xref: /titanic_52/usr/src/uts/i86pc/os/lgrpplat.c (revision 06e46062ef4f5f4b687cbafb4518fb123fe23920)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 
30 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
31 #include <sys/cmn_err.h>
32 #include <sys/cpupart.h>
33 #include <sys/cpuvar.h>
34 #include <sys/lgrp.h>
35 #include <sys/machsystm.h>
36 #include <sys/memlist.h>
37 #include <sys/memnode.h>
38 #include <sys/mman.h>
39 #include <sys/pci_cfgspace.h>
40 #include <sys/pci_impl.h>
41 #include <sys/param.h>
42 #include <sys/promif.h>		/* for prom_printf() */
43 #include <sys/systm.h>
44 #include <sys/thread.h>
45 #include <sys/types.h>
46 #include <sys/var.h>
47 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
48 #include <vm/hat_i86.h>
49 #include <vm/seg_kmem.h>
50 #include <vm/vm_dep.h>
51 
52 
53 
54 /*
55  * lgroup platform support for x86 platforms.
56  */
57 
58 #define	MAX_NODES		8
59 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
60 
61 #define	LGRP_PLAT_CPU_TO_NODE(cpu)	(chip_plat_get_chipid(cpu))
62 
63 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
64 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
65 
66 
67 /*
68  * Multiprocessor Opteron machines have Non Uniform Memory Access (NUMA).
69  *
70  * Until System Affinity Resource Table (SRAT) becomes part of ACPI standard,
71  * we need to examine registers in PCI configuration space to determine how
72  * many nodes are in the system and which CPUs and memory are in each node.
73  * This could be determined by probing all memory from each CPU, but that is
74  * too expensive to do while booting the kernel.
75  *
76  * NOTE: Using these PCI configuration space registers to determine this
77  *       locality info is Opteron K8 specific and not guaranteed to work on
78  *       the next generation Opteron processor.  Furthermore, we assume that
79  *	 there is one CPU per node and CPU 0 is in node 0, CPU 1 is in node 1,
80  *	 etc. which should be true for Opteron K8....
81  */
82 
83 /*
84  * Opteron DRAM Address Map in PCI configuration space gives base and limit
85  * of physical memory in each node for Opteron K8.  The following constants
86  * and macros define their contents, structure, and access.
87  */
88 
89 /*
90  * How many bits to shift Opteron DRAM Address Map base and limit registers
91  * to get actual value
92  */
93 #define	OPT_DRAMADDR_LSHIFT_ADDR	8	/* shift left for address */
94 
95 #define	OPT_DRAMADDR_MASK_OFF	0xFFFFFF	/* offset for address */
96 
97 /*
98  * Bit masks defining what's in Opteron DRAM Address Map base register
99  */
100 #define	OPT_DRAMBASE_MASK_RE		0x1	/* read enable */
101 #define	OPT_DRAMBASE_MASK_WE		0x2	/* write enable */
102 #define	OPT_DRAMBASE_MASK_INTRLVEN	0x700	/* interleave */
103 
104 #define	OPT_DRAMBASE_MASK_ADDR	0xFFFF0000	/* address bits 39-24 */
105 
106 /*
107  * Macros to get values from Opteron DRAM Address Map base register
108  */
109 #define	OPT_DRAMBASE(reg) \
110 	(((u_longlong_t)reg & OPT_DRAMBASE_MASK_ADDR) << \
111 	    OPT_DRAMADDR_LSHIFT_ADDR)
112 
113 
114 /*
115  * Bit masks defining what's in Opteron DRAM Address Map limit register
116  */
117 #define	OPT_DRAMLIMIT_MASK_DSTNODE	0x7		/* destination node */
118 #define	OPT_DRAMLIMIT_MASK_INTRLVSEL	0x70		/* interleave select */
119 #define	OPT_DRAMLIMIT_MASK_ADDR		0xFFFF0000	/* addr bits 39-24 */
120 
121 /*
122  * Macros to get values from Opteron DRAM Address Map limit register
123  */
124 #define	OPT_DRAMLIMIT(reg) \
125 	(((u_longlong_t)reg & OPT_DRAMLIMIT_MASK_ADDR) << \
126 	    OPT_DRAMADDR_LSHIFT_ADDR)
127 
128 
129 /*
130  * Opteron Node ID register in PCI configuration space contains
131  * number of nodes in system, etc. for Opteron K8.  The following
132  * constants and macros define its contents, structure, and access.
