xref: /linux/lib/cpu_rmap.c (revision 24bce201d79807b668bf9d9e0aca801c5c0d5f78)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * cpu_rmap.c: CPU affinity reverse-map support
4  * Copyright 2011 Solarflare Communications Inc.
5  */
6 
7 #include <linux/cpu_rmap.h>
8 #include <linux/interrupt.h>
9 #include <linux/export.h>
10 
11 /*
12  * These functions maintain a mapping from CPUs to some ordered set of
13  * objects with CPU affinities.  This can be seen as a reverse-map of
14  * CPU affinity.  However, we do not assume that the object affinities
15  * cover all CPUs in the system.  For those CPUs not directly covered
16  * by object affinities, we attempt to find a nearest object based on
17  * CPU topology.
18  */
19 
20 /**
21  * alloc_cpu_rmap - allocate CPU affinity reverse-map
22  * @size: Number of objects to be mapped
23  * @flags: Allocation flags e.g. %GFP_KERNEL
24  */
25 struct cpu_rmap *alloc_cpu_rmap(unsigned int size, gfp_t flags)
26 {
27 	struct cpu_rmap *rmap;
28 	unsigned int cpu;
29 	size_t obj_offset;
30 
31 	/* This is a silly number of objects, and we use u16 indices. */
32 	if (size > 0xffff)
33 		return NULL;
34 
35 	/* Offset of object pointer array from base structure */
36 	obj_offset = ALIGN(offsetof(struct cpu_rmap, near[nr_cpu_ids]),
37 			   sizeof(void *));
38 
39 	rmap = kzalloc(obj_offset + size * sizeof(rmap->obj[0]), flags);
40 	if (!rmap)
41 		return NULL;
42 
43 	kref_init(&rmap->refcount);
44 	rmap->obj = (void **)((char *)rmap + obj_offset);
45 
46 	/* Initially assign CPUs to objects on a rota, since we have
47 	 * no idea where the objects are.  Use infinite distance, so
48 	 * any object with known distance is preferable.  Include the
49 	 * CPUs that are not present/online, since we definitely want
50 	 * any newly-hotplugged CPUs to have some object assigned.
51 	 */
52 	for_each_possible_cpu(cpu) {
53 		rmap->near[cpu].index = cpu % size;
54 		rmap->near[cpu].dist = CPU_RMAP_DIST_INF;
55 	}
56 
57 	rmap->size = size;
58 	return rmap;
59 }
60 EXPORT_SYMBOL(alloc_cpu_rmap);
61 
62 /**
63  * cpu_rmap_release - internal reclaiming helper called from kref_put
64  * @ref: kref to struct cpu_rmap
65  */
66 static void cpu_rmap_release(struct kref *ref)
67 {
68 	struct cpu_rmap *rmap = container_of(ref, struct cpu_rmap, refcount);
69 	kfree(rmap);
70 }
71 
72 /**
73  * cpu_rmap_get - internal helper to get new ref on a cpu_rmap
74  * @rmap: reverse-map allocated with alloc_cpu_rmap()
75  */
76 static inline void cpu_rmap_get(struct cpu_rmap *rmap)
77 {
78 	kref_get(&rmap->refcount);
79 }
80 
81 /**
82  * cpu_rmap_put - release ref on a cpu_rmap
83  * @rmap: reverse-map allocated with alloc_cpu_rmap()
84  */
85 int cpu_rmap_put(struct cpu_rmap *rmap)
86 {
87 	return kref_put(&rmap->refcount, cpu_rmap_release);
88 }
89 EXPORT_SYMBOL(cpu_rmap_put);
90 
91 /* Reevaluate nearest object for given CPU, comparing with the given
92  * neighbours at the given distance.
93  */
94 static bool cpu_rmap_copy_neigh(struct cpu_rmap *rmap, unsigned int cpu,
95 				const struct cpumask *mask, u16 dist)
96 {
97 	int neigh;
98 
99 	for_each_cpu(neigh, mask) {
100 		if (rmap->near[cpu].dist > dist &&
101 		    rmap->near[neigh].dist <= dist) {
102 			rmap->near[cpu].index = rmap->near[neigh].index;
103 			rmap->near[cpu].dist = dist;
104 			return true;
105 		}
106 	}
107 	return false;
108 }
109 
110 #ifdef DEBUG
111 static void debug_print_rmap(const struct cpu_rmap *rmap, const char *prefix)
112 {
113 	unsigned index;
114 	unsigned int cpu;
115 
116 	pr_info("cpu_rmap %p, %s:\n", rmap, prefix);
117 
118 	for_each_possible_cpu(cpu) {
119 		index = rmap->near[cpu].index;
120 		pr_info("cpu %d -> obj %u (distance %u)\n",
121 			cpu, index, rmap->near[cpu].dist);
122 	}
123 }
124 #else
125 static inline void
126 debug_print_rmap(const struct cpu_rmap *rmap, const char *prefix)
127 {
128 }
129 #endif
130 
131 /**
132  * cpu_rmap_add - add object to a rmap
133  * @rmap: CPU rmap allocated with alloc_cpu_rmap()
134  * @obj: Object to add to rmap
135  *
136  * Return index of object.
137  */
138 int cpu_rmap_add(struct cpu_rmap *rmap, void *obj)
139 {
140 	u16 index;
141 
142 	BUG_ON(rmap->used >= rmap->size);
143 	index = rmap->used++;
144 	rmap->obj[index] = obj;
145 	return index;
146 }
147 EXPORT_SYMBOL(cpu_rmap_add);
148 
149 /**
150  * cpu_rmap_update - update CPU rmap following a change of object affinity
151  * @rmap: CPU rmap to update
152  * @index: Index of object whose affinity changed
153  * @affinity: New CPU affinity of object
154  */
155 int cpu_rmap_update(struct cpu_rmap *rmap, u16 index,
156 		    const struct cpumask *affinity)
157 {
158 	cpumask_var_t update_mask;
159 	unsigned int cpu;
160 
161 	if (unlikely(!zalloc_cpumask_var(&update_mask, GFP_KERNEL)))
162 		return -ENOMEM;
163 
164 	/* Invalidate distance for all CPUs for which this used to be
165 	 * the nearest object.  Mark those CPUs for update.
