xref: /linux/arch/powerpc/platforms/cell/spu_base.c (revision c0c914eca7f251c70facc37dfebeaf176601918d)
1 /*
2  * Low-level SPU handling
3  *
4  * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
5  *
6  * Author: Arnd Bergmann <arndb@de.ibm.com>
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; either version 2, or (at your option)
11  * any later version.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License
19  * along with this program; if not, write to the Free Software
20  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21  */
22 
23 #undef DEBUG
24 
25 #include <linux/interrupt.h>
26 #include <linux/list.h>
27 #include <linux/module.h>
28 #include <linux/ptrace.h>
29 #include <linux/slab.h>
30 #include <linux/wait.h>
31 #include <linux/mm.h>
32 #include <linux/io.h>
33 #include <linux/mutex.h>
34 #include <linux/linux_logo.h>
35 #include <linux/syscore_ops.h>
36 #include <asm/spu.h>
37 #include <asm/spu_priv1.h>
38 #include <asm/spu_csa.h>
39 #include <asm/xmon.h>
40 #include <asm/prom.h>
41 #include <asm/kexec.h>
42 
43 const struct spu_management_ops *spu_management_ops;
44 EXPORT_SYMBOL_GPL(spu_management_ops);
45 
46 const struct spu_priv1_ops *spu_priv1_ops;
47 EXPORT_SYMBOL_GPL(spu_priv1_ops);
48 
49 struct cbe_spu_info cbe_spu_info[MAX_NUMNODES];
50 EXPORT_SYMBOL_GPL(cbe_spu_info);
51 
52 /*
53  * The spufs fault-handling code needs to call force_sig_info to raise signals
54  * on DMA errors. Export it here to avoid general kernel-wide access to this
55  * function
56  */
57 EXPORT_SYMBOL_GPL(force_sig_info);
58 
59 /*
60  * Protects cbe_spu_info and spu->number.
61  */
62 static DEFINE_SPINLOCK(spu_lock);
63 
64 /*
65  * List of all spus in the system.
66  *
67  * This list is iterated by callers from irq context and callers that
68  * want to sleep.  Thus modifications need to be done with both
69  * spu_full_list_lock and spu_full_list_mutex held, while iterating
70  * through it requires either of these locks.
71  *
72  * In addition spu_full_list_lock protects all assignmens to
73  * spu->mm.
74  */
75 static LIST_HEAD(spu_full_list);
76 static DEFINE_SPINLOCK(spu_full_list_lock);
77 static DEFINE_MUTEX(spu_full_list_mutex);
78 
79 void spu_invalidate_slbs(struct spu *spu)
80 {
81 	struct spu_priv2 __iomem *priv2 = spu->priv2;
82 	unsigned long flags;
83 
84 	spin_lock_irqsave(&spu->register_lock, flags);
85 	if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
86 		out_be64(&priv2->slb_invalidate_all_W, 0UL);
87 	spin_unlock_irqrestore(&spu->register_lock, flags);
88 }
89 EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
90 
91 /* This is called by the MM core when a segment size is changed, to
92  * request a flush of all the SPEs using a given mm
93  */
94 void spu_flush_all_slbs(struct mm_struct *mm)
95 {
96 	struct spu *spu;
97 	unsigned long flags;
98 
99 	spin_lock_irqsave(&spu_full_list_lock, flags);
100 	list_for_each_entry(spu, &spu_full_list, full_list) {
101 		if (spu->mm == mm)
102 			spu_invalidate_slbs(spu);
103 	}
104 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
105 }
106 
107 /* The hack below stinks... try to do something better one of
108  * these days... Does it even work properly with NR_CPUS == 1 ?
