xref: /illumos-gate/usr/src/uts/i86pc/os/lgrpplat.c (revision dd4eeefdb8e4583c47e28a7f315db6087931ef06)
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 #if defined(__xpv)
530 	/*
531 	 * XXPV	For now, the hypervisor treats all memory equally.
532 	 */
533 	lgrp_plat_node_cnt = max_mem_nodes = 1;
534 #else	/* __xpv */
535 	uint_t		bus;
536 	uint_t		dev;
537 	uint_t		node;
538 	uint_t		off_hi;
539 	uint_t		off_lo;
540 	uint64_t	nb_cfg_reg;
541 
542 	extern lgrp_load_t	lgrp_expand_proc_thresh;
543 	extern lgrp_load_t	lgrp_expand_proc_diff;
544 
545 	/*
546 	 * Initialize as a UMA machine if this isn't an Opteron
547 	 */
548 	if (!is_opteron() || lgrp_topo_ht_limit() == 1) {
549 		lgrp_plat_node_cnt = max_mem_nodes = 1;
550 		return;
551 	}
552 
553 	/*
554 	 * Read configuration registers from PCI configuration space to
555 	 * determine node information, which memory is in each node, etc.
556 	 *
557 	 * Write to PCI configuration space address register to specify
558 	 * which configuration register to read and read/write PCI
559 	 * configuration space data register to get/set contents
560 	 */
561 	bus = OPT_PCS_BUS_CONFIG;
562 	dev = OPT_PCS_DEV_NODE0;
563 	off_hi = OPT_PCS_OFF_DRAMBASE_HI;
564 	off_lo = OPT_PCS_OFF_DRAMBASE_LO;
565 
566 	/*
567 	 * Read node ID register for node 0 to get node count
568 	 */
569 	opt_node_info[0] = pci_getl_func(bus, dev, OPT_PCS_FUNC_HT,
570 	    OPT_PCS_OFF_NODEID);
571 	lgrp_plat_node_cnt = OPT_NODE_CNT(opt_node_info[0]) + 1;
572 
573 	/*
574 	 * For Greyhound, PCI Extended Configuration Space must be enabled to
575 	 * read high DRAM address map base and limit registers
576 	 */
577 	if (opt_family == AMD_FAMILY_GREYHOUND) {
578 		nb_cfg_reg = rdmsr(MSR_AMD_NB_CFG);
579 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
580 			wrmsr(MSR_AMD_NB_CFG,
581 			    nb_cfg_reg | AMD_GH_NB_CFG_EN_ECS);
582 	}
583 
584 	for (node = 0; node < lgrp_plat_node_cnt; node++) {
585 		uint32_t	base_hi;
586 		uint32_t	base_lo;
587 		uint32_t	limit_hi;
588 		uint32_t	limit_lo;
589 
590 		/*
591 		 * Read node ID register (except for node 0 which we just read)
592 		 */
593 		if (node > 0) {
594 			opt_node_info[node] = pci_getl_func(bus, dev,
595 			    OPT_PCS_FUNC_HT, OPT_PCS_OFF_NODEID);
596 		}
597 
598 		/*
599 		 * Read DRAM base and limit registers which specify
600 		 * physical memory range of each node
601 		 */
602 		if (opt_family != AMD_FAMILY_GREYHOUND)
603 			base_hi = 0;
604 		else {
605 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
606 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
607 			base_hi = opt_dram_map[node].base_hi =
608 			    inl(PCI_CONFDATA);
609 		}
610 		base_lo = opt_dram_map[node].base_lo = pci_getl_func(bus, dev,
611 		    OPT_PCS_FUNC_ADDRMAP, off_lo);
612 
613 		if (opt_dram_map[node].base_lo & OPT_DRAMBASE_LO_MASK_INTRLVEN)
614 			lgrp_plat_mem_intrlv++;
615 
616 		off_hi += 4;	/* high limit register offset */
617 		if (opt_family != AMD_FAMILY_GREYHOUND)
618 			limit_hi = 0;
619 		else {
620 			outl(PCI_CONFADD, OPT_PCI_ECS_ADDR(bus, dev,
621 			    OPT_PCS_FUNC_ADDRMAP, off_hi));
622 			limit_hi = opt_dram_map[node].limit_hi =
623 			    inl(PCI_CONFDATA);
624 		}
625 
626 		off_lo += 4;	/* low limit register offset */
627 		limit_lo = opt_dram_map[node].limit_lo = pci_getl_func(bus,
628 		    dev, OPT_PCS_FUNC_ADDRMAP, off_lo);
629 
630 		/*
631 		 * Increment device number to next node and register offsets
632 		 * for DRAM base register of next node
633 		 */
634 		off_hi += 4;
635 		off_lo += 4;
636 		dev++;
637 
638 		/*
639 		 * Both read and write enable bits must be enabled in DRAM
640 		 * address map base register for physical memory to exist in
641 		 * node
642 		 */
643 		if ((base_lo & OPT_DRAMBASE_LO_MASK_RE) == 0 ||
644 		    (base_lo & OPT_DRAMBASE_LO_MASK_WE) == 0) {
645 			/*
646 			 * Mark node memory as non-existent and set start and
647 			 * end addresses to be same in lgrp_plat_node_memory[]
648 			 */
649 			lgrp_plat_node_memory[node].exists = 0;