133  */
134 
135 /*
136  * Bit masks defining what's in Opteron Node ID register
137  */
138 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
139 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
140 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
141 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
142 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
143 
144 /*
145  * How many bits in Opteron Node ID register to shift right to get actual value
146  */
147 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
148 
149 /*
150  * Macros to get values from Opteron Node ID register
151  */
152 #define	OPT_NODE_CNT(reg) \
153 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
154 
155 
156 /*
157  * PCI configuration space registers accessed by specifying
158  * a bus, device, function, and offset.  The following constants
159  * define the values needed to access Opteron K8 configuration
160  * info to determine its node topology
161  */
162 
163 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
164 
165 /*
166  * Opteron PCI configuration space register function values
167  */
168 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
169 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
170 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
171 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
172 
173 /*
174  * PCI Configuration Space register offsets
175  */
176 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
177 #define	OPT_PCS_OFF_DRAMBASE	0x40	/* DRAM Base register (node 0) */
178 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
179 
180 /*
181  * Opteron PCI Configuration Space device IDs for nodes
182  */
183 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
184 
185 
186 /*
187  * Bookkeeping for latencies seen during probing (used for verification)
188  */
189 typedef	struct lgrp_plat_latency_acct {
190 	hrtime_t	la_value;	/* latency value */
191 	int		la_count;	/* occurrences */
192 } lgrp_plat_latency_acct_t;
193 
194 
195 /*
196  * Choices for probing to determine lgroup topology
197  */
198 typedef	enum lgrp_plat_probe_op {
199 	LGRP_PLAT_PROBE_PGCPY,		/* Use page copy */
200 	LGRP_PLAT_PROBE_VENDOR		/* Read vendor ID on Northbridge */
201 } lgrp_plat_probe_op_t;
202 
203 
204 /*
205  * Opteron DRAM address map gives base and limit for physical memory in a node
206  */
207 typedef	struct opt_dram_addr_map {
208 	uint32_t	base;
209 	uint32_t	limit;
210 } opt_dram_addr_map_t;
211 
212 
213 /*
214  * Starting and ending page for physical memory in node
215  */
216 typedef	struct phys_addr_map {
217 	pfn_t	start;
218 	pfn_t	end;
219 	int	exists;
220 } phys_addr_map_t;
221 
222 
223 /*
224  * Opteron DRAM address map for each node
225  */
226 struct opt_dram_addr_map	opt_dram_map[MAX_NODES];
227 
228 /*
229  * Node ID register contents for each node
230  */
231 uint_t				opt_node_info[MAX_NODES];
232 
233 /*
234  * Whether memory is interleaved across nodes causing MPO to be disabled
235  */
236 int			lgrp_plat_mem_intrlv = 0;
237 
238 /*
239  * Number of nodes in system
240  */
241 uint_t			lgrp_plat_node_cnt = 1;
242 
243 /*
244  * Physical address range for memory in each node
245  */
246 phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
247 
248 /*
249  * Probe costs (individual and total) and flush cost
250  */
251 hrtime_t		lgrp_plat_flush_cost = 0;
252 hrtime_t		lgrp_plat_probe_cost = 0;
253 hrtime_t		lgrp_plat_probe_cost_total = 0;
254 
255 /*
256  * Error code for latency adjustment and verification
257  */
258 int			lgrp_plat_probe_error_code = 0;
259 
260 /*
261  * How much latencies were off from minimum values gotten
262  */
263 hrtime_t		lgrp_plat_probe_errors[MAX_NODES][MAX_NODES];
264 
265 /*
266  * Unique probe latencies and number of occurrences of each
267  */
268 lgrp_plat_latency_acct_t	lgrp_plat_probe_lat_acct[MAX_NODES];
269 
270 /*
271  * Size of memory buffer in each node for probing
272  */
273 size_t			lgrp_plat_probe_memsize = 0;
274 
275 /*
276  * Virtual address of page in each node for probing
277  */
278 caddr_t			lgrp_plat_probe_memory[MAX_NODES];
279 
280 /*
281  * Number of unique latencies in probe times
282  */
283 int			lgrp_plat_probe_nlatencies = 0;
284 
285 /*
286  * How many rounds of probing to do
287  */
288 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
289 
290 /*
291  * Number of samples to take when probing each node
292  */
293 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
294 
295 /*
296  * How to probe to determine lgroup topology
297  */
298 lgrp_plat_probe_op_t	lgrp_plat_probe_op = LGRP_PLAT_PROBE_VENDOR;
299 
300 /*
301  * PFN of page in each node for probing
302  */
303 pfn_t			lgrp_plat_probe_pfn[MAX_NODES];
304 
305 /*
306  * Whether probe time was suspect (ie. not within tolerance of value that it
307  * should match)
308  */
309 int			lgrp_plat_probe_suspect[MAX_NODES][MAX_NODES];
310 
311 /*
312  * How long it takes to access memory from each node
313  */
314 hrtime_t		lgrp_plat_probe_times[MAX_NODES][MAX_NODES];
315 
316 /*
317  * Min and max node memory probe times seen
318  */
319 hrtime_t		lgrp_plat_probe_time_max = 0;
320 hrtime_t		lgrp_plat_probe_time_min = -1;
321 hrtime_t		lgrp_plat_probe_max[MAX_NODES][MAX_NODES];
322 hrtime_t		lgrp_plat_probe_min[MAX_NODES][MAX_NODES];
323 
324 
325 /*
326  * Allocate lgrp and lgrp stat arrays statically.
327  */
328 static lgrp_t	lgrp_space[NLGRP];
329 static int	nlgrps_alloc;
330 
331 struct lgrp_stats lgrp_stats[NLGRP];
332 
333 #define	CPUID_FAMILY_OPTERON	15
334 
335 uint_t	opt_family = 0;
336 uint_t	opt_model = 0;
337 uint_t	opt_probe_func = OPT_PCS_FUNC_DRAM;
338 
339 
340 /*
341  * Determine whether we're running on an AMD Opteron K8 machine
342  */
343 int
344 is_opteron(void)
345 {
346 	if (x86_vendor != X86_VENDOR_AMD)
347 		return (0);
348 
349 	if (cpuid_getfamily(CPU) == CPUID_FAMILY_OPTERON)
350 		return (1);
351 	else
352 		return (0);
353 }
354 
355 int
356 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
357 {
358 	if (max_mem_nodes == 1)
359 		return (0);
360 
361 	return ((int)hand);
362 }
363 
364 lgrp_handle_t
365 plat_mem_node_to_lgrphand(int mnode)
366 {
367 	if (max_mem_nodes == 1)
368 		return (LGRP_DEFAULT_HANDLE);
369 
370 	return ((lgrp_handle_t)mnode);
371 }
372 
373 int
374 plat_pfn_to_mem_node(pfn_t pfn)
375 {
376 	int	node;
377 
378 	if (max_mem_nodes == 1)
379 		return (0);
380 
381 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
382 		/*
383 		 * Skip nodes with no memory
384 		 */
385 		if (!lgrp_plat_node_memory[node].exists)
386 			continue;
387 
388 		if (pfn >= lgrp_plat_node_memory[node].start &&
389 		    pfn <= lgrp_plat_node_memory[node].end)
390 			return (node);
391 	}
392 
393 	ASSERT(node < lgrp_plat_node_cnt);
394 	return (-1);
395 }
396 
397 /*
398  * Configure memory nodes for machines with more than one node (ie NUMA)
399  */
400 void
401 plat_build_mem_nodes(struct memlist *list)
402 {
403 	pfn_t		cur_start;	/* start addr of subrange */
404 	pfn_t		cur_end;	/* end addr of subrange */
405 	pfn_t		start;		/* start addr of whole range */
406 	pfn_t		end;		/* end addr of whole range */
407 
408 	/*
409 	 * Boot install lists are arranged <addr, len>, ...