109  */
110 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
111 {
112 	int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
113 
114 	/* Global TLBIE broadcast required with SPEs. */
115 	bitmap_fill(cpumask_bits(mm_cpumask(mm)), nr);
116 }
117 
118 void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
119 {
120 	unsigned long flags;
121 
122 	spin_lock_irqsave(&spu_full_list_lock, flags);
123 	spu->mm = mm;
124 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
125 	if (mm)
126 		mm_needs_global_tlbie(mm);
127 }
128 EXPORT_SYMBOL_GPL(spu_associate_mm);
129 
130 int spu_64k_pages_available(void)
131 {
132 	return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
133 }
134 EXPORT_SYMBOL_GPL(spu_64k_pages_available);
135 
136 static void spu_restart_dma(struct spu *spu)
137 {
138 	struct spu_priv2 __iomem *priv2 = spu->priv2;
139 
140 	if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
141 		out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
142 	else {
143 		set_bit(SPU_CONTEXT_FAULT_PENDING, &spu->flags);
144 		mb();
145 	}
146 }
147 
148 static inline void spu_load_slb(struct spu *spu, int slbe, struct copro_slb *slb)
149 {
150 	struct spu_priv2 __iomem *priv2 = spu->priv2;
151 
152 	pr_debug("%s: adding SLB[%d] 0x%016llx 0x%016llx\n",
153 			__func__, slbe, slb->vsid, slb->esid);
154 
155 	out_be64(&priv2->slb_index_W, slbe);
156 	/* set invalid before writing vsid */
157 	out_be64(&priv2->slb_esid_RW, 0);
158 	/* now it's safe to write the vsid */
159 	out_be64(&priv2->slb_vsid_RW, slb->vsid);
160 	/* setting the new esid makes the entry valid again */
161 	out_be64(&priv2->slb_esid_RW, slb->esid);
162 }
163 
164 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
165 {
166 	struct copro_slb slb;
167 	int ret;
168 
169 	ret = copro_calculate_slb(spu->mm, ea, &slb);
170 	if (ret)
171 		return ret;
172 
173 	spu_load_slb(spu, spu->slb_replace, &slb);
174 
175 	spu->slb_replace++;
176 	if (spu->slb_replace >= 8)
177 		spu->slb_replace = 0;
178 
179 	spu_restart_dma(spu);
180 	spu->stats.slb_flt++;
181 	return 0;
182 }
183 
184 extern int hash_page(unsigned long ea, unsigned long access,
185 		     unsigned long trap, unsigned long dsisr); //XXX
186 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
187 {
188 	int ret;
189 
190 	pr_debug("%s, %llx, %lx\n", __func__, dsisr, ea);
191 
192 	/*
193 	 * Handle kernel space hash faults immediately. User hash
194 	 * faults need to be deferred to process context.
195 	 */
196 	if ((dsisr & MFC_DSISR_PTE_NOT_FOUND) &&
197 	    (REGION_ID(ea) != USER_REGION_ID)) {
198 
199 		spin_unlock(&spu->register_lock);
200 		ret = hash_page(ea, _PAGE_PRESENT, 0x300, dsisr);
201 		spin_lock(&spu->register_lock);
202 
203 		if (!ret) {
204 			spu_restart_dma(spu);
205 			return 0;
206 		}
207 	}
208 
209 	spu->class_1_dar = ea;
210 	spu->class_1_dsisr = dsisr;
211 
212 	spu->stop_callback(spu, 1);
213 
214 	spu->class_1_dar = 0;
215 	spu->class_1_dsisr = 0;
216 
217 	return 0;
218 }
219 
220 static void __spu_kernel_slb(void *addr, struct copro_slb *slb)
221 {
222 	unsigned long ea = (unsigned long)addr;
223 	u64 llp;
224 
225 	if (REGION_ID(ea) == KERNEL_REGION_ID)
226 		llp = mmu_psize_defs[mmu_linear_psize].sllp;
227 	else
228 		llp = mmu_psize_defs[mmu_virtual_psize].sllp;
229 
230 	slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
231 		SLB_VSID_KERNEL | llp;
232 	slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
233 }
234 
235 /**
236  * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
237  * address @new_addr is present.
