xref: /titanic_51/usr/src/uts/i86pc/os/lgrpplat.c (revision 678453a8ed49104d8adad58f3ba591bdc39883e8)
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[]		APIC ID to node ID mapping table
89  *					indexed by hashed APIC ID (only used
90  *					for SRAT)
91  *
92  * - lgrp_plat_lat_stats.latencies[][]	Table of latencies between same and
93  *					different nodes indexed by node ID
94  *
95  * - lgrp_plat_node_cnt			Number of NUMA nodes in system
96  *
97  * - lgrp_plat_node_domain[]		Node ID to proximity domain ID mapping
98  *					table indexed by node ID (only used
99  *					for SRAT)
100  *
101  * - lgrp_plat_node_memory[]		Table with physical address range for
102  *					each node indexed by node ID
103  *
104  * The code is implemented to make the following always be true:
105  *
106  *	lgroup platform handle == node ID == memnode ID
107  *
108  * Moreover, it allows for the proximity domain ID to be equal to all of the
109  * above as long as the proximity domains IDs are numbered from 0 to <number of
110  * nodes - 1>.  This is done by hashing each proximity domain ID into the range
111  * from 0 to <number of nodes - 1>.  Then proximity ID N will hash into node ID
112  * N and proximity domain ID N will be entered into lgrp_plat_node_domain[N]
113  * and be assigned node ID N.  If the proximity domain IDs aren't numbered
114  * from 0 to <number of nodes - 1>, then hashing the proximity domain IDs into
115  * lgrp_plat_node_domain[] will still work for assigning proximity domain IDs
116  * to node IDs.  However, the proximity domain IDs may not map to the
117  * equivalent node ID since we want to keep the node IDs numbered from 0 to
118  * <number of nodes - 1> to minimize cost of searching and potentially space.
119  */
120 
121 
122 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
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 CPU APIC ID into CPU to node mapping table using max_ncpus
169  * to minimize span of entries used
170  */
171 #define	CPU_NODE_HASH(apicid)		((apicid) % max_ncpus)
172 
173 /*
174  * Hash proximity domain ID into node to domain mapping table using to minimize
175  * span of entries used
176  */
177 #define	NODE_DOMAIN_HASH(domain)	((domain) % lgrp_plat_node_cnt)
178 
179 
180 /*
181  * CPU APIC ID to node ID mapping structure (only used with SRAT)
182  */
183 typedef	struct cpu_node_map {
184 	int		exists;
185 	uint_t		node;
186 	uint32_t	apicid;
187 	uint32_t	prox_domain;
188 } cpu_node_map_t;
189 
190 /*
191  * Latency statistics
192  */
193 typedef struct lgrp_plat_latency_stats {
194 	hrtime_t	latencies[MAX_NODES][MAX_NODES];
195 	hrtime_t	latency_max;
196 	hrtime_t	latency_min;
197 } lgrp_plat_latency_stats_t;
198 
199 /*
200  * Memory configuration for probing
201  */
202 typedef struct lgrp_plat_probe_mem_config {
203 	size_t	probe_memsize;		/* how much memory to probe per node */
204 	caddr_t	probe_va[MAX_NODES];	/* where memory mapped for probing */
205 	pfn_t	probe_pfn[MAX_NODES];	/* physical pages to map for probing */
206 } lgrp_plat_probe_mem_config_t;
207 
208 /*
209  * Statistics kept for probing
210  */
211 typedef struct lgrp_plat_probe_stats {
212 	hrtime_t	flush_cost;
213 	hrtime_t	probe_cost;
214 	hrtime_t	probe_cost_total;
215 	hrtime_t	probe_error_code;
216 	hrtime_t	probe_errors[MAX_NODES][MAX_NODES];
217 	int		probe_suspect[MAX_NODES][MAX_NODES];
218 	hrtime_t	probe_max[MAX_NODES][MAX_NODES];
219 	hrtime_t	probe_min[MAX_NODES][MAX_NODES];
220 } lgrp_plat_probe_stats_t;
221 
222 /*
223  * Node to proximity domain ID mapping structure (only used with SRAT)
224  */
225 typedef	struct node_domain_map {
226 	int		exists;
227 	uint32_t	prox_domain;
228 } node_domain_map_t;
229 
230 /*
231  * Node ID and starting and ending page for physical memory in node
232  */
233 typedef	struct node_phys_addr_map {
234 	pfn_t		start;
235 	pfn_t		end;
236 	int		exists;
237 	uint32_t	prox_domain;
238 } node_phys_addr_map_t;
239 
240 
241 /*
242  * CPU APIC ID to node ID mapping table (only used for SRAT)
243  */
244 static cpu_node_map_t			lgrp_plat_cpu_node[NCPU];
245 
246 /*
247  * Latency statistics
248  */
249 lgrp_plat_latency_stats_t		lgrp_plat_lat_stats;
250 
251 /*
252  * Whether memory is interleaved across nodes causing MPO to be disabled
253  */
254 static int				lgrp_plat_mem_intrlv = 0;
255 
256 /*
257  * Node ID to proximity domain ID mapping table (only used for SRAT)
258  */
259 static node_domain_map_t		lgrp_plat_node_domain[MAX_NODES];
260 
261 /*
262  * Physical address range for memory in each node
263  */
264 static node_phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
265 
266 /*
267  * Statistics gotten from probing
268  */
269 static lgrp_plat_probe_stats_t		lgrp_plat_probe_stats;
270 
271 /*
272  * Memory configuration for probing
273  */
274 static lgrp_plat_probe_mem_config_t	lgrp_plat_probe_mem_config;
275 
276 /*
277  * Error code from processing ACPI SRAT
278  */
279 static int				lgrp_plat_srat_error = 0;
280 
281 /*
282  * Error code from processing ACPI SLIT
283  */
284 static int				lgrp_plat_slit_error = 0;
285 
286 /*
287  * Allocate lgroup array statically
288  */
289 static lgrp_t				lgrp_space[NLGRP];
290 static int				nlgrps_alloc;
291 
292 
293 /*
294  * Number of nodes in system
295  */
296 uint_t			lgrp_plat_node_cnt = 1;
297 
298 /*
299  * Configuration Parameters for Probing
300  * - lgrp_plat_probe_flags	Flags to specify enabling probing, probe
301  *				operation, etc.
302  * - lgrp_plat_probe_nrounds	How many rounds of probing to do
303  * - lgrp_plat_probe_nsamples	Number of samples to take when probing each
304  *				node
305  * - lgrp_plat_probe_nreads	Number of times to read vendor ID from
306  *				Northbridge for each probe
307  */
308 uint_t			lgrp_plat_probe_flags = 0;
309 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
310 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
311 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
312 
313 /*
314  * Enable use of ACPI System Resource Affinity Table (SRAT) and System
315  * Locality Information Table (SLIT)
316  */
317 int			lgrp_plat_srat_enable = 1;
318 int			lgrp_plat_slit_enable = 1;
319 
320 /*
321  * Static array to hold lgroup statistics
322  */
323 struct lgrp_stats	lgrp_stats[NLGRP];
324 
325 
326 /*
327  * Forward declarations of platform interface routines
328  */
329 void		plat_build_mem_nodes(struct memlist *list);
330 
331 int		plat_lgrphand_to_mem_node(lgrp_handle_t hand);
332 
333 lgrp_handle_t	plat_mem_node_to_lgrphand(int mnode);
334 
335 int		plat_mnode_xcheck(pfn_t pfncnt);
336 
337 int		plat_pfn_to_mem_node(pfn_t pfn);
338 
339 /*
340  * Forward declarations of lgroup platform interface routines
341  */
342 lgrp_t		*lgrp_plat_alloc(lgrp_id_t lgrpid);
343 
344 void		lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg);
345 
346 lgrp_handle_t	lgrp_plat_cpu_to_hand(processorid_t id);
347 
348 void		lgrp_plat_init(void);
349 
350 int		lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to);
351 
352 void		lgrp_plat_main_init(void);
353 
354 int		lgrp_plat_max_lgrps(void);
355 
356 pgcnt_t		lgrp_plat_mem_size(lgrp_handle_t plathand,
357     lgrp_mem_query_t query);
358 
359 lgrp_handle_t	lgrp_plat_pfn_to_hand(pfn_t pfn);
360 
361 void		lgrp_plat_probe(void);
362 
363 lgrp_handle_t	lgrp_plat_root_hand(void);
364 
365 
366 /*
367  * Forward declarations of local routines
368  */
369 static int	is_opteron(void);
370 
371 static int	lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node);
372 
373 static int	lgrp_plat_domain_to_node(node_domain_map_t *node_domain,
374     uint32_t domain);
375 
376 static void	lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
377     lgrp_plat_latency_stats_t *lat_stats,
378     lgrp_plat_probe_stats_t *probe_stats);
379 
380 static int	lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
381     lgrp_plat_latency_stats_t *lat_stats);
382 
383 static pgcnt_t	lgrp_plat_mem_size_default(lgrp_handle_t, lgrp_mem_query_t);
384 
385 static int	lgrp_plat_node_domain_update(node_domain_map_t *node_domain,
386     uint32_t domain);
387 
388 static int	lgrp_plat_node_memory_update(node_domain_map_t *node_domain,
389     node_phys_addr_map_t *node_memory, uint64_t start, uint64_t end,
390     uint32_t domain);
391 
392 static hrtime_t	lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
393     lgrp_plat_probe_mem_config_t *probe_mem_config,
394     lgrp_plat_latency_stats_t *lat_stats,
395     lgrp_plat_probe_stats_t *probe_stats);
396 
397 static int	lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
398     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats);
399 
400 static int	lgrp_plat_process_srat(struct srat *tp, uint_t *node_cnt,
401     node_domain_map_t *node_domain, cpu_node_map_t *cpu_node,
402     node_phys_addr_map_t *node_memory);
403 
404 static int	lgrp_plat_srat_domains(struct srat *tp);
405 
406 static void	lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
407     lgrp_plat_latency_stats_t *lat_stats);
408 
409 static void	opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
410     node_phys_addr_map_t *node_memory);
411 
412 static hrtime_t	opt_probe_vendor(int dest_node, int nreads);
413 
414 
415 /*
416  * PLATFORM INTERFACE ROUTINES
417  */
418 
419 /*
420  * Configure memory nodes for machines with more than one node (ie NUMA)
421  */
422 void
423 plat_build_mem_nodes(struct memlist *list)
424 {
425 	pfn_t		cur_start;	/* start addr of subrange */
426 	pfn_t		cur_end;	/* end addr of subrange */
427 	pfn_t		start;		/* start addr of whole range */
428 	pfn_t		end;		/* end addr of whole range */
429 
430 	/*
431 	 * Boot install lists are arranged <addr, len>, ...