410 	 */
411 	while (list) {
412 		int	node;
413 
414 		start = list->address >> PAGESHIFT;
415 		end = (list->address + list->size - 1) >> PAGESHIFT;
416 
417 		if (start > physmax) {
418 			list = list->next;
419 			continue;
420 		}
421 		if (end > physmax)
422 			end = physmax;
423 
424 		/*
425 		 * When there is only one memnode, just add memory to memnode
426 		 */
427 		if (max_mem_nodes == 1) {
428 			mem_node_add_slice(start, end);
429 			list = list->next;
430 			continue;
431 		}
432 
433 		/*
434 		 * mem_node_add_slice() expects to get a memory range that
435 		 * is within one memnode, so need to split any memory range
436 		 * that spans multiple memnodes into subranges that are each
437 		 * contained within one memnode when feeding them to
438 		 * mem_node_add_slice()
439 		 */
440 		cur_start = start;
441 		do {
442 			node = plat_pfn_to_mem_node(cur_start);
443 
444 			/*
445 			 * Panic if DRAM address map registers or SRAT say
446 			 * memory in node doesn't exist or address from
447 			 * boot installed memory list entry isn't in this node.
448 			 * This shouldn't happen and rest of code can't deal
449 			 * with this if it does.
450 			 */
451 			if (node < 0 || node >= lgrp_plat_node_cnt ||
452 			    !lgrp_plat_node_memory[node].exists ||
453 			    cur_start < lgrp_plat_node_memory[node].start ||
454 			    cur_start > lgrp_plat_node_memory[node].end) {
455 				cmn_err(CE_PANIC, "Don't know which memnode "
456 				    "to add installed memory address 0x%lx\n",
457 				    cur_start);
458 			}
459 
460 			/*
461 			 * End of current subrange should not span memnodes
462 			 */
463 			cur_end = end;
464 			if (lgrp_plat_node_memory[node].exists &&
465 			    cur_end > lgrp_plat_node_memory[node].end)
466 				cur_end = lgrp_plat_node_memory[node].end;
467 
468 			mem_node_add_slice(cur_start, cur_end);
469 
470 			/*
471 			 * Next subrange starts after end of current one
472 			 */
473 			cur_start = cur_end + 1;
474 		} while (cur_end < end);
475 
476 		list = list->next;
477 	}
478 	mem_node_physalign = 0;
479 	mem_node_pfn_shift = 0;
480 }
481 
482 
483 /*
484  * Platform-specific initialization of lgroups
485  */
486 void
487 lgrp_plat_init(void)
488 {
489 	uint_t		bus;
490 	uint_t		dev;
491 	uint_t		node;
492 	uint_t		off;
493 
494 	extern lgrp_load_t	lgrp_expand_proc_thresh;
495 	extern lgrp_load_t	lgrp_expand_proc_diff;
496 
497 	/*
498 	 * Initialize as a UMA machine if this isn't an Opteron
499 	 */
500 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
501 		lgrp_plat_node_cnt = max_mem_nodes = 1;
502 		return;
503 	}
504 
505 	/*
506 	 * Read configuration registers from PCI configuration space to
507 	 * determine node information, which memory is in each node, etc.
508 	 *
509 	 * Write to PCI configuration space address register to specify
510 	 * which configuration register to read and read/write PCI
511 	 * configuration space data register to get/set contents
512 	 */
513 	bus = OPT_PCS_BUS_CONFIG;
514 	dev = OPT_PCS_DEV_NODE0;
515 	off = OPT_PCS_OFF_DRAMBASE;
516 
517 	/*
518 	 * Read node ID register for node 0 to get node count
519 	 */
520 	opt_node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
521 	    OPT_PCS_OFF_NODEID);
522 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
523 
524 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
525 		/*
526 		 * Read node ID register (except for node 0 which we just read)
527 		 */
528 		if (node > 0) {
529 			opt_node_info[node] = pci_getl_func(bus, dev,
530 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
531 		}
532 
533 		/*
534 		 * Read DRAM base and limit registers which specify
535 		 * physical memory range of each node
536 		 */
537 		opt_dram_map[node].base = pci_getl_func(bus, dev,
538 		    OPT_PCS_FUNC_ADDRMAP, off);
539 		if (opt_dram_map[node].base & OPT_DRAMBASE_MASK_INTRLVEN)
540 			lgrp_plat_mem_intrlv++;
541 
542 		off += 4;	/* limit register offset */
543 		opt_dram_map[node].limit = pci_getl_func(bus, dev,
544 		    OPT_PCS_FUNC_ADDRMAP, off);
545 
546 		/*
547 		 * Increment device number to next node and register offset for
548 		 * DRAM base register of next node
549 		 */
550 		off += 4;
551 		dev++;
552 
553 		/*
554 		 * Both read and write enable bits must be enabled in DRAM
555 		 * address map base register for physical memory to exist in
556 		 * node
557 		 */
558 		if ((opt_dram_map[node].base & OPT_DRAMBASE_MASK_RE) == 0 ||
559 		    (opt_dram_map[node].base & OPT_DRAMBASE_MASK_WE) == 0) {
560 			/*
561 			 * Mark node memory as non-existent and set start and
562 			 * end addresses to be same in lgrp_plat_node_memory[]
563 			 */
564 			lgrp_plat_node_memory[node].exists = 0;
565 			lgrp_plat_node_memory[node].start =
566 			    lgrp_plat_node_memory[node].end = (pfn_t)-1;
567 			continue;
568 		}
569 
570 		/*
571 		 * Get PFN for first page in each node,
572 		 * so we can probe memory to determine latency topology
573 		 */
574 		lgrp_plat_probe_pfn[node] =
575 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
576 
577 		/*
578 		 * Mark node memory as existing and remember physical address
579 		 * range of each node for use later
580 		 */
581 		lgrp_plat_node_memory[node].exists = 1;
582 		lgrp_plat_node_memory[node].start =
583 		    btop(OPT_DRAMBASE(opt_dram_map[node].base));
584 		lgrp_plat_node_memory[node].end =
585 		    btop(OPT_DRAMLIMIT(opt_dram_map[node].limit) |
586 		    OPT_DRAMADDR_MASK_OFF);
587 	}
588 
589 	/*
590 	 * Only use one memory node if memory is interleaved between any nodes
591 	 */
592 	if (lgrp_plat_mem_intrlv) {
593 		lgrp_plat_node_cnt = max_mem_nodes = 1;
594 		(void) lgrp_topo_ht_limit_set(1);
595 	} else {
596 		max_mem_nodes = lgrp_plat_node_cnt;
597 
598 		/*
599 		 * Probing errors can mess up the lgroup topology and force us
600 		 * fall back to a 2 level lgroup topology.  Here we bound how
601 		 * tall the lgroup topology can grow in hopes of avoiding any
602 		 * anamolies in probing from messing up the lgroup topology
603 		 * by limiting the accuracy of the latency topology.