238  */
239 static inline int __slb_present(struct copro_slb *slbs, int nr_slbs,
240 		void *new_addr)
241 {
242 	unsigned long ea = (unsigned long)new_addr;
243 	int i;
244 
245 	for (i = 0; i < nr_slbs; i++)
246 		if (!((slbs[i].esid ^ ea) & ESID_MASK))
247 			return 1;
248 
249 	return 0;
250 }
251 
252 /**
253  * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
254  * need to map both the context save area, and the save/restore code.
255  *
256  * Because the lscsa and code may cross segment boundaires, we check to see
257  * if mappings are required for the start and end of each range. We currently
258  * assume that the mappings are smaller that one segment - if not, something
259  * is seriously wrong.
260  */
261 void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
262 		void *code, int code_size)
263 {
264 	struct copro_slb slbs[4];
265 	int i, nr_slbs = 0;
266 	/* start and end addresses of both mappings */
267 	void *addrs[] = {
268 		lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
269 		code, code + code_size - 1
270 	};
271 
272 	/* check the set of addresses, and create a new entry in the slbs array
273 	 * if there isn't already a SLB for that address */
274 	for (i = 0; i < ARRAY_SIZE(addrs); i++) {
275 		if (__slb_present(slbs, nr_slbs, addrs[i]))
276 			continue;
277 
278 		__spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
279 		nr_slbs++;
280 	}
281 
282 	spin_lock_irq(&spu->register_lock);
283 	/* Add the set of SLBs */
284 	for (i = 0; i < nr_slbs; i++)
285 		spu_load_slb(spu, i, &slbs[i]);
286 	spin_unlock_irq(&spu->register_lock);
287 }
288 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
289 
290 static irqreturn_t
291 spu_irq_class_0(int irq, void *data)
292 {
293 	struct spu *spu;
294 	unsigned long stat, mask;
295 
296 	spu = data;
297 
298 	spin_lock(&spu->register_lock);
299 	mask = spu_int_mask_get(spu, 0);
300 	stat = spu_int_stat_get(spu, 0) & mask;
301 
302 	spu->class_0_pending |= stat;
303 	spu->class_0_dar = spu_mfc_dar_get(spu);
304 	spu->stop_callback(spu, 0);
305 	spu->class_0_pending = 0;
306 	spu->class_0_dar = 0;
307 
308 	spu_int_stat_clear(spu, 0, stat);
309 	spin_unlock(&spu->register_lock);
310 
311 	return IRQ_HANDLED;
312 }
313 
314 static irqreturn_t
315 spu_irq_class_1(int irq, void *data)
316 {
317 	struct spu *spu;
318 	unsigned long stat, mask, dar, dsisr;
319 
320 	spu = data;
321 
322 	/* atomically read & clear class1 status. */
323 	spin_lock(&spu->register_lock);
324 	mask  = spu_int_mask_get(spu, 1);
325 	stat  = spu_int_stat_get(spu, 1) & mask;
326 	dar   = spu_mfc_dar_get(spu);
327 	dsisr = spu_mfc_dsisr_get(spu);
328 	if (stat & CLASS1_STORAGE_FAULT_INTR)
329 		spu_mfc_dsisr_set(spu, 0ul);
330 	spu_int_stat_clear(spu, 1, stat);
331 
332 	pr_debug("%s: %lx %lx %lx %lx\n", __func__, mask, stat,
333 			dar, dsisr);
334 
335 	if (stat & CLASS1_SEGMENT_FAULT_INTR)
336 		__spu_trap_data_seg(spu, dar);
337 
338 	if (stat & CLASS1_STORAGE_FAULT_INTR)
339 		__spu_trap_data_map(spu, dar, dsisr);
340 
341 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
342 		;
343 
344 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
