xref: /openbmc/linux/arch/sparc/kernel/cpumap.c (revision 8fdff1dc)
1 /* cpumap.c: used for optimizing CPU assignment
2  *
3  * Copyright (C) 2009 Hong H. Pham <hong.pham@windriver.com>
4  */
5 
6 #include <linux/export.h>
7 #include <linux/slab.h>
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10 #include <linux/cpumask.h>
11 #include <linux/spinlock.h>
12 #include <asm/cpudata.h>
13 #include "cpumap.h"
14 
15 
16 enum {
17 	CPUINFO_LVL_ROOT = 0,
18 	CPUINFO_LVL_NODE,
19 	CPUINFO_LVL_CORE,
20 	CPUINFO_LVL_PROC,
21 	CPUINFO_LVL_MAX,
22 };
23 
24 enum {
25 	ROVER_NO_OP              = 0,
26 	/* Increment rover every time level is visited */
27 	ROVER_INC_ON_VISIT       = 1 << 0,
28 	/* Increment parent's rover every time rover wraps around */
29 	ROVER_INC_PARENT_ON_LOOP = 1 << 1,
30 };
31 
32 struct cpuinfo_node {
33 	int id;
34 	int level;
35 	int num_cpus;    /* Number of CPUs in this hierarchy */
36 	int parent_index;
37 	int child_start; /* Array index of the first child node */
38 	int child_end;   /* Array index of the last child node */
39 	int rover;       /* Child node iterator */
40 };
41 
42 struct cpuinfo_level {
43 	int start_index; /* Index of first node of a level in a cpuinfo tree */
44 	int end_index;   /* Index of last node of a level in a cpuinfo tree */
45 	int num_nodes;   /* Number of nodes in a level in a cpuinfo tree */
46 };
47 
48 struct cpuinfo_tree {
49 	int total_nodes;
50 
51 	/* Offsets into nodes[] for each level of the tree */
52 	struct cpuinfo_level level[CPUINFO_LVL_MAX];
53 	struct cpuinfo_node  nodes[0];
54 };
55 
56 
57 static struct cpuinfo_tree *cpuinfo_tree;
58 
59 static u16 cpu_distribution_map[NR_CPUS];
60 static DEFINE_SPINLOCK(cpu_map_lock);
61 
62 
63 /* Niagara optimized cpuinfo tree traversal. */
64 static const int niagara_iterate_method[] = {
65 	[CPUINFO_LVL_ROOT] = ROVER_NO_OP,
66 
67 	/* Strands (or virtual CPUs) within a core may not run concurrently
68 	 * on the Niagara, as instruction pipeline(s) are shared.  Distribute
69 	 * work to strands in different cores first for better concurrency.
70 	 * Go to next NUMA node when all cores are used.
71 	 */
72 	[CPUINFO_LVL_NODE] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
73 
74 	/* Strands are grouped together by proc_id in cpuinfo_sparc, i.e.
75 	 * a proc_id represents an instruction pipeline.  Distribute work to
76 	 * strands in different proc_id groups if the core has multiple
77 	 * instruction pipelines (e.g. the Niagara 2/2+ has two).
78 	 */
79 	[CPUINFO_LVL_CORE] = ROVER_INC_ON_VISIT,
80 
81 	/* Pick the next strand in the proc_id group. */
82 	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT,
83 };
84 
85 /* Generic cpuinfo tree traversal.  Distribute work round robin across NUMA
86  * nodes.
87  */
88 static const int generic_iterate_method[] = {
89 	[CPUINFO_LVL_ROOT] = ROVER_INC_ON_VISIT,
90 	[CPUINFO_LVL_NODE] = ROVER_NO_OP,
91 	[CPUINFO_LVL_CORE] = ROVER_INC_PARENT_ON_LOOP,
92 	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
93 };
94 
95 
96 static int cpuinfo_id(int cpu, int level)
97 {
98 	int id;
99 
100 	switch (level) {
101 	case CPUINFO_LVL_ROOT:
102 		id = 0;
103 		break;
104 	case CPUINFO_LVL_NODE:
105 		id = cpu_to_node(cpu);
106 		break;
107 	case CPUINFO_LVL_CORE:
108 		id = cpu_data(cpu).core_id;
109 		break;
110 	case CPUINFO_LVL_PROC:
111 		id = cpu_data(cpu).proc_id;
112 		break;
113 	default:
114 		id = -EINVAL;
115 	}
116 	return id;
117 }
118 
119 /*
120  * Enumerate the CPU information in __cpu_data to determine the start index,
121  * end index, and number of nodes for each level in the cpuinfo tree.  The
122  * total number of cpuinfo nodes required to build the tree is returned.
