xref: /titanic_51/usr/src/uts/i86pc/os/lgrpplat.c (revision a776e90690d98f629ab57f6493ff5949af9d65bd)
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 #include <sys/archsystm.h>	/* for {in,out}{b,w,l}() */
31 #include <sys/cmn_err.h>
32 #include <sys/controlregs.h>
33 #include <sys/cpupart.h>
34 #include <sys/cpuvar.h>
35 #include <sys/lgrp.h>
36 #include <sys/machsystm.h>
37 #include <sys/memlist.h>
38 #include <sys/memnode.h>
39 #include <sys/mman.h>
40 #include <sys/pci_cfgspace.h>
41 #include <sys/pci_impl.h>
42 #include <sys/param.h>
43 #include <sys/pghw.h>
44 #include <sys/promif.h>		/* for prom_printf() */
45 #include <sys/systm.h>
46 #include <sys/thread.h>
47 #include <sys/types.h>
48 #include <sys/var.h>
49 #include <sys/x86_archext.h>	/* for x86_feature and X86_AMD */
50 #include <sys/sysmacros.h>
51 #include <vm/hat_i86.h>
52 #include <vm/seg_kmem.h>
53 #include <vm/vm_dep.h>
54 
55 
56 /*
57  * lgroup platform support for x86 platforms.
58  */
59 
60 #define	MAX_NODES		8
61 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
62 
63 #define	LGRP_PLAT_CPU_TO_NODE(cpu) (pg_plat_hw_instance_id(cpu, PGHW_CHIP))
64 
65 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
66 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
67 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
68 
69 /*
70  * Multiprocessor Opteron machines have Non Uniform Memory Access (NUMA).
71  *
72  * Until this code supports reading System Resource Affinity Table (SRAT),
73  * we need to examine registers in PCI configuration space to determine how
74  * many nodes are in the system and which CPUs and memory are in each node.
75  * This could be determined by probing all memory from each CPU, but that is
76  * too expensive to do while booting the kernel.
77  *
78  * NOTE: Using these PCI configuration space registers to determine this
79  *       locality info is not guaranteed to work on future generations of
80  *	 Opteron processor.
81  */
82 
83 /*
84  * Opteron DRAM Address Map in PCI configuration space gives base and limit
85  * of physical memory in each node.  The following constants and macros define
86  * 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_HI_LSHIFT_ADDR	40	/* shift left for address */
94 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
95 
96 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
97 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
98 
99 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
100 
101 /*
102  * Macros to derive addresses from Opteron DRAM Address Map registers
103  */
104 #define	OPT_DRAMADDR_HI(reg) \
105 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
106 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
107 
108 #define	OPT_DRAMADDR_LO(reg) \
109 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
110 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
111 
112 #define	OPT_DRAMADDR(high, low) \
113 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
114 
115 /*
116  * Bit masks defining what's in Opteron DRAM Address Map base register
117  */
118 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
119 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
120 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
121 
122 /*
123  * Bit masks defining what's in Opteron DRAM Address Map limit register
124  */
125 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
126 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
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  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
157  * "in/out" instructions
158  *
159  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
160  *	 other uses should just do MMIO to access PCI ECS.
161  *	 Must enable special bit in Northbridge Configuration Register on
162  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
163  *	 using "in/out" instructions and restore special bit after done
164  *	 accessing PCI ECS.
165  */
166 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
167 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
168 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
169 	    ((((reg) >> 8) & 0xf) << 24))
170 
171 /*
172  * PCI configuration space registers accessed by specifying
173  * a bus, device, function, and offset.  The following constants
174  * define the values needed to access Opteron K8 configuration
175  * info to determine its node topology
176  */
177 
178 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
179 
180 /*
181  * Opteron PCI configuration space register function values
182  */
183 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
184 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
185 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
186 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
187 
188 /*
189  * PCI Configuration Space register offsets
190  */
191 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
192 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
193 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
194 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
195 
196 /*
197  * Opteron PCI Configuration Space device IDs for nodes
198  */
199 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
200 
201 
202 /*
203  * Bookkeeping for latencies seen during probing (used for verification)
204  */
205 typedef	struct lgrp_plat_latency_acct {
206 	hrtime_t	la_value;	/* latency value */
207 	int		la_count;	/* occurrences */
208 } lgrp_plat_latency_acct_t;
209 
210 
211 /*
212  * Choices for probing to determine lgroup topology
213  */
214 typedef	enum lgrp_plat_probe_op {
215 	LGRP_PLAT_PROBE_PGCPY,		/* Use page copy */
216 	LGRP_PLAT_PROBE_VENDOR		/* Read vendor ID on Northbridge */
217 } lgrp_plat_probe_op_t;
218 
219 
220 /*
221  * Opteron DRAM address map gives base and limit for physical memory in a node
222  */
223 typedef	struct opt_dram_addr_map {
224 	uint32_t	base_hi;
225 	uint32_t	base_lo;
226 	uint32_t	limit_hi;
227 	uint32_t	limit_lo;
228 } opt_dram_addr_map_t;
229 
230 
231 /*
232  * Starting and ending page for physical memory in node
233  */
234 typedef	struct phys_addr_map {
235 	pfn_t	start;
236 	pfn_t	end;
237 	int	exists;
238 } phys_addr_map_t;
239 
240 
241 /*
242  * Opteron DRAM address map for each node
243  */
244 struct opt_dram_addr_map	opt_dram_map[MAX_NODES];
245 
246 /*
247  * Node ID register contents for each node
248  */
249 uint_t				opt_node_info[MAX_NODES];
250 
251 /*
252  * Whether memory is interleaved across nodes causing MPO to be disabled
253  */
254 int			lgrp_plat_mem_intrlv = 0;
255 
256 /*
257  * Number of nodes in system
258  */
259 uint_t			lgrp_plat_node_cnt = 1;
260 
261 /*
262  * Physical address range for memory in each node
263  */
264 phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
265 
266 /*
267  * Probe costs (individual and total) and flush cost
268  */
269 hrtime_t		lgrp_plat_flush_cost = 0;
270 hrtime_t		lgrp_plat_probe_cost = 0;
271 hrtime_t		lgrp_plat_probe_cost_total = 0;
272 
273 /*
274  * Error code for latency adjustment and verification
275  */
276 int			lgrp_plat_probe_error_code = 0;
277 
278 /*
279  * How much latencies were off from minimum values gotten
280  */
281 hrtime_t		lgrp_plat_probe_errors[MAX_NODES][MAX_NODES];
282 
283 /*
284  * Unique probe latencies and number of occurrences of each
285  */
286 lgrp_plat_latency_acct_t	lgrp_plat_probe_lat_acct[MAX_NODES];
287 
288 /*
289  * Size of memory buffer in each node for probing
290  */
291 size_t			lgrp_plat_probe_memsize = 0;
292 
293 /*
294  * Virtual address of page in each node for probing
295  */
296 caddr_t			lgrp_plat_probe_memory[MAX_NODES];
297 
298 /*
299  * Number of unique latencies in probe times
300  */
301 int			lgrp_plat_probe_nlatencies = 0;
302 
303 /*
304  * How many rounds of probing to do
305  */
306 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
307 
308 /*
309  * Number of samples to take when probing each node
310  */
311 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
312 
313 /*
314  * Number of times to read vendor ID from Northbridge for each probe.
