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 <asm/spu.h>
36 #include <asm/spu_priv1.h>
37 #include <asm/spu_csa.h>
38 #include <asm/xmon.h>
39 #include <asm/prom.h>
40 
41 const struct spu_management_ops *spu_management_ops;
42 EXPORT_SYMBOL_GPL(spu_management_ops);
43 
44 const struct spu_priv1_ops *spu_priv1_ops;
45 EXPORT_SYMBOL_GPL(spu_priv1_ops);
46 
47 struct cbe_spu_info cbe_spu_info[MAX_NUMNODES];
48 EXPORT_SYMBOL_GPL(cbe_spu_info);
49 
50 /*
51  * The spufs fault-handling code needs to call force_sig_info to raise signals
52  * on DMA errors. Export it here to avoid general kernel-wide access to this
53  * function
54  */
55 EXPORT_SYMBOL_GPL(force_sig_info);
56 
57 /*
58  * Protects cbe_spu_info and spu->number.
59  */
60 static DEFINE_SPINLOCK(spu_lock);
61 
62 /*
63  * List of all spus in the system.
64  *
65  * This list is iterated by callers from irq context and callers that
66  * want to sleep.  Thus modifications need to be done with both
67  * spu_full_list_lock and spu_full_list_mutex held, while iterating
68  * through it requires either of these locks.
69  *
70  * In addition spu_full_list_lock protects all assignmens to
71  * spu->mm.
72  */
73 static LIST_HEAD(spu_full_list);
74 static DEFINE_SPINLOCK(spu_full_list_lock);
75 static DEFINE_MUTEX(spu_full_list_mutex);
76 
77 struct spu_slb {
78 	u64 esid, vsid;
79 };
80 
81 void spu_invalidate_slbs(struct spu *spu)
82 {
83 	struct spu_priv2 __iomem *priv2 = spu->priv2;
84 
85 	if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
86 		out_be64(&priv2->slb_invalidate_all_W, 0UL);
87 }
88 EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
89 
90 /* This is called by the MM core when a segment size is changed, to
91  * request a flush of all the SPEs using a given mm
92  */
93 void spu_flush_all_slbs(struct mm_struct *mm)
94 {
95 	struct spu *spu;
96 	unsigned long flags;
97 
98 	spin_lock_irqsave(&spu_full_list_lock, flags);
99 	list_for_each_entry(spu, &spu_full_list, full_list) {
100 		if (spu->mm == mm)
101 			spu_invalidate_slbs(spu);
102 	}
103 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
104 }
105 
106 /* The hack below stinks... try to do something better one of
107  * these days... Does it even work properly with NR_CPUS == 1 ?
108  */
109 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
110 {
111 	int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
112 
113 	/* Global TLBIE broadcast required with SPEs. */
114 	__cpus_setall(&mm->cpu_vm_mask, nr);
115 }
116 
117 void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
118 {
119 	unsigned long flags;
120 
121 	spin_lock_irqsave(&spu_full_list_lock, flags);
122 	spu->mm = mm;
123 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
124 	if (mm)
125 		mm_needs_global_tlbie(mm);
126 }
127 EXPORT_SYMBOL_GPL(spu_associate_mm);
128 
129 int spu_64k_pages_available(void)
130 {
131 	return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
132 }
133 EXPORT_SYMBOL_GPL(spu_64k_pages_available);
134 
135 static void spu_restart_dma(struct spu *spu)
136 {
137 	struct spu_priv2 __iomem *priv2 = spu->priv2;
138 
139 	if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
140 		out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
141 }
142 
143 static inline void spu_load_slb(struct spu *spu, int slbe, struct spu_slb *slb)
144 {
145 	struct spu_priv2 __iomem *priv2 = spu->priv2;
146 
147 	pr_debug("%s: adding SLB[%d] 0x%016lx 0x%016lx\n",
148 			__func__, slbe, slb->vsid, slb->esid);
149 
150 	out_be64(&priv2->slb_index_W, slbe);
151 	out_be64(&priv2->slb_vsid_RW, slb->vsid);
152 	out_be64(&priv2->slb_esid_RW, slb->esid);
153 }
154 
155 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
156 {
157 	struct mm_struct *mm = spu->mm;
158 	struct spu_slb slb;
159 	int psize;
160 
161 	pr_debug("%s\n", __FUNCTION__);
162 
163 	if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
164 		/* SLBs are pre-loaded for context switch, so
165 		 * we should never get here!