650 			lgrp_plat_node_memory[node].start =
651 			    lgrp_plat_node_memory[node].end = (pfn_t)-1;
652 			continue;
653 		}
654 
655 		/*
656 		 * Get PFN for first page in each node,
657 		 * so we can probe memory to determine latency topology
658 		 */
659 		lgrp_plat_probe_pfn[node] =
660 		    btop(OPT_DRAMADDR(base_hi, base_lo));
661 
662 		/*
663 		 * Mark node memory as existing and remember physical address
664 		 * range of each node for use later
665 		 */
666 		lgrp_plat_node_memory[node].exists = 1;
667 
668 		lgrp_plat_node_memory[node].start =
669 		    btop(OPT_DRAMADDR(base_hi, base_lo));
670 
671 		lgrp_plat_node_memory[node].end =
672 		    btop(OPT_DRAMADDR(limit_hi, limit_lo) |
673 		    OPT_DRAMADDR_LO_MASK_OFF);
674 	}
675 
676 	/*
677 	 * Restore PCI Extended Configuration Space enable bit
678 	 */
679 	if (opt_family == AMD_FAMILY_GREYHOUND) {
680 		if ((nb_cfg_reg & AMD_GH_NB_CFG_EN_ECS) == 0)
681 			wrmsr(MSR_AMD_NB_CFG, nb_cfg_reg);
682 	}
683 
684 	/*
685 	 * Only use one memory node if memory is interleaved between any nodes
686 	 */
687 	if (lgrp_plat_mem_intrlv) {
688 		lgrp_plat_node_cnt = max_mem_nodes = 1;
689 		(void) lgrp_topo_ht_limit_set(1);
690 	} else {
691 		max_mem_nodes = lgrp_plat_node_cnt;
692 
693 		/*
694 		 * Probing errors can mess up the lgroup topology and force us
695 		 * fall back to a 2 level lgroup topology.  Here we bound how
696 		 * tall the lgroup topology can grow in hopes of avoiding any
697 		 * anamolies in probing from messing up the lgroup topology
698 		 * by limiting the accuracy of the latency topology.
699 		 *
700 		 * Assume that nodes will at least be configured in a ring,
701 		 * so limit height of lgroup topology to be less than number
702 		 * of nodes on a system with 4 or more nodes
703 		 */
704 		if (lgrp_plat_node_cnt >= 4 &&
705 		    lgrp_topo_ht_limit() == lgrp_topo_ht_limit_default())
706 			(void) lgrp_topo_ht_limit_set(lgrp_plat_node_cnt - 1);
707 	}
708 
709 	/*
710 	 * Lgroups on Opteron architectures have but a single physical
711 	 * processor. Tune lgrp_expand_proc_thresh and lgrp_expand_proc_diff
712 	 * so that lgrp_choose() will spread things out aggressively.
713 	 */
714 	lgrp_expand_proc_thresh = LGRP_LOADAVG_THREAD_MAX / 2;
715 	lgrp_expand_proc_diff = 0;
716 #endif	/* __xpv */
717 }
718 
719 
720 /*
721  * Latencies must be within 1/(2**LGRP_LAT_TOLERANCE_SHIFT) of each other to
722  * be considered same
723  */
724 #define	LGRP_LAT_TOLERANCE_SHIFT	4
725 
726 int	lgrp_plat_probe_lt_shift = LGRP_LAT_TOLERANCE_SHIFT;
727 
728 
729 /*
730  * Adjust latencies between nodes to be symmetric, normalize latencies between
731  * any nodes that are within some tolerance to be same, and make local
732  * latencies be same
733  */
734 static void
735 lgrp_plat_latency_adjust(void)
736 {
737 	int				i;
738 	int				j;
739 	int				k;
740 	int				l;
741 	u_longlong_t			max;
742 	u_longlong_t			min;
743 	u_longlong_t			t;
744 	u_longlong_t			t1;
745 	u_longlong_t			t2;
746 	const lgrp_config_flag_t	cflag = LGRP_CONFIG_LAT_CHANGE_ALL;
747 	int				lat_corrected[MAX_NODES][MAX_NODES];
748 
749 	/*
750 	 * Nothing to do when this is an UMA machine
751 	 */
752 	if (max_mem_nodes == 1)
753 		return;
754 
755 	/*
756 	 * Make sure that latencies are symmetric between any two nodes
757 	 * (ie. latency(node0, node1) == latency(node1, node0))
758 	 */
759 	for (i = 0; i < lgrp_plat_node_cnt; i++)
760 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
761 			t1 = lgrp_plat_probe_times[i][j];
762 			t2 = lgrp_plat_probe_times[j][i];
763 
764 			if (t1 == 0 || t2 == 0 || t1 == t2)
765 				continue;
766 
767 			/*
768 			 * Latencies should be same
769 			 * - Use minimum of two latencies which should be same
770 			 * - Track suspect probe times not within tolerance of
771 			 *   min value
772 			 * - Remember how much values are corrected by
773 			 */
774 			if (t1 > t2) {
775 				t = t2;
776 				lgrp_plat_probe_errors[i][j] += t1 - t2;
777 				if (t1 - t2 > t2 >> lgrp_plat_probe_lt_shift) {
778 					lgrp_plat_probe_suspect[i][j]++;
