xref: /illumos-gate/usr/src/uts/i86pc/os/lgrpplat.c (revision 1a578a15d3f76161f037cd99883a1f54a9eda785)
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 2007 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 <vm/hat_i86.h>
51 #include <vm/seg_kmem.h>
52 #include <vm/vm_dep.h>
53 
54 
55 /*
56  * lgroup platform support for x86 platforms.
57  */
58 
59 #define	MAX_NODES		8
60 #define	NLGRP			(MAX_NODES * (MAX_NODES - 1) + 1)
61 
62 #define	LGRP_PLAT_CPU_TO_NODE(cpu) (pg_plat_hw_instance_id(cpu, PGHW_CHIP))
63 
64 #define	LGRP_PLAT_PROBE_NROUNDS		64	/* default laps for probing */
65 #define	LGRP_PLAT_PROBE_NSAMPLES	1	/* default samples to take */
66 #define	LGRP_PLAT_PROBE_NREADS		256	/* number of vendor ID reads */
67 
68 /*
69  * Multiprocessor Opteron machines have Non Uniform Memory Access (NUMA).
70  *
71  * Until this code supports reading System Resource Affinity Table (SRAT),
72  * we need to examine registers in PCI configuration space to determine how
73  * many nodes are in the system and which CPUs and memory are in each node.
74  * This could be determined by probing all memory from each CPU, but that is
75  * too expensive to do while booting the kernel.
76  *
77  * NOTE: Using these PCI configuration space registers to determine this
78  *       locality info is not guaranteed to work on future generations of
79  *	 Opteron processor.
80  */
81 
82 /*
83  * Opteron DRAM Address Map in PCI configuration space gives base and limit
84  * of physical memory in each node.  The following constants and macros define
85  * their contents, structure, and access.
86  */
87 
88 /*
89  * How many bits to shift Opteron DRAM Address Map base and limit registers
90  * to get actual value
91  */
92 #define	OPT_DRAMADDR_HI_LSHIFT_ADDR	40	/* shift left for address */
93 #define	OPT_DRAMADDR_LO_LSHIFT_ADDR	8	/* shift left for address */
94 
95 #define	OPT_DRAMADDR_HI_MASK_ADDR	0x000000FF /* address bits 47-40 */
96 #define	OPT_DRAMADDR_LO_MASK_ADDR	0xFFFF0000 /* address bits 39-24 */
97 
98 #define	OPT_DRAMADDR_LO_MASK_OFF	0xFFFFFF /* offset for address */
99 
100 /*
101  * Macros to derive addresses from Opteron DRAM Address Map registers
102  */
103 #define	OPT_DRAMADDR_HI(reg) \
104 	(((u_longlong_t)reg & OPT_DRAMADDR_HI_MASK_ADDR) << \
105 	    OPT_DRAMADDR_HI_LSHIFT_ADDR)
106 
107 #define	OPT_DRAMADDR_LO(reg) \
108 	(((u_longlong_t)reg & OPT_DRAMADDR_LO_MASK_ADDR) << \
109 	    OPT_DRAMADDR_LO_LSHIFT_ADDR)
110 
111 #define	OPT_DRAMADDR(high, low) \
112 	(OPT_DRAMADDR_HI(high) | OPT_DRAMADDR_LO(low))
113 
114 /*
115  * Bit masks defining what's in Opteron DRAM Address Map base register
116  */
117 #define	OPT_DRAMBASE_LO_MASK_RE		0x1	/* read enable */
118 #define	OPT_DRAMBASE_LO_MASK_WE		0x2	/* write enable */
119 #define	OPT_DRAMBASE_LO_MASK_INTRLVEN	0x700	/* interleave */
120 
121 /*
122  * Bit masks defining what's in Opteron DRAM Address Map limit register
123  */
124 #define	OPT_DRAMLIMIT_LO_MASK_DSTNODE	0x7		/* destination node */
125 #define	OPT_DRAMLIMIT_LO_MASK_INTRLVSEL	0x700		/* interleave select */
126 
127 
128 /*
129  * Opteron Node ID register in PCI configuration space contains
130  * number of nodes in system, etc. for Opteron K8.  The following
131  * constants and macros define its contents, structure, and access.
132  */
133 
134 /*
135  * Bit masks defining what's in Opteron Node ID register
136  */
137 #define	OPT_NODE_MASK_ID	0x7	/* node ID */
138 #define	OPT_NODE_MASK_CNT	0x70	/* node count */
139 #define	OPT_NODE_MASK_IONODE	0x700	/* Hypertransport I/O hub node ID */
140 #define	OPT_NODE_MASK_LCKNODE	0x7000	/* lock controller node ID */
141 #define	OPT_NODE_MASK_CPUCNT	0xF0000	/* CPUs in system (0 means 1 CPU)  */
142 
143 /*
144  * How many bits in Opteron Node ID register to shift right to get actual value
145  */
146 #define	OPT_NODE_RSHIFT_CNT	0x4	/* shift right for node count value */
147 
148 /*
149  * Macros to get values from Opteron Node ID register
150  */
151 #define	OPT_NODE_CNT(reg) \
152 	((reg & OPT_NODE_MASK_CNT) >> OPT_NODE_RSHIFT_CNT)
153 
154 /*
155  * Macro to setup PCI Extended Configuration Space (ECS) address to give to
156  * "in/out" instructions
157  *
158  * NOTE: Should only be used in lgrp_plat_init() before MMIO setup because any
159  *	 other uses should just do MMIO to access PCI ECS.
160  *	 Must enable special bit in Northbridge Configuration Register on
161  *	 Greyhound for extended CF8 space access to be able to access PCI ECS
162  *	 using "in/out" instructions and restore special bit after done
163  *	 accessing PCI ECS.