166 		 */
167 		printk("%s: invalid access during switch!\n", __func__);
168 		return 1;
169 	}
170 	slb.esid = (ea & ESID_MASK) | SLB_ESID_V;
171 
172 	switch(REGION_ID(ea)) {
173 	case USER_REGION_ID:
174 #ifdef CONFIG_PPC_MM_SLICES
175 		psize = get_slice_psize(mm, ea);
176 #else
177 		psize = mm->context.user_psize;
178 #endif
179 		slb.vsid = (get_vsid(mm->context.id, ea, MMU_SEGSIZE_256M)
180 				<< SLB_VSID_SHIFT) | SLB_VSID_USER;
181 		break;
182 	case VMALLOC_REGION_ID:
183 		if (ea < VMALLOC_END)
184 			psize = mmu_vmalloc_psize;
185 		else
186 			psize = mmu_io_psize;
187 		slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
188 				<< SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
189 		break;
190 	case KERNEL_REGION_ID:
191 		psize = mmu_linear_psize;
192 		slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
193 				<< SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
194 		break;
195 	default:
196 		/* Future: support kernel segments so that drivers
197 		 * can use SPUs.
198 		 */
199 		pr_debug("invalid region access at %016lx\n", ea);
200 		return 1;
201 	}
202 	slb.vsid |= mmu_psize_defs[psize].sllp;
203 
204 	spu_load_slb(spu, spu->slb_replace, &slb);
205 
206 	spu->slb_replace++;
207 	if (spu->slb_replace >= 8)
208 		spu->slb_replace = 0;
209 
210 	spu_restart_dma(spu);
211 	spu->stats.slb_flt++;
212 	return 0;
213 }
214 
215 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
216 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
217 {
218 	pr_debug("%s, %lx, %lx\n", __FUNCTION__, dsisr, ea);
219 
220 	/* Handle kernel space hash faults immediately.
221 	   User hash faults need to be deferred to process context. */
222 	if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
223 	    && REGION_ID(ea) != USER_REGION_ID
224 	    && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
225 		spu_restart_dma(spu);
226 		return 0;
227 	}
228 
229 	if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
230 		printk("%s: invalid access during switch!\n", __func__);
231 		return 1;
232 	}
233 
234 	spu->class_0_pending = 0;
235 	spu->dar = ea;
236 	spu->dsisr = dsisr;
237 
238 	spu->stop_callback(spu);
239 
240 	return 0;
241 }
242 
243 static void __spu_kernel_slb(void *addr, struct spu_slb *slb)
244 {
245 	unsigned long ea = (unsigned long)addr;
246 	u64 llp;
247 
248 	if (REGION_ID(ea) == KERNEL_REGION_ID)
249 		llp = mmu_psize_defs[mmu_linear_psize].sllp;
250 	else
251 		llp = mmu_psize_defs[mmu_virtual_psize].sllp;
252 
253 	slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
254 		SLB_VSID_KERNEL | llp;
255 	slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
256 }
257 
258 /**
259  * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
260  * address @new_addr is present.
261  */
262 static inline int __slb_present(struct spu_slb *slbs, int nr_slbs,
263 		void *new_addr)
264 {
265 	unsigned long ea = (unsigned long)new_addr;
266 	int i;
267 
268 	for (i = 0; i < nr_slbs; i++)
269 		if (!((slbs[i].esid ^ ea) & ESID_MASK))
270 			return 1;
271 
272 	return 0;
273 }
274 
275 /**
276  * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
277  * need to map both the context save area, and the save/restore code.
278  *
279  * Because the lscsa and code may cross segment boundaires, we check to see
280  * if mappings are required for the start and end of each range. We currently
281  * assume that the mappings are smaller that one segment - if not, something
282  * is seriously wrong.