779 					lgrp_plat_probe_suspect[j][i]++;
780 				}
781 			} else if (t2 > t1) {
782 				t = t1;
783 				lgrp_plat_probe_errors[j][i] += t2 - t1;
784 				if (t2 - t1 > t1 >> lgrp_plat_probe_lt_shift) {
785 					lgrp_plat_probe_suspect[i][j]++;
786 					lgrp_plat_probe_suspect[j][i]++;
787 				}
788 			}
789 
790 			lgrp_plat_probe_times[i][j] =
791 			    lgrp_plat_probe_times[j][i] = t;
792 			lgrp_config(cflag, t1, t);
793 			lgrp_config(cflag, t2, t);
794 		}
795 
796 	/*
797 	 * Keep track of which latencies get corrected
798 	 */
799 	for (i = 0; i < MAX_NODES; i++)
800 		for (j = 0; j < MAX_NODES; j++)
801 			lat_corrected[i][j] = 0;
802 
803 	/*
804 	 * For every two nodes, see whether there is another pair of nodes which
805 	 * are about the same distance apart and make the latencies be the same
806 	 * if they are close enough together
807 	 */
808 	for (i = 0; i < lgrp_plat_node_cnt; i++)
809 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
810 			/*
811 			 * Pick one pair of nodes (i, j)
812 			 * and get latency between them
813 			 */
814 			t1 = lgrp_plat_probe_times[i][j];
815 
816 			/*
817 			 * Skip this pair of nodes if there isn't a latency
818 			 * for it yet
819 			 */
820 			if (t1 == 0)
821 				continue;
822 
823 			for (k = 0; k < lgrp_plat_node_cnt; k++)
824 				for (l = 0; l < lgrp_plat_node_cnt; l++) {
825 					/*
826 					 * Pick another pair of nodes (k, l)
827 					 * not same as (i, j) and get latency
828 					 * between them
829 					 */
830 					if (k == i && l == j)
831 						continue;
832 
833 					t2 = lgrp_plat_probe_times[k][l];
834 
835 					/*
836 					 * Skip this pair of nodes if there
837 					 * isn't a latency for it yet
838 					 */
839 
840 					if (t2 == 0)
841 						continue;
842 
843 					/*
844 					 * Skip nodes (k, l) if they already
845 					 * have same latency as (i, j) or
846 					 * their latency isn't close enough to
847 					 * be considered/made the same
848 					 */
849 					if (t1 == t2 || (t1 > t2 && t1 - t2 >
850 					    t1 >> lgrp_plat_probe_lt_shift) ||
851 					    (t2 > t1 && t2 - t1 >
852 					    t2 >> lgrp_plat_probe_lt_shift))
853 						continue;
854 
855 					/*
856 					 * Make latency(i, j) same as
857 					 * latency(k, l), try to use latency
858 					 * that has been adjusted already to get
859 					 * more consistency (if possible), and
860 					 * remember which latencies were
861 					 * adjusted for next time
862 					 */
863 					if (lat_corrected[i][j]) {
864 						t = t1;
865 						lgrp_config(cflag, t2, t);
866 						t2 = t;
867 					} else if (lat_corrected[k][l]) {
868 						t = t2;
869 						lgrp_config(cflag, t1, t);
870 						t1 = t;
871 					} else {
872 						if (t1 > t2)
873 							t = t2;
874 						else
875 							t = t1;
876 						lgrp_config(cflag, t1, t);
877 						lgrp_config(cflag, t2, t);
878 						t1 = t2 = t;
879 					}
880 
881 					lgrp_plat_probe_times[i][j] =
882 					    lgrp_plat_probe_times[k][l] = t;
883 
884 					lat_corrected[i][j] =
885 					    lat_corrected[k][l] = 1;
886 				}
887 		}
888 
889 	/*
890 	 * Local latencies should be same
891 	 * - Find min and max local latencies
892 	 * - Make all local latencies be minimum
893 	 */
894 	min = -1;
895 	max = 0;
896 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
897 		t = lgrp_plat_probe_times[i][i];
898 		if (t == 0)
899 			continue;
900 		if (min == -1 || t < min)
901 			min = t;
902 		if (t > max)
903 			max = t;
904 	}
905 	if (min != max) {
906 		for (i = 0; i < lgrp_plat_node_cnt; i++) {
907 			int	local;
908 
909 			local = lgrp_plat_probe_times[i][i];
910 			if (local == 0)
911 				continue;
912 
913 			/*
914 			 * Track suspect probe times that aren't within
915 			 * tolerance of minimum local latency and how much
916 			 * probe times are corrected by
917 			 */
918 			if (local - min > min >> lgrp_plat_probe_lt_shift)
919 				lgrp_plat_probe_suspect[i][i]++;
920 
921 			lgrp_plat_probe_errors[i][i] += local - min;
922 
923 			/*
924 			 * Make local latencies be minimum