164  */
165 #define	OPT_PCI_ECS_ADDR(bus, device, function, reg) \
166 	(PCI_CONE | (((bus) & 0xff) << 16) | (((device & 0x1f)) << 11)  | \
167 	    (((function) & 0x7) << 8) | ((reg) & 0xfc) | \
168 	    ((((reg) >> 8) & 0xf) << 24))
169 
170 /*
171  * PCI configuration space registers accessed by specifying
172  * a bus, device, function, and offset.  The following constants
173  * define the values needed to access Opteron K8 configuration
174  * info to determine its node topology
175  */
176 
177 #define	OPT_PCS_BUS_CONFIG	0	/* Hypertransport config space bus */
178 
179 /*
180  * Opteron PCI configuration space register function values
181  */
182 #define	OPT_PCS_FUNC_HT		0	/* Hypertransport configuration */
183 #define	OPT_PCS_FUNC_ADDRMAP	1	/* Address map configuration */
184 #define	OPT_PCS_FUNC_DRAM	2	/* DRAM configuration */
185 #define	OPT_PCS_FUNC_MISC	3	/* Miscellaneous configuration */
186 
187 /*
188  * PCI Configuration Space register offsets
189  */
190 #define	OPT_PCS_OFF_VENDOR	0x0	/* device/vendor ID register */
191 #define	OPT_PCS_OFF_DRAMBASE_HI	0x140	/* DRAM Base register (node 0) */
192 #define	OPT_PCS_OFF_DRAMBASE_LO	0x40	/* DRAM Base register (node 0) */
193 #define	OPT_PCS_OFF_NODEID	0x60	/* Node ID register */
194 
195 /*
196  * Opteron PCI Configuration Space device IDs for nodes
197  */
198 #define	OPT_PCS_DEV_NODE0		24	/* device number for node 0 */
199 
200 
201 /*
202  * Bookkeeping for latencies seen during probing (used for verification)
203  */
204 typedef	struct lgrp_plat_latency_acct {
205 	hrtime_t	la_value;	/* latency value */
206 	int		la_count;	/* occurrences */
207 } lgrp_plat_latency_acct_t;
208 
209 
210 /*
211  * Choices for probing to determine lgroup topology
212  */
213 typedef	enum lgrp_plat_probe_op {
214 	LGRP_PLAT_PROBE_PGCPY,		/* Use page copy */
215 	LGRP_PLAT_PROBE_VENDOR		/* Read vendor ID on Northbridge */
216 } lgrp_plat_probe_op_t;
217 
218 
219 /*
220  * Opteron DRAM address map gives base and limit for physical memory in a node
221  */
222 typedef	struct opt_dram_addr_map {
223 	uint32_t	base_hi;
224 	uint32_t	base_lo;
225 	uint32_t	limit_hi;
226 	uint32_t	limit_lo;
227 } opt_dram_addr_map_t;
228 
229 
230 /*
231  * Starting and ending page for physical memory in node
232  */
233 typedef	struct phys_addr_map {
234 	pfn_t	start;
235 	pfn_t	end;
236 	int	exists;
237 } phys_addr_map_t;
238 
239 
240 /*
241  * Opteron DRAM address map for each node
242  */
243 struct opt_dram_addr_map	opt_dram_map[MAX_NODES];
244 
245 /*
246  * Node ID register contents for each node
247  */
248 uint_t				opt_node_info[MAX_NODES];
249 
250 /*
251  * Whether memory is interleaved across nodes causing MPO to be disabled
252  */
253 int			lgrp_plat_mem_intrlv = 0;
254 
255 /*
256  * Number of nodes in system
257  */
258 uint_t			lgrp_plat_node_cnt = 1;
259 
260 /*
261  * Physical address range for memory in each node
262  */
263 phys_addr_map_t		lgrp_plat_node_memory[MAX_NODES];
264 
265 /*
266  * Probe costs (individual and total) and flush cost
267  */
268 hrtime_t		lgrp_plat_flush_cost = 0;
269 hrtime_t		lgrp_plat_probe_cost = 0;
270 hrtime_t		lgrp_plat_probe_cost_total = 0;
271 
272 /*
273  * Error code for latency adjustment and verification
274  */
275 int			lgrp_plat_probe_error_code = 0;
276 
277 /*
278  * How much latencies were off from minimum values gotten
279  */
280 hrtime_t		lgrp_plat_probe_errors[MAX_NODES][MAX_NODES];
281 
282 /*
283  * Unique probe latencies and number of occurrences of each
284  */
285 lgrp_plat_latency_acct_t	lgrp_plat_probe_lat_acct[MAX_NODES];
286 
287 /*
288  * Size of memory buffer in each node for probing
289  */
290 size_t			lgrp_plat_probe_memsize = 0;
291 
292 /*
293  * Virtual address of page in each node for probing
294  */
295 caddr_t			lgrp_plat_probe_memory[MAX_NODES];
296 
297 /*
298  * Number of unique latencies in probe times
299  */
300 int			lgrp_plat_probe_nlatencies = 0;
301 
302 /*
303  * How many rounds of probing to do
304  */
305 int			lgrp_plat_probe_nrounds = LGRP_PLAT_PROBE_NROUNDS;
306 
307 /*
308  * Number of samples to take when probing each node
309  */
310 int			lgrp_plat_probe_nsamples = LGRP_PLAT_PROBE_NSAMPLES;
311 
312 /*
313  * Number of times to read vendor ID from Northbridge for each probe.
314  */
315 int			lgrp_plat_probe_nreads = LGRP_PLAT_PROBE_NREADS;
316 
317 /*
318  * How to probe to determine lgroup topology
319  */
320 lgrp_plat_probe_op_t	lgrp_plat_probe_op = LGRP_PLAT_PROBE_VENDOR;
321 
322 /*
323  * PFN of page in each node for probing
324  */
325 pfn_t			lgrp_plat_probe_pfn[MAX_NODES];
326 
327 /*
328  * Whether probe time was suspect (ie. not within tolerance of value that it
329  * should match)
330  */
331 int			lgrp_plat_probe_suspect[MAX_NODES][MAX_NODES];
332 
333 /*
334  * How long it takes to access memory from each node
335  */
336 hrtime_t		lgrp_plat_probe_times[MAX_NODES][MAX_NODES];
337 
338 /*
339  * Min and max node memory probe times seen
340  */
341 hrtime_t		lgrp_plat_probe_time_max = 0;
342 hrtime_t		lgrp_plat_probe_time_min = -1;
343 hrtime_t		lgrp_plat_probe_max[MAX_NODES][MAX_NODES];
344 hrtime_t		lgrp_plat_probe_min[MAX_NODES][MAX_NODES];
345 
346 
347 /*
348  * Allocate lgrp and lgrp stat arrays statically.
349  */
350 static lgrp_t	lgrp_space[NLGRP];
351 static int	nlgrps_alloc;
352 
353 struct lgrp_stats lgrp_stats[NLGRP];
354 
355 /*
356  * Supported AMD processor families
357  */
358 #define	AMD_FAMILY_HAMMER	15
359 #define	AMD_FAMILY_GREYHOUND	16
360 
361 /*
362  * Whether to have is_opteron() return 1 even when processor isn't
363  * supported
364  */
365 uint_t	is_opteron_override = 0;
366 
367 /*
368  * AMD processor family for current CPU
369  */
370 uint_t	opt_family = 0;
371 
372 uint_t	opt_probe_func = OPT_PCS_FUNC_DRAM;
373 
374 
375 /*
376  * Determine whether we're running on a supported AMD Opteron since reading
377  * node count and DRAM address map registers may have different format or
378  * may not be supported in future processor families
379  */
380 int
381 is_opteron(void)
382 {
383 
384 	if (x86_vendor != X86_VENDOR_AMD)
385 		return (0);
386 
387 	opt_family = cpuid_getfamily(CPU);
388 	if (opt_family == AMD_FAMILY_HAMMER ||
389 	    opt_family == AMD_FAMILY_GREYHOUND || is_opteron_override)
390 		return (1);
391 	else
392 		return (0);
393 }
394 
395 int
396 plat_lgrphand_to_mem_node(lgrp_handle_t hand)
397 {
398 	if (max_mem_nodes == 1)
399 		return (0);
400 
401 	return ((int)hand);
402 }
403 
404 lgrp_handle_t
405 plat_mem_node_to_lgrphand(int mnode)
406 {
407 	if (max_mem_nodes == 1)
408 		return (LGRP_DEFAULT_HANDLE);
409 
410 	return ((lgrp_handle_t)mnode);
411 }
412 
413 int
414 plat_pfn_to_mem_node(pfn_t pfn)
415 {
416 	int	node;
417 
418 	if (max_mem_nodes == 1)
419 		return (0);
420 
421 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
422 		/*
423 		 * Skip nodes with no memory
424 		 */
425 		if (!lgrp_plat_node_memory[node].exists)
426 			continue;
427 
428 		if (pfn >= lgrp_plat_node_memory[node].start &&
429 		    pfn <= lgrp_plat_node_memory[node].end)
430 			return (node);
431 	}
432 
433 	ASSERT(node < lgrp_plat_node_cnt);
434 	return (-1);
435 }
436 
437 /*
438  * Configure memory nodes for machines with more than one node (ie NUMA)
439  */
440 void
441 plat_build_mem_nodes(struct memlist *list)
442 {
443 	pfn_t		cur_start;	/* start addr of subrange */
444 	pfn_t		cur_end;	/* end addr of subrange */
445 	pfn_t		start;		/* start addr of whole range */
446 	pfn_t		end;		/* end addr of whole range */
447 
448 	/*
449 	 * Boot install lists are arranged <addr, len>, ...
450 	 */
451 	while (list) {
452 		int	node;
453 
454 		start = list->address >> PAGESHIFT;
455 		end = (list->address + list->size - 1) >> PAGESHIFT;
456 
457 		if (start > physmax) {
458 			list = list->next;
459 			continue;
460 		}
461 		if (end > physmax)
462 			end = physmax;
463 
464 		/*
465 		 * When there is only one memnode, just add memory to memnode
466 		 */
467 		if (max_mem_nodes == 1) {
468 			mem_node_add_slice(start, end);
469 			list = list->next;
470 			continue;
471 		}
472 
473 		/*
474 		 * mem_node_add_slice() expects to get a memory range that
475 		 * is within one memnode, so need to split any memory range
476 		 * that spans multiple memnodes into subranges that are each
477 		 * contained within one memnode when feeding them to
478 		 * mem_node_add_slice()
479 		 */
480 		cur_start = start;
481 		do {
482 			node = plat_pfn_to_mem_node(cur_start);
483 
484 			/*
485 			 * Panic if DRAM address map registers or SRAT say
486 			 * memory in node doesn't exist or address from
487 			 * boot installed memory list entry isn't in this node.
488 			 * This shouldn't happen and rest of code can't deal
489 			 * with this if it does.
490 			 */
491 			if (node < 0 || node >= lgrp_plat_node_cnt ||
492 			    !lgrp_plat_node_memory[node].exists ||
493 			    cur_start < lgrp_plat_node_memory[node].start ||
494 			    cur_start > lgrp_plat_node_memory[node].end) {
495 				cmn_err(CE_PANIC, "Don't know which memnode "
496 				    "to add installed memory address 0x%lx\n",
497 				    cur_start);
498 			}
499 
500 			/*
501 			 * End of current subrange should not span memnodes
502 			 */
503 			cur_end = end;
504 			if (lgrp_plat_node_memory[node].exists &&
505 			    cur_end > lgrp_plat_node_memory[node].end)
506 				cur_end = lgrp_plat_node_memory[node].end;
507 
508 			mem_node_add_slice(cur_start, cur_end);
509 
510 			/*
511 			 * Next subrange starts after end of current one
512 			 */
513 			cur_start = cur_end + 1;
514 		} while (cur_end < end);
515 
516 		list = list->next;
517 	}
518 	mem_node_physalign = 0;
519 	mem_node_pfn_shift = 0;
520 }
521 
522 
523 /*
524  * Platform-specific initialization of lgroups
525  */
526 void
527 lgrp_plat_init(void)
528 {
529 	uint_t		bus;
530 	uint_t		dev;
531 	uint_t		node;
532 	uint_t		off_hi;
533 	uint_t		off_lo;
534 	uint64_t	nb_cfg_reg;
535 
536 	extern lgrp_load_t	lgrp_expand_proc_thresh;
537 	extern lgrp_load_t	lgrp_expand_proc_diff;
538 
539 	/*
540 	 * Initialize as a UMA machine if this isn't an Opteron
541 	 */
542 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
543 		lgrp_plat_node_cnt = max_mem_nodes = 1;
544 		return;
545 	}
546 
547 	/*
548 	 * Read configuration registers from PCI configuration space to
549 	 * determine node information, which memory is in each node, etc.
550 	 *
551 	 * Write to PCI configuration space address register to specify
552 	 * which configuration register to read and read/write PCI
553 	 * configuration space data register to get/set contents
554 	 */
555 	bus = OPT_PCS_BUS_CONFIG;
556 	dev = OPT_PCS_DEV_NODE0;
557 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
558 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
559 
560 	/*
561 	 * Read node ID register for node 0 to get node count
562 	 */
563 	opt_node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
564 	    OPT_PCS_OFF_NODEID);
565 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
566 
567 	/*
568 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
569 	 * read high DRAM address map base and limit registers
570 	 */
571 	if (opt_family == AMD_FAMILY_GREYHOUND) {
572 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
573 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
574 			wrmsr(MSR_AMD_NB_CFG,
575 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
576 	}
577 
578 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
579 		uint32_t	base_hi;
580 		uint32_t	base_lo;
581 		uint32_t	limit_hi;
582 		uint32_t	limit_lo;
583 
584 		/*
585 		 * Read node ID register (except for node 0 which we just read)
586 		 */
587 		if (node > 0) {
588 			opt_node_info[node] = pci_getl_func(bus, dev,
589 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
590 		}
591 
592 		/*
593 		 * Read DRAM base and limit registers which specify
594 		 * physical memory range of each node
595 		 */
596 		if (opt_family != AMD_FAMILY_GREYHOUND)
597 			base_hi = 0;
598 		else {
599 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
600 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
601 			base_hi = opt_dram_map[node].base_hi =
602 			    inl(PCI_CONFDATA);
603 		}
604 		base_lo = opt_dram_map[node].base_lo = pci_getl_func(bus, dev,
605 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
606 
607 		if (opt_dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN)
608 			lgrp_plat_mem_intrlv++;
609 
610 		off_hi += 4;	/* high limit register offset */
611 		if (opt_family != AMD_FAMILY_GREYHOUND)
612 			limit_hi = 0;
613 		else {
614 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
615 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
616 			limit_hi = opt_dram_map[node].limit_hi =
617 			    inl(PCI_CONFDATA);
618 		}
619 
620 		off_lo += 4;	/* low limit register offset */
621 		limit_lo = opt_dram_map[node].limit_lo = pci_getl_func(bus,
622 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
623 
624 		/*
625 		 * Increment device number to next node and register offsets
626 		 * for DRAM base register of next node
627 		 */
628 		off_hi += 4;
629 		off_lo += 4;
630 		dev++;
631 
632 		/*
633 		 * Both read and write enable bits must be enabled in DRAM
634 		 * address map base register for physical memory to exist in
635 		 * node
636 		 */
637 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
638 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
639 			/*
640 			 * Mark node memory as non-existent and set start and
641 			 * end addresses to be same in lgrp_plat_node_memory[]
642 			 */
643 			lgrp_plat_node_memory[node].exists = 0;
644 			lgrp_plat_node_memory[node].start =
645 			    lgrp_plat_node_memory[node].end = (pfn_t)-1;
646 			continue;
647 		}
648 
649 		/*
650 		 * Get PFN for first page in each node,
651 		 * so we can probe memory to determine latency topology
652 		 */
653 		lgrp_plat_probe_pfn[node] =
654 		    btop(OPT_DRAMADDR(base_hi, base_lo));
655 
656 		/*
657 		 * Mark node memory as existing and remember physical address
658 		 * range of each node for use later
659 		 */
660 		lgrp_plat_node_memory[node].exists = 1;
661 
662 		lgrp_plat_node_memory[node].start =
663 		    btop(OPT_DRAMADDR(base_hi, base_lo));
664 
665 		lgrp_plat_node_memory[node].end =
666 		    btop(OPT_DRAMADDR(limit_hi, limit_lo) |
667 		    OPT_DRAMADDR_LO_MASK_OFF);
668 	}
669 
670 	/*
671 	 * Restore PCI Extended Configuration Space enable bit
672 	 */
673 	if (opt_family == AMD_FAMILY_GREYHOUND) {
674 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
675 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
676 	}
677 
678 	/*
679 	 * Only use one memory node if memory is interleaved between any nodes
680 	 */
681 	if (lgrp_plat_mem_intrlv) {
682 		lgrp_plat_node_cnt = max_mem_nodes = 1;
683 		(void) lgrp_topo_ht_limit_set(1);
684 	} else {
685 		max_mem_nodes = lgrp_plat_node_cnt;
686 
687 		/*
688 		 * Probing errors can mess up the lgroup topology and force us
689 		 * fall back to a 2 level lgroup topology.  Here we bound how
690 		 * tall the lgroup topology can grow in hopes of avoiding any
691 		 * anamolies in probing from messing up the lgroup topology
692 		 * by limiting the accuracy of the latency topology.
693 		 *
694 		 * Assume that nodes will at least be configured in a ring,
695 		 * so limit height of lgroup topology to be less than number
696 		 * of nodes on a system with 4 or more nodes
697 		 */
698 		if (lgrp_plat_node_cnt >= 4 &&
699 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
700 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
701 	}
702 
703 	/*
704 	 * Lgroups on Opteron architectures have but a single physical
705 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
706 	 * so that lgrp_choose() will spread things out aggressively.
707 	 */
708 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
709 	lgrp_expand_proc_diff = 0;
710 }
711 
712 
713 /*
714  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
715  * be considered same
716  */
717 #define	LGRP_LAT_TOLERANCE_SHIFT	4
718 
719 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
720 
721 
722 /*
723  * Adjust latencies between nodes to be symmetric, normalize latencies between
724  * any nodes that are within some tolerance to be same, and make local
725  * latencies be same
726  */
727 static void
728 lgrp_plat_latency_adjust(void)
729 {
730 	int				i;
731 	int				j;
732 	int				k;
733 	int				l;
734 	u_longlong_t			max;
735 	u_longlong_t			min;
736 	u_longlong_t			t;
737 	u_longlong_t			t1;
738 	u_longlong_t			t2;
739 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
740 	int				lat_corrected[MAX_NODES][MAX_NODES];
741 
742 	/*
743 	 * Nothing to do when this is an UMA machine
744 	 */
745 	if (max_mem_nodes == 1)
746 		return;
747 
748 	/*
749 	 * Make sure that latencies are symmetric between any two nodes
750 	 * (ie. latency(node0, node1) == latency(node1, node0))
751 	 */
752 	for (i = 0; i < lgrp_plat_node_cnt; i++)
753 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
754 			t1 = lgrp_plat_probe_times[i][j];
755 			t2 = lgrp_plat_probe_times[j][i];
756 
757 			if (t1 == 0 || t2 == 0 || t1 == t2)
758 				continue;
759 
760 			/*
761 			 * Latencies should be same
762 			 * - Use minimum of two latencies which should be same
763 			 * - Track suspect probe times not within tolerance of
764 			 *   min value
765 			 * - Remember how much values are corrected by
766 			 */
767 			if (t1 > t2) {
768 				t = t2;
769 				lgrp_plat_probe_errors[i][j] += t1 - t2;
770 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
771 					lgrp_plat_probe_suspect[i][j]++;
772 					lgrp_plat_probe_suspect[j][i]++;
773 				}
774 			} else if (t2 > t1) {
775 				t = t1;
776 				lgrp_plat_probe_errors[j][i] += t2 - t1;
777 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
778 					lgrp_plat_probe_suspect[i][j]++;
779 					lgrp_plat_probe_suspect[j][i]++;
780 				}
781 			}
782 
783 			lgrp_plat_probe_times[i][j] =
784 			    lgrp_plat_probe_times[j][i] = t;
785 			lgrp_config(cflag, t1, t);
786 			lgrp_config(cflag, t2, t);
787 		}
788 
789 	/*
790 	 * Keep track of which latencies get corrected
791 	 */
792 	for (i = 0; i < MAX_NODES; i++)
793 		for (j = 0; j < MAX_NODES; j++)
794 			lat_corrected[i][j] = 0;
795 
796 	/*
797 	 * For every two nodes, see whether there is another pair of nodes which
798 	 * are about the same distance apart and make the latencies be the same
799 	 * if they are close enough together
800 	 */
801 	for (i = 0; i < lgrp_plat_node_cnt; i++)
802 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
803 			/*
804 			 * Pick one pair of nodes (i, j)
805 			 * and get latency between them
806 			 */
807 			t1 = lgrp_plat_probe_times[i][j];
808 
809 			/*
810 			 * Skip this pair of nodes if there isn't a latency
811 			 * for it yet
812 			 */
813 			if (t1 == 0)
814 				continue;
815 
816 			for (k = 0; k < lgrp_plat_node_cnt; k++)
817 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
818 					/*
819 					 * Pick another pair of nodes (k, l)
820 					 * not same as (i, j) and get latency
821 					 * between them
822 					 */
823 					if (k == i && l == j)
824 						continue;
825 
826 					t2 = lgrp_plat_probe_times[k][l];
827 
828 					/*
829 					 * Skip this pair of nodes if there
830 					 * isn't a latency for it yet
831 					 */
832 
833 					if (t2 == 0)
834 						continue;
835 
836 					/*
837 					 * Skip nodes (k, l) if they already
838 					 * have same latency as (i, j) or
839 					 * their latency isn't close enough to
840 					 * be considered/made the same
841 					 */
842 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
843 					    t1 >> lgrp_plat_probe_lt_shift) ||
844 					    (t2 > t1 && t2 - t1 >
845 					    t2 >> lgrp_plat_probe_lt_shift))
846 						continue;
847 
848 					/*
849 					 * Make latency(i, j) same as
850 					 * latency(k, l), try to use latency
851 					 * that has been adjusted already to get
852 					 * more consistency (if possible), and
853 					 * remember which latencies were
854 					 * adjusted for next time
855 					 */
856 					if (lat_corrected[i][j]) {
857 						t = t1;
858 						lgrp_config(cflag, t2, t);
859 						t2 = t;
860 					} else if (lat_corrected[k][l]) {
861 						t = t2;
862 						lgrp_config(cflag, t1, t);
863 						t1 = t;
864 					} else {
865 						if (t1 > t2)
866 							t = t2;
867 						else
868 							t = t1;
869 						lgrp_config(cflag, t1, t);
870 						lgrp_config(cflag, t2, t);
871 						t1 = t2 = t;
872 					}
873 
874 					lgrp_plat_probe_times[i][j] =
875 					    lgrp_plat_probe_times[k][l] = t;
876 
877 					lat_corrected[i][j] =
878 					    lat_corrected[k][l] = 1;
879 				}
880 		}
881 
882 	/*
883 	 * Local latencies should be same
884 	 * - Find min and max local latencies
885 	 * - Make all local latencies be minimum
886 	 */
887 	min = -1;
888 	max = 0;
889 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
890 		t = lgrp_plat_probe_times[i][i];
891 		if (t == 0)
892 			continue;
893 		if (min == -1 || t < min)
894 			min = t;
895 		if (t > max)
896 			max = t;
897 	}
898 	if (min != max) {
899 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
900 			int	local;
901 
902 			local = lgrp_plat_probe_times[i][i];
903 			if (local == 0)
904 				continue;
905 
906 			/*
907 			 * Track suspect probe times that aren't within
908 			 * tolerance of minimum local latency and how much
909 			 * probe times are corrected by
910 			 */
911 			if (local - min > min >> lgrp_plat_probe_lt_shift)
912 				lgrp_plat_probe_suspect[i][i]++;
913 
914 			lgrp_plat_probe_errors[i][i] += local - min;
915 
916 			/*
917 			 * Make local latencies be minimum
918 			 */
919 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
920 			lgrp_plat_probe_times[i][i] = min;
921 		}
922 	}
923 
924 	/*
925 	 * Determine max probe time again since just adjusted latencies
926 	 */
927 	lgrp_plat_probe_time_max = 0;
928 	for (i = 0; i < lgrp_plat_node_cnt; i++)
929 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
930 			t = lgrp_plat_probe_times[i][j];
931 			if (t > lgrp_plat_probe_time_max)
932 				lgrp_plat_probe_time_max = t;
933 		}
934 }
935 
936 
937 /*
938  * Verify following about latencies between nodes:
939  *
940  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
941  * - Local latencies same
942  * - Local < remote
943  * - Number of latencies seen is reasonable
944  * - Number of occurrences of a given latency should be more than 1
945  *
946  * Returns:
947  *	0	Success
948  *	-1	Not symmetric
949  *	-2	Local latencies not same
950  *	-3	Local >= remote
951  *	-4	Wrong number of latencies
952  *	-5	Not enough occurrences of given latency
953  */
954 static int
955 lgrp_plat_latency_verify(void)
956 {
957 	int				i;
958 	int				j;
959 	lgrp_plat_latency_acct_t	*l;
960 	int				probed;
961 	u_longlong_t			t1;
962 	u_longlong_t			t2;
963 
964 	/*
965 	 * Nothing to do when this is an UMA machine, lgroup topology is
966 	 * limited to 2 levels, or there aren't any probe times yet
967 	 */
968 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
969 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
970 		return (0);
971 
972 	/*
973 	 * Make sure that latencies are symmetric between any two nodes
974 	 * (ie. latency(node0, node1) == latency(node1, node0))
975 	 */
976 	for (i = 0; i < lgrp_plat_node_cnt; i++)
977 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
978 			t1 = lgrp_plat_probe_times[i][j];
979 			t2 = lgrp_plat_probe_times[j][i];
980 
981 			if (t1 == 0 || t2 == 0 || t1 == t2)
982 				continue;
983 
984 			return (-1);
985 		}
986 
987 	/*
988 	 * Local latencies should be same
989 	 */
990 	t1 = lgrp_plat_probe_times[0][0];
991 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
992 		t2 = lgrp_plat_probe_times[i][i];
993 		if (t2 == 0)
994 			continue;
995 
996 		if (t1 == 0) {
997 			t1 = t2;
998 			continue;
999 		}
1000 
1001 		if (t1 != t2)
1002 			return (-2);
1003 	}
1004 
1005 	/*
1006 	 * Local latencies should be less than remote
1007 	 */
1008 	if (t1) {
1009 		for (i = 0; i < lgrp_plat_node_cnt; i++)
1010 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1011 				t2 = lgrp_plat_probe_times[i][j];
1012 				if (i == j || t2 == 0)
1013 					continue;
1014 
1015 				if (t1 >= t2)
1016 					return (-3);
1017 			}
1018 	}
1019 
1020 	/*
1021 	 * Rest of checks are not very useful for machines with less than
1022 	 * 4 nodes (which means less than 3 latencies on Opteron)
1023 	 */
1024 	if (lgrp_plat_node_cnt < 4)
1025 		return (0);
1026 
1027 	/*
1028 	 * Need to see whether done probing in order to verify number of
1029 	 * latencies are correct
1030 	 */
1031 	probed = 0;
1032 	for (i = 0; i < lgrp_plat_node_cnt; i++)
1033 		if (lgrp_plat_probe_times[i][i])
1034 			probed++;
1035 
1036 	if (probed != lgrp_plat_node_cnt)
1037 		return (0);
1038 
1039 	/*
1040 	 * Determine number of unique latencies seen in probe times,
1041 	 * their values, and number of occurrences of each
1042 	 */
1043 	lgrp_plat_probe_nlatencies = 0;
1044 	bzero(lgrp_plat_probe_lat_acct,
1045 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
1046 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1047 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1048 			int	k;
1049 
1050 			/*
1051 			 * Look at each probe time
1052 			 */
1053 			t1 = lgrp_plat_probe_times[i][j];
1054 			if (t1 == 0)
1055 				continue;
1056 
1057 			/*
1058 			 * Account for unique latencies
1059 			 */
1060 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1061 				l = &lgrp_plat_probe_lat_acct[k];
1062 				if (t1 == l->la_value) {
1063 					/*
1064 					 * Increment number of occurrences
1065 					 * if seen before
1066 					 */
1067 					l->la_count++;
1068 					break;
1069 				} else if (l->la_value == 0) {
1070 					/*
1071 					 * Record latency if haven't seen before
1072 					 */
1073 					l->la_value = t1;
1074 					l->la_count++;
1075 					lgrp_plat_probe_nlatencies++;
1076 					break;
1077 				}
1078 			}
1079 		}
1080 	}
1081 
1082 	/*
1083 	 * Number of latencies should be relative to number of
1084 	 * nodes in system:
1085 	 * - Same as nodes when nodes <= 2
1086 	 * - Less than nodes when nodes > 2
1087 	 * - Greater than 2 when nodes >= 4
1088 	 */
1089 	if ((lgrp_plat_node_cnt <= 2 &&
1090 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
1091 	    (lgrp_plat_node_cnt > 2 &&
1092 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
1093 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
1094 	    lgrp_plat_probe_nlatencies <= 2))
1095 		return (-4);
1096 
1097 	/*
1098 	 * There should be more than one occurrence of every latency
1099 	 * as long as probing is complete
1100 	 */
1101 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
1102 		l = &lgrp_plat_probe_lat_acct[i];
1103 		if (l->la_count <= 1)
1104 			return (-5);
1105 	}
1106 	return (0);
1107 }
1108 
1109 
1110 /*
1111  * Set lgroup latencies for 2 level lgroup topology
1112  */
1113 static void
1114 lgrp_plat_2level_setup(void)
1115 {
1116 	int	i;
1117 
1118 	if (lgrp_plat_node_cnt >= 4)
1119 		cmn_err(CE_NOTE,
1120 		    "MPO only optimizing for local and remote\n");
1121 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1122 		int	j;
1123 
1124 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1125 			if (i == j)
1126 				lgrp_plat_probe_times[i][j] = 2;
1127 			else
1128 				lgrp_plat_probe_times[i][j] = 3;
1129 		}
1130 	}
1131 	lgrp_plat_probe_time_min = 2;
1132 	lgrp_plat_probe_time_max = 3;
1133 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1134 }
1135 
1136 
1137 /*
1138  * Return time needed to probe from current CPU to memory in given node
1139  */
1140 static hrtime_t
1141 lgrp_plat_probe_time(int to)
1142 {
1143 	caddr_t		buf;
1144 	uint_t		dev;
1145 	/* LINTED: set but not used in function */
1146 	volatile uint_t	dev_vendor;
1147 	hrtime_t	elapsed;
1148 	hrtime_t	end;
1149 	int		from;
1150 	int		i;
1151 	int		ipl;
1152 	hrtime_t	max;
1153 	hrtime_t	min;
1154 	hrtime_t	start;
1155 	int		cnt;
1156 	extern int	use_sse_pagecopy;
1157 
1158 	/*
1159 	 * Determine ID of node containing current CPU
1160 	 */
1161 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1162 
1163 	/*
1164 	 * Do common work for probing main memory
1165 	 */
1166 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1167 		/*
1168 		 * Skip probing any nodes without memory and
1169 		 * set probe time to 0
1170 		 */
1171 		if (lgrp_plat_probe_memory[to] == NULL) {
1172 			lgrp_plat_probe_times[from][to] = 0;
1173 			return (0);
1174 		}
1175 
1176 		/*
1177 		 * Invalidate caches once instead of once every sample
1178 		 * which should cut cost of probing by a lot
1179 		 */
1180 		lgrp_plat_flush_cost = gethrtime();
1181 		invalidate_cache();
1182 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1183 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1184 	}
1185 
1186 	/*
1187 	 * Probe from current CPU to given memory using specified operation
1188 	 * and take specified number of samples
1189 	 */
1190 	max = 0;
1191 	min = -1;
1192 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1193 		lgrp_plat_probe_cost = gethrtime();
1194 
1195 		/*
1196 		 * Can't measure probe time if gethrtime() isn't working yet
1197 		 */
1198 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1199 			return (0);
1200 
1201 		switch (lgrp_plat_probe_op) {
1202 
1203 		case LGRP_PLAT_PROBE_PGCPY:
1204 		default:
1205 			/*
1206 			 * Measure how long it takes to copy page
1207 			 * on top of itself
1208 			 */
1209 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1210 
1211 			kpreempt_disable();
1212 			ipl = splhigh();
1213 			start = gethrtime();
1214 			if (use_sse_pagecopy)
1215 				hwblkpagecopy(buf, buf);
1216 			else
1217 				bcopy(buf, buf, PAGESIZE);
1218 			end = gethrtime();
1219 			elapsed = end - start;
1220 			splx(ipl);
1221 			kpreempt_enable();
1222 			break;
1223 
1224 		case LGRP_PLAT_PROBE_VENDOR:
1225 			/*
1226 			 * Measure how long it takes to read vendor ID from
1227 			 * Northbridge
1228 			 */
1229 			dev = OPT_PCS_DEV_NODE0 + to;
1230 			kpreempt_disable();
1231 			ipl = spl8();
1232 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1233 			    OPT_PCS_OFF_VENDOR));
1234 			start = gethrtime();
1235 			for (cnt = 0; cnt < lgrp_plat_probe_nreads; cnt++)
1236 				dev_vendor = inl(PCI_CONFDATA);
1237 			end = gethrtime();
1238 			elapsed = (end - start) / lgrp_plat_probe_nreads;
1239 			splx(ipl);
1240 			kpreempt_enable();
1241 			break;
1242 		}
1243 
1244 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1245 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1246 
1247 		if (min == -1 || elapsed < min)
1248 			min = elapsed;
1249 		if (elapsed > max)
1250 			max = elapsed;
1251 	}
1252 
1253 	/*
1254 	 * Update minimum and maximum probe times between
1255 	 * these two nodes
1256 	 */
1257 	if (min < lgrp_plat_probe_min[from][to] ||
1258 	    lgrp_plat_probe_min[from][to] == 0)
1259 		lgrp_plat_probe_min[from][to] = min;
1260 
1261 	if (max > lgrp_plat_probe_max[from][to])
1262 		lgrp_plat_probe_max[from][to] = max;
1263 
1264 	return (min);
1265 }
1266 
1267 
1268 /*
1269  * Probe memory in each node from current CPU to determine latency topology
1270  */
1271 void
1272 lgrp_plat_probe(void)
1273 {
1274 	int		from;
1275 	int		i;
1276 	hrtime_t	probe_time;
1277 	int		to;
1278 
1279 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1280 		return;
1281 
1282 	/*
1283 	 * Determine ID of node containing current CPU
1284 	 */
1285 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1286 
1287 	/*
1288 	 * Don't need to probe if got times already
1289 	 */
1290 	if (lgrp_plat_probe_times[from][from] != 0)
1291 		return;
1292 
1293 	/*
1294 	 * Read vendor ID in Northbridge or read and write page(s)
1295 	 * in each node from current CPU and remember how long it takes,
1296 	 * so we can build latency topology of machine later.
1297 	 * This should approximate the memory latency between each node.
1298 	 */
1299 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1300 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1301 			/*
1302 			 * Get probe time and bail out if can't get it yet
1303 			 */
1304 			probe_time = lgrp_plat_probe_time(to);
1305 			if (probe_time == 0)
1306 				return;
1307 
1308 			/*
1309 			 * Keep lowest probe time as latency between nodes
1310 			 */
1311 			if (lgrp_plat_probe_times[from][to] == 0 ||
1312 			    probe_time < lgrp_plat_probe_times[from][to])
1313 				lgrp_plat_probe_times[from][to] = probe_time;
1314 
1315 			/*
1316 			 * Update overall minimum and maximum probe times
1317 			 * across all nodes
1318 			 */
1319 			if (probe_time < lgrp_plat_probe_time_min ||
1320 			    lgrp_plat_probe_time_min == -1)
1321 				lgrp_plat_probe_time_min = probe_time;
1322 			if (probe_time > lgrp_plat_probe_time_max)
1323 				lgrp_plat_probe_time_max = probe_time;
1324 		}
1325 
1326 	/*
1327 	 * - Fix up latencies such that local latencies are same,
1328 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1329 	 *
1330 	 * - Verify that latencies look ok
1331 	 *
1332 	 * - Fallback to just optimizing for local and remote if
1333 	 *   latencies didn't look right
1334 	 */
1335 	lgrp_plat_latency_adjust();
1336 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1337 	if (lgrp_plat_probe_error_code)
1338 		lgrp_plat_2level_setup();
1339 }
1340 
1341 
1342 /*
1343  * Platform-specific initialization
1344  */
1345 void
1346 lgrp_plat_main_init(void)
1347 {
1348 	int	curnode;
1349 	int	ht_limit;
1350 	int	i;
1351 
1352 	/*
1353 	 * Print a notice that MPO is disabled when memory is interleaved
1354 	 * across nodes....Would do this when it is discovered, but can't
1355 	 * because it happens way too early during boot....
1356 	 */
1357 	if (lgrp_plat_mem_intrlv)
1358 		cmn_err(CE_NOTE,
1359 		    "MPO disabled because memory is interleaved\n");
1360 
1361 	/*
1362 	 * Don't bother to do any probing if there is only one node or the
1363 	 * height of the lgroup topology less than or equal to 2
1364 	 */
1365 	ht_limit = lgrp_topo_ht_limit();
1366 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1367 		/*
1368 		 * Setup lgroup latencies for 2 level lgroup topology
1369 		 * (ie. local and remote only) if they haven't been set yet
1370 		 */
1371 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1372 		    lgrp_plat_probe_time_max == 0)
1373 			lgrp_plat_2level_setup();
1374 		return;
1375 	}
1376 
1377 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1378 		/*
1379 		 * Should have been able to probe from CPU 0 when it was added
1380 		 * to lgroup hierarchy, but may not have been able to then
1381 		 * because it happens so early in boot that gethrtime() hasn't
1382 		 * been initialized.  (:-(
1383 		 */
1384 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1385 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1386 			lgrp_plat_probe();
1387 
1388 		return;
1389 	}
1390 
1391 	/*
1392 	 * When probing memory, use one page for every sample to determine
1393 	 * lgroup topology and taking multiple samples
1394 	 */
1395 	if (lgrp_plat_probe_memsize == 0)
1396 		lgrp_plat_probe_memsize = PAGESIZE *
1397 		    lgrp_plat_probe_nsamples;
1398 
1399 	/*
1400 	 * Map memory in each node needed for probing to determine latency
1401 	 * topology
1402 	 */
1403 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1404 		int	mnode;
1405 
1406 		/*
1407 		 * Skip this node and leave its probe page NULL
1408 		 * if it doesn't have any memory
1409 		 */
1410 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1411 		if (!mem_node_config[mnode].exists) {
1412 			lgrp_plat_probe_memory[i] = NULL;
1413 			continue;
1414 		}
1415 
1416 		/*
1417 		 * Allocate one kernel virtual page
1418 		 */
1419 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1420 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1421 		if (lgrp_plat_probe_memory[i] == NULL) {
1422 			cmn_err(CE_WARN,
1423 			    "lgrp_plat_main_init: couldn't allocate memory");
1424 			return;
1425 		}
1426 
1427 		/*
1428 		 * Map virtual page to first page in node
1429 		 */
1430 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1431 		    lgrp_plat_probe_memsize,
1432 		    lgrp_plat_probe_pfn[i],
1433 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1434 		    HAT_LOAD_NOCONSIST);
1435 	}
1436 
1437 	/*
1438 	 * Probe from current CPU
1439 	 */
1440 	lgrp_plat_probe();
1441 }
1442 
1443 /*
1444  * Allocate additional space for an lgroup.
1445  */
1446 /* ARGSUSED */
1447 lgrp_t *
1448 lgrp_plat_alloc(lgrp_id_t lgrpid)
1449 {
1450 	lgrp_t *lgrp;
1451 
1452 	lgrp = &lgrp_space[nlgrps_alloc++];
1453 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1454 		return (NULL);
1455 	return (lgrp);
1456 }
1457 
1458 /*
1459  * Platform handling for (re)configuration changes
1460  */
1461 /* ARGSUSED */
1462 void
1463 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1464 {
1465 }
1466 
1467 /*
1468  * Return the platform handle for the lgroup containing the given CPU
1469  */
1470 /* ARGSUSED */
1471 lgrp_handle_t
1472 lgrp_plat_cpu_to_hand(processorid_t id)
1473 {
1474 	if (lgrp_plat_node_cnt == 1)
1475 		return (LGRP_DEFAULT_HANDLE);
1476 
1477 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1478 }
1479 
1480 /*
1481  * Return the platform handle of the lgroup that contains the physical memory
1482  * corresponding to the given page frame number
1483  */
1484 /* ARGSUSED */
1485 lgrp_handle_t
1486 lgrp_plat_pfn_to_hand(pfn_t pfn)
1487 {
1488 	int	mnode;
1489 
1490 	if (max_mem_nodes == 1)
1491 		return (LGRP_DEFAULT_HANDLE);
1492 
1493 	if (pfn > physmax)
1494 		return (LGRP_NULL_HANDLE);
1495 
1496 	mnode = plat_pfn_to_mem_node(pfn);
1497 	if (mnode < 0)
1498 		return (LGRP_NULL_HANDLE);
1499 
1500 	return (MEM_NODE_2_LGRPHAND(mnode));
1501 }
1502 
1503 /*
1504  * Return the maximum number of lgrps supported by the platform.
1505  * Before lgrp topology is known it returns an estimate based on the number of
1506  * nodes. Once topology is known it returns the actual maximim number of lgrps
1507  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1508  * may not grow during system lifetime.
1509  */
1510 int
1511 lgrp_plat_max_lgrps()
1512 {
1513 	return (lgrp_topo_initialized ?
1514 	    lgrp_alloc_max + 1 :
1515 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1516 }
1517 
1518 /*
1519  * Return the number of free, allocatable, or installed
1520  * pages in an lgroup
1521  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1522  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1523  */
1524 /* ARGSUSED */
1525 static pgcnt_t
1526 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1527 {
1528 	struct memlist *mlist;
1529 	pgcnt_t npgs = 0;
1530 	extern struct memlist *phys_avail;
1531 	extern struct memlist *phys_install;
1532 
1533 	switch (query) {
1534 	case LGRP_MEM_SIZE_FREE:
1535 		return ((pgcnt_t)freemem);
1536 	case LGRP_MEM_SIZE_AVAIL:
1537 		memlist_read_lock();
1538 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1539 			npgs += btop(mlist->size);
1540 		memlist_read_unlock();
1541 		return (npgs);
1542 	case LGRP_MEM_SIZE_INSTALL:
1543 		memlist_read_lock();
1544 		for (mlist = phys_install; mlist; mlist = mlist->next)
1545 			npgs += btop(mlist->size);
1546 		memlist_read_unlock();
1547 		return (npgs);
1548 	default:
1549 		return ((pgcnt_t)0);
1550 	}
1551 }
1552 
1553 /*
1554  * Return the number of free pages in an lgroup.
1555  *
1556  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1557  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1558  * number of allocatable base pagesize pages corresponding to the
1559  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1560  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1561  * memory installed, regardless of whether or not it's usable.
1562  */
1563 pgcnt_t
1564 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1565 {
1566 	int	mnode;
1567 	pgcnt_t npgs = (pgcnt_t)0;
1568 	extern struct memlist *phys_avail;
1569 	extern struct memlist *phys_install;
1570 
1571 
1572 	if (plathand == LGRP_DEFAULT_HANDLE)
1573 		return (lgrp_plat_mem_size_default(plathand, query));
1574 
1575 	if (plathand != LGRP_NULL_HANDLE) {
1576 		mnode = plat_lgrphand_to_mem_node(plathand);
1577 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1578 			switch (query) {
1579 			case LGRP_MEM_SIZE_FREE:
1580 				npgs = MNODE_PGCNT(mnode);
1581 				break;
1582 			case LGRP_MEM_SIZE_AVAIL:
1583 				npgs = mem_node_memlist_pages(mnode,
1584 				    phys_avail);
1585 				break;
1586 			case LGRP_MEM_SIZE_INSTALL:
1587 				npgs = mem_node_memlist_pages(mnode,
1588 				    phys_install);
1589 				break;
1590 			default:
1591 				break;
1592 			}
1593 		}
1594 	}
1595 	return (npgs);
1596 }
1597 
1598 /*
1599  * Return latency between "from" and "to" lgroups
1600  *
1601  * This latency number can only be used for relative comparison
1602  * between lgroups on the running system, cannot be used across platforms,
1603  * and may not reflect the actual latency.  It is platform and implementation
1604  * specific, so platform gets to decide its value.  It would be nice if the
1605  * number was at least proportional to make comparisons more meaningful though.
1606  */
1607 /* ARGSUSED */
1608 int
1609 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1610 {
1611 	lgrp_handle_t	src, dest;
1612 
1613 	if (max_mem_nodes == 1)
1614 		return (0);
1615 
1616 	/*
1617 	 * Return max latency for root lgroup
1618 	 */
1619 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1620 		return (lgrp_plat_probe_time_max);
1621 
1622 	src = from;
1623 	dest = to;
1624 
1625 	/*
1626 	 * Return 0 for nodes (lgroup platform handles) out of range
1627 	 */
1628 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1629 		return (0);
1630 
1631 	/*
1632 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1633 	 * and we are trying to get latency from current CPU to some node
1634 	 */
1635 	if (lgrp_plat_probe_times[src][src] == 0 &&
1636 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1637 		lgrp_plat_probe();
1638 
1639 	return (lgrp_plat_probe_times[src][dest]);
1640 }
1641 
1642 /*
1643  * Return platform handle for root lgroup
1644  */
1645 lgrp_handle_t
1646 lgrp_plat_root_hand(void)
1647 {
1648 	return (LGRP_DEFAULT_HANDLE);
1649 }
1650