315  */
316 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
317 
318 /*
319  * How to probe to determine lgroup topology
320  */
321 lgrp_plat_probe_op_t	lgrp_plat_probe_op = LGRP_PLAT_PROBE_VENDOR;
322 
323 /*
324  * PFN of page in each node for probing
325  */
326 pfn_t			lgrp_plat_probe_pfn[MAX_NODES];
327 
328 /*
329  * Whether probe time was suspect (ie. not within tolerance of value that it
330  * should match)
331  */
332 int			lgrp_plat_probe_suspect[MAX_NODES][MAX_NODES];
333 
334 /*
335  * How long it takes to access memory from each node
336  */
337 hrtime_t		lgrp_plat_probe_times[MAX_NODES][MAX_NODES];
338 
339 /*
340  * Min and max node memory probe times seen
341  */
342 hrtime_t		lgrp_plat_probe_time_max = 0;
343 hrtime_t		lgrp_plat_probe_time_min = -1;
344 hrtime_t		lgrp_plat_probe_max[MAX_NODES][MAX_NODES];
345 hrtime_t		lgrp_plat_probe_min[MAX_NODES][MAX_NODES];
346 
347 
348 /*
349  * Allocate lgrp and lgrp stat arrays statically.
350  */
351 static lgrp_t	lgrp_space[NLGRP];
352 static int	nlgrps_alloc;
353 
354 struct lgrp_stats lgrp_stats[NLGRP];
355 
356 /*
357  * Supported AMD processor families
358  */
359 #define	AMD_FAMILY_HAMMER	15
360 #define	AMD_FAMILY_GREYHOUND	16
361 
362 /*
363  * Whether to have is_opteron() return 1 even when processor isn't
364  * supported
365  */
366 uint_t	is_opteron_override = 0;
367 
368 /*
369  * AMD processor family for current CPU
370  */
371 uint_t	opt_family = 0;
372 
373 uint_t	opt_probe_func = OPT_PCS_FUNC_DRAM;
374 
375 
376 /*
377  * Determine whether we're running on a supported AMD Opteron since reading
378  * node count and DRAM address map registers may have different format or
379  * may not be supported in future processor families
380  */
381 int
382 is_opteron(void)
383 {
384 
385 	if (x86_vendor != X86_VENDOR_AMD)
386 		return (0);
387 
388 	opt_family = cpuid_getfamily(CPU);
389 	if (opt_family == AMD_FAMILY_HAMMER ||
390 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
391 		return (1);
392 	else
393 		return (0);
394 }
395 
396 int
397 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
398 {
399 	if (max_mem_nodes == 1)
400 		return (0);
401 
402 	return ((int)hand);
403 }
404 
405 lgrp_handle_t
406 plat_mem_node_to_lgrphand(int mnode)
407 {
408 	if (max_mem_nodes == 1)
409 		return (LGRP_DEFAULT_HANDLE);
410 
411 	return ((lgrp_handle_t)mnode);
412 }
413 
414 
415 /*
416  * plat_mnode_xcheck: checks the node memory ranges to see if there is a pfncnt
417  * range of pages aligned on pfncnt that crosses an node boundary. Returns 1 if
418  * a crossing is found and returns 0 otherwise.
419  */
420 int
421 plat_mnode_xcheck(pfn_t pfncnt)
422 {
423 	int	node, prevnode = -1, basenode;
424 	pfn_t	ea, sa;
425 
426 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
427 
428 		if (lgrp_plat_node_memory[node].exists == 0)
429 			continue;
430 
431 		if (prevnode == -1) {
432 			prevnode = node;
433 			basenode = node;
434 			continue;
435 		}
436 
437 		/* assume x86 node pfn ranges are in increasing order */
438 		ASSERT(lgrp_plat_node_memory[node].start >
439 		    lgrp_plat_node_memory[prevnode].end);
440 
441 		/*
442 		 * continue if the starting address of node is not contiguous
443 		 * with the previous node.
444 		 */
445 
446 		if (lgrp_plat_node_memory[node].start !=
447 		    (lgrp_plat_node_memory[prevnode].end + 1)) {
448 			basenode = node;
449 			prevnode = node;
450 			continue;
451 		}
452 
453 		/* check if the starting address of node is pfncnt aligned */
454 		if ((lgrp_plat_node_memory[node].start & (pfncnt - 1)) != 0) {
455 
456 			/*
457 			 * at this point, node starts at an unaligned boundary
458 			 * and is contiguous with the previous node(s) to
459 			 * basenode. Check if there is an aligned contiguous
460 			 * range of length pfncnt that crosses this boundary.
461 			 */
462 
463 			sa = P2ALIGN(lgrp_plat_node_memory[prevnode].end,
464 			    pfncnt);
465 			ea = P2ROUNDUP((lgrp_plat_node_memory[node].start),
466 			    pfncnt);
467 
468 			ASSERT((ea - sa) == pfncnt);
469 			if (sa >= lgrp_plat_node_memory[basenode].start &&
470 			    ea <= (lgrp_plat_node_memory[node].end + 1))
471 				return (1);
472 		}
473 		prevnode = node;
474 	}
475 	return (0);
476 }
477 
478 int
479 plat_pfn_to_mem_node(pfn_t pfn)
480 {
481 	int	node;
482 
483 	if (max_mem_nodes == 1)
484 		return (0);
485 
486 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
487 		/*
488 		 * Skip nodes with no memory
489 		 */
490 		if (!lgrp_plat_node_memory[node].exists)
491 			continue;
492 
493 		if (pfn >= lgrp_plat_node_memory[node].start &&
494 		    pfn <= lgrp_plat_node_memory[node].end)
495 			return (node);
496 	}
497 
498 	ASSERT(node < lgrp_plat_node_cnt);
499 	return (-1);
500 }
501 
502 /*
503  * Configure memory nodes for machines with more than one node (ie NUMA)
504  */
505 void
506 plat_build_mem_nodes(struct memlist *list)
507 {
508 	pfn_t		cur_start;	/* start addr of subrange */
509 	pfn_t		cur_end;	/* end addr of subrange */
510 	pfn_t		start;		/* start addr of whole range */
511 	pfn_t		end;		/* end addr of whole range */
512 
513 	/*
514 	 * Boot install lists are arranged <addr, len>, ...
515 	 */
516 	while (list) {
517 		int	node;
518 
519 		start = list->address >> PAGESHIFT;
520 		end = (list->address + list->size - 1) >> PAGESHIFT;
521 
522 		if (start > physmax) {
523 			list = list->next;
524 			continue;
525 		}
526 		if (end > physmax)
527 			end = physmax;
528 
529 		/*
530 		 * When there is only one memnode, just add memory to memnode
531 		 */
532 		if (max_mem_nodes == 1) {
533 			mem_node_add_slice(start, end);
534 			list = list->next;
535 			continue;
536 		}
537 
538 		/*
539 		 * mem_node_add_slice() expects to get a memory range that
540 		 * is within one memnode, so need to split any memory range
541 		 * that spans multiple memnodes into subranges that are each
542 		 * contained within one memnode when feeding them to
543 		 * mem_node_add_slice()
544 		 */
545 		cur_start = start;
546 		do {
547 			node = plat_pfn_to_mem_node(cur_start);
548 
549 			/*
550 			 * Panic if DRAM address map registers or SRAT say
551 			 * memory in node doesn't exist or address from
552 			 * boot installed memory list entry isn't in this node.
553 			 * This shouldn't happen and rest of code can't deal
554 			 * with this if it does.
555 			 */
556 			if (node < 0 || node >= lgrp_plat_node_cnt ||
557 			    !lgrp_plat_node_memory[node].exists ||
558 			    cur_start < lgrp_plat_node_memory[node].start ||
559 			    cur_start > lgrp_plat_node_memory[node].end) {
560 				cmn_err(CE_PANIC, "Don't know which memnode "
561 				    "to add installed memory address 0x%lx\n",
562 				    cur_start);
563 			}
564 
565 			/*
566 			 * End of current subrange should not span memnodes
567 			 */
568 			cur_end = end;
569 			if (lgrp_plat_node_memory[node].exists &&
570 			    cur_end > lgrp_plat_node_memory[node].end)
571 				cur_end = lgrp_plat_node_memory[node].end;
572 
573 			mem_node_add_slice(cur_start, cur_end);
574 
575 			/*
576 			 * Next subrange starts after end of current one
577 			 */
578 			cur_start = cur_end + 1;
579 		} while (cur_end < end);
580 
581 		list = list->next;
582 	}
583 	mem_node_physalign = 0;
584 	mem_node_pfn_shift = 0;
585 }
586 
587 
588 /*
589  * Platform-specific initialization of lgroups
590  */
591 void
592 lgrp_plat_init(void)
593 {
594 #if defined(__xpv)
595 	/*
596 	 * XXPV	For now, the hypervisor treats all memory equally.
597 	 */
598 	lgrp_plat_node_cnt = max_mem_nodes = 1;
599 #else	/* __xpv */
600 	uint_t		bus;
601 	uint_t		dev;
602 	uint_t		node;
603 	uint_t		off_hi;
604 	uint_t		off_lo;
605 	uint64_t	nb_cfg_reg;
606 
607 	extern lgrp_load_t	lgrp_expand_proc_thresh;
608 	extern lgrp_load_t	lgrp_expand_proc_diff;
609 
610 	/*
611 	 * Initialize as a UMA machine if this isn't an Opteron
612 	 */
613 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
614 		lgrp_plat_node_cnt = max_mem_nodes = 1;
615 		return;
616 	}
617 
618 	/*
619 	 * Read configuration registers from PCI configuration space to
620 	 * determine node information, which memory is in each node, etc.
621 	 *
622 	 * Write to PCI configuration space address register to specify
623 	 * which configuration register to read and read/write PCI
624 	 * configuration space data register to get/set contents
625 	 */
626 	bus = OPT_PCS_BUS_CONFIG;
627 	dev = OPT_PCS_DEV_NODE0;
628 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
629 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
630 
631 	/*
632 	 * Read node ID register for node 0 to get node count
633 	 */
634 	opt_node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
635 	    OPT_PCS_OFF_NODEID);
636 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
637 
638 	/*
639 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
640 	 * read high DRAM address map base and limit registers
641 	 */
642 	if (opt_family == AMD_FAMILY_GREYHOUND) {
643 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
644 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
645 			wrmsr(MSR_AMD_NB_CFG,
646 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
647 	}
648 
649 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
650 		uint32_t	base_hi;
651 		uint32_t	base_lo;
652 		uint32_t	limit_hi;
653 		uint32_t	limit_lo;
654 
655 		/*
656 		 * Read node ID register (except for node 0 which we just read)
657 		 */
658 		if (node > 0) {
659 			opt_node_info[node] = pci_getl_func(bus, dev,
660 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
661 		}
662 
663 		/*
664 		 * Read DRAM base and limit registers which specify
665 		 * physical memory range of each node
666 		 */
667 		if (opt_family != AMD_FAMILY_GREYHOUND)
668 			base_hi = 0;
669 		else {
670 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
671 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
672 			base_hi = opt_dram_map[node].base_hi =
673 			    inl(PCI_CONFDATA);
674 		}
675 		base_lo = opt_dram_map[node].base_lo = pci_getl_func(bus, dev,
676 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
677 
678 		if (opt_dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN)
679 			lgrp_plat_mem_intrlv++;
680 
681 		off_hi += 4;	/* high limit register offset */
682 		if (opt_family != AMD_FAMILY_GREYHOUND)
683 			limit_hi = 0;
684 		else {
685 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
686 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
687 			limit_hi = opt_dram_map[node].limit_hi =
688 			    inl(PCI_CONFDATA);
689 		}
690 
691 		off_lo += 4;	/* low limit register offset */
692 		limit_lo = opt_dram_map[node].limit_lo = pci_getl_func(bus,
693 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
694 
695 		/*
696 		 * Increment device number to next node and register offsets
697 		 * for DRAM base register of next node
698 		 */
699 		off_hi += 4;
700 		off_lo += 4;
701 		dev++;
702 
703 		/*
704 		 * Both read and write enable bits must be enabled in DRAM
705 		 * address map base register for physical memory to exist in
706 		 * node
707 		 */
708 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
709 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
710 			/*
711 			 * Mark node memory as non-existent and set start and
712 			 * end addresses to be same in lgrp_plat_node_memory[]
713 			 */
714 			lgrp_plat_node_memory[node].exists = 0;
715 			lgrp_plat_node_memory[node].start =
716 			    lgrp_plat_node_memory[node].end = (pfn_t)-1;
717 			continue;
718 		}
719 
720 		/*
721 		 * Get PFN for first page in each node,
722 		 * so we can probe memory to determine latency topology
723 		 */
724 		lgrp_plat_probe_pfn[node] =
725 		    btop(OPT_DRAMADDR(base_hi, base_lo));
726 
727 		/*
728 		 * Mark node memory as existing and remember physical address
729 		 * range of each node for use later
730 		 */
731 		lgrp_plat_node_memory[node].exists = 1;
732 
733 		lgrp_plat_node_memory[node].start =
734 		    btop(OPT_DRAMADDR(base_hi, base_lo));
735 
736 		lgrp_plat_node_memory[node].end =
737 		    btop(OPT_DRAMADDR(limit_hi, limit_lo) |
738 		    OPT_DRAMADDR_LO_MASK_OFF);
739 	}
740 
741 	/*
742 	 * Restore PCI Extended Configuration Space enable bit
743 	 */
744 	if (opt_family == AMD_FAMILY_GREYHOUND) {
745 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
746 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
747 	}
748 
749 	/*
750 	 * Only use one memory node if memory is interleaved between any nodes
751 	 */
752 	if (lgrp_plat_mem_intrlv) {
753 		lgrp_plat_node_cnt = max_mem_nodes = 1;
754 		(void) lgrp_topo_ht_limit_set(1);
755 	} else {
756 		max_mem_nodes = lgrp_plat_node_cnt;
757 
758 		/*
759 		 * Probing errors can mess up the lgroup topology and force us
760 		 * fall back to a 2 level lgroup topology.  Here we bound how
761 		 * tall the lgroup topology can grow in hopes of avoiding any
762 		 * anamolies in probing from messing up the lgroup topology
763 		 * by limiting the accuracy of the latency topology.
764 		 *
765 		 * Assume that nodes will at least be configured in a ring,
766 		 * so limit height of lgroup topology to be less than number
767 		 * of nodes on a system with 4 or more nodes
768 		 */
769 		if (lgrp_plat_node_cnt >= 4 &&
770 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
771 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
772 	}
773 
774 	/*
775 	 * Lgroups on Opteron architectures have but a single physical
776 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
777 	 * so that lgrp_choose() will spread things out aggressively.
778 	 */
779 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
780 	lgrp_expand_proc_diff = 0;
781 #endif	/* __xpv */
782 }
783 
784 
785 /*
786  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
787  * be considered same
788  */
789 #define	LGRP_LAT_TOLERANCE_SHIFT	4
790 
791 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
792 
793 
794 /*
795  * Adjust latencies between nodes to be symmetric, normalize latencies between
796  * any nodes that are within some tolerance to be same, and make local
797  * latencies be same
798  */
799 static void
800 lgrp_plat_latency_adjust(void)
801 {
802 	int				i;
803 	int				j;
804 	int				k;
805 	int				l;
806 	u_longlong_t			max;
807 	u_longlong_t			min;
808 	u_longlong_t			t;
809 	u_longlong_t			t1;
810 	u_longlong_t			t2;
811 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
812 	int				lat_corrected[MAX_NODES][MAX_NODES];
813 
814 	/*
815 	 * Nothing to do when this is an UMA machine
816 	 */
817 	if (max_mem_nodes == 1)
818 		return;
819 
820 	/*
821 	 * Make sure that latencies are symmetric between any two nodes
822 	 * (ie. latency(node0, node1) == latency(node1, node0))
823 	 */
824 	for (i = 0; i < lgrp_plat_node_cnt; i++)
825 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
826 			t1 = lgrp_plat_probe_times[i][j];
827 			t2 = lgrp_plat_probe_times[j][i];
828 
829 			if (t1 == 0 || t2 == 0 || t1 == t2)
830 				continue;
831 
832 			/*
833 			 * Latencies should be same
834 			 * - Use minimum of two latencies which should be same
835 			 * - Track suspect probe times not within tolerance of
836 			 *   min value
837 			 * - Remember how much values are corrected by
838 			 */
839 			if (t1 > t2) {
840 				t = t2;
841 				lgrp_plat_probe_errors[i][j] += t1 - t2;
842 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
843 					lgrp_plat_probe_suspect[i][j]++;
844 					lgrp_plat_probe_suspect[j][i]++;
845 				}
846 			} else if (t2 > t1) {
847 				t = t1;
848 				lgrp_plat_probe_errors[j][i] += t2 - t1;
849 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
850 					lgrp_plat_probe_suspect[i][j]++;
851 					lgrp_plat_probe_suspect[j][i]++;
852 				}
853 			}
854 
855 			lgrp_plat_probe_times[i][j] =
856 			    lgrp_plat_probe_times[j][i] = t;
857 			lgrp_config(cflag, t1, t);
858 			lgrp_config(cflag, t2, t);
859 		}
860 
861 	/*
862 	 * Keep track of which latencies get corrected
863 	 */
864 	for (i = 0; i < MAX_NODES; i++)
865 		for (j = 0; j < MAX_NODES; j++)
866 			lat_corrected[i][j] = 0;
867 
868 	/*
869 	 * For every two nodes, see whether there is another pair of nodes which
870 	 * are about the same distance apart and make the latencies be the same
871 	 * if they are close enough together
872 	 */
873 	for (i = 0; i < lgrp_plat_node_cnt; i++)
874 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
875 			/*
876 			 * Pick one pair of nodes (i, j)
877 			 * and get latency between them
878 			 */
879 			t1 = lgrp_plat_probe_times[i][j];
880 
881 			/*
882 			 * Skip this pair of nodes if there isn't a latency
883 			 * for it yet
884 			 */
885 			if (t1 == 0)
886 				continue;
887 
888 			for (k = 0; k < lgrp_plat_node_cnt; k++)
889 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
890 					/*
891 					 * Pick another pair of nodes (k, l)
892 					 * not same as (i, j) and get latency
893 					 * between them
894 					 */
895 					if (k == i && l == j)
896 						continue;
897 
898 					t2 = lgrp_plat_probe_times[k][l];
899 
900 					/*
901 					 * Skip this pair of nodes if there
902 					 * isn't a latency for it yet
903 					 */
904 
905 					if (t2 == 0)
906 						continue;
907 
908 					/*
909 					 * Skip nodes (k, l) if they already
910 					 * have same latency as (i, j) or
911 					 * their latency isn't close enough to
912 					 * be considered/made the same
913 					 */
914 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
915 					    t1 >> lgrp_plat_probe_lt_shift) ||
916 					    (t2 > t1 && t2 - t1 >
917 					    t2 >> lgrp_plat_probe_lt_shift))
918 						continue;
919 
920 					/*
921 					 * Make latency(i, j) same as
922 					 * latency(k, l), try to use latency
923 					 * that has been adjusted already to get
924 					 * more consistency (if possible), and
925 					 * remember which latencies were
926 					 * adjusted for next time
927 					 */
928 					if (lat_corrected[i][j]) {
929 						t = t1;
930 						lgrp_config(cflag, t2, t);
931 						t2 = t;
932 					} else if (lat_corrected[k][l]) {
933 						t = t2;
934 						lgrp_config(cflag, t1, t);
935 						t1 = t;
936 					} else {
937 						if (t1 > t2)
938 							t = t2;
939 						else
940 							t = t1;
941 						lgrp_config(cflag, t1, t);
942 						lgrp_config(cflag, t2, t);
943 						t1 = t2 = t;
944 					}
945 
946 					lgrp_plat_probe_times[i][j] =
947 					    lgrp_plat_probe_times[k][l] = t;
948 
949 					lat_corrected[i][j] =
950 					    lat_corrected[k][l] = 1;
951 				}
952 		}
953 
954 	/*
955 	 * Local latencies should be same
956 	 * - Find min and max local latencies
957 	 * - Make all local latencies be minimum
958 	 */
959 	min = -1;
960 	max = 0;
961 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
962 		t = lgrp_plat_probe_times[i][i];
963 		if (t == 0)
964 			continue;
965 		if (min == -1 || t < min)
966 			min = t;
967 		if (t > max)
968 			max = t;
969 	}
970 	if (min != max) {
971 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
972 			int	local;
973 
974 			local = lgrp_plat_probe_times[i][i];
975 			if (local == 0)
976 				continue;
977 
978 			/*
979 			 * Track suspect probe times that aren't within
980 			 * tolerance of minimum local latency and how much
981 			 * probe times are corrected by
982 			 */
983 			if (local - min > min >> lgrp_plat_probe_lt_shift)
984 				lgrp_plat_probe_suspect[i][i]++;
985 
986 			lgrp_plat_probe_errors[i][i] += local - min;
987 
988 			/*
989 			 * Make local latencies be minimum
990 			 */
991 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
992 			lgrp_plat_probe_times[i][i] = min;
993 		}
994 	}
995 
996 	/*
997 	 * Determine max probe time again since just adjusted latencies
998 	 */
999 	lgrp_plat_probe_time_max = 0;
1000 	for (i = 0; i < lgrp_plat_node_cnt; i++)
1001 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1002 			t = lgrp_plat_probe_times[i][j];
1003 			if (t > lgrp_plat_probe_time_max)
1004 				lgrp_plat_probe_time_max = t;
1005 		}
1006 }
1007 
1008 
1009 /*
1010  * Verify following about latencies between nodes:
1011  *
1012  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
1013  * - Local latencies same
1014  * - Local < remote
1015  * - Number of latencies seen is reasonable
1016  * - Number of occurrences of a given latency should be more than 1
1017  *
1018  * Returns:
1019  *	0	Success
1020  *	-1	Not symmetric
1021  *	-2	Local latencies not same
1022  *	-3	Local >= remote
1023  *	-4	Wrong number of latencies
1024  *	-5	Not enough occurrences of given latency
1025  */
1026 static int
1027 lgrp_plat_latency_verify(void)
1028 {
1029 	int				i;
1030 	int				j;
1031 	lgrp_plat_latency_acct_t	*l;
1032 	int				probed;
1033 	u_longlong_t			t1;
1034 	u_longlong_t			t2;
1035 
1036 	/*
1037 	 * Nothing to do when this is an UMA machine, lgroup topology is
1038 	 * limited to 2 levels, or there aren't any probe times yet
1039 	 */
1040 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
1041 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
1042 		return (0);
1043 
1044 	/*
1045 	 * Make sure that latencies are symmetric between any two nodes
1046 	 * (ie. latency(node0, node1) == latency(node1, node0))
1047 	 */
1048 	for (i = 0; i < lgrp_plat_node_cnt; i++)
1049 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1050 			t1 = lgrp_plat_probe_times[i][j];
1051 			t2 = lgrp_plat_probe_times[j][i];
1052 
1053 			if (t1 == 0 || t2 == 0 || t1 == t2)
1054 				continue;
1055 
1056 			return (-1);
1057 		}
1058 
1059 	/*
1060 	 * Local latencies should be same
1061 	 */
1062 	t1 = lgrp_plat_probe_times[0][0];
1063 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
1064 		t2 = lgrp_plat_probe_times[i][i];
1065 		if (t2 == 0)
1066 			continue;
1067 
1068 		if (t1 == 0) {
1069 			t1 = t2;
1070 			continue;
1071 		}
1072 
1073 		if (t1 != t2)
1074 			return (-2);
1075 	}
1076 
1077 	/*
1078 	 * Local latencies should be less than remote
1079 	 */
1080 	if (t1) {
1081 		for (i = 0; i < lgrp_plat_node_cnt; i++)
1082 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1083 				t2 = lgrp_plat_probe_times[i][j];
1084 				if (i == j || t2 == 0)
1085 					continue;
1086 
1087 				if (t1 >= t2)
1088 					return (-3);
1089 			}
1090 	}
1091 
1092 	/*
1093 	 * Rest of checks are not very useful for machines with less than
1094 	 * 4 nodes (which means less than 3 latencies on Opteron)
1095 	 */
1096 	if (lgrp_plat_node_cnt < 4)
1097 		return (0);
1098 
1099 	/*
1100 	 * Need to see whether done probing in order to verify number of
1101 	 * latencies are correct
1102 	 */
1103 	probed = 0;
1104 	for (i = 0; i < lgrp_plat_node_cnt; i++)
1105 		if (lgrp_plat_probe_times[i][i])
1106 			probed++;
1107 
1108 	if (probed != lgrp_plat_node_cnt)
1109 		return (0);
1110 
1111 	/*
1112 	 * Determine number of unique latencies seen in probe times,
1113 	 * their values, and number of occurrences of each
1114 	 */
1115 	lgrp_plat_probe_nlatencies = 0;
1116 	bzero(lgrp_plat_probe_lat_acct,
1117 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
1118 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1119 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1120 			int	k;
1121 
1122 			/*
1123 			 * Look at each probe time
1124 			 */
1125 			t1 = lgrp_plat_probe_times[i][j];
1126 			if (t1 == 0)
1127 				continue;
1128 
1129 			/*
1130 			 * Account for unique latencies
1131 			 */
1132 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1133 				l = &lgrp_plat_probe_lat_acct[k];
1134 				if (t1 == l->la_value) {
1135 					/*
1136 					 * Increment number of occurrences
1137 					 * if seen before
1138 					 */
1139 					l->la_count++;
1140 					break;
1141 				} else if (l->la_value == 0) {
1142 					/*
1143 					 * Record latency if haven't seen before
1144 					 */
1145 					l->la_value = t1;
1146 					l->la_count++;
1147 					lgrp_plat_probe_nlatencies++;
1148 					break;
1149 				}
1150 			}
1151 		}
1152 	}
1153 
1154 	/*
1155 	 * Number of latencies should be relative to number of
1156 	 * nodes in system:
1157 	 * - Same as nodes when nodes <= 2
1158 	 * - Less than nodes when nodes > 2
1159 	 * - Greater than 2 when nodes >= 4
1160 	 */
1161 	if ((lgrp_plat_node_cnt <= 2 &&
1162 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
1163 	    (lgrp_plat_node_cnt > 2 &&
1164 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
1165 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
1166 	    lgrp_plat_probe_nlatencies <= 2))
1167 		return (-4);
1168 
1169 	/*
1170 	 * There should be more than one occurrence of every latency
1171 	 * as long as probing is complete
1172 	 */
1173 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
1174 		l = &lgrp_plat_probe_lat_acct[i];
1175 		if (l->la_count <= 1)
1176 			return (-5);
1177 	}
1178 	return (0);
1179 }
1180 
1181 
1182 /*
1183  * Set lgroup latencies for 2 level lgroup topology
1184  */
1185 static void
1186 lgrp_plat_2level_setup(void)
1187 {
1188 	int	i;
1189 
1190 	if (lgrp_plat_node_cnt >= 4)
1191 		cmn_err(CE_NOTE,
1192 		    "MPO only optimizing for local and remote\n");
1193 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1194 		int	j;
1195 
1196 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1197 			if (i == j)
1198 				lgrp_plat_probe_times[i][j] = 2;
1199 			else
1200 				lgrp_plat_probe_times[i][j] = 3;
1201 		}
1202 	}
1203 	lgrp_plat_probe_time_min = 2;
1204 	lgrp_plat_probe_time_max = 3;
1205 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1206 }
1207 
1208 
1209 /*
1210  * Return time needed to probe from current CPU to memory in given node
1211  */
1212 static hrtime_t
1213 lgrp_plat_probe_time(int to)
1214 {
1215 	caddr_t		buf;
1216 	uint_t		dev;
1217 	/* LINTED: set but not used in function */
1218 	volatile uint_t	dev_vendor;
1219 	hrtime_t	elapsed;
1220 	hrtime_t	end;
1221 	int		from;
1222 	int		i;
1223 	int		ipl;
1224 	hrtime_t	max;
1225 	hrtime_t	min;
1226 	hrtime_t	start;
1227 	int		cnt;
1228 	extern int	use_sse_pagecopy;
1229 
1230 	/*
1231 	 * Determine ID of node containing current CPU
1232 	 */
1233 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1234 
1235 	/*
1236 	 * Do common work for probing main memory
1237 	 */
1238 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1239 		/*
1240 		 * Skip probing any nodes without memory and
1241 		 * set probe time to 0
1242 		 */
1243 		if (lgrp_plat_probe_memory[to] == NULL) {
1244 			lgrp_plat_probe_times[from][to] = 0;
1245 			return (0);
1246 		}
1247 
1248 		/*
1249 		 * Invalidate caches once instead of once every sample
1250 		 * which should cut cost of probing by a lot
1251 		 */
1252 		lgrp_plat_flush_cost = gethrtime();
1253 		invalidate_cache();
1254 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1255 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1256 	}
1257 
1258 	/*
1259 	 * Probe from current CPU to given memory using specified operation
1260 	 * and take specified number of samples
1261 	 */
1262 	max = 0;
1263 	min = -1;
1264 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1265 		lgrp_plat_probe_cost = gethrtime();
1266 
1267 		/*
1268 		 * Can't measure probe time if gethrtime() isn't working yet
1269 		 */
1270 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1271 			return (0);
1272 
1273 		switch (lgrp_plat_probe_op) {
1274 
1275 		case LGRP_PLAT_PROBE_PGCPY:
1276 		default:
1277 			/*
1278 			 * Measure how long it takes to copy page
1279 			 * on top of itself
1280 			 */
1281 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1282 
1283 			kpreempt_disable();
1284 			ipl = splhigh();
1285 			start = gethrtime();
1286 			if (use_sse_pagecopy)
1287 				hwblkpagecopy(buf, buf);
1288 			else
1289 				bcopy(buf, buf, PAGESIZE);
1290 			end = gethrtime();
1291 			elapsed = end - start;
1292 			splx(ipl);
1293 			kpreempt_enable();
1294 			break;
1295 
1296 		case LGRP_PLAT_PROBE_VENDOR:
1297 			/*
1298 			 * Measure how long it takes to read vendor ID from
1299 			 * Northbridge
1300 			 */
1301 			dev = OPT_PCS_DEV_NODE0 + to;
1302 			kpreempt_disable();
1303 			ipl = spl8();
1304 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1305 			    OPT_PCS_OFF_VENDOR));
1306 			start = gethrtime();
1307 			for (cnt = 0; cnt < lgrp_plat_probe_nreads; cnt++)
1308 				dev_vendor = inl(PCI_CONFDATA);
1309 			end = gethrtime();
1310 			elapsed = (end - start) / lgrp_plat_probe_nreads;
1311 			splx(ipl);
1312 			kpreempt_enable();
1313 			break;
1314 		}
1315 
1316 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1317 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1318 
1319 		if (min == -1 || elapsed < min)
1320 			min = elapsed;
1321 		if (elapsed > max)
1322 			max = elapsed;
1323 	}
1324 
1325 	/*
1326 	 * Update minimum and maximum probe times between
1327 	 * these two nodes
1328 	 */
1329 	if (min < lgrp_plat_probe_min[from][to] ||
1330 	    lgrp_plat_probe_min[from][to] == 0)
1331 		lgrp_plat_probe_min[from][to] = min;
1332 
1333 	if (max > lgrp_plat_probe_max[from][to])
1334 		lgrp_plat_probe_max[from][to] = max;
1335 
1336 	return (min);
1337 }
1338 
1339 
1340 /*
1341  * Probe memory in each node from current CPU to determine latency topology
1342  */
1343 void
1344 lgrp_plat_probe(void)
1345 {
1346 	int		from;
1347 	int		i;
1348 	hrtime_t	probe_time;
1349 	int		to;
1350 
1351 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1352 		return;
1353 
1354 	/*
1355 	 * Determine ID of node containing current CPU
1356 	 */
1357 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1358 
1359 	/*
1360 	 * Don't need to probe if got times already
1361 	 */
1362 	if (lgrp_plat_probe_times[from][from] != 0)
1363 		return;
1364 
1365 	/*
1366 	 * Read vendor ID in Northbridge or read and write page(s)
1367 	 * in each node from current CPU and remember how long it takes,
1368 	 * so we can build latency topology of machine later.
1369 	 * This should approximate the memory latency between each node.
1370 	 */
1371 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1372 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1373 			/*
1374 			 * Get probe time and bail out if can't get it yet
1375 			 */
1376 			probe_time = lgrp_plat_probe_time(to);
1377 			if (probe_time == 0)
1378 				return;
1379 
1380 			/*
1381 			 * Keep lowest probe time as latency between nodes
1382 			 */
1383 			if (lgrp_plat_probe_times[from][to] == 0 ||
1384 			    probe_time < lgrp_plat_probe_times[from][to])
1385 				lgrp_plat_probe_times[from][to] = probe_time;
1386 
1387 			/*
1388 			 * Update overall minimum and maximum probe times
1389 			 * across all nodes
1390 			 */
1391 			if (probe_time < lgrp_plat_probe_time_min ||
1392 			    lgrp_plat_probe_time_min == -1)
1393 				lgrp_plat_probe_time_min = probe_time;
1394 			if (probe_time > lgrp_plat_probe_time_max)
1395 				lgrp_plat_probe_time_max = probe_time;
1396 		}
1397 
1398 	/*
1399 	 * - Fix up latencies such that local latencies are same,
1400 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1401 	 *
1402 	 * - Verify that latencies look ok
1403 	 *
1404 	 * - Fallback to just optimizing for local and remote if
1405 	 *   latencies didn't look right
1406 	 */
1407 	lgrp_plat_latency_adjust();
1408 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1409 	if (lgrp_plat_probe_error_code)
1410 		lgrp_plat_2level_setup();
1411 }
1412 
1413 
1414 /*
1415  * Platform-specific initialization
1416  */
1417 void
1418 lgrp_plat_main_init(void)
1419 {
1420 	int	curnode;
1421 	int	ht_limit;
1422 	int	i;
1423 
1424 	/*
1425 	 * Print a notice that MPO is disabled when memory is interleaved
1426 	 * across nodes....Would do this when it is discovered, but can't
1427 	 * because it happens way too early during boot....
1428 	 */
1429 	if (lgrp_plat_mem_intrlv)
1430 		cmn_err(CE_NOTE,
1431 		    "MPO disabled because memory is interleaved\n");
1432 
1433 	/*
1434 	 * Don't bother to do any probing if there is only one node or the
1435 	 * height of the lgroup topology less than or equal to 2
1436 	 */
1437 	ht_limit = lgrp_topo_ht_limit();
1438 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1439 		/*
1440 		 * Setup lgroup latencies for 2 level lgroup topology
1441 		 * (ie. local and remote only) if they haven't been set yet
1442 		 */
1443 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1444 		    lgrp_plat_probe_time_max == 0)
1445 			lgrp_plat_2level_setup();
1446 		return;
1447 	}
1448 
1449 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1450 		/*
1451 		 * Should have been able to probe from CPU 0 when it was added
1452 		 * to lgroup hierarchy, but may not have been able to then
1453 		 * because it happens so early in boot that gethrtime() hasn't
1454 		 * been initialized.  (:-(
1455 		 */
1456 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1457 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1458 			lgrp_plat_probe();
1459 
1460 		return;
1461 	}
1462 
1463 	/*
1464 	 * When probing memory, use one page for every sample to determine
1465 	 * lgroup topology and taking multiple samples
1466 	 */
1467 	if (lgrp_plat_probe_memsize == 0)
1468 		lgrp_plat_probe_memsize = PAGESIZE *
1469 		    lgrp_plat_probe_nsamples;
1470 
1471 	/*
1472 	 * Map memory in each node needed for probing to determine latency
1473 	 * topology
1474 	 */
1475 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1476 		int	mnode;
1477 
1478 		/*
1479 		 * Skip this node and leave its probe page NULL
1480 		 * if it doesn't have any memory
1481 		 */
1482 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1483 		if (!mem_node_config[mnode].exists) {
1484 			lgrp_plat_probe_memory[i] = NULL;
1485 			continue;
1486 		}
1487 
1488 		/*
1489 		 * Allocate one kernel virtual page
1490 		 */
1491 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1492 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1493 		if (lgrp_plat_probe_memory[i] == NULL) {
1494 			cmn_err(CE_WARN,
1495 			    "lgrp_plat_main_init: couldn't allocate memory");
1496 			return;
1497 		}
1498 
1499 		/*
1500 		 * Map virtual page to first page in node
1501 		 */
1502 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1503 		    lgrp_plat_probe_memsize,
1504 		    lgrp_plat_probe_pfn[i],
1505 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1506 		    HAT_LOAD_NOCONSIST);
1507 	}
1508 
1509 	/*
1510 	 * Probe from current CPU
1511 	 */
1512 	lgrp_plat_probe();
1513 }
1514 
1515 /*
1516  * Allocate additional space for an lgroup.
1517  */
1518 /* ARGSUSED */
1519 lgrp_t *
1520 lgrp_plat_alloc(lgrp_id_t lgrpid)
1521 {
1522 	lgrp_t *lgrp;
1523 
1524 	lgrp = &lgrp_space[nlgrps_alloc++];
1525 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1526 		return (NULL);
1527 	return (lgrp);
1528 }
1529 
1530 /*
1531  * Platform handling for (re)configuration changes
1532  */
1533 /* ARGSUSED */
1534 void
1535 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1536 {
1537 }
1538 
1539 /*
1540  * Return the platform handle for the lgroup containing the given CPU
1541  */
1542 /* ARGSUSED */
1543 lgrp_handle_t
1544 lgrp_plat_cpu_to_hand(processorid_t id)
1545 {
1546 	if (lgrp_plat_node_cnt == 1)
1547 		return (LGRP_DEFAULT_HANDLE);
1548 
1549 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1550 }
1551 
1552 /*
1553  * Return the platform handle of the lgroup that contains the physical memory
1554  * corresponding to the given page frame number
1555  */
1556 /* ARGSUSED */
1557 lgrp_handle_t
1558 lgrp_plat_pfn_to_hand(pfn_t pfn)
1559 {
1560 	int	mnode;
1561 
1562 	if (max_mem_nodes == 1)
1563 		return (LGRP_DEFAULT_HANDLE);
1564 
1565 	if (pfn > physmax)
1566 		return (LGRP_NULL_HANDLE);
1567 
1568 	mnode = plat_pfn_to_mem_node(pfn);
1569 	if (mnode < 0)
1570 		return (LGRP_NULL_HANDLE);
1571 
1572 	return (MEM_NODE_2_LGRPHAND(mnode));
1573 }
1574 
1575 /*
1576  * Return the maximum number of lgrps supported by the platform.
1577  * Before lgrp topology is known it returns an estimate based on the number of
1578  * nodes. Once topology is known it returns the actual maximim number of lgrps
1579  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1580  * may not grow during system lifetime.
1581  */
1582 int
1583 lgrp_plat_max_lgrps()
1584 {
1585 	return (lgrp_topo_initialized ?
1586 	    lgrp_alloc_max + 1 :
1587 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1588 }
1589 
1590 /*
1591  * Return the number of free, allocatable, or installed
1592  * pages in an lgroup
1593  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1594  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1595  */
1596 /* ARGSUSED */
1597 static pgcnt_t
1598 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1599 {
1600 	struct memlist *mlist;
1601 	pgcnt_t npgs = 0;
1602 	extern struct memlist *phys_avail;
1603 	extern struct memlist *phys_install;
1604 
1605 	switch (query) {
1606 	case LGRP_MEM_SIZE_FREE:
1607 		return ((pgcnt_t)freemem);
1608 	case LGRP_MEM_SIZE_AVAIL:
1609 		memlist_read_lock();
1610 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1611 			npgs += btop(mlist->size);
1612 		memlist_read_unlock();
1613 		return (npgs);
1614 	case LGRP_MEM_SIZE_INSTALL:
1615 		memlist_read_lock();
1616 		for (mlist = phys_install; mlist; mlist = mlist->next)
1617 			npgs += btop(mlist->size);
1618 		memlist_read_unlock();
1619 		return (npgs);
1620 	default:
1621 		return ((pgcnt_t)0);
1622 	}
1623 }
1624 
1625 /*
1626  * Return the number of free pages in an lgroup.
1627  *
1628  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1629  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1630  * number of allocatable base pagesize pages corresponding to the
1631  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1632  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1633  * memory installed, regardless of whether or not it's usable.
1634  */
1635 pgcnt_t
1636 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1637 {
1638 	int	mnode;
1639 	pgcnt_t npgs = (pgcnt_t)0;
1640 	extern struct memlist *phys_avail;
1641 	extern struct memlist *phys_install;
1642 
1643 
1644 	if (plathand == LGRP_DEFAULT_HANDLE)
1645 		return (lgrp_plat_mem_size_default(plathand, query));
1646 
1647 	if (plathand != LGRP_NULL_HANDLE) {
1648 		mnode = plat_lgrphand_to_mem_node(plathand);
1649 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1650 			switch (query) {
1651 			case LGRP_MEM_SIZE_FREE:
1652 				npgs = MNODE_PGCNT(mnode);
1653 				break;
1654 			case LGRP_MEM_SIZE_AVAIL:
1655 				npgs = mem_node_memlist_pages(mnode,
1656 				    phys_avail);
1657 				break;
1658 			case LGRP_MEM_SIZE_INSTALL:
1659 				npgs = mem_node_memlist_pages(mnode,
1660 				    phys_install);
1661 				break;
1662 			default:
1663 				break;
1664 			}
1665 		}
1666 	}
1667 	return (npgs);
1668 }
1669 
1670 /*
1671  * Return latency between "from" and "to" lgroups
1672  *
1673  * This latency number can only be used for relative comparison
1674  * between lgroups on the running system, cannot be used across platforms,
1675  * and may not reflect the actual latency.  It is platform and implementation
1676  * specific, so platform gets to decide its value.  It would be nice if the
1677  * number was at least proportional to make comparisons more meaningful though.
1678  */
1679 /* ARGSUSED */
1680 int
1681 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1682 {
1683 	lgrp_handle_t	src, dest;
1684 
1685 	if (max_mem_nodes == 1)
1686 		return (0);
1687 
1688 	/*
1689 	 * Return max latency for root lgroup
1690 	 */
1691 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1692 		return (lgrp_plat_probe_time_max);
1693 
1694 	src = from;
1695 	dest = to;
1696 
1697 	/*
1698 	 * Return 0 for nodes (lgroup platform handles) out of range
1699 	 */
1700 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1701 		return (0);
1702 
1703 	/*
1704 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1705 	 * and we are trying to get latency from current CPU to some node
1706 	 */
1707 	if (lgrp_plat_probe_times[src][src] == 0 &&
1708 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1709 		lgrp_plat_probe();
1710 
1711 	return (lgrp_plat_probe_times[src][dest]);
1712 }
1713 
1714 /*
1715  * Return platform handle for root lgroup
1716  */
1717 lgrp_handle_t
1718 lgrp_plat_root_hand(void)
1719 {
1720 	return (LGRP_DEFAULT_HANDLE);
1721 }
1722