1 /*
2 * spu management operations for of based platforms
3 *
4 * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
5 * Copyright 2006 Sony Corp.
6 * (C) Copyright 2007 TOSHIBA CORPORATION
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; version 2 of the License.
11 *
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License along
18 * with this program; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
20 */
21
22 #include <linux/interrupt.h>
23 #include <linux/list.h>
24 #include <linux/module.h>
25 #include <linux/ptrace.h>
26 #include <linux/wait.h>
27 #include <linux/mm.h>
28 #include <linux/io.h>
29 #include <linux/mutex.h>
30 #include <linux/device.h>
31
32 #include <asm/spu.h>
33 #include <asm/spu_priv1.h>
34 #include <asm/firmware.h>
35 #include <asm/prom.h>
36
37 #include "spufs/spufs.h"
38 #include "interrupt.h"
39
spu_devnode(struct spu * spu)40 struct device_node *spu_devnode(struct spu *spu)
41 {
42 return spu->devnode;
43 }
44
45 EXPORT_SYMBOL_GPL(spu_devnode);
46
find_spu_unit_number(struct device_node * spe)47 static u64 __init find_spu_unit_number(struct device_node *spe)
48 {
49 const unsigned int *prop;
50 int proplen;
51
52 /* new device trees should provide the physical-id attribute */
53 prop = of_get_property(spe, "physical-id", &proplen);
54 if (proplen == 4)
55 return (u64)*prop;
56
57 /* celleb device tree provides the unit-id */
58 prop = of_get_property(spe, "unit-id", &proplen);
59 if (proplen == 4)
60 return (u64)*prop;
61
62 /* legacy device trees provide the id in the reg attribute */
63 prop = of_get_property(spe, "reg", &proplen);
64 if (proplen == 4)
65 return (u64)*prop;
66
67 return 0;
68 }
69
spu_unmap(struct spu * spu)70 static void spu_unmap(struct spu *spu)
71 {
72 if (!firmware_has_feature(FW_FEATURE_LPAR))
73 iounmap(spu->priv1);
74 iounmap(spu->priv2);
75 iounmap(spu->problem);
76 iounmap((__force u8 __iomem *)spu->local_store);
77 }
78
spu_map_interrupts_old(struct spu * spu,struct device_node * np)79 static int __init spu_map_interrupts_old(struct spu *spu,
80 struct device_node *np)
81 {
82 unsigned int isrc;
83 const u32 *tmp;
84 int nid;
85
86 /* Get the interrupt source unit from the device-tree */
87 tmp = of_get_property(np, "isrc", NULL);
88 if (!tmp)
89 return -ENODEV;
90 isrc = tmp[0];
91
92 tmp = of_get_property(np->parent->parent, "node-id", NULL);
93 if (!tmp) {
94 printk(KERN_WARNING "%s: can't find node-id\n", __func__);
95 nid = spu->node;
96 } else
97 nid = tmp[0];
98
99 /* Add the node number */
100 isrc |= nid << IIC_IRQ_NODE_SHIFT;
101
102 /* Now map interrupts of all 3 classes */
103 spu->irqs[0] = irq_create_mapping(NULL, IIC_IRQ_CLASS_0 | isrc);
104 spu->irqs[1] = irq_create_mapping(NULL, IIC_IRQ_CLASS_1 | isrc);
105 spu->irqs[2] = irq_create_mapping(NULL, IIC_IRQ_CLASS_2 | isrc);
106
107 /* Right now, we only fail if class 2 failed */
108 return spu->irqs[2] == NO_IRQ ? -EINVAL : 0;
109 }
110
spu_map_prop_old(struct spu * spu,struct device_node * n,const char * name)111 static void __iomem * __init spu_map_prop_old(struct spu *spu,
112 struct device_node *n,
113 const char *name)
114 {
115 const struct address_prop {
116 unsigned long address;
117 unsigned int len;
118 } __attribute__((packed)) *prop;
119 int proplen;
120
121 prop = of_get_property(n, name, &proplen);
122 if (prop == NULL || proplen != sizeof (struct address_prop))
123 return NULL;
124
125 return ioremap(prop->address, prop->len);
126 }
127
spu_map_device_old(struct spu * spu)128 static int __init spu_map_device_old(struct spu *spu)
129 {
130 struct device_node *node = spu->devnode;
131 const char *prop;
132 int ret;
133
134 ret = -ENODEV;
135 spu->name = of_get_property(node, "name", NULL);
136 if (!spu->name)
137 goto out;
138