925 			 */
926 			lgrp_config(LGRP_CONFIG_LAT_CHANGE, i, min);
927 			lgrp_plat_probe_times[i][i] = min;
928 		}
929 	}
930 
931 	/*
932 	 * Determine max probe time again since just adjusted latencies
933 	 */
934 	lgrp_plat_probe_time_max = 0;
935 	for (i = 0; i < lgrp_plat_node_cnt; i++)
936 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
937 			t = lgrp_plat_probe_times[i][j];
938 			if (t > lgrp_plat_probe_time_max)
939 				lgrp_plat_probe_time_max = t;
940 		}
941 }
942 
943 
944 /*
945  * Verify following about latencies between nodes:
946  *
947  * - Latencies should be symmetric (ie. latency(a, b) == latency(b, a))
948  * - Local latencies same
949  * - Local < remote
950  * - Number of latencies seen is reasonable
951  * - Number of occurrences of a given latency should be more than 1
952  *
953  * Returns:
954  *	0	Success
955  *	-1	Not symmetric
956  *	-2	Local latencies not same
957  *	-3	Local >= remote
958  *	-4	Wrong number of latencies
959  *	-5	Not enough occurrences of given latency
960  */
961 static int
962 lgrp_plat_latency_verify(void)
963 {
964 	int				i;
965 	int				j;
966 	lgrp_plat_latency_acct_t	*l;
967 	int				probed;
968 	u_longlong_t			t1;
969 	u_longlong_t			t2;
970 
971 	/*
972 	 * Nothing to do when this is an UMA machine, lgroup topology is
973 	 * limited to 2 levels, or there aren't any probe times yet
974 	 */
975 	if (max_mem_nodes == 1 || lgrp_topo_levels < 2 ||
976 	    (lgrp_plat_probe_time_max == 0 && lgrp_plat_probe_time_min == -1))
977 		return (0);
978 
979 	/*
980 	 * Make sure that latencies are symmetric between any two nodes
981 	 * (ie. latency(node0, node1) == latency(node1, node0))
982 	 */
983 	for (i = 0; i < lgrp_plat_node_cnt; i++)
984 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
985 			t1 = lgrp_plat_probe_times[i][j];
986 			t2 = lgrp_plat_probe_times[j][i];
987 
988 			if (t1 == 0 || t2 == 0 || t1 == t2)
989 				continue;
990 
991 			return (-1);
992 		}
993 
994 	/*
995 	 * Local latencies should be same
996 	 */
997 	t1 = lgrp_plat_probe_times[0][0];
998 	for (i = 1; i < lgrp_plat_node_cnt; i++) {
999 		t2 = lgrp_plat_probe_times[i][i];
1000 		if (t2 == 0)
1001 			continue;
1002 
1003 		if (t1 == 0) {
1004 			t1 = t2;
1005 			continue;
1006 		}
1007 
1008 		if (t1 != t2)
1009 			return (-2);
1010 	}
1011 
1012 	/*
1013 	 * Local latencies should be less than remote
1014 	 */
1015 	if (t1) {
1016 		for (i = 0; i < lgrp_plat_node_cnt; i++)
1017 			for (j = 0; j < lgrp_plat_node_cnt; j++) {
1018 				t2 = lgrp_plat_probe_times[i][j];
1019 				if (i == j || t2 == 0)
1020 					continue;
1021 
1022 				if (t1 >= t2)
1023 					return (-3);
1024 			}
1025 	}
1026 
1027 	/*
1028 	 * Rest of checks are not very useful for machines with less than
1029 	 * 4 nodes (which means less than 3 latencies on Opteron)
1030 	 */
1031 	if (lgrp_plat_node_cnt < 4)
1032 		return (0);
1033 
1034 	/*
1035 	 * Need to see whether done probing in order to verify number of
1036 	 * latencies are correct
1037 	 */
1038 	probed = 0;
1039 	for (i = 0; i < lgrp_plat_node_cnt; i++)
1040 		if (lgrp_plat_probe_times[i][i])
1041 			probed++;
1042 
1043 	if (probed != lgrp_plat_node_cnt)
1044 		return (0);
1045 
1046 	/*
1047 	 * Determine number of unique latencies seen in probe times,
1048 	 * their values, and number of occurrences of each
1049 	 */
1050 	lgrp_plat_probe_nlatencies = 0;
1051 	bzero(lgrp_plat_probe_lat_acct,
1052 	    MAX_NODES * sizeof (lgrp_plat_latency_acct_t));
1053 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1054 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1055 			int	k;
1056 
1057 			/*
1058 			 * Look at each probe time
1059 			 */
1060 			t1 = lgrp_plat_probe_times[i][j];
1061 			if (t1 == 0)
1062 				continue;
1063 
1064 			/*
1065 			 * Account for unique latencies
1066 			 */
1067 			for (k = 0; k < lgrp_plat_node_cnt; k++) {
1068 				l = &lgrp_plat_probe_lat_acct[k];
1069 				if (t1 == l->la_value) {
1070 					/*
1071 					 * Increment number of occurrences
1072 					 * if seen before
1073 					 */
