xref: /openbmc/linux/arch/sparc/kernel/cpumap.c (revision 2c684d89)
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/cpumask.h>
10 #include <linux/spinlock.h>
11 #include <asm/cpudata.h>
12 #include "cpumap.h"
13 
14 
15 enum {
16 	CPUINFO_LVL_ROOT = 0,
17 	CPUINFO_LVL_NODE,
18 	CPUINFO_LVL_CORE,
19 	CPUINFO_LVL_PROC,
20 	CPUINFO_LVL_MAX,
21 };
22 
23 enum {
24 	ROVER_NO_OP              = 0,
25 	/* Increment rover every time level is visited */
26 	ROVER_INC_ON_VISIT       = 1 << 0,
27 	/* Increment parent's rover every time rover wraps around */
28 	ROVER_INC_PARENT_ON_LOOP = 1 << 1,
29 };
30 
31 struct cpuinfo_node {
32 	int id;
33 	int level;
34 	int num_cpus;    /* Number of CPUs in this hierarchy */
35 	int parent_index;
36 	int child_start; /* Array index of the first child node */
37 	int child_end;   /* Array index of the last child node */
38 	int rover;       /* Child node iterator */
39 };
40 
41 struct cpuinfo_level {
42 	int start_index; /* Index of first node of a level in a cpuinfo tree */
43 	int end_index;   /* Index of last node of a level in a cpuinfo tree */
44 	int num_nodes;   /* Number of nodes in a level in a cpuinfo tree */
45 };
46 
47 struct cpuinfo_tree {
48 	int total_nodes;
49 
50 	/* Offsets into nodes[] for each level of the tree */
51 	struct cpuinfo_level level[CPUINFO_LVL_MAX];
52 	struct cpuinfo_node  nodes[0];
53 };
54 
55 
56 static struct cpuinfo_tree *cpuinfo_tree;
57 
58 static u16 cpu_distribution_map[NR_CPUS];
59 static DEFINE_SPINLOCK(cpu_map_lock);
60 
61 
62 /* Niagara optimized cpuinfo tree traversal. */
63 static const int niagara_iterate_method[] = {
64 	[CPUINFO_LVL_ROOT] = ROVER_NO_OP,
65 
66 	/* Strands (or virtual CPUs) within a core may not run concurrently
67 	 * on the Niagara, as instruction pipeline(s) are shared.  Distribute
68 	 * work to strands in different cores first for better concurrency.
69 	 * Go to next NUMA node when all cores are used.
70 	 */
71 	[CPUINFO_LVL_NODE] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
72 
73 	/* Strands are grouped together by proc_id in cpuinfo_sparc, i.e.
74 	 * a proc_id represents an instruction pipeline.  Distribute work to
75 	 * strands in different proc_id groups if the core has multiple
76 	 * instruction pipelines (e.g. the Niagara 2/2+ has two).
77 	 */
78 	[CPUINFO_LVL_CORE] = ROVER_INC_ON_VISIT,
79 
80 	/* Pick the next strand in the proc_id group. */
81 	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT,
82 };
83 
84 /* Generic cpuinfo tree traversal.  Distribute work round robin across NUMA
85  * nodes.
86  */
87 static const int generic_iterate_method[] = {
88 	[CPUINFO_LVL_ROOT] = ROVER_INC_ON_VISIT,
89 	[CPUINFO_LVL_NODE] = ROVER_NO_OP,
90 	[CPUINFO_LVL_CORE] = ROVER_INC_PARENT_ON_LOOP,
91 	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
92 };
93 
94 
95 static int cpuinfo_id(int cpu, int level)
96 {
97 	int id;
98 
99 	switch (level) {
100 	case CPUINFO_LVL_ROOT:
101 		id = 0;
102 		break;
103 	case CPUINFO_LVL_NODE:
104 		id = cpu_to_node(cpu);
105 		break;
106 	case CPUINFO_LVL_CORE:
107 		id = cpu_data(cpu).core_id;
108 		break;
109 	case CPUINFO_LVL_PROC:
110 		id = cpu_data(cpu).proc_id;
111 		break;
112 	default:
113 		id = -EINVAL;
114 	}
115 	return id;
116 }
117 
118 /*
119  * Enumerate the CPU information in __cpu_data to determine the start index,
120  * end index, and number of nodes for each level in the cpuinfo tree.  The
121  * total number of cpuinfo nodes required to build the tree is returned.