604 		 *
605 		 * Assume that nodes will at least be configured in a ring,
606 		 * so limit height of lgroup topology to be less than number
607 		 * of nodes on a system with 4 or more nodes
608 		 */
609 		if (lgrp_plat_node_cnt >= 4 &&
610 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
611 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
612 	}
613 
614 	/*
615 	 * Lgroups on Opteron architectures have but a single physical
616 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
617 	 * so that lgrp_choose() will spread things out aggressively.
618 	 */
619 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
620 	lgrp_expand_proc_diff = 0;
621 }
622 
623 
624 /*
625  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
626  * be considered same
627  */
628 #define	LGRP_LAT_TOLERANCE_SHIFT	4
629 
630 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
631 
632 
633 /*
634  * Adjust latencies between nodes to be symmetric, normalize latencies between
635  * any nodes that are within some tolerance to be same, and make local
636  * latencies be same
637  */
638 static void
639 lgrp_plat_latency_adjust(void)
640 {
641 	int				i;
642 	int				j;
643 	int				k;
644 	int				l;
645 	u_longlong_t			max;
646 	u_longlong_t			min;
647 	u_longlong_t			t;
648 	u_longlong_t			t1;
649 	u_longlong_t			t2;
650 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LATENCY_CHANGE;
651 	int				lat_corrected[MAX_NODES][MAX_NODES];
652 
653 	/*
654 	 * Nothing to do when this is an UMA machine
655 	 */
656 	if (max_mem_nodes == 1)
657 		return;
658 
659 	/*
660 	 * Make sure that latencies are symmetric between any two nodes
661 	 * (ie. latency(node0, node1) == latency(node1, node0))
662 	 */
663 	for (i = 0; i < lgrp_plat_node_cnt; i++)
664 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
665 			t1 = lgrp_plat_probe_times[i][j];
666 			t2 = lgrp_plat_probe_times[j][i];
667 
668 			if (t1 == 0 || t2 == 0 || t1 == t2)
669 				continue;
670 
671 			/*
672 			 * Latencies should be same
673 			 * - Use minimum of two latencies which should be same
674 			 * - Track suspect probe times not within tolerance of
675 			 *   min value
676 			 * - Remember how much values are corrected by
677 			 */
678 			if (t1 > t2) {
679 				t = t2;
680 				lgrp_plat_probe_errors[i][j] += t1 - t2;
681 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
682 					lgrp_plat_probe_suspect[i][j]++;
683 					lgrp_plat_probe_suspect[j][i]++;
684 				}
685 			} else if (t2 > t1) {
686 				t = t1;
687 				lgrp_plat_probe_errors[j][i] += t2 - t1;
688 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
689 					lgrp_plat_probe_suspect[i][j]++;
690 					lgrp_plat_probe_suspect[j][i]++;
691 				}
692 			}
693 
694 			lgrp_plat_probe_times[i][j] =
695 			    lgrp_plat_probe_times[j][i] = t;
696 			lgrp_config(cflag, t1, t);
697 			lgrp_config(cflag, t2, t);
698 		}
699 
700 	/*
701 	 * Keep track of which latencies get corrected
702 	 */
703 	for (i = 0; i < MAX_NODES; i++)
704 		for (j = 0; j < MAX_NODES; j++)
705 			lat_corrected[i][j] = 0;
706 
707 	/*
708 	 * For every two nodes, see whether there is another pair of nodes which
709 	 * are about the same distance apart and make the latencies be the same
710 	 * if they are close enough together
711 	 */
712 	for (i = 0; i < lgrp_plat_node_cnt; i++)
713 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
714 			/*
715 			 * Pick one pair of nodes (i, j)
716 			 * and get latency between them
717 			 */
718 			t1 = lgrp_plat_probe_times[i][j];
719 
720 			/*
721 			 * Skip this pair of nodes if there isn't a latency
722 			 * for it yet
723 			 */
724 			if (t1 == 0)
725 				continue;
726 
727 			for (k = 0; k < lgrp_plat_node_cnt; k++)
728 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
729 					/*
730 					 * Pick another pair of nodes (k, l)
731 					 * not same as (i, j) and get latency
732 					 * between them
733 					 */
734 					if (k == i && l == j)
735 						continue;
736 
737 					t2 = lgrp_plat_probe_times[k][l];
738 
739 					/*
740 					 * Skip this pair of nodes if there
741 					 * isn't a latency for it yet
742 					 */
743 
744 					if (t2 == 0)
745 						continue;
746 
747 					/*
748 					 * Skip nodes (k, l) if they already
749 					 * have same latency as (i, j) or
750 					 * their latency isn't close enough to
751 					 * be considered/made the same
752 					 */
753 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
754 					    t1 >> lgrp_plat_probe_lt_shift) ||
755 					    (t2 > t1 && t2 - t1 >
756 					    t2 >> lgrp_plat_probe_lt_shift))
757 						continue;
758 
759 					/*
760 					 * Make latency(i, j) same as
761 					 * latency(k, l), try to use latency
762 					 * that has been adjusted already to get
763 					 * more consistency (if possible), and
764 					 * remember which latencies were
765 					 * adjusted for next time
766 					 */
767 					if (lat_corrected[i][j]) {
768 						t = t1;
769 						lgrp_config(cflag, t2, t);
770 						t2 = t;
771 					} else if (lat_corrected[k][l]) {
772 						t = t2;
773 						lgrp_config(cflag, t1, t);
774 						t1 = t;
775 					} else {
776 						if (t1 > t2)
777 							t = t2;
778 						else
779 							t = t1;
780 						lgrp_config(cflag, t1, t);
781 						lgrp_config(cflag, t2, t);
782 						t1 = t2 = t;
783 					}
784 
785 					lgrp_plat_probe_times[i][j] =
786 					    lgrp_plat_probe_times[k][l] = t;
787 
788 					lat_corrected[i][j] =
789 					    lat_corrected[k][l] = 1;
790 				}
791 		}
792 
793 	/*