283  */
284 void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
285 		void *code, int code_size)
286 {
287 	struct spu_slb slbs[4];
288 	int i, nr_slbs = 0;
289 	/* start and end addresses of both mappings */
290 	void *addrs[] = {
291 		lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
292 		code, code + code_size - 1
293 	};
294 
295 	/* check the set of addresses, and create a new entry in the slbs array
296 	 * if there isn't already a SLB for that address */
297 	for (i = 0; i < ARRAY_SIZE(addrs); i++) {
298 		if (__slb_present(slbs, nr_slbs, addrs[i]))
299 			continue;
300 
301 		__spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
302 		nr_slbs++;
303 	}
304 
305 	/* Add the set of SLBs */
306 	for (i = 0; i < nr_slbs; i++)
307 		spu_load_slb(spu, i, &slbs[i]);
308 }
309 EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
310 
311 static irqreturn_t
312 spu_irq_class_0(int irq, void *data)
313 {
314 	struct spu *spu;
315 	unsigned long stat, mask;
316 
317 	spu = data;
318 
319 	spin_lock(&spu->register_lock);
320 	mask = spu_int_mask_get(spu, 0);
321 	stat = spu_int_stat_get(spu, 0) & mask;
322 
323 	spu->class_0_pending |= stat;
324 	spu->dsisr = spu_mfc_dsisr_get(spu);
325 	spu->dar = spu_mfc_dar_get(spu);
326 	spin_unlock(&spu->register_lock);
327 
328 	spu->stop_callback(spu);
329 
330 	spu_int_stat_clear(spu, 0, stat);
331 
332 	return IRQ_HANDLED;
333 }
334 
335 static irqreturn_t
336 spu_irq_class_1(int irq, void *data)
337 {
338 	struct spu *spu;
339 	unsigned long stat, mask, dar, dsisr;
340 
341 	spu = data;
342 
343 	/* atomically read & clear class1 status. */
344 	spin_lock(&spu->register_lock);
345 	mask  = spu_int_mask_get(spu, 1);
346 	stat  = spu_int_stat_get(spu, 1) & mask;
347 	dar   = spu_mfc_dar_get(spu);
348 	dsisr = spu_mfc_dsisr_get(spu);
349 	if (stat & CLASS1_STORAGE_FAULT_INTR)
350 		spu_mfc_dsisr_set(spu, 0ul);
351 	spu_int_stat_clear(spu, 1, stat);
352 	spin_unlock(&spu->register_lock);
353 	pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__, mask, stat,
354 			dar, dsisr);
355 
356 	if (stat & CLASS1_SEGMENT_FAULT_INTR)
357 		__spu_trap_data_seg(spu, dar);
358 
359 	if (stat & CLASS1_STORAGE_FAULT_INTR)
360 		__spu_trap_data_map(spu, dar, dsisr);
361 
362 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
363 		;
364 
365 	if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
366 		;
367 
368 	return stat ? IRQ_HANDLED : IRQ_NONE;
369 }
370 
371 static irqreturn_t
372 spu_irq_class_2(int irq, void *data)
373 {
374 	struct spu *spu;
375 	unsigned long stat;
376 	unsigned long mask;
377 	const int mailbox_intrs =
378 		CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
379 
380 	spu = data;
381 	spin_lock(&spu->register_lock);
382 	stat = spu_int_stat_get(spu, 2);
383 	mask = spu_int_mask_get(spu, 2);
384 	/* ignore interrupts we're not waiting for */
385 	stat &= mask;
386 
387 	/* mailbox interrupts are level triggered. mask them now before
388 	 * acknowledging */
389 	if (stat & mailbox_intrs)
390 		spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
391 	/* acknowledge all interrupts before the callbacks */
392 	spu_int_stat_clear(spu, 2, stat);
393 	spin_unlock(&spu->register_lock);
394 
395 	pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
396 
397 	if (stat & CLASS2_MAILBOX_INTR)
398 		spu->ibox_callback(spu);
399 
400 	if (stat & CLASS2_SPU_STOP_INTR)
401 		spu->stop_callback(spu);
402 
403 	if (stat & CLASS2_SPU_HALT_INTR)
404 		spu->stop_callback(spu);
405 
406 	if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
407 		spu->mfc_callback(spu);
408 
409 	if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
