xref: /linux/arch/powerpc/platforms/cell/spu_manage.c (revision a1c613ae4c322ddd58d5a8539dbfba2a0380a8c0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * spu management operations for of based platforms
4  *
5  * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
6  * Copyright 2006 Sony Corp.
7  * (C) Copyright 2007 TOSHIBA CORPORATION
8  */
9 
10 #include <linux/interrupt.h>
11 #include <linux/list.h>
12 #include <linux/export.h>
13 #include <linux/ptrace.h>
14 #include <linux/wait.h>
15 #include <linux/mm.h>
16 #include <linux/io.h>
17 #include <linux/mutex.h>
18 #include <linux/device.h>
19 #include <linux/of_address.h>
20 #include <linux/of_irq.h>
21 
22 #include <asm/spu.h>
23 #include <asm/spu_priv1.h>
24 #include <asm/firmware.h>
25 
26 #include "spufs/spufs.h"
27 #include "interrupt.h"
28 #include "spu_priv1_mmio.h"
29 
spu_devnode(struct spu * spu)30 struct device_node *spu_devnode(struct spu *spu)
31 {
32 	return spu->devnode;
33 }
34 
35 EXPORT_SYMBOL_GPL(spu_devnode);
36 
find_spu_unit_number(struct device_node * spe)37 static u64 __init find_spu_unit_number(struct device_node *spe)
38 {
39 	const unsigned int *prop;
40 	int proplen;
41 
42 	/* new device trees should provide the physical-id attribute */
43 	prop = of_get_property(spe, "physical-id", &proplen);
44 	if (proplen == 4)
45 		return (u64)*prop;
46 
47 	/* celleb device tree provides the unit-id */
48 	prop = of_get_property(spe, "unit-id", &proplen);
49 	if (proplen == 4)
50 		return (u64)*prop;
51 
52 	/* legacy device trees provide the id in the reg attribute */
53 	prop = of_get_property(spe, "reg", &proplen);
54 	if (proplen == 4)
55 		return (u64)*prop;
56 
57 	return 0;
58 }
59 
spu_unmap(struct spu * spu)60 static void spu_unmap(struct spu *spu)
61 {
62 	if (!firmware_has_feature(FW_FEATURE_LPAR))
63 		iounmap(spu->priv1);
64 	iounmap(spu->priv2);
65 	iounmap(spu->problem);
66 	iounmap((__force u8 __iomem *)spu->local_store);
67 }
68 
spu_map_interrupts_old(struct spu * spu,struct device_node * np)69 static int __init spu_map_interrupts_old(struct spu *spu,
70 	struct device_node *np)
71 {
72 	unsigned int isrc;
73 	const u32 *tmp;
74 	int nid;
75 
76 	/* Get the interrupt source unit from the device-tree */
77 	tmp = of_get_property(np, "isrc", NULL);
78 	if (!tmp)
79 		return -ENODEV;
80 	isrc = tmp[0];
81 
82 	tmp = of_get_property(np->parent->parent, "node-id", NULL);
83 	if (!tmp) {
84 		printk(KERN_WARNING "%s: can't find node-id\n", __func__);
85 		nid = spu->node;
86 	} else
87 		nid = tmp[0];
88 
89 	/* Add the node number */
90 	isrc |= nid << IIC_IRQ_NODE_SHIFT;
91 
92 	/* Now map interrupts of all 3 classes */
93 	spu->irqs[0] = irq_create_mapping(NULL, IIC_IRQ_CLASS_0 | isrc);
94 	spu->irqs[1] = irq_create_mapping(NULL, IIC_IRQ_CLASS_1 | isrc);
95 	spu->irqs[2] = irq_create_mapping(NULL, IIC_IRQ_CLASS_2 | isrc);
96 
97 	/* Right now, we only fail if class 2 failed */
98 	if (!spu->irqs[2])
99 		return -EINVAL;
100 
101 	return 0;
102 }
103 