1074 					l->la_count++;
1075 					break;
1076 				} else if (l->la_value == 0) {
1077 					/*
1078 					 * Record latency if haven't seen before
1079 					 */
1080 					l->la_value = t1;
1081 					l->la_count++;
1082 					lgrp_plat_probe_nlatencies++;
1083 					break;
1084 				}
1085 			}
1086 		}
1087 	}
1088 
1089 	/*
1090 	 * Number of latencies should be relative to number of
1091 	 * nodes in system:
1092 	 * - Same as nodes when nodes <= 2
1093 	 * - Less than nodes when nodes > 2
1094 	 * - Greater than 2 when nodes >= 4
1095 	 */
1096 	if ((lgrp_plat_node_cnt <= 2 &&
1097 	    lgrp_plat_probe_nlatencies != lgrp_plat_node_cnt) ||
1098 	    (lgrp_plat_node_cnt > 2 &&
1099 	    lgrp_plat_probe_nlatencies >= lgrp_plat_node_cnt) ||
1100 	    (lgrp_plat_node_cnt >= 4 && lgrp_topo_levels >= 3 &&
1101 	    lgrp_plat_probe_nlatencies <= 2))
1102 		return (-4);
1103 
1104 	/*
1105 	 * There should be more than one occurrence of every latency
1106 	 * as long as probing is complete
1107 	 */
1108 	for (i = 0; i < lgrp_plat_probe_nlatencies; i++) {
1109 		l = &lgrp_plat_probe_lat_acct[i];
1110 		if (l->la_count <= 1)
1111 			return (-5);
1112 	}
1113 	return (0);
1114 }
1115 
1116 
1117 /*
1118  * Set lgroup latencies for 2 level lgroup topology
1119  */
1120 static void
1121 lgrp_plat_2level_setup(void)
1122 {
1123 	int	i;
1124 
1125 	if (lgrp_plat_node_cnt >= 4)
1126 		cmn_err(CE_NOTE,
1127 		    "MPO only optimizing for local and remote\n");
1128 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1129 		int	j;
1130 
1131 		for (j = 0; j < lgrp_plat_node_cnt; j++) {
1132 			if (i == j)
1133 				lgrp_plat_probe_times[i][j] = 2;
1134 			else
1135 				lgrp_plat_probe_times[i][j] = 3;
1136 		}
1137 	}
1138 	lgrp_plat_probe_time_min = 2;
1139 	lgrp_plat_probe_time_max = 3;
1140 	lgrp_config(LGRP_CONFIG_FLATTEN, 2, 0);
1141 }
1142 
1143 
1144 /*
1145  * Return time needed to probe from current CPU to memory in given node
1146  */
1147 static hrtime_t
1148 lgrp_plat_probe_time(int to)
1149 {
1150 	caddr_t		buf;
1151 	uint_t		dev;
1152 	/* LINTED: set but not used in function */
1153 	volatile uint_t	dev_vendor;
1154 	hrtime_t	elapsed;
1155 	hrtime_t	end;
1156 	int		from;
1157 	int		i;
1158 	int		ipl;
1159 	hrtime_t	max;
1160 	hrtime_t	min;
1161 	hrtime_t	start;
1162 	int		cnt;
1163 	extern int	use_sse_pagecopy;
1164 
1165 	/*
1166 	 * Determine ID of node containing current CPU
1167 	 */
1168 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1169 
1170 	/*
1171 	 * Do common work for probing main memory
1172 	 */
1173 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_PGCPY) {
1174 		/*
1175 		 * Skip probing any nodes without memory and
1176 		 * set probe time to 0
1177 		 */
1178 		if (lgrp_plat_probe_memory[to] == NULL) {
1179 			lgrp_plat_probe_times[from][to] = 0;
1180 			return (0);
1181 		}
1182 
1183 		/*
1184 		 * Invalidate caches once instead of once every sample
1185 		 * which should cut cost of probing by a lot
1186 		 */
1187 		lgrp_plat_flush_cost = gethrtime();
1188 		invalidate_cache();
1189 		lgrp_plat_flush_cost = gethrtime() - lgrp_plat_flush_cost;
1190 		lgrp_plat_probe_cost_total += lgrp_plat_flush_cost;
1191 	}
1192 
1193 	/*
1194 	 * Probe from current CPU to given memory using specified operation
1195 	 * and take specified number of samples
1196 	 */
1197 	max = 0;
1198 	min = -1;
1199 	for (i = 0; i < lgrp_plat_probe_nsamples; i++) {
1200 		lgrp_plat_probe_cost = gethrtime();
1201 
1202 		/*
1203 		 * Can't measure probe time if gethrtime() isn't working yet
1204 		 */
1205 		if (lgrp_plat_probe_cost == 0 && gethrtime() == 0)
1206 			return (0);
1207 
1208 		switch (lgrp_plat_probe_op) {
1209 
1210 		case LGRP_PLAT_PROBE_PGCPY:
1211 		default:
1212 			/*
1213 			 * Measure how long it takes to copy page
1214 			 * on top of itself
1215 			 */
1216 			buf = lgrp_plat_probe_memory[to] + (i * PAGESIZE);
1217 
1218 			kpreempt_disable();
1219 			ipl = splhigh();
1220 			start = gethrtime();
1221 			if (use_sse_pagecopy)
1222 				hwblkpagecopy(buf, buf);