794 	 * Local latencies should be same
795 	 * - Find min and max local latencies
796 	 * - Make all local latencies be minimum
797 	 */
798 	min = -1;
799 	max = 0;
800 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
801 		t = lgrp_plat_probe_times[i][i];
802 		if (t == 0)
803 			continue;
804 		if (min == -1 || t < min)
805 			min = t;
806 		if (t > max)
807 			max = t;
808 	}
809 	if (min != max) {
810 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
811 			int	local;
812 
813 			local = lgrp_plat_probe_times[i][i];
814 			if (local == 0)
815 				continue;
816 
817 			/*
818 			 * Track suspect probe times that aren't within
819 			 * tolerance of minimum local latency and how much
820 			 * probe times are corrected by
821 			 */
822 			if (local - min > min >> lgrp_plat_probe_lt_shift)
823 				lgrp_plat_probe_suspect[i][i]++;
824 
825 			lgrp_plat_probe_errors[i][i] += local - min;
826 
827 			/*
828 			 * Make local latencies be minimum
829 			 */
830 			lgrp_config(cflag, local, min);
831 			lgrp_plat_probe_times[i][i] = min;
832 		}
833 	}
834 
835 	/*
836 	 * Determine max probe time again since just adjusted latencies
837 	 */
838 	lgrp_plat_probe_time_max = 0;
839 	for (i = 0; i < lgrp_plat_node_cnt; i++)
840 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
841 			t = lgrp_plat_probe_times[i][j];
842 			if (t > lgrp_plat_probe_time_max)
843 				lgrp_plat_probe_time_max = t;
844 		}
845 }
846 
847 
848 /*
849  * Verify following about latencies between nodes:
850  *
851  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
852  * - Local latencies same
853  * - Local < remote
854  * - Number of latencies seen is reasonable
855  * - Number of occurrences of a given latency should be more than 1
856  *
857  * Returns:
858  *	0	Success
859  *	-1	Not symmetric
860  *	-2	Local latencies not same
861  *	-3	Local >= remote
862  *	-4	Wrong number of latencies
863  *	-5	Not enough occurrences of given latency
864  */
865 static int
866 lgrp_plat_latency_verify(void)
867 {
868 	int				i;
869 	int				j;
870 	lgrp_plat_latency_acct_t	*l;
871 	int				probed;
872 	u_longlong_t			t1;
873 	u_longlong_t			t2;
874 
875 	/*
876 	 * Nothing to do when this is an UMA machine, lgroup topology is
877 	 * limited to 2 levels, or there aren't any probe times yet
878 	 */
879 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
880 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
881 		return (0);
882 
883 	/*
884 	 * Make sure that latencies are symmetric between any two nodes
885 	 * (ie. latency(node0, node1) == latency(node1, node0))
886 	 */
887 	for (i = 0; i < lgrp_plat_node_cnt; i++)
888 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
889 			t1 = lgrp_plat_probe_times[i][j];
890 			t2 = lgrp_plat_probe_times[j][i];
891 
892 			if (t1 == 0 || t2 == 0 || t1 == t2)
893 				continue;
894 
895 			return (-1);
896 		}
897 
898 	/*
899 	 * Local latencies should be same
900 	 */
901 	t1 = lgrp_plat_probe_times[0][0];
902 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
903 		t2 = lgrp_plat_probe_times[i][i];
904 		if (t2 == 0)
905 			continue;
906 
907 		if (t1 == 0) {
908 			t1 = t2;
909 			continue;
910 		}
911 
912 		if (t1 != t2)
913 			return (-2);
914 	}
915 
916 	/*
917 	 * Local latencies should be less than remote
918 	 */
919 	if (t1) {
920 		for (i = 0; i < lgrp_plat_node_cnt; i++)
921 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
922 				t2 = lgrp_plat_probe_times[i][j];
923 				if (i == j || t2 == 0)
924 					continue;
925 
926 				if (t1 >= t2)
927 					return (-3);
928 			}
929 	}
930 
931 	/*
932 	 * Rest of checks are not very useful for machines with less than
933 	 * 4 nodes (which means less than 3 latencies on Opteron)
934 	 */
935 	if (lgrp_plat_node_cnt < 4)
936 		return (0);
937 
938 	/*
939 	 * Need to see whether done probing in order to verify number of
940 	 * latencies are correct
941 	 */
942 	probed = 0;
943 	for (i = 0; i < lgrp_plat_node_cnt; i++)
944 		if (lgrp_plat_probe_times[i][i])
945 			probed++;
946 
947 	if (probed != lgrp_plat_node_cnt)
948 		return (0);
949 
950 	/*
951 	 * Determine number of unique latencies seen in probe times,
952 	 * their values, and number of occurrences of each
953 	 */
954 	lgrp_plat_probe_nlatencies = 0;
955 	bzero(lgrp_plat_probe_lat_acct,
956 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
957 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
958 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
959 			int	k;
960 
961 			/*
962 			 * Look at each probe time
963 			 */
964 			t1 = lgrp_plat_probe_times[i][j];
965 			if (t1 == 0)
966 				continue;
967 
968 			/*
969 			 * Account for unique latencies
970 			 */
971 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
972 				l = &lgrp_plat_probe_lat_acct[k];
973 				if (t1 == l->la_value) {
974 					/*
975 					 * Increment number of occurrences
976 					 * if seen before
977 					 */
978 					l->la_count++;
979 					break;
980 				} else if (l->la_value == 0) {
981 					/*
982 					 * Record latency if haven't seen before
983 					 */
984 					l->la_value = t1;
985 					l->la_count++;
986 					lgrp_plat_probe_nlatencies++;
987 					break;
988 				}
989 			}
990 		}
991 	}
992 
993 	/*
994 	 * Number of latencies should be relative to number of
995 	 * nodes in system:
996 	 * - Same as nodes when nodes <= 2
997 	 * - Less than nodes when nodes > 2
998 	 * - Greater than 2 when nodes >= 4
999 	 */
1000 	if ((lgrp_plat_node_cnt <= 2 &&
1001 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
