939 }
1940 
1941 
1942 /*
1943  * Read ACPI System Locality Information Table (SLIT) to determine how far each
1944  * NUMA node is from each other
1945  */
1946 static int
1947 lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
1948     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats)
1949 {
1950 	int		i;
1951 	int		j;
1952 	int		localities;
1953 	hrtime_t	max;
1954 	hrtime_t	min;
1955 	int		retval;
1956 	uint8_t		*slit_entries;
1957 
1958 	if (tp == NULL || !lgrp_plat_slit_enable)
1959 		return (1);
1960 
1961 	if (lat_stats == NULL)
1962 		return (2);
1963 
1964 	localities = tp->number;
1965 	if (localities != node_cnt)
1966 		return (3);
1967 
1968 	min = lat_stats->latency_min;
1969 	max = lat_stats->latency_max;
1970 
1971 	/*
1972 	 * Fill in latency matrix based on SLIT entries
1973 	 */
1974 	slit_entries = tp->entry;
1975 	for (i = 0; i < localities; i++) {
1976 		for (j = 0; j < localities; j++) {
1977 			uint8_t	latency;
1978 
1979 			latency = slit_entries[(i * localities) + j];
1980 			lat_stats->latencies[i][j] = latency;
1981 			if (latency < min || min == -1)
1982 				min = latency;
1983 			if (latency > max)
1984 				max = latency;
1985 		}
1986 	}
1987 
1988 	/*
1989 	 * Verify that latencies/distances given in SLIT look reasonable
1990 	 */
1991 	retval = lgrp_plat_latency_verify(node_memory, lat_stats);
1992 
1993 	if (retval) {
1994 		/*
1995 		 * Reinitialize (zero) latency table since SLIT doesn't look
1996 		 * right
1997 		 */
1998 		for (i = 0; i < localities; i++) {
1999 			for (j = 0; j < localities; j++)
2000 				lat_stats->latencies[i][j] = 0;
2001 		}
2002 	} else {
2003 		/*
2004 		 * Update min and max latencies seen since SLIT looks valid
2005 		 */
2006 		lat_stats->latency_min = min;
2007 		lat_stats->latency_max = max;
2008 	}
2009 
2010 	return (retval);
2011 }
2012 
2013 
2014 /*
2015  * Read ACPI System Resource Affinity Table (SRAT) to determine which CPUs
2016  * and memory are local to each other in the same NUMA node
2017  */
2018 static int
2019 lgrp_plat_process_srat(struct srat *tp, int cpu_count, uint_t *node_cnt,
2020     node_domain_map_t *node_domain, cpu_node_map_t *cpu_node,
2021     node_phys_addr_map_t *node_memory)
2022 {
2023 	struct srat_item	*srat_end;
2024 	int			i;
2025 	struct srat_item	*item;
2026 	int			proc_entry_count;
2027 
2028 	if (tp == NULL || !lgrp_plat_srat_enable)
2029 		return (1);
2030 
2031 	/*
2032 	 * Determine number of nodes by counting number of proximity domains in
2033 	 * SRAT
2034 	 */
2035 	if (node_cnt) {
2036 		int	nodes;
2037 
2038 		nodes = lgrp_plat_srat_domains(tp);
2039 		if (nodes < 0) {
2040 			*node_cnt = 1;
2041 			return (2);
2042 		}
2043 		*node_cnt = nodes;
2044 	}
2045 
2046 	/*
2047 	 * Walk through SRAT, examining each CPU and memory entry to determine
2048 	 * which CPUs and memory belong to which node.