345 		;
346 
347 	spu->class_1_dsisr = 0;
348 	spu->class_1_dar = 0;
349 
350 	spin_unlock(&spu->register_lock);
351 
352 	return stat ? IRQ_HANDLED : IRQ_NONE;
353 }
354 
355 static irqreturn_t
356 spu_irq_class_2(int irq, void *data)
357 {
358 	struct spu *spu;
359 	unsigned long stat;
360 	unsigned long mask;
361 	const int mailbox_intrs =
362 		CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
363 
364 	spu = data;
365 	spin_lock(&spu->register_lock);
366 	stat = spu_int_stat_get(spu, 2);
367 	mask = spu_int_mask_get(spu, 2);
368 	/* ignore interrupts we're not waiting for */
369 	stat &= mask;
370 	/* mailbox interrupts are level triggered. mask them now before
371 	 * acknowledging */
372 	if (stat & mailbox_intrs)
373 		spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
374 	/* acknowledge all interrupts before the callbacks */
375 	spu_int_stat_clear(spu, 2, stat);
376 
377 	pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
378 
379 	if (stat & CLASS2_MAILBOX_INTR)
380 		spu->ibox_callback(spu);
381 
382 	if (stat & CLASS2_SPU_STOP_INTR)
383 		spu->stop_callback(spu, 2);
384 
385 	if (stat & CLASS2_SPU_HALT_INTR)
386 		spu->stop_callback(spu, 2);
387 
388 	if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
389 		spu->mfc_callback(spu);
390 
391 	if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
392 		spu->wbox_callback(spu);
393 
394 	spu->stats.class2_intr++;
395 
396 	spin_unlock(&spu->register_lock);
397 
398 	return stat ? IRQ_HANDLED : IRQ_NONE;
399 }
400 
401 static int spu_request_irqs(struct spu *spu)
402 {
403 	int ret = 0;
404 
405 	if (spu->irqs[0] != NO_IRQ) {
406 		snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
407 			 spu->number);
408 		ret = request_irq(spu->irqs[0], spu_irq_class_0,
409 				  0, spu->irq_c0, spu);
410 		if (ret)
411 			goto bail0;
412 	}
413 	if (spu->irqs[1] != NO_IRQ) {
414 		snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
415 			 spu->number);
416 		ret = request_irq(spu->irqs[1], spu_irq_class_1,
417 				  0, spu->irq_c1, spu);
418 		if (ret)
419 			goto bail1;
420 	}
421 	if (spu->irqs[2] != NO_IRQ) {
422 		snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
423 			 spu->number);
424 		ret = request_irq(spu->irqs[2], spu_irq_class_2,
425 				  0, spu->irq_c2, spu);
426 		if (ret)
427 			goto bail2;
428 	}
429 	return 0;
430 
431 bail2:
432 	if (spu->irqs[1] != NO_IRQ)
433 		free_irq(spu->irqs[1], spu);
434 bail1:
435 	if (spu->irqs[0] != NO_IRQ)
436 		free_irq(spu->irqs[0], spu);
437 bail0:
438 	return ret;
439 }
440 
441 static void spu_free_irqs(struct spu *spu)
442 {
443 	if (spu->irqs[0] != NO_IRQ)
444 		free_irq(spu->irqs[0], spu);
445 	if (spu->irqs[1] != NO_IRQ)
446 		free_irq(spu->irqs[1], spu);
447 	if (spu->irqs[2] != NO_IRQ)
448 		free_irq(spu->irqs[2], spu);
449 }
450 
451 void spu_init_channels(struct spu *spu)
452 {
453 	static const struct {
454 		 unsigned channel;
455 		 unsigned count;
456 	} zero_list[] = {
457 		{ 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
458 		{ 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
459 	}, count_list[] = {