432 	 */
433 	while (list) {
434 		int	node;
435 
436 		start = list->address >> PAGESHIFT;
437 		end = (list->address + list->size - 1) >> PAGESHIFT;
438 
439 		if (start > physmax) {
440 			list = list->next;
441 			continue;
442 		}
443 		if (end > physmax)
444 			end = physmax;
445 
446 		/*
447 		 * When there is only one memnode, just add memory to memnode
448 		 */
449 		if (max_mem_nodes == 1) {
450 			mem_node_add_slice(start, end);
451 			list = list->next;
452 			continue;
453 		}
454 
455 		/*
456 		 * mem_node_add_slice() expects to get a memory range that
457 		 * is within one memnode, so need to split any memory range
458 		 * that spans multiple memnodes into subranges that are each
459 		 * contained within one memnode when feeding them to
460 		 * mem_node_add_slice()
461 		 */
462 		cur_start = start;
463 		do {
464 			node = plat_pfn_to_mem_node(cur_start);
465 
466 			/*
467 			 * Panic if DRAM address map registers or SRAT say
468 			 * memory in node doesn't exist or address from
469 			 * boot installed memory list entry isn't in this node.
470 			 * This shouldn't happen and rest of code can't deal
471 			 * with this if it does.
472 			 */
473 			if (node < 0 || node >= lgrp_plat_node_cnt ||
474 			    !lgrp_plat_node_memory[node].exists ||
475 			    cur_start < lgrp_plat_node_memory[node].start ||
476 			    cur_start > lgrp_plat_node_memory[node].end) {
477 				cmn_err(CE_PANIC, "Don't know which memnode "
478 				    "to add installed memory address 0x%lx\n",
479 				    cur_start);
480 			}
481 
482 			/*
483 			 * End of current subrange should not span memnodes
484 			 */
485 			cur_end = end;
486 			if (lgrp_plat_node_memory[node].exists &&
487 			    cur_end > lgrp_plat_node_memory[node].end)
488 				cur_end = lgrp_plat_node_memory[node].end;
489 
490 			mem_node_add_slice(cur_start, cur_end);
491 
492 			/*
493 			 * Next subrange starts after end of current one
494 			 */
495 			cur_start = cur_end + 1;
496 		} while (cur_end < end);
497 
498 		list = list->next;
499 	}
500 	mem_node_physalign = 0;
501 	mem_node_pfn_shift = 0;
502 }
503 
504 
505 int
506 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
507 {
508 	if (max_mem_nodes == 1)
509 		return (0);
510 
511 	return ((int)hand);
512 }
513 
514 
515 /*
516  * plat_mnode_xcheck: checks the node memory ranges to see if there is a pfncnt
517  * range of pages aligned on pfncnt that crosses an node boundary. Returns 1 if
518  * a crossing is found and returns 0 otherwise.
519  */
520 int
521 plat_mnode_xcheck(pfn_t pfncnt)
522 {
523 	int	node, prevnode = -1, basenode;
524 	pfn_t	ea, sa;
525 
526 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
527 
528 		if (lgrp_plat_node_memory[node].exists == 0)
529 			continue;
530 
531 		if (prevnode == -1) {
532 			prevnode = node;
533 			basenode = node;
534 			continue;
535 		}
536 
537 		/* assume x86 node pfn ranges are in increasing order */
538 		ASSERT(lgrp_plat_node_memory[node].start >
539 		    lgrp_plat_node_memory[prevnode].end);
540 
541 		/*
542 		 * continue if the starting address of node is not contiguous
543 		 * with the previous node.
544 		 */
545 
546 		if (lgrp_plat_node_memory[node].start !=
547 		    (lgrp_plat_node_memory[prevnode].end + 1)) {
548 			basenode = node;
549 			prevnode = node;
550 			continue;
551 		}
552 
553 		/* check if the starting address of node is pfncnt aligned */
554 		if ((lgrp_plat_node_memory[node].start & (pfncnt - 1)) != 0) {
555 
556 			/*
557 			 * at this point, node starts at an unaligned boundary
558 			 * and is contiguous with the previous node(s) to
559 			 * basenode. Check if there is an aligned contiguous
560 			 * range of length pfncnt that crosses this boundary.
561 			 */
562 
563 			sa = P2ALIGN(lgrp_plat_node_memory[prevnode].end,
564 			    pfncnt);
565 			ea = P2ROUNDUP((lgrp_plat_node_memory[node].start),
566 			    pfncnt);
567 
568 			ASSERT((ea - sa) == pfncnt);
569 			if (sa >= lgrp_plat_node_memory[basenode].start &&
570 			    ea <= (lgrp_plat_node_memory[node].end + 1))
571 				return (1);
572 		}
573 		prevnode = node;
574 	}
575 	return (0);
576 }
577 
578 
579 lgrp_handle_t
580 plat_mem_node_to_lgrphand(int mnode)
581 {
582 	if (max_mem_nodes == 1)
583 		return (LGRP_DEFAULT_HANDLE);
584 
585 	return ((lgrp_handle_t)mnode);
586 }
587 
588 
589 int
590 plat_pfn_to_mem_node(pfn_t pfn)
591 {
592 	int	node;
593 
594 	if (max_mem_nodes == 1)
595 		return (0);
596 
597 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
598 		/*
599 		 * Skip nodes with no memory
600 		 */
601 		if (!lgrp_plat_node_memory[node].exists)
602 			continue;
603 
604 		if (pfn >= lgrp_plat_node_memory[node].start &&
605 		    pfn <= lgrp_plat_node_memory[node].end)
606 			return (node);
607 	}
608 
609 	/*
610 	 * Didn't find memnode where this PFN lives which should never happen
611 	 */
612 	ASSERT(node < lgrp_plat_node_cnt);
613 	return (-1);
614 }
615 
616 
617 /*
618  * LGROUP PLATFORM INTERFACE ROUTINES
619  */
620 
621 /*
622  * Allocate additional space for an lgroup.
623  */
624 /* ARGSUSED */
625 lgrp_t *
626 lgrp_plat_alloc(lgrp_id_t lgrpid)
627 {
628 	lgrp_t *lgrp;
629 
630 	lgrp = &lgrp_space[nlgrps_alloc++];
631 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
632 		return (NULL);
633 	return (lgrp);
634 }
635 
636 
637 /*
638  * Platform handling for (re)configuration changes
639  */
640 /* ARGSUSED */
641 void
642 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
643 {
644 }
645 
646 
647 /*
648  * Return the platform handle for the lgroup containing the given CPU
649  */
650 /* ARGSUSED */
651 lgrp_handle_t
652 lgrp_plat_cpu_to_hand(processorid_t id)
653 {
654 	lgrp_handle_t	hand;
655 
656 	if (lgrp_plat_node_cnt == 1)
657 		return (LGRP_DEFAULT_HANDLE);
658 
659 	hand = (lgrp_handle_t)lgrp_plat_cpu_to_node(cpu[id],
660 	    lgrp_plat_cpu_node);
661 
662 	ASSERT(hand != (lgrp_handle_t)-1);
663 	if (hand == (lgrp_handle_t)-1)
664 		return (LGRP_NULL_HANDLE);
665 
666 	return (hand);
667 }
668 
669 
670 /*
671  * Platform-specific initialization of lgroups
672  */
673 void
674 lgrp_plat_init(void)
675 {
676 #if defined(__xpv)
677 	/*
678 	 * XXPV	For now, the hypervisor treats all memory equally.
679 	 */
680 	lgrp_plat_node_cnt = max_mem_nodes = 1;
681 #else	/* __xpv */
682 	uint_t	probe_op;
683 
684 	/*
685 	 * Initialize as a UMA machine
686 	 */
687 	if (lgrp_topo_ht_limit() == 1) {
688 		lgrp_plat_node_cnt = max_mem_nodes = 1;
689 		return;
690 	}
691 
692 	/*
693 	 * Determine which CPUs and memory are local to each other and number
694 	 * of NUMA nodes by reading ACPI System Resource Affinity Table (SRAT)
695 	 */
696 	lgrp_plat_srat_error = lgrp_plat_process_srat(srat_ptr,
697 	    &lgrp_plat_node_cnt, lgrp_plat_node_domain, lgrp_plat_cpu_node,
698 	    lgrp_plat_node_memory);
699 
700 	/*
701 	 * Try to use PCI config space registers on Opteron if SRAT doesn't
702 	 * exist or there is some error processing the SRAT
703 	 */
704 	if (lgrp_plat_srat_error != 0 && is_opteron())
705 		opt_get_numa_config(&lgrp_plat_node_cnt, &lgrp_plat_mem_intrlv,
706 		    lgrp_plat_node_memory);
707 
708 	/*
709 	 * Don't bother to setup system for multiple lgroups and only use one
710 	 * memory node when memory is interleaved between any nodes or there is
711 	 * only one NUMA node
712 	 *
713 	 * NOTE: May need to change this for Dynamic Reconfiguration (DR)
714 	 *	 when and if it happens for x86/x64
715 	 */
716 	if (lgrp_plat_mem_intrlv || lgrp_plat_node_cnt == 1) {
717 		lgrp_plat_node_cnt = max_mem_nodes = 1;
718 		(void) lgrp_topo_ht_limit_set(1);
719 		return;
720 	}
721 
722 	/*
723 	 * Leaf lgroups on x86/x64 architectures contain one physical
724 	 * processor chip. Tune lgrp_expand_proc_thresh and
725 	 * lgrp_expand_proc_diff so that lgrp_choose() will spread
726 	 * things out aggressively.