1223 			else
1224 				bcopy(buf, buf, PAGESIZE);
1225 			end = gethrtime();
1226 			elapsed = end - start;
1227 			splx(ipl);
1228 			kpreempt_enable();
1229 			break;
1230 
1231 		case LGRP_PLAT_PROBE_VENDOR:
1232 			/*
1233 			 * Measure how long it takes to read vendor ID from
1234 			 * Northbridge
1235 			 */
1236 			dev = OPT_PCS_DEV_NODE0 + to;
1237 			kpreempt_disable();
1238 			ipl = spl8();
1239 			outl(PCI_CONFADD, PCI_CADDR1(0, dev, opt_probe_func,
1240 			    OPT_PCS_OFF_VENDOR));
1241 			start = gethrtime();
1242 			for (cnt = 0; cnt < lgrp_plat_probe_nreads; cnt++)
1243 				dev_vendor = inl(PCI_CONFDATA);
1244 			end = gethrtime();
1245 			elapsed = (end - start) / lgrp_plat_probe_nreads;
1246 			splx(ipl);
1247 			kpreempt_enable();
1248 			break;
1249 		}
1250 
1251 		lgrp_plat_probe_cost = gethrtime() - lgrp_plat_probe_cost;
1252 		lgrp_plat_probe_cost_total += lgrp_plat_probe_cost;
1253 
1254 		if (min == -1 || elapsed < min)
1255 			min = elapsed;
1256 		if (elapsed > max)
1257 			max = elapsed;
1258 	}
1259 
1260 	/*
1261 	 * Update minimum and maximum probe times between
1262 	 * these two nodes
1263 	 */
1264 	if (min < lgrp_plat_probe_min[from][to] ||
1265 	    lgrp_plat_probe_min[from][to] == 0)
1266 		lgrp_plat_probe_min[from][to] = min;
1267 
1268 	if (max > lgrp_plat_probe_max[from][to])
1269 		lgrp_plat_probe_max[from][to] = max;
1270 
1271 	return (min);
1272 }
1273 
1274 
1275 /*
1276  * Probe memory in each node from current CPU to determine latency topology
1277  */
1278 void
1279 lgrp_plat_probe(void)
1280 {
1281 	int		from;
1282 	int		i;
1283 	hrtime_t	probe_time;
1284 	int		to;
1285 
1286 	if (max_mem_nodes == 1 || lgrp_topo_ht_limit() <= 2)
1287 		return;
1288 
1289 	/*
1290 	 * Determine ID of node containing current CPU
1291 	 */
1292 	from = LGRP_PLAT_CPU_TO_NODE(CPU);
1293 
1294 	/*
1295 	 * Don't need to probe if got times already
1296 	 */
1297 	if (lgrp_plat_probe_times[from][from] != 0)
1298 		return;
1299 
1300 	/*
1301 	 * Read vendor ID in Northbridge or read and write page(s)
1302 	 * in each node from current CPU and remember how long it takes,
1303 	 * so we can build latency topology of machine later.
1304 	 * This should approximate the memory latency between each node.
1305 	 */
1306 	for (i = 0; i < lgrp_plat_probe_nrounds; i++)
1307 		for (to = 0; to < lgrp_plat_node_cnt; to++) {
1308 			/*
1309 			 * Get probe time and bail out if can't get it yet
1310 			 */
1311 			probe_time = lgrp_plat_probe_time(to);
1312 			if (probe_time == 0)
1313 				return;
1314 
1315 			/*
1316 			 * Keep lowest probe time as latency between nodes
1317 			 */
1318 			if (lgrp_plat_probe_times[from][to] == 0 ||
1319 			    probe_time < lgrp_plat_probe_times[from][to])
1320 				lgrp_plat_probe_times[from][to] = probe_time;
1321 
1322 			/*
1323 			 * Update overall minimum and maximum probe times
1324 			 * across all nodes
1325 			 */
1326 			if (probe_time < lgrp_plat_probe_time_min ||
1327 			    lgrp_plat_probe_time_min == -1)
1328 				lgrp_plat_probe_time_min = probe_time;
1329 			if (probe_time > lgrp_plat_probe_time_max)
1330 				lgrp_plat_probe_time_max = probe_time;
1331 		}
1332 
1333 	/*
1334 	 * - Fix up latencies such that local latencies are same,
1335 	 *   latency(i, j) == latency(j, i), etc. (if possible)
1336 	 *
1337 	 * - Verify that latencies look ok
1338 	 *
1339 	 * - Fallback to just optimizing for local and remote if
1340 	 *   latencies didn't look right
1341 	 */
1342 	lgrp_plat_latency_adjust();
1343 	lgrp_plat_probe_error_code = lgrp_plat_latency_verify();
1344 	if (lgrp_plat_probe_error_code)
1345 		lgrp_plat_2level_setup();
1346 }
1347 
1348 
1349 /*
1350  * Platform-specific initialization
1351  */
1352 void
1353 lgrp_plat_main_init(void)
1354 {
1355 	int	curnode;
1356 	int	ht_limit;
1357 	int	i;
1358 
1359 	/*
1360 	 * Print a notice that MPO is disabled when memory is interleaved
1361 	 * across nodes....Would do this when it is discovered, but can't
1362 	 * because it happens way too early during boot....