139 prop = of_get_property(node, "local-store", NULL);
140 if (!prop)
141 goto out;
142 spu->local_store_phys = *(unsigned long *)prop;
143
144 /* we use local store as ram, not io memory */
145 spu->local_store = (void __force *)
146 spu_map_prop_old(spu, node, "local-store");
147 if (!spu->local_store)
148 goto out;
149
150 prop = of_get_property(node, "problem", NULL);
151 if (!prop)
152 goto out_unmap;
153 spu->problem_phys = *(unsigned long *)prop;
154
155 spu->problem = spu_map_prop_old(spu, node, "problem");
156 if (!spu->problem)
157 goto out_unmap;
158
159 spu->priv2 = spu_map_prop_old(spu, node, "priv2");
160 if (!spu->priv2)
161 goto out_unmap;
162
163 if (!firmware_has_feature(FW_FEATURE_LPAR)) {
164 spu->priv1 = spu_map_prop_old(spu, node, "priv1");
165 if (!spu->priv1)
166 goto out_unmap;
167 }
168
169 ret = 0;
170 goto out;
171
172 out_unmap:
173 spu_unmap(spu);
174 out:
175 return ret;
176 }
177
spu_map_interrupts(struct spu * spu,struct device_node * np)178 static int __init spu_map_interrupts(struct spu *spu, struct device_node *np)
179 {
180 struct of_irq oirq;
181 int ret;
182 int i;
183
184 for (i=0; i < 3; i++) {
185 ret = of_irq_map_one(np, i, &oirq);
186 if (ret) {
187 pr_debug("spu_new: failed to get irq %d\n", i);
188 goto err;
189 }
190 ret = -EINVAL;
191 pr_debug(" irq %d no 0x%x on %s\n", i, oirq.specifier[0],
192 oirq.controller->full_name);
193 spu->irqs[i] = irq_create_of_mapping(oirq.controller,
194 oirq.specifier, oirq.size);
195 if (spu->irqs[i] == NO_IRQ) {
196 pr_debug("spu_new: failed to map it !\n");
197 goto err;
198 }
199 }
200 return 0;
201
202 err:
203 pr_debug("failed to map irq %x for spu %s\n", *oirq.specifier,
204 spu->name);
205 for (; i >= 0; i--) {
206 if (spu->irqs[i] != NO_IRQ)
207 irq_dispose_mapping(spu->irqs[i]);
208 }
209 return ret;
210 }
211
spu_map_resource(struct spu * spu,int nr,void __iomem ** virt,unsigned long * phys)212 static int spu_map_resource(struct spu *spu, int nr,
213 void __iomem** virt, unsigned long *phys)
214 {
215 struct device_node *np = spu->devnode;
216 struct resource resource = { };
217 unsigned long len;
218 int ret;
219
220 ret = of_address_to_resource(np, nr, &resource);
221 if (ret)
222 return ret;
223 if (phys)
224 *phys = resource.start;
225 len = resource.end - resource.start + 1;
226 *virt = ioremap(resource.start, len);
227 if (!*virt)
228 return -EINVAL;
229 return 0;
230 }
231
spu_map_device(struct spu * spu)232 static int __init spu_map_device(struct spu *spu)
233 {
234 struct device_node *np = spu->devnode;
235 int ret = -ENODEV;
236
237 spu->name = of_get_property(np, "name", NULL);
238 if (!spu->name)
239 goto out;
240
241 ret = spu_map_resource(spu, 0, (void __iomem**)&spu->local_store,
242 &spu->local_store_phys);
243 if (ret) {
244 pr_debug("spu_new: failed to map %s resource 0\n",
245 np->full_name);
246 goto out;
247 }
248 ret = spu_map_resource(spu, 1, (void __iomem**)&spu->problem,
249 &spu->problem_phys);
250 if (ret) {
251 pr_debug("spu_new: failed to map %s resource 1\n",
252 np->full_name);
253 goto out_unmap;
254 }
255 ret = spu_map_resource(spu, 2, (void __iomem**)&spu->priv2, NULL);
256 if (ret) {
257 pr_debug("spu_new: failed to map %s resource 2\n",
258 np->full_name);
259 goto out_unmap;
260 }
261 if (!firmware_has_feature(FW_FEATURE_LPAR))
262 ret = spu_map_resource(spu, 3,
263 (void __iomem**)&spu->priv1, NULL);
264 if (ret) {
265 pr_debug("spu_new: failed to map %s resource 3\n",
266 np->full_name);
267 goto out_unmap;
268 }
269 pr_debug("spu_new: %s maps:\n", np->full_name);
270 pr_debug(" local store : 0x%016lx -> 0x%p\n",
271 spu->local_store_phys, spu->local_store);
272 pr_debug(" problem state : 0x%016lx -> 0x%p\n",
273 spu->problem_phys, spu->problem);
274 pr_debug(" priv2 : 0x%p\n", spu->priv2);