460 		{ 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
461 		{ 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
462 		{ 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
463 	};
464 	struct spu_priv2 __iomem *priv2;
465 	int i;
466 
467 	priv2 = spu->priv2;
468 
469 	/* initialize all channel data to zero */
470 	for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
471 		int count;
472 
473 		out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
474 		for (count = 0; count < zero_list[i].count; count++)
475 			out_be64(&priv2->spu_chnldata_RW, 0);
476 	}
477 
478 	/* initialize channel counts to meaningful values */
479 	for (i = 0; i < ARRAY_SIZE(count_list); i++) {
480 		out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
481 		out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
482 	}
483 }
484 EXPORT_SYMBOL_GPL(spu_init_channels);
485 
486 static struct bus_type spu_subsys = {
487 	.name = "spu",
488 	.dev_name = "spu",
489 };
490 
491 int spu_add_dev_attr(struct device_attribute *attr)
492 {
493 	struct spu *spu;
494 
495 	mutex_lock(&spu_full_list_mutex);
496 	list_for_each_entry(spu, &spu_full_list, full_list)
497 		device_create_file(&spu->dev, attr);
498 	mutex_unlock(&spu_full_list_mutex);
499 
500 	return 0;
501 }
502 EXPORT_SYMBOL_GPL(spu_add_dev_attr);
503 
504 int spu_add_dev_attr_group(struct attribute_group *attrs)
505 {
506 	struct spu *spu;
507 	int rc = 0;
508 
509 	mutex_lock(&spu_full_list_mutex);
510 	list_for_each_entry(spu, &spu_full_list, full_list) {
511 		rc = sysfs_create_group(&spu->dev.kobj, attrs);
512 
513 		/* we're in trouble here, but try unwinding anyway */
514 		if (rc) {
515 			printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
516 					__func__, attrs->name);
517 
518 			list_for_each_entry_continue_reverse(spu,
519 					&spu_full_list, full_list)
520 				sysfs_remove_group(&spu->dev.kobj, attrs);
521 			break;
522 		}
523 	}
524 
525 	mutex_unlock(&spu_full_list_mutex);
526 
527 	return rc;
528 }
529 EXPORT_SYMBOL_GPL(spu_add_dev_attr_group);
530 
531 
532 void spu_remove_dev_attr(struct device_attribute *attr)
533 {
534 	struct spu *spu;
535 
536 	mutex_lock(&spu_full_list_mutex);
537 	list_for_each_entry(spu, &spu_full_list, full_list)
538 		device_remove_file(&spu->dev, attr);
539 	mutex_unlock(&spu_full_list_mutex);
540 }
541 EXPORT_SYMBOL_GPL(spu_remove_dev_attr);
542 
543 void spu_remove_dev_attr_group(struct attribute_group *attrs)
544 {
545 	struct spu *spu;
546 
547 	mutex_lock(&spu_full_list_mutex);
548 	list_for_each_entry(spu, &spu_full_list, full_list)
549 		sysfs_remove_group(&spu->dev.kobj, attrs);
550 	mutex_unlock(&spu_full_list_mutex);
551 }
552 EXPORT_SYMBOL_GPL(spu_remove_dev_attr_group);
553 
554 static int spu_create_dev(struct spu *spu)
555 {
556 	int ret;
557 
558 	spu->dev.id = spu->number;
559 	spu->dev.bus = &spu_subsys;
560 	ret = device_register(&spu->dev);
561 	if (ret) {
562 		printk(KERN_ERR "Can't register SPU %d with sysfs\n",
563 				spu->number);
564 		return ret;
565 	}
566 
567 	sysfs_add_device_to_node(&spu->dev, spu->node);
568 
569 	return 0;
570 }
571 