122  */
123 static int enumerate_cpuinfo_nodes(struct cpuinfo_level *tree_level)
124 {
125 	int prev_id[CPUINFO_LVL_MAX];
126 	int i, n, num_nodes;
127 
128 	for (i = CPUINFO_LVL_ROOT; i < CPUINFO_LVL_MAX; i++) {
129 		struct cpuinfo_level *lv = &tree_level[i];
130 
131 		prev_id[i] = -1;
132 		lv->start_index = lv->end_index = lv->num_nodes = 0;
133 	}
134 
135 	num_nodes = 1; /* Include the root node */
136 
137 	for (i = 0; i < num_possible_cpus(); i++) {
138 		if (!cpu_online(i))
139 			continue;
140 
141 		n = cpuinfo_id(i, CPUINFO_LVL_NODE);
142 		if (n > prev_id[CPUINFO_LVL_NODE]) {
143 			tree_level[CPUINFO_LVL_NODE].num_nodes++;
144 			prev_id[CPUINFO_LVL_NODE] = n;
145 			num_nodes++;
146 		}
147 		n = cpuinfo_id(i, CPUINFO_LVL_CORE);
148 		if (n > prev_id[CPUINFO_LVL_CORE]) {
149 			tree_level[CPUINFO_LVL_CORE].num_nodes++;
150 			prev_id[CPUINFO_LVL_CORE] = n;
151 			num_nodes++;
152 		}
153 		n = cpuinfo_id(i, CPUINFO_LVL_PROC);
154 		if (n > prev_id[CPUINFO_LVL_PROC]) {
155 			tree_level[CPUINFO_LVL_PROC].num_nodes++;
156 			prev_id[CPUINFO_LVL_PROC] = n;
157 			num_nodes++;
158 		}
159 	}
160 
161 	tree_level[CPUINFO_LVL_ROOT].num_nodes = 1;
162 
163 	n = tree_level[CPUINFO_LVL_NODE].num_nodes;
164 	tree_level[CPUINFO_LVL_NODE].start_index = 1;
165 	tree_level[CPUINFO_LVL_NODE].end_index   = n;
166 
167 	n++;
168 	tree_level[CPUINFO_LVL_CORE].start_index = n;
169 	n += tree_level[CPUINFO_LVL_CORE].num_nodes;
170 	tree_level[CPUINFO_LVL_CORE].end_index   = n - 1;
171 
172 	tree_level[CPUINFO_LVL_PROC].start_index = n;
173 	n += tree_level[CPUINFO_LVL_PROC].num_nodes;
174 	tree_level[CPUINFO_LVL_PROC].end_index   = n - 1;
175 
176 	return num_nodes;
177 }
178 
179 /* Build a tree representation of the CPU hierarchy using the per CPU
180  * information in __cpu_data.  Entries in __cpu_data[0..NR_CPUS] are
181  * assumed to be sorted in ascending order based on node, core_id, and
182  * proc_id (in order of significance).