410 		spu->wbox_callback(spu);
411 
412 	spu->stats.class2_intr++;
413 	return stat ? IRQ_HANDLED : IRQ_NONE;
414 }
415 
416 static int spu_request_irqs(struct spu *spu)
417 {
418 	int ret = 0;
419 
420 	if (spu->irqs[0] != NO_IRQ) {
421 		snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
422 			 spu->number);
423 		ret = request_irq(spu->irqs[0], spu_irq_class_0,
424 				  IRQF_DISABLED,
425 				  spu->irq_c0, spu);
426 		if (ret)
427 			goto bail0;
428 	}
429 	if (spu->irqs[1] != NO_IRQ) {
430 		snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
431 			 spu->number);
432 		ret = request_irq(spu->irqs[1], spu_irq_class_1,
433 				  IRQF_DISABLED,
434 				  spu->irq_c1, spu);
435 		if (ret)
436 			goto bail1;
437 	}
438 	if (spu->irqs[2] != NO_IRQ) {
439 		snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
440 			 spu->number);
441 		ret = request_irq(spu->irqs[2], spu_irq_class_2,
442 				  IRQF_DISABLED,
443 				  spu->irq_c2, spu);
444 		if (ret)
445 			goto bail2;
446 	}
447 	return 0;
448 
449 bail2:
450 	if (spu->irqs[1] != NO_IRQ)
451 		free_irq(spu->irqs[1], spu);
452 bail1:
453 	if (spu->irqs[0] != NO_IRQ)
454 		free_irq(spu->irqs[0], spu);
455 bail0:
456 	return ret;
457 }
458 
459 static void spu_free_irqs(struct spu *spu)
460 {
461 	if (spu->irqs[0] != NO_IRQ)
462 		free_irq(spu->irqs[0], spu);
463 	if (spu->irqs[1] != NO_IRQ)
464 		free_irq(spu->irqs[1], spu);
465 	if (spu->irqs[2] != NO_IRQ)
466 		free_irq(spu->irqs[2], spu);
467 }
468 
469 void spu_init_channels(struct spu *spu)
470 {
471 	static const struct {
472 		 unsigned channel;
473 		 unsigned count;
474 	} zero_list[] = {
475 		{ 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
476 		{ 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
477 	}, count_list[] = {
478 		{ 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
479 		{ 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
480 		{ 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
481 	};
482 	struct spu_priv2 __iomem *priv2;
483 	int i;
484 
485 	priv2 = spu->priv2;
486 
487 	/* initialize all channel data to zero */
488 	for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
489 		int count;
490 
491 		out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
492 		for (count = 0; count < zero_list[i].count; count++)
493 			out_be64(&priv2->spu_chnldata_RW, 0);
494 	}
495 
496 	/* initialize channel counts to meaningful values */
497 	for (i = 0; i < ARRAY_SIZE(count_list); i++) {
498 		out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
499 		out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
500 	}
501 }
502 EXPORT_SYMBOL_GPL(spu_init_channels);
503 
504 static int spu_shutdown(struct sys_device *sysdev)
505 {
506 	struct spu *spu = container_of(sysdev, struct spu, sysdev);
507 
508 	spu_free_irqs(spu);
509 	spu_destroy_spu(spu);
510 	return 0;
511 }
512 
513 static struct sysdev_class spu_sysdev_class = {
514 	.name = "spu",
515 	.shutdown = spu_shutdown,
516 };
517 
518 int spu_add_sysdev_attr(struct sysdev_attribute *attr)
519 {
520 	struct spu *spu;
521 
522 	mutex_lock(&spu_full_list_mutex);
523 	list_for_each_entry(spu, &spu_full_list, full_list)
524 		sysdev_create_file(&spu->sysdev, attr);
525 	mutex_unlock(&spu_full_list_mutex);
526 
527 	return 0;
528 }
529 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr);
530 
531 int spu_add_sysdev_attr_group(struct attribute_group *attrs)
532 {
533 	struct spu *spu;
534 	int rc = 0;
535 
536 	mutex_lock(&spu_full_list_mutex);
537 	list_for_each_entry(spu, &spu_full_list, full_list) {
538 		rc = sysfs_create_group(&spu->sysdev.kobj, attrs);
539 