572 static int __init create_spu(void *data)
573 {
574 	struct spu *spu;
575 	int ret;
576 	static int number;
577 	unsigned long flags;
578 
579 	ret = -ENOMEM;
580 	spu = kzalloc(sizeof (*spu), GFP_KERNEL);
581 	if (!spu)
582 		goto out;
583 
584 	spu->alloc_state = SPU_FREE;
585 
586 	spin_lock_init(&spu->register_lock);
587 	spin_lock(&spu_lock);
588 	spu->number = number++;
589 	spin_unlock(&spu_lock);
590 
591 	ret = spu_create_spu(spu, data);
592 
593 	if (ret)
594 		goto out_free;
595 
596 	spu_mfc_sdr_setup(spu);
597 	spu_mfc_sr1_set(spu, 0x33);
598 	ret = spu_request_irqs(spu);
599 	if (ret)
600 		goto out_destroy;
601 
602 	ret = spu_create_dev(spu);
603 	if (ret)
604 		goto out_free_irqs;
605 
606 	mutex_lock(&cbe_spu_info[spu->node].list_mutex);
607 	list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
608 	cbe_spu_info[spu->node].n_spus++;
609 	mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
610 
611 	mutex_lock(&spu_full_list_mutex);
612 	spin_lock_irqsave(&spu_full_list_lock, flags);
613 	list_add(&spu->full_list, &spu_full_list);
614 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
615 	mutex_unlock(&spu_full_list_mutex);
616 
617 	spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
618 	spu->stats.tstamp = ktime_get_ns();
619 
620 	INIT_LIST_HEAD(&spu->aff_list);
621 
622 	goto out;
623 
624 out_free_irqs:
625 	spu_free_irqs(spu);
626 out_destroy:
627 	spu_destroy_spu(spu);
628 out_free:
629 	kfree(spu);
630 out:
631 	return ret;
632 }
633 
634 static const char *spu_state_names[] = {
635 	"user", "system", "iowait", "idle"
636 };
637 
638 static unsigned long long spu_acct_time(struct spu *spu,
639 		enum spu_utilization_state state)
640 {
641 	unsigned long long time = spu->stats.times[state];
642 
643 	/*
644 	 * If the spu is idle or the context is stopped, utilization
645 	 * statistics are not updated.  Apply the time delta from the
646 	 * last recorded state of the spu.
647 	 */
648 	if (spu->stats.util_state == state)
649 		time += ktime_get_ns() - spu->stats.tstamp;
650 
651 	return time / NSEC_PER_MSEC;
652 }
653 
654 
655 static ssize_t spu_stat_show(struct device *dev,
656 				struct device_attribute *attr, char *buf)
657 {
658 	struct spu *spu = container_of(dev, struct spu, dev);
659 
660 	return sprintf(buf, "%s %llu %llu %llu %llu "
661 		      "%llu %llu %llu %llu %llu %llu %llu %llu\n",
662 		spu_state_names[spu->stats.util_state],
663 		spu_acct_time(spu, SPU_UTIL_USER),
664 		spu_acct_time(spu, SPU_UTIL_SYSTEM),
665 		spu_acct_time(spu, SPU_UTIL_IOWAIT),
666 		spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
667 		spu->stats.vol_ctx_switch,
668 		spu->stats.invol_ctx_switch,
669 		spu->stats.slb_flt,
670 		spu->stats.hash_flt,
671 		spu->stats.min_flt,
672 		spu->stats.maj_flt,
673 		spu->stats.class2_intr,
674 		spu->stats.libassist);
675 }
676 
677 static DEVICE_ATTR(stat, 0444, spu_stat_show, NULL);
678 
679 #ifdef CONFIG_KEXEC
680 
681 struct crash_spu_info {
682 	struct spu *spu;
683 	u32 saved_spu_runcntl_RW;
684 	u32 saved_spu_status_R;
685 	u32 saved_spu_npc_RW;
686 	u64 saved_mfc_sr1_RW;
687 	u64 saved_mfc_dar;