123  */
124 static int enumerate_cpuinfo_nodes(struct cpuinfo_level *tree_level)
125 {
126 	int prev_id[CPUINFO_LVL_MAX];
127 	int i, n, num_nodes;
128 
129 	for (i = CPUINFO_LVL_ROOT; i < CPUINFO_LVL_MAX; i++) {
130 		struct cpuinfo_level *lv = &tree_level[i];
131 
132 		prev_id[i] = -1;
133 		lv->start_index = lv->end_index = lv->num_nodes = 0;
134 	}
135 
136 	num_nodes = 1; /* Include the root node */
137 
138 	for (i = 0; i < num_possible_cpus(); i++) {
139 		if (!cpu_online(i))
140 			continue;
141 
142 		n = cpuinfo_id(i, CPUINFO_LVL_NODE);
143 		if (n > prev_id[CPUINFO_LVL_NODE]) {
144 			tree_level[CPUINFO_LVL_NODE].num_nodes++;
145 			prev_id[CPUINFO_LVL_NODE] = n;
146 			num_nodes++;
147 		}
148 		n = cpuinfo_id(i, CPUINFO_LVL_CORE);
149 		if (n > prev_id[CPUINFO_LVL_CORE]) {
150 			tree_level[CPUINFO_LVL_CORE].num_nodes++;
151 			prev_id[CPUINFO_LVL_CORE] = n;
152 			num_nodes++;
153 		}
154 		n = cpuinfo_id(i, CPUINFO_LVL_PROC);
155 		if (n > prev_id[CPUINFO_LVL_PROC]) {
156 			tree_level[CPUINFO_LVL_PROC].num_nodes++;
157 			prev_id[CPUINFO_LVL_PROC] = n;
158 			num_nodes++;
159 		}
160 	}
161 
162 	tree_level[CPUINFO_LVL_ROOT].num_nodes = 1;
163 
164 	n = tree_level[CPUINFO_LVL_NODE].num_nodes;
165 	tree_level[CPUINFO_LVL_NODE].start_index = 1;
166 	tree_level[CPUINFO_LVL_NODE].end_index   = n;
167 
168 	n++;
169 	tree_level[CPUINFO_LVL_CORE].start_index = n;
170 	n += tree_level[CPUINFO_LVL_CORE].num_nodes;
171 	tree_level[CPUINFO_LVL_CORE].end_index   = n - 1;
172 
173 	tree_level[CPUINFO_LVL_PROC].start_index = n;
174 	n += tree_level[CPUINFO_LVL_PROC].num_nodes;
175 	tree_level[CPUINFO_LVL_PROC].end_index   = n - 1;
176 
177 	return num_nodes;
178 }
179 
180 /* Build a tree representation of the CPU hierarchy using the per CPU
181  * information in __cpu_data.  Entries in __cpu_data[0..NR_CPUS] are
182  * assumed to be sorted in ascending order based on node, core_id, and
183  * proc_id (in order of significance).