275 pr_debug(" priv1 : 0x%p\n", spu->priv1);
276
277 return 0;
278
279 out_unmap:
280 spu_unmap(spu);
281 out:
282 pr_debug("failed to map spe %s: %d\n", spu->name, ret);
283 return ret;
284 }
285
of_enumerate_spus(int (* fn)(void * data))286 static int __init of_enumerate_spus(int (*fn)(void *data))
287 {
288 int ret;
289 struct device_node *node;
290 unsigned int n = 0;
291
292 ret = -ENODEV;
293 for (node = of_find_node_by_type(NULL, "spe");
294 node; node = of_find_node_by_type(node, "spe")) {
295 ret = fn(node);
296 if (ret) {
297 printk(KERN_WARNING "%s: Error initializing %s\n",
298 __func__, node->name);
299 break;
300 }
301 n++;
302 }
303 return ret ? ret : n;
304 }
305
of_create_spu(struct spu * spu,void * data)306 static int __init of_create_spu(struct spu *spu, void *data)
307 {
308 int ret;
309 struct device_node *spe = (struct device_node *)data;
310 static int legacy_map = 0, legacy_irq = 0;
311
312 spu->devnode = of_node_get(spe);
313 spu->spe_id = find_spu_unit_number(spe);
314
315 spu->node = of_node_to_nid(spe);
316 if (spu->node >= MAX_NUMNODES) {
317 printk(KERN_WARNING "SPE %s on node %d ignored,"
318 " node number too big\n", spe->full_name, spu->node);
319 printk(KERN_WARNING "Check if CONFIG_NUMA is enabled.\n");
320 ret = -ENODEV;
321 goto out;
322 }
323
324 ret = spu_map_device(spu);
325 if (ret) {
326 if (!legacy_map) {
327 legacy_map = 1;
328 printk(KERN_WARNING "%s: Legacy device tree found, "
329 "trying to map old style\n", __func__);
330 }
331 ret = spu_map_device_old(spu);
332 if (ret) {
333 printk(KERN_ERR "Unable to map %s\n",
334 spu->name);
335 goto out;
336 }
337 }
338
339 ret = spu_map_interrupts(spu, spe);
340 if (ret) {
341 if (!legacy_irq) {
342 legacy_irq = 1;
343 printk(KERN_WARNING "%s: Legacy device tree found, "
344 "trying old style irq\n", __func__);
345 }
346 ret = spu_map_interrupts_old(spu, spe);
347 if (ret) {
348 printk(KERN_ERR "%s: could not map interrupts\n",
349 spu->name);
350 goto out_unmap;
351 }
352 }
353
354 pr_debug("Using SPE %s %p %p %p %p %d\n", spu->name,
355 spu->local_store, spu->problem, spu->priv1,
356 spu->priv2, spu->number);
357 goto out;
358
359 out_unmap:
360 spu_unmap(spu);
361 out:
362 return ret;
363 }
364
of_destroy_spu(struct spu * spu)365 static int of_destroy_spu(struct spu *spu)
366 {
367 spu_unmap(spu);
368 of_node_put(spu->devnode);
369 return 0;
370 }
371
enable_spu_by_master_run(struct spu_context * ctx)372 static void enable_spu_by_master_run(struct spu_context *ctx)
373 {
374 ctx->ops->master_start(ctx);
375 }
376
disable_spu_by_master_run(struct spu_context * ctx)377 static void disable_spu_by_master_run(struct spu_context *ctx)
378 {
379 ctx->ops->master_stop(ctx);
380 }
381
382 /* Hardcoded affinity idxs for qs20 */
383 #define QS20_SPES_PER_BE 8
384 static int qs20_reg_idxs[QS20_SPES_PER_BE] = { 0, 2, 4, 6, 7, 5, 3, 1 };
385 static int qs20_reg_memory[QS20_SPES_PER_BE] = { 1, 1, 0, 0, 0, 0, 0, 0 };
386
spu_lookup_reg(int node,u32 reg)387 static struct spu *spu_lookup_reg(int node, u32 reg)
388 {
389 struct spu *spu;
390 const u32 *spu_reg;
391
392 list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
393 spu_reg = of_get_property(spu_devnode(spu), "reg", NULL);
394 if (*spu_reg == reg)
395 return spu;
396 }
397 return NULL;
398 }
399
init_affinity_qs20_harcoded(void)400 static void init_affinity_qs20_harcoded(void)
401 {
402 int node, i;
403 struct spu *last_spu, *spu;
404 u32 reg;
405
406 for (node = 0; node < MAX_NUMNODES; node++) {
407 last_spu = NULL;
408 for (i = 0; i < QS20_SPES_PER_BE; i++) {
409 reg = qs20_reg_idxs[i];
410 spu = spu_lookup_reg(node, reg);
411 if (!spu)
412 continue;
413 spu->has_mem_affinity = qs20_reg_memory[reg];
414 if (last_spu)