688 	u64 saved_mfc_dsisr;
689 };
690 
691 #define CRASH_NUM_SPUS	16	/* Enough for current hardware */
692 static struct crash_spu_info crash_spu_info[CRASH_NUM_SPUS];
693 
694 static void crash_kexec_stop_spus(void)
695 {
696 	struct spu *spu;
697 	int i;
698 	u64 tmp;
699 
700 	for (i = 0; i < CRASH_NUM_SPUS; i++) {
701 		if (!crash_spu_info[i].spu)
702 			continue;
703 
704 		spu = crash_spu_info[i].spu;
705 
706 		crash_spu_info[i].saved_spu_runcntl_RW =
707 			in_be32(&spu->problem->spu_runcntl_RW);
708 		crash_spu_info[i].saved_spu_status_R =
709 			in_be32(&spu->problem->spu_status_R);
710 		crash_spu_info[i].saved_spu_npc_RW =
711 			in_be32(&spu->problem->spu_npc_RW);
712 
713 		crash_spu_info[i].saved_mfc_dar    = spu_mfc_dar_get(spu);
714 		crash_spu_info[i].saved_mfc_dsisr  = spu_mfc_dsisr_get(spu);
715 		tmp = spu_mfc_sr1_get(spu);
716 		crash_spu_info[i].saved_mfc_sr1_RW = tmp;
717 
718 		tmp &= ~MFC_STATE1_MASTER_RUN_CONTROL_MASK;
719 		spu_mfc_sr1_set(spu, tmp);
720 
721 		__delay(200);
722 	}
723 }
724 
725 static void crash_register_spus(struct list_head *list)
726 {
727 	struct spu *spu;
728 	int ret;
729 
730 	list_for_each_entry(spu, list, full_list) {
731 		if (WARN_ON(spu->number >= CRASH_NUM_SPUS))
732 			continue;
733 
734 		crash_spu_info[spu->number].spu = spu;
735 	}
736 
737 	ret = crash_shutdown_register(&crash_kexec_stop_spus);
738 	if (ret)
739 		printk(KERN_ERR "Could not register SPU crash handler");
740 }
741 
742 #else
743 static inline void crash_register_spus(struct list_head *list)
744 {
745 }
746 #endif
747 
748 static void spu_shutdown(void)
749 {
750 	struct spu *spu;
751 
752 	mutex_lock(&spu_full_list_mutex);
753 	list_for_each_entry(spu, &spu_full_list, full_list) {
754 		spu_free_irqs(spu);
755 		spu_destroy_spu(spu);
756 	}
757 	mutex_unlock(&spu_full_list_mutex);
758 }
759 
760 static struct syscore_ops spu_syscore_ops = {
761 	.shutdown = spu_shutdown,
762 };
763 
764 static int __init init_spu_base(void)
765 {
766 	int i, ret = 0;
767 
768 	for (i = 0; i < MAX_NUMNODES; i++) {
769 		mutex_init(&cbe_spu_info[i].list_mutex);
770 		INIT_LIST_HEAD(&cbe_spu_info[i].spus);
771 	}
772 
773 	if (!spu_management_ops)
774 		goto out;
775 
776 	/* create system subsystem for spus */
777 	ret = subsys_system_register(&spu_subsys, NULL);
778 	if (ret)
779 		goto out;
780 
781 	ret = spu_enumerate_spus(create_spu);
782 
783 	if (ret < 0) {
784 		printk(KERN_WARNING "%s: Error initializing spus\n",
785 			__func__);
786 		goto out_unregister_subsys;
787 	}
788 
789 	if (ret > 0)
790 		fb_append_extra_logo(&logo_spe_clut224, ret);
791 
792 	mutex_lock(&spu_full_list_mutex);
793 	xmon_register_spus(&spu_full_list);
794 	crash_register_spus(&spu_full_list);
795 	mutex_unlock(&spu_full_list_mutex);
796 	spu_add_dev_attr(&dev_attr_stat);
797 	register_syscore_ops(&spu_syscore_ops);
798 
799 	spu_init_affinity();
800 
801 	return 0;
802 
803  out_unregister_subsys:
804 	bus_unregister(&spu_subsys);
805  out:
806 	return ret;
807 }
808 module_init(init_spu_base);
809 
810 MODULE_LICENSE("GPL");
811 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");
812