540 		/* we're in trouble here, but try unwinding anyway */
541 		if (rc) {
542 			printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
543 					__func__, attrs->name);
544 
545 			list_for_each_entry_continue_reverse(spu,
546 					&spu_full_list, full_list)
547 				sysfs_remove_group(&spu->sysdev.kobj, attrs);
548 			break;
549 		}
550 	}
551 
552 	mutex_unlock(&spu_full_list_mutex);
553 
554 	return rc;
555 }
556 EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group);
557 
558 
559 void spu_remove_sysdev_attr(struct sysdev_attribute *attr)
560 {
561 	struct spu *spu;
562 
563 	mutex_lock(&spu_full_list_mutex);
564 	list_for_each_entry(spu, &spu_full_list, full_list)
565 		sysdev_remove_file(&spu->sysdev, attr);
566 	mutex_unlock(&spu_full_list_mutex);
567 }
568 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr);
569 
570 void spu_remove_sysdev_attr_group(struct attribute_group *attrs)
571 {
572 	struct spu *spu;
573 
574 	mutex_lock(&spu_full_list_mutex);
575 	list_for_each_entry(spu, &spu_full_list, full_list)
576 		sysfs_remove_group(&spu->sysdev.kobj, attrs);
577 	mutex_unlock(&spu_full_list_mutex);
578 }
579 EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group);
580 
581 static int spu_create_sysdev(struct spu *spu)
582 {
583 	int ret;
584 
585 	spu->sysdev.id = spu->number;
586 	spu->sysdev.cls = &spu_sysdev_class;
587 	ret = sysdev_register(&spu->sysdev);
588 	if (ret) {
589 		printk(KERN_ERR "Can't register SPU %d with sysfs\n",
590 				spu->number);
591 		return ret;
592 	}
593 
594 	sysfs_add_device_to_node(&spu->sysdev, spu->node);
595 
596 	return 0;
597 }
598 
599 static int __init create_spu(void *data)
600 {
601 	struct spu *spu;
602 	int ret;
603 	static int number;
604 	unsigned long flags;
605 	struct timespec ts;
606 
607 	ret = -ENOMEM;
608 	spu = kzalloc(sizeof (*spu), GFP_KERNEL);
609 	if (!spu)
610 		goto out;
611 
612 	spu->alloc_state = SPU_FREE;
613 
614 	spin_lock_init(&spu->register_lock);
615 	spin_lock(&spu_lock);
616 	spu->number = number++;
617 	spin_unlock(&spu_lock);
618 
619 	ret = spu_create_spu(spu, data);
620 
621 	if (ret)
622 		goto out_free;
623 
624 	spu_mfc_sdr_setup(spu);
625 	spu_mfc_sr1_set(spu, 0x33);
626 	ret = spu_request_irqs(spu);
627 	if (ret)
628 		goto out_destroy;
629 
630 	ret = spu_create_sysdev(spu);
631 	if (ret)
632 		goto out_free_irqs;
633 
634 	mutex_lock(&cbe_spu_info[spu->node].list_mutex);
635 	list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
636 	cbe_spu_info[spu->node].n_spus++;
637 	mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
638 
639 	mutex_lock(&spu_full_list_mutex);
640 	spin_lock_irqsave(&spu_full_list_lock, flags);
641 	list_add(&spu->full_list, &spu_full_list);
642 	spin_unlock_irqrestore(&spu_full_list_lock, flags);
643 	mutex_unlock(&spu_full_list_mutex);
644 
645 	spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
646 	ktime_get_ts(&ts);
647 	spu->stats.tstamp = timespec_to_ns(&ts);
648 
649 	INIT_LIST_HEAD(&spu->aff_list);
650 
651 	goto out;
652 
653 out_free_irqs:
654 	spu_free_irqs(spu);
655 out_destroy:
656 	spu_destroy_spu(spu);
657 out_free:
658 	kfree(spu);
659 out:
660 	return ret;
661 }
662 
663 static const char *spu_state_names[] = {
664 	"user", "system", "iowait", "idle"
665 };
666 
667 static unsigned long long spu_acct_time(struct spu *spu,
668 		enum spu_utilization_state state)
669 {
670 	struct timespec ts;
671 	unsigned long long time = spu->stats.times[state];
672 
673 	/*
674 	 * If the spu is idle or the context is stopped, utilization
675 	 * statistics are not updated.  Apply the time delta from the
676 	 * last recorded state of the spu.