1002 	    (lgrp_plat_node_cnt > 2 &&
1003 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
1004 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
1005 	    lgrp_plat_probe_nlatencies <= 2))
1006 		return (-4);
1007 
1008 	/*
1009 	 * There should be more than one occurrence of every latency
1010 	 * as long as probing is complete
1011 	 */
1012 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
1013 		l = &lgrp_plat_probe_lat_acct[i];
1014 		if (l->la_count <= 1)
1015 			return (-5);
1016 	}
1017 	return (0);
1018 }
1019 
1020 
1021 /*
1022  * Set lgroup latencies for 2 level lgroup topology
1023  */
1024 static void
1025 lgrp_plat_2level_setup(void)
1026 {
1027 	int	i;
1028 
1029 	if (lgrp_plat_node_cnt >= 4)
1030 		cmn_err(CE_NOTE,
1031 		    "MPO only optimizing for local and remote\n");
1032 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1033 		int	j;
1034 
1035 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1036 			if (i == j)
1037 				lgrp_plat_probe_times[i][j] = 2;
1038 			else
1039 				lgrp_plat_probe_times[i][j] = 3;
1040 		}
1041 	}
1042 	lgrp_plat_probe_time_min = 2;
1043 	lgrp_plat_probe_time_max = 3;
1044 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1045 }
1046 
1047 
1048 /*
1049  * Return time needed to probe from current CPU to memory in given node
1050  */
1051 static hrtime_t
1052 lgrp_plat_probe_time(int to)
1053 {
1054 	caddr_t		buf;
1055 	uint_t		dev;
1056 	/* LINTED: set but not used in function */
1057 	volatile uint_t	dev_vendor;
1058 	hrtime_t	elapsed;
1059 	hrtime_t	end;
1060 	int		from;
1061 	int		i;
1062 	int		ipl;
1063 	hrtime_t	max;
1064 	hrtime_t	min;
1065 	hrtime_t	start;
1066 	extern int	use_sse_pagecopy;
1067 
1068 	/*
1069 	 * Determine ID of node containing current CPU
1070 	 */
1071 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1072 
1073 	/*
1074 	 * Do common work for probing main memory
1075 	 */
1076 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1077 		/*
1078 		 * Skip probing any nodes without memory and
1079 		 * set probe time to 0
1080 		 */
1081 		if (lgrp_plat_probe_memory[to] == NULL) {
1082 			lgrp_plat_probe_times[from][to] = 0;
1083 			return (0);
1084 		}
1085 
1086 		/*
1087 		 * Invalidate caches once instead of once every sample
1088 		 * which should cut cost of probing by a lot
1089 		 */
1090 		lgrp_plat_flush_cost = gethrtime();
1091 		invalidate_cache();
1092 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1093 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1094 	}
1095 
1096 	/*
1097 	 * Probe from current CPU to given memory using specified operation
1098 	 * and take specified number of samples
1099 	 */
1100 	max = 0;
1101 	min = -1;
1102 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1103 		lgrp_plat_probe_cost = gethrtime();
1104 
1105 		/*
1106 		 * Can't measure probe time if gethrtime() isn't working yet
1107 		 */
1108 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1109 			return (0);
1110 
1111 		switch (lgrp_plat_probe_op) {
1112 
1113 		case LGRP_PLAT_PROBE_PGCPY:
1114 		default:
1115 			/*
1116 			 * Measure how long it takes to copy page
1117 			 * on top of itself
1118 			 */
1119 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1120 
1121 			kpreempt_disable();
1122 			ipl = splhigh();
1123 			start = gethrtime();
1124 			if (use_sse_pagecopy)
1125 				hwblkpagecopy(buf, buf);
1126 			else
1127 				bcopy(buf, buf, PAGESIZE);
1128 			end = gethrtime();
1129 			elapsed = end - start;
1130 			splx(ipl);
1131 			kpreempt_enable();
1132 			break;
1133 
1134 		case LGRP_PLAT_PROBE_VENDOR:
1135 			/*
1136 			 * Measure how long it takes to read vendor ID from
1137 			 * Northbridge
1138 			 */
1139 			dev = OPT_PCS_DEV_NODE0 + to;
1140 			kpreempt_disable();
1141 			ipl = spl8();
1142 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1143 			    OPT_PCS_OFF_VENDOR));
1144 			start = gethrtime();
1145 			dev_vendor = inl(PCI_CONFDATA);
1146 			end = gethrtime();
1147 			elapsed = end - start;
1148 			splx(ipl);
1149 			kpreempt_enable();
1150 			break;
1151 		}
1152 
1153 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1154 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1155 
1156 		if (min == -1 || elapsed < min)
1157 			min = elapsed;
1158 		if (elapsed > max)
1159 			max = elapsed;
1160 	}
1161 
1162 	/*
1163 	 * Update minimum and maximum probe times between
1164 	 * these two nodes
1165 	 */
1166 	if (min < lgrp_plat_probe_min[from][to] ||
1167 	    lgrp_plat_probe_min[from][to] == 0)
1168 		lgrp_plat_probe_min[from][to] = min;
1169 
1170 	if (max > lgrp_plat_probe_max[from][to])
1171 		lgrp_plat_probe_max[from][to] = max;
1172 
1173 	return (min);
1174 }
1175 
1176 
1177 /*
1178  * Probe memory in each node from current CPU to determine latency topology
1179  */
1180 void
1181 lgrp_plat_probe(void)
1182 {
1183 	int		from;
1184 	int		i;
1185 	hrtime_t	probe_time;
1186 	int		to;
1187 
1188 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1189 		return;
1190 
1191 	/*
1192 	 * Determine ID of node containing current CPU
1193 	 */
1194 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1195 
1196 	/*
1197 	 * Don't need to probe if got times already
1198 	 */
1199 	if (lgrp_plat_probe_times[from][from] != 0)
1200 		return;
1201 
1202 	/*
1203 	 * Read vendor ID in Northbridge or read and write page(s)
1204 	 * in each node from current CPU and remember how long it takes,
1205 	 * so we can build latency topology of machine later.