183  */
184 static struct cpuinfo_tree *build_cpuinfo_tree(void)
185 {
186 	struct cpuinfo_tree *new_tree;
187 	struct cpuinfo_node *node;
188 	struct cpuinfo_level tmp_level[CPUINFO_LVL_MAX];
189 	int num_cpus[CPUINFO_LVL_MAX];
190 	int level_rover[CPUINFO_LVL_MAX];
191 	int prev_id[CPUINFO_LVL_MAX];
192 	int n, id, cpu, prev_cpu, last_cpu, level;
193 
194 	n = enumerate_cpuinfo_nodes(tmp_level);
195 
196 	new_tree = kzalloc(sizeof(struct cpuinfo_tree) +
197 	                   (sizeof(struct cpuinfo_node) * n), GFP_ATOMIC);
198 	if (!new_tree)
199 		return NULL;
200 
201 	new_tree->total_nodes = n;
202 	memcpy(&new_tree->level, tmp_level, sizeof(tmp_level));
203 
204 	prev_cpu = cpu = cpumask_first(cpu_online_mask);
205 
206 	/* Initialize all levels in the tree with the first CPU */
207 	for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT; level--) {
208 		n = new_tree->level[level].start_index;
209 
210 		level_rover[level] = n;
211 		node = &new_tree->nodes[n];
212 
213 		id = cpuinfo_id(cpu, level);
214 		if (unlikely(id < 0)) {
215 			kfree(new_tree);
216 			return NULL;
217 		}
218 		node->id = id;
219 		node->level = level;
220 		node->num_cpus = 1;
221 
222 		node->parent_index = (level > CPUINFO_LVL_ROOT)
223 		    ? new_tree->level[level - 1].start_index : -1;
224 
225 		node->child_start = node->child_end = node->rover =
226 		    (level == CPUINFO_LVL_PROC)
227 		    ? cpu : new_tree->level[level + 1].start_index;
228 
229 		prev_id[level] = node->id;
230 		num_cpus[level] = 1;
231 	}
232 
233 	for (last_cpu = (num_possible_cpus() - 1); last_cpu >= 0; last_cpu--) {
234 		if (cpu_online(last_cpu))
235 			break;
236 	}
237 
238 	while (++cpu <= last_cpu) {
239 		if (!cpu_online(cpu))
240 			continue;
241 
242 		for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT;
243 		     level--) {
244 			id = cpuinfo_id(cpu, level);
245 			if (unlikely(id < 0)) {
246 				kfree(new_tree);
247 				return NULL;
248 			}
249 
250 			if ((id != prev_id[level]) || (cpu == last_cpu)) {
251 				prev_id[level] = id;
252 				node = &new_tree->nodes[level_rover[level]];
253 				node->num_cpus = num_cpus[level];
254 				num_cpus[level] = 1;
255 
256 				if (cpu == last_cpu)
257 					node->num_cpus++;
258 
259 				/* Connect tree node to parent */
260 				if (level == CPUINFO_LVL_ROOT)
261 					node->parent_index = -1;
262 				else
263 					node->parent_index =
264 					    level_rover[level - 1];
265 
266 				if (level == CPUINFO_LVL_PROC) {
267 					node->child_end =
268 					    (cpu == last_cpu) ? cpu : prev_cpu;
269 				} else {
270 					node->child_end =
271 					    level_rover[level + 1] - 1;
272 				}
273 
274 				/* Initialize the next node in the same level */
275 				n = ++level_rover[level];
276 				if (n <= new_tree->level[level].end_index) {
277 					node = &new_tree->nodes[n];
278 					node->id = id;
279 					node->level = level;
280 
281 					/* Connect node to child */
282 					node->child_start = node->child_end =
283 					node->rover =
284 					    (level == CPUINFO_LVL_PROC)
285 					    ? cpu : level_rover[level + 1];
286 				}
287 			} else
288 				num_cpus[level]++;
289 		}
290 		prev_cpu = cpu;
291 	}
292 
293 	return new_tree;
294 }
295 
296 static void increment_rover(struct cpuinfo_tree *t, int node_index,
297                             int root_index, const int *rover_inc_table)
298 {
299 	struct cpuinfo_node *node = &t->nodes[node_index];
300 	int top_level, level;
301 
302 	top_level = t->nodes[root_index].level;
303 	for (level = node->level; level >= top_level; level--) {