727 	 */
728 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
729 	lgrp_expand_proc_diff = 0;
730 
731 	/*
732 	 * There should be one memnode (physical page free list(s)) for
733 	 * each node
734 	 */
735 	max_mem_nodes = lgrp_plat_node_cnt;
736 
737 	/*
738 	 * Determine how far each NUMA node is from each other by
739 	 * reading ACPI System Locality Information Table (SLIT) if it
740 	 * exists
741 	 */
742 	lgrp_plat_slit_error = lgrp_plat_process_slit(slit_ptr,
743 	    lgrp_plat_node_cnt, lgrp_plat_node_memory,
744 	    &lgrp_plat_lat_stats);
745 	if (lgrp_plat_slit_error == 0)
746 		return;
747 
748 	/*
749 	 * Probe to determine latency between NUMA nodes when SLIT
750 	 * doesn't exist or make sense
751 	 */
752 	lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_ENABLE;
753 
754 	/*
755 	 * Specify whether to probe using vendor ID register or page copy
756 	 * if hasn't been specified already or is overspecified
757 	 */
758 	probe_op = lgrp_plat_probe_flags &
759 	    (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
760 
761 	if (probe_op == 0 ||
762 	    probe_op == (LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR)) {
763 		lgrp_plat_probe_flags &=
764 		    ~(LGRP_PLAT_PROBE_PGCPY|LGRP_PLAT_PROBE_VENDOR);
765 		if (is_opteron())
766 			lgrp_plat_probe_flags |=
767 			    LGRP_PLAT_PROBE_VENDOR;
768 		else
769 			lgrp_plat_probe_flags |= LGRP_PLAT_PROBE_PGCPY;
770 	}
771 
772 	/*
773 	 * Probing errors can mess up the lgroup topology and
774 	 * force us fall back to a 2 level lgroup topology.
775 	 * Here we bound how tall the lgroup topology can grow
776 	 * in hopes of avoiding any anamolies in probing from
777 	 * messing up the lgroup topology by limiting the
778 	 * accuracy of the latency topology.
779 	 *
780 	 * Assume that nodes will at least be configured in a
781 	 * ring, so limit height of lgroup topology to be less
782 	 * than number of nodes on a system with 4 or more
783 	 * nodes
784 	 */
785 	if (lgrp_plat_node_cnt >= 4 && lgrp_topo_ht_limit() ==
786 	    lgrp_topo_ht_limit_default())
787 		(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
788 #endif	/* __xpv */
789 }
790 
791 
792 /*
793  * Return latency between "from" and "to" lgroups
794  *
795  * This latency number can only be used for relative comparison
796  * between lgroups on the running system, cannot be used across platforms,
797  * and may not reflect the actual latency.  It is platform and implementation
798  * specific, so platform gets to decide its value.  It would be nice if the
799  * number was at least proportional to make comparisons more meaningful though.
800  */
801 /* ARGSUSED */
802 int
803 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
804 {
805 	lgrp_handle_t	src, dest;
806 	int		node;
807 
808 	if (max_mem_nodes == 1)
809 		return (0);
810 
811 	/*
812 	 * Return max latency for root lgroup
813 	 */
814 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
815 		return (lgrp_plat_lat_stats.latency_max);
816 
817 	src = from;
818 	dest = to;
819 
820 	/*
821 	 * Return 0 for nodes (lgroup platform handles) out of range
822 	 */
823 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
824 		return (0);
825 
826 	/*
827 	 * Probe from current CPU if its lgroup latencies haven't been set yet
828 	 * and we are trying to get latency from current CPU to some node
829 	 */
830 	node = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
831 	ASSERT(node >= 0 && node < lgrp_plat_node_cnt);
832 	if (lgrp_plat_lat_stats.latencies[src][src] == 0 && node == src)
833 		lgrp_plat_probe();
834 
835 	return (lgrp_plat_lat_stats.latencies[src][dest]);
836 }
837 
838 
839 /*
840  * Platform-specific initialization
841  */
842 void
843 lgrp_plat_main_init(void)
844 {
845 	int	curnode;
846 	int	ht_limit;
847 	int	i;
848 
849 	/*
850 	 * Print a notice that MPO is disabled when memory is interleaved
851 	 * across nodes....Would do this when it is discovered, but can't
852 	 * because it happens way too early during boot....
853 	 */
854 	if (lgrp_plat_mem_intrlv)
855 		cmn_err(CE_NOTE,
856 		    "MPO disabled because memory is interleaved\n");
857 
858 	/*
859 	 * Don't bother to do any probing if it is disabled, there is only one
860 	 * node, or the height of the lgroup topology less than or equal to 2
861 	 */
862 	ht_limit = lgrp_topo_ht_limit();
863 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
864 	    max_mem_nodes == 1 || ht_limit <= 2) {
865 		/*
866 		 * Setup lgroup latencies for 2 level lgroup topology
867 		 * (ie. local and remote only) if they haven't been set yet
868 		 */
869 		if (ht_limit == 2 && lgrp_plat_lat_stats.latency_min == -1 &&
870 		    lgrp_plat_lat_stats.latency_max == 0)
871 			lgrp_plat_2level_setup(lgrp_plat_node_memory,
872 			    &lgrp_plat_lat_stats);
873 		return;
874 	}
875 
876 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
877 		/*
878 		 * Should have been able to probe from CPU 0 when it was added
879 		 * to lgroup hierarchy, but may not have been able to then
880 		 * because it happens so early in boot that gethrtime() hasn't
881 		 * been initialized.  (:-(
882 		 */
883 		curnode = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
884 		ASSERT(curnode >= 0 && curnode < lgrp_plat_node_cnt);
885 		if (lgrp_plat_lat_stats.latencies[curnode][curnode] == 0)
886 			lgrp_plat_probe();
887 
888 		return;
889 	}
890 
891 	/*
892 	 * When probing memory, use one page for every sample to determine
893 	 * lgroup topology and taking multiple samples
894 	 */
895 	if (lgrp_plat_probe_mem_config.probe_memsize == 0)
896 		lgrp_plat_probe_mem_config.probe_memsize = PAGESIZE *
897 		    lgrp_plat_probe_nsamples;
898 
899 	/*
900 	 * Map memory in each node needed for probing to determine latency
901 	 * topology
902 	 */
903 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
904 		int	mnode;
905 
906 		/*
907 		 * Skip this node and leave its probe page NULL
908 		 * if it doesn't have any memory
909 		 */
910 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
911 		if (!mem_node_config[mnode].exists) {
912 			lgrp_plat_probe_mem_config.probe_va[i] = NULL;
913 			continue;
914 		}
915 
916 		/*
917 		 * Allocate one kernel virtual page
918 		 */
919 		lgrp_plat_probe_mem_config.probe_va[i] = vmem_alloc(heap_arena,
920 		    lgrp_plat_probe_mem_config.probe_memsize, VM_NOSLEEP);
921 		if (lgrp_plat_probe_mem_config.probe_va[i] == NULL) {
922 			cmn_err(CE_WARN,
923 			    "lgrp_plat_main_init: couldn't allocate memory");
924 			return;
925 		}
926 
927 		/*
928 		 * Get PFN for first page in each node
929 		 */
930 		lgrp_plat_probe_mem_config.probe_pfn[i] =
931 		    mem_node_config[mnode].physbase;
932 
933 		/*
934 		 * Map virtual page to first page in node
935 		 */
936 		hat_devload(kas.a_hat, lgrp_plat_probe_mem_config.probe_va[i],
937 		    lgrp_plat_probe_mem_config.probe_memsize,
938 		    lgrp_plat_probe_mem_config.probe_pfn[i],
939 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
940 		    HAT_LOAD_NOCONSIST);
941 	}
942 
943 	/*
944 	 * Probe from current CPU
945 	 */
946 	lgrp_plat_probe();
947 }
948 
949 
950 /*
951  * Return the maximum number of lgrps supported by the platform.
952  * Before lgrp topology is known it returns an estimate based on the number of
953  * nodes. Once topology is known it returns the actual maximim number of lgrps
954  * created. Since x86/x64 doesn't support Dynamic Reconfiguration (DR) and
955  * dynamic addition of new nodes, this number may not grow during system
956  * lifetime (yet).
957  */
958 int
959 lgrp_plat_max_lgrps(void)
960 {
961 	return (lgrp_topo_initialized ?
962 	    lgrp_alloc_max + 1 :
963 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
964 }
965 
966 
967 /*
968  * Return the number of free pages in an lgroup.
969  *
970  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
971  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
972  * number of allocatable base pagesize pages corresponding to the
973  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
974  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
975  * memory installed, regardless of whether or not it's usable.
976  */
977 pgcnt_t
978 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
979 {
980 	int	mnode;
981 	pgcnt_t npgs = (pgcnt_t)0;
982 	extern struct memlist *phys_avail;
983 	extern struct memlist *phys_install;
984 
985 
986 	if (plathand == LGRP_DEFAULT_HANDLE)
987 		return (lgrp_plat_mem_size_default(plathand, query));
988 
989 	if (plathand != LGRP_NULL_HANDLE) {
990 		mnode = plat_lgrphand_to_mem_node(plathand);
991 		if (mnode >= 0 && mem_node_config[mnode].exists) {
992 			switch (query) {
993 			case LGRP_MEM_SIZE_FREE:
994 				npgs = MNODE_PGCNT(mnode);
995 				break;
996 			case LGRP_MEM_SIZE_AVAIL:
997 				npgs = mem_node_memlist_pages(mnode,
998 				    phys_avail);
999 				break;
1000 			case LGRP_MEM_SIZE_INSTALL:
1001 				npgs = mem_node_memlist_pages(mnode,
1002 				    phys_install);
1003 				break;
1004 			default:
1005 				break;
1006 			}
1007 		}
1008 	}
1009 	return (npgs);
1010 }
1011 
1012 
1013 /*
1014  * Return the platform handle of the lgroup that contains the physical memory
1015  * corresponding to the given page frame number
1016  */
1017 /* ARGSUSED */
1018 lgrp_handle_t
1019 lgrp_plat_pfn_to_hand(pfn_t pfn)
1020 {
1021 	int	mnode;
1022 
1023 	if (max_mem_nodes == 1)
1024 		return (LGRP_DEFAULT_HANDLE);
1025 
1026 	if (pfn > physmax)
1027 		return (LGRP_NULL_HANDLE);
1028 
1029 	mnode = plat_pfn_to_mem_node(pfn);
1030 	if (mnode < 0)
1031 		return (LGRP_NULL_HANDLE);
1032 
1033 	return (MEM_NODE_2_LGRPHAND(mnode));
1034 }
1035 
1036 
1037 /*
1038  * Probe memory in each node from current CPU to determine latency topology
1039  *
1040  * The probing code will probe the vendor ID register on the Northbridge of
1041  * Opteron processors and probe memory for other processors by default.