415 list_add_tail(&spu->aff_list,
416 &last_spu->aff_list);
417 last_spu = spu;
418 }
419 }
420 }
421
of_has_vicinity(void)422 static int of_has_vicinity(void)
423 {
424 struct device_node *dn;
425
426 for_each_node_by_type(dn, "spe") {
427 if (of_find_property(dn, "vicinity", NULL)) {
428 of_node_put(dn);
429 return 1;
430 }
431 }
432 return 0;
433 }
434
devnode_spu(int cbe,struct device_node * dn)435 static struct spu *devnode_spu(int cbe, struct device_node *dn)
436 {
437 struct spu *spu;
438
439 list_for_each_entry(spu, &cbe_spu_info[cbe].spus, cbe_list)
440 if (spu_devnode(spu) == dn)
441 return spu;
442 return NULL;
443 }
444
445 static struct spu *
neighbour_spu(int cbe,struct device_node * target,struct device_node * avoid)446 neighbour_spu(int cbe, struct device_node *target, struct device_node *avoid)
447 {
448 struct spu *spu;
449 struct device_node *spu_dn;
450 const phandle *vic_handles;
451 int lenp, i;
452
453 list_for_each_entry(spu, &cbe_spu_info[cbe].spus, cbe_list) {
454 spu_dn = spu_devnode(spu);
455 if (spu_dn == avoid)
456 continue;
457 vic_handles = of_get_property(spu_dn, "vicinity", &lenp);
458 for (i=0; i < (lenp / sizeof(phandle)); i++) {
459 if (vic_handles[i] == target->phandle)
460 return spu;
461 }
462 }
463 return NULL;
464 }
465
init_affinity_node(int cbe)466 static void init_affinity_node(int cbe)
467 {
468 struct spu *spu, *last_spu;
469 struct device_node *vic_dn, *last_spu_dn;
470 phandle avoid_ph;
471 const phandle *vic_handles;
472 const char *name;
473 int lenp, i, added;
474
475 last_spu = list_first_entry(&cbe_spu_info[cbe].spus, struct spu,
476 cbe_list);
477 avoid_ph = 0;
478 for (added = 1; added < cbe_spu_info[cbe].n_spus; added++) {
479 last_spu_dn = spu_devnode(last_spu);
480 vic_handles = of_get_property(last_spu_dn, "vicinity", &lenp);
481
482 /*
483 * Walk through each phandle in vicinity property of the spu
484 * (tipically two vicinity phandles per spe node)
485 */
486 for (i = 0; i < (lenp / sizeof(phandle)); i++) {
487 if (vic_handles[i] == avoid_ph)
488 continue;
489
490 vic_dn = of_find_node_by_phandle(vic_handles[i]);
491 if (!vic_dn)
492 continue;
493
494 /* a neighbour might be spe, mic-tm, or bif0 */
495 name = of_get_property(vic_dn, "name", NULL);
496 if (!name)
497 continue;
498
499 if (strcmp(name, "spe") == 0) {
500 spu = devnode_spu(cbe, vic_dn);
501 avoid_ph = last_spu_dn->phandle;
502 } else {
503 /*
504 * "mic-tm" and "bif0" nodes do not have
505 * vicinity property. So we need to find the
506 * spe which has vic_dn as neighbour, but
507 * skipping the one we came from (last_spu_dn)
508 */
509 spu = neighbour_spu(cbe, vic_dn, last_spu_dn);
510 if (!spu)
511 continue;
512 if (!strcmp(name, "mic-tm")) {
513 last_spu->has_mem_affinity = 1;
514 spu->has_mem_affinity = 1;
515 }
516 avoid_ph = vic_dn->phandle;
517 }
518
519 list_add_tail(&spu->aff_list, &last_spu->aff_list);
520 last_spu = spu;
521 break;
522 }
523 }
524 }
525
init_affinity_fw(void)526 static void init_affinity_fw(void)
527 {
528 int cbe;
529
530 for (cbe = 0; cbe < MAX_NUMNODES; cbe++)
531 init_affinity_node(cbe);
532 }
533
init_affinity(void)534 static int __init init_affinity(void)
535 {
536 if (of_has_vicinity()) {
537 init_affinity_fw();
538 } else {
539 long root = of_get_flat_dt_root();
540 if (of_flat_dt_is_compatible(root, "IBM,CPBW-1.0"))
541 init_affinity_qs20_harcoded();
542 else
543 printk("No affinity configuration found\n");
544 }
545
546 return 0;
547 }
548
549 const struct spu_management_ops spu_management_of_ops = {
550 .enumerate_spus = of_enumerate_spus,
551 .create_spu = of_create_spu,
552 .destroy_spu = of_destroy_spu,
553 .enable_spu = enable_spu_by_master_run,
554 .disable_spu = disable_spu_by_master_run,
555 .init_affinity = init_affinity,
556 };
557