1363 	 */
1364 	if (lgrp_plat_mem_intrlv)
1365 		cmn_err(CE_NOTE,
1366 		    "MPO disabled because memory is interleaved\n");
1367 
1368 	/*
1369 	 * Don't bother to do any probing if there is only one node or the
1370 	 * height of the lgroup topology less than or equal to 2
1371 	 */
1372 	ht_limit = lgrp_topo_ht_limit();
1373 	if (max_mem_nodes == 1 || ht_limit <= 2) {
1374 		/*
1375 		 * Setup lgroup latencies for 2 level lgroup topology
1376 		 * (ie. local and remote only) if they haven't been set yet
1377 		 */
1378 		if (ht_limit == 2 && lgrp_plat_probe_time_min == -1 &&
1379 		    lgrp_plat_probe_time_max == 0)
1380 			lgrp_plat_2level_setup();
1381 		return;
1382 	}
1383 
1384 	if (lgrp_plat_probe_op == LGRP_PLAT_PROBE_VENDOR) {
1385 		/*
1386 		 * Should have been able to probe from CPU 0 when it was added
1387 		 * to lgroup hierarchy, but may not have been able to then
1388 		 * because it happens so early in boot that gethrtime() hasn't
1389 		 * been initialized.  (:-(
1390 		 */
1391 		curnode = LGRP_PLAT_CPU_TO_NODE(CPU);
1392 		if (lgrp_plat_probe_times[curnode][curnode] == 0)
1393 			lgrp_plat_probe();
1394 
1395 		return;
1396 	}
1397 
1398 	/*
1399 	 * When probing memory, use one page for every sample to determine
1400 	 * lgroup topology and taking multiple samples
1401 	 */
1402 	if (lgrp_plat_probe_memsize == 0)
1403 		lgrp_plat_probe_memsize = PAGESIZE *
1404 		    lgrp_plat_probe_nsamples;
1405 
1406 	/*
1407 	 * Map memory in each node needed for probing to determine latency
1408 	 * topology
1409 	 */
1410 	for (i = 0; i < lgrp_plat_node_cnt; i++) {
1411 		int	mnode;
1412 
1413 		/*
1414 		 * Skip this node and leave its probe page NULL
1415 		 * if it doesn't have any memory
1416 		 */
1417 		mnode = plat_lgrphand_to_mem_node((lgrp_handle_t)i);
1418 		if (!mem_node_config[mnode].exists) {
1419 			lgrp_plat_probe_memory[i] = NULL;
1420 			continue;
1421 		}
1422 
1423 		/*
1424 		 * Allocate one kernel virtual page
1425 		 */
1426 		lgrp_plat_probe_memory[i] = vmem_alloc(heap_arena,
1427 		    lgrp_plat_probe_memsize, VM_NOSLEEP);
1428 		if (lgrp_plat_probe_memory[i] == NULL) {
1429 			cmn_err(CE_WARN,
1430 			    "lgrp_plat_main_init: couldn't allocate memory");
1431 			return;
1432 		}
1433 
1434 		/*
1435 		 * Map virtual page to first page in node
1436 		 */
1437 		hat_devload(kas.a_hat, lgrp_plat_probe_memory[i],
1438 		    lgrp_plat_probe_memsize,
1439 		    lgrp_plat_probe_pfn[i],
1440 		    PROT_READ | PROT_WRITE | HAT_PLAT_NOCACHE,
1441 		    HAT_LOAD_NOCONSIST);
1442 	}
1443 
1444 	/*
1445 	 * Probe from current CPU
1446 	 */
1447 	lgrp_plat_probe();
1448 }
1449 
1450 /*
1451  * Allocate additional space for an lgroup.
1452  */
1453 /* ARGSUSED */
1454 lgrp_t *
1455 lgrp_plat_alloc(lgrp_id_t lgrpid)
1456 {
1457 	lgrp_t *lgrp;
1458 
1459 	lgrp = &lgrp_space[nlgrps_alloc++];
1460 	if (lgrpid >= NLGRP || nlgrps_alloc > NLGRP)
1461 		return (NULL);
1462 	return (lgrp);
1463 }
1464 
1465 /*
1466  * Platform handling for (re)configuration changes
1467  */
1468 /* ARGSUSED */
1469 void
1470 lgrp_plat_config(lgrp_config_flag_t flag, uintptr_t arg)
1471 {
1472 }
1473 
1474 /*
1475  * Return the platform handle for the lgroup containing the given CPU
1476  */
1477 /* ARGSUSED */
1478 lgrp_handle_t
1479 lgrp_plat_cpu_to_hand(processorid_t id)
1480 {
1481 	if (lgrp_plat_node_cnt == 1)
1482 		return (LGRP_DEFAULT_HANDLE);
1483 
1484 	return ((lgrp_handle_t)LGRP_PLAT_CPU_TO_NODE(cpu[id]));
1485 }
1486 
1487 /*
1488  * Return the platform handle of the lgroup that contains the physical memory
1489  * corresponding to the given page frame number
1490  */
1491 /* ARGSUSED */
1492 lgrp_handle_t
1493 lgrp_plat_pfn_to_hand(pfn_t pfn)
1494 {
1495 	int	mnode;
1496 
1497 	if (max_mem_nodes == 1)
1498 		return (LGRP_DEFAULT_HANDLE);
1499 
1500 	if (pfn > physmax)
1501 		return (LGRP_NULL_HANDLE);
1502 
1503 	mnode = plat_pfn_to_mem_node(pfn);
1504 	if (mnode < 0)
1505 		return (LGRP_NULL_HANDLE);
1506 
1507 	return (MEM_NODE_2_LGRPHAND(mnode));
1508 }
1509 
1510 /*
1511  * Return the maximum number of lgrps supported by the platform.
1512  * Before lgrp topology is known it returns an estimate based on the number of
1513  * nodes. Once topology is known it returns the actual maximim number of lgrps
1514  * created. Since x86 doesn't support dynamic addition of new nodes, this number
1515  * may not grow during system lifetime.
1516  */
1517 int
1518 lgrp_plat_max_lgrps()
1519 {
1520 	return (lgrp_topo_initialized ?