1042  *
1043  * Since probing is inherently error prone, the code takes laps across all the
1044  * nodes probing from each node to each of the other nodes some number of
1045  * times.  Furthermore, each node is probed some number of times before moving
1046  * onto the next one during each lap.  The minimum latency gotten between nodes
1047  * is kept as the latency between the nodes.
1048  *
1049  * After all that,  the probe times are adjusted by normalizing values that are
1050  * close to each other and local latencies are made the same.  Lastly, the
1051  * latencies are verified to make sure that certain conditions are met (eg.
1052  * local < remote, latency(a, b) == latency(b, a), etc.).
1053  *
1054  * If any of the conditions aren't met, the code will export a NUMA
1055  * configuration with the local CPUs and memory given by the SRAT or PCI config
1056  * space registers and one remote memory latency since it can't tell exactly
1057  * how far each node is from each other.
1058  */
1059 void
1060 lgrp_plat_probe(void)
1061 {
1062 	int				from;
1063 	int				i;
1064 	lgrp_plat_latency_stats_t	*lat_stats;
1065 	hrtime_t			probe_time;
1066 	int				to;
1067 
1068 	if (!(lgrp_plat_probe_flags & LGRP_PLAT_PROBE_ENABLE) ||
1069 	    max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1070 		return;
1071 
1072 	/*
1073 	 * Determine ID of node containing current CPU
1074 	 */
1075 	from = lgrp_plat_cpu_to_node(CPU, lgrp_plat_cpu_node);
1076 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
1077 	if (srat_ptr && lgrp_plat_srat_enable && !lgrp_plat_srat_error)
1078 		ASSERT(lgrp_plat_node_domain[from].exists);
1079 
1080 	/*
1081 	 * Don't need to probe if got times already
1082 	 */
1083 	lat_stats = &lgrp_plat_lat_stats;
1084 	if (lat_stats->latencies[from][from] != 0)
1085 		return;
1086 
1087 	/*
1088 	 * Read vendor ID in Northbridge or read and write page(s)
1089 	 * in each node from current CPU and remember how long it takes,
1090 	 * so we can build latency topology of machine later.
1091 	 * This should approximate the memory latency between each node.
1092 	 */
1093 	for (i = 0; i < lgrp_plat_probe_nrounds; i++) {
1094 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1095 			/*
1096 			 * Get probe time and bail out if can't get it yet
1097 			 */
1098 			probe_time = lgrp_plat_probe_time(to,
1099 			    lgrp_plat_cpu_node, &lgrp_plat_probe_mem_config,
1100 			    &lgrp_plat_lat_stats, &lgrp_plat_probe_stats);
1101 			if (probe_time == 0)
1102 				return;
1103 
1104 			/*
1105 			 * Keep lowest probe time as latency between nodes
1106 			 */
1107 			if (lat_stats->latencies[from][to] == 0 ||
1108 			    probe_time < lat_stats->latencies[from][to])
1109 				lat_stats->latencies[from][to] = probe_time;
1110 
1111 			/*
1112 			 * Update overall minimum and maximum probe times
1113 			 * across all nodes
1114 			 */
1115 			if (probe_time < lat_stats->latency_min ||
1116 			    lat_stats->latency_min == -1)
1117 				lat_stats->latency_min = probe_time;
1118 			if (probe_time > lat_stats->latency_max)
1119 				lat_stats->latency_max = probe_time;
1120 		}
1121 	}
1122 
1123 	/*
1124 	 * - Fix up latencies such that local latencies are same,
1125 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1126 	 *
1127 	 * - Verify that latencies look ok
1128 	 *
1129 	 * - Fallback to just optimizing for local and remote if
1130 	 *   latencies didn't look right
1131 	 */
1132 	lgrp_plat_latency_adjust(lgrp_plat_node_memory, &lgrp_plat_lat_stats,
1133 	    &lgrp_plat_probe_stats);
1134 	lgrp_plat_probe_stats.probe_error_code =
1135 	    lgrp_plat_latency_verify(lgrp_plat_node_memory,
1136 	    &lgrp_plat_lat_stats);
1137 	if (lgrp_plat_probe_stats.probe_error_code)
1138 		lgrp_plat_2level_setup(lgrp_plat_node_memory,
1139 		    &lgrp_plat_lat_stats);
1140 }
1141 
1142 
1143 /*
1144  * Return platform handle for root lgroup
1145  */
1146 lgrp_handle_t
1147 lgrp_plat_root_hand(void)
1148 {
1149 	return (LGRP_DEFAULT_HANDLE);
1150 }
1151 
1152 
1153 /*
1154  * INTERNAL ROUTINES
1155  */
1156 
1157 
1158 /*
1159  * Update CPU to node mapping for given CPU and proximity domain (and returns
1160  * negative numbers for errors and positive ones for success)
1161  */
1162 static int
1163 lgrp_plat_cpu_node_update(node_domain_map_t *node_domain,
1164     cpu_node_map_t *cpu_node, uint32_t apicid, uint32_t domain)
1165 {
1166 	uint_t	i;
1167 	uint_t	start;
1168 	int	node;
1169 
1170 	/*
1171 	 * Get node number for proximity domain
1172 	 */
1173 	node = lgrp_plat_domain_to_node(node_domain, domain);
1174 	if (node == -1) {
1175 		node = lgrp_plat_node_domain_update(node_domain, domain);
1176 		if (node == -1)
1177 			return (-1);
1178 	}
1179 
1180 	/*
1181 	 * Hash given CPU APIC ID into CPU to node mapping table/array and
1182 	 * enter it and its corresponding node and proximity domain IDs into
1183 	 * first non-existent or matching entry
1184 	 */
1185 	i = start = CPU_NODE_HASH(apicid);
1186 	do {
1187 		if (cpu_node[i].exists) {
1188 			/*
1189 			 * Update already existing entry for CPU
1190 			 */
1191 			if (cpu_node[i].apicid == apicid) {
1192 				/*
1193 				 * Just return when everything same
1194 				 */
1195 				if (cpu_node[i].prox_domain == domain &&
1196 				    cpu_node[i].node == node)
1197 					return (1);
1198 
1199 				/*
1200 				 * Assert that proximity domain and node IDs
1201 				 * should be same and return error on non-debug
1202 				 * kernel
1203 				 */
1204 				ASSERT(cpu_node[i].prox_domain == domain &&
1205 				    cpu_node[i].node == node);
1206 				return (-1);
1207 			}
1208 		} else {
1209 			/*
1210 			 * Create new entry for CPU
1211 			 */
1212 			cpu_node[i].exists = 1;
1213 			cpu_node[i].apicid = apicid;
1214 			cpu_node[i].prox_domain = domain;
1215 			cpu_node[i].node = node;
1216 			return (0);
1217 		}
1218 		i = CPU_NODE_HASH(i + 1);
1219 	} while (i != start);
1220 
1221 	/*
1222 	 * Ran out of supported number of entries which shouldn't happen....
1223 	 */
1224 	ASSERT(i != start);
1225 	return (-1);
1226 }
1227 
1228 
1229 /*
1230  * Get node ID for given CPU ID
1231  */
1232 static int
1233 lgrp_plat_cpu_to_node(cpu_t *cp, cpu_node_map_t *cpu_node)
1234 {
1235 	uint32_t	apicid;
1236 	uint_t		i;
1237 	uint_t		start;
1238 
1239 	if (cp == NULL)
1240 		return (-1);
1241 
1242 	/*
1243 	 * SRAT doesn't exist, isn't enabled, or there was an error processing
1244 	 * it, so return chip ID for Opteron and -1 otherwise.
1245 	 */
1246 	if (srat_ptr == NULL || !lgrp_plat_srat_enable ||
1247 	    lgrp_plat_srat_error) {
1248 		if (is_opteron())
1249 			return (pg_plat_hw_instance_id(cp, PGHW_CHIP));
1250 		return (-1);
1251 	}
1252 
1253 	/*
1254 	 * SRAT does exist, so get APIC ID for given CPU and map that to its
1255 	 * node ID
1256 	 */
1257 	apicid = cpuid_get_apicid(cp);
1258 	i = start = CPU_NODE_HASH(apicid);
1259 	do {
1260 		if (cpu_node[i].apicid == apicid && cpu_node[i].exists)
1261 			return (cpu_node[i].node);
1262 		i = CPU_NODE_HASH(i + 1);
1263 	} while (i != start);
1264 	return (-1);
1265 }
1266 
1267 
1268 /*
1269  * Return node number for given proximity domain/system locality
1270  */
1271 static int
1272 lgrp_plat_domain_to_node(node_domain_map_t *node_domain, uint32_t domain)
1273 {
1274 	uint_t	node;
1275 	uint_t	start;
1276 
1277 	/*
1278 	 * Hash proximity domain ID into node to domain mapping table (array),
1279 	 * search for entry with matching proximity domain ID, and return index
1280 	 * of matching entry as node ID.