304 		node->rover++;
305 		if (node->rover <= node->child_end)
306 			return;
307 
308 		node->rover = node->child_start;
309 		/* If parent's rover does not need to be adjusted, stop here. */
310 		if ((level == top_level) ||
311 		    !(rover_inc_table[level] & ROVER_INC_PARENT_ON_LOOP))
312 			return;
313 
314 		node = &t->nodes[node->parent_index];
315 	}
316 }
317 
318 static int iterate_cpu(struct cpuinfo_tree *t, unsigned int root_index)
319 {
320 	const int *rover_inc_table;
321 	int level, new_index, index = root_index;
322 
323 	switch (sun4v_chip_type) {
324 	case SUN4V_CHIP_NIAGARA1:
325 	case SUN4V_CHIP_NIAGARA2:
326 	case SUN4V_CHIP_NIAGARA3:
327 	case SUN4V_CHIP_NIAGARA4:
328 	case SUN4V_CHIP_NIAGARA5:
329 	case SUN4V_CHIP_SPARC_M6:
330 	case SUN4V_CHIP_SPARC_M7:
331 	case SUN4V_CHIP_SPARC64X:
332 		rover_inc_table = niagara_iterate_method;
333 		break;
334 	default:
335 		rover_inc_table = generic_iterate_method;
336 	}
337 
338 	for (level = t->nodes[root_index].level; level < CPUINFO_LVL_MAX;
339 	     level++) {
340 		new_index = t->nodes[index].rover;
341 		if (rover_inc_table[level] & ROVER_INC_ON_VISIT)
342 			increment_rover(t, index, root_index, rover_inc_table);
343 
344 		index = new_index;
345 	}
346 	return index;
347 }
348 
349 static void _cpu_map_rebuild(void)
350 {
351 	int i;
352 
353 	if (cpuinfo_tree) {
354 		kfree(cpuinfo_tree);
355 		cpuinfo_tree = NULL;
356 	}
357 
358 	cpuinfo_tree = build_cpuinfo_tree();
359 	if (!cpuinfo_tree)
360 		return;
361 
362 	/* Build CPU distribution map that spans all online CPUs.  No need
363 	 * to check if the CPU is online, as that is done when the cpuinfo
364 	 * tree is being built.
365 	 */
366 	for (i = 0; i < cpuinfo_tree->nodes[0].num_cpus; i++)
367 		cpu_distribution_map[i] = iterate_cpu(cpuinfo_tree, 0);
368 }
369 
370 /* Fallback if the cpuinfo tree could not be built.  CPU mapping is linear
371  * round robin.
372  */
373 static int simple_map_to_cpu(unsigned int index)
374 {
375 	int i, end, cpu_rover;
376 
377 	cpu_rover = 0;
378 	end = index % num_online_cpus();
379 	for (i = 0; i < num_possible_cpus(); i++) {
380 		if (cpu_online(cpu_rover)) {
381 			if (cpu_rover >= end)
382 				return cpu_rover;
383 
384 			cpu_rover++;
385 		}
386 	}
387 
388 	/* Impossible, since num_online_cpus() <= num_possible_cpus() */
389 	return cpumask_first(cpu_online_mask);
390 }
391 
392 static int _map_to_cpu(unsigned int index)
393 {
394 	struct cpuinfo_node *root_node;
395 
396 	if (unlikely(!cpuinfo_tree)) {
397 		_cpu_map_rebuild();
398 		if (!cpuinfo_tree)
399 			return simple_map_to_cpu(index);
400 	}
401 
402 	root_node = &cpuinfo_tree->nodes[0];
403 #ifdef CONFIG_HOTPLUG_CPU
404 	if (unlikely(root_node->num_cpus != num_online_cpus())) {
405 		_cpu_map_rebuild();
406 		if (!cpuinfo_tree)
407 			return simple_map_to_cpu(index);
408 	}
409 #endif
410 	return cpu_distribution_map[index % root_node->num_cpus];
411 }
412 
413 int map_to_cpu(unsigned int index)
414 {
415 	int mapped_cpu;
416 	unsigned long flag;
417 
418 	spin_lock_irqsave(&cpu_map_lock, flag);
419 	mapped_cpu = _map_to_cpu(index);
420 
421 #ifdef CONFIG_HOTPLUG_CPU
422 	while (unlikely(!cpu_online(mapped_cpu)))
423 		mapped_cpu = _map_to_cpu(index);
424 #endif
425 	spin_unlock_irqrestore(&cpu_map_lock, flag);
426 	return mapped_cpu;
427 }
428 EXPORT_SYMBOL(map_to_cpu);
429 
430 void cpu_map_rebuild(void)
431 {
432 	unsigned long flag;
433 
434 	spin_lock_irqsave(&cpu_map_lock, flag);
435 	_cpu_map_rebuild();
436 	spin_unlock_irqrestore(&cpu_map_lock, flag);
437 }
438