spu_map_prop_old(struct spu * spu,struct device_node * n,const char * name)104 static void __iomem * __init spu_map_prop_old(struct spu *spu,
105 					      struct device_node *n,
106 					      const char *name)
107 {
108 	const struct address_prop {
109 		unsigned long address;
110 		unsigned int len;
111 	} __attribute__((packed)) *prop;
112 	int proplen;
113 
114 	prop = of_get_property(n, name, &proplen);
115 	if (prop == NULL || proplen != sizeof (struct address_prop))
116 		return NULL;
117 
118 	return ioremap(prop->address, prop->len);
119 }
120 
spu_map_device_old(struct spu * spu)121 static int __init spu_map_device_old(struct spu *spu)
122 {
123 	struct device_node *node = spu->devnode;
124 	const char *prop;
125 	int ret;
126 
127 	ret = -ENODEV;
128 	spu->name = of_get_property(node, "name", NULL);
129 	if (!spu->name)
130 		goto out;
131 
132 	prop = of_get_property(node, "local-store", NULL);
133 	if (!prop)
134 		goto out;
135 	spu->local_store_phys = *(unsigned long *)prop;
136 
137 	/* we use local store as ram, not io memory */
138 	spu->local_store = (void __force *)
139 		spu_map_prop_old(spu, node, "local-store");
140 	if (!spu->local_store)
141 		goto out;
142 
143 	prop = of_get_property(node, "problem", NULL);
144 	if (!prop)
145 		goto out_unmap;
146 	spu->problem_phys = *(unsigned long *)prop;
147 
148 	spu->problem = spu_map_prop_old(spu, node, "problem");
149 	if (!spu->problem)
150 		goto out_unmap;
151 
152 	spu->priv2 = spu_map_prop_old(spu, node, "priv2");
153 	if (!spu->priv2)
154 		goto out_unmap;
155 
156 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
157 		spu->priv1 = spu_map_prop_old(spu, node, "priv1");
158 		if (!spu->priv1)
159 			goto out_unmap;
160 	}
161 
162 	ret = 0;
163 	goto out;
164 
165 out_unmap:
166 	spu_unmap(spu);
167 out:
168 	return ret;
169 }
170 
spu_map_interrupts(struct spu * spu,struct device_node * np)171 static int __init spu_map_interrupts(struct spu *spu, struct device_node *np)
172 {
173 	int i;
174 
175 	for (i=0; i < 3; i++) {
176 		spu->irqs[i] = irq_of_parse_and_map(np, i);
177 		if (!spu->irqs[i])
178 			goto err;
179 	}
180 	return 0;
181 
182 err:
183 	pr_debug("failed to map irq %x for spu %s\n", i, spu->name);
184 	for (; i >= 0; i--) {
185 		if (spu->irqs[i])
186 			irq_dispose_mapping(spu->irqs[i]);
187 	}
188 	return -EINVAL;
189 }
190 
spu_map_resource(struct spu * spu,int nr,void __iomem ** virt,unsigned long * phys)191 static int __init spu_map_resource(struct spu *spu, int nr,
192 			    void __iomem** virt, unsigned long *phys)
193 {
194 	struct device_node *np = spu->devnode;
195 	struct resource resource = { };
196 	unsigned long len;
197 	int ret;
198 
199 	ret = of_address_to_resource(np, nr, &resource);
200 	if (ret)
201 		return ret;
202 	if (phys)
203 		*phys = resource.start;
204 	len = resource_size(&resource);
205 	*virt = ioremap(resource.start, len);
206 	if (!*virt)
207 		return -EINVAL;
208 	return 0;
209 }
210 
spu_map_device(struct spu * spu)211 static int __init spu_map_device(struct spu *spu)
212 {
213 	struct device_node *np = spu->devnode;