184  */
185 static struct cpuinfo_tree *build_cpuinfo_tree(void)
186 {
187 	struct cpuinfo_tree *new_tree;
188 	struct cpuinfo_node *node;
189 	struct cpuinfo_level tmp_level[CPUINFO_LVL_MAX];
190 	int num_cpus[CPUINFO_LVL_MAX];
191 	int level_rover[CPUINFO_LVL_MAX];
192 	int prev_id[CPUINFO_LVL_MAX];
193 	int n, id, cpu, prev_cpu, last_cpu, level;
194 
195 	n = enumerate_cpuinfo_nodes(tmp_level);
196 
197 	new_tree = kzalloc(sizeof(struct cpuinfo_tree) +
198 	                   (sizeof(struct cpuinfo_node) * n), GFP_ATOMIC);
199 	if (!new_tree)
200 		return NULL;
201 
202 	new_tree->total_nodes = n;
203 	memcpy(&new_tree->level, tmp_level, sizeof(tmp_level));
204 
205 	prev_cpu = cpu = cpumask_first(cpu_online_mask);
206 
207 	/* Initialize all levels in the tree with the first CPU */
208 	for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT; level--) {
209 		n = new_tree->level[level].start_index;
210 
211 		level_rover[level] = n;
212 		node = &new_tree->nodes[n];
213 
214 		id = cpuinfo_id(cpu, level);
215 		if (unlikely(id < 0)) {
216 			kfree(new_tree);
217 			return NULL;
218 		}
219 		node->id = id;
220 		node->level = level;
221 		node->num_cpus = 1;
222 
223 		node->parent_index = (level > CPUINFO_LVL_ROOT)
224 		    ? new_tree->level[level - 1].start_index : -1;
225 
226 		node->child_start = node->child_end = node->rover =
227 		    (level == CPUINFO_LVL_PROC)
228 		    ? cpu : new_tree->level[level + 1].start_index;
229 
230 		prev_id[level] = node->id;
231 		num_cpus[level] = 1;
232 	}
233 
234 	for (last_cpu = (num_possible_cpus() - 1); last_cpu >= 0; last_cpu--) {
235 		if (cpu_online(last_cpu))
236 			break;
237 	}
238 
239 	while (++cpu <= last_cpu) {
240 		if (!cpu_online(cpu))
241 			continue;
242 
243 		for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT;
244 		     level--) {
245 			id = cpuinfo_id(cpu, level);
246 			if (unlikely(id < 0)) {
247 				kfree(new_tree);
248 				return NULL;
249 			}
250 
251 			if ((id != prev_id[level]) || (cpu == last_cpu)) {
252 				prev_id[level] = id;
253 				node = &new_tree->nodes[level_rover[level]];
254 				node->num_cpus = num_cpus[level];
255 				num_cpus[level] = 1;
256 
257 				if (cpu == last_cpu)
258 					node->num_cpus++;
259 
260 				/* Connect tree node to parent */
261 				if (level == CPUINFO_LVL_ROOT)
262 					node->parent_index = -1;
263 				else
264 					node->parent_index =
265 					    level_rover[level - 1];
266 
267 				if (level == CPUINFO_LVL_PROC) {
268 					node->child_end =
269 					    (cpu == last_cpu) ? cpu : prev_cpu;
270 				} else {
271 					node->child_end =
272 					    level_rover[level + 1] - 1;
273 				}
274 
275 				/* Initialize the next node in the same level */
276 				n = ++level_rover[level];
277 				if (n <= new_tree->level[level].end_index) {
278 					node = &new_tree->nodes[n];
279 					node->id = id;
280 					node->level = level;
281 
282 					/* Connect node to child */
283 					node->child_start = node->child_end =
284 					node->rover =
285 					    (level == CPUINFO_LVL_PROC)
286 					    ? cpu : level_rover[level + 1];
287 				}
288 			} else
289 				num_cpus[level]++;
290 		}
291 		prev_cpu = cpu;
292 	}
293 
294 	return new_tree;
295 }
296 
297 static void increment_rover(struct cpuinfo_tree *t, int node_index,
298                             int root_index, const int *rover_inc_table)
299 {
300 	struct cpuinfo_node *node = &t->nodes[node_index];
301 	int top_level, level;
302 
303 	top_level = t->nodes[root_index].level;