1281 	 */
1282 	node = start = NODE_DOMAIN_HASH(domain);
1283 	do {
1284 		if (node_domain[node].prox_domain == domain &&
1285 		    node_domain[node].exists)
1286 			return (node);
1287 		node = NODE_DOMAIN_HASH(node + 1);
1288 	} while (node != start);
1289 	return (-1);
1290 }
1291 
1292 
1293 /*
1294  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
1295  * be considered same
1296  */
1297 #define	LGRP_LAT_TOLERANCE_SHIFT	4
1298 
1299 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
1300 
1301 
1302 /*
1303  * Adjust latencies between nodes to be symmetric, normalize latencies between
1304  * any nodes that are within some tolerance to be same, and make local
1305  * latencies be same
1306  */
1307 static void
1308 lgrp_plat_latency_adjust(node_phys_addr_map_t *node_memory,
1309     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1310 {
1311 	int				i;
1312 	int				j;
1313 	int				k;
1314 	int				l;
1315 	u_longlong_t			max;
1316 	u_longlong_t			min;
1317 	u_longlong_t			t;
1318 	u_longlong_t			t1;
1319 	u_longlong_t			t2;
1320 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
1321 	int				lat_corrected[MAX_NODES][MAX_NODES];
1322 
1323 	/*
1324 	 * Nothing to do when this is an UMA machine or don't have args needed
1325 	 */
1326 	if (max_mem_nodes == 1)
1327 		return;
1328 
1329 	ASSERT(node_memory != NULL && lat_stats != NULL &&
1330 	    probe_stats != NULL);
1331 
1332 	/*
1333 	 * Make sure that latencies are symmetric between any two nodes
1334 	 * (ie. latency(node0, node1) == latency(node1, node0))
1335 	 */
1336 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1337 		if (!node_memory[i].exists)
1338 			continue;
1339 
1340 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1341 			if (!node_memory[j].exists)
1342 				continue;
1343 
1344 			t1 = lat_stats->latencies[i][j];
1345 			t2 = lat_stats->latencies[j][i];
1346 
1347 			if (t1 == 0 || t2 == 0 || t1 == t2)
1348 				continue;
1349 
1350 			/*
1351 			 * Latencies should be same
1352 			 * - Use minimum of two latencies which should be same
1353 			 * - Track suspect probe times not within tolerance of
1354 			 *   min value
1355 			 * - Remember how much values are corrected by
1356 			 */
1357 			if (t1 > t2) {
1358 				t = t2;
1359 				probe_stats->probe_errors[i][j] += t1 - t2;
1360 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
1361 					probe_stats->probe_suspect[i][j]++;
1362 					probe_stats->probe_suspect[j][i]++;
1363 				}
1364 			} else if (t2 > t1) {
1365 				t = t1;
1366 				probe_stats->probe_errors[j][i] += t2 - t1;
1367 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
1368 					probe_stats->probe_suspect[i][j]++;
1369 					probe_stats->probe_suspect[j][i]++;
1370 				}
1371 			}
1372 
1373 			lat_stats->latencies[i][j] =
1374 			    lat_stats->latencies[j][i] = t;
1375 			lgrp_config(cflag, t1, t);
1376 			lgrp_config(cflag, t2, t);
1377 		}
1378 	}
1379 
1380 	/*
1381 	 * Keep track of which latencies get corrected
1382 	 */
1383 	for (i = 0; i < MAX_NODES; i++)
1384 		for (j = 0; j < MAX_NODES; j++)
1385 			lat_corrected[i][j] = 0;
1386 
1387 	/*
1388 	 * For every two nodes, see whether there is another pair of nodes which
1389 	 * are about the same distance apart and make the latencies be the same
1390 	 * if they are close enough together
1391 	 */
1392 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1393 		if (!node_memory[i].exists)
1394 			continue;
1395 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1396 			if (!node_memory[j].exists)
1397 				continue;
1398 			/*
1399 			 * Pick one pair of nodes (i, j)
1400 			 * and get latency between them
1401 			 */
1402 			t1 = lat_stats->latencies[i][j];
1403 
1404 			/*
1405 			 * Skip this pair of nodes if there isn't a latency
1406 			 * for it yet
1407 			 */
1408 			if (t1 == 0)
1409 				continue;
1410 
1411 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1412 				if (!node_memory[k].exists)
1413 					continue;
1414 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
1415 					if (!node_memory[l].exists)
1416 						continue;
1417 					/*
1418 					 * Pick another pair of nodes (k, l)
1419 					 * not same as (i, j) and get latency
1420 					 * between them
1421 					 */
1422 					if (k == i && l == j)
1423 						continue;
1424 
1425 					t2 = lat_stats->latencies[k][l];
1426 
1427 					/*
1428 					 * Skip this pair of nodes if there
1429 					 * isn't a latency for it yet
1430 					 */
1431 
1432 					if (t2 == 0)
1433 						continue;
1434 
1435 					/*
1436 					 * Skip nodes (k, l) if they already
1437 					 * have same latency as (i, j) or
1438 					 * their latency isn't close enough to
1439 					 * be considered/made the same
1440 					 */
1441 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
1442 					    t1 >> lgrp_plat_probe_lt_shift) ||
1443 					    (t2 > t1 && t2 - t1 >
1444 					    t2 >> lgrp_plat_probe_lt_shift))
1445 						continue;
1446 
1447 					/*
1448 					 * Make latency(i, j) same as
1449 					 * latency(k, l), try to use latency
1450 					 * that has been adjusted already to get
1451 					 * more consistency (if possible), and
1452 					 * remember which latencies were
1453 					 * adjusted for next time
1454 					 */
1455 					if (lat_corrected[i][j]) {
1456 						t = t1;
1457 						lgrp_config(cflag, t2, t);
1458 						t2 = t;
1459 					} else if (lat_corrected[k][l]) {
1460 						t = t2;
1461 						lgrp_config(cflag, t1, t);
1462 						t1 = t;
1463 					} else {
1464 						if (t1 > t2)
1465 							t = t2;
1466 						else
1467 							t = t1;
1468 						lgrp_config(cflag, t1, t);
1469 						lgrp_config(cflag, t2, t);
1470 						t1 = t2 = t;
1471 					}
1472 
1473 					lat_stats->latencies[i][j] =
1474 					    lat_stats->latencies[k][l] = t;
1475 
1476 					lat_corrected[i][j] =
1477 					    lat_corrected[k][l] = 1;
1478 				}
1479 			}
1480 		}
1481 	}
1482 
1483 	/*
1484 	 * Local latencies should be same
1485 	 * - Find min and max local latencies
1486 	 * - Make all local latencies be minimum
1487 	 */
1488 	min = -1;
1489 	max = 0;
1490 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1491 		if (!node_memory[i].exists)
1492 			continue;
1493 		t = lat_stats->latencies[i][i];
1494 		if (t == 0)
1495 			continue;
1496 		if (min == -1 || t < min)
1497 			min = t;
1498 		if (t > max)
1499 			max = t;
1500 	}
1501 	if (min != max) {
1502 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1503 			int	local;
1504 
1505 			if (!node_memory[i].exists)
1506 				continue;
1507 
1508 			local = lat_stats->latencies[i][i];
1509 			if (local == 0)
1510 				continue;
1511 
1512 			/*
1513 			 * Track suspect probe times that aren't within
1514 			 * tolerance of minimum local latency and how much
1515 			 * probe times are corrected by
1516 			 */
1517 			if (local - min > min >> lgrp_plat_probe_lt_shift)
1518 				probe_stats->probe_suspect[i][i]++;
1519 
1520 			probe_stats->probe_errors[i][i] += local - min;
1521 
1522 			/*
1523 			 * Make local latencies be minimum
1524 			 */
1525 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
1526 			lat_stats->latencies[i][i] = min;
1527 		}
1528 	}
1529 
1530 	/*
1531 	 * Determine max probe time again since just adjusted latencies
1532 	 */
1533 	lat_stats->latency_max = 0;
1534 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1535 		if (!node_memory[i].exists)
1536 			continue;
1537 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1538 			if (!node_memory[j].exists)
1539 				continue;
1540 			t = lat_stats->latencies[i][j];
1541 			if (t > lat_stats->latency_max)
1542 				lat_stats->latency_max = t;
1543 		}
1544 	}
1545 }
1546 
1547 
1548 /*
1549  * Verify following about latencies between nodes:
1550  *
1551  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
1552  * - Local latencies same
1553  * - Local < remote
1554  * - Number of latencies seen is reasonable
1555  * - Number of occurrences of a given latency should be more than 1
1556  *
1557  * Returns:
1558  *	0	Success
1559  *	-1	Not symmetric
1560  *	-2	Local latencies not same
1561  *	-3	Local >= remote
1562  */
1563 static int
1564 lgrp_plat_latency_verify(node_phys_addr_map_t *node_memory,
1565     lgrp_plat_latency_stats_t *lat_stats)
1566 {
1567 	int				i;
1568 	int				j;
1569 	u_longlong_t			t1;
1570 	u_longlong_t			t2;
1571 
1572 	ASSERT(node_memory != NULL && lat_stats != NULL);
1573 
1574 	/*
1575 	 * Nothing to do when this is an UMA machine, lgroup topology is
1576 	 * limited to 2 levels, or there aren't any probe times yet
1577 	 */
1578 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
1579 	    lat_stats->latencies[0][0] == 0)
1580 		return (0);
1581 
1582 	/*
1583 	 * Make sure that latencies are symmetric between any two nodes
1584 	 * (ie. latency(node0, node1) == latency(node1, node0))
1585 	 */
1586 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1587 		if (!node_memory[i].exists)
1588 			continue;
1589 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1590 			if (!node_memory[j].exists)
1591 				continue;
1592 			t1 = lat_stats->latencies[i][j];
1593 			t2 = lat_stats->latencies[j][i];
1594 
1595 			if (t1 == 0 || t2 == 0 || t1 == t2)
1596 				continue;
1597 
1598 			return (-1);
1599 		}
1600 	}
1601 
1602 	/*
1603 	 * Local latencies should be same
1604 	 */
1605 	t1 = lat_stats->latencies[0][0];
1606 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
1607 		if (!node_memory[i].exists)
1608 			continue;
1609 
1610 		t2 = lat_stats->latencies[i][i];
1611 		if (t2 == 0)
1612 			continue;
1613 
1614 		if (t1 == 0) {
1615 			t1 = t2;
1616 			continue;
1617 		}
1618 
1619 		if (t1 != t2)
1620 			return (-2);
1621 	}
1622 
1623 	/*
1624 	 * Local latencies should be less than remote
1625 	 */
1626 	if (t1) {
1627 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
1628 			if (!node_memory[i].exists)
1629 				continue;
1630 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1631 				if (!node_memory[j].exists)
1632 					continue;
1633 				t2 = lat_stats->latencies[i][j];
1634 				if (i == j || t2 == 0)
1635 					continue;
1636 
1637 				if (t1 >= t2)
1638 					return (-3);
1639 			}
1640 		}
1641 	}
1642 
1643 	return (0);
1644 }
1645 
1646 
1647 /*
1648  * Return the number of free, allocatable, or installed
1649  * pages in an lgroup
1650  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1651  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1652  */
1653 /* ARGSUSED */
1654 static pgcnt_t
1655 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1656 {
1657 	struct memlist *mlist;
1658 	pgcnt_t npgs = 0;
1659 	extern struct memlist *phys_avail;
1660 	extern struct memlist *phys_install;
1661 
1662 	switch (query) {
1663 	case LGRP_MEM_SIZE_FREE:
1664 		return ((pgcnt_t)freemem);
1665 	case LGRP_MEM_SIZE_AVAIL:
1666 		memlist_read_lock();
1667 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1668 			npgs += btop(mlist->size);
1669 		memlist_read_unlock();
1670 		return (npgs);
1671 	case LGRP_MEM_SIZE_INSTALL:
1672 		memlist_read_lock();
1673 		for (mlist = phys_install; mlist; mlist = mlist->next)
1674 			npgs += btop(mlist->size);
1675 		memlist_read_unlock();
1676 		return (npgs);
1677 	default:
1678 		return ((pgcnt_t)0);
1679 	}
1680 }
1681 
1682 
1683 /*
1684  * Update node to proximity domain mappings for given domain and return node ID
1685  */
1686 static int
1687 lgrp_plat_node_domain_update(node_domain_map_t *node_domain, uint32_t domain)
1688 {
1689 	uint_t	node;
1690 	uint_t	start;
1691 
1692 	/*
1693 	 * Hash proximity domain ID into node to domain mapping table (array)
1694 	 * and add entry for it into first non-existent or matching entry found
1695 	 */
1696 	node = start = NODE_DOMAIN_HASH(domain);
1697 	do {
1698 		/*
1699 		 * Entry doesn't exist yet, so create one for this proximity
1700 		 * domain and return node ID which is index into mapping table.