1521 	    lgrp_alloc_max + 1 :
1522 	    lgrp_plat_node_cnt * (lgrp_plat_node_cnt - 1) + 1);
1523 }
1524 
1525 /*
1526  * Return the number of free, allocatable, or installed
1527  * pages in an lgroup
1528  * This is a copy of the MAX_MEM_NODES == 1 version of the routine
1529  * used when MPO is disabled (i.e. single lgroup) or this is the root lgroup
1530  */
1531 /* ARGSUSED */
1532 static pgcnt_t
1533 lgrp_plat_mem_size_default(lgrp_handle_t lgrphand, lgrp_mem_query_t query)
1534 {
1535 	struct memlist *mlist;
1536 	pgcnt_t npgs = 0;
1537 	extern struct memlist *phys_avail;
1538 	extern struct memlist *phys_install;
1539 
1540 	switch (query) {
1541 	case LGRP_MEM_SIZE_FREE:
1542 		return ((pgcnt_t)freemem);
1543 	case LGRP_MEM_SIZE_AVAIL:
1544 		memlist_read_lock();
1545 		for (mlist = phys_avail; mlist; mlist = mlist->next)
1546 			npgs += btop(mlist->size);
1547 		memlist_read_unlock();
1548 		return (npgs);
1549 	case LGRP_MEM_SIZE_INSTALL:
1550 		memlist_read_lock();
1551 		for (mlist = phys_install; mlist; mlist = mlist->next)
1552 			npgs += btop(mlist->size);
1553 		memlist_read_unlock();
1554 		return (npgs);
1555 	default:
1556 		return ((pgcnt_t)0);
1557 	}
1558 }
1559 
1560 /*
1561  * Return the number of free pages in an lgroup.
1562  *
1563  * For query of LGRP_MEM_SIZE_FREE, return the number of base pagesize
1564  * pages on freelists.  For query of LGRP_MEM_SIZE_AVAIL, return the
1565  * number of allocatable base pagesize pages corresponding to the
1566  * lgroup (e.g. do not include page_t's, BOP_ALLOC()'ed memory, ..)
1567  * For query of LGRP_MEM_SIZE_INSTALL, return the amount of physical
1568  * memory installed, regardless of whether or not it's usable.
1569  */
1570 pgcnt_t
1571 lgrp_plat_mem_size(lgrp_handle_t plathand, lgrp_mem_query_t query)
1572 {
1573 	int	mnode;
1574 	pgcnt_t npgs = (pgcnt_t)0;
1575 	extern struct memlist *phys_avail;
1576 	extern struct memlist *phys_install;
1577 
1578 
1579 	if (plathand == LGRP_DEFAULT_HANDLE)
1580 		return (lgrp_plat_mem_size_default(plathand, query));
1581 
1582 	if (plathand != LGRP_NULL_HANDLE) {
1583 		mnode = plat_lgrphand_to_mem_node(plathand);
1584 		if (mnode >= 0 && mem_node_config[mnode].exists) {
1585 			switch (query) {
1586 			case LGRP_MEM_SIZE_FREE:
1587 				npgs = MNODE_PGCNT(mnode);
1588 				break;
1589 			case LGRP_MEM_SIZE_AVAIL:
1590 				npgs = mem_node_memlist_pages(mnode,
1591 				    phys_avail);
1592 				break;
1593 			case LGRP_MEM_SIZE_INSTALL:
1594 				npgs = mem_node_memlist_pages(mnode,
1595 				    phys_install);
1596 				break;
1597 			default:
1598 				break;
1599 			}
1600 		}
1601 	}
1602 	return (npgs);
1603 }
1604 
1605 /*
1606  * Return latency between "from" and "to" lgroups
1607  *
1608  * This latency number can only be used for relative comparison
1609  * between lgroups on the running system, cannot be used across platforms,
1610  * and may not reflect the actual latency.  It is platform and implementation
1611  * specific, so platform gets to decide its value.  It would be nice if the
1612  * number was at least proportional to make comparisons more meaningful though.
1613  */
1614 /* ARGSUSED */
1615 int
1616 lgrp_plat_latency(lgrp_handle_t from, lgrp_handle_t to)
1617 {
1618 	lgrp_handle_t	src, dest;
1619 
1620 	if (max_mem_nodes == 1)
1621 		return (0);
1622 
1623 	/*
1624 	 * Return max latency for root lgroup
1625 	 */
1626 	if (from == LGRP_DEFAULT_HANDLE || to == LGRP_DEFAULT_HANDLE)
1627 		return (lgrp_plat_probe_time_max);
1628 
1629 	src = from;
1630 	dest = to;
1631 
1632 	/*
1633 	 * Return 0 for nodes (lgroup platform handles) out of range
1634 	 */
1635 	if (src < 0 || src >= MAX_NODES || dest < 0 || dest >= MAX_NODES)
1636 		return (0);
1637 
1638 	/*
1639 	 * Probe from current CPU if its lgroup latencies haven't been set yet
1640 	 * and we are trying to get latency from current CPU to some node
1641 	 */
1642 	if (lgrp_plat_probe_times[src][src] == 0 &&
1643 	    LGRP_PLAT_CPU_TO_NODE(CPU) == src)
1644 		lgrp_plat_probe();
1645 
1646 	return (lgrp_plat_probe_times[src][dest]);
1647 }
1648 
1649 /*
1650  * Return platform handle for root lgroup
1651  */
1652 lgrp_handle_t
1653 lgrp_plat_root_hand(void)
1654 {
1655 	return (LGRP_DEFAULT_HANDLE);
1656 }
1657