1206 	 * This should approximate the memory latency between each node.
1207 	 */
1208 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1209 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1210 			/*
1211 			 * Get probe time and bail out if can't get it yet
1212 			 */
1213 			probe_time = lgrp_plat_probe_time(to);
1214 			if (probe_time == 0)
1215 				return;
1216 
1217 			/*
1218 			 * Keep lowest probe time as latency between nodes
1219 			 */
1220 			if (lgrp_plat_probe_times[from][to] == 0 ||
1221 			    probe_time < lgrp_plat_probe_times[from][to])
1222 				lgrp_plat_probe_times[from][to] = probe_time;
1223 
1224 			/*
1225 			 * Update overall minimum and maximum probe times
1226 			 * across all nodes
1227 			 */
1228 			if (probe_time < lgrp_plat_probe_time_min ||
1229 			    lgrp_plat_probe_time_min == -1)
1230 				lgrp_plat_probe_time_min = probe_time;
1231 			if (probe_time > lgrp_plat_probe_time_max)
1232 				lgrp_plat_probe_time_max = probe_time;
1233 		}
1234 
1235 	/*
1236 	 * - Fix up latencies such that local latencies are same,
1237 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1238 	 *
1239 	 * - Verify that latencies look ok
1240 	 *
1241 	 * - Fallback to just optimizing for local and remote if
1242 	 *   latencies didn't look right
1243 	 */
1244 	lgrp_plat_latency_adjust();
1245 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1246 	if (lgrp_plat_probe_error_code)
1247 		lgrp_plat_2level_setup();
1248 }
1249 
1250 
1251 /*
1252  * Platform-specific initialization
1253  */
1254 void
1255 lgrp_plat_main_init(void)
1256 {
1257 	int	curnode;
1258 	int	ht_limit;
1259 	int	i;
1260 
1261 	/*
1262 	 * Print a notice that MPO is disabled when memory is interleaved
1263 	 * across nodes....Would do this when it is discovered, but can't
1264 	 * because it happens way too early during boot....
1265 	 */
1266 	if (lgrp_plat_mem_intrlv)
1267 		cmn_err(CE_NOTE,
1268 		    "MPO disabled because memory is interleaved\n");
1269 
1270 	/*
1271 	 * Don't bother to do any probing if there is only one node or the
1272 	 * height of the lgroup topology less than or equal to 2
1273 	 */
1274 	ht_limit = lgrp_topo_ht_limit();
1275 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1276 		/*
1277 		 * Setup lgroup latencies for 2 level lgroup topology
1278 		 * (ie. local and remote only) if they haven't been set yet
1279 		 */
1280 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1281 		    lgrp_plat_probe_time_max == 0)
1282 			lgrp_plat_2level_setup();
1283 		return;
1284 	}
1285 
1286 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1287 		/*
1288 		 * Should have been able to probe from CPU 0 when it was added
1289 		 * to lgroup hierarchy, but may not have been able to then
1290 		 * because it happens so early in boot that gethrtime() hasn't
1291 		 * been initialized.  (:-(
1292 		 */
1293 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1294 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1295 			lgrp_plat_probe();
1296 
1297 		return;
1298 	}
1299 
1300 	/*
1301 	 * When probing memory, use one page for every sample to determine
1302 	 * lgroup topology and taking multiple samples
1303 	 */
1304 	if (lgrp_plat_probe_memsize == 0)
1305 		lgrp_plat_probe_memsize = PAGESIZE *
1306 		    lgrp_plat_probe_nsamples;
1307 
1308 	/*
1309 	 * Map memory in each node needed for probing to determine latency
1310 	 * topology
1311 	 */
1312 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1313 		int	mnode;
1314 
1315 		/*
1316 		 * Skip this node and leave its probe page NULL
1317 		 * if it doesn't have any memory
1318 		 */
1319 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1320 		if (!mem_node_config[mnode].exists) {
1321 			lgrp_plat_probe_memory[i] = NULL;
1322 			continue;
1323 		}
1324 
1325 		/*
1326 		 * Allocate one kernel virtual page
1327 		 */
1328 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1329 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1330 		if (lgrp_plat_probe_memory[i] == NULL) {
1331 			cmn_err(CE_WARN,
1332 			    "lgrp_plat_main_init: couldn't allocate memory");
1333 			return;
1334 		}
1335 
1336 		/*
1337 		 * Map virtual page to first page in node
1338 		 */
1339 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1340 		    lgrp_plat_probe_memsize,
1341 		    lgrp_plat_probe_pfn[i],
1342 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1343 		    HAT_LOAD_NOCONSIST);
1344 	}
1345 
1346 	/*
1347 	 * Probe from current CPU
1348 	 */
1349 	lgrp_plat_probe();
1350 }
1351 
1352 /*
1353  * Allocate additional space for an lgroup.