1701 		 */
1702 		if (!node_domain[node].exists) {
1703 			node_domain[node].exists = 1;
1704 			node_domain[node].prox_domain = domain;
1705 			return (node);
1706 		}
1707 
1708 		/*
1709 		 * Entry exists for this proximity domain already, so just
1710 		 * return node ID (index into table).
1711 		 */
1712 		if (node_domain[node].prox_domain == domain)
1713 			return (node);
1714 		node = NODE_DOMAIN_HASH(node + 1);
1715 	} while (node != start);
1716 
1717 	/*
1718 	 * Ran out of supported number of entries which shouldn't happen....
1719 	 */
1720 	ASSERT(node != start);
1721 	return (-1);
1722 }
1723 
1724 
1725 /*
1726  * Update node memory information for given proximity domain with specified
1727  * starting and ending physical address range (and return positive numbers for
1728  * success and negative ones for errors)
1729  */
1730 static int
1731 lgrp_plat_node_memory_update(node_domain_map_t *node_domain,
1732     node_phys_addr_map_t *node_memory, uint64_t start, uint64_t end,
1733     uint32_t domain)
1734 {
1735 	int	node;
1736 
1737 	/*
1738 	 * Get node number for proximity domain
1739 	 */
1740 	node = lgrp_plat_domain_to_node(node_domain, domain);
1741 	if (node == -1) {
1742 		node = lgrp_plat_node_domain_update(node_domain, domain);
1743 		if (node == -1)
1744 			return (-1);
1745 	}
1746 
1747 	/*
1748 	 * Create entry in table for node if it doesn't exist
1749 	 */
1750 	if (!node_memory[node].exists) {
1751 		node_memory[node].exists = 1;
1752 		node_memory[node].start = btop(start);
1753 		node_memory[node].end = btop(end);
1754 		node_memory[node].prox_domain = domain;
1755 		return (0);
1756 	}
1757 
1758 	/*
1759 	 * Entry already exists for this proximity domain
1760 	 *
1761 	 * There may be more than one SRAT memory entry for a domain, so we may
1762 	 * need to update existing start or end address for the node.
1763 	 */
1764 	if (node_memory[node].prox_domain == domain) {
1765 		if (btop(start) < node_memory[node].start)
1766 			node_memory[node].start = btop(start);
1767 		if (btop(end) > node_memory[node].end)
1768 			node_memory[node].end = btop(end);
1769 		return (1);
1770 	}
1771 	return (-2);
1772 }
1773 
1774 
1775 /*
1776  * Return time needed to probe from current CPU to memory in given node
1777  */
1778 static hrtime_t
1779 lgrp_plat_probe_time(int to, cpu_node_map_t *cpu_node,
1780     lgrp_plat_probe_mem_config_t *probe_mem_config,
1781     lgrp_plat_latency_stats_t *lat_stats, lgrp_plat_probe_stats_t *probe_stats)
1782 {
1783 	caddr_t			buf;
1784 	hrtime_t		elapsed;
1785 	hrtime_t		end;
1786 	int			from;
1787 	int			i;
1788 	int			ipl;
1789 	hrtime_t		max;
1790 	hrtime_t		min;
1791 	hrtime_t		start;
1792 	extern int		use_sse_pagecopy;
1793 
1794 	/*
1795 	 * Determine ID of node containing current CPU
1796 	 */
1797 	from = lgrp_plat_cpu_to_node(CPU, cpu_node);
1798 	ASSERT(from >= 0 && from < lgrp_plat_node_cnt);
1799 
1800 	/*
1801 	 * Do common work for probing main memory
1802 	 */
1803 	if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_PGCPY) {
1804 		/*
1805 		 * Skip probing any nodes without memory and
1806 		 * set probe time to 0
1807 		 */
1808 		if (probe_mem_config->probe_va[to] == NULL) {
1809 			lat_stats->latencies[from][to] = 0;
1810 			return (0);
1811 		}
1812 
1813 		/*
1814 		 * Invalidate caches once instead of once every sample
1815 		 * which should cut cost of probing by a lot
1816 		 */
1817 		probe_stats->flush_cost = gethrtime();
1818 		invalidate_cache();
1819 		probe_stats->flush_cost = gethrtime() -
1820 		    probe_stats->flush_cost;
1821 		probe_stats->probe_cost_total += probe_stats->flush_cost;
1822 	}
1823 
1824 	/*
1825 	 * Probe from current CPU to given memory using specified operation
1826 	 * and take specified number of samples
1827 	 */
1828 	max = 0;
1829 	min = -1;
1830 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1831 		probe_stats->probe_cost = gethrtime();
1832 
1833 		/*
1834 		 * Can't measure probe time if gethrtime() isn't working yet
1835 		 */
1836 		if (probe_stats->probe_cost == 0 && gethrtime() == 0)
1837 			return (0);
1838 
1839 		if (lgrp_plat_probe_flags & LGRP_PLAT_PROBE_VENDOR) {
1840 			/*
1841 			 * Measure how long it takes to read vendor ID from
1842 			 * Northbridge
1843 			 */
1844 			elapsed = opt_probe_vendor(to, lgrp_plat_probe_nreads);
1845 		} else {
1846 			/*
1847 			 * Measure how long it takes to copy page
1848 			 * on top of itself
1849 			 */
1850 			buf = probe_mem_config->probe_va[to] + (i * PAGESIZE);
1851 
1852 			kpreempt_disable();
1853 			ipl = splhigh();
1854 			start = gethrtime();
1855 			if (use_sse_pagecopy)
1856 				hwblkpagecopy(buf, buf);
1857 			else
1858 				bcopy(buf, buf, PAGESIZE);
1859 			end = gethrtime();
1860 			elapsed = end - start;
1861 			splx(ipl);
1862 			kpreempt_enable();
1863 		}
1864 
1865 		probe_stats->probe_cost = gethrtime() -
1866 		    probe_stats->probe_cost;
1867 		probe_stats->probe_cost_total += probe_stats->probe_cost;
1868 
1869 		if (min == -1 || elapsed < min)
1870 			min = elapsed;
1871 		if (elapsed > max)
1872 			max = elapsed;
1873 	}
1874 
1875 	/*
1876 	 * Update minimum and maximum probe times between
1877 	 * these two nodes
1878 	 */
1879 	if (min < probe_stats->probe_min[from][to] ||
1880 	    probe_stats->probe_min[from][to] == 0)
1881 		probe_stats->probe_min[from][to] = min;
1882 
1883 	if (max > probe_stats->probe_max[from][to])
1884 		probe_stats->probe_max[from][to] = max;
1885 
1886 	return (min);
1887 }
1888 
1889 
1890 /*
1891  * Read ACPI System Locality Information Table (SLIT) to determine how far each
1892  * NUMA node is from each other
1893  */
1894 static int
1895 lgrp_plat_process_slit(struct slit *tp, uint_t node_cnt,
1896     node_phys_addr_map_t *node_memory, lgrp_plat_latency_stats_t *lat_stats)
1897 {
1898 	int		i;
1899 	int		j;
1900 	int		localities;
1901 	hrtime_t	max;
1902 	hrtime_t	min;
1903 	int		retval;
1904 	uint8_t		*slit_entries;
1905 
1906 	if (tp == NULL || !lgrp_plat_slit_enable)
1907 		return (1);
1908 
1909 	if (lat_stats == NULL)
1910 		return (2);
1911 
1912 	localities = tp->number;
1913 	if (localities != node_cnt)
1914 		return (3);
1915 
1916 	min = lat_stats->latency_min;
1917 	max = lat_stats->latency_max;
1918 
1919 	/*
1920 	 * Fill in latency matrix based on SLIT entries
1921 	 */
1922 	slit_entries = tp->entry;
1923 	for (i = 0; i < localities; i++) {
1924 		for (j = 0; j < localities; j++) {
1925 			uint8_t	latency;
1926 
1927 			latency = slit_entries[(i * localities) + j];
1928 			lat_stats->latencies[i][j] = latency;
1929 			if (latency < min)
1930 				min = latency;
1931 			if (latency > max)
1932 				max = latency;
1933 		}
1934 	}
1935 
1936 	/*
1937 	 * Verify that latencies/distances given in SLIT look reasonable
1938 	 */
1939 	retval = lgrp_plat_latency_verify(node_memory, lat_stats);
1940 
1941 	if (retval) {
1942 		/*
1943 		 * Reinitialize (zero) latency table since SLIT doesn't look
1944 		 * right
1945 		 */
1946 		for (i = 0; i < localities; i++) {
1947 			for (j = 0; j < localities; j++)
1948 				lat_stats->latencies[i][j] = 0;
1949 		}
1950 	} else {
1951 		/*
1952 		 * Update min and max latencies seen since SLIT looks valid
1953 		 */
1954 		lat_stats->latency_min = min;
1955 		lat_stats->latency_max = max;
1956 	}
1957 
1958 	return (retval);
1959 }
1960 
1961 
1962 /*
1963  * Read ACPI System Resource Affinity Table (SRAT) to determine which CPUs
1964  * and memory are local to each other in the same NUMA node
1965  */
1966 static int
1967 lgrp_plat_process_srat(struct srat *tp, uint_t *node_cnt,
1968     node_domain_map_t *node_domain, cpu_node_map_t *cpu_node,
1969     node_phys_addr_map_t *node_memory)
1970 {
1971 	struct srat_item	*end;
1972 	int			i;
1973 	struct srat_item	*item;
1974 
1975 	if (tp == NULL || !lgrp_plat_srat_enable)
1976 		return (1);
1977 
1978 	/*
1979 	 * Determine number of nodes by counting number of proximity domains in
1980 	 * SRAT
1981 	 */
1982 	if (node_cnt) {
1983 		int	nodes;
1984 
1985 		nodes = lgrp_plat_srat_domains(tp);
1986 		if (nodes < 0) {
1987 			*node_cnt = 1;
1988 			return (2);
1989 		}
1990 		*node_cnt = nodes;
1991 	}
1992 
1993 	/*
1994 	 * Walk through SRAT, examining each CPU and memory entry to determine
1995 	 * which CPUs and memory belong to which node.