2049 	 */
2050 	item = tp->list;
2051 	srat_end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2052 	proc_entry_count = 0;
2053 	while (item < srat_end) {
2054 		uint32_t	apic_id;
2055 		uint32_t	domain;
2056 		uint64_t	end;
2057 		uint64_t	length;
2058 		uint64_t	start;
2059 
2060 		switch (item->type) {
2061 		case SRAT_PROCESSOR:	/* CPU entry */
2062 			if (!(item->i.p.flags & SRAT_ENABLED) ||
2063 			    cpu_node == NULL)
2064 				break;
2065 
2066 			/*
2067 			 * Calculate domain (node) ID and fill in APIC ID to
2068 			 * domain/node mapping table
2069 			 */
2070 			domain = item->i.p.domain1;
2071 			for (i = 0; i < 3; i++) {
2072 				domain += item->i.p.domain2[i] <<
2073 				    ((i + 1) * 8);
2074 			}
2075 			apic_id = item->i.p.apic_id;
2076 
2077 			if (lgrp_plat_cpu_node_update(node_domain, cpu_node,
2078 			    cpu_count, apic_id, domain) < 0)
2079 				return (3);
2080 
2081 			proc_entry_count++;
2082 			break;
2083 
2084 		case SRAT_MEMORY:	/* memory entry */
2085 			if (!(item->i.m.flags & SRAT_ENABLED) ||
2086 			    node_memory == NULL)
2087 				break;
2088 
2089 			/*
2090 			 * Get domain (node) ID and fill in domain/node
2091 			 * to memory mapping table
2092 			 */
2093 			domain = item->i.m.domain;
2094 			start = item->i.m.base_addr;
2095 			length = item->i.m.len;
2096 			end = start + length - 1;
2097 
2098 			if (lgrp_plat_node_memory_update(node_domain,
2099 			    node_memory, start, end, domain) < 0)
2100 				return (4);
2101 			break;
2102 
2103 		default:
2104 			break;
2105 		}
2106 
2107 		item = (struct srat_item *)((uintptr_t)item + item->len);
2108 	}
2109 
2110 	/*
2111 	 * Should have seen at least as many SRAT processor entries as CPUs
2112 	 */
2113 	if (proc_entry_count >= cpu_count)
2114 		return (5);
2115 
2116 	return (0);
2117 }
2118 
2119 
2120 /*
2121  * Return number of proximity domains given in ACPI SRAT
2122  */
2123 static int
2124 lgrp_plat_srat_domains(struct srat *tp)
2125 {
2126 	int			domain_cnt;
2127 	struct srat_item	*end;
2128 	int			i;
2129 	struct srat_item	*item;
2130 	node_domain_map_t	node_domain[MAX_NODES];
2131 
2132 
2133 	if (tp == NULL || !lgrp_plat_srat_enable)
2134 		return (1);
2135 
2136 	/*
2137 	 * Walk through SRAT, examining each CPU and memory entry to determine
2138 	 * proximity domain ID for each.
2139 	 */
2140 	domain_cnt = 0;
2141 	item = tp->list;
2142 	end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2143 	bzero(node_domain, MAX_NODES * sizeof (node_domain_map_t));
2144 	while (item < end) {
2145 		uint32_t	domain;
2146 		boolean_t	overflow;
2147 		uint_t		start;
2148 
2149 		switch (item->type) {
2150 		case SRAT_PROCESSOR:	/* CPU entry */
2151 			if (!(item->i.p.flags & SRAT_ENABLED))
2152 				break;
2153 			domain = item->i.p.domain1;
2154 			for (i = 0; i < 3; i++) {
2155 				domain += item->i.p.domain2[i] <<
2156 				    ((i + 1) * 8);
2157 			}
2158 			break;
2159 
2160 		case SRAT_MEMORY:	/* memory entry */
2161 			if (!(item->i.m.flags & SRAT_ENABLED))
2162 				break;
2163 			domain = item->i.m.domain;
2164 			break;
2165 
2166 		default:
2167 			break;
2168 		}
2169 
2170 		/*
2171 		 * Count and keep track of which proximity domain IDs seen
2172 		 */
2173 		start = i = domain % MAX_NODES;
2174 		overflow = B_TRUE;
2175 		do {
2176 			/*
2177 			 * Create entry for proximity domain and increment
2178 			 * count when no entry exists where proximity domain
2179 			 * hashed
2180 			 */
2181 			if (!node_domain[i].exists) {
2182 				node_domain[i].exists = 1;
2183 				node_domain[i].prox_domain = domain;
2184 				domain_cnt++;
2185 				overflow = B_FALSE;
2186 				break;
2187 			}
2188 
2189 			/*
2190 			 * Nothing to do when proximity domain seen already
2191 			 * and its entry exists
2192 			 */
2193 			if (node_domain[i].prox_domain == domain) {
2194 				overflow = B_FALSE;
2195 				break;
2196 			}
2197 
2198 			/*
2199 			 * Entry exists where proximity domain hashed, but for
2200 			 * different proximity domain so keep search for empty
2201 			 * slot to put it or matching entry whichever comes
2202 			 * first.