304 	for (level = node->level; level >= top_level; level--) {
305 		node->rover++;
306 		if (node->rover <= node->child_end)
307 			return;
308 
309 		node->rover = node->child_start;
310 		/* If parent's rover does not need to be adjusted, stop here. */
311 		if ((level == top_level) ||
312 		    !(rover_inc_table[level] & ROVER_INC_PARENT_ON_LOOP))
313 			return;
314 
315 		node = &t->nodes[node->parent_index];
316 	}
317 }
318 
319 static int iterate_cpu(struct cpuinfo_tree *t, unsigned int root_index)
320 {
321 	const int *rover_inc_table;
322 	int level, new_index, index = root_index;
323 
324 	switch (sun4v_chip_type) {
325 	case SUN4V_CHIP_NIAGARA1:
326 	case SUN4V_CHIP_NIAGARA2:
327 	case SUN4V_CHIP_NIAGARA3:
328 	case SUN4V_CHIP_NIAGARA4:
329 	case SUN4V_CHIP_NIAGARA5:
330 		rover_inc_table = niagara_iterate_method;
331 		break;
332 	default:
333 		rover_inc_table = generic_iterate_method;
334 	}
335 
336 	for (level = t->nodes[root_index].level; level < CPUINFO_LVL_MAX;
337 	     level++) {
338 		new_index = t->nodes[index].rover;
339 		if (rover_inc_table[level] & ROVER_INC_ON_VISIT)
340 			increment_rover(t, index, root_index, rover_inc_table);
341 
342 		index = new_index;
343 	}
344 	return index;
345 }
346 
347 static void _cpu_map_rebuild(void)
348 {
349 	int i;
350 
351 	if (cpuinfo_tree) {
352 		kfree(cpuinfo_tree);
353 		cpuinfo_tree = NULL;
354 	}
355 
356 	cpuinfo_tree = build_cpuinfo_tree();
357 	if (!cpuinfo_tree)
358 		return;
359 
360 	/* Build CPU distribution map that spans all online CPUs.  No need
361 	 * to check if the CPU is online, as that is done when the cpuinfo
362 	 * tree is being built.
363 	 */
364 	for (i = 0; i < cpuinfo_tree->nodes[0].num_cpus; i++)
365 		cpu_distribution_map[i] = iterate_cpu(cpuinfo_tree, 0);
366 }
367 
368 /* Fallback if the cpuinfo tree could not be built.  CPU mapping is linear
369  * round robin.
370  */
371 static int simple_map_to_cpu(unsigned int index)
372 {
373 	int i, end, cpu_rover;
374 
375 	cpu_rover = 0;
376 	end = index % num_online_cpus();
377 	for (i = 0; i < num_possible_cpus(); i++) {
378 		if (cpu_online(cpu_rover)) {
379 			if (cpu_rover >= end)
380 				return cpu_rover;
381 
382 			cpu_rover++;
383 		}
384 	}
385 
386 	/* Impossible, since num_online_cpus() <= num_possible_cpus() */
387 	return cpumask_first(cpu_online_mask);
388 }
389 
390 static int _map_to_cpu(unsigned int index)
391 {
392 	struct cpuinfo_node *root_node;
393 
394 	if (unlikely(!cpuinfo_tree)) {
395 		_cpu_map_rebuild();
396 		if (!cpuinfo_tree)
397 			return simple_map_to_cpu(index);
398 	}
399 
400 	root_node = &cpuinfo_tree->nodes[0];
401 #ifdef CONFIG_HOTPLUG_CPU
402 	if (unlikely(root_node->num_cpus != num_online_cpus())) {
403 		_cpu_map_rebuild();
404 		if (!cpuinfo_tree)
405 			return simple_map_to_cpu(index);
406 	}
407 #endif
408 	return cpu_distribution_map[index % root_node->num_cpus];
409 }
410 
411 int map_to_cpu(unsigned int index)
412 {
413 	int mapped_cpu;
414 	unsigned long flag;
415 
416 	spin_lock_irqsave(&cpu_map_lock, flag);
417 	mapped_cpu = _map_to_cpu(index);
418 
419 #ifdef CONFIG_HOTPLUG_CPU
420 	while (unlikely(!cpu_online(mapped_cpu)))
421 		mapped_cpu = _map_to_cpu(index);
422 #endif
423 	spin_unlock_irqrestore(&cpu_map_lock, flag);
424 	return mapped_cpu;
425 }
426 EXPORT_SYMBOL(map_to_cpu);
427 
428 void cpu_map_rebuild(void)
429 {
430 	unsigned long flag;
431 
432 	spin_lock_irqsave(&cpu_map_lock, flag);
433 	_cpu_map_rebuild();
434 	spin_unlock_irqrestore(&cpu_map_lock, flag);
435 }
436