1996 	 */
1997 	item = tp->list;
1998 	end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
1999 	while (item < end) {
2000 		uint32_t	apic_id;
2001 		uint32_t	domain;
2002 		uint64_t	end;
2003 		uint64_t	length;
2004 		uint64_t	start;
2005 
2006 		switch (item->type) {
2007 		case SRAT_PROCESSOR:	/* CPU entry */
2008 			if (!(item->i.p.flags & SRAT_ENABLED) ||
2009 			    cpu_node == NULL)
2010 				break;
2011 
2012 			/*
2013 			 * Calculate domain (node) ID and fill in APIC ID to
2014 			 * domain/node mapping table
2015 			 */
2016 			domain = item->i.p.domain1;
2017 			for (i = 0; i < 3; i++) {
2018 				domain += item->i.p.domain2[i] <<
2019 				    ((i + 1) * 8);
2020 			}
2021 			apic_id = item->i.p.apic_id;
2022 
2023 			if (lgrp_plat_cpu_node_update(node_domain, cpu_node,
2024 			    apic_id, domain) < 0)
2025 				return (3);
2026 			break;
2027 
2028 		case SRAT_MEMORY:	/* memory entry */
2029 			if (!(item->i.m.flags & SRAT_ENABLED) ||
2030 			    node_memory == NULL)
2031 				break;
2032 
2033 			/*
2034 			 * Get domain (node) ID and fill in domain/node
2035 			 * to memory mapping table
2036 			 */
2037 			domain = item->i.m.domain;
2038 			start = item->i.m.base_addr;
2039 			length = item->i.m.len;
2040 			end = start + length - 1;
2041 
2042 			if (lgrp_plat_node_memory_update(node_domain,
2043 			    node_memory, start, end, domain) < 0)
2044 				return (4);
2045 			break;
2046 
2047 		default:
2048 			break;
2049 		}
2050 
2051 		item = (struct srat_item *)((uintptr_t)item + item->len);
2052 	}
2053 	return (0);
2054 }
2055 
2056 
2057 /*
2058  * Return number of proximity domains given in ACPI SRAT
2059  */
2060 static int
2061 lgrp_plat_srat_domains(struct srat *tp)
2062 {
2063 	int			domain_cnt;
2064 	struct srat_item	*end;
2065 	int			i;
2066 	struct srat_item	*item;
2067 	node_domain_map_t	node_domain[MAX_NODES];
2068 
2069 
2070 	if (tp == NULL || !lgrp_plat_srat_enable)
2071 		return (1);
2072 
2073 	/*
2074 	 * Walk through SRAT, examining each CPU and memory entry to determine
2075 	 * proximity domain ID for each.
2076 	 */
2077 	domain_cnt = 0;
2078 	item = tp->list;
2079 	end = (struct srat_item *)(tp->hdr.len + (uintptr_t)tp);
2080 	bzero(node_domain, MAX_NODES * sizeof (node_domain_map_t));
2081 	while (item < end) {
2082 		uint32_t	domain;
2083 		boolean_t	overflow;
2084 		uint_t		start;
2085 
2086 		switch (item->type) {
2087 		case SRAT_PROCESSOR:	/* CPU entry */
2088 			if (!(item->i.p.flags & SRAT_ENABLED))
2089 				break;
2090 			domain = item->i.p.domain1;
2091 			for (i = 0; i < 3; i++) {
2092 				domain += item->i.p.domain2[i] <<
2093 				    ((i + 1) * 8);
2094 			}
2095 			break;
2096 
2097 		case SRAT_MEMORY:	/* memory entry */
2098 			if (!(item->i.m.flags & SRAT_ENABLED))
2099 				break;
2100 			domain = item->i.m.domain;
2101 			break;
2102 
2103 		default:
2104 			break;
2105 		}
2106 
2107 		/*
2108 		 * Count and keep track of which proximity domain IDs seen
2109 		 */
2110 		start = i = domain % MAX_NODES;
2111 		overflow = B_TRUE;
2112 		do {
2113 			/*
2114 			 * Create entry for proximity domain and increment
2115 			 * count when no entry exists where proximity domain
2116 			 * hashed
2117 			 */
2118 			if (!node_domain[i].exists) {
2119 				node_domain[i].exists = 1;
2120 				node_domain[i].prox_domain = domain;
2121 				domain_cnt++;
2122 				overflow = B_FALSE;
2123 				break;
2124 			}
2125 
2126 			/*
2127 			 * Nothing to do when proximity domain seen already
2128 			 * and its entry exists
2129 			 */
2130 			if (node_domain[i].prox_domain == domain) {
2131 				overflow = B_FALSE;
2132 				break;
2133 			}
2134 
2135 			/*
2136 			 * Entry exists where proximity domain hashed, but for
2137 			 * different proximity domain so keep search for empty
2138 			 * slot to put it or matching entry whichever comes
2139 			 * first.
2140 			 */
2141 			i = (i + 1) % MAX_NODES;
2142 		} while (i != start);
2143 
2144 		/*
2145 		 * Didn't find empty or matching entry which means have more
2146 		 * proximity domains than supported nodes (:-(
2147 		 */
2148 		ASSERT(overflow != B_TRUE);
2149 		if (overflow == B_TRUE)
2150 			return (-1);
2151 
2152 		item = (struct srat_item *)((uintptr_t)item + item->len);
2153 	}
2154 	return (domain_cnt);
2155 }
2156 
2157 
2158 /*
2159  * Set lgroup latencies for 2 level lgroup topology
2160  */
2161 static void
2162 lgrp_plat_2level_setup(node_phys_addr_map_t *node_memory,
2163     lgrp_plat_latency_stats_t *lat_stats)
2164 {
2165 	int	i;
2166 
2167 	ASSERT(node_memory != NULL && lat_stats != NULL);
2168 
2169 	if (lgrp_plat_node_cnt >= 4)
2170 		cmn_err(CE_NOTE,
2171 		    "MPO only optimizing for local and remote\n");
2172 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
2173 		int	j;
2174 
2175 		if (!node_memory[i].exists)
2176 			continue;
2177 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
2178 			if (!node_memory[j].exists)
2179 				continue;
2180 			if (i == j)
2181 				lat_stats->latencies[i][j] = 2;
2182 			else
2183 				lat_stats->latencies[i][j] = 3;
2184 		}
2185 	}
2186 	lat_stats->latency_min = 2;
2187 	lat_stats->latency_max = 3;
2188 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
2189 }
2190 
2191 
2192 /*
2193  * The following Opteron specific constants, macros, types, and routines define
2194  * PCI configuration space registers and how to read them to determine the NUMA
2195  * configuration of *supported* Opteron processors.  They provide the same
2196  * information that may be gotten from the ACPI System Resource Affinity Table
2197  * (SRAT) if it exists on the machine of interest.
2198  *
2199  * The AMD BIOS and Kernel Developer's Guide (BKDG) for the processor family
2200  * of interest describes all of these registers and their contents.  The main
2201  * registers used by this code to determine the NUMA configuration of the
2202  * machine are the node ID register for the number of NUMA nodes and the DRAM
2203  * address map registers for the physical address range of each node.
2204  *
2205  * NOTE: The format and how to determine the NUMA configuration using PCI
2206  *	 config space registers may change or may not be supported in future
2207  *	 Opteron processor families.