166 	 */
167 	for_each_online_cpu(cpu) {
168 		if (rmap->near[cpu].index == index) {
169 			rmap->near[cpu].dist = CPU_RMAP_DIST_INF;
170 			cpumask_set_cpu(cpu, update_mask);
171 		}
172 	}
173 
174 	debug_print_rmap(rmap, "after invalidating old distances");
175 
176 	/* Set distance to 0 for all CPUs in the new affinity mask.
177 	 * Mark all CPUs within their NUMA nodes for update.
178 	 */
179 	for_each_cpu(cpu, affinity) {
180 		rmap->near[cpu].index = index;
181 		rmap->near[cpu].dist = 0;
182 		cpumask_or(update_mask, update_mask,
183 			   cpumask_of_node(cpu_to_node(cpu)));
184 	}
185 
186 	debug_print_rmap(rmap, "after updating neighbours");
187 
188 	/* Update distances based on topology */
189 	for_each_cpu(cpu, update_mask) {
190 		if (cpu_rmap_copy_neigh(rmap, cpu,
191 					topology_sibling_cpumask(cpu), 1))
192 			continue;
193 		if (cpu_rmap_copy_neigh(rmap, cpu,
194 					topology_core_cpumask(cpu), 2))
195 			continue;
196 		if (cpu_rmap_copy_neigh(rmap, cpu,
197 					cpumask_of_node(cpu_to_node(cpu)), 3))
198 			continue;
199 		/* We could continue into NUMA node distances, but for now
200 		 * we give up.
201 		 */
202 	}
203 
204 	debug_print_rmap(rmap, "after copying neighbours");
205 
206 	free_cpumask_var(update_mask);
207 	return 0;
208 }
209 EXPORT_SYMBOL(cpu_rmap_update);
210 
211 /* Glue between IRQ affinity notifiers and CPU rmaps */
212 
213 struct irq_glue {
214 	struct irq_affinity_notify notify;
215 	struct cpu_rmap *rmap;
216 	u16 index;
217 };
218 
219 /**
220  * free_irq_cpu_rmap - free a CPU affinity reverse-map used for IRQs
221  * @rmap: Reverse-map allocated with alloc_irq_cpu_map(), or %NULL
222  *
223  * Must be called in process context, before freeing the IRQs.
224  */
225 void free_irq_cpu_rmap(struct cpu_rmap *rmap)
226 {
227 	struct irq_glue *glue;
228 	u16 index;
229 
230 	if (!rmap)
231 		return;
232 
233 	for (index = 0; index < rmap->used; index++) {
234 		glue = rmap->obj[index];
235 		irq_set_affinity_notifier(glue->notify.irq, NULL);
236 	}
237 
238 	cpu_rmap_put(rmap);
239 }
240 EXPORT_SYMBOL(free_irq_cpu_rmap);
241 
242 /**
243  * irq_cpu_rmap_notify - callback for IRQ subsystem when IRQ affinity updated
244  * @notify: struct irq_affinity_notify passed by irq/manage.c
245  * @mask: cpu mask for new SMP affinity
246  *
247  * This is executed in workqueue context.
248  */
249 static void
250 irq_cpu_rmap_notify(struct irq_affinity_notify *notify, const cpumask_t *mask)
251 {
252 	struct irq_glue *glue =
253 		container_of(notify, struct irq_glue, notify);
254 	int rc;
255 
256 	rc = cpu_rmap_update(glue->rmap, glue->index, mask);
257 	if (rc)
258 		pr_warn("irq_cpu_rmap_notify: update failed: %d\n", rc);
259 }
260 
261 /**
262  * irq_cpu_rmap_release - reclaiming callback for IRQ subsystem
263  * @ref: kref to struct irq_affinity_notify passed by irq/manage.c
264  */
265 static void irq_cpu_rmap_release(struct kref *ref)
266 {
267 	struct irq_glue *glue =
268 		container_of(ref, struct irq_glue, notify.kref);
269 
270 	cpu_rmap_put(glue->rmap);
271 	kfree(glue);
272 }
273 
274 /**
275  * irq_cpu_rmap_add - add an IRQ to a CPU affinity reverse-map
276  * @rmap: The reverse-map
277  * @irq: The IRQ number
278  *
279  * This adds an IRQ affinity notifier that will update the reverse-map
280  * automatically.
281  *
282  * Must be called in process context, after the IRQ is allocated but
283  * before it is bound with request_irq().
284  */
285 int irq_cpu_rmap_add(struct cpu_rmap *rmap, int irq)
286 {
287 	struct irq_glue *glue = kzalloc(sizeof(*glue), GFP_KERNEL);
288 	int rc;
289 
290 	if (!glue)
291 		return -ENOMEM;
292 	glue->notify.notify = irq_cpu_rmap_notify;
293 	glue->notify.release = irq_cpu_rmap_release;
294 	glue->rmap = rmap;
295 	cpu_rmap_get(rmap);
296 	glue->index = cpu_rmap_add(rmap, glue);
297 	rc = irq_set_affinity_notifier(irq, &glue->notify);
298 	if (rc) {
299 		cpu_rmap_put(glue->rmap);
300 		kfree(glue);
301 	}
302 	return rc;
303 }
304 EXPORT_SYMBOL(irq_cpu_rmap_add);
305