677 	 */
678 	if (spu->stats.util_state == state) {
679 		ktime_get_ts(&ts);
680 		time += timespec_to_ns(&ts) - spu->stats.tstamp;
681 	}
682 
683 	return time / NSEC_PER_MSEC;
684 }
685 
686 
687 static ssize_t spu_stat_show(struct sys_device *sysdev, char *buf)
688 {
689 	struct spu *spu = container_of(sysdev, struct spu, sysdev);
690 
691 	return sprintf(buf, "%s %llu %llu %llu %llu "
692 		      "%llu %llu %llu %llu %llu %llu %llu %llu\n",
693 		spu_state_names[spu->stats.util_state],
694 		spu_acct_time(spu, SPU_UTIL_USER),
695 		spu_acct_time(spu, SPU_UTIL_SYSTEM),
696 		spu_acct_time(spu, SPU_UTIL_IOWAIT),
697 		spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
698 		spu->stats.vol_ctx_switch,
699 		spu->stats.invol_ctx_switch,
700 		spu->stats.slb_flt,
701 		spu->stats.hash_flt,
702 		spu->stats.min_flt,
703 		spu->stats.maj_flt,
704 		spu->stats.class2_intr,
705 		spu->stats.libassist);
706 }
707 
708 static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
709 
710 static int __init init_spu_base(void)
711 {
712 	int i, ret = 0;
713 
714 	for (i = 0; i < MAX_NUMNODES; i++) {
715 		mutex_init(&cbe_spu_info[i].list_mutex);
716 		INIT_LIST_HEAD(&cbe_spu_info[i].spus);
717 	}
718 
719 	if (!spu_management_ops)
720 		goto out;
721 
722 	/* create sysdev class for spus */
723 	ret = sysdev_class_register(&spu_sysdev_class);
724 	if (ret)
725 		goto out;
726 
727 	ret = spu_enumerate_spus(create_spu);
728 
729 	if (ret < 0) {
730 		printk(KERN_WARNING "%s: Error initializing spus\n",
731 			__FUNCTION__);
732 		goto out_unregister_sysdev_class;
733 	}
734 
735 	if (ret > 0) {
736 		/*
737 		 * We cannot put the forward declaration in
738 		 * <linux/linux_logo.h> because of conflicting session type
739 		 * conflicts for const and __initdata with different compiler
740 		 * versions
741 		 */
742 		extern const struct linux_logo logo_spe_clut224;
743 
744 		fb_append_extra_logo(&logo_spe_clut224, ret);
745 	}
746 
747 	mutex_lock(&spu_full_list_mutex);
748 	xmon_register_spus(&spu_full_list);
749 	crash_register_spus(&spu_full_list);
750 	mutex_unlock(&spu_full_list_mutex);
751 	spu_add_sysdev_attr(&attr_stat);
752 
753 	spu_init_affinity();
754 
755 	return 0;
756 
757  out_unregister_sysdev_class:
758 	sysdev_class_unregister(&spu_sysdev_class);
759  out:
760 	return ret;
761 }
762 module_init(init_spu_base);
763 
764 MODULE_LICENSE("GPL");
765 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");
766