2208  */
2209 
2210 /*
2211  * How many bits to shift Opteron DRAM Address Map base and limit registers
2212  * to get actual value
2213  */
2214 #define	OPT_DRAMADDR_HI_LSHIFT_ADDR	40	/* shift left for address */
2215 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
2216 
2217 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
2218 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
2219 
2220 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
2221 
2222 /*
2223  * Macros to derive addresses from Opteron DRAM Address Map registers
2224  */
2225 #define	OPT_DRAMADDR_HI(reg) \
2226 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
2227 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
2228 
2229 #define	OPT_DRAMADDR_LO(reg) \
2230 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
2231 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
2232 
2233 #define	OPT_DRAMADDR(high, low) \
2234 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
2235 
2236 /*
2237  * Bit masks defining what's in Opteron DRAM Address Map base register
2238  */
2239 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
2240 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
2241 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
2242 
2243 /*
2244  * Bit masks defining what's in Opteron DRAM Address Map limit register
2245  */
2246 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
2247 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
2248 
2249 
2250 /*
2251  * Opteron Node ID register in PCI configuration space contains
2252  * number of nodes in system, etc. for Opteron K8.  The following
2253  * constants and macros define its contents, structure, and access.
2254  */
2255 
2256 /*
2257  * Bit masks defining what's in Opteron Node ID register
2258  */
2259 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
2260 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
2261 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
2262 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
2263 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
2264 
2265 /*
2266  * How many bits in Opteron Node ID register to shift right to get actual value
2267  */
2268 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
2269 
2270 /*
2271  * Macros to get values from Opteron Node ID register
2272  */
2273 #define	OPT_NODE_CNT(reg) \
2274 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
2275 
2276 /*
2277  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
2278  * "in/out" instructions
2279  *
2280  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
2281  *	 other uses should just do MMIO to access PCI ECS.
2282  *	 Must enable special bit in Northbridge Configuration Register on
2283  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
2284  *	 using "in/out" instructions and restore special bit after done
2285  *	 accessing PCI ECS.
2286  */
2287 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
2288 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
2289 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
2290 	    ((((reg) >> 8) & 0xf) << 24))
2291 
2292 /*
2293  * PCI configuration space registers accessed by specifying
2294  * a bus, device, function, and offset.  The following constants
2295  * define the values needed to access Opteron K8 configuration
2296  * info to determine its node topology
2297  */
2298 
2299 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
2300 
2301 /*
2302  * Opteron PCI configuration space register function values
2303  */
2304 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
2305 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
2306 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
2307 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
2308 
2309 /*
2310  * PCI Configuration Space register offsets
2311  */
2312 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
2313 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
2314 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
2315 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
2316 
2317 /*
2318  * Opteron PCI Configuration Space device IDs for nodes
2319  */
2320 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
2321 
2322 
2323 /*
2324  * Opteron DRAM address map gives base and limit for physical memory in a node
2325  */
2326 typedef	struct opt_dram_addr_map {
2327 	uint32_t	base_hi;
2328 	uint32_t	base_lo;
2329 	uint32_t	limit_hi;
2330 	uint32_t	limit_lo;
2331 } opt_dram_addr_map_t;
2332 
2333 
2334 /*
2335  * Supported AMD processor families
2336  */
2337 #define	AMD_FAMILY_HAMMER	15
2338 #define	AMD_FAMILY_GREYHOUND	16
2339 
2340 /*
2341  * Whether to have is_opteron() return 1 even when processor isn't supported
2342  */
2343 uint_t	is_opteron_override = 0;
2344 
2345 /*
2346  * AMD processor family for current CPU
2347  */
2348 uint_t	opt_family = 0;
2349 
2350 
2351 /*
2352  * Determine whether we're running on a supported AMD Opteron since reading
2353  * node count and DRAM address map registers may have different format or
2354  * may not be supported across processor families
2355  */
2356 static int
2357 is_opteron(void)
2358 {
2359 
2360 	if (x86_vendor != X86_VENDOR_AMD)
2361 		return (0);
2362 
2363 	opt_family = cpuid_getfamily(CPU);
2364 	if (opt_family == AMD_FAMILY_HAMMER ||
2365 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
2366 		return (1);
2367 	else
2368 		return (0);
2369 }
2370 
2371 
2372 /*
2373  * Determine NUMA configuration for Opteron from registers that live in PCI
2374  * configuration space
2375  */
2376 static void
2377 opt_get_numa_config(uint_t *node_cnt, int *mem_intrlv,
2378     node_phys_addr_map_t *node_memory)
2379 {
2380 	uint_t				bus;
2381 	uint_t				dev;
2382 	struct opt_dram_addr_map	dram_map[MAX_NODES];
2383 	uint_t				node;
2384 	uint_t				node_info[MAX_NODES];
2385 	uint_t				off_hi;
2386 	uint_t				off_lo;
2387 	uint64_t			nb_cfg_reg;
2388 
2389 	/*
2390 	 * Read configuration registers from PCI configuration space to
2391 	 * determine node information, which memory is in each node, etc.
2392 	 *
2393 	 * Write to PCI configuration space address register to specify
2394 	 * which configuration register to read and read/write PCI
2395 	 * configuration space data register to get/set contents
2396 	 */
2397 	bus = OPT_PCS_BUS_CONFIG;
2398 	dev = OPT_PCS_DEV_NODE0;
2399 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
2400 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
2401 
2402 	/*
2403 	 * Read node ID register for node 0 to get node count
2404 	 */
2405 	node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
2406 	    OPT_PCS_OFF_NODEID);
2407 	*node_cnt = OPT_NODE_CNT(node_info[0]) + 1;
2408 
2409 	/*
2410 	 * If number of nodes is more than maximum supported, then set node
2411 	 * count to 1 and treat system as UMA instead of NUMA.
2412 	 */
2413 	if (*node_cnt > MAX_NODES) {
2414 		*node_cnt = 1;
2415 		return;
2416 	}
2417 
2418 	/*
2419 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
2420 	 * read high DRAM address map base and limit registers
2421 	 */
2422 	if (opt_family == AMD_FAMILY_GREYHOUND) {
2423 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
2424 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
2425 			wrmsr(MSR_AMD_NB_CFG,
2426 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
2427 	}
2428 
2429 	for (node = 0; node < *node_cnt; node++) {
2430 		uint32_t	base_hi;
2431 		uint32_t	base_lo;
2432 		uint32_t	limit_hi;
2433 		uint32_t	limit_lo;
2434 
2435 		/*
2436 		 * Read node ID register (except for node 0 which we just read)
2437 		 */
2438 		if (node > 0) {
2439 			node_info[node] = pci_getl_func(bus, dev,
2440 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
2441 		}
2442 
2443 		/*
2444 		 * Read DRAM base and limit registers which specify
2445 		 * physical memory range of each node
2446 		 */
2447 		if (opt_family != AMD_FAMILY_GREYHOUND)
2448 			base_hi = 0;
2449 		else {
2450 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2451 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2452 			base_hi = dram_map[node].base_hi =
2453 			    inl(PCI_CONFDATA);
2454 		}
2455 		base_lo = dram_map[node].base_lo = pci_getl_func(bus, dev,
2456 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
2457 
2458 		if ((dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN) &&
2459 		    mem_intrlv)
2460 			*mem_intrlv = *mem_intrlv + 1;
2461 
2462 		off_hi += 4;	/* high limit register offset */
2463 		if (opt_family != AMD_FAMILY_GREYHOUND)
2464 			limit_hi = 0;
2465 		else {
2466 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
2467 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
2468 			limit_hi = dram_map[node].limit_hi =
2469 			    inl(PCI_CONFDATA);
2470 		}
2471 
2472 		off_lo += 4;	/* low limit register offset */
2473 		limit_lo = dram_map[node].limit_lo = pci_getl_func(bus,
2474 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
2475 
2476 		/*
2477 		 * Increment device number to next node and register offsets
2478 		 * for DRAM base register of next node
2479 		 */
2480 		off_hi += 4;
2481 		off_lo += 4;
2482 		dev++;
2483 
2484 		/*
2485 		 * Both read and write enable bits must be enabled in DRAM
2486 		 * address map base register for physical memory to exist in
2487 		 * node
2488 		 */
2489 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
2490 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
2491 			/*
2492 			 * Mark node memory as non-existent and set start and
2493 			 * end addresses to be same in node_memory[]
2494 			 */
2495 			node_memory[node].exists = 0;
2496 			node_memory[node].start = node_memory[node].end =
2497 			    (pfn_t)-1;
2498 			continue;
2499 		}
2500 
2501 		/*
2502 		 * Mark node memory as existing and remember physical address
2503 		 * range of each node for use later
2504 		 */
2505 		node_memory[node].exists = 1;
2506 
2507 		node_memory[node].start = btop(OPT_DRAMADDR(base_hi, base_lo));
2508 
2509 		node_memory[node].end = btop(OPT_DRAMADDR(limit_hi, limit_lo) |
2510 		    OPT_DRAMADDR_LO_MASK_OFF);
2511 	}
2512 
2513 	/*
2514 	 * Restore PCI Extended Configuration Space enable bit
2515 	 */
2516 	if (opt_family == AMD_FAMILY_GREYHOUND) {
2517 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
2518 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
2519 	}
2520 }
2521 
2522 
2523 /*
2524  * Return average amount of time to read vendor ID register on Northbridge
2525  * N times on specified destination node from current CPU
2526  */
2527 static hrtime_t
2528 opt_probe_vendor(int dest_node, int nreads)
2529 {
2530 	int		cnt;
2531 	uint_t		dev;
2532 	/* LINTED: set but not used in function */
2533 	volatile uint_t	dev_vendor;
2534 	hrtime_t	elapsed;
2535 	hrtime_t	end;
2536 	int		ipl;
2537 	hrtime_t	start;
2538 
2539 	dev = OPT_PCS_DEV_NODE0 + dest_node;
2540 	kpreempt_disable();
2541 	ipl = spl8();
2542 	outl(PCI_CONFADD, PCI_CADDR1(0, dev, OPT_PCS_FUNC_DRAM,
2543 	    OPT_PCS_OFF_VENDOR));
2544 	start = gethrtime();
2545 	for (cnt = 0; cnt < nreads; cnt++)
2546 		dev_vendor = inl(PCI_CONFDATA);
2547 	end = gethrtime();
2548 	elapsed = (end - start) / nreads;
2549 	splx(ipl);
2550 	kpreempt_enable();
2551 	return (elapsed);
2552 }
2553