2203 			 */
2204 			i = (i + 1) % MAX_NODES;
2205 		} while (i != start);
2206 
2207 		/*
2208 		 * Didn't find empty or matching entry which means have more
2209 		 * proximity domains than supported nodes (:-(
2210 		 */
2211 		ASSERT(overflow != B_TRUE);
2212 		if (overflow == B_TRUE)
2213 			return (-1);
2214 
2215 		item = (struct srat_item *)((uintptr_t)item + item->len);
2216 	}
2217 	return (domain_cnt);
2218 }
2219 
2220 
2221 /*
2222  * Set lgroup latencies for 2 level lgroup topology
2223  */
2224 static void
2225 lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
2226     lgrp_plat_latency_stats_t *lat_stats)
2227 {
2228 	int	i;
2229 
2230 	ASSERT(node_memory != NULL && lat_stats != NULL);
2231 
2232 	if (lgrp_plat_node_cnt >= 4)
2233 		cmn_err(CE_NOTE,
2234 		    "MPO only optimizing for local and remote\n");
2235 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
2236 		int	j;
2237 
2238 		if (!node_memory[i].exists)
2239 			continue;
2240 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
2241 			if (!node_memory[j].exists)
2242 				continue;
2243 			if (i == j)
2244 				lat_stats->latencies[i][j] = 2;
2245 			else
2246 				lat_stats->latencies[i][j] = 3;
2247 		}
2248 	}
2249 	lat_stats->latency_min = 2;
2250 	lat_stats->latency_max = 3;
2251 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
2252 }
2253 
2254 
2255 /*
2256  * The following Opteron specific constants, macros, types, and routines define
2257  * PCI configuration space registers and how to read them to determine the NUMA
2258  * configuration of *supported* Opteron processors.  They provide the same
2259  * information that may be gotten from the ACPI System Resource Affinity Table
2260  * (SRAT) if it exists on the machine of interest.
2261  *
2262  * The AMD BIOS and Kernel Developer's Guide (BKDG) for the processor family
2263  * of interest describes all of these registers and their contents.  The main
2264  * registers used by this code to determine the NUMA configuration of the
2265  * machine are the node ID register for the number of NUMA nodes and the DRAM
2266  * address map registers for the physical address range of each node.
2267  *
2268  * NOTE: The format and how to determine the NUMA configuration using PCI
2269  *	 config space registers may change or may not be supported in future
2270  *	 Opteron processor families.
2271  */
2272 
2273 /*
2274  * How many bits to shift Opteron DRAM Address Map base and limit registers
2275  * to get actual value
2276  */
2277 #define	OPT_DRAMADDR_HI_LSHIFT_ADDR	40	/* shift left for address */
2278 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
2279 
2280 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
2281 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
2282 
2283 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
2284 
2285 /*
2286  * Macros to derive addresses from Opteron DRAM Address Map registers
2287  */
2288 #define	OPT_DRAMADDR_HI(reg) \
2289 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
2290 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
2291 
2292 #define	OPT_DRAMADDR_LO(reg) \
2293 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
2294 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
2295 
2296 #define	OPT_DRAMADDR(high, low) \
2297 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
2298 
2299 /*
2300  * Bit masks defining what's in Opteron DRAM Address Map base register
2301  */
2302 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
2303 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
2304 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
2305 
2306 /*
2307  * Bit masks defining what's in Opteron DRAM Address Map limit register
2308  */
2309 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
2310 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
2311 
2312 
2313 /*
2314  * Opteron Node ID register in PCI configuration space contains
2315  * number of nodes in system, etc. for Opteron K8.  The following
2316  * constants and macros define its contents, structure, and access.
2317  */
2318 
2319 /*
2320  * Bit masks defining what's in Opteron Node ID register
2321  */
2322 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
2323 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
2324 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
2325 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
2326 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
2327 
2328 /*
2329  * How many bits in Opteron Node ID register to shift right to get actual value
2330  */
2331 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
2332 
2333 /*
2334  * Macros to get values from Opteron Node ID register
2335  */
2336 #define	OPT_NODE_CNT(reg) \
2337 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
2338 
2339 /*
2340  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
2341  * "in/out" instructions
2342  *
2343  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
2344  *	 other uses should just do MMIO to access PCI ECS.
2345  *	 Must enable special bit in Northbridge Configuration Register on
2346  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
2347  *	 using "in/out" instructions and restore special bit after done
2348  *	 accessing PCI ECS.
2349  */
2350 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
2351 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
2352 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
2353 	    ((((reg) >> 8) & 0xf) << 24))
2354 
2355 /*
2356  * PCI configuration space registers accessed by specifying
2357  * a bus, device, function, and offset.  The following constants
2358  * define the values needed to access Opteron K8 configuration
2359  * info to determine its node topology
2360  */
2361 
2362 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
2363 
2364 /*
2365  * Opteron PCI configuration space register function values
2366  */
2367 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
2368 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
2369 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
2370 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
2371 
2372 /*
2373  * PCI Configuration Space register offsets
2374  */
2375 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
2376 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
2377 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
2378 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
2379 
2380 /*
2381  * Opteron PCI Configuration Space device IDs for nodes
2382  */
2383 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
2384 
2385 
2386 /*
2387  * Opteron DRAM address map gives base and limit for physical memory in a node
2388  */
2389 typedef	struct opt_dram_addr_map {
2390 	uint32_t	base_hi;
2391 	uint32_t	base_lo;
2392 	uint32_t	limit_hi;
2393 	uint32_t	limit_lo;
2394 } opt_dram_addr_map_t;
2395 
2396 
2397 /*
2398  * Supported AMD processor families
2399  */
2400 #define	AMD_FAMILY_HAMMER	15
2401 #define	AMD_FAMILY_GREYHOUND	16
2402 
2403 /*
2404  * Whether to have is_opteron() return 1 even when processor isn't supported
2405  */
2406 uint_t	is_opteron_override = 0;
2407 
2408 /*
2409  * AMD processor family for current CPU
2410  */
2411 uint_t	opt_family = 0;
2412 
2413 
2414 /*
2415  * Determine whether we're running on a supported AMD Opteron since reading
2416  * node count and DRAM address map registers may have different format or
2417  * may not be supported across processor families
2418  */
2419 static int
2420 is_opteron(void)
2421 {
2422 
2423 	if (x86_vendor != X86_VENDOR_AMD)
2424 		return (0);
2425 
2426 	opt_family = cpuid_getfamily(CPU);
2427 	if (opt_family == AMD_FAMILY_HAMMER ||
2428 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
2429 		return (1);
2430 	else
2431 		return (0);
2432 }
2433 
2434 
2435 /*
2436  * Determine NUMA configuration for Opteron from registers that live in PCI
2437  * configuration space
2438  */
2439 static void
2440 opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
2441     node_phys_addr_map_t *node_memory)
2442 {
2443 	uint_t				bus;
2444 	uint_t				dev;
2445 	struct opt_dram_addr_map	dram_map[MAX_NODES];
2446 	uint_t				node;
2447 	uint_t				node_info[MAX_NODES];
2448 	uint_t				off_hi;
2449 	uint_t				off_lo;
2450 	uint64_t			nb_cfg_reg;
2451 
2452 	/*
2453 	 * Read configuration registers from PCI configuration space to
2454 	 * determine node information, which memory is in each node, etc.
2455 	 *
2456 	 * Write to PCI configuration space address register to specify
2457 	 * which configuration register to read and read/write PCI
2458 	 * configuration space data register to get/set contents
2459 	 */
2460 	bus = OPT_PCS_BUS_CONFIG;
2461 	dev = OPT_PCS_DEV_NODE0;
2462 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
2463 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
2464 
2465 	/*
2466 	 * Read node ID register for node 0 to get node count
2467 	 */
2468 	node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
2469 	    OPT_PCS_OFF_NODEID);
2470 	*node_cnt = OPT_NODE_CNT(node_info[0]) + 1;
2471 
2472 	/*
2473 	 * If number of nodes is more than maximum supported, then set node
2474 	 * count to 1 and treat system as UMA instead of NUMA.
2475 	 */
2476 	if (*node_cnt > MAX_NODES) {
2477 		*node_cnt = 1;
2478 		return;
2479 	}
2480 
2481 	/*
2482 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
2483 	 * read high DRAM address map base and limit registers
2484 	 */
2485 	if (opt_family == AMD_FAMILY_GREYHOUND) {
2486 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
2487 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
2488 			wrmsr(MSR_AMD_NB_CFG,
2489 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
2490 	}
2491 
2492 	for (node = 0; node < *node_cnt; node++) {
2493 		uint32_t	base_hi;
2494 		uint32_t	base_lo;
2495 		uint32_t	limit_hi;
2496 		uint32_t	limit_lo;
2497 
2498 		/*
2499 		 * Read node ID register (except for node 0 which we just read)
2500 		 */
2501 		if (node > 0) {
2502 			node_info[node] = pci_getl_func(bus, dev,
2503 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
2504 		}
2505 
2506 		/*
2507 		 * Read DRAM base and limit registers which specify
2508 		 * physical memory range of each node
2509 		 */
2510 		if (opt_family != AMD_FAMILY_GREYHOUND)
2511 			base_hi = 0;
2512 		else {
2513 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2514 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2515 			base_hi = dram_map[node].base_hi =
2516 			    inl(PCI_CONFDATA);
2517 		}
2518 		base_lo = dram_map[node].base_lo = pci_getl_func(bus, dev,
2519 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
2520 
2521 		if ((dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN) &&
2522 		    mem_intrlv)
2523 			*mem_intrlv = *mem_intrlv + 1;
2524 
2525 		off_hi += 4;	/* high limit register offset */
2526 		if (opt_family != AMD_FAMILY_GREYHOUND)
2527 			limit_hi = 0;
2528 		else {
2529 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2530 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2531 			limit_hi = dram_map[node].limit_hi =
2532 			    inl(PCI_CONFDATA);
2533 		}
2534 
2535 		off_lo += 4;	/* low limit register offset */
2536 		limit_lo = dram_map[node].limit_lo = pci_getl_func(bus,
2537 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
2538 
2539 		/*
2540 		 * Increment device number to next node and register offsets
2541 		 * for DRAM base register of next node
2542 		 */
2543 		off_hi += 4;
2544 		off_lo += 4;
2545 		dev++;
2546 
2547 		/*
2548 		 * Both read and write enable bits must be enabled in DRAM
2549 		 * address map base register for physical memory to exist in
2550 		 * node
2551 		 */
2552 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
2553 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
2554 			/*
2555 			 * Mark node memory as non-existent and set start and
2556 			 * end addresses to be same in node_memory[]
2557 			 */
2558 			node_memory[node].exists = 0;
2559 			node_memory[node].start = node_memory[node].end =
2560 			    (pfn_t)-1;
2561 			continue;
2562 		}
2563 
2564 		/*
2565 		 * Mark node memory as existing and remember physical address
2566 		 * range of each node for use later
2567 		 */
2568 		node_memory[node].exists = 1;
2569 
2570 		node_memory[node].start = btop(OPT_DRAMADDR(base_hi, base_lo));
2571 
2572 		node_memory[node].end = btop(OPT_DRAMADDR(limit_hi, limit_lo) |
2573 		    OPT_DRAMADDR_LO_MASK_OFF);
2574 	}
2575 
2576 	/*
2577 	 * Restore PCI Extended Configuration Space enable bit
2578 	 */
2579 	if (opt_family == AMD_FAMILY_GREYHOUND) {
2580 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
2581 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
2582 	}
2583 }
2584 
2585 
2586 /*
2587  * Return average amount of time to read vendor ID register on Northbridge
2588  * N times on specified destination node from current CPU
2589  */
2590 static hrtime_t
2591 opt_probe_vendor(int dest_node, int nreads)
2592 {
2593 	int		cnt;
2594 	uint_t		dev;
2595 	/* LINTED: set but not used in function */
2596 	volatile uint_t	dev_vendor;
2597 	hrtime_t	elapsed;
2598 	hrtime_t	end;
2599 	int		ipl;
2600 	hrtime_t	start;
2601 
2602 	dev = OPT_PCS_DEV_NODE0 + dest_node;
2603 	kpreempt_disable();
2604 	ipl = spl8();
2605 	outl(PCI_CONFADD, PCI_CADDR1(0, dev, OPT_PCS_FUNC_DRAM,
2606 	    OPT_PCS_OFF_VENDOR));
2607 	start = gethrtime();
2608 	for (cnt = 0; cnt < nreads; cnt++)
2609 		dev_vendor = inl(PCI_CONFDATA);
2610 	end = gethrtime();
2611 	elapsed = (end - start) / nreads;
2612 	splx(ipl);
2613 	kpreempt_enable();
2614 	return (elapsed);
2615 }
2616