1354  */
1355 /* ARGSUSED */
1356 lgrp_t *
1357 lgrp_plat_alloc(lgrp_id_t lgrpid)
1358 {
1359 	lgrp_t *lgrp;
1360 
1361 	lgrp = &lgrp_space[nlgrps_alloc++];
1362 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1363 		return (NULL);
1364 	return (lgrp);
1365 }
1366 
1367 /*
1368  * Platform handling for (re)configuration changes
1369  */
1370 /* ARGSUSED */
1371 void
1372 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1373 {
1374 }
1375 
1376 /*
1377  * Return the platform handle for the lgroup containing the given CPU
1378  */
1379 /* ARGSUSED */
1380 lgrp_handle_t
1381 lgrp_plat_cpu_to_hand(processorid_t id)
1382 {
1383 	if (lgrp_plat_node_cnt == 1)
1384 		return (LGRP_DEFAULT_HANDLE);
1385 
1386 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1387 }
1388 
1389 /*
1390  * Return the platform handle of the lgroup that contains the physical memory
1391  * corresponding to the given page frame number
1392  */
1393 /* ARGSUSED */
1394 lgrp_handle_t
1395 lgrp_plat_pfn_to_hand(pfn_t pfn)
1396 {
1397 	int	mnode;
1398 
1399 	if (max_mem_nodes == 1)
1400 		return (LGRP_DEFAULT_HANDLE);
1401 
1402 	mnode = plat_pfn_to_mem_node(pfn);
1403 	return (MEM_NODE_2_LGRPHAND(mnode));
1404 }
1405 
1406 /*
1407  * Return the maximum number of lgrps supported by the platform.
1408  * Before lgrp topology is known it returns an estimate based on the number of
1409  * nodes. Once topology is known it returns the actual maximim number of lgrps
1410  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1411  * may not grow during system lifetime.
1412  */
1413 int
1414 lgrp_plat_max_lgrps()
1415 {
1416 	return (lgrp_topo_initialized ?
1417 	    lgrp_alloc_max + 1 :
1418 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1419 }
1420 
1421 /*
1422  * Return the number of free, allocatable, or installed
1423  * pages in an lgroup
1424  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1425  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1426  */
1427 /* ARGSUSED */
1428 static pgcnt_t
1429 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1430 {
1431 	struct memlist *mlist;
1432 	pgcnt_t npgs = 0;
1433 	extern struct memlist *phys_avail;
1434 	extern struct memlist *phys_install;
1435 
1436 	switch (query) {
1437 	case LGRP_MEM_SIZE_FREE:
1438 		return ((pgcnt_t)freemem);
1439 	case LGRP_MEM_SIZE_AVAIL:
1440 		memlist_read_lock();
1441 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1442 			npgs += btop(mlist->size);
1443 		memlist_read_unlock();
1444 		return (npgs);
1445 	case LGRP_MEM_SIZE_INSTALL:
1446 		memlist_read_lock();
1447 		for (mlist = phys_install; mlist; mlist = mlist->next)
1448 			npgs += btop(mlist->size);
1449 		memlist_read_unlock();
1450 		return (npgs);
1451 	default:
1452 		return ((pgcnt_t)0);
1453 	}
1454 }
1455 
1456 /*
1457  * Return the number of free pages in an lgroup.
1458  *
1459  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1460  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1461  * number of allocatable base pagesize pages corresponding to the
1462  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1463  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1464  * memory installed, regardless of whether or not it's usable.
1465  */
1466 pgcnt_t
1467 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1468 {
1469 	int	mnode;
1470 	pgcnt_t npgs = (pgcnt_t)0;
1471 	extern struct memlist *phys_avail;
1472 	extern struct memlist *phys_install;
1473 
1474 
1475 	if (plathand == LGRP_DEFAULT_HANDLE)
1476 		return (lgrp_plat_mem_size_default(plathand, query));
1477 
1478 	if (plathand != LGRP_NULL_HANDLE) {
1479 		mnode = plat_lgrphand_to_mem_node(plathand);
1480 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1481 			switch (query) {
1482 			case LGRP_MEM_SIZE_FREE:
1483 				npgs = MNODE_PGCNT(mnode);
1484 				break;
1485 			case LGRP_MEM_SIZE_AVAIL:
1486 				npgs = mem_node_memlist_pages(mnode,
1487 				    phys_avail);
1488 				break;
1489 			case LGRP_MEM_SIZE_INSTALL:
1490 				npgs = mem_node_memlist_pages(mnode,
1491 				    phys_install);
1492 				break;
1493 			default:
1494 				break;
1495 			}
1496 		}
1497 	}
1498 	return (npgs);
1499 }
1500 
1501 /*
1502  * Return latency between "from" and "to" lgroups
1503  *
1504  * This latency number can only be used for relative comparison
1505  * between lgroups on the running system, cannot be used across platforms,
1506  * and may not reflect the actual latency.  It is platform and implementation
1507  * specific, so platform gets to decide its value.  It would be nice if the
1508  * number was at least proportional to make comparisons more meaningful though.
1509  */
1510 /* ARGSUSED */
1511 int
1512 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1513 {
1514 	lgrp_handle_t	src, dest;
1515 
1516 	if (max_mem_nodes == 1)
1517 		return (0);
1518 
1519 	/*
1520 	 * Return max latency for root lgroup
1521 	 */
1522 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1523 		return (lgrp_plat_probe_time_max);
1524 
1525 	src = from;
1526 	dest = to;
1527 
1528 	/*
1529 	 * Return 0 for nodes (lgroup platform handles) out of range
1530 	 */
1531 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1532 		return (0);
1533 
1534 	/*
1535 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1536 	 * and we are trying to get latency from current CPU to some node
1537 	 */
1538 	if (lgrp_plat_probe_times[src][src] == 0 &&
1539 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1540 		lgrp_plat_probe();
1541 
1542 	return (lgrp_plat_probe_times[src][dest]);
1543 }
1544 
1545 /*
1546  * Return platform handle for root lgroup
1547  */
1548 lgrp_handle_t
1549 lgrp_plat_root_hand(void)
1550 {
1551 	return (LGRP_DEFAULT_HANDLE);
1552 }
1553