214 	int ret = -ENODEV;
215 
216 	spu->name = of_get_property(np, "name", NULL);
217 	if (!spu->name)
218 		goto out;
219 
220 	ret = spu_map_resource(spu, 0, (void __iomem**)&spu->local_store,
221 			       &spu->local_store_phys);
222 	if (ret) {
223 		pr_debug("spu_new: failed to map %pOF resource 0\n",
224 			 np);
225 		goto out;
226 	}
227 	ret = spu_map_resource(spu, 1, (void __iomem**)&spu->problem,
228 			       &spu->problem_phys);
229 	if (ret) {
230 		pr_debug("spu_new: failed to map %pOF resource 1\n",
231 			 np);
232 		goto out_unmap;
233 	}
234 	ret = spu_map_resource(spu, 2, (void __iomem**)&spu->priv2, NULL);
235 	if (ret) {
236 		pr_debug("spu_new: failed to map %pOF resource 2\n",
237 			 np);
238 		goto out_unmap;
239 	}
240 	if (!firmware_has_feature(FW_FEATURE_LPAR))
241 		ret = spu_map_resource(spu, 3,
242 			       (void __iomem**)&spu->priv1, NULL);
243 	if (ret) {
244 		pr_debug("spu_new: failed to map %pOF resource 3\n",
245 			 np);
246 		goto out_unmap;
247 	}
248 	pr_debug("spu_new: %pOF maps:\n", np);
249 	pr_debug("  local store   : 0x%016lx -> 0x%p\n",
250 		 spu->local_store_phys, spu->local_store);
251 	pr_debug("  problem state : 0x%016lx -> 0x%p\n",
252 		 spu->problem_phys, spu->problem);
253 	pr_debug("  priv2         :                       0x%p\n", spu->priv2);
254 	pr_debug("  priv1         :                       0x%p\n", spu->priv1);
255 
256 	return 0;
257 
258 out_unmap:
259 	spu_unmap(spu);
260 out:
261 	pr_debug("failed to map spe %s: %d\n", spu->name, ret);
262 	return ret;
263 }
264 
of_enumerate_spus(int (* fn)(void * data))265 static int __init of_enumerate_spus(int (*fn)(void *data))
266 {
267 	int ret;
268 	struct device_node *node;
269 	unsigned int n = 0;
270 
271 	ret = -ENODEV;
272 	for_each_node_by_type(node, "spe") {
273 		ret = fn(node);
274 		if (ret) {
275 			printk(KERN_WARNING "%s: Error initializing %pOFn\n",
276 				__func__, node);
277 			of_node_put(node);
278 			break;
279 		}
280 		n++;
281 	}
282 	return ret ? ret : n;
283 }
284 
of_create_spu(struct spu * spu,void * data)285 static int __init of_create_spu(struct spu *spu, void *data)
286 {
287 	int ret;
288 	struct device_node *spe = (struct device_node *)data;
289 	static int legacy_map = 0, legacy_irq = 0;
290 
291 	spu->devnode = of_node_get(spe);
292 	spu->spe_id = find_spu_unit_number(spe);
293 
294 	spu->node = of_node_to_nid(spe);
295 	if (spu->node >= MAX_NUMNODES) {
296 		printk(KERN_WARNING "SPE %pOF on node %d ignored,"
297 		       " node number too big\n", spe, spu->node);
298 		printk(KERN_WARNING "Check if CONFIG_NUMA is enabled.\n");
299 		ret = -ENODEV;
300 		goto out;
301 	}
302 
303 	ret = spu_map_device(spu);
304 	if (ret) {
305 		if (!legacy_map) {
306 			legacy_map = 1;
307 			printk(KERN_WARNING "%s: Legacy device tree found, "
308 				"trying to map old style\n", __func__);
309 		}
310 		ret = spu_map_device_old(spu);
311 		if (ret) {
312 			printk(KERN_ERR "Unable to map %s\n",
313 				spu->name);
314 			goto out;
315 		}
316 	}
317 
318 	ret = spu_map_interrupts(spu, spe);
319 	if (ret) {
320 		if (!legacy_irq) {
321 			legacy_irq = 1;
322 			printk(KERN_WARNING "%s: Legacy device tree found, "
323 				"trying old style irq\n", __func__);
324 		}
325 		ret = spu_map_interrupts_old(spu, spe);
326 		if (ret) {
327 			printk(KERN_ERR "%s: could not map interrupts\n",
328 				spu->name);
329 			goto out_unmap;
330 		}
331 	}
332 
333 	pr_debug("Using SPE %s %p %p %p %p %d\n", spu->name,
334 		spu->local_store, spu->problem, spu->priv1,
335 		spu->priv2, spu->number);
336 	goto out;
337 
338 out_unmap:
339 	spu_unmap(spu);
340 out:
341 	return ret;
342 }
343 
of_destroy_spu(struct spu * spu)344 static int of_destroy_spu(struct spu *spu)
345 {
346 	spu_unmap(spu);
347 	of_node_put(spu->devnode);
348 	return 0;
349 }
350 
enable_spu_by_master_run(struct spu_context * ctx)351 static void enable_spu_by_master_run(struct spu_context *ctx)
352 {
353 	ctx->ops->master_start(ctx);
354 }
355 
disable_spu_by_master_run(struct spu_context * ctx)356 static void disable_spu_by_master_run(struct spu_context *ctx)
357 {
358 	ctx->ops->master_stop(ctx);
359 }
360 
361 /* Hardcoded affinity idxs for qs20 */
362 #define QS20_SPES_PER_BE 8
363 static int qs20_reg_idxs[QS20_SPES_PER_BE] =   { 0, 2, 4, 6, 7, 5, 3, 1 };
364 static int qs20_reg_memory[QS20_SPES_PER_BE] = { 1, 1, 0, 0, 0, 0, 0, 0 };
365 
spu_lookup_reg(int node,u32 reg)366 static struct spu *__init spu_lookup_reg(int node, u32 reg)
367 {
368 	struct spu *spu;
369 	const u32 *spu_reg;
370 
371 	list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
372 		spu_reg = of_get_property(spu_devnode(spu), "reg", NULL);
373 		if (*spu_reg == reg)
374 			return spu;
375 	}
376 	return NULL;
377 }
378 
init_affinity_qs20_harcoded(void)379 static void __init init_affinity_qs20_harcoded(void)
380 {
381 	int node, i;
382 	struct spu *last_spu, *spu;
383 	u32 reg;
384 
385 	for (node = 0; node < MAX_NUMNODES; node++) {
386 		last_spu = NULL;
387 		for (i = 0; i < QS20_SPES_PER_BE; i++) {
388 			reg = qs20_reg_idxs[i];
389 			spu = spu_lookup_reg(node, reg);
390 			if (!spu)
391 				continue;
392 			spu->has_mem_affinity = qs20_reg_memory[reg];
393 			if (last_spu)
394 				list_add_tail(&spu->aff_list,
395 						&last_spu->aff_list);
396 			last_spu = spu;
397 		}
398 	}
399 }
400 
of_has_vicinity(void)401 static int __init of_has_vicinity(void)
402 {
403 	struct device_node *dn;
404 
405 	for_each_node_by_type(dn, "spe") {
406 		if (of_property_present(dn, "vicinity"))  {
407 			of_node_put(dn);
408 			return 1;
409 		}
410 	}
411 	return 0;
412 }
413 
devnode_spu(int cbe,struct device_node * dn)414 static struct spu *__init devnode_spu(int cbe, struct device_node *dn)
415 {
416 	struct spu *spu;
417 
418 	list_for_each_entry(spu, &cbe_spu_info[cbe].spus, cbe_list)
419 		if (spu_devnode(spu) == dn)
420 			return spu;
421 	return NULL;
422 }
423 
424 static struct spu * __init
neighbour_spu(int cbe,struct device_node * target,struct device_node * avoid)425 neighbour_spu(int cbe, struct device_node *target, struct device_node *avoid)
426 {
427 	struct spu *spu;
428 	struct device_node *spu_dn;
429 	const phandle *vic_handles;
430 	int lenp, i;
431 
432 	list_for_each_entry(spu, &cbe_spu_info[cbe].spus, cbe_list) {
433 		spu_dn = spu_devnode(spu);
434 		if (spu_dn == avoid)
435 			continue;
436 		vic_handles = of_get_property(spu_dn, "vicinity", &lenp);
437 		for (i=0; i < (lenp / sizeof(phandle)); i++) {
438 			if (vic_handles[i] == target->phandle)
439 				return spu;
440 		}
441 	}
442 	return NULL;
443 }
444 
init_affinity_node(int cbe)445 static void __init init_affinity_node(int cbe)
446 {
447 	struct spu *spu, *last_spu;
448 	struct device_node *vic_dn, *last_spu_dn;
449 	phandle avoid_ph;
450 	const phandle *vic_handles;
451 	int lenp, i, added;
452 
453 	last_spu = list_first_entry(&cbe_spu_info[cbe].spus, struct spu,
454 								cbe_list);
455 	avoid_ph = 0;
456 	for (added = 1; added < cbe_spu_info[cbe].n_spus; added++) {
457 		last_spu_dn = spu_devnode(last_spu);
458 		vic_handles = of_get_property(last_spu_dn, "vicinity", &lenp);
459 
460 		/*
461 		 * Walk through each phandle in vicinity property of the spu
462 		 * (typically two vicinity phandles per spe node)
463 		 */
464 		for (i = 0; i < (lenp / sizeof(phandle)); i++) {
465 			if (vic_handles[i] == avoid_ph)
466 				continue;
467 
468 			vic_dn = of_find_node_by_phandle(vic_handles[i]);
469 			if (!vic_dn)
470 				continue;
471 
472 			if (of_node_name_eq(vic_dn, "spe") ) {
473 				spu = devnode_spu(cbe, vic_dn);
474 				avoid_ph = last_spu_dn->phandle;
475 			} else {
476 				/*
477 				 * "mic-tm" and "bif0" nodes do not have
478 				 * vicinity property. So we need to find the
479 				 * spe which has vic_dn as neighbour, but
480 				 * skipping the one we came from (last_spu_dn)
481 				 */
482 				spu = neighbour_spu(cbe, vic_dn, last_spu_dn);
483 				if (!spu)
484 					continue;
485 				if (of_node_name_eq(vic_dn, "mic-tm")) {
486 					last_spu->has_mem_affinity = 1;
487 					spu->has_mem_affinity = 1;
488 				}
489 				avoid_ph = vic_dn->phandle;
490 			}
491 
492 			of_node_put(vic_dn);
493 
494 			list_add_tail(&spu->aff_list, &last_spu->aff_list);
495 			last_spu = spu;
496 			break;
497 		}
498 	}
499 }
500 
init_affinity_fw(void)501 static void __init init_affinity_fw(void)
502 {
503 	int cbe;
504 
505 	for (cbe = 0; cbe < MAX_NUMNODES; cbe++)
506 		init_affinity_node(cbe);
507 }
508 
init_affinity(void)509 static int __init init_affinity(void)
510 {
511 	if (of_has_vicinity()) {
512 		init_affinity_fw();
513 	} else {
514 		if (of_machine_is_compatible("IBM,CPBW-1.0"))
515 			init_affinity_qs20_harcoded();
516 		else
517 			printk("No affinity configuration found\n");
518 	}
519 
520 	return 0;
521 }
522 
523 const struct spu_management_ops spu_management_of_ops = {
524 	.enumerate_spus = of_enumerate_spus,
525 	.create_spu = of_create_spu,
526 	.destroy_spu = of_destroy_spu,
527 	.enable_spu = enable_spu_by_master_run,
528 	.disable_spu = disable_spu_by_master_run,
529 	.init_affinity = init